Eve library

V1 Domain Workbenches — SOTA Build-Out Master Checklist

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

  • Opened: 2026-07-23
  • Last scope-and-coverage audit: 2026-07-23
  • Last execution-granularity audit: 2026-07-23
  • Audit pass: third pass; bidirectional feature coverage plus manual semantic execution decomposition against current code, contracts, domain inventories, deep dives, walkthroughs, and specialist plans
  • Status: proposed execution ledger; no unchecked item is implied complete

This is the implementation ledger for seven related authoring and operations initiatives: Isis, Metis, Yemaya, Veritas, Euterpe, Aja, and Bellona. It expands the original outline into independently verifiable work units and corrects its greenfield assumption. The repository already contains substantial domain libraries, BFF routes, standalone applications, and tests. The goal is to integrate, harden, and complete those systems without duplicating them or violating the product-surface boundaries already ratified elsewhere.

The second audit explicitly expanded capability families that the first deep ledger left implicit: Isis production asset/control matrices; Metis credentials, lecture/media, discipline, and correctness systems; Yemaya's complete existing AAA suite; Veritas investigative, channel, regulatory, community, B2B, USSD, commercial-boundary, and multimedia operations; Euterpe's structural DAW, plugin/interchange, AI-audio, spatial, stage, and DJ workflows; Aja live and multimodal capture plus vertical/CG/generative pipelines; and Bellona render, gameplay, virtual-production, XR, mocap, MetaHuman, specialist DCC, remote-agent, package, and event operations. This third pass reviews every parent outcome and manually decomposes compound work into domain-specific, independently provable child tasks wherever that creates a more executable ledger. Section 1.2 makes atomic, bidirectional feature disposition—not the presence of a broad phase—the permanent coverage invariant.

Tara remains the quality exemplar:

Tara is a quality floor, not a UI template to clone blindly. Every initiative must meet its domain's professional workflow, safety model, interoperability standards, latency profile, and evidence burden.


1. Authoritative scope correction#

The previous revision said that all seven domains had rich backends but zero interactive authoring UI. That statement is not supportable against the worktree. The 2026-07-23 audit found mature UI code in apps/isis/web, apps/yemaya/studio-web, apps/yemaya/studio-desktop, apps/euterpe/studio-web, apps/bellona/control-room, apps/metis/admin, apps/metis/web, and apps/veritas/web. The real gap is fragmented coverage, uncertain live wiring, duplicated route trees, incomplete cross-surface workflows, and inconsistent production proof.

The canonical boundary in docs/oshun/studio-vs-yemaya-studio.md also forbids rebuilding raw ComfyUI, model, LoRA, GPU, voice-cloning, full 3D, or music-stem controls inside contemplative Oshun Studio. Those tools belong to Yemaya Studio or operator applications.

A source-mounted page is not proof that a surface is authorized or reachable. For example, /studio/isis/nvidia-rtx-processing has an interactive page module but remains AAA-only under the proxy because STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS is empty. The baseline must therefore test effective middleware behavior as well as count files and routes.

Initiative Product classification Canonical host to ratify in P0 Existing implementation to integrate Boundary that must not regress
Isis V1 governed-generation substrate plus operator factory apps/isis/web for operators; curated cards/gallery in Oshun Studio; AAA controls in Yemaya Studio apps/isis/web, 70+ libs/isis/*, V1 Isis adapter/BFF No raw provider/model/workflow/GPU controls under /studio/isis/*
Metis V1 customer domain plus institutional authoring Oshun Studio workbench backed by Metis services; operator review remains /operator/metis apps/metis/{admin,web}, libs/metis/*, Metis contracts/BFF Do not conflate learner delivery, institutional admin, and authoring roles
Yemaya AAA creative-production suite and V1 rendering substrate apps/yemaya/studio-{web,desktop} existing Studio apps, study workspace, case pipeline, asset/production systems Do not present the AAA suite as a new V1 customer domain tab
Veritas V1 customer domain plus newsroom/editorial authoring Oshun Studio newsroom integrated with Veritas CMS/services apps/veritas/{web,cms}, libs/veritas/*, canonical V1 contracts Reader surfaces and protected newsroom surfaces remain separate
Euterpe Broader-platform music product, not yet ratified as an Oshun V1 domain apps/euterpe/studio-web existing DAW, audio engine, mastering, rights, distribution systems No automatic V1 shell registration without an explicit product decision
Aja Cross-domain motion subsystem and operator pipeline Host decision between an Aja control room and a Studio operator route Aja services/libs plus existing static Studio route inventory Metis reaches Aja through the canonical adapter, not private internals
Bellona Studio-only V1 engine/build substrate apps/bellona/control-room with a governed Oshun Studio entry/facade existing control room, bridge hosts, build services, Studio routes No customer-domain route; no duplicate control room implementation

1.1 Baseline evidence to refresh before implementation#

These values are discovery evidence only. They are not readiness claims.

Domain Top-level library directories Test files under libs/<domain> Existing application evidence
Isis 77 1,206 apps/isis/web: 69 TSX files and Playwright coverage
Metis 27 365 apps/metis/admin and apps/metis/web
Yemaya 61 1,118 studio-web: 90 TSX; studio-desktop: 12 TSX
Veritas 65 98 apps/veritas/web: 27 TSX; CMS and service apps
Euterpe 52 336 apps/euterpe/studio-web: 115 TSX and 200+ tests
Aja 42 65 service/CLI apps; static Oshun Studio route inventory
Bellona 41 493 apps/bellona/control-room: 10 TSX and Playwright
  • BASE.1 Regenerate the package, route, application, and test inventory from the current branch. Evidence: a committed machine-readable snapshot and the exact command used to produce it. <!-- evidence: evidence/v1-workbenches/inventory/ (manifest/packages/applications/routes/tests/reconciliation.json + README.md); generator scripts/v1-workbenches/generate-inventory.mjs; command pnpm inventory:v1-workbenches; verify pnpm verify:inventory:v1-workbenches (drift-check + 8 known-value tests, all green at SHA 050e5d67). 473 packages, 48 apps, 329 routes, 4812 tests, 0 unexplained records. -->

    • BASE.1.a Pin the audited commit SHA, workspace roots, ignored paths, and tool versions before scanning. Evidence: inventory manifest records all four inputs. <!-- evidence: manifest.json inputs.{auditedCommitSha, workspaceRoots, ignoredPathGlobs, toolVersions} -->
    • BASE.1.b Enumerate workspace packages and top-level domain libraries from manifests and project configuration. Evidence: package rows retain manifest path, project name, domain, and target metadata. <!-- evidence: packages.json — 473 rows with manifests[], projectName, domain, targets[], exports[], language (npm/Nx/Python/C++ union so no project is hidden) -->
    • BASE.1.c Enumerate deployable applications, services, workers, CLIs, desktop shells, and bridge hosts. Evidence: application rows identify runtime type, entry point, and owning project. <!-- evidence: applications.json — 48 rows with runtimeType (web-frontend/service/worker/cli/desktop-shell/engine-bridge-host/gpu-worker/remote-agent-host/api-gateway/mobile-app/agent-runtime), entryPoint, projectName, root -->
    • BASE.1.d Enumerate filesystem, framework, proxy, redirect, and middleware-derived routes independently. Evidence: route rows retain route source and effective host. <!-- evidence: routes.json — framework-aware: Next.js filesystem routes (app-router-page/app-router-handler/pages-router-page/pages-api-route/middleware) are gated on an ancestor next.config so Vite apps' src/pages are not miscounted; React Router <Route path=> framework routes captured as source react-router; config-redirect/config-rewrite from next.config; test/util files excluded. 125 routes (was 329 with Vite/test false positives). manifest.proxyGovernance records the external Studio proxy (apps/oshun/web/src/proxy.ts) + verified fail-closed STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS=[]. -->
    • BASE.1.e Enumerate unit, component, contract, integration, browser, mobile, desktop, accessibility, performance, security, and evaluation test files. Evidence: test rows identify harness and owning target. <!-- evidence: tests.json — 4812 rows with harness (vitest/playwright/maestro/node:test), category (unit/component/contract/integration/accessibility/performance/security/evaluation/browser-or-device), owningTarget -->
    • BASE.1.f Reconcile generated records with workspace configuration so aliases neither duplicate nor hide a package, application, route, or test. Evidence: reconciliation report has zero unexplained records. <!-- evidence: reconciliation.json unexplainedRecords=0; surfaced 11 cross-domain name leaks (stale lilith->aja) + non-npm projects as owner-routed findings; benign scope-strip conventions classified -->
    • BASE.1.g Commit the normalized snapshot together with its deterministic command and input SHA. Evidence: a clean regeneration produces no diff. <!-- evidence: pnpm verify:inventory:v1-workbenches → "inventory: no drift."; no timestamps in output; git ls-files driven; determinism test green -->
  • BASE.2 Classify every existing surface as live, fixture-backed, documentation-only, dead, redirected, or provider-gated. Evidence: every discovered route has exactly one classification.

    • BASE.2.a Define mutually exclusive classification criteria, including how mixed live/fixture and partially wired surfaces are represented. Evidence: reviewed classification schema and examples. <!-- evidence: evidence/v1-workbenches/surface-classification-schema.md — 6 mutually-exclusive primary classes with a precedence decision order (one class per surface), sub-state fields (wiringCompleteness/fixtureLeakage/providerStates/roleMatrix/dataStates/evidenceRefs) for mixed & partial surfaces, and repo-grounded positive + negative examples (Studio-Isis fail-closed provider-gating, comfy-cloud provider_not_configured, metis NextResponse.redirect, veritas reading-list live). Runtime application is BASE.2.b-g. -->
    • BASE.2.b Resolve each route through its real host, middleware, feature flags, and authorization policy rather than inferring reachability from a source file. Evidence: effective-route results for every route row.
    • BASE.2.c Trace each reachable surface to its actual data source and record whether fixtures, hard-coded values, or fabricated adapters can enter the result. Evidence: route-to-data-source links.
    • BASE.2.d Exercise provider-gated surfaces with provider configuration present and absent. Evidence: both configured and not_configured results are captured.
    • BASE.2.e Verify redirects at the final destination and distinguish intentional compatibility redirects from abandoned route shells. Evidence: redirect-chain and destination-status records.
    • BASE.2.f Assign an owner and remediation or retirement action to every dead, ambiguous, or mixed-classification surface. Evidence: zero ownerless exception rows.
    • BASE.2.g Independently review the completed classification against the discovered route inventory. Evidence: signed review with zero unclassified routes.
  • BASE.3 Map every top-level library to reuse, refactor, retire, or explicitly out-of-scope. Evidence: no library directory is absent from the signed inventory.

    • BASE.3.a Record every library's manifest, exported entry points, public symbols, and current domain ownership. Evidence: source-linked library rows. <!-- evidence: evidence/v1-workbenches/inventory/library-graph.json (per-package manifest/root/domain/exports/publicSymbolCount) + library-public-symbols.json (96,094 symbols across 412/415 libs; barrel-following through .js/.d.ts/dist entry indirection). 3 zero-symbol libs surfaced (2 Unity C# plugins + @oshun/documentation broken entry). command pnpm inventory:v1-workbenches. -->
    • BASE.3.b Identify all application, service, worker, CLI, test, and library consumers for each library. Evidence: inbound-consumer map. <!-- evidence: library-graph.json packages[].inboundConsumers (distinct consumer projects, whole-repo git-grep scan so consumers outside the 7 domains are counted; multi-line } from '@scope/x' handled; string-literal mentions excluded). Guarded by generate-library-graph.test.mjs. -->
    • BASE.3.c Identify each library's runtime and build-time dependencies, including cross-domain and private-internal imports. Evidence: outbound-dependency map. <!-- evidence: library-graph.json packages[].outboundDependencies with crossDomain + deep (private-internal, past-public-export) flags per edge; outboundCrossDomainCount/outboundDeepImportCount rollups. -->
    • BASE.3.d Assign exactly one reuse, refactor, retire, or out-of-scope disposition with rationale and owner. Evidence: zero ambiguous or placeholder dispositions. <!-- partial: library-graph.json packages[].proposedDisposition + dispositionRationale provide a machine recommendation (258 reuse / 191 retire-candidate / 12 refactor-candidate); OWNER assignment is a human decision and is not fabricated, so this stays unchecked. -->
    • BASE.3.e For refactor rows, name the target boundary, migration order, compatibility period, and removal condition. Evidence: executable migration references.
    • BASE.3.f For retire or out-of-scope rows, identify every affected consumer and the replacement, compatibility, or removal path. Evidence: consumer-complete impact records.
    • BASE.3.g Reconcile the signed mapping back to the filesystem. Evidence: a path-level check reports neither missing nor extra libraries.
  • BASE.4 Record route ownership and canonical host decisions before adding a route. Evidence: approved decision records resolve every row above.

    • BASE.4.a Identify the audience, trust boundary, data classification, latency profile, and operator/customer role for every candidate route. Evidence: completed route-context fields.
    • BASE.4.b Enumerate all plausible existing hosts and the reuse, duplication, navigation, and deployment consequences of each. Evidence: host-option matrix.
    • BASE.4.c Select one canonical host and owning team for every retained route. Evidence: approved route-to-host and route-to-owner mappings.
    • BASE.4.d Decide whether each noncanonical route is removed, redirected, proxied, or retained for compatibility. Evidence: migration behavior and end date per noncanonical route.
    • BASE.4.e Record cross-host authentication, authorization, deep-link, and data-handoff requirements. Evidence: approved boundary contract.
    • BASE.4.f Obtain product, architecture, security, and domain-owner approval where the decision changes an existing product boundary. Evidence: named approval records.
  • BASE.5 Reconcile this ledger with domains.json, V1 architecture, DOMAINS.md, and the Studio/Yemaya boundary. Evidence: the domain registry checker passes and no documentation contradiction remains.

    • BASE.5.a Extract domain classification, host, audience, ownership, and boundary assertions from each authoritative source. Evidence: source-and-heading-indexed assertion table. <!-- evidence: registry-reconciliation.json authoritativeAssertions — per-initiative table indexed by source + heading provenance across domains.json (status/npmScope/owner/projects), ledger §1 (classification + boundary), and DOMAINS.md (tagline + line). e.g. Euterpe: DOMAINS.md:178 "Music & Sonic Creation" vs domains.json active vs ledger "not yet ratified". -->
    • BASE.5.b Compare domains.json identifiers and lifecycle fields with the ledger's seven initiatives. Evidence: field-level registry diff. <!-- evidence: evidence/v1-workbenches/registry-reconciliation.json + scripts/v1-workbenches/generate-registry-reconciliation.mjs — per-initiative field checks (npmScope/libsPath/appsPath/projects vs inventory reality) and 3 surfaced contradictions for BASE.5.f: euterpe ratification (domains.json status=active vs ledger "not yet ratified as an Oshun V1 domain"), bellona + yemaya project-count drift. Deterministic + 5 tests. -->
    • BASE.5.c Compare V1 architecture host and dependency rules with every route and integration proposed here. Evidence: architecture-rule diff.
    • BASE.5.d Compare DOMAINS.md product classifications and ownership with the ledger. Evidence: classification-and-owner diff.
    • BASE.5.e Compare every Studio and Yemaya exposure with the ratified Studio/Yemaya boundary. Evidence: route-and-control boundary matrix. <!-- evidence: evidence/v1-workbenches/studio-boundary-matrix.json + scripts/v1-workbenches/generate-studio-boundary-matrix.mjs — verifies the 31 AAA-only segments in docs/oshun/studio-vs-yemaya-studio.md agree EXACTLY with AAA_ONLY_STUDIO_ROUTES in libs/isis/entitlements/src/studio-boundary.ts (docCodeAgree, 0 diff), STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS is empty (failClosed), and every AAA control (ComfyUI/LoRA/GPU/voice/3D/music) is blocked on the contemplative product. --check fails on any boundary regression. 5 tests. -->
    • BASE.5.f Resolve each contradiction in its authoritative source rather than documenting competing truths. Evidence: zero unresolved diff rows.
    • BASE.5.g Run the registry checker and validate every changed internal document link. Evidence: passing command outputs. <!-- evidence: node tools/domains/check-registry.mjs → "✓ domains.json in sync (58 domains, 63 scope tags)" (exit 0). Every relative link + referenced file path in the changed docs (evidence/v1-workbenches/README.md, inventory/README.md, and the ledger's markdown links) resolves to a real path (link-check clean, 0 broken). Note: this validates link/registry integrity; the Euterpe classification contradiction is an inter-source contradiction for BASE.5.f (human), not a broken link. -->
  • BASE.6 Capture a browser walkthrough of every existing candidate host. Evidence: dated screenshots, network traces, console logs, and a route coverage matrix are linked from the inventory.

    • BASE.6.a Define the supported browser, viewport, locale, tenant, role, feature-flag, provider, and data-state matrix for the walkthrough. Evidence: reproducible walkthrough manifest.
    • BASE.6.b Provision nonfixture accounts and deterministic live-data records for every authorized role. Evidence: fixture-independent setup record with secret-safe identifiers.
    • BASE.6.c Visit every inventory route through normal navigation and by direct deep link. Evidence: navigation and deep-link result per route.
    • BASE.6.d Capture the initial, loaded, empty, partial, unauthorized, unavailable, and error states that the route can produce. Evidence: state-indexed screenshots or approved nonapplicability.
    • BASE.6.e Archive request/response metadata, redirects, console errors, failed resources, and runtime warnings for each route. Evidence: timestamped trace and console bundle.
    • BASE.6.f Record broken navigation, unreachable source-mounted pages, fixture leakage, stale content, and host-boundary violations as owned findings. Evidence: finding IDs linked to route rows.
    • BASE.6.g Reconcile captured routes with the candidate-host inventory. Evidence: matrix reports zero unvisited reachable routes.
  • BASE.7 Probe each current API from its real UI with fixtures disabled. Evidence: the matrix distinguishes live data, empty data, 503 not_configured, and accidental mock data.

    • BASE.7.a Enumerate every API request initiated by each current UI and map it to the initiating user action. Evidence: action-to-request map.
    • BASE.7.b Disable documented fixture, demo, mock-service, interception, and development-fallback switches. Evidence: captured runtime configuration with all fixture paths off.
    • BASE.7.c Exercise an authorized live-data response and verify displayed values against persisted or provider source data. Evidence: source/UI value comparison.
    • BASE.7.d Exercise a legitimate empty result and verify the UI does not substitute sample content. Evidence: empty-state trace and screenshot.
    • BASE.7.e Remove required provider configuration and verify the API and UI preserve explicit not_configured semantics. Evidence: 503 body, trace, and visible state.
    • BASE.7.f Exercise authentication, authorization, and tenant-isolation failures for each protected API family. Evidence: negative matrix with no data disclosure.
    • BASE.7.g Detect responses originating from hard-coded adapters, request interception, seeded demos, or fallback fixtures. Evidence: zero unexplained mock signatures.
    • BASE.7.h Link every API finding to the owning route, service, and remediation task. Evidence: zero ownerless findings.
  • BASE.8 Establish the baseline accessibility, performance, reliability, and security results for every host. Evidence: raw reports are archived before feature work begins.

    • BASE.8.a Pin the host build, environment, accounts, datasets, browsers, devices, clocks, and network profiles used by all baseline runs. Evidence: baseline environment manifest.
    • BASE.8.b Run automated accessibility scans over every reachable route and state in the walkthrough matrix. Evidence: raw route-level results.
    • BASE.8.c Run keyboard and screen-reader smoke journeys for each host's critical workflow. Evidence: journey-level findings and recordings.
    • BASE.8.d Measure navigation, interaction, rendering, memory, and transfer budgets under the supported device/network profiles. Evidence: raw performance traces and budget comparison.
    • BASE.8.e Measure API availability, latency, error rate, queue age, retry, and recovery behavior for exercised dependencies. Evidence: reliability baseline report.
    • BASE.8.f Run dependency, secret, configuration, header, transport, authorization, upload, and remote-fetch security checks as applicable. Evidence: raw security reports.
    • BASE.8.g Deduplicate findings without discarding route/state provenance, then assign severity and owner. Evidence: triaged baseline register.
    • BASE.8.h Archive immutable raw reports before remediation begins. Evidence: checksummed evidence index with capture timestamps.

1.2 Atomic feature-completeness rule#

The audit unit is an atomic capability, not a package, route, page, or broad phase label. A capability is in scope for reconciliation when it is asserted or implemented by any current application route, component, library, public contract, API/worker/CLI command, schema or migration, event, test, feature inventory, domain deep dive, walkthrough, proposal, or release specification. Every discovered capability must have one explicit disposition:

  1. exposed in the named workbench for an authorized audience;
  2. exposed in a separate operator, AAA, consumer, institutional, or developer surface with a governed workbench handoff where appropriate;
  3. runtime-only but observable and controllable through an authorized operational projection;
  4. retained for compatibility with no new exposure;
  5. retired through a dated migration; or
  6. excluded by a named product/safety decision with rationale, owner, compensating control, and review date.

Broad tasks such as "map every library" do not by themselves prove feature coverage. The machine-readable inventory must split multi-feature packages and pages into atomic rows and map each row to its contract, owner, audience, host, route or entry point, backing API/job/store, permissions, safety and quality gates, failure states, tests, documentation, and disposition evidence.

  • BASE.9 Generate an atomic capability inventory from every source class named above and retain source file plus symbol/route/heading provenance. Evidence: deterministic committed inventory and generator command.
    • BASE.9.a Define the canonical capability-row schema, identifier rules, and the boundary between one atomic behavior and a compound feature. Evidence: reviewed schema with positive and negative examples. <!-- evidence: evidence/v1-workbenches/capability-inventory.manifest.json — rowSchema (id cap:<kind>:<domain>:<slug>, fields, sourceFields, kinds) + atomicityBoundary with the one-behavior-per-row rule and positive/negative examples (single route/operation-symbol/entity/CLI-verb = atomic; package/dual-purpose page/getter/factory = split or excluded). -->
    • BASE.9.b Extract capability candidates from application routes and interactive components. Evidence: source path, route, export, and interaction provenance per candidate. <!-- evidence: capability-inventory.json kind=surface (110), one row per route from BASE.1 routes.json with source.{file,ref=routePath,provenance=route:<source>,host}. -->
    • BASE.9.c Extract capability candidates from library public exports and externally consumed internal symbols. Evidence: symbol-level provenance. <!-- evidence: capability-inventory.json kind=operation (3222) from BASE.3 library-public-symbols.json with a tight action-verb filter (excludes getters/factories/pure helpers); source.{file=pkg,ref=symbol,verb}. Deep externally-consumed-internal imports are captured at package granularity in library-graph.json (outboundDeepImportCount). -->
    • BASE.9.d Extract capability candidates from public contracts, API operations, workers, jobs, events, and CLI commands. Evidence: operation-level provenance. <!-- evidence: capability-inventory.json — 654 kind=api-operation (Express/Fastify METHOD /path handlers, provenance=http-route), 2227 kind=event (emit/publish/dispatch string event names, provenance=event-emit), 37 kind=command (CLI, provenance=cli-command), 2 kind=job (worker/queue, provenance=worker-queue). Contract operations are represented by their concrete HTTP api-operation rows. -->
    • BASE.9.e Extract capability candidates from schemas, migrations, persisted entities, state machines, and policy definitions. Evidence: entity/transition/rule provenance. <!-- partial: 162 persisted entities (Prisma models + SQL tables, deduped) extracted with provenance; state-machine transitions and policy-rule definitions are NOT yet extracted, so this stays unchecked. -->
    • BASE.9.f Extract capability candidates from tests, feature inventories, deep dives, walkthroughs, proposals, and release specifications. Evidence: test or document heading provenance.
    • BASE.9.g Split candidates that combine distinct audiences, mutations, lifecycle transitions, policies, failure modes, or outputs. Evidence: one-behavior-per-row review results.
    • BASE.9.h Preserve every original source reference when normalizing candidates into canonical rows. Evidence: zero source records without a canonical row. <!-- evidence: every capability-inventory.json row carries source.{file,ref,provenance}; the generate-capability-inventory.test.mjs schema test asserts no row lacks provenance or ref. -->
    • BASE.9.i Commit the deterministic inventory and exact extraction command. Evidence: rerunning from the pinned SHA produces no diff. <!-- evidence: node scripts/v1-workbenches/generate-capability-inventory.mjs --check → "capability-inventory: no drift."; sorted output, no timestamps; command wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches. --> <!-- BASE.9.f (documented from feature-inventories/deep-dives/proposals beyond tests) and BASE.9.g (active compound-candidate splitting review) remain unchecked; the parent BASE.9 stays unchecked until d/e/f/g complete. -->
  • BASE.10 Reconcile aliases and duplicates without collapsing distinct audience, lifecycle, policy, or data behaviors. Evidence: every source capability maps to exactly one canonical capability ID and all aliases remain searchable.
    • BASE.10.a Normalize spelling, casing, route aliases, deprecated names, and equivalent public/internal terminology. Evidence: searchable alias candidates with source provenance.
    • BASE.10.b Group only candidates whose actors, inputs, outputs, mutations, policies, lifecycle, and failure semantics are equivalent. Evidence: equivalence criteria recorded for every merge.
    • BASE.10.c Keep separate rows when audience, lifecycle, policy, data classification, side effects, or recovery behavior differs. Evidence: reviewed false-merge challenge set.
    • BASE.10.d Select a stable canonical identifier and preferred label for every accepted group. Evidence: canonical-ID assignment table.
    • BASE.10.e Retain bidirectional links from every alias and duplicate source to its canonical row. Evidence: alias lookup tests.
    • BASE.10.f Resolve ambiguous candidate groups with the named domain owner. Evidence: zero unresolved merge decisions.
    • BASE.10.g Check that every source candidate maps to exactly one canonical ID and no canonical row is source-empty unless labeled net-new. Evidence: passing cardinality report.
  • BASE.11 Assign every canonical capability one of the six dispositions above. Evidence: zero missing, ambiguous, or generic "covered elsewhere" dispositions.
    • BASE.11.a Record the authorized audience and canonical surface for every capability proposed for direct workbench exposure. Evidence: audience-and-surface fields.
    • BASE.11.b Record the separate surface and governed handoff for every operator, AAA, consumer, institutional, or developer exposure. Evidence: handoff contract or approved no-handoff rationale.
    • BASE.11.c Record the observability and control projection for every runtime-only capability. Evidence: operational projection references.
    • BASE.11.d Record compatibility owner, supported consumers, and exit condition for every compatibility-only capability. Evidence: compatibility register.
    • BASE.11.e Record migration, consumer impact, removal date, and rollback for every retired capability. Evidence: dated retirement plan.
    • BASE.11.f Record approver, rationale, compensating control, and review date for every excluded capability. Evidence: approved exclusion record.
    • BASE.11.g Reject placeholder dispositions such as "covered elsewhere" until the exact capability and destination are linked. Evidence: lint reports zero generic dispositions.
    • BASE.11.h Obtain domain-owner review of all assignments and resolve conflicts between source intent and product scope. Evidence: signed disposition review.
  • BASE.12 For every exposed or operationally projected capability, record create/read/update/delete or invoke/cancel/retry semantics, bulk behavior, import/export, search/discovery, history/diff, collaboration/review, accessibility, localization, offline/degraded behavior, observability, recovery, and release/retirement applicability. Evidence: applicability matrix contains an implementation/test link or approved N/A record in every cell.
    • BASE.12.a Record create, read, update, delete, and lifecycle-transition applicability per capability. Evidence: mutation/read cells contain implementation and test links or reviewed N/A records.
    • BASE.12.b Record invoke, cancel, retry, resume, timeout, and compensation applicability for command and job capabilities. Evidence: command-state cells are complete.
    • BASE.12.c Record single-item and bulk eligibility, limits, partial success, and rollback behavior. Evidence: bulk-operation cells are complete.
    • BASE.12.d Record import, export, interchange version, validation, round-trip, and unsupported-format behavior. Evidence: interchange cells are complete.
    • BASE.12.e Record search, filter, sort, pagination, saved view, catalog, and discovery behavior. Evidence: discovery cells are complete.
    • BASE.12.f Record immutable revision, history, diff, restore, fork, and provenance behavior. Evidence: revision cells are complete.
    • BASE.12.g Record collaboration, comment, assignment, review, approval, override, and notification behavior. Evidence: governance cells are complete.
    • BASE.12.h Record keyboard, screen-reader, nonvisual alternative, responsive, locale, time-zone, and bidirectional-text behavior. Evidence: accessibility/localization cells are complete.
    • BASE.12.i Record offline, reconnect, partial, stale, conflict, degraded, unavailable, and recovery behavior. Evidence: resilience-state cells are complete.
    • BASE.12.j Record audit, trace, metric, log, cost, freshness, support, and incident-diagnostic behavior. Evidence: operability cells are complete.
    • BASE.12.k Record release, migration, rollback, compatibility, deprecation, retirement, retention, and deletion behavior. Evidence: lifecycle-release cells are complete.
    • BASE.12.l Validate that every matrix cell has a precise implementation and test reference or an approved child-specific N/A record. Evidence: zero blank, generic, or unapproved cells.
  • BASE.13 Diff the atomic capability inventory in CI. Evidence: a new route, exported symbol, contract operation, job, schema entity, CLI command, documented feature, or walkthrough fails the gate until dispositioned.
    • BASE.13.a Serialize the canonical inventory in deterministic source and identifier order. Evidence: repeated generation is byte-stable. <!-- evidence: capability-inventory.json is stable-stringified (sorted keys) and ID-sorted; generate-capability-inventory.mjs --check → "no drift". The gate compares by semantic id with key-canonicalized source. -->
    • BASE.13.b Detect additions, removals, and semantic changes in route and interactive-component candidates. Evidence: route/component mutation tests. <!-- evidence: scripts/v1-workbenches/capability-diff-gate.mjs + capability-diff-gate.test.mjs — injecting a net-new surface capability is detected as ADDED; removing/moving one is detected as REMOVED/CHANGED. -->
    • BASE.13.c Detect additions, removals, and semantic changes in exported symbols and packages. Evidence: library mutation tests. <!-- evidence: gate test injects a net-new operation capability (library export) and asserts it fails the gate. -->
    • BASE.13.d Detect additions, removals, and semantic changes in contracts, API operations, jobs, events, and CLI commands. Evidence: interface mutation tests. <!-- evidence: the capability diff gate now covers api-operation, event, command, and job kinds; the gate test injects a net-new capability of every kind (incl. api-operation and event) and asserts each fails the gate. -->
    • BASE.13.e Detect additions, removals, and semantic changes in schemas, migrations, and persisted entities. Evidence: data-model mutation tests. <!-- evidence: gate test injects a net-new entity capability (persisted schema/table) and asserts it fails the gate. -->
    • BASE.13.f Detect additions, removals, and semantic changes in documented features and walkthrough findings. Evidence: documentation mutation tests. <!-- not covered: the gate operates on the core capability set; documented (test/doc heading) rows and walkthrough findings are not gated yet. -->
    • BASE.13.g Fail with the exact undispositioned source and remediation command rather than an opaque snapshot mismatch. Evidence: reviewed failure output. <!-- evidence: the gate prints ADDED/REMOVED/CHANGED capability IDs grouped by kind plus the remediation command (pnpm inventory:v1-workbenches) and exits 1, instead of a byte-diff. -->
    • BASE.13.h Prove that an explicitly dispositioned change passes while unrelated formatting changes do not create false capability churn. Evidence: positive and noise-control CI tests. <!-- evidence: gate test asserts committed baseline == freshly built passes (dispositioned), and reordering source object keys registers NO change (ids are semantic kind:domain:behavior, path-free). --> <!-- BASE.13.d/f and the parent BASE.13 stay unchecked (events/contracts + documented sources not yet gated). -->
  • BASE.14 Run a bidirectional audit: every inventory row maps to a ledger requirement and every domain ledger requirement maps back to one or more authoritative inventory rows or a clearly labeled net-new requirement. Evidence: zero orphan source capabilities and zero unexplained checklist capabilities.
    • BASE.14.a Link every canonical inventory row to its exact parent or child ledger identifier. Evidence: zero inventory rows without ledger links.
    • BASE.14.b Link every ledger capability requirement to one or more canonical inventory identifiers. Evidence: zero ledger requirements without inventory or net-new links.
    • BASE.14.c Label net-new requirements explicitly and record their product, architecture, and domain authority. Evidence: approved net-new register.
    • BASE.14.d Detect many-to-one mappings that conceal distinct audience, policy, lifecycle, or failure behaviors. Evidence: reviewed mapping cardinality exceptions.
    • BASE.14.e Detect inventory capabilities linked only to broad phase labels instead of executable requirements. Evidence: zero phase-only links.
    • BASE.14.f Resolve orphan source capabilities to exposure, separate surface, runtime projection, compatibility, retirement, or exclusion. Evidence: zero source-side orphans.
    • BASE.14.g Resolve unexplained checklist capabilities to authoritative source evidence or approved net-new status. Evidence: zero ledger-side orphans.
    • BASE.14.h Publish the final forward and reverse coverage reports. Evidence: checksummed reports generated from the same inventory version.
  • BASE.15 Require named domain-owner review of the final atomic inventory, including consumer/runtime-only exclusions. Evidence: signed review records for all seven domains; an engineering agent cannot self-approve product scope.
    • BASE.15.a Prepare one domain-specific review packet containing source coverage, canonical capabilities, aliases, dispositions, exclusions, mappings, and unresolved findings. Evidence: seven versioned packets.
    • BASE.15.b Obtain Isis owner review and resolve every Isis comment. Evidence: named Isis approval and closed comment log.
    • BASE.15.c Obtain Metis owner review and resolve every Metis comment. Evidence: named Metis approval and closed comment log.
    • BASE.15.d Obtain Yemaya owner review and resolve every Yemaya comment. Evidence: named Yemaya approval and closed comment log.
    • BASE.15.e Obtain Veritas owner review and resolve every Veritas comment. Evidence: named Veritas approval and closed comment log.
    • BASE.15.f Obtain Euterpe owner review and resolve every Euterpe comment. Evidence: named Euterpe approval and closed comment log.
    • BASE.15.g Obtain Aja owner review and resolve every Aja comment. Evidence: named Aja approval and closed comment log.
    • BASE.15.h Obtain Bellona owner review and resolve every Bellona comment. Evidence: named Bellona approval and closed comment log.
    • BASE.15.i Re-run both coverage directions after all owner-requested changes. Evidence: final zero-orphan reports at the approved revision.

2. Execution contract#

2.1 Checkbox and evidence rules#

  • [ ] and [x] are the sole task-completion source of truth.

  • Each checkbox represents one independently reviewable outcome.

  • A checkbox may flip only when the evidence named by the item exists.

  • A phase heading is organizational; it is never a completion signal.

  • An item marked not applicable must remain unchecked until a decision record identifies the approver, rationale, replacement control, and review date.

  • Existing code is evidence only after it has been inspected and its relevant verification has passed on the current branch.

  • A unit test does not prove a browser journey; a fixture does not prove live integration; a screenshot does not prove authorization; a narrow build does not prove a release gate. <!-- AND A SPECIFICATION DOES NOT PROVE A PRODUCT — measured 2026-08-16. evidence/v1-workbenches/v1-reference-implementation-reach.json — generate-v1-reference-implementation-reach.mjs + .test.mjs (8 tests, 2 ratchets, 3 controls), wired into inventory + verify.

    A large share of this checklist's closed cells cite a "COMPILING reference" under evidence/v1-workbenches/<name>/*.ts. There are 88 such directories, 692 typed source files and 8,715 exported symbols. Every one carries its own strict tsconfig, most carry a check.ts, and the verify lane runs them, so they compile and their tests pass. NOT ONE LINE UNDER libs/ OR apps/ IMPORTS A MODULE FROM ANY OF THEM. The mentions that exist are JSON artifact paths and prose in comments, which is why the scan matches an IMPORT SPECIFIER and not a path string.

    693 CHECKLIST CELLS CITE ONE OF THESE REFERENCES AND 683 OF THEM ARE CHECKED. That is not automatically wrong. A cell whose evidence clause asks for a schema, a contract, a taxonomy or a row is legitimately satisfied by an executable specification — that is what a specification is, and these are unusually good ones. A cell whose clause asks for TESTS, fixtures, a probe, a gate or a report is asking for something EXECUTED, and then what matters is what it was executed against: the answer is the specification, because nothing imports it.

    CLASSIFIED BY EACH CELL'S OWN EVIDENCE SENTENCE, executed evidence winning when a clause names both: 588 behavioural, 51 specifiable, 54 unclassified — and 582 CHECKED cells whose evidence asks for something executed cite a reference no product code imports. A first cut of the classifier matched only named phrases like "integration tests" and left 530 of the 693 clauses unclassified, which would have made this split meaningless; most cells say "lifecycle tests", "eligibility tests", "race tests".

    This artifact draws a distinction the checklist itself does not, because a cell records only that its evidence exists. -->

  • Human governance approvals remain unchecked until performed by the named human. An agent may prepare an engineering-readiness packet but may not fabricate approval.

  • A top-level identifier such as I3.4 remains the accountable outcome. When that outcome contains separately buildable behavior, sequential gates, independently owned work, multiple artifacts, or multiple proof obligations, nested .a, .b, and later identifiers define its executable child tasks.

  • Child tasks must be written from the parent's domain semantics. Do not add generic lifecycle boilerplate merely to increase the checkbox count, and do not derive work mechanically from punctuation or keywords.

  • A parent may flip to [x] only after every applicable nested checkbox is complete and linked from the parent's rollup evidence. The parent must remain unchecked while any child is unchecked, failed, stale, or awaiting approval.

  • A nested task may be marked not applicable only through the same reviewed N/A record required above. A broad N/A decision may cover multiple children only when it names each child identifier and gives a child-specific rationale.

  • Child identifiers are stable evidence keys and follow execution order where ordering exists. Renaming or resequencing a child requires updating every evidence and dependency reference to it.

  • An already atomic parent needs no ceremonial children. During implementation, newly discovered semantic work must become a new child or parent when it has independent behavior, ownership, sequencing, failure handling, or approval.

  • Automation may identify review candidates and validate checklist structure, identifiers, evidence clauses, and parent/child completion consistency. It must not author or regenerate the semantic task text.

2.2 Quality and honesty rules#

  • Every scorer, gate, editor, dispatcher, and pipeline stage ships real domain logic with known-value tests or fails loudly behind an explicit not_configured or unavailable result.
  • Generated, inferred, scored, or processed values must never be invented in UI adapters.
  • Provider errors, partial results, stale reads, degraded dependencies, retries, and cancellations must be visible and machine-readable.
  • Fixtures must carry a persistent fixture badge and must be impossible to enable accidentally in production.
  • All mutations are authenticated, authorized at object and property level, audited, idempotent where retryable, and durable before acknowledgement.
  • All publish, release, export, and promotion paths are fail-closed.
  • All high-risk AI operations preserve input, prompt/template version, model identity, parameters, safety results, output hashes, cost, actor, and timestamps.
  • Destructive operations expose impact previews, require typed confirmation or step-up authorization proportionate to risk, and provide a documented recovery path.

2.3 UI and interaction thesis#

The shared visual thesis is a calm, professional production surface: strong typography, dense but readable information, restrained color, and one dominant working canvas. The content plan is workspace first, context/inspector second, status and evidence always visible, with no marketing hero inside an operator tool. The interaction thesis is:

  1. fast workspace-to-inspector transitions that preserve selection and focus;
  2. visible, non-ornamental state transitions for jobs, reviews, and publishes;
  3. keyboard-first command, comparison, and triage flows with reduced-motion equivalents.
  • UX.1 Define a shared utility-first information architecture: global navigation, primary workspace, contextual inspector, and status region. Evidence: approved responsive wireframes for both archetypes.
    • UX.1.a Define the global navigation hierarchy, route labels, current location treatment, and cross-workbench exits. Evidence: annotated navigation map.
    • UX.1.b Define the primary workspace's canvas, editor, table, graph, timeline, or queue responsibilities for each workbench archetype. Evidence: archetype-by-region responsibility matrix.
    • UX.1.c Define contextual inspector opening, pinning, replacement, dismissal, and focus-return behavior. Evidence: annotated inspector interaction flow.
    • UX.1.d Define the persistent status region for save, sync, job, review, gate, publish, fixture, degraded, and connectivity state. Evidence: status-content specification.
    • UX.1.e Define how selection and working context survive transitions between navigation, workspace, inspector, and status regions. Evidence: state-preservation interaction cases.
    • UX.1.f Produce wide and compact wireframes for editor-centric and operations-centric archetypes. Evidence: four approved responsive wireframes.
  • UX.2 Define density modes without card mosaics or decorative dashboards. Evidence: the design review confirms cards exist only when the card is the interaction.
    • UX.2.a Define comfortable and compact spacing, row-height, control-size, and information-priority tokens. Evidence: density token table.
    • UX.2.b Identify the default density for editors, tables, queues, inspectors, timelines, and touch-oriented review. Evidence: surface-density matrix.
    • UX.2.c Define which information may collapse, abbreviate, wrap, or move to an inspector at each density. Evidence: content-priority rules.
    • UX.2.d Define when a card is itself a selectable, draggable, comparable, or previewable object and prohibit decorative container cards elsewhere. Evidence: reviewed card-use inventory.
    • UX.2.e Preserve the user's density preference per workbench and provide an accessible reset. Evidence: preference contract and interaction test.
    • UX.2.f Validate both densities at 200% zoom and with long translated labels. Evidence: no-clipping visual test matrix.
  • UX.3 Define one accent color and semantic status tokens that remain distinguishable without color. Evidence: token contrast tests pass.
    • UX.3.a Select the accent token and specify permitted emphasis, selection, focus, and action uses. Evidence: approved accent-use rules.
    • UX.3.b Define independent semantic tokens for informational, success, warning, failure, destructive, stale, fixture, degraded, and pending states. Evidence: token inventory.
    • UX.3.c Pair every semantic state with text, iconography, pattern, shape, or position so color is never the sole signal. Evidence: noncolor cue matrix.
    • UX.3.d Define foreground, background, border, focus, chart, and disabled token pairs for every supported theme. Evidence: theme token mappings.
    • UX.3.e Test text, nontext controls, focus indicators, and adjacent status colors against required contrast ratios. Evidence: automated contrast report with zero violations.
  • UX.4 Define at most two type families and a type scale optimized for long-form editing, tables, graphs, code, and metadata. Evidence: rendered specimen passes at 200% zoom.
    • UX.4.a Choose the primary interface/editor family and any secondary monospace or specialist family with licensed font sources and fallbacks. Evidence: typography decision record.
    • UX.4.b Define display, heading, body, label, metadata, table, code, and numeric styles with size, weight, line height, and tracking. Evidence: typed token scale.
    • UX.4.c Define tabular-number, code, long-form prose, dense metadata, and graph-label treatments. Evidence: content-type specimen.
    • UX.4.d Validate fallbacks, missing glyphs, diacritics, bidirectional text, vertical-script metadata, and long translations. Evidence: multilingual glyph specimen.
    • UX.4.e Validate reflow, clipping, truncation disclosure, and reading order at 200% and 400% text zoom. Evidence: zoom test report.
  • UX.5 Define focus, selection, hover, active, disabled, stale, fixture, degraded, conflict, and destructive states. Evidence: Storybook or equivalent state fixtures cover each state.
    • UX.5.a Define keyboard focus appearance, focus-visible behavior, focus order, and focus restoration. Evidence: focus-state fixtures and tests.
    • UX.5.b Define single, multiple, range, remote-user, and unavailable selection appearance. Evidence: selection-state fixtures.
    • UX.5.c Define hover behavior that adds no hover-only information or action. Evidence: pointer/no-pointer comparison fixtures.
    • UX.5.d Define pressed, active-route, active-tool, and in-progress action treatment. Evidence: active-state fixtures.
    • UX.5.e Define disabled versus read-only semantics, explanations, and keyboard exposure. Evidence: disabled/read-only fixtures and semantic assertions.
    • UX.5.f Define stale-data indicators, freshness details, and refresh or recompute actions. Evidence: stale-state fixtures.
    • UX.5.g Define persistent fixture labeling and prevent fixture state from visually resembling live production data. Evidence: fixture-state fixtures and production guard test.
    • UX.5.h Define degraded and unavailable states with affected capability, retained capability, cause, and recovery action. Evidence: degradation-state fixtures.
    • UX.5.i Define local/remote edit conflicts, comparison, resolution, and unresolved-state treatment. Evidence: conflict-state fixtures.
    • UX.5.j Define destructive action hierarchy, impact preview, typed or step-up confirmation, completion, and recovery treatment. Evidence: destructive-state fixtures.
  • UX.6 Define three responsive arrangements: wide three-region workspace, compact two-region workspace, and single-region mobile/tablet review. Evidence: no horizontal page overflow at supported viewports.
    • UX.6.a Define wide-layout minimum dimensions and the allocation, resizing, and collapse rules for navigation, workspace, and inspector. Evidence: wide-layout specification.
    • UX.6.b Define compact-layout region pairing, inspector overlay/drawer behavior, and selection persistence. Evidence: compact-layout specification.
    • UX.6.c Define single-region review navigation, context breadcrumbs, inspector replacement, and back behavior. Evidence: single-region flow.
    • UX.6.d Define content-driven transition thresholds without assuming device labels or orientation. Evidence: breakpoint rationale and tests.
    • UX.6.e Define minimum and maximum pane sizes plus keyboard-accessible resizing and reset. Evidence: pane-resize interaction tests.
    • UX.6.f Test every arrangement with zoom, translated strings, persistent scrollbars, and software keyboards. Evidence: zero page-level overflow or obscured-action failures.
  • UX.7 Define motion durations, easing, interruption, and prefers-reduced-motion behavior. Evidence: automated reduced-motion assertions cover all shared transitions.
    • UX.7.a Define duration tiers for immediate feedback, region transition, route transition, progress, and explanatory motion. Evidence: duration token table.
    • UX.7.b Define easing and direction rules that communicate spatial or lifecycle relationships without decorative movement. Evidence: approved motion examples.
    • UX.7.c Define interruption, reversal, cancellation, rapid-repeat, and focus behavior during animation. Evidence: interruption test cases.
    • UX.7.d Define reduced-motion replacements that preserve state change and orientation without translation, scale, parallax, or flashing. Evidence: transition-by-transition alternative matrix.
    • UX.7.e Enforce flashing, autoplay, and vestibular-safety limits. Evidence: automated motion-safety assertions.
    • UX.7.f Test every shared transition with reduced motion enabled before route load and toggled at runtime. Evidence: passing browser automation.
  • UX.8 Define keyboard maps for navigation, selection, compare, command palette, save, undo, redo, comment, and escape. Evidence: collision audit covers browser, OS, screen-reader, and editor shortcuts.
    • UX.8.a Define landmark, region, tab, panel, and focus-mode navigation commands. Evidence: navigation key map and journey test.
    • UX.8.b Define single, range, additive, parent/child, and clear-selection commands. Evidence: selection key map and tests.
    • UX.8.c Define compare entry, side switching, synchronized navigation, difference traversal, and compare exit. Evidence: compare key map.
    • UX.8.d Define command-palette open, search, group navigation, execution, disabled-command explanation, and dismissal. Evidence: palette key map.
    • UX.8.e Define save, save-as-version, undo, redo, and history behavior across text, structured edits, and remote conflicts. Evidence: editing-command key map.
    • UX.8.f Define comment creation, anchored navigation, submit, cancel, and resolve shortcuts. Evidence: review-command key map.
    • UX.8.g Define layered Escape behavior for composition, menus, dialogs, modes, selections, inspectors, and routes. Evidence: ordered Escape contract.
    • UX.8.h Audit every proposed chord against supported browsers, operating systems, assistive technologies, international layouts, and embedded editors. Evidence: collision matrix with resolved conflicts.
    • UX.8.i Publish discoverable shortcut help and allow remapping where platform or accessibility conflicts remain. Evidence: help and preference interaction tests.
  • UX.9 Define touch and pen alternatives for drag, resize, graph pan, and timeline trim. Evidence: every drag operation has a single-pointer, keyboard, and non-drag alternative.
    • UX.9.a Inventory every drag, reorder, resize, pan, scrub, trim, connect, lasso, and freehand interaction by surface. Evidence: complete gesture inventory.
    • UX.9.b Define tap/select plus explicit move controls for reordering and drag placement. Evidence: non-drag reorder flows.
    • UX.9.c Define numeric, stepper, preset, and reset controls for pane, object, and region resizing. Evidence: non-drag resize flows.
    • UX.9.d Define directional, minimap, search, and fit controls for graph, canvas, waveform, and timeline panning. Evidence: non-gesture navigation flows.
    • UX.9.e Define timecode/numeric in-out fields and frame/sample stepping for timeline and media trimming. Evidence: non-drag trim flow.
    • UX.9.f Define touch/pen hit targets, palm rejection, pressure/tilt applicability, cancellation, and accidental-gesture recovery. Evidence: input-modality specification.
    • UX.9.g Test every inventoried gesture with touch, pen, keyboard, switch, and a single pointer without path-based movement. Evidence: complete modality matrix.
  • UX.10 Define plain-language utility copy for loading, empty, offline, unauthorized, not configured, failed, partial, stale, and retrying states. Evidence: copy review finds no marketing or false-success language.
    • UX.10.a Define loading copy that names the requested object or operation and does not imply success. Evidence: reviewed loading strings.
    • UX.10.b Define empty-state copy that distinguishes no records, no matches, no permission-visible records, and first-use setup. Evidence: reviewed empty-state variants.
    • UX.10.c Define offline copy that states readable, editable, queued, and unavailable capabilities plus reconnection behavior. Evidence: reviewed offline strings.
    • UX.10.d Define unauthorized and forbidden copy without disclosing hidden object existence or sensitive policy detail. Evidence: security-reviewed authorization strings.
    • UX.10.e Define not-configured copy that names the missing capability, authorized resolver, and safe next step. Evidence: reviewed configuration strings.
    • UX.10.f Define failure copy with affected action, retained data, correlation/support reference, retry eligibility, and recovery action. Evidence: reviewed failure strings.
    • UX.10.g Define partial-result copy that quantifies completed, failed, skipped, and still-running work. Evidence: reviewed partial-result strings.
    • UX.10.h Define stale-state copy with evidence age, invalidating change, impact, and refresh/recompute action. Evidence: reviewed stale strings.
    • UX.10.i Define retrying copy with attempt state, backoff or next-attempt timing, cancellation, and terminal-failure transition. Evidence: reviewed retry strings.
    • UX.10.j Test utility copy with screen readers, narrow layouts, long translations, and deliberately induced failures. Evidence: content QA matrix with no false-success language.

2.4 Current external standards baseline#

The versions below are a planning baseline, not a license to hard-code a version forever. Each domain must pin the implemented version, validate against the real conformance tooling, and record its upgrade policy.

Concern Baseline
Web accessibility WCAG 2.2 Level AA, including focus-not-obscured, dragging alternatives, target size, redundant entry, and accessible authentication
Media provenance C2PA Content Credentials with signed manifests, validation, trust, hard/soft bindings, ingredient lineage, and UX guidance
3D interchange OpenUSD validation plus usdchecker; explicit profile/version per target
Education interchange QTI 3, QTI accessibility, LTI 1.3/LTI Advantage, OneRoster, Caliper, Open Badges
News exchange IPTC NewsML-G2 current supported release, Media Topics, correction/update metadata, rights, packages, and planning items
Music delivery DDEX ERN implemented profile; ITU-R BS.1770 loudness/true peak; EBU R 128 where broadcast delivery applies
API security OWASP API Security Top 10 2023 plus repository security/privacy controls
Observability OpenTelemetry semantic conventions for traces, metrics, logs, profiles, and resources
  • STD.1 Record the exact WCAG 2.2 conformance target and supported assistive-technology/browser matrix. Evidence: accessibility test plan.
    • STD.1.a Record the required WCAG 2.2 level, included success criteria, allowed exemptions, and exception approval process. Evidence: conformance policy.
    • STD.1.b Enumerate supported browsers, engines, operating systems, viewport classes, input modalities, and zoom levels. Evidence: supported platform matrix.
    • STD.1.c Enumerate supported screen reader/browser/OS combinations and other required assistive technologies. Evidence: assistive-technology matrix.
    • STD.1.d Map automated, manual, and journey-level checks to each applicable success criterion. Evidence: criterion-to-test index.
    • STD.1.e Define severity, waiver, remediation deadline, and release-block rules for accessibility failures. Evidence: approved exception policy.
  • STD.2 Record the supported C2PA specification and trust-list policy. Evidence: provenance compatibility decision record.
    • STD.2.a Pin supported C2PA and Content Credentials specification versions and compatibility behavior for older/newer claims. Evidence: version-support table.
    • STD.2.b Define accepted manifest stores, hard and soft bindings, ingredient assertions, actions, and custom assertion policy. Evidence: manifest profile.
    • STD.2.c Define signer identity, certificate-chain, timestamp, revocation, trust-list source, update, and offline-cache policy. Evidence: cryptographic trust policy.
    • STD.2.d Define valid, invalid, untrusted, unknown, revoked, expired, missing, and partially recoverable UI/API results. Evidence: validation result taxonomy.
    • STD.2.e Define signing-key custody, rotation, compromise, revocation, and audit ownership. Evidence: approved key operations record.
    • STD.2.f Validate the profile with official or independently compatible fixtures. Evidence: interoperability report.
  • STD.3 Record the OpenUSD profile, schema set, resolver policy, and target DCC/engine version matrix. Evidence: interchange compatibility record.
    • STD.3.a Pin supported USD file formats, core version, composition features, layer conventions, and stage metadata. Evidence: OpenUSD profile.
    • STD.3.b Enumerate required, optional, extended, and forbidden schemas with version constraints. Evidence: schema compatibility table.
    • STD.3.c Define asset resolver schemes, search roots, package behavior, credentials, network access, and unresolved-asset failure semantics. Evidence: resolver policy.
    • STD.3.d Define units, axes, frame rates, color management, material, texture, animation, skeleton, and variant conventions. Evidence: semantic interchange rules.
    • STD.3.e Record import/export capability and known deltas for every target DCC, engine, plugin, and supported version. Evidence: target matrix.
    • STD.3.f Pin usdchecker and target-host validation commands plus accepted warning policy. Evidence: executable conformance plan.
    • STD.3.g Validate canonical round trips and unsupported-feature failures across the matrix. Evidence: checksummed compatibility fixtures and results.
  • STD.4 Record QTI, LTI, OneRoster, Caliper, Open Badges, SCORM, and xAPI versions that Metis imports or exports. Evidence: standards matrix.
    • STD.4.a Record QTI version, profile, accessibility extensions, interaction support, import/export direction, and lossy mappings. Evidence: QTI support row and fixtures.
    • STD.4.b Record LTI Core and Advantage services, message flows, security profile, platform/tool roles, and unsupported claims. Evidence: LTI support row and launch fixtures.
    • STD.4.c Record OneRoster version, CSV/REST modes, resource coverage, identifier mapping, and delta-sync behavior. Evidence: OneRoster support row and fixtures.
    • STD.4.d Record Caliper version, profiles, event coverage, actor/entity privacy, transport, and retry semantics. Evidence: Caliper support row and fixtures.
    • STD.4.e Record Open Badges version, assertion/profile fields, evidence, signing/verification, revocation, and wallet interoperability. Evidence: Open Badges support row and fixtures.
    • STD.4.f Record SCORM versions, packaging/runtime coverage, sequencing limits, import/export direction, and compatibility policy. Evidence: SCORM support row and fixtures.
    • STD.4.g Record xAPI version, statement profiles, endpoint/authentication, identifier mapping, and delivery/retry behavior. Evidence: xAPI support row and fixtures.
    • STD.4.h Define precedence and conflict handling when the same course, roster, activity, result, or credential crosses multiple standards. Evidence: cross-standard mapping decisions.
    • STD.4.i Validate declared support using conformance tooling or independently produced provider/consumer fixtures. Evidence: complete standards conformance report.
  • STD.5 Record the NewsML-G2 conformance level, catalog, vocabulary, and update/correction policy. Evidence: Veritas exchange profile.
    • STD.5.a Pin the supported NewsML-G2 release, conformance level, item types, package structures, and processing model. Evidence: versioned exchange profile.
    • STD.5.b Define provider, service, product, subject, genre, role, urgency, and Media Topics catalog/vocabulary mappings. Evidence: catalog mapping.
    • STD.5.c Define identifier, version, timestamp, language, location, relationship, and package-reference rules. Evidence: item identity and linkage rules.
    • STD.5.d Define rights, usage terms, embargo, source, credit, and confidential/protected metadata mappings. Evidence: rights mapping.
    • STD.5.e Define update, correction, replacement, cancellation, retraction, and tombstone semantics. Evidence: lifecycle transition profile.
    • STD.5.f Define planning-item ingest/export and assignment/calendar mappings. Evidence: planning interoperability profile.
    • STD.5.g Validate producer and consumer behavior with canonical, third-party, invalid, and version-transition fixtures. Evidence: conformance results.
  • STD.6 Record the DDEX ERN profile and every distributor-specific delta. Evidence: Euterpe delivery compatibility matrix.
    • STD.6.a Pin the ERN version, implementation profile, release types, resource types, deal model, and allowed message flows. Evidence: DDEX profile.
    • STD.6.b Define party, contributor, rights-controller, identifier, territory, date, takedown, and update mappings. Evidence: semantic mapping table.
    • STD.6.c Define audio, image, lyric, metadata, checksum, naming, directory, and transport package rules. Evidence: delivery-package profile.
    • STD.6.d Inventory every distributor's required, optional, forbidden, transformed, and defaulted fields. Evidence: distributor delta rows.
    • STD.6.e Define validation, rejection, warning, partial-acceptance, redelivery, acknowledgement, and reconciliation behavior. Evidence: delivery state model.
    • STD.6.f Validate canonical and distributor-specific packages against available schemas, validators, and sandbox endpoints. Evidence: destination-level compatibility results.
  • STD.7 Record loudness, peak, sample-rate, bit-depth, channel-layout, and codec targets per delivery destination. Evidence: audio delivery profile.
    • STD.7.a Inventory streaming, download, broadcast, cinema, immersive, social, preview, archive, and distributor destinations. Evidence: destination register.
    • STD.7.b Record integrated, short-term, and momentary loudness targets and measurement gating per destination. Evidence: loudness rows with units and tolerance.
    • STD.7.c Record true-peak and sample-peak ceilings plus measurement oversampling policy. Evidence: peak rows with units and tolerance.
    • STD.7.d Record sample rate, bit depth, dither, channel layout, bed/object, and downmix requirements. Evidence: PCM/layout rows.
    • STD.7.e Record codec, profile, bitrate or quality mode, container, metadata, gapless, and artwork requirements. Evidence: encoded-delivery rows.
    • STD.7.f Record normalization, limiting, resampling, encoding, and metadata-transformation rules without overwriting the source master. Evidence: derivative-generation policy.
    • STD.7.g Validate boundary, wrong-unit, clipping, silence, short-program, multichannel, and codec round-trip fixtures. Evidence: known-value delivery test report.
  • STD.8 Map every workbench API to the OWASP API risk controls. Evidence: threat model and automated negative-test index.
    • STD.8.a Inventory every public, internal, bridge, webhook, callback, streaming, and administrative API operation. Evidence: operation-level API register. <!-- workbench-api-register.json: 1,053 operations across two surfaces, source-derived (balanced-paren scan of every app.<method>() in the Oshun BFF + Electron ipcMain.handle + Euterpe Tauri #[command]). 755 HTTP (640 administrative under /v1/admin/{isis 164, aja 163, yemaya 113, bellona 100, studio 100}; 115 public studio-authoring incl. 70 Tara + isis/metis/veritas/ aja adapters) + 298 bridge (294 Yemaya Electron IPC incl. the Bellona bridge, 4 Euterpe Tauri commands). Classified by the STD.8.a 7-kind transport taxonomy; streaming=0/internal=0/webhook=0 are HONEST findings (no SSE producers or websocket upgrades; the 5 webhook-* paths are admin CONFIG APIs, not inbound receivers, so correctly administrative). Scope is the 7 V1 workbench domains + shared studio + Tara authoring; Hathor (not a V1 phase) and consumer /v1 routes excluded. generate-workbench-api-register.mjs + 8 tests, wired into inventory/verify + CI. STD.8.b-h (per-operation OWASP control mapping, negative tests, exception register) are security-analysis work and stay unchecked; the parent STD.8 stays [ ]. -->
    • STD.8.b Map broken object/property/function authorization controls and tenant boundaries to each operation. Evidence: authorization-control rows.
    • STD.8.c Map authentication, token/session lifecycle, sensitive business flow, and automated-abuse controls to each operation. Evidence: identity/abuse-control rows.
    • STD.8.d Map resource consumption, pagination, query complexity, payload, upload, storage, compute, queue, and spend controls. Evidence: consumption-control rows.
    • STD.8.e Map SSRF, unsafe downstream API consumption, redirect, DNS, egress, timeout, and response-validation controls. Evidence: remote-interaction rows.
    • STD.8.f Map inventory management, version retirement, documentation, environment exposure, configuration, and secret controls. Evidence: management-control rows.
    • STD.8.g Add operation-specific negative tests for every mapped control, including authenticated cross-tenant and property-level cases. Evidence: passing negative-test index.
    • STD.8.h Record residual risk, compensating control, owner, expiry, and retest date for every unautomated or unmet control. Evidence: approved exception register.
  • STD.9 Define OpenTelemetry attribute names, cardinality limits, trace propagation, and sensitive-data exclusions. Evidence: observability schema review.
    • STD.9.a Define canonical resource attributes for service, deployment, domain, host, tenant-safe scope, worker, and provider identity. Evidence: resource schema.
    • STD.9.b Define span names and attributes for HTTP, RPC, database, queue, job, model/provider, render, publish, and bridge operations. Evidence: trace schema.
    • STD.9.c Define metric instruments, units, aggregation, histogram boundaries, and allowed dimensions. Evidence: metric schema.
    • STD.9.d Define structured log event names, severity, correlation, error, retry, and audit-link fields. Evidence: log schema.
    • STD.9.e Set cardinality budgets and prohibit raw IDs, unbounded labels, content, prompts, source text, and stack traces as dimensions. Evidence: attribute allow/deny rules.
    • STD.9.f Define propagation through browser, API, worker, queue, provider, desktop, bridge, and callback boundaries. Evidence: propagation map.
    • STD.9.g Define PII, learner, source-identity, biometric, credential, rights, prompt, and customer-content redaction or exclusion. Evidence: privacy-reviewed telemetry policy.
    • STD.9.h Define sampling, retention, regional storage, access, deletion, and support-safe export behavior. Evidence: telemetry lifecycle policy.
    • STD.9.i Validate schema enforcement, propagation continuity, cardinality bounds, and sensitive-data absence. Evidence: automated observability contract tests.
  • STD.10 Establish quarterly standards-review ownership. Evidence: named owner, review cadence, compatibility policy, and backlog link.
    • STD.10.a Name the primary and backup owner for accessibility and assistive-technology support. Evidence: ownership register entry.
    • STD.10.b Name the primary and backup owner for C2PA provenance and trust policy. Evidence: ownership register entry.
    • STD.10.c Name the primary and backup owner for OpenUSD and target-host interoperability. Evidence: ownership register entry.
    • STD.10.d Name the primary and backup owner for education interchange standards. Evidence: ownership register entry.
    • STD.10.e Name the primary and backup owner for NewsML-G2 and newsroom exchange. Evidence: ownership register entry.
    • STD.10.f Name the primary and backup owner for DDEX and audio-delivery standards. Evidence: ownership register entry.
    • STD.10.g Name the primary and backup owner for OWASP API controls and OpenTelemetry conventions. Evidence: ownership register entries.
    • STD.10.h Define the quarterly review agenda, required evidence, change classification, compatibility window, deprecation notice, and emergency review trigger. Evidence: approved review procedure.
    • STD.10.i Link discovered changes to owned backlog items and verify closure at the next review. Evidence: standards-review issue template and cadence record.

2.5 Binding execution notes for the remaining work (added 2026-09-11)#

These notes come from a review of every open cell against the tree at commit 48f4bc8ddb. They bind every remaining phase; a per-cell <!-- SPEC 2026-09-11 --> comment below repeats only what is specific to that cell. Nothing in this section flips a checkbox.

  1. Name the host before building or testing UI. Every cell whose verb is build, run, exercise, or verify against a page names the host application, its port, and the route. Metis today: authoring UI is split between apps/metis/web (/create/* inside the LEARNER app, port 3020, login-only middleware with no role gate) and apps/metis/admin (hub, workspaces, psychometrics, credentials, integrations; port 3021), both backed by the Python service apps/metis/service (FastAPI, port 8000, env prefix METIS_). There is NO /studio/metis route in apps/oshun/web; the Section 1 table's "Oshun Studio workbench" host is unratified (M0.1.c is open) and libs/oshun/navigation/src/release-scope.ts scopes Metis to V1.2. Port 3020 is also apps/oshun/admin's port: never run both. For every other domain the N0.1 host reconciliation cell (V0.1, Y0.1, E0.1, A0.1, B0.1) is a precondition for that domain's UI cells; when the choice is a product decision, draft the ADR as Proposed, build against the existing host as-is, and say so in the cell. Never build a third host. Veritas has no authoring host at all today; the Phase V header note records what exists instead.
  2. "Live backing" means no request interception. The existing Metis Playwright suites are mocked: 64 of 65 specs in apps/metis/web/e2e and 20 of 68 in apps/metis/admin/e2e (plus every e2e/support/*-fixtures.ts) intercept **/api/** with page.route, and admin-fixtures.ts signs in with a literal test-admin-token. A full-journey cell (M10.9–M10.12, M10.21, M10.22, and every later N9.x journey) passes only when its spec contains no page.route, no fixture token, and its assertions read state back from the service or the database. Reusing those fixtures satisfies a UI-state cell, never a journey cell.
  3. A skipped real-database test is not a pass. The service's *_real_database.py and *_postgres_integration.py suites call pytest.skip when METIS_TEST_DATABASE_URL is unset and shell out to apps/metis/service/.venv/bin/alembic upgrade head. That .venv does not exist on this Mac (system python is 3.9; the service needs 3.11+). Recipe: /opt/homebrew/bin/python3.11 -m venv apps/metis/service/.venv, then apps/metis/service/.venv/bin/pip install -e 'apps/metis/service[dev]', docker compose -f docker/docker-compose.dev.yml up -d postgres minio redis, and METIS_TEST_DATABASE_URL=postgresql+asyncpg://oshun:oshun_dev@127.0.0.1:5432/<isolated db>. Every evidence block reports skipped counts beside passed counts.
  4. Model-backed evaluation binds the cheap route, per feature. The service defaults are wrong for tests: ai_provider="openai" (that account has no credits) and the floating deepseek/deepseek-v4-flash. Every evaluator has its own settings (METIS_<FEATURE>_MODEL_PROVIDER, _MODEL, _BASE_URL, _API_KEY for SOURCE_CLAIM, STANDARDS_ALIGNMENT, COURSE_OUTLINE, LESSON_ASSIST, and the rest in config.py): set provider openrouter, base URL https://openrouter.ai/api/v1, model deepseek/deepseek-v4-flash-0731, OPENROUTER_PROVIDER_SORT=price, and the key from ~/Desktop/workspace/drawer/env-master.env through a throwaway /tmp/<task>-env.sh. Never openai or anthropic; never a key in the repo.
  5. Corpora and labels may not be invented. A "licensed/consented representative held-out corpus" (M10.3.a, M10.23, I4.4.b, I9.6, and every later calibration cell) does not exist in this repository and cannot be written by an agent. An agent may build a repo-owned synthetic corpus whose labels derive from ground truth the fixture itself fixes (a known key, a planted contradiction, an executable worked solution), state in the manifest that it is synthetic, unrepresentative and un-adjudicated, and report thresholds as provisional. The parent stays open until a named human supplies or approves the corpus.
  6. External conformance receipts are one gate wherever they appear. M1.11.g/M10.5.d (QTI: 1edtech-qti3-xsd, longsight-qti3-cli, pinned in libs/contracts/src/metis/qti-package.ts), M1.12.g/M10.6 (LTI: the 1EdTech certification suite plus one independent LTI 1.3 platform), M1.13.g/M10.7 (OneRoster, Caliper, Open Badges, SCORM, xAPI validators pinned in interoperability-profiles.ts). A receipt is the tool's own output over our artifact, retained under evidence/v1-workbenches/metis-conformance/<tool>/ with the tool version and the artifact hash. Our own simulator, contract test, or schema is never a receipt. Install the tool when it is publicly installable and record the exact command and version; when it needs an account, a licence, or a hosted platform, record the attempt and leave the cell open. Docker-heavy tools (a Moodle or Canvas platform) run on the Linux dev server, not this Mac.
  7. Load, performance, and full Playwright sweeps obey the machine rules. memory_pressure | tail -2 (Mac) or free -m (Linux) before every heavy invocation. 68 admin specs and 65 web specs are sweeps, so next build && next start, never next dev; --project=chromium only (both configs also list firefox and webkit — do not run them); one Next server at a time; kill your PIDs at session end and verify with lsof. Load tests (M10.15 and every later N.15) run on the Linux dev server with a stop threshold written into the plan; a load run on this Mac is a recorded failed attempt, not evidence.
  8. WCAG 2.2, not the 2.1 helper. @oshun/testing/playwright's axeAfterEach runs WCAG_21_AA_TAGS. A cell that says 2.2 uses the kit's @oshun/workbench-kit/axe-policy and wcag22-coverage (S11.3) or passes tags including wcag22aa explicitly, and the evidence says which.
  9. Human approvals stay open, and so do their parents. Every cell whose verb is obtain, ratify, approve, or sign (M0.5.e, M0.11.b/f, M0.12.g, M0.13.f, M0.14.g, M10.20, S0.12, S11.24.h, DEP.2, and the same cells in every later phase) gets a readiness packet with the approver's name left blank. A generator that reports "0 of N earned" needs no re-run; the missing input is a person.
  10. Re-measure a blocker comment before acting on it. Several <!-- blocker --> notes predate later commits (M0.15.d's first vector changed on 2026-09-06). Read the code the note names, restate the current state with a date, then work.
  11. Structural invariants. One checkbox per identifier (a rollup note is prose, never a second - [x] <id> line); a parent is [ ] while any child is [ ]; every cell carries an Evidence: clause; child letters are contiguous; an identifier at end of line is still an identifier. node scripts/v1-workbenches/checklist-structure-lint.mjs checks these and exits non-zero; run it before every commit that touches this file.
  12. Ground a phase before building it (added 2026-09-18). Phases Y, V, E, A and B were enumerated, not decomposed: measured that day, more than half of the open cells in Y and E pack six or more slash-separated behaviours into one box, none names a file, package or route, and the evidence clause averages 2.4 words ("task tests."). Such a cell is not yet a task. The first work in a phase, after its N0 decisions, is a grounding pass over that phase only: for each cell, name the existing file, package, contract or route it changes (or say that none exists and where the new one goes), and split every cell whose behaviours have independent failure modes into children (.a, .b, …) that each carry their own specific Evidence: clause — the test file, the journey, the matrix row. Do it when the phase starts, never for a phase that is not next: a split made months ahead is written against code that will have moved. TODOS/registry.json lists these phases as deferred for that reason, so the task board does not offer their cells until the phase before them is done.
  13. Work order (added 2026-09-18). The task board's queue for this file is: section 1 (BASE), section 4 (DEP), the Phase S residue, the Phase M residue, then Phase I starting at I0 — with I0.20, the crosswalk against ISIS_CHROMA_RUNPOD_MVP_TODOS_2026-09-11.md, before any I6–I9 cell. Sections 2.3, 2.4 and 3 (UX., STD., CAP.*) are acceptance criteria that every domain phase must meet; they are checked when DONE.3 can be, not first, and a phase cell that satisfies one should name its id in its evidence. node tools/todos-board.mjs --next v1-domain-workbenches prints the next cells in that order.
  14. Who owns and who reviews (decision taken under the owner's delegation, 2026-09-18). This estate has one accountable person, the repository owner, and its delivery work is done by coding agents, so the words "owner", "independent review" and "signed" in a cell mean this. The accountable owner of every phase is the repository owner. The delivery owner of a cell is the agent session that holds its task on the Eve Task Board. The independent reviewer is a second agent session that did not author the work; its review is "signed" when it is recorded with the session, the date and the commit it read, through ./eve verify once ETB.4.07 exists and as a dated review note in the cell until then. That is separation of duties for engineering review, and DEP.2.a–f are worked against it. It does not stand in for the named human approvals of rule 9 (legal, privacy, education, accessibility, product sign-off). Those stay open; once the packet exists the cell carries the human blocked tag, which takes it out of the agent queue and puts it on the owner's list in TODOS/PARKED.md. Seven such cells were tagged on 2026-09-18 (M0.5.e, M0.11.b, M0.11.f, M0.12.g, M0.13.f, M0.14.g, M10.20).

3. Cross-cutting SOTA capability contract#

Every applicable row below must have domain-specific implementation and proof. Rejecting a row requires a P0 decision record and a compensating-control review.

3.1 Authoring and operations capability#

  • CAP.1 Frictionless capture supports keyboard command, paste, upload, import, deep link, and contextual creation where applicable. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.1.a Provide a discoverable keyboard command that opens capture with focus in the first required field. Evidence: per-domain shortcut and focus journey.
    • CAP.1.b Accept pasted text, URLs, files, and supported structured clipboard types without silently discarding formatting or provenance. Evidence: per-domain clipboard fixture matrix.
    • CAP.1.c Accept supported uploads with type, size, checksum, quarantine, progress, cancellation, and retry behavior. Evidence: per-domain upload journey and negative tests.
    • CAP.1.d Import each supported domain interchange format with preview, validation, conflict, partial-failure, and rollback behavior. Evidence: per-domain import conformance tests.
    • CAP.1.e Open capture from a safe deep link while preserving authorized target, tenant, and return context. Evidence: per-domain deep-link tests.
    • CAP.1.f Create an object contextually from its parent, selection, timeline position, graph node, source, or review item as applicable. Evidence: per-domain contextual-creation journeys.
    • CAP.1.g Record a reviewed N/A decision for each capture entry point a domain does not support. Evidence: capability matrix has no implicit omissions.
  • CAP.2 Capture preserves source metadata, rights hints, locale, actor, tenant, timestamp, and import checksum. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.2.a Preserve original source identifier, URI/path, filename, media type, size, and supplied descriptive metadata. Evidence: source metadata round-trip tests.
    • CAP.2.b Preserve supplied license, copyright, attribution, embargo, consent, and usage-condition hints without treating hints as verified rights. Evidence: rights-hint fixtures and labeling assertions.
    • CAP.2.c Preserve declared and detected language, script, region, and direction independently. Evidence: locale round-trip fixtures.
    • CAP.2.d Record the authenticated actor and effective delegated actor at the capture boundary. Evidence: actor attribution assertions.
    • CAP.2.e Record tenant and authorized workspace/project context without accepting client-selected cross-tenant ownership. Evidence: tenant-spoofing negative tests.
    • CAP.2.f Record source-observed, client-supplied, server-received, and persisted timestamps with explicit semantics. Evidence: clock-controlled timestamp tests.
    • CAP.2.g Compute the import checksum from received bytes and retain the algorithm and value through later revisions. Evidence: known-byte hash and mutation tests.
  • CAP.3 Structured editors use domain-typed blocks rather than an unvalidated generic JSON textarea. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.3.a Inventory the domain object and block types authors must create, inspect, reorder, nest, reference, or remove. Evidence: per-domain editor type register.
    • CAP.3.b Define a versioned schema and stable identifier for every editable block type. Evidence: machine-readable block contracts.
    • CAP.3.c Provide type-specific authoring controls, labels, help, constraints, and nonvisual semantics. Evidence: component fixtures and accessibility assertions per type.
    • CAP.3.d Validate field and cross-block invariants during editing and again at the service boundary. Evidence: shared invalid-fixture tests.
    • CAP.3.e Preserve unknown forward-compatible fields without exposing unsafe arbitrary mutation. Evidence: version-skew round-trip tests.
    • CAP.3.f Define block-version migration, unsupported-version read-only behavior, and lossless export. Evidence: migration and compatibility fixtures.
    • CAP.3.g Remove or confine raw JSON editing to an explicitly authorized developer tool with schema validation and audit. Evidence: route and authorization audit.
  • CAP.4 Autosave is durable, observable, recoverable, and explicit about offline or failed writes. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.4.a Define save triggers, debounce, maximum unsaved duration, coalescing, and flush-on-navigation behavior. Evidence: clock-controlled save scheduling tests.
    • CAP.4.b Acknowledge saved state only after the authoritative durable write and required outbox/event transaction commit. Evidence: forced commit-failure tests.
    • CAP.4.c Display dirty, saving, saved-at, queued-offline, retrying, failed, conflict, and recovered states. Evidence: state-transition component and browser tests.
    • CAP.4.d Persist offline edits locally with tenant/user binding, encryption where required, size limits, and expiry. Evidence: offline-store security and quota tests.
    • CAP.4.e Reconcile queued offline edits against the current server revision without last-write-wins data loss. Evidence: reconnect conflict tests.
    • CAP.4.f Preserve failed payloads and expose retry, copy/export, discard, and support-reference actions. Evidence: failure-recovery journey.
    • CAP.4.g Recover unsaved work after refresh, crash, tab suspension, and expired authentication. Evidence: browser lifecycle recovery matrix.
    • CAP.4.h Emit bounded save latency, failure, retry, conflict, and recovery telemetry without content. Evidence: telemetry contract tests.
  • CAP.5 Undo and redo cover local edits, structured operations, and conflict resolution without rewriting server history. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.5.a Define transaction boundaries for text entry, composition, paste, formatting, and multi-field local edits. Evidence: local undo-grouping tests.
    • CAP.5.b Implement inverse operations for add, remove, move, reorder, nest, link, split, merge, resize, and domain-specific structured edits. Evidence: operation-by-operation round-trip tests.
    • CAP.5.c Preserve selection, focus, viewport, and inspector context across undo and redo. Evidence: keyboard journey assertions.
    • CAP.5.d Define undo behavior after autosave, remote edits, validation, and server normalization. Evidence: save/remote boundary tests.
    • CAP.5.e Represent conflict-resolution choices as new reversible local operations rather than mutations of historical revisions. Evidence: conflict-resolution history tests.
    • CAP.5.f Clear or branch invalid undo history on restore, permission loss, schema migration, and destructive lifecycle transitions. Evidence: boundary-condition tests.
    • CAP.5.g Ensure redo never duplicates server commands or side effects such as publish, dispatch, render, or notification. Evidence: idempotency and side-effect tests.
  • CAP.6 Version history supports immutable revisions, semantic diff, authorship, restore-as-new-version, and durable audit linkage. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.6.a Persist immutable revision identity, parent/base revision, schema version, actor, timestamp, reason, and content hash. Evidence: revision persistence assertions.
    • CAP.6.b Prevent update or deletion of retained revision content outside an explicit governed retention process. Evidence: authorization and database immutability tests.
    • CAP.6.c Produce domain-semantic diffs for fields, blocks, relations, ranges, media/time data, graphs, or artifacts as applicable. Evidence: known-change diff fixtures.
    • CAP.6.d Attribute each change to the authenticated or delegated actor without inferring authorship from current ownership. Evidence: multi-actor attribution tests.
    • CAP.6.e Restore a historical revision by creating a new head revision that references the restored source. Evidence: restore lineage tests.
    • CAP.6.f Link revision creation, restore, migration, merge, and lifecycle events to immutable audit records. Evidence: revision-to-audit integrity tests.
    • CAP.6.g Paginate and authorize revision history without leaking deleted, protected, or cross-tenant content. Evidence: history query negative tests.
  • CAP.7 Concurrent editing has presence, selection awareness, merge rules, and an explicit conflict surface. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.7.a Define session identity, heartbeat, expiry, reconnect, duplicate tab, and privacy behavior for collaborator presence. Evidence: clock-controlled presence tests.
    • CAP.7.b Expose active collaborator identity and location only to authorized participants. Evidence: role and cross-tenant negative tests.
    • CAP.7.c Represent remote cursors, selections, object focus, and edit intent without color-only distinction. Evidence: concurrent-selection accessibility tests.
    • CAP.7.d Define merge semantics for text, typed blocks, ordered collections, references, metadata, and domain-specific structures. Evidence: deterministic concurrent-operation fixtures.
    • CAP.7.e Detect incompatible base revisions, delete/edit races, schema skew, and permission changes. Evidence: conflict-detection tests.
    • CAP.7.f Provide compare, choose, combine, duplicate-as-variant, retry, and preserve-for-support actions on the conflict surface. Evidence: conflict-resolution journey.
    • CAP.7.g Preserve unsent local work across reconnect and prevent remote updates from being overwritten silently. Evidence: network-partition browser tests.
  • CAP.8 Divergence supports best-of-N generation, manual variants, stable comparison, score provenance, and recorded selection rationale. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.8.a Define which domain objects may branch, the permitted variant count, and whether variants share or copy dependencies. Evidence: per-domain divergence contract.
    • CAP.8.b Create best-of-N candidates with stable candidate and originating run identifiers. Evidence: deterministic run-to-variant linkage tests.
    • CAP.8.c Create, duplicate, rename, archive, and delete manual variants without mutating sibling history. Evidence: manual variant lifecycle tests.
    • CAP.8.d Freeze comparison inputs and render semantically aligned differences across supported variant types. Evidence: stable comparison fixtures.
    • CAP.8.e Attach raw score, units, evaluator/corpus/model version, threshold policy, confidence, and timestamp to each scored candidate. Evidence: score provenance assertions.
    • CAP.8.f Allow selection only among authorized, nonstale, gate-eligible candidates or surface the exact blocking reason. Evidence: selection guard tests.
    • CAP.8.g Record selected candidate, considered alternatives, actor, rationale, evidence, and supersession. Evidence: decision-record integration tests.
  • CAP.9 Decision records capture considered options, evidence, actor, rationale, consequences, and reversal or supersession. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.9.a Assign stable identity, decision type, scope, owner, status, and effective version. Evidence: decision contract tests.
    • CAP.9.b Record every considered option, including retain-current-state, rejection rationale, and links to dependent decisions. Evidence: multi-option fixtures.
    • CAP.9.c Link immutable source, evaluation, review, experiment, and risk evidence with captured versions. Evidence: evidence-reference integrity tests.
    • CAP.9.d Record proposer, decider, reviewers, delegated authority, and decision timestamp. Evidence: actor/authority assertions.
    • CAP.9.e Record rationale, trade-offs, known consequences, affected audiences, and accepted residual risks. Evidence: required-field validation tests.
    • CAP.9.f Define proposed, under-review, approved, rejected, effective, superseded, and reversed transitions with authorization. Evidence: state-machine tests.
    • CAP.9.g Reverse or supersede through a new linked decision without rewriting the original record. Evidence: immutable lineage tests.
  • CAP.10 Maturity or stage transitions use server-enforced guards with exact, actionable refusal reasons. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.10.a Define the versioned state machine, permitted transitions, and terminal states per domain object type. Evidence: machine-readable transition definitions.
    • CAP.10.b Define required fields, dependencies, gate results, approvals, rights, consent, and provenance for each transition. Evidence: transition precondition table.
    • CAP.10.c Enforce object, transition, property, tenant, and delegated-role authorization at the service boundary. Evidence: transition authorization tests.
    • CAP.10.d Evaluate guards against the authoritative current revision in a concurrency-safe transaction. Evidence: stale-client and race tests.
    • CAP.10.e Return stable refusal codes plus the exact failed conditions, affected references, and permitted remediation. Evidence: negative contract fixtures.
    • CAP.10.f Prevent clients from bypassing guards through direct field writes, bulk operations, imports, jobs, or compatibility APIs. Evidence: alternate-path negative tests.
    • CAP.10.g Audit requested, refused, completed, reversed, and overridden transitions with correlated evidence. Evidence: audit assertions.
  • CAP.11 Review comments anchor to stable entities, fields, blocks, ranges, bars, frames, nodes, or timecodes as the domain requires. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.11.a Define stable anchor identity and revision semantics for whole entities and fields. Evidence: entity/field anchor contracts.
    • CAP.11.b Define stable block and text-range anchors that survive edits, splits, merges, and reordering where resolvable. Evidence: edit-rebasing fixtures.
    • CAP.11.c Define measure, beat, bar, sample, frame, and timecode anchors with units and frame-rate/time-base identity. Evidence: temporal anchor fixtures.
    • CAP.11.d Define graph node, edge, port, canvas object, scene object, and hierarchy anchors. Evidence: structural anchor fixtures.
    • CAP.11.e Detect deleted, replaced, migrated, permission-hidden, and ambiguously rebased anchors. Evidence: orphan-anchor taxonomy tests.
    • CAP.11.f Preserve the original excerpt, thumbnail, waveform, or structural context without leaking protected content. Evidence: context-snapshot security tests.
    • CAP.11.g Provide accessible navigation from comment to current anchor and from anchored content back to its comment thread. Evidence: keyboard/screen-reader review journey.
  • CAP.12 Review supports assignments, due dates, quorum, required roles, abstention, conflict of interest, escalation, and delegated authority. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.12.a Assign individual reviewers and role/group queues with acceptance, reassignment, and workload visibility. Evidence: assignment lifecycle tests.
    • CAP.12.b Define due dates with time zone, reminders, overdue state, extension, and service-level policy. Evidence: clock-controlled due-date tests.
    • CAP.12.c Define quorum calculation for eligible, completed, recused, abstaining, replaced, and delegated reviewers. Evidence: known-value quorum tests.
    • CAP.12.d Require named specialist roles for domain, safety, rights, legal, accessibility, security, or release reviews where applicable. Evidence: missing-role gate tests.
    • CAP.12.e Record abstention reason without counting abstention as approval, rejection, or silent absence. Evidence: verdict aggregation tests.
    • CAP.12.f Declare, review, enforce, and audit conflicts of interest and recusals. Evidence: conflicted-reviewer negative tests.
    • CAP.12.g Escalate overdue, blocked, disputed, unavailable-reviewer, and policy-triggered reviews to the configured authority. Evidence: escalation routing tests.
    • CAP.12.h Delegate authority with explicit scope, start, expiry, delegator, delegate, nondelegable actions, and revocation. Evidence: delegation authorization tests.
  • CAP.13 Search covers full text, structured fields, status, taxonomy, rights, quality, ownership, dates, and stable IDs. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.13.a Define indexed full-text fields, language analysis, stemming, phrase, typo, and protected-content behavior. Evidence: multilingual known-result tests.
    • CAP.13.b Index supported structured fields with typed equality, range, existence, and relationship filters. Evidence: field-query fixtures.
    • CAP.13.c Index lifecycle, review, gate, publish, job, stale, and failure statuses with current-state semantics. Evidence: status-filter tests.
    • CAP.13.d Index hierarchical taxonomy and aliases without collapsing distinct concepts. Evidence: taxonomy search fixtures.
    • CAP.13.e Index rights, consent, embargo, provenance, and usage eligibility without leaking protected detail. Evidence: rights-filter authorization tests.
    • CAP.13.f Index raw quality values, verdicts, evaluator versions, and freshness separately. Evidence: quality-query fixtures.
    • CAP.13.g Index owner, contributor, assignee, reviewer, team, and delegated actor with tenant-aware identity resolution. Evidence: ownership-query tests.
    • CAP.13.h Index created, updated, captured, effective, due, embargo, publish, and event dates with explicit time-zone semantics. Evidence: clock-boundary tests.
    • CAP.13.i Resolve exact stable IDs and supported external identifiers without fuzzy ambiguity. Evidence: exact-identifier tests.
    • CAP.13.j Apply object/property authorization, tombstone, retention, and tenant filters before counts, facets, snippets, and results. Evidence: search disclosure negative tests.
  • CAP.14 Saved views preserve query, columns, sort, grouping, density, and sharing scope. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.14.a Persist the normalized full-text and structured query, including relative-date and current-user semantics. Evidence: query round-trip tests.
    • CAP.14.b Persist visible, hidden, ordered, pinned, sized, and type-specific column configuration. Evidence: column-layout round-trip tests.
    • CAP.14.c Persist multi-key sort with direction, null ordering, locale, and deterministic tie-breaker. Evidence: stable-sort tests.
    • CAP.14.d Persist grouping, expansion, aggregation, and group ordering without storing unauthorized result data. Evidence: grouping tests.
    • CAP.14.e Persist density and supported display mode independently from the result query. Evidence: presentation-preference tests.
    • CAP.14.f Define private, named-user, team, role, tenant, and public sharing scopes with owner-controlled permissions. Evidence: share-authorization tests.
    • CAP.14.g Handle renamed/removed fields, taxonomy changes, authorization loss, and schema version skew. Evidence: saved-view migration tests.
    • CAP.14.h Expose clone, rename, make-default, transfer, revoke, and delete operations with audit. Evidence: saved-view lifecycle tests.
  • CAP.15 Bulk actions expose eligibility counts, skipped reasons, impact preview, progress, per-item results, cancellation, and safe retry. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.15.a Freeze the candidate set by stable IDs or a versioned query and display total selected versus currently matched. Evidence: concurrent-query selection tests.
    • CAP.15.b Evaluate object/property authorization, state, dependency, policy, and capacity eligibility per item before confirmation. Evidence: known-eligibility tests.
    • CAP.15.c Display eligible, ineligible, unknown, and already-satisfied counts with item-specific skipped reasons. Evidence: preview assertions.
    • CAP.15.d Preview resulting state, dependent impact, notifications, external side effects, cost, irreversibility, and rollback availability. Evidence: action-specific impact fixtures.
    • CAP.15.e Require confirmation proportional to impact and bind the approved action parameters to execution. Evidence: confirmation-tamper negative tests.
    • CAP.15.f Execute large batches as durable bounded-concurrency jobs with live completed/failed/skipped/pending progress. Evidence: batch job integration tests.
    • CAP.15.g Persist one result, reason, side-effect reference, and audit event per item. Evidence: per-item result completeness tests.
    • CAP.15.h Cancel unstarted work, report in-flight semantics, and preserve completed results without claiming atomic rollback. Evidence: cancellation tests.
    • CAP.15.i Retry only eligible failed items using stable idempotency keys and the original approved parameters. Evidence: partial-retry duplication tests.

3.2 Trust, AI, and provenance capability#

  • CAP.16 Every AI affordance is assistive by default and requires explicit human acceptance before it changes authored truth. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.16.a Inventory every generation, completion, rewrite, extraction, classification, scoring, recommendation, translation, and automation affordance. Evidence: per-domain AI affordance register.
    • CAP.16.b Define the authored source, proposed output, diff, confidence or uncertainty, provenance, and limitations visible before acceptance. Evidence: affordance-specific review contract.
    • CAP.16.c Store AI results as proposals, variants, or suggestions that do not mutate the authoritative revision on arrival. Evidence: persistence-state tests.
    • CAP.16.d Require an authorized human to accept each proposed change or an explicitly bounded group after reviewing its exact diff. Evidence: acceptance authorization and UI tests.
    • CAP.16.e Record accepted, partially accepted, edited-before-acceptance, rejected, expired, and superseded outcomes without treating absence as acceptance. Evidence: proposal lifecycle tests.
    • CAP.16.f Prevent automated publish, gate override, rights/consent attestation, source verification, grade, or other high-impact truth mutation. Evidence: prohibited-action negative tests.
    • CAP.16.g Link accepted content to the exact proposal and human actor while preserving later human edits independently. Evidence: authored lineage assertions.
  • CAP.17 AI output is labeled with provider/model/template versions and carries the originating run ID through publication. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.17.a Persist provider, model/deployment identifier, immutable model version or digest, and runtime region with every run. Evidence: run provenance assertions.
    • CAP.17.b Persist system/policy prompt, template, tool, retrieval, and evaluator versions or immutable hashes. Evidence: template/tool provenance assertions.
    • CAP.17.c Persist request parameters, safety policy, seed when supported, dependency versions, start/end time, actor, tenant, and cost. Evidence: complete run-record tests.
    • CAP.17.d Assign one stable originating run ID to each output and derivative proposal. Evidence: run-to-output cardinality tests.
    • CAP.17.e Display AI origin and accessible provenance detail in editor, compare, review, history, search, and published-preview surfaces. Evidence: surface-by-surface labeling tests.
    • CAP.17.f Preserve the run reference through copy, merge, restore, translation, export, publication, correction, and retirement without misattributing later edits. Evidence: lineage-transition tests.
    • CAP.17.g Reject or visibly flag an output whose referenced run record is missing, inaccessible, or integrity-invalid. Evidence: broken-lineage negative tests.
  • CAP.18 Prompt injection defenses fence user and retrieved content and test indirect injection through titles, metadata, attachments, and URLs. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.18.a Separate trusted system/policy instructions, developer templates, user requests, retrieved sources, tool results, and prior generated content in the execution contract. Evidence: role-tagged request fixtures.
    • CAP.18.b Mark retrieved and user-controlled content as untrusted data and prohibit it from granting tools, changing policy, or redefining output authority. Evidence: instruction-precedence tests.
    • CAP.18.c Minimize enabled tools and constrain each tool by typed arguments, allowlisted resources, authorization, tenant, and side-effect policy. Evidence: tool-capability matrix and negative tests.
    • CAP.18.d Normalize and inspect filenames, titles, descriptions, EXIF, document metadata, hidden text, captions, comments, and archive entries. Evidence: metadata injection corpus.
    • CAP.18.e Fetch URLs through the governed remote-fetch boundary and treat redirects, page text, headers, fragments, and linked attachments as untrusted. Evidence: URL injection and SSRF tests.
    • CAP.18.f Parse attachments in isolation with size/depth/type limits and retain source boundaries in the model input. Evidence: malicious attachment fixtures.
    • CAP.18.g Detect or constrain exfiltration requests for secrets, protected sources, learner/biometric data, unreleased assets, and cross-tenant context. Evidence: data-exfiltration negative tests.
    • CAP.18.h Test direct, indirect, encoded, multilingual, split-content, multimodal, tool-result, and multi-turn injections. Evidence: versioned adversarial evaluation report.
    • CAP.18.i Fail closed or require bounded human review when instruction trust cannot be determined. Evidence: ambiguous-trust behavior tests.
  • CAP.19 Safety classifiers expose model version, thresholds, confidence, uncertainty, and appeal or override policy. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.19.a Define the exact safety classes, applicability, input units, possible raw outputs, and non-pass states. Evidence: classifier contract.
    • CAP.19.b Persist provider, model, immutable version/digest, corpus, preprocessing, and runtime version with each result. Evidence: classifier provenance tests.
    • CAP.19.c Separate raw class scores from versioned policy thresholds and composed verdicts. Evidence: recomputation tests under changed policy.
    • CAP.19.d Expose value, unit, threshold, confidence/calibration, and uncertainty or insufficient-data status. Evidence: API and accessible UI assertions.
    • CAP.19.e Define allow, warn, block, quarantine, human-review, unavailable, unsupported, and evaluator-failure behavior. Evidence: state taxonomy tests.
    • CAP.19.f Define who may appeal, the submission evidence, reviewer independence, deadline, status, and notification behavior. Evidence: appeal lifecycle tests.
    • CAP.19.g Define override authority, scope, reason, expiry, downstream impact, and non-overridable classes. Evidence: override policy tests.
    • CAP.19.h Calibrate thresholds and subgroup behavior on representative, licensed, versioned corpora before release. Evidence: calibration and fairness report.
  • CAP.20 High-impact gate overrides require a reason, scoped authority, expiry where appropriate, and immutable audit. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.20.a Enumerate overrideable gates, nonoverrideable gates, allowable verdicts, and maximum risk scope. Evidence: approved override registry.
    • CAP.20.b Require a structured reason, supporting evidence, affected object/revision, intended consequence, and acknowledged residual risk. Evidence: request-validation tests.
    • CAP.20.c Enforce actor role, tenant, object, gate, target environment, lifecycle stage, and delegated-authority scope. Evidence: authorization matrix and negative tests.
    • CAP.20.d Require step-up authentication, separation of duties, quorum, or named specialist approval according to impact. Evidence: high-impact approval tests.
    • CAP.20.e Set effective time, expiry or one-use semantics, revocation, renewal, and automatic stale behavior. Evidence: clock-controlled tests.
    • CAP.20.f Bind an override to the exact gate result, inputs, evaluator, policy, object revision, and release target. Evidence: changed-input invalidation tests.
    • CAP.20.g Record request, evidence, approvals, refusals, activation, use, expiry, revocation, and downstream actions immutably. Evidence: append-only audit assertions.
    • CAP.20.h Surface active overrides and their scope in review, compare, publish, release, history, and incident views. Evidence: visibility journey tests.
  • CAP.21 Source, rights, canon, consent, or provenance dependencies form a traversable impact graph. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.21.a Define stable node types for source objects, rights records, canon facts, consent grants, provenance claims, authored revisions, generated artifacts, gates, and releases. Evidence: graph schema.
    • CAP.21.b Define typed, directional, versioned edges such as derives-from, cites, depends-on, licenses, consents-to, contradicts, supersedes, and published-as. Evidence: edge vocabulary.
    • CAP.21.c Create graph edges transactionally when ingest, generation, edit, merge, gate, export, publish, or correction creates a dependency. Evidence: mutation-to-edge integration tests.
    • CAP.21.d Preserve exact dependency version or revision instead of resolving only to a mutable latest object. Evidence: historical graph tests.
    • CAP.21.e Traverse upstream ingredients and downstream dependents with bounded depth, cycle handling, pagination, and stable ordering. Evidence: graph traversal tests.
    • CAP.21.f Apply tenant, object, property, protected-source, and rights/consent authorization to nodes, edges, counts, and paths. Evidence: graph disclosure negative tests.
    • CAP.21.g Expose impact paths, affected gates/releases, and uncertainty without overstating missing or inaccessible edges. Evidence: impact-view tests.
    • CAP.21.h Reconcile orphaned, dangling, duplicated, and cycle-invalid dependencies. Evidence: graph integrity checker.
  • CAP.22 Dependency changes automatically mark affected gates stale and enqueue or require recomputation. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.22.a Define which create, update, delete, revoke, expire, supersede, correction, and policy/version changes invalidate each gate type. Evidence: dependency-to-gate invalidation table.
    • CAP.22.b Emit a durable versioned dependency-change event in the same transaction as the authoritative mutation. Evidence: outbox atomicity tests.
    • CAP.22.c Traverse direct and transitive affected objects with cycle, depth, fan-out, and authorization-safe handling. Evidence: known-impact graph tests.
    • CAP.22.d Atomically mark affected gate results stale with invalidating event, old/new dependency versions, timestamp, and reason. Evidence: stale-record assertions.
    • CAP.22.e Enqueue idempotent recomputation where policy permits automatic work and record priority and cost/capacity decisions. Evidence: queue deduplication tests.
    • CAP.22.f Require explicit recomputation where consent, cost, policy, provider availability, or human judgment prevents automatic work. Evidence: manual-recompute state tests.
    • CAP.22.g Surface pending, running, failed, blocked, cancelled, and completed recomputation without restoring freshness prematurely. Evidence: recomputation lifecycle tests.
    • CAP.22.h Recover missed or failed invalidation through periodic reconciliation against dependency versions. Evidence: fault-injection and repair report.
  • CAP.23 Publish uses current gate evidence; previously passing stale evidence never silently authorizes release. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.23.a Define the complete required-gate set for object type, revision, audience, channel, territory, environment, and release action. Evidence: publish gate policy.
    • CAP.23.b Resolve gate results by exact input object/revision, dependency versions, evaluator, corpus/model, policy/threshold, and target. Evidence: exact-match selection tests.
    • CAP.23.c Reject missing, stale, expired, overridden-out-of-scope, unavailable, failed, or unsupported evidence with stable reasons. Evidence: publish refusal fixtures.
    • CAP.23.d Recheck dependencies and gate freshness in the authoritative publish transaction or reservation boundary. Evidence: publish/invalidate race tests.
    • CAP.23.e Freeze the accepted gate-evidence set and hashes into the release candidate. Evidence: release manifest assertions.
    • CAP.23.f Invalidate or abort a reserved candidate when a required input changes before final publication. Evidence: mid-release change tests.
    • CAP.23.g Audit the exact evidence consulted, refusals, overrides, actor, target, and final outcome. Evidence: publish audit integrity tests.
  • CAP.24 Artifact hashes are computed from stored bytes and verified again at import, promotion, publish, and retrieval boundaries. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.24.a Define approved hash algorithms, canonical identifier format, chunk/multipart behavior, and algorithm migration policy. Evidence: artifact-integrity profile.
    • CAP.24.b Compute the authoritative digest while ingesting or after reading the exact durably stored bytes, not client metadata. Evidence: client-hash spoofing tests.
    • CAP.24.c Store byte length, media type, storage version, hash algorithm, digest, and immutable blob reference transactionally. Evidence: metadata/blob consistency tests.
    • CAP.24.d Recompute and compare hashes at every external import and quarantine mismatches before parsing or use. Evidence: import corruption tests.
    • CAP.24.e Recompute and compare hashes when promoting between quarantine, draft, approved, release, or storage tiers. Evidence: promotion corruption tests.
    • CAP.24.f Recompute and compare hashes immediately before publish/export package signing or transfer. Evidence: prepublication tamper tests.
    • CAP.24.g Verify stored bytes during retrieval according to an explicit read-time, sampled-scrub, or full-scrub policy and fail closed on mismatch. Evidence: retrieval/scrub fault tests.
    • CAP.24.h Record corruption findings, affected derivatives/releases, quarantine, recovery source, repair, and revalidation. Evidence: integrity incident workflow test.
  • CAP.25 Signed provenance records preserve ingredients, transformations, tools/models, actors, policies, and validation status. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.25.a Assign a signed provenance record to the exact artifact bytes and immutable artifact identifier. Evidence: artifact-manifest binding tests.
    • CAP.25.b Enumerate each ingredient with identifier, hash, relationship, origin, rights/consent hints, and accessible or redacted status. Evidence: multi-ingredient fixtures.
    • CAP.25.c Record ordered capture, edit, transform, generation, merge, validation, approval, export, and publish actions with parameters. Evidence: action-history fixtures.
    • CAP.25.d Record application, library, plugin, tool, provider, model, template, policy, and environment versions used by each action. Evidence: tool/model provenance assertions.
    • CAP.25.e Record authenticated actor, delegated role, organization, and automation identity without exposing protected identities. Evidence: actor provenance and redaction tests.
    • CAP.25.f Record safety, rights, consent, quality, and release policy versions plus result references without presenting verdicts as facts beyond their scope. Evidence: policy linkage tests.
    • CAP.25.g Sign with governed keys, timestamp where required, preserve the certificate chain, and support rotation/revocation. Evidence: cryptographic lifecycle tests.
    • CAP.25.h Validate binding, signature, trust, revocation, ingredient availability, assertion support, and policy status independently. Evidence: provenance validation matrix.
  • CAP.26 Soft-binding and external-ledger recovery are tested when embedded provenance is stripped by a delivery channel. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.26.a Identify delivery channels and transformations that preserve, modify, or strip each embedded provenance mechanism. Evidence: channel-transformation matrix.
    • CAP.26.b Define privacy-reviewed soft bindings appropriate to media type, such as robust fingerprint, watermark, perceptual hash, or lookup hint. Evidence: binding design record.
    • CAP.26.c Register external-ledger records by immutable artifact/provenance identity with authorized lookup and retention. Evidence: ledger persistence and access tests.
    • CAP.26.d Recover the correct record from an unmodified derivative after embedded metadata removal. Evidence: positive recovery fixtures.
    • CAP.26.e Test resize, crop, transcode, recompress, remix, screenshot, re-record, channel normalization, and metadata rewriting as applicable. Evidence: transformation robustness matrix.
    • CAP.26.f Measure false positive, false negative, ambiguity, collision, and insufficient-evidence outcomes without claiming certainty. Evidence: known-corpus evaluation.
    • CAP.26.g Handle missing, multiple, revoked, private, deleted, and integrity-invalid ledger matches safely. Evidence: recovery negative tests.
    • CAP.26.h Display recovered, partially recovered, unresolved, and untrusted status with the method and limitations. Evidence: accessible recovery UX tests.
  • CAP.27 Secrets, protected source identities, biometric data, learner records, and unreleased assets never enter prompts or telemetry by default. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.27.a Classify credential material, tokens, keys, connection strings, and secret-bearing configuration at input and logging boundaries. Evidence: secret classification tests.
    • CAP.27.b Classify protected source names, contact details, documents, relationship graphs, and identifying metadata. Evidence: protected-source fixtures.
    • CAP.27.c Classify raw and derived face, voice, body, motion, health, and other biometric identifiers. Evidence: biometric classification fixtures.
    • CAP.27.d Classify learner identity, enrollment, activity, assessment, accommodation, integrity, and credential records. Evidence: learner-record fixtures.
    • CAP.27.e Classify unreleased media, scripts, builds, research, models, product plans, and embargoed artifacts. Evidence: unreleased-asset fixtures.
    • CAP.27.f Exclude classified values from prompt text, retrieval context, tool arguments, model metadata, traces, metrics, logs, profiles, and support exports by default. Evidence: boundary-by-boundary leakage tests.
    • CAP.27.g Permit exceptional processing only through an approved provider, purpose, minimization, region, retention, consent/legal basis, and explicit actor action. Evidence: exception-policy enforcement tests.
    • CAP.27.h Redact before emission, detect post-emission leakage, block or quarantine unsafe runs, and trigger incident handling. Evidence: canary-secret and sensitive-data tests.
    • CAP.27.i Verify provider no-training/no-retention controls and telemetry exporter filtering rather than relying on UI claims. Evidence: configuration and integration audit.

3.3 Production engineering capability#

  • CAP.28 Contracts are versioned, validated at every trust boundary, and backward compatibility is tested. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.28.a Inventory public, interservice, event, job, persistence, provider, desktop/bridge, import/export, and UI contracts. Evidence: per-domain contract register.
    • CAP.28.b Assign canonical ownership, namespace, identifier, semantic version, and source-of-truth representation to each contract. Evidence: ownership/version declarations.
    • CAP.28.c Define required, optional, nullable, default, invariant, enum, unknown-field, and extension semantics. Evidence: machine-readable schemas and contract notes.
    • CAP.28.d Validate untrusted input before domain use and validate output before crossing each external or asynchronous boundary. Evidence: boundary validation tests.
    • CAP.28.e Define version negotiation, unsupported-version failure, deprecation, compatibility window, and consumer migration behavior. Evidence: compatibility policy.
    • CAP.28.f Retain canonical valid, invalid, boundary, stale, unauthorized, and old/new-version fixtures. Evidence: versioned fixture corpus.
    • CAP.28.g Test current provider against supported consumers and current consumer against supported provider versions. Evidence: bidirectional contract test matrix.
    • CAP.28.h Detect schema/source/generated-client drift in CI. Evidence: passing drift gate and deliberate-mutation test.
  • CAP.29 Persistence uses explicit queryable columns for operationally important fields instead of opaque payload blobs. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.29.a Inventory fields used for authorization, tenancy, lifecycle, uniqueness, ordering, filtering, sorting, joining, retention, audit, freshness, and operations. Evidence: operational-field register.
    • CAP.29.b Map each operational field to a typed column with unit, format, nullability, default, constraint, and owning schema version. Evidence: logical-to-physical mapping.
    • CAP.29.c Normalize relationships that require referential integrity or independent lifecycle rather than embedding them in payload blobs. Evidence: relationship design review.
    • CAP.29.d Retain flexible payloads only for genuinely nonoperational, versioned extension data with size and schema limits. Evidence: approved payload-field inventory.
    • CAP.29.e Add indexes and query plans for supported access patterns without indexing sensitive content indiscriminately. Evidence: production-shape query-plan report.
    • CAP.29.f Backfill explicit columns from existing payloads with malformed, missing, conflicting, and version-skew handling. Evidence: rehearsal and reconciliation report.
    • CAP.29.g Prevent writes that update a blob without keeping explicit authoritative columns consistent. Evidence: dual-write invariant tests.
  • CAP.30 Every mutable table carries tenant, actor, created/updated, revision, tombstone, and concurrency metadata as applicable. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.30.a Record authoritative tenant and owning scope with foreign-key or equivalent isolation constraints. Evidence: cross-tenant write/read negative tests.
    • CAP.30.b Record creating actor, last-mutating actor, delegated actor, and automation/service identity as applicable. Evidence: actor attribution tests.
    • CAP.30.c Record server-created and server-updated timestamps with consistent precision and clock source. Evidence: timestamp invariant tests.
    • CAP.30.d Record monotonic revision or concurrency token and enforce it on conflicting mutations. Evidence: stale-write rejection tests.
    • CAP.30.e Define tombstone timestamp, actor, reason, retention, visibility, restoration, and purge linkage. Evidence: soft-delete lifecycle tests.
    • CAP.30.f Distinguish tables for immutable events/revisions, global reference data, ephemeral leases, and mutable tenant data with explicit N/A rationale. Evidence: table-by-table applicability review.
    • CAP.30.g Enforce metadata through shared persistence boundaries and block bypass by bulk/import/job paths. Evidence: alternate-write-path tests.
  • CAP.31 Migrations have forward, rollback-or-compensating, backfill, compatibility-window, and production-size rehearsal evidence. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.31.a Define the forward schema/data change, prerequisites, transactional boundaries, locks, and expected duration. Evidence: reviewed forward plan.
    • CAP.31.b Define exact rollback when safe or an explicit compensating migration when rollback would lose accepted writes. Evidence: rollback/compensation plan.
    • CAP.31.c Implement resumable, idempotent, observable, and rate-limited backfill with checkpoint and quarantine behavior. Evidence: interruption and resume tests.
    • CAP.31.d Define expand/migrate/contract compatibility windows for old and new application, worker, event, and schema versions. Evidence: version coexistence matrix.
    • CAP.31.e Validate preconditions and postconditions, row counts, hashes, invariants, nulls, duplicates, orphans, and sampled semantic results. Evidence: reconciliation queries.
    • CAP.31.f Rehearse forward, concurrent traffic, backfill, rollback or compensation, and redeploy on production-size representative data. Evidence: timed rehearsal report.
    • CAP.31.g Define monitoring, abort thresholds, operator commands, ownership, communication, and evidence capture for production execution. Evidence: migration runbook.
  • CAP.32 Retryable commands use idempotency keys and return the original result rather than creating duplicates. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.32.a Inventory retryable create, update, transition, job, publish, provider, webhook, import/export, and bulk commands. Evidence: idempotent-command register.
    • CAP.32.b Define idempotency-key scope by tenant, actor/client, operation, target, and compatibility version. Evidence: key-scope contract.
    • CAP.32.c Persist key, canonical request hash, state, authoritative result reference, error class, and expiry atomically with command acceptance. Evidence: persistence integration tests.
    • CAP.32.d Return the original success or still-running result for an identical request without repeating side effects. Evidence: concurrent duplicate-request tests.
    • CAP.32.e Reject reuse of a key with different parameters, target, identity, or contract version. Evidence: key-reuse negative tests.
    • CAP.32.f Define behavior for transient failure before acceptance, terminal failure after acceptance, timeout with unknown outcome, and expired records. Evidence: failure-state fixtures.
    • CAP.32.g Reconcile uncertain provider or downstream outcomes before retrying an irreversible side effect. Evidence: fault-injection tests.
  • CAP.33 Multi-system writes use an outbox, inbox, or documented saga with replay and poison-message handling. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.33.a Inventory mutations that span a database, blob store, queue, search index, provider, ledger, notification, or another domain. Evidence: multi-system mutation register.
    • CAP.33.b Choose transactional outbox, idempotent inbox, orchestrated saga, choreographed saga, or explicit reconciliation per mutation. Evidence: pattern decision rows.
    • CAP.33.c Persist authoritative mutation and outbound intent atomically where a local transaction exists. Evidence: forced-rollback tests.
    • CAP.33.d Deduplicate inbound messages by tenant, producer, event or command identity, and version. Evidence: duplicate-delivery tests.
    • CAP.33.e Define step order, preconditions, timeouts, retries, compensation, irreversible steps, and point of no return. Evidence: saga state-machine tests.
    • CAP.33.f Preserve ordering or explicitly tolerate reordering, duplication, late delivery, and concurrent consumers. Evidence: delivery-order fault tests.
    • CAP.33.g Quarantine poison messages with error class, payload-safe diagnostics, retry history, ownership, and authorized replay. Evidence: dead-letter journey.
    • CAP.33.h Reconcile missing, stuck, duplicated, and partially compensated work against authoritative state. Evidence: reconciliation repair tests.
  • CAP.34 Long-running work uses durable jobs with heartbeat, lease, cancellation, retry policy, checkpoint, and dead-letter semantics. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.34.a Define job type, versioned parameters, immutable input references, priority, queue, tenant, actor, cost estimate, and idempotency scope. Evidence: job contract.
    • CAP.34.b Persist queued, leased, running, waiting, cancelling, cancelled, retrying, failed, dead-lettered, and succeeded transitions. Evidence: job state-machine tests.
    • CAP.34.c Acquire exclusive time-bounded leases and renew heartbeats with worker identity and fencing token. Evidence: competing-worker and expired-lease tests.
    • CAP.34.d Define cooperative cancellation checkpoints, noncancellable regions, external cancellation, compensation, and terminal result. Evidence: stage-by-stage cancellation tests.
    • CAP.34.e Classify retryable, terminal, policy, authorization, invalid, provider, resource, timeout, and unknown errors. Evidence: error-to-retry table and tests.
    • CAP.34.f Persist progress and resumable checkpoints without reporting uncommitted work as complete. Evidence: crash/resume fault tests.
    • CAP.34.g Apply bounded exponential backoff, jitter, maximum attempts, age deadline, capacity, and spend limits. Evidence: deterministic retry-schedule tests.
    • CAP.34.h Dead-letter terminal/exhausted jobs with safe diagnostics, retained inputs, owner, remediation, and authorized replay or abandon. Evidence: dead-letter operations test.
    • CAP.34.i Reconcile orphan leases, silent workers, missing results, and side effects with unknown outcomes. Evidence: worker-loss repair test.
  • CAP.35 APIs paginate deterministically and define stable sort and cursor semantics under concurrent mutation. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.35.a Define default and maximum page sizes plus server behavior for zero, negative, excessive, or malformed values. Evidence: limit boundary tests.
    • CAP.35.b Define an explicit stable default sort and permitted client-selected sorts with deterministic unique tie-breakers. Evidence: equal-key fixtures.
    • CAP.35.c Encode sort, filter, direction, position, schema/version, and authorization-safe snapshot context in an opaque integrity-protected cursor. Evidence: cursor contract tests.
    • CAP.35.d Reject a cursor reused with incompatible filters, sort, identity, tenant, API version, or expired snapshot. Evidence: cursor misuse negative tests.
    • CAP.35.e Define insertion, update, deletion, tombstone, and permission change semantics between page requests. Evidence: concurrent-mutation pagination tests.
    • CAP.35.f Return next/previous or continuation metadata consistently without expensive or disclosure-prone total counts by default. Evidence: response contract tests.
    • CAP.35.g Verify no duplicate or skipped records within the declared consistency model. Evidence: randomized pagination property tests.
  • CAP.36 API rate, concurrency, payload-size, storage, compute, and spend limits are enforced per actor and tenant. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.36.a Define request-rate and burst budgets by actor, tenant, operation, client class, and trust tier. Evidence: rate policy.
    • CAP.36.b Define simultaneous request, job, stream, provider, render, bridge-session, and expensive-query limits. Evidence: concurrency policy.
    • CAP.36.c Define compressed and decompressed request, response, upload, archive-entry, field, and batch-size limits. Evidence: payload policy.
    • CAP.36.d Define object count, metadata, blob, revision, cache, log, and retained-output storage quotas. Evidence: storage policy.
    • CAP.36.e Define CPU, memory, GPU, wall-time, query complexity, token, frame/sample, and provider-use budgets. Evidence: compute policy.
    • CAP.36.f Define estimated, reserved, actual, daily/monthly, and exceptional spend limits with currency and attribution. Evidence: spend policy.
    • CAP.36.g Enforce limits at the earliest authoritative boundary and again before delayed or externally priced work. Evidence: bypass and reservation-race tests.
    • CAP.36.h Return stable quota category, current use where safe, reset or retry time, remediation, and correlation metadata. Evidence: limit-response fixtures.
    • CAP.36.i Prevent actors from evading tenant budgets and tenants from consuming another tenant's capacity. Evidence: identity/tenant abuse tests.
  • CAP.37 Remote fetches defend against SSRF, DNS rebinding, redirect escapes, oversized content, decompression bombs, and malicious media. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.37.a Parse and canonicalize scheme, user info, hostname, port, Unicode/IDNA, IP literals, and path before policy evaluation. Evidence: URL parser adversarial fixtures.
    • CAP.37.b Allowlist schemes and ports and block loopback, link-local, private, reserved, multicast, metadata, unix/file, and internal service destinations. Evidence: address-range negative tests.
    • CAP.37.c Resolve through a controlled resolver, validate every returned address, pin the connection target, and revalidate on reconnect. Evidence: DNS rebinding tests.
    • CAP.37.d Disable redirects by default or validate every redirect hop, target, scheme, port, credential, and maximum count. Evidence: redirect-escape tests.
    • CAP.37.e Apply connect, TLS, header, idle, body, total, retry, and bandwidth limits without trusting Content-Length. Evidence: slow/large response tests.
    • CAP.37.f Bound compressed size, decompressed size, ratio, archive entries, nesting depth, dimensions, frames, duration, and parser resources. Evidence: decompression and media-bomb corpus.
    • CAP.37.g Sniff actual content type, quarantine mismatches, and parse untrusted media in isolated constrained workers. Evidence: polyglot and malformed-media tests.
    • CAP.37.h Strip outbound credentials and sensitive headers, restrict egress, and prevent response data from becoming privileged instructions. Evidence: credential-leak and injection tests.
    • CAP.37.i Log safe policy decisions and correlation metadata without leaking signed URLs, credentials, protected locations, or content. Evidence: telemetry redaction assertions.
  • CAP.38 Uploads use quarantine, content sniffing, malware/moderation scans, checksum verification, retention, and orphan cleanup. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.38.a Create a tenant-bound upload intent with allowed types, maximum bytes, part count, expiry, actor, destination purpose, and idempotency key. Evidence: upload-intent contract tests.
    • CAP.38.b Write received bytes only to nonpublic quarantine and prevent downstream parsing, preview, indexing, or serving before promotion. Evidence: quarantine isolation tests.
    • CAP.38.c Enforce streaming byte/part limits and compute server-observed size and cryptographic checksum. Evidence: truncation/oversize/hash tests.
    • CAP.38.d Sniff actual type and structure, compare declared metadata, and reject polyglot, malformed, executable, or unsupported content. Evidence: adversarial type corpus.
    • CAP.38.e Run malware, content-safety/moderation, archive, and domain-specific validation with versioned results. Evidence: scanner result matrix.
    • CAP.38.f Promote only when every required scan is current and passing, binding the promoted blob to the scanned hash. Evidence: scan/promote race tests.
    • CAP.38.g Define quarantine, rejected, infected, appealed, expired, and retained-for-incident access and deletion policy. Evidence: retention lifecycle tests.
    • CAP.38.h Reconcile and remove abandoned multipart uploads, orphan quarantine blobs, stale intents, and failed promotions without deleting referenced evidence. Evidence: cleanup reconciliation tests.
  • CAP.39 Caches include tenant, authorization, schema, and dependency versions in their keys and expose invalidation reasons. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.39.a Inventory browser, application, HTTP/CDN, query, object, evaluation, render, provider, and search caches. Evidence: cache register.
    • CAP.39.b Include tenant and owning scope in keys or prove the cached data is globally public and immutable. Evidence: cross-tenant cache tests.
    • CAP.39.c Include actor/role/policy/permission version or cache only data whose representation is authorization-independent. Evidence: permission-change tests.
    • CAP.39.d Include contract/schema, object revision, dependency, evaluator, model/corpus, policy/threshold, locale, and target versions as applicable. Evidence: stale-version cache tests.
    • CAP.39.e Define TTL, stale-while-revalidate, negative caching, unavailable/failure caching, and stampede control. Evidence: cache-state timing tests.
    • CAP.39.f Emit durable invalidation from authoritative mutations and support bounded tag/key eviction. Evidence: mutation-to-invalidation tests.
    • CAP.39.g Surface hit/miss, stored/served age, reason, dependency versions, invalidation source, and refresh outcome without sensitive key data. Evidence: cache diagnostic assertions.
    • CAP.39.h Reconcile cache divergence and prove permission revocation, deletion, rights/consent change, and emergency purge take effect within policy. Evidence: emergency-invalidation tests.
  • CAP.40 Feature flags have owner, purpose, audience, default, expiry, rollback behavior, and removal task. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.40.a Assign a stable flag key, type, owning team/person, creation date, linked decision, and change authority. Evidence: flag registry.
    • CAP.40.b State the temporary purpose, hypothesis or migration, affected capability, success criteria, and non-goals. Evidence: reviewed purpose fields.
    • CAP.40.c Define eligible environments, tenants, roles, users, cohorts, devices, regions, and dependency conditions. Evidence: targeting contract and isolation tests.
    • CAP.40.d Define safe code/config default, missing-service behavior, evaluation-failure behavior, and new-tenant behavior. Evidence: default/failure tests.
    • CAP.40.e Set rollout start, review date, expiry, maximum lifetime, and alert before expiry. Evidence: clock-controlled expiry tests.
    • CAP.40.f Define rollback trigger, kill-switch authority, propagation budget, cached-evaluation invalidation, and state/data compatibility. Evidence: rollback rehearsal.
    • CAP.40.g Audit flag evaluations and changes with bounded cardinality and no sensitive targeting leakage. Evidence: telemetry/audit tests.
    • CAP.40.h Create and link the code, schema, data, test, documentation, and registry removal task at flag creation. Evidence: no flag lacks removal criteria.
    • CAP.40.i Block expired permanent-unknown flags and contradictory nested flag combinations in CI. Evidence: flag lifecycle lint tests.

3.4 Accessibility, localization, and resilient UX capability#

  • CAP.41 Every route passes automated WCAG 2.2 AA checks in each supported theme, density, viewport, and locale class. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.41.a Derive the reachable route and critical-state matrix from the authoritative route inventory. Evidence: no route/state is omitted or implicitly sampled.
    • CAP.41.b Run the configured rules in every supported light, dark, high-contrast, or branded theme. Evidence: theme-indexed scan results.
    • CAP.41.c Run each route in comfortable, compact, touch, and other supported density modes. Evidence: density-indexed scan results.
    • CAP.41.d Run at minimum supported narrow, medium, wide, zoomed/reflow, and orientation-dependent viewports. Evidence: viewport-indexed results.
    • CAP.41.e Run representative Latin, long-string, bidirectional, CJK or other supported locale/script classes. Evidence: locale-indexed results.
    • CAP.41.f Scan loaded, empty, validation-error, unavailable, conflict, dialog/drawer, and destructive-confirmation states where present. Evidence: state-indexed results.
    • CAP.41.g Triage every violation to a fix, explicit duplicate, tool false positive, or approved temporary waiver with owner and expiry. Evidence: zero unexplained findings.
    • CAP.41.h Keep the matrix in browser automation and fail releases on new unwaived AA violations. Evidence: deliberate-regression gate test.
  • CAP.42 Keyboard-only users can complete every critical journey. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.42.a Enumerate critical create/edit, review, gate, bulk, recover, publish/release, search, and administrative journeys per workbench. Evidence: approved keyboard journey inventory.
    • CAP.42.b Reach every actionable control through logical focus order without pointer, touch, voice, or direct accessibility-tree invocation. Evidence: focus traversal tests.
    • CAP.42.c Operate menus, dialogs, tabs, trees, grids, lists, editors, sliders, splitters, canvases, graphs, and timelines with documented keys. Evidence: widget-pattern interaction tests.
    • CAP.42.d Provide keyboard equivalents for drag, lasso, pan, zoom, trim, resize, reorder, connect, scrub, and spatial placement. Evidence: nonpointer journey tests.
    • CAP.42.e Preserve visible focus, selection, scroll position, working context, and escape route throughout each journey. Evidence: step-level assertions.
    • CAP.42.f Avoid traps except intentional modal focus with reachable dismissal and correct focus return. Evidence: trap and restoration tests.
    • CAP.42.g Test supported browser/OS/editor/screen-reader shortcut collisions and international keyboard layouts. Evidence: collision matrix.
  • CAP.43 Screen-reader announcements cover validation, save, job progress, gate changes, comments, conflict, publish, and destructive results. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.43.a Announce field and form validation at the affected control with error summary, count, and focus/navigation behavior. Evidence: validation announcement tests.
    • CAP.43.b Announce dirty, saving, saved, queued, retrying, failed, conflict, and recovered save states without excessive repetition. Evidence: save-state speech log.
    • CAP.43.c Announce job queued/running stage, meaningful progress, cancellation, retry, completion, partial result, and failure. Evidence: job speech log.
    • CAP.43.d Announce gate pending, running, stale, pass, warn, fail, unavailable, overridden, and recomputation changes. Evidence: gate speech log.
    • CAP.43.e Announce new comment, reply, mention, assignment, resolve, reopen, and anchor navigation in review context. Evidence: comment speech log.
    • CAP.43.f Announce edit conflict, affected content, comparison entry, resolution result, and preserved local work. Evidence: conflict speech log.
    • CAP.43.g Announce publish/release preparation, refusal, approval, dispatch, completion, partial delivery, correction, rollback, and failure. Evidence: release speech log.
    • CAP.43.h Announce destructive impact, confirmation requirements, progress, per-item result, recovery, and irreversible completion. Evidence: destructive-action speech log.
    • CAP.43.i Validate announcement priority, deduplication, interruption, language, accessible names, and real screen-reader behavior. Evidence: assistive-technology test matrix.
  • CAP.44 Graphs, canvases, timelines, waveforms, 3D previews, and charts provide a semantically equivalent navigable data view. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.44.a Inventory every non-DOM visual surface and define the domain facts, relationships, states, and operations it communicates. Evidence: visual-to-semantic content map.
    • CAP.44.b Provide a structured table/tree/list/transcript view with the same entities, values, units, ordering, hierarchy, and status. Evidence: semantic-view component tests.
    • CAP.44.c Synchronize selection, focus, filters, time/range, zoomed context, and active object bidirectionally between visual and semantic views. Evidence: synchronization tests.
    • CAP.44.d Provide nonvisual create, edit, connect, reorder, position, trim, annotate, compare, and delete operations where the visual surface is editable. Evidence: equivalent-operation journeys.
    • CAP.44.e Describe color, shape, line, texture, spatial, animation, waveform, and depth encodings in text and expose exact raw values. Evidence: encoding legend and accessible descriptions.
    • CAP.44.f Preserve performance through pagination, virtualization, range selection, and on-demand detail without hiding total/position semantics. Evidence: large-dataset accessibility tests.
    • CAP.44.g Verify the semantic view supports the complete critical journey with keyboard and screen reader. Evidence: nonvisual end-to-end journey.
  • CAP.45 Audio and video have captions, transcripts, visual alerts, and non-audio alternatives where content permits. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.45.a Inventory prerecorded, live, generated, uploaded, embedded, preview, monitoring, tutorial, and communication media. Evidence: media-accessibility register.
    • CAP.45.b Provide synchronized captions for dialogue and meaningful sound, with language, speaker, timing, correction, and quality review. Evidence: caption conformance tests.
    • CAP.45.c Provide navigable transcripts with speakers, timestamps, chapters/segments, search, download policy, and media synchronization. Evidence: transcript interaction tests.
    • CAP.45.d Provide audio description or an equivalent descriptive media alternative for essential visual information. Evidence: content-by-content applicability and proof.
    • CAP.45.e Convert alarms, clipping, cues, countdowns, stage signals, connection loss, and other audio-only status to visible and haptic alternatives where supported. Evidence: alert modality tests.
    • CAP.45.f Provide text, notation, data, waveform metrics, event lists, or controllable replay for domain information conveyed primarily by sound. Evidence: non-audio workflow tests.
    • CAP.45.g Expose caption/transcript availability, processing, failure, partial, language, and stale states without false completeness. Evidence: media-state UI tests.
  • CAP.46 Focus is restored predictably after dialogs, drawers, route transitions, virtualization, and deleted items. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.46.a Capture the invoking control and restore focus to it after modal dialog completion or cancellation when it remains available. Evidence: dialog focus tests.
    • CAP.46.b Restore focus after nonmodal drawer/inspector dismissal without moving the user's workspace selection. Evidence: drawer focus tests.
    • CAP.46.c Place focus at the route's meaningful heading, restored selection, validation target, or requested deep-link anchor after navigation. Evidence: route transition tests.
    • CAP.46.d Restore focus to a virtualized item by stable identity and scroll it into view without relying on recycled DOM position. Evidence: virtualization focus tests.
    • CAP.46.e After deletion, move focus by documented rule to next item, previous item, parent, collection heading, or recovery action. Evidence: first/middle/last deletion tests.
    • CAP.46.f Handle disabled, permission-hidden, filtered, moved, or concurrently deleted restoration targets with a safe announced fallback. Evidence: unavailable-target tests.
    • CAP.46.g Prevent asynchronous loading, toasts, status updates, and background rerenders from stealing focus. Evidence: focus stability tests.
  • CAP.47 Virtualized content preserves accessible row counts, position, selection, and keyboard navigation. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.47.a Expose total row/item count or an honest unknown/loading state independently from rendered DOM count. Evidence: collection semantic tests.
    • CAP.47.b Expose each item's stable position, set size, level, grouping, expansion, and relationship semantics. Evidence: screen-reader tree/grid logs.
    • CAP.47.c Preserve single, multiple, range, select-all-filtered, and offscreen selection by stable identity. Evidence: scroll/recycle selection tests.
    • CAP.47.d Implement arrow, page, home/end, type-ahead, group, and focus navigation according to the selected widget pattern. Evidence: keyboard navigation tests.
    • CAP.47.e Retain focused item and logical scroll anchor through prepend, append, resort, filter, resize, refresh, and concurrent mutation. Evidence: dynamic collection tests.
    • CAP.47.f Announce loading, loaded ranges, result-count changes, moved focus, and unavailable items without flooding live regions. Evidence: announcement tests.
    • CAP.47.g Test empty, one-row, huge, variable-height, nested, grouped, and permission-changing collections. Evidence: virtualization boundary matrix.
  • CAP.48 Drag-and-drop always has non-drag controls and does not require fine motor precision. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.48.a Inventory reorder, move, upload, place, connect, trim, resize, pan, scrub, lasso, and domain-specific drag interactions. Evidence: per-domain drag inventory.
    • CAP.48.b Provide select-plus-explicit-action controls for moving between parents, groups, lanes, tracks, stages, or collections. Evidence: non-drag move journeys.
    • CAP.48.c Provide move-up/down/start/end and direct-position controls for ordering. Evidence: non-drag reorder tests.
    • CAP.48.d Provide numeric, stepper, preset, snapping, and reset controls for position, duration, trim, size, and value changes. Evidence: precision-independent operation tests.
    • CAP.48.e Provide source/target selectors and create/remove actions for graph, routing, dependency, or relationship connections. Evidence: non-drag connection tests.
    • CAP.48.f Meet target-size/spacing requirements, support cancellation, avoid path-dependent gestures, and tolerate coarse pointers and tremor. Evidence: motor accessibility tests.
    • CAP.48.g Keep drag and non-drag paths behaviorally equivalent for validation, authorization, impact preview, audit, undo, and result. Evidence: path-equivalence tests.
  • CAP.49 Locale-aware formatting covers dates, times, time zones, numbers, units, currency, pluralization, sorting, and collation. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.49.a Format calendar dates with explicit locale, calendar system, granularity, and unambiguous machine value. Evidence: locale date fixtures.
    • CAP.49.b Format time, duration, relative time, time zone, daylight-saving transition, and ambiguous/nonexistent local time correctly. Evidence: clock and zone boundary tests.
    • CAP.49.c Format decimal, integer, percentage, scientific, compact, and significant-digit numbers without changing stored values. Evidence: numeric locale fixtures.
    • CAP.49.d Format domain units with conversion policy, displayed unit, precision, rounding, and source-unit preservation. Evidence: known-unit tests.
    • CAP.49.e Format currency with ISO identity, minor units, exchange-rate source/time when converted, and accounting behavior. Evidence: currency fixtures.
    • CAP.49.f Use locale plural categories, select/gender rules where needed, and parameterized messages without string concatenation. Evidence: pluralization tests.
    • CAP.49.g Apply locale-aware sort and collation with numeric, case, accent, punctuation, script, and stable tie-break behavior. Evidence: multilingual ordering fixtures.
    • CAP.49.h Keep canonical identifiers, hashes, code, paths, standards values, and protocol payloads locale-independent. Evidence: locale-switch invariant tests.
  • CAP.50 Bidirectional text, vertical-script metadata, long translations, and font fallback are visual-regression tested where applicable. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.50.a Test full RTL layout mirroring, navigation order, panel placement, icons, directional controls, and embedded LTR technical content. Evidence: RTL route snapshots and interaction tests.
    • CAP.50.b Test mixed-direction titles, names, URLs, paths, numbers, timecodes, code, identifiers, and punctuation with isolation controls. Evidence: bidi adversarial fixtures.
    • CAP.50.c Test vertical writing-mode metadata and distinguish content direction from interface direction. Evidence: vertical-script fixtures.
    • CAP.50.d Expand translatable strings to representative long forms and test wrapping, reflow, control growth, tables, dialogs, toasts, and narrow layouts. Evidence: pseudo-localization snapshots.
    • CAP.50.e Test primary and fallback fonts across supported scripts, combining marks, emoji, symbols, notation, and specialist glyphs. Evidence: glyph coverage snapshots.
    • CAP.50.f Detect tofu, clipped marks, baseline shifts, broken ligatures, reordered text, overlap, truncation without disclosure, and hidden controls. Evidence: reviewed visual-diff results.
    • CAP.50.g Run the matrix in each supported theme, density, viewport, and zoom level. Evidence: cross-dimension visual regression report.
  • CAP.51 Localizable content and non-localizable identifiers are separated in contracts and editors. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.51.a Classify each field as invariant identifier, source-authored content, localizable content, locale-neutral metadata, or derived display value. Evidence: field localization schema.
    • CAP.51.b Store stable IDs, keys, hashes, enum codes, standard vocabularies, paths, URIs, and protocol values independently from labels. Evidence: contract separation tests.
    • CAP.51.c Store localized values by canonical locale with source locale, fallback, translation status, translator, revision, and provenance. Evidence: localized-value contract tests.
    • CAP.51.d Prevent locale switching from rewriting identifiers, source-authored truth, references, search keys, or interoperability values. Evidence: locale-switch mutation tests.
    • CAP.51.e Present identifier and localizable label controls distinctly with appropriate validation, read-only state, and accessible help. Evidence: editor component tests.
    • CAP.51.f Define missing, fallback, stale, machine-suggested, human-reviewed, approved, and obsolete translation behavior. Evidence: translation lifecycle tests.
    • CAP.51.g Migrate legacy combined identifier/label fields without breaking external references or silently inventing translations. Evidence: migration reconciliation report.
  • CAP.52 Offline, reconnect, tab suspension, expired auth, and mid-operation network loss have deterministic recovery. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.52.a Define which reads, edits, commands, jobs, reviews, approvals, publishes, and destructive actions are available, queued, or prohibited offline. Evidence: per-domain offline capability policy.
    • CAP.52.b Detect offline, captive, degraded, high-latency, and server-unavailable states without equating every request failure to offline. Evidence: connectivity classification tests.
    • CAP.52.c Persist allowed local work with user/tenant binding, encryption, schema/version, size quota, expiry, and support-safe export. Evidence: offline-store tests.
    • CAP.52.d Reauthenticate before replay after expired auth and recheck current tenant, permissions, policy, object revision, and dependencies. Evidence: expired-auth replay tests.
    • CAP.52.e Resume or reconcile work after tab suspension, process freeze, browser discard, desktop sleep, and mobile backgrounding. Evidence: lifecycle automation matrix.
    • CAP.52.f Handle network loss before acceptance, after acceptance with unknown result, during upload, during job, and during streamed output. Evidence: boundary-specific fault tests.
    • CAP.52.g Replay only idempotent queued work in dependency order and surface conflict, rejection, partial completion, and stale intent. Evidence: reconnect replay tests.
    • CAP.52.h Preserve recovery evidence and never claim success until the authoritative state is read back. Evidence: false-success negative tests.
  • CAP.53 Deep links restore route, selected entity, tab, filter, inspector, and safe share context. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.53.a Define canonical versioned URL/path/query/state encoding for each restorable route context. Evidence: deep-link contract.
    • CAP.53.b Restore the authorized entity by stable identifier and handle missing, moved, deleted, retained, cross-tenant, or forbidden targets. Evidence: entity-resolution tests.
    • CAP.53.c Restore valid tab/subview and fall back predictably when a tab is removed, unsupported, or unauthorized. Evidence: tab compatibility tests.
    • CAP.53.d Restore typed filters, sort, grouping, pagination anchor, and saved-view references without accepting unsafe query expressions. Evidence: query-state round-trip tests.
    • CAP.53.e Restore inspector mode, selected subobject/anchor, compare target, and focus without opening hidden protected context. Evidence: inspector restoration tests.
    • CAP.53.f Separate shareable state from local-only drafts, secrets, protected identities, signed URLs, transient tokens, and oversized selection data. Evidence: URL leakage tests.
    • CAP.53.g Recheck authentication, tenancy, object/property authorization, lifecycle, and link expiry at resolution time. Evidence: deep-link security negative tests.
    • CAP.53.h Preserve a safe return location and provide intelligible recovery for stale, malformed, incompatible, or inaccessible links. Evidence: failure journey tests.
  • CAP.54 Browser refresh and back/forward navigation do not discard unsaved work or mutate state unexpectedly. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.54.a Classify route state as canonical URL state, server-persisted state, recoverable local draft, ephemeral UI state, or nonrestorable secret state. Evidence: route-state ownership map.
    • CAP.54.b Flush, queue, or persist allowed unsaved work before reload and warn only when data loss truly remains possible. Evidence: refresh recovery tests.
    • CAP.54.c Restore local drafts only for the same authenticated user, tenant, object, base revision, and compatible schema. Evidence: draft-isolation tests.
    • CAP.54.d Make route rendering and history restoration read-only unless the user explicitly repeats a mutation. Evidence: no-mutation navigation assertions.
    • CAP.54.e Restore route, selection, filters, inspector, focus, and reasonable scroll position on Back and Forward. Evidence: history navigation tests.
    • CAP.54.f Handle POST/redirect, pending upload, running job, open dialog, in-progress composition, and conflict states without duplicate commands. Evidence: lifecycle boundary tests.
    • CAP.54.g Detect server changes, deletion, permission loss, expired auth, and stale local base on restoration. Evidence: stale-restoration tests.
    • CAP.54.h Test hard refresh, soft reload, crash recovery, tab duplicate, tab restore, and browser session restore. Evidence: browser lifecycle matrix.

3.5 Operability and release capability#

  • CAP.55 Every user mutation emits a correlated audit event and trace. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.55.a Inventory create, edit, delete, restore, transition, review, approval, override, bulk, import, publish, configuration, and administrative mutations. Evidence: mutation-to-audit register.
    • CAP.55.b Assign stable mutation, command/idempotency, request, trace, actor/session, and resulting audit event identifiers. Evidence: correlation contract tests.
    • CAP.55.c Record authenticated and delegated actor, tenant, target, authorized action, prior/new revision, timestamp, reason, result, and policy decision. Evidence: audit schema assertions.
    • CAP.55.d Link to content hashes, diffs, gate/provenance evidence, and side-effect references without duplicating protected content in audit. Evidence: audit redaction and linkage tests.
    • CAP.55.e Persist the audit event in the same transaction or guaranteed outbox as the accepted mutation. Evidence: mutation/audit atomicity tests.
    • CAP.55.f Propagate trace context across API, database, queue, worker, provider, desktop/bridge, webhook, and notification boundaries. Evidence: end-to-end trace continuity tests.
    • CAP.55.g Record refused and failed mutation attempts at a safe detail level without claiming an authoritative state change. Evidence: negative audit tests.
    • CAP.55.h Enforce append-only integrity, retention, access, legal hold, export, and tenant isolation for audit events. Evidence: audit lifecycle tests.
  • CAP.56 Every job exposes queue delay, execution time, retries, resource use, provider/model, cost, outcome, and failure class. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.56.a Record accepted, eligible, queued, leased, started, stage, completed, and terminal timestamps with one clock/unit contract. Evidence: job timing assertions.
    • CAP.56.b Derive queue delay, scheduling delay, active execution, external wait, paused time, retry backoff, and total elapsed time separately. Evidence: known-timeline tests.
    • CAP.56.c Record attempt number, retry reason, policy, next-attempt time, worker, lease, checkpoint, and prior attempt linkage. Evidence: retry-history tests.
    • CAP.56.d Record bounded CPU, memory, GPU type/time/memory, storage, network, token, frame/sample, and provider usage as applicable. Evidence: resource instrumentation tests.
    • CAP.56.e Record provider, model/deployment, immutable version/digest, region, request reference, and provider operation/reference. Evidence: provider job provenance tests.
    • CAP.56.f Record estimated, reserved, actual, refunded, and unallocated cost with currency, rate-card version, and attribution. Evidence: cost reconciliation tests.
    • CAP.56.g Record succeeded, partial, cancelled, compensated, failed, dead-lettered, abandoned, and unknown outcomes without collapsing them. Evidence: terminal state tests.
    • CAP.56.h Classify input, authorization, policy, dependency, provider, timeout, capacity, resource, internal, cancellation, and unknown failure classes. Evidence: error classification tests.
    • CAP.56.i Expose support-safe job diagnostics and evidence links without content, secrets, signed URLs, protected identities, or unbounded labels. Evidence: telemetry privacy tests.
  • CAP.57 Metrics have bounded cardinality and exclude content, PII, protected-source data, and secrets. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.57.a Inventory each metric name, type, unit, owner, purpose, aggregation, retention, and allowed dimensions. Evidence: metric registry.
    • CAP.57.b Set explicit cardinality budgets per instrument and service plus warning and hard-fail behavior. Evidence: budget policy and load tests.
    • CAP.57.c Allowlist bounded domain, operation, state, error-class, provider-class, model-family, region, environment, and cohort labels as required. Evidence: attribute schema tests.
    • CAP.57.d Prohibit raw user/tenant/object/job/request IDs, URLs, filenames, query text, prompts, content, stack traces, and exception messages as labels. Evidence: forbidden-label lint tests.
    • CAP.57.e Redact or exclude direct and quasi-identifying PII, learner records, protected-source identity, biometric values, rights details, credentials, and secrets. Evidence: canary-data leakage tests.
    • CAP.57.f Handle unknown/new enum values through bounded fallback categories without silently dropping the event. Evidence: unknown-value tests.
    • CAP.57.g Monitor series count, label churn, dropped samples, exporter rejection, cost, and schema violations. Evidence: metric pipeline observability tests.
  • CAP.58 SLOs cover availability, latency, durability, queue age, publish correctness, and critical gate freshness. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.58.a Define user-visible service availability SLIs, qualifying requests, exclusions, target, window, and error budget. Evidence: approved availability SLO.
    • CAP.58.b Define latency SLIs and percentile targets for critical reads, mutations, searches, saves, jobs, reviews, and publish actions. Evidence: latency SLOs.
    • CAP.58.c Define acknowledged-write, revision, blob, event, audit, and restore durability/correctness SLIs. Evidence: durability SLOs.
    • CAP.58.d Define queue age, start delay, completion deadline, stuck job, and dead-letter SLOs by priority and job class. Evidence: queue SLOs.
    • CAP.58.e Define publish correctness for artifact bytes, metadata, audience/target, rights/consent, gates, provenance, delivery, and read-back. Evidence: publish SLO.
    • CAP.58.f Define maximum age and invalidation/recomputation time for safety, rights, consent, canon, quality, and release-critical gate evidence. Evidence: freshness SLOs.
    • CAP.58.g Assign owner, measurement source, dashboard, burn alerts, incident trigger, error-budget policy, and review cadence to each SLO. Evidence: operational SLO register.
    • CAP.58.h Validate SLIs against controlled success/failure events and prove missing telemetry cannot appear as success. Evidence: synthetic SLI tests.
  • CAP.59 Alerts are actionable, deduplicated, routed by ownership, and linked to a tested runbook. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.59.a Define the user/business impact, triggering SLI or invariant, severity, threshold, window, and recovery condition for each alert. Evidence: alert catalog.
    • CAP.59.b Include affected domain/service/tenant-safe scope, environment, start time, current value, recent change, dashboard, logs/traces, and correlation data. Evidence: notification template tests.
    • CAP.59.c Group duplicate symptoms by incident key and suppress dependent alerts without hiding independent failures. Evidence: deduplication scenario tests.
    • CAP.59.d Route primary, backup, specialist, security/privacy/safety, and vendor escalation by service ownership and time. Evidence: routing tests.
    • CAP.59.e Define acknowledgement, paging, escalation, silence, maintenance, resolve, reopen, and handoff behavior. Evidence: alert-lifecycle tests.
    • CAP.59.f Link a versioned runbook with validation, mitigation, rollback/kill switch, recovery, evidence preservation, and escalation. Evidence: runbook schema review.
    • CAP.59.g Exercise alert firing, notification, ownership, runbook steps, recovery detection, and post-incident evidence in a game day. Evidence: timed rehearsal report.
    • CAP.59.h Review noisy, unactionable, stale-owner, missing-data, and never-fired alerts on a defined cadence. Evidence: alert hygiene report.
  • CAP.60 Backup and restore prove both metadata and blob consistency, including provenance and audit records. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.60.a Inventory authoritative databases, blobs, manifests, search rebuild inputs, queues/outboxes, keys/config references, audit, and external-ledger data. Evidence: backup scope register.
    • CAP.60.b Define full/incremental cadence, point-in-time coverage, retention, region, immutability, encryption, key separation, and access. Evidence: backup policy.
    • CAP.60.c Capture a transactionally consistent recovery point or explicit ordering markers across metadata, blobs, events, audit, and provenance. Evidence: consistency design.
    • CAP.60.d Verify backup completeness, checksums, decryptability, schema version, catalog, and restore prerequisites automatically. Evidence: backup verification report.
    • CAP.60.e Restore into an isolated environment using documented credentials and infrastructure without depending on the failed primary. Evidence: independent restore rehearsal.
    • CAP.60.f Reconcile rows to blobs, revisions to content, provenance to ingredients, audit to mutations, outbox/inbox positions, tombstones, legal holds, and tenant counts. Evidence: post-restore integrity report.
    • CAP.60.g Rehash restored bytes and validate representative read, search, history, review, gate, export, and publish-preparation journeys. Evidence: restored-system verification.
    • CAP.60.h Measure achieved recovery point/time and record gaps, owners, remediation, and next rehearsal. Evidence: signed restore report.
  • CAP.61 Disaster recovery proves RPO/RTO and safe replay of jobs, outbox events, and pending releases. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.61.a Define disaster scenarios, affected regions/services/data, dependency assumptions, authority, and per-capability RPO/RTO. Evidence: approved DR scope.
    • CAP.61.b Provision isolated recovery infrastructure, networking, secrets/keys, providers, storage, queues, observability, and access. Evidence: recovery-environment readiness check.
    • CAP.61.c Restore authoritative state to the selected recovery point and measure actual data loss and elapsed time. Evidence: timed recovery log.
    • CAP.61.d Fence old writers, leases, schedulers, webhooks, bridges, and provider callbacks before activating recovery writes. Evidence: split-brain prevention tests.
    • CAP.61.e Reconcile queued/running/retrying/unknown jobs before replay and preserve idempotency, checkpoints, external outcome, and cost state. Evidence: job replay report.
    • CAP.61.f Replay outbox/inbox/events in safe order with duplicate, out-of-order, poison, and consumer-version handling. Evidence: event replay report.
    • CAP.61.g Revalidate candidate inputs, gates, rights/consent, provenance, approvals, idempotency, and destinations before resuming pending releases. Evidence: pending-release reconciliation.
    • CAP.61.h Verify user journeys, integrations, observability, alerts, backups, and support procedures before failover completion. Evidence: recovery acceptance matrix.
    • CAP.61.i Rehearse failback or establish the recovery environment as the new primary with data reconciliation and rollback. Evidence: failback plan and result.
  • CAP.62 Cost controls cover estimation, reservation, actual spend, variance, quotas, anomaly detection, and kill switch. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.62.a Define versioned unit prices for provider/model, tokens, GPU/CPU, storage, transfer, render frames/samples, licenses, and other billable resources. Evidence: rate-card registry.
    • CAP.62.b Estimate cost from validated inputs before acceptance and expose assumptions, currency, uncertainty, and expiry. Evidence: known-price estimate tests.
    • CAP.62.c Atomically reserve actor/project/tenant/provider budget for accepted work and handle concurrent reservations. Evidence: oversubscription race tests.
    • CAP.62.d Record actual metered usage and provider charges per attempt, job, derivative, retry, partial result, and cancellation. Evidence: usage attribution tests.
    • CAP.62.e Reconcile estimate, reservation, actual, refund/release, and invoice/exported cost with explicit variance reasons. Evidence: cost reconciliation report.
    • CAP.62.f Enforce per-operation, actor, project, tenant, day/month, provider, resource, and exceptional approval quotas. Evidence: quota boundary and isolation tests.
    • CAP.62.g Detect spend rate, variance, volume, provider, resource, retry, and unallocated-cost anomalies with bounded dimensions. Evidence: anomaly scenario tests.
    • CAP.62.h Provide scoped pause/kill controls with named authority, propagation target, in-flight policy, recovery, and audit. Evidence: kill-switch rehearsal.
    • CAP.62.i Prevent cost-control failure from becoming unbounded execution or fabricated zero cost. Evidence: unavailable-metering fail-closed tests.
  • CAP.63 Release supports canary or tenant cohort rollout, rollback, compatibility monitoring, and data-migration observation. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.63.a Define immutable release candidate, artifacts, contracts, schemas, flags/config, migrations, dependency versions, and evidence. Evidence: release manifest.
    • CAP.63.b Define canary or cohort selection by environment, tenant, role, region, platform, data shape, and risk without leaking targeting. Evidence: rollout targeting policy.
    • CAP.63.c Define rollout stages, minimum observation, success/error/SLO thresholds, approval points, automatic pause, and abort. Evidence: staged rollout plan.
    • CAP.63.d Verify old/new API, event, job, schema, cache, client, desktop/bridge, provider, and artifact compatibility during coexistence. Evidence: compatibility dashboards and synthetic tests.
    • CAP.63.e Observe migration progress, lock/latency, backfill, validation, drift, old/new writer behavior, and contract completion. Evidence: migration telemetry.
    • CAP.63.f Define code/config/flag rollback, database rollback or compensation, queued-job policy, artifact withdrawal, and irreversible boundary. Evidence: rollback runbook.
    • CAP.63.g Rehearse rollback from each stage with realistic traffic and in-flight mutations/jobs. Evidence: stage-level rollback results.
    • CAP.63.h Promote only on fresh evidence and record actor, cohort, timestamps, observations, exceptions, and final state. Evidence: release progression audit.
    • CAP.63.i Remove temporary compatibility, dual-write/read, and rollout flags after the defined observation window. Evidence: closure checklist.
  • CAP.64 A release evidence bundle links requirements, code, tests, scans, migrations, screenshots, performance results, runbooks, and approvals. Evidence: per-domain capability-matrix row linked to implementation and current proof.
    • CAP.64.a Pin release identifier, commit SHA, build inputs, dependencies, generated artifacts, configuration, target environment, and artifact hashes. Evidence: immutable release identity record.
    • CAP.64.b Link every in-scope requirement and ledger identifier to its implementation, disposition, or approved N/A evidence. Evidence: requirement traceability report.
    • CAP.64.c Link source diffs, review records, generated-code drift results, build outputs, and reproducibility or provenance references. Evidence: code/build evidence index.
    • CAP.64.d Link raw unit, property, contract, integration, end-to-end, mobile/desktop, evaluation, accessibility, resilience, and release test reports. Evidence: test evidence index.
    • CAP.64.e Link dependency, license, secret, static/dynamic security, authorization, privacy, safety, rights, consent, and provenance scans. Evidence: trust scan index.
    • CAP.64.f Link migration plans, rehearsals, compatibility windows, reconciliations, rollback/compensation results, and production observation queries. Evidence: migration evidence index.
    • CAP.64.g Link dated screenshots, recordings, network traces, console logs, accessibility outputs, and performance/load/cost/SLO results. Evidence: user/runtime evidence index.
    • CAP.64.h Link deployment, migration, rollback, incident, DR, support, provider, bridge/host, and kill-switch runbooks with rehearsal state. Evidence: operations evidence index.
    • CAP.64.i Link named technical, product, domain, security, privacy, safety, rights/legal, accessibility, operations, and human governance approvals without fabricating missing signoff. Evidence: approval register.
    • CAP.64.j Validate every link, hash, version, timestamp, owner, expiry, waiver, scope, and target before signing the bundle. Evidence: release-bundle integrity check.

4. Dependency and sequencing map#

  1. Complete BASE, STD, and Phase S architecture decisions before new workbench packages or routes.
  2. Migrate Tara to the shared kit without behavioral regression.
  3. Integrate existing standalone surfaces before creating replacements.
  4. Build domain contracts and known-value gate logic before persistence.
  5. Build persistence and service authorization before mutation UI.
  6. Build read-only UI against live contracts before enabling mutations.
  7. Build publish/release only after trust, provenance, and review gates.
  8. Run focused verification throughout; run each expensive gate sequentially under the repository resource-safety policy.
  9. Prepare engineering-readiness packets before human launch decisions.
  • DEP.1 Encode the phase dependency graph in a machine-readable file. Evidence: a cycle detector and unknown-ID validator pass. <!-- evidence: evidence/v1-workbenches/dependency-graph.json (23 nodes, 66 edges) + scripts/v1-workbenches/generate-dependency-graph.mjs; --check runs the cycle detector + unknown-ID validator and exits non-zero on failure → validation.passed=true. Command pnpm inventory:v1-workbenches; verify pnpm verify:inventory:v1-workbenches. -->
    • DEP.1.a Define the graph schema for phase/task node, dependency edge, edge reason, blocking strength, evidence gate, owner, and source. Evidence: versioned graph schema. <!-- evidence: schema 'v1-workbenches-dependency-graph/1'; nodes carry id/kind/label; edges carry from/to/class/sourceRule/reason/blockingStrength/evidenceGate/owner (evidenceGate & owner are null slots pending DEP.2). -->
    • DEP.1.b Encode every sequencing rule numbered above as explicit graph edges rather than prose-only constraints. Evidence: nine rules have machine-readable coverage. <!-- evidence: coverage.sequencingRulesCovered=[1..9], allNineRulesCovered=true; each edge records its sourceRule. -->
    • DEP.1.c Encode shared-to-domain, domain-to-domain, host, migration, contract, infrastructure, evidence, and human-approval dependencies. Evidence: dependency-class coverage report. <!-- evidence: coverage.dependencyClasses covers all 8 classes; domain-to-domain edges derived from §5 X.1-X.5 cross-domain integrations. -->
    • DEP.1.d Validate every node against an existing ledger identifier and every referenced artifact/environment against its registry. Evidence: zero unknown references. <!-- evidence: validation.unknownNodes=[] and danglingEdges=[]; phase nodes validated against ## Phase X headings, meta/integration nodes against task-ID prefixes parsed live from the ledger. No node references an external artifact/environment, so that clause is vacuously clean. -->
    • DEP.1.e Detect direct and transitive cycles and print the complete minimal cycle path. Evidence: cycle unit tests. <!-- evidence: findCycle() DFS returns the minimal a->b->c->a path; generate-dependency-graph.test.mjs asserts a positive 3-node cycle control + the real graph is acyclic (validation.cycle=null). -->
    • DEP.1.f Detect redundant, contradictory, self, orphaned, and unsatisfiable dependencies. Evidence: graph integrity report. <!-- evidence: validation.{selfEdges,redundantEdges,contradictoryEdges,orphanNodes,danglingEdges} all empty; dangling-edge check covers unsatisfiable (a dependency on a non-existent node). -->
    • DEP.1.g Generate stable critical-path and parallel-ready projections from the committed graph. Evidence: deterministic projection check. <!-- evidence: projections.criticalPath (8-stage build-order spine) + parallelReady (S, STAGE.contracts); determinism test asserts byte-stable output across runs. -->
  • DEP.2 Assign an owner and reviewer to every phase. Evidence: no phase has an unowned critical path.
    • DEP.2.a Enumerate shared, domain, cross-domain, rollout, and final-gate phases requiring accountable ownership. Evidence: phase ownership rows.
    • DEP.2.b Assign one accountable delivery owner with team and escalation path to each phase. Evidence: zero missing owners.
    • DEP.2.c Assign an independent reviewer with the required domain, platform, safety, rights, security, accessibility, or operations expertise. Evidence: reviewer qualification rows.
    • DEP.2.d Prevent self-review where the phase contains high-impact, regulated, trust, release, or product-scope decisions. Evidence: separation-of-duties validation.
    • DEP.2.e Assign backup ownership for critical-path phases and record handoff/escalation timing. Evidence: critical-path coverage report.
    • DEP.2.f Validate that every critical-path node inherits or declares both owner and reviewer. Evidence: graph ownership check.
  • DEP.3 Record cross-domain contract dependencies and compatibility windows. Evidence: producer/consumer matrix covers every integration.
    • DEP.3.a Inventory every synchronous API, asynchronous event, job, provider adapter, schema, blob/bundle, bridge, and embedded UI handoff across domains. Evidence: cross-domain integration register.
    • DEP.3.b Record producer, owner, contract identifier/version, transport, data classification, and availability/SLO for each integration. Evidence: producer-side matrix fields.
    • DEP.3.c Record every consumer, accepted versions, optional/required use, failure behavior, retry, and fallback. Evidence: consumer-side matrix fields.
    • DEP.3.d Define additive, breaking, deprecated, unsupported, and unknown version behavior. Evidence: compatibility policy per integration.
    • DEP.3.e Set coexistence start/end, producer-first or consumer-first ordering, feature/config flag, backfill, and removal conditions. Evidence: compatibility-window rows.
    • DEP.3.f Link provider and consumer contract, version-skew, replay, and failure tests. Evidence: no integration lacks bidirectional proof.
    • DEP.3.g Resolve producer/consumer owner conflicts and orphan consumers before scheduling implementation. Evidence: zero unresolved matrix rows.
  • DEP.4 Record required infrastructure, credentials, fixtures, corpora, media, DCC versions, GPU capacity, and test tenants. Evidence: environment readiness checklist.
    • DEP.4.a Inventory databases, blob stores, queues, caches, search, telemetry, signing, model/provider, rendering, bridge, and delivery infrastructure by environment. Evidence: infrastructure rows.
    • DEP.4.b Inventory service accounts, OAuth clients, certificates, signing keys, provider secrets, DCC/engine licenses, and rotation owners without recording secret values. Evidence: credential-reference rows.
    • DEP.4.c Inventory valid, invalid, boundary, adversarial, version-skew, tenant-isolation, and migration fixtures with generation/source provenance. Evidence: fixture rows.
    • DEP.4.d Inventory licensed evaluation corpora, expected outputs, subgroup coverage, access policy, retention, and version. Evidence: corpus rows.
    • DEP.4.e Inventory source, generated, malformed, rights/consent, multilingual, accessibility, performance, and delivery media assets. Evidence: media fixture rows.
    • DEP.4.f Record supported DCC, engine, plugin, codec, browser, device, desktop runtime, and driver versions plus installation source. Evidence: target-version rows.
    • DEP.4.g Record GPU/CPU/memory/storage/network capacity, concurrency, reservation, quotas, cost, and fallback. Evidence: capacity rows.
    • DEP.4.h Provision isolated tenants, users, roles, cross-tenant collision data, and provider states for every integration and journey. Evidence: test-tenant rows.
    • DEP.4.i Validate availability, permission, version, quota, health, and cleanup for every readiness row. Evidence: zero unverified prerequisites.
  • DEP.5 Define a critical-path release order with explicit parallel-safe slices. Evidence: the plan respects shared migrations and host ownership.
    • DEP.5.a Calculate the transitive critical path from the committed dependency graph with duration/confidence inputs. Evidence: generated critical-path projection.
    • DEP.5.b Identify architecture, product-boundary, contract, schema, security, infrastructure, and human-decision gates that block downstream starts. Evidence: blocking-gate list.
    • DEP.5.c Define the producer/consumer and expand/backfill/verify/contract order for every shared or cross-domain migration. Evidence: migration-sequencing rows.
    • DEP.5.d Define host-owner approval and deployment order for reused, proxied, redirected, embedded, or retired surfaces. Evidence: host-sequencing rows.
    • DEP.5.e Mark slices parallel-safe only when they share no incompatible contract, schema, data, host, environment, capacity, or approval mutation. Evidence: independence rationale per slice.
    • DEP.5.f Record integration, evidence-refresh, merge, and release synchronization points between parallel slices. Evidence: explicit join gates.
    • DEP.5.g Define replan behavior when an estimate, dependency, blocker, standard, provider, or product decision changes. Evidence: versioned schedule update procedure.
  • DEP.6 Define stop-the-line conditions for security, provenance, consent, data loss, gate correctness, and cross-tenant leakage. Evidence: release policy links each condition to an owner and rollback.
    • DEP.6.a Define credential exposure, exploitable authorization bypass, remote code execution, unsafe fetch/upload, dependency compromise, and critical unpatched vulnerability stops. Evidence: security stop policy.
    • DEP.6.b Define invalid/missing signature, artifact hash mismatch, broken ingredient lineage, untrusted required provenance, and signing key compromise stops. Evidence: provenance stop policy.
    • DEP.6.c Define missing/revoked/expired/out-of-scope consent, protected identity exposure, and nonconsensual biometric/voice/body use stops. Evidence: consent stop policy.
    • DEP.6.d Define acknowledged-write loss, corruption, unrecoverable revision/blob mismatch, failed restore, migration reconciliation gap, and unsafe replay stops. Evidence: data-loss stop policy.
    • DEP.6.e Define fabricated, nondeterministic, stale-authorizing, miscalibrated, bypassable, or unavailable fail-open critical gate stops. Evidence: gate-correctness stop policy.
    • DEP.6.f Define any cross-tenant read, write, count, search, cache, export, event, telemetry, or provider-context disclosure as a stop. Evidence: tenancy stop policy.
    • DEP.6.g Assign incident commander, technical owner, specialist reviewer, communication path, evidence custodian, and restart authority by stop class. Evidence: named responsibility matrix.
    • DEP.6.h Define automatic pause/kill action, affected targets, in-flight behavior, rollback/containment, recovery proof, and approval to resume. Evidence: stop-and-resume runbooks.
    • DEP.6.i Rehearse at least one trigger per class and prove the release cannot proceed while the stop remains active. Evidence: game-day report.

Phase S — Shared workbench platform#

Phase S builds @oshun/workbench-kit only for genuinely domain-neutral behavior. Domain rules remain in domain packages. Tara's current behavior and tests are the extraction acceptance gate.

S0 — Extraction audit and decisions#

  • S0.1 Inventory Tara domain-library exports consumed by web, BFF, tests, migrations, and content publishing. Evidence: import graph snapshot. <!-- evidence: evidence/v1-workbenches/tara-import-graph.json — generator scripts/v1-workbenches/generate-tara-import-graph.mjs (deterministic, git-grep driven, --check drift gate), 14 known-value tests, wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches + CI. -->

    • S0.1.a Enumerate Tara public entry points and exported symbols from package metadata and source. Evidence: export-level inventory. <!-- evidence: .exportInventory — 209 symbols resolved from the src/index.ts barrel (star + named re-exports followed), each with origin module id and declaration kind (const/function/class/type/interface/enum). -->
    • S0.1.b Map web application imports to exact exports and consuming routes/components. Evidence: web consumer edges. <!-- evidence: .consumersByArea.web = [] — no web surface imports the domain library directly; apps/oshun/web/.../tara-workbench reaches Tara through the BFF over HTTP, not via @oshun/tara-workbench. Honest zero. -->
    • S0.1.c Map BFF imports to handlers, middleware, stores, and runtime paths. Evidence: BFF consumer edges. <!-- evidence: .consumersByArea.bff — 7 production BFF files (routes, server/app wiring, tara-workbench/{evidence,plane-categories,review-verdict}, agentic bundle publisher), each edge lists exact imported symbols. -->
    • S0.1.d Map unit/integration/browser test, fixture, and builder imports without treating test-only use as production reuse. Evidence: test consumer edges. <!-- evidence: .consumerEdges[].isTest — 2 test edges segregated (routes + bundle-publisher specs); excluded from every production-reuse count and from the unused-export computation. -->
    • S0.1.e Map migration, schema, code-generation, and persistence registration imports. Evidence: data-path consumer edges. <!-- evidence: .consumersByArea.dataPath = [] — no data-path import edge. libs/oshun/persistence contract-persistence-registry.ts references Tara only in a comment (no import); classifier + comment-stripping confirm. -->
    • S0.1.f Map content build, validation, bundle, and publishing imports. Evidence: publishing consumer edges. <!-- evidence: .consumersByArea.publishing — tara-bundle-publisher.ts (assembleMeditationBundle + MeditationBundle* types), the one production publishing consumer. -->
    • S0.1.g Flag unused exports, deep imports, circular imports, and exports with only one real consumer. Evidence: reviewed import graph findings. <!-- evidence: .findings — 151 unused exports (validated: sampled 12, all 0 non-lib/non-test hits), 2 test-only exports, 56 single-consumer exports (all @oshun/bff), 0 deep imports, 0 intra-package cycles. -->
  • S0.2 Inventory Tara web components and classify domain-neutral shell behavior separately from meditation-specific behavior. Evidence: component extraction matrix. <!-- evidence: inventory (a,b) = tara-component-inventory.json; classification (c-f) = tara-component-extraction-matrix.md + component-extraction-matrix.json, the full extraction matrix (29 components → 4 neutral / 13 Tara-owned / 9 split / 2 reject / 1 shared-shell; 13 extractable contracts two-archetype validated). Named architecture ratification of the matrix stays a human step (see S0 signoff); the classification is coherence-gated in CI (generate-component-extraction-matrix.mjs --check, 11 tests). --> <!-- adr: docs/adr/ADR-S1-workbench-kit-package-ownership.md carries the ownership + dependency direction this matrix's extraction targets must respect. -->

    • S0.2.a Enumerate route, layout, workspace, inspector, editor, review, gate, publish, history, and utility components. Evidence: source-linked component rows. <!-- evidence: .components — 29 rows (28 StudioTaraWorkbench* owned + shared ShellLayout), each source-linked (.file) and role-tagged across the full taxonomy, route-linked to the page(s) that render it (multi-component pages captured). generate-tara-component-inventory.mjs, --check gate, tests. -->
    • S0.2.b Record each component's props, state, context, hooks, data requests, mutation paths, styles, tokens, and tests. Evidence: dependency columns are complete. <!-- evidence: per-row columns props{hasProps,propNames,propsType}, localStateHooks, contextHooks, hooks[], dataRequests[] (GET), mutationPaths[] (POST/PUT/PATCH/DELETE, BASE-const resolved), inlineStyleBlocks, designTokens[] (var(--…)), hasTest/testFile. 24 read + 27 mutation endpoints; 25 owned components have no co-located test (surfaced coverage gap). -->
    • S0.2.c Identify behavior whose contract and interaction semantics are independently usable by at least one non-Tara archetype. Evidence: domain-neutral rationale per candidate. <!-- evidence: evidence/v1-workbenches/tara-component-extraction-matrix.md + component-extraction-matrix.json .buckets.neutralCandidate (4 zero-signal) plus the 9 mixed-split neutral contracts = 13 extractable contracts, each with a per-component domain-neutral rationale grounded in its real routes. generator generate-component-extraction-matrix.mjs (--check gate), 11 tests. -->
    • S0.2.d Identify meditation vocabulary, stages, content blocks, gates, scores, routes, providers, copy, and policy that must remain Tara-owned. Evidence: Tara-specific rationale per retained component. <!-- evidence: .buckets.taraOwned (13) + the 2 mixed-reject rows, each with a Tara-specific rationale citing its meditation-vocabulary terms (invitational/ breath/grounding/spark/cadence/seasonal/plane/steering/meditation). Coherence- checked: every highMeditationSignalComponent is Tara-owned. -->
    • S0.2.e Split mixed components at an explicit seam or reject extraction when the seam would leak Tara semantics. Evidence: split/reject decision per mixed row. <!-- evidence: .buckets.mixedSplit (9, each with a named neutral contract + seam) and .buckets.mixedReject (2, residue too thin). Seam = the interaction contract over a domain-supplied adapter; Tara keeps routes/copy. Every mixed row carries a split|reject verb (test-enforced). -->
    • S0.2.f Validate proposed neutral components against both an authoring and an operations-console consumer. Evidence: two-archetype usage sketches. <!-- evidence: .archetypeSketches — all 13 extractable neutral contracts have a concrete authoring-console AND operations-console usage sketch (e.g. CommentThread: draft comments vs incident-timeline comments). Test asserts both archetypes present for every contract. EvidenceGroundingPanel/SourceLibrary name Veritas as the near-at-hand real 2nd consumer (S0.11 gate). -->
  • S0.3 Inventory Tara BFF handlers, middleware, stores, and provider seams. Evidence: route-to-capability matrix. <!-- evidence: evidence/v1-workbenches/tara-bff-matrix.json — generator scripts/v1-workbenches/generate-tara-bff-matrix.mjs (string/comment/template- aware balanced-paren route scanner), 14 known-value tests, wired into inventory/verify:inventory:v1-workbenches + CI. 70 routes fully covered. -->

    • S0.3.a Enumerate all mounted and source-defined Tara BFF routes with method, path, handler, middleware chain, and reachability. Evidence: effective route rows. <!-- evidence: .routes — 70 rows (GET 28 / POST 36 / PATCH 5 / PUT 1), each with resolved path, guard wrapper (all guarded), and middleware chain ([abuseProtection, authProtection]). Reachability = the guarded auth+scope+ store gate (absent store ⇒ 503 not_configured); 0 ungated routes. -->
    • S0.3.b Trace each handler to command/query, domain service, store, provider, job, blob, and external side effect. Evidence: handler-dependency rows. <!-- evidence: per-route storeMethods[] (47 distinct), domainCalls[] (@oshun/tara-workbench symbols actually invoked), auditEvents[] (32 distinct), externalEffects[] (publish-manifest/model-provider/http-fetch/ blob). -->
    • S0.3.c Record actor resolution, object/property/function authorization, tenancy, validation, idempotency, concurrency, rate/size limits, and audit per route. Evidence: boundary-control rows. <!-- evidence: per-route guard (actor+tenancy+scope), middleware (auth + abuse/rate/size), permissions[] + dynamicPermissions[] + hasPermissionCheck (function authz, literal AND dynamic e.g. stage.requiredPermission), validationSchemas[], usesIdempotency, concurrencyControls[] (revision-version/supersede-in-place/409-conflict), auditEvents[]. Surfaces 7 auth-only mutations (guarded but no function-permission, e.g. script/compile binds _actor). -->
    • S0.3.d Classify stores as durable production, in-memory production, fixture/demo, cache, or adapter and record lifecycle semantics. Evidence: store classification rows. <!-- evidence: .stores — TaraWorkbenchStore = durable-production, backed by @oshun/persistence ContractPersistenceService (durable-before-ack writes, typed read client); new Map() usages are local read-aggregation, not the backend. -->
    • S0.3.e Classify provider seams as configured, not configured, fallback, fixture, or accidental mock with error/cancellation/cost behavior. Evidence: provider classification rows. <!-- evidence: .providers — TaraWorkbenchModelProvider = configured-or-fail-closed: a null LLM provider makes every capability (diverge/extractClaims/assistScript) throw TaraWorkbenchNotConfiguredError (honest fail-loud seam; no fabrication, fixture, or accidental mock). -->
    • S0.3.f Map each route to the smallest shared or Tara-owned capability and its current tests. Evidence: zero unmapped reachable routes. <!-- evidence: every route carries a capability id (tara-workbench.<segs>.<verb>) — zero unmapped reachable routes — plus a tested flag from the spec's inject() paths. 59/70 tested; the 11 untested routes are surfaced honestly in coverageGaps.untestedRoutes (a coverage gap, not an unmapped route). -->
  • S0.4 Inventory @oshun/studio-authoring overlap. Evidence: duplication report identifies reuse, consolidation, and intentionally separate logic. <!-- evidence: inventory + duplication/cycle checks (a,f) = tara-studio-authoring-overlap.json; the reuse/consolidate/separate SELECTION (b-e) = studio-authoring-overlap-selection.json + tara-studio-authoring-overlap-selection.md (0 exact dups; 4 reuse / 0 consolidate / 5 separate over the 9 SA concern categories; DECISIVE FINDING: the two packages are already correctly factored — no consolidation debt). Named architecture ratification stays a human step (S0 signoff); the selection is coherence-gated in CI (11 tests). -->

    • S0.4.a Enumerate public and internally consumed studio-authoring contracts, actor/role logic, components, hooks, stores, utilities, and tests. Evidence: source-linked inventory. <!-- evidence: .studioAuthoringInventory — 315 exports across 9 barrel categories, each source-linked (.module) + kind-tagged; surfaceBuckets show actor-role=36, contracts-domain-logic=279, components/hooks/stores=0 (pure logic library — Tara's 28 UI components can NOT consolidate into it without a new UI layer), tests=6. generate-tara-studio-authoring-overlap.mjs, tests. -->
    • S0.4.b Match exact and near-equivalent Tara capabilities by input, output, lifecycle, policy, audience, and failure semantics. Evidence: candidate overlap pairs. <!-- evidence: studio-authoring-overlap-selection.json .overlap — recomputed from the two committed inventories: 0 exact-name duplicates + 117/209 Tara exports sharing a CONCERN token (source/manifest/gate/block/review/asset/ transition/stage/publish/concept/decision/divergence) with an SA symbol = concern-level candidate pairs, NOT impl duplicates. generator generate-studio-authoring-overlap-selection.mjs (--check gate), 11 tests. -->
    • S0.4.c Select direct reuse where ownership and behavior already match. Evidence: reuse rows with consumer migration path. <!-- evidence: .selection.reuse (4) = versioning/collaboration/asset-metadata/ localization-workflow — neutral concerns Tara should consume from the lower layer; each .dispositions[].migrationPath gives the consumer migration (e.g. reference SA versioning contracts, drop Tara-local version helpers). -->
    • S0.4.d Select consolidation where two implementations should move to one authoritative package. Evidence: target owner and removal sequence. <!-- evidence: .selection.consolidate = [] — NONE. With 0 exact duplicates + a clean one-way acyclic layer boundary (Tara → studio-authoring; SA does not import Tara), there is no consolidation debt: no two live implementations to collapse. A real selection outcome (empty + justified), coherence-gated against the 0-duplicate fact, not a deferral. -->
    • S0.4.e Retain separate logic when audience, domain semantics, trust, lifecycle, dependency direction, or release cadence differs. Evidence: reviewed separation rationale. <!-- evidence: .selection.separate (5) = creator-roles/editorial-lifecycle/ taxonomy-curation/authoring-blocks/templates — meditation-specialized (invitational language, seasonal registry, program cadence, plane taxonomy, review-ladder grades); differ by domain semantics + audience. Each row carries the retained Tara counterpart module + rationale. --> <!-- doc: evidence/v1-workbenches/tara-studio-authoring-overlap-selection.md -->
    • S0.4.f Check that the final package graph has no duplicate exported contract or dependency cycle. Evidence: boundary and duplication checks. <!-- evidence: .dependencyChecks — 0 duplicate exported contract names across studio-authoring (315) and tara-workbench (209); dependency direction is acyclic (Tara → studio-authoring; studio-authoring does NOT import Tara), cycle=false. -->
  • S0.5 Inventory shared contract packages and generated persistence registration. Evidence: authoritative owner is named for every shared type. <!-- partial: inventory + lineage + consumer graph + integrity (a-d) done for the Tara contract slice in evidence/v1-workbenches/tara-contract-inventory.json. Owner assignment (e) is a human decision and per-type owners are not yet named, so the parent stays unchecked. -->

    • S0.5.a Enumerate shared schemas, TypeScript/Python/Rust types, protobuf, OpenAPI, events, generated clients, database models, and registry rows. Evidence: contract artifact inventory. <!-- evidence: .contractSchemas (98 Zod schemas, 44 enums) + .inferredTypes (98 z.infer types) + .persistenceRows (22 objectContract entities) + driftGuardSpecs (4), all source-linked. Scoped to the Tara contract slice libs/contracts/src/tara (the shared types the workbench-kit extraction touches). generate-tara-contract-inventory.mjs, --check gate, tests. -->
    • S0.5.b Trace each generated artifact to its source schema and generation command. Evidence: source-to-generated lineage. <!-- evidence: .lineage — each of the 22 persistence rows resolves to its source Zod schema (sourceFound=true for ALL — 0 source-less rows) and the generation command (generate-contract-persistence-schema.ts → Prisma). -->
    • S0.5.c Trace each shared type to all producers, consumers, persisted tables, events, and public docs. Evidence: type consumer graph. <!-- evidence: .consumers — 14 real importers of @oshun/contracts/tara categorized (bff=6, domain-lib=5, web=3); producer = libs/contracts/src/tara, persisted tables = the 22 objectContract rows. -->
    • S0.5.d Identify duplicate, forked, unregistered, source-less, and generated-but-hand-edited types. Evidence: contract integrity findings. <!-- evidence: .findings — 0 source-less registry rows (healthy lineage); 76 unregistered schemas surfaced (mostly composed sub-schemas/enums, noted); the contracts↔domain-lib direction is recorded (contracts is the lower layer, the domain lib CONSUMES it — the enum "mirror" is reuse, not a fork). generated-but-hand-edited detection needs a live generation (see S0.5.f). -->
    • S0.5.e Assign one authoritative owner, source, namespace, version policy, compatibility window, and deprecation path per type. Evidence: zero ownerless shared types. <!-- human: owner/version-policy assignment is a governance decision. -->
    • S0.5.f Validate persistence/model registration from a clean generation and real database metadata. Evidence: generation drift and registration checks. <!-- runtime: needs a running database + Prisma generation (generate-contract-persistence-schema.ts / generate:v3-schema --check); not run on this 16GB Mac without the dev DB up. -->
  • S0.6 Record ADR-S1 for package ownership and dependency direction. Evidence: no proposed dependency cycle. <!-- partial: ADR-S1 recorded (Status: Proposed) at docs/adr/ADR-S1-workbench-kit-package-ownership.md with the context + graph snapshot (a) and options trade-off (b). The current graph is proven acyclic (S0.4). Parent stays unchecked: ownership assignment (c), the chosen direction's rules/validation (d,e), and owner approval (f) are open human decisions. -->

    • S0.6.a Define the decision questions and current package/consumer graph. Evidence: ADR context and graph snapshot. <!-- evidence: ADR-S1 §"Context and Problem Statement" (3 decision questions) + §"Current package and consumer graph" (snapshot table from the 7 committed evidence artifacts; direction proven acyclic). -->
    • S0.6.b Compare retain-current, extend studio-authoring, create workbench-kit, and other viable ownership options. Evidence: option trade-off table. <!-- evidence: ADR-S1 §"Considered Options" — 4-option trade-off table (retain / extend studio-authoring / new workbench-kit / staged hybrid), each row grounded in the evidence (studio-authoring has no UI; S0.11 shows 0/209 exports clear the reuse bar). -->
    • S0.6.c Assign ownership for domain-neutral contracts, UI, server boundaries, persistence helpers, testing, and domain plugins. Evidence: responsibility map. <!-- evidence: workbench-kit-boundary.json .responsibilityMap + ADR-S1 Decision Outcome — concern→layer: contracts→@oshun/contracts, neutral logic →@oshun/studio-authoring, neutral UI shell/primitives + server/testing helpers→@oshun/workbench-kit (proposed, S0.11-gated), domain semantics + plugins→domain packages, transport→BFF, composition→apps. Named-team owners are S0.6.f (human). generate-workbench-kit-boundary.mjs + 5 tests. -->
    • S0.6.d Specify allowed and forbidden dependencies among applications, shared kit, studio-authoring, domain libraries, contracts, BFF, and adapters. Evidence: normative dependency rules. <!-- evidence: .dependencyRules — 6 layers (contracts<shared-logic< workbench-kit<domain<bff<app); dependencies may point ONLY downward (15 allowed edges); upward + same-layer edges forbidden (21), as Nx-tag-style rules. -->
    • S0.6.e Validate the chosen direction with a cycle simulation and Nx/tag rule prototype. Evidence: passing boundary proof. <!-- evidence: .cycleSimulation — the proposed layered graph is ACYCLIC, cross-checked against the current observed graph proven acyclic in S0.4 (tara→studio-authoring→contracts). boundaryProofPasses=true; a deliberately inserted upward edge is caught as a cycle (regression test). --check gates it. -->
    • S0.6.f Obtain architecture and affected package-owner approval. Evidence: ADR approval record. <!-- human: named architecture + owner sign-off on ADR-S1 (now records the staged-hybrid direction as agent-adopted PENDING this named ratification). -->
  • S0.7 Record ADR-S2 for generic parameters and extension points. Evidence: examples cover content-authoring and operations-console archetypes. <!-- evidence: docs/adr/ADR-S2-generic-parameters-and-extension-points.md + evidence/v1-workbenches/workbench-kit-extension-model.json + the compiling reference evidence/v1-workbenches/workbench-kit-extension-model/ (both archetypes as COMPILING plugins). generator generate-workbench-kit-extension-model.mjs (--check gate + tsc --noEmit gate in verify), 9 tests. Named architecture ratification stays a human step. -->

    • S0.7.a Inventory variability in identity, stage, block, score, gate, artifact, preview, inspector, import, publish, and route semantics. Evidence: variability table. <!-- evidence: .variability — all 11 dimensions tabled, each with a mechanism + rationale; 6/11 ground to an S0.2-extracted neutral contract. -->
    • S0.7.b Define which variability belongs in generic type parameters, discriminated interfaces, plugin capabilities, composition slots, or domain-owned adapters. Evidence: extension mechanism mapping. <!-- evidence: .variability[].mechanism — each dimension mapped to exactly one of the 5 mechanisms (2 generic-type-param, 1 discriminated-interface, 1 plugin- capability, 3 composition-slot, 4 domain-adapter); all 5 mechanisms used. -->
    • S0.7.c Define invariants the shared kit owns and behavior plugins may not override. Evidence: normative core contract. <!-- evidence: .invariants (6) = authorization/audit/idempotency/honest-capability/ no-escape-hatch/deletion-threshold, each sourced to S0.3/S0.8/S0.11; encoded in kit-contract.ts enforceInvariants() + a closed MutationEnvelope. -->
    • S0.7.d Define optional capability discovery and honest unsupported/not- configured behavior. Evidence: capability negotiation examples. <!-- evidence: .capabilityModel — capabilities()/supports() discovery + typed CapabilityResult<T> ({configured:false,reason}); the ops plugin honestly declares canPublish=false (index.ts), proving negotiation is real. -->
    • S0.7.e Model one Tara-like content-authoring plugin without leaking meditation types into the shared API. Evidence: compiling example. <!-- evidence: workbench-kit-extension-model/plugin-content-authoring.ts — compiles under strict tsc --noEmit; leak check enforces 0 meditation/spark/plane/ cadence/invitational tokens in the CODE (comments-stripped). -->
    • S0.7.f Model one Aja/Bellona-like operations plugin with jobs, queues, resources, and host targets. Evidence: compiling example. <!-- evidence: workbench-kit-extension-model/plugin-operations.ts — compiles under strict tsc; ProcessingJob models host target + resourceUnits + queueDepth; disjoint entity/stage/artifact types from the authoring plugin. -->
    • S0.7.g Reject generic parameters or slots that have only one concrete use or permit unsafe arbitrary escape hatches. Evidence: abstraction challenge review. <!-- evidence: .rejections (3) = single-use domain generic (TMeditation*), arbitrary escape hatch (extraData:any/raw:unknown bypassing invariants), kit-level score() computing a domain number. --> <!-- doc: evidence/v1-workbenches/workbench-kit-extension-model/ (compiling reference) -->
  • S0.8 Record ADR-S3 for event, audit, idempotency, and transaction boundaries. Evidence: failure scenarios are diagrammed. <!-- evidence: docs/adr/ADR-S3-event-audit-idempotency-transaction-boundaries.md + event-audit-boundaries.json. a=mutation-effect-matrix, b/c=transaction- boundaries (34 local-ACID + 8 outbox, 0 saga + identity contract), d/e= delivery/audit semantics, f=7 mermaid failure diagrams (all coherence-gated). Named architecture ratification stays a human step. -->

    • S0.8.a Inventory shared commands and mutations that produce persistence, event, job, blob, provider, publish, notification, or audit effects. Evidence: mutation-effect matrix. <!-- evidence: .mutations — all 42 Tara BFF mutations classified across the 8 effect types (persistenceWrite 32, event/audit 34, provider 4, publish 3, blob 2, notification 2, job 0). crossBoundaryMutations (8) isolates the ones touching an external boundary (the S0.8 outbox/saga candidates); every write mutation audits (templated event names handled). generate-tara-mutation- effect-matrix.mjs + 6 tests, wired into inventory/verify. -->
    • S0.8.b Define authoritative transaction boundaries and when outbox, inbox, or saga behavior is required. Evidence: boundary decision table. <!-- evidence: tara-transaction-boundaries.json .boundaryDecisionTable — the standard transactional-outbox rule applied to the S0.8.a effect matrix: 34 local ACID transactions (persistence+audit only) + 8 transactional-outbox mutations (persistence then provider/publish/blob/notification), 0 sagas (no external effect must be atomic with persistence). Per-route pattern + rationale. generate-tara-transaction-boundaries.mjs + 5 tests. -->
    • S0.8.c Define command, idempotency, event, trace, audit, job, and external-operation identity and correlation rules. Evidence: identity contract. <!-- evidence: .identityContract — command id (randomUUID/server id; idempotency-capable routes carry a randomUUID command id), event id (audit event name + target id per write), correlation (per-request id through the guarded handler + audit), external-operation id (outbox row correlates the durable write with its async dispatch). Derived per-mutation from the matrix. -->
    • S0.8.d Define event ordering, delivery, duplication, replay, compatibility, and poison handling. Evidence: delivery semantics. <!-- evidence: event-audit-boundaries.json .deliverySemantics (6) — at-least-once, per-aggregate ordering, dedup-by-externalOperationId, durable replay, versioned compatibility, dead-letter poison. Over the 8 outbox mutations (S0.8.b). generate-event-audit-boundaries.mjs (--check gate), 8 tests. -->
    • S0.8.e Define audit atomicity, refusal/failure capture, immutability, redaction, retention, and evidence linkage. Evidence: audit semantics. <!-- evidence: .auditSemantics (6) — audit atomic-in-tx with the write (S0.8.b local-ACID), refusals/failures captured, append-only immutable, redact-at- write, RetentionPolicy-governed, evidence-linked actor→command→effect (S0.8.c). -->
    • S0.8.f Diagram failure before commit, after commit/before dispatch, duplicate delivery, worker loss, external success/local timeout, compensation failure, and replay. Evidence: reviewed failure diagrams. <!-- evidence: .failureScenarios (7) + docs/adr/ADR-S3-*.md mermaid sequence diagrams for each. compensation-failure is NOT APPLICABLE (sagaRequired=0 — coherence-checked); the only durability seam is after-commit/before-dispatch, closed by the durable outbox; dedup-by-externalOperationId makes 3/5/7 safe. --> <!-- doc: docs/adr/ADR-S3-event-audit-idempotency-transaction-boundaries.md -->
  • S0.9 Record ADR-S4 for UI composition and theming. Evidence: existing Oshun, Isis, Yemaya, Euterpe, and Bellona design systems are reconciled. <!-- evidence: docs/adr/ADR-S4-ui-composition-and-theming.md + workbench-kit-theme-contract.json + the workbench-kit-theme/ prototype. a=design-system-comparison, b=ui-composition-boundary, c=9-category token contract (6 mapped/3 introduced), d=scoped isolation prototype, e=4 archetype compositions (no clone), f=6-condition a11y matrix. All coherence-gated. Named architecture ratification stays a human step. -->

    • S0.9.a Inventory shells, navigation, layout primitives, tokens, themes, typography, icons, component libraries, overlays, and interaction patterns in all five systems. Evidence: design-system comparison. <!-- evidence: .systems — per-system token defs/uses, css/inline/styled approach, shell/layout components, theme signals, interaction primitives for all 5 (oshun/isis/yemaya/euterpe/bellona). Key finding: the 5 do NOT share a styling approach — token OWNERS oshun(600)/yemaya(67), token CONSUMER isis(87 uses, 0 local defs), non-token euterpe(inline)/bellona — so a uniform shared token contract (S0.9.c) is not adoptable as-is. generate-design-system-comparison.mjs + 7 tests, wired into inventory/verify. -->
    • S0.9.b Define shared shell/primitive responsibilities versus host/domain-owned composition, vocabulary, information architecture, and specialist visualization. Evidence: ownership boundary. <!-- ui-composition-boundary.json: 10 UI concerns assigned kit-owned (5: neutral app shell/nav, layout primitives, interaction primitives, the token CONTRACT/role-names, and the theme-injection seam) vs host/domain-owned (5: domain vocabulary, information architecture, domain composition, specialist visualization e.g. DAW/3D/mocap, and the token VALUES). Derived from the S0.6 layering fact (kit = neutral UI layer) + the S0.9.a divergence (oshun/yemaya own tokens, isis consumes, euterpe/bellona non-token — so token VALUES cannot be shared and stay host-owned; only the naming CONTRACT is neutral). generate-ui-composition-boundary.mjs + 6 tests, wired into inventory/verify + CI. S0.9.c token contract (creative), S0.9.d isolation prototype (code), S0.9.f a11y matrix (runtime) remain unchecked. -->
    • S0.9.c Define token contracts for color, typography, spacing, density, elevation, motion, focus, semantic states, and data visualization. Evidence: token mapping. <!-- evidence: workbench-kit-theme-contract.json .tokenContract — 9 categories of --wk-_ role names; 6 MAPPED to real oshun --l-_ tokens (color/typography/ elevation/motion/semantic-states/data-viz), 3 CONTRACT-INTRODUCES (spacing/ density/focus, which oshun does not tokenize). generate-workbench-kit-theme- contract.mjs (--check gate), 9 tests. -->
    • S0.9.d Define host theme injection without global CSS leakage, runtime class collisions, duplicated providers, or inaccessible overrides. Evidence: theming isolation prototype. <!-- evidence: prototype workbench-kit-theme/tokens.css — all tokens scoped to [data-workbench-theme] (0 --wk-_ in a bare :root block, checked), --wk-_ names not classes (no collision), single scoped provider (nesting re-scopes via the cascade), reduced-motion + forced-colors handled at the contract level so a host override cannot defeat them. .isolation (4 props) validated against the CSS. -->
    • S0.9.e Reconcile Oshun calm authoring, Isis operator, Yemaya/Euterpe creative production, and Bellona control-room density without forcing one visual clone. Evidence: archetype compositions. <!-- evidence: prototype workbench-kit-theme/archetypes.css — 4 archetypes (calm-authoring/operator/creative-production/control-room) consume the SAME contract with divergent density/motion/contrast values (≥2 distinct --wk-density, checked) — reconciled, NOT a visual clone. -->
    • S0.9.f Validate themes, densities, zoom, reduced motion, and forced colors with shared components. Evidence: cross-host visual/accessibility matrix. <!-- evidence: .a11yMatrix — 6 conditions (light/dark theme, compact/comfortable density, zoom, reduced-motion, forced-colors) each mapped to a contract mechanism ASSERTED PRESENT in the prototype CSS (data-density/rem/prefers-reduced-motion/ forced-colors). This validates the a11y CONTRACT coverage; live per-pixel cross-host screenshots are the runtime layer (noted honestly in the finding). --> <!-- doc: docs/adr/ADR-S4-ui-composition-and-theming.md + workbench-kit-theme/ prototype -->
  • S0.10 Record ADR-S5 for compatibility and incremental migration. Evidence: Tara can migrate without a flag day. <!-- evidence: docs/adr/ADR-S5-compatibility-and-incremental-migration.md + migration-plan.json. a=migration-surface register, b=additive coexistence seams, c=13-slice ordered list (S0.2 contracts), d=slice transition contract (no dual-write), e=mixed-version journey (routes/store/deep-links unchanged), f=milestone-gated retirement (S0.11). Tara migrates WITHOUT a flag day; all coherence-gated. Named architecture ratification stays a human step. -->

    • S0.10.a Inventory Tara public imports, routes, state stores, contracts, persisted data, events, generated artifacts, and current consumers. Evidence: migration surface register. <!-- evidence: .register — consolidates all 8 named dimensions from the six committed inventories (publicImports 209, routes 70, durableStores 1, contracts 98, persistedEntities 22, auditEvents 32, generatedArtifactRows 22, domainLibConsumers 1), each with source provenance; migrationRiskFlags surface 151 unused exports + 11 untested routes + 25 untested UI components. generate-tara-migration-surface.mjs + 7 tests, wired into inventory/verify. -->
    • S0.10.b Define additive compatibility seams, adapters, aliases, and version negotiation needed while old and new packages coexist. Evidence: coexistence architecture. <!-- evidence: migration-plan.json .coexistenceSeams (4) — adapter (S0.7 domain adapters), alias (re-export kit under Tara's names), version-negotiation (S0.7.d capability discovery), no-breaking-change (Tara keeps all 70 routes). generate-migration-plan.mjs (--check gate), 9 tests. -->
    • S0.10.c Slice migration by independently verifiable capability rather than package-wide replacement. Evidence: ordered migration slice list. <!-- evidence: .slices — the 13 S0.2 neutral contracts as a gapless 1..13 total order by risk: tier1 EvidenceGroundingPanel/SourceLibrary (Veritas 2nd consumer clears S0.11 first), tier2 the 0-signal neutral contracts, tier3 the mixed-split ones needing an adapter. Capability-sliced, not package-wide. -->
    • S0.10.d Define per-slice enablement, dual-read/write avoidance or reconciliation, rollback, evidence, and removal conditions. Evidence: slice transition contract. <!-- evidence: .sliceTransitionContract — lifecycle enable→verify→coexist→remove (flag-gated, rollback = keep Tara inline default, evidence = slice tests pass, removal = S0.11 ≥2 consumers). NO dual-write: UI extraction over the SAME Tara store (S0.8 keeps the store Tara-owned) — nothing to reconcile. -->
    • S0.10.e Define route/state/deep-link compatibility and preserve current Tara behavior for partially migrated users. Evidence: mixed-version journey plan. <!-- evidence: .mixedVersionJourney — 70 routes + 1 durable store + the /studio/tara-workbench/* deep-link namespace UNCHANGED; the migration swaps a component implementation, not a route/store/URL, so a partially-migrated user sees identical behavior (the S0.2 seam is the interaction contract). -->
    • S0.10.f Define the final contract period, legacy consumer migration, and old implementation deletion gate. Evidence: dated retirement plan. <!-- evidence: .retirementPlan — milestone-gated T0..T3 (T0=first Veritas 2nd consumer clears S0.11), contract period = per-slice until S0.11, single legacy consumer @oshun/bff re-points, deletion gate = S0.11 (≥2 consumers + passing tests). Absolute calendar dates are a named-ratification input (honest scope). --> <!-- doc: docs/adr/ADR-S5-compatibility-and-incremental-migration.md -->
  • S0.11 Define a deletion threshold: no shared abstraction ships with only one real consumer unless the reuse is already proven. Evidence: each exported module lists at least two consumers or an approved exception. <!-- partial: the threshold, definition, report, eligibility check, and CI regression (a,b,c,f) are implemented in evidence/v1-workbenches/deletion-threshold.json + deletion-threshold-gate.mjs. The parent stays unchecked because d (exception records) and e (delete/return one-consumer abstractions) are human decisions not yet executed — no shared kit has shipped, so 0/209 candidates currently clear the bar. -->

    • S0.11.a Define a real consumer as a production capability with distinct domain ownership or archetype, excluding tests, demos, fixtures, examples, aliases, and reexports. Evidence: consumer definition. <!-- evidence: .summary.realConsumerDefinition (REAL_CONSUMER_DEFINITION) — production import by a DISTINCT consumer project (domain/archetype); tests, fixtures, demos, aliases, re-exports excluded; identity = importing package. Implemented in countRealConsumers/realConsumersByExport, unit-tested. -->
    • S0.11.b Count consumers per proposed exported module and symbol from the authoritative import/capability graph. Evidence: export-to-consumer report. <!-- evidence: .report — per exported symbol, its distinct real consumer projects + realConsumerCount, computed from tara-import-graph.json (the authoritative import graph, S0.1). -->
    • S0.11.c Require two implemented consumers before extraction where behavior is not already domain-neutral and stable. Evidence: extraction eligibility check. <!-- evidence: .report[].eligibility (threshold=2): eligible / single-consumer / zero-consumer. Result: 0/209 Tara exports eligible (56 single, 153 zero) — none extractable as-is without a second-domain consumer or an exception. -->
    • S0.11.d For exceptions, record imminent second consumer, proven prior reuse, cost of delay, abstraction boundary, owner, and expiry. Evidence: approved exception record. <!-- human: an approved, time-boxed exception record per exception. -->
    • S0.11.e Delete or return one-consumer abstractions to the owning domain before shared-kit release. Evidence: zero unexplained one-consumer exports. <!-- human/impl: acts on S0.11.c results at shared-kit release; no kit shipped yet, so nothing to delete/return. -->
    • S0.11.f Recheck consumer counts on every shared export addition/removal in CI. Evidence: deliberate one-consumer regression test. <!-- evidence: deletion-threshold-gate.mjs --check runs in verify:inventory:v1-workbenches (CI) — drift + enforcement (exit 1 once a package is in SHARED_KIT_PACKAGES); deletion-threshold-gate.test.mjs has a deliberate single-consumer + test-only-excluded regression case. -->
  • S0.12 Ratify the shared capability matrix with domain owners. Evidence: review records list accepted, rejected, and domain-owned rows.

    • S0.12.a Prepare per-domain rows with capability semantics, proposed shared boundary, extension point, consumer, dependency, and evidence. Evidence: seven review packets. <!-- shared-capability-review-packets.json: SEVEN packets, one per owner S0.12.c names, each carrying all 13 neutral contracts from the S0.2 extraction matrix — 91 rows. Every row carries the six things this cell asks for and nothing it cannot know: semantics verbatim from the extraction matrix, the kit layer and the edges it may and may not use from the S0.6 boundary proof, the ADR-S2 variability dimension mapped to that contract, the consumer, the invariants adoption inherits, and source-linked evidence. THE COLUMN THAT MAKES A PACKET WORTH AN OWNER'S TIME is the consumer, and it is COMPUTED: does this domain already have something of that name? 52 rows ask a domain to REPLACE OR ALIGN something it has and 39 to ADOPT something new, and the two are different decisions with different costs — a matrix that did not separate them would ask seven owners the same question about materially different work. PipelineFunnel is the only contract all seven domains already have, which makes it the strongest extraction candidate and the one most likely to mean seven different things. AgentAutonomyBudget is the only one NO domain has, so it is Tara-only today and cannot clear the S0.11 two-consumer bar on this matrix alone. 42 of 91 rows carry no ADR-S2 variability dimension at all and ask the owner to accept the contract as-is or name the variation they need. ALL 91 DECISIONS ARE PENDING with a null decider, and --check exits non-zero if that ever stops being true. This cell PREPARES the review; S0.12.b-g are named owner signoffs, and an agent that filled them in would be forging the exact artifact the cell exists to produce. generate-shared-capability-review-packets.mjs + 24 tests, wired into inventory/verify + CI. --> <!-- The one judgment is the recognition VOCABULARY, so it is published per contract with its reason and is disputable by the owner reading it. Two of its terms were removed after the first run reported overlap that was not there, and both are now regression tests: library matched every symbol in @isis/3d-asset-library and reported a licence-assignment function as a catalogue browser — a fact about a DIRECTORY NAME, fixed by matching the symbol rather than the package-qualified behaviour string; and the bare verb assign fired on assigning a licence, a colour or a bitrate, none of which is reviewer assignment. Prisma MIGRATIONS are excluded beside the schema for the same reason: a migration is the history of a table and the schema is its current truth, so counting both inflates a domain's apparent overlap in proportion to how long it has existed. Terms match WORDS, not substrings — gate does not fire on aggregate, which is tested — and each row publishes its match count by capability kind and whether the matches are concentrated in one or two projects or scattered, so a weak match is visible as one. notDerivableFromSource says the thing that matters: a vocabulary finds capabilities CALLED the same, and whether they ARE the same is the judgment the review exists to make. An adopt row is an offer, not a gap. -->
    • S0.12.b Obtain Tara owner review as the extraction source and behavioral regression authority. Evidence: signed Tara comments.
    • S0.12.c Obtain Isis, Metis, Yemaya, Veritas, Euterpe, Aja, and Bellona owner decisions for their candidate rows. Evidence: seven named domain review records.
    • S0.12.d Record accepted shared capabilities with named consumers and invariants. Evidence: accepted row register.
    • S0.12.e Record rejected capabilities with rationale and any retirement or redesign action. Evidence: rejected row register.
    • S0.12.f Record domain-owned capabilities with authoritative package and explicitly forbidden shared semantics. Evidence: domain-owned row register.
    • S0.12.g Resolve conflicting owner decisions through product/architecture governance rather than averaging incompatible requirements. Evidence: zero unresolved matrix rows.

S1 — Package, contracts, and extension model#

  • S1.1 Scaffold libs/oshun/workbench-kit using repository ESM, Nx, TypeScript, lint, and Vitest conventions. Evidence: targeted lint, typecheck, test, and build pass. <!-- libs/oshun/workbench-kit, mirroring its sibling @oshun/studio-authoring (the other half of ADR-S1's layering) rather than inventing a shape. All five targets pass on this branch: lint, typecheck, test (23), build, and a smoke target that consumes the BUILT package. -->

    • S1.1.a Create the package manifest, project configuration, source/test roots, README, ownership metadata, and initial public entry point. Evidence: scaffold file review. <!-- package.json / project.json / src / README with the ownership table (layer, scope, owning ADRs, ledger cells, and S0.6.f approval recorded as PENDING rather than implied). The initial entry point is the SIX kit-owned invariants from ADR-S2, as a runtime-checked closed mutation envelope rather than a document a plugin is trusted to have read — the load-bearing one being no-escape-hatch, since every other invariant is defeated by a single raw/any field, so the envelope refuses UNKNOWN keys as well as missing ones. That is the part of the kit that does not wait on S0.12: the export map (S1.2) is drawn from the RATIFIED capability matrix and all 91 of its rows are still pending, so shipping a capability now would prejudge the decision the review exists to make. The rules are restated in source (a library must not read an audit artifact at runtime) and a test asserts they still equal the artifact verbatim, so the restating cannot drift. -->
    • S1.1.b Configure ESM and TypeScript module resolution, project references, declaration output, source maps, and repository aliases. Evidence: emitted declaration and import inspection. <!-- type: module; moduleResolution bundler; tsconfig.json references lib+spec; tsconfig.lib.json emits declaration + declarationMap; the alias @oshun/workbench-kit is in tsconfig.base.json. Emitted output inspected: index.js/.d.ts/.js.map/.d.ts.map + invariants.*. -->
    • S1.1.c Configure Nx tags, named inputs, cacheable lint/typecheck/test/build targets, and affected-file inputs. Evidence: project target inspection. <!-- tags scope:oshun / layer:workbench-kit / type:lib — the layer tag is what makes ADR-S1's direction enforceable by a boundary check instead of by review. Targets: build (@nx/js:tsc with declared outputs, so it caches), lint, test, typecheck, plus smoke dependsOn build. -->
    • S1.1.d Inherit repository lint/format rules without blanket local disables. Evidence: lint rejects a deliberate invalid fixture. <!-- No local eslint config and no file-level disables: the package inherits the root flat config. Proven both ways — the real source lints clean, and a deliberate invalid fixture (any params + an unused local) was rejected with 1 error and 2 warnings before being deleted. A lint that passes everything is indistinguishable from a lint that is not running. -->
    • S1.1.e Configure Vitest environment, setup, coverage boundaries, and test-only entry points. Evidence: sample unit and coverage run. <!-- vitest.config.ts: node environment, root-pinned, include src/**/*.{test,spec}.ts, v8 coverage. passWithNoTests is FALSE in the Nx target, so an empty suite fails rather than reporting green. 23 tests cover every invariant branch, including the ones that make the closed envelope worth having: a truthy value where a boolean verdict belongs, an allowed decision that names no grant, a refusal with no reason, expectedRevision undefined vs explicit null, and the same consumer twice counting as one. -->
    • S1.1.f Configure deterministic build output and validate consumption from an external ESM fixture. Evidence: packed-consumer smoke test. <!-- tools/packed-consumer-smoke.mjs, wired as its own Nx target that dependsOn build. THIS CAUGHT A REAL DEFECT the other four targets could not: the manifest declared main/types/exports as ./src/index.ts (copied from the sibling, where nothing exercises them), so the built package was one Node could not import at all. Lint, typecheck, test and build were all green over it. The entries now name the emitted output and the manifest records why; in-workspace consumers resolve through the tsconfig path alias and never read those fields. The smoke test imports the emitted entry, checks every declared export is there, and then EXERCISES it — a valid envelope accepted and an escape-hatch field refused — because an emitted module that enforced nothing would pass an export check. -->
    • S1.1.g Run the four targeted targets sequentially on the current branch. Evidence: archived passing command output. <!-- lint PASS, typecheck PASS, test PASS (23), build PASS, run one after another via npx nx <target> @oshun/workbench-kit; smoke PASS on top. Archived as a repeatable TARGET rather than a transcript: a pasted log proves a machine once, and pnpm nx smoke @oshun/workbench-kit proves it every time. --> <!-- CORRECTION (S1.4): the typecheck PASS recorded above and in S1.1.b was VACUOUS. tsc --noEmit against a solution-style tsconfig with empty files/include compiles nothing and exits 0 — the target was green over a deliberately injected const x: number = 'nope'. Fixed in project.json to tsc --noEmit -p tsconfig.lib.json && tsc --noEmit -p tsconfig.spec.json, and verified to FAIL on an injected error in src and again in a spec. The four targets do now all check something; three of them always did. -->
  • S1.2 Add the package export map without deep-import escape hatches. Evidence: public API test rejects unexported imports. <!-- evidence/v1-workbenches/workbench-kit-export-map.json, generated and gated by scripts/v1-workbenches/generate-workbench-kit-export-map.mjs (--check wired into verify:inventory:v1-workbenches + CI, 31 tests). The escape hatch this guards against is not mainly ./src/* in a manifest — it is the slower one where the kit grows an entry point nobody ratified, because editing exports is easier than getting seven owners to agree. So the approved set is DERIVED and the manifest is checked against it. -->

    • S1.2.a Enumerate approved root and feature subpath entry points from the ratified shared capability matrix. Evidence: public API list. <!-- .publicApi — three entry points, each naming what authorizes it. A capability subpath exists only when the S0.12 matrix has ACCEPTED that contract for >=2 domains (S0.11: one domain accepting a shared contract is a domain abstraction with extra steps). All 91 rows are pending, so the capability-derived set is EMPTY and the map contains only what a kit-owned ADR decision authorizes: the ADR-S2 invariants, the invariants subpath, and the test corpus. The gate fails the moment a row is accepted with no entry point, OR an entry point exists with nothing authorizing it — both directions, because a map that only caught the second would let ratified capabilities go unshipped silently. -->
    • S1.2.b Map each entry point to ESM implementation and types without wildcard or source-directory exports. Evidence: manifest review. <!-- checkManifest: exports must EQUAL the derived set (both directions), no * key, no bare-string target (which resolves no types), import ends .js and types ends .d.ts, and no target that is TypeScript source or reaches into a src/ directory. That last check turned up a real defect: .d.ts ends with .ts, so the first version flagged every conforming entry point — the one-character difference is now its own test. The build had to change with it: @nx/js:tsc emitted under dist/.../src/, so a manifest could only name ./src/index.js; project.json now sets the executor's rootDir and output is flat. -->
    • S1.2.c Separate browser, React, server, contract, and test-only entry points so incompatible runtimes cannot import one another. Evidence: runtime boundary tests. <!-- RUNTIME_CLASSES: a six-class lattice (contract/neutral/browser/ react/server/test) with an explicit may-import set, forbidden module specifiers and forbidden ambient globals, enforced over the REAL module graph via the TypeScript AST — transitively, so a violation two hops from the entry point is still the entry point's. The two edges that matter are browser-may-not-reach-server and its mirror. Today's kit occupies two classes (contract, test), so the lattice is proven on deliberate violating fixtures as well: a browser entry reaching node:fs, a server entry touching document, a contract entry importing react, and a browser->server edge, each fed a MIXED graph so a gate that rejects everything is distinguishable from one that rejects the right thing. The DOM-global check is a documented approximation (free identifiers, not scope resolution) and both its directions — flagged bare document, unflagged ctx.document/local/parameter/object-key — are tested. -->
    • S1.2.d Reject internal source paths and undeclared subpaths in workspace and packed-package consumer fixtures. Evidence: negative import tests. <!-- BOTH resolution modes, because they are different mechanisms. Packed: tools/packed-consumer-smoke.mjs builds a real install (a temp dir whose node_modules/@oshun/workbench-kit symlinks to dist) and requires ERR_PACKAGE_PATH_NOT_EXPORTED for ./src/index.js, ./invariants.js, ./testing.js, ./package.json and a never-existed subpath — Node's exports resolution answers, not this script. Proven by mutation: adding ./* to the built manifest made all five resolve. Workspace: an in-workspace consumer NEVER READS package.json at all, so a sealed manifest plus one @oshun/workbench-kit/* alias is a package that is closed to outsiders and wide open to the seven domains that use it — checkAliases + public-api.spec assert the alias table has no wildcard and equals the approved specifiers exactly. Also proven by mutation: adding the wildcard alias to tsconfig.base.json failed the gate and two spec cases. The sibling @oshun/studio-authoring ships exactly that alias, which is why absence is asserted, not assumed. -->
    • S1.2.e Import every declared subpath from both workspace and built package modes. Evidence: positive public API matrix. <!-- Workspace: public-api.spec imports every subpath from the artifact (not a hand-written list) through the tsconfig aliases, which vitest.config.ts derives from tsconfig.base.json rather than duplicating — a duplicated alias table would let the spec pass against aliases that do not exist, which is the one thing it is for. Packed: the smoke test imports all three from the symlinked install. Both then assert the root and the /invariants subpath are the SAME module instance, because two copies is how instanceof KitInvariantError starts failing for one consumer and not another with no message that explains why. Ordering gotcha found live: Vite alias matching is first-prefix-wins, so a root-first table rewrote /invariants to .../index.ts/invariants; the aliases are sorted longest-first. -->
    • S1.2.f Scan production bundles for test builders, fixtures, server code, and unintended side effects. Evidence: bundle inspection. <!-- Over the EMITTED dist, not source. (1) No production entry point's emitted graph may reach the test-only module — this is the check that stops an actor-decision forger shipping in a bundle, and it is why testing is a separate entry point rather than part of the root. (2) No production module may do work at import time, which is what sideEffects: false promises the bundler that will drop it; deliberately stricter than the ECMAScript notion (any top-level call, including one that only computes a constant), stated as such, and the test-only entry is exempt because it is never in a production bundle to be dropped from. Both proven by mutation on the real package: making index.ts re-export a builder produced the leak violation from two checks, and adding one computed const to invariants.ts failed the scan. That second run reported the same defect twice — invariants is reached from both production entry points — so the scan now dedupes by module; one defect reported twice reads as two defects. --> <!-- Server code in a production bundle is covered by the S1.2.c lattice rather than a second scan: server is a runtime class, so a production browser/react entry that reached one would fail forbidden-class-edge and forbidden-node-builtin. No server entry point exists today. -->
  • S1.3 Add path mapping and dependency tags. Evidence: Nx boundary checks pass. <!-- evidence/v1-workbenches/workbench-kit-dependency-constraints.json + generate-workbench-kit-dependency-constraints.mjs (--check in verify:inventory:v1-workbenches + CI, 20 tests), plus two live Nx targets: alias-resolution (tsc) and boundary-fixtures (the real ESLint). S0.6 DECIDED the layering; the gap between decided and enforced is where layering dies, so this cell measures that gap instead of asserting it closed. HEADLINE: of the six S0.6 layers, layer:bff has ZERO carrier projects, so a constraint naming it passes every build without ever matching an import — the report says unenforceable-no-carriers for those two edges rather than counting them enforced. layer:shared-logic had zero too until @oshun/studio-authoring was tagged as the layer S0.6 names it. -->

    • S1.3.a Add canonical TypeScript path aliases for approved package entry points only. Evidence: resolution tests. <!-- tools/alias-resolution.mjs, its own Nx target. The S1.2 spec matrix runs under Vitest, which resolves through VITE's alias plugin — a different resolver from the tsc every consumer actually uses, and the two disagree in ways that matter (Vite takes the longest prefix, TypeScript matches paths by specificity then checks the file exists). So this asks TypeScript, from a consumer position, both ways: every approved specifier resolves to the exact file its alias names AND a generated consumer program importing all three typechecks clean; every deep specifier (/src/..., an emitted-file path, the pre-S1.2 layout, a plausible internal module) does NOT resolve. Proven by mutation — re-adding @oshun/workbench-kit/* made two of the six resolve straight into src. The probe config EXTENDS tsconfig.base.json so the alias table under test is the repository's, not a copy. -->
    • S1.3.b Add package scope, type, domain, and layer tags consistent with the repository taxonomy. Evidence: project metadata review. <!-- scope:oshun / layer:workbench-kit / type:lib / domain:neutral. The domain tag is the one worth arguing: the kit is domain-NEUTRAL, and saying so as a tag is what lets the claim be checked rather than believed. The gate publishes a census over all 3502 tracked project.json files: layer:contracts 19, layer:domain 1790, layer:app 11, layer:shared-logic 1, layer:workbench-kit 1, layer:bff 0. A layer with no carriers is reported enforceable: false — the failure mode this cell exists to surface. What is NOT derivable is where the other 40-odd legacy layer tags (layer:service 229, layer:infra 124, layer:studio…) belong in the six-layer model; that is an architecture decision, not a lookup, and the summary says so. -->
    • S1.3.c Define allowed edges from hosts, BFFs, domain packages, studio-authoring, contracts, shared kit, and adapters. Evidence: dependency constraint table. <!-- All 15 S0.6 allowed edges replayed against the depConstraints eslint.config.js actually exports — IMPORTED, not regexed, because a flat config builds its rules in JavaScript and a regex would agree with a comment while disagreeing with the code. THIS FOUND A REAL BLOCKER: layer:domain -> layer:workbench-kit and -> layer:shared-logic were both REFUSED, so a domain package could not import the kit at all — the layering was decided and unusable, which is precisely what S1.3.f means by resolving violations in proposed consumers. Both are now in the layer:domain allow-list and enforced. -->
    • S1.3.d Forbid app-internal imports, domain-to-host reverse edges, browser-to-server imports, server-to-React imports, and private deep imports. Evidence: boundary rules. <!-- All 21 S0.6 forbidden edges replayed the same way. The kit's own rules are a PAIR — allow-list (contracts, shared-logic) plus deny-list (domain, bff, app) — because a single allow-list is defeated by the one edit anybody would make. Neither names its own layer, so the same-layer edge S0.6 also forbids is refused: two peer kits importing each other are one kit with a seam nobody can cut. Browser-to-server and server-to-React are NOT expressible as Nx tags (tags are per PROJECT and these are per MODULE within one package) — they are the S1.2.c runtime lattice, enforced over the module graph, and this cell does not double-count them. Private deep imports were already refused by Nx's relative-import rule, which the fixtures confirm rather than claim. THREE PRE-EXISTING DIVERGENCES are reported, never folded into "enforced": layer:app/contracts/domain each permit their own layer, which S0.6 forbids. Closing that is a 1820-project migration, not a cell in S1.3, so the gate names and counts them on every run and holds only the layers Phase S introduced to the full bar. -->
    • S1.3.e Add deliberate allowed and forbidden consumer fixtures for each dependency class. Evidence: rule tests reject only forbidden edges. <!-- tools/boundary-fixtures.mjs writes five real TypeScript files into the package, runs the repository's real ESLint over them, and deletes them in a finally with a preflight sweep and a .gitignore entry for a killed run. BOTH halves are asserted: 2 permitted edges must be ACCEPTED and 3 forbidden ones REJECTED, because a rule that also refuses the permitted edges is a rule nobody can adopt. Proven by mutation: removing the new constraints made two of the three forbidden edges pass. THE CHECK CAUGHT ITS OWN VACUITY. Excluding the fixture directory from tsconfig.lib.json looked tidier and silently defeated everything: typescript-eslint's project service could not find the files, every rule was replaced by a parsing error, and the boundary rule never ran — all three forbidden fixtures started "passing". The tool now fails on a parsing error naming that exact cause, and the tsconfig comment records why the directory stays included. -->
    • S1.3.f Resolve all violations in proposed consumers without disabling the constraint. Evidence: clean targeted boundary check. <!-- The two real violations were layer:domain -> layer:workbench-kit and -> layer:shared-logic (S1.3.c). Resolved by adding the two layers to the layer:domain allow-list — the edge S0.6 ratified — rather than by disabling the rule, adding an eslint-disable, or allow-listing the package name. nx lint is clean on @oshun/workbench-kit, @oshun/studio-authoring and @oshun/domain-tara (the retagged and newly-permitted projects), and the gate reports 0 contradictions with 22 edges enforced. --> <!-- Still open by design: 11 of the 36 S0.6 edges are unenforceable-no-carriers because layer:bff has no tagged project. That is a fact about the repository, not a gap in this cell, and it is published on every run rather than rounded up. -->
  • S1.4 Define WorkbenchId, TenantId, ActorRef, ArtifactRef, RevisionRef, and StableAnchor branded types. Evidence: invalid cross-assignment fails type tests. <!-- libs/oshun/workbench-kit/src/identity.ts + identity.spec.ts (82 tests), published as the ./identity entry point (S1.4 is what authorizes it in the S1.2 derived export map) and re-exported from the root. THREE DECISIONS shape the module and each is stated in source: parsing is STRICT and normalising is EXPLICIT; there is NO unsafe cast; parsers return RESULTS rather than throwing (unlike ./invariants, because a malformed id at a boundary is expected input while an unauthorised mutation is a bug). FOUND AND FIXED A REAL DEFECT ON THE WAY IN: the typecheck target ran tsc --noEmit against a solution-style tsconfig with empty files/include, so it compiled ZERO files and was green over a deliberate const x: number = 'nope'. S1.1.b/g recorded it as PASS; it passed because it did nothing. Now tsc --noEmit -p tsconfig.lib.json && -p tsconfig.spec.json, proven to fail on an injected error in src AND in a spec. S1.4.f depends on that being true. -->

    • S1.4.a Define canonical string format, namespace, parser, serializer, equality, and validation for WorkbenchId and TenantId. Evidence: valid/invalid/round-trip fixtures. <!-- <namespace>:<segment>, segment 1-63 lowercase alphanumerics with internal hyphens, over seven namespaces (tenant/workbench/actor/artifact/ entity/revision/block). LOWERCASE-ONLY IS THE LOAD-BEARING CHOICE: it is what lets equality be byte equality instead of a case-folding function whose locale behaviour seven codebases would each get subtly wrong. So parseTenantId('tenant:ACME') FAILS with its own code (not-canonical-case) rather than folding — a parser that quietly lowercased would mean two strings both "are" the tenant and neither is the one in storage. normalizeIdCandidate does the folding, reports every change it made, and is opted into at an input boundary; an HTTP form may use it, a stored id must not, and the report is what lets the two callers differ. formatId deliberately returns string (widening), because a value going into a URL or a column typed as branded invites skipping the parse on the way back. 10 rejection codes, each tested. -->
    • S1.4.b Define ActorRef for human, delegated, service, system, and redacted identities without using display data as identity. Evidence: actor variant tests. <!-- Five variants, closed-shape parser. "Without using display data as identity" is enforced TWO ways rather than asserted: structurally, since no email-shaped string can satisfy actor:<segment>, so the prohibition needs no denylist of address formats; and by the closed shape, which refuses unknown keys and gives name/email/avatarUrl/... their own code (display-data-in-identity) with the reason — a name changes, an email is reassigned, and either used as identity turns an audit trail into a guess. delegated carries BOTH actorId and onBehalfOf because an audit row with one cannot answer either "who did this" or "whose authority was used"; authorityOf returns the PRINCIPAL for it, which is the question every authorization check is really asking. A delegation to oneself is refused: it records an authority chain of length zero while claiming one exists. redacted exists so an erasure leaves the audit row intact — a nullable actorId cannot distinguish "erased" from "never recorded" — and authorityOf returns null for it, not the surrogate, because the surrogate identifies the ROW. -->
    • S1.4.c Define ArtifactRef with artifact identity, immutable byte version/hash, media/profile, and resolution semantics. Evidence: artifact reference tests. <!-- Pinned and floating are SEPARATE TYPES, not one type with a nullable digest, so "did we pin this?" is a question the compiler answers. That is the resolution semantics: resolutionContract reports deterministic/cacheable/signable, and a floating ref is none of the three — signing one commits to nothing and caching one caches a lie. Media is REQUIRED on a pinned ref (the bytes are fixed, so what they are is knowable) and nullable on a floating one. The digest carries its algorithm rather than assuming sha256, so adding a second is a widening instead of a silent reinterpretation of old rows. Uppercase hex is rejected for the same reason uppercase ids are. -->
    • S1.4.d Define RevisionRef with entity/workbench scope, revision ID, schema version, and optional parent/base. Evidence: revision reference tests. <!-- tenant + workbench + entityId + revisionId + schemaVersion, and parent/base as two fields because they answer different questions and diverge exactly when it matters: after a rebase, parent is where the content came from and base is what the change was expressed against. A three-way merge needs both; a model with one silently picks a side. Both must be a revision id or explicitly NULL — undefined is refused, because it makes "first revision" and "we did not record it" the same value. A self-PARENT is refused (a lineage cycle) while a self-BASE is ALLOWED (an empty diff, a real thing to express); the asymmetry is deliberate and tested. Wrong-kind ids are caught in all four slots. -->
    • S1.4.e Define StableAnchor variants for entity, field, block, range, temporal, graph, canvas, and hierarchy targets with revision semantics. Evidence: anchor variant tests. <!-- All eight, each carrying the RevisionRef it was TAKEN AT — that is the revision semantics, and without it an offset into a document nobody can date is a number, not an anchor. It is what makes AnchorResolution expressible: resolved / rebased-with-confidence / orphaned with CAP.11.e's five causes kept distinct, because the product answers differ (deleted = the comment is history; permission-hidden = this reader may not know it exists; ambiguous = a human could rescue it but nothing may move it automatically). Real domain content rather than a shape: FRAME RATES ARE RATIONALS, since 29.97 is 30000/1001 and comparing floats is how a note lands one frame off after a JSON round trip — and unreduced rates are refused so equality is field equality. Ranges are half-open over CODE POINTS, not UTF-16 code units, which can split a surrogate pair into a range that cannot render; start === end is a caret, not an error. A fractional frame or sample position is refused (rounding it silently moves the anchor) while a fractional seconds position is fine. A port anchor MUST name its owning node, because port ids are only unique within a node; a node anchor must NOT carry ownerId, or somebody will read it as meaningful. Paths are ARRAYS so a segment containing a dot cannot split the path. -->
    • S1.4.f Prohibit unchecked string casts and cross-kind ID assignment at compile time and parsing boundaries. Evidence: type-negative and runtime tests. <!-- COMPILE TIME: a declare const ... : unique symbol that is never defined and never exported, so a consumer cannot name it and therefore cannot construct a branded value. Seven @ts-expect-error blocks are the assertions — an UNUSED @ts-expect-error is itself a TypeScript error, so a brand that stops working fails nx typecheck on the line that expected it to work. PROVEN BY MUTATION: making the brand optional broke one directive, removing it entirely broke three. idEquals uses NoInfer on its second parameter so comparing a WorkbenchId with a TenantId is an error rather than inferring the union and comparing anyway. PARSING BOUNDARY: wrong-namespace is its own code — a tenant id used as a workbench id would address a different object with the same segment — and there is no exported unsafe cast at all, because an escape hatch is used exactly once and then everywhere. S1.2.f ALSO CHANGED THE CODE rather than being worked around: the strict import-time-work scan flagged three module-level new Set([...]) constructions, so they are frozen arrays and the module is genuinely inert at import — which is what sideEffects: false already promised. -->
  • S1.5 Define versioned envelopes for commands, queries, events, job updates, and capability probes. Evidence: golden wire fixtures. <!-- libs/oshun/workbench-kit/src/envelopes.ts, published as ./envelopes. Evidence is evidence/v1-workbenches/workbench-kit-envelope-goldens.json — 57 cases, 22 that must parse and 35 that must be refused with a NAMED CODE — generated by generate-workbench-kit-envelope-goldens.mjs (11 structural tests) and consumed by envelopes.spec.ts (77 tests) running the real parsers. The corpus is committed JSON rather than built by the code under test, because a test that constructs an envelope with the same helpers the parser uses proves the two agree with EACH OTHER and says nothing about the bytes another producer sent. THE CORPUS EARNED ITS KEEP ON THE FIRST RUN, finding three real defects: the extension-key pattern rejected isis.retryHint (it forced kebab-case on the name segment, so every realistic extension would have been refused); a succeeded outcome with no output field at all reported "ArtifactRef must be an object", sending the reader to the wrong field; and older-compatible had no coverage at all. MUTATION-TESTED three ways — dropping the deadline invariant, the namespace check, and the terminal-checkpoint rule each fails exactly the one case that claims to cover it. -->

    • S1.5.a Define common envelope ID, kind, schema version, timestamp, actor/tenant/workbench scope, correlation, trace, and extension fields. Evidence: common wire schema. <!-- Ten common fields, closed shape, ONE open field. extensions is the whole compatibility story: every other field is refused if undeclared (the same no-escape-hatch rule ./invariants.ts enforces on mutations), which would normally make forward compatibility impossible — a producer on a newer schema adds a field, an older reader rejects it, and the rollout is a flag day. Keys are <vendor>.<name>, the kit never reads them, and a key that SHADOWS a declared field is refused because x.actor reads as authoritative in a log and is validated by nobody. That combination is what lets versionCompatibility() tolerate a NEWER envelope: if it parses under the current shape the producer only added extensions, so nothing this reader needed was dropped; if it does not, the addition was a real field, and honouring a field you do not implement is worse than refusing the message. No hand-maintained version ceiling to bump per producer release. correlationId AND traceId are both present and are not the same thing: a correlation id is a BUSINESS transaction that survives retries and appears in the audit record; a trace id is an OBSERVABILITY span tree that is sampled and may be absent. Using the trace for correlation means "show me everything in this publish" silently returns a subset whenever tracing is off, so traceId is nullable and correlationId is not. -->
    • S1.5.b Define command envelope for idempotency, expected revision, deadline, authorization intent, and typed payload. Evidence: command golden fixtures. <!-- 17 golden command cases. authorizationIntent is the grant the caller CLAIMS to need — an intent, not a verdict; the verdict is ./invariants.ts's ActorDecision. Keeping them apart is what makes commandAuthorizationMismatch() possible: a command asking for workbench:read that was authorised under workbench:publish is privilege confusion, and one combined field cannot notice it. A deadline at or before occurredAt is refused as a CLOCK FAULT rather than accepted as an expired command, because treating it as a timeout hides the fault. expectedRevision must be a revision or explicitly null — undefined is the ambiguity that makes a lost update silent. -->
    • S1.5.c Define query envelope for typed filters, pagination, consistency, locale, field projection, and payload. Evidence: query golden fixtures. <!-- Cursor and offset pagination are separate SHAPES, so a query carrying both is structurally impossible rather than caught by a check — a shape with both has no defined meaning and whichever the server reads first becomes the answer. bounded-staleness REQUIRES its maxStalenessSeconds, because without the number it is "eventual" with a name that sounds like a guarantee. pageSize is bounded at 200, since an unbounded page size is how one read becomes a full table scan behind an innocent-looking query. locale is REQUIRED and BCP-47: "the server's locale" is nobody's locale and produces dates no reader asked for. projection is [] for the default rather than omitted, so there is no third state. -->
    • S1.5.d Define event envelope for producer, subject, sequence, causation, correlation, delivery compatibility, and typed payload. Evidence: event golden fixtures. <!-- producer is separate from actor because a retry by a different component has the same actor. delivery is carried rather than assumed so a consumer knows whether it must deduplicate instead of guessing and being wrong once. dedupeKey() is (subject, sequence), NOT envelopeId: a redelivery may legitimately carry a fresh envelope id while a genuine second event never reuses a sequence for one subject — tested with a redelivery that must produce the same key. Self-causation is refused because a causation walk over it would not terminate; causationId must be an id or explicitly null, since undefined cannot distinguish "externally triggered" from "we lost the chain". -->
    • S1.5.e Define job update envelope for job/attempt/stage identity, progress, checkpoint, usage/cost, output/error, and timestamp. Evidence: job-update golden fixtures. <!-- 12 golden cases. COST IS INTEGER MINOR UNITS with an explicit ISO-4217 currency, because a float cost accumulates error over a million updates and then disagrees with the invoice. progress is 0..1 or explicitly null (indeterminate) — a percentage and a fraction in one field is how a bar shows 4500%. A TERMINAL outcome may not carry a checkpoint: a resumable-looking terminal update is how a finished job gets run twice. A succeeded outcome must name its output (a success with nothing to show cannot be told apart from a job that silently produced nothing) and may not report progress below 1. A failed outcome must state retryable, because a caller cannot infer that from an error code and guessing wrong either loses work or retries forever. isProgressRegression() deliberately returns false across ATTEMPTS (a retry legitimately starts over) and for a null (a producer saying it no longer knows). -->
    • S1.5.f Define capability probe envelope for capability/version, dependency status, state, freshness, limitations, and safe remediation. Evidence: probe golden fixtures. <!-- "SAFE remediation" is made STRUCTURAL rather than asserted: a closed seven-action enum plus an https docs URL and a requiresOperator flag, with no prose field at all. Free text cannot be made safe — the first operator to paste a working curl command puts a bearer token in a probe response that gets logged — and the golden corpus includes a docsUrl carrying api_key=sk-live-... that must be refused. Two consistency rules make the answer worth reading: a configured capability may not offer remediation (if something needs fixing the state is degraded), and a non-configured one may not offer none ("broken, and there is nothing you can do" is almost never true and is never actionable). freshnessSeconds 0 means never-cache, which a nullable field could not say; isProbeFresh() also refuses a clock that went backwards, because that is a fault rather than freshness. CAPABILITY_STATES is ONE constant shared with S1.6 so the probe and the state record cannot drift into disagreeing about what states exist. -->
    • S1.5.g Test malformed, unknown-kind, older/newer compatible, unsupported-version, unauthorized, and oversized envelopes. Evidence: negative/compatibility fixtures. <!-- All seven classes declared; SIX covered by 35 cases and older-compatible reported UNREACHABLE with its reason: MIN_SUPPORTED_SCHEMA_VERSION === CURRENT_SCHEMA_VERSION, so version 1 has no predecessor and a fixture claiming to cover it would have to invent a version this build does not support. That exemption EXPIRES on its own — the generator stops declaring it unreachable the moment the two constants diverge, and the uncovered-class gate then fails. Oversized is checked on the TEXT before parsing, because a 4 MB envelope should be refused before it becomes 4 MB of objects; the 256 KiB limit is the smallest payload cap among the transports these envelopes cross, so fitting here means fitting everywhere. The fixture describes the oversized case (fillerBytes) rather than committing 256 KiB of JSON, because a fixture nobody can open is a fixture nobody reviews. THE S1.3 ALIAS CHECK CAUGHT A CONTRACT VIOLATION HERE: the byte count used TextEncoder, an ambient global, in a module whose declared runtime class is contract — importable from every runtime, depending on nothing that is not. A consumer compiling with types: [] had no declaration for it. Replaced with utf8ByteLength(), eight lines, cross-checked against Node's TextEncoder in the spec including a 4-byte emoji that a naive charCodeAt walk would count as 6. -->
  • S1.6 Define discriminated configured, not_configured, degraded, unavailable, and failed capability states. Evidence: exhaustive switch tests. <!-- libs/oshun/workbench-kit/src/capability-state.ts + 54 tests, published as ./capability-state. ./invariants.ts gave honest-capability a minimal TWO-state shape so the rule could ship with the scaffold; this is the full vocabulary, and the reason it needs five is that the three in the middle are the ones a caller gets wrong. The state NAMES and the Remediation shape are IMPORTED from ./envelopes.js rather than redeclared — a transport module owning a domain vocabulary looks backwards and is deliberate, because S1.5 shipped the probe envelope first and two lists of five strings in one package would eventually disagree about what states exist. -->

    • S1.6.a Define common capability identity, checked-at time, freshness, dependency, owner, limitation, remediation, and safe-detail fields. Evidence: common state schema. <!-- Nine common fields on every state. safeDetail is not a naming convention: parseSafeDetail() runs the same credential scan over the detail, every limitation, every dependency detail and the remediation URL, because the failure mode is convenience rather than malice — somebody pastes the request that reproduced the problem, headers included, into a field a dashboard renders and a log aggregator retains. owner.contactPath must be an https rota or runbook URL and NOT an email address, for the same reason S1.4.b keeps display data out of identity: an address goes stale and names a person rather than a responsibility. One consistency rule is stated once and applies to all five: only configured may offer remediation none. -->
    • S1.6.b Define configured only when required configuration, dependencies, permissions, and supported versions are verified. Evidence: configured-state tests. <!-- configured carries four VERIFICATION RECORDS, not a boolean, and the parser refuses the state unless all four subjects are covered. The four fail differently and independently — credentials present but the account lacks the scope, every dependency up but the provider is a major version ahead, config complete but the key is for the wrong tenant — so a single verified: true cannot say which one was checked, which means it cannot say which one was SKIPPED. Each of the four has its own rejection test. A duplicate subject is refused (it hides which check the second entry reported), and so is a configured state with a dependency that is not itself configured: a capability is not more available than the things it needs. -->
    • S1.6.c Define not_configured with missing configuration category and authorized resolution path but no secret values. Evidence: not-configured fixtures. <!-- MissingConfiguration is {category, settingName} and has NOWHERE to put a value — not "we redact it", which is the version of that promise that does not survive contact with a hurry. The setting NAME is carried because ISIS_PROVIDER_KEY is exactly what an operator needs to hear. resolutionAuthority (self-service / tenant-admin / platform-operator) makes the answer actionable by the person reading it rather than by somebody they have to go and find, and it is cross-checked against remediation.requiresOperator: a platform-operator resolution that claims no operator is required is telling a self-service user to fix something they cannot reach. An empty missing list is refused — "not configured, and we cannot say what is missing" leaves the reader no move. -->
    • S1.6.d Define degraded with affected and retained behavior, cause, limitations, recheck/expiry, and recovery. Evidence: degraded fixtures. <!-- BOTH halves are required and non-empty, and that is the whole distinction between degraded and its neighbours: nothing affected is configured, nothing RETAINED is unavailable. A caller told only what broke disables a surface that is still half usable; a caller told only that it works hits the part that does not. A behaviour listed in both is refused, since a reader of either list alone would get the opposite answer. Automatic recovery with no recheckAfterSeconds is refused — nothing would ever look again, so the degradation is permanent in practice — while MANUAL recovery with no recheck is fine, because a human is the trigger. An expiresAt at or before checkedAt is refused: a degradation that expired before it was observed cannot be acted on. -->
    • S1.6.e Define unavailable with transient/terminal/unknown dependency state, retry/circuit metadata, and no fabricated fallback. Evidence: unavailable fixtures. <!-- unknown is a real disposition and not a cop-out: a timeout says nothing about whether the far side is restarting or gone, and forcing a guess produces retry policy built on invented information. FIVE coherence rules, each with its own test: terminal-but-retryable (a client retries a permanent failure until its own budget runs out and reports the wrong cause), transient-but-not-retryable (work discarded although the state says it would have succeeded shortly), retryable-with-no-delay (a client with no backoff is the thing keeping the dependency down), its inverse, and an open circuit with no cooldown — a busy loop wearing a circuit breaker. "No fabricated fallback" is enforced by NAMING the ten keys such a value would arrive under, checked before the generic unknown-field pass. A generic message is the one somebody works around; this refusal says that a stale answer presented as a live one IS the fabricated success the honest-capability invariant exists to prevent. -->
    • S1.6.f Define failed for completed probe/evaluation failure with stable error class, correlation, safe detail, and remediation. Evidence: failed fixtures. <!-- failed means an evaluation COMPLETED and returned "this is broken"; unavailable means no answer was obtained. The two have opposite retry semantics, so collapsing them means retrying a permanent failure forever or giving up on a blip — and the shapes make the confusion impossible: failed has no retry metadata and unavailable has no failureClass, both tested. failureClass is a closed seven-value enum because a free-text class changes when somebody improves the wording and every dashboard grouped on it silently splits into two series; the message lives in safeDetail, where it is allowed to change. correlationId is REQUIRED: a completed verdict nobody can find the trace of is indistinguishable from a guess. -->
    • S1.6.g Require exhaustive handling and prohibit boolean/truthy collapse in typed and UI consumers. Evidence: compile-time negative and consumer tests. <!-- matchCapabilityState() is TOTAL: the handler record requires all five keys, so adding a sixth state breaks every consumer at compile time instead of falling into somebody's default: and being treated as whichever neighbour that branch already handled. Four @ts-expect-error blocks are the assertions (an unused one is itself an error) — a partial handler record, reading a variant field off the union, assigning a state NAME where the union belongs, and a handler typed for the wrong variant. PROVEN BY MUTATION: making one handler optional broke the directive AND the implementation. THE BOOLEAN PROHIBITION IS ENFORCED, NOT DOCUMENTED. usability() is THREE-valued (full / reduced / none) because a boolean cannot represent degraded without rounding it to a neighbour — round up and a caller uses the broken half, round down and a caller disables a surface that still works. And a test checks by REFLECTION that the module exports no is/has/can predicate over the union except isStateFresh (which is about the answer's AGE, not its meaning). Both mutations caught: collapsing usability to two values, and adding an isUsable boolean. -->
  • S1.7 Define a typed workbench capability registry. Evidence: duplicate capability IDs and missing version metadata fail validation. <!-- libs/oshun/workbench-kit/src/registry.ts + registry.spec.ts (50 tests), published as ./registry. Evidence is evidence/v1-workbenches/workbench-kit-registry-fixtures.json — a realistic SIX-capability registry across five domains plus 30 cases (17 descriptor rejections, 10 registry-level checks, 3 merge cases), generated by generate-workbench-kit-registry-fixtures.mjs (12 structural tests) and run through the real parser/validator/merge/projections by the kit spec. The two the cell names explicitly both have fixtures: registry/duplicate-id and descriptor/missing-version. MUTATION-TESTED four ways — last-write- wins merge, cycles following optional edges, a projection dropping a field, and the duplicate-id check removed — each failing exactly the case that claims to cover it. -->

    • S1.7.a Define stable capability ID, version, owner, domain, audience, host, runtime, lifecycle, and description. Evidence: registry schema. <!-- All ten, none optional. The version is a STRICT semver with no build metadata and no leading zeroes: semver ignores build metadata in precedence, so two entries could differ in a field no comparison can see, and 1.02.0/1.2.0 would be two spellings of one version. compareSemanticVersions implements real precedence including the rule people get backwards — a numeric prerelease identifier ranks BELOW an alphanumeric one — tested over an eight-version ladder. The lifecycle rule is bidirectional: a deprecated or retired capability MUST name its successor (a deprecation with no replacement is an instruction to stop with no instruction to start, which is why nobody migrates), and a live one must NOT (it reads as deprecated to anyone scanning the registry). -->
    • S1.7.b Define required/optional dependencies, provided contracts, configuration references, provider/resource needs, and health probes. Evidence: dependency metadata tests. <!-- A dependency listed as both required and optional is refused: a caller reading either list alone gets the opposite answer about whether it can run without it. Self-dependency is refused (the walk would not terminate). Every configuration reference must state required, because an optional setting treated as required makes a working deployment look broken. A probe interval below 5s is refused — a health check costing more than the information it produces is how the check becomes the load — and a duplicate probe name is refused because a runtime would run it twice and report two answers to one question. -->
    • S1.7.c Define supported commands, queries, jobs, routes, blocks, gates, import/export, and publisher features by typed references. Evidence: capability metadata tests. <!-- Eight kinds as <kind>:<dotted.name> rather than eight arrays of bare strings. The kind travels WITH the name so a consumer holding one reference in isolation still knows which of eight registries to look it up in; a registry full of "run.publish" cannot say. Round-trip tested, duplicates refused, and the sample registry exercises at least six of the eight kinds so the parser is not proven correct on command: alone. -->
    • S1.7.d Define authorization scope, data classification, trust controls, observability, SLO, cost, and fixture policy. Evidence: operational metadata tests. <!-- Four rules that make the operational metadata load-bearing rather than decorative. (1) An SLO availability target of exactly 1 is REFUSED: a 100% target leaves no error budget, so every incident is a breach and the number stops informing any decision. (2) At least one metric is required, because an unmeasured capability cannot have its SLO evaluated, which makes the SLO decorative. (3) restricted data requires synthetic-only fixtures and personal data forbids a production copy — otherwise the capability's own test suite holds the data its classification forbids copying, and anonymisation is a claim about re-identification risk that nothing in this registry can verify. (4) Admin-only scopes with requiresElevation false is refused: one of the two is wrong and the permissive reading is the one that ships. Cost is integer minor units, since a float estimate accumulates error the moment anyone multiplies it by a volume. -->
    • S1.7.e Enforce unique IDs/aliases, semantic versions, known owners, known contracts, declared dependencies, and acyclic graph. Evidence: registry negative tests. <!-- Nine violation kinds, each with a fixture, and --check fails if a kind's bucket empties. The graph checks run over a real six-node registry rather than a two-node toy. THE MOST INSTRUCTIVE CASE IS A NEGATIVE OF A NEGATIVE: a cycle through an OPTIONAL dependency must NOT be reported, because an absent optional dependency satisfies the loop — without that case, a validator that reported every kind on every input would pass all nine positive cases. Cycle detection is iterative with an explicit stack (a plugin-supplied registry should produce a violation, not a stack overflow) and reports one cycle once however many nodes it is entered from. retired-still-required is called out separately: retiring something its dependents still need is the failure retirement exists to prevent, and the registry says so before the deploy does. An empty knownOwners list SKIPS the check rather than failing everything — an empty register means "we do not have that list", not "nothing is valid". -->
    • S1.7.f Merge/discover shared and domain plugin entries deterministically and reject collisions rather than last-write-wins. Evidence: multi-plugin registry tests. <!-- A collision is an ERROR naming EVERY source that claimed the id, and the failed result carries no entries at all — so there is no winner for a caller to reach for by accident. Last-write-wins is the obvious implementation and it is how a domain plugin silently replaces a shared capability, in an order that depends on plugin discovery. THE PAIR THAT MATTERS: two sources claiming one id, and ONE source claiming an id twice. Both are collisions, because deduplicating the second would be last-write-wins wearing a different hat. Determinism is asserted by merging the same sources in reverse order and comparing bytes. -->
    • S1.7.g Generate runtime probe and human/API reference projections from the same registry. Evidence: source-to-output drift test. <!-- Both projections read the same entries. The drift test does NOT compare against a committed copy of the output, because a copy is a second artifact to maintain and it drifts in the other direction — instead every projection row is traced back to the registry field it came from, which catches an invented value and a stale one, and is the property a copy would only approximate. Byte-stable across calls and independent of input order. The two projections deliberately DISAGREE on one thing: a RETIRED capability is dropped from the probe list (a runtime should not execute it) but kept in the reference (a page that forgets retired entries cannot tell anyone what replaced them) — two different questions, tested as such. -->
  • S1.8 Define a route descriptor with host, audience, scope, navigation, deep-link, and feature-flag metadata. Evidence: route integrity tests. <!-- libs/oshun/workbench-kit/src/routes.ts + routes.spec.ts (53 tests), published as ./routes. Evidence is evidence/v1-workbenches/workbench-kit-route-fixtures.json — a six-route set across three hosts plus 37 cases (24 descriptor rejections, 10 set-level checks, 3 projection cases), generated by generate-workbench-kit-route-fixtures.mjs (13 structural tests). A route is the one artifact three audiences read: a router mounts it, a menu renders it, and a share link resolves it. Those three disagree constantly — a page in the menu the router never mounted, a link that leaks a tenant's data, a flag that hides the button but not the URL — so all three are DERIVED from one declaration. MUTATION-TESTED four ways: dropping the scope cross-check, dropping the method from the effective path, dropping a route from navigation without recording it, and dropping the share-safety check. -->

    • S1.8.a Define canonical route ID, path/pattern, method or UI kind, owning capability, contract version, and lifecycle. Evidence: route descriptor schema. <!-- "method OR ui kind" is a DISCRIMINATED UNION, so a descriptor carrying a GET and a dialog is unrepresentable rather than caught — a single optional-both shape leaves every consumer deciding which half to believe, and both directions are fixtured. A trailing slash is refused because /runs and /runs/ would be two spellings of one route and the duplicate-path check could not see the collision. A parameter used twice in one path is refused: the two positions can disagree and the router keeps one. A deprecated or retired route MUST name a redirect, or every link and bookmark to it becomes a 404 with no way to find the replacement. -->
    • S1.8.b Define canonical host, allowed host aliases, audience, role, tenant/workspace scope, and data classification. Evidence: host/audience validation tests. <!-- THE LOAD-BEARING RULE OF THE WHOLE CELL: a tenant- or workspace-scoped route must say WHERE the scope comes from, and the two answers are cross-checked against the path. path-parameter requires the path to actually contain :tenantId; session requires it NOT to. A route that takes a tenant from both is one where the two can disagree and whichever the handler reads first becomes the answer — which is a tenant-isolation bug that no amount of handler review finds, because each half looks correct alone. All four directions have fixtures (missing binding, platform route with a binding, and the two mismatches). hostAliases repeating the canonical host is refused: the alias list is what a route is ALSO reachable from, and including the canonical one makes the duplicate-path check see two. -->
    • S1.8.c Define navigation placement, label/icon keys, order, parent, breadcrumb, discoverability, and external/embedded behavior. Evidence: navigation metadata tests. <!-- labelKey and iconKey are i18n KEYS, validated as dotted keys — a literal label in a route descriptor is a string no translator will ever see, and the descriptor is exactly where one gets typed because it is the one place the words are convenient to write. A hidden route declaring a menu placement is refused: a menu slot for something the menu must not show is a contradiction one of the two consumers resolves wrongly. An API route carrying navigation is refused — a menu entry that resolves to a JSON endpoint is a link nobody can follow. Self-parenting is refused (the breadcrumb walk would not terminate). -->
    • S1.8.d Define deep-link entity, subview, filter, inspector, anchor, expiry, and safe-share state. Evidence: deep-link descriptor tests. <!-- SHARE SAFETY IS CHECKED AGAINST THE DATA CLASSIFICATION: a publicly shareable link into a confidential, restricted or personal route is refused. A descriptor is precisely where somebody sets public for convenience, months after the classification was decided somewhere else, and the two would never be read together by a human. The expiry rule runs both ways: a never-share link may NOT carry one (there is nothing to expire, and the number reads as permission to share it for that long) and a shareable one MUST (a URL that keeps working after the person it was sent to has left). The entity parameter must appear in the path, or a link built from it resolves to the route's default entity rather than the one that was shared. Anchor kinds reuse the S1.4 vocabulary rather than inventing a second one. -->
    • S1.8.e Define feature/config flag identity, default, targeting, dependency, expiry, and unavailable behavior. Evidence: flag metadata tests. <!-- A FLAG IS TEMPORARY UNLESS SOMEBODY ARGUES OTHERWISE: exactly one of expiresAt and permanentReason, never both and never neither. Neither is the flag nobody removes because nobody remembers whether it is still load-bearing; both is a contradiction. Both failure modes are fixtured, and the shipped route set contains one of each so the rule is exercised in both directions rather than only refused. fail-open is REFUSED on a confidential/restricted/personal route: an outage in the flag service would grant access the flag exists to withhold. Every flag must state its default, because a route whose default nobody wrote down behaves differently the first time the flag service is unreachable. Self-dependency is refused. -->
    • S1.8.f Validate unique IDs and effective paths, known host/capability, authorized audience/scope, navigable parents, and no redirect cycles. Evidence: route integrity negative tests. <!-- Nine violation kinds, each fixtured, --check failing on an empty bucket. The effective path is HOST + METHOD + PATH, and both halves of that matter: a host alias is enough to collide two routes that look distinct in isolation, and the same path with two different methods is two handlers rather than a collision. THAT SECOND CASE IS THE NEGATIVE-OF-A-NEGATIVE — without it, a check that reported every path twice would pass both collision fixtures. unnavigable-parent is separate from unknown-parent because they fail differently: a breadcrumb through a route that appears in no menu is a link to a page the user cannot get back to. audience-exceeds-capability catches a route widening the audience of the thing it exposes; the capability audiences come from the S1.7 registry rather than being inferred, since inferring them would make any route correct by construction. An empty capability register SKIPS the check rather than failing everything. -->
    • S1.8.g Generate route inventory and host navigation projections from the descriptor without silently mounting unauthorized source pages. Evidence: generated route snapshot and drift test. <!-- The navigation projection RETURNS ITS EXCLUSIONS with a reason each (no-navigation-placement / different-host / audience-not-permitted / not-discoverable / retired), and the test that matters is navigationAccountsForEveryRoute: every route is either in the tree or in the exclusion list, never in neither and never in both. A projection that quietly returns a shorter list is indistinguishable from one that never knew about the missing routes, and "without silently mounting unauthorized source pages" needs that honesty in BOTH directions. Three projection cases across two audiences and two hosts, so the filter is proven to filter rather than to return everything. The inventory and the menu deliberately disagree on a search-only route: a menu shows what is browsable, an inventory records what is mounted. Both projections are byte-stable and independent of input order. --> <!-- Note for S1.9+: hostNavigation promotes a child whose parent was excluded to a root rather than dropping it, so an audience that may see a leaf but not its parent still reaches the leaf. That is a deliberate choice and the alternative (hiding the subtree) is defensible; it is called out here because it is the kind of decision a later cell may need to revisit once real menus exist. -->
  • S1.9 Define domain plugin interfaces for stage models, gate evaluators, structured blocks, previews, inspectors, imports, and publishers. Evidence: minimal A and B archetype test plugins compile. <!-- libs/oshun/workbench-kit/src/plugins.ts, published as @oshun/workbench-kit/plugins (S1.2's derived export map authorizes it from this cell), plus the two archetypes under src/archetypes/ and evidence/v1-workbenches/workbench-kit-plugin-fixtures.json (73 cases, 45 violation kinds, generator + 12 structural tests wired into verify:inventory:v1-workbenches). 491 tests in the kit, all seven Nx targets green. THE LINE THE WHOLE MODULE IS DRAWN AGAINST: a FORMAT is a spelling rule (a stage id is lowercase-kebab so it survives a URL and a JSON key); a VOCABULARY is a decision about the domain's model (a stage is called review). The kit enforces the first and refuses to hold an opinion about the second. That is not provable by a sentence, so the evidence is two archetypes with DISJOINT stage vocabularies that both validate, and a --check that fails the moment they share a name. Four rules carry the module: a declaration and a contribution must agree in BOTH directions; a measurement is not a verdict; nobody verifies their own work; absence is a value, never an exception. -->

    • S1.9.a Define plugin identity, supported kit/contract versions, capability declarations, dependencies, configuration, and lifecycle. Evidence: base plugin contract. <!-- PluginDescriptor + parsePluginDescriptor, 18 committed rejections across 8 error codes. The descriptor is DATA and the contributions are code, split deliberately: a host decides whether to LOAD a plugin before it can run any of it, and a system that must execute a plugin to learn whether it is admissible has already given it the machine. Kit support is a HALF-OPEN semantic range (>=min, <below) — an inclusive maximum of 1.4.0 excludes 1.4.1, so every honestly pinned plugin breaks on a patch release and the pressure becomes to stop pinning. The wire schema range is a separate INTEGER type, because modelling it as semver would let a plugin declare ^1.2 support for a thing that has no minor version. declares is a discriminated implemented/degraded/unsupported and is closed PER VARIANT: retains on an unsupported row would describe behaviour the row just denied, and a reader believes the more generous half. -->
    • S1.9.b Define typed stage model and transition-guard registration with no shared assumption about domain stage names. Evidence: stage plugin fixtures. <!-- StageModel<TStage extends string> / TransitionGuard / validateStageModel, 10 fixtures. The validator never compares a stage against a vocabulary — only against the model's own stages — and the proof is clean/two-plugins-with-disjoint-stage-vocabularies plus the archetypes. GuardOutcome has THREE members: an indeterminate guard whose dependency is unreachable must not answer allow, and transitionDecision reports the first block AND the first indeterminate separately, because a caller that learns only "blocked" cannot tell whether to show a remediation or page whoever owns the unreachable guard. Reachability is computed rather than asserted: an unreachable stage is either dead or the transition that reaches it was never written, and both are worth saying out loud. -->
    • S1.9.c Define gate/evaluator input, raw result, policy verdict, provenance, staleness, remediation, and failure interfaces. Evidence: gate plugin fixtures. <!-- GateInput / GateMeasurement / GateProvenance / GateFailure / PolicyDecision / GateVerdict + gateVerdict, 7 verdict fixtures and 2 registration fixtures. THE SEPARATION IS THE CELL: the evaluator MEASURES and the policy DECIDES, so one measurement can be read against two tenants' thresholds — and staleness becomes expressible at all, because a stale measurement is a real number that must not become a pass. The kit resolves absence, failure and staleness BEFORE the policy is consulted, and NonPassingVerdict excludes pass and warn by type, so a domain cannot write a policy mapping "the gate did not run" to "the gate passed". decidedBy names which of the two decided, so a green gate is always traceable to the thing that greened it — and the assertion that matters in the spec is not the verdict but that the policy's decide was NOT CALLED on the three kit-decided paths. The freshness window is half-open and a future timestamp is not fresh: a clock skew read as extra validity extends every cached pass by the skew. -->
    • S1.9.d Define block schema, validator, editor, semantic renderer, migration, diff, anchor, import/export, and accessibility interfaces. Evidence: block plugin fixtures. <!-- BlockContribution (validate / editor / toSemanticNodes / migrationSteps / diff / anchors / interchange / accessibility) + validateBlockContribution, 9 fixtures. The renderer returns SEMANTICS rather than markup — a contract-class module cannot bind a view library, and a block returning markup makes every non-visual consumer (screen reader, export, plain-text digest, diff) parse HTML to recover meaning the block already had. The editor is a real contract minus the component: surface, typed commands with per-command scope and a reversible flag, and a REQUIRED read-only fallback, because the alternative a missing one invites is showing the editor and refusing on save, which teaches people their work can vanish. applyEdit is pure and re-validates, since an editor is the one place a block is built from untyped input. The migration chain must be CONTIGUOUS from its lowest version to the current one: a gap means every document at that version is unreadable and nothing reports it until one shows up. diff and anchors take the revision they are taken AT rather than manufacturing one — an anchor without it is an offset into a document nobody can date. -->
    • S1.9.e Define preview and inspector contributions with typed selection, authorization, loading/state, nonvisual alternative, and host isolation. Evidence: projection plugin fixtures. <!-- ProjectionContribution + SelectionContract + ContributionAuthorization + ProjectionState + NonvisualAlternative + HostIsolation + validateProjectionContribution, 6 fixtures. ready is the ONLY state carrying a model, and empty and unauthorized are deliberately different members: collapsing them is how a permission defect renders as an empty state for months, because the panel looks the same and nobody files anything. The non-visual alternative is REQUIRED rather than a courtesy — a projection is often the only place a piece of data is shown at all, and if its only rendering is visual the data has no API, no export and no screen-reader path. Isolation rules with teeth: an iframe with no permitted origin can be framed by anything that can then post to it, * is the absence of the check written as a value, and an in-process contribution over confidential/restricted/personal data may not reach the network, because in-process means it carries the host's credentials off-origin. -->
    • S1.9.f Define importer plan/validate/execute/reconcile and publisher plan/execute/verify/compensate interfaces. Evidence: interchange plugin fixtures. <!-- Importer / PublisherRuntime / PublisherMetadata + executionMayProceed + verificationIsIndependent, 4 fixtures and the archetype behaviour tests. Approval is bound to a PLAN DIGEST, so the bug this exists for cannot happen: validate approves plan A, the source changes, execute runs plan B under A's approval, and both halves look correct in review because neither can see the other's input. The archetype test runs exactly that and asserts 0 applied. VerificationSource has three members and the executor's own report is not one of them — "the thing that did the work says it worked" is the most common shape of a fabricated success in a publishing pipeline, and it is the one that survives review because the code reporting it is the code that would have failed; a matching verification with no evidence is refused for the same reason. Reversibility is a discriminated PAIR: reversible implies a compensation, irreversible implies a stated reason, and neither can be omitted quietly. -->
    • S1.9.g Compile and runtime-validate one content-authoring and one operations-console plugin with different types and unsupported capabilities. Evidence: A/B archetype tests. <!-- src/archetypes/story-desk.archetype.ts (A, content authoring) and incident-console.archetype.ts (B, operations). COMPILE is proven by the typecheck target: the archetypes are excluded from tsconfig.lib.json and included in tsconfig.spec.json, so tsc --noEmit -p tsconfig.spec.json compiles them while nothing reaches dist — an archetype in dist is a fixture shipped to every consumer. (They stay inside a tsconfig, which is the src/__boundary-fixtures__/ lesson: excluding files from every project takes them out of typescript-eslint's project service and replaces every rule with a parse error.) RUNTIME-VALIDATE is validatePlugins([A, B]) === [] plus admitPlugin returning three DIFFERENT answers — activated, refused (kit version) and not-configured (missing setting) — because a host that collapsed the first two would show "not configured" for a plugin built against a kit that no longer exists. DIFFERENT TYPES, made checkable: disjoint stage vocabularies, different extension-point SETS (B contributes no block and no inspector), and opposite publisher reversibility with two different verification sources. UNSUPPORTED CAPABILITIES: A declares image generation unsupported (no model access by policy) and the archive degraded (with what it retains); B declares structured authoring unsupported, because an incident record is a fixed schema written by detection systems and an editor there would offer a surface no responder may use mid-incident. Both are cross-checked against the fixture's recorded claims, so converging the two archetypes fails a test rather than quietly weakening the evidence. The archetypes are real logic over injected PORTS, not sketches: the attribution gate parses the draft and cross-references its footnotes, the importer parses markdown, the migrations transform real documents, and the publisher verifies by reading the target back through the same port it wrote through. A test supplies in-memory ports — that is the dependency boundary, and the only place a double belongs. -->
  • S1.10 Define stable error codes and safe public error envelopes. Evidence: internal causes are logged but secrets and stack traces are absent from client payloads. <!-- libs/oshun/workbench-kit/src/errors.ts, published as @oshun/workbench-kit/errors, plus evidence/v1-workbenches/workbench-kit-error-fixtures.json (63 cases, 21 violation kinds, 11 leak classes, the 28-code transport table) and a generator + 12 structural tests wired into verify:inventory:v1-workbenches. 571 tests in the kit. THE HEADLINE EVIDENCE, as one test: a cause holding a stack frame, an INSERT statement, a JWT and another tenant's id produces an envelope that repeats NONE of them, while an authorized operator gets all of it back through the correlation id alone. The design decision the whole module rests on: an error crosses the boundary TWICE, and the safe crossing is CONSTRUCTED, never derived by deleting fields from the rich one. Deletion is a blocklist, and a blocklist is wrong the first time somebody adds a field — the leak then arrives with a feature that had nothing to do with errors, and nothing fails. -->

    • S1.10.a Define namespaces and stable codes for validation, authentication, authorization, tenancy, conflict, state, dependency, provider, job, policy, quota, and internal errors. Evidence: error code registry. <!-- ERROR_REGISTRY: 28 codes across all twelve namespaces, SHIPPED by the kit rather than left to each host — the whole value of a stable code is that a client written against one workbench handles the same failure the same way in another, and a per-domain registry is a per-domain vocabulary with extra ceremony. The namespace set is CLOSED because a namespace is what an old client falls back to when it meets a code it has never seen; if the set can grow, the fallback has nothing to fall back to. validateErrorRegistry reports 21 kinds and the shipped registry is one of the corpus's clean cases, so a rule that became too strict fails here rather than being discovered by a domain. -->
    • S1.10.b Define envelope fields for code, safe message key/parameters, affected field/reference, retryability, remediation, correlation, and optional details. Evidence: public error schema. <!-- PublicErrorEnvelope + publicError. Parameters are SCALARS only (an object carries whatever its author put on it, and a scanner can read a string but cannot know what a nested field MEANS) and only under keys the CODE declared: an open parameter bag is the leak that arrives without a code change, when a handler passes the object it already had and a field added to that object six months later goes out with it. The message is an i18n KEY, never text — a literal is both untranslatable and a string a client will end up matching on. Two retry rules run in opposite directions: a non-retryable code may not carry a retry-after, and a code that says "retry after backoff" MUST, because otherwise every client picks its own delay and the ones that pick badly are indistinguishable from an attack. -->
    • S1.10.c Map domain/internal/provider/database/queue/bridge exceptions to public codes without string matching fragile messages. Evidence: mapper unit tests. <!-- mapCause + DEFAULT_CAUSE_MAPPINGS (SQLSTATEs, provider error codes, domain error names) + validateMappings. The guarantee is structural: mapCause has NO message parameter, so there is nothing to match on. validateMappings then refuses a classifier that reads as prose — a space, a trailing full stop, over 64 characters, a capitalised sentence — because message-matching does not arrive as a decision, it arrives as one mapping somebody added under pressure with the provider's sentence pasted in. An unmapped cause falls back to internal.unexpected rather than improvising a code: an unmapped provider failure that inherited the provider's own retry semantics is a retry storm nobody chose. Five cause cases including two that must NOT match — a classifier colliding across kinds, and a source-agnostic rule that must still match a source it never named. -->
    • S1.10.d Define HTTP/RPC/job/event status and retry semantics per code. Evidence: transport mapping table. <!-- transportTable() over the registry, and the committed table in the fixture is a hand-MIRRORED claim the spec asserts equality against. Deriving it would make it agree by construction and prove nothing; mirroring means a change to any code's status or retry semantics has to be made twice, in two files, by somebody who noticed. Four cross-checks run inside the registry validator, each catching a pair that reads as reasonable in isolation: backoff advice on a status that will never succeed, "change your request" on a server fault, a worker retrying what the code says cannot be retried, and a job parked for an operator without saying an operator has to act. The rpc column is checked against the http column too, so the same failure cannot mean two things depending on which transport a client used. -->
    • S1.10.e Redact secrets, tokens, signed URLs, stack traces, SQL, protected identities, content, prompts, and cross-tenant object detail. Evidence: canary-data negative tests. <!-- scanEnvelope over 11 leak classes, with one canary each, and publicError REFUSES rather than strips: a caller whose parameter silently disappeared ships a message with a hole in it and learns nothing. The corpus's sharpest half is the five LOOKALIKES — a field path containing "select", a short sk-2026, prose containing the word "at", a plain docs URL, and the addressed tenant's OWN id — all of which must pass, or the scanner is not detecting secrets, it is refusing to emit anything. DOCUMENTED APPROXIMATION, stated rather than hidden: a prompt is prose, and so is a document and a comment, so no pattern distinguishes one from an error message somebody wrote. Prompts (and SQL, stacks, tokens) are therefore ALSO refused by the NAME of the field carrying them, and the limit is that a prompt under an undeclared key would not be caught by shape. What catches that case is the S1.10.b rule that a parameter key must be declared by the code — the two checks cover each other and neither is asked to do the other's job. Cross-tenant detection fails CLOSED: with no addressed tenant given, every tenant id in the payload is foreign, because "we do not know who this is for" is not a reason to send them somebody else's id. -->
    • S1.10.f Correlate the safe envelope to access-controlled internal cause, trace, log, and audit context. Evidence: support lookup integration test. <!-- lookupCause + envelopeMatchesCause, and the integration test is the cell's headline: build an envelope from a cause carrying a stack frame, an INSERT statement, a JWT and tenant:globex, assert the serialized envelope contains none of them, then resolve the correlation id and assert an authorized operator gets every one back. Four outcomes, and two of the distinctions are the point. expired is separate from not-found — "we deleted it on this date, because that is the retention we committed to" and "we have no record of this" are different answers, and collapsing them turns every retention deletion into a suspected data-loss incident. And an unauthorized caller is TOLD SO rather than told not-found: the usual advice is the opposite, and the argument for this direction is recorded in notDerivableFromSource — the correlation id came from the customer, who already knows the error happened, so hiding existence protects nothing and costs an hour. The lookup RETURNS the audit row rather than writing it: a function that both reads and audits in one place is one where the audit can be skipped by calling the read directly. envelopeMatchesCause checks quoting with a sliding 24-character window rather than whole-string equality, because the interesting failure is an envelope that quotes PART of the cause — a clause pasted into a parameter because it read well — and a whole-string check never sees it. -->
    • S1.10.g Version and deprecate codes without reusing semantics. Evidence: compatibility and unknown-code consumer tests. <!-- resolveCode returns five statuses and every one has a case. retired is deliberately distinct from unknown: "this used to mean something else" and "I have never heard of this" are different answers and only the first tells a maintainer what happened. A retired code may never be re-registered (retired-code-reregistered), because an old client still has the old meaning compiled in and would handle the new failure as the old one, silently and correctly-looking. The unknown-code path is the compatibility claim: a code from a NEWER registry still parses into a namespace, so an old client degrades to that namespace's retry semantics rather than crashing or — worse — treating an unrecognised error as a success. The per-namespace fallbacks are deliberately the cautious answer in each case and are recorded as a judgement rather than a derivation. A deprecation must name a live replacement with the SAME retryability, since a consumer that migrates on the strength of the pointer would otherwise silently change its retry behaviour, and a deprecated code stays emittable so the consumers the deprecation exists to give time to keep working. -->
  • S1.11 Define schema-version negotiation and supported-version discovery. Evidence: backward-compatible and unsupported-version contract tests. <!-- libs/oshun/workbench-kit/src/negotiation.ts, published as @oshun/workbench-kit/negotiation, plus evidence/v1-workbenches/workbench-kit-negotiation-fixtures.json (34 cases, 19 violation kinds, the 6-combination skew matrix) and a generator + 10 structural tests wired into verify:inventory:v1-workbenches. 615 tests in the kit. THE ASYMMETRY THE MODULE IS BUILT AROUND: negotiation needs a live counterparty. HTTP and the desktop bridge have one; an event, a job and a bundle do not — the reader may not exist yet and may not exist for months. So only two of the six transports may declare themselves negotiable and the other four have to be conservative instead. This is the thing a single "we negotiate versions" story gets wrong: it is written for request/response and then applied to a queue, where there is nobody to negotiate with. -->

    • S1.11.a Define how clients request or declare contract versions for HTTP, events, jobs, bundles, desktop/bridge, and generated clients. Evidence: per-transport version carrier rules. <!-- VERSION_CARRIERS: one carrier per transport, each naming the mechanism (header / envelope field / manifest field / handshake / compiled constant), whether a message with no version is refused, whether the receiver may answer with a different one, and an EXAMPLE — a carrier rule with no example is one every team implements slightly differently, and the differences only show up across a boundary. validateCarriers refuses a second carrier for one transport (two places a version can be written are two places it can disagree, and the handler reads whichever it reads first), a transport with none, and an asynchronous transport claiming to negotiate. The generated-client carrier is required and NOT negotiable on purpose: a generated client cannot decide at runtime to speak a version it has no types for. -->
    • S1.11.b Define server default selection, minimum/maximum and exact supported versions, compatible ranges, and preference ordering. Evidence: negotiation algorithm tests. <!-- CapabilitySupport + negotiate + validateSupport. Supported is a set of EXACT versions, never a range: a range admits versions nobody implemented and the gap only appears when a client asks for one. The SERVER's preference wins over the client's, recorded in notDerivableFromSource as a decision with a cost — the server knows which of its implementations is healthy and which is being drained, and letting the client pick keeps the oldest version alive as long as one client asks for it, which is the state a sunset date exists to end. Four reasons are reported (exact-single-offer, server-preference, highest-common, server-default) and the spec asserts the REASON as well as the version, because a right answer reached the wrong way stops covering what it claims the moment the algorithm changes. A default that is not in the supported set is refused: every client that declares nothing would get a version that does not exist. -->
    • S1.11.c Expose authenticated capability/version discovery with owner, lifecycle, deprecation, sunset, and documentation links. Evidence: discovery contract tests. <!-- discoveryDocument + validateDiscovery. An unauthenticated caller is REFUSED rather than served a redacted document: an inventory of every capability and exactly which old contracts are still served is a map of what to try, and a redacted one has informative gaps. Each entry must name an owner (a row nobody owns is one a caller cannot ask a question about) and a docs link (without it the entry says a version exists and nothing about how to move off it). Two deprecation checks matter: a deprecated version missing from the supported list makes a client on it conclude it is already off — the opposite of what a deprecation notice is for — and a sunset date that has passed while the version is still served teaches clients to ignore the next one. -->
    • S1.11.d Return stable unsupported/too-old/too-new/unknown-version errors with supported alternatives and no silent downgrade. Evidence: refusal fixtures. <!-- Five codes added to the S1.10 registry — validation.version_missing / _too_old / _too_new / _unsupported / _unknown — so a version refusal is a registry code with transport and retry semantics like any other, and the spec asserts every code the negotiator can return resolves in that registry. The three no-overlap answers are deliberately DISTINCT: only "too old" and "too new" tell anybody which side to upgrade, and collapsing them is how a version mismatch becomes a support ticket instead of a one-line fix. Every refusal carries the supported set, so a client is told what it could have asked for. NO SILENT DOWNGRADE is data rather than prose: a negotiated result whose version is not the client's first choice reports departedFrom AND direction, and the two directions are separate events — an upgrade means the client under-asked and should raise its floor, a downgrade is the one with a sunset attached. -->
    • S1.11.e Test old client/new server, new client/old server, overlapping ranges, no overlap, missing declaration, and disabled capability. Evidence: version-skew matrix. <!-- All six combinations are a CHECKED list in the fixture and --check fails on an uncovered one, because a per-case spec assertion would be perfectly green with three of the six missing — nothing in a per-case assertion knows what the other cases were supposed to be. 11 skew cases in total, including the two that are easy to get wrong: an old client offering only v1 must get v1 even though the server prefers v2 (a preference must not override an offer set that does not contain it), and a missing declaration on a REQUIRED carrier is refused while on an optional one the default applies AND the reason records that it did, so a caller can tell a chosen version from an assumed one. A disabled capability returns unavailable, not a version refusal: it is not a version problem, and a version code would send somebody to change their client for no reason. -->
    • S1.11.f Bind negotiated version to validation, serialization, cache, idempotency, audit, and telemetry. Evidence: end-to-end version propagation tests. <!-- bindVersion derives all six from the negotiated version, and validateBinding refuses a binding that lost it. Derived rather than configured so the version cannot be present in one key and absent from another. The two that carry the cell: a cache key without the version serves a v1 body to a v2 client, and an idempotency key without it dedupes a v2 retry against the v1 original and returns the OLD shape with a 200 — the failure a client is least likely to detect. Both look exactly like a working system until somebody upgrades. The propagation test negotiates, derives, asserts every derived key carries the version, and asserts a different negotiated version produces different keys so nothing can be shared between the two by accident. -->
  • S1.12 Define deprecation metadata and telemetry. Evidence: deprecated contract/route use is observable without breaking consumers. <!-- libs/oshun/workbench-kit/src/deprecation.ts, published as @oshun/workbench-kit/deprecation, plus evidence/v1-workbenches/workbench-kit-deprecation-fixtures.json (37 cases, 21 violation kinds, the 9-surface coverage matrix) and a generator + 11 structural tests wired into verify:inventory:v1-workbenches. 663 tests in the kit. THE TENSION THE CELL IS ABOUT: the two halves pull against each other, because anything loud enough for a consumer to notice is loud enough to change what their code does, and a warning that is genuinely safe is one nobody reads. The resolution is that the warning is always out-of-band and the OBSERVATION is the server's — usage is counted, alerted on and acted on here, so a consumer that never reads a header still shows up in the telemetry. That is the only way a sunset date is ever decided from evidence rather than from optimism. -->

    • S1.12.a Define deprecated-since, replacement, migration guide, warning, owner, supported-through, sunset, removal version, and exception fields. Evidence: deprecation schema. <!-- DeprecationNotice + validateNotice, 10 fixture cases. Two design points beyond the field list. Replacement is discriminated so no-replacement is a REAL answer with a required reason — some surfaces are withdrawn because the thing they did was a mistake, and pretending there is a migration target sends people looking for one. And the three dates are ordered against each other: supportedThrough before deprecatedSince means consumers were never given a window at all, and a sunset before the supported-through means the two dates describe different promises and a consumer believes the later one. The migration guide is required, because a deprecation without one tells somebody to stop and not what to do instead. -->
    • S1.12.b Attach deprecation metadata to registry capability, contract, field/enum, route, event/job, package export, bundle profile, and plugin versions. Evidence: coverage matrix. <!-- SURFACE_ATTACHMENTS + validateAttachments, nine surfaces — event and job split rather than the ledger's "event/job", because they deprecate differently: an event has consumers the producer cannot enumerate and a job has exactly one worker fleet. THE MATRIX IS HONEST RATHER THAN FLATTERING. Five surfaces have a descriptor field or a version range the kit already owns and names it (CapabilityDescriptor.lifecycle, RouteDescriptor.lifecycle, PluginDescriptor.lifecycle, the envelope schemaVersion ranges). The other four have neither, so they attach through a NAMED sidecar register and the matrix says sidecar-register instead of claiming a descriptor field that does not exist — which would make the matrix green and the deprecation unattachable. Every sidecar row is required to name its register, because those are the rows that rot first: nothing about the surface itself points at them. -->
    • S1.12.c Emit protocol-appropriate warnings without changing successful response meaning or exposing other consumers. Evidence: compatibility tests. <!-- warningFor over five protocols, each with an out-of-band channel: HTTP response headers, envelope extensions (never the payload), job-update extensions (never the result), a separate bridge frame, and stderr (never stdout). The success payload never changes, so a client that parses the body sees exactly what it saw before. "WITHOUT EXPOSING OTHER CONSUMERS" is made checkable by making the warning CALLER-INDEPENDENT: every field is derived from the notice and nothing from the request, and warningsAreCallerIndependent proves it by COMPARING two warnings rather than by reading the implementation — a reading stops being true the first time somebody adds a helpful field and nothing fails. The spec also proves the check can FAIL, by handing it a hand-built warning carrying alsoStillUsedBy. A warning that named who else was still on the old version would be a migration lever and a disclosure at the same time. -->
    • S1.12.d Record bounded usage by deprecated ID/version, consumer class, environment, and tenant-safe cohort. Evidence: telemetry schema tests. <!-- UsageRecord + validateUsage. TENANT-SAFE is enforced twice: a cohort containing tenant: is refused outright, and a cohort with fewer than COHORT_ANONYMITY_THRESHOLD (5) distinct tenants must report as the below-threshold bucket rather than as itself — a cohort that small is a tenant id with extra steps, and the record is read by every dashboard the surface appears on. The escape is the CORRECT behaviour rather than a suppression: the count is still published, under a bucket that names nobody. BOUNDED means the window: a record covers at most 24 hours, because a counter that only grows cannot be alerted on — the number goes up forever and never crosses a threshold that means anything. An inverted window is refused too, since it still sums into a dashboard. Usage against a surface no notice covers is reported, and the surface check is SKIPPED when the host supplied no register, for the same reason the plugin capability check is: a check that failed every record because the host has no list is one somebody disables. -->
    • S1.12.e Alert on usage after migration deadline and block removal while supported consumers remain without approved exception. Evidence: lifecycle gate tests. <!-- removalDecision (four outcomes) + deprecationAlerts (two severities and silence). The alert line falls where it does for a reason: past supportedThrough the surface still answers and nobody will fix it, which is a warning; past sunsetOn it was supposed to be gone, which is either a migration that did not happen or a removal that did not, and both are somebody failing to do a thing they said they would. Inside the supported window nothing is said at all, because alerting on use of something still supported trains everybody to ignore the alert. TWO CASES CARRY THE CELL and they look alike. An EXPIRED exception stops PERMITTING a removal rather than continuing to block it — an exception with no expiry is a decision nobody revisits and the surface outlives everybody who agreed to it. And development/staging usage does not block a removal, because that traffic is ours and blocking on our own smoke tests is how a surface outlives its replacement. The blocked outcome names WHICH consumer classes remain, since "somebody is using it" is not an answer anybody can act on. -->
    • S1.12.f Remove telemetry and compatibility adapters with the deprecated surface. Evidence: closure checklist. <!-- closureChecklist + validateClosure, six items, and the two that get left are on the list from the start: a telemetry counter for a removed thing sits at zero forever and reads as success, and a compatibility adapter nobody removed is a second implementation of a contract that no longer exists — which the next person to read the code will assume is load-bearing. The violation is claiming to be FINISHED with items open, not being unfinished: being mid-removal is normal, and a checklist that complained about it would be one people stop opening. A checklist that never mentions an item is reported separately, because an item nobody listed is one nobody forgot. -->
  • S1.13 Export test builders separately from production entry points. Evidence: production bundle scan contains no fixture factories. <!-- libs/oshun/workbench-kit/src/testing/ (six test-only subpaths, 51 utilities) plus libs/oshun/workbench-kit/src/synthetic.ts published as @oshun/workbench-kit/synthetic, the evidence artifact evidence/v1-workbenches/workbench-kit-test-utility-register.json, two new Nx targets (test-only-imports, production-scan) and a generator + 20 structural tests wired into verify:inventory:v1-workbenches. 811 tests in the kit; all nine Nx targets green. THE SHAPE OF THE ANSWER: four mechanisms, because each covers what the others structurally cannot — the export map's runtime lattice reads the SOURCE graph, the lint ban reads the IMPORT a consumer wrote, the emitted scan reads DIST, and the runtime guard is what happens when a fixture arrives anyway. Three of the four are static, and the static ones all assume the fixture travelled as code. The one that matters in an incident is the fourth, because a fixture usually arrives as DATA — a promoted seed, a container that resolved a double because it was registered first. -->

    • S1.13.a Inventory builders, factories, fixtures, fake clocks/IDs, in-memory stores, mock providers, and assertion helpers. Evidence: test-utility register. <!-- TEST_UTILITY_REGISTER in src/testing/register.ts — 51 rows over nine kinds (builder, violation-builder, fixture-corpus, clock, id-source, in-memory-store, assertion, builder-contract, constant) — and generate-workbench-kit-test-utility-register.mjs, which DERIVES the same inventory from the TypeScript AST and fails when the two disagree. An inventory is the easiest artifact in this ledger to fake: write a list, ship it, and it is wrong by the second commit. So it is checked in BOTH directions and by two independent readers. An export nobody registered fails; a row naming a symbol nobody exports fails — that is the direction that rots quietly, because a renamed builder leaves behind a row that reads as coverage. The generator evaluates the register as a literal and THROWS on a computed row rather than reading it as undefined, which would produce a clean report over an inventory describing nothing. The 30 type-only exports are reported separately, because the runtime completeness check in register.spec.ts cannot see them and a .d.ts leak is still a leak. -->
    • S1.13.b Define explicit test subpaths by runtime and capability without reexport from production roots. Evidence: export-map review. <!-- Six subpaths: ./testing (aggregate), /assertions, /doubles, /envelopes, /identity, /invariants, each added to KIT_OWNED_ENTRY_POINTS, package.json exports and the tsconfig alias table, all three derived from the same gate. The single test runtime class became two: test-contract (runtime-agnostic, and genuinely constrained — no Node built-in, no DOM global, no React) and test-server (declared with NO members, so the first helper needing a temp directory is placed rather than smuggled into the agnostic class). The old test class permitted everything, which is a lattice entry that checks nothing. THE ONE DECISION HERE IS THE OMISSION: ./testing re-exports every pure subpath and NOT ./testing/doubles. Everything else is a function that returns a value; the doubles hold mutable state that outlives the call, and a suite that asked for a builder and received a shared clock with it is one stray reference from a test that passes alone and fails in a run — which gets labelled flaky and retried rather than fixed. Checked three ways: the register spec, the generator over index.ts's export clauses, and the packed-consumer smoke test in the resolution mode a domain actually uses. -->
    • S1.13.c Ensure builders produce schema-valid defaults while requiring explicit values for identity, tenant, rights/consent, or other fields whose defaults would hide test intent. Evidence: builder contract tests. <!-- requireExplicit in src/testing/contract.ts plus src/testing/builders.spec.ts: a probe table covering all 28 builders, checked against the register for completeness, so "add a builder, forget the probe" fails here rather than shipping an unchecked factory. Every builder validates its own output through the kit's REAL parser before returning it — a factory that can silently emit an invalid value hands the suites that expect a rejection a pass for free. THE HONEST HALF IS THE OTHER COLUMN. A register where every builder requires everything is easy to write and impossible to use, so each sensitive field a builder DOES default is listed with the argument for it, and the gate requires the argument to be longer than a word. The reviewer question becomes "is that reason true?" rather than "did anybody think about this?". One field has no defensible default anywhere and the gate refuses it outright however good the prose: tenant. There is no aTenantId(), because one shared tenant makes an entire suite of isolation tests pass against a query that never filtered by tenant. -->
    • S1.13.d Mark fixture/synthetic values and prevent builders from being selected by production dependency injection. Evidence: production guard tests. <!-- src/synthetic.ts, published as @oshun/workbench-kit/synthetic — the PRODUCTION half, which therefore cannot live in a test-only module. Two marks, because neither survives what the other does: a global-registry symbol (enumerable, so a spread keeps it; symbol-keyed, so Object.keys and the closed-shape parsers never see it) survives in-process and dies in JSON, and a reserved id form survives JSON and a database column and says nothing about objects. THE RESERVATION IS STRUCTURAL, NOT LEXICAL, AND THE SCAN IS WHY. The first version reserved a bare synthetic- prefix on any string, and the S1.13.f scan immediately reported registry.ts shipping fixture data — it was the ratified S1.7 FixturePolicy value 'synthetic-only'. A lexical reservation over an open vocabulary collides. So a fixture id now always carries a colon (tenant:synthetic-acme, synthetic:idempotency-1), a form an enum member and a stage name never take, and nobody maintains a list of words the scan should ignore. FOR DI: selectProductionImplementation treats "the only registered candidate is a double" as an ERROR, never a fallback — the double satisfies the interface and the process boots, and that boot is the incident. It also refuses an ambiguous choice rather than letting registration order decide. The scan reports truncation as a finding, so a clean result from a walk that stopped early cannot be read as clean. assertTestRuntime is the fourth layer and its LIMIT is stated in the code and covered by a test: it detects a declared configuration, not a deployment. -->
    • S1.13.e Reject production source imports from test-only subpaths through boundary lint rules. Evidence: deliberate invalid import test. <!-- @typescript-eslint/no-restricted-imports in eslint.config.js — two blocks, one banning @oshun/workbench-kit/testing* repository-wide and one banning the relative spelling inside the kit, which the package pattern cannot see. Deliberately the typescript-eslint rule and not the core one: the core no-restricted-imports is already configured per-directory (the V1-P2-3500 axe contract, an off for the V4/V5 suites) and flat config resolves by last match, so a repository-wide core block would have silently replaced them. tools/test-only-import-fixtures.mjs writes eight real TypeScript files, runs the repository's real ESLint, and deletes them: five forbidden imports must be rejected and three permitted ones accepted, because a rule that also refuses the permitted imports is one the first person to hit it turns off. It fails on a PARSING error rather than counting it as a pass — the sibling boundary-fixtures tool documents that exact near-miss. The refusal message is part of the contract and is asserted: it must name @oshun/workbench-kit/synthetic as the production alternative, or the next reader reaches for eslint-disable. NOT COVERED, AND SAID RATHER THAN IMPLIED: the root config globally ignores libs/**/*.spec.ts, so a spec is never linted and this rule never runs on one. That is the intended outcome, but it means the ban rests on "every linted file is production source" rather than on a carve-out. -->
    • S1.13.f Scan built exports, bundles, source maps, and transitive modules for test utilities and fixture data. Evidence: clean production scan. <!-- tools/production-scan.mjs (Nx target production-scan, dependsOn build). Five scans over dist: reachability over the emitted .js AND .d.ts graphs separately (different graphs — a type-only leak still ships the names), a token scan, a fixture-identifier scan, a source-map scan, and a re-read of the emitted manifest. Result: 12 production entry points, 24 production files, 24 source maps, 72 forbidden tokens, zero findings. Tokens are read from the AST rather than by grep, so a doc comment mentioning a builder is not a finding and a renamed identifier still is. The token list is DERIVED from the S1.13.a artifact, so a builder added tomorrow is scanned for without anybody remembering to add it — and the fixture id forms are read from src/synthetic.ts through the same artifact, so the guard, the register and the scanner cannot hold three opinions about what a fixture id looks like. THE NEGATIVE CONTROL IS THE PART THAT MAKES THE RESULT WORTH ANYTHING: the same scan runs over the emitted TEST files and must find hits (264 today). Without it every assertion passes when the scan is looking at nothing — a renamed dist directory, an empty token list, an AST walk that returns no identifiers. The scan has already earned its place twice: it found the 'synthetic-only' collision that redesigned S1.13.d, and the packed-consumer smoke test's flat module resolution, which stopped at the new directory boundary and reported clean. -->

    Two further defects this cell surfaced, fixed here rather than filed: the export map's module graph read src/ non-recursively and resolved relative specifiers by stripping ./, so a nested module was invisible and ../x.js matched nothing — every check downstream saw a smaller graph than exists. And assertTestRuntime read the process global directly, which compiles only where @types/node is installed; the alias-resolution gate caught it by compiling a consumer program with types: [], which is what a test-contract subpath claiming to be runtime-agnostic has to survive.

  • S1.14 Generate API reference from source contracts. Evidence: generation is deterministic and CI detects drift. <!-- evidence/v1-workbenches/workbench-kit-api/ — eleven pages, 501 exported symbols, 749 fields, 45 examples drawn from 181 committed fixture cases — plus workbench-kit-api-reference.json (the model + violations), the generator, and 29 structural tests wired into verify:inventory:v1-workbenches. THE REASON IT IS GENERATED AND NOT WRITTEN: a hand-written reference for eighteen entry points is wrong within a week and nobody finds out, because the only reader who would notice already knew the answer. So every fact on every page is read from the TypeScript AST of the module the page owns, the export map, or a committed fixture corpus, and the only authored text in the whole output is the one-line intro per page, which lives in the generator beside the page definition so it changes with it. -->

    • S1.14.a Define authoritative contract/registry sources and exclude internal, test-only, unsupported, and secret configuration detail. Evidence: documented generation inputs. <!-- The PAGES table names, per page, the modules it owns, the fixture corpus it draws from and the subpaths it documents; the index page prints the inputs and the exclusions rather than leaving them in the generator. Four exclusions, each enforced rather than described. INTERNAL: only exported declarations are read, because a private helper in a reference is an invitation to depend on it. TEST-ONLY: a page owning a testing/ module is a violation, and a test tries four real builder names against every rendered page. UNSUPPORTED: no capability page exists, because all 91 S0.12 rows are still pending and the index says so with the number. SECRET: five credential shapes are scanned for. THE SECRET RULE EARNED ITS PLACE ON THE FIRST RUN. It found three hits in error.md — the S1.10.e leak corpus deliberately carries a JWT, a provider key and a DSN with a password, because a leak scan nobody can watch detect anything is not a scan. Exempting the page would have been the easy fix and the wrong one: it is the page most likely to carry a real leak. So the examples are REDACTED to their shape, the page says how many and why, and the scan runs AFTER the redaction — a credential arriving by a path the redactor does not cover still fails the build. -->
    • S1.14.b Generate capability, route, command/query, event/job, schema, error, version, lifecycle, and plugin reference pages. Evidence: complete reference index. <!-- All nine, plus identity and invariants, because those two carry the types every other page links to and a reference that omitted them would have 200 dangling links. COMPLETENESS IS A PROPERTY, NOT A COUNT: every production module is owned by exactly ONE page, and checkCoverage fails on a module no page owns, a module two pages own, a page that documents nothing, and a required page that was not generated. So "is the reference complete?" is answered by the gate rather than by reading the index. command-query and event-job are VIEWS over envelopes rather than owners — those four envelope kinds are what a reader looks up, and the shared schema is one page down — which is why ownership and presentation are separate ideas here. -->
    • S1.14.c Include valid and invalid examples, authorization/scope, idempotency/concurrency, pagination, state/failure, compatibility, deprecation, and evidence links. Evidence: section completeness check. <!-- Twelve sections, identical on every page, and a test asserts every one is present on all eleven. A reader who learns where the idempotency answer lives on one page knows where it lives on all of them, and a missing section is a hole rather than a formatting choice. THE SIX CROSS-CUTTING SECTIONS ARE DERIVED, not written: a field or symbol appears when its NAME matches a named vocabulary. That is mechanical on purpose, and both directions of the trade are stated on the page — it can list a symbol whose name suggests a concern it does not carry, and it cannot go stale, because a new field carrying that concern appears without anybody remembering to mention it. A section with no matches prints why it does not apply (20 of them across the set) rather than being omitted, because an omitted section is indistinguishable from an unanswered one. Examples are split accepted/refused by ONE uniform rule over seven differently-shaped corpora, each refusal naming the code or violation kinds it must produce, and a page shows at most three of each WITH the full count — a truncated view must never read as coverage. -->
    • S1.14.d Sort output and serialize examples deterministically with pinned tool versions. Evidence: repeat generation is byte-identical. <!-- Symbols sorted by name, pages by id, examples by case id, payloads through the repository's stable stringifier; a test builds twice and compares every page byte for byte. THE PINNED TOOL IS PRETTIER, AND IT HAD TO BE. lint-staged runs prettier --write on every staged .md, so generated markdown that was not already Prettier-shaped would be reformatted at commit time and the drift gate would fail on every commit forever. The generator therefore formats each page through the repository's own resolved Prettier config and RECORDS the version in the index. A Prettier upgrade then shows up as drift, which is the correct outcome: the formatter is part of the output contract, and the alternative is a silent reflow in an unrelated pull request. -->
    • S1.14.e Validate internal links, anchors, schema refs, example parsing, and source provenance. Evidence: documentation validation output. <!-- validatePages over the RENDERED markdown, not over the model: every [text](#anchor) and [text](./page.md#anchor) must resolve to a heading that exists (slugged the way the renderer slugs them), every fenced JSON block must parse, no page may match a credential shape, and every documented symbol must name a source file the page claims. Schema refs are the interesting one — a field typed ArtifactRef is rendered as a LINK to the page that owns it, so the anchor check turns "the type reference resolves" into something the build proves rather than something the reader discovers is false. Each violation kind has a deliberate failing fixture, including the one that matters most: a cross-page anchor that DOES resolve must not be reported, or a validator that rejects everything would look identical to one that works. -->
    • S1.14.f Add CI check that regenerates in memory or a temp directory and prints actionable drift. Evidence: deliberate source-change failure. <!-- --check rebuilds the whole model and every page in memory, and compares against what is committed. Actionable means three things it prints and most drift gates do not: WHICH file differs, the first differing LINE with both versions quoted, and the command that fixes it. A gate that says "output differs" sends somebody to a 700-line diff. It also fails on an ORPHAN — a committed page the generator no longer produces — which a file-by-file comparison alone would never notice, and which is exactly what a renamed page leaves behind. The deliberate-source-change evidence runs the real reader over a changed contract and asserts the output changes, rather than editing the kit and putting it back; an assertion that a gate fires must not depend on somebody remembering to undo the thing that made it fire. -->

S2 — Lifecycle, revisions, decisions, and gates#

  • S2.1 Extract the generic stage-machine engine with typed stage unions. Evidence: illegal transitions cannot be expressed without an explicit guard bypass type. <!-- libs/oshun/workbench-kit/src/stages.ts, published as @oshun/workbench-kit/stages, with 43 cases in stages.spec.ts. 855 tests in the kit; all nine Nx targets green. WHY IT IS A SEPARATE MODULE FROM S1.9: that cell lets a plugin DECLARE a stage model and has the host validate it at boot; this one decides whether a transition may happen. They fail differently — a bad model is a boot error somebody fixes before shipping, and a bad transition is a published document nobody can unpublish — so they are separate surfaces with one shared implementation of the two graph analyses (below). THE PROPERTY THE MODULE EXISTS FOR: an illegal transition cannot be EXPRESSED. Not rejected at runtime — expressed. plan() takes only a transition id the graph contains, and anything else needs PrivilegedBypass. There is deliberately no third path, because the third path is what every workflow engine grows under deadline: a force boolean, then a force boolean with a comment. -->

    • S2.1.a Define typed stage identity, stage metadata, initial/terminal flags, and domain-owned stage union registration. Evidence: stage definition type tests. <!-- StageSpec<TStage> + defineStages, returning a BRANDED StageUniverse. The brand is load-bearing: without it a caller can assemble a StageSpec[] inline and skip the function where the one-initial-stage and duplicate-name rules live, and a universe that was never checked looks exactly like one that was. Two refusals worth naming. Zero initial stages means nothing can be created; TWO means the answer to "where does a new object start?" depends on which row the reader saw first, and both readers are confident. meaning is required on every stage — a sentence saying what being in it means — because that is the sentence the next person adding a stage needs and the one nobody writes down. -->
    • S2.1.b Define typed transition identity, source set, target, trigger, actor scope, guard set, side-effect intent, and reversal metadata. Evidence: transition schema tests. <!-- Transition<TStage> carries all eight. from is a SET rather than one stage: "cancel" is one transition available from three stages, and modelling it as three transitions is three places to add the fourth guard to. THE FIELD THAT DOES REAL WORK IS effects. It is declared rather than discovered, because it decides whether a transition can be undone by writing the old stage back — publish sends email, and unpublishing does not unsend it. The graph therefore REFUSES a transition that claims reversible: true while causing a notification, a publish or a payment: a reversal that only looks complete is worse than none. reversal also names the transition that undoes it, and the pair is checked to actually lead back, so "reversible" is not a boolean somebody set once. -->
    • S2.1.c Build a transition graph that rejects unknown stages, duplicate transitions, unreachable stages, unintended cycles, and missing initial state. Evidence: graph validation tests. <!-- buildStageGraph over twenty violation kinds. Every check is one a reviewer would otherwise do by hand over a diagram, and the two a diagram makes hardest are the ones that matter: REACHABILITY (a stage nothing can get to is either dead or missing an edge, and both look like a complete diagram) and CYCLES (a loop is invisible in a list of transitions and obvious only once an object is stuck in one). A cycle is legitimate — rework, reopen, re-review — and is declared by marking the edge that CLOSES it, not every edge on it; requiring all of them would mean marking most of a healthy graph, which is how a rule stops being read. A self-loop needs no revisits at all, because reentrant on the stage is already that declaration, and a self-loop on a stage that did not declare itself gets the better error (self-transition-not-reentrant) rather than "there is a cycle somewhere". Cycles are reported per COMPONENT: an eight-edge cycle reported eight times reads as eight problems. -->
    • S2.1.d Expose allowed-transition queries with exact guard requirements but no mutation. Evidence: query known-value tests. <!-- allowedTransitions, isLegalTransition, deadEnds. The queries return the EXACT guard ids each transition demands, not "some guards apply" — a UI built on this can answer "why is this disabled?" in the client instead of by attempting the transition and reading the error. deadEnds is the query nobody asks for and everybody needs: a stage with no exit that never declared itself terminal reads as "still in progress" on every dashboard forever. NoInfer on the stage parameter is what makes the typing real. Without it TypeScript takes a candidate from both arguments and widens the union, so querying one domain's graph with another domain's stage typechecks — the exact mistake S2.1.g exists to prevent. The @ts-expect-error in the spec was UNUSED until this was added, which is how it was found. -->
    • S2.1.e Execute transitions only through a typed authorized command and expected revision. Evidence: direct-state-write negative tests. <!-- TransitionCommand requires a non-null expectedRevision: unlike a create, a transition always has a previous state, and one applied without it is a lost update that leaves the object in a stage nobody chose. The module exports NO writer — plan() returns a decision and a planned write, and the caller's persistence layer performs it. A module that can both decide and write is one where the decision can be skipped by calling the writer, and the writer is always the more convenient of the two; the negative test asserts the surface contains no set/apply/commit/write/ force/move symbol at all. The check order is deliberate: authorization first (an unauthorized caller learns nothing about which transitions exist), then existence, then GUARD COVERAGE, then the verdicts. Coverage before verdicts is the one that matters — a command carrying two verdicts for a three-guard transition has a guard nobody ran, and folding the verdicts it did carry reports a pass. Missing verdicts come back indeterminate, which is a different outcome from blocked: one is a policy message for the user, the other is an incident for whoever owns the unreachable dependency. -->
    • S2.1.f Require an explicit privileged bypass type carrying authority, reason, scope, evidence, and audit linkage. Evidence: compile/runtime bypass tests. <!-- PrivilegedBypass + planBypass, a SEPARATE entry point rather than a parameter on plan. Sharing one would make the override reachable by passing one more argument, and every override in every workflow engine started as one more argument. All five fields are required and an empty string is not a value — a reason of "" is worse than no reason, because it makes the audit row look complete. Six runtime cases, one per field plus a whitespace-only evidence entry. The compile-time half is that TransitionCommand has no bypass field at all, asserted with a @ts-expect-error. A bypass overrides the WORKFLOW and not the vocabulary: a target stage the universe does not define is still refused. -->
    • S2.1.g Instantiate two domains with disjoint stage unions to prove no shared hard-coded stages. Evidence: cross-domain compile fixtures. <!-- A story desk (draft/review/legal/published/retracted) and an incident console (triage/mitigating/monitoring/resolved) — no shared stage name, no shared transition, no shared guard. Both build a graph and both answer the same queries, which is what "the kit assumes no domain vocabulary" has to mean. A kit that quietly assumed 'draft' | 'review' | 'published' would pass every runtime test in the file and be a content-authoring kit wearing a general name. Two @ts-expect-error directives make it a COMPILE fixture rather than a naming convention, and they are checked by tsc --noEmit on tsconfig.spec.json — an unused directive is itself an error, so the fixtures cannot rot into assertions that no longer assert. -->

    The extraction the cell names is the two graph analyses. plugins.ts carried its own reachability loop for the declarative model a plugin contributes, and two implementations of "can you get there from here?" is one that disagrees with the other the first time a source set or a self-loop is involved. Both now call unreachableFrom/cyclesIn over an untyped EdgeSet — the plugin contract is a WEAKER shape than an engine transition (no trigger, no effects, no reversal), so it lifts into that rather than into a full graph, and nothing has to invent the fields the plugin never stated.

  • S2.2 Make transition guards pure and evidence-driven. Evidence: known-value tests cover pass, fail, stale, missing, and unavailable evidence. <!-- libs/oshun/workbench-kit/src/guards.ts, published as @oshun/workbench-kit/guards, 40 cases. 896 tests in the kit; all nine Nx targets green. PURITY IS ENFORCED, NOT REQUESTED. The whole value of a guard's answer rests on it being reproducible: a guard that reads the clock says yes on Tuesday and no on Wednesday about the same object, and one that reads randomness cannot be replayed at all — so "the publish gate passed" stops being a claim anybody can check. evaluateGuard therefore takes the ambient world away for the duration of the call rather than asking authors not to use it. -->

    • S2.2.a Define immutable guard input containing current object/revision, actor/scope, target transition, versioned evidence refs, policy, and clock snapshot. Evidence: guard input contract. <!-- GuardInput carries all seven. THE DESIGN IS THE ABSENCE: there is no client, no repository, no config object and no clock — only a clock SNAPSHOT, which is a value. A guard cannot reach for what it was never handed, and adding a field here is a kit decision a reviewer sees. EvidenceRef is versioned because "the rights check passed" is meaningless without which rights record it passed against: the record changes and the decision it justified does not. The input is deep-frozen before the call, so a guard that writes to its subject fails instead of making its own answer depend on whether it ran first. -->
    • S2.2.b Prohibit database, network, provider, current-time, random, global configuration, and mutation access inside guard evaluation. Evidence: boundary lint/type tests. <!-- A RUNTIME HARNESS rather than a lint rule, because guards live in domain packages the kit's lint config does not reach, and because a lint rule is advice while a throwing stub is a failing test. Date is replaced wholesale (a guard reaches the clock through the constructor as readily as through now(), and stubbing only the static leaves the more common spelling working), Math.random and fetch likewise; Date.parse and Date.UTC stay, because a guard has to read the instant it was given. Everything is restored in a finally — a harness that could leave Math.random replaced would take the process down the first time a guard failed, and that is asserted. ANY throw becomes evaluator-error, including a plain TypeError: the alternative is a harness that classifies some failures and lets the rest escape into the caller, where a broken gate becomes a 500 instead of a gate outcome. THE LIMIT, STATED IN THE MODULE: a guard that captured a database client in a closure before evaluation is invisible to a harness that only controls globals. That half belongs to the package hosting the guard; this half covers what people reach for without noticing. -->
    • S2.2.c Define pass, fail, stale, missing, unavailable, unsupported, and evaluator-error results with stable reason/remediation fields. Evidence: result union tests. <!-- Seven statuses and a closed seven-value remediation vocabulary, with a default mapping per status and a per-answer override. The reason there are seven and not two: pass/fail are decisions about the SUBJECT and the other five are about the evidence or the guard, so collapsing any of them into fail tells a user to fix something that is not wrong with their work — refresh evidence, produce evidence, retry later, this does not apply, and page the team that owns the broken guard are five different actions. evaluator-error must never read as a fail, because a broken guard reporting failure is a broken guard nobody fixes. Remediation is a closed vocabulary rather than prose for the reason the error module gives: the first person to paste a working command into a free-text field puts a credential in a log line. -->
    • S2.2.d Compose multiple guards deterministically with explicit blocker/warning and short-circuit policy. Evidence: composition truth-table tests. <!-- composeGuards sorts by guard id before folding, so a decision does not depend on the order a host registered guards in — asserted by composing the same results forwards and backwards and comparing fingerprints. A seven-row truth table covers blocking vs advisory against pass/fail/unavailable/unsupported. Two decisions worth arguing with. An unsupported guard ABSTAINS: not a pass, because "this guard does not apply" is not evidence the transition is safe, and not a block, because a guard that does not apply cannot refuse. And short-circuit is OFF by default — it saves work and produces a UI that reveals one problem at a time, so a user fixes four things in four round trips. When it is on, the guards that never ran are NAMED in the decision, so one reported problem does not read as one problem. An unregistered guard is treated as blocking: a result from a guard nobody registered is more suspicious than one from a guard that was. -->
    • S2.2.e Preserve every consulted evidence version and policy version in the result. Evidence: provenance assertions. <!-- The HARNESS attaches provenance, not the guard, so a guard cannot forget to. Consulted evidence is sorted, so two runs record it identically, and the policy id and version are on every result including the ones that consulted nothing. The staleness budget is applied here for the same reason: a guard that checks its own evidence age is one that can forget to, and forgetting produces a PASS — the direction a missing check must never fail in. It applies only to a pass; a fail stays a fail however old the evidence is, because the subject is wrong either way and "refresh evidence" would send the user to fix the wrong thing. -->
    • S2.2.f Re-evaluate identical input byte-for-byte to the same semantic result. Evidence: determinism/property tests. <!-- decisionFingerprint folds every guard id, status, reason key, remediation, policy version and consulted evidence version into one string. Comparing fingerprints rather than results is the point: comparing results passes whenever both runs happen to be wrong in the same way, and a fingerprint that never changed would make every re-evaluation agree — so the suite also asserts it CHANGES when a single evidence version or the policy version moves. -->

    The purity harness is the part worth reusing elsewhere in this ledger. Every other "must be deterministic" cell in Phase S is currently a convention; this is the shape that makes one enforceable — take the capability away for the duration of the call, restore it in a finally, and turn the resulting throw into a typed outcome rather than an exception.

  • S2.3 Add optimistic concurrency with expected revision. Evidence: concurrent updates return a typed conflict and never last-write-win silently. <!-- evidence/partial: a–c, e and f are done and checked below; libs/oshun/workbench-kit/src/concurrency.ts, published as @oshun/workbench-kit/concurrency, 70 cases plus a 10-case race harness and 13 committed conflict goldens. 980 tests in the kit at that point; all ten Nx targets green (the tenth is new — see S2.3.b). The PARENT stays unchecked because S2.3.d's named evidence is a browser/BFF journey and the kit has no host to run one in until S7. THE THREE HOLES A PRECONDITION DOES NOT CLOSE BY ITSELF, which is what this cell is actually about. (1) It is optional — a nullable expectedRevision makes "I mean to create this" and "I do not know what revision I am editing" the same value, and the second one is a lost update. (2) The conflict is a dead end: a 409 with no payload tells a user to reload, which throws away what they typed. (3) The retry helpfully refreshes the revision, which defeats the check on exactly the requests the check existed for. Each is closed by a type rather than a convention. One collateral find, in S1.10's leak scan rather than here: the api-key canary detector required an unbroken alphanumeric body after sk-, so it matched only the unsegmented form no issuer mints and passed every real sk-live-… / sk_test_… / sk-proj-… key. Found by planting a realistic key in a DECLARED conflict field, where the closed-shape rule cannot see it and only the value scan can. Detector broadened, boundary pinned by a new lookalike case. -->

    • S2.3.a Require expected revision or an explicitly authorized create-if-absent precondition on every mutable command. Evidence: command schema tests. <!-- Precondition is a two-variant union — a revision, or a create-if-absent carrying the grant that authorized it, the natural key that makes absence checkable and idempotent, and a reason. THE DESIGN IS THAT THERE IS NO THIRD VALUE and no absence: parsePrecondition(null) is refused rather than defaulted to a create, because defaulting is the repair that turns every malformed update into a brand new object. The "on every mutable command" half is a SCHEMA audit rather than a runtime check, because the question is "is there a mutable command anywhere in this workbench that can be sent without one?" and no amount of testing the happy path asks it. auditCommandSchemas also runs the cross-checks that only exist between two columns: a command declaring itself safe to replay while carrying no idempotency key is the combination that turns a network timeout into a double application, and each column reads as fine alone. -->
    • S2.3.b Compare expected revision and mutate/increment revision atomically in persistence. Evidence: real-database race tests. <!-- pnpm nx revision-race @oshun/workbench-kit — tools/revision-race.mjs, a tenth Nx target. THREE real concurrent PostgreSQL connections, run green on PostgreSQL 16.14: an atomic UPDATE … WHERE revision = $expected produced exactly one winner and no lost update, and INSERT … ON CONFLICT (natural_key) DO NOTHING produced exactly one creator. THE TWO CONTROLS ARE WHY THAT MEANS ANYTHING. The same race runs against a read-then-write UPDATE and a check-then-INSERT, and both MUST come back red — they did, losing an update and creating a duplicate. Without them a green run is indistinguishable from contenders that never overlapped, or from a harness that measures nothing. The harness itself is @oshun/workbench-kit/testing/concurrency, so the property is defined once for all seven domains rather than seven times. Its two calibration stores live in ./testing/doubles because they hold state, which is the kit's own rule about test utilities applied to the kit's own test utilities. And the target FAILS rather than skips with no database: a skipped race that reports success would let "S2.3.b is green" mean "nobody has a database". It is therefore not in the CI inventory gate, which has no database service. -->
    • S2.3.c Return typed current/expected revision, affected target, safe current metadata, and compare/reload remediation on mismatch. Evidence: conflict envelope fixtures. <!-- 13 golden envelopes + 3 refusals in evidence/v1-workbenches/workbench-kit-conflict-fixtures.json, written by hand in the generator and rebuilt by conflictEnvelope() in the spec. Two independent statements of the same fact: a corpus derived from the code it checks agrees with any bug that code has. THE AFFECTED TARGET is the field that makes a conflict actionable on the six non-single paths. "Your import failed" versus "row sku-40199 was edited by someone else at 14:02" is the whole difference, and the goldens cover all seven paths so a corpus of single-object conflicts cannot pass for coverage. SAFE METADATA is a closed set of scalars checked as a closed set — the union of detail keys across every golden must be inside it, which is the assertion that fails the first time somebody adds a helpful currentTitle. updatedBy is an actor id and never a name, inherited from ActorRef refusing display data rather than re-decided here. Two registry changes came with it. reload-and-compare is a new remediation action and the first SELF-SERVICE one, so publicError's requiresOperator default now reads off a named list instead of action !== 'none' — routing a routine conflict to an operator buries it in a support queue. And state.subject_absent is a new code because an absent subject is NOT a stale revision: there is nothing to reload, and "reload and compare" sends the user to something that will 404. It also gave the registry its first 404, so the transport cross-check row mapping rpc not-found to HTTP 404 can now fire at all. -->
    • S2.3.d Preserve the rejected client intent for compare, retry, copy, or support recovery without applying it. Evidence: browser/BFF conflict journey. <!-- evidence/partial: the SUBSTANCE is implemented and covered by 9 cases — preserveIntent (payload verbatim, precondition left STALE), compareIntent, copyableIntent, supportViewOfIntent (scoped like the S1.10.f cause lookup, expiry checked so the two refusals do not leak existence) and resumeIntent, which is the ONLY route from an intent to a write and requires a fresh observation, an explicitly restated precondition and a new idempotency key. Restating the rejected precondition is itself refused: resuming with it either fails again or, if the head happens to have returned, applies against a state the author never saw. The checkbox stays UNCHECKED because the named evidence is a browser/BFF journey and there is no host to run one in: the kit has no consumers yet (S1.2's export map is still empty of ratified capabilities) and the shared application shell is S7. Writing a throwaway app to satisfy the wording would be a fixture asserting itself. Re-check this when the first workbench adopts the kit. -->
    • S2.3.e Apply the same precondition to bulk, import, job, restore, transition, and compatibility mutation paths. Evidence: alternate-path negative tests. <!-- Seven named paths, ONE applyPrecondition. Seven copies of the comparison would be six chances to write >= where === belongs, and the copy that drifts is always the one nobody reads — so the path is a parameter that a conflict reports, never a reason to fork the logic. Asserted by classifying the same conflict on all seven and comparing. Four negatives, each a real per-path bug: a bulk request with ONE request-level revision (right about one subject at most, so the rest are unchecked writes); a job whose precondition is checked when it is queued rather than when it runs (the gap between the two IS the window the check exists for); a restore whose precondition names its SOURCE revision rather than the current head (the mistake is nearly invisible — the source is the revision the request is about, and using it makes the restore succeed no matter what has happened since, which is exactly the change the user needed to be shown); and an automatically replayed restore, which writes a second revision every time. Coverage is reported HONESTLY: an audit of one command returns no violations AND names the six paths it did not look at, because an empty violation list otherwise reads as full coverage. preconditionOfTransition bridges S2.1 rather than duplicating it, so the transition path goes through the same function as the other six. -->
    • S2.3.f Prevent automatic retry from changing the expected revision or overwriting a remote update. Evidence: retry conflict tests. <!-- THE SHARPEST CELL IN S2.3. An automatic retry that re-reads the current revision and tries again is a last-write-win wearing an optimistic-concurrency costume: it defeats the precondition on exactly the requests the precondition was carried for, using the mechanism meant to be safe. So retryDecision has no code path that produces a refreshed attempt — asserted exhaustively over every failure kind and every budget, not just the obvious one — and a conflict returns stop-conflict, a decision distinct from stop so a caller routes it to the recovery flow rather than to an error toast. Only unknown-outcome may be retried, and the safety argument is a PAIR: the idempotency key covers "it already applied" and the precondition covers "somebody else applied something". Either alone is insufficient, which is why retryViolations reports a changed key and a changed precondition as separate findings, along with a re-minted next revision and a counter that did not move (a budget that never advances is infinite). All four are reported at once rather than first-wins. Backoff and jitter are deliberately absent and said so: timing belongs to the transport, and this module answers only "what may change between attempts?". --> <!-- evidence: libs/oshun/workbench-kit/src/concurrency.ts + concurrency.spec.ts (70 cases), src/testing/concurrency.ts + concurrency.spec.ts (10), tools/revision-race.mjs, scripts/v1-workbenches/generate-workbench-kit-conflict-fixtures.mjs (+12 structural tests), evidence/v1-workbenches/ workbench-kit-conflict-fixtures.json. -->

    The reusable shape here is the CALIBRATED HARNESS. S2.3.b is the first cell in Phase S whose property lives outside the code — no type makes a write atomic — and the answer was to ship the race in the kit and run it against a real engine with deliberately broken controls beside it. A conformance harness with no negative control proves that it ran, not that it can fail; every later cell that has to check something a database, a browser or a provider does should carry one the same way.

  • S2.4 Add immutable revision metadata and parent references. Evidence: history integrity test detects gaps, cycles, and hash mismatch. <!-- libs/oshun/workbench-kit/src/revisions.ts, published as @oshun/workbench-kit/revisions, 49 cases; plus tools/history-constraints.mjs, an eleventh Nx target running against a real PostgreSQL. 1,031 tests in the kit; all eleven targets green. A REVISION HISTORY IS THE PART OF A WORKBENCH EVERYBODY READS AND NOTHING CHECKS, because append-only things are assumed to be fine. They are not: a parent pointer into a hard-deleted row, a sequence number two writers allocated, a merge with no recorded base, a content hash left behind by an in-place typo fix. Each is silent, and each turns the history from a record into a story. So the module is mostly checkHistoryIntegrity — seventeen violation kinds, every one of them reachable, asserted as a set rather than as a count so a kind nothing can produce fails the suite. A NINE-MUTANT RUN over the checker killed nine of nine after it killed one of my own: Object.is(input, -0) ? '0' : JSON.stringify(input) in the canonicalizer is DEAD, because JSON.stringify(-0) is already "0". The branch was removed and the assertion kept — the property is what matters, and a branch nothing can reach is a claim the reader believes and the code does not keep. -->

    • S2.4.a Persist stable revision ID, entity/workbench scope, schema version, author/delegate, timestamp, reason, and content hash. Evidence: revision metadata tests. <!-- RevisionRecord carries all eight plus sequence. The sequence is there AS WELL AS the lineage because the two catch different faults: a broken parent pointer is a structural break, and a reused or skipped sequence is a concurrency break that leaves the structure intact and looking fine. reason is required and non-empty, enforced by the integrity check and again by a database CHECK. A history whose entries say nothing about why they exist is a list of timestamps, and "why is this like this?" is the only question anybody actually brings to one. The author is an ActorRef, so a delegated author carries both identities — a row with one of the two can answer neither "who did this" nor "whose authority was used". -->
    • S2.4.b Persist one parent for linear history and explicit multiple parents/base for merge or divergence semantics. Evidence: lineage fixture tests. <!-- Lineage is three variants rather than a parents[] array: the array turns "is this the first revision?" and "is this a merge?" into arithmetic on a length that every reader has to redo, and it cannot make base required only where it means something. A MERGE REQUIRES ITS BASE. Without the common ancestor the three-way comparison was computed against, "which side won this field, and why?" has no answer for the rest of the object's life. Refused rather than defaulted, because the default would have to be a guess at an ancestor. merge-base-not-an-ancestor catches the base that is merely plausible. DIVERGENCE IS NOT A FAULT. Two revisions off one parent is a legitimate state; the integrity check stays silent about it and the reconstruction reports two heads and a branch point, for a UI to resolve. And the cycle walk is a THREE-COLOUR walk rather than a visited set, because a visited set cannot tell a cycle from a diamond — and a diamond is the normal shape of every correct merge, so the naive check would fire on all of them. -->
    • S2.4.c Prevent update/deletion of revision content and metadata outside governed retention/tombstone operations. Evidence: immutability negative tests. <!-- Three governed operations and nothing else: tombstone the content, place a legal hold, release one. immutabilityBreaches compares two states of the same revision and names every immutable field that moved, which is the check for the edit that happens OUTSIDE the module — a migration, a support script, an ORM save() on a loaded entity. A TOMBSTONE NEVER CLEARS THE HASH. That single decision is what lets an erasure request and a verifiable chain both hold: the content goes, the metadata and the original digest stay, and the history now says "this was removed", which is different from "this was never here" and different again from a gap nobody can explain. A legal hold BLOCKS a tombstone rather than losing to it. The collision between "delete my data" and "preserve this for litigation" is routine and a person resolves it; what the type does is make the collision visible at the moment of removal instead of letting whichever code path ran last decide. Releasing is by hold IDENTITY, never by position. THE SAME RULE AT THE LAYER THAT CAN HOLD IT: a BEFORE UPDATE trigger in tools/history-constraints.mjs refuses any change to an immutable column, verified against a real PostgreSQL — six in-place edits rejected (reason, timestamp, content hash, re-parent, un-tombstone, tombstone under hold) while the governed tombstone and the legal hold still go through. An application-level rule runs where somebody called it; this runs where the write lands. -->
    • S2.4.d Enforce unique revision sequence/identity and detect missing parent, cross-entity parent, self-parent, and cycle. Evidence: database and graph integrity tests. <!-- BOTH halves, and the split between them is the finding. DATABASE (pnpm nx history-constraints, real PostgreSQL 16.14): the engine refused 12 classes — duplicate identity, duplicate sequence (raced on two real connections, because a unique constraint is only interesting under concurrency), dangling parent, self-parent, empty reason, deleting a parent that still has children (no cascade: a cascade deletes history to keep a pointer tidy), and the six immutability edits above. GRAPH: the same run then inserts a history every column constraint accepts and hands it to checkHistoryIntegrity, which found 6 kinds the engine structurally cannot see — cycle, multiple-roots, schema-version-regression, sequence-gap, sequence-not-after-parent, timestamp-before-parent. Each is a property of the GRAPH rather than of a row. Either half alone would read as coverage. Learned in the doing: the checker HOLDS its cycle walk while a structural fault is outstanding, so a calibration fixture sliced out of the valid history produced missing-parent and no cycle. Suppressing the walk is right — a walk over dangling edges reports a loop that is really a missing row — but it means a fixture has to be built whole. -->
    • S2.4.e Compute/verify content hash from canonical stored content or referenced artifact bytes. Evidence: tamper detection tests. <!-- THE KIT DOES NOT HASH, and says so. Its contract runtime class may not reach a Node built-in and the Web Crypto digest is async, which would make every caller of a pure comparison async; the host has a real hash function. What the host does NOT have is a canonical form both ends agree on, and that is the half that produces mismatching digests for identical content, so it lives here. canonicalize REFUSES rather than coerces, and every refusal is a value whose JSON round trip is lossy: NaN and Infinity both serialize to null, so two different contents would hash the same; undefined vanishes from an object and becomes null in an array, so the same content hashes differently depending on where it sits. It also names the PATH — blocks[1].text — and refuses a circular reference while accepting the same object twice, because a shared child is not a cycle. Six verdicts, not two. Collapsing them sends somebody to investigate a tamper that is a privacy request being honoured: tombstoned is not tampered, and an artifact-backed revision returns artifact-pin-verified — a COMPARISON against the digest the reference pins, because a function claiming to have verified bytes it never read would be the fabricated success this module exists to prevent. -->
    • S2.4.f Reconstruct heads, branches, ancestry, and chronological audit links without mutable denormalized history. Evidence: reconstruction tests. <!-- NOTHING IS STORED. No isHead, no childCount, no depth. The property is asserted the way it has to be: append a child and the head moves, with nothing written to the old head. A denormalized head is a second source of truth, and the failure it produces is an object nobody can edit because the head it points at was rolled back — unfixable from the application, because the application believes the column. clockDisagreesWithSequence is the honest part. Two orderings exist and can differ: the sequence is what the database allocated, the timestamp is what a machine's clock said. Both are returned along with the fact that they disagree, rather than silently preferring one — when they do disagree, an audit rendered in timestamp order shows events in an order that did not happen. ancestryOf is breadth-first, so a merge shows both sides before their shared parent rather than one side's whole history first, and it returns a typed MISS for a revision that is not in the set: an empty ancestry is what a root has, so returning it for "not loaded" renders "this is the first version" over a revision that simply was not there. --> <!-- evidence: libs/oshun/workbench-kit/src/revisions.ts + revisions.spec.ts (49 cases, 9/9 mutants killed), libs/oshun/workbench-kit/tools/history-constraints.mjs (Nx target history-constraints, real PostgreSQL). -->

    Worth carrying forward: this cell and S2.3.b are the two where the kit stopped being able to prove its own claim alone, and both were answered the same way — put the rule in the layer that can hold it, then run a harness that shows what the OTHER layer catches that this one cannot. A constraint set with no graph check beside it, or a graph check with no constraints beneath it, each reads as complete on its own.

  • S2.5 Add semantic diff adapters for scalar, list, tree, graph, timeline, and binary-metadata changes. Evidence: golden diff fixtures. <!-- libs/oshun/workbench-kit/src/diff.ts, published as @oshun/workbench-kit/diff, 78 cases; 21 golden scenarios in evidence/v1-workbenches/workbench-kit-diff-fixtures.json written by hand in the generator and replayed by the spec against the real adapters, plus 12 structural tests over the corpus. 1,110 tests in the kit; all eleven Nx targets green. THE CORPUS IS WEIGHTED TOWARDS WHAT A STRUCTURAL COMPARISON GETS WRONG rather than what it gets right, because the wrong answers are all defensible: a reparented section reported as a delete and an add, a rerouted edge that loses its identity, a conform to a different frame rate that reports nothing at all, one cut reported as three changes, and identical bytes whose metadata disagrees about them. A TWELVE-MUTANT RUN killed 10 of 12 on the first pass. Both survivors were in the split detector and both were masked by a later check that happened to catch the same fixtures; the two cases that distinguish them — pieces that OVERLAP rather than abut, and a re-identified clip covering the same span — were added and both mutants then died. Neither case was obvious from reading the code, which is the argument for running mutants at all. Coverage is three matrices rather than a count: every adapter, every change type, and every non-value DiffValue kind. The refusal states are what the contract exists for, and a corpus of ordinary changes proves none of them. -->

    • S2.5.a Define common diff operation identity, path/anchor, before/after revision, change type, actor/provenance, and inaccessible/unsupported states. Evidence: shared diff contract. <!-- DiffOperation: a derived id (adapter, path, change type — so two runs agree and a cursor can address one), a sortable path, a translation KEY with scalar parameters, both StableAnchors, and provenance carrying both revisions and the adapter version. THE ANCHORS ARE S1.4.e's, not a diff-only path. A client that can already jump to an anchor can jump to a change, and the time-base and graph-target vocabulary arrives with it instead of being re-modelled slightly differently here. Both sides are nullable and at least one is set: an added thing has no before-anchor. DiffValue is eight variants and every one of them is load-bearing. absent/null/default are three different facts — not sent, sent as nothing, not sent and the schema fills it in — and collapsing them makes "somebody cleared this" and "somebody never touched this" the same row. redacted/inaccessible/unsupported are three different refusals: it changed and you may not see to what; it may have changed and you cannot be told; nothing here can compare these two, which is a gap in the adapter rather than a fact about the content. opaque is bytes. provenance.actor is NULL between two revisions with several authors in between. Naming one of them would be a guess presented as a fact. -->
    • S2.5.b Implement scalar/object-field comparison with absent/null/default, unit, locale-neutral, enum, and sensitive-value semantics. Evidence: scalar golden fixtures. <!-- Five semantics, and three of them are judgements rather than mechanics. UNITS: 1000 m and 1 km are one distance, so the magnitudes are compared after conversion and the result is a PRESENTATION operation naming the respelling. Reporting a change would send somebody to review a unit preference; reporting nothing would hide that the preference moved. A change of DIMENSION is replaced, not a resize. An unknown unit is a gap in the conversion table and the side carrying it says so, rather than the field being reported as changed. LOCALE-NEUTRAL: comparison is over typed values, never over formatted text — "1,234.5" and "1.234,5" are one number, and a diff that compared the strings would report every locale switch as an edit. ENUM: a value that is no longer a member is not a change to the field, it is the vocabulary having moved underneath it. Reported as unsupported, because an ordinary change sends somebody to fix content that was valid when it was written. SENSITIVE: the comparison still runs and only the values are withheld, parameters included. A diff that skipped sensitive fields would let a change hide behind its own sensitivity. -->
    • S2.5.c Implement ordered/unordered list insert, delete, move, replace, duplicate identity, and item-field changes. Evidence: list golden fixtures. <!-- THE LIMIT IS THE POINT. With item identity a move is a move and an edit is one row; WITHOUT it a move is indistinguishable from a delete plus an add — not because this implementation is weak but because the information is not there. So listCanDetectMoves reports the capability and the unkeyed path runs a real longest-common-subsequence rather than a similarity heuristic that guesses. The LCS is O(n*m) and that tradeoff is written down: right for the tens-to-hundreds a workbench field holds, wrong for a hundred-thousand-line file — and a field that large is an artifact, which goes through the binary adapter. A DUPLICATED KEY is reported and then skipped. Which of two same-keyed items is being compared with which is a guess, and a guessed pairing produces a diff that is confidently wrong. An UNORDERED list is compared as a set, where a move cannot exist: reordering a permission list produces nothing at all. -->
    • S2.5.d Implement tree add/remove/move/reparent/reorder/field changes with stable node identity and cycle-invalid state. Evidence: tree fixtures. <!-- A reparent is moved and a sibling shuffle is reordered, and they are separate change types because they call for different review attention — one collapsed "moved" makes the cheap one look expensive. A node that moved to a new parent ALSO lands at a new position, and the reorder is deliberately suppressed for it: one action shown as two rows is how a review of six edits reports twelve. A CYCLE IS NOT A TREE. The nodes in the loop are reported unsupported and the walk does not happen — comparing a structure that cannot exist either loops or produces an answer about it. -->
    • S2.5.e Implement graph node/edge/port/property add/remove/reconnect and layout-versus-semantic change separation. Evidence: graph fixtures. <!-- An edge whose endpoints moved is changed, not remove-plus-add: the edge has an identity, and a review that loses it cannot say "this connection was rerouted", which is the only useful sentence about it. A retyped node IS a replacement, because nothing about a gain node survives becoming a reverb. LAYOUT IS SEPARATED FROM MEANING. A node's position is a real stored property and a change to it is a presentation operation, so a client renders "3 changes, and 44 layout changes" rather than "47 changes" — the second is a review nobody reads, and the reason they stop is that an afternoon of tidying looks exactly like an afternoon of work. -->
    • S2.5.f Implement timeline/temporal add/remove/move/trim/split/merge, time-base conversion, track/lane, and parameter changes. Evidence: timeline fixtures. <!-- THE TIME BASE IS PART OF THE POSITION. Frame 300 is 10s at 30fps and 10.01s at 29.97, so a comparison of the numbers reports "unchanged" for a conform that moved every clip in the sequence. Positions convert through the exact rationals S1.4.e already carries and are compared by cross-multiplication — 29.97 is 30000/1001 and comparing it as a float is how a cut lands one frame out after a round trip. The converse is asserted too: 10 seconds IS frame 300 at 30fps, and a rebase that changed nothing must not report a move. ONE CUT IS ONE ACTION. A split is detected on source identity plus CONTIGUITY plus exact span coverage, so it is not claimed whenever somebody adds a clip from the same source nearby; a merge is the same detector run backwards. A trim names WHICH edge moved, because "the head was trimmed" and "the tail was trimmed" are different edits. A MUSICAL time base is refused rather than converted: seconds per pulse depends on a tempo map the timeline does not carry, and converting would mean inventing a tempo and reporting the invention as a position. -->
    • S2.5.g Implement binary metadata comparison for identity/hash, size, media type, dimensions/duration, codec, channels, embedded metadata, provenance, and availability without pretending to diff opaque bytes. Evidence: binary-metadata fixtures. <!-- The bytes are opaque and the twelve things around them are compared properly, which is both the honest answer and the useful one: "the codec changed and the duration did not" is actionable and "the file is different" is not. THE INTERESTING CASE IS THE INCONSISTENCY. Identical digests with different dimensions means one side's metadata was written by something that did not read the bytes — a finding rather than a change. Three fields are excluded from it with reasons: provenance is a record ABOUT the bytes rather than a property of them, mediaType is a label a re-ingest can legitimately correct, and embedded metadata may live in a sidecar the digest does not cover — flagging that would report an inconsistency every time an archive was re-catalogued. An unretrievable artifact is inaccessible, never removed: the second sends somebody to find out who deleted it. -->
    • S2.5.h Provide stable ordering, accessible summaries, pagination, and redaction across all adapters. Evidence: cross-adapter contract tests. <!-- All four live OUTSIDE the adapters, in assembleDiff. An adapter that ordered its own output would be a second ordering, and two orderings disagree the first time somebody adds a change type. ORDERING is total — path, then change type, then id — and asserted by assembling the same operations forwards and backwards and comparing bytes. SUMMARIES are counts plus a per-operation translation key with scalar parameters; never a prose sentence, for the reason the error module gives. Redacted, inaccessible and unsupported operations are COUNTED: a diff over restricted content that looked empty would say "nothing changed", which is the one wrong answer a reviewer cannot recover from. PAGINATION is by cursor on the last operation id, not by offset. A diff is recomputed per request, so an offset skips or repeats rows — that is the failure mode here rather than an edge case. REDACTION hides VALUES and never rows, parameters included, and is prefix-matched on path SEGMENTS so payroll is not caught by a rule about pay. The cross-adapter test runs all six through the same assembly and asserts every one produces at least one operation — a contract test over five would pass with the sixth unwired. --> <!-- evidence: libs/oshun/workbench-kit/src/diff.ts + diff.spec.ts (78 cases, 12/12 mutants killed), scripts/v1-workbenches/generate-workbench-kit-diff-fixtures.mjs (+12 structural tests), evidence/v1-workbenches/workbench-kit-diff-fixtures.json (21 scenarios, 29 operations). -->

    A method note for the rest of the ledger: the two mutants that survived here both survived because a LATER check happened to reject the same fixtures, not because the code was right. That is the class of bug a passing suite cannot show you and a mutant can — a guard that has never been the reason anything failed. Worth running a short mutation pass on any cell whose value is a decision procedure rather than a shape.

  • S2.6 Add restore-as-new-revision. Evidence: restoring never mutates or deletes history. <!-- libs/oshun/workbench-kit/src/restore.ts, published as @oshun/workbench-kit/restore, 33 cases; plus tools/restore-transaction.mjs, a twelfth Nx target running against a real PostgreSQL. 1,144 tests in the kit; all twelve targets green. "RESTORE THIS OLD VERSION" IS THE FEATURE MOST LIKELY TO BE IMPLEMENTED AS AN UPDATE, and that implementation is correct exactly once: after it, the thing it overwrote is the only record of what was there and it is gone. So the cell's own check is historyHarms, which compares the record set before and after and reports anything that was not a pure append — over the SETS rather than over the code, so it catches the update however it was performed, including by a migration that never went near this module. This cell is also where the pieces built in S2.3–S2.5 are used rather than restated: the precondition and its restore-specific check come from S2.3, the record and lineage from S2.4, the preview diff from S2.5. The only new vocabulary is what a restore adds — the refusals, the downstream impact, and the outbox. -->

    • S2.6.a Authorize restore against object, source revision, current head, target stage, and protected-field policy. Evidence: restore authorization tests. <!-- Ten refusal kinds, every one reachable and asserted as a set. THE ONE THAT MATTERS IS source-is-tombstoned. A revision whose content was removed under an S2.4 governed operation cannot be restored, refused by IDENTITY rather than by policy so no configuration can enable it: a restore that reinstated erased content would undo a legal obligation through a feature nobody thinks of as a data path. grant-missing is a separate grant from editing, because restoring reinstates content the current editors chose to move away from. source-is-the-head refuses a no-op that would mint a history entry recording that nothing happened. A diverged history has no single head, so actualHead returns null rather than picking a branch — choosing which branch is current is not this function's decision. Authorization runs BEFORE anything about the content, so a caller with no grant learns nothing about which fields the object protects — asserted by checking that the unauthorized case returns exactly one refusal even when a protected field would also have been touched. The preview operations are a PARAMETER rather than something this recomputes, so the decision is made against the same preview a human was shown rather than against content the approver never saw. -->
    • S2.6.b Preview semantic differences, dependency/gate staleness, rights or consent impact, and irreversible downstream effects. Evidence: restore preview fixtures. <!-- THE PREVIEW IS NOT THE DIFF. What changed is the easy half and the S2.5 adapters already answer it; the half that decides whether a restore is safe is everything downstream. The case that makes the point is a rights withdrawal: the CONTENT is identical to what was approved, the permission that made it publishable is gone, and a preview showing only the diff shows a green screen. Gate staleness keys on whether the restore changes anything SEMANTIC, not on whether it changes anything at all. A restore that only respells a unit leaves every verdict standing, and saying otherwise trains people to re-run gates for nothing — and a gate people re-run for nothing is a gate people click through. Irreversible effects are filtered through S2.1's EFFECT_KINDS: writing an old stage back undoes an internal write and does not unsend an email, so only the second kind belongs in a warning. Everything is sorted, so two previews of one restore read identically. -->
    • S2.6.c Require expected current revision and reject changed head before commit. Evidence: restore race test. <!-- pnpm nx restore-transaction @oshun/workbench-kit — two REAL concurrent PostgreSQL transactions restoring off the same head: exactly one commits and one is refused as head-moved. A WRONG TURN WORTH RECORDING, and it is now written in the tool. The obvious implementation is SELECT … FROM head WHERE revision_id = $expected FOR UPDATE, and it does not work: with a DERIVED head (S2.4.f) there is no row the winner modifies — appending a child does not touch its parent — so the lock serializes the transactions and their SNAPSHOTS still both predate the other's insert. The first run had one commit and one duplicate-key ERROR rather than one clean refusal. What actually decides an append-only race is the UNIQUE (entity, sequence) constraint: the next sequence is derived from the head the caller named, so two restores off one head necessarily collide, and 23505 IS "the head moved". SERIALIZABLE would also work, at the cost of a retry loop. In-kit, the same precondition is refused before it reaches a database: S2.3's restoreTargetsHead catches the precondition that names the SOURCE — the mistake that makes a restore succeed no matter what has happened since. -->
    • S2.6.d Create a new immutable revision whose content derives from and links to the selected historical revision. Evidence: restore lineage assertions. <!-- PARENT = THE HEAD, BASE = THE SOURCE, and that is exactly why S1.4.d has two fields. The new revision COMES FROM the head (that is where it sits in the history) and its content was EXPRESSED AGAINST the revision being restored. A model with one field has to choose: with parent = source the history forks and the intervening revisions fall off the main line; with base = head, "what is this a copy of?" is gone and the restore is indistinguishable from somebody retyping the old text. The schema version comes from the HEAD, not from the source. Content cannot un-migrate — S2.4's schema-version-regression refuses exactly that — so a restore across a migration restores content that has been brought forward, and recording the old version would make the history claim otherwise. Asserted by planning a restore over a migrated head. The planned append is fed straight back into checkHistoryIntegrity and must produce zero violations: the two cells check each other rather than each asserting its own shape is fine. -->
    • S2.6.e Preserve intervening revisions and current history; never reassign old heads or delete descendants. Evidence: history integrity tests. <!-- Six harm kinds, every one reachable: a removed revision, a mutated one, a re-pointed parent, more than one revision for one restore, nothing added at all, and an appended revision that does not follow the head. The last two are the interesting pair. nothing-added is the SHAPE AN IN-PLACE UPDATE LEAVES BEHIND — the object changed and the history does not say so — and it is reported alongside revision-mutated for the same input, which is what that implementation looks like from the outside. added-revision-does-not-follow-the-head catches the restore that forks: it appends honestly and leaves the intervening revisions on a branch nothing points at, which reads as history preservation and is not. Checked against the DATABASE too: the restore harness loads the rows before and after a real committed restore and runs the kit's own historyHarms over them, so the property is asserted against what the engine holds rather than against what the planner returned. -->
    • S2.6.f Invalidate dependent gates/projections and emit audit/outbox events transactionally. Evidence: restore integration tests. <!-- An OUTBOX rather than a publish. A restore that writes the revision and then publishes is two operations, and the gap between them is where a crash leaves a restored revision nothing downstream heard about; rows in the same transaction as the revision cannot be half-written. Proved against a real engine with an INJECTED FAILURE after the revision insert — a genuine constraint violation rather than a thrown error in application code, because what is being tested is whether the ENGINE unwinds. Neither the revision nor the outbox rows survive: an orphan revision is a restore nothing heard about, and an orphan event republishes a revision that does not exist. The same restore then runs WITHOUT the injected failure and must succeed — a rollback proves nothing if the happy path never writes anything either. One event per invalidated gate and per dependent projection, plus the restore event itself even when nothing downstream is affected: a restore that emitted nothing would be invisible to anything not watching the revision table. And a replayed restore under one idempotency key writes ONE revision, or two identical revisions land and "who restored this?" has two answers. --> <!-- evidence: libs/oshun/workbench-kit/src/restore.ts + restore.spec.ts (33 cases), libs/oshun/workbench-kit/tools/restore-transaction.mjs (Nx target restore-transaction, real PostgreSQL: race, atomicity with an injected failure, append-only via the kit's own checker, idempotent replay). -->
  • S2.7 Add domain-neutral decision records. Evidence: rationale, actor, options, evidence refs, and supersession invariants are tested. <!-- libs/oshun/workbench-kit/src/decisions.ts, published as @oshun/workbench-kit/decisions, 43 cases, 12/12 mutants killed. 1,188 tests in the kit; all twelve Nx targets green. A WIKI PAGE HAS THREE PROPERTIES THAT MAKE IT USELESS WITHIN A YEAR: nobody can tell whether it is still in force, the alternatives that were rejected are gone, and the evidence it rested on is a link to something that has changed since. Each of the three is a field here with a rule attached, and 23 violation kinds — every one reachable, asserted as a set. THE SMOKE TARGET CAUGHT A REAL DEFECT. DECISION_GRAPH was a module-scope IIFE, and S1.2.f refused it: work at import time is exactly what "sideEffects": false tells a bundler it may drop, and a throw at import would take down a consumer that merely imported a type. Now a memoized decisionGraph(). Worth remembering as a class — a validated CONSTANT is the natural way to write this and it is a side effect. -->

    • S2.7.a Define decision ID/type/scope, owner, status, effective version, review date, and affected capability/object references. Evidence: decision schema tests. <!-- Six types and five scopes, both closed. The temptation is an open string so a domain can name its own kinds; the cost is that "who must approve an editorial decision?" becomes unanswerable at the kit level, and the approval rules are the reason the record exists. DecisionReview IS THE CELL'S BEST FIELD. Two variants rather than a nullable date, because a null means both "never needs review" and "nobody set one" — and the second is how a temporary exception becomes permanent. A standing decision is allowed and has to say why it stands. A scheduled review date that was already past when the record was written is refused: a decision that arrives overdue is one nobody notices is overdue. -->
    • S2.7.b Define considered option identity, description, status, benefits/costs, dependencies, risks, and rejection reason. Evidence: multi-option fixtures. <!-- THE OPTIONS THAT LOST ARE THE RECORD. A decision listing only what was chosen cannot answer the only question anybody brings to it later — "did they think of X?" — and "we did, and here is why not" is the entire value of having written it down. Exactly one chosen, every rejected or deferred option carrying a reason, and no chosen option carrying one (if it does, one of the two fields is wrong and a reader cannot tell which). A SINGLE-OPTION DECISION IS ALLOWED and has to say so. "We considered nothing else, because the vendor contract permits one shape" is a different claim from "we considered three and this won", and a schema that refused the first would just get three fabricated alternatives. -->
    • S2.7.c Define immutable evidence references with version/hash and access-safe summary. Evidence: evidence integrity tests. <!-- A reference to "the latency report" is a reference to whatever that document says today; a version plus a digest is a reference to what its author actually read, which is what makes a decision re-checkable. The safeSummary is scanned with S1.10's OWN detectors, reached through a new scanText export rather than a second implementation — the whole argument for a shared kit, and the alternative is two sets of detectors one of which is always behind the other. The case that motivates it: a decision archive is exported, read by people who were not there, and kept for years, so the first connection string pasted into a summary field outlives everyone who could recognise it. -->
    • S2.7.d Record proposer, decider, reviewers, delegated authority, timestamp, rationale, consequences, and residual risk. Evidence: actor/required-field tests. <!-- delegatedAuthority is separate from a delegated ActorRef because the two answer different questions: the ref says who was at the keyboard and the field says whose authority the decision carries, and a record with only the first cannot say whether the decision binds. SEPARATION OF DUTY IS A POLICY, NOT A UNIVERSAL. Architecture, rights and exception decisions bind other people, so the decider may not be the proposer; an operational decision an engineer makes and records is the normal case rather than a governance failure. Both directions asserted. Same shape for reviewers: a reviewed TYPE that left proposed with none named fails, because "the review did not happen" and "nobody recorded who did it" look identical afterwards. -->
    • S2.7.e Define propose/review/approve/reject/effective/superseded/reversed transitions and authorization. Evidence: decision state tests. <!-- BUILT ON THE S2.1 ENGINE, not as a switch. That is reuse in both directions: the decision lifecycle inherits reachability, terminality, cycle and effect-versus-reversal checking — a switch would have had none of them — and the engine gets its first consumer that is not a test or an archetype. The checks it inherits are asserted rather than assumed: no dead ends; the three terminal stages have no way out; a proposed decision cannot jump straight to effective, skipping the review nobody did; an approval by somebody holding only the propose grant is unauthorized; and a transition whose guard was never evaluated is indeterminate rather than a pass, which is the engine's coverage-before-verdicts rule. Every transition declares itself irreversible with a reason, because an approval is a record of what somebody decided and reversing one is a new decision that should read as one. -->
    • S2.7.f Supersede or reverse only by creating a new linked record while retaining original content. Evidence: lineage immutability tests. <!-- Exactly TWO fields may move on the original: its stage, to superseded, and its link. Everything else — the rationale, the options, the evidence, the people — is the record of what was believed at the time, and a supersession that edited it would make the archive agree with the present, which is the one thing an archive must not do. Six harm kinds, all reachable. The constructor and the checker are SEPARATE on purpose: a system where only supersede() can produce a supersession is a system that has not met a migration, so supersessionHarms compares two states and does not care how they were produced. resolveInForce walks the chain and returns a typed MISS rather than the last record it could reach, because both ways this goes wrong in a real archive produce a plausible answer otherwise: a successor that is not in the set (never written, or in a tenant this reader cannot see) and a loop (two records superseding each other, which a migration produces and nothing else notices). -->
  • S2.8 Add score-vector provenance without a default fabricated score. Evidence: absent scores remain absent end-to-end. <!-- libs/oshun/workbench-kit/src/scores.ts, published as @oshun/workbench-kit/scores, 49 cases, 11/11 mutants killed. 1,238 tests in the kit; all twelve Nx targets green. THE CHAIN THIS CELL IS ABOUT: the evaluator did not run, so the row is missing; the aggregate needs a number, so missing becomes zero; the UI needs a bar, so zero becomes a bar at the bottom; the gate needs a comparison, so a bar at the bottom becomes a fail. Nobody lied. Four reasonable local decisions produce a confident number describing nothing, and the suite is organised link by link along it. -->

    • S2.8.a Define score dimension ID, raw value, unit, direction, scale, applicability, and optional uncertainty independently from verdict. Evidence: score dimension schema. <!-- "score" is refused as a unit, by name, along with "points", "value", "rating" and the empty string: each is a refusal to name one, and a dimension without a unit can be averaged with anything. THREE DIRECTIONS, NOT TWO. target-is-best is the one that gets forgotten, and forgetting it inverts every comparison on that dimension — loudness has a target and a track further from it is worse on either side. A target-direction with no target and a target on a non-target direction are both refused, because one of the two fields is then wrong and a comparison picks the wrong one. The ordinal scale exists so a rubric CANNOT be averaged later. And uncertainty is separate from the value and checked against it: an interval that excludes its own measurement means one of the three numbers came from a different run, and a standard deviation over n=1 came from somewhere else entirely. -->
    • S2.8.b Persist evaluator/corpus/model/provider, preprocessing, policy, threshold, input revision/hash, run, and timestamp versions. Evidence: score provenance tests. <!-- Sixteen fields, long on purpose: every one has been the reason two scores that looked comparable were not. The cross-checks are the useful part — a model with no provider (the same model name behind two providers is two models, and the swap is invisible), a corpus with no version (a corpus that gained rows is a different ruler), a reading whose inputRevision is not the vector's subject (a stale reading presented as current is the fabrication that looks most like data). comparabilityFaults turns all of it into a refusal rather than a warning, and the sharp one is different-evaluator-version: the number moved because the RULER moved, and a chart of the two is a chart of a release note. A POLICY change is deliberately NOT a comparability fault — it makes two verdicts incomparable, not two measurements, and that is S2.8.d's problem. Both directions asserted. -->
    • S2.8.c Represent absent, not applicable, insufficient data, unsupported, unavailable, failed, stale, and redacted explicitly without numeric sentinels. Evidence: nonvalue state fixtures. <!-- Eight states, each a different action: run the evaluator, do nothing, get more input, retry, page the owner, re-run, ask for access. The expensive collapse is not-applicable into absent — the first must not count against a total and the second must. THE SENTINEL IS NOT REFUSED AT RUNTIME, IT IS UNTYPEABLE. A reading is a measured number or a named state and there is no third shape, so -1 is not something the type can hold. What IS checked at runtime is the wire: parseScoreReading refuses a payload carrying both a raw and a state, which is exactly the shape a fabricating adapter emits — it filled the number in and left the state behind. adoptSentinel is the one honest seam. A legacy feed that writes -1 exists; pretending otherwise means the -1 reaches an average. It converts at the boundary, refuses to guess (the caller declares the mapping) and RECORDS every conversion, so "900 of 4,000 rows were sentinels" is a fact somebody can see rather than a silence. -->
    • S2.8.d Preserve raw values when policy/threshold changes and recompute derived verdicts separately. Evidence: policy recomputation tests. <!-- applyThresholds returns the SAME reading object it was given — asserted by reference identity rather than by comparing fields, because a copy that is equal today can stop being equal. Three policies over one measurement produce pass, warn and fail, and the measurement never moves. Every verdict is stamped with the policy id and version it came from; without it, a verdict computed under v2 and one under v3 are the same row and a threshold change silently rewrites history. NO THRESHOLD IS NOT A PASS. A dimension the policy says nothing about is undetermined, because "nothing was said" is not "it was fine". -->
    • S2.8.e Serialize, persist, query, display, compare, export, and publish an absent score without creating zero, neutral, average, or pass. Evidence: end-to-end absence tests. <!-- Every hop, asserted: a non-value round-trips through JSON as a non-value; it is undetermined at the gate for all eight states (never a pass, which ships, and never a fail, which blocks a release for something nobody measured); it exports as a null value BESIDE a state column, so a spreadsheet summing the values gets nothing from the row; and the whole chain is run end to end with a final assertion that no zero appeared anywhere in the serialized output. aggregate reports n and excluded by state, because a mean of 0.9 over two measurements and eleven absences is a different fact from a mean of 0.9 over thirteen — and a return value carrying only the number cannot tell them apart. A mean of nothing is REFUSED rather than returned as zero, and so is a mean across units, across directions, or over an ordinal scale. -->
    • S2.8.f Reject UI/BFF adapters that fill missing dimensions or average incompatible units. Evidence: adversarial adapter tests. <!-- probeAdapter RUNS the adapter rather than reading a declaration. "This adapter does not fill missing values" is a sentence in a review; finding a 0 in the output is a fact. The corpus contains all eight non-values deliberately, because the one an adapter mishandles is never the one it was written against. Five fabrication kinds, all reachable: inventing a number, dropping the row (a dimension that vanishes reads as one nobody asked for), changing a measured value, turning a non-value into a verdict, and sorting the absences among the numbers. A throw counts too — an adapter that cannot represent a non-value fails the same way as one that invents a number. THE ORDERING CHECK IS WHERE THE MUTATION PASS EARNED ITS KEEP. The first version asked "is a non-value between two measured rows?" and flagged an HONEST adapter that merely preserved input order — a check that fires on the correct implementation gets suppressed, and the suppression covers the real case too. It now asks whether the output equals what sorting by value ?? 0 would produce AND differs from the input order, so an alphabetical table is not a finding and a default value ?? 0 sort is. A mutant removing that guard survived until a "reorders for another reason" case was added. --> <!-- evidence: libs/oshun/workbench-kit/src/scores.ts + scores.spec.ts (49 cases, 11/11 mutants killed, 19 violation kinds and 5 fabrication kinds each reachable and asserted as a set). -->
  • S2.9 Add gate definitions with severity, owner, applicability, evidence version, expiry, and remediation metadata. Evidence: schema tests. <!-- libs/oshun/workbench-kit/src/gate-definitions.ts, published as @oshun/workbench-kit/gate-definitions, 44 cases, 13/13 mutants killed. 1,283 tests in the kit; all twelve Nx targets green. A DEFINITION IS NOT A CONTRIBUTION, and the two fail differently. S1.9 already gave a plugin a way to contribute a gate — evaluator, policy, three fallbacks. A bad contribution is a bug somebody fixes; a gate with no owner, no appeal path and a severity nobody agreed is a governance hole that shows up the first time it blocks a release at 2am and there is nobody to ask. So conformsToDefinition is the cell's reason for existing: the catalogue says what a stale gate does and the contribution decides what it ACTUALLY does, a release manager reads the first and a workflow obeys the second, and when they differ NOTHING FAILS — the gate simply behaves differently from its documentation, which is discovered during an incident review. 27 definition violation kinds and 5 registry ones, every one reachable. -->

    • S2.9.a Define stable gate ID/type/version, domain/object applicability, lifecycle stage, audience/target, and enabled/retired state. Evidence: gate identity tests. <!-- Applicability carries AUDIENCE as well as object kind: a rights gate on something going to a regulator is not the gate on an internal preview, and a definition that says neither has to be strict enough for the regulator everywhere. THE LIFECYCLE HAS THREE STATES, NOT TWO. shadow is the one most gate systems lack, and its absence is why new gates either block on their first day or never ship: a gate that runs, records and blocks nothing is how anybody finds out what it would have done. It carries a since and a REASON, so a shadow that has been running for a year is visible as the decision nobody made. canBlock returns false for a shadow gate whatever its severity says — asserted, because that is the whole point of the state. -->
    • S2.9.b Define blocker/warning/informational severity independently from raw evaluator result. Evidence: severity fixtures. <!-- SEVERITY VARIES BY LIFECYCLE STAGE, which a single field cannot express. The same failing measurement is a blocker on a published asset and a warning on a draft; one field forces the strictest reading everywhere, and that is how a drafting surface acquires eleven blockers and people learn to work around gates. A default plus per-stage overrides, and an override naming a stage the gate does not run at is refused — an override that can never apply reads as coverage. Independence from the verdict is asserted in both directions: severity says what the platform DOES and the verdict says what the evaluator FOUND, and no severity name appears in the verdict vocabulary. -->
    • S2.9.c Assign accountable owner, evaluator/provider owner, remediation owner, reviewer/override authority, and escalation. Evidence: ownership validation. <!-- Five roles because in practice they are five teams: who owns whether the gate should exist, who owns the code that answers it, who helps somebody past a failure, who may waive it, and where a disagreement goes. Collapsing them produces the failure message that tells an editor to contact the platform team. Two separation-of-duty rules, and both are the same shape as S2.7's: the override authority may not be the accountable owner (otherwise the gate constrains nobody who can reach that team) and neither may the escalation path (otherwise a disagreement about the gate escalates to the people who own it). -->
    • S2.9.d Define required evidence types, exact version constraints, freshness/expiry, invalidating dependencies, and unavailable behavior. Evidence: evidence policy tests. <!-- whenEvidenceUnavailable HAS NO pass OPTION — as a type, not as a rule. A gate that passes when its evidence is missing passes when the thing it checks is broken in exactly the way that hides the evidence. The three are block, warn and skip-with-record, and the third still writes a row, so the absence is countable rather than silent. invalidatedBy is the field that stops a gate being green forever: it names what makes the evidence stale REGARDLESS of age, so a rights check passed against last month's consent record does not survive the record moving. Without it, freshness is a timer, and a timer only knows how long ago the answer was given. Zero and infinite budgets are both refused. A BLOCKING GATE WITH NO REQUIRED EVIDENCE is refused: "it was green" is not a checkable claim about a gate with no declared inputs. A gate that only warns is permitted one — both directions asserted. -->
    • S2.9.e Define pass/warn/fail/review/unsupported policy, thresholds, and deterministic composition references. Evidence: policy schema tests. <!-- THE CELL'S WORDING NAMES THE VERDICTS THE KIT ALREADY HAS. pass/warn/fail/review/unsupported and S1.9's pass/warn/block/needs-human/indeterminate are the same five outcomes, so LEDGER_OUTCOME_TO_VERDICT publishes the mapping instead of a parallel enum — two words for "block" is how a client switches on one of them, handles four cases and falls through on the fifth. Asserted to be a bijection onto the shipped vocabulary. A fallback verdict that PASSES is refused, for the same reason the evidence behaviour has no pass. Composition here is the REFERENCE and the shape, not the aggregation (that is S2.10): standalone with references, a composite with none, a quorum with no number, and a quorum larger than its references are all refused, and the last is the one that reads as configured and can never be met. -->
    • S2.9.f Define user-facing reason, affected fields/refs, remediation, documentation, recompute, appeal, and waiver metadata. Evidence: remediation metadata tests. <!-- A BLOCKER WITHOUT AN APPEAL IS A DEAD END, and a dead end produces exactly one behaviour: somebody turns the gate off. So a blocker must carry a remediation or a document, an appeal path, and an appeal must have a clock — an appeal with no time to answer in is a refusal with extra steps. A gate that only warns needs none of it, and that is asserted too, or the rule would make every informational gate carry an escalation nobody uses. The waiver may not be granted by the accountable owner: otherwise the gate is advisory to the team that owns it and mandatory to everybody else. And a waiver with no expiry is a gate that was removed by somebody who did not have to say so. reasonKey is a translation key, because the one sentence a blocked person reads is the last one that should be untranslatable. --> <!-- evidence: libs/oshun/workbench-kit/src/gate-definitions.ts + gate-definitions.spec.ts (44 cases, 13/13 mutants killed). The registry check finds the cross-gate faults a single definition cannot show — a composition cycle, a composite still depending on a retired gate, a retirement pointing at a replacement nobody wrote — using the same three-colour walk S2.4 uses, because a visited set cannot tell a cycle from a diamond and two gates composing a third is a diamond. -->
  • S2.10 Add gate-run results with deterministic aggregate policy. Evidence: blocker, warning, waived, unavailable, and stale combinations have known outcomes. <!-- libs/oshun/workbench-kit/src/gate-runs.ts, published as @oshun/workbench-kit/gate-runs, 108 cases, 39/39 mutants killed. 1,392 tests in the kit; all twelve Nx targets green. THE CELL NAMES TEN WORDS AND THEY ARE FOUR DIFFERENT KINDS OF THING. blocker/warning are SEVERITIES and belong to the definition; pass/fail are VERDICTS and belong to the evaluator; unavailable/unsupported/stale/ missing/error are NON-VALUES and belong to the measurement; and waived belongs to none of them, because it is a decision taken after all three. A tenth enum containing all ten is how a switch ends up handling blocker and stale in the same statement, so ComponentState is a union of the three vocabularies the kit already has (S1.9 verdicts, S2.8 non-values) and severity and waiver are applied to it rather than mixed into it. THE FOLD STARTS FROM THE APPLICABLE DEFINITIONS, NOT FROM THE ROWS THAT EXIST. A fold over the results in hand cannot tell "every gate passed" from "the gate crashed before it wrote a row", and under that fold the cheapest way to pass a release is for a gate to fail early enough. Six headlines, because a console that renders four of them as one green is the reason they exist: everything passed, a check did not run, somebody signed for a failure, and a sibling covered a failure are four different conversations and only the first is uneventful. -->

    • S2.10.a Persist run ID, gate definition/version, target revision, dependency/evidence versions, evaluator/policy, actor/job, and timestamps. Evidence: result provenance tests. <!-- The trigger is a union — actor, job or schedule — because "no actor and no job" is the state a row lands in when the writer forgot, and it reads in a table exactly like a system run. FRESHNESS IS COMPUTED, NOT STAMPED. A stored stale: false is true exactly once and stays in the column afterwards, so evidenceFreshness derives it at read time from the recorded timestamps, the definition's budget, and the CURRENT dependency versions. That last is what makes S2.9's invalidatedBy operative: an EvidenceRecord carries the versions it was computed over, so a rights check answered four minutes ago against a consent record that has since moved comes back stale — which freshness-by-age cannot see, because the answer really was recent. evidence-recorded-after-the-run-finished is the quiet one: evidence stamped later than finishedAt is not what the gate read, it is what somebody attached afterwards. -->
    • S2.10.b Persist raw measurements and nonvalue states separately from policy verdict and aggregate release eligibility. Evidence: separation tests. <!-- Three fields, and the cell's wording ("raw measurements and nonvalue states") IS S2.8's ScoreReading, so the observation reuses it rather than restating an absence vocabulary. GateObservation is evaluated-with-readings or not-evaluated-with-a-named-state; the verdict is S1.9's, derived under the policy named in the provenance; the effect is what the platform does about it. passed-without-evaluating is the rule this cell exists for, and it covers warn as well as pass because warn is in the kit's PASSING set — a row that says "did not run" and "warn" ships exactly as readily. It is written by ordinary code: a catch that defaults the verdict, a provider timeout falling through to the happy path, a migration that backfilled a column. -->
    • S2.10.c Define precedence and aggregation for blocker, warning, pass, fail, unavailable, unsupported, stale, missing, waived, and error. Evidence: exhaustive truth table. <!-- 336 combinations generated rather than a chosen few (14 states x 2 blockable x 3 evidence behaviours x 2 gap policies x 2 ledger states), with the count asserted so that adding a state without revisiting the suite fails there rather than silently halving coverage. A WAIVER'S ONLY POWER IS TO TURN A BLOCK INTO A WAIVED BLOCK, and that is checked by generating the table TWICE — with and without a waiver — and diffing. The diff must be exactly blocks -> waived; a waiver that could promote a gap or satisfy anything appears as a cell that moved and should not have. Reading the function could not establish that. Nothing is satisfied except an actual pass. An ambiguous ledger (two results for one gate) is fail-closed INDEPENDENTLY of the evidence policy, because it is not a missing input — it is a question about which row is the answer, and choosing by iteration order is the order dependence S2.10.d forbids. A deliberate divergence from S2.8: unsupported is not-counted for a score so it cannot drag an average, and for a gate it means "the check did not happen" — treating that as not-applicable is how a new asset kind ships unchecked. -->
    • S2.10.d Define all-of, any-of, quorum, conditional, target-specific, and domain plugin composition without order dependence. Evidence: aggregation strategy tests. <!-- Kleene three-valued, and the third value is load-bearing: any-of over a failure and a check that could not run is neither satisfied nor refuted, because the one that did not run might have passed. A quorum is refuted only when trues + unknowns < need — a two-valued quorum reports "not enough approvals" for a panel where one reviewer merely has not finished, and the remediation for those two is not the same. conditional and target-specific dispatch on RECORDED facts rather than closures: a predicate given as a function cannot be stored, replayed or compared between two policy versions, so a strategy built from closures is unrecomputable by construction. A DOMAIN FOLD IS RUN TWICE — over the members and over the same members reversed — and disbelieved if it disagrees with itself, which catches the actual bug (a strategy that reads members[0]) more cheaply than a code review. The members are FROZEN rather than compared afterwards: taking the capability away is stronger than checking whether it was used, and the mutation pass found that freezing only one direction lets a strategy that writes on one particular ordering through. An advisory gate can never satisfy a strategy: otherwise any-of over a blocker and an informational gate ships whenever the informational one is green. -->
    • S2.10.e Preserve every component result and reason; never hide a blocker behind an average or passing sibling. Evidence: component visibility tests. <!-- THE STRATEGY DECIDES SATISFACTION AND MAY NOT DECIDE BLOCKING, which makes this structural rather than a rule somebody has to remember. A component an any-of tolerated keeps its blocks effect, stays in components, appears in tolerated, and forces the headline to eligible-by-composition — a different sentence from eligible. And there is no average to hide behind: the aggregate carries counts and a headline, never a pass rate. visibilityFaults is the other half, the same shape as S2.8.f's adapter probe and for the same reason — the fold can be correct and the thing a person reads can still be wrong, and no test of the fold sees that. waived-shown-as-passing is the one that matters most: rendering a signed exception in the same green as a pass is precisely how waivers become routine, and it is the most common way a gate stops meaning anything without anybody deciding to remove it. A headline may be more cautious than the fold and never less. -->
    • S2.10.f Recompute identical input deterministically and version aggregate policy changes. Evidence: property and version-change tests. <!-- Order independence is asserted over every permutation of the results, of the definitions, and of the waivers, by comparing serialised aggregates — not by reading the folds. Instants are the kit's UTC-with-Z text, so string order is time order and no date arithmetic enters the comparison. A RELEASE POLICY CAN ONLY EVER TIGHTEN. gapsBlockFor lets a deployment refuse a gate's own record-and-continue for one audience; waivers are NARROWED per audience and never granted, so a gate whose definition carries no waiver policy is not made waivable by any deployment — that is a recorded fault. Waivers for the regulator audience are refused outright, the same shape as S2.9's refusal to let the accountable owner grant one: a waiver is an internal decision and that audience is the one place an internal decision cannot stand in for the check. explainPolicyChange ships a fold change with the list of decisions it reverses, because loosened is the half that needs a human, and changedWithoutAVersionBump is the unambiguous defect: every aggregate already recorded carries the old version and would now recompute differently under it, so the history of what was approved is rewritten with nothing saying it moved. --> <!-- evidence: libs/oshun/workbench-kit/src/gate-runs.ts + gate-runs.spec.ts (108 cases, 39/39 mutants killed; 12 run violation kinds, 12 aggregate fault kinds, 8 policy violation kinds and 7 visibility fault kinds each reachable and asserted as a set). The mutation pass earned its keep three times: it found a dead any-of empty-list branch, an unreachable post-hoc mutation check the freeze had already made impossible, and two guards (a blocker returning warn, an evidence policy of warn) that no test covered. --> <!-- Not in scope here, on purpose: the waiver REQUEST, approval, renewal and revocation history are S2.11, so this cell defines the minimum WaiverRecord the fold needs and no more, rather than two cells each defining a waiver. -->
  • S2.11 Add authorized, time-bounded gate waivers. Evidence: expiry automatically makes the release ineligible. <!-- libs/oshun/workbench-kit/src/gate-waivers.ts, published as @oshun/workbench-kit/gate-waivers, 61 cases, 54/54 mutants killed. 1,454 tests in the kit; all twelve Nx targets green. THE HEADLINE IS PROVED THROUGH A REAL FOLD, not asserted about a status field: toWaiverRecord returns null for anything but an active standing, so the S2.10 aggregate goes from eligible-with-waivers to blocked across the expiry instant with nothing else moving, no job having run and no row updated. The suite drives four instants either side of it. A REQUEST IS NOT A GRANT. One row with a status column makes "who asked" and "who agreed" the same record and leaves the request editable after approval, so the grant restates the request's four identifying fields and carries a digest of what was approved — checked against a recomputation the CALLER supplies, because a kit contract entry point may not reach a Node builtin and comparing a stored digest with itself passes by construction. That test conflating two clocks was the one real bug this cell surfaced: crossing the waiver's expiry also crossed the evidence's freshness budget, and the release came back undetermined for a reason that had nothing to do with waivers. Both readings are now pinned separately. -->

    • S2.11.a Define waiver request for exact gate result, target revision, audience/channel/environment, reason, evidence, and risk. Evidence: waiver schema. <!-- The request names a RUN, never a gate: a waiver against a gate rather than a result is an exemption, and it will outlive every reason it was granted for. Channel and environment are separate fields from audience because they fail differently — an exception agreed for staging that silently covers production is a deployment mistake, and one agreed for a newsletter that covers the wire is an editorial one. A scope naming neither is refused, because it then covers whatever the reader assumes. Risk is ordinal so the interesting comparison can be made: an approval that accepted LESS than the request declared, with no reassessment recorded, means the approver agreed to a smaller thing than the one described and the record shows only the smaller one. A severe residual risk needs a second signature, and that signature may not be the approver's own. -->
    • S2.11.b Enforce gate-specific waiverability and actor/role/tenant/object scope plus separation of duties. Evidence: authorization negative tests. <!-- Four independent narrowings a waiver has to survive, and the gate's own definition is the first: a gate whose presentation.waiver is null cannot be made waivable by any deployment, any policy, or any approver. Three separations of duty, all negative-tested: the approver may not be the requester (a waiver anybody can grant themselves is not a control, it is a field), may not be the gate's accountable owner (S2.9's rule, re-checked at USE time because ownership changes after a waiver is signed), and may not hold a role the gate never delegated to. Scope is checked as seven independent misses — tenant, object, audience, channel, environment, revision, run — each asserted on its own, and scope is checked BEFORE binding because "this is not your story" and "the evaluator moved" are different conversations. -->
    • S2.11.c Require start, expiry or one-use, review date, revocation, and renewal as a new decision. Evidence: clock-controlled lifecycle tests. <!-- ONE-USE IS NOT A SHORT EXPIRY. "An exception for this one publish" and "an exception for everything that shipped in that hour" are different decisions, and modelling the first as a one-hour window is how it becomes the second. A one-use grant records what spent it and stops authorising every other release, inside a window that still bounds it — an unused one-use waiver from last year is not live either. A backdated start is refused: a waiver authorising a period before anybody agreed to it is how a release that already shipped acquires cover. A review date after expiry is refused: a review scheduled for after the thing has lapsed is one nobody does, and it is what makes a renewal automatic. A zero-length window is refused alongside an inverted one — the mutation pass found <= untested at the equal instant, which is the nastier half, since it reads in a table as a waiver with dates. RENEWAL IS A NEW DECISION — there is no renew(), only a grant that supersedes, in the same shape S2.7 uses. renewalChain walks back to the origin and ADDS THE INSTALMENTS UP, which is the finding the function exists for: eight consecutive forty-eight-hour waivers are a two-week exemption no single approval authorised, every link reads as within the gate's limit, and nothing sees it unless something sums them. It also catches a renewal that widened the scope (a new exception wearing a renewal's name) and one that shaved the accepted risk down a level at a time until a severe exception is on the books as a low one. -->
    • S2.11.d Bind the waiver to gate/input/evaluator/policy/dependency versions and invalidate it on any out-of-scope change. Evidence: binding invalidation tests. <!-- A GATE VERSION BUMP CAN NEVER BE TOLERATED, as a type rather than a rule: REBINDABLE_DIMENSIONS is input/evaluator/policy/dependency and has no member an approver could write for the gate version, because S2.9 bumps that when the gate's MEANING changes and an approval of the old meaning is not one of the new. Asserted by granting a waiver that tolerates every dimension there is and watching it still come unbound. The tolerance list is the approver's judgement and the default is empty, which is the strictest reading. It earns its existence in the other direction too: a waiver that survives nothing at all gets re-requested constantly, and people learn to ask for a week instead of an hour. A dependency that stops being REPORTED is unbound rather than tolerated — there is nothing left to compare against, so the check cannot be shown to still hold. -->
    • S2.11.e Include active waiver scope, reason, approver, expiry, and residual risk in review/publish/release surfaces. Evidence: visibility tests. <!-- The five fields are five different objections, and the check names each: without the scope you cannot tell whether it covers what you are looking at, without the approver there is nobody to ask, without the expiry you cannot tell a decision from a habit, without the reason there is nothing to disagree with, and without the residual risk the release reads as fine. Two faults run the other way and both matter more than an omission. A rendered expiry that is not the grant's own is worse than none, because it is believed. And a revoked or expired waiver drawn as live tells a release manager they are covered when they are not — the same shape as S2.10.e's waived-shown-as-passing, one layer up. The five fields are demanded only of a waiver that AUTHORISES something; a lapsed one shown as lapsed needs none of them, and asserted, or the rule would make every historical exception carry a full disclosure. -->
    • S2.11.f Make expired, revoked, stale, or out-of-scope waivers nonauthorizing without deleting their audit history. Evidence: release-eligibility tests. <!-- Eight standings, EXACTLY ONE OF WHICH AUTHORISES — asserted as authorizes === (status === 'active') over the whole vocabulary, with every status reached and exactly one authorizing standing in the set. The precedence is deliberate and is the answer to "which reason wins": revoked, then superseded, then the clock, then scope, then binding. Reporting "expired" for a waiver somebody revoked would put the wrong sentence in front of the person asking why the release stopped. NOTHING IS DELETED. The approver, the approval date, the expiry, the accepted risk and the scope sit OUTSIDE the status, so a revoked waiver stays fully readable — a revoked waiver that vanishes takes with it the fact that somebody once thought this was acceptable, which is the fact an incident review is looking for. --> <!-- evidence: libs/oshun/workbench-kit/src/gate-waivers.ts + gate-waivers.spec.ts (61 cases, 54/54 mutants killed; 22 grant violation kinds, 8 statuses, 6 renewal finding kinds and 9 disclosure fault kinds each reachable and asserted as a set). The mutation pass found two real gaps: a zero-length window at the <= boundary, and three of the four grant/request identity fields never varied on their own — a check that fired on one of them would have passed the suite and let the other three through. --> <!-- Deliberately still S2.10's: the fold itself. This cell decides authorization and hands the aggregate a record or nothing; it does not re-implement precedence, and the two agree because the record carries the grant's own window rather than a copy that can drift. -->
  • S2.12 Add dependency-triggered staleness propagation. Evidence: a source, model, consent, canon, or asset update marks every dependent gate stale. <!-- libs/oshun/workbench-kit/src/staleness.ts, published as @oshun/workbench-kit/staleness, 59 cases, 48/48 mutants killed, plus a new Postgres target outbox-atomicity. 1,514 tests in the kit; all THIRTEEN Nx targets green. S2.9 declared what makes evidence stale and S2.10 computed it for ONE result against one set of current versions. Neither can answer the question this cell exists for: a consent record moved — WHAT ELSE is now wrong. That is a graph walk, and it is the difference between a gate that goes red when somebody looks at it and a platform that says so. The headline is driven end to end: one withdrawal, through a three-deep graph, produces marks on the rights, canon and likeness gates — and every one of them then REFUSES to be recomputed automatically, because the thing that moved was a withdrawal. -->

    • S2.12.a Define dependency node/edge and invalidating mutation types for sources, models, consent, canon, assets, policies, evaluators, and target configurations. Evidence: invalidation registry. <!-- NOT EVERY MUTATION INVALIDATES, AND THE REGISTRY HAS TO SAY WHY NOT. A source whose title was corrected does not invalidate a rights check; one whose licence changed does. A registry where everything invalidates makes the recompute queue the whole corpus every night, and the response to that has never once been to tune the registry — it is to turn propagation off. So why is required on BOTH answers and the negative one is the more important: an unexplained false is a judgement somebody made to stop the queue filling up, and there is no way to tell it from a real one afterwards. Two structural rules pointing opposite ways: a kind where NOTHING invalidates will never propagate and reads exactly like a covered kind, and a kind where EVERYTHING invalidates is the registry nobody kept. Both refused, both asserted. 27 shipped entries across the eight kinds, and model/alias-repointed is the one that justifies the subsystem: a provider moves a stable name onto new weights, nothing here changed, no revision moved, no editor did anything, and every answer that name gave is now an answer to a different question. Nothing inside a single result can see it. An UNREGISTERED mutation is treated as invalidating, and says so — a mutation nobody classified propagating nothing is the exact failure this cell exists to prevent, and a noisy queue is the cheaper mistake. -->
    • S2.12.b Emit versioned dependency-change events transactionally with authoritative mutations. Evidence: outbox atomicity tests. <!-- "The mutation and its event commit together or neither commits" is a property of a TRANSACTION, and a test with a fake store proves the fake. So tools/outbox-atomicity.mjs builds both tables and drives a real PostgreSQL, and THE NEGATIVE CONTROL IS THE POINT: the same failure is injected twice, once inside one transaction (both tables must come back unchanged) and once with the two writes committed separately (which MUST leave an authoritative row with no event beside it). If the control does not break, the harness cannot tell the two designs apart and every green it has reported is worthless. Calibrated by splitting the atomic write into two transactions and confirming the run goes red. The kit half is the one no engine can hold: UNIQUE (node, sequence) refuses two events claiming position 3 and cheerfully accepts 1, 2, 4 — and 1, 2, 4 is exactly the shape a lost event leaves. validateEventStream finds the gap, the broken version chain and the backwards version in a stream the constraints accepted. The harness caught its own fixture: the first draft's backwards-version row was 100 -> 7, which the from_precedes_to CHECK rejects — a row-level fault, and the wrong kind to calibrate a stream checker against. -->
    • S2.12.c Traverse direct and transitive dependents with bounded fan-out, cycles, depth, pagination, and version checks. Evidence: graph traversal tests. <!-- A TRUNCATED WALK IS NOT A COMPLETED ONE, as a type. Fan-out caps, depth limits and pages are all necessary and are also how a propagation silently leaves half its dependents fresh, so PropagationOutcome is a union whose incomplete arm carries the truncation notes and a resume cursor and cannot be read for affected without the caller seeing that it stopped. Cycles are REPORTED AND SURVIVED rather than refused: a canon graph genuinely has mutually-referential entities, and a traversal that threw on one would make the whole propagation depend on the cleanliness of somebody else's data. Breadth-first, so depth means what it says and a limit cuts the far edge rather than one arbitrary branch. The version check is what makes a replayed event cheap instead of a second sweep: a result already at or past the new version is not affected. The mutation pass found TWO GUARDS MASKING EACH OTHER — dropping the visited set still produced the right affected, because the result dedup caught the duplicate on the way out. Separated by counting expansions (a wide diamond re-expanded per path is how this becomes exponential) and by a result reachable from two genuinely distinct nodes, which the visited set cannot deduplicate. -->
    • S2.12.d Mark affected results stale with old/new dependency version, event, reason, and time before downstream authorization. Evidence: stale-state tests. <!-- Every field the cell names, plus one it does not: the PATH. "Your rights check is stale" is unactionable; "your rights check is stale because the consent record under the canon entity it cites was withdrawn" is a task, and the path is the only place that sentence can come from. Omitted for a direct dependency, where it would be noise. "Before downstream authorization" is made checkable rather than asserted: unappliedMarks compares two timestamps that already exist — a mark stamped BEFORE an S2.10 aggregate ran, against a component that aggregate did not treat as stale. That combination is a release decided over a staleness somebody had already detected. A mark stamped AFTER the aggregate is reported as its own kind and is not a fault: it is the world moving on, and treating it as a violation would make every propagation indict every earlier release. -->
    • S2.12.e Queue idempotent recomputation or require explicit action according to policy/cost/consent/provider constraints. Evidence: recomputation routing tests. <!-- WITHDRAWN CONSENT IS NEVER AUTO-RECOMPUTED, and it is checked FIRST. Re-running a model over material whose consent was withdrawn, because an automated staleness sweep noticed the evidence was old, is the worst automatic action this subsystem could take — and any ordering that lets a budget or provider check reach it first is one where a configuration change makes it possible. Asserted with every other constraint simultaneously hostile. Three explicit-action routes rather than a queue: a human-answered gate (a queue that re-ran it would be re-recording a judgement nobody made twice), a provider that is down (a job in front of a dependency that is down retries until somebody notices the queue rather than the outage), and a cost past budget (a sweep that overspends gets switched off entirely rather than tuned). THE IDEMPOTENCY KEY MAY NOT CONTAIN THE EVENT OR THE CLOCK. It names what the job would DO — target, dependency, version being caught up to, evaluator — so a replayed event collapses onto the job already queued. A key carrying the event id makes every replay a new job, which is the incident that makes people delete the outbox. Asserted in both directions: a replay matches, and a version or evaluator move does not. -->
    • S2.12.f Reconcile missed events and partial propagation against stored dependency versions. Evidence: fault-injection repair tests. <!-- RECONCILIATION COMPARES STORED VERSIONS, NOT EVENT LOGS. An event that was never emitted leaves no trace in the log that would have carried it, so a repair pass built on the log finds a consistent log and reports nothing. The only durable evidence is a result whose observed version is behind its source and which carries no mark — a comparison between two pieces of state, neither of which is the thing that failed. partial-propagation is the half that says the walk RAN: some results under a node caught the change and others did not, which is a walk truncated, timed out or crashed halfway, and it looks nothing like one that never started. Only the mixture distinguishes them, and a fully missed node is asserted NOT to be reported as partial. Two findings about wrong state rather than missing propagation: observed-version-ahead-of-the-source (the observation was written from somewhere that is not the source of truth) and mark-for-a-version-that-never-arrived (an event emitted and rolled back, whose consequence outlived it). The second needed a markedToVersion field to be reachable at all — it was declared and unreachable in the first draft, which is the shape this ledger refuses, so it was made real rather than deleted. repairPlan re-emits only for the two findings that are propagation to redo; the other three are state to fix. --> <!-- evidence: libs/oshun/workbench-kit/src/staleness.ts + staleness.spec.ts (59 cases, 48/48 mutants killed; 5 registry violation kinds, 8 event violation kinds, 3 application fault kinds and 5 reconciliation finding kinds each reachable and asserted as a set), plus libs/oshun/workbench-kit/tools/outbox-atomicity.mjs driving a real PostgreSQL with a negative control that must come back broken. -->
  • S2.13 Add a human-signoff gate type that automation cannot self-complete. Evidence: system actor attempts are rejected. <!-- libs/oshun/workbench-kit/src/signoff.ts, published as @oshun/workbench-kit/signoff, 55 cases, 55/55 mutants killed. 1,570 tests in the kit; all thirteen Nx targets green. Every other gate in the kit is a MEASUREMENT — something runs, something is compared against a threshold, a verdict comes out. This one is a person taking responsibility, and the whole cell is one word of its title: SELF-COMPLETE. Automation may schedule it, remind about it, escalate it, show it and record it. It may not produce one. The failure guarded against is almost never an attacker. It is an agent holding a service token that happens to carry the editor-in-chief claim, a seeded environment where every reviewer is a fixture, a support engineer's impersonation session left open in another tab — and each of those satisfies every role check ever written. -->

    • S2.13.a Define required human role, named-person or eligible-pool scope, quorum, sequence, independence, and expiry. Evidence: human gate schema. <!-- A QUORUM WITHOUT INDEPENDENCE IS ONE SIGNATURE WITH A COUNT, and the count is what a release console renders. So a quorum above one that declares no independence is refused, and the dedup by signer is UNCONDITIONAL — the mutation pass found that a gate which happens not to declare previous-signer independence could otherwise reach a quorum of two with one person. Independence is DECLARED, not inferred. same-team is reported as undecidable rather than silently passed: the kit has no org chart, and a membership check invented from an id prefix is wrong in exactly the cases it matters. A REJECTION STANDS. It is not outvoted by later approvals and cannot be cleared by collecting more — a quorum reachable by asking until somebody says yes has no lower bound, and it reads in an audit as "three of four approved". sequential exists because some reviews are only meaningful in an order: legal reading a piece editorial has not finished is legal reading a draft. An order over a quorum of one is refused, because it reads as a stronger control and is the same one. -->
    • S2.13.b Require authenticated human presence and step-up authentication where risk policy demands. Evidence: identity/step-up tests. <!-- A STEP-UP REUSED FROM LOGIN IS NOT A STEP-UP. It has to have happened AFTER the review opened and inside a fifteen-minute window, or it is a property of the session rather than of the judgement — which is exactly what makes a stolen session enough. Both bounds asserted, including the edge of the window in each direction. WEAK_STEP_UP_METHODS names password and SMS, and a gate whose only accepted methods are weak is refused: re-entering the session credential proves the session is still the session, which is precisely what somebody holding it can also prove. -->
    • S2.13.c Reject service, system, agent, model, fixture, unauthenticated, and impersonated actors even when they possess a generic role claim. Evidence: actor-type negative tests. <!-- THE ROLE CLAIM IS NOT THE CONTROL — THE ACTOR KIND IS, AND IT IS CHECKED FIRST. The suite's central case is a service token holding editor-in-chief: it satisfies every role check ever written and is refused before the role is read. Where it can be a type it is one — a completed record carries a HumanPresence, and only admitSigner produces one, so a service actor has no path to a valid signoff at all. The cell's list maps onto S1.4's actor model plus S1.13.d's mark: service and system are not-a-human-actor, an impersonation is delegated (a real person, at a real keyboard, whose signature would land the responsibility on somebody who never saw the thing), and a FIXTURE is caught by BOTH halves of the synthetic mark — the symbol the builders stamp, which does not survive JSON, and the reserved id prefix, which survives JSON and a database column. A check on one of them passes for exactly the fixtures the other catches, and both are asserted. The strongest evidence here is self-referential: no actor this repository's own test builders can produce is able to complete a human gate. -->
    • S2.13.d Record reviewed target/evidence versions, decision, comment, timestamp, authority, conflict/recusal, and signature or equivalent. Evidence: signoff record tests. <!-- The VERSIONS are the point. "They approved it" is unfalsifiable a month later; "they approved revision 7 against consent record v3 and rights record v2" can be checked against what those things are now, which is what makes S2.13.e possible at all — so a record with no evidence versions is refused, on the grounds that nothing could ever make it stale. Conflict and recusal are two fields rather than a boolean, because "declared a conflict and recused" and "declared a conflict and signed anyway" are the two outcomes and only one is a problem. Signing over an unrecused conflict is refused — a conflict declaration is only worth collecting if it is, otherwise it is a field somebody fills in and nothing reads — and so is a recusal that still produced a decision. role-does-not-match-what-was-held catches the recorded authority drifting from the authority that was actually checked. -->
    • S2.13.e Invalidate signoff on expiry or relevant target/evidence/policy change. Evidence: staleness tests. <!-- A signoff is EVIDENCE, and it goes stale in exactly the way S2.12 makes everything else go stale: a person approved specific versions of specific things, and any of them moving means they approved something else. Seven standings, exactly one of which stands, with the signer, the time and the decision held OUTSIDE the status so a lapsed signoff stays fully readable. scope-widened is the one a version comparison misses entirely. Nothing about the content changed; the audience did. An approval given for an internal preview that now covers a regulator submission is the same bytes reviewed for a different question. Narrowing is deliberately NOT invalidating — the person already approved the wider case, and invalidating on it would mean every audience restriction costs a new review. A dropped evidence entry is distinguished from a moved one, because the remediations differ: re-review against the new version, versus find out why the thing the reviewer looked at is gone. -->
    • S2.13.f Require a new human action for renewal, reversal, or materially changed scope. Evidence: no-automation lifecycle tests. <!-- There is NO ARM of planRenewal that produces a valid signoff, and that is asserted over all three intents. There is no renew() and no extend() — an extend is how a person who reviewed something once is recorded as having reviewed it every quarter since — and a reversal is a person deciding the other way, so a system that can flip a decision without one has a decision field rather than a signoff. "Materially changed scope" is COMPUTED by comparing what was reviewed against what is now proposed, never taken from a flag, because the flag is the thing an automated pipeline learns to leave false. The type-level half holds the rest: the only thing a renewal could produce is a HumanPresence, and nothing but admitSigner can build one. --> <!-- evidence: libs/oshun/workbench-kit/src/signoff.ts + signoff.spec.ts (55 cases, 55/55 mutants killed; 12 refusal reasons, 10 requirement violation kinds, 9 record violation kinds, 7 quorum fault kinds and 7 standings each reachable and asserted as a set). The mutation pass found two guards masked by later checks: the unconditional signer dedup, and a dropped evidence entry that fell through to the moved-version branch and produced the same standing with the wrong sentence. --> <!-- Not done here: wiring the human gate into S2.10's fold as a gate CONTRIBUTION. That is a plugin registration rather than a contract, and S2.10 already routes a requiresHuman recomputation to a person instead of a queue; a domain wires the two together when it ships a real one. -->
  • S2.14 Add stage/gate audit projections. Evidence: the projection can be rebuilt from events and matches stored read models. <!-- libs/oshun/workbench-kit/src/audit.ts, published as @oshun/workbench-kit/audit, 46 cases, 49/49 mutants killed. 1,617 tests in the kit; all thirteen Nx targets green. THIS CLOSES PHASE S2. A PROJECTION IS A CACHE. That sentence is either true or it is a lie somebody tells about a second source of truth, and the difference is entirely whether the thing can be thrown away and rebuilt — so the cell's headline evidence is not one test among several here, it is the property the rest of the module exists to make checkable. -->

    • S2.14.a Define immutable event types for transition request/refusal/ completion/bypass and gate definition/run/stale/waive/recompute/signoff lifecycle. Evidence: audit event schemas. <!-- THE VOCABULARY IS DERIVED FROM THE CELLS, NOT INVENTED. Fifteen kinds, each naming a decision S2.1–S2.13 already makes, and uncoveredSurfaces checks the mapping in the direction that matters: an unused event kind is dead weight, and a decision the platform MAKES and does not record is a thing that happened and cannot be shown to have happened. transition-bypassed is the kind people leave out and the first one an incident review looks for. -->
    • S2.14.b Include target/revision, prior/new state, actor/delegate, reason, evidence/policy versions, correlation/trace, and safe error refs. Evidence: event completeness tests. <!-- Actor and principal are two fields, because a row recording one of them answers neither "who did this" nor "whose authority was used" — the two questions an incident review asks, in that order — and both directions are refused (a delegated actor with no principal, and a principal on an actor that is not delegated). The error ref is typed as a REFERENCE and prose in it is refused: the field exists so a failure can be joined to an incident, and text is how a stack frame ends up in an export. detail-carries-a-leak BORROWS RATHER THAN REINVENTS — S1.10 already argued about what a credential, an email address or a cross-tenant id looks like in free text, an audit detail is free text crossing the same boundary, and a second set of detectors is one that is always behind the first. -->
    • S2.14.c Build per-object timeline, current stage, available/refused transitions, gate status/freshness, waiver, and signoff projections. Evidence: projection reducer tests. <!-- One reducer, total over the kind union with no default arm, so adding an event kind is a compile error rather than a fact the projection quietly drops. Refused transitions keep their reason, bypasses are COUNTED separately (of every number in the projection that is the one worth reconciling first), gate rows carry staleness and waiver together because a release console reads them together, and a revoked or expired waiver clears the row rather than leaving it green. -->
    • S2.14.d Apply deterministic ordering, idempotency, versioning, out-of-order handling, poison quarantine, and replay. Evidence: event-delivery tests. <!-- ORDER COMES FROM THE SEQUENCE, NEVER FROM THE TIMESTAMP. Two writers a hundred milliseconds apart on skewed clocks produce a trail that reorders itself, and BOTH ORDERS READ AS PLAUSIBLE — asserted with a stream whose timestamps are deliberately out of order and whose fold is not. Out-of-order delivery is HELD, not applied: folding event 5 before 4 produces a state that never existed and produces it silently. Duplicates by event id are dropped (that IS idempotency here); duplicates by POSITION are quarantined, because two events in one place is a story with a hole whichever is chosen. A REAL BUG THIS CELL SURFACED: the stall only fired when a LATER event revealed the gap, so a stream whose trailing events were all quarantined came back COMPLETE. Fixed by making quarantine imply incomplete independently of position, with redelivery the one exception since dropping a duplicate removes nothing from the story. Both directions are now pinned. -->
    • S2.14.e Rebuild projections from an empty store and compare to transactional read models and authoritative state. Evidence: full-replay parity test. <!-- FIELD BY FIELD RATHER THAN BY A DIGEST, deliberately. A checksum answers "do they differ", which is the question you already suspected the answer to; the question worth asking is WHICH FIELD, because that is what distinguishes a drifted reducer from a lost event from a stored copy something wrote to directly. completeness-disagrees is the sharpest of the nine: a stored projection that believes it is whole while a replay stalls is one that stepped over something. -->
    • S2.14.f Apply tenant/object/property authorization and redaction to audit queries and exports. Evidence: disclosure negative tests. <!-- REDACTION IS APPLIED TO THE PROJECTION, NOT TO THE QUERY — a filter in a query is one the next endpoint forgets, and every audit export is a next endpoint. A cross-tenant viewer gets visible: false rather than an empty projection, because "there is an object here and you may see none of it" is itself the disclosure, and returning an empty result feels like a refusal already, which is why people ship it. Three capabilities, each removing something different, and disclosureFaults is the adversarial half in the shape S2.8.f and S2.10.e use: the narrowing can be correct and what reached the wire can still be wrong. leak-in-a-disclosed-detail runs the details a viewer IS permitted through the kit's own S1.10 detectors — permission to read the reason is not permission to read a bearer token somebody pasted into it. --> <!-- evidence: libs/oshun/workbench-kit/src/audit.ts + audit.spec.ts (46 cases, 49/49 mutants killed; 13 event violation kinds, 7 delivery fault kinds, 9 parity fault kinds and 5 disclosure fault kinds each reachable and asserted as a set). The mutation pass found two guards masked by sharing a fault KIND across two loops — the timeline actor check and the signoff signer check each satisfied a toContain the other should have — now separated by asserting the fault's SUBJECT. --> <!-- PHASE S2 COMPLETE: S2.1–S2.14, with only S2.3.d outstanding (browser-gated). -->

S3 — Persistence, tenancy, events, and jobs#

  • S3.1 Write the shared persistence conventions document. Evidence: it covers explicit columns, DateTime, migrations, indexes, generated clients, and real-DB verification. <!-- PARENT HELD UNCHECKED: S3.1.h needs two named human approvals (data platform and security) and an agent may prepare the packet but may not approve it. Everything else is done. docs/conventions/shared-persistence-conventions.md, with a machine-checked companion: scripts/v1-workbenches/generate-persistence-conformance.mjs writes evidence/v1-workbenches/persistence-conformance.json, registered in inventory:v1-workbenches, verify:inventory:v1-workbenches --check, and a new gate:persistence-ratchet. 16 tests. A CONVENTIONS DOCUMENT NOBODY CAN MEASURE AGAINST IS ALREADY OUT OF DATE AND THE ONLY QUESTION IS BY HOW MUCH. So the document ships with a scanner over all 16 Prisma schemas / 439 models, and the report is EXPECTED TO BE MOSTLY RED — those models predate the document, and a scanner tuned until it came back green would be measuring the tuning. The enforcement that works on a tree this size is the RATCHET: a schema whose conforming count goes DOWN fails, new shared entities conform, existing ones do not get worse. -->

    • S3.1.a Define entity/table naming, stable IDs, tenant/workspace scope, actor/revision/tombstone metadata, relationships, and extension fields. Evidence: normative data-model section. <!-- §1. Identity is the kit's @oshun/workbench-kit/identity vocabulary rather than bare UUIDs, so idEquals(workbench, tenant) is a COMPILE ERROR rather than a comparison that returns false; sequential integer keys are prohibited on anything a client sees, because a number that reveals how many rows exist is a business fact shipped in a URL. Scope is a COLUMN, never a join — an isolation rule enforced by a join is one a missing join disables — and a global reference table's omission must be DECLARED, because an undeclared omission is indistinguishable from a forgotten one. updatedById is a column for the same reason: an audit trail that joins to answer "who" stops answering when the join breaks, and it breaks exactly when somebody is looking. ON DELETE CASCADE is prohibited on anything carrying history: a cascade deletes an audit trail to keep a pointer tidy. -->
    • S3.1.b Define which operational fields require explicit typed columns and the limited allowed uses of JSON/blob payloads. Evidence: column-versus-payload decision rules. <!-- §2. Six conditions that force a column (queried, constrained, authorized on, part of a lifecycle/version/identity, computed from by a retention rule, grepped for during an incident) and exactly THREE permitted payload uses: an opaque third-party response kept verbatim for provenance, schema-versioned document content owned by a domain plugin, and a rebuildable read-model projection. Everything else in a payload is a field somebody needs to query within a year, and the extraction migration then is strictly more expensive than the column now. The section says out loud that the scanner catches only the crudest version of this (a majority-JSON model, 439/439 passing today) and that whether a payload is genuinely opaque is a review obligation — listed in §11 rather than left as silently-unchecked prose. -->
    • S3.1.c Define UTC instant, local date/time, duration, time zone, calendar, precision, server-clock, and serialization rules for DateTime values. Evidence: temporal semantics section. <!-- §3. Instants are timestamptz, always UTC, never String — a timestamp in a text column sorts as text and two spellings of one instant both exist. Serialisation is the kit's Instant shape (literal Z), so STRING ORDER IS TIME ORDER, which every ordered log and keyset cursor depends on. A LOCAL DATE IS A DIFFERENT TYPE, not an instant: a date plus an IANA zone column, because a local date coerced to an instant moves when the server does. Durations carry the unit in the NAME. Every platform timestamp comes from the database transaction or one injected clock, and §S2.14 is cited for why — two writers on skewed clocks produce an audit trail that reorders itself and both orders read as plausible. -->
    • S3.1.d Define constraints, indexes, pagination order, query-plan review, row-level authorization, encryption, and sensitive-field storage. Evidence: integrity/performance section. <!-- §4. CONSTRAINTS BELONG IN THE DATABASE, because application checks run where somebody remembered to call them and a migration, a support script and a bulk importer all reach the table directly — with the kit's own four Postgres harnesses named as the template. Uniqueness that matters is a UNIQUE constraint, not a SELECT before an INSERT (the kit's revision race exists because that lesson cost a run). Pagination is keyset with the id as tie break, because offset pagination over a table receiving writes returns duplicates and skips rows silently and only under load. An index declares the query it serves, or nobody can remove it. Personal data goes in NAMED COLUMNS so a retention rule and an erasure request can find it — an identifier inside a JSON payload is one no erasure request will reach. Secrets are references, never columns. -->
    • S3.1.e Define migration expand/backfill/verify/contract, compatibility, rollback/compensation, production rehearsal, and evidence conventions. Evidence: migration section. <!-- §5. Four SEPARATE DEPLOYABLE STEPS, and the compatibility rule is the one that makes them necessary: every intermediate state must be one where the PREVIOUSLY DEPLOYED version still works, because that is the state a rollback lands in. A migration only correct after the deploy completes cannot be rolled back. A backfill that cannot be resumed cannot be run twice, and it will be. Destructive steps ship alone and late: there is no compensating transaction for a dropped column, there is a restore from backup and an outage. Rehearsal records WALL-CLOCK LOCK DURATION, which is the number that matters rather than row count. 13/16 schemas have a migrations directory; the three without are named as the first to fix. -->
    • S3.1.f Define generated client/model ownership, commands, checked-in artifacts, drift, version skew, and hand-edit prohibition. Evidence: code-generation section. <!-- §6. Generated output IS checked in (460 tracked files today) and the section defends that choice for a monorepo where a consumer must typecheck without running a generator — then names the obligation it creates: A CHECKED-IN ARTIFACT WITH NO DRIFT CHECK IS A CACHE WITH NO INVALIDATION, so CI regenerates and fails on a diff. Hand-editing is PROHIBITED rather than discouraged: a hand edit survives until the next regeneration and fails silently at that moment, in a diff nobody reads. Generator versions are pinned in the pnpm catalog because a client generated by a newer CLI fails at the first query, in production, with an error message about a field. -->
    • S3.1.g Define required real-database unit-of-work, concurrency, constraint, migration, query-plan, isolation, backup/restore, and compatibility verification. Evidence: verification section. <!-- §7. Eight required real-Postgres verifications, each with the reason a fake cannot answer it. The rule extracted from the kit's four existing harnesses and stated normatively: A HARNESS MUST CONTAIN A NEGATIVE CONTROL THAT FAILS — outbox-atomicity injects the same failure inside one transaction and across two, and the two-transaction control MUST leave an orphan, because a harness whose control does not break has proved that it ran and not that it can fail. And they FAIL RATHER THAN SKIP with no database, because a skipped check reporting success is the evidence shape this ledger refuses. -->
    • S3.1.h Review the document against each shared entity and obtain data- platform/security ownership approval. Evidence: applicability and approval record. <!-- §8 ships the applicability review: a table of the nine shared entities S2 introduced, each with its owner scope, whether it is append-only, and whether it soft-deletes — and the observation that every one but the projection is append-only, which is the largest difference between them and most of the 439 models already in the tree. THE APPROVAL HALF IS GENUINELY BLOCKED. §10 names the two signatures outstanding (data platform for the identity scheme, payload policy, migration process and generated-artifact policy; security for the scope columns, row-level authorization, encryption and the verification list) and records that an agent may prepare the packet and may not approve it. This stays unchecked until both are recorded with a name, a date and a review date. --> <!-- evidence: docs/conventions/shared-persistence-conventions.md (11 sections) + evidence/v1-workbenches/persistence-conformance.json + a 16-test calibration suite. EVERY SCANNER RULE IS FED A MODEL THAT FAILS IT: a rule with only conforming fixtures cannot be told from one whose regular expression never matches, which matters most for the JSON rule that comes back 439/439. The calibration found a real parser defect before the baseline was committed — a } inside a // comment ended a model early and silently dropped every field after it. §11 lists the six things the document says and the scanner cannot check, so the gap between what is written and what is enforced is itself written down. -->
  • S3.2 Define tenant and workspace ownership for every shared entity. Evidence: row-access policy matrix. <!-- libs/oshun/workbench-kit/src/tenancy.ts, published as @oshun/workbench-kit/tenancy, 49 cases, 40/40 mutants killed, plus a new Postgres target tenant-isolation. 1,667 tests in the kit; all FOURTEEN Nx targets green. ISOLATION IS THE ONE PROPERTY IN THIS KIT WHERE BEING RIGHT 99% OF THE TIME IS INDISTINGUISHABLE FROM BEING WRONG: a single query that forgets its predicate is the whole failure, and it fails silently until the day two tenants happen to collide. So the module makes the scope of a row a DECLARED FACT a machine can check rather than a convention every call site has to remember. -->

    • S3.2.a Inventory shared authoritative, revision, evidence, job, event, audit, blob, review, taxonomy, calendar, catalog, and projection entities. Evidence: entity register. <!-- SHARED_ENTITY_REGISTER ships sixteen entries covering all twelve kinds the cell names, and the register is NOT decoration: the suite runs every entry through scopeFaults with an empty predicate list and asserts a fault for each scoped one and none for each unscoped one, so an entry that stopped being enforceable fails there rather than sitting in a list. -->
    • S3.2.b Classify each entity as tenant-owned, workspace-owned, cross-workspace tenant-owned, global reference, public immutable, or operational system state. Evidence: ownership classification. <!-- SIX CLASSES, BECAUSE COLLAPSING THEM IS WHAT PRODUCES THE LEAK, and the pair that matters is cross-workspace-tenant against workspace-owned: a taxonomy shared across a tenant's workspaces is not workspace-owned, and modelling it as one means either duplicating it per workspace or carrying a workspaceId that lies. A cross-workspace entity carrying one anyway is REFUSED — a column that must never be filtered on is one that will be, and the day it is, half the tenant's rows vanish for reasons nobody can reproduce. An unscoped class must DECLARE why, because an undeclared omission is indistinguishable from a forgotten one. A projection declared append-only is refused, since that is what turns S2.14's rebuildable cache into a second source of truth. All six classes are asserted to be in use, so none is decoration. -->
    • S3.2.c Define creator, owner, contributor, reviewer, delegate, operator, service, and support access per operation and lifecycle state. Evidence: row-access matrix. <!-- The tenant check is FIRST and unconditional, which is the only ordering no later rule can undo — asserted for every principal. AN OPERATOR MAY READ OPERATIONAL STATE ACROSS TENANTS AND MAY NOT READ TENANT CONTENT. Somebody debugging a stuck queue needs the queue, not the story in it, and the reason every platform eventually grants both is that the matrix had one operator column. Support reads and never writes, and only while a time-boxed grant is live — the same principal S2.13 refuses a SIGNATURE from, for the same reason seen from the other side: it may see everything and take responsibility for nothing. An export counts as a write, because it is a write to somewhere else. A service with no capability is a token that can do whatever the code around it does. A reviewer who can edit is an author whose review nobody else performed. -->
    • S3.2.d Define ownership inheritance and validation for child, join, revision, event, job, blob, cache, and projection rows. Evidence: relationship policy. <!-- A CHILD'S SCOPE MUST EQUAL ITS PARENT'S, AND A JOIN ROW HAS TWO PARENTS. The join is the case a foreign key cannot catch: each side is a perfectly valid reference to a row that exists, and it is only the PAIR that crosses a boundary, so neither table's constraints can see the other end. A link between one tenant's story and another's asset passes every per-table check ever written. A NULL scope is its own fault kind, because null matches nothing and is therefore filtered out of every isolation test that goes looking for the wrong tenant. The mutation pass caught the workspace half of that check being masked by the tenant half — same fault kind, two branches — now separated by asserting the subject. -->
    • S3.2.e Define cross-workspace share, transfer, fork/copy, export/import, support access, and tenant migration behavior. Evidence: boundary transition rules. <!-- AN EXPORT CARRIES THE SCOPE IT CAME FROM, SO AN IMPORT IS A RE-SCOPE AND NEVER A RESTORE. That is the highest-consequence rule in the module and the least likely to be caught by anything else, because it is performed DELIBERATELY, by an administrator, using a SUPPORTED FEATURE. A share across tenants is refused outright: sharing grants read without copying, which inside a tenant is a permission and across tenants is one tenant reading another's rows — the honest operations across that boundary are a fork (copy and re-scope) or a transfer (move). A transfer moves the whole subtree or none: a parent in one place and its revisions in another is a history nobody can read. Operational state may not be exported at all, because it discloses volume and timing. A tenant migration needs an approver, being the largest scope change the platform can make and the one with no partial rollback. -->
    • S3.2.f Enforce scope in schema/queries and test colliding IDs, counts, search, joins, caches, jobs, events, exports, and audit. Evidence: entity-complete isolation tests. <!-- The static half is scopeFaults, and the rule it exists for is workspace-predicate-without-a-tenant-predicate: a workspace id is unique inside a tenant AND NOWHERE ELSE, so where workspaceId = ? is correct right up until two tenants mint the same one — and it looks MORE careful than an unfiltered query, which is why it survives review. Joins are checked separately, because the predicate on the driving table says nothing about the table it was joined to. The runtime half is tools/tenant-isolation.mjs, and the reason it cannot be a unit test is the word COLLIDING: isolation bugs do not show up on distinct data, every query returns the right rows because there is only one tenant's worth of them. So two tenants are seeded with DELIBERATELY IDENTICAL workspace ids, slugs, external references and idempotency keys, and nine query shapes (count, the workspace-only filter, search, external lookup, join, aggregate, keyset page, job queue, audit trail) each run TWICE — the scoped form must return one tenant's rows and the unscoped form MUST return both. A fixture that fails to collide makes every green above it meaningless, so all nine controls are required to leak. Calibrated in both directions by hand: a scoped query made to leak is caught, and a control made not to leak is caught. The two layers are checked AGAINST EACH OTHER rather than each alone: the same shapes the engine runs are handed to scopeFaults, which must be silent on every scoped one and must complain about every unscoped one. The engine also holds the composite FK that makes a workspace id safe to filter on beside a tenant, and refuses a join row whose ends sit in different tenants. -->
  • S3.3 Define soft-delete, retention, legal-hold, and purge semantics. Evidence: lifecycle tests cover all transitions. <!-- libs/oshun/workbench-kit/src/deletion.ts, published as @oshun/workbench-kit/deletion, 45 cases, 38/38 mutants killed. 1,713 tests in the kit; all fourteen Nx targets green. EVERY OTHER LIFECYCLE IN THIS KIT IS RECOVERABLE. This one ends in an irreversible step performed by a machine at a scheduled time against data nobody is looking at, which is why almost all of the code is about refusing to take it. -->

    • S3.3.a Define active, deletion-requested, soft-deleted, retained, legal-held, purge-eligible, purging, purged, restored, and failed states. Evidence: deletion lifecycle machine. <!-- Ten states, and the table is worth reading for what is ABSENT. There is no edge out of purged: a purge is the one operation in this kit with no compensating action, and an unpurge edge would be a lie in a diagram. There is no edge from legal-held straight to purge-eligible either — a hold is released first, by a person, and only then does the clock resume. Every state is asserted reachable from active, so none is a state nothing can enter. -->
    • S3.3.b Define deletion authorization, impact preview, confirmation, grace period, tombstone visibility, undo/restore, and notification. Evidence: soft-delete transition tests. <!-- The legal-hold check runs BEFORE the edge table, because a hold is not a state the subject is in — it is a condition over it, and a subject can be held while active, soft-deleted or retained. Checking the edge first lets a perfectly legal-looking soft-deleted -> purge-eligible through while a hold is in force, which is a mutant this suite kills. The grace period is the only recovery that needs no operator: it exists so an accidental deletion can be undone by the person who made it. Every step requires an actor and a reason, because this is the lifecycle where "who did this and why" is actually asked. -->
    • S3.3.c Define retention duration/source, min/max precedence, category, tenant policy, regulatory/contractual override, and review. Evidence: retention policy tests. <!-- The strictest FLOOR and the strictest CEILING both win — the maximum of the minimums and the minimum of the maximums — and AN UNSATISFIABLE POLICY IS REFUSED RATHER THAN RESOLVED. A regulatory minimum of seven years and a contractual maximum of two cannot both be met, and settling it at purge time means choosing which obligation to breach, at 3am, in a job, with nobody told. It is a policy error and it belongs in front of whoever wrote the policies. The resolution names WHICH rule decided each bound, so a dispute has an answer, and ties resolve by source rather than by input order — asserted by reversing the input. An empty rule set is refused too: a category with no rule is not "delete whenever". -->
    • S3.3.d Define legal-hold scope, authority, reason, evidence, start/end, protected data, secrecy, release, and audit. Evidence: hold lifecycle tests. <!-- THE HARDEST CASE IN THE CELL, AND THE ONE USUALLY HANDLED BY DOING NOTHING: a subject asks for erasure while litigation requires the data kept. The hold wins — and the request is SUSPENDED, not refused, which is the whole of the person's remaining rights, because a suspended request has a review date and resumes on its own. SECRECY IS A SEPARATE FIELD AND A SEPARATE MESSAGE KEY. A confidential hold still suspends the request, and the subject is told it is under review WITHOUT being told why — a different sentence, which is why it is a different key rather than a conditional inside one, and the hold id is withheld with it. A disclosable hold is preferred when both apply. no-review-date-on-an-open-hold is the rule that matters over years: an open-ended hold is normal (litigation does not schedule itself) and one with no end AND no review is an indefinite suspension of every deletion right the subject has, created by somebody who has since left. revisionHoldMarker is the seam to S2.4's narrower per-revision hold, so that marker cannot be minted from nothing. -->
    • S3.3.e Define dependency order for metadata, blobs, revisions, events, caches, indexes, derivatives, providers, exports, and ledger records. Evidence: purge graph. <!-- PURGE ORDER IS NOT AN OPTIMISATION. It runs leaves-first because deleting the metadata first destroys the map to everything else — the blobs, the derivatives and the provider-side copies are then unreachable and become the permanent residue nobody can find to delete. Asserted positionally rather than described: every one of search-index, cache, derivative, provider-copy and blob must precede metadata, and the ledger REDACTION comes after even that, because the ledger is amended rather than removed (S3.3.g). -->
    • S3.3.f Make purge durable, idempotent, resumable, observable, and fail-closed around held/referenced data. Evidence: partial-failure purge tests. <!-- failed IS A STATE, NOT AN EXCEPTION, because the alternative is the worst outcome available: a purge that half-succeeded and reported success leaves the data partly gone, so it cannot be restored, and recorded as gone, so nobody looks. A failed run carries exactly which steps completed and which were never attempted. Resuming is re-running with the same completed set — every step is idempotent by contract, so a step that ran and failed to record is safe to run again, and a resumed purge does not start from the top and re-delete an index that no longer exists. A completed purge re-run attempts nothing. A step with nothing to do is skipped rather than done, so a purge over an entity with no derivatives is not recorded as having deleted some. -->
    • S3.3.g Preserve allowed noncontent audit/tombstone proof without retaining prohibited personal or protected content. Evidence: post-purge privacy review. <!-- A CONTENT HASH IS NOT ANONYMISATION. The surviving tombstone deliberately has NOWHERE TO PUT ONE — asserted structurally, by checking that no field name contains digest, hash or content — because a hash of a name, an email or a phone number is a lookup key for anybody who can guess the input, and guessing the input is the entire threat model for personal data. A digest would let a holder of the tombstone confirm whether a specific person was in the system, which is precisely the fact the purge removed. residueIn is the adversarial pass, in the shape S2.8.f, S2.10.e and S2.14.f use: a purge can run every step correctly and still leave the answer in the tombstone. It catches a free-text reason ("removed at the request of <name>" survives the purge that deleted <name>), an identifier smuggled into a type field, a tombstone keyed by the id it replaced (which re-identifies what was erased whenever the id came from the subject), and a step list that claims more than ran — the record somebody relies on to say the data is gone. The digest scan deliberately EXCLUDES the tombstone's own id, and the mutation pass forced that exclusion to be justified: an opaque random id rendered as hex is indistinguishable from a digest by shape, and scanning it would make every correctly-minted tombstone report a residue, which is how the check gets switched off. Pinned in both directions. -->
  • S3.4 Define revision and idempotency unique constraints. Evidence: concurrent duplicate writes collapse safely. <!-- libs/oshun/workbench-kit/src/idempotency.ts, published as @oshun/workbench-kit/idempotency, 29 cases, 25/25 mutants killed, plus a new Postgres target idempotency-contention. 1,743 tests in the kit; all FIFTEEN Nx targets green. AN IDEMPOTENCY KEY IS THE THING EVERYBODY AGREES THEY HAVE AND ALMOST NOBODY SCOPES CORRECTLY, and the failure is asymmetric: scoped too widely it returns one caller another's result, scoped too narrowly it runs an irreversible operation twice. Both are silent. -->

    • S3.4.a Define unique entity revision ID and monotonic sequence/head constraints scoped by tenant/workspace and object. Evidence: schema constraints. <!-- Eight named constraints, each PLACED at the layer that can hold it, with the reason written beside it. The placement is the content: a constraint list is otherwise a list of things somebody BELIEVES are enforced. unique-sequence-per-entity is called out as the one that decides an append-only race, and a SELECT-then-INSERT does not — the same lesson the kit's revision-race harness cost a run to learn. The database and application halves are asserted to PARTITION the set, so nothing is in both or neither. -->
    • S3.4.b Define parent/base revision referential integrity and prevent self, cross-object, missing, and cyclic lineage. Evidence: constraint and integrity tests. <!-- TWO OF THE EIGHT CANNOT BE DATABASE CONSTRAINTS AT ALL, and saying so is the honest half. A FOREIGN KEY PROVES THE PARENT EXISTS AND NOT THAT IT BELONGS TO THIS ENTITY — comparing the two entity ids is a join the constraint does not perform — and a single head is a count over siblings rather than a property of a row. They are marked application rather than written down as if a CHECK covered them. Cyclic lineage is S2.4's checkHistoryIntegrity and is proved against a real engine by history-constraints, which is cited rather than duplicated: the same three-colour walk, in the one place it lives. -->
    • S3.4.c Define idempotency record uniqueness by tenant, actor/client, operation/version, key, and target where required. Evidence: key scope constraints. <!-- THE KEY IS NOT THE CONSTRAINT. Five fields plus the target, and dropping any one is a bug somebody has shipped: without the tenant two customers collide, without the client one caller's retry returns another's result, without the operation a reused key crosses operations, WITHOUT THE OPERATION VERSION a retry arriving after the contract changed hands a v1 result to a v2 caller who will parse it. THE ENCODING IS INJECTIVE, and the suite found the first draft was not: a - sentinel for a null target collided with a real target id of -, and a separator inside an id let a|b and a + |b produce one key. A key two scopes can forge returns one caller another's result, which is the exact failure this module is about, so it is JSON rather than a join. key-is-not-opaque looks fussy and is not: a key derived from an email or a URL leaks that value into every log line, error message and support ticket that quotes it back. -->
    • S3.4.d Store canonical request hash and reject same key with different input, target, identity, or contract. Evidence: mismatch tests. <!-- THE DIGEST IS COMPARED BEFORE THE STATE, and the ordering is the rule. A stored record in completed whose digest differs is NOT a replay — it is a caller reusing a key for something else, and returning the stored result reports SUCCESS FOR SOMETHING THAT NEVER RAN, to somebody who will never look again. Every other branch returns a result, so the comparison has to come first; asserted across all four states. -->
    • S3.4.e Atomically return existing accepted/running/completed result under concurrent identical requests. Evidence: high-contention real-DB tests. <!-- THE ONE CLAIM IN S3.4 NO UNIT TEST CAN MAKE, and the wrong implementation — SELECT, then INSERT if absent — passes every serialized test ever written and fails only under exactly this load. So tools/idempotency-contention.mjs opens 32 real connections and fires them at one key four ways: the claim (exactly one wins, and the 31 losers RE-READ and must find the winner's row — returning NOTHING to the loser is the second-most-common bug here and looks like a transient failure rather than a race), the replay (all 32 get the same result reference), the conflict (all 32 sending a different digest are refused without disturbing the stored record), and THE NEGATIVE CONTROL: the naive read-then-write everybody writes first, raced identically, which MUST perform the effect many times. It performed it 32 times. Calibrated by hand in both directions — an upserting claim is caught, and a control repaired so it cannot double-effect is caught. -->
    • S3.4.f Define expiry/partition/archive behavior without allowing late duplicates of irreversible commands. Evidence: retention boundary tests. <!-- AN IDEMPOTENCY RECORD'S RETENTION IS A PROPERTY OF THE OPERATION'S REVERSIBILITY, NOT A GLOBAL TTL. Expiring the record for an irreversible command re-enables a duplicate of it — a second charge, a second publish, a second delete — and a 24-hour default applied uniformly across every operation is how that reaches production. An expired irreversible record REFUSES the request rather than letting it through, because letting it through performs the operation a second time, which is exactly what the record existed to prevent. earliestExpiry takes the irreversible FLOOR and survives a misconfigured policy where somebody set the retry window shorter than the reversible cap — asserted, because that is the configuration a partition job would otherwise use to drop a record early. A record expired the moment it is written reports that ALONE, because every other check would be reasoning about a lifetime that does not exist. -->
  • S3.5 Define searchable indexes for stage, status, owner, dates, taxonomy, gate status, and stable identifiers. Evidence: query plans on realistic volume avoid sequential scans for critical paths. <!-- src/indexes.ts (36th entry point) + tools/query-plans.mjs (15th Nx target). AN INDEX LIST IS THE EASIEST ARTIFACT IN A SCHEMA TO WRITE AND THE HARDEST TO ARGUE WITH A YEAR LATER, because by then nobody remembers which query each one was for and removing any of them is a gamble — so the workload ships first and every index names the queries it exists for, in both directions. Verified on a real Postgres 16 at 948,000 rows: 24 queries judged, every interactive one blocking without its index, three separate controls. The static model and the engine calibrate each other — coverageFor predicts a sort or its absence BEFORE each query runs and the harness fails if the plan disagrees, which is how the module's central rule got corrected: the first version carried an "ordering is broken" flag that a mutation pass showed was dead (masked by the alignment check), and asking Postgres showed the real rule is that A LIST AND AN ORDERING ON THE SAME COLUMN ARE MUTUALLY EXCLUSIVE — the planner keeps the ordering and the list narrows nothing. 100 tests in the cell, 20 mutations killed. -->

    • S3.5.a Enumerate critical filter/sort/pagination queries and production- shape row/cardinality distributions. Evidence: query workload manifest. <!-- SHARED_WORKLOAD in libs/oshun/workbench-kit/src/indexes.ts: 24 queries, each with its predicates (equality/membership/range/text-prefix/ is-null), sort keys, pagination mode and the person waiting for it. SHARED_VOLUME_PROFILES carries the distribution for all 8 entities — volume class, write class, p50/p99 rows per tenant, LARGEST TENANT SHARE, per-column distinct/null/sensitive, and the column sets unique within scope. distributionFaults reports 7 kinds, including a column any query filters on with no declared cardinality. THERE IS NO offset MEMBER of PAGINATION_MODES: a rule in a document can be forgotten and a type cannot. And a workload with no declared skew is one that will be measured on whatever the fixture contained — uniform data makes every index look excellent. -->

    • S3.5.b Add tenant/scope-prefixed indexes for lifecycle stage and domain status combinations. Evidence: stage/status query plans. <!-- LIFECYCLE_INDEXES: wbe_board_stage (tenant, workspace, stage, updatedAt DESC, id DESC) and wbe_review_queue, partial on the one stage a review queue is ever about. The scope prefix is DERIVED from the ownership class S3.2 assigned rather than restated, and narrowed to the columns every served query actually pins — an inbox is a tenant-wide read of workspace-owned rows, so demanding a workspace prefix would demand an index it cannot enter. Plans measured by pnpm nx query-plans: the board pages with no Sort node. The same index for the same board filtered to SEVERAL stages sorts its whole match to return 25 rows, which is the one declared limitation below. -->

    • S3.5.c Add owner, actor, assignee, reviewer, team, and delegated-scope indexes required by inbox/search. Evidence: ownership query plans. <!-- OWNERSHIP_INDEXES plus OWNERSHIP_DIMENSIONS, which gives each of the six words its own row: a plan covering five of six reads identically to one covering six. assignee → wbe_assignee_open (partial on the open flag, not on a stage list, because an index predicate rewrite is a rebuild); reviewer → wbe_review_queue; team/owner → wbe_team_unassigned, partial on assignee_id IS NULL, the one case where indexing a mostly-null column is the cheap option. ACTOR IS ANSWERED ON THE AUDIT STREAM (audit_actor_activity), not on the entity: "what did this person do" is a question about the trail, and the entity table can only answer it for the last person who touched each row — and audit_event has index budget while the central authoring table does not. DELEGATED SCOPE IS THE SAME INDEX with a different principal bound, so it is a separate ACCESS PATH in the manifest and not a second index; what differs is the authorization, and S3.2 owns that. Search: wbe_title_prefix. Plans measured. -->

    • S3.5.d Add created/updated/due/effective/expiry/embargo/publish/event time indexes with deterministic ID tie-breakers. Evidence: date query plans. <!-- TEMPORAL_INDEXES + TEMPORAL_DIMENSIONS, one row per kind of time. Five are kit columns with an index (created, updated, due, expiry, publish, event); effective and embargo are declared DOMAIN-OWNED with the rule any plugin's version must follow, because an effective date belongs to the thing that is effective and an embargo is a publication rule a newsroom has and a courseware workbench does not. sortIsTotal is the tie-break check and it counts what the query FIXES, not whether the sort ends in id: the audit stream fixes objectId and sorts by sequence, and (objectId, sequence) is unique, so bolting an id on would buy a wider index for nothing. A TIE BREAK IS A CORRECTNESS PROPERTY, NOT A PERFORMANCE ONE, and the harness proves it rather than asserting it: 3,000 rows stamped with one instant, then the same keyset walked three ways. An exclusive cursor on the timestamp alone returned 25 of 150 rows and stopped; an inclusive one returned 75 rows of which 25 were distinct, handing back the same page; with the id in the cursor, 150 rows in exactly the total order. -->

    • S3.5.e Add taxonomy membership/path/alias indexes for supported hierarchical and facet queries. Evidence: taxonomy query plans. <!-- TAXONOMY_INDEXES + TAXONOMY_DIMENSIONS. Membership in BOTH directions (tax_link_term, tax_link_entity) — a join table gets one index by accident, the one its key gives it, and the reverse lookup is the one that turns into a scan on the page showing a record's tags. tax_path_prefix makes a subtree a prefix range. tax_alias_exact is UNIQUE on the raw alias: an index on lower(alias) resolves two aliases that differ only in case to one term, and nobody sees the merge until the wrong term is applied. THE C COLLATION IS NOT AN OPTIMISATION HERE. In the database collation a b-tree cannot serve LIKE 'x%' at all — the plan reads the whole scope and filters, which is not a sequential scan and passes any gate looking only for one. Probed against Postgres 16 before the model was written: the other spelling, text_pattern_ops, makes the prefix work and then answers ORDER BY path with an incremental sort, because they are different operator families. COLLATE "C" gives both. -->

    • S3.5.f Add gate ID/verdict/severity/freshness/evaluator/policy/expiry indexes for eligibility and triage. Evidence: gate query plans. <!-- GATE_INDEXES + GATE_DIMENSIONS, one row per property. gate id + freshness → gate_freshness; verdict → gate_triage (the console reads one verdict at a time, worst first, so it is an equality and not a filter); the S2.10 fold's read → gate_target_history; expiry → waiver_expiry, partial on active because S2.11 keeps revoked and superseded grants readable forever and none of them can expire into anything. SEVERITY IS DELIBERATELY NOT A COLUMN ON THE RUN. It is a property of the gate DEFINITION at a lifecycle stage (S2.9) — the same failing run is a blocker before release and a warning in draft — so denormalising it onto the run would make the copy authoritative, which is the exact drift S2.9 exists to name. evaluator/policy are a DECLINED index with the trade-off stated: gate_run is append-only-hot with a budget of three, and the comparability report runs on the scale of a policy review rather than a page load, so it is declared full-scan-batch and the gate is told to expect a scan rather than to fail one. -->

    • S3.5.g Add exact internal/external stable identifier uniqueness and lookup indexes without fuzzy collision. Evidence: identifier plans. <!-- IDENTIFIER_INDEXES + SHARED_IDENTIFIER_LOOKUPS + identifierFaults, which reports two things and neither of them shows up as slow: an identifier lookup with no UNIQUE index (the second row arrives without complaint and the lookup starts answering with whichever one the index reaches first), and uniqueness that omits the ISSUING SYSTEM. AN EXTERNAL IDENTIFIER IS UNIQUE PER ISSUER AND NOWHERE ELSE, so wbe_external_ref is unique over (tenant, issuer, ref): uniqueness over (tenant, ref) holds exactly until a second upstream is connected, and then two unrelated records collide on one string and one of them stops being reachable. No lower(), no trim, no normalisation — a fuzzy identifier index is a merge nobody asked for. -->

    • S3.5.h Measure write/storage cost, redundant indexes, partial-index predicates, sensitive-data exposure, and plan stability. Evidence: index trade-off report. <!-- MEASURED, not estimated: evidence/v1-workbenches/ workbench-kit-query-plans.json, written by the harness from a real run. Storage: 95.5 MiB of indexes against 82.1 MiB of table, per index. Write cost: 20,000 inserts took 337ms with the eight workbench_entity indexes and 44ms without them — 7.7x, which is the argument for a hard budget rather than a preference. Static half in indexFaults: redundantIndexes (a left prefix with no narrower partial predicate — and a UNIQUE index is never redundant however wide the other one is, because it is a constraint wearing an index's clothes); partialIsImplied (a partial predicate the query never states is an index the planner cannot prove contains the rows, so it will not use it); sensitive-column-indexed over columns AND include payloads (an index on personal data is a second copy an index-only scan reads and a base backup carries); INDEX_BUDGET per write class. Plan stability is estimate-off-by-an-order-of-magnitude, warn rather than block because it is a leading indicator — and it is suppressed under a Limit, since a node told to stop early produces fewer rows than planned BECAUSE it was told to. DECLINED_INDEXES is the mirror of "an index declares the query it serves": a query with a named reason for having no index, so the next person does not add one on an append-only table forever to speed up a twice-a-year report. -->

    • S3.5.i Fail performance checks when critical realistic queries regress to unsafe sequential scan or unstable sort. Evidence: query-plan gate. <!-- judgePlan + parseExplainPlan in src/indexes.ts, driven by the new Nx target query-plans (tools/query-plans.mjs, 15th target). Seven plan faults, each with a verdict conditional on the entity's volume class and the query's criticality — a sequential scan over a reference table is the correct plan and failing it teaches people to add indexes that cost writes and buy nothing. THE MOST DANGEROUS PLAN IS NOT A SEQUENTIAL SCAN. It is filter-discarded-most-of-what-it-read: an index scan that enters on the tenant, reads 40,000 rows and returns 22, because the predicate that discriminates is not one the index could apply. Every gate that greps for Seq Scan passes it. parseExplainPlan THROWS on input it cannot read rather than returning an empty plan, and a missing plan is indeterminate — an absent measurement is not a passing one. THE HARNESS, at 948,000 rows across 8 tables: 24 queries judged, 1 blocking and it is the declared limitation, asserted to be exactly the plan fault the declaration NAMES so a stale declaration cannot go on excusing something else. NEGATIVE CONTROL: with every index dropped, all 18 interactive queries must come back blocking — if one passed, the fixture could not tell a good plan from a bad one. THIRD CONTROL: the same gate on a 500-row copy WITH every index blocks 6 of 18 correct plans, so the harness refuses to report anything below a 50,000-row floor. A toy fixture does not give a weaker answer, it gives a wrong one. Two more measured: the small tenant's board through an index led by stage discards 28,572 rows and blocks, and through the tenant-leading one discards 0 and passes — neither is a sequential scan; and a list plus an ordering on one column read 8,572 rows with no Sort node, confirming the planner keeps the ordering and drops the array. -->

  • S3.6 Implement a transaction-bound outbox. Evidence: commit-with-event and rollback-without-event integration tests. <!-- src/outbox.ts (37th entry point) + tools/outbox-dispatch.mjs (16th Nx target, outbox-dispatch). 53 tests in the cell, 1,898 in the kit, 24 mutations killed with no survivors. The evidence line asks for two integration tests and the harness runs FIVE properties with FIVE controls, because each of these is a claim about a database and a scheduler that a fake store would only prove about the fake: the atomic write (control: two transactions, which MUST orphan), the concurrent claim (control: unlocked SELECT-then-UPDATE, which MUST double-claim), per-subject order (control: unpartitioned claiming, which MUST invert), the crash window (control: mark-then-publish, which MUST lose the event), and lease recovery. Two of the six decisions are enforced by TYPE rather than by rule: PayloadRef has no member for a pointer to mutable state, and markPublished cannot be reached without a branded receipt only a real broker acknowledgement produces — so the publish order that loses events cannot be written at all. -->

    • S3.6.a Define outbox row identity, tenant, producer, aggregate/subject, event kind/version, sequence, causation/correlation/trace, payload ref, availability, attempt, and state. Evidence: outbox schema. <!-- OutboxRow in libs/oshun/workbench-kit/src/outbox.ts, with rowFaults reporting 14 kinds, every one of them fed a failing row. THE OUTBOX ID IS THE PUBLISHED EVENT'S ID, never re-minted per attempt: an outbox is at-least-once by construction, so the second delivery has to be recognisable as the same event or every consumer's dedup is decoration, and its key is exactly S3.4's idempotency scope. PayloadRef HAS NO lookup MEMBER — an event carrying only {id} publishes whatever that row holds when the consumer gets there, which is not what happened, it is what happened LAST; leaving the member out means no producer can write one. CAUSATION IS NOT CORRELATION: with only a correlation you know forty events belong together and not which produced which, which is what an incident asks. leased is a STATE, not a flag on pending — collapsing them is how a dispatcher that died leaves rows nothing will pick up. -->

    • S3.6.b Insert authoritative mutation and outbox intent in one database transaction through the shared unit of work. Evidence: atomic commit/ rollback tests. <!-- commitWithEvents is the unit of work, and THERE IS NO EXPORTED FUNCTION THAT COMMITS A MUTATION ALONE. That is the design: the guarantee is not "remember to append the outbox row", it is that a signature which lets you forget does not exist. emissionFaults runs BEFORE the transaction opens, so a commit carrying an event about a different aggregate — atomic with the wrong thing, and reading as more thorough rather than less — never reaches the database. Proved twice. Unit: the call order is asserted exactly (begin, mutate, append, commit), and an append that throws rolls the mutation back. Real engine (pnpm nx outbox-dispatch, phase 1): commit wrote the row and its event; an injected failure rolled both back; and THE CONTROL — the same failure across two transactions — left an authoritative row with no event beside it, which is the orphan this design exists to prevent. -->

    • S3.6.c Claim unpublished rows with bounded batches, ordering/fairness, leases or skip-locking, and concurrent dispatcher safety. Evidence: dispatcher contention tests. <!-- ClaimPolicy + claimFaults (6 kinds) + selectBatch + claimOrder + reclaimable. PER-SUBJECT ORDER IS THE ONLY ORDER THERE IS AND CONCURRENCY BREAKS IT BY DEFAULT, so subjectPartitioned holds one event per subject in flight; the sequence numbers otherwise record an order nobody observed. Fairness is a PER-TENANT CAP inside the batch, because one tenant's backlog is otherwise every dispatcher's whole batch. A lease that is not longer than the publish timeout is a fault: it expires while the publish is still in flight. Real engine, phase 2: 8 dispatchers over 240 rows. FOR UPDATE SKIP LOCKED claimed every row exactly once; THE CONTROL — the unlocked SELECT-then-UPDATE everybody writes first — claimed ~900 rows more than once (and deadlocks outright when it updates a multi-row set, which is the same defect wearing a louder hat). Phase 2 also runs selectBatch from the BUILT package and the SQL claim over the same rows and fails if they choose different subjects: a selection rule and a statement that disagree is how a policy becomes a comment. LANDMINE: Postgres refuses FOR UPDATE beside DISTINCT, so the partition is two NOT EXISTS (no leased sibling, no earlier pending sibling). -->

    • S3.6.d Publish with stable event ID and destination/partition semantics, then mark outcome without losing an externally published event. Evidence: publish/ack crash-window tests. <!-- PUBLISH, THEN MARK — NEVER THE REVERSE. The crash window between them cannot be closed, only pointed in a direction, and the tidier-looking order is the one that loses data. So markPublished takes a DeliveryReceipt, the receipt is branded with a type-only symbol, and the ONLY function returning one is receiveAck — which refuses an acknowledgement with no broker offset, because a call that returned is not an acknowledgement. The unsafe order cannot be written with these types. The partition key is the SUBJECT and never the tenant: a broker preserves order within a partition and nowhere else, so partitioning by anything coarser than the unit of ordering throws that order away at the last possible moment. Real engine, phase 4: publishing first and crashing delivered the event TWICE UNDER ONE ID, which every consumer deduplicates. THE CONTROL, mark-then-publish, delivered it ZERO times with the row saying published — no retry will find it and no alarm will fire. Phase 3 measured the ordering half: partitioned claiming published 1,2 and the unpartitioned control published 2,1. -->

    • S3.6.e Apply retry classification, backoff, max age/attempts, circuit, and poison quarantine. Evidence: delivery fault tests. <!-- decideDelivery + backoffSeconds + circuitFor. UNKNOWN IS NOT RETRYABLE AND NOT PERMANENT: folding it into the first turns one poison message into an infinite loop that starves everything behind it; folding it into the second discards an event because DNS was slow. It gets its own lower attempt ceiling and is then a person's problem. Order is not arbitrary and the mutation pass checks each step: a permanent failure quarantines whatever its attempt count OR its age (retrying a rejected schema forty times produces forty identical rejections and one very confident engineer); age is checked before attempts, because the row that has been failing for a week on a long backoff is the one nobody is looking at. MAX AGE AND MAX ATTEMPTS ARE BOTH NEEDED — attempts alone lets a row on a long backoff live for weeks. Jitter is a FRACTION of the delay, not a constant, because a constant added to a capped backoff stops spreading exactly when the queue is worst. The circuit counts a failure RATE over a window, not consecutive failures: a destination failing every other message is broken, and a consecutive counter never trips on it because every success resets it. It says nothing below a minimum sample count, because opening a circuit on two failures is how a deploy takes a destination down. -->

    • S3.6.f Expose queue age, attempts, lag, failures, poison, throughput, correlation, and safe replay operations. Evidence: outbox operability tests. <!-- queueHealth + byCorrelation + decideReplay (5 refusals). DEPTH IS NOT THE ALARM; AGE IS. A queue holding a hundred thousand rows draining at ten thousand a second is healthy and a queue holding eleven rows that has not moved in an hour is an incident — depth alarms on the first and says nothing about the second, which is why every dashboard that alarms on depth has a story about the outage it missed. The suite pins that with the two queues side by side. A REPLAY IS SAFE FOR EXACTLY ONE REASON: it is the same event id, so every consumer that already saw it drops it. replay-would-mint-a-new- event-id is the refusal that matters — under a new id it is not a replay, it is a second event, and it is the shape a well-meaning operator reaches for when the original row looks stale. Replaying a quarantined row without resolving what quarantined it repeats the failure; past the retention window consumers have discarded their dedup state, so the replay arrives as new. -->

    • S3.6.g Retain/archive/delete dispatched rows according to replay, audit, privacy, and incident policy. Evidence: outbox retention tests. <!-- OutboxRetentionRule + outboxRetentionFaults (4 kinds) + outboxRetentionAction. TWO CLOCKS, NOT ONE, and that is the whole section. The ROW is the only record that something was published, so deleting it on success makes the first incident question — "did we send it?" — permanently unanswerable; it is kept for the incident window. The BODY may hold personal data that S3.3's erasure has to reach, so it is redacted on its own shorter schedule. Collapsing them means choosing between an unanswerable incident and an unerasable payload, and personal-data-kept-as-long-as-the-row is the fault that names it. AGE IS NEVER APPLIED TO AN UNPUBLISHED ROW: a stuck row is exactly the evidence of why it is stuck, and a quarantined one is an open question whose answer is in the payload. A retention window shorter than the replay window the console offers is a fault too — the operation it lists would fail on rows it still shows. -->

  • S3.7 Implement an idempotent inbox. Evidence: duplicate and out-of-order events produce one correct projection. <!-- src/inbox.ts (38th entry point) + tools/inbox-idempotency.mjs (17th Nx target). 44 tests in the cell, 1,943 in the kit, 20 mutations killed. THE ASYMMETRY THE WHOLE MODULE FOLLOWS FROM: an event processed twice is a VISIBLE bug, and an event suppressed as a duplicate and never applied is an INVISIBLE one — because the suppression is exactly the mechanism that would otherwise have fixed it. Every decision here points away from the second. Two of them are enforced by vocabulary rather than by rule: there is no receipt-first binding strategy, and the dedup key has no shape that omits the consumer. -->

    • S3.7.a Define inbox identity by tenant, consumer/projection, producer, event ID/version, partition/subject, and payload hash. Evidence: inbox schema. <!-- InboxKey + inboxKey + InboxReceipt + validateKey (7 kinds) in libs/oshun/workbench-kit/src/inbox.ts. THE DEDUPLICATION KEY CARRIES THE CONSUMER. One event delivered to three projections has to be processed three times, and a receipt keyed on the event id alone — which is what "have we seen this event?" naturally becomes — lets the first consumer's receipt silently suppress the other two. Proved rather than asserted: the harness runs both keys and the event-id-only one suppresses the second projection with no error, no row, nothing to find. The key is JSON.stringify of the tuple, for the reason S3.4 found the hard way: any separator can appear inside a field. consumer === producer is refused — a service consuming its own event deduplicates against its own write and stops processing anything. -->

    • S3.7.b Validate envelope/version/authorization/scope and persist receipt before or with side effects. Evidence: invalid-event tests. <!-- decideIntake with 6 refusals, each fed a failing request. THE ORDER IS THE POINT: the tenant is checked before anything about the content, because an event that crossed a boundary is a security question and not a routing one — and a version check that refused first would give a probe a way to learn which versions exist. The suite pins that with a request that is BOTH foreign-tenant and unsupported-version and asserts which answer comes back. AN UNKNOWN KIND IS REFUSED, NOT IGNORED: silently dropping an event a consumer does not recognise is indistinguishable from the outside from processing it, and it is what happens on the day a producer ships a kind this consumer was supposed to handle. -->

    • S3.7.c Deduplicate concurrent and repeated delivery and return prior processing outcome without repeating effects. Evidence: duplicate contention tests. <!-- decideClaim returning 6 results. THE BODY IS COMPARED BEFORE THE STATE: a stored receipt whose payload hash differs is not a duplicate at all, it is one event id carrying two events, and answering it with the stored outcome reports success for something that never ran. A DUPLICATE IS ANSWERED WITH THE PRIOR OUTCOME, never with "already processed" — a consumer that acknowledges without knowing what happened cannot tell an applied event from an ignored one, and those want different acks. A processed receipt carrying no outcome is refused rather than answered. Real engine (pnpm nx inbox-idempotency, phase 1): 24 simultaneous deliveries of one event applied it ONCE through a unique constraint and exactly one caller believed it ran the work. THE CONTROL — the read-then-write dedup everybody writes first — applied it 24 times. -->

    • S3.7.d Define per-subject sequence, gap, late, out-of-order, and commutative-event handling. Evidence: ordering fixture matrix. <!-- decideOrdering over 5 dispositions + releasable + foldDeliveries. A GAP IS NOT A LOSS UNTIL IT IS OLD: position 5 arriving with 4 unseen is almost always 4 still in flight, so it is HELD; treating the gap as a loss immediately turns every reordering into an incident, and holding forever is a projection that silently stops advancing — so past a tolerance it becomes gap-is-a-loss, a third answer that is neither. WHETHER A KIND COMMUTES IS DECLARED, NEVER INFERRED, with a reason: a tag added and a tag removed commute, a balance credited and a balance multiplied do not, and the difference is not visible in the payload. The fold's high-water mark is the largest CONTIGUOUS position and never the largest applied — for a commutative kind those are different numbers, and using the wrong one drops a late-arriving 1 and 2 as superseded. A mutation pass caught exactly that, and the case that kills it is now in the suite. -->

    • S3.7.e Apply inbox receipt and projection/domain mutation in one transaction or documented saga. Evidence: crash atomicity tests. <!-- BINDING_STRATEGIES HAS NO receipt-first MEMBER, and that is the sharpest thing in the cell. On the producing side, writing the receipt before the effect costs a duplicate; HERE IT COSTS THE EVENT PERMANENTLY, because the receipt is what suppresses the redelivery that would have repaired it — the deduplication is what makes the loss permanent. Leaving the option out of the vocabulary is the only way to make it unavailable rather than merely discouraged. A saga is allowed and must name BOTH its compensation and its reconciliation; a compensation beside a real transaction is refused as code that will never run and will be trusted anyway. Real engine, phase 3: one transaction, crash, redelivery → the effect landed exactly once. THE CONTROL, receipt-first → the effect landed 0 times AND the redelivery was suppressed. The event is simply gone. -->

    • S3.7.f Persist processing version/result/error/attempt and support safe retry after transient failure. Evidence: retry tests. <!-- The receipt carries processingVersion, attempt, outcome and lastError; decideRetry reads them. THE PROCESSING VERSION IS PART OF THE RECEIPT: a receipt written by v1 of a projection does not mean the event has been processed by v2, and treating it as one is how a fixed reducer never runs on the history it was fixed for — decideClaim returns stale-processing-version for a newer consumer and, deliberately, does NOT reprocess for an older one. A non-transient failure is quarantined on the first attempt, because the same reducer on the same body fails the same way and more attempts produce a longer log and nothing else. -->

    • S3.7.g Rebuild a projection from replayed duplicate/out-of-order history and match authoritative expected state. Evidence: full replay parity. <!-- replayParityFaults compares FIELD BY FIELD, never by a digest — S2.14's rule, for the same reason: a checksum answers "do they differ", and the question worth asking is which field, because a drifted reducer, a lost event and a row somebody wrote to directly all produce one bit and three different investigations. It distinguishes a field the replay lost from one the store never had, which are opposite diagnoses. foldDeliveries is the reference, and the suite asserts the property the whole module is for: five shufflings of one history, duplicates included, all produce the same projection. Real engine, phase 4: the shuffled delivery [3,1,1,5,2,4,3] left the projection at [1,2,3,4,5] and the control that applies on arrival left it at [3,1,5,2,4]; the kit's fold and the engine's projection were then compared field by field. -->

  • S3.8 Implement poison-event quarantine and replay. Evidence: operator workflow test preserves cause and audit. <!-- src/quarantine.ts (39th entry point) + tools/quarantine-blast-radius.mjs (18th Nx target). 39 tests in the cell, 1,983 in the kit, 22 mutations killed with no survivors. This cell is where S3.6 and S3.7 both stop. The reusable shape is the THREE-WAY comparison: blast radius is a property where both errors are expensive and only one of them is loud, so a harness with a single control cannot show it — too wide is an outage somebody notices in minutes, too narrow is a projection that silently applied position 4 without 3. -->

    • S3.8.a Classify malformed, unsupported, unauthorized, invariant, dependency, deterministic-handler, repeated transient, and unknown poison conditions. Evidence: classification tests. <!-- classifyPoison over all 8 conditions in libs/oshun/workbench-kit/src/quarantine.ts, each fed signals that produce it. ORDER IS CHEAPEST AND MOST CERTAIN FIRST: a message that would not parse cannot have violated an invariant, and an UNAUTHORIZED event is refused WITHOUT being inspected for what it would have done. REPEATED TRANSIENT IS ITS OWN CONDITION — a timeout that has failed twelve times is not a timeout, and filing it under unknown hides the one fact that matters: the network was fine and the handler is not. UNKNOWN IS NOT A BUCKET: it carries unclassified: true, because a growing unknown pile is a taxonomy that has stopped describing reality and nothing about it looks wrong on a dashboard. dependency-missing is kept separate from deterministic-handler-failure — one is usually an ordering problem and the other is a bug, and folding them applies one's fix to the other. -->
    • S3.8.b Quarantine original immutable event reference/hash, safe payload access, consumer/version, attempts, cause chain, trace, and timestamps. Evidence: quarantine schema tests. <!-- QuarantineRecord + recordFaults (8 kinds). THE RECORD HOLDS A REFERENCE AND A HASH, NEVER A SECOND COPY OF THE TRUTH: a copy diverges and then two records disagree about what happened with no way to tell which is the event; the hash is what makes the reference checkable, and S3.8.e uses it to refuse a replay of something that has since moved. THE PAYLOAD IS NOT RAW BY DEFAULT. A quarantine console is where every failed message in the system ends up, which makes it the one screen where personal data from everywhere is displayed together, to whoever is on call — so authorized-raw requires a named assignee and the reading is attributable. The cause CHAIN is stored newest-first and checked for it: the last error is the symptom. -->
    • S3.8.c Stop hot-loop delivery while isolating only the affected event, subject, partition, or consumer according to ordering semantics. Evidence: blast-radius tests. <!-- isolationFor + deliveryIsBlocked, and pnpm nx quarantine-blast-radius. THE BLAST RADIUS IS THE SMALLEST SCOPE THAT PRESERVES THE ORDER, AND BOTH DIRECTIONS ARE WRONG — which is why this is the rare harness with THREE configurations and two of them required to be visibly wrong rather than one control: event scope left the projection at [1,2,4,5] — position 4 applied over a missing 3. Nothing downstream reports it. This is the dangerous direction because it is the quiet one. consumer scope left an UNRELATED subject with nothing applied at all: an outage with a different name. subject scope left the poisoned subject waiting and the unrelated one moving. The only answer that is both correct and available. The scope is DERIVED from the kind's commutativity (S3.7.d's declaration) rather than chosen per incident, and the harness checks the kit derives the middle answer. partition is in the vocabulary because an isolation can be declared there, and nothing DERIVES it — a partition holds many subjects, so choosing it stops streams that were never affected. -->
    • S3.8.d Expose authorized inspect, assign, comment, link incident/fix, suppress, transform/migrate, replay, skip, and archive actions. Evidence: operator API/UI journey. <!-- OPERATOR_ACTIONS (9) + authorizeAction with 7 refusals, each fed a request that produces it. SKIP AND SUPPRESS ARE DIFFERENT ACTIONS: skipping abandons ONE event and the position it held, suppressing stops a CLASS of them, and a suppression scoped to a single event is REFUSED — because calling a one-off decision a standing rule is how it stops being reviewed. A skip needs a reason, because the gap it leaves in the sequence will outlive the person who made the decision. A TRANSFORM MUST RETAIN THE ORIGINAL: a repaired event replayed with no record of what it was is a forgery, however well-intentioned the repair. A closed record may be read and commented on and nothing else — reopening is a new decision, not a continuation. -->
    • S3.8.e Revalidate event, handler version, dependencies, authorization, sequence, and dry-run impact before replay. Evidence: replay guard tests. <!-- replayGuards runs all 7 and failedGuards names the ones that did not pass. The one that surprises people: REPLAYING WITHOUT A CHANGE REPEATS THE FAILURE. If the event is the same, the handler is the same and nothing around it moved, the replay is a retry with a longer audit trail — so it is refused unless the handler version moved OR the cause was environmental. The others: the reference still resolves; the HASH is unchanged (otherwise the replay runs something other than what was quarantined); dependencies are present; the producer is still authorized (a quarantined event does not carry a grant forward); and the SEQUENCE POSITION IS STILL MEANINGFUL — if the subject has applied past it, replaying now applies an old state over a newer one. -->
    • S3.8.f Replay idempotently with new attempt identity while preserving original event identity and complete audit. Evidence: replay integration tests. <!-- planReplay. THE EVENT IDENTITY IS PRESERVED AND THE ATTEMPT IDENTITY IS NEW, and both halves are load-bearing: preserved so S3.7's receipt recognises it — which makes replaying something that DID apply a harmless no-op — and new so the audit has two runs to point at. An attempt id equal to the event id is refused, as is a replay on a failed guard or on a PARTIAL checklist (a replay authorized on fewer guard results than there are guards is authorized on the ones somebody remembered). Proved at a real engine: republished under its own id the receipt recognised it and the effect stayed at 1; re-minted as a new event — the shape "just re-emit it" takes at 3am — the receipt missed it and the effect was applied twice. -->
    • S3.8.g Alert and reconcile growing quarantine, blocked partitions, stale owner, and unresolved age. Evidence: poison operations tests. <!-- quarantineHealth + quarantineAlerts over 5 alert kinds. A BLOCKED PARTITION IS AN OUTAGE WITH A DIFFERENT NAME: the quarantine count is a workload number and the blocked-partition AGE is an availability one, and a console showing only the first is one where a stopped stream looks like a busy week. taxonomy-drifting fires on the FRACTION of open records that are unclassified rather than the count, because the question is whether the conditions still describe what is failing. staleOwner counts assigned-and-untouched separately from unassigned — the second is a queue and the first is a false sense that somebody has it. -->
  • S3.9 Implement durable jobs with lease, heartbeat, progress, checkpoint, cancellation, retry, and dead-letter state. Evidence: clock-controlled integration tests. <!-- src/jobs.ts (34th production entry point) + tools/job-fencing.mjs (20th Nx target). 112 tests in the cell, 2,096 in the kit, 43 mutations killed with no survivors. S3.6 and S3.7 move an event between two processes that are both ALIVE. This cell is the case where one of them is not — and, worse, does not know it yet. Every bug here reads as correct code: expiry reads as safety, a checkpoint reads as progress, a reconciler reads as repair. So the harness carries FOUR controls rather than one, and every one of them has to come back visibly wrong before the answers beside them mean anything. -->

    • S3.9.a Persist job identity/type/version, tenant/actor, immutable inputs, priority/queue, idempotency, budget, deadline, and initial queued state. Evidence: job creation tests. <!-- JobRecord + jobCreationFaults (14 kinds), all 14 produced by a test. THE TWO THAT ACTUALLY BITE: idempotency-key-not-tenant-scoped — a key namespace shared across tenants means tenant B's enqueue can be answered with tenant A's job, silently, and the first symptom is a result in the wrong account; and deadline-inside-one-attempt, where the job is enqueued ALREADY unable to finish, so it burns an attempt, a lease and a slot to discover what was knowable at creation. JOB_STATES has no failed: a failure with attempts left is queued with a later runnableAt and one without them is dead-lettered; a state in between is a shelf where jobs sit while nobody is either retrying them or triaging them. Priority is a declared BAND rather than a free integer because the failure is social — with an open range the first caller in a hurry passes 1000 and the queue has a private VIP lane nobody agreed to. The inputs are a REFERENCE plus a hash (S3.8.b's rule): "immutable" is a promise until something checks it, and S3.9.h's replay is what checks it, months later. -->
    • S3.9.b Claim with exclusive time-bounded lease, worker ID, attempt, and fencing token. Evidence: competing-worker tests. <!-- decideJobClaim (10 refusals) + JobLease + FENCING_SOURCES, and the first two controls of pnpm nx job-fencing. A LEASE IS NOT A LOCK AND WHAT MAKES IT SAFE IS THE FENCING TOKEN. Exclusivity belongs to the store's conditional UPDATE, and the harness shows the difference: read-then-write let BOTH workers believe they had claimed the job; the conditional UPDATE produced one winner and minted the token in the same statement. FENCING_SOURCES has no wall-clock member and that absence is the point — two claims inside one millisecond produce EQUAL tokens and a stepped-back clock produces one the store has already beaten, and in both cases token > highest passes a zombie write through while looking exactly like a correct implementation. attempt-counter qualifies only because the attempt is incremented inside the same atomic claim — and the claim DECLARES its source and is held to it, so a token claiming to come from the counter that is not the incremented attempt is refused rather than trusted. The claim also refuses the three things a worker would otherwise discover by doing the work: a deadline that will pass mid-attempt, a budget that cannot afford the attempt, and a token that does not beat the last one. -->
    • S3.9.c Renew heartbeat and lease on a controlled cadence and prevent expired workers from committing. Evidence: lease-loss tests. <!-- heartbeatFaults (4 kinds), workerHoldsLease, leaseReclaimableAt, fenceWrite -> FencedWrite. THE MARGINS POINT AWAY FROM EACH OTHER: the worker gives the lease up EARLY (expiresAt - skew) and the reconciler takes it away LATE (expiresAt + skew), so the window between them belongs to nobody. A test walks every second across the lease and asserts there is no instant where both believe they hold it. Symmetric margins would leave one, and everything written inside it is a zombie write. one-missed-beat-loses-the-lease is what keeps a healthy worker alive: with under three beats per lease a single dropped packet evicts a worker that is doing fine, and the operator's conclusion is "leases are flaky". fenceWrite is the ONLY producer of FencedWrite, so a commit that did not check the fence cannot be written down. It refuses a superseded token EVEN WHILE THE LEASE LOOKS LIVE, which is the case a lease alone cannot see and CONTROL 2 of the harness demonstrates: with expiry and no fence at the write, the paused worker's result overwrote the live one's; with the token compared at the write, the same UPDATE matched 0 rows. -->
    • S3.9.d Persist stage-specific completed/total progress, message key, output refs, and monotonic timestamp without false completion. Evidence: progress invariant tests. <!-- StageProgress + progressFaults (8 kinds) + jobProgress + completionFaults (3 kinds). PROGRESS IS THE ONE FIELD WHOSE FAILURE MODE READS AS SUCCESS, so three separate rules point the same way. (1) jobProgress returns NO done — a test asserts the exact key set — because completion is a STATE decided by whatever commits the last effect; derive it from the progress table and a worker that dies after its final report has "finished". (2) An unknown total is null and never 0: a total of zero renders as a FULL bar in every progress widget ever written. (3) A denominator with a hole in it is invented, so one indeterminate stage makes the whole fold indeterminate rather than summing what is known and reporting a percentage of it. And the direction that is explicitly NOT a fault: every stage at its total while the job is still running is normal — the work is done and the result is not committed. Flagging it is how a system ends up wiring completion to the bar. observation-went-backwards exists because two reports that cross in flight land in the order they land, and without a monotonic observation the older one overwrites the newer. -->
    • S3.9.e Persist versioned resumable checkpoints transactionally with authoritative stage effects. Evidence: crash/resume tests. <!-- CHECKPOINT_ORDERINGS + planStageCommit + decideResume, and CONTROL 3 of the harness. A CHECKPOINT WRITTEN BEFORE ITS EFFECT LOSES THE EFFECT FOREVER — the same asymmetry as S3.6/S3.7 and the same answer: there is no checkpoint-first member. Effect-first can only REPEAT, which idempotence absorbs; checkpoint-first SKIPS, and nothing anywhere reports a stage that was never run. The harness kills a five-unit stage during unit 3: checkpoint-first resumed from 3 and finished [1,2,4,5] with unit 3 never applied by ANYBODY and the job reporting success; effect-first resumed from 2, redid unit 3, and the key made it 5 effects for 5 units. planStageCommit refuses the two combinations that read as correct: a NON-transactional effect declared part of the store transaction (the code looks right and the external call sits outside the transaction supposedly protecting it), and effect-first on a non-idempotent effect (effect-first exists precisely so the resume may redo it). decideResume's unobvious answer is refuse for an unreadable checkpoint on a non-idempotent stage: restarting looks helpful and redoes effects the checkpoint knows about and this worker cannot see — "how far did it get" is exactly the question the unreadable cursor holds. A checkpoint at a HIGHER fence than your lease means you are the zombie, not the resumer. -->
    • S3.9.f Implement requested/cancelling/cancelled semantics, cooperative checkpoints, noncancellable regions, external cancel, and compensation. Evidence: stage cancellation tests. <!-- requestCancel (4 outcomes), regionFaults (3 kinds), cancellationPoint (3 actions), compensationPlan. CANCELLATION IS A REQUEST AND A JOB IS NOT CANCELLED BECAUSE SOMEBODY ASKED. A QUEUED job has nobody to ask and is cancelled here and now; a RUNNING one enters cancelling and stays there until the worker acknowledges. Collapsing the two tells the operator the effects have stopped at the exact moment they have not, and they act on that. region-longer-than-the-lease turns the safety mechanism INTO the failure it was protecting against: inside a region longer than the lease the worker is guaranteed to become a zombie before it comes out, so the uninterruptible section is precisely the window in which somebody else claims the job. escalate-and-fence is the third answer that stops cancelling being a state jobs live in — a region past its own bound is a cooperative cancel that has ALREADY failed, and waiting longer for it is what leaves a queue full of half-cancelled jobs. compensationPlan undoes in reverse order and reports partiallyApplied: a cancelled job with an uncompensable effect behind it left real state in the world, and reporting cancelled with nothing beside it describes a world that does not exist. -->
    • S3.9.g Classify retryability and apply bounded attempts, exponential backoff/jitter, maximum age, capacity, and spend policy. Evidence: deterministic retry tests. <!-- JOB_FAILURE_KINDS + classifyJobFailure + jobBackoffSeconds + decideJobRetry (3 decisions). Reuses S1.10's Retryability vocabulary rather than inventing a parallel one — these are the same four answers a public error already has to give. cancelled is deliberately NOT a failure kind: the moment it joins the taxonomy some branch of the retry decision picks it up and retries a job somebody asked to stop. THE DEADLINE AND THE BUDGET ARE DIFFERENT REFUSALS. Time may have been the network; money was spent and spending it again is a decision a person makes. The spend check runs BEFORE the attempt against the attempt's own estimate, because one that runs after is an accounting entry rather than a control. unknown-outcome's answer DEPENDS ON THE EFFECT and carries its own LOWER attempt ceiling, since each attempt may be a second effect — a test pins that the third attempt is refused by the unknown-outcome ceiling and not by the ordinary one wearing its coat. defer is a third answer and not a retry: a queue at capacity has not refused the attempt, it has not OFFERED one, so a deferred job burns neither an attempt nor a budget reservation. Folding capacity into retry is how a busy hour quietly consumes every job's allowance. Jitter is a FRACTION of the delay (S3.6's rule): a constant window stops spreading exactly when every worker's backoff has converged on the cap, which is when the spread is the only thing preventing a synchronized retry. -->
    • S3.9.h Dead-letter exhausted/terminal jobs with safe diagnostics, owner, retained inputs, remediation, and authorized replay/abandon. Evidence: dead-letter workflow test. <!-- DeadLetter + deadLetterFaults (9 kinds) + authorizeDeadLetterAction (7 refusals) + planDeadLetterReplay. diagnostic-is-a-serialized-payload is S3.8's problem arriving by a different road: a dead-letter console shows every failed job from every part of the product to whoever is on call, and the easiest way to write a useful diagnostic is to serialize the thing that failed — which puts the inputs, and whatever personal data was in them, on that screen. The inputs are already retained BY REFERENCE for the replay; the diagnostic does not need a second copy. Detected syntactically (a JSON signature and a length bound), not by a redacted: true promise. THE JOB IDENTITY IS PRESERVED — the mirror of S3.8.f, and the argument is the idempotency key. A new job id forces one of two wrong moves: reuse the key and the enqueue is answered with the dead job's outcome, or drop it and two runs of the same work exist with nothing relating them. planDeadLetterReplay CONSTRUCTS the continuation, so a caller cannot mint a new id, and refuses an approval granted for a different action. inputs-moved is the jobs-specific half of "nothing has changed": the inputs are declared immutable, so a reference that now hashes to something else is not a changed job — it is a DIFFERENT job wearing this one's id. purge-would-strand-an-open-record refuses to remove the only remaining action on a record nobody has closed. -->
    • S3.9.i Reconcile orphan leases and external outcomes unknown after worker death. Evidence: fault-injection repair test. <!-- reconcileOrphan (6 repairs) + orphanSweep, and CONTROL 4 of the harness. AN ORPHANED LEASE IS NOT A FAILED JOB. When a worker dies, what happened to its external effect is UNKNOWN, and unknown is not failed: calling it failed and retrying doubles the payment, calling it done drops the work. CONTROL 4 makes the first one visible — reconciling an unknown at-most-once outcome as a failure and requeuing left TWO charges for one job — and the kit's answer left one, with the job parked for a person. With a probe available the same input is ASKED rather than assumed (probe-the-external-system), and a probe that comes back committed marks the job succeeded rather than redoing the work. The reconciler waits the SKEW out (S3.9.c) before touching anything: reclaiming at the lease's own expiry races the holder's last heartbeat. A checkpoint at a fence above the lease means a newer holder already has the job and this lease is the stale record, so the repair is to leave it alone. Every repair that takes the lease away sets fenceOut, because the previous holder may be alive and about to write. -->
  • S3.10 Implement job parent/child graphs and aggregate progress. Evidence: partial failure semantics are deterministic. <!-- src/job-graphs.ts (35th production entry point) + tools/graph-admission.mjs (21st Nx target). 42 tests in the cell, 2,139 in the kit, 38 mutations killed with no survivors. S3.9 is one job and one worker. This is the shape real work has, and the whole cell follows from one thing: A PARENT'S ANSWER IS NOT ANY SINGLE CHILD'S ANSWER. Cycle detection is the S2.1 stage engine's Tarjan REUSED over a lifted edge set rather than a second implementation. -->

    • S3.10.a Define parent/child edge type, ordering, dependency, optional/required, weight/unit, failure policy, and cancellation propagation. Evidence: graph schema. <!-- JobEdge + FAILURE_POLICIES + CANCELLATION_PROPAGATIONS. TWO ABSENT MEMBERS. FAILURE_POLICIES has no ignore: a policy that makes a failure invisible is the one thing a graph must not offer, because the child that is safe to ignore today is the one whose silent failure is diagnosed six weeks later as "the thumbnails stopped working at some point". CANCELLATION_PROPAGATIONS has no orphan-children: stopping the parent and leaving the children running is how a cancelled job goes on spending with nobody reading the results, and let-children-finish is the same intent stated honestly. REQUIRED/OPTIONAL IS THE FAILURE POLICY rather than a second field beside it — two fields that must agree are two places to get it wrong, and this is the one that decides anything. isRequired derives it. Ordering is a declared ordinal, so it does not depend on insertion. -->
    • S3.10.b Validate known jobs, same tenant/workbench, acyclic graph, maximum depth/fan-out, and immutable edges after execution begins. Evidence: graph constraint tests. <!-- graphFaults (14 kinds, all 14 produced by a test) + asEdgeSet + graphDepth + graphFanOut. THE REUSE: cyclesIn/unreachableFrom from src/stages.ts (S2.1) over a lifted edge set. It carries BOTH edge kinds — parent-to-child AND dependency-to-dependent — because a cycle among SIBLINGS is the one a parent/child check alone would miss and the easy one to write. initial is the root, which cyclesIn ignores and unreachableFrom uses, so a node no parent edge reaches is reported. edges-changed-after-execution-began is the immutability rule: once a node has left queued, a new dependency describes an execution that did not happen, because the child it constrains may have already run without waiting for it. zero-weight-child is the quiet one — a child with no weight contributes nothing to the bar and can still fail the parent, so the bar reads 100% on a graph that cannot finish. mixed-weight-units refuses adding seconds to megabytes. -->
    • S3.10.c Schedule ready children only after required dependencies and capacity/budget gates. Evidence: dependency scheduling tests. <!-- readyChildren -> {ready, blocked, deferred, unreachable}, and pnpm nx graph-admission. THE BUDGET IS CHECKED OVER THE ADMITTED SET, NOT PER CHILD: five children each individually affordable are not five children the budget can afford, and a per-child check passes all five and finds out afterwards — S3.9's "an accounting entry rather than a control" one level up. The harness shows it costs 600 against a ceiling of 250 WITH NO CONCURRENCY AT ALL. unreachable is separated from blocked because they need opposite actions: a blocked child is fine and waiting, and one whose dependency dead-lettered is a graph that cannot complete and will otherwise sit in waiting forever with nothing left to wait for. A child whose cost nobody estimated is deferred rather than started. -->
    • S3.10.d Aggregate heterogeneous progress through explicit weights/units and preserve unknown/indeterminate work honestly. Evidence: known-value aggregate tests. <!-- aggregateProgress -> {totalWeight, knownWeight, unknownWeight, lowerBound, upperBound, fraction}. AN UNKNOWN CHILD IS A RANGE, NOT A ZERO — and this is where the parent knows MORE than S3.9.d did about a single job. S3.9 could only answer null for an indeterminate stage; a parent knows the unknown child's WEIGHT, so it can state exactly how much of the answer is missing. Known-value test: weights 3600/60/300 at 0.5/null/1 gives [0.530303, 0.545455] and a fraction of null, because a single number here would be a guess dressed as a measurement. A child with NO report is unknown rather than zero: zero is a claim about the work, absent is a claim about the report. And a parent with no weight reports null rather than a full bar. -->
    • S3.10.e Define parent queued/running/waiting/partial/failed/cancelled/ succeeded outcome for every required/optional child combination. Evidence: exhaustive state truth table. <!-- parentOutcome over PARENT_OUTCOMES, with a test that enumerates ALL 144 combinations of two children across 6 job states x 2 failure policies and asserts the properties on every one, including both directions of "succeeded means exactly one thing" and "the loss list exists iff there is a loss". partial EXISTS SO THE FOLD NEVER HAS TO CHOOSE BETWEEN TWO LIES: failed throws away required work that did finish, succeeded records the loss nowhere. ORDER: failed is decided before cancelled, because a dead-lettered required child needs somebody and a cancellation is a decision already made — reporting the cancellation first buries the fault under an intention. waiting vs running distinguishes "in flight with nothing executing" from "something executing", and its detail separates a STARVED QUEUE (children could start, none is) from a dependency. failed exists here and deliberately NOT in JOB_STATES (S3.9.a): a leaf's failure is an attempt to retry or triage, and a parent's is a FACT about its children. -->
    • S3.10.f Propagate cancellation and retry without duplicating completed children or invalidating reusable output incorrectly. Evidence: cancellation/retry graph tests. <!-- propagateCancellation + planGraphRetry. BOTH DEFAULTS ARE WRONG, SO NEITHER IS THE DEFAULT: re-running everything duplicates work that succeeded, and reusing everything that succeeded reuses output produced by the code that has just been fixed — which is WORSE, because it looks like a successful retry. The input hash and the handler version decide, per child. AND REUSE IS TRANSITIVE. A child downstream of one being re-run is re-run too, however unchanged its own inputs look, because its inputs were the upstream child's output; reusing it produces a result that is half old and half new with nothing saying so. Computed as a closure over dependsOn, and a test pins that one changed handler version re-runs the whole chain behind it. Cancellation stops the NEWEST work first and never touches a finished child — "cancel the graph" is a statement about work that has not happened, and applying it to work that HAS gives a completed child an outcome it never had. -->
    • S3.10.g Expose traversable graph, critical path, child results, blockers, costs, and safe diagnostics. Evidence: graph query/UI tests. <!-- graphView + rootBlockers + criticalPath. ROOT BLOCKERS, NOT IMMEDIATE ONES. A parent waiting on a child that is itself waiting on two others has three names it could report, and only the ones not waiting on anything are worth an operator's attention — the rest are consequences, and reporting them is how an incident spends twenty minutes walking a dependency chain by hand. THE CRITICAL PATH IS OVER DECLARED DEPENDENCIES AND EXPECTED DURATIONS, never over what happened to be slow: a child that took three hours because the queue was full is not on the critical path, and putting it there sends people to optimise a scheduling problem as though it were a work problem. Enforced by construction — criticalPath takes only the graph, so an actual elapsed time cannot reach it. The view carries a failureCode and has NO message field: a graph view shows every child of every tenant's job to whoever is on call, which makes it the same screen as S3.8's quarantine console and S3.9's dead letter, and a child's failure MESSAGE is exactly where a serialized input rides in. A test pins the node's key set so a message cannot be added without somebody deciding to. -->
  • S3.11 Implement blob references with size, media type, checksum, encryption, retention, and provenance metadata. Evidence: stored-byte rehash integration test. <!-- src/blobs.ts (36th production entry point) + tools/blob-storage.mjs (22nd Nx target). 61 tests in the cell, 2,200 in the kit, 51 mutations killed with no survivors, plus 7 mutations of the BUILT file that the storage harness had to catch. Every other module in this kit describes something this system decided. This one describes bytes SOMEBODY ELSE produced, in a store SOMEBODY ELSE runs, and almost every rule follows from not believing either of them. -->

    • S3.11.a Define blob reference ID, tenant/scope, immutable storage locator/version, state, created actor/time, and owning references. Evidence: blob schema. <!-- BlobReference + blobFaults (9 kinds, all 9 produced) + locatorKey. THE LOCATOR IS IMMUTABLE AND THE VERSION IS PART OF IT rather than a field beside it: overwriting the bytes under a locator silently re-points every reference that pinned it, which turns a content-addressed reference (S1.4.c) back into a floating one without changing its type. version-not-recorded in S3.11.g is the same rule defeated by an omission instead of an overwrite. The three active-without-… faults are separate rather than one, because they fail for different reasons — the scanner was down, the read-back was skipped, the owning row was rolled back — and each describes a blob a reader would treat as real. -->
    • S3.11.b Record server-observed byte size and sniffed media type separately from untrusted supplied metadata. Evidence: spoofing tests. <!-- DeclaredMetadata + ObservedMetadata + sniffMediaType + spoofFaults (6 kinds). THE DECLARED AND THE OBSERVED NEVER MERGE: there is deliberately no single mediaType or byteSize on the record, and a test pins the key sets, because the moment there is one field the uploader's claim quietly becomes the server's observation and an executable is served as an image with everything in the log looking correct. THE SNIFFER RETURNS EVERY MATCH, NOT THE FIRST. A real magic-number table (PNG/JPEG/GIF/WebP at offset 8/MP4 ftyp at offset 4/PDF/ZIP/GZIP/OGG plus ELF, Mach-O, PE and a shebang), and the zip END-OF-CENTRAL-DIRECTORY is scanned in the TAIL — which is what makes a GIF/zip polyglot detectable at all. Two decoders will disagree about that file and both will be right, so returning one answer hands each of them a different file while reporting one. Known-byte tests, not shape tests. A disagreement is not an error: none of these faults means the bytes are bad, they mean the DESCRIPTION was wrong — and a system that reconciles them by overwriting the description has destroyed the evidence that anything was ever wrong. -->
    • S3.11.c Compute approved hash from stored bytes and persist algorithm, digest, and verification time. Evidence: known-byte and tamper tests. <!-- APPROVED_DIGEST_ALGORITHMS + DIGEST_VERIFICATION_SOURCES + verifyDigest (5 outcomes) + verificationIsStale, and pnpm nx blob-storage. THE DIGEST IS OF THE STORED BYTES READ BACK. Hashing the inbound stream proves what you were SENT; hashing what the store returns proves what you HAVE, and the two differ in exactly the cases that matter. TWO NAMED WAYS TO PRODUCE A GREEN TICK OVER NOTHING, both of which the harness demonstrates against a real object store: metadata compares the stored digest to ITSELF and passed while the bytes underneath had been replaced; a prefix read passed because the corruption was in the last 16 of 2,248 bytes, which is where a truncated write leaves it. Both are outcomes with names rather than silent passes. md5 and sha1 are absent and stay absent: S3 hands back an MD5 as the ETag, an ETag is a cache token, and the way that mistake enters a codebase is a field that will hold either. -->
    • S3.11.d Define encryption mode/key reference/version, region, transport, rotation, and decrypt authorization without storing keys. Evidence: encryption metadata tests. <!-- EncryptionMetadata + encryptionFaults (7 kinds) + authorizeDecrypt (4 refusals). KEY_MATERIAL_LOCATIONS HAS NO inline: a record that CAN hold a key eventually does — at the moment somebody is debugging why decryption fails, in a row that is then backed up, replicated and shipped to a support console. The absence is the control; a policy saying "do not put keys here" is not. no-key-version is the one that costs a weekend: an unversioned key reference resolves to whatever is current, so the FIRST ROTATION makes every object encrypted before it undecryptable — and the failure appears only when somebody reads an old blob, which is never near the rotation in a log. key-region-differs-from-blob-region is not a storage detail: a key that crosses a region crosses a jurisdiction. authorizeDecrypt returns a NAME and a VERSION and a test pins the key set. Whoever holds the decision still has to go to the key service and be told no again; a function that returned material would be making that service's decision on its behalf. -->
    • S3.11.e Define quarantine/active/rejected/retained/held/purge states, retention source/expiry, legal hold, and deletion semantics. Evidence: blob lifecycle tests. <!-- BLOB_STATES + admitBlob + deletionStateFor + purgeRefusals. THIS MODULE DOES NOT OWN A SECOND RETENTION MACHINE. Retention sources, the strictest-floor/strictest-ceiling resolution, legal holds, purge steps, tombstones and residue are all S3.3's (src/deletion.ts), and deletionStateFor is the ENTIRE interface between the two — a mapping, not a reimplementation. Two retention implementations is two answers to "may this be deleted", and the one that gets asked is whichever module the caller happened to import. A test walks every blob state and asserts the mapping lands in a state DELETION_TRANSITIONS actually has. A BLOB IS ADMITTED TO NOTHING BY DEFAULT: unavailable is not passed, so a check that could not answer holds the blob in QUARANTINE while one that answered no REJECTS it. The difference between "nothing objected" and "everything approved" is invisible while the scanner is healthy and is the entire outcome when it is not. purgeRefusals carries only what deletion.ts cannot know: that these bytes are pointed at by three other rows, and that an authorized upload may still be in flight toward them — purging then races the bytes, and the upload wins by recreating an object nothing points at. MUTATION-PASS CATCH: the executable and polyglot admission guards were both MASKED by the permitted-types check (an ELF is not on the permitted list, so the earlier rule refused first and the guards were untested). Fixed with three isolating cases — a permitted-but-executable type, a permitted-but-polyglot pair, and an ordinary unpermitted type that is neither. All three guards then died under mutation. -->
    • S3.11.f Link provenance manifest/ledger, ingredients, generating run, transformations, source revision, and derived artifacts. Evidence: provenance linkage tests. <!-- ProvenanceManifest + provenanceFaults (7 kinds) + stalenessOf (4 verdicts). The ingredients are references AND digests: a reference alone names something that may have moved, and the digest is what makes the ancestry checkable later — which is the only time anybody looks. transformation-without-a-tool-version decides whether the manifest is worth keeping at all. "Resized with imagemagick" cannot be reproduced and cannot be audited; a provenance record that cannot answer "what exactly did this" is a story about the file rather than a record of it. A DERIVED ARTIFACT WHOSE SOURCE MOVED IS NOT "THE SAME ARTIFACT, UPDATED". The bytes are fine and the DESCRIPTION has stopped being true, and the manifest is the only thing that can say so — everything downstream will happily keep serving it. And an ingredient that cannot be resolved at all is unknown, never current: not the same as unchanged. -->
    • S3.11.g Re-read and rehash the exact stored bytes across local/object storage adapters and detect corruption. Evidence: real-storage integration tests. <!-- BlobStore + adapterFaults (4 kinds) and pnpm nx blob-storage, against a LOCAL FILESYSTEM adapter and a REAL S3-compatible object store (MinIO from the dev compose) side by side. BlobStore.read returns the WHOLE object and takes no range, because a partial read is the second way a verification passes without checking anything (S3.11.c) and an interface with an optional range invites it. THE THIRD CONTROL IS AN ADAPTER THAT TRANSFORMS: a line-ending normaliser, which is a real thing that has happened to real files, and which returned 2,207 bytes where the honest adapters returned 2,248. That class of bug is invisible until two copies of the same object are hashed and compared, and by then the original is usually gone. The sniffer is round-tripped too: an ELF stored under a .png key and a GIF/zip polyglot both come back out of the object store still reading as what they are. The polyglot is the harder one — its zip record is in the TAIL, which is exactly what a store that pads or truncates destroys first — and its answer is only worth anything if it survives the storage layer. CALIBRATION: seven mutations of the BUILT blobs.js were run against this harness. The first pass killed four; the executable and polyglot sniffs survived because the harness only checked the leading type, and the round-trip cases above were added until they died. -->
  • S3.12 Implement orphan-blob discovery and safe reclamation. Evidence: referenced, legal-held, and pending-upload blobs are never removed. <!-- src/reclamation.ts (37th production entry point) + tools/reclamation-race.mjs (23rd Nx target). 50 tests in the cell, 2,251 in the kit, 48 mutations killed with no survivors. EVERY OTHER MODULE IN THIS KIT IS CAREFUL ABOUT BEING WRONG. THIS ONE IS CAREFUL ABOUT BEING WRONG IN ONE DIRECTION: a blob wrongly kept costs storage, a blob wrongly deleted is gone — no compensating action, no retry, and usually no way to find out until somebody opens the thing it was attached to months later. -->

    • S3.12.a Enumerate authoritative references from entities, revisions, jobs, outbox/inbox, evidence, provenance, bundles, uploads, releases, audit, and legal holds. Evidence: reference-source register. <!-- REFERENCE_SOURCES (12) + ReferenceRegister + registerFaults (5 kinds) + referencedBlobIds. "UNREFERENCED" IS A CLAIM ABOUT WHAT YOU LOOKED AT and never a property of the blob, so the register is a CLOSED list and a sweep that could not reach every source does not run over what it could. The failure is not exotic: a team ships a feature that references blobs, nobody adds it here, and six weeks later the sweep deletes their bytes. source-missing-from-the-register is checked against the DECLARED vocabulary rather than against what the caller supplied — a register that validates itself against its own contents cannot detect the thing it exists to detect. A PARTIAL scan is a separate field and the same refusal as an unreachable one. The union across sources, never an intersection. THE HARNESS MAKES THIS CONCRETE: a blob referenced only from provenance, with that source missing from the register, came out classified unreferenced-active with action delete-bytes. -->
    • S3.12.b Detect metadata-without-bytes, bytes-without-metadata, unreferenced active, abandoned upload/quarantine, failed derivative, and dangling-reference classes. Evidence: orphan classification tests. <!-- ORPHAN_CLASSES (8) + ORPHAN_ACTIONS (4) + classifyOrphan. THERE IS NO GENERIC delete ACTION, because the classes fail in OPPOSITE directions: a row pointing at bytes that are gone is repaired by deleting the ROW and nothing else, and an object nothing knows about is the only class where deleting BYTES is the repair. One verb over both is how a reclaimer deletes bytes a live row still points at. ORDER: dangling-reference is checked FIRST, because it is the one class where the fault is in the pointer rather than in the blob, and treating it as an orphan deletes the wrong end of a broken link. abandoned-quarantine is leave: a blob held pending a check that never answered is exactly what an incident would want to look at, so it needs a person rather than a sweep — and SWEEPABLE_BLOB_STATES says so. -->
    • S3.12.c Apply age/grace, upload/lease state, retention, hold, quarantine/incident, release, and provider reconciliation gates. Evidence: eligibility policy tests. <!-- ELIGIBILITY_GATES (7) + eligibility + isEligible. EVERY GATE RUNS AND ALL SEVEN RESULTS ARE RETURNED — no short-circuit on the first refusal. An operator needs to know a blob is held for three reasons, because releasing one of them does not make it deletable and a sweep that reports only the first teaches otherwise. retention-elapsed reads the CLOCK AND the S3.3 machine, because that machine is the authority (S3.11's decision 6) and a reclaimer consulting only the timestamp would be a second answer to a question deletion.ts owns. provider-reconciled is the gate that stops "no row for this object" meaning "the listing is stale". isEligible requires the result list to be COMPLETE, so a caller that ran five gates cannot pass. MUTATION-PASS CATCH: the retention CLOCK was masked by the deletion-state check beside it — the fixture's window had already elapsed, so flipping the clock changed nothing. An isolating case (state clean, clock not yet elapsed) was added and it died. -->
    • S3.12.d Produce a dry-run candidate manifest with size/cost, reason, reference search, confidence, and sampled verification. Evidence: operator preview. <!-- ReclamationCandidate + confidenceOf + dryRunManifest. CONFIDENCE IS A GATE, NOT A LABEL: markCandidate refuses anything below high without a sampled verification, and confidence is low whenever the register was incomplete — which is exactly the case where the verdict is about what was searched rather than about the blob. notExamined is REPORTED rather than implied. A sweep that lists two hundred candidates out of a bucket it stopped walking at ten thousand objects has described a sample, and a preview that does not say so reads as a complete account of what is reclaimable. Every candidate carries the sources searched AND the register version it was judged at. -->
    • S3.12.e Mark candidates with a reclamation token and recheck references atomically before deletion. Evidence: new-reference race tests. <!-- ReclamationToken (branded; only markCandidate produces one and planDeletion requires one) + MARK_REFUSALS (4) + DELETION_REFUSALS (5), and pnpm nx reclamation-race. THE RECHECK HAS TO BE IN THE SAME STATEMENT AS THE DELETE, so planDeletion RETURNS A PREDICATE RATHER THAN A PERMISSION: everything it checked can change between returning and the statement landing, and a caller that evaluates the conditions and then deletes unconditionally has reproduced the race the token exists to close. THE HARNESS SHOWS IT IS NOT THEORETICAL. Three controls, each deleting a referenced blob: mark-then-delete with no recheck; RECHECK-THEN-DELETE, the conscientious version, where the reference lands in the window the separate SELECT opened; and the incomplete register from S3.12.a. Beside them, one statement carrying reclamation_token = $2 AND NOT EXISTS(...) matched 0 rows and the blob survived — and the same statement DOES delete a genuinely unreferenced one, so the guard is not simply always refusing. register-version-moved is the subtle refusal: a source added between the dry run and the mark is a place this blob has never been searched for. -->
    • S3.12.f Delete idempotently in bounded batches, preserve audit/tombstone where permitted, and report partial/unknown outcomes. Evidence: deletion fault tests. <!-- DELETION_OUTCOMES (5) + foldBatch + batchFaults (3 kinds). already-absent IS A SUCCESS and unknown IS NOT A FAILURE, and both are load-bearing. A reclaimer runs twice — after a crash, after an operator re-runs it — so an object already gone must be an ordinary outcome rather than something somebody investigates. And an unknown outcome RECORDED AS DELETED leaks the bytes forever with the row saying otherwise, which is the one state nothing later can detect; unknown-recorded-as-deleted is the fault that catches it. notAttempted is stated rather than implied, and an unbounded batch is refused: a reclamation job that dies halfway has deleted an arbitrary prefix of an unknown set. tombstone-dropped-where-required is the audit half — a reclamation that leaves no record is indistinguishable from data loss six months later. -->
    • S3.12.g Reconcile deletion and restore/relink false positives before permanently closing findings. Evidence: reclamation report. <!-- FALSE_POSITIVE_SIGNALS (4) + FINDING_STATES + reconcileSweep. A FINDING IS NOT CLOSED BY DELETING. A restore request, a reference created for a blob that is gone, a broken render, a derivative regenerated from scratch — these are the ONLY evidence that ever arrives that a reclaimer was wrong, and they arrive AFTER. Closing on delete is what makes a reclamation bug permanent: nothing is watching by the time the symptom shows up, and the sweep that caused it has been green for weeks. The observation window is a PARAMETER and not a constant, because it is a property of how quickly this product's users would notice and nobody in this module knows that. A source added to the register since the sweep keeps the finding open too — there is a place it never looked, so the absence of a signal does not cover it. And prematurelyClosed makes decision 6 checkable after the fact: a finding closed before its window elapsed is one whose false positives were never going to be found. -->
  • S3.13 Implement encrypted sensitive-field storage or vault references. Evidence: database dumps and logs do not expose protected values. <!-- src/protection.ts (38th production entry point) + tools/canary-leak.mjs (24th Nx target). 178 tests in the cell, 2,430 in the kit. The mutation pass ran 246 mutations over 123 guards: 236 killed, TEN SURVIVORS, all closed — nine by tests that were missing and one by deleting a guard that turned out to change nothing. Re-running those sites killed all twelve occurrences. TWO OF THE TEN WERE DESIGN FINDINGS RATHER THAN TEST GAPS. revocationPlan survived BOTH directions at one branch, which is the mutation pass saying the two crypto-shred arms were the same answer in different words — one column and a whole tenant, distinguished only by prose an operator does not read. blastRadius is now a FIELD. And the cycle-detection walk carried a continue that the while condition already expressed; two spellings of one condition are a place they can later disagree, so it is gone. A THIRD LESSON IS ABOUT THE HARNESS: removing the cycle break turns the walk into an infinite synchronous loop, which no vitest timeout can fire against — the event loop never yields. A mutation harness with no per-run timeout does not report that as a survivor, it hangs, and the first attempt burned 25 CPU-minutes on one mutant before anybody looked. S3.11 DESCRIBES AN ENCRYPTION ARRANGEMENT SOMEBODY ELSE PERFORMS, on bytes in a store somebody else runs. This is the other half and it is not the same problem: the plaintext is in this process, the sealing happens before the INSERT, and the thing being protected is a COLUMN — queried, indexed, logged, echoed back in constraint violations, exported to support, and copied into every backup. The blob half can describe; this half has to bind. -->

    • S3.13.a Inventory sensitive fields and classify encryption-at-rest, tokenization, one-way hash, vault reference, or prohibited persistence. Evidence: field protection matrix. <!-- SENSITIVITY_CLASSES (9) + FIELD_OPERATIONS (9) + PROTECTION_STRATEGIES (5) + STRATEGY_CAPABILITIES + matrixFaults (12 kinds) + strategiesFor. A CLASSIFICATION IS A CONSTRAINT ON WHAT THE FIELD MAY BE USED FOR, not a label on how secret it feels. Everybody agrees a national identifier is sensitive; the argument is always about whether the support console may search by it. So the matrix is DERIVED from the declared operations, and STRATEGY_CAPABILITIES is the whole design: one-way-hash has no equality-match (a hash that answers "are these the same" is salted per-field, which is a token with a different name and a different leak) and encrypt-at-rest has none either (a randomized AEAD is what stops the comparison, and removing the nonce to restore it is becoming a deterministic token without anybody deciding to accept the leak). THREE OPERATIONS ARE SUPPORTED BY NOTHING. Ordering and substring search over a protected field are not hard problems somebody should go and solve: order-preserving encryption leaks the order, which for a birth date or a balance is most of the value, and an encrypted prefix index leaks the prefix. Both are sold as encryption. The OPERATION is refused. THE FIELD IS NOT THE ONLY COPY — an index, a search vector, a "last four", an export column, a metric label, a cache entry — so derivatives are declared and checked: an attacker reads the weakest copy, so a protected column with a plaintext derivative is a plaintext column with a ceremony in front of it. Checked against a COLUMN CENSUS rather than against itself, for S3.12.a's reason: an unclassified sensitive column does not appear as an unprotected entry, it appears as nothing at all, and stores plaintext by default. token-over-a-low-cardinality-field takes its floor as a PARAMETER, and the harness inverted a real one: three countries behind three tokens, one GROUP BY, and the 31/12/4 ranking hands back the mapping. -->
    • S3.13.b Define key hierarchy, KMS/vault ownership, tenant/context binding, algorithm/version, rotation, revocation, and break-glass access. Evidence: cryptographic design review. <!-- KEY_TIERS (3) + keyDesignFaults (14 kinds) + revocationPlan + authorizeShred (4 refusals) + authorizeBreakGlass (8 refusals). ROTATION AND REVOCATION ARE NOT DEGREES OF THE SAME ACTION and the function exists so a runbook cannot use the words interchangeably. Rotating a KEK re-wraps what is under it and touches no rows; destroying one so the material can never be used again destroys everything beneath it, and at the DEK tier that IS the data. SO CRYPTO-SHREDDING IS A DELETION WEARING KEY-MANAGEMENT CLOTHES, and authorizeShred asks S3.3's legal holds the question a purge is asked — a hold that stops a purge and not a key destruction is a hold with a hole in it shaped exactly like the operation nobody classified as deletion. Break-glass borrows validateHumanGate rather than building a second quorum machine (S2.13 already argued that a quorum without independence is one signature with a count beside it); what it ADDS is the three rules specific to keys — THE AUDIT RECORD IS A PRECONDITION, because an audit written afterwards is prevented by whatever made the emergency access necessary; the emergency identity is not the routine one, or no log distinguishes a break-glass read from ordinary traffic; and a grant over every key is not an emergency, it is a second production credential with a form in front of it. wrapping-key-is-exportable is checked against the GRAPH, not the tier: the property that matters is having children. A data key is EXPECTED to be exportable — the application holds it to run the AEAD, and that is the envelope. rotation-slower-than-the-data-it-protects-is-retained reads like bookkeeping and is not: a key rotated every 400 days over rows kept for 90 means every row that ever existed lived and died under one key, so the window the rotation bounds is the entire dataset. -->
    • S3.13.c Encrypt before persistence with authenticated associated context and prevent plaintext indexing/querying unless explicitly tokenized. Evidence: tamper and query tests. <!-- associatedData + contextFaults + SealedField + sealField (7 refusals) + openField (8 failures) + queryFaults (6 kinds), and pnpm nx canary-leak. THE BINDING GOES IN THE AAD, NOT IN THE WHERE CLAUSE. openField builds its associated data from expected — where the row actually IS — and NEVER from sealed.context. A ciphertext copied into another row arrives with its own envelope, and opening against the envelope's own claim authenticates the copy and hands back the value. The stored context is a DIAGNOSTIC that is never allowed to become the control, and the harness performs the wrong version beside the right one so the check is not vacuous: with the AAD from the envelope, the row copied into tenant-b opened cleanly. The AAD is LENGTH-PREFIXED because the collision is real, not hypothetical: with a plain separator, tenant a|b + entity c and tenant a + entity b|c are the same string, so a ciphertext written for one opens in the other. The part ORDER is fixed in code rather than read from Object.keys — key order is an implementation detail a spread or a JSON round trip may change, and changing it makes every existing ciphertext fail to open with no diagnosis at all. value-is-already-sealed is STRUCTURAL and saves the migration: re-sealing an envelope produces ciphertext that opens to a JSON document, and nothing looks wrong until somebody reads the field. unique-index-over-a-randomized-ciphertext is the query fault that reads as a working constraint — and the harness demonstrated it, one value sealed twice inserting twice. THE ENGINE THEN VOLUNTEERED THE OTHER HALF: the same index REFUSED a verbatim byte copy. It constrains the nonce, not the value, so it is useless against the case it was added for and effective only against a duplication nobody performs by accident. -->
    • S3.13.d Resolve vault references through least-privilege identity, purpose, short-lived access, caching prohibition, and audit. Evidence: vault integration tests. <!-- referenceFaults (4 kinds) + authorizeResolution (8 refusals) + SecretLease + usageFaults (4 kinds). A RESOLVED SECRET IS A SECRET THAT NOW LIVES SOMEWHERE ELSE, and this runtime cannot wipe a string. So the module does not pretend to erase one: what it offers is a BOUNDED LEASE plus a usage record that makes "it was not cached" a checkable claim rather than a sentence in a design document. Saying which of the two you have is the whole point, and usageFaults is the other half of the sentence — value-was-persisted NAMES the destinations rather than counting them, because a cache is one problem and a log is a worse one. audit-sink-unavailable is fail-closed and it is the right way round: an unauditable secret read is exactly the read somebody will later need the record of, and degrading to "resolve anyway, log later" trades the only evidence a protected value was ever produced for the availability of whatever needed it. PURPOSE is the only part of the request that narrows a LEGITIMATE caller — least privilege on the identity alone permits every read that identity was ever needed for. reference-path-carries-a-value defeats the arrangement quietly: a path built from what it points at — secret/users/ada@example.com/ssn — puts the identifier in every row, log line and trace attribute. The value is in the vault and the answer to "who has one" is in the column. -->
    • S3.13.e Redact ORM errors, SQL parameters, logs, traces, metrics, audit summaries, test snapshots, support exports, and backups. Evidence: canary-secret scans. <!-- REDACTION_SURFACES (9) + redactionFaults (7 kinds) + unscannedSurfaces + REDACTION_TOKEN, calibrated by pnpm nx canary-leak. REDACTION IS DRIVEN BY THE MATRIX, NOT BY WHAT THE VALUE LOOKS LIKE. A diagnosis, a home address and a source's name have no signature, and a redactor that greps for things that look like secrets passes all three. So the field matrix is the MECHANISM and S1.10's scanText is the CALIBRATION — borrowed rather than reimplemented, because a second set of detectors is one that is always behind the first. orm-error is listed first because it is the leak nobody writes down, and the harness proves it against the real engine rather than against a message written here to be caught: Postgres answered a unique violation with Key (contact_plain)=(canary-…@leak.invalid) already exists. and redactionFaults came back with all three of canary-survived, protected-field-name-carries-a-value and the S1.10 shape hit. The regex was then tightened so the CONSTRAINT NAME in the same string — account_tax_id_key — does not match on the column. REDACTION_TOKEN is a CONSTANT and there is deliberately no redact(value) helper: anything taking the secret as an argument to decide what to print is a redaction whose output is a function of the secret, and a same-length mask publishes the length, which for a PIN or a postcode is most of the value space. presence-disclosed-by-the-key-alone is the Veritas-grade one and reaches S2.7's shape from a different direction — for a protected source or a health field the key alone answers "is there a value", and that answer IS the disclosure, so the surface emits nothing rather than a redacted something. It is checked against UNAUTHORIZED_READER_SURFACES rather than all nine: a backup IS the database (presence is in it by construction, and the control there is that the artifact is encrypted and held away from the keys), and a parameter list belongs to a query that was already authorized. -->
    • S3.13.f Migrate plaintext values safely with resumable backfill, verification, compatibility, rollback/compensation, and plaintext removal. Evidence: migration rehearsal. <!-- MIGRATION_PHASES (6) + backfillFaults (6 kinds) + PLAINTEXT_RESIDUES (10) + residualPlaintext + advanceMigration (5 refusals). read-switch IS ITS OWN PHASE, and collapsing it into backfill is the mistake: reading from the sealed column while the plaintext is still there is the ONLY configuration in which the new path runs on production traffic AND the old value is still available to fall back to. That window is where every problem with the sealing shows up. DROPPING THE COLUMN REMOVES IT FROM QUERIES, NOT FROM DISK. Postgres's DROP COLUMN does not rewrite the heap, so the values sit in the existing tuples; and the WAL, a physical replica, an inactive replication slot pinning the log before the migration, the index leaves, the backups, the search index, the warehouse, the caches and whatever support was sent all have their own copy on their own schedule. Ten residues, ten INDEPENDENT facts — the spec flips each evidence flag one at a time and requires exactly one residue back, because a shared flag would let clearing the WAL silently clear the warehouse. backfill-without-dual-write is the silent one: rows written after the cursor passes are written in plaintext and never seen again, the migration completes, and the verification counts agree with the rows the backfill touched. cursor-over-a-mutable-column is the same problem from the other side. Rolling BACKWARD is allowed everywhere except from contract, which is the phase that deleted the thing it would roll back to — after it, the recovery is a restore from a backup that still contains exactly what the migration existed to remove. -->
    • S3.13.g Test backup/restore, rotation, unavailable KMS/vault, corrupted ciphertext, deleted key, and unauthorized access. Evidence: failure matrix. <!-- FAILURE_MODES (7) + FAILURE_EXPECTATIONS + failureModeFor + behaviourFaults (6 kinds) + restoreFaults (3 kinds). THE MODES MUST BE DISTINGUISHABLE, and retryable is the load-bearing column. A caller holding one "decrypt failed" retries everything, which is correct for an unavailable KMS and catastrophic for a destroyed key: the permanent loss is recorded as a series of transient errors, the alert is a rate rather than an event, and the row is discovered unreadable months later by a person rather than by a system. Exactly two modes are retryable and exactly one is undisclosable, and the spec asserts both sets rather than sampling them. modes-share-one-code is the finding the section exists for, and the pair that collapses first is always kms-unavailable and key-destroyed, because both arrive as an exception from the same client library. THIS IS WHY THE ADAPTER CONTRACT MAKES THE KEY SERVICE'S ANSWER SEPARATE FROM THE TAG CHECK — an AEAD alone returns one error for a wrong key, a wrong context and a corrupted byte, and an adapter that collapses the key service's answer into a boolean has thrown away the material the whole matrix is made of. It is also why the key id, version and algorithm sit in the CLEAR on the envelope: a deliberate metadata disclosure, bought for exactly this. OpenResult has no string | null success. That signature is how a failure becomes an empty field: the caller writes the null back, the row is now genuinely blank, and the record of what was there is a backup nobody has restored — which is what failed-open catches. restore-predates-a-crypto-shred turns a recovery into an incident: a key destroyed to satisfy an erasure request does not stop the rows existing in the backup. key-state-not-restored is the quiet one — a database that passes every check with unreadable protected columns, which survives a restore drill because a drill counts rows. -->
  • S3.14 Implement database health and schema-compatibility probes. Evidence: incompatible schema returns fail-closed readiness. <!-- src/probes.ts (39th production entry point) + tools/probe-controls.mjs (25th Nx target). 97 tests in the cell, 2,528 in the kit. The mutation pass ran 106 mutations over 53 guards and left TWO survivors, both closed; the re-run is 104 killed with none surviving. ONE OF THE TWO WAS A DESIGN FINDING. SchemaRequirement.exact had its own comparison branch in compareSchema, and the branch was FULLY MASKED by the range — for any pin worth having exact === minimum === maximum, so the range already produced the same verdict and the branch could not change an answer. It is gone. exact is now a DECLARATION that the one-version range was deliberate, which is the only job it was ever doing: it lets no-room-for-a-migration-deployed-ahead-of-this-build tell an intentional pin from an accidentally narrow range, and a pin that disagrees with its range is refused. A PROBE IS THE SMALLEST PIECE OF CODE IN A DEPLOYMENT AND IT DECIDES WHETHER EVERY OTHER PIECE RECEIVES TRAFFIC. That asymmetry is the whole subject: almost every failure here is a probe that was RIGHT about something unimportant and, by being right, took a working service out of rotation or kept a broken one in. -->

    • S3.14.a Separate liveness from readiness and define safe database connection/timeout/pool checks. Evidence: probe contract. <!-- DATABASE_PROBE_KINDS (3) + PROBE_CHECKS (10) + LIVENESS_PERMITTED_CHECKS + contractFaults (9 kinds). LIVENESS MUST NOT TOUCH THE DATABASE, and the ABSENCE of every database check from the permitted set is the control. A liveness probe that queries the database restarts every process the moment the database has a hiccup — turning a thirty-second blip into an outage, and making recovery SLOWER, because each new process opens fresh connections into a database that is already struggling. A probe sharing the request pool is the first thing to fail under load, which removes a healthy-but-BUSY process from rotation and sends its traffic to the ones that are next. And probe-timeout-leaves-no-headroom is the quiet one: a probe killed before it answers reports a TIMEOUT, which is the only probe result that carries no reason at all — the failure mode that produces the longest incidents. Headroom is a PARAMETER, because how variable the hop is belongs to the deployment. WRITING THIS TEST FOUND A REAL FAULT-MASKING BUG: a liveness probe that reached the database also reported probe-has-no-connection-budget, which sends somebody to provision a pool for a check that should not exist. The pool and budget rules now apply only to the kinds ALLOWED to reach the database. -->
    • S3.14.b Verify read and, where required, transactional write capability without mutating user data. Evidence: probe integration tests. <!-- WRITE_PROBE_METHODS (6) + SAFE_WRITE_PROBE_METHODS (3) + WRITE_BLOCKING_STATES (4) + probeMethodFaults (5 kinds), and pnpm nx probe-controls. WRITE_BLOCKING_STATES IS THE ARGUMENT FOR HAVING A WRITE PROBE AT ALL: four states that leave reads working perfectly and writes failing, so a service probed only by reading stays in rotation and fails every mutation — and the first thing to notice is a user. The harness proves it against the engine rather than describing it: in a read-only session SELECT 1 returned 1 and the INSERT came back cannot execute INSERT in a read-only transaction. rolled-back-write-to-a-user-table is LISTED rather than omitted, because it is the one somebody reaches for — it looks safe, and it works until the first exception path that skips the rollback. CONTROL 3 shows the other half: a concurrent user UPDATE waited behind the probe's open transaction until its lock timeout. THE ROLLBACK ENDS THE WRITE AND NOT THE LOCK. Missed states are named in FULL rather than the first, because they fail independently: a promoted replica and a full disk are unrelated events with the same symptom. -->
    • S3.14.c Read authoritative migration/schema version and compare it with application minimum, maximum, exact, dirty, in-progress, and incompatible states. Evidence: version matrix tests. <!-- SCHEMA_STATES (6) + compareSchema + requirementFaults (3). THE COMPARISON IS NOT SYMMETRIC. Too old is fatal — the code expects objects that are not there, so reads break as surely as writes. TOO NEW IS EXPAND/CONTRACT WORKING: the migration is deployed ahead of the code ON PURPOSE, so a build declaring maximum === minimum reads as a strict careful requirement and is in fact one that refuses readiness for the length of EVERY rollout. That is a fault in the requirement, not the database. ORDER IS LOAD-BEARING: unknown first (nothing below it can be evaluated without a version), then dirty (a half-applied migration means the schema is neither the old one nor the new one, so the number it reports describes nothing), then in-progress — the one state where WAITING is right and failing is wrong, because a startup probe that fails during a migration crash-loops the process waiting for that migration. A NULL VERSION IS NOT ZERO. "No migration tracking" and "no migrations applied" have opposite responses, and CONTROL 4 shows the cost: a probe that catches the missing-table error and defaults to 0 gets too-old — run the migrations — against a database that is probably somebody else's. -->
    • S3.14.d Check required extensions, tables, columns, constraints, indexes, triggers/policies, generated model registry, and permissions. Evidence: compatibility assertions. <!-- REQUIREMENT_KINDS (9) + SILENT_WHEN_ABSENT + OBSERVATION_RESULTS (3) + compatibilityFindings (5 kinds). not-checkable IS A SEPARATE RESULT FROM absent, and collapsing them is the mistake. A catalogue query refused for lack of permission says NOTHING about whether the object exists, and reporting it as missing sends an operator to create a table that is already there — at best an hour, at worst against the wrong database. It still blocks; it blocks with a different remediation. SILENT_WHEN_ABSENT is the sharper half. A missing TABLE errors the first time anything touches it. A missing ROW POLICY raises nothing at all — it returns MORE ROWS to a query that looks exactly like yesterday's. THE HARNESS MAKES THIS CONCRETE AND IT IS WORSE THAN EXPECTED: with RLS disabled the policy is STILL LISTED IN pg_policies, so a compatibility check that looks for the POLICY passes while the tenant reads both tenants' rows. Only pg_class.relrowsecurity — a different catalogue column — catches it. requirement-never-observed covers the gap between the two lists: a requirement nothing looked for is not satisfied, and a report that silently omits it reads identically to one where it passed. -->
    • S3.14.e Return configured/degraded/unavailable/failed status with safe reason and remediation but no DSN, credentials, SQL, or schema secrets. Evidence: public response redaction tests. <!-- PublicProbeView + publicProbeView + PROBE_CODES + disclosures (5 kinds) + safeDetailIsUsable. A CONSTRUCTED OBJECT, NEVER THE REPORT WITH FIELDS DELETED — S1.10's argument, for S1.10's reason: deletion is a blocklist and a blocklist is wrong the first time somebody adds a field. There is nowhere in the shape to put an object name, a version comparison, a SQL fragment or a DSN, and the spec asserts the KEY SET rather than trusting it. A probe body listing the missing tables and columns is a schema dump on an endpoint the orchestrator reaches WITHOUT CREDENTIALS, and it is the most useful possible document for anybody deciding what to try next. The object-name check is the only detector this module owns: the rest is S1.10's scanText, which already argues about credentials, connection strings and SQL and already has the patterns. code-is-a-sentence borrows S1.6's argument about stable classes — every dashboard that grouped on a prose code splits into two series the first time somebody improves the wording. -->
    • S3.14.f Fail readiness before serving writes/jobs when schema is too old, too new, dirty, partially migrated, or missing required policy. Evidence: startup/readiness tests. <!-- SERVED_CAPABILITIES (4) + SERVE_DECISIONS (4) + SCHEMA_STATE_POLICY + readinessDecision + readinessState. FAIL-CLOSED IS ABOUT THE WRITE PATH, NOT ABOUT THE PROCESS. A process that refuses to start cannot serve the reads it could have served and cannot tell anybody why: it is in a crash loop with the explanation in a log nobody is tailing. So readiness is decided PER CAPABILITY. migration-admin is on the list for the reason that makes the whole design work: every other capability can be refused, and if the one that would FIX the situation is refused with them, the deployment has locked itself out of its own repair. It survives too-old, too-new, dirty and in-progress, and is refused only under unknown — where nothing here is a claim about this database, INCLUDING the claim that it is the right one. too-new gives reads serve-degraded, which is the only honest answer: refusing them turns a recoverable rollout or rollback into an outage, and serving them silently hides that a contract-phase migration may have removed something this build still reads. in-progress is wait and not refuse — a refusal tells the orchestrator this process is broken, and it is replaced by another one that will also be waiting. THERE IS NO SECOND STATUS LADDER: readinessState returns an S1.6 CapabilityState, so usability() answers for it. S1.6 already argued that a boolean cannot represent degraded without rounding it to a neighbour and that whichever way it rounds is wrong — and a probe is exactly where that rounding happens. The owner is TAKEN rather than invented; a default would be a plausible-looking rota that pages nobody. -->
    • S3.14.g Instrument latency/failure without high-cardinality database detail and alert on sustained incompatibility. Evidence: probe observability tests. <!-- FORBIDDEN_METRIC_LABELS (13) + metricFaults (3 kinds) + ALERT_VERDICTS (4) + alertDecision. A metric labelled by TABLE grows one series per table and publishes the table list to everyone with dashboard access; a metric labelled by SCHEMA VERSION grows a new series on every migration and never retires the old ones, which is the shape that quietly doubles a metrics bill over a year. no-outcome-label is the one that makes the metric misleading rather than expensive: with success and failure averaged together, A FAST FAILURE LOOKS LIKE AN IMPROVEMENT. THE ALERT HAS THREE WAYS TO BE USELESS AND ALL THREE ARE CHECKED. A threshold shorter than the deploy window fires on every release, and an alert that fires on every release is one somebody mutes — after which it never fires for the reason it was written; that is checked FIRST, because a decision made with a broken threshold is not worth computing. A GAP in the readings is NOT continuity: the incompatibility may have ended and restarted inside the hole, and a duration computed across it invents the part nobody measured. And the window is measured from the current RUN of incompatibility rather than from the first incompatible reading ever seen. -->
  • S3.15 Write migration templates for expand/backfill/verify/contract. Evidence: a sample zero-downtime migration rehearsal passes. <!-- docs/migrations/ (README + four templates) + scripts/v1-workbenches/generate-migration-templates.mjs (+ 10 tests) + scripts/v1-workbenches/migration-rehearsal.mjs. Follows S3.1's shape: a document with a machine-checked companion, rather than a kit entry point — S3.15 does not introduce a runtime contract anybody imports. A SCHEMA CHANGE THAT CAN BE DEPLOYED WITHOUT DOWNTIME IS NOT ONE MIGRATION. It is four, separated by DEPLOYMENTS, and the separation is the whole mechanism: between any two there is a moment where the old code and the new code are both live against the same database and both have to work. The flag day fails in a way nobody can reproduce afterwards — a few minutes of errors during a rollout, after which the fleet is consistent again. -->

    • S3.15.a Create expand template for additive nullable/default-safe schema, indexes/concurrent build, permissions, generated client update, and compatibility checks. Evidence: expand template. <!-- docs/migrations/expand.template.sql, 7 checked requirements. THE COMPATIBILITY CHECK IS RUN AGAINST THE OLD BINARY, and running it against the new one would be checking the wrong thing — the old binary is the one still deployed, and this phase's entire claim is that it cannot tell anything changed. A VOLATILE DEFAULT IS REFUSED and directed to the backfill: now() and gen_random_uuid() need a distinct value per row, so the engine rewrites the whole table under ACCESS EXCLUSIVE — on EVERY version, including the 11+ the templates target, where a non-volatile default is only a pg_attribute entry. The same statement is a metadata change or an exclusive lock over a hundred million rows depending on a version number nothing in the statement mentions. CREATE INDEX CONCURRENTLY is followed by an INVALID-index check, because that is the failure that leaves no error anywhere: the statement reported failure once in a terminal, and the index sits in the catalogue being ignored by the planner while everybody assumes it is there. Permissions are verified with has_column_privilege AS THE APPLICATION ROLE — the migration runs as an owner, so everything works for it. -->
    • S3.15.b Create backfill template with immutable input query, batches, checkpoints, idempotency, rate/lock limits, quarantine, progress, and resume. Evidence: backfill template. <!-- docs/migrations/backfill.template.md, 14 checked requirements. THE CHECKPOINT IS WRITTEN IN THE SAME TRANSACTION AS THE BATCH. Written after, a crash re-runs a batch — survivable, because the AND new_column IS NULL predicate makes it a no-op. Written BEFORE, a crash SKIPS one, and nothing later detects it. Idempotency is two clauses covering different races: the NULL predicate stops a retry overwriting a value the application has since changed, and FOR UPDATE SKIP LOCKED lets two workers run without either waiting — SKIP rather than NOWAIT, because a locked row is one somebody is editing and that is a reason to come back rather than to fail the batch. §4b IS THERE BECAUSE THE REHEARSAL FOUND THE BUG IN ITS OWN TRIGGER. A dual-write trigger written as "fill whichever column is NULL" is correct on INSERT and DOES NOTHING ON UPDATE — both columns already hold a value, so neither branch fires and the two diverge from the first edit onward. Nothing detects it until the verify phase's write-parity check, which is exactly what caught it. The template now keys on TG_OP and propagates the side that CHANGED, and is BEFORE rather than AFTER (an AFTER trigger needs a second UPDATE, which fires the trigger again). -->
    • S3.15.c Create verify template for counts, nulls, constraints, hashes, duplicates, orphans, semantic samples, performance, and old/new read/write parity. Evidence: verify template. <!-- docs/migrations/verify.template.sql, 9 checked requirements, and SET default_transaction_read_only = on is enforced by the checker rather than intended — this phase has no business writing, and a template that says so only in a comment is one somebody edits in a hurry. EVERY QUERY STATES ITS EXPECTED RESULT. A verify template returning numbers with nothing to compare them against is a report somebody skims. IS DISTINCT FROM rather than <>: NULL <> NULL is NULL, which is not true, so a plain <> silently excludes exactly the rows the null-count section just told you to look at. invented is the column worth reading twice — a value in the new column where the source had none is a default applied somewhere nobody intended. Duplicates and orphans are checked HERE because phase 4 tightens those constraints, and a violation found under an ACCESS EXCLUSIVE lock is found in the worst possible place. §9 checks the plan for the same reason: phase 4 removes the column the current queries use, and a Seq Scan here means it replaces a fast query with a slow one at the moment the fallback disappears. -->
    • S3.15.d Create contract template for precondition proof, old writer/ reader absence, removal/constraint tightening, cleanup, and rollback or compensation. Evidence: contract template. <!-- docs/migrations/contract.template.sql, 10 checked requirements. THE ONE-WAY DOOR, so it opens with a PROOF rather than a checklist — a checklist is a list of things somebody believed, and §1 is four queries whose answers are facts about the running system, each filed WITH the time it was taken, because a proof from yesterday is a proof about yesterday's fleet. NO OLD READER is the fact people skip, because "we deployed it last week" feels like the same statement. It is not: a job runner, a cron pod, a scaled-to-zero worker and a paused canary are all processes that can still be running last month's build. §0 checks for a long-running query BEFORE the DDL, because the DDL will wait for its lock and WHILE IT WAITS IT BLOCKS EVERYTHING ELSE — a 3-second migration started behind a 40-minute analytics query is a 40-minute outage. NOT NULL is tightened through a NOT VALID check plus VALIDATE so the exclusive lock is brief and the scan runs under SHARE UPDATE EXCLUSIVE. And §7 lists what DROP COLUMN did NOT remove — for an ordinary column that is bookkeeping, for a classified one it is the difference between "removed" and "removed from queries". -->
    • S3.15.e Include owner, approvals, maintenance/lock budget, observability, abort threshold, incident/rollback, evidence paths, and completion criteria in every phase. Evidence: template completeness test. <!-- COMMON_SECTIONS (8) enforced across all four templates by generate-migration-templates.mjs, with a NEGATIVE CONTROL: the safety scan is re-run over a deliberately-broken template built in memory, and if that produces no findings the scan is not working and a clean result over the real files means nothing. It produces 23. THE CHECKER FOUND TWO REAL GAPS ON ITS FIRST RUN — the backfill template had no ## Incident and rollback section at all, and the CREATE INDEX scan fired on the SENTENCE explaining why the plain form is dangerous. The second is the more interesting fix: a statement is distinguished from a mention by its ON <target> clause, because without that distinction the way to quiet the checker is to DELETE THE EXPLANATION. Each section is a question asked during an incident at the point where nobody has time to work out the answer: the owner is a ROTA URL and never a name; the lock budget is enforced with lock_timeout in the template rather than in somebody's intention; the abort threshold is a NUMBER agreed in advance, because the decision to abort is made by a tired person at 03:00 who has already invested two hours. -->
    • S3.15.f Implement a sample rename/type-change migration using dual- version coexistence rather than flag-day deployment. Evidence: sample migration artifacts. <!-- The worked rename (story_source.contact -> story_source.contact_email) lives INSIDE scripts/v1-workbenches/migration-rehearsal.mjs rather than in a sample/ directory beside the templates, because a sample that is only READ decays and one that is EXECUTED on every run cannot. A rename is the worked example precisely because it is the change that most invites a flag day: RENAME COLUMN is instant, takes a trivial lock, and looks free — and is the single statement that guarantees every process built against the old name starts failing the moment it commits. TWO CONTROLS PROVE THAT, and both must break the old writer. CONTROL A commits the rename and the old writer's next INSERT fails with column "contact" ... does not exist. Committed rather than held open, and that distinction cost a hung run: an UNCOMMITTED rename holds ACCESS EXCLUSIVE so the other session's INSERT WAITS rather than failing, which demonstrates lock contention and not the flag day. CONTROL B found something better than what was written for it: the engine REFUSES ADD COLUMN ... NOT NULL on a non-empty table outright, so the form that actually reaches production is a column added WITH a default and the default dropped in a later migration — and that one breaks the old writer with a not-null violation on an INSERT that does not name it. -->
    • S3.15.g Rehearse on production-size representative data with old/new binaries, live writes, interruption, rollback/compensation, and final query plans. Evidence: timed zero-downtime report. <!-- scripts/v1-workbenches/migration-rehearsal.mjs: 10,000 rows, four phases, 0.9s. THE OLD BINARY WRITES THROUGHOUT, which is the only version of this rehearsal worth running — "no downtime" is a statement about the minutes when both binaries are live, and a rehearsal with one writer rehearses the situation that was never in doubt. An old writer (INSERT naming only contact) and a new writer (naming both) insert between every backfill batch, and the old writer's UPDATE is what the parity check runs against. The backfill is INTERRUPTED at cursor row-002000 by ending the connection, then resumed from the checkpoint by a fresh one; nothing special happens, which is the property. The ROLLBACK REHEARSAL is done from that half-converted state and shows it is LEGAL: every row still carries the old column, so rolling the new binary back loses nothing. Plans are captured before and after the drop and compared — a plan that CHANGED across the drop means the planner was using something now gone — and the report ends by stating the 10,001 dead tuples still holding the old values until a rewrite, which is the part that reads as done and is not. -->
  • S3.16 Add tenant-isolation integration tests using two tenants and colliding object IDs. Evidence: every read, write, search, export, and event path remains isolated. <!-- tools/tenant-isolation-matrix.mjs (26th Nx target). 6 matrices, 34 probes, 6 negative controls. IT DIFFERS FROM S3.2.f's tenant-isolation IN ITS ENFORCEMENT MECHANISM, which is the whole reason it is a second harness rather than more cases in the first. S3.2.f checks that the queries CARRY their tenant predicates. This one TAKES THE PREDICATES AWAY: every statement runs as a role that cannot bypass row-level security, with no tenant clause in it at all, so what is under test is the DATABASE's isolation rather than the application's discipline. An isolation rule enforced by a predicate is one a missing predicate disables, and the missing predicate is the entire failure mode. -->

    • S3.16.a Provision two tenants with identical user-facing identifiers, object IDs where storage permits, names, taxonomy, timestamps, and content markers. Evidence: collision fixture setup. <!-- Both tenants hold workspace newsroom, slug quarterly-review, term investigations, the same created_at and the same content marker; only the id carries a tenant prefix, which is what a real system does. The fixture asserts the collision — 2 rows across 2 tenants share the natural key — before any probe runs, because isolation bugs do not show up on distinct data. -->
    • S3.16.b Test direct get/list/filter/paginate/count/facet/join/history/ revision reads across both identities and roles. Evidence: read isolation matrix. <!-- 8 shapes plus the count, run from BOTH identities so a policy that happens to name one tenant is caught, and every one written without a tenant predicate. The COUNT is its own case: it returns no row and still answers how many the other tenant has. Roles are the kit's half — decideAccess must refuse a cross-tenant contributor read with different-tenant — so the engine and the contract are checked against each other rather than each asserted alone. -->
    • S3.16.c Test create/update/delete/restore/transition/bulk/import/review/ waiver/publish mutations with cross-tenant targets and relationships. Evidence: write isolation matrix. <!-- 10 mutations, including the two that are not obvious: an INSERT with a FORGED tenant_id, which the policy's WITH CHECK refuses, and an UPDATE that tries to MOVE a row between tenants, which is the mutation that would otherwise pass every read check afterwards. THE BULK CASE NEEDED ITS OWN RULE. UPDATE workbench SET stage=... with no predicate SHOULD affect rows — the asking tenant's own — and the first version of the harness failed it for that. The invariant is not "touched nothing", it is "did not reach past its own", which is why every write is verified against the OTHER tenant's row. -->
    • S3.16.d Test full-text/structured search, suggestions, saved views, caches, indexes, and unauthorized result counts/snippets. Evidence: search/cache isolation matrix. <!-- Real tsvector + GIN, so the full-text path is the engine's rather than a LIKE standing in for it. The two that leak WITHOUT RETURNING A ROW are checked separately: the result COUNT over a marker both tenants carry must be 1, and ts_headline must not produce a fragment of the other tenant's body. A search surface can be perfectly scoped in its rows and still answer "how many" and "roughly what". -->
    • S3.16.e Test blob upload/download, derivative, signed URL, bundle/export, backup/restore slice, audit export, and data-subject flows. Evidence: artifact/export isolation matrix. <!-- 7 paths. The bundle export is the interesting one — a LEFT JOIN from workbench to blobs, which is the shape where a scoped driving table carries an unscoped join and returns the other tenant's rows in its second column. The backup SLICE is checked as a slice rather than a dump: a per-tenant copy taken as the tenant. -->
    • S3.16.f Test command/event/job/outbox/inbox/provider/callback/notification identity and context propagation. Evidence: asynchronous isolation matrix. <!-- The asynchronous paths are where the tenant is CARRIED rather than joined, so the two failures are a drain that reads another tenant's pending work and a CLAIM that marks it processed. Both are checked, and so is an inbox insert with a forged context — a provider callback is exactly where a tenant id arrives from outside. -->
    • S3.16.g Test error, trace, metric, log, support, and health diagnostics for cross-tenant leakage. Evidence: observability isolation matrix. <!-- The assertion here is about ISOLATION rather than about how the engine formats a message: the unique violation must fire the TENANT-SCOPED constraint. A global unique over (workspace, slug) would have collided with the OTHER tenant's row, and the refusal would then be an oracle — "somebody else already has this slug" — returned to a caller not allowed to know it. The health count is checked to be strictly smaller than the unrestricted one, so a diagnostic cannot quietly become a cross-tenant total. -->
    • S3.16.h Run tests against the real supported database with row-level policies/constraints enabled and fail on any disclosure or mutation. Evidence: passing real-DB report. <!-- RLS is ENABLED and FORCED on all five tables, the app role is NOBYPASSRLS, and the harness asserts that before running anything — a role that can bypass makes every probe pass whatever the policies say, which is the failure S3.14's CONTROL 2 already ran into once. fail on any MUTATION is a separate rule from fail on any disclosure: every write is compared before and after against the target row, so a statement that was refused and mutated anyway still fails. CALIBRATED BY BREAKING IT. A policy of USING (true) produces 28 findings. Removing WITH CHECK produces none — because Postgres reuses the USING expression as the check when none is given, which is worth knowing and is not a weakness. And a unique key without the tenant makes the colliding fixture impossible to SEED, which the harness reports rather than works around. -->

S4 — API, authorization, audit, and abuse controls#

  • S4.1 Implement createWorkbenchRouter without encoding Tara-specific routes or roles. Evidence: both archetype fixtures mount successfully. <!-- src/router.ts + src/router.spec.ts (46 cases, all green with the kit's 2575). New public subpath ./router, authorized by S4.1 in KIT_OWNED_ENTRY_POINTS; the export map regenerates to 47 entry points with no lattice violation, which is also what proves the router imports no Node built-in, no DOM global and nothing above its class. THE FAILURE THIS IS BUILT AGAINST IS NOT A MISSING MIDDLEWARE, IT IS ONE IN THE WRONG PLACE. Authorization after the handler; the body parsed before the size was checked; an idempotency key claimed by a caller who was about to be refused. Each reads as correct in review because each middleware is individually right. So {@link MIDDLEWARE_STEPS} is DATA, the response carries the steps that actually ran, and every order case asserts a prefix property — a request that fails at step N ran exactly 1..N and nothing after. That is checkable for every step and every failure; "the middlewares are registered in the right order" is checkable by reading the code that is already wrong. -->

    • S4.1.a Define router inputs for plugin registry, contracts, actor resolver, authorization, stores, jobs, providers, audit, tracing, limits, and capability probes. Evidence: constructor type tests. <!-- Eleven seams, and the test loops over all ten optional ones removing each in turn: a missing seam is one missing-seam violation naming it, refused at construction rather than at the request that first reaches the step needing it. EACH SEAM IS USED BY THE ROUTER AND NOT MERELY PASSED THROUGH, which is the difference between a seam and a decoration. jobs is asked whether an accepted job exists before a 202 goes out. providers is WRAPPED, so an operation the deployment does not offer returns a typed not-configured without the call reaching the host — asserted by counting host calls: 1 for the known operation, 0 for the typo. stores is read by both the idempotency claim and the precondition. probes decides unprobed versus mounted. requestDigest returns a plain string and is PARSED rather than typed as Sha256Hex and trusted: the seam is implemented by a host the compiler never sees, and a truncated digest would silently become a key that never matches its own retry. -->
    • S4.1.b Register only capability-declared routes and versions with deterministic collision refusal. Evidence: route registration tests. <!-- A route mounts only when exactly one loaded plugin declares its owning capability as implemented or degraded, AND the deployment's CapabilitySupport serves its contract major. Undeclared, declared- unsupported (the route says the plugin does what the plugin says it does not), declared by two plugins, and unserved-version are four separate refusals with their own kinds. THE DETERMINISM CLAIM IS ASSERTED AS AN EQUALITY, NOT AS NON-EMPTINESS: the colliding pair is passed forwards and backwards and the two violation lists must be toEqual. Collision is on the NORMALISED path — /things/:thingId and /things/:id are one route with two spellings — while /things/new against /things/:id is not a collision but a specificity question, resolved by a computed weight rather than by insertion order, and tested in both directions. Excluded routes are REPORTED (other-host, ui-surface, retired-with-its- redirect) for S1.8.g's reason read backwards: a router that quietly mounts fewer routes than it was given is indistinguishable from one that never knew about them. -->
    • S4.1.c Apply ordered authentication, scope resolution, validation, authorization, idempotency/concurrency, limits, tracing, audit, and safe-error middleware. Evidence: middleware-order tests. <!-- Eleven ordered steps plus two finalisers. Authorization is SPLIT (route roles need no body, so a wrong-role caller never has one parsed; object authorization needs the target validation produces) and limits are split (a size cap before anything reads, rate and spend once the caller is known and permitted). Both splits are justified by a case: bodyReads is asserted 0 at the size cap and at the role check. THE PLACEMENT THAT MATTERS MOST IS THE IDEMPOTENCY CLAIM, which is AFTER authorization — the test counts record lookups and asserts 0 for a request authorization refuses, because a key claimed earlier lets an unauthorized caller burn another caller's key and learn from the replay that one exists. Scope resolution does the check no descriptor can make statically: S1.8 already refuses a route taking its tenant from both the path and the session, but a VALID session for one tenant on a path naming another is a request-time comparison, and it is refused rather than resolved in either direction. Concealment renders as 404 and not 403, because a 403 confirms the row exists to somebody not allowed to know that. A thrown handler becomes internal.unexpected and the thrown text is asserted absent from the envelope; the audit row still records the step it died at, for a request that never reached a handler at all. -->
    • S4.1.d Keep stage names, roles, content types, gates, routes, copy, provider operations, and publish targets in domain plugins. Evidence: shared source scan. <!-- tools/router-vocabulary-scan.mjs (27th Nx target). 30 terms, 40 shared modules, 0 findings. THE CORPUS IS DERIVED, NEVER AUTHORED: every term reads a live exported value — the stage model's own stages, the block's own media types, the routes' own paths and roles — so a word added to a plugin next year is guarded the day it is added. The archetypes are excluded from dist on purpose (a worked example in the shipped package is a fixture handed to every consumer), so the tool transpiles their source in memory and imports it as a data URL with specifiers rewritten to the BUILT kit. THE HARD PART WAS NOT FINDING TERMS, IT WAS NOT DROWNING IN THEM. A domain may name a stage with an ordinary English word, and draft, published, resolved, triage and monitoring are all real archetype stages AND all words this kit uses for unrelated things — an outbox row is published, an anchor is resolved, an artifact channel is draft. The first version reported 52 findings, 51 of them false, which is how this kind of check gets switched off. So a term is reported only where the collision cannot be coincidence: a literal COMPARED against a vocabulary-shaped subject (lifecycleStage, not state), a COMPOUND term as an identifier, or a declared route path anywhere. IT FOUND A REAL ONE. tenancy.ts read request.lifecycleStage === 'published' — the kit deciding what a terminal stage is called, which protects a workbench whose stage happens to be published and silently protects nothing in one whose stage is released, aired or filed. AccessRequest.lifecycleStage is now stageIsImmutable: boolean, answered by the plugin's own stage model, and the denial is immutable-stage-rows-are-not-editable. Calibrated in the other direction too: every run first scans a synthetic module built FROM the corpus and exits non-zero if it finds fewer than 4 planted terms, because a clean report from a scan that cannot see is the worst possible outcome. -->
    • S4.1.e Mount content-authoring and operations-console fixtures with disjoint types and routes. Evidence: archetype integration tests. <!-- Both archetypes gained real routes, bindings and command schemas — in the PLUGINS, which is S4.1.d holding: 4 desk routes under /desks, 3 console routes under /consoles, and the test asserts the two share no path segment, no role and no stage. The handlers run the archetypes' OWN logic rather than returning a shape resembling what it would have produced: the attribution measure over a real draft (1 claim, 1 attributed), the runbook freshness measure (181 days, one unverified step), the markdown reader (slug quarterly-review, 1 paragraph), the pull-quote editor. If the pipeline hands a handler the wrong body, parameters or scope, the domain function refuses and the test sees it. The console is TENANT-scoped where the desk is WORKSPACE-scoped, and that is a real difference rather than variety for its own sake: an incident crosses every workspace in a tenant, and scoping it to one would hide it from the responders in the others. -->
    • S4.1.f Refuse missing/duplicate/incompatible plugins and expose honest capability state. Evidence: startup negative tests. <!-- Duplicate plugin id, kit-version-unsupported, no plugins at all, orphan route, orphan binding — and every violation is collected rather than returned at the first, so a host with three mistakes finds three in one run instead of three deploys (asserted: two independent mistakes, two violations). THE CAPABILITY ROWS ARE WHERE THE HONESTY LIVES. unprobed is its own disposition and is NEVER rounded up to mounted: a capability nothing measured is not a working one, and a mounted route over an unmeasured capability is the exact shape in which an absent measurement renders as a good one. A declaration of unsupported outranks any probe and is reported as declared-unsupported with no routes — the archetypes supply two real cases, the desk's capability:image-generation and the console's capability:structured-authoring. --> <!-- S4.1 refactor note: AccessRequest.lifecycleStage became stageIsImmutable (S4.1.d above). Callers updated: tenancy.spec.ts and tools/tenant-isolation-matrix.mjs. No consumer outside the kit. --> <!-- S4.1 rollup: pnpm nx build/lint/test @oshun/workbench-kit green; 2575 kit tests pass; node tools/router-vocabulary-scan.mjs and node tools/tenant-isolation-matrix.mjs both green after the tenancy change; generate-workbench-kit-export-map.mjs --check reports no drift and no violations at 47 entry points. -->
  • S4.2 Reuse the canonical Studio actor resolver rather than inventing a second role system. Evidence: role parity tests against @oshun/studio-authoring. <!-- src/actors.ts + src/actors.spec.ts (31 cases). New ./actors subpath, 48 entry points, no lattice violation. THE FIRST HONEST FINDING IS THAT HALF THE THING TO REUSE DOES NOT EXIST. @oshun/studio-authoring — the one authoring package this layer may depend on, per the ADR-S1 constraint @nx/enforce-module-boundaries enforces — is canonical for the ROLE TAXONOMY and its permission matrix. It has no session, no token, no claim set and no resolveActor. The packages that do are layer:auth, which layer:workbench-kit may not reach, and reaching sideways would be the second identity system this item exists to prevent. AND IT MAY NOT BE IMPORTED EITHER. The kit is a BUILDABLE package emitting under rootDir: src, so importing another workspace library's source puts that library's files outside its own root — TS6059, the compiler saying the emitted package would be incomplete. Which produced the better design rather than a workaround: the kit holds NO role list, NO permission table and NO role-to-principal map, so there is nothing in it that CAN drift. The taxonomy arrives as a CanonicalRoleModel, and the spec binds the real one — which is also exactly where the item asks for it ("role parity tests against @oshun/studio-authoring"). -->

    • S4.2.a Inventory canonical claims, sessions/tokens, human/service/system variants, tenant/workspace membership, roles, delegations, and support access. Evidence: resolver contract map. <!-- CanonicalClaims, closed by parseCanonicalClaims over 16 named fields, covering every item in the cell: subject/audience/issuer/ issuedAt/expiresAt/sessionId/sessionStrength/revoked (claims and session), tenant + workspaces (membership), roles, onBehalfOf + delegationGrant (delegation), serviceName (the machine variant), supportGrant, sessionCorrelation. THE CLOSED SHAPE IS THE POINT. A claim set is exactly the object somebody widens by passing the whole decoded token through, and rejectExtraFields refuses the display-data class BY NAME — the test adds an email and gets "carries display data", not a generic unknown field. Three invariants are refused rather than tolerated: half a delegation (the audit row could answer "who did this" or "whose authority", never both), a window that expires no later than it was issued (a clock or signing bug, not an expiry), and an ABSENT revocation flag — which would read as "not revoked" and is indistinguishable from a revocation check nobody performed. -->
    • S4.2.b Adapt the canonical resolver without copying role strings or reinterpreting claims. Evidence: adapter source review. <!-- The source review is short because there is nothing to review: grep src/actors.ts for a canonical role name and there is not one. The taxonomy is a three-member CanonicalRoleModel — roles, permissions and a holds FUNCTION, not a matrix, because a matrix crossing this boundary would be a copy taken at a moment of something allowed to change. THE TEMPTATION WAS THE PRINCIPAL MAPPING. A twelve-row Record<StudioCreatorRole, Principal> would have been convenient, would have been TOTAL (so a thirteenth canonical role would break the build), and would have been a transcription of the role names in the one package that must not hold them. It is PrincipalMapping, supplied by the deployment, and the compile-time guarantee it would have bought is bought instead by S4.2.f's drift check — which is strictly better, because it also catches the version skew a compile-time check cannot see. -->
    • S4.2.c Preserve authenticated actor, delegated actor, tenant/workspace, session strength, and correlation. Evidence: actor context tests. <!-- ResolvedActor gained sessionStrength, sessionCorrelation, supportGrantLive, canonicalRoles and unrecognisedRoles. SESSION_STRENGTHS is four values rather than a boolean because S4.6 has to be able to require a hardware key and strong: boolean cannot express the middle — and service-key is deliberately NOT the strongest, since a long-lived key is not somebody who just touched a security device, and ranking it as one is how a service account satisfies a re-authentication requirement. The delegated case carries BOTH identities into ActorRef; the support grant is live only while it is, so a session outliving its grant is standing access with an audit trail rather than access. -->
    • S4.2.d Reject missing, expired, revoked, malformed, wrong-audience, cross-tenant, and unsupported identity state. Evidence: negative tests. <!-- Seven reasons in REJECTION_REASONS, each asserted to arrive as itself, plus a TOTAL REJECTION_CODES record in the router so adding a reason without deciding what a client is told is a compile error. THE ORDER OF THE CHECKS IS ARGUED RATHER THAN INCIDENTAL. Shape first, because nothing can be read off a malformed record. Audience next, so a token minted for another service is not inspected at all. Then REVOCATION BEFORE EXPIRY — asserted directly: a session that is both reports revoked, because reporting the expiry would tell an operator it simply timed out, which is a different incident from one somebody revoked. cross-tenant maps to tenancy.tenant_mismatch and not to an authentication code: the credential was perfectly valid and was presented against somebody else's data, which is a security event rather than a reason to show a login prompt. A token carrying two identities at once is unsupported rather than resolved, because picking one means picking which half of a compromised token to believe. -->
    • S4.2.e Compare every relevant role, claim, and delegation result between Studio and the shared router. Evidence: parity fixture matrix. <!-- parityMatrix over 12 roles x 10 permissions = 120 cells. THE FIRST VERSION OF THIS WAS A TAUTOLOGY AND IS WORTH RECORDING. It asked the canonical model for both sides, so it agreed with itself unconditionally and could not have failed — a test that proves nothing while reading as thorough. The answerer is now a PARAMETER, so the matrix has something it can genuinely disagree with, and the negative control passes a transcribed table with one cell wrong (author gains approve-publish) and asserts exactly one disagreement is reported, naming both sides. Only then does the real run — the kit's own answerer against the canonical model, zero disagreements — mean anything. A second guard for vacuity: the canonical model is asserted to contain both answers, because a matrix over a model that said true everywhere would agree with anything. -->
    • S4.2.f Detect new canonical roles/claims as drift rather than silently treating them as unprivileged or privileged. Evidence: drift test. <!-- Two halves, because the failure has two shapes. AT STARTUP: roleDrift(model, mapping) runs BOTH directions — a canonical role nothing maps (the identity service deployed ahead of the workbench) and a mapping for a role no longer in the taxonomy (a grant nobody can find by reading the taxonomy). It REPORTS and never resolves, which is the whole item: treating an unmapped role as unprivileged silently removes access from whoever holds it — an outage with no error in it — and treating it as privileged makes an unknown string a grant. PER REQUEST: an unrecognised role in a token is CARRIED on the resolution rather than filtered out, and classifyRoles splits a token's roles three ways against the vocabulary the router's own mounted routes name: canonical, domain, and unattributed. The third bucket is the one worth having — a version skew and a forged claim both land in it, and neither may be silently honoured. --> <!-- S4.2 rollup: 2607 kit tests pass (31 new); tsc -p tsconfig.lib.json and -p tsconfig.spec.json both clean; export map regenerates to 48 entry points with --check reporting no drift and no violations; tools/router-vocabulary-scan.mjs still clean with actors.ts in scope. @oshun/studio-authoring is a DEV dependency and a spec-only vitest alias, so the emitted package does not carry it. -->
  • S4.3 Define object-level authorization hooks. Evidence: an authorized role cannot access another tenant or unassigned protected object. <!-- src/authorization.ts + src/authorization.spec.ts (44 cases). New ./authorization subpath, 49 entry points, no lattice violation. THE ITEM'S OWN SENTENCE CONTAINS THE PROBLEM: the caller is AUTHORIZED. They passed authentication, they hold a real role on a real route, and every check that looks at the REQUEST says yes. Only a check that looks at the OBJECT can refuse them, and the reason platforms ship without one is that every individual endpoint looks fine. -->

    • S4.3.a Define hook input for actor/delegate, tenant/workspace, action, object type/ID, lifecycle, owner/assignment, classification, policy, and request context. Evidence: hook contract. <!-- AuthorizationInput = actor + resolved principal + roles + tenant/workspace + action + ObjectMetadata + PolicyContext + RequestContext. The delegate half comes through ActorRef's delegated variant, which already carries both identities. ObjectMetadata HAS NOWHERE TO PUT CONTENT. No title, no summary, no body, no displayName — and the test asserts the exact key set rather than describing the rule, so a field added for convenience fails. Not "we redact it before use": there is no field, which is the only version of that promise that survives somebody in a hurry. -->
    • S4.3.b Load authoritative minimal object metadata through a tenant-scoped resolver before decision. Evidence: loader isolation tests. <!-- loadObject passes the scope to the host's resolver AND compares the tenant on what comes back. THAT SECOND HALF LOOKS REDUNDANT AND IS THE MOST VALUABLE LINE IN THE MODULE. The most ordinary bug in a resolver is WHERE id = $1 with the tenant predicate left off — invisible to every test that reads its own tenant's data, and a cross-tenant read the first time an id collides or is guessed. One comparison turns it from a leak into a faulty outcome an operator sees. faulty is its OWN result rather than a thrown error or a silent conceal: concealing it would make the leak look like a missing row and every request would keep being served by a resolver that does not scope. The caller still learns nothing — they get unavailable — and the fault goes to the log. -->
    • S4.3.c Compose domain policy, ownership, assignment, team/role, protected-object, legal-hold, lifecycle, and support rules. Evidence: policy truth tables. <!-- Eight families plus a NINTH, parent-visibility, because the eight are all answerable from the object alone and that one needs a second object; folding it in would put a two-object question inside a truth table over one, where it would eventually be simplified away. Domain policy DELEGATES to S3.2's decideAccess rather than re-deciding — a second implementation would be a second chance to write >= where === belongs, in the module least likely to be read again — and the one translation it performs is which denials become CONCEAL. The rules that are narrower than they look: ownership speaks only about administer, because a story's author does not get to bypass review; a legal hold PRESERVES and does not hide, so it refuses only the operations policy names and a read is untouched (a hold that blocked reads would stop the review it exists to enable); assignment compares against the CURRENT list on the loaded metadata, never anything the request carried, so a reassignment two minutes ago takes effect now. -->
    • S4.3.d Return allow, deny, concealed-not-found, step-up, or unavailable with stable safe reason/remediation. Evidence: result fixtures. <!-- OUTCOME_PRECEDENCE is DATA and the tests assert it rather than describing it, because "which refusal does the caller see" is a disclosure decision and an ordering accident is exactly how a 403 replaces a 404 for a cross-tenant id. CONCEAL > DENY: a 403 is an answer — it says the object exists. DENY > UNAVAILABLE: when something already said no, no is the complete and more useful answer. UNAVAILABLE > STEP-UP: telling somebody "authenticate again and you are in" when the next attempt still cannot be decided is a promise nothing keeps. A concealed object and an absent one are asserted INDISTINGUISHABLE in everything the caller can observe — same disposition, same message key, same remediation — while the logged reason differs, because an operator has to tell "no such row" from "not yours". And every rule runs even after one refuses, so contributing answers "you were added to the assignment and it is still refused" without a rerun. A final test produces every declared reason from real code, so a reason nothing can produce cannot sit in the catalogue looking like coverage. -->
    • S4.3.e Apply hooks to reads, counts, search, history, blobs, exports, jobs/events, mutations, and indirect parent/child access. Evidence: endpoint coverage matrix. <!-- ACCESS_PATHS (11) + coverageGaps, checked in the direction that matters: not "is every hook applied to something" but "is there a path with no hook on it". A count, a search snippet and a signed blob URL are three doors into the same room and two of them answer without ever returning the object. Two gap kinds with different sentences, because they are different failures: a path nothing declares is one nobody has looked at; a path whose surfaces all skip the hook is one somebody looked at and left open. toRouterDecision closes the router path — the mapping lives here rather than in each host because unavailable carries a retry-after and the other refusals must not, and a host that guessed would either hammer a failing dependency or give up on a blip. -->
    • S4.3.f Test colliding cross-tenant IDs, unassigned protected objects, hidden parents, indirect refs, and stale assignments. Evidence: disclosure negative tests. <!-- Six negatives, every one of them a WELL-FORMED request from an authorized caller: a colliding id identical in every field but the tenant (which is how a collision actually looks); a protected object with an EMPTY assignee list, reachable by nobody — reading empty as "no restriction" would invert the control; a reassignment the caller's request still claims; an object legitimately inside the tenant but under a parent the caller cannot see, which is "I cannot list your folders but I can open a file by id"; a support session WITH a live grant that still may not see a protected source; and an unreachable metadata store, where defaulting to visible leaks and defaulting to hidden breaks every object whose lookup was merely slow, so neither answer is given. --> <!-- S4.3 rollup: 44 new cases, kit suite green; tsc -p tsconfig.lib.json and -p tsconfig.spec.json clean; export map at 49 entry points with --check reporting no drift and no violations. -->
  • S4.4 Define property-level authorization hooks. Evidence: mass-assignment tests cannot modify owner, tenant, approval, gate, provenance, or audit fields. <!-- src/properties.ts + src/properties.spec.ts (47 cases). New ./properties subpath, 50 entry points, no lattice violation. S4.3 answers "may you touch this row"; this answers the question that is left and the one mass assignment exploits — you may update the row, so WHICH FIELDS may you set. A system with only the first check lets an author who may edit their own draft set ownerId, approvedBy and tenantId in the request that fixes a typo. The evidence clause names six fields and the interesting half is not the names: each must stay unwritable through every mutation shape AND every spelling a client can produce. The corpus asserts 11 protected classes x 5 shapes = 55 refusals, plus 6 spellings of the owner field. -->

    • S4.4.a Define readable, createable, updateable, clearable, sortable, filterable, searchable, exportable, and audit-visible permissions per field/path and object state. Evidence: property policy schema. <!-- Nine permissions, five write-resistance classes, plus per-field writableInStates and requiredScope. server-owned and write-once are kept apart because the difference is real: a server-owned field is refused at CREATION too, a write-once one is settable exactly then — asserted both ways. A BUG THIS FOUND IN ITSELF, worth recording: canonicalPath stripped * as punctuation, so every wildcard policy row was silently dropped from the index and blocks.0.text came back unknown-field while the policy file still read as covering it. A control that fails open on its own configuration is worse than an absent one. Fixed, with a regression test that also proves a non-wildcard punctuation segment is still refused. -->
    • S4.4.b Normalize patches, nested objects, arrays, JSON Patch, forms, import mappings, and generated proposals before policy. Evidence: mutation-shape tests. <!-- normalizeMutation is the ONLY entry, so no shape has its own check and therefore no shape has its own bypass. Five declared shapes reduce to one FieldWrite list. THREE THAT A NAIVE IMPLEMENTATION GETS WRONG. A JSON Patch move writes its target AND removes its source, so both ends are authorised — the source end is the one skipped because the operation "carries no value". A SPARSE array is walked by index and not by forEach, which skips holes, so [,,{ownerId}] would otherwise be authorised at the wrong path from the one the store applies. And a generated proposal goes through the same normalisation as a request body, because "we generated it" is not a provenance claim about what the prompt contained. -->
    • S4.4.c Reject unknown, immutable, server-owned, write-once, conditionally protected, and unauthorized fields before mutation. Evidence: mass-assignment corpus. <!-- Nine refusal kinds, and a test that produces every one of them from real code so none can sit in the catalogue looking like coverage. UNKNOWN IS REFUSED, NEVER DROPPED. Dropping is what makes mass assignment survivable: the request succeeds, the caller believes the field was set, and the day somebody adds a policy row the "harmless" ignored field becomes a live write. THE REFUSAL IS WHOLE-MUTATION. Applying the permitted half of a patch leaves the object in a state the caller never asked for and cannot predict — a headline changed without the summary that explained it — so one unauthorised field means nothing is applied, and the caller is told which field. -->
    • S4.4.d Protect owner, tenant, approval, gate, provenance, audit, role, lifecycle, revision, rights, and consent fields explicitly. Evidence: protected-field tests. <!-- Eleven named classes, not one protected: true flag. The flag cannot say WHICH control somebody removed, and its removal is a one-character diff nobody reads twice. The corpus runs every class through every shape. -->
    • S4.4.e Filter reads, errors, diffs, history, search snippets/facets, exports, audit, and provider context at property level. Evidence: field disclosure tests. <!-- Nine surfaces, each mapped to the permission it needs. Two mappings are the point: an ERROR needs read, because naming a field the caller may not read tells them it exists and is the cheapest field-enumeration oracle there is; and PROVIDER CONTEXT needs export, because everything sent to a third party is an export whoever wrote the integration did not call one — asserted with an internal note that is readable and not exportable. filterForSurface returns what was WITHHELD as well as what is visible, so a partial object can say "3 fields are not shown" instead of presenting a complete-looking record with holes in it. -->
    • S4.4.f Reevaluate after defaults/normalization and test aliases, duplicate keys, Unicode/case variants, magic fields, sparse arrays, and version skew. Evidence: bypass negative tests. <!-- ownerId, owner_id, OWNER_ID, Owner-Id, ownerid, OwnerID and the FULLWIDTH ownerId all reach the same policy row; the Cyrillic-о оwnerId is REFUSED rather than folded, because transliterating would make the policy decide which of two different-looking fields it was protecting. Policy is compared against the CANONICAL path — the one the store will apply — and not the text that arrived. DUPLICATE KEYS NEEDED THEIR OWN SCANNER. JSON.parse keeps the last occurrence and reports nothing, so {"ownerId":"me","ownerId":"you"} is indistinguishable after parsing from a body that only said you, while every proxy, log and validator in front of the handler may have seen the first. duplicateKeys reads the TEXT, tracks string literals and escapes so a brace inside a value cannot open a scope, and keeps a container stack so a repeated string in an ARRAY is not reported as a duplicate key — which is how a real duplicate check gets switched off. __proto__, constructor and $-prefixed keys are refused BY NAME so the refusal explains itself, and the test builds them through JSON.parse because a __proto__ key in an object literal sets the prototype instead of creating an own property. --> <!-- S4.4 rollup: 47 new cases, kit suite green; export map at 50 entry points, --check clean; tsc clean on both projects. -->
  • S4.5 Define function-level authorization for capture, edit, review, waive, publish, unpublish, purge, and administration. Evidence: role matrix contract tests. <!-- src/functions.ts + src/functions.spec.ts (18 cases over a register of all eight named acts). New ./functions subpath, 51 entry points. S4.3 and S4.4 both answer questions about a ROW. This answers one about an ACT — may you waive a gate, sign off a release, purge a tenant, flip a kill switch — and some of those touch no field at all, so a platform that only checks rows has no check on them. roleMatrix is GENERATED from decideFunction rather than authored beside it, so it cannot say something the code does not do, which is the failure mode of every permission matrix maintained in a document. It also shows the case a role-only matrix cannot express: a publisher who holds the role and is still refused, because they wrote the thing. --> <!-- S4.5 note: public-api.spec.ts's subpath-resolution case got an explicit 30s timeout. The root specifier pulls the whole barrel through the transform and crossed the 5s default once the kit passed fifty entry points; the case measures whether a subpath RESOLVES, not how fast, and letting it fail under parallel load would make a strict check flaky. --> <!-- S4.5.a–g are covered by the same module and spec: (a) capture authority with quota/rights/consent preconditions, each asserted independently; (b) edit by scope; (c) review refusing the author; (d) a waiver refusing whoever ran the gate, which is the rule every incident produces because the person best placed to waive a gate is the person who ran it; (e) publish requiring two DISTINCT approvers — "approvals >= 2" is satisfied by one person clicking twice, and the test asserts that case fails; (f) purge and kill switch as irreversible administrative acts needing a hardware key and a permitted classification; (g) below. S4.5.g IS STRUCTURAL RATHER THAN DISCIPLINED. decideFunction takes NO invocation path, so the six paths in INVOCATION_PATHS — including the compatibility shim, where the old decision usually survives, and direct-service, which skips the router entirely — cannot disagree. A decideFunction(path, …) signature would have been the obvious design and would have made the property untestable, because the thing under test would be the caller's care. UI discovery is DERIVED from the same decision: hidden for what the caller cannot change, disabled for what they can, so a menu neither teaches nothing nor lists things to file tickets about. Two more decisions with cases. A precondition nobody could answer is indeterminate and never a refusal — a caller told "no" for a quota service that was merely down goes and finds somebody to raise their quota, and the quota was never the problem. And preconditions run LAST, so an unauthorised attempt never reaches a downstream system: asking a quota service about a request that was never going to be permitted costs money and tells that service the attempt happened. validateFunctions holds the defaults a flat register loses one function at a time: a release function with no separation rule is a publish somebody can perform on their own work; an irreversible one with no session requirement is one stolen cookie away; an administration function with no scope is reachable by any authenticated session. --> <!-- S4.5 rollup: 2719 kit tests pass; tsc clean on both projects; export map at 51 entry points, --check clean. -->

    • S4.5.a Define capture/create authority by source type, target parent, object type, quota, classification, rights, and consent. Evidence: capture permission tests.
    • S4.5.b Define edit authority by revision, lifecycle, field/block class, ownership/assignment, lock, and conflict. Evidence: edit permission tests.
    • S4.5.c Define comment, suggestion, assignment, verdict, approval, abstention, recusal, escalation, and delegation authority. Evidence: review permission tests.
    • S4.5.d Define gate run/recompute, appeal, waiver/override, signoff, and policy configuration authority with separation of duties. Evidence: gate permission tests.
    • S4.5.e Define preview, publish/release, correction, supersession, unpublish, recall, and rollback authority per target. Evidence: release permission tests.
    • S4.5.f Define purge/restore, secret/key, replay, feature/config, support, kill switch, and other administrative authority. Evidence: high-risk permission tests.
    • S4.5.g Apply identical decisions to UI discovery, API, job, event, compatibility, and direct service paths. Evidence: alternate-path tests.
  • S4.6 Add step-up authorization support for high-risk mutations. Evidence: stale or missing step-up tokens are rejected. <!-- src/stepup.ts + src/stepup.spec.ts (33 cases) + a real-DB harness tools/step-up-race.mjs (28th Nx target). New ./stepup subpath, 52 entry points. A step-up challenge is the one credential-adjacent object a workbench holds, and what it is for is narrow: proving the person performing THIS act, on THIS target, with THESE parameters, is present right now. Almost every way of getting it wrong comes from making it slightly more general than that. -->

    • S4.6.a Define risk combinations requiring step-up and accepted assurance methods. Evidence: step-up policy. <!-- Risk is a COMBINATION — irreversible, classification, blast radius, externally visible — rather than a flag, because no one of them is sufficient and every pair is argued about: an irreversible act on public data is a smaller problem than a reversible one over a whole tenant. THE LADDER IS THE POINT. ASSURANCE_METHODS is ordered weakest-first with recovery-code at the bottom: a bearer secret written down somewhere is not a security key, and systems that model step-up as verified: true accept the weakest ENROLLED method for the strongest act — precisely the path an attacker holding the password takes. There is a case asserting enrolment does not weaken the requirement. When several rules apply the STRONGEST assurance and the SHORTEST lifetime win. First-match would mean inserting a rule at the top silently weakens every act below it, invisibly in the diff. -->
    • S4.6.b Create a short-lived one-time challenge bound to actor/session, tenant, exact action, target, parameters/hash, reason, and policy. Evidence: challenge schema tests. <!-- Challenge binds seven things and CHALLENGE_BINDINGS is DATA that verifyChallenge iterates, so a binding that stops being checked is a deletion from a list rather than a missing if. Each of the seven has its own rebinding case: obtain a challenge cheaply, spend it expensively — approve a small refund, reuse the token for a large one. parseChallenge refuses a credential field BY NAME (password, assertion, signature, otp) rather than with a generic unknown-field message, because the generic message is the one somebody works around. And a window that expires no later than it was issued is refused as the clock bug it is rather than treated as merely expired. -->
    • S4.6.c Complete supported WebAuthn/MFA/reauth through the canonical identity service without handling credentials in the workbench. Evidence: identity integration tests. <!-- The identity service is a seam returning an Attestation — nonce, actor, session, method name, completion time and an opaque signature the kit compares and never parses. There is no field in any type here a WebAuthn assertion, a TOTP code or a password could travel in, asserted by pinning the attestation's key set. unavailable is kept apart from failed: reporting an unreachable identity service as a failed authentication tells a user their security key was refused, which is a different thing to go and investigate. -->
    • S4.6.d Verify signature, audience, binding, session, assurance, nonce, issued/expiry time, and unused state at mutation. Evidence: tamper/stale/replay tests. <!-- Eight verification failures, every one producible, in the order signature -> bindings -> assurance -> time -> use. Signature first because an unsigned attestation says nothing about the rest; USE LAST because it is the one fact that costs a store read, and a request that was going to be rejected anyway should not cause one — asserted. completed-before-issued looks pedantic and is the replay that passes every other check: an attestation minted for an earlier, legitimate challenge, with every other field copied. -->
    • S4.6.e Consume the challenge atomically with command acceptance and reject concurrent reuse. Evidence: real-DB race test. <!-- tools/step-up-race.mjs, four properties over a real engine, in 0.9s. The claim no unit test can make: verifyChallenge is handed alreadyUsed as a boolean, and everything interesting happens in the gap between reading it and writing the command. RACE: 32 connections consume one challenge and perform the command in one transaction — 1 commits, 1 command row, and the challenge names the command that exists. ROLLBACK: the opposite failure, and the one people ship. A command that fails after consuming must leave the challenge UNUSED and spendable; burning it means the user re-authenticates for a request that never happened every time the store hiccups, which is how step-up gets removed from the flow that needed it. CONTROL: the check-then-use implementation, raced identically, let 32 of 32 through. Without it a passing race proves only that the harness ran. EXPIRY: the predicate lives in the consuming UPDATE, so a challenge that expired between verification and the write is refused by the DATABASE rather than by a clock the application read a moment earlier. -->
    • S4.6.f Audit request, completion/failure, use, expiry, and actor without exposing authentication secrets. Evidence: audit tests. <!-- Five events and an audit row with nowhere to put a secret: auditFields() is exported so the assertion is "these and no others" rather than a test asserting its own copy. The nonce IS carried, because it is the join key an incident review needs and it is safe once consumed — which is what consuming it is for. auditLeaks scans VALUES as well as keys, because a caller that puts an attestation into correlationId has broken nothing structurally and leaked it anyway — the same argument S1.10 makes about a declared key given an undeclared value. Tested with a JWT-shaped blob and with an undeclared field. --> <!-- S4.6 rollup: 2753 kit tests pass (33 new); tsc clean on both projects; export map at 52 entry points, --check clean; node tools/step-up-race.mjs green against the dev Postgres. -->
  • S4.7 Add idempotency middleware for retryable commands. Evidence: repeated keys with identical bodies replay; mismatched bodies conflict. <!-- src/retries.ts + src/retries.spec.ts (32 cases). New ./retries subpath, 53 entry points. The replay/conflict decision itself is S3.4's decideClaim and the pipeline placement is S4.1's — this is the part neither covered: what a route DECLARES about being retried, what the digest is taken OVER, and what happens when the first attempt neither succeeded nor failed. THE CASE THAT DECIDES WHETHER AN IMPLEMENTATION IS RIGHT is neither of the two the evidence clause names. It is the retry carrying one extra field the caller was never allowed to set: identical in effect, different in bytes. A digest over the raw body makes it a CONFLICT, a conflict tells the client to mint a new key, and minting a new key for a request that already ran is how a retry becomes a double charge. -->

    • S4.7.a Mark retryable routes with key requirement, scope, request canonicalizer, retention, and unknown-outcome policy. Evidence: route metadata tests. <!-- RouteRetryPolicy carries the five things the cell names, and validateRetryPolicy refuses six incoherent combinations, every one producible. The two worth reading: an irreversible effect declaring retry-safely for an unknown outcome, which is the combination that turns one network timeout into two of whatever the route does; and RETENTION TIED TO REVERSIBILITY rather than a global TTL — an irreversible record must outlive every retry a client will make, because the moment it expires the second attempt is indistinguishable from a first. A reversible one may expire in an hour, and that is asserted too. reconcile-without-a-probe catches the policy that names a step nothing can perform, which in practice becomes a retry. -->
    • S4.7.b Validate and scope keys by actor/tenant, route/contract, target, and operation version. Evidence: key validation tests. <!-- The five-field SCOPE is S3.4's validateScope, delegated rather than re-decided and asserted here so the two halves cannot drift. What this adds is the client's half — the key VALUE. The length floor is load-bearing: a short key from a client that mints them per screen rather than per request collides between two users, and the second is handed the first's result. A key over 255 characters is one somebody is using as storage, and a key derived from an email or a URL leaks that value into every log line and support ticket that quotes it back. -->
    • S4.7.c Canonicalize authorized parameters and persist request hash, state, and result reference atomically with acceptance. Evidence: integration tests. <!-- canonicalizeAuthorized takes the writes S4.4 PERMITTED and nothing else. Stable across key order, across nested object key order, and across the spelling of a field — headline and HEADLINE are one field, so a client that changed its serialiser has not changed its request. Separated by control characters that cannot occur in a canonical path, an operation name or JSON, with a case proving {headline:"x set y"} and two fields do not collide. The load-bearing case is the last one: an unauthorized field cannot change the digest, so the retry that carries it is still a replay rather than the conflict that poisons the key. -->
    • S4.7.d Replay prior accepted, running, success, or terminal outcome for identical retries without invoking the handler. Evidence: concurrent duplicate tests. <!-- Every state replays or re-claims correctly: completed replays with its result reference, claimed and running report already-running rather than executing a second time or returning null, and FAILED is a genuine retry rather than a duplicate — the record is re-claimed. The handler is not invoked in any of them, which the router's own case asserts by counting handler calls. -->
    • S4.7.e Return a typed conflict for changed body/query/target/identity/ version and do not reserve keys for prevalidation failures. Evidence: mismatch tests. <!-- A changed digest conflicts, and the digest is compared BEFORE the state — asserted across all four states, because every other branch returns a result and returning the stored one would report success for something that never ran to a caller who will never look again. An expired record for an irreversible operation refuses rather than re-claiming. "Do not reserve keys for prevalidation failures" is STRUCTURAL: the router runs validate and object-authorize before idempotency, so a request failing either never reaches the claim. router.spec.ts asserts it by counting record lookups — 0 for a request authorization refuses. -->
    • S4.7.f Reconcile timeout/connection-loss outcomes before repeating irreversible provider or publish effects. Evidence: fault tests. <!-- The unknown outcome is a third answer and the dangerous one. A connection lost after the request left is neither a success nor a failure, and both obvious policies are wrong for some route: retrying repeats an irreversible effect, refusing strands a caller who was fine. The table has four cells and one of them is the whole item: a reconciliation that comes back INDETERMINATE refuses and asks for a person. That is the cell people fill in with "retry" because the alternative is inconvenient, and it is the cell that produces the duplicate charge. The probe is keyed by the IDEMPOTENCY KEY and not by any local id, because after a connection loss the local id may never have been written — which is the state being reconciled. Asserted by capturing what the probe was asked. --> <!-- S4.7 rollup: 2786 kit tests pass (32 new); tsc clean on both projects; export map at 53 entry points, --check clean. -->
  • S4.8 Add optimistic-concurrency preconditions. Evidence: stale revisions return 409 or 412 with merge context. <!-- src/preconditions.ts + src/preconditions.spec.ts (18 cases) plus three interaction cases appended to router.spec.ts. New ./preconditions subpath, 54 entry points. The DECISION is S2.3's applyPrecondition — one total function every path runs — and the atomic write is already proved against a real engine by tools/revision-race.mjs. This is everything between an HTTP request and that function, which S2.3 deliberately did not reach. -->

    • S4.8.a Declare expected-revision requirement and accepted field/header per mutable route. Evidence: route contracts. <!-- RoutePrecondition declares the path, whether it is required, the CARRIER (an if-match header or a body field) and the field name, plus perItem for a multi-subject request. preconditionGaps refuses a multi-subject route carrying one revision for many subjects: a single revision can only describe one of them and the rest are unchecked. -->
    • S4.8.b Validate expected revision against authorized current object and reject missing/malformed conditions. Evidence: input tests. <!-- evaluateAgainstAuthorized takes S4.3's AuthorizedLoad and REFUSES TO WORK FROM ANYTHING ELSE, which is the signature doing the arguing. There is no way to call it with a revision from a second query — and a second query is the one somebody writes without the tenant predicate, so the object whose revision decided the conflict would not be the object authorization approved. It also closes the narrower hole: between two reads the row can change, and the precondition would be compared against a revision nobody was authorized to see. Cases cover a concealed object and a load for a DIFFERENT entity. A missing condition is REFUSED, never defaulted: no if-match: *, no last-write-wins fallback, because a default is last-write-wins with extra steps and it is off for whoever forgets the header. -->
    • S4.8.c Pass the precondition to an atomic conditional persistence write. Evidence: real-DB race test. <!-- tools/revision-race.mjs (S2.3.b), re-run green on this branch against PostgreSQL 16.14: 3 concurrent connections, atomic UPDATE and INSERT each producing exactly one winner with no lost update, and BOTH read-then-write controls losing one — so the harness is known to be capable of failing. Nothing new was needed here; what was needed was to confirm the claim still holds, which the execution contract requires of existing code used as evidence. -->
    • S4.8.d Return 409/412, safe current/expected revision, conflict reason, compare/reload action, and correlation. Evidence: response fixtures. <!-- 409 when the object is there and has moved on; 412 when the condition itself could not be evaluated. The distinction is not decoration: a 409 says "reload and compare" and a client can act on it, while a client treating a 412 the same way would reload forever against a request that will never be conditional. subject-absent is the one conflict kind that is a 412 and not retryable — telling a user to reload an object that has been deleted sends them somewhere that will 404. The envelope is S2.3's conflictEnvelope, so the conflict inherits the canary scan and the declared-key rule; the response asserts both revisions, the mutation path, the reload-and-compare remediation, the correlation id and the preserved intent token — which is what makes it different from a bare 409, because the client can offer compare, retry and copy instead of "your changes were lost". Both authorization-shaped input faults produce the SAME public answer as a malformed condition, so a caller cannot learn the object exists. -->
    • S4.8.e Apply to edit, transition, restore, review, waiver, bulk, import, job, and publish paths. Evidence: coverage matrix. <!-- PRECONDITION_PATHS is the item's NINE and deliberately not S2.3's seven MutationPaths, which are shaped by how storage sees a write: a review, a waiver and a publish are all single there and are entirely different things to leave unprotected. storagePathFor maps between them, so the coverage question is asked in the vocabulary a reviewer thinks in while the decision stays in the one the engine needs. The gap check runs in the direction that matters — not "does every declared route have a precondition" but "is there a path nothing declares one for", which is invisible to any check iterating the routes that exist. -->
    • S4.8.f Verify idempotent retry replays an accepted result while a new stale intent conflicts. Evidence: interaction tests. <!-- Three cases at the router, where the two steps actually meet. A RETRY of an accepted request replays even though the object has since moved to a newer revision: it already happened, and re-deciding it against a revision that changed after it succeeded would refuse a request that cannot be refused. The trail proves it never reached the concurrency step. A NEW intent carrying the same stale revision conflicts, running the full prefix through concurrency. And a retry whose DIGEST differs conflicts on the key BEFORE the precondition is consulted, so an accepted key cannot be reused to smuggle a different mutation past a stale revision. --> <!-- S4.8 rollup: 2808 kit tests pass (18 + 3 new); tsc clean on both projects; export map at 54 entry points, --check clean; node tools/revision-race.mjs green against PostgreSQL 16.14. -->
  • S4.9 Add deterministic cursor pagination. Evidence: no duplicate or missing records while concurrent inserts occur. <!-- src/cursors.ts + src/cursors.spec.ts (40 cases). New ./cursors subpath, 55 entry points. indexes.ts (S3.x) already refuses offset as a pagination MODE and holds sortIsTotal; this is the cursor itself. THE CASE THAT ACTUALLY DECIDES IT IS NOT THE CONCURRENT INSERT. It is the TIE — a hundred rows sharing one updatedAt, which is what an import produces — with the page boundary falling inside it. The property runs over 300 rows across 3 distinct timestamps, at seven page sizes including 1 and n+1, forward and backward. -->

    • S4.9.a Define default/maximum limit and stable default/allowed sorts with unique tie-breakers per collection. Evidence: pagination registry. <!-- validateCollection refuses seven incoherent declarations, every one producible. Two are the item: a sort whose last key is not unique within scope, and a unique key that is NOT LAST — which looks like pedantry and is not, because every key after a unique one is dead (the rows are already totally ordered by then), so [id, updatedAt] is [id] with a decoration a reviewer believes is doing something. The two are reported as DIFFERENT faults with different fixes: telling somebody to add a tie-breaker that is already there sends them the wrong way. A requested limit is CLAMPED rather than refused. -->
    • S4.9.b Encode sort, filter, direction, position, version, and safe consistency context in an opaque protected cursor. Evidence: cursor tests. <!-- CursorSeal is a seam the host implements with authenticated encryption. The kit will not offer base64 as a substitute: a client that decodes a cursor and edits the position is a client reading rows by editing a string, and the test asserts the seal REFUSES a token it did not produce rather than decoding one and finding it wrong. The payload carries sort, direction, position (VALUES, not an index — an index describes a position in a result set that no longer exists the moment anything is inserted before it), version, query digest and an opaque consistency marker. parseCursorPosition refuses anything but scalars: a row fragment travelling in a token is one the next caller may not be allowed to read. -->
    • S4.9.c Bind cursors to actor/tenant scope, route/contract, normalized query, and lifetime. Evidence: misuse tests. <!-- Nine rejections, every one producible: tenant, actor, collection, route, contract version, cursor version, sort, expiry and query digest. THE QUERY DIGEST IS CHECKED LAST, deliberately and with its own case. "Your cursor is for a different query" tells a caller their cursor WAS valid for something, which is worth withholding from somebody holding a token that is not theirs — so the scope bindings answer first. Without the digest at all, the second page of a search for status=draft becomes the second page of a search for everything, starting at a position that means nothing there. An EMPTY query gets a digest of its own rather than a null, or a cursor for it would match any query. -->
    • S4.9.d Query through authorization using strict before/after keyset conditions rather than offset. Evidence: query plan and disclosure tests. <!-- keysetCondition returns a described ROW-VALUE comparison — (a, b) > (?, ?) is one index range scan and the OR expansion is several, only one of which the planner will use — and it is STRICT always, because a >= returns the row the previous page ended on, which is the duplicate the whole scheme exists to avoid and the mistake made when somebody "fixes" a missing row that was actually a tie. A BACKWARD PAGE IS THE FORWARD QUERY WITH EVERY DIRECTION FLIPPED and the results reversed afterwards, not a separate query with < swapped in by hand — which is where the two directions drift and one starts skipping ties. Asserted, along with the absence of any offset in the condition's own shape. -->
    • S4.9.e Define concurrent insert/update/delete/tombstone/permission-change behavior and any snapshot guarantee. Evidence: scenario tests. <!-- Seven concurrent events, each mapped to what the traversal promises, with and without a snapshot. THE HONEST ADMISSIONS ARE THE UPDATE ROWS: a row that moves ACROSS the cursor position mid-traversal can be missed or repeated, and no keyset scheme prevents that without a snapshot. Claiming otherwise would be the fabricated guarantee. What it DOES promise is that a row nobody touched is returned exactly once, which offset cannot promise even with one. And a snapshot fixes the data and never the permissions: a row the caller lost access to mid-traversal stops being returned either way, which is the one cell that does not move. -->
    • S4.9.f Prove forward/backward traversal has no duplicates or missing items within the declared model across equal sort values. Evidence: property tests. <!-- The traversal runs to completion over 300 rows across 3 distinct sort values — a hundred-way tie at every position — at page sizes 1, 2, 7, 25, 299, 300 and 301, forward and backward, with zero duplicates and zero missing. Reversing the backward PAGES (not the rows) reproduces the forward order exactly; the page boundaries differ between directions and that is correct, since the remainder is at the start one way and the end the other. Then the evidence clause literally: a row inserted after the cursor between every page, still zero and zero. AND THE CONTROL. The offset implementation, raced identically with one insert BEFORE the cursor per page, is asserted to duplicate or skip. Without it a green traversal proves only that the harness ran. --> <!-- S4.9 rollup: 2849 kit tests pass (40 new); tsc clean on both projects; export map at 55 entry points, --check clean. -->
  • S4.10 Add payload, upload, query-complexity, rate, concurrency, and spend limits. Evidence: limit tests return actionable retry metadata. <!-- src/limits.ts + src/limits.spec.ts (33 cases). New ./limits subpath, 56 entry points. S4.1 put the two-phase LimitSeam in the pipeline; this is the policy behind it. EVERY LIMIT HERE HAS A BYPASS, AND THE BYPASSES ARE WHY THIS IS A MODULE RATHER THAN A FEW NUMBERS. A suite that sends one oversized body from one identity in one tenant confirms the numbers are wired up and misses all six of them. -->

    • S4.10.a Enforce compressed/decompressed body, field, array, batch, response, and archive limits before expensive parsing. Evidence: payload boundary tests. <!-- The wire cap, the DECOMPRESSED cap and the RATIO are three limits, not one. A 4 MB body cap enforced on the wire is not a cap against Content-Encoding — a few hundred kilobytes of zeroes expands to a gigabyte — and a body inside BOTH byte caps that expands 3000:1 is still a bomb aimed at a smaller target, which is its own case. Every breach is reported rather than the first, because a client fixing one and resubmitting to discover the next is making the problem worse one round trip at a time. -->
    • S4.10.b Enforce upload bytes, parts, files, dimensions, duration, frames/pages, nesting, and decompression ratio. Evidence: upload limit corpus. <!-- PIXELS is the limit people leave out and the one that decides an image endpoint: a 64,000 x 64,000 PNG of one colour is twelve kilobytes on the wire and gigabytes once decoded, so every byte limit passes and the decoder is the target. Asserted as its own case with a 12 KB upload. declared-size-mismatch is not a size limit at all — it is the check that the declared and received lengths agree, because a limit enforced against a DECLARED size is a limit against a number the client chose. Eight breach kinds, every one producible. -->
    • S4.10.c Score and limit filter depth, predicates, joins, expansion, requested fields, sort, page size, graph traversal, and search cost. Evidence: query complexity tests. <!-- Scored from the query's own shape BEFORE it runs, because a complexity budget checked afterwards is a timeout with extra steps: the work is done, the connection was held, and the limit only threw the result away. COMPLEXITY_WEIGHTS is DECLARED so a refusal can be explained — "your query scored 248" is not an explanation — and the verdict names its drivers largest-first. Joins and expansions outweigh predicates because both multiply rows; graph traversal outweighs both because its cost is exponential in depth rather than linear. -->
    • S4.10.d Enforce sustained/burst request rate by actor, tenant, route, client class, and trust tier. Evidence: rate-limit tests. <!-- Rules are keyed by DIMENSION and BUCKET, and consume charges every bucket a request falls into rather than the one it names. The tenant bucket is checked first and its breach reported in preference, because a tenant breach is the one another caller in the same tenant can do nothing about and is therefore the more useful thing to be told. -->
    • S4.10.e Enforce concurrent requests, streams, uploads, jobs, provider calls, and expensive operations by actor/tenant/resource. Evidence: concurrency tests. <!-- Concurrency dimensions carry a window of zero — they are instantaneous, not windowed — and a breach still returns a retry-after of one second, because no reset exists to name and "try again when something finishes" is the honest answer. Five concurrency dimensions are declared separately (requests, streams, uploads, jobs, provider calls) so dimension-with-no-rule can find the ones nothing limits. -->
    • S4.10.f Estimate/reserve and enforce per-operation, actor, tenant, period, provider, and exceptional spend budgets. Evidence: spend tests. <!-- RESERVE BEFORE THE WORK, SETTLE AFTER. Charging afterwards is a budget that can only be exceeded and the amount is unbounded — one runaway job spends a month's allowance and the limit reports it afterwards. Reservations count against the ceiling, or two concurrent expensive operations each see room for one. An operation whose cost cannot be ESTIMATED is unestimable and blocks; treating it as zero is the shape of every "we could not price it so we ran it". An exception carries an expiry, because an exceptional allowance with none is a permanent one somebody described as exceptional. And settle REPORTS the overrun rather than absorbing it: a reservation regularly under the actual is an estimator wrong in the one direction that matters. -->
    • S4.10.g Return stable limit category, safe current/max where permitted, reset/retry time, remediation, and correlation. Evidence: response fixtures. <!-- current and maximum are disclosed for rate, payload, upload, complexity and concurrency, and WITHHELD for spend. The asymmetry is deliberate: a rate ceiling is a property of the API and telling a client makes them a better client, while a spend ceiling is a commercial fact about the tenant that the person hitting it may not be entitled to know. retryAfterSeconds is present wherever a reset exists, because a limit with no reset time is one every client guesses at — and the ones that guess badly are indistinguishable from an attack. parsePositiveLimit refuses zero and infinity, which are usually a unit conversion that went wrong and are both silent: one refuses everything and the other limits nothing. -->
    • S4.10.h Prevent identity rotation, multiple clients, retries, or batch subdivision from bypassing tenant/actor limits. Evidence: abuse tests. <!-- Four bypass findings over a RULE SET rather than over traffic, because none of them shows up in a test that sends requests from one identity in one tenant. no-tenant-bucket: ten identities in one tenant get ten times the allowance, and minting identities is cheaper than the limit. no-actor-bucket: one caller consumes everybody else's allowance and the tenant experiences it as an outage. dimension-with-no-rule: the dimension nothing limits at all — the test finds concurrent-uploads and concurrent-provider-calls unlimited in a plausible-looking rule set. cost-not-charged-per-item: a request costs one whatever it carries, so one request split into ten batches costs the same as one and the limit is a limit on round trips rather than on work. consume charges a batch of fifty as fifty, asserted. --> <!-- S4.10 rollup: 2883 kit tests pass (33 new); tsc clean on both projects; export map at 56 entry points, --check clean. -->
  • S4.11 Add guarded remote-resource fetching. Evidence: SSRF, redirect, private-IP, DNS-rebinding, oversize, timeout, and decompression tests. <!-- src/fetching.ts + src/fetching.spec.ts (27 cases). New ./fetching subpath, 57 entry points. THE REPOSITORY ALREADY HAS A REAL SSRF GUARD — @oshun/http-client's assertUrlAllowed / classifyIp / makeSsrfSafeLookup — and it is layer:infra, which layer:workbench-kit may not depend on. Same wall S4.2 hit with @oshun/auth, and it points the same way: a second address classifier here would be the duplicate this ledger exists to avoid, and it would be the WORSE of the two because it is the one nobody runs against real traffic. So this module classifies NO address. It owns the policy, the seam, and a CORPUS — and fetching.spec.ts points that corpus at the real guard the way S4.2's parity matrix points at the real role taxonomy. The guard refuses all 21 targets, including every encoded bypass. -->

    • S4.11.a Canonicalize scheme, credentials, IDNA/host, port, IP literal, path, and fragment before policy. Evidence: URL parser fixtures. <!-- Not a URL parser — the canonical guard has one and it is better placed. What is here is the POLICY the parser is run under: schemes, ports, redirect handling, the six timeouts, the byte and archive bounds, the accepted media types, and two booleans about credentials. validateFetchPolicy refuses ten incoherent combinations. -->
    • S4.11.b Allowlist schemes/ports and block loopback, link-local, private, reserved, multicast, metadata, internal service, file, and unix targets. Evidence: address negative tests. <!-- SSRF_CORPUS: 21 targets, each with the reason it exists, run against @oshun/http-client in the spec. Loopback, link-local, private, reserved, multicast, four cloud metadata endpoints, credentials in the authority — and the ENCODED BYPASSES, which are the interesting half: 127.1, the decimal 2130706433, the octal 0177.0.0.1, and ::ffff:169.254.169.254, where an IPv6 classifier that does not unwrap the mapping applies IPv6 rules to an IPv4 address and the metadata service is an IPv4 address. Two controls make the green run mean something: a guard that refuses NOTHING is reported in full, and an ordinary public target is asserted still permitted — a guard that refuses everything passes every negative test and is useless. internal-service is deliberately EMPTY with a stated reason: a deployment's own service names are its highest-value targets and the kit knows none of them, so inventing plausible ones would produce a corpus that passes everywhere and protects nowhere. extendCorpus is how a deployment adds its own, and it refuses an entry with no reason. -->
    • S4.11.c Resolve through controlled DNS, validate every address, pin the connection target, and revalidate on reconnect. Evidence: DNS-rebinding tests. <!-- connectionWasPinned compares the address the guard VALIDATED against the address the socket CONNECTED to. That comparison is the whole of DNS rebinding: the guard resolves a name and approves the answer, the HTTP client resolves the name again, and the record's TTL was one second because the attacker owns the zone. The only defence is to connect to the address that was checked, and the only way to know it happened is to compare them. -->
    • S4.11.d Disable redirects or validate every hop/target/scheme/port/ credential with a strict maximum. Evidence: redirect escape tests. <!-- A redirect limit that contradicts the redirect policy is refused in BOTH directions — redirects refused with a limit above zero, and validated with a limit of zero — because one of the two is what somebody will believe. A chain longer than ten hops is refused on the grounds that nobody is reading them and every real target is reachable in fewer. FetchHop carries the resolved address per hop, so a redirect chain is auditable after the fact rather than only bounded. -->
    • S4.11.e Enforce connection/TLS/header/idle/body/total timeout, bandwidth, response, retry, and content limits. Evidence: slow/oversize tests. <!-- Six timeouts, and the TOTAL is required separately with its own case: the phase timeouts cannot substitute for it, because a server that sends one byte before each idle timeout holds the connection for ever without ever being idle. A phase timeout above the total is refused too — it can never fire, so the total is the only one that ever does, and the phase a reviewer believes is bounding something is bounding nothing. -->
    • S4.11.f Bound compressed/decompressed bytes, archive entries/depth, dimensions, frames, duration, and parser resources. Evidence: bomb fixtures. <!-- Compressed bytes, decompressed bytes, expansion ratio, archive entries and archive depth, with decideContent refusing an oversize response and an over-expanding one separately. The ratio matters on its own: a response inside both byte caps that expands ninety-fold is still a bomb aimed at the fetcher. -->
    • S4.11.g Sniff content, quarantine mismatch/malicious media, isolate parsing, and prevent privileged interpretation of fetched text. Evidence: polyglot/injection tests. <!-- The SNIFFED type decides, not the declared one, because the declared type is the far side's claim — asserted with a response labelled text/markdown whose bytes are an ELF binary. A MISMATCH IS QUARANTINED RATHER THAN REFUSED OR TRUSTED. It is often innocent (a server that labels everything application/octet-stream) and it is also exactly what a polyglot file looks like, so it is isolated and reported instead of either being handed to a parser or thrown away with the evidence. -->
    • S4.11.h Strip outbound credentials and redact signed URLs/protected locations from logs and public errors. Evidence: leakage tests. <!-- withoutUserinfo removes the OUTER authority's credentials and leaves an inner URL alone — asserted with https://example.com/?next=https://a:b@c, which a naive pattern eats, changing the request it was meant to sanitise. redactUrl replaces the WHOLE query rather than known parameter names: a per-parameter redaction has to know every provider's parameter, and the one it does not know is the one that goes out. The path is kept, which is what somebody debugging actually needs. Unparseable input returns (unparseable url) rather than throwing, because a redactor that throws is one somebody wraps in a try/catch that logs the original. --> <!-- S4.11 rollup: 2911 kit tests pass (27 new); tsc clean on both projects; export map at 57 entry points, --check clean. @oshun/http-client is a DEV dependency and a spec-only vitest alias, so the emitted package does not carry it — same arrangement as S4.2's @oshun/studio-authoring. -->
  • S4.12 Add audit capture for every mutation outcome, including refusals and partial failures. Evidence: route inventory has no mutation without an audit assertion. <!-- src/mutation-audit.ts + src/mutation-audit.spec.ts (28 cases). New ./mutation-audit subpath, 58 entry points. S2.14 built the LIFECYCLE audit — fifteen kinds from the stage engine, the gates and the signoffs, folded into a projection. This is the other one, and the evidence clause says what makes it different: "route inventory has no mutation without an audit assertion". The subject is COVERAGE, not shape. -->

    • S4.12.a Derive the mutation route/command/job inventory and assign required audit event types. Evidence: route-to-audit matrix. <!-- auditCoverage runs from the INVENTORY toward the assertions, and the direction is the whole design. Asking "does every audit row have a route" finds nothing; asking "does every mutation have an assertion" finds the route somebody added last month. Four surfaces — route, command, job, event handler — because the last two are the ones that never appear in a route table and mutate anyway. Six gap kinds, every one producible, including the one that inflates a coverage number: an assertion for a mutation that no longer exists, which passes against nothing and is counted. -->
    • S4.12.b Capture actor/delegate, tenant, target, action, authorized parameters/hash, prior/new revision, reason, policy, correlation, and time. Evidence: audit schema assertions. <!-- Actor with its delegate half, tenant, target, action, the parameter DIGEST, reason key, policy version, correlation and time. THE DIGEST AND NOT THE PARAMETERS. It answers "was this the same request" without answering "what did it say", which is the question an audit trail must not answer to whoever reads it — and it is the same digest S4.7 takes, so a retry and its original are visibly one act. priorRevision and newRevision are both required and non-null on a mutation record: a record of a change that cannot say what it changed FROM is one nobody can reconstruct a history out of. -->
    • S4.12.c Persist accepted/completed audit atomically with authoritative mutation or guaranteed outbox intent. Evidence: commit/rollback tests. <!-- Two lawful designs and one that resembles both. assessPersistence accepts the audit row in the mutation transaction, and accepts an outbox INTENT in that transaction delivered afterwards — which is what survives an audit sink being down. It refuses the third: writing the audit row after the transaction commits, which is what "we log after we save" produces, where every crash between the two leaves a mutation nobody can attribute and the window is exactly as long as the audit write takes. An outbox intent written OUTSIDE the transaction is refused as the same design with an extra table in it. The outbox half is proved against a real engine by tools/outbox-atomicity.mjs, re-run green on this branch: one transaction left both tables unchanged, and the negative control — the same failure across two transactions — produced the orphaned authoritative row, which is what makes the first result mean anything. -->
    • S4.12.d Record validation, authentication, authorization, policy, conflict, limit, and dependency refusals without claiming mutation. Evidence: refusal audit tests. <!-- TWO RECORD TYPES, NOT ONE WITH A NULLABLE FIELD. The refusal record has no newRevision at all — not "we leave it null", there is nowhere to put one — because a single type with an optional revision is how a refused request ends up in a report of what changed. Asserted both ways: the shape has no such key, and a record that acquired one through an untyped boundary is reported as refusal-claiming-a-mutation. Seven refusal classes covering validation, authentication, authorization, policy, conflict, limit and dependency, each carrying the stable code and never a message. -->
    • S4.12.e Record started/running/partial/completed/failed/cancelled/ compensated/unknown outcomes and per-item bulk results. Evidence: asynchronous/partial audit tests. <!-- Ten outcomes including compensated and unknown. THE unknown MEMBER IS THE POINT: a worker that died mid-flight has an outcome nobody knows, and leaving the member out does not remove the state — it records it as failed or completed, one of which is a lie in every case, and S4.7's reconciliation exists because it is real. summarizeBulk has NO THRESHOLD: one failure in ten thousand is partial, asserted at exactly that scale, because the person whose row failed is not helped by a percentage and "the import succeeded" is the report this prevents. -->
    • S4.12.f Redact content/secrets/protected data, enforce tenant/property access, and preserve immutable integrity/retention. Evidence: audit security tests. <!-- AUDIT_RECORD_FIELDS is exported so the assertion is "these and no others" rather than a copy of the list beside the list, and mutationAuditLeaks scans VALUES as well as keys: a caller who puts the record's content into action has broken nothing structurally and leaked it anyway. Prose in an audit field is flagged because the audit trail is read by more people than the record is. The parameter digest is exempted by name — 64 hex characters would otherwise trip the credential pattern, and being an unreadable identifier is its whole purpose. Retention refuses two things: a trail that can be UPDATED, which is a record of what somebody last wanted it to say; and refusals kept longer than mutations, which leaves an investigation into a breach finding every failed attempt and none of the successful ones. -->
    • S4.12.g Fail the inventory check when a mutation lacks an outcome audit assertion. Evidence: deliberate uncovered-route failure. <!-- The negative control, and the load-bearing case of the whole item: a deliberately uncovered route that the gate must fail on, asserted down to the single gap and its mutation id. A coverage gate that has never failed is a gate nobody has proved is connected to anything, and every other assertion in this file is conditional on this one. --> <!-- S4.12 rollup: 2940 kit tests pass (28 new); tsc clean on both projects; export map at 58 entry points, --check clean; node tools/outbox-atomicity.mjs green with its negative control. -->
  • S4.13 Add request/trace correlation to audit and job IDs. Evidence: one journey can be reconstructed across web, BFF, worker, storage, and publisher. <!-- src/correlation.ts + src/correlation.spec.ts (22 cases). New ./correlation subpath, 59 entry points. WHAT MAKES THE EVIDENCE CLAUSE HARD IS NOT THE PLUMBING. A journey is reconstructable only if every hop wrote down the SAME identifier, and the hops that forget are always the asynchronous ones — the job attempt, the outbox delivery, the provider callback — because those are where the request that started it has already returned. -->

    • S4.13.a Generate or validate request and W3C trace context at the browser or trusted ingress without accepting spoofed privileged correlation. Evidence: ingress tests. <!-- THE ASYMMETRY IS THE RULE: an inbound TRACE is continued and an inbound CORRELATION is refused. A trace id is a join key with no authority attached and the worst a forged one does is join two unrelated traces; a correlation is what support SEARCHES BY, so accepting a claimed one lets a caller file their request under another tenant's reference and the person who searches it is shown somebody else's request. A trusted internal ingress may supply one; a browser may not. The refusal is RECORDED rather than silent, because a client that keeps sending one is a client somebody should talk to. parseTraceparent refuses an all-zero trace id by specification and by argument: accepting it makes every request from every tenant appear to be the same trace, which is worse than no tracing because it looks like tracing. A malformed inbound trace does not refuse the request — it is a reason not to CONTINUE a trace that was never valid. -->
    • S4.13.b Create command/idempotency and audit IDs and link them to request, trace, actor, tenant, target, and revision. Evidence: mutation correlation tests. <!-- Thirteen link kinds, each node naming its parent plus trace, correlation, tenant, target and revision. linkFaults reports eight ways a set of nodes fails to be one journey, every one producible, including a CYCLE — reported rather than walked, with a bound, so a bad graph is a failing check rather than a hanging one. -->
    • S4.13.c Propagate correlation through database transaction/outbox, event, inbox, job parent/child/attempt, blob, provider, and external operation. Evidence: async propagation tests. <!-- The asynchronous kinds are named as a set, and a node of one of them with NO parent at all is its own fault kind: the request that produced it has already returned, which is exactly why this is the hop that forgets. A dangling parent is the hop that will simply be absent from the reconstruction, and nobody finds out until they try — usually during an incident. -->
    • S4.13.d Propagate allowed trace/correlation through publisher, callback/webhook, desktop/bridge, and read-back verification. Evidence: boundary integration tests. <!-- Publish, callback, bridge-call and read-back are link kinds, so a boundary crossing is a node in the same graph rather than a separate story. The tenant check across a parent edge looks redundant beside the correlation check and is not: a node that inherited the right correlation and the WRONG tenant is a row returned to whoever searches that correlation, and they are in the other tenant. -->
    • S4.13.e Index support-safe IDs for reconstruction without using raw high-cardinality IDs as metric labels. Evidence: observability schema tests. <!-- Checked in BOTH directions, which is the part usually missed. An id used as a metric LABEL is one series per request: the scrape gets slower every day, the dashboards stop loading, and the fix is to delete the metric — taking the signal it was added for with it. And an id NOT in the support index is the opposite failure: the journey is fully recorded and unfindable, which fails "reconstructable" just as completely as not recording it. A secret in the support index is refused too, since indexing it means storing it. -->
    • S4.13.f Reconstruct a complete web-to-publisher journey with retries, partial failure, and compensation from stored evidence. Evidence: correlation acceptance report. <!-- A fourteen-node journey across the five tiers the clause names, with a failed first attempt, a RETRY, a provider call, a publish, a read-back and a compensation, rebuilt from the stored nodes alone. A retry appears as two ATTEMPTS OF ONE JOB rather than two jobs — asserted through attemptsByJob and the per-attempt ordinal — because three attempts of one job is a retry story and three jobs is a different and much worse one. Steps are ordered by TIME and carry DEPTH, because the two answer different questions: time is what happened next, depth is what caused what, and a reconstruction with only one is either a list nobody can read as a tree or a tree nobody can read as a sequence. And it REFUSES to reconstruct from nodes that do not form one journey, rather than returning a partial one nobody can tell is partial. --> <!-- S4.13 rollup: 2963 kit tests pass (22 new); tsc clean on both projects; export map at 59 entry points, --check clean. -->
  • S4.14 Add explicit capability and dependency-health endpoints. Evidence: client states derive from real probes, not build-time guesses. <!-- src/health.ts + src/health.spec.ts (29 cases). New ./health subpath, 60 entry points. The pieces existed — S1.6's five states, S3.14's probes, S1.11's discovery document. What was missing is the ENDPOINT, which is where the decisions none of the pieces can make alone get made. THE EVIDENCE CLAUSE IS MADE STRUCTURAL RATHER THAN ASPIRATIONAL. deriveCapabilityState(capability, probes, missingConfiguration) has no parameter a build-time guess could enter through — asserted on its arity — so a fast path cannot be added without editing the signature. -->

    • S4.14.a Define authenticated capability discovery by workbench/domain, contract/version, audience, host, feature/config state, and lifecycle. Evidence: capability endpoint contract. <!-- discoveryFaults over rows carrying workbench, domain, contract version, audiences, host, lifecycle, live flag values and state. unauthenticated-discovery is the one worth arguing about: a capability list is a map of what this deployment can do and which tenants have which features enabled, so served openly it is a reconnaissance endpoint with an OpenAPI-shaped response. And a row reporting configured that nothing probed is reported as a build-time guess wearing a runtime answer's clothes. -->
    • S4.14.b Implement real database/schema, blob, queue/worker, cache/search, provider/model, signing/trust, delivery/publisher, and bridge/host probes as applicable. Evidence: probe inventory. <!-- Fourteen dependency kinds — database, schema, blob store, queue, worker, cache, search, provider, model, signing, trust store, delivery, publisher, bridge host — so a probe inventory has a vocabulary to be complete against rather than a list somebody extends when they remember. The kit does not implement the probes: they reach real systems, which is the host's, and S3.14 already holds the rules a database probe has to obey. What is here is the shape their results arrive in. -->
    • S4.14.c Return configured, not-configured, degraded, unavailable, or failed state with checked time, freshness, affected/retained capability, safe reason, and remediation. Evidence: state fixtures. <!-- deriveCapabilityState returns S1.6's five states with the message key, the affected dependencies and the retained behaviour. S1.6's degraded distinction is HELD rather than restated: nothing retained is unavailable and not degraded, so a client is never told part of something works when none of it does. And a capability nothing probed is unavailable rather than configured, which is the same fabricated-success refusal the rest of the kit makes. -->
    • S4.14.d Separate liveness/readiness from user-facing dependency state and avoid probing through build-time environment assumptions. Evidence: runtime configuration-change tests. <!-- Four endpoint kinds with endpointFaults keeping them apart. liveness-depends-on-a-dependency is S3.14's rule restated where it is actually violated: a liveness probe that checks the database restarts a healthy process every time the database is briefly slow, and the restart makes the database slower. readiness-with-no-dependencies is the opposite — it answers yes while the process cannot serve a request. The build-time half is the derivation's signature (see the parent). -->
    • S4.14.e Authorize and redact probe details so ordinary clients receive no secret, topology, internal host, protected tenant, or exploit detail. Evidence: role response tests. <!-- Three audiences, and the client view is CONSTRUCTED rather than filtered. Deleting fields from a richer object is a blocklist, and a blocklist is wrong the first time somebody adds a field — the leak then arrives with a feature that had nothing to do with health. The client shape has four fields and nowhere to put a hostname, a driver message, a query, a version or a duration; the test asserts the key set and that 10.0.4.19 and SELECT 1 are absent from the serialised view. healthDisclosureFaults covers what the shape cannot see: a message key or capability name somebody built by interpolating a hostname, a credential or a driver error into it. -->
    • S4.14.f Add timeout/cache/staleness/circuit behavior and distinguish probe failure from dependency failure. Evidence: fault tests. <!-- Freshness is THREE-VALUED because stale is the one that gets served. Under load a cached result is the right thing to return; what is not right is returning it as though it had just been taken, and a two-valued answer forces either that or withholding it entirely. An EXPIRED cache derives unavailable rather than its last answer, because serving the last known good state is how a dependency that has been down for an hour reads as healthy. And unhealthy (the dependency answered that it is broken) is kept apart from unreachable (the probe did not complete). The two have opposite retry semantics, and reporting the second as the first pages somebody about a healthy database. -->
    • S4.14.g Make UI capability banners and controls consume these endpoints exclusively. Evidence: browser network assertion. <!-- controlFor(view) takes a ClientHealthView and NOTHING ELSE, asserted on its arity. "Consume these endpoints exclusively" is therefore a property of the type rather than a rule somebody follows: there is no parameter a build-time flag could enter through. BANNER_FOR_STATE is total over the five states, so a sixth breaks every consumer at compile time rather than falling into a default: and rendering nothing. degraded is a WARNING and not blocking, which is the whole reason the state exists — a blocking banner over a partially working capability stops the part that does. not_configured is informational rather than a failure, because nothing is broken and telling a user their system is sends them to the wrong person. And a STALE answer still enables a working capability, since refusing to act on one is how a cache miss becomes an outage. THE CHECKBOX STAYS UNCHECKED. The item's named evidence is a BROWSER NETWORK ASSERTION, and no application mounts these endpoints yet — the contract they would be mounted against is what this commit delivers. The controlFor signature makes "exclusively" true by construction for any surface that uses it, and that is a different claim from having watched a browser make no other request. Under §2.1 a checkbox flips only when the evidence NAMED by the item exists, so this one waits for the app work, and S4.14 waits with it. --> <!-- S4.14 status: a–f are built and green (2993 kit tests, 29 new; tsc clean on both projects; export map at 60 entry points, --check clean). The parent stays unchecked because S4.14.g does, per §2.1. -->
  • S4.15 Add safe error mapping. Evidence: expected domain errors remain stable while unexpected errors produce an opaque ID and server trace. <!-- src/error-mapping.ts + src/error-mapping.spec.ts (23 cases). New ./error-mapping subpath, 61 entry points. S1.10 built the substrate — registry, constructed envelope, canary scan, classifier-not-message rule, transport table — and its own validateErrorRegistry is re-run green here, since the execution contract requires existing code used as evidence to have passed on this branch. What this adds is the three things a register needs beyond the codes in it: a check for HOLES, a boundary, and a degradation. -->

    • S4.15.a Inventory validation, auth, tenancy, conflict, lifecycle, policy/gate, dependency, provider, job, quota, and internal error types. Evidence: error mapping register. <!-- familyGaps asks which of the eleven families has NO code, because a family with none is one whose failures all arrive as internal.unexpected — a 500 for something the caller could have fixed, and a page for somebody who cannot. Invisible to any check that iterates the codes that exist. The eleven families map onto S1.10's twelve namespaces, and the mapping is asserted TOTAL so no namespace escapes. A RULE WAS TRIED HERE AND REMOVED, which is worth recording. "Every code in this family says the same thing about retrying" sounds like a smell and is not: validation should be entirely non-retryable, because waiting never fixes a bad request, and conflict entirely after-change, because waiting never resolves a conflict. It fired on both — on correct configuration — which is a check somebody switches off within a week, the same failure mode S4.1.d's scan was narrowed to avoid. It was replaced by family-served-only-by-deprecated-codes, which is sound, and by an assertion that the shipped registry produces NO findings at all — the thing that keeps the replacement honest. -->
    • S4.15.b Map expected errors to stable code, status, message key/params, field/ref details, retryability, remediation, and correlation. Evidence: mapping tests. <!-- The envelope is S1.10's publicError, exercised here end to end: a stale-revision conflict carrying its key, its details, its reload-and-compare remediation and its correlation, with scanEnvelope finding no leak. And the negative: a parameter the code never declared is refused, because an undeclared parameter is how a field added to some handler's context object six months from now goes out with it. -->
    • S4.15.c Preserve discriminated domain semantics across API, job update, event, and UI without relying on message parsing. Evidence: end-to-end error fixtures. <!-- All four transports branch on the CODE — that is the claim. A job update does not say "failed: not authorized" for a UI to parse; it carries the code, and the UI renders from the code and the message key. The four projections are built from ONE registry row, so agreement is derived rather than asserted twice, and three drifts are caught: a code that differs (one failure with two identities), a retryability that differs (the same failure telling a worker to retry and a client to give up), and a SIGNAL CONTAINING A SPACE — a sentence, which a consumer branching on breaks the day somebody improves the wording. That last one is the message-parsing this item exists to stop, arriving through the field nobody looked at. -->
    • S4.15.d Map unexpected errors to an opaque public ID/message and retain access-controlled cause/stack/trace internally. Evidence: unexpected error tests. <!-- mapSafely gives the caller an opaque internal.unexpected with a support reference, and retains the cause internally behind S1.10's support scope. Three ways in, one way out: no mapping, a mapper that THREW, and a code the registry does not know — the last because a code nobody registered has no status, no retryability and no declared parameters, so emitting it invents all three. The caller cannot tell the three apart, which is correct; the log can, which is the point — both halves asserted. -->
    • S4.15.e Redact secrets, SQL, stack, internal URLs, signed URLs, content, prompts, protected identities, and cross-tenant references. Evidence: canary leakage tests. <!-- S1.10's canary scan, exercised on a real constructed envelope. The stronger guarantee is structural and already held there: the envelope is CONSTRUCTED from declared scalars rather than filtered from a rich cause, so a secret, a SQL fragment, a stack frame or a signed URL has no field to travel in. The thrown mapper case additionally asserts that the exception's own text does not appear in the result. -->
    • S4.15.f Handle mapper failure, unknown future code, nested aggregate error, and partial bulk result safely. Evidence: resilience tests. <!-- Four resilience cases, all real. A MAPPER THAT THROWS DOES NOT BECOME THE ERROR — when it does, the original failure disappears and everybody spends the incident debugging the error handler. An UNKNOWN FUTURE CODE degrades by NAMESPACE and never by a default: a client compiled last year meeting policy.consent_withdrawn should treat it as a policy failure, and "unknown error, retry" is how a permanent refusal becomes a retry loop. S1.10 closed the namespace set precisely so this fallback has something to fall back to. An AGGREGATE reports every distinct member code rather than the first, because "the import failed: field_invalid" over a run where half the rows hit a quota sends the operator to fix the wrong thing. And a run with any success is PARTIAL with its own code, because a client that reads a bulk failure as total failure retries the whole batch and duplicates everything that worked. --> <!-- S4.15 rollup: 3017 kit tests pass (23 new); tsc clean on both projects; export map at 61 entry points, --check clean. -->
  • S4.16 Generate OpenAPI or equivalent route contracts. Evidence: request, response, auth, error, and idempotency semantics are complete. <!-- src/openapi.ts, generated from the route descriptors rather than written beside them. A hand-maintained specification describes what somebody believed the API did on the day they last edited it, and the gap between that and the code is invisible from either side. THE DECISION THAT SHAPES EVERYTHING ELSE: completeness is measured FROM THE ROUTES TOWARD THE DOCUMENT. A specification is trivially made complete by deleting the operations that are hard to describe, and a check that iterated the document would find nothing wrong with having done it. -->

    • S4.16.a Derive operation ID, method/path, owner/capability, summary, lifecycle, version, and tags from route descriptors. Evidence: operation inventory parity. <!-- The operation id IS the route id. The descriptor already carries a dotted, lowercase, unique, stable identifier, and inventing a second one creates a mapping nobody maintains and a generated client whose method names drift from the routes they call. Parity is asserted in both directions over the two archetypes' REAL routes: a mounted route with no operation is a client missing a call the server answers, and an operation with no route is a client with a method that 404s. The summary is a translation KEY and never a sentence — a summary written in one language is one a generated client shows to everybody — and summary-is-a-sentence is a gap kind for exactly that. -->
    • S4.16.b Generate path/query/header/body/upload schemas with units, formats, limits, defaults, and examples. Evidence: request schema tests. <!-- Every :param in the declared path must appear in the contract, or a generated client cannot build the URL — asserted against /desks/:workspaceId/stories/:storyId/quotes/:blockId, which has three. number-without-a-unit is the fault that causes real incidents: a timeout documented as integer is one every client guesses the unit of, and half guess seconds where the server means milliseconds — a thousand-fold error in the direction that looks like a hang. required-with-a-default is the contradiction where the client picks one answer to "what happens when it is absent" and the server picks the other. A body with no documented limit is one every client discovers as a 413. And a parameter can be PRESENT in the document and still undocumented in the only sense that matters, so parameterFaults feeds contractGaps rather than sitting beside it. -->
    • S4.16.c Generate success/async/partial/paginated/stream responses and capability/job references. Evidence: response schema tests. <!-- A success and an error response are both required: without the error, every generated client treats a failure as an unexpected exception and every one of them handles it differently. A 202 must name where the caller polls — an acceptance nobody can follow up on is worse than a refusal — and a paginated response must name the cursor parameter it accepts back, or it is a first page and nothing else. Both are gap kinds, both produced by a test, and the archetype's 201-returning import is carried through the real handler in the gate. -->
    • S4.16.d Document authentication, tenant/workspace, object/property/ function authorization, step-up, and data classification. Evidence: security completeness check. <!-- Security is PER OPERATION and never only at the document root. "This API uses bearer tokens" is true and useless; which roles, which scope, whether the object is authorized separately, whether a step-up is required and what classification the data carries are the things that DIFFER between operations, and the ones that differ are the ones worth documenting. The console's intake names system among its audiences and therefore authenticates as a service rather than a session — derived, not asserted twice. security-not-documented fires on a scoped operation that lists no roles. -->
    • S4.16.e Document validation/auth/conflict/limit/dependency/provider/ internal error envelopes and statuses. Evidence: error coverage check. <!-- errorCoverageGaps checks S1.10's register in BOTH directions against the document. Forward: each of the seven families the item names must have a documented envelope, spelled in the register's own vocabulary — "limit" is quota there, and a second word for it would be two codes for one failure with different retry advice. Backward, and this is the direction that matters more: a document that names a code the register does not have, or gives it a status the server never sends, is worse than one that says nothing — a client branching on it takes the default branch nobody tested. Both directions are exercised, and the shipped register produces zero findings. -->
    • S4.16.f Document idempotency key scope/retention/replay/conflict and expected-revision preconditions. Evidence: mutation semantics check. <!-- The mutation section is derived from S2.3's CommandSchema, so the document cannot claim semantics the command does not have: the key's scope, its retention, the if-match precondition and the 409 it can return. A mutation with neither is two gap kinds, not one, because they fail differently — a client with no idempotency key discovers retry semantics by causing a duplicate, and a client with no precondition writes last-write-wins by default. A READ carries no mutation section at all rather than one full of nulls. -->
    • S4.16.g Validate spec, examples, generated clients, server conformance, deterministic output, and drift. Evidence: contract generation gate. <!-- tools/route-contract-gate.mjs, an Nx target (nx route-contract-gate workbench-kit), running the WHOLE chain: route descriptors → contract → generated client → mounted router → archetype handlers. A specification validates; that is the problem. A document can be syntactically perfect, internally consistent and complete against its own schema while describing an API that does not exist, because every check that reads only the document agrees with it. So the gate never reads the document alone. The routes are the archetypes' real ones, the router is createWorkbenchRouter with all ten seams, the handlers are the archetypes' own bindings over in-memory ports, and the client is GENERATED from the contract and IMPORTED as a module rather than compared as text. Six checks: spec (from the routes), examples (parsed and checked against their own schemas and limits — an example is the part of a document that gets copied), errors, clients, conformance (every generated call must reach the route the contract names, with a status the contract documents), and drift against a committed tools/route-contract.json, byte-compared after regenerating twice and in reverse route order. FIVE NEGATIVE CONTROLS, every one verified to fire by inverting it: a document with a route deleted, a body example that is not JSON, a status the server never sends, a client built from a contract with the wrong path (which must NOT reach the route, or the conformance check proves nothing), and a changed status code that must show up as drift. Two real defects were found by running it rather than reading it: the pull-quote body the document described was not the block the desk's own parser accepts, and properties.js initialised a Set at import time under a "sideEffects": false promise — caught by the packed-consumer smoke once the new entry point widened its scan. --> <!-- S4.16 rollup: 3060 kit tests pass (43 new); tsc clean on lib and spec projects; export map at 62 entry points, --check clean; the route-contract, vocabulary-scan and packed-consumer gates all pass. -->
  • S4.17 Add consumer-driven contract tests for every host application. Evidence: incompatible BFF changes fail before merge. <!-- src/consumers.ts plus tools/consumer-contract-gate.mjs (nx consumer-contract-gate workbench-kit). A route contract says what the provider OFFERS and cannot say what anybody NEEDS, which is the whole difference between a change that is free and one that takes a host down: removing a response field is safe unless somebody reads it, and no amount of reading the provider's own document says which. THE TWO SIDES ARE PRODUCED DIFFERENTLY ON PURPOSE. The provider surface (tools/provider-surface.json) is OBSERVED — every guaranteed field is one a real handler returned through the mounted router, and every error code is one a real deliberate mistake provoked. The consumer manifests (tools/consumer-contracts.json) are DECLARED by the hosts, because a manifest derived from the provider would agree by construction and prove nothing; the entire value is in the disagreement. -->

    • S4.17.a Inventory each host application's consumed routes, operations, versions, fields, enum/error/state variants, pagination, and streaming assumptions. Evidence: host consumer manifests. <!-- Three manifests over the archetypes' real routes: studio-web (studio-guild), operator-console (reliability-guild) and cli (platform-tools). consumes-nothing is the fault that catches the real omission — a host with an empty manifest is not one with no dependencies, it is one nobody has inventoried, and it passes every check downstream precisely because it asked for nothing. The five non-success STATES are enumerated and all five must be answered for: a host that handles error and nothing else renders not-configured as something to retry, degraded as fresh and version-unsupported as an outage, and all three look like the product working. Also refused: a window containing no version (the skew matrix comes out empty rather than red), an operation that both streams and paginates (two answers to "where does the next item come from"), and variants declared for a field the host does not read. A FALSE POSITIVE WAS FOUND AND FIXED here by running it: duplicate detection folded the request and response field lists together, so feedId — sent in the intake and returned in the plan — was reported as listed twice. -->
    • S4.17.b Capture provider-independent expectations as source-controlled contracts with valid/invalid request and response examples. Evidence: per-host contract suites. <!-- Sixteen committed examples, two per consumed operation, and the checks are on the LABELS rather than on the shapes. A "valid" example missing a field the manifest says the host reads means one of the two is a lie and the suite will pass against a provider that omits it. An "invalid" example that satisfies every declared requirement means the suite passes whatever the provider does — worse than having no suite, because it is counted as coverage. Both are checked by evaluating the example against the manifest, so the label has to earn itself. An operation with only happy paths is also a finding: a suite with no refusals passes against a provider that has stopped validating anything. -->
    • S4.17.c Verify current BFF/provider satisfies each consumer contract without permissive unknown fallbacks hiding required behavior. Evidence: provider verification reports. <!-- verifyProvider reports per host with its owner, and code-handled-only-by-an-unknown-fallback is the fault the clause is about: a consumer whose default branch shows a generic error technically handles every code that will ever exist, and counting that as coverage would make every consumer compatible with every provider — a suite that cannot fail. The code is reported whatever the policy says; the policy only decides how bad it is. providerSurfaceGaps ties the surface to S4.16's GENERATED contract in both directions, so "the provider offers this" is a claim about the mounted routes rather than about a second hand-maintained document, and a surface listing variants for a field it does not guarantee is refused — otherwise every "the provider cannot emit that" finding is decided by a list nothing keeps true. -->
    • S4.17.d Verify consumers tolerate declared additive fields/variants and handle not-configured/degraded/unavailable/error/version states. Evidence: consumer compatibility tests. <!-- surfaceChanges classifies and impactedHosts decides, and the split is the point: whether a change breaks anybody is NOT a property of the change. An added enum variant is additive for a host with a default branch and a crash for one that switches exhaustively; an added error code is additive for a host that degrades on unknown codes and a refusal for one that does not. Both pairs are asserted on the SAME change with opposite verdicts. A rule that guessed would be wrong for half the fleet, and wrong in the reassuring direction. The five states are required by the manifest check in (a). -->
    • S4.17.e Test old/new host and BFF versions across the supported compatibility window. Evidence: version-skew matrix. <!-- A cell per host per release, each unsupported one carrying the window that excludes it. The RELEASE window is kept apart from the per-operation contract versions on purpose: an operation's version says what shape one call has, a release version says which deployment is in front of it, and a host can be current on every operation and still unable to run against next week's release. commonVersions answers the question anybody actually asks — "can we deploy this?" — and an empty answer is a state the fleet can be in with no single host at fault. The gate asserts the release actually being served is among them. -->
    • S4.17.f Add CI ownership and affected execution so a breaking BFF change fails before merge with the exact impacted host. Evidence: deliberate breaking-change test. <!-- Every manifest must name an owner (no-owner is a fault: a report that cannot say who to tell is one somebody routes by asking around), and breakingChangeReport emits one build-failing line per impacted host naming the host, the owner, the operation and the exact field — the difference between a gate people fix and a gate people disable. affectedHosts is deliberately WIDER than impactedHosts: selection picks whose tests could possibly notice, not who a static rule already calls broken, because selecting the latter would make the suites unable to find anything the rule missed. THE DELIBERATE BREAKING CHANGE RUNS IN THE GATE: a field a host reads is removed from the observed surface and the gate fails unless the report is breaking, names that host's owner, and quotes the field. EIGHT NEGATIVE CONTROLS, every one verified to fire by inverting it. A REAL DEFECT WAS FOUND BY RUNNING THIS RATHER THAN READING IT: every archetype refusal came back as a bare status with NO envelope — no code, no message key, no correlation — because their message keys (storyDesk.error.…) have an uppercase letter in the first segment, which the kit's own translation-key rule rejects, so publicError refused and the router fell back to a status alone. Every existing test asserted the status and passed straight through it. The keys are now domain-prefixed (veritas.storyDesk.error.…) and router.spec.ts asserts a domain refusal carries a real envelope. --> <!-- S4.17 rollup: 3096 kit tests pass (36 new); tsc clean on lib and spec; export map at 63 entry points, --check clean; route-contract, consumer-contract, vocabulary-scan and packed-consumer gates all pass. -->
  • S4.18 Add an API inventory drift gate. Evidence: undocumented and unowned routes fail CI. <!-- src/inventory.ts plus tools/api-inventory-gate.mjs (nx api-inventory-gate workbench-kit). S1.8's routeInventory projects the routes somebody DECLARED; this inventories the routes that EXIST, and the difference between the two sets is the whole item. The ones nobody declared are exactly the ones nobody reviews. EVERYTHING IS OBSERVED. The inventory is built by mounting the real router and reading what it did — which routes it mounted, and, by driving each one, the actual pipeline steps each request ran. -->

    • S4.18.a Enumerate framework, router factory, domain plugin, proxy, redirect, webhook/callback, streaming, health, and administrative routes. Evidence: deterministic API inventory. <!-- Nine origins, because nine different things mount HTTP surface and only one of them goes through the descriptor table: a gate that enumerated the router factory alone would report a clean inventory for a deployment whose entire admin surface is unowned. originCoverage reports every origin INCLUDING the zeroes — "we found no webhooks" and "we never looked for webhooks" are the same row in a report that drops empties, and only one of them is good news. Deterministic by construction: the same routes walked in any order produce one digest, asserted by reversing the list. Every entry names its file and symbol, because a finding that cannot say where sends somebody grepping. -->
    • S4.18.b Reconcile mounted operations with route descriptors, generated contracts, capability registry, owner, audience/host, and lifecycle. Evidence: bidirectional mapping report. <!-- Both directions, and they are different failures: an entry with no descriptor is a route nobody declared — no owner, no audience, no lifecycle, nothing that would put it in front of a reviewer — and a descriptor with no entry is a declaration nothing serves, which is a 404 for whoever followed the documentation. Ten faults, every one producible, including method/path/host/audiences/lifecycle/version disagreeing with the descriptor. descriptorlessOrigins is SUPPLIED rather than assumed: whether a health endpoint is declared is a deployment's decision, and a kit that decided would produce either a permanent false finding or a permanent blind spot. A LOGIC BUG WAS FOUND BY THE PRODUCIBILITY TEST here — the capability check only ran when the entry and the descriptor DISAGREED, so a route and a descriptor both pointing at a capability the registry never declared passed clean. -->
    • S4.18.c Detect new, removed, moved, renamed, method/path/version-changed, undocumented, unowned, duplicate, and unreachable operations. Evidence: drift classifications. <!-- Eleven classes. MOVED and RENAMED are kept apart because the fix is opposite: a moved operation kept its identity and changed its address, so every client needs a new URL and the calling code is fine; a renamed one kept its address and changed its identity, so every client's URL still works and every generated method name is wrong. Collapsing them into "changed" produces a report whose remediation is a guess. The static classes hold whether or not anything changed — undocumented, unowned, duplicate (two registrations on one address, of which one is unreachable and which one depends on registration order), and unreachable (the inventory has it, the router did not mount it, so it is documented, owned, and answers nothing). -->
    • S4.18.d Ignore comments/formatting/generation noise while retaining semantic contract and middleware changes. Evidence: noise-control tests. <!-- The digest covers the semantics and excludes the source location and the walk order. A gate that fired on a handler moving file is one somebody adds --no-verify for, and after that it never fires again — so the gate proves it by relocating EVERY handler to a generated path and requiring zero findings. Middleware stays in, and that is the half that matters: a route that quietly stopped running its authorization step is byte-identical from the path, and the negative control removes object-authorize from one route and fails the gate if it is not reported. Audiences are compared as a SET, so reordering is noise and adding one is not. -->
    • S4.18.e Fail CI with exact route source, missing metadata, affected consumers, and remediation command. Evidence: deliberate route addition test. <!-- REMEDIATION_FOR_CLASS is a total map held next to the classes, so a new class cannot be added without a command to fix it: a failure that says what is wrong and not what to do is one people route to whoever last worked it out. Affected consumers come from S4.17's manifests — the kit already knows who reads what, so leaving it to the reader would be withholding it. THE DELIBERATE ADDITION RUNS IN THE GATE: an undeclared administrative route is added to the observed inventory and the gate fails unless it is reported as new, undocumented AND unowned, with apps/oshun/bff/src/admin.ts:reindexRoute on every line. -->
    • S4.18.f Require explicit disposition/deprecation/migration for removals and compatibility routes. Evidence: removal mutation tests. <!-- Four dispositions — deprecated, migrated, compatibility-route, never-shipped — and a removal with none is refused, because "it is gone" is the absence of a decision and is indistinguishable from a route deleted by a bad merge. Each is checked against something real: a deprecation must cite a notice the register actually holds, a migration must point at an operation something still serves, a compatibility route must carry an end date (one without is a permanent route nobody decided to keep), and never-shipped-claim-for-a-shipped-operation catches the convenient answer — it is the cheapest disposition to write and the inventory knows the lifecycle, so the claim is checkable. A disposition recorded for an operation still being served is also a fault: a stale record silently excuses the next removal. THE REMOVAL MUTATION RUNS IN THE GATE, including the false "never shipped" claim. SIX NEGATIVE CONTROLS, every one verified to fire by inverting it. --> <!-- S4.18 rollup: 3125 kit tests pass (29 new); tsc clean on lib and spec; export map at 64 entry points, --check clean; route-contract, consumer-contract, api-inventory, vocabulary-scan and packed-consumer gates all pass. §S4 is complete except S4.14/S4.14.g, which stay open: their evidence is a browser network assertion and no app mounts the capability-health endpoints yet. -->

S5 — Provider, grounding, trust, and policy seams#

  • S5.1 Extract WorkbenchModelProvider with explicit operations rather than a generic text-completion escape hatch. Evidence: unregistered operations fail typecheck. <!-- src/model-provider.ts. The failure this exists to prevent is ONE GENERIC METHOD. A port with complete(prompt: string): Promise<string> has no safety policy, no grounding requirement, no acceptance criterion, no cost model and no version — because none of those can be stated about whatever the caller put in the prompt. Every one of them then moves into the callers, where it is implemented differently in each and reviewed in none. So the port takes REGISTERED OPERATIONS, and the kit holds the eight domain-neutral CLASSES while the domain holds the operations. -->
    • S5.1.a Inventory domain model operations with distinct inputs, outputs, safety, grounding, latency, streaming, tool, cost, and acceptance semantics. Evidence: operation register. <!-- Eleven register faults, every one producible. The load-bearing one is generation-with-neither-grounding-nor-human-acceptance: an operation that generates free output, grounds it in nothing and is accepted by nobody IS a text completion, and giving it a name does not change what it is — which is how the escape hatch comes back after being removed. The classes are named by what the OUTPUT is rather than by the task, because the output decides everything downstream: an extraction can be checked against its source and a generation against none of those, which is exactly why a generation must state its acceptance criterion. Also refused: a tool WILDCARD (blast radius is the whole tool registry and it grows every time somebody adds one), a moderation whose failure only annotates (no effect unless every caller remembers to read it), a zero latency budget (nothing ever fires and it reads as green), and personal data admitted with nothing screened for. -->
    • S5.1.b Define stable operation IDs and typed versioned request/result/ error contracts for each operation class. Evidence: provider contract types. <!-- OperationRegister<TId> is keyed by the operation id, and operationRef is generic over that key union — which is what makes S5.1.e's compile-time refusal a type error rather than a lookup that returns undefined. The version is read from the DECLARATION and never passed in: a caller that could name the version could name one the provider never served. resolveOperation keeps the two runtime failures apart — an unregistered id is a caller using something that does not exist, and a version mismatch is a caller that is simply older than this deployment. The result is a three-way ModelOutcome, and the third arm is the point: an UNKNOWN outcome is a result, not an error, because the work may have run and the money may have been spent. -->
    • S5.1.c Define provider capability discovery and operation-specific model, region, modality, tool, context, output, and policy support. Evidence: capability fixtures. <!-- Eight capability faults checked against the register, all producible. policy-the-operation-requires-is-unsupported is the one worth having: a provider serving a screened operation without being able to screen for the category returns output that looks exactly like screened output, and nothing downstream can tell. configured: false is a STATE rather than a failure, and a provider that reports itself configured while offering nothing is a routing destination that refuses everything sent to it. -->
    • S5.1.d Define invoke/stream/cancel/reconcile interfaces that preserve idempotency, correlation, usage/cost, provenance, and unknown outcome. Evidence: lifecycle interface tests. <!-- reconcile exists because of the unknown outcome: without it every caller resolves "did that run?" by retrying, which is the one thing an unknown outcome makes unsafe. Six request faults and six outcome faults. The correlation is echoed in the PROVENANCE and not only carried on the request — a provider log and an application log that cannot be joined are two records of one event. The quiet outcome fault is completed-with-unmeasured-usage-reported-as-zero: a provider that cannot report usage and returns zeroes produces a cost dashboard that is confidently wrong, and the most expensive operation is the one that looks free. refused-not-configured-naming-a-credential refuses a refusal that carries a credential VALUE where a setting NAME belongs. -->
    • S5.1.e Register domain adapters explicitly and prevent generic prompt/ text completion or arbitrary tool invocation through the shared API. Evidence: compile-time negative tests. <!-- Two @ts-expect-error calls inside a function the spec never invokes: the assertion is made by tsc, and an UNUSED @ts-expect-error is itself a compile error — so if either call ever started type-checking, the spec project would fail to build. Alongside it, modelAdapterFaults refuses an adapter registered for an operation nothing declares (reachable surface with no safety policy), two adapters for one operation (which one serves a call depends on registration order), and an operation whose last id segment is complete, prompt, raw, chat or passthrough — names for "run whatever I send you", because the escape hatch always returns as a legitimately registered operation whose input is a prompt. -->
    • S5.1.f Implement two providers with different supported operation sets and honest unsupported/not-configured results. Evidence: adapter contract suite. <!-- src/testing/model-provider-conformance.ts, SHIPPED as a test-contract subpath rather than kept in the kit's own spec, because the adapters worth checking are the ones domains write next year against providers this repository has never heard of. Two adapters exercise it with disjoint operation sets, and both really compute: the extraction scans sentences for claims, the classification counts terms against a closed severity set, and the embedding is a real hashing projection whose unit norm is asserted to five places. The two probes that carry the suite have no convenient correct answer — an operation the adapter does not list must come back unsupported-operation and NOT retryable (answering it is output invented for work it cannot do; marking it retryable turns a permanent gap into a retry loop), and an unconfigured provider must refuse everything while naming what is missing. The suite is shown to CATCH all three: an adapter that answers unlisted work, an unconfigured one that answers anyway, and a cancel that always succeeds. --> <!-- S5.1 rollup: 3156 kit tests pass (31 new); tsc clean on lib and spec; export map at 66 entry points, --check clean; the packed-consumer and vocabulary-scan gates pass. -->
  • S5.2 Implement not_configured as the production default. Evidence: no provider binding results in an exact 503 and zero fabricated output. <!-- src/provider-registry.ts. The registry starts EMPTY — not "empty in production", not "empty unless a fallback is registered", empty in every composition until a binding has passed three probes. Every other arrangement has the same failure: a deployment ships with no credentials and answers anyway, because something plausible was registered as a default and nothing in the request path can tell a real answer from a convincing one. The 503 is dependency.not_configured, added to S1.10's register in the same change: after-remediation and park-for-operator, because waiting never fixes a missing binding and neither does changing the request. -->
    • S5.2.a Initialize provider registries empty unless an approved runtime binding passes configuration/credential/capability probes. Evidence: default registry test. <!-- emptyRegistry binds nothing and resolveCapability answers not-configured, which is the only state a deployment can be in that cannot answer with something it made up. admitAll is a FOLD rather than a mutable registry, and it returns the refusals alongside the result: a deployment that silently admitted two of three providers looks exactly like one that was only given two. Eight refusals, every one producible, and the one that matters is probe-never-ran — an absent probe result is refused rather than read as a pass, because the opposite default is how a credential check skipped for a release becomes a credential check nobody runs. capability-already-bound is refused for the ordering reason the router uses: which provider serves a call must not depend on registration order. -->
    • S5.2.b Remove or isolate demo, fake, deterministic, fallback, and heuristic providers behind explicit test/dev-only composition. Evidence: production dependency scan. <!-- Two layers. At admission, a fixture-class binding is refused outright in a production composition, and any binding is refused for a composition it was not approved for. At the artifact level, S1.13.f's production-scan target already walks the BUILT package — the emitted .js and .d.ts graphs separately, the source maps, and the exports manifest — for test utilities and reserved fixture identifiers, with a negative control that re-runs the whole scan over the emitted test files and fails if it finds nothing there. It is green at 59 production entry points against 85 derived test-utility tokens, with 299 control hits. -->
    • S5.2.c Return typed not_configured with capability, safe missing category, owner/remediation, and correlation as HTTP 503. Evidence: contract fixture. <!-- notConfiguredResponse projects S1.6.c's NotConfiguredState onto the wire, and the projection is where the decision is: it carries the missing CATEGORIES and never the setting names. CMS_API_KEY is exactly what an operator needs to hear and exactly what a public error must not say — a setting name is a map of the deployment, and the caller who triggered this is not the person who can act on it. The owner, the resolution authority and the correlation go instead, so the answer is actionable by whoever reads it. The status and the code are asserted against the shipped register rather than restated. -->
    • S5.2.d Prevent UI/BFF adapters from substituting sample output, empty success, cached unrelated output, or inferred scores. Evidence: end-to-end negative tests. <!-- Four named substitutions, all producible, and all four turn "we cannot do this" into something that renders as the product working. empty-success is the one that gets written by accident: a 200 with an empty list reads as "there is nothing here", which is a true statement about a different situation. cached-unrelated-output compares the correlation the body was produced for against the one it is being served for — a cache keyed loosely enough to do that is keyed loosely enough to cross a tenant. And the checker says NOTHING about a capability that is actually bound, so it cannot be satisfied by refusing everything. -->
    • S5.2.e Keep fixture providers visibly labeled and impossible to enable in production build/runtime configuration. Evidence: production guard tests. <!-- fixture-approved-for-production checks the DECLARATION and not only the registry, because approvedFor: ['production'] on a fixture is one line in a file nobody re-reads and it is the line that makes the admission check pass. The labelling half runs both ways over S1.13.d's synthetic mark: fixture output without it is indistinguishable from a real answer once it has been passed on twice, and RUNTIME output carrying it is a fixture wired in behind a real provider's name — which is the direction a one-way check would miss. -->
    • S5.2.f Exercise configured, missing credentials, unsupported operation, disabled provider, unhealthy provider, and probe failure separately. Evidence: capability state matrix. <!-- Six situations, asserted DISTINCT in one matrix built from real admission refusals rather than from labels. The remedies differ and only one of them is the caller's: a missing credential is an operator's afternoon, a disabled provider is a decision somebody made, an unhealthy one may recover on its own, and an unsupported operation never will — so retryHelps is false for the disabled provider and true for the unhealthy one, and a single "unavailable" would give one retry policy that is wrong for one of them. The ORDER is the order of certainty: an unsupported operation is refused before health is consulted, because a provider that is up cannot start supporting an operation it does not implement. And healthy: null is probe-failed rather than configured, since an unprobed provider is not a healthy one. --> <!-- S5.2 rollup: 3184 kit tests pass (27 new); tsc clean on lib and spec; export map at 67 entry points, --check clean; the production-scan, packed-consumer, vocabulary-scan, route-contract, consumer-contract and api-inventory gates all pass. -->
  • S5.3 Fence instructions, user content, retrieved sources, tool output, and attachments into distinct trust zones. Evidence: indirect-injection eval. <!-- src/trust-zones.ts and src/injection-eval.ts. The failure this prevents is a prompt that is ONE STRING: once the policy, the template, the question, a retrieved page and a tool's JSON have been concatenated, nothing downstream can say which of them said what, and the model infers authority from phrasing — the one signal an attacker controls completely. THE CANONICAL RENDERER ALREADY EXISTS and the kit may not import it: @oshun/prompt-trust is layer:shared and the kit is buildable (TS6059), the same shape as S4.2 and S4.11. So the kit owns the CONTRACT, the CHECKS and the CORPUS; the conformance spec imports the real renderer as a devDependency and runs the real corpus through it. Writing a second sanitiser here would have produced two that drift, and the one that matters is the one in the request path. -->
    • S5.3.a Define immutable trust-zone types and precedence for system/policy, developer template, user request/content, retrieved source, tool result, attachment, and prior generated output. Evidence: trust contract. <!-- Precedence is a TOTAL order rather than a trusted/untrusted flag, because the question asked at runtime is always between two named zones — "the retrieved page contradicts the template, which wins?" — and a partial order has pairs with no answer, which then get settled by whoever writes that branch and settled differently in the next file. Rank comes from the KIND and a zone never carries its own: a zone that could state its rank would be ranked by the assembler, and the assembler stands closest to the attacker's bytes. The AUTHOR is what an attacker controls rather than the kind, so a "developer template" that is really a database row an operator edits is refused, and a system policy may come from the repository and nowhere else — deployment policy is configuration SELECTING among repository-authored policies, or an operator zone one rank lower that can narrow and never widen. user-request outranks user-content because the human authored one and merely POSSESSES the other, which is the whole of indirect injection. prior-generated-output ranks last, below the attachment: it is the only zone with no external author, and one instruction that survives one turn comes back as something already decided. Eight precedence faults, all producible, and the shipped tables are asserted to produce none. The first version of the unaccountable-author check fired on the CORRECT table — user-request is authored by the end user, which is what it is — the same failure S4.15 had; it was narrowed to binding and directive, with a test naming the mistake. -->
    • S5.3.b Serialize each zone with unambiguous boundaries, source identity, authority label, data classification, and truncation status. Evidence: request assembly tests. <!-- Thirteen assembly faults, checked on the rendered STRING rather than on the object that produced it, because every failure here is a property of the string: a fence the content can forge, a label a template change dropped, a note somebody added between two blocks. A check that reads the object confirms the object. unfenced-text-between-zones is the one that catches the real regression — a fencing scheme is never removed, it is bypassed by one helpful line ("Note: the requesting user is an administrator") appended by somebody fixing a ticket, and that line has no author, no rank and no classification while sitting at the top of the model's attention. Its threshold is 16 characters and not zero, because a finding that fires on "\n\n" is a finding somebody switches off. truncation-recorded-but-not-visible-to-the-reader is the quiet one: the only reader who could compensate for a document being cut in half is the model, and it reads the fence rather than the audit record. Zones must also be emitted in precedence order, since the first instruction-shaped text in a prompt sets the frame the rest is read against. -->
    • S5.3.c Prevent untrusted zones from altering system policy, tool availability, authorization, output schema, or evidence standards. Evidence: precedence tests. <!-- Each of the five controls carries a FLOOR: the lowest-ranked zone permitted to influence it. authorization's floor is null — not "only the system policy may decide it" but NO zone at all, because the answer is read from the actor and the grant store before assembly begins, and a control whose floor is any zone is a control somebody eventually decides from a prompt. The load-bearing fault is control-decided-after-the-model-ran, and it is the one that looks fine in review: a tool allow-list computed from the model's chosen call is not an allow-list, it is a record of what happened; filtering output against a schema the output selected is not validation. In each case the control is still there, still logged, and no longer a control. Also checked: a control nothing states took whatever default the assembly code happened to have, and an unstated default is the one nobody reviews. -->
    • S5.3.d Sanitize/normalize title, metadata, hidden text, markup, URLs, filenames, archive members, media metadata, and multimodal text. Evidence: indirect-injection corpus. <!-- Ten channels with a disposition each, because body-text is the channel everybody sanitises and the one attacks stopped using: the other nine reach the model through code that FORMATS rather than fences — Document: ${doc.title} — and a title is attacker-controlled in exactly the way the body is, with none of the handling. hidden-text is removed rather than sanitised, since there is no legitimate reason for text a human reader cannot see to reach a model that acts on it, and the removal must be DISCLOSED or the model summarises a document that says something different from the one the user is looking at. The kit does not normalise — it VERIFIES, given what arrived and what is about to be emitted, which composes with any sanitiser including one written next year. The real algorithm it does own is the invisible-character scan: Unicode tag characters (U+E0000–U+E007F, a complete ASCII alphabet that renders as nothing and is the case an eyeball review cannot perform), interlinear markers, and a bidi BALANCE check rather than a presence check — the 2021 trojan-source result turns on a run that is never popped, and balance is evaluated per LINE because the bidi algorithm resets at a paragraph separator, so a whole-string counter would report every multi-paragraph RTL document. Zero-width joiners are allowed in prose (load-bearing in Devanagari and in emoji sequences) and refused in identifier channels, where their only effect is to make two different names render identically. -->
    • S5.3.e Minimize tools and bind typed arguments to actor/tenant/resource/ side-effect policy outside the model. Evidence: tool authorization tests. <!-- The model chooses WHICH tool and never WHOSE data. A context-bound argument's value is taken from the request and the model's value for it is DISCARDED rather than validated — because validating a model-supplied tenant id against the actor's tenants is a correct check that still leaves the model picking which of the actor's three tenants gets written to, and the test asserts exactly that difference. identity-argument-left-model-supplied is the confused deputy in one line: the model holds the deployment's credentials while reading the attacker's document. external-effect-without-confirmation-beside-inert-zones is compound on purpose — the tool alone is fine, the untrusted zone alone is fine, and the two in one call with no human between them is the arrangement where a page decides what gets sent and to whom. minimizeTools is an intersection computed BEFORE assembly, since minimisation applied afterwards is minimisation of a list the model has already been shown. -->
    • S5.3.f Evaluate direct, encoded, multilingual, split, multi-turn, retrieved, tool-result, attachment, URL, and multimodal attacks. Evidence: versioned adversarial report. <!-- 23 cases across all ten vectors and all five controls, graded by the kit's own production checks and run in injection-eval.spec.ts against the repository's REAL renderer. THE CORPUS IS NOT A FILTER: nothing in the request path consults it, and it would be worthless there, because a defence that works by recognising the attack string fails on the string it has not seen. Four gradings carry the report. A case that did not RUN is graded as harshly as one that escaped, or a suite improves its score by erroring. A payload that never reached the prompt is inconclusive rather than a pass — a pipeline that drops the attachment on a parse error would otherwise sweep the attachment vector, and that vector is about the attachment being READ. Coverage is gated on BOTH margins, so forty direct cases and no attachment cases is not 97% covered; the vector×control matrix is published rather than gated, since fifty pairs is a gate somebody deletes. And the corpus pins a LITERAL contract fingerprint: the first version evaluated trustContractFingerprint() here, which equals the live value by construction and made the staleness check unable to fire in the one situation it exists for. Two tests are the evidence the composition is load-bearing rather than decorative: with the invisible-strip step removed and nothing else changed, exactly encoded-003, encoded-004 and multilingual-002 escape — @oshun/prompt-trust normalises, neutralises encoded delimiters and bounds length, and does not strip tag characters or balance bidi runs; and with a pipeline that lets the payload's zone influence its target control, all 23 escape. --> <!-- S5.3 rollup: 3266 kit tests pass (81 new); tsc clean on lib and spec; export map at 69 entry points, --check clean; the production-scan, packed-consumer, vocabulary-scan, route-contract, consumer-contract and api-inventory gates all pass. The packed-consumer gate caught real work at import time in both new modules (Object.freeze, a Symbol() brand, two new Sets and a tagEncode call) — sideEffects: false tells a bundler it may drop exactly that, so the tables are plain literals, the brand is minted on first use behind a declare const, and the tag-block payload is an escaped literal whose plaintext a test asserts. -->
  • S5.4 Require structured-output schema validation. Evidence: malformed, truncated, extra-field, and wrong-version outputs are rejected. <!-- src/structured-output.ts. The failure this prevents is JSON.parse(response) — not because parsing is wrong, but because of everything that sentence skips: no declared schema, no bound checked, a parser that accepts a duplicate key and silently keeps the last one, bytes that may be the first half of an answer, and the original discarded the moment it succeeds. Each is a separate failure and each presents as the same thing, an object that looks right. Six fault families, 45 faults, all producible. -->
    • S5.4.a Select exact operation/result schema and version before provider invocation and include provider-supported structured mode where available. Evidence: request contract tests. <!-- The schema is pinned BEFORE the call, because choosing the validator once the output is in hand means output that happens to match a laxer schema is accepted — S5.3.c's control-decided-after-the-model-ran with a JSON parser attached. A provider's structured mode is a guarantee about SYNTAX and nothing else: a grammar can force {"approved": <boolean>} and cannot make the boolean right. So the strongest available mode is requested (declining it means paying for a retry to learn something the provider would have refused to emit) AND structured-mode-used-instead-of-validation is a fault — it is the only one in this section that makes something PASS rather than fail, and the team that turns on JSON-schema mode and deletes their validator has not moved validation to the provider; the provider is checking a shape and nobody is checking what is in it. -->
    • S5.4.b Bound raw bytes/tokens/depth/arrays/strings/numbers and detect incomplete/stream-truncated output before parsing. Evidence: limit tests. <!-- All of it runs BEFORE the parse, because a truncated response can still parse: {"total": 4} is valid JSON and valid against any schema whose remaining fields are optional, so the object is complete, the validation passes and nine tenths of the answer is missing with nothing anywhere saying so — the test asserts exactly that pair. The finish reason is the only place that fact exists, and unknown-finish-reason-treated-as-complete is the quiet version: the default reading of an absent measurement must never be the good one. content-filter is kept separate from length, because partial text after a refusal is not a partial answer. The default depth budget is 32: a document nested 200 000 deep is ~400 kB of brackets, comfortably inside any byte budget. -->
    • S5.4.c Parse strictly and reject malformed syntax, duplicate keys, invalid Unicode/numbers, schema mismatch, unknown forbidden fields, and wrong discriminator/version. Evidence: adversarial fixtures. <!-- scanJsonSource is a real strict scanner over the SOURCE, and it has to be, because three of its findings do not survive parsing. JSON.parse('{"approved":false,"approved":true}') returns true — the last one wins, the first leaves no trace, and anything able to append to a response has a rewrite primitive; the spec asserts that V8 behaviour rather than describing it. 9007199254740993 parses to ...992 with no error at all, so an identifier arriving as a large integer is silently a different identifier. And a lone surrogate becomes U+FFFD in whichever component re-encodes first. The scanner keeps an EXPLICIT STACK rather than recursing, which is not style: a recursive-descent parser meets a depth bomb by exhausting the call stack, so the check written to report the attack is what the attack breaks. The test shows the hazard is real — V8's own JSON.parse is iterative and survives 200 000 levels, and the first ordinary recursive walk over the object it returns throws RangeError. Duplicate detection is scoped per OBJECT, since the same key in two sibling objects is ordinary and a check that fired on it would be switched off in a week. The envelope is deliberately not a schema language — the kit shipping one would be a second validator beside the domain's, and two validators disagreeing is what this module is about; what the kit owns is the discriminator, the version and the field allowlist, because those decide WHICH validator runs. An unknown field is refused rather than ignored: a field nobody declared is a field nobody reviewed, and in a model's output it is additionally a place to put things. Object.hasOwn rather than truthiness, so "verdict": null is an answer that was given rather than one that was missing. -->
    • S5.4.d Validate cross-field/domain invariants, stable references, units, authorization scope, and prohibited instructions/URLs after schema validation. Evidence: semantic validation tests. <!-- citation-to-a-source-that-was-never-supplied catches a fabrication rather than a bug: an id resolving to nothing is the shape a made-up reference takes once the format is strict enough to stop made-up prose. reference-to-an-entity-outside-the-actor-scope is the cross-tenant leak, checked here rather than at render time because the id is in the result the moment it validates and everything downstream treats a validated result as a checked one. An invariant that THROWS is reported as violated rather than swallowed — a check that errors is not a check that passed. And the ledger's "prohibited instructions" is answered WITHOUT scanning the output for instructions, which would have been the S4.15 mistake again: whether the output contains an instruction does not matter if it returns in an inert zone, and no amount of scanning saves it if it returns above one. So the check is output-readmitted-above-a-generated-zone, read off S5.3's rank. -->
    • S5.4.e Preserve raw response hash/ref and validation errors; never repair silently into a passing result. Evidence: invalid-output provenance tests. <!-- The line is drawn on what a repair CHANGES, not on its size: trimming a trailing newline and supplying a missing required field are both one line of code. Whitespace, encoding normalisation and stripping a ```json fence remove ambiguity about what the model said; adding a field, removing one or coercing a value WRITE what it said and file the result under its name. accepted-with-faults-recorded is the silent one and the reason the record exists: somewhere between "the validator found three problems" and "the result was used" there is a branch that decided they did not matter, and nothing in the accepted object shows it — a validator whose findings can be recorded and ignored is a logging call. A rejection must keep the raw bytes, since the one case where the exact bytes are certainly needed is the one where validation failed. -->
    • S5.4.f Define bounded retry-with-correction only for approved error classes and retain every attempt/cost/provenance. Evidence: retry tests. <!-- Only FORM is correctable: lexical and envelope are statements about the shape of the answer and a model can be told which shape to use. semantic is deliberately excluded — a model told "that citation did not resolve" returns one that does, and the second answer is worse than the first because it now looks correct. selection and pre-parse are excluded because neither is the model's doing: retrying a truncation at the same token budget truncates again. A MIXED set is refused outright, since fixing the syntax of an output that also cites a non-existent source leaves the content problem behind a response that now parses — and correctionAllowed returns the blocking families, because "no" without a reason is a decision somebody works around. only-the-final-attempt-retained is checked by requiring the attempt numbers to be exactly 1..n: a pipeline that keeps the attempt that worked reports a cost short by the retry factor, and the operations that retry most are the ones being under-reported. schema-loosened-between-attempts is the escape hatch reopening — the bar moved after somebody saw the output — and prior-output-fed-back-above-a-generated-zone is the S5.3 tie: a correction prompt is the shortest path from "the model said something wrong once" to "the model was instructed to say it". --> <!-- S5.4 rollup: 3323 kit tests pass (57 new); tsc clean on lib and spec; export map at 70 entry points, --check clean; all seven gates pass. Two barrel collisions caught by tsc only after the index edit (semanticFaults with error-mapping.ts, provenanceFaults with blobs.ts, plus Quantity with diff.ts) — renamed to the result*/validation*/Output* families. The packed-consumer gate again caught import-time work (new Set for the whitespace table). -->
  • S5.5 Preserve provider response provenance before normalization. Evidence: raw response hash and normalized projection are linked. <!-- src/response-provenance.ts. The failure this prevents is a normalized record that is the ONLY record. Normalization is lossy and not reversible: once a response has been mapped into the shape the rest of the system wants, "what did the provider return?" can only be answered from that shape — a projection made by an adapter whose version nobody wrote down. The day the adapter turns out to have been wrong, every artifact downstream is wrong and there is nothing left to re-derive them from. Six fault families, 44 faults, all producible. -->
    • S5.5.a Capture provider operation/request IDs, provider/model/deployment/ version/region, parameters, tool/template/policy versions, actor, and timestamps before normalization. Evidence: run metadata tests. <!-- captured-after-normalization is load-bearing and invisible in a working system: everything reconciles until the first response that makes the adapter throw, which produces no record at all — so the population with no provenance is exactly the population somebody would want to look at. model-without-a-deployment is the quiet one: on every hosted platform the model NAME is not the thing serving the request, two deployments of one name differ in version, region, quota and content filter, and a record naming only the model cannot tell a regression in one from a regression in the other. The tool-version check is scoped to the tools the run could REACH rather than the ones it called, since a tool that changed behaviour makes two runs incomparable whether or not it fired. Parameters are recorded as a DIGEST so two runs can be compared without copying temperature and seed into every row. -->
    • S5.5.b Store or securely reference exact raw response bytes/event stream and compute hash/length/media type under retention policy. Evidence: raw response integrity tests. <!-- A hash without the bytes is an assertion, not evidence: a digest says whether bytes you already hold are the recorded ones and cannot say what they were. So an absent artifact must carry a REASON, and the four are kept apart because "deleted on schedule" and "the write silently failed" look identical in every listing and the difference is discovered during an incident. A reference AND an absence reason together is refused, since one of the two is stale and nothing says which. Zero length is checked against the SHA-256 of zero bytes rather than treated as empty, because a length taken from one place and a hash from another have already disagreed. Stream events and single bodies are checked in both directions. Digest algorithms come from blobs.ts's approved set rather than a second list. -->
    • S5.5.c Record transport/status/headers allowlist, finish reason, usage, safety, provider errors, cancellation, and partial/truncation state. Evidence: provider outcome fixtures. <!-- Headers are an ALLOWLIST and never a capture: storing them wholesale stores the authorization echo, the set-cookie, the rate-limit token and whatever the provider starts sending next quarter, into the record most likely to be exported, replicated and kept longest. A separate FORBIDDEN set is checked as well as the allowlist, because the allowlist is a list somebody edits and the forbidden set is what must survive the edit. blocked-safety-verdict-on-a-complete-outcome is the contradiction worth catching: the artifact will be used and the safety record will say it never should have been, and both are true in the store. A 200 carrying a provider error code is kept apart from a 500 with none, since a reader that switches on the status has already treated the first as a result. -->
    • S5.5.d Normalize through a versioned adapter and store adapter version, schema, validation result, warnings, and normalized hash. Evidence: normalization provenance tests. <!-- The adapter version is part of the OUTPUT, because two deployments running different mappings over identical bytes produce different records and nothing can otherwise tell which produced which. normalized-digest-equal-to-the-raw-digest catches an adapter that did nothing, which is not hypothetical: a passthrough left from a migration, a branch that returns its input on an unrecognised shape, and a digest computed over the wrong variable all produce it and all three look like a normalization that ran. A failed validation whose projection was KEPT is a fault and one whose projection was discarded is not — the failure is a bad projection sitting in the store beside the good ones, where the next query does not join on validation status. -->
    • S5.5.e Link every normalized field/output/score to originating response or mark derived/defaulted/rejected explicitly. Evidence: lineage tests. <!-- Four lineages because they are four different things. defaulted must not fold into supplied: a confidence of 0.5 the adapter filled in because the provider sent none reads, in every chart and every threshold, exactly like one the model computed — so a default may name no source and must give a reason. supplied-field-with-more-than-one-source makes the supplied/derived line enforceable: a value assembled from two places is a computation however simple, and an audit tracing it lands on one input and concludes the provider said something it did not. A rejected field must be ABSENT from the projection, since every reader takes the projection. And coverage runs the other way too — a projection field nothing traces may have come from the response, a computation or a constant, and all three read the same. -->
    • S5.5.f Enforce authorization, encryption, redaction, retention, legal hold, and deletion for raw provider content. Evidence: provenance data lifecycle tests. <!-- Retention, holds, encryption and purge are NOT reimplemented here: deletion.ts, protection.ts and blobs.ts already hold every other artifact in the workbench to them, and a second copy for provider content would be the one that misses the next hold. What this owns is the composition and the two rules specific to this data. Raw provider content takes the STRICTEST classification of everything on both sides of normalization, because it holds what the redactor removed on the way to the projection — classifying it from the projection classifies it from the copy with the sensitive parts already taken out. raw-readable-by-more-roles-than-the-projection is the first specific rule: access is granted on the projection and then extended to "the underlying record" by whoever is debugging, and the raw response is the unredacted original. projection-outliving-the-raw-it-came-from is the second and is the inversion people expect — keeping the derived record after the original is gone leaves an artifact nothing can substantiate. A purge under an open hold is refused, since a hold a document purge respects and a transcript purge does not is a hold with a gap in it. --> <!-- S5.5 rollup: 3366 kit tests pass (43 new); tsc clean on lib and spec; export map at 71 entry points, --check clean; all seven gates pass. Four more barrel collisions caught only after the index edit (ProviderOutcome with router.ts, NormalizationFault with properties.ts, LINEAGE_KINDS/LineageKind with revisions.ts) — renamed to *Record/FIELD_LINEAGE_*. -->
  • S5.6 Implement cancellation, timeout, retry classification, and non-retryable safety refusals. Evidence: deterministic fault-injection tests. <!-- src/provider-lifecycle.ts. The failure this prevents is a system where every one of these is PRESENT and none of them reaches the provider: cancellation stops the spinner while the tokens keep being generated, a client timeout fires while the call runs to completion on the far side, and a retry loop treats a safety refusal as transient and asks eleven more times. All three look correct in the code that contains them; what is wrong is what they do not reach. Seven fault families, 45 faults and refusal reasons, all producible; the deterministic part of "deterministic fault-injection" is that the jitter is a PARAMETER. -->
    • S5.6.a Propagate user/job cancellation through queue, worker, provider request/stream, tools, and downstream side effects. Evidence: cancellation boundary tests. <!-- Cancellation propagates SOURCE-FIRST, and the order is read from CANCELLATION_STAGES rather than restated in the check. The other order has a race with a name: stopping the worker while the queue is still live hands the same job to the next worker, and the cancellation is now racing the redelivery it caused. A stage that cannot be cancelled must NAME ITS COMPENSATION — an email has been sent, a webhook delivered — because "not reachable" with nothing beside it is a cancellation that reported success and changed nothing, which is worse than one that failed since nobody looks again. reported-cancelled-while-a-stage-is-still-running is the one users meet: the request leaves the interface, the provider call runs to completion and is billed, and the only visible consequence of cancelling was that nobody is watching. Timestamps are checked as well as ranks, because a stage can be listed in order and reached out of it. -->
    • S5.6.b Define connect, first-byte, idle, total, tool, stage, and job deadlines and preserve partial/unknown outcome. Evidence: clock tests. <!-- All seven, and DEADLINES NEST — the nesting is the check. A connect timeout of 60s under a total of 30s means the total can never fire, because the thing it bounds is still inside a connect that has not given up; every layer here has a version of that bug. An expiry is an UNKNOWN OUTCOME and never a failure (S5.1's third decision reached from the other side): a deadline says how long WE waited and nothing whatever about what the far side did, so recording it as failed is how a charged request is retried. succeeded-with-partial is refused for the opposite reason — the part that arrived before the timeout is being handed on as the answer, and every reader downstream takes it for a complete one. And the partial must be RETAINED, since it is the only evidence of how far the call got and the retry that follows starts from nothing. -->
    • S5.6.c Classify throttle/capacity, transient network, timeout, unavailable, malformed output, policy refusal, safety refusal, authorization, invalid input, quota, and unknown errors. Evidence: classification table tests. <!-- All eleven as ONE TABLE with a retryability (from errors.ts), a family (from error-mapping.ts), whether the outcome is unknown and whether it may already have been charged. A classification spread across eleven if statements in six adapters is eleven decisions nobody can read together, and they will disagree. malformed-output is retryable and is not a refusal, because the model produced something and it was the wrong shape — the one failure a second attempt genuinely fixes. The test names the three cells where a retry can duplicate a charge (timeout, transient-network, unknown), which is the whole reason S5.6.f exists. AN EARLIER CHECK WAS REMOVED for the S4.15 reason: "a family shared by more than one class" fires on the CORRECT table, since transient-network, timeout and unavailable are all rightly dependency — the remedy for all three is to wait. -->
    • S5.6.d Retry only approved transient classes with idempotency, bounded attempts/age/backoff/jitter, budget, and cancellation. Evidence: retry schedule tests. <!-- Full-jitter backoff — a uniform draw across the whole window rather than the window plus a wobble, because equal jitter leaves every caller retrying in the same half-second and the thundering herd survives it — with the jitter supplied as a PARAMETER so the schedule is testable and the function stays deterministic. The budget's fourth bound is the one that matters: a per-WINDOW retry share ceiling, because per-request bounds are what turn a partial outage into a full one (every caller retries three times, the provider sees four times the load at exactly the moment it is failing, and what was degraded is now down). The refusal ORDER is the order of certainty: cancellation first, since a retry after the user walked away is work nobody will read; then the refusal classes, which no budget makes retryable; then the unknown-outcome cells last, because those need the reconciliation to have run. -->
    • S5.6.e Never automatically retry or prompt-around safety/policy refusal, unauthorized action, invalid consent/rights, or nonretryable validation. Evidence: refusal negative tests. <!-- prompt-changed-after-a-refusal is the whole point, and it is why an attempt record carries a PROMPT DIGEST: an automatic re-ask with different wording after a safety refusal is indistinguishable from a legitimate retry unless the two prompts are compared. It is not error handling — it is the system looking for a phrasing that gets the answer that was declined, run at machine speed with nobody in the sequence, and it is reached by ordinary-looking code that regenerates on failure. The humanInitiated exception is narrow and deliberate: a person who reads a refusal and edits their request is doing something the system should support, and the fault is the loop that does it without them. refusal-reclassified-as-transient-on-a-later-attempt catches the other route — whichever classification is wrong, a retry loop takes the permissive one. -->
    • S5.6.f Reconcile external request status after local timeout/worker loss before repeating costly or irreversible operations. Evidence: unknown-outcome tests. <!-- retries.ts already decides this for an effect, on REVERSIBILITY. What is added here is COST as a separate axis: a model call is usually reversible in the sense that nothing durable changed, and never reversible in the sense that matters to a budget, so a pipeline reasoning only about reversibility repeats an expensive generation freely and discovers it at the end of the month. A repeat with no provider request id is refused outright, since there is no handle to ask with and the only available answer is a guess that costs money. And the indeterminate cell refuses rather than retrying — it is the cell people fill in with "retry" because the alternative needs a person, and it is the one that produces the duplicate charge. -->
    • S5.6.g Persist attempt history, partial outputs, costs, reason, and final state without false success. Evidence: fault-injection report. <!-- succeeded-with-no-successful-attempt is the false success in its plainest form and it is reached honestly: a catch falling through to a default result, a state machine whose terminal transition is unconditional, a reconciliation that filled in a value — all three produce a run that reads as having worked. total-cost-below-the-sum-of-attempts is the same shape applied to money, since the attempts that failed cost exactly what the one that worked cost. attempt-numbers-not-contiguous is how a dropped attempt is detected at all: its cost and its partial output went with it. And a run may not be closed while it still contains an attempt whose outcome nobody knows. --> <!-- S5.6 rollup: 3414 kit tests pass (48 new); tsc clean on lib and spec; export map at 72 entry points, --check clean; all seven gates pass. Four more barrel collisions (REFUSAL_CLASSES, retryDecision, RetryDecision, StageEffect) renamed to the provider*/Cancellation* families. The public-api spec catches a stale generated map before the --check does, so regenerate before running the suite. -->
  • S5.7 Implement per-run cost estimate, reservation, actual cost, and budget variance. Evidence: over-budget dispatch is blocked before provider invocation. <!-- src/cost-accounting.ts, the MONEY layer above limits.ts — which already owns SpendBudget, available, reserve and settle and is scope-agnostic on purpose. The failure this prevents is a budget that can only be EXCEEDED: cost measured after the work arrives too late to change anything, and the amount by which it is wrong is unbounded, because one runaway job spends a month's allowance and the limit reports it the following morning. Seven fault families, 47 faults and refusal reasons, all producible. -->
    • S5.7.a Maintain versioned provider/model/operation/tool/input/output/ region unit prices, currency, effective period, and source. Evidence: rate-card fixtures. <!-- Money is INTEGER MINOR UNITS at a stated scale, because a price of $0.000003 per token is not representable in cents and is exact in millionths of one, and floating-point money accumulates a discrepancy across a million calls that nobody can reconcile afterwards. Currency travels with every amount rather than being a deployment constant, since two providers billing in two currencies summed into one "cost" produces a number in no currency that will be compared against a budget in a third. gap-between-periods-for-one-key is the one that costs money quietly: an overlap is noticed because two prices disagree, and a gap is not, because a lookup that finds nothing returns nothing and the caller adds nothing to the total — so the run in the gap is free, on the day the price changed, which is the day the totals are being watched. An abutting pair (one period ending exactly where the next begins) is correct and is not reported. And a price must name a SOURCE, because one nobody can trace to a page, a contract or an invoice is a number somebody typed and it is what every total is built on. -->
    • S5.7.b Estimate cost from validated request, context, output limits, retries/tools/derivatives, and uncertainty before acceptance. Evidence: known-price tests. <!-- An estimate is an INTERVAL and its upper bound is RE-DERIVED by the checker rather than trusted, which is the only way this is worth having: an estimator that quietly uses the expected output length produces an interval that looks exactly like a correct one and is wrong by the ratio between typical and maximum — for a generation, often a factor of ten. A budget sized on the typical case is exceeded by a run that behaved perfectly legally, and that run is not a bug anybody can fix. The derivation must cover the retry bound and the tool bound as separate terms, both checked by re-deriving with that bound removed and asserting the stated ceiling is larger. Derivative operations (an embedding of the output, a moderation pass) must appear in a term — matched on a WHOLE TOKEN after a bare includes was found to let an id of x match maxToolCalls, which is a check reporting coverage it does not have. And an operation with no declared output ceiling is refused rather than estimated: the upper bound of an unbounded output is not a number. -->
    • S5.7.c Reserve actor/project/tenant/provider budget atomically with run acceptance and reject concurrent oversubscription. Evidence: budget race tests. <!-- limits.ts decides whether an estimate FITS, and that decision is taken against a snapshot — so between reading availability and writing the reservation, any number of concurrent runs each individually fit, and the budget is oversubscribed by exactly the number of workers. read-then-write-without-a-version is that race named, and the shipped contract is a conditional write carrying the version each scope was read at. partial-reservation-across-scopes is the other half: a run draws on four budgets, and if two are written and two fail, budget is held against a run that was not accepted and nothing will release it because nothing recorded it. A reservation must also EXPIRE, or a worker that crashes between reserving and settling shrinks the budget by one run permanently, with nothing in the system spending it. -->
    • S5.7.d Recheck reservation immediately before each provider/tool dispatch and prevent invocation when insufficient or expired. Evidence: predispatch failure tests. <!-- The gap between reserving and dispatching is where the queue is, and it can be hours; a reservation taken at acceptance and never rechecked is a reservation against a budget since spent by everything that queued behind it, and the first anybody hears of it is the invoice. An expired reservation is REFUSED rather than renewed — renewing it silently grants budget that was not available at acceptance, on the strength of the run having waited. The comparison is against the dispatch's UPPER bound rather than its expected cost, because dispatching on the expected cost permits a legal run to overrun the check that just passed. -->
    • S5.7.e Capture provider-reported and independently derived actual usage/ cost per attempt/tool and handle missing/disputed billing. Evidence: metering tests. <!-- TWO measurements, because the provider's number is simultaneously the one you are billed on and the one you cannot check; the independent derivation from the recorded request and response is not a better number — it is the only way to notice that the billed one moved. Missing provider usage recorded as zero is a fault rather than a value. tolerance-wide-enough-to-hide-a-doubling exists because a tolerance is where a disagreement goes to be forgiven, and it is set by whoever is tired of the alert: at 100% it silences a bill that doubled, which is the one event the comparison exists for. -->
    • S5.7.f Release/refund unused reservation and reconcile estimate, reserved, actual, cancelled, failed, retried, and unallocated values. Evidence: reconciliation report. <!-- Reconciliation is an EQUATION: reserved = settled + released + still-held. A residual nobody can name is budget that has leaked, and it leaks in the direction of LESS apparent spend, because it is money that was held and is now attributed to nothing. budget-still-held-on-a-finished-run is how the leak usually happens — the happy path settles and releases and the failure path returns early — and held budget on a run that is still RUNNING is correct and is not reported. A cancelled run that released nothing shrinks the budget every time somebody presses cancel. -->
    • S5.7.g Record absolute/percentage variance and stable reason and alert on threshold anomaly. Evidence: variance scenario tests. <!-- BOTH thresholds must be crossed. A one-penny estimate that came in at nine pence is 900% over and worth nobody's morning; a five-percent overrun on a four-thousand-pound batch is worth somebody's — and a channel that fires on the first is muted before the second happens, so an alert on the relative threshold alone is a fault in its own right. The reasons are a stable closed set so variance can be grouped, and unexplained on a variance past both thresholds is reported, because that is the one worth investigating and it is the one the record gives an investigator nothing to group by. A zero estimate yields 0 or Infinity rather than a NaN somebody charts. --> <!-- S5.7 rollup: 3471 kit tests pass (57 new); tsc clean on lib and spec; export map at 73 entry points, --check clean; all seven gates pass. Only one barrel collision this time (ReconciliationFinding with staleness.ts), caught by grepping the planned export names BEFORE writing the index block. -->
  • S5.8 Extract the generic dependency/claim graph without hard-coding Tara's claim kinds. Evidence: Metis evidence, Veritas sources, Yemaya canon, Isis provenance, and Euterpe rights fixtures compile. <!-- src/dependency-graph.ts, and the module is not called claim-graph.ts for the reason the item exists. The failure this prevents is a shared library extracted from one domain that still speaks its language: every other domain then either translates itself into the first domain's words or forks, and both are worse than not sharing. The tell is always in the NAMES. The kit owns the SHAPE and the plugins own the TYPES — a node kind is a string a plugin registers, and what is fixed here is structural: which endpoint depends on which, whether a kind is versioned, whether an edge may leave its scope, whether cycles over it are a fault, and what a bounded traversal means. Five fault families, 33 faults, all producible. -->
    • S5.8.a Define plugin-owned node/edge type registry atop shared stable identity, version, direction, source, classification, and lifecycle. Evidence: graph extension contract. <!-- DIRECTION is a property of the edge KIND rather than of the traversal: if each caller decides which way is "upstream", the word has as many meanings as there are call sites, and an impact query in one screen answers a different question from the identically named one in the next. edge-into-a-versioned-kind-that-does-not-pin is the fault that produces wrong ANSWERS rather than errors — an edge to "the current version" is still an edge after the target changes, nothing fails, nothing is emitted, and every conclusion drawn through it now rests on something the graph never said. neither is a real value for a symmetric relation, and symmetric-edge-declared-acyclic exists because such an edge is a cycle by construction, so the acyclicity check would fire on correct data — the S4.15 failure, refused in advance. Every kind must name an OWNER, since an unowned kind is one the kit is assumed to define, which is exactly how a domain's vocabulary reaches a shared surface. -->
    • S5.8.b Define generic add/remove/version/supersede/resolve and bounded upstream/downstream/path/impact query interfaces. Evidence: operation tests. <!-- The traversal is BREADTH-first, and that is not a preference: a depth-limited depth-first walk returns one deep path and none of the near neighbours, which is the opposite of what anybody asking "what does this affect" wants back. Truncation is part of the ANSWER and is reported by which bound was hit, because a bounded query that does not say it stopped early returns a small result that reads as a complete one and the reader has no way to tell. A neither edge is traversed in neither direction, so an impact query cannot report an impact through a relation that asserts no dependency. Paths are bounded on count as well as depth, since the number of paths between two nodes in a dense graph is exponential and the query that discovers this is the one running in a request. -->
    • S5.8.c Enforce node/edge schema, same-scope rules, duplicate/cycle/ dangling policy, and immutable historical version references. Evidence: graph integrity tests. <!-- Cycles are checked PER EDGE KIND, and that distinction is the difference between a useful finding and a permanent one: real graphs contain cycles across kinds — A annotates B while B supersedes A — and reporting those would report every real graph. What is a fault is a cycle within a kind that declared itself acyclic, because that kind's own traversals will not terminate and the BOUND is what will stop them rather than the data. The detector is iterative with an explicit stack, since a graph deep enough to matter is deep enough to exhaust a recursive walk. pinned-version-that-is-not-present is the case pinning exists for — a reference to something that has since been rewritten. And a duplicate edge is a fault because every count through it doubles, which puts every de-duplicating reader in disagreement with every counting one. -->
    • S5.8.d Apply tenant/object/property/protected-identity authorization to nodes, edges, counts, paths, and graph errors. Evidence: disclosure negative tests. <!-- An edge survives the projection only when BOTH endpoints do: an edge with one visible end is not a partially visible relation, it is the statement "this thing you can see is connected to something you cannot", which is precisely the fact being withheld. count-computed-before-projection is the leak nobody notices, because the node LIST is correct — fourteen reported, three returned, and the difference is eleven facts about a graph this reader has no access to. path-returned-with-its-middle-elided is the same idea one level up: the elision asserts the connection. And S5.8.d is the one place in the kit where NOT-FOUND and FORBIDDEN are deliberately the SAME answer — everywhere else they are kept apart, and here they cannot be, because the only thing separating them is the existence of the node and existence is what is being denied. -->
    • S5.8.e Build Metis evidence, Veritas source/claim, Yemaya canon, Isis artifact/provenance, and Euterpe rights fixtures using only plugin types. Evidence: five compiling fixtures. <!-- Five registries in dependency-graph.spec.ts, each written as its own domain would write it: Metis exhibits and findings, Veritas assertions and attributions, Yemaya works and editions, Isis renders and input assets, Euterpe recordings and licences. No two share a node kind or an edge kind, and the test proves it pairwise. The claim is scoped to KINDS and deliberately NOT to lifecycle states: the first version of that test failed for the right reason, because Metis withdraws an exhibit and Euterpe withdraws a recording — English having a word for it is not two domains sharing a model, the kit never reads a state name, and a test demanding five disjoint English vocabularies would be enforcing a thesaurus. Yemaya's sits-beside is carried through as a symmetric edge without a direction being invented for it. -->
    • S5.8.f Verify shared source/export names contain no Tara-specific claim, meditation, stage, authority, or policy semantics. Evidence: API review. <!-- A CHECK rather than a reading. domainVocabularyFindings splits an identifier the way a reviewer reads it (camelCase, PascalCase, SCREAMING_SNAKE) and matches WHOLE WORDS, so reclaimed is not a claim and stagedFile is a stage — the first is the false positive that gets a check deleted and the second is the real one. The vocabulary is a PARAMETER and not shipped, because a kit holding a list of one domain's words is a kit that knows them; the spec derives the corpus from the five fixtures' own registries plus the words this item names by hand. And the check runs against Object.keys(import * as dependencyGraph) — the module's real runtime exports rather than a list somebody maintains. --> <!-- S5.8 rollup: 3509 kit tests pass (38 new); tsc clean on lib and spec; export map at 74 entry points, --check clean; all seven gates pass. Pre-checking the planned export names caught nine collisions before they were written; DependencyNode slipped through against diff.ts and became GraphEntity. -->
  • S5.9 Implement authority/severity policies as domain plugins. Evidence: one domain cannot silently use another domain's floors. <!-- src/policy-plugins.ts. The failure never arrives as a decision; it arrives as a FALLBACK. Somebody's lookup returns undefined, the crash is fixed with ?? defaultPolicy, and from then on a domain with a two-independent-source rule is judged by whichever policy happened to be registered first — producing verdicts that are the right shape, pass every schema, and answer a question nobody asked. Five fault families, 36 faults and refusal reasons, all producible. -->
    • S5.9.a Define plugin identity/version/domain, applicable node/edge/gate/ action types, inputs, outputs, and ownership. Evidence: policy plugin contract. <!-- applies-to-nothing is the one that matters: a policy declaring no scope can be selected for anything, which defeats the applicability rule by OMISSION rather than by a mistake anybody would review. The version must be pinned because an evaluation records which policy judged it, and an unpinned record names a moving target that no past verdict can be reproduced against. domain-prefixed-capability is small and saves a class of confusion — veritas.sourcing inside the veritas domain produces the key veritas/veritas.sourcing, and the day somebody normalises those keys two domains collide. -->
    • S5.9.b Define authority ranking/independence, source/evidence sufficiency, conflict, protected/anonymous source, and unavailable semantics per domain. Evidence: authority policy fixtures. <!-- Independence is an AFFILIATION PARTITION rather than a pairwise relation: counting a maximum independent set over an arbitrary relation is intractable, and the domain question is not arbitrary — two sources from one press office, one lab or one rights holder are one source, and a rule that counts documents is satisfied by one office sending two letters. UNAVAILABLE is not the lowest rank: ranking it bottom turns "we could not check" into "we checked and it was weak" and lets the computation proceed. It produces UNDETERMINED — but only when it could have changed the answer, because sufficient without it stays sufficient and insufficient counting every unreachable source optimistically stays insufficient, which is what makes a third value survivable in a system that has outages. A protected source counts and is never named, and neither is the COUNT of them, since "met, on three sources, one withheld" has disclosed that a protected source exists. conflict-rule-with-no-tie-rule is the branch nobody writes: higher-rank-wins cannot resolve two sources at the SAME rank saying opposite things, which is the situation a conflict rule exists for. -->
    • S5.9.c Define severity classes, thresholds, escalation, blocker/warning, target/audience, and remediation semantics per domain. Evidence: severity policy fixtures. <!-- blocking-class-below-a-non-blocking-one is the shape that reads as fine and is not: if a middling finding blocks and a worse one only warns, then the way to ship is to make the problem worse, so blocking has to be monotone in severity or the policy rewards the wrong move. A blocking class must name what CLEARS it, or the only way forward anybody finds is the override. Escalation is walked rather than assumed, so a downward escalation (an unresolved finding getting quieter over time) and a cycle are both reported. And classifyMeasurement returns the HIGHEST class cleared rather than the first matched — a scan that stops at the first match reports a critical finding as a warning whenever the classes are read in the order they are declared. -->
    • S5.9.d Register exact policy by domain/capability/version and reject missing, duplicate, wrong-domain, or unsupported policy. Evidence: registry negative tests. <!-- No default and no nearest match. Every refusal here is a case where some implementation somewhere would have returned SOMETHING — the policy from the neighbouring domain, the previous version, the first of two registrations — and every one of those produces a verdict of the right shape about a question that was not asked. The wrong-domain refusal is detected rather than merely missed: when the capability exists but under another domain, the refusal SAYS so, because "not found" would send the next person to register a duplicate. -->
    • S5.9.e Persist selected policy/version with each evaluation and prevent caller-supplied weakening below code/domain floors. Evidence: downgrade tests. <!-- A caller may NARROW and may never widen. The direction of strictness is read from the setting rather than assumed, because a maximum-staleness threshold widens by going UP while a required-source count widens by going down — a checker that assumed one direction would wave the other through. Both floors are checked because they fail differently: the code floor is what this system will not do at all, and the domain floor is what this domain decided for itself. A domain floor found BELOW the code floor is a configuration that was accepted once and has been quietly ineffective since, while whoever set it believes it is in force. Refused rather than logged, because a logged weakening is a weakening. -->
    • S5.9.f Run cross-domain fixtures that deliberately attempt policy substitution and verify refusal. Evidence: isolation tests. <!-- Two registered domains that genuinely DISAGREE about the same evidence — Veritas needs two independent affiliations before an assertion is published, Metis needs one because an exhibit is examined rather than corroborated — and the spec asserts the disagreement first, because without it "substitution is refused" is a rule with nothing at stake. Four routes are then attempted and all four refused: asking as one domain for a subject only the other declares, asking as one domain at the other's version, asking under a domain nobody registered, and asking for the right domain with the wrong policy kind. A fifth case asserts Metis can still reach its OWN policy, so the refusals are a boundary rather than a wall around everything. --> <!-- S5.9 rollup: 3547 kit tests pass (38 new); tsc clean on lib and spec; export map at 75 entry points, --check clean; all seven gates pass. Pre-checking the export names against src/*.ts before writing the barrel found the one collision (classifySeverity) in advance; it became classifyMeasurement. -->
  • S5.10 Implement recomputation and blast-radius planning. Evidence: preview is read-only; apply creates jobs and audit. <!-- src/recomputation.ts, composing S5.8's bounded traversal and S5.7's money. The failure this prevents is a preview that is not one: a blast-radius view is the thing a user opens most often — on every edit, on every hover, from a table of two hundred rows — and every write it performs happens at that rate. Six fault families, 34 faults and refusal reasons, all producible. -->
    • S5.10.a Accept exact changed node/version/event and policy context and resolve affected direct/transitive results, gates, artifacts, reviews, releases, and projections. Evidence: impact query tests. <!-- The query names a VERSION of the node and a VERSION of the policy, because a radius computed from "the node" describes whatever the node is when somebody reads the answer, and an unpinned policy lets the plan and the run be judged differently. Six affected KINDS, structural rather than any domain's nouns. direct-impact-at-distance-zero catches the subject listed as its own consequence — or, more often, a distance nobody computed. -->
    • S5.10.b Bound traversal by tenant, authorization, edge type, depth, fan-out, cycles, and pagination while representing inaccessible/unknown impact. Evidence: traversal safety tests. <!-- The bounds and the cycle handling come from S5.8's traverseDependencies rather than a second traversal. What is added here is REPRESENTATION: inaccessible impact is one disclosed bit rather than an omission, and the argument for disclosing it is the alternative — a user applying a change believing it touches three things when it touches thirty has been misled by a system that was being careful. A truncated radius presented as the radius is the same failure as S5.8's, one level up, and worse here because the number is about to decide whether to proceed. -->
    • S5.10.c Compute deduplicated dependency order, required versus optional work, stale-only versus recompute, provider/resource needs, cost/time, blockers, and releases at risk. Evidence: plan fixtures. <!-- FULL RECOMPUTE IS NOT THE SAFE DEFAULT. It costs more and it discards human work: recomputing a gate somebody signed off throws away the signature, and re-obtaining it is asking a person to agree again to something they have already agreed to, which is how a sign-off becomes a formality. So stale-only is the default and a full recompute that discards a decision has to be ASKED FOR. required-step-depending-on-an-optional-one is the ordering mistake that survives review: an optional step may fail without failing the run, so the required step behind it either waits for something that never arrives or runs on an input that was never produced. A step listed twice for one item is two jobs racing each other for one output row, and a total below the sum of the steps matters because the headline is what gets approved. -->
    • S5.10.d Return an immutable versioned preview with no persistence, provider invocation, queue write, notification, or audit-as-mutation. Evidence: read-only side-effect tests. <!-- Eight forbidden effects enumerated, including the two the ledger does not name — a cache entry and a lock — because those are the writes that get called not-writes. The AUDIT write is reported as its own fault kind rather than folded in with the rest, and that separation is the point: the others reach production as bugs, and the audit row reaches it as a decision somebody defended in review. A preview that records that it happened has happened, and it happens at the rate a user hovers over a row. The returned plan must also be immutable, or the caller can edit it between preview and apply and the digest apply revalidates describes a plan nobody produced. -->
    • S5.10.e Revalidate source/dependency/policy versions, authorization, budget/capacity, and plan hash when applying. Evidence: stale-plan tests. <!-- The ORDER of the refusals is the order of certainty and it is checked: the digest first, because every later check would be answered about the wrong plan; authorization next, so a caller who may not do this does not learn the plan's contents from the refusal; then freshness, then budget and capacity, then the radius approval — last, because it is the only one a person can resolve in the moment. A plan older than its freshness bound is refused EVEN WHEN THE DIGEST MATCHES, since a matching digest only says nothing changed that the digest covers. And refusing on budget is refusing to leave a graph half recomputed, which is the state with no correct reading at all. -->
    • S5.10.f Create idempotent parent/child durable jobs in dependency order and a correlated audit/notification record. Evidence: apply integration tests. <!-- The idempotency key is derived from the plan digest and the step, so a re-apply of one plan is one job. job-order-not-matching-the-plan compares the job's wait-list against the plan's dependency edges rather than trusting that they agree. And apply-with-no-audit-record is the deliberate MIRROR of S5.10.d: the preview must not audit and the apply must, and a system with those two backwards has a trail of everything nobody did and none of what they did. -->
    • S5.10.g Expose partial, blocked, failed, cancelled, superseded, and completed recomputation and updated release eligibility. Evidence: lifecycle tests. <!-- The two central faults are one distinction from opposite ends: required work outstanding must not read as COMPLETED, and optional work failing must not read as FAILED. A system that collapses them has one of two behaviours — it blocks releases on work nobody needed, or it releases on work that did not happen — and which one it has is an accident of which branch was written first. A superseded run must name its successor or the work cannot be followed, and a release whose eligibility moved must say why, because whoever is waiting on it has nothing else to act on. --> <!-- S5.10 rollup: 3585 kit tests pass (38 new); tsc clean on lib and spec; export map at 76 entry points, --check clean; all seven gates pass. Four barrel collisions (OUTCOME_FAULTS, outcomeFaults, OutcomeFinding, PlanFault) renamed into the recomputation* family. -->
  • S5.11 Add policy version and evaluation timestamp to every trust result. Evidence: policy changes invalidate dependent results. <!-- src/trust-results.ts. The failure this prevents is a system that gets MORE WRONG as its policy gets better: an unstamped result is read a year later as though it had been evaluated under today's rules, so every improvement silently reinterprets a body of old results as agreeing with it — and the better the new policy is, the more confidently wrong that reinterpretation becomes. Nothing fails; the record simply means something it never meant. Five fault families, 31 faults, all producible. -->
    • S5.11.a Inventory all gate, safety, authority, rights, consent, canon, provenance, quality, fairness, and release trust result types. Evidence: result coverage register. <!-- All ten families, and the register is checked in BOTH directions: a family nothing produces a result for answers the coverage question with "none" rather than "checked", and a declared type that carries no stamp is named. persisted-type-whose-row-omits-the-stamp is the one that produces the problem years later — the in-memory result carries its policy version, the write drops it because the column was added afterwards, and every reader of the table sees results that appear to have been evaluated under the current policy. The schema is right and the data is not. A type that is never persisted is deliberately not reported for it. -->
    • S5.11.b Require stable policy ID and immutable semantic version/hash in each result schema and persistence row. Evidence: schema tests. <!-- The HASH is required beside the version and is not redundant: a version is what somebody remembered to bump, and two deployments can hold different bytes under one number for as long as it takes anybody to notice — the version is for a human deciding whether the rule changed and the hash is for a machine deciding whether the bytes did. The version must be SEMANTIC, or a reader cannot tell a clarification from a change of rule and every policy edit invalidates everything or nothing. The EVALUATOR version is stamped too, because a bug fixed in the evaluator changes results the policy did not, and without it those results are indistinguishable from the corrected ones. -->
    • S5.11.c Record evaluation start/completion/reference clock, expiry, and source/evidence observation time with explicit semantics. Evidence: temporal tests. <!-- FOUR times, not one. A result computed at 10:00, about evidence seen at 09:00, evaluated as of 08:00 has three different answers to "when", and a system with one field has picked one of them for everybody. The REFERENCE clock is load-bearing: a rule like "within thirty days" needs an anchor, and an anchor of "now" makes the same evidence pass today and fail on Thursday — so reference-clock-equal-to-completion is reported, and reported ONLY when the result declares a freshness window, since a result with no window is not made wrong by anchoring on itself. EXPIRED IS NOT FAILED: an expired result says the answer is too old to use and nothing about the subject, and collapsing the two turns a stale freshness window into an accusation. -->
    • S5.11.d Include evaluator/corpus/model/threshold/input/dependency versions needed to reproduce or explain the result. Evidence: provenance completeness tests. <!-- The check takes what was USED as a separate argument rather than reading it back out of the record, and that is the whole design: a completeness check that reads only the record can report internal consistency and nothing else, so a corpus that was consulted and never recorded is invisible to it. The reverse is checked too — a threshold in the provenance that no code applies is not pedantry, it is a number somebody will tune expecting an effect. And a dependency read without a version is what breaks S5.11.e: a change to it cannot mark this result stale. -->
    • S5.11.e Emit policy-change events and mark all affected current results stale before downstream authorization. Evidence: invalidation tests. <!-- The ORDER is the guarantee. Marking results stale is easy; marking them before anything reads them is the problem — a gate that read a result in the window between the policy changing and the sweep finishing passed on an answer already known to be wrong, and it will not be asked again. rule-change-recorded-as-a-clarification is how the sweep gets skipped legitimately: a clarification invalidates nothing, so labelling a rule change as one is the cheapest possible way to avoid re-evaluating a large population, and the label is chosen by whoever is shipping the change. The over-wide sweep is a fault too, since re-evaluating work that was correct costs money and re-opens decisions somebody had closed. -->
    • S5.11.f Reject or migrate legacy results missing policy/time rather than assuming current policy. Evidence: legacy compatibility tests. <!-- Three answers and never a fourth. migrate requires an EXTERNAL source that establishes which policy version was live during the evaluation window — a deployment log, a release record — because a migration that infers the version from the current one is exactly the assumption this section exists to refuse, performed once and then invisible. Everything else is reject, and rejecting is not the destructive option: a rejected legacy result is re-evaluated and becomes correct, while an accepted one is wrong and looks identical to a right one for ever. A stamp that is PRESENT and incomplete is rejected rather than accepted. --> <!-- S5.11 rollup: 3620 kit tests pass (35 new); tsc clean on lib and spec; export map at 77 entry points, --check clean; all seven gates pass. No barrel collisions — the pre-write name grep is now part of the recipe. -->
  • S5.12 Add safe human override hooks. Evidence: authorization, rationale, scope, expiry, notification, and audit are enforced. <!-- src/overrides.ts. The failure this prevents is a control plane that is never removed and stops applying anyway: overridden often enough, the override IS the process and the check is a formality somebody clicks past. The load-bearing decision is that an override DOES NOT CHANGE THE ANSWER — there is no requested outcome meaning "make it pass", because a system that rewrites the result destroys the only record the check ever fired, and six months later the release reads as one where everything passed. Eight fault families, 71 faults, all producible; 96 tests. -->
    • S5.12.a Register overrideable and nonoverrideable result/gate/policy types and maximum scope. Evidence: override registry. <!-- Two decisions the registry exists for. Overrideability is NOT decided by the team that owns the check: governedBy is separate from producedBy, or the check is advisory to its own owner and mandatory to everybody else, and nobody outside that team can tell which of the two they are looking at. And the lookup DEFAULT-DENIES — an unregistered subject is refused, since the other default ships every result type ever added as overrideable with nobody ever deciding. A non-overrideable entry must name an obligation OUTSIDE this system (a statute, a licence, a contract) so reversing it is an argument rather than a one-line diff, and one that still carries an approver count and an assurance level is reported: the machinery is built and a single boolean in the same file stands between it and being used. FAMILY_MAX_BREADTH is a ceiling rather than a default, and safety/rights/consent/authority/provenance sit at single-object because breadth changes the KIND of those decisions — a platform-wide safety override is not an exception, it is the check switched off by somebody who did not have to say so. -->
    • S5.12.b Require exact result/input/policy versions, requested outcome, structured rationale, evidence, residual risk, and intended consequence. Evidence: request validation tests. <!-- The rationale is STRUCTURED because a free-text reason box collects "approved, see thread" and makes three different decisions indistinguishable. the-check-is-wrong is a bug report and must carry a defect, or the same override is filed again every week for ever and the check stays wrong because nothing ever recorded that it was; the-risk-is-real-and-accepted is the only claim where a residual-risk statement means anything; the-check-could-not-decide is an absence of information somebody could go and get, so it has the shortest window in the system (72h). The binding is handed to S5.11's own stampFaults rather than restated — an override is an exception to one evaluation and the definition of "which evaluation" belongs to the module that stamps them. argument-that-cites-authority-instead-of-a-reason is deliberately narrow and has a counterexample test: a rationale that names the VP AND makes the argument passes. -->
    • S5.12.c Enforce actor/role/tenant/object/target/stage scope, step-up, separation of duties, quorum, and delegated authority. Evidence: authorization matrix. <!-- A REFUSAL IS NOT A VOTE. Systems model approvals as a count, so a refusal is a missing approval, so the remedy for a refusal is to ask somebody else — and a threshold of two is met by the third and fourth people asked while the first one's objection sits unanswered. overrideAuthorization refuses whatever the count says. Separation of duties is about the BENEFICIARY and not only the requester, who is very often a release engineer filing on another team's behalf; the reporting line is WALKED, with a cycle guard and a test that proves termination. one-actor-counted-twice is how a quorum of three is met by two people, one of whom holds two roles — ordinary in a small team and never re-counted. Delegated authority is bounded three ways: it may not compound (onward-delegation), may not hand over more than the delegator held, and past 720h is a role change that never got a role change's review. -->
    • S5.12.d Define start, expiry/one-use, review, renewal, revocation, and invalidation on dependency/version change. Evidence: clock and binding tests. <!-- The override TOLERATES NOTHING, and that is the deliberate contrast with S2.11's waiver: a waiver may declare which changes it survives because it waives a gate's verdict on a class of failure, while an override disputes ONE computed answer — if the policy, its bytes, the evaluator, the input or a dependency moved, there is a different answer nobody has read, and an exception that survives into it approves a conclusion no human has seen. amended is a STANDING rather than a validation violation and is ordered ahead of the clock: a grant whose request has been edited since approval is not a live exception with a paperwork problem. renewal-on-the-same-evidence and renewed-before-the-review-it-was-granted-with are the two ways a seventy-two-hour exception becomes a standing arrangement no single approval ever authorised. overrideGrantFaults closes the other direction from the digest: a grant that was never an answer to its request, including one that quietly lowered the severity — which understates what was accepted AND, because S5.12.e routes on that exact field, decides that fewer people hear about it. -->
    • S5.12.e Notify affected owners/reviewers/publishers/security/safety/ rights roles according to severity without leaking protected detail. Evidence: routing tests. <!-- ONE severity scale doing double duty as the residual risk and the routing key, because two scales means somebody writes a mapping and the mapping is where a critical override becomes a routine notification. The notice is part of the CONTROL: at serious and above it must land before the override takes effect, or the people who would have objected learned about it from the thing that had already happened and the decision was unilateral. The two leak faults point in OPPOSITE directions on purpose — sending the finding to the publisher broadcasts the thing the result was protecting, and withholding it from the safety desk asks them to review a decision they may not see, which is a review in name only. protectedTerms are supplied by the caller because a redaction rule that guesses redacts the wrong things and then gets switched off. The routing table is checked for MONOTONICITY, since a table that drops a recipient as things get worse makes escalating an override the way to stop that role hearing about it; the shipped table produces zero findings, and over-notification has a legitimate case (additionalRecipients) with its own test. -->
    • S5.12.f Record request, approvals/refusals, activation/use, effect, expiry/revocation, and downstream actions immutably. Evidence: audit tests. <!-- refusal-absent-from-the-trail is the fault worth the function: a trail that records the approvals and drops the refusals turns a contested decision into a unanimous one, and the reader most likely to be misled is the incident review asking whether anybody saw this coming — the answer was yes and the record says no. effect-applied-with-no-approval-before-it is the ordering that makes the rest true, because an effect applied first and approved afterwards produces a record in which every field lines up perfectly. Immutability is a digest CHAIN plus sequence continuity, so an entry edited or inserted afterwards cannot leave the trail reading as it did. And downstream-action-that-names-no-override is the question an audit is actually asked — not "was this approved" but "what shipped under it", which an action that recorded no override id makes unanswerable. -->
    • S5.12.g Surface active override and residual risk wherever the result authorizes review, publish, release, export, or operation. Evidence: end-to-end visibility tests. <!-- result-rendered-as-passing-under-an-override is decision 1 arriving at the screen: the record can be immaculate — result unchanged, override adjacent, trail intact — and a surface that draws the pair as a green tick has undone all of it for every person who will ever look, because nobody reads the record and everybody reads the screen. artifact-that-left-without-the-override is the one that outlives the system: a risk accepted inside the workbench travels with the decision, while an export or a release hands somebody a FILE, and a manifest that does not carry the exception gives the recipient a work product asserting more than anybody agreed to with no way to find out. The four disclosure fields are four different objections — no scope and you cannot tell if it covers you, no approver and there is nobody to ask, no expiry and you cannot tell a decision from a habit, no residual risk and it reads as fine — and a rendered expiry that is not the real one is worse than none, because it is believed. --> <!-- S5.12 rollup: 3717 kit tests pass (96 new); tsc clean on lib and spec; export map at 78 entry points, --check clean; all six kit gates pass. No barrel collisions — the pre-write name grep caught none of the 93 new exports. -->

S6 — Bundle assembly, provenance, publish, and verification#

  • S6.1 Define a schema-parameterized bundle builder with no arbitrary file writes. Evidence: output layout is constrained by a domain profile. <!-- src/bundles.ts plus tools/bundle-confinement.mjs (new Nx target bundle-confinement). The failure this prevents is a builder that takes an output directory and a callback — a general-purpose file writer with a manifest attached. Every domain that needs one more file adds one more callback, and within a year the "profile" describes none of what actually ships; the part that makes it expensive is that it STILL VALIDATES, because validation was written against the profile too. Six fault families, 61 faults, all producible; 43 tests. -->

    • S6.1.a Define domain profile ID/version, bundle manifest schema, allowed entry kinds, required/optional paths, multiplicity, media/schema constraints, and target compatibility. Evidence: profile contract. <!-- The fault worth the function is limits-that-cannot-hold-the-required-slots: a profile whose entry cap is below the sum of its own required minimums (plus the manifest, which is an entry too) describes a bundle that can NEVER validate, and nobody finds out until a real assembly fails at the end, having done all the work, with an error about a limit rather than about the profile. The quiet one is multi-entry-slot-whose-template-has-no-ordinal — every entry after the first resolves to the path the last one took, the assembly succeeds, the manifest lists one file, and the bundle is short the rest with nothing saying so. No slot may reach the manifest, by path or by kind, because a slot that produces the manifest lets the caller write the description of what the caller produced. Version ranges are compared NUMERICALLY, or 1.10.0 sorts inside a range ending at 1.9.9. -->
    • S6.1.b Accept immutable artifact/reference inputs and logical profile slots rather than caller-selected filesystem destinations. Evidence: builder input schema. <!-- ArtifactInput has no destination field and that absence IS the contract — the only strings a caller controls are identifiers, so slot-id-that-is-really-a-path defends the one place the decision can be attacked: a caller who wants a destination puts a path into the identifier the builder carries through, it works, and nothing else looks. An external REFERENCE contributes a manifest row and no bytes, so external-reference-with-a-byte-length matters in both directions: it makes a bundle fail a limit over data it does not contain, and it makes a bundle that points at a hundred gigabytes look like one that holds them. A reference in a byte-bearing slot is a bundle claiming to contain what it points at, and there is a counterexample test where a reference in an index slot is accepted. -->
    • S6.1.c Resolve every logical slot to a normalized profile-owned relative path and reject unmapped or duplicate slots. Evidence: layout tests. <!-- resolveLayout returns entries OR faults and never both, because a partially resolved layout is the thing a caller writes anyway; a test asserts the refusal carries no entries key at all. The manifest is not among the entries — it is written from them, so a layout containing it would be describing its own description. input-with-no-slot is reported here rather than left to assembly: by the time validation runs, the difference between "not written" and "written and not listed" has stopped being visible, and an artifact that resolves to no path is a file somebody believes is in the bundle. -->
    • S6.1.d Write only through an isolated temporary assembly root using create-new semantics and bounded resources. Evidence: filesystem confinement tests. <!-- NOT PROVABLE IN MEMORY, so tools/bundle-confinement.mjs builds a real 0700 root, performs real open(2) writes, resolves every destination with realpathSync, and runs the SHIPPED checker over what the kernel actually did — with four controls that must each be caught. CONTROL 1 is a sibling directory sharing the root's prefix (<root>-evil), which startsWith accepts and the separator-aware isInsideRoot refuses; the tool ASSERTS the naive check accepts it, so the control cannot go vacuous. CONTROL 2 is a symlink out of the root: the bytes land outside and a checker comparing the path it was ASKED to write sees a perfectly confined relative path. CONTROL 3 is a symlink WITHIN the root — containment still passes and the bytes are not where the manifest says, which is why write-that-followed-a-symlink is a separate fault from being outside. CONTROL 4 truncates over an existing file, and the run asserts the same write with wx raises EEXIST from the kernel. Verified adversarially: with isInsideRoot replaced by startsWith in the built output, the tool fails on CONTROL 1. -->
    • S6.1.e Validate complete layout, manifest, refs, hashes, profile invariants, and unused/missing inputs before archive creation. Evidence: assembly validation tests. <!-- Every digest goes through S3.11's own verifyDigest with the source the assembly REPORTS, and the outcome is carried on the finding rather than flattened, because "the file is corrupt", "only a prefix was hashed" and "the digest came from the metadata" are three different conversations with three different people. entry-on-disk-that-the-layout-does-not-name is the fault that catches everything this module cannot see — a symlink, a leftover root, a temporary file the writer forgot, a traversal that landed inside the tree — and once the archive is made it is indistinguishable from content. And validation gates the ARCHIVE rather than the publish, because a half-assembled archive that failed validation is still a file, and a file that exists is a file somebody picks up. -->
    • S6.1.f Prove domain profiles can differ without arbitrary callback/file write escape hatches. Evidence: two-profile compile/runtime fixtures. <!-- Two shipped profiles as DATA: STORY_PACKAGE_PROFILE (an optional multi-entry media slot, a signature, two targets) and DATASET_RELEASE_PROFILE (a required multi-entry dataset slot, a deeper tree, no media, two-of-eight attestations, one target whose version range excludes the other profile entirely). Both produce zero profile findings and one resolveLayout produces both layouts, asserted path by path. The escape-hatch claim is CHECKED rather than asserted: profileEscapeHatches walks a profile at runtime and reports any function, absolute path or traversal in it, takes unknown because the value that arrives from JSON, a plugin or a database has not been through the type system, and terminates on a cycle. Both shipped profiles return nothing. -->
  • S6.2 Reject absolute paths, traversal, ambiguous Unicode paths, symlinks, case collisions, duplicate refs, and reserved names. Evidence: adversarial path suite. <!-- src/bundle-paths.ts plus tools/bundle-path-adversarial.mjs (new Nx target). Every classic archive vulnerability is ONE shape: a check ran on a string and then something transformed the string. Percent-decoding, Unicode normalization, separator translation, extraction and the target's own folding are five transformations, so the pipeline is DATA, the same path is checked at each stage, and every finding carries the EARLIEST stage that saw it — a vector visible only after decoding is a demonstration that the check on the supplied string agrees with an attacker. Eight fault families, 46 faults, all producible; a 27-vector shipped corpus; 44 tests. -->

    • S6.2.a Reject POSIX, Windows drive/UNC/device, URL/file-scheme, rooted, and platform-dependent absolute paths. Evidence: absolute path corpus. <!-- windows-drive-relative (C:entry.json) is the one that gets through review: no separator after the colon, so it reads as a filename with an odd prefix and every "starts with a separator" check passes it — and on Windows it resolves against the current directory OF DRIVE C, which the process never chose. windows-device-namespace (\\?\, \\.\) is separate from UNC because the prefix exists to hand the rest of the string to the object manager UNPROCESSED: a path that would have been rejected after normalization is never normalized. And a file URL is separated from any other scheme, because the remedies differ — one is an absolute path wearing a URL and the other is content fetched at extraction time, pinned to nothing. -->
    • S6.2.b Reject dot/dot-dot traversal before and after percent/Unicode/ separator normalization and archive extraction. Evidence: traversal corpus. <!-- traversal-created-by-a-single-pass-filter is the fault worth the function: a/....//x contains no ../ and ONE pass of removing ../ produces one, so a filter written as a single replace emits the exact string it was written to remove. escapes-its-own-root is kept separate from dot-dot-segment because a/../b resolves inside and ../b does not, and telling somebody their path is an attack when it is merely untidy is how a warning gets switched off — dot-segment is reported in those words. The corpus carries a vector (a/%2e%2e/%2e%2e/x) whose traversal is INVISIBLE on the raw string and visible one stage later, and a test asserts it is invisible before decoding: that is the proof that a check on the supplied string agrees with an attacker. -->
    • S6.2.c Normalize Unicode by one policy and reject confusable, invisible, bidi-control, invalid, or normalization-colliding paths. Evidence: Unicode path corpus. <!-- ONE normal form (NFC), and a path not already in it is REPORTED rather than silently converted, because silent conversion is how the manifest and the filesystem come to disagree about which file this is. separator-introduced-by-a-wider-normalization is the deep one and is not theoretical: U+FF0F is one component under NFC and two under NFKC, so the check compares the separator COUNT under the two forms rather than keeping a list of characters current. Bidi marks are held OUT of the invisible set on purpose — a directional override is not a character somebody cannot see, it is one that reorders the others, and reporting it as merely invisible puts the weaker sentence in front of the reader. The confusable check is per-COMPONENT, since Latin filenames under Cyrillic directories are ordinary and one name holding both an a and an а is not; the counterexample is a test. -->
    • S6.2.d Reject symlink, hardlink, junction, reparse, device, FIFO/socket, and other nonregular file entries. Evidence: special-file corpus. <!-- file-with-more-than-one-link is the fault worth having, and it is PROVEN ON A REAL FILESYSTEM by tools/bundle-path-adversarial.mjs: a hardlink reports as regular from lstat, reads back byte-identical to a file outside the root, and would hash correctly. The link count is the only tell, and nothing looks at the link count unless it was written to. A hardlink means another name outside the bundle reaches the same inode, so writing through it writes somewhere the assembly never chose and deleting the assembly root does not remove the data. -->
    • S6.2.e Detect case-fold, Unicode-fold, trailing dot/space, separator, and target-filesystem collisions. Evidence: cross-platform collision tests. <!-- A collision is a property of the TARGET, not of the path, so it is computed against a declared filesystem and three real ones ship (POSIX_EXT4, WINDOWS_NTFS, APPLE_APFS) that disagree about case, normal form, trailing dots and path length. The real-filesystem tool states the claim the only way it can be stated: README.md/readme.md and an NFC/NFD pair are FOUR REAL DISTINCT FILES on this box, the local target reports zero findings, and NTFS and APFS report a case fold and a Unicode fold — the bundle is fine on the machine that built it and loses a file on the machine that opens it, which is never the same machine. The fault kind names WHICH fold collapsed the pair, asked one transform at a time, because the remedies differ. -->
    • S6.2.f Reject duplicate logical/artifact/manifest references and conflicting byte identities. Evidence: duplicate ref tests. <!-- Three faults, not one. one-digest-at-two-paths is deliberately NOT a fault — it is deduplication, and a rule refusing it is switched off by the first bundle that legitimately ships one image twice (there is a test asserting it produces nothing). one-path-with-two-digests is never legitimate: the archive holds one file and the manifest describes two, so a verifier passes on whichever row it reaches first. And one-artifact-id-with-two-digests is the conflicting byte identity that survives review, because each half is correct on its own and only the pair is wrong. A reference that points at the manifest is refused separately: the document describing the bundle would be an item inside its own description, and a verifier walking it does not terminate. -->
    • S6.2.g Reject reserved device names, manifest/control paths, forbidden extensions, and overlong path/component/depth. Evidence: reserved/limit tests. <!-- The device-name check is on the STEM and case-insensitive, because CON.txt opens the console; a test walks CON/con.txt/NUL.json/lpt9.bin and asserts console.json is untouched. Control paths are an ARGUMENT rather than a constant, since what counts as one is a property of the profile — the manifest, a signature directory, whatever the importer reads as metadata — and an entry landing there is content that arrives as instructions. Forbidden extensions are the formats that execute or launch when opened, and a bundle is a thing people unpack without reading. -->
    • S6.2.h Revalidate normalized paths at final open/write/extract time to prevent race or adapter reinterpretation. Evidence: TOCTOU tests. <!-- Everything before the open describes a STRING; only the open describes a FILE. The real-filesystem tool performs the swap: between the check and the open, entries/ stops being a directory and becomes a symlink to an attacker-controlled directory, the string never changes, and the bytes land outside. It also recorded something worth keeping — O_NOFOLLOW guards the FINAL component only, so the O_NOFOLLOW open SUCCEEDED here; usedNoFollow in the contract therefore means "no link followed on ANY component", which needs per-component opens. open-without-nofollow is reported even when nothing changed, because a run in which the race did not fire is not a run in which it cannot, and a check that only reports observed divergence reports nothing on the first day somebody wins. The tool carries the counter-control too: an unraced open must report NOTHING. --> <!-- S6.2 rollup: 3806 kit tests pass (44 new); tsc clean on lib and spec; export map at 80 entry points, --check clean; all six kit gates plus both new real-filesystem targets pass. The corpus is 27 shipped vectors, each declaring the fault it must produce AND the stage at which it becomes visible, and both halves are asserted. -->
  • S6.3 Validate manifest schemas from their authoritative files. Evidence: tests fail when schema and builder drift. <!-- schemas/bundle-manifest.schema.json (a real JSON Schema 2020-12 file) + src/manifest-schema.ts + tools/manifest-schema-drift.mjs (new Nx target). The failure this prevents is the one every schema-shaped system arrives at: the document and the code that produces the document drift, and nothing notices because the tests were written against whichever of the two the author had open. The bundle STILL VALIDATES — the validator was compiled from the schema, the builder emits what the builder emits — so the only party who finds out is an importer on the other side of an organisation, months later, holding a file that says it conforms to something it does not. Five fault families, 38 faults, all producible; 29 tests. -->

    • S6.3.a Register authoritative manifest schema path, ID/version, owner, code-generation command, and supported bundle profiles. Evidence: schema registry. <!-- validator-version-that-is-a-range is the fault worth the function: ^8.20.0 reads as a pin to everybody who skims it and means two CI runs, or CI and a laptop, compile the SAME schema with different code — and the day they disagree is the day one of them starts accepting something, and the run that accepted it is the one that shipped. generationCommand is not decoration either: a derived artifact whose regeneration command lives only in somebody's shell history is regenerated differently by the next person, and the difference is invisible in the diff because both outputs are generated. profile-whose-manifest-no-schema-governs is checked from the OTHER end, with the known profile ids as an argument, because a profile nothing validates is invisible from inside a register that contains only the schemas somebody remembered to add. -->
    • S6.3.b Compile validators directly from authoritative schema with pinned tool/version and strict unknown/format/reference behavior. Evidence: validator generation. <!-- The spec COMPILES THE REAL FILE with the real pinned ajv and asserts the installed version IS the pinned one — not that the register says so. schema-changed-since-the-validator-was-compiled compares a recorded digest against a caller-supplied RECOMPUTATION, since comparing a stored digest with itself passes by construction. Three leniencies are named because each turns a rule into a comment: a non-strict validator IGNORES an unrecognised keyword, so a misspelled requred reads as present and constrains nothing; an unknown FORMAT treated as satisfied leaves a field unconstrained while looking constrained; and a remote $ref makes validation a network call, so the rules are whatever that host serves today and the document deciding whether a bundle is well-formed is not in the repository. A test greps the shipped file for a remote ref. -->
    • S6.3.c Validate builder input, assembled manifest, serialized bytes, archived manifest, importer parse, and migrated output. Evidence: boundary coverage tests. <!-- VALIDATING THE OBJECT IS NOT VALIDATING THE BYTES, and byte-boundary-validated-as-an-object is the fault that says so. Serialization is where a Date becomes a string or an [object Object], an undefined property disappears, a large integer loses precision and a NaN becomes null; a boundary that validated the in-memory value has checked the one representation that was never in doubt, and the remedy is one line nobody writes because the check they already have passes. The coverage row therefore records parsedFromBytes, and the spec actually crosses the boundary — stringify, parse, validate THAT — rather than asserting a flag. A boundary with no row at all is reported as unvalidated, because silence is not "no". -->
    • S6.3.d Maintain canonical minimal/full valid and malformed/boundary/ version-skew invalid fixtures. Evidence: schema fixture corpus. <!-- 16 fixtures, and 13 of them MUST FAIL: a set of valid documents proves a validator accepts and only a rejected one proves it still refuses, so a test asserts the failing half outnumbers the passing half. Every fixture records WHY, so one that starts passing names the rule that stopped applying instead of producing a diff that says a test changed. version-skew is its own class because it is the failure nobody writes a test for — the document is well-formed in every respect except that it says it conforms to a version this is not, and a validator ignoring that applies the wrong rules to a document that told it not to. Every verdict is asserted twice: once on the object and once after a real round trip through bytes. -->
    • S6.3.e Compare generated types, builder fields/defaults, docs/examples, and validator output with schema. Evidence: cross-artifact drift check. <!-- schemaFacts DERIVES what it needs from the parsed document rather than restating it, because a TypeScript constant listing the required fields would be a third artifact and the third artifact is always the one that drifts — nothing validates against it. The sample is a manifest the BUILDER actually produced (S6.1's resolveLayout over the shipped profile), not a declaration of what it produces. Both directions are walked, since a field the schema requires that nothing emits and a field the builder emits that the schema never declared are different bugs with different owners. schema-enum-value-nothing-can-produce is the direction nobody writes: a kind left in the enum after the code emitting it was deleted is a value an importer is entitled to handle, so somebody writes a branch that is never taken and never removed. -->
    • S6.3.f Deliberately mutate a required field/version/enum in schema and builder to prove CI catches both directions of drift. Evidence: mutation tests. <!-- tools/manifest-schema-drift.mjs. A drift check that has never seen drift is a function returning an empty array, so this runs SEVEN deliberate mutations — three on the schema, four on the builder — and each one names the mechanism that must catch it, because "something failed" is also what a broken suite says. Before any of them it runs the control that makes the rest meaningful: the unmutated pair must produce zero drift findings and the real manifest must validate. The interesting result is M3: opening additionalProperties makes the compiled validator ACCEPT what it used to refuse, so the cross-artifact check is the only thing left that can see it — which is the argument for having both. M6 and M7 are the mirror, caught by the validator and not by the drift check. Run: pnpm nx manifest-schema-drift @oshun/workbench-kit. -->
  • S6.4 Rehash every referenced artifact from bytes during assembly. Evidence: metadata-only checksum spoof is rejected. <!-- src/artifact-rehash.ts + tools/artifact-rehash-race.mjs (new Nx target). A manifest is a set of assertions about bytes, and the only question that matters about it is where each assertion came from. Copy a digest out of the store metadata and the manifest certifies the metadata; take it from the caller and it certifies the caller. Both produce a document indistinguishable from one that was earned — same fields, same shape, same green tick — believed by everything downstream for longer than anybody involved will be around to correct it. Six fault families, 43 faults, all producible; 38 tests. -->

    • S6.4.a Resolve artifact refs through authorized immutable storage and reject mutable/unknown/nonfinal references. Evidence: resolver tests. <!-- A REFERENCE THAT CAN CHANGE IS NOT A REFERENCE: assembled twice, a tag or a "latest" produces two different bundles under one id, and the second is the one nobody diffed. version-pinned-reference-with-no-version is a mutable reference with a reassuring name. reference-to-something-not-yet-final looks like a corner case and makes the strangest bugs — an object whose upload is still open reads, returns fewer bytes than it will have, and hashes to something that will never be true again, so the bundle fails verification for reasons nobody can reproduce because by the time anybody looks the upload finished. And reading across a tenant boundary is checked here because an assembly is a very quiet way to copy bytes across one: the output is a file rather than a response. -->
    • S6.4.b Stream exact stored bytes under size/time/resource limits and compute approved digest and length. Evidence: known-byte tests. <!-- A LIMIT ENFORCED AGAINST THE METADATA IS NOT A LIMIT. Reading contentLength and deciding the object is small enough bounds the number the object itself supplied — which is exactly the number under dispute whenever a size limit matters. The enforcement costs one counter in the read loop and nobody adds it, because the version they have never fails. digest-over-fewer-bytes-than-were-read is S3.11's partial read caught at the source: a hash finalised early gives a digest for a prefix and a length for the whole, and the pair reads as verified. An abort WITH a stated reason is deliberately not a fault — that is a bounded read doing its job, and reporting it would be reporting the control working. -->
    • S6.4.c Compare computed digest/length/media type to artifact metadata, manifest expectation, and provenance binding. Evidence: mismatch tests. <!-- Three expectations, checked SEPARATELY because they fail for different reasons and are owned by different people: the store metadata disagreeing means the object changed or the metadata was written wrong; the manifest disagreeing means the assembly is building something other than what was planned; the provenance disagreeing means these are not the bytes a chain of custody was established for, and that is the only one of the three not fixed by rebuilding. nothing-to-compare-against is reported, because a rehash with no expectation is an expensive way to record a number while the log says "verified". Media type comes from S3.11's sniffer, which returns a LIST — a GIF header with a zip's central directory at the end is a valid image to a decoder and a valid archive to an unzipper, and a bundle carrying one hands each consumer a different file while the manifest describes one. -->
    • S6.4.d Copy/link into assembly only after verification and bind manifest entry to computed values rather than caller metadata. Evidence: spoof-prevention tests. <!-- The headline. BINDING_SOURCES has four values and a test asserts exactly ONE of them produces no findings. manifest-bound-to-store-metadata IS the metadata-only checksum spoof this cell exists to reject, and it passes every check that reads the manifest. manifest-bound-to-the-source-bytes-rather-than-the-written-ones is the near-miss that survives review, because it IS a computed digest — of the wrong bytes; the archive contains the copy. The order is data (BINDING_ORDER), and rehash-written sits between the copy and the binding because after a successful verify and a successful copy there is nothing left that FEELS unchecked, which is why it is the step every implementation omits. -->
    • S6.4.e Detect storage changes between read/hash/copy through immutable version/fencing or hash the exact written bytes. Evidence: race tests. <!-- Exactly TWO ways to close the window, and one must hold: either the store gives a handle that changes whenever the object does and both ends are compared, or the written bytes are hashed. A system with neither is not slightly weaker — it has no means by which it could ever notice, which is what no-fencing-and-no-rehash-of-the-written-bytes says. A race has no in-memory representation, so tools/artifact-rehash-race.mjs performs it on real files: the source is replaced after being verified and before being copied, and the run shows the written-bytes rehash is what catches it while a checker that hashed the source sees a perfectly consistent story. It carries the counter-control too — an unraced copy must report nothing. -->
    • S6.4.f Quarantine/report corruption and affected derivatives/releases without publishing partial assembly. Evidence: corruption workflow test. <!-- CORRUPTION IS A FACT ABOUT A POPULATION. The artifact was corrupt BEFORE the rehash noticed, so failing this assembly answers the smallest version of the question and nothing else will ask the larger one. derivative-left-out-of-the-enumeration is checked against a list the caller supplies from the dependency graph, because an enumeration that walks the same index the corruption came through finds whatever that index says — which is the thing in doubt. partial-assembly-published is reported first and alone: every other failure here can be repaired late and that one cannot be undone at all. A counterexample test asserts no hold is demanded when nothing depends on the artifact. -->
  • S6.5 Produce deterministic archives. Evidence: identical inputs yield byte-identical output and hash. <!-- src/determinism.ts + tools/archive-determinism.mjs (new Nx target). A non-reproducible archive breaks exactly one thing, and it is the thing every later cell depends on: you can no longer tell "the inputs changed" from "the build ran again". A rebuild whose hash differs is then indistinguishable from a tampered artifact — and because the first explanation is almost always the right one, the response to both becomes re-run it and move on, which is correct ninety-nine times and catastrophic once, with nothing about the ninety-nine to say which run was the hundredth. Four fault families, 39 faults, all producible; 24 tests. -->

    • S6.5.a Define normalized entry ordering, path encoding, timestamps, owner/group, permissions, extended attributes, line endings, and manifest serialization. Evidence: reproducibility profile. <!-- CANONICAL_REPRODUCIBILITY_PROFILE ships and is held to zero findings. SORT ORDER IS BYTE ORDER: localeCompare puts "a.txt" before "B.txt" and byte order puts "B.txt" first — asserted in the spec rather than described — so a collated archive's entry order is a function of LANG, which is the sort every reviewer reads as correct and the one input nobody records. The only admissible timestamp came from an INPUT: an mtime is a property of a CHECKOUT, so the same commit cloned twice makes two archives and neither is wrong. And line-endings: preserved is deliberately NOT a fault — normalizing line endings rewrites CONTENT and moves every digest the manifest records, so the dimension must be DECIDED and preserved is a decision while platform is not. -->
    • S6.5.b Pin archive format/version, compression algorithm/level/library, dictionary, and platform-independent flags. Evidence: toolchain lock. <!-- THE CONTAINER IS PART OF THE OUTPUT, and the three gzip header fields are named separately: FNAME carries the path of the temporary file the compressor was handed, MTIME is four bytes of build clock in a header nobody reads, and byte 9 is the OPERATING SYSTEM the compressor ran on — it encodes nothing about the content, decompresses to the same bytes whatever it says, and makes the CI archive differ from the laptop one in a way no diff of the contents will ever show. The tool demonstrates exactly that by flipping the byte and asserting the decompressed output is unchanged while the hash is not. The format/compression pair is also checked for internal agreement in both directions. -->
    • S6.5.c Exclude temporary paths, host metadata, absolute source paths, random IDs, current clock, process order, and nondeterministic maps. Evidence: nondeterminism scan. <!-- The scan runs over what the archive ACTUALLY contains — paths and text samples — rather than over a declaration, and the entry ORDER is rechecked here because a profile that declares byte order and a builder that sorts by collation is precisely the case a declaration cannot detect; this is the only place the real order is visible. Host facts are supplied by the caller for the same reason S6.2's protected terms are: a contract entry point has no environment, and a scan that guessed at hostnames would report the wrong ones and be switched off. timestamp-that-is-not-the-declared-epoch takes the declared instant as an argument, because the field is almost always called something innocuous like "generatedAt" and moves the hash every second. -->
    • S6.5.d Build from the same inputs on supported operating systems/runtimes and compare bytes/hash/entry metadata. Evidence: cross-platform result. <!-- buildComparisonFaults compares archive digest, manifest digest, entry ORDER and per-entry mode/uid/gid/mtime, because "the bytes differ" is a useless message and "the mtime of one entry differs" is a fix. every-build-on-one-platform is reported honestly: path encoding, line endings and the container OS byte are exactly the differences a single platform cannot show, and this box is one platform — so the tool asserts the checker STILL reports platform as an untested axis while agreeing on the bytes. A green single-platform run that claimed cross-platform reproducibility would be the failure this item exists to prevent. -->
    • S6.5.e Build under shuffled input order, clocks, locale, time zone, and concurrency and compare byte identity. Evidence: perturbation tests. <!-- tools/archive-determinism.mjs contains a real ~60-line ustar writer whose every non-content field comes from the profile, so a claim in the profile and a byte in the output are the same statement. It builds one input four times — shuffled order, every source file touched an hour forward, a Turkish locale, a Kiritimati time zone and a 0077 umask, all set for real in the process — and the four archives are byte-identical. Beside them four CONTROLS that must each produce a difference, because a determinism suite where nothing can ever differ would pass over a builder returning a constant: collation ordering, a real mtime, the build user's uid, and gzip at two levels plus the OS byte. only-one-build and axis-never-perturbed are the two ways this claim gets made without being tested, and both report green today. -->
  • S6.6 Generate SBOM-like component and dependency inventory for each release bundle. Evidence: every file and external ref has provenance. <!-- src/bill-of-materials.ts + tools/bom-reconciliation.mjs (new Nx target). A bill of materials generated from the same declaration the bundle was built from is a photocopy: it agrees with the manifest for the reason a photocopy agrees with the original, it can never disagree, and the disagreement is the only thing an inventory was ever for. So the failure this item guards against is not a missing inventory — it is one that exists, lists every file, reads as evidence, and is a second copy of the claim it was supposed to check. Six fault families, 60 faults, all producible; 59 tests. -->

    • S6.6.a Inventory every archive entry with logical role, path, media type, bytes/hash, source artifact/revision, generated/third-party status, and provenance ref. Evidence: file component rows. <!-- FileComponent + fileComponentFaults. role is S6.1's BundleEntryKind rather than a second vocabulary for the same thing, and mediaType goes through S3.11's sniffMediaType where bytes are supplied — the module decides neither question twice. digestHex is a plain string and NOT the branded Sha256Hex deliberately: a row is an OBSERVATION, so it has to be able to hold what was observed, and a branded field would make row-with-a-digest-that-is-not-a-digest unrepresentable, which improves the type and not the archive. "Generated" is split three ways because a build tool, a model and a person are three answers to who is answerable, and generated-row-with-no-source-revision is the row that reads as complete and is not — regenerating the file is the only way to check the digest, and that needs the revision. sampledBytes is optional because an inventory is usually assembled from an index, and a check that demanded the bytes would be run without them. -->
    • S6.6.b Inventory external URLs/resolvers/providers/models/tools/plugins/ packages/licenses that the bundle references or requires. Evidence: external component rows. <!-- ExternalComponent + externalComponentFaults, ten kinds because ten different things get referenced and exactly one of them arrives through a package manager — a bill generated from a lockfile lists that one and none of the other nine. The most common way a release stops working is a URL in a configuration file nothing counted as a dependency. locator-that-carries-a-credential is the fault that turns a compliance document into an incident: a bill of materials is written to be PUBLISHED, and a signed URL pasted into a config travels into it without anybody deciding to publish a token. license-document-referenced-by-url-alone applies the same idea to the one external everybody assumes is safe to leave remote — a license reachable only by URL can change or stop resolving, and the release then cannot show what it agreed to when it shipped. A locator with no scheme stops the checks for that row rather than reporting five consequences of one cause. -->
    • S6.6.c Record dependency edges, version/range/digest, required/optional, runtime/build-only, target/environment, and resolution state. Evidence: dependency graph. <!-- BomDependency + dependencyEdgeFaults. A RANGE IS NOT A DEPENDENCY OF THIS BUNDLE: the bundle contains one resolved thing, and a range describes the family of bundles the build was permitted to produce, exactly one of which was produced. The mirror-image fault is dynamic-edge-carrying-a-digest — an answer that arrives at runtime by design has no digest by nature, so recording the one observed during the build states a fact about every future run from one moment. Both read as thorough. vendored is kept apart from resolved because the bytes are in the bundle and the range no longer applies to anything, and dynamic-at-runtime apart from unresolved because nothing is missing. Cycles are NOT re-decided here: S5.8 already owns what a cycle means per edge kind, and a second answer would be a second answer rather than a check. edge-for-an-environment-the-bundle-does-not-target reports nothing when the caller passes no targets, so a caller that does not know them is not told every edge is wrong. -->
    • S6.6.d Record license/rights/consent/export/security metadata and unresolved/unknown status without fabricating values. Evidence: compliance fields. <!-- ComplianceRecord + complianceFaults. UNKNOWN IS A VALUE AND IT IS THREE VALUES: "nobody looked", "somebody looked and there is none" and "the question does not apply" have three different consequences, and the field that collapses them is the field that makes an unread license look permissive. dimension-not-recorded-at-all is the difference between a shorter document and an honest one — an absent dimension is not unknown, it is a document that never asked, and it reads as one with nothing to report. license-inherited-from-the-bundle is the sharpest instance of fabrication and the reason the worked example gives its third-party row OFL-1.1 against a bundle at CC-BY-4.0: a default that fills empty license fields with the publisher's own expression produces an inventory with no blanks and every third-party row wrong in the same direction, and an example where every expression agreed could not tell that from a correct inventory. not-applicable is required to carry a basis because it is a conclusion somebody reached and otherwise the state that costs nothing to write and closes the question. The consent fault does not object to unknown being RECORDED — that is the decision working; it objects to shipping a release whose model-generated component is the one nobody assessed. -->
    • S6.6.e Link components to generation/transformation actions, actors, tools/models/policies, validation, and release candidate. Evidence: provenance linkage. <!-- GenerationAction + bomLinkageFaults. action-for-another-release-candidate is the fault worth building this for: a component carried over from a previous build arrives with a complete, internally consistent chain — actor, tool, policies, a passing validation run — and every check that reads the chain on its own terms approves it. The only thing wrong with it is which release it belongs to, and that is visible from nowhere else; the spec asserts it is the ONLY finding on such an action. action-consuming-something-the-bill-does-not- list is the input side of the same idea and is where a training corpus, a reference dataset or a licensed font goes missing — the output is inventoried and what it was made from is not. action-with-an-unattributed -actor treats a build service as an actor with an identity and only unattributed as an absence, with a counterexample test, because a rule that fired on a service-run build would be switched off in a week. kind is a plugin-owned string for S5.8's reason. -->
    • S6.6.f Reconcile manifest/archive/provenance/importer refs so no file or external ref is absent, duplicated, or orphaned. Evidence: completeness check. <!-- bomReconciliationFaults over four witnesses, compared pairwise in BOTH directions for S4.18's reason: an entry with no declaration and a declaration with no entry are wrong in opposite ways. ABSENT, DUPLICATED AND ORPHANED ARE THREE FAULTS AND A SET SEES TWO — so archivePaths is a LIST that is COUNTED, never a set that is tested for membership. tools/bom-reconciliation.mjs settles that with real GNU tar: it builds the archive, tar --appends a second member at a path the archive already carries, extracts it, and shows the recipient holds bytes whose digest is not the one the bill recorded — while asserting BOTH set-membership sweeps report it clean FIRST, so the control cannot go vacuous. Verified by patching tally(witnesses.archivePaths) in the BUILT dist to dedupe and confirming the tool fails. Three more controls: a member appended after the manifest was written (bill and manifest agree perfectly and only the archive witness sees it — the argument for the archive being a witness at all), the PHOTOCOPY (every-row-from-one-observer fires on an inventory that reconciles against all four witnesses and has nothing else wrong with it), and an ELF binary of exactly the same length described as an image with a digest that verifies, so the size check and the digest check both pass. The importer comparison covers only externals a RECIPIENT reaches, since a build-time model endpoint is legitimately absent from an importer's list. --> <!-- S6.6 rollup: 3960 kit tests pass (59 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 84 entry points with no drift and no violations. -->
  • S6.7 Add C2PA signing/validation adapters where the asset format supports them. Evidence: valid, tampered, unknown-trust, expired, and stripped credentials are distinguished. <!-- src/content-credentials.ts + tools/c2pa-signing.mjs (new Nx target). The evidence line names five situations together because a validator returning { valid: boolean } gives the SAME answer for four of them: ask only whether the signature verifies and a credential signed by an expired certificate says yes, one whose chain reaches no anchor says yes, and an asset with NO credential says yes because nothing failed. One of those four is fine, one is a policy question, one is a key-lifecycle problem, and one is an asset with no provenance being displayed as though it had some. Five fault families, 47 faults, all producible; ten independent checks; nine states; 60 tests. The kit owns the CONTRACT and the state mapping — @oshun/content-signing is the repository's one real Ed25519 signer and is resolved from the SPEC (devDependency + ALIASED_PACKAGES), so nothing here is a second implementation of something that already exists. -->

    • S6.7.a Register supported asset formats, C2PA spec/profile, embedding/ sidecar mechanism, hard/soft bindings, assertions/actions, and unsupported behavior. Evidence: adapter profile registry. <!-- AdapterProfile + adapterProfileFaults, and the shipped WORKED_ADAPTER_PROFILE is held to zero findings. embedded-manifest-with-no-hard-binding is the fault that produces a LABEL rather than a credential: a manifest carried inside the file, signed, listing every action and naming no digest of the content asserts that somebody said these things and not that they are about THIS asset — and it copies into another file unchanged. fallback-that-ships-the-asset-silently is decision-by-default: refusing, writing a sidecar and shipping it labelled unsigned are three decisions somebody made, and shipping it silently is the absence of one. ship-silently is IN the vocabulary for that reason — a registry that could not express it would report every deployment as compliant. The soft-binding rule is scoped to formats where a hard binding is available, with a counterexample test for a live stream, because a rule that fires on the only configuration a format allows is one somebody switches off. -->
    • S6.7.b Assemble ingredient/action/tool/actor/policy assertions from authoritative bundle/provenance records. Evidence: manifest construction fixtures. <!-- CredentialAssertion + assertionAssemblyFaults. assertion-the-caller-supplied is the whole item in one fault: a credential is a set of statements somebody signed, and the only question worth asking about each is where it came from. Read the actions out of the provenance ledger and the credential certifies what happened; take them from the caller and it certifies what the caller typed, in a document that looks exactly the same and is believed by everything downstream. caller-supplied is a value in ASSERTION_SOURCES for the same reason ship-silently is one. The reconciliation with the bundle runs in BOTH directions, and bundle-ingredient-with-no-assertion is the direction people forget: an ingredient list that omits an input is not a smaller truth, because a consumer reads it as the complete set of things the asset was made from and the omitted one is usually the one somebody would have asked about. -->
    • S6.7.c Sign with governed key/cert chain/timestamp and bind to exact bytes without exposing key material. Evidence: signing integration tests. <!-- SigningRequest + signingFaults + signerPortEscapeHatches. signing-key-that-leaves-its-boundary and key-material-in-the-request are the same failure at two distances — a key that CAN be copied and a key that already has been — and the second is found by scanning every string field for a PEM private-key header, because key material never arrives in a request deliberately, it arrives in a field somebody added for debugging. The finding NEVER echoes what it found, asserted by a test, or the conformance check would put the key in the log of every deployment that ran it. signerPortEscapeHatches is shaped after S6.1's profileEscapeHatches and takes unknown; the spec runs it against the REAL Ed25519ClaimSigner, which holds its PEM in a #private field and therefore reports clean, and against a port with the key on a public field, which reports twice — once for the field NAME and once for the PEM in its value. digest-that-came-from-the-caller is S6.4 arriving here unchanged, and no-timestamp-authority is decision 5 as a requirement: without one the credential's lifetime becomes the certificate's, and an archive signed today stops validating on a date chosen by a certificate authority's renewal policy. -->
    • S6.7.d Validate manifest structure, binding, signature, chain, trust list, revocation, timestamp, assertion support, and ingredient availability separately. Evidence: validation matrix. <!-- VALIDATION_CHECKS (TEN — certificate-validity is split out of chain, because "no path reaches any anchor" and "the path exists and a certificate on it was outside its window" are different failures with different responses and only the second is answerable by a timestamp) + validationReportFaults + validityAtSigning. The item says "separately" and that is decision 3: short-circuiting returns "invalid" when the useful sentence was "the signature is perfect and the certificate lapsed last Tuesday", so a report that stopped early is itself a finding. NOT-PERFORMED IS NOT PASS — a revocation responder nobody could reach has established nothing, and unavailable-revocation-treated-as-a-result fires on the DETAIL reading like an answer rather than on the outage, with a counterexample test for an honestly-described outage. credential-absent-and-checks-reported-as- passing is the shape of the boolean this module replaces: with nothing to check, every question passes by having nothing to fail. validityAtSigning takes all three instants as arguments (a function that read the clock would answer differently every run and could not be tested at a boundary) and refuses to be rescued by an UNVERIFIED timestamp, because that is a string the signer chose and honouring it lets an expired key sign anything by backdating. -->
    • S6.7.e Represent valid/trusted, cryptographically valid-untrusted, tampered, revoked/expired, unknown/unsupported, missing, stripped, and partially recoverable states. Evidence: result union tests. <!-- CREDENTIAL_STATES + credentialState. NINE states, not eight: revoked is separated from expired because a lapsed certificate is a calendar problem whose content is probably fine and a revoked one puts everything that key ever signed in doubt — different people, different urgency, and a system reporting them alike routes the second to whoever handles renewals. The precedence is stated and tested: TAMPERING OUTRANKS TRUST, since reporting the trust problem first sends an incident to the trust-list maintainer. MISSING AND STRIPPED ARE DIFFERENT AND THE ASSET CANNOT TELL YOU WHICH — an image with no manifest looks identical whether one was never made or a platform removed it on upload — so stripped requires ExternalWitness, and a validator claiming it from the asset alone would be inventing history. That gap is what S6.8 exists to close. The spec drives the signature check through the repository's REAL Ed25519 signer: a real signature verifies and gives valid-and-trusted, one flipped byte verifies false and gives tampered, and a signature from another key verifies false in exactly the same way — which is the fact that makes the chain and trust checks necessary rather than decorative. -->
    • S6.7.f Test official/third-party fixtures, post-sign byte mutation, metadata stripping, certificate lifecycle, and offline trust cache. Evidence: interoperability report. <!-- INTEROPERABILITY_CASES (ten) + interoperabilityFaults, and tools/c2pa-signing.mjs for the parts that need real cryptography. self-produced-fixture-presented-as-third-party is the fault this exists for: an adapter validated only against bytes it produced demonstrates that it agrees with itself, which every implementation does including the wrong ones, and generating the fixture is the cheapest way to appear to have tested interoperability. state-never-demonstrated closes the other end — a suite can cover every case in the list and never once produce revoked or stripped, and a state no test has seen is a branch nobody has executed (S6.5's axis-never-perturbed in another domain). The tool produces all five situations with REAL OpenSSL certificate authorities (two independent roots, three leaves, real openssl verify path building) and real P-256 signatures, and CONTROL 1 asserts the naive boolean answers FOUR of them TRUE before the state machine answers them four ways — three of those four trues are wrong. Verified by patching credentialState's stripped branch in the BUILT dist and confirming the tool fails. --> <!-- S6.7 rollup: 4021 kit tests pass (60 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan, packed-consumer-smoke and test-only-import-fixtures all pass over the rebuilt dist; the export map regenerates to 85 entry points with no drift and no violations. -->
  • S6.8 Add a provenance-ledger fallback for formats or channels that strip embedded credentials. Evidence: soft-binding recovery test. <!-- src/soft-bindings.ts + tools/soft-binding-recovery.mjs (new Nx target). A soft binding returns a DISTANCE, and the danger is that "close" is a number and a number invites a verdict. A perceptual hash's false-positive rate is a property of the CORPUS SIZE rather than of the algorithm, so a threshold that separates fifty images perfectly returns matches for unrelated photographs all day at a million — and a system that turns the best candidate into an answer produces confident, specific, wrong attributions, most often for the most-published images. Six fault families, 45 faults, all producible; seven outcomes; 53 tests. -->

    • S6.8.a Register formats/channels/transforms that strip or alter embedded provenance and select approved binding/recovery method. Evidence: channel profile. <!-- ChannelProfile + channelProfileFaults. Which platforms remove which carriers is a fact somebody can go and MEASURE, so a channel with no registered behaviour is a finding rather than a default — the alternative is that a credential's survival is discovered by a reader who does not find it. behaviour-recorded-with-no-observation is the softer version: a behaviour written down from a blog post is a guess with a table around it, and it stops being true the next time the platform ships. transforming-channel-relying-on-a-hard-binding-alone is the arithmetic of the hard hash in the other direction — it will not match after a re-encode, so a transcoding channel with only that has no recovery at all. A counterexample test covers a channel that preserves the credential and needs no fallback. -->
    • S6.8.b Persist external ledger record keyed by immutable provenance/ artifact identity with hard hash, approved soft bindings, manifest/ref, trust, retention, and access policy. Evidence: ledger schema tests. <!-- LedgerRecord + ledgerRecordFaults. THE HARD HASH STAYS: an unmodified copy has to be recognised exactly, and a record holding only a soft binding answers "probably" to a question that has a yes. content-derived-record-that-is-public is the fault people argue with — a perceptual hash is not a copy of the image so publishing one feels harmless, and it is a durable identifier that SURVIVES CROPPING, which is the operation somebody performs to remove identifying context; a public index of them is a reverse image search over everything anybody registered. Every threshold carries measuredAtCorpusSize, because the expected number of false positives grows with the number of records searched and nothing in the number itself says so. And binding-computed-below-the-signal-floor closes a hole my own probe found: a featureless asset produces sixty-four bits of rounding error that index and match exactly like a real hash, so the strength the value was computed at travels on the record. -->
    • S6.8.c Generate media-appropriate fingerprint/watermark/perceptual hash/ lookup hint with privacy and collision review. Evidence: binding tests. <!-- A REAL DCT-II perceptual hash (perceptualHash), a real gradient hash (differenceHash), a real hammingDistance, plus BindingReview + bindingReviewFaults. The standard construction — 32x32 box reduction, separable DCT-II, the low-frequency 8x8 block against its own median — with ONE STATED MODIFICATION: the DC term is dropped, because brightness is the transformation applied most often and least deliberately, and the next coefficient along one axis is added so the hash is 64 bits of which 64 are informative rather than 63. Measured, not described: half-scale moves it 0 bits, a brightness shift 0, an eight-pixel crop 2, and two different images 30. The two methods are shipped because they FAIL DIFFERENTLY, and the spec asserts the directions rather than claiming them. binding-narrower-than-its-corpus-needs takes the MEASURED collision rate rather than a formula, because the effective width depends on how the corpus is distributed and real image corpora are nothing like uniform. -->
    • S6.8.d Recover exact/robust matches after representative resize, crop, transcode, recompress, remix, screenshot, rerecord, or normalization. Evidence: transformation matrix. <!-- TransformationOutcome + transformationMatrixFaults, and tools/soft-binding-recovery.mjs for the measurements. The tool runs SIX REAL ffmpeg transformations over a real frame — a bicubic downscale, a ten per cent border crop, a JPEG round trip at q25, a screenshot round trip (down to 40% and back through a lossy encode), a tone curve, and an H.264 pass — and decodes each back to raw gray. Five keep the record (0, 0, 0, 2, 0 bits of 64); the CROP MOVES IT SIXTEEN AND LOSES IT. That is a real property of the method, so the tool asserts the matrix REPORTS the loss rather than that no loss occurred, and fails if nothing is ever lost — a tool measuring only what survives would be describing its corpus. transformation-measured-on-one-sample is the other way a matrix reads as complete: the variance between items under one transformation is larger than the difference between transformations. -->
    • S6.8.e Measure false positive/negative, ambiguity, insufficient evidence, collision, and adversarial removal. Evidence: known-corpus evaluation. <!-- CorpusEvaluation + corpusEvaluationFaults. evaluation-with-no-negative-pairs is the first thing to look for: measuring only true pairs gives RECALL, which answers "does a transformed copy find its record" and says nothing at all about "does an unrelated item find one" — and the second is where the harm is, because a false positive attributes somebody's work to somebody else with a number beside it. threshold-measured-on-a-corpus-smaller-than-the-live-index is the fault that survives every review: the threshold was correct when it was chosen, the index grew, and nothing anywhere is attached to the growth. evaluation-that-never-produced-insufficient-evidence and ambiguity-never-observed are the S6.5 axis-never-perturbed idea here — a suite where every query resolved has either a threshold that admits everything or a corpus with nothing hard in it, and the multiple-match path that produces the wrong attribution has never run. AN ADVERSARY IS NOT A TRANSFORMATION: resize and recompress happen, an adversary picks what maximises the distance while keeping the content worth publishing, and a recall figure over benign transforms only is true and irrelevant. -->
    • S6.8.f Handle missing/multiple/private/deleted/revoked/invalid matches and display method, confidence/limitations, and trust safely. Evidence: negative and accessible UX tests. <!-- recoveryOutcome (seven outcomes) + presentationFaults. The exact hard match is decided FIRST and no distance is consulted for it. multiple-soft-matches and insufficient-evidence are first-class answers, and the tool's CONTROL 2 shows why: a nearest-candidate matcher named a record at 28 bits for an image that is not in the corpus at all, every time it was asked, while the strict recovery returned no-match. A WITHHELD RECORD IS DELIBERATELY INDISTINGUISHABLE FROM NO MATCH — the inverse of the kit's usual rule and of S5.8's, because telling a searcher a record exists and may not be shown tells them the item is registered, and a reverse lookup is the query where the existence of the answer IS the answer. At the surface, soft-match-presented-as-an-identification is where the whole chain is usually lost (one sentence reading "verified as" turns every careful layer beneath it into a claim) and distance-rendered-as-a-probability is the same failure in numeric form: 6 of 64 bits is not "91% confident", the conversion invents a probability from a quantity that has none, and the invented number is the one people quote. Colour-alone and text-alternative faults cover the accessible-UX half. --> <!-- S6.8 rollup: 4075 kit tests pass (53 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 86 entry points with no drift and no violations. The tool was verified to FAIL by patching usable to a constant true in the built dist, which made CONTROL 3 report the real flat-frame signal strengths of 1.4e-13 and 1.3e-12. -->
  • S6.9 Add publisher adapters with plan, execute, verify, and compensate phases. Evidence: partial external failure is recoverable and audited. <!-- src/publishers.ts + tools/publish-saga.mjs (new Nx target). The shape almost every publisher ships is one code path with a dryRun flag, which makes the plan a MODE of publishing rather than a document: it cannot be stored, diffed, reviewed or approved, it goes stale with nothing to notice, and the one case the four phases exist for — the target changing between deciding and doing — is invisible by construction, because the deciding happened inside the doing. Seven fault families, 59 faults, all producible; six reconcile actions; 44 tests. Composes S2.5's PublisherRuntime port (which already refuses the executor's own report as a verification source) and S3.9's ExternalOutcome, rather than restating either. -->

    • S6.9.a Define adapter identity/version, target capability/config probe, supported bundle/profile/action, authorization, idempotency, and external reference contract. Evidence: publisher interface. <!-- PublisherAdapter + adapterContractFaults. adapter-that-returns-no-external-reference is the fault that makes every later phase impossible: with no identifier the target assigned there is nothing to read back, nothing to compensate and nothing to reconcile, so the pipeline necessarily falls back to trusting the response — which S2.5 already refuses. IDEMPOTENCY_HONOURING carries ignores-the-key as a real value because plenty of real targets do, and a registry that could not express it would let every one of them look safe to retry; declaring it is NOT a fault, and unknown is, because the alternative to declaring it is discovering it as a duplicate publication. -->
    • S6.9.b Implement read-only plan resolving destinations, transforms, eligibility, cost, impact, conflicts, irreversible points, and rollback/ compensation. Evidence: plan fixtures. <!-- PublicationPlan/PublicationStep + publicationPlanFaults. plan-that-performed-something is the decision at the top of the module: every property a plan is supposed to have — reviewable, storable, comparable, approvable — depends on it having written nothing. plan-with-no-irreversible-point-identified fires on a plan whose every external step claims full reversibility, because a plan that is genuinely safe and one nobody thought about look identical, and the point of no return is the single fact an approver most needs. Cost and reach are required per step for the same reason: a plan that says what it will do and not what it will spend or who it will reach leaves the only reviewable facts as the ones nobody was worried about. -->
    • S6.9.c Revalidate plan hash, release lock, authorization, gates, provenance/rights/consent, target health, and budget at execute. Evidence: stale-plan tests. <!-- ExecuteConditions + executeAdmissionFaults, twelve checks. The point is that all of them were checked BEFORE and that is not enough: each is a fact with a timestamp, and the interval between planning and executing is exactly the interval an approval workflow introduces. Rechecking is not redundancy — the first check answered a different question about a different moment. plan-executed-under-a-digest-nobody-approved catches the edited plan: the plan is stored, somebody amends a destination, the digest moves, and the approval still points at the old bytes. target-health-never-probed is kept apart from target-unhealthy-at-execute because taking the health from the plan is a statement about a moment that has passed and a system nobody here controls. -->
    • S6.9.d Execute idempotently with durable per-step state, retries, external IDs, request/response hashes, partial result, and audit. Evidence: execution fault tests. <!-- StepExecution + executionFaults. retry-that-changed-the-idempotency-key is the fault that turns a retry into a second publication — it happens whenever the key is derived from something that moves (a timestamp, an attempt number, a fresh id) and the resulting duplicate is indistinguishable at the target from a deliberate second publish. settled-step-with-no-request-digest is what a dispute needs: six months later the question is what was SENT, and a system holding only its own summary of what it meant to send cannot answer it. sent-step-that-was-never-recorded-as-started is the ordering bug — a record written after the request returns leaves a publication nothing knows about if the process dies in between. -->
    • S6.9.e Verify by target read-back of bytes/hash, metadata, audience, state, permissions, discoverability, and external acknowledgements. Evidence: verification tests. <!-- VerificationReadback + verificationFaults, seven dimensions read SEPARATELY. published-but-not-discoverable gets its own fault rather than being one more mismatch, because it is the partial success that reaches production most often and answers 200 to everything: the object exists, the bytes are right, the permissions are right, and nobody can find it, and the failure surfaces weeks later as an absence. The tool produces exactly that against a real HTTP target — a real 201, a real read-back where six of seven dimensions match, and a real search returning zero hits. dimension-never-read-back treats an unread dimension as a finding, since an unread one reads as a matching one, and verification-that-is-only-the-publish-response is S2.5's decision arriving here as a checks array where no source is the target. -->
    • S6.9.f Compensate reversible steps in safe order, preserve irreversible effects, and expose manual remediation for failed/partial compensation. Evidence: compensation scenarios. <!-- compensationOrder (what should happen) and compensationFaults (what did) are DELIBERATELY separate functions: a system where the same code does both can only ever agree with itself, which is S6.6's photocopy in another shape. The order is the reverse of execution and stops at the first irreversible step counting backwards, and everything before that point is PRESERVED — removing a step something irreversible depends on leaves a state neither the plan nor the target's history describes. Every preserved effect and every failed compensation becomes an instruction naming the step, the external reference and the destination, because a system that cannot undo what it did and only logs that fact is how an asset stays published for a year after the release that put it there was withdrawn. Verified by patching the blocked propagation out of the BUILT dist and confirming the tool reports that the upload would be undone behind an announcement the target refuses to recall. -->
    • S6.9.g Reconcile unknown outcomes after timeout/callback loss/worker death before retry. Evidence: external outcome tests. <!-- reconcilePublish + publishReconciliationFaults, composing S3.9's ExternalOutcome rather than inventing a second vocabulary for the same four answers. The branch this exists for is an unknown outcome against a target that IGNORES the idempotency key: retrying is a second publication and not retrying may be a dropped release, both wrong in ways nobody can see from here, so it is settled by ASKING when a probe exists and by a person when it does not — never by the reconciler picking whichever error it finds less embarrassing. mark-failed-from-the-probe is in the vocabulary and is never CHOSEN, which the spec asserts: a probe reporting "failed" means nothing landed, which is a reason to retry. It exists so the fault checker can report a system that took it anyway, and callback-loss-treated-as-a-failure is the expensive, reasonable-looking version — a webhook that never arrived is a statement about the callback and says nothing about whether the publish happened. --> <!-- S6.9 rollup: 4120 kit tests pass (44 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 87 entry points with no drift and no violations. -->
  • S6.10 Add immutable release IDs and source revision locks. Evidence: post-approval edits cannot mutate a release candidate. <!-- src/release-locks.ts + tools/release-lock-race.mjs (new Nx target, races a REAL PostgreSQL). Address a candidate as "project 42's current release" and editing project 42 changes what the release is: every gate result, every signature and every credential downstream then attests to different content under the same name, and NONE OF THEM IS WRONG, because each was true when it was made. There is no moment at which anything failed, which is why nothing catches it. Six fault families, 38 faults, all producible; 34 tests. -->

    • S6.10.a Create release ID independently from mutable object/project IDs and bind target/profile/environment/locale/territory/audience. Evidence: release identity tests. <!-- ReleaseIdentity + releaseIdentityFaults. The id is MINTED and carries no object id, no slug and nothing that can be renamed; what binds it to the world is six recorded dimensions, because "the German release" and "the German release for the EU store" are different things a derived id cannot tell apart. release-id-that-contains-the-subject-reference is the version that survives a review of the minting code — generated randomly and then FORMATTED as project-42-release-3, which gives back every property the minting was for. two-releases-binding-the-same-dimensions is the other direction, with a counterexample test showing a different territory is a different release and not a finding. -->
    • S6.10.b Freeze exact source revisions, artifact byte refs/hashes, dependencies, bundle profile, manifests, contracts, policies/gates, approvals, and publisher config. Evidence: release lock manifest. <!-- ReleaseLock + releaseLockFaults over all ten inputs. THE LOCK NAMES REVISIONS, NEVER REFERENCES: "the policy set" is not frozen and "policy set at revision 41" is, and the test for every entry is whether two people resolving it a year apart get the same bytes — which a reference with no version fails while looking exactly like one that does not. The three inputs people leave out are policies, gates and publisher-config, and the detail says why: they feel like environment rather than content, and a gate definition that changed after approval means the release that shipped passed a check nobody can re-run. A pin without a digest is refused wherever the input has bytes, since a re-publish under the same reference is otherwise invisible. -->
    • S6.10.c Create the lock atomically with approval/candidate transition and reject stale expected source revisions. Evidence: lock race tests. <!-- LockAdmission + lockAdmissionFaults, and tools/release-lock-race.mjs against a real PostgreSQL 16.14. lock-written-outside-the-candidate-transition is easy to ship precisely because the two writes are adjacent IN THE CODE and not in the database; the milliseconds between them are exactly the window in which the content moves. THE TOOL FOUND A DEFECT IN MY OWN FIRST DESIGN: the atomic insert carried NOT EXISTS (SELECT 1 FROM release_lock WHERE held) and produced THREE simultaneous winners, because that clause is a READ and reads do not conflict with another transaction's uncommitted write under read-committed. The fix is a partial unique index, and held-lock-uniqueness-left-to-a-predicate is a shipped fault kind because of it. CONTROL 1 now drops the index and reproduces the three winners from the same statement, so the finding is demonstrated rather than described. -->
    • S6.10.d Prevent edits, restores, recomputation, metadata changes, or target changes from mutating the locked candidate. Evidence: post-approval mutation tests. <!-- MutationAttempt + lockedMutationFaults over eight mutation kinds. REFUSING IS HALF THE ANSWER: a conflict and nothing else sends somebody to edit the underlying object instead, which is the same mutation by a route the lock cannot see, so refusal-with-no-candidate-to-go-to demands the refusal name where the change belongs. mutation-refused-against-a-lock-that-is-not-held is the counterexample rule, with a test — a released, cancelled or expired lock freezes nothing, and a system that keeps refusing after it is gone is one people route around permanently. The database side is CONTROL 3 in the tool: a guarded UPDATE changes 0 rows while the lock is held, 1 row after it is released, and the same statement without the guard changes 1 row either way. -->
    • S6.10.e Require changed input to create a new release candidate/version linked as replacement/supersession. Evidence: new candidate lineage tests. <!-- CandidateLineage + candidateLineageFaults. changed-input-with-no-new-candidate takes the inputs that ACTUALLY changed and asks whether a candidate exists for them, because the failure mode is not somebody bypassing the lock — it is somebody changing an input the lock froze and everybody agreeing afterwards that it was a small change. two-candidates-superseding-one is a fork presented as a line, resolved by whichever query sorts first, and supersession-that-changed-nothing catches the lock being worked around by re-creating the release. The cycle check walks the chain, because a history that loops has no first version and no current one. -->
    • S6.10.f Audit lock creation, reads, failed mutation attempts, cancellation, expiry, and release outcome. Evidence: lock audit tests. <!-- LOCK_AUDIT_EVENTS + lockAuditFaults. reads-not-audited and refusals-not-audited are split out of the generic missing-event fault because they are the two everybody omits and for OPPOSITE reasons: a read feels like nothing happened, and a refusal feels like nothing happened either. A refused mutation is the only evidence the lock ever did anything, and a system that audits what succeeded and drops what it refused cannot show it was ever holding; a read matters because when a lock turns out to have been wrong the question is who acted on it, and a trail of writes alone names everything the lock did and nobody it reached. expiry-recorded-with-no-outcome is the quiet one — a lock that lapsed and a release that was abandoned become the same row. --> <!-- S6.10 rollup: 4155 kit tests pass (34 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 88 entry points with no drift and no violations. -->
  • S6.11 Add publish eligibility recomputation inside the transaction or release lock. Evidence: time-of-check/time-of-use race test. <!-- src/publish-eligibility.ts + tools/eligibility-toctou.mjs (new Nx target, races a REAL PostgreSQL 16.14). "Inside the transaction" turns out not to be a property of the transaction: with the recomputation and the lock write in ONE transaction, a consent revoked between them is still published at read-committed, because a transaction that only READ a row does not conflict with the transaction that changes it — and it is still published at repeatable-read and at serializable, where the re-read answers from the snapshot the transaction opened with. Raising the isolation level HIDES the change read-committed would have caught. Six fault families, 47 faults, all producible; 54 tests. -->

    • S6.11.a Resolve required gates/approvals/rights/consent/provenance/ dependencies/target policy for exact release candidate and destination. Evidence: eligibility policy tests. <!-- RequirementResolution + requirementResolutionFaults over the item's own seven kinds. Decision 1: eligibility is a function of (candidate, DESTINATION), because the destination decides which rights clearance applies and which gates run at all, so requirement-set-reused-across-destinations means one publish was authorized by checks resolved for the other and the store rules nobody read are the ones that come back. evidence-kind-neither-required-nor- excused is S6.6's "unknown is a value" from the other side: a destination that genuinely needs no rights clearance is real and common, so the rule is not that every kind is required but that every kind is either required or EXCUSED WITH A REASON — "rights do not apply to an internal preview" and "nobody thought about rights" are otherwise the same empty set. There is a counterexample test for exactly that case. -->
    • S6.11.b Re-read current source/dependency/evidence/policy/waiver/signoff versions rather than trusting preview/UI state. Evidence: stale input tests. <!-- ConsultedEvidence + staleInputFaults. Every value carries the ORIGIN it was read from, and only one of the five is evidence; evidence-taken-from-the-request is the publisher supplying the evidence that authorizes the publisher, which arrives looking like a saved round trip. Decision 2 is that a claimed version which does not match the read is a REFUSAL AND NOT A CORRECTION: taking the newer value publishes content nobody looked at under an approval that named different content, which is S6.10's failure with a shorter window and no step at which anything failed. status-recorded-from-a-probe-that-did-not-complete covers both directions of S6.7's not-performed-is-not-pass — a satisfied from a call that threw is the fail-open, and an invalid from a call that timed out is a fabricated failure that sends the evidence owner to investigate a defect that does not exist. -->
    • S6.11.c Recompute eligibility inside the same transaction as release lock or under an equivalent fenced reservation. Evidence: atomicity integration tests. <!-- AtomicSite + eligibilityAtomicityFaults, and every rule here was MEASURED against PostgreSQL 16.14 rather than reasoned about. THE TOOL FOUND THE DESIGN, NOT JUST A DEFECT IN IT. consulted-input-with-neither- a-row-conflict-nor-a-write-predicate fires even when the recomputation and the write share one transaction, because reads do not conflict. And write-predicate-relied-on-under-snapshot-isolation INVERTS the usual advice: the identical INSERT ... WHERE version = $expected refuses correctly at read-committed, where the predicate is evaluated against the current row, and silently accepts at repeatable-read and serializable, where it is evaluated against the snapshot — and serializable does not abort, because two transactions are a dependency and not a cycle, so "publish then revoke" is a legal serial order. A system hardened by raising the level gets quietly worse. What works at every level is a ROW CONFLICT on every consulted input, which CONTROL 5 shows ordering a concurrent revoke after the lock write and raising "could not serialize access due to concurrent update" against one that already committed. The fence branch exists because the item allows an equivalent reservation, and fence-token-the-writer-validates-itself is the S6.10.c finding in a new shape: a caller comparing a token to its own copy before writing is a read-then-write with a comparison in between. -->
    • S6.11.d Reject missing/stale/expired/invalid/unavailable/out-of-scope/ unauthorized evidence with exact safe reasons. Evidence: refusal matrix. <!-- EvidenceRefusal + evidenceRefusalFaults over a shipped STATUS_DISPOSITIONS table, with a test asserting the shipped table produces zero findings for every status. Decision 4: the seven are seven, and their remedies are obtain, re-run, renew, investigate, wait, resolve-for-this-destination and ask-somebody-else — seven different people doing seven different things. unavailable-reported-as-a-failed- check is the one everybody ships: nobody could ask the question, and calling that a failure sends a release owner to argue with a gate owner about an outage. invalid-treated-as-retryable is the opposite mistake — evidence that does not verify verifies no better the second time, and a retry there is the loop a pipeline sits in overnight. unauthorized-refusal-that-discloses-the-evidence is the leak: a refusal that explains which assertion failed turns a publish attempt into a way to read evidence the caller was refused. every-refusal-given-one- disposition is the S6.7 boolean-answers-N-situations check as an aggregate, and it is narrowed against the S4.15 trap — four of the seven legitimately share the plain refuse disposition, so it fires only on a matrix that covers ALL eight statuses and answers them identically, with a counterexample test for the legitimate four. -->
    • S6.11.e Freeze consulted evidence/versions/hashes into the release lock. Evidence: lock provenance assertions. <!-- EligibilityProvenance + eligibilityProvenanceFaults, which takes the consultation as a SECOND argument so the two sides come from different places — a provenance record generated by the pass that wrote the lock agrees with it by construction, which is S6.6's photocopy in a third shape, and frozen-evidence-that-was-never-consulted is what catches it. Decision 5: the lock freezes what was CONSULTED rather than what was required, because "gate G was required" cannot be re-checked a year later and "gate G run 88 over digest X reported pass" can. The consultations that found NOTHING are frozen too, and freeze-that-kept-only-what-was-found is why: keep the satisfied rows and drop the rest and the lock truthfully says the release was published with the evidence it names, which is also exactly what a release published with four of nine pieces of evidence says. Reuses S2.11's ReleaseEligibility for the outcome rather than inventing a second vocabulary for the same six answers. -->
    • S6.11.f Race a source edit, consent revoke, policy change, gate stale, approval expiry, and target change against publication and prove none silently authorize. Evidence: TOCTOU tests. <!-- RaceObservation + eligibilityRaceFaults, and tools/eligibility-toctou.mjs runs all six against a real database: CONTROL 1 asserts the NAIVE path publishes over every one of them before anything else is measured, since a refusal of a race that never happened is not evidence. The tool reports the OBSERVED order rather than the intended one, because a row conflict turns an in-window change into an after-the-write one and that is the mechanism — reporting the intention would hide it. change-before-the-recomputation-that-published is the quieter fault and in one way the worse: a change that landed in the window is a timing problem, and one that landed BEFORE the recomputation and published anyway means that input is not in the decision at all. refusal-that-does-not-name-the-change is why a stopped publish is not enough — the operator presses retry, the second attempt succeeds, and nobody learns a consent was revoked. event-never-raced is S6.5's axis-never-perturbed here, and the tool proves it fires by dropping consent-revoke from its own matrix. A change landing AFTER the write is deliberately not a fault: the publish happened, the candidate is locked by S6.10.d, and refusing there would refuse a legitimate later edit. --> <!-- S6.11 rollup: 4210 kit tests pass (54 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 89 entry points with no drift and no violations. tools/eligibility-toctou.mjs passes every control against PostgreSQL 16.14, and was verified by patching the BUILT dist twice — emptying SNAPSHOT_ISOLATIONS fails CONTROL 4 and disabling the no-conflict-no-predicate branch fails CONTROL 2. -->
  • S6.12 Add verified-in-target state based on read-back, not a successful write response. Evidence: corrupted or missing target content fails. <!-- src/target-verification.ts + tools/target-readback.mjs (new Nx target, runs a REAL S3-compatible object store beside a REAL http target that misbehaves on purpose). A 2xx is produced before the write is durable, before it is visible, before it is indexed, and by the component whose correctness is the question. Six fault families, 43 faults, all producible; 45 tests. Composes S6.9.e's read-back rather than restating it: that owned the saga's verify step, this owns the contract, the binding strength, the seven states and the journeys. -->

    • S6.12.a Define target-specific read-back endpoint/operation, authorization, eventual-consistency delay, retry budget, and stable external identity. Evidence: verification contract. <!-- TargetVerificationContract + verificationContractFaults. retry-budget-shorter-than-the-consistency-delay is the one that survives review, because every field is plausible on its own: ninety seconds of declared eventual consistency, three attempts over thirty, and the two numbers in different files. Every publish to that target then reports pending, the state is dismissed as noise, and the one that never landed looks like all the others — CONTROL 4 measures a real 404 for 400ms and a real 200 afterwards, so the arithmetic is against a clock rather than a comment. read-back-through-a-callback-that-cannot-be- repeated is decision 1: a callback arrives once, and a verification that cannot be re-run is the one thing verification is not. identity-constructed-rather-than-returned is decision 5, and CONTROL 5 makes it concrete by serving a complete, internally valid object at a path built from our own inputs. list-and-filter is deliberately NOT a fault, with a counterexample test — it is the only read some targets have. -->
    • S6.12.b Read back target existence/state and exact bytes/hash or strongest supported binding. Evidence: content integrity tests. <!-- ContentReadback + contentIntegrityFaults, and the tool's CONTROL 0 is the argument. Truncate an object in place and MinIO recomputes its entity tag over what it now holds, so "ask the target for its checksum and compare it to the target's content" reports SUCCESS over a half-written release — asserted BEFORE the real check, both before and after the corruption, so the control cannot go vacuous. That is digest-taken-from-the-target-rather-than-from-its-bytes, S6.6's photocopy in its most convincing disguise, and byte-length-that-does-not-match is kept separate because a short write is the corruption real targets produce and the one a self-consistent checksum cannot see. Decision 3 is the binding LADDER: exact bytes, an entity tag and "something exists here" are three different claims, and if all three write verified the column means the weakest of them everywhere it is read. Both directions are faults — a claim that degraded silently and one somebody inflated. -->
    • S6.12.c Verify metadata, version, audience/permissions, locale/territory, rights/embargo, provenance, relations, and discoverability. Evidence: target field matrix. <!-- TARGET_FIELDS + fieldMatrixFaults over the item's own eight. rights-or-embargo-never-verified and discoverability-never-verified are split out of the generic missing-field fault because their failures are not defects: an embargo that did not take at the target is a release published early — a legal event against a target answering 200 to everything — and an object with the right bytes and the right permissions that nobody can find is the partial success that surfaces weeks later as an absence. CONTROL 6 produces exactly that against a real listing request. field-compared-against-what-we-sent is the quiet one: the observed value came from the plan or the publish response, so the comparison is a value against itself and every field passes on a target that stored nothing. -->
    • S6.12.d Distinguish pending visibility, verified, mismatched, missing, unauthorized-to-verify, target unavailable, and verification failed. Evidence: state union tests. <!-- VERIFICATION_STATES + verificationStateFaults over a shipped STATES_FROM_READ table, so the state has to FOLLOW FROM what the read attempt did rather than merely be a member of the union. verified-without-a-completed-read is decision 4 and the reason the seven exist: CONTROL 3 accepts three real writes and refuses three real reads with 403, which is an ordinary target ACL, and the failure is STABLE — a verifier that swallows it verifies every release to that target forever with nothing ever read. refusal-to-read-recorded-as-a-target-outage keeps a 403 apart from a 503: one is permanent and about us, the other transient and about them, and merging them retries a permission problem forever while reporting somebody else's incident. unresolved-state-with-nobody-to-own-it is where an exhausted retry budget has to land, with a counterexample test showing that naming an owner resolves it. -->
    • S6.12.e Prevent publisher 2xx/callback/queued acknowledgment from setting verified state. Evidence: false-success tests. <!-- VERIFIED_EVIDENCE + falseSuccessFaults: six things get used as proof and one of them is proof. verified-set-by-elapsed-time is the worst because it ALWAYS SUCCEEDS — "we waited five minutes and nobody complained" is how a target down for a week reports every release as verified, and it degrades silently, since the longer the outage the more confident the record. verified-written-outside-the-state-machine is how all the others survive being forbidden: the transition refuses them, the column is still there, and a migration or a support script sets it directly. An operator assertion is a fault too rather than an exemption — S5.12's rule that an override records a decision beside the result and never replaces it. -->
    • S6.12.f Trigger compensation/incident/manual reconciliation on corrupted, partial, wrong-target, or unknown read-back and audit every attempt. Evidence: failure journey tests. <!-- VerificationJourney + verificationJourneyFaults over a shipped REQUIRED_ACTIONS table, with a test asserting the table itself produces zero findings. Two rows are the interesting ones. wrong-target-compensated-automatically: compensation means writing to the target again, usually a delete, and the whole finding is that we do not know whose object is there — undoing a mistake at an address that is not ours is a second, larger mistake made faster. And an unknown read-back is reconciled and never retried, which is S6.9.g arriving here. attempts-not-all-audited is decision 6: a trail holding only the successful verifications describes a system that verifies when things work, and the reads that returned 404, 403 and a timeout are the only evidence the verification could ever fail — exactly the ones a log-on- success path drops. --> <!-- S6.12 rollup: 4256 kit tests pass (45 new plus one for the binding order); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 90 entry points with no drift and no violations. tools/target-readback.mjs passes every control against MinIO and a real http target, and was verified by patching the BUILT dist twice — disabling the target-computed-digest branch fails CONTROL 0 and disabling the verified-without-a-read branch fails CONTROL 3. -->
  • S6.13 Add correction, supersession, unpublish, recall, and tombstone semantics. Evidence: downstream state and audit remain coherent. <!-- src/withdrawal.ts + tools/withdrawal-fanout.mjs (new Nx target, four REAL http targets over real sockets and a REAL PostgreSQL). The five are not severity levels of one action — they answer what was wrong, what to use instead, what people may still see, what has to stop right now, and what is left where the thing was. A single withdraw button collapses them the same way every time, into unpublish, because unpublish is the one with no follow-up work. Seven fault families, 55 faults, all producible; 50 tests. -->

    • S6.13.a Define correction as a linked new version preserving original, correction reason/scope, changed facts/artifacts, notice, and target update. Evidence: correction lifecycle tests. <!-- Correction + correctionFaults. Decision 1: correction-that-edited-the-original-in-place is the whole item, because the correction exists precisely BECAUSE somebody already acted on the wrong version, and editing in place leaves them nothing to compare against — the notice then announces a difference that cannot be observed. CONTROL 6 makes that literal against a real database: after the UPDATE, 0 rows hold the value somebody acted on and 1 holds the corrected one, while an append-only version table holds both. notice-that-is-not-where-the-original-was is the version everybody ships — a notice on the new version reaches people who found the new version, and everybody following the old link sees the wrong content with no sign that it is wrong, which is the population the notice is for. -->
    • S6.13.b Define supersession as a new release identity replacing prior use while retaining navigable lineage and compatibility. Evidence: supersession tests. <!-- Supersession + supersessionFaults, reusing S6.10's SUPERSESSION_REASONS rather than inventing a second vocabulary. Decision 2 is that supersession moves USE, not existence, and the two faults are opposite mistakes about the same reference. superseded-release-no-longer-resolvable breaks every citation made while it was current; supersession-that-redirects-silently is worse, because the citation still resolves — to different content, under the same name — so anybody checking a quotation finds it absent from the thing it is attributed to and nothing anywhere reports an error. That is S6.10's derived-identity failure arriving at the other end of the life cycle. compatibility-not-stated treats unknown as a real value with a counterexample test, because a supersession silent about compatibility is read as drop-in by every consumer that automates the upgrade. -->
    • S6.13.c Define unpublish as removal from active/public presentation with target-specific reversibility, cache/search/catalog cleanup, and no history erasure. Evidence: unpublish tests. <!-- Unpublish + unpublishFaults over six presentation surfaces. search-index-not-cleared and cache-not-invalidated are split out of the generic surface fault for OPPOSITE reasons: the index is a SEPARATE SYSTEM, so removing the object does nothing to it and the withdrawn release stays the first result for its own title; the cache is the SAME SYSTEM WITH A DELAY, so it keeps serving for as long as its lifetime says with nothing anywhere in an error state. CONTROL 2 demonstrates the second against a real PostgreSQL materialized view — the table returns 0 and the view returns 1 for the same deleted row, until an explicit refresh. Decision 3 is reversibility-claimed-over-an-irreversible- target: whether an unpublish can be undone belongs to the target, and a console offering restore over a store that burns the identifier is promising what nobody can deliver. -->
    • S6.13.d Define recall as urgent multi-target withdrawal with severity, authority, notifications, stop-distribution, affected derivative tracking, and incident linkage. Evidence: recall rehearsal. <!-- Recall + recallFaults. Decision 4: reconciliation-before-distribution-stopped is the fault that READS AS DILIGENCE — work out exactly which targets hold it, produce a list, then act — and every minute of that inventory is distribution of the thing being recalled, while the inventory is the part that can be done afterwards. recall-with-no-authority is separated from recall-authority-not-verified because an unauthenticated urgent withdrawal is a denial of service with a polite name: the action is designed to be fast, to bypass review and to reach every target, which is exactly what somebody malicious wants from it. derivative-not-tracked decides whether a recall works at all, and CONTROL 4 shows why against a real database: a one-level WHERE parent = $1 finds one derivative and a recursive CTE finds three, the two it missed being a mirror of an excerpt and a set built from the mirror. -->
    • S6.13.e Define tombstone content, stable identity, reason/category, visibility, rights/privacy-safe detail, replacement link, and retention. Evidence: tombstone tests. <!-- ReleaseTombstone + tombstoneFaults. Decision 5 is that a tombstone is CONTENT and what it does not say is part of it. address-that-answers-not-found is the default everybody gets for free, and CONTROL 3 proves the problem rather than describing it: the withdrawn address and one that never existed returned BYTE-IDENTICAL responses (404, 21 bytes each), so a deliberate withdrawal renders as a broken link that people retry and file bugs about. tombstone-that-names-a-person is the opposite failure and the more serious one — a removal made on somebody's request, explained in public at the address everybody already has, performs exactly the exposure the removal was for, so the category is what has to be public and the particulars are what must not be. tombstone-with-no-retention is the quiet one: a cleanup turns the address back into a plain not-found and the withdrawal becomes invisible again. The replacement requirement is narrowed to the two categories that imply one exists, with a counterexample test for expired. -->
    • S6.13.f Propagate each action to targets, catalog/search/cache, provenance/ledger, dependencies, analytics, notifications, and future eligibility. Evidence: downstream integration tests. <!-- Propagation + propagationFaults over the item's own nine channels. future-eligibility-not-updated is the one that UNDOES THE WHOLE ACTION, and CONTROL 5 produces it against a real database: with a recall row present, the eligibility query STILL returned the release as publishable, because nothing joined the recall table — so the next release train re-publishes recalled content while working exactly as designed, and no step in it is wrong. Adding the join returns 0. provenance-ledger-not-updated is the quiet one: the world says withdrawn and the ledger says published, and the ledger is what an audit a year from now reads. -->
    • S6.13.g Preserve immutable audit and distinguish partial/failed/unknown target outcomes with retry/compensation. Evidence: multi-target fault tests. <!-- WithdrawalRecord + withdrawalAuditFaults, with rollUpOf as a SEPARATE exported function so one piece of code computes what the roll-up should be and another compares it to what was recorded — a single function doing both can only agree with itself (S6.6's photocopy again). Decision 7, and CONTROL 1 is the argument: four real targets, two accepted the withdrawal, and only ONE actually stopped serving. The target that answered 200 and kept serving is the outcome no response code can distinguish, which is why the record is per target and the roll-up is derived. partial-rolled-up-as-complete and unknown-rolled-up-as-complete are split because they are wrong differently: a partial withdrawal is KNOWN to be still serving somewhere and an unknown one MIGHT be, and the second is the one people defend — nothing came back saying it failed. audit-that-can-be-rewritten is here because who ordered it, when distribution stopped and which targets confirmed is the evidence a dispute turns on. --> <!-- S6.13 rollup: 4307 kit tests pass (50 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 91 entry points with no drift and no violations. tools/withdrawal-fanout.mjs passes every control against four real http targets and PostgreSQL 16.14, and was verified by patching the BUILT dist twice — removing rollUpOf's partial branch fails CONTROL 1 and disabling the answers-not-found branch fails CONTROL 3. -->
  • S6.14 Add bundle importer compatibility and migration hooks. Evidence: supported old versions round-trip; unsupported versions fail with a migration path. <!-- src/importer.ts + tools/importer-roundtrip.mjs (new Nx target: REAL tar archives, REAL ajv-compiled schemas at two versions, a REAL P-256 key, REAL gzip). A version we do not know must be REFUSED, not read partially — an importer is the one place where content arrives that we did not produce, at a version we did not choose, from a writer we cannot ask. Seven fault families, 49 faults, all producible; 54 tests. Composes S2.5's importer port (plan/validate/execute/reconcile) rather than restating it. -->

    • S6.14.a Detect archive/profile/manifest/contract versions and validate path/resource/security limits before parsing content. Evidence: importer preflight tests. <!-- ImportPreflight + preflightFaults over ten checks. Decision 1: parsing IS the attack surface, so check-performed-after-parsing is the one that reads as fine in review — the check is right there in the code and it does refuse, after the parser has allocated for the sizes the archive declared and followed the paths it named. compression-ratio-not-checked is split out because it is the limit on nothing you can see, and CONTROL 3 measures it rather than asserting it: a real gzip compressed 8MB to 8.0KB, a ratio of 1026:1, so every bound on the archive passes and the expansion is bounded by nothing. observed-value-not-recorded is the quiet one — a check that passed and recorded nothing cannot be tuned, and a bound nothing approaches reads exactly like one nothing measures. limit-not-declared is narrowed to the five numeric checks, with a counterexample test for a version check that legitimately has no bound. -->
    • S6.14.b Negotiate exact supported, backward-compatible, migratable, too-old, too-new, unknown, and unsupported versions without silent downgrade. Evidence: version matrix. <!-- SupportWindow + an exported versionVerdict that the checker COMPARES AGAINST rather than derives from — two functions, so a recorded decision is judged by something that did not make it. The order is argued: unknown first (a version we have never heard of is not too old or too new, it is off the axis), then withdrawn, and migratable applies only BELOW the floor, because inside the window we read it directly and a migration we could apply but do not need answers nothing. too-new-read-anyway is decision 2 and CONTROL 2 produces it: a v5 reader over a real v6 manifest returns a complete-looking object with no error, and the field the writer marked as one a reader may not skip was never asked for. fields-ignored-with-no-record is what makes a legitimate partial read reviewable at all. -->
    • S6.14.c Define pure ordered migration hooks with from/to version, preconditions, transformed fields/files/refs, warnings/loss report, and provenance. Evidence: migration interface. <!-- MigrationChain + migrationHookFaults, and observedLosses as a REAL exported key-path diff. Decision 3: the loss report is COMPUTED, not declared, because a migration asked what it dropped answers from the same assumptions that made it drop it — S6.6's photocopy in the one place where the thing being audited is a function somebody wrote in a hurry. The diff compares PRESENCE, not equality: a rewritten value is the migration doing its job and a removed one needs somebody to have agreed, and it reports a key removed outright, a subtree that became a scalar, and a value nulled where it held content. CONTROL 4 runs a real 2→4 migration that drops legacyProfile and declares nothing; the diff finds it. hook-that-is-not-pure matters more than it looks: a migration reading a clock produces a different bundle every run, so S6.5's determinism is undone by the one step between the archive and the import. -->
    • S6.14.d Preserve unknown compatible extension data and reject unknown required semantics. Evidence: forward compatibility tests. <!-- ExtensionField + forwardCompatibilityFaults. Decision 4 is that this is TWO RULES, and a reader with one always has exactly one of the two failures. Drop everything unknown and a round trip deletes an extension somebody else depends on, surfacing months later as a difference nobody can account for — CONTROL 6 does exactly that and shows the preserving reader returning the field byte-for-byte. Accept everything unknown and a field whose whole purpose is to change how the document is interpreted is ignored by the one component that had to honour it. field-with-no-must-understand-declaration is what makes two rules possible at all: without the flag the reader must pick one policy for every unknown field, and whichever it picks is wrong for half of them. unknown-optional-field-rejected is the over-strict direction, because a reader that refuses every addition means nobody can extend the format. -->
    • S6.14.e Validate and rehash after each migration, produce new bundle/ manifest identity, and link source/migration lineage. Evidence: migration integrity tests. <!-- MigratedBundle + migrationIntegrityFaults. signature-carried-over-a-migration is the one worth stating loudest, and CONTROL 5 settles it with real cryptography: a real P-256 signature VERIFIED over the v2 archive and FAILED over the migrated one, so reattaching it is not a policy choice — it files a real signature by a real key against bytes the signer never saw, and it survives review because everything about it still looks valid. migrated-bundle-not-revalidated-against-the-target-schema is the quiet one, and CONTROL 1 shows why with two real ajv-compiled schemas: the v2 manifest FAILS the v4 schema and the migrated one passes it, so validating the output against the source schema tests the wrong contract — the fields the migration existed for are exactly the ones the source schema does not know about. -->
    • S6.14.f Import as plan/preview then idempotent authorized execution with conflict, partial failure, rollback/cleanup, and audit. Evidence: importer journey tests. <!-- BundleImportJourney + bundleImportFaults, composing S2.5's ImportPlan/ImportValidation/ImportExecution port rather than inventing a second vocabulary — this records the facts about a particular run that the port cannot express. partial-import-reported-as-a-failure is decision 6 and the reason partially-imported exists as an outcome: a run that created half its refs and reports "failed" tells the caller to retry, and the retry either creates the other half twice or refuses every ref because it is already there, with neither recoverable without somebody reading the store. execution-against-a-plan-that-moved is S6.9's rule arriving here — the digest approval was bound to is the digest execute has to see. unresolved-conflict-executed is narrowed with a counterexample: refusing on an unresolved conflict is the correct outcome, not a finding. -->
    • S6.14.g Return exact supported target/migration path or upgrade requirement for unsupported input. Evidence: failure response tests. <!-- ImportRefusal + refusalResponseFaults. Decision 7: the caller is on the far side of a version boundary — they cannot inspect this reader, and the only three actions that exist are produce a different version, apply a migration, or upgrade the reader. migratable-refused-without-offering-the-migration is the expensive one because the answer was ALREADY COMPUTED by the verdict, and the refusal that omits it sends somebody to rebuild a bundle that did not need rebuilding. too-new-refusal-with-no-required-reader-version exists because the caller cannot make their bundle older, so the reader version is the only actionable fact that case has. migration-path-that-does-not-end-at-the-supported-target catches the refusal that is internally inconsistent — follow it and land somewhere still refused. And refusal-for-a-version-that-is-supported is the counterexample rule with a test. --> <!-- S6.14 rollup: 4362 kit tests pass (54 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 92 entry points with no drift and no violations. tools/importer-roundtrip.mjs passes every control against real tar, real ajv, real gzip and a real P-256 key, and was verified by patching the BUILT dist twice — silencing observedLosses fails CONTROL 4 and disabling the too-new branch of versionVerdict fails CONTROL 2. -->
  • S6.15 Add golden round-trip tests for every domain bundle profile. Evidence: build, archive, import, verify, and rebuild are deterministic. <!-- src/golden-round-trip.ts + tools/golden-round-trip.mjs (new Nx target, builds and archives every SHIPPED profile with real GNU tar). This is the last item in S6 and the one that keeps the rest true — and the easiest in the phase to write so that it can never fail, which is why every control starts from the exported WORKED_BUNDLE_PROFILES registry rather than a list in the tool. Seven fault families, 40 faults, all producible; 41 tests. S6 IS NOW CLOSED. -->

    • S6.15.a Register minimal, full, boundary, localized, provenance-rich, legacy-supported, and invalid fixtures per domain profile. Evidence: profile fixture manifest. <!-- ProfileFixture + fixtureRegistryFaults over the item's own seven classes, checked per registered profile. profile-with-no-refused-fixture is stated separately from the missing- class fault because it changes what the suite PROVES: everything else measures whether the pipeline still accepts what it accepted yesterday, and only a fixture that must be refused measures whether it still refuses — a corpus without one goes green against a build that validates nothing. legacy-fixture-that-declares-the-current-version is the quiet one: a legacy fixture whose content says it is current exercises no migration, so it is a second copy of the full fixture under another name while the corpus reports seven classes. -->
    • S6.15.b Build logical bundle from authoritative inputs and compare manifest/layout/component inventory/provenance to goldens. Evidence: build golden tests. <!-- BuildGolden + buildGoldenFaults. Decision 1 is the whole item: a golden is a RECORDED answer, and the three ways of losing that are named separately because they arrive by different routes. regenerated-at-test-time is the comparison agreeing with itself; captured-on-failure is the same thing one step removed, where the first red build wrote the expectation so whatever happened that day is now the definition of correct; and rewrites-on-mismatch is the mechanism that turns a checked-in literal into either of the other two the first time it goes red. aspect-not-pinned-by-the-golden covers the four the item names — an unpinned aspect can change in any way at all and the golden still passes. -->
    • S6.15.c Archive twice under perturbed order/clock/locale/platform inputs and compare exact bytes/hash/entry metadata. Evidence: archive reproducibility tests. <!-- ArchiveRun + archiveReproducibilityFaults, iterating S6.5's PERTURBATION_AXES IMPORTED rather than restated: a second list of the things that vary between two machines would drift, and the one that drifted would be the one nobody was reading. The tool varies SIX of the seven for real — shuffled input order, touched mtimes, a Turkish locale, a Kiritimati time zone, a 0077 umask and interleaved writes — and asserts FIRST that localeCompare under Turkish disagrees with byte order, so the sort being byte order is load-bearing rather than incidental. The seventh, platform, cannot be varied on one box, and the tool asserts the checker REPORTS it as never perturbed rather than narrowing the axis list to hide it. entry-metadata-never-compared closes the other direction: two archives hash alike when a normaliser flattens modes, uids and mtimes, so the tool reads the tar headers and compares what they were flattened TO. two-baselines-that-disagree fires before any axis is judged, because a build that is not deterministic against itself makes every axis noise. -->
    • S6.15.d Import and migrate where needed, then compare semantic objects, artifact bytes/hashes, refs, rights/consent, gates, and provenance. Evidence: import parity tests. <!-- ParityComparison + importParityFaults over the item's own eight dimensions. comparison-with-one-side-derived-from-the-other is the photocopy where it is hardest to see: comparing the imported object to the object the importer just produced from the same parse tests that a function is deterministic, not that the bundle survived, so both sides have to be read. rights-or-consent-never-compared is split out because those two are the dimensions whose loss is not a defect — a bundle that arrives with its rights position flattened is content being used on terms nobody agreed to, and it imports perfectly. legacy-fixture-imported-without-a-migration closes the loop with S6.15.a: a legacy fixture that took the current path exercised nothing. -->
    • S6.15.e Verify schema/path/hash/signature/trust/profile/target semantics and expected invalid fixtures. Evidence: verification matrix. <!-- VerificationRow + verificationMatrixFaults over seven aspects in BOTH directions. expected-refusal-that-was-accepted is what the matrix exists for: everything else says the pipeline still handles what it handled, and only an invalid fixture that got through says a check stopped running — which produces no error anywhere else, ever. aspect-that-did-not-run is kept apart from a mismatch for the S6.7 reason: an unreachable trust store and a valid chain both leave the signature aspect with nothing to report, and only one of them is fine. CONTROL 5 in the tool produces the refusal against real profiles, with the naive "did we write every file we meant to" check asserted to accept the same input first. -->
    • S6.15.f Rebuild from imported state and require byte identity when the profile/version promises it or documented semantic identity otherwise. Evidence: rebuild comparison. <!-- RebuildComparison + rebuildFaults. Decision 4: neither answer works everywhere, because a profile whose archive embeds a signature cannot promise byte identity after a re-sign, and one promising only semantic identity with nothing written down has promised nothing a test can check. So the promise is per profile and profile-that-promises-no-identity-at-all is the finding — a rebuild comparison against no promise passes whatever comes out. CONTROL 6 does the real version: each archive is extracted and rebuilt FROM THE EXTRACTED BYTES, and the two are byte-identical for both shipped profiles. -->
    • S6.15.g Fail when a registered domain profile lacks current golden coverage. Evidence: registry coverage gate. <!-- goldenCoverageFaults(profiles, coverage) — and the argument order is the entire item. It walks the SHIPPED registry and asks what coverage exists for each, rather than walking the coverage and reporting what it found; a report built the other way round finds every golden covered by definition, and a profile added tomorrow with no fixtures is precisely what it cannot see. CONTROL 1 proves that against the live registry by dropping one profile's coverage and requiring the finding to NAME bundle.dataset-release. coverage-for-a-profile-that-is-not-registered is the other direction and is not symmetric noise: goldens for a profile nobody ships spend the suite on something that cannot break, and they are why a coverage percentage looks healthy while a live profile has nothing. This is the mechanism by which S6.15 stays true after it is marked. --> <!-- S6.15 rollup: 4404 kit tests pass (41 new); tsc clean on lib and spec; nx build clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist; the export map regenerates to 93 entry points with no drift and no violations. tools/golden-round-trip.mjs passes every control over both shipped profiles, and was verified by patching the BUILT dist twice — silencing registered-profile-with-no-coverage fails CONTROL 1 and silencing axis-never-perturbed fails CONTROL 3. §S6 is closed with this item. -->

S7 — Shared application shell and working surfaces#

  • S7.1 Build a restrained workbench shell with navigation, primary canvas, inspector, status region, and command palette. Evidence: responsive browser snapshots. <!-- Six regions and exactly one main. The kit owns everything that is not a matter of taste — libs/oshun/workbench-kit/src/shell.ts, the 94th entry point — and apps/oshun/web/src/components/workbench/ renders it. That split is the point: the shell would otherwise be re-implemented once per workbench, and the parts that must not vary (which routes an actor may see, where focus goes when a panel closes, what a status slot is allowed to say, which arrangement a width gets) are computed rather than described. The parent's "responsive browser snapshots" are met by a real Chromium driving the shipped components at 640/1100/1600px, not by an image: a screenshot proves a machine once and pnpm nx workbench-shell-browser @oshun/web proves it every time. All seven children complete. --> <!-- CORRECTION to the naive reading of this item: jsdom cannot settle any of the three claims a shell actually makes. It has no layout engine, so "the arrangement fits" is unmeasurable; no Tab traversal, so "the keyboard reaches the work" is unmeasurable; and no computed style, so "the surface stays inside its budget" is unmeasurable. A shell verified only in jsdom is verified against a simulator of the only thing it exists to do — which is why the browser tool exists alongside the 33 component tests rather than instead of them. -->

    • S7.1.a Build landmark-based global/workbench navigation from authorized route descriptors with current location, breadcrumbs, collapse, and keyboard traversal. Evidence: navigation tests. <!-- evidence: libs/oshun/workbench-kit/src/shell.ts §S7.1.a + apps/oshun/web/src/components/workbench/WorkbenchNavigation.tsx. authorizedNavigation COMPOSES S1.8's hostNavigation (host, audience, discoverability, lifecycle) and adds the two filters a menu cannot skip: the roles the route declares, and its flags. Flag handling finally reads whenUnavailable, which S1.8 made every route declare and nothing has ever consumed — and it separates "the flag service answered and knows nothing about this flag" (take defaultEnabled) from "the flag service did not answer" (take hide / disable-with-a-reason / fail-closed / fail-open). Collapsing those two makes every flag-service outage either a silent feature launch or a silent feature removal. breadcrumbTrail is the first reader of S1.8's breadcrumb mode, and it walks the AUTHORIZED tree so a trail cannot name an ancestor the actor may not open — a breadcrumb is a disclosure surface, and "Clients / Acme / Brief" tells you Acme exists whether or not the link works. navigationTraversal owns the single roving tab stop, and navigationTraversalFaults reports the collapsed subtree that stayed tabbable and the current location hidden behind a collapsed ancestor. navigationFaults carries S1.8.g's honesty check up a level: a route in neither the tree nor the exclusion list has been silently dropped. Proof: 135 kit tests (shell.spec.ts, every route fixture built through parseRouteDescriptor) + 33 jsdom tests + real Chromium CONTROL 2, which walks the document with real Tab presses and asserts the navigation contributes exactly ONE stop for four routes; FAULT 2 makes every item tabbable and the same walk reports 4. -->
    • S7.1.b Build the dominant primary workspace region with domain plugin composition, loading boundary, selection context, and no decorative wrapper cards. Evidence: two-archetype fixtures. <!-- evidence: workspaceFaults in shell.ts + the editor-centric and operations-centric fixtures in apps/oshun/web/src/components/workbench/__fixtures__/workbenchShellFixture.tsx, read by BOTH the jsdom tests and the browser proof — one fixture, because a browser check with its own host measures a page nothing else renders. Faults: two surfaces claiming the primary region (two dominant canvases is two half-sized ones), a plugin surface composed by no declared extension point, a loading placeholder that shows values which look final, a placeholder that does not hold the layout, a loading state that is visual only, a cancel control for work that keeps running, a selection the inspector does not read, a decorative wrapper card, and a workspace below the 50% width share the shell thesis requires. Real-browser CONTROL 3 measures that share from getBoundingClientRect at 640/1100/1600px rather than from the stylesheet. -->
    • S7.1.c Build contextual inspector placement, pin/replace/close, resizable width, overflow, and focus restoration. Evidence: inspector interaction tests. <!-- evidence: openInspector/pinInspector/closeInspector in shell.ts + WorkbenchInspector.tsx. Replace-unless-pinned, and re-opening a subject keeps its ORIGINAL opener — the second opener is where the actor is now, the first is where they were when the entry appeared, and restoring to the newer one sends them somewhere they never left. THE FOCUS LADDER IS THE ITEM: opener, then the entry that took its place, then the region — never the document body, which is the default every framework gives you. closeInspector takes liveElementIds because element.focus() on a detached node moves focus to the body and reports nothing; the component checks document.getElementById before it hands the id over. Keyboard resizing is not an alternative to the pointer one, it is the only one a keyboard user has. Proof: jsdom restores focus to the opening row and drives the resize to its bound; real-browser CONTROL 5 confirms it with a real click, and FAULT 5 removes the opener from the document first and confirms focus lands in the inspector region rather than on the body. -->
    • S7.1.d Build persistent compact status for save/sync, connectivity, capability/fixture, job, gate, review, publish, and stale state. Evidence: status matrix. <!-- evidence: STATUS_SLOTS + STATUS_VOCABULARY + statusFaults in shell.ts + WorkbenchStatusRegion.tsx. Eight slots, each with its own closed vocabulary, and EVERY vocabulary ends in unknown — a slot that disappears when it has nothing to report renders identically to one saying everything is fine, so slot-missing-from-the-region is a fault and the region always renders all eight. state-asserted-with-no- observation refuses the shell asserting a fact nobody measured; unknown-state-carrying-an-observation refuses the opposite, a measurement taken and discarded. fixture-source-without-a-fixture-state is the §2.2 rule: fixture data may not present itself as live. Every state carries a noncolour glyph, and the kit test walks all 44 states in all 8 vocabularies. -->
    • S7.1.e Build command palette discovery, typed capability commands, disabled explanations, search/groups, keyboard behavior, and recent commands without bypassing authorization. Evidence: palette tests. <!-- evidence: paletteProjection + paletteFaults in shell.ts + WorkbenchCommandPalette.tsx. A command the actor may not run is DISABLED WITH A REASON or NOT THERE — the choice between them is a disclosure decision (concealedCommandIds), and the third option, offered-enabled-then-refused-by-the-server, is what available-command-the-actor-may-not-run exists to stop. Recency is applied AFTER authorization: a recent list records what was allowed once, and re-offering it is the most ordinary way a revoked grant keeps working (recent-command-offered-without-reauthorization). Role and capability are separate refusals with separate reason keys, so holding the role while the capability is not configured still says why. Search matches TRANSLATED labels and keywords — matching the command id makes every command findable in exactly one language. RESERVED_CHORDS refuses a binding the browser takes first, and a destructive command with no deliberate modifier is a fault. Proof: 15 kit tests + 8 jsdom tests including the id-only search returning nothing. -->
    • S7.1.f Compose wide three-region, compact two-region, and single-region review arrangements with preserved selection/focus. Evidence: responsive browser tests. <!-- evidence: arrangementForViewport + ARRANGEMENT_REGIONS + arrangementFaults in shell.ts, and WorkbenchShell.tsx choosing the arrangement from that function rather than from a CSS breakpoint — so the layout and the check cannot drift apart, which a second copy of the numbers in a media query guarantees they would. Preservation is only observable ACROSS a transition, so arrangementFaults takes two observations: a single snapshot of a compact layout is indistinguishable from one that just threw the actor's selection away. A region a narrower arrangement drops keeps a disclosure control (dropped-region-with-no-alternate-affordance). Real-browser CONTROL 3 renders all three at real viewport widths and FAULT 4 forces the wide columns at 600px. THE FINDING FROM THAT FAULT: the shell sets overflow: hidden so the page can never scroll sideways, which means an arrangement that does not fit is CLIPPED, not scrolled — and clipped content is unreachable by any means. The document-level check reported a clean layout over it, so the tool measures the shell's own scrollWidth/clientWidth as well, and asserts the document-level check stays silent to prove the shell measurement is the one doing the work. -->
    • S7.1.g Apply restrained tokens: two type families max, one accent, semantic status cues, minimal borders/shadows, and cards only where the card is the interaction. Evidence: visual conformance snapshots. <!-- evidence: restraintFaults in shell.ts + WorkbenchShell.module.css, which IS the budget: two families, one accent, nine semantic status colours, one shadow reserved for the palette. Written as COUNTS OVER WHAT A BROWSER COMPUTED rather than as adjectives, because a restraint rule written as an adjective is a design review nobody can fail. Real-browser CONTROL 4 reads computed font-family, the --wb-accent custom property, the computed color of each rendered status slot and every non-none box-shadow; FAULT 3 injects a third family and a second accent and both are reported. TWO DECISIONS THE REAL BROWSER FORCED. (1) An INSET shadow is a line, not an elevation: the 2px inset rule marking the current route would otherwise count as a second elevation and fail the check on the one thing it exists to allow. (2) A CARD MUST DECLARE THE INTERACTION THAT MAKES IT ONE (data-card-interaction): inferring "selectable" from aria-current scored every row that was not selected RIGHT NOW as decorative, which is the opposite of what it is. Status tokens are read from the RENDERED slots, weakest instance winning per tone (S6.12's ladder), and FAULT 6 strips a slot back to a coloured dot and confirms status-token-signalled-by-colour-alone fires. --> <!-- S7.1 rollup: 4540 kit tests (135 new) + 33 jsdom component tests + apps/oshun/web/tools/workbench-shell-browser.mjs (5 CONTROLs, 7 FAULTs) driving real Chromium 1223 over the SHIPPED components and stylesheet, bundled from source by esbuild with the checkers imported from the BUILT dist. tsc clean on kit lib and spec; nx build @oshun/workbench-kit clean; production-scan, route-contract-gate, consumer-contract-gate, api-inventory-gate, router-vocabulary-scan and packed-consumer-smoke all pass over the rebuilt dist (94 subpaths); the export map regenerates to 94 entry points with no drift and no violations; eslint clean on the new app components including jsx-a11y. The tool was verified by patching the BUILT dist twice — silencing more-than-one-main-landmark fails FAULT 1, and silencing status-token-signalled-by-colour-alone fails FAULT 6 — and passes again over the restored build. --> <!-- REGISTRATION (five places): src/index.ts, package.json exports, tsconfig.base.json alias, the KIT_OWNED_ENTRY_POINTS list in scripts/v1-workbenches/generate-workbench-kit-export-map.mjs, and the workbench-shell-browser target on apps/oshun/web/project.json (the tool drives the app, so the target lives with the app and dependsOn the kit build). The app also gains @oshun/workbench-kit aliases in its tsconfig.json and vitest.config.ts. -->
  • S7.2 Provide route-level loading, empty, unauthorized, not-configured, degraded, offline, and failure states. Evidence: component and Playwright state matrix. <!-- Seven states are named here and they are usually written as seven independent components. THEY ARE NOT INDEPENDENT. A route load asks several questions at once — who is this, may they, is the capability configured, is the network there, did the query return anything — and the answers arrive out of order and disagree. Exactly one thing renders, and which one is a decision that in most codebases is made by whichever if somebody wrote first. libs/oshun/workbench-kit/src/route-states.ts (the 95th entry point) makes it explicit: resolveRouteState orders the answers, BLOCKER_PRECEDENCE writes the order down as data, and RouteStateBoundary in the app renders what it decides. Two more distinctions carry the rest: BLOCKED and ADVISED are different answers (degraded is content plus a limitation, not a state instead of it), and PENDING IS AN ANSWER TOO AND IT IS NOT "NO" — a pending authorization read as an allow is optimistic disclosure, and a spinner over a refusal that is already known makes the refusal slower than the success. All eight children complete. --> <!-- THE FINDING THAT JUSTIFIES WRITING THE ORDER DOWN: written by hand, the route precedence put step-up-required ABOVE unauthorized. That reads sensibly — offer the stronger session first — and it disagrees with S4's OUTCOME_PRECEDENCE, which already decided that a denial beats a step-up: sending somebody through an authentication they will be refused after both wastes the authentication and hints the object is reachable. routePrecedenceFaults() now composes S4's list and fails on any disagreement, so the two orders cannot drift again, and S4's unavailable (authorization could not be evaluated) got its own blocker rather than sharing a word with a capability that is unavailable — two states with different causes, different remedies and different people to chase. -->

    • S7.2.a Provide skeleton/progress loading that preserves layout, announces status, supports cancellation where real, and never displays fake final values. Evidence: loading tests. <!-- evidence: routeLoadingFaults in libs/oshun/workbench-kit/src/route-states.ts, composing S7.1's LoadingBoundary rather than restating it. THE ITEM-SPECIFIC FAULT IS determinate-progress-that-nothing-measured: a bar that reaches 90% on a timer is a number the code never computed, which is the one thing §2.2 forbids everywhere else, and the honest control for unmeasured progress is an indeterminate one. Also reported: an indeterminate control carrying a value nothing will render, a skeleton still showing the route being LEFT (the actor reads the old page believing it is the new one), an ASSERTIVE loading announcement (loading is not an emergency, and this is the setting that makes a workbench unusable), a silent one, and a cancel control the request outlives. The shipped boundary renders a skeleton with no digits in it, asserted in jsdom. -->
    • S7.2.b Distinguish first-use, no records, no matches, filtered-out, retention/deletion, and no permission-visible records. Evidence: empty state fixtures. <!-- evidence: supportedEmptyKind + emptyStateFaults + seven EMPTY_KINDS. THE CLAIM MUST BE SUPPORTED BY THE EVIDENCE THAT PRODUCED IT — emptyStateFaults compares what the surface rendered to what the counts allow, so no-records asserted while a filter is active is a fault rather than a wording choice. The ORDER inside supportedEmptyKind is the load-bearing part: concealment is checked BEFORE absence, because "there are none" is false the moment some exist that this actor may not see, and that particular falsehood is the one a shell tells by accident. concealed-count-disclosed refuses the other direction: how many records are hidden is enough to confirm one exists, so the count never renders — proved by a jsdom assertion that the none-visible page contains no digit at all, and again in the browser, where FAULT 2 injects the count and confirms the check can see it. first-use-with-no-way-to-create-anything applies to exactly one kind, with a counterexample test that the other six are not required to offer an action. -->
    • S7.2.c Provide unauthenticated, unauthorized, forbidden-concealed, step- up-required, and expired-session behavior without protected disclosure. Evidence: authorization state tests. <!-- evidence: authorizationStateFaults + the boundary's concealed branch. CONCEALMENT IS A RENDERING, NOT A STATUS CODE: the boundary renders the REAL not-found component through renderNotFound, not a second page built from the same strings, so "indistinguishable" is a fact about one component instead of a promise about two. The kit takes the not-found baseline as an INPUT for the same reason — a comparison against a constant written in the checker would pass while the actual not-found page said something else. Browser CONTROL 3 compares the two ACCESSIBILITY TREES byte for byte (the genuine one is rendered outside the boundary by its own scenario), and FAULT 1 changes one heading and confirms the comparison sees it. Also: a refusal may not name the object it refuses, a step-up with no challenge is a refusal with no way forward, an expiry the actor did not choose may not destroy their work, and return-url-carries-protected-values refuses the return-to that travels through the identity provider, the history and the referrer. -->
    • S7.2.d Render server not_configured capability, missing category, allowed resolver, and safe next step without setup success claims. Evidence: not-configured route tests. <!-- evidence: notConfiguredFaults(state, presentation). S1.6 already made not_configured carry missing[] and resolutionAuthority, and NOTHING HAS EVER READ EITHER — this is their first reader, and the check is a comparison rather than a restatement: the page and the server state are supplied separately, because a page generated from the state agrees with it for every input (S6.6's photocopy). Faults in both directions: a category the server reported and the page dropped (the fixer fixes the ones they were told about and the capability stays down) and one the page invented (somebody configures what was already configured). next-step-addressed-to-an-actor-who-may-not-take-it uses AUTHORITY_ROLES, where self-service requires none — that is the whole difference, and the counterexample test proves the same page is fine when it tells this actor somebody else will do it. SECRET_SHAPES scans the rendered detail for a URL with credentials, an internal host, a bearer token and a private address, because a configuration error message is the most natural place in the system for a connection string to reach a screen. -->
    • S7.2.e Render affected/retained capability, limitation, cause, freshness, and recovery for degraded state. Evidence: degradation tests. <!-- evidence: degradationFaults(state, presentation, now), again a comparison rather than a rendering. DEGRADED WITH NOTHING RETAINED IS NOT DEGRADED — usability-disagrees-with-the-state refuses a degradation notice over a configured or unavailable capability outright, composing S1.6's three-valued usability (the function that already refuses to round degraded to a boolean) instead of a fourth opinion. The retained-half fault is the one people omit: a notice carrying only what broke reads as an outage and stops people using what still works, so retained-behaviour-dropped-from-the-notice is a fault and the shipped banner renders BOTH lists. Freshness composes isStateFresh rather than recomputing an age — two freshness rules eventually disagree and the one on screen wins — and expired-degradation-presented-as-current catches a banner still steering people away from work that now succeeds. -->
    • S7.2.f Render offline versus server/provider unavailable distinctly and state readable/editable/queued/prohibited behavior. Evidence: connectivity tests. <!-- evidence: connectivityFaults + the resolver's connectivity rung. OFFLINE IS ABOUT US AND UNAVAILABLE IS ABOUT THEM — S6.12's 403/503 split one layer up, and it decides whether the actor restarts their router or waits for somebody else. Both directions are faults. Offline BLOCKS only what cannot be served from what is already held: a surface that can be read offline is content plus an offline-readonly advisory, because hiding readable work behind a network notice is the offline behaviour people complain about most. The four abilities are checked against what is actually offered: an edit control under readable, editing against a queue that does not survive a reload, writes held in memory only (a silent data-loss event the actor was told would not happen), and a queue promised where nothing could ever be queued. Browser CONTROL 5 is a REAL disconnection: context.setOffline(true) flips navigator.onLine, the live-connectivity scenario reads it rather than a fixture, the offline state renders, and reconnecting restores the content with no reload. -->
    • S7.2.g Render validation, dependency, provider, conflict, limit, partial, internal, and unknown failure with preserved work, retry eligibility, remediation, and correlation. Evidence: failure tests. <!-- evidence: RETRYABLE_FAILURES + failureFaults. RETRY ELIGIBILITY IS A PROPERTY OF THE CLASS, decided once, not a button each surface chooses: offering retry on a validation failure teaches people that retry is decoration, and withholding it on a dependency failure makes them redo work that would have succeeded. conflict is the interesting entry — not retryable, because the state it collided with has moved, so the honest affordance is reload-and-reapply. A failure that discards the actor's input is a data-loss event wearing an error message. partial-failure-presented-as-total stops the actor repeating an operation and doing the successful part twice, and partial is resolved as CONTENT plus an advisory rather than as a blocker for the same reason. unknown-failure-claiming-a-cause refuses a guess that sends somebody to investigate a defect that may not exist, and the secret scan runs on internal/unknown only — with a counterexample proving a validation message containing an IP-shaped string is not flagged. -->
    • S7.2.h Test direct load, client navigation, refresh, back/forward, reconnect, locale/theme/density, keyboard, and screen reader for every state. Evidence: Playwright matrix. <!-- evidence: apps/oshun/web/tools/route-states-browser.mjs — 8 CONTROLs and 4 FAULTs over real Chromium, 23 scenarios covering all 12 ROUTE_BLOCKERS. SIX OF THE EIGHT JOURNEYS DO NOT EXIST IN JSDOM: no history, no reload, no network stack to take away, no prefers-color-scheme, no Tab, no accessibility tree. THE FINDING THAT MADE THIS TOOL WORTH WRITING: page.goto('...#name') from a page already on that origin is a SAME-DOCUMENT navigation — nothing reloads, React never remounts, and focus stays where the last keystroke left it. Half the matrix was measuring an in-page state change while calling it a direct load; directLoad() adds a query so each one is a real document load, and the hash stays what the fixture reads. CONTROL 1 also compares the page against a SECOND evaluation of the kit in the harness process over the same inputs (exposed as window.routeStateInputs), which is why weakening the built dist is detectable at all. CONTROL 6 sweeps dark theme, reduced motion, an RTL locale, 200% text and a 360px viewport over every scenario and asserts no state changes and nothing overflows; CONTROL 7 walks every state's actions with real Tab presses and asserts that a state which REPLACED the content took focus while one that only added an advisory did not. --> <!-- S7.2 rollup: 4634 kit tests (94 new) + 23 jsdom boundary tests + the browser matrix, all green. Verified by patching the BUILT dist twice: mapping conceal to unauthorized fails CONTROL 1's page/kit agreement, and silencing the concealment-beats-absence rule in supportedEmptyKind fails CONTROL 8 — both pass again over the restored build. tsc clean on kit lib and spec and on the new app components (via a temporary noCheck: false project, proved non-vacuous by injecting a type error); nx build, the export map at 95 entry points, and all six kit gates clean. Registration: src/index.ts, package.json exports, tsconfig.base.json, the export-map generator, and the route-states-browser target on apps/oshun/web/project.json. --> <!-- TRAP HIT: PRECEDENCE_FAULTS was invented AFTER the pre-write export-name grep and collided with trust-zones.ts. The kit typecheck passed and public-api.spec.ts failed with a rolldown parse error on the generated index. Renamed to ROUTE_PRECEDENCE_FAULTS; the grep must be re-run over the finished file, not the planned one. --> <!-- CORRECTION: resolutionFaults, ResolutionFault, FAILURE_CLASSES and the over-generic known/pending/Unknown were also renamed for the same reason (artifact-rehash, capability-state). -->
  • S7.3 Build a live capability banner driven by server probes. Evidence: fixture mode cannot masquerade as live. <!-- libs/oshun/workbench-kit/src/capability-banner.ts (96th entry point, 56 tests) + apps/oshun/web/src/components/workbench/ CapabilityBanner.tsx + tools/capability-banner-browser.mjs, which runs a REAL node:http probe endpoint and lets real Chromium actually fetch it — 9 controls, 2 faults, 23 real requests over a socket. The evidence clause is a claim about what happens when a REQUEST is made, and there is exactly one way to check it: make the request. All six children complete. --> <!-- THE LOAD-BEARING TYPE IS NOT THE BANNER STATE, IT IS THE ORIGIN OF THE CLAIM. Nine named origins a "live" assertion can come from, and LIVE_ASSERTING_ORIGINS has exactly one entry. Each of the other eight records WHY it is not evidence, as data, because a rule with no argument attached is one somebody removes the first time it is inconvenient — and origin-with-no-recorded-rejection makes the missing argument a fault. This is S6.11's "ask where the evidence came from" applied to the surface a person actually reads. -->

    • S7.3.a Consume authenticated runtime capability/dependency endpoints with version, checked-at, freshness, affected/retained behavior, owner, and safe remediation. Evidence: network/contract tests. <!-- evidence: probeContractFaults(requested, response, expectedState) + useCapabilityProbe, which parses the response with S1.6's own parseCapabilityState. A hand-rolled body.state === 'configured' accepts {"state":"configured"} with no owner, no freshness and no verifications — the exact shape a stub endpoint returns — so the browser tool serves that body from a real socket and CONTROL 5 asserts the page refuses it. requested is a separate parameter from response.capability because a probe that answers about the wrong capability is indistinguishable from a healthy one, which is what a shared cache key produces. THREE REAL KIT RULES THE FIXTURES HAD TO SATISFY, each caught by running the parser rather than by reading it: a contactPath must be an https rota or runbook URL (an email names a person rather than a responsibility), a configured capability must offer remediation none (the kit refuses "nothing is wrong, and here is how to fix it"), and failed requires a correlationId because a completed evaluation nobody can find the trace of is a guess. -->
    • S7.3.b Represent configured, not-configured, degraded, unavailable, failed, stale-probe, probe-loading, and probe-error independently. Evidence: banner state fixtures. <!-- evidence: BANNER_STATES (8) + PROBE_STATES (3) + bannerState(outcome, now) + bannerStateFaults. THE SPLIT IS THE ITEM: unavailable says the dependency is down, probe-error says we could not ask. Collapsing them is S6.11's fail-open in one direction and a fabricated outage in the other, so probe-state-rendered-as-a-capability-state and its mirror are separate faults. STALENESS IS CHECKED BEFORE THE STATE IS READ — a stale configured is not a configured capability, it is an answer that has stopped being one — and rendering it in a quieter colour would keep the claim while looking careful. Browser CONTROL 3 gets a real 500 and CONTROL 4 a real three-hour-old checkedAt. -->
    • S7.3.c Display fixture source/version and persistent unmistakable fixture labeling in banner and working context. Evidence: fixture UI tests. <!-- evidence: REQUIRED_LABEL_SURFACES + fixtureLabelFaults + FixtureBadge. THE BADGE BELONGS IN THE WORKING CONTEXT, NOT ONLY IN THE BANNER: a banner scrolls away, gets dismissed, and never appears in the screenshot somebody pastes into a ticket, so fixture-labelled-only-in-the-banner is a fault and browser CONTROL 6 counts the badge in both surfaces. The source NAME and VERSION are part of the label — "fixture" alone does not say which one — and the badge is not dismissible (§2.2 says persistent, and a dismissible badge is absent for everybody after a week). fixture-label-rendered-while-live closes the other direction: a badge over live data teaches the reader that the badge means nothing. -->
    • S7.3.d Prevent build variables, client heuristics, successful route load, cached content, or demo data from asserting live state. Evidence: false-live negative tests. <!-- evidence: LIVE_CLAIM_ORIGINS / LIVE_ASSERTING_ORIGINS / ORIGIN_REJECTIONS / liveClaimFaults, with a negative test per origin. The component decides the origin from the REQUEST, not from the answer — a 200 says nothing about whether a credential was sent — which is why browser CONTROL 2 is the sharpest check in this item: the endpoint returns a BYTE-IDENTICAL 200 body whether or not it saw an authorization header, the page still records unauthenticated-probe-response, and the kit still refuses a live claim on it. An endpoint that answered differently would have proved the rule by accident. -->
    • S7.3.e Provide accessible detail and authorized setup/status link without exposing secrets or internal topology. Evidence: role/redaction tests. <!-- evidence: bannerDisclosureFaults, which COMPOSES S1.6's parseSafeDetail for credentials rather than inventing a second check, and adds what that check does not look for: TOPOLOGY. An internal hostname, a private address, a pod name and a bare host:port are not credentials — parseSafeDetail passes all four — and each is a free map of the estate for anybody reading the banner. The setup link is authorized BEFORE it is offered, in the kit and in the component, with a jsdom test proving an author does not see the settings button a tenant admin does; a link that produces a refusal has turned one problem into two and the second looks like a defect in the shell. -->
    • S7.3.f Refresh on dependency/flag/provider change with bounded polling or event update and preserve last-known age honestly. Evidence: transition tests. <!-- evidence: refreshPolicyFaults + MIN_POLL_INTERVAL_SECONDS + the hook's scheduler. Faults: an unbounded poll, one below the floor, one that keeps running in a hidden tab, no backoff after repeated probe errors (client load rises exactly when the dependency can least take it), a poll where every invalidating change already arrives as an event, an invalidating change with nothing listening, an age computed from render time, and a failed refresh that leaves the previous answer on screen as current — the last of which hides precisely the outage the refresh existed to detect. THE HIDDEN-TAB PAUSE IS PROVED IN JSDOM, NOT IN THE BROWSER, and the reason is recorded in the test: headless Chromium reports every page as visible, opening a second tab does not make the first hidden, and Emulation.setVisibilityStateOverride does not exist in this Chromium — all three were tried. The browser proves what it can prove: real requests arriving at a real endpoint at the floored rate. --> <!-- THE FINDING FROM THE REAL FETCH: a now captured when the component MOUNTED is EARLIER than the timestamp of a probe that answers afterwards, so the age is negative, isStateFresh correctly refuses a negative age, and a perfectly fresh answer renders as stale-probe. Same family as an age computed from render time, in the opposite direction and far harder to notice, because the banner looks conservative rather than wrong. The reading now carries its own ticking now, read at comparison time rather than at mount or from the last tick. Nothing in jsdom would have surfaced this: it needs a response that genuinely arrives after the render. --> <!-- S7.3 rollup: 4691 kit tests (57 new) + 12 jsdom banner tests + the real-endpoint browser tool. Verified by patching the BUILT dist twice — adding unauthenticated-probe-response to LIVE_ASSERTING_ORIGINS fails CONTROL 2, and silencing the staleness check in bannerState fails CONTROL 9 — and passing again over the restored build. CONTROL 9 exists BECAUSE of that second patch: the first attempt at it changed nothing the tool read, since every other control routes through the page (which runs the kit compiled from source), so the tool now runs the BUILT kit over the same probe body fetched separately and compares. Export map at 96 entry points, six kit gates clean, eslint and a non-vacuous tsc clean on the new app files. Registration adds the capability-banner-browser target on apps/oshun/web/project.json. -->
  • S7.4 Build a filter/search grammar with URL serialization. Evidence: copy/paste and back/forward restore the exact view. <!-- The evidence clause sounds like a serialization problem and is not. It is a problem about MEANING: a string somebody pastes into a chat window has to mean the same thing when it is opened, by a different person, in a different locale, a week later, against a workbench that has shipped twice since. Every part of query-grammar.ts (97th entry point, 49 tests) exists because one of those five words breaks something — the canonical form for "the same thing", the locale-neutral identifiers for "a different locale", the preserved relative date for "a week later", the grammar version for "shipped twice since", and the same-error-for-every-hidden-field rule for "a different person". All seven children complete. --> <!-- A QUERY IS A TREE AND THE STRING IS A RENDERING OF IT. The URL, the migration and the editor all operate on the tree, so there is one place where meaning lives, and the contract is that rendering a parsed tree and parsing it again yields the same tree — run over the whole corpus, not over the entries that happened to be convenient. -->

    • S7.4.a Define typed field, operator, value, grouping, negation, range, relative date, taxonomy, rights, quality, ownership, ID, and free-text grammar. Evidence: grammar specification. <!-- evidence: libs/oshun/workbench-kit/src/query-grammar.ts — twelve FIELD_KINDS, twelve OPERATORS, and OPERATORS_FOR_KIND as a CLOSED TABLE keyed by kind rather than a per-field list, because which comparisons make sense is a property of the TYPE. > over a taxonomy term is not a stricter filter, it is a question with no answer, and allowing it returns an empty result set whose emptiness looks like a fact about the data — so the six unordered kinds admit no gt and no between, asserted directly. Values cover literal, number-with-unit, absolute date, RELATIVE date, boolean, list and range; grouping is parenthesised; negation is NOT or a leading hyphen; bare words are free text. -->
    • S7.4.b Implement tokenizer/parser with positions and stable syntax errors for quotes, escapes, Unicode, nesting, ambiguity, and limits. Evidence: parser corpus. <!-- evidence: tokenize + parseQuery + 14 SYNTAX_ERRORS, each with a positive test, over a 27-entry corpus of queries somebody would actually type. EVERY NODE CARRIES A HALF-OPEN SOURCE RANGE in UTF-16 code units, which is what a text input's selection API uses — an error without a position is one nobody can fix, and an editor that finds it by re-scanning has a second parser that will disagree with the first. Unicode is not special-cased anywhere: an Arabic tag, Cyrillic text and an emoji are ordinary term content. PRECEDENCE IS WRITTEN DOWN AND TESTED — NOT > AND > OR, juxtaposition is AND — because every search box implements some version of this and almost none document it, which is how a OR b AND c comes to mean two different things in two surfaces. THE TOKENIZER BUG THE CORPUS CAUGHT: marking a token as quoted because a quote appeared ANYWHERE in it turned title~"two words" into a free-text search, which is the failure that looks like the filter simply not working. leadingQuote and valueQuoted are now separate. -->
    • S7.4.c Validate fields/operators/values against domain capability, contract version, authorization, locale-neutral units, and complexity budget. Evidence: semantic validation tests. <!-- evidence: validateQuery + 8 SEMANTIC_ERRORS. THE ONE THAT WOULD BE A SECURITY DEFECT RATHER THAN A USABILITY ONE: a field that does not exist, one belonging to a capability this workbench cannot use, and one the actor may not filter on all produce unknown-field with the SAME words — asserted character for character in the spec. If they differed, the search box would be an oracle for the schema and anybody could enumerate it one word at a time. This is S4's conceal-versus-deny split inside a parser. Also: an operator the kind does not admit, a value of the wrong kind, a value outside the declared set, a field that arrived in a newer contract, AMBIGUOUS_UNITS (a duration written 3mo is three months to one reader and three minutes to another), a range whose bounds are the wrong way round, and a COMPLEXITY BUDGET spent while the query is still a tree — refusing it there is cheaper than discovering it as a timeout after the work is done. -->
    • S7.4.d Serialize a canonical normalized query deterministically without changing meaning or localizing identifiers/operators. Evidence: parse/serialize property tests. <!-- evidence: serializeQuery + queriesEqual, with the round-trip property run over EVERY corpus entry rather than a chosen few — a canonical form that survives the examples somebody picked is a canonical form for those examples. Also asserted: serializing twice changes nothing; parentheses appear only where precedence needs them; a relative date stays relative (created:>now-7d never becomes a date); equivalent relative spellings (now-168h, now-7d, now-1w) collapse to ONE form; and no identifier or operator is ever non-ASCII. WHAT IT DELIBERATELY DOES NOT DO IS REORDER: a AND b and b AND a mean the same thing and serialize differently, because a canonical form that shuffles somebody's filters is one they stop trusting the first time they paste a link. Determinism means the same tree always renders the same string, not that equivalent trees do. -->
    • S7.4.e Encode safe shareable query, sort, grouping, view, and pagination anchor in URL while excluding protected/local-only state. Evidence: URL leakage tests. <!-- evidence: encodeViewState / decodeViewState / urlFaults, with VIEW_PARAMS as the allow-list and LOCAL_ONLY_KEYS as the explicit refusal — selection, scroll, draft, token, impersonation, session id and cursor secret — because "do not put secrets in the URL" is a habit and habits do not fail builds. protected-value-in-the-url catches the subtler one: a filter on a protected field discloses the value even when the results do not, since the FILTER is in the link. anchor-is-an-offset refuses a row number as a pagination anchor — anything inserted before it shifts the page, so the link points at different rows every time it is opened — and url-longer-than-the-safe- limit names the point where proxies and mail clients truncate, after which a truncated query parses into a different one. -->
    • S7.4.f Restore exact valid view on paste/direct link/refresh/back/forward and provide migration/fallback for older or invalid grammar versions. Evidence: navigation tests. <!-- evidence: apps/oshun/web/tools/query-view-browser.mjs — 6 controls, 3 faults, real Chromium. The evidence clause names three things jsdom does not have: an address bar, a history stack and a keyboard, so a test that calls decodeViewState and compares trees proves the codec and nothing about the sentence. CONTROL 3 opens the shared link in a SECOND real page and checks the canonical query and the editor contents; CONTROL 4 presses real Back, real Forward and a real reload. A GRAMMAR VERSION WE DO NOT KNOW IS REFUSED, NOT READ PARTIALLY (S6.14's rule, one layer up), and the fault run caught the host doing exactly that: it was echoing the refused q back into the editor, which would re-parse a foreign grammar with this one. A query that failed to PARSE is handed back — it is in a grammar we understand and the reader can fix it — and a query in a version we do not know is not. -->
    • S7.4.g Provide keyboard editor, suggestions, accessible token/error semantics, plain form alternative, and no search-on-every-keystroke abuse. Evidence: accessibility/performance tests. <!-- evidence: apps/oshun/web/src/components/workbench/QueryEditor.tsx — a combobox with aria-activedescendant, a listbox of field suggestions accepted from the keyboard, aria-invalid plus a polite error region that names the character RANGE (from the parser, never from a second scan), and a <details> plain form that reaches the same canonical query without the grammar, for anybody who does not know it and anybody driving the page by voice. NO SEARCH ON EVERY KEYSTROKE IS A NUMBER, NOT A CLAIM: browser CONTROL 1 types 29 real characters and counts 1 search. A box wired straight to a request sends one per character and the answers arrive out of order, so the list settles on whichever prefix the network happened to return last. CONTROL 5 confirms the underlined range is the quote (21-34) and not the whole string, by comparing it to the BUILT kit's own range for the same input. --> <!-- S7.4 rollup: 4741 kit tests (50 new) + the browser tool. Verified by patching the BUILT dist twice — widening the unterminated-quote range to the whole query fails CONTROL 5, and making queriesEqual always disagree fails CONTROL 2 — and passing again over the restored build. THE FIRST ATTEMPT AT A PATCH FAILED TO FAIL: weakening SUPPORTED_GRAMMAR_VERSIONS changed nothing, because the page runs the kit compiled from source and only the symbols the HARNESS imports are under test. Same lesson as S7.3, and the reason both tools cross-check against the built kit rather than only driving the page. Export map at 97 entry points, six kit gates, eslint and a non-vacuous tsc clean. Registration adds the query-view-browser target. -->
  • S7.5 Build accessible data tables with column control, density, pinning, resizing, sort, selection, and virtualization. Evidence: keyboard and screen-reader tests. <!-- A table is the surface people spend the most time in and the one with the most ways to be quietly wrong, and almost none of them are visual: a sort that disagrees with the server, a selection that means something different from what it says, a total that counts the rendered window. Each looks correct on screen and produces the wrong answer downstream. data-tables.ts (98th entry point, 52 tests) names them; apps/oshun/web/src/components/workbench/DataTable.tsx renders a real one over 12,431 rows. All eight children complete. --> <!-- The browser proof earned its place three times over. Reading the source would not have shown that the touch density left its controls at 24px, that the pinned column shared a sticky offset with the selection column and slid under it, or that a width constant disagreed with the rendered border box by 16px. All three are invisible in a screenshot and all three are one horizontal scroll away from being obvious. -->

    • S7.5.a Define semantic table/grid pattern by interaction need, accessible name/description, captions, headers, relationships, row identity, and loading/empty/error state. Evidence: base table tests. <!-- evidence: libs/oshun/workbench-kit/src/data-tables.ts — patternForInteraction DERIVES the pattern rather than letting a surface pick one: grid exists because cells take focus, and a table whose cells do not take focus gains nothing from it but the obligation. grid-without-arrow-key-navigation is the fault that matters — the role tells assistive technology to switch to its own navigation mode, and without arrow-key cell movement the keys then do nothing, which is worse than never having claimed it. Also: an unnamed table (a page with two announces "table" twice), one with no caption or description (a nonvisual reader has the columns and no idea what filtered them), rows addressed by position rather than identity, and each of loading, empty and error unannounced, reported separately. -->
    • S7.5.b Implement visible/hidden/ordered column control with required- column rules, reset, persistence, and authorization-aware fields. Evidence: column control tests. <!-- evidence: columnControlFaults. A COLUMN THE ACTOR MAY NOT SEE IS NOT A HIDDEN COLUMN — hidden means "you turned it off", unauthorized means it is not yours to turn on, and if they share a control the column picker becomes a list of the fields somebody is not allowed to read. Both directions are faults: offered in the picker, and actually rendered. A hidden-but-permitted column stays in the ORDER so turning it back on lands it where the actor left it, and dropping it is hidden-column-forgotten-instead-of-remembered. Required columns cannot be hidden, the order may not name a column that does not exist, and a table somebody has rearranged into uselessness has a reset. -->
    • S7.5.c Implement comfortable/compact/touch density without clipping, inaccessible targets, or page overflow. Evidence: density/zoom tests. <!-- evidence: MIN_ROW_HEIGHT_PX + MIN_TARGET_PX + densityFaults, and browser CONTROL 4 MEASURES all three densities with getBoundingClientRect rather than reading the stylesheet. THAT MEASUREMENT FOUND A REAL DEFECT: the touch density enlarged its rows and left its checkbox at 24px, so the one thing a touch density exists for was the one thing it did not do. The control size is now derived from MIN_TARGET_PX[density]. The floors are WCAG 2.2's 24x24 minimum, with touch held to 44. Page overflow is checked as a document-level fact: the table scrolls inside its own region or it is a defect. -->
    • S7.5.d Implement keyboard and numeric resizing, min/max/reset, persisted width, overflow disclosure, and screen-reader feedback. Evidence: resize tests. <!-- evidence: resizeFaults + the separator handle on every column header, with aria-valuenow/min/max. Keyboard resizing is not an alternative to the pointer one, it is the ONLY one a keyboard user has, so resize-is-pointer-only is a fault rather than an enhancement. Browser CONTROL 6 presses real ArrowRight, then asserts the RENDERED column width matches aria-valuenow to within a pixel — a handle that reports a number the layout does not agree with is a control that lies to exactly the reader who cannot see the result. End and Home clamp at the declared bounds rather than running away. -->
    • S7.5.e Implement left/right pinning, pinned boundaries, keyboard navigation, reorder interaction, and narrow-view fallback. Evidence: pinning tests. <!-- evidence: pinningFaults + position: sticky on header and body cells. THE BROWSER FOUND TWO BUGS HERE THAT READING COULD NOT. First, the start offsets were computed from the data columns alone, so the first pinned column claimed inset-inline-start: 0 — the offset the selection column already had — and slid underneath it on a real horizontal scroll. Then the constant said 40px and the cell rendered at 56px, because it was not border-box, so it stuck 16px from where it started. Both look correct in a screenshot; CONTROL 5 scrolls for real and asserts the pinned column moved by at most one pixel WHILE an unpinned one moved 285. FAULT 2 removes the sticky positioning and confirms the same column then travels 100+px, so the control is not passing by accident. The kit also refuses a pinned region past half the viewport, a boundary the keyboard cannot cross, and a pinned column that is not contiguous with its edge — on both sides. -->
    • S7.5.f Implement single/multi stable sort with direction/null/locale/ tie-break semantics and server parity. Evidence: sort tests. <!-- evidence: sortFaults, and this is the child where a table can be perfectly usable, perfectly accessible, and still produce an order in which one row appears on two pages. FOUR RULES ABOUT AGREEING WITH THE SERVER. A client-side sort over a server-PAGINATED list orders the rows that happen to be loaded, not the list — the first row of page two can belong before the last row of page one — and the counterexample test confirms the same client sort is fine when the server is not paginating. Nulls have a position and it must be the server's, or the two pages are ordered by different rules. Collation is part of the contract: an undeclared one is the runtime default and changes with the reader's locale, and one that disagrees with the database means sorting is not a total order. A sort with no unique tie-break is not stable across pages, and a tie-break that is not the LAST key decides comparisons it was not meant to. -->
    • S7.5.g Implement single/range/additive/select-all-filtered/offscreen selection with eligible/hidden/permission-changing behavior. Evidence: selection tests. <!-- evidence: SelectionSet + rowSelectionFaults. "SELECT ALL FILTERED" IS A QUERY, NOT A LIST OF IDS — ten thousand ids do not fit in a request and the set moves while somebody looks at it — so it is a frozen query plus the count observed at freeze, and anything downstream reports the difference rather than pretending there was none. A count taken from the DOM is a lie virtualization tells, so count-derived-from-the-rendered-rows is a fault on its own, and browser CONTROL 3 selects a row, scrolls 300,000px until that row has left the DOM ENTIRELY, and asserts the count is unchanged. Ineligible rows counted as selected make the preview promise more than the result delivers; a selection hidden by the current filter is the surprise this exists to prevent. -->
    • S7.5.h Implement virtualization with total/position semantics, stable focus/selection/scroll anchor, dynamic rows, and semantic nonvirtual fallback for print/export. Evidence: large dataset a11y tests. <!-- evidence: virtualizationFaults + DataTable.tsx, proved over 12,431 rows because a virtualized table over twenty rows proves nothing about virtualization. aria-rowcount IS THE TOTAL and aria-rowindex is ABSOLUTE: browser CONTROL 2 scrolls deep and confirms the first rendered row announces a position in the LIST (row 4,800-odd of 12,431) rather than in the window, which is the difference between describing the data and describing the implementation. PRINT DROPS THE WINDOW — page.emulateMedia({media:'print'}) is a real print rendering, and CONTROL 7 asserts 18 rows on screen become 400 in print and that the pinned column stops being sticky so it is not printed over the rest. Without the beforeprint/matchMedia listener the print stylesheet alone cannot bring back rows React never rendered. --> <!-- S7.5 rollup: 4794 kit tests (52 new) + apps/oshun/web/tools/data-table-browser.mjs (7 CONTROLs, 2 FAULTs) over 12,431 rows in real Chromium. THE BROWSER FOUND THREE DEFECTS THE SOURCE DID NOT SHOW: the touch density leaving its controls at 24px, the pinned column sharing an offset with the selection column, and a width constant that disagreed with the rendered box. Verified by patching the BUILT dist twice — silencing the aria-rowcount check fails FAULT 1, and raising the comfortable row floor to 120px fails a measured CONTROL 4 assertion — and passing again over the restored build. Export map at 98 entry points, six kit gates, eslint and a non-vacuous tsc clean. Registration adds the data-table-browser target. -->
  • S7.6 Build kanban or stage-board primitives only where stage movement is the core interaction. Evidence: non-drag move controls and transition refusal UI. <!-- A board is the most requested workbench surface and the least often earned. It is a good surface for exactly one job — moving objects between stages — and a bad one for everything else it gets used for, because the thing it is best at is a column per stage, which looks like insight and is only ever a GROUP BY. So the item's own first child is a GATE, and this is the first S7 parent where the leading question is whether to build the thing at all. stage-boards.ts (99th entry point, 88 tests) answers it from MEASURED movement over the S2.1 stage graph rather than from a request, and apps/oshun/web/src/components/workbench/StageBoard.tsx renders one. All seven children complete. --> <!-- The claims here are facts about an interaction OVER TIME, which is why the proof is apps/oshun/web/tools/stage-board-browser.mjs (12 CONTROLs, 3 FAULTs) rather than a jsdom render: whether a card is still in its source column WHILE a request is in flight, whether a real HTML5 drag lands on a column the server did not declare, whether the keyboard path produces a byte-identical board to the pointer path, where focus is afterwards, and what a genuinely offline browser does with a move. Verified by patching the BUILT dist twice — flipping conflict.retryAllowed to true, and raising the CONFIRMATION_BANDS floors — and confirming CONTROL 7 and CONTROL 8 fail, then pass again over the restored build. -->

    • S7.6.a Require a domain applicability decision proving stage movement, not monitoring or decorative status, is the primary object interaction. Evidence: board registration rationale. <!-- evidence: boardApplicabilityFaults(decision, graph) in libs/oshun/workbench-kit/src/stage-boards.ts. It takes the rationale AND the S2.1 stage graph SEPARATELY, because they can disagree and the disagreement is the finding — S6.6's photocopy rule applied to a UI registration, so a claim cannot agree with itself. THE SHARPEST RULE IS THE GRAPH ONE: a registration that says people move objects, over a graph in which every Transition.trigger is system/schedule/ upstream-event, is a request for drag handles on a report (no-actor-triggered-transition), and a graph where fewer than half the non-terminal stages have an actor-triggered way out is a set of places objects arrive at (most-stages-cannot-be-moved-from). The counts are refused unless they are measured-from-the-audit-log, because the belief that people work the board is exactly the belief a board creates. Floors, each with its reason: 14 days, 50 objects, MIN_ACTOR_MOVES_PER_OBJECT = 2 (one move is a lifecycle and the surface for a lifecycle is a status column), and half the sessions moving something (below that people come to look). MAX_BOARD_STAGES is DERIVED — BOARD_REFERENCE_VIEWPORT_PX / BOARD_COLUMN_MIN_WIDTH_PX = 1440/220 = 6 — and browser CONTROL 2 MEASURES the derivation: six columns fit and a seventh overflows the track. -->
    • S7.6.b Render typed stages, allowed/terminal state, counts, WIP/capacity, ordered items, freshness, and authorization from server contracts. Evidence: stage board fixtures. <!-- evidence: boardColumnFaults(columns, graph, now), 19 faults, and browser CONTROL 1 feeds it the columns read straight out of the DOM. A COLUMN COUNT MUST BE THE STAGE'S TOTAL, NOT THE ITEMS LOADED INTO IT — S7.5's aria-rowcount defect one level up and worse, because the count is the number people plan with, so a column saying 12 while holding the first 12 of 900 is the difference between a WIP limit and a WIP suggestion. A STAGE THE ACTOR MAY NOT SEE IS NOT AN EMPTY STAGE: an empty column is a claim about the data and the truth is a claim about the reader, so deny renders "not yours to read" and conceal renders NOTHING (browser CONTROL 12 drives both). Terminality is READ from the graph, never restated by the column, and a column that disagrees is itself the finding; dead ends come from deadEnds(graph). The truncation note fires in BOTH directions, because a note that is always there is furniture and stops working on the column where it is true. The clock is read at COMPARISON time (S7.3): an observation timestamped after now is observation-is-in-the-future, not a fresh column. -->
    • S7.6.c Implement drag preview/drop only for server-declared candidate transitions and preserve source until accepted. Evidence: drag tests. <!-- evidence: dragFaults(interaction, now), 15 faults, and browser CONTROLs 3 and 4 with a REAL HTML5 drag. THE SOURCE COLUMN KEEPS THE ITEM UNTIL THE SERVER ACCEPTS: the component writes an item's stage in exactly one place, the branch handling an accepted result, so a refusal needs no compensation — the board never claimed the move happened. CONTROL 3 proves it with a 400ms server: mid-flight the card still reports data-stage="packing", data-board-state is unchanged, and it is marked pending; then the accept moves it. source-removed-at: drag-start is the defect that looks fine in every screenshot and is indistinguishable, when the refusal lands, from somebody else moving the object back. DROP IS PERMITTED EXACTLY WHERE THE SERVER SAID: CONTROL 4 holds a drag open and reads the columns that opted in — exactly the three the server declared — then drops on a fourth and asserts the board is byte-identical afterwards. candidates-recomputed-on-the-client is a second implementation of the transition graph that drifts on the first guard nobody mirrored, and a candidate list is perishable (CANDIDATE_FRESHNESS_SECONDS = 60). -->
    • S7.6.d Provide select-plus-move command, target list, keyboard shortcuts, and accessible move dialog as fully equivalent non-drag paths. Evidence: keyboard/screen-reader tests. <!-- evidence: nonDragParityFaults(reaches, dialog, shortcuts) and browser CONTROL 5. PARITY IS CLAIMED PER OPERATION, NOT PER SURFACE: BOARD_OPERATIONS names all seven, and "the keyboard can move a card" is true of a board where the keyboard cannot choose the position, cannot move a selection and cannot read the refusal. The equivalence is STRUCTURAL rather than a pair of paths somebody keeps in sync — the drop handler and the keyboard command call one performMove, and CONTROL 5 makes the same move both ways and compares data-board-state character for character. The target list applies S4's conceal/deny split: a target refused for a reason the actor may know MUST be listed WITH the reason (omitting it sends them back to try the drag again) and a concealed one must not be listed at all, because a greyed row carrying its name is the disclosure. Shortcuts compose S7.1's RESERVED_CHORDS; the one-key binding lives on the card rather than on the document, so CONTROL 10 types "order 42m0" into the filter and no dialog opens. -->
    • S7.6.e Preview guard/evidence/approval/impact/cost requirements and require confirmation proportionate to transition risk. Evidence: transition preview tests. <!-- evidence: transitionRiskScore + RISK_POINTS + CONFIRMATION_BANDS + requiredConfirmation + transitionPreviewFaults (16 faults). CONFIRMATION IS PROPORTIONATE IN BOTH DIRECTIONS. Under-confirming an irreversible publish is obvious; OVER-confirming a routine move is the same defect with a longer fuse, because a board that asks "are you sure" for every drag trains everybody to dismiss the one that mattered. The score is a published table rather than a chain of ifs, so a reviewer argues with a number, and blast radius is logarithmic (blastRadiusPoints) so a large routine bulk move does not outrank an irreversible single one. Guards compose S2.2's NON_VERDICT_STATUSES: a guard that could not run rendering as a tick is S6.11's fail-open in a preview. Browser CONTROL 8 runs requiredConfirmation from the BUILT dist over the same risk the page rendered and compares — a routine move asks for nothing, a dispatch demands the strength the kit computes, the button stays disabled until the challenge is met, a refused challenge leaves the object where it was, and a workbench with NO configured challenge says so and refuses rather than downgrading to the confirm button it does have. -->
    • S7.6.f Handle accepted, conflict, refused-with-reasons, stale evidence, unauthorized, partial bulk, offline, and failed transitions without optimistic false state. Evidence: result matrix. <!-- evidence: MOVE_OUTCOME_RULES (the matrix as data) + moveResultFaults + moveMatrixFaults, and browser CONTROL 7 drives SEVEN of the eight through the real component and CONTROL 9 drives the eighth with context.setOffline(true). THE EIGHT ARE NOT FOUR PAIRS. conflict is a successful move by somebody else, so offering a retry offers to overwrite it and the honest offer is the re-read; unauthorized requires a re-read even though nothing about the object changed, because the move was OFFERED, which means either the grants changed or the candidate list was wrong. The component reads retryAllowed/refreshRequired/announcement FROM the rule, so patching conflict.retryAllowed in the BUILT dist makes CONTROL 7 fail (verified). Politeness is checked in both directions. offline must not become a background intention: CONTROL 9 asserts the board is unchanged, a retry is offered because the request never left, and RECONNECTING DOES NOT SILENTLY DELIVER THE MOVE. moveMatrixFaults is separate because a board can render every outcome it has ever SEEN perfectly and fall through to the generic failure toast for the two it has not. -->
    • S7.6.g Preserve focus/selection after move and expose semantic table/list parity for every board fact and operation. Evidence: nonvisual journey. <!-- evidence: focusPreservationFaults + boardParityFaults, browser CONTROL 6 and CONTROL 11, and apps/oshun/web/src/components/__tests__/StageBoard.spec.tsx. FOCUS IS RESTORED BY IDENTITY, NEVER BY INDEX — after a move the object at that position is a different one, so the next keystroke acts on something nobody chose. CONTROL 6 reads document.activeElement for real: after an accept it is the same card in its new column, after a refusal it is the refusal message (which is where the object is NOT), and never the body. THE REAL BROWSER FOUND A BUG IN THE RULE ITSELF: the first refusal it drove had no selection at all, and refused-item-dropped-from-the-selection fired over an empty set — a rule asserting about a set nobody had built. Fixed and covered by a counterexample test. Parity is a table over BOARD_FACTS: a fact conveyed only by WHICH COLUMN a card sits in is a fact a list does not have, so CONTROL 11 asserts the stage is a VALUE in the row, that all eight other facts are columns in the list, and that a move can be started from it — an equivalent view that cannot act is a read-only copy. --> <!-- S7.6 rollup: 4883 kit tests (88 new, 99th entry point) + 12 jsdom tests + apps/oshun/web/tools/stage-board-browser.mjs (12 CONTROLs, 3 FAULTs) in real Chromium. THE BROWSER EARNED ITS PLACE ON THE FIRST RUN: refused-item-dropped-from-the-selection fired over a move made with no selection. packed-consumer-smoke earned its place too — Math.floor at module scope is work at import time that "sideEffects": false tells a bundler it may drop, so MAX_BOARD_STAGES subtracts the remainder before dividing instead. Six kit gates, eslint and a non-vacuous tsc (proved by injecting a type error) clean. Registration adds the stage-board-browser target. -->
  • S7.7 Build a stable inspector router. Evidence: selection persists across refresh and invalid/deleted selections recover safely. <!-- An inspector looks like a detail view and behaves like a second application: it has its own routing, its own loading, its own authorization, its own history and its own focus, and every one of those is somewhere a workbench quietly loses what somebody was looking at. inspector-routes.ts (100th entry point, 54 tests) names the rules and apps/oshun/web/src/components/workbench/InspectorRouter.tsx obeys them structurally rather than by checking: ONE function builds the URL and it walks the kit's five-key allow-list, so a draft or a token has no path into the address bar; ONE branch produces the concealed refusal, so the three answers that must be identical cannot drift apart. All seven children complete. --> <!-- Half of this item is REQUESTS and half is NAVIGATION, and jsdom has neither, so the proof is apps/oshun/web/tools/inspector-router-browser.mjs (10 CONTROLs, 3 FAULTs) against a real node:http server: page.reload() throws the tree away for real, page.goBack() is a real popstate, and page.on('request') counts what left the browser rather than what a mock was asked for. THE BROWSER FOUND TWO REAL DEFECTS — a malformed id was left in the address bar forever because the early return skipped the clearing rule, and a concealed refusal erased its own explanation one render after clearing the selection that produced it. Verified by patching the BUILT dist twice: RECOVERY_RULES.filtered.clearsSelection and dropping subview from ROUTE_STATE_KEYS each fail a MEASURED control, and both pass again over the restored build. -->

    • S7.7.a Register inspector routes by typed object/subobject/anchor, capability, authorized data loader, component, and compatibility version. Evidence: inspector registry tests. <!-- evidence: inspectorRegistryFaults(routes, anchorKindsByObject) in libs/oshun/workbench-kit/src/inspector-routes.ts. The anchorable parts of an object arrive SEPARATELY from the routes, so a route anchoring at a block in an object that has no blocks is a finding rather than a self-consistent pair — S6.6's photocopy rule. The sharpest rules are about the LOADER: one that is registered as not applying authorization relies on its caller having checked, and A DEEP LINK HAS NO CALLER; and one loader serving several object kinds means one authorization rule covers all of them, so the narrowest kind sets the rule for the widest. Two routes resolving the same object and subview is a finding because which one a URL opens then depends on registration order, so the same link shows different people different panels. The compatibility window is checked in both directions: inverted, and a window one major wide, which makes every link shared before the last payload change unopenable rather than degraded. -->
    • S7.7.b Encode only safe workbench/object/subview/anchor/compare state in route and keep drafts/secrets/transient credentials local. Evidence: URL state tests. <!-- evidence: ROUTE_STATE_KEYS + inspectorUrlFaults, and browser CONTROL 1. THE URL IS A COPY — pasted into chat, kept in a bookmark, logged by every proxy, sent as a Referer, read by whoever finds it — so the route may carry WHICH object is being looked at and nothing about what is in it. The vocabulary is an ALLOW-LIST, because the failure mode is always one more key that seemed harmless and a deny-list only catches the ones somebody already thought of. permission-decision-in-the-url is the one worth naming twice: ?canEdit=1 reads as state and behaves as an authorization anybody can type. Sensitivity composes S3's PRESENCE_DISCLOSING_CLASSES — for protected-source the KEY BEING PRESENT is the disclosure, whatever the value says. CONTROL 1 loads a link carrying token and rogue, and the address bar afterwards holds neither; a draft typed into the panel survives a subview change and does NOT survive a reload, which is what local-only means. -->
    • S7.7.c Load by stable identity with tenant/object/property authorization and handle loading/partial/stale/conflict/degraded state. Evidence: loader tests. <!-- evidence: inspectorLoaderFaults(load, now), 12 faults. IDENTITY IS STABLE OR THE LINK ROTS: a slug from a title breaks every shared link the first time somebody fixes a typo, and an index into a list makes the link open a DIFFERENT object without failing, which is worse. AUTHORIZATION IS THREE QUESTIONS — tenant, object, and property — and the third is the leak nobody sees, because a payload carrying fields the actor may not read looks exactly right on screen; the only thing stopping the disclosure is that the panel does not draw them. A check after the read is a finding at both levels: the data has already crossed the boundary and how long the refusal takes answers whether the id exists. partial-rendered-as-complete turns an outage into an edit — the section that failed to load renders as a section somebody emptied. Browser CONTROL 10 drives a partial and a stale load through the real server and reads the notices back. -->
    • S7.7.d Preserve inspector selection, tab, scroll, and focus on refresh, back/forward, workspace navigation, and compatible revision update. Evidence: navigation tests. <!-- evidence: PRESERVATION_RULES + preservationFaults(before, after) + HISTORY_RULES + historyDispositionFaults, and browser CONTROLs 2 and 3. Preservation is only observable ACROSS a transition, so the finder takes a before AND an after (S7.1's lesson): one snapshot of a panel that threw its selection away looks the same as one that never had it. THE ROW WORTH ARGUING ABOUT IS THE INCOMPATIBLE REVISION UPDATE, which keeps the object and must NOT keep the subview — a panel holding the name of a section it can no longer render is a heading over nothing, which is worse than moving. scroll-restored-before-the-content is the silent one: applied before the content that gives the panel its height, the position clamps to zero and the restoration reports success. CONTROL 2 scrolls to 420, reloads for real, and reads 420 back. CONTROL 3 measures history.length: one object selection adds one entry, three tab changes add none, Back returns to the previous SELECTION, and the dispositions handed to the kit are DERIVED from those measurements rather than restated. -->
    • S7.7.e Recover from invalid, malformed, unsupported, moved, deleted, filtered, archived, tombstoned, permission-hidden, and cross-tenant selection. Evidence: recovery matrix. <!-- evidence: RECOVERY_RULES (ten problems as data) + recoveryFaults + indistinguishabilityFaults + recoveryMatrixFaults, and browser CONTROLs 4, 5 and 8. THE TEN ARE NOT SEVERITIES, and "clear the selection" is the WRONG recovery for most of them: a filtered object exists and the current view hides it, so clearing loses what somebody was working on to a control they could have changed; an archived one is readable and not editable; a tombstoned one has the RECORD OF ITS REMOVAL as its content; a moved one is FOLLOWED, with permanent-redirect semantics and a replace rather than a push, because a location that no longer resolves is not a place to go Back to. DELETED, PERMISSION-HIDDEN AND CROSS-TENANT ANSWER IDENTICALLY, and CONTROL 5 proves it twice: the three HTTP responses are asserted byte-identical (status, content-length, body) AND the three rendered panels are compared character for character, because either alone leaves the other free to be an oracle. CONTROL 8 types Robert; DROP TABLE stories into the route and asserts the server received NOTHING — an id that is not an id can only produce a log line and an error rate somebody investigates. -->
    • S7.7.f Restore focus to invoker/selected object/fallback after close or invalidation and announce the change. Evidence: focus tests. <!-- evidence: inspectorFocusFaults and browser CONTROL 9. The panel is removed from the document, so focus() on a node inside it silently moves focus to the body and looks in every log like a successful restoration — the component checks isConnected before calling it and falls back when the invoker is gone. focus-returned-to-an-invoker-that- is-gone is the case the declared fallback exists for: the row that opened the panel is exactly the row a deletion removes. announcement-does-not-name-the-object is the other half — "panel closed" arriving after a list has re-rendered says nothing about which of several panels or which object. CONTROL 9 reads document.activeElement after a real close and finds the invoking row, not the body. -->
    • S7.7.g Prevent inspector routes from mutating state on resolution and avoid hidden N+1/unbounded fetch behavior. Evidence: network and no-mutation assertions. <!-- evidence: inspectorResolutionFaults + PERMITTED_RESOLUTION_WRITES, and browser CONTROLs 6 and 7 counting REAL requests. RESOLVING A ROUTE MAY NOT WRITE: if it does, the Back button is a mutation, a link preview is a mutation and a crawler is an editor. Exactly one write is permitted and it is named — an append-only read audit, which is not the object, does not change what any other reader sees and is not covered by the object's entity tag; a "last viewed" stamp is all three. CONTROL 6 drives a load, a tab change, a reload, a selection and a Back and asserts ZERO non-GET requests, at the browser AND at the server. CONTROL 7 renders the same panel over 5 items and over 500 and asserts the request count is IDENTICAL — the shape is the finding, not the figure, which is why the ceiling is a baseline plus MAX_REQUESTS_OVER_BASELINE rather than zero. --> <!-- S7.7 rollup: 4938 kit tests (54 new, 100th entry point) + 11 jsdom tests + apps/oshun/web/tools/inspector-router-browser.mjs (10 CONTROLs, 3 FAULTs) against a real node:http server in real Chromium. The browser found two defects reading the source did not show. A dist patch is only a verification when it changes a MEASURED verdict: WIDENING ROUTE_STATE_KEYS proved nothing and NARROWING it failed the control, which is the same asymmetry S7.5 found with a floor. Six kit gates, eslint and a non-vacuous tsc clean. Registration adds the inspector-router-browser target. -->
  • S7.8 Build structured field and block-editor primitives with validation, dirty state, autosave, undo/redo, and revision metadata. Evidence: conflict journey. <!-- This is the surface where a piece of work lives between the moment somebody type it and the moment it is somewhere durable, so every defect in it costs them something they wrote — and almost none of them look like defects: a save indicator a quarter of a second early, a dirty flag that counts the server trimming a space, an undo that reaches back past somebody else. block-editors.ts (101st entry point, 62 tests) names them and apps/oshun/web/src/components/workbench/BlockEditor.tsx makes four of the rules STRUCTURAL rather than checked: dirty is serialize(local) !== serialize(base) with no counter anywhere, "saved" is written in the one branch that handles a 200, the queue is in localStorage, and undo pushes a new operation against a stack fenced at every remote edit. All eight children complete. --> <!-- The evidence asked for is a CONFLICT JOURNEY, which is a sequence of things happening to one document over time, so the proof is apps/oshun/web/tools/block-editor-browser.mjs (10 CONTROLs, 3 FAULTs) against a real node:http server holding a real revision. THE BROWSER FOUND A REAL DEFECT: the undo entry captured its selection inside the CHANGE handler, by which time the browser had already replaced the selected text, so undo restored a caret at 9 where the actor had a range at 8..13. Reading it in beforeinput is the fix. Verified by patching the BUILT dist twice — dropping local-storage from DURABLE_STORES, and adding a revision fact the banner does not render — each failing a measured control, and both passing again over the restored build. -->

    • S7.8.a Register versioned typed field/block schema, stable identity, editor renderer, semantic renderer, validation, diff, migration, anchor, and serialization. Evidence: block registry tests. <!-- evidence: blockRegistryFaults in libs/oshun/workbench-kit/src/block-editors.ts. A BLOCK WITHOUT A SEMANTIC RENDERER IS CONTENT THAT ONLY EXISTS IN THE EDITOR — export, print, search, the feed and the API all read the semantic form, and a block that has only an editor form is invisible to every one of them. IDENTITY IS NOT DERIVED FROM CONTENT: a block id computed from its text changes when somebody fixes a typo, and every anchor, comment, reference and deep link pointing at it dies silently. The migration check WALKS the chain from oldestStoredVersion rather than counting migrations, because the failure is always a gap — two migrations look like plenty and neither starts at the version a stored document is saved at. Also: one renderer serving both views, an undeclared round trip, anchors the serializer drops, and a validator shared between two block types (the rules for the strictest apply to the loosest and the codes name the wrong type). -->
    • S7.8.b Build type-appropriate controls with labels/help, units/formats, required/nullable/default, read-only/unauthorized, and accessible semantics. Evidence: field fixture matrix. <!-- evidence: CONTROL_FOR_KIND + fieldControlFaults, and browser CONTROL 1 reads the rendered control back for every kind. The control is DERIVED from the kind, and a date in a free-text box is the most expensive instance: it parses differently in two locales and what comes back is a different day rather than an error. AN UNAUTHORIZED FIELD IS NOT A READ-ONLY FIELD — read-only shows the value and refuses the edit, unauthorized must not show it — and the component renders a DIFFERENT ELEMENT rather than a disabled control, so CONTROL 1 asserts the value appears nowhere in the document at all. A number states its unit (millilitres and litres look identical in a text box), a date states its format, required and nullable together is a contradiction one side will enforce, a default that fails its own validation makes a new block invalid before anybody touches it, and the validation message is ASSOCIATED with the control rather than merely drawn near it. -->
    • S7.8.c Implement add/remove/move/reorder/nest/split/merge/reference operations through domain commands rather than arbitrary JSON mutation. Evidence: structured operation tests. <!-- evidence: BLOCK_OPERATIONS + INVERSE_OPERATION + blockCommandFaults(commands, rules, rawMutations). The raw mutations arrive as their own argument, so a change the editor made outside the vocabulary is the finding rather than an absence: a patch cannot be inverted for undo, cannot be authorized per operation, cannot carry the anchors a split has to divide, and cannot be replayed against a document that moved. split-or-merge-loses-an-anchor is the one that costs a newsroom a comment thread — the anchor is still in the database and points at a block that no longer exists. The allowed parents and the depth come from the document model passed in, not from the command, and a cross-document reference with no pinned revision quietly re-quotes whatever its author last wrote. -->
    • S7.8.d Run local field/cross-block validation for feedback and authoritative service validation on save; preserve stable reason/ remediation. Evidence: shared invalid fixtures. <!-- evidence: validationParityFaults(fixtureIds, local, service, decisions) and browser CONTROL 8. THIS IS THE ONE THAT SETTLES THE QUESTION: two validators tested separately both pass their own tests, and whether they say the same thing about the same document is a different question. The finder runs BOTH over ONE corpus and compares the REASON CODES, because a client reporting title-too-long where the service reports headline-length cannot point the refusal at the field somebody has to fix. Both failure directions are faults: a client pass treated as the decision (the client is the part an attacker controls), and a save the client refused to send that the service would have accepted (somebody cannot save valid work and the workaround is another browser). A reason code must be a slug — prose cannot be matched, counted or translated. -->
    • S7.8.e Track dirty state against base revision across local changes, server normalization, remote edits, save, failure, and restore. Evidence: dirty-state tests. <!-- evidence: dirtyStateFaults(observation) and browser CONTROLs 2, 4 and 10. DIRTY IS A COMPARISON AGAINST THE BASE REVISION, NOT A COUNT OF KEYSTROKES: CONTROL 2 types a character and deletes it and reads data-dirty="false" back, which a counter cannot produce. THE SERVER NORMALIZING WHAT IT STORED IS NOT A LOCAL CHANGE — CONTROL 4 saves "Bridge vote carried ", the server stores it single-spaced, and the editor shows the stored form and reports NO changes; counting that difference makes the next autosave send the normalization back to be normalized again. A failed save restores the flag and leaves the base revision alone (CONTROL 10), a commit advances it, and somebody else saving must not clear a dirty flag this editor is holding. -->
    • S7.8.f Implement durable autosave/offline queue/failure recovery with explicit state and no saved indicator before commit. Evidence: autosave fault tests. <!-- evidence: autosaveFaults + DURABLE_STORES, and browser CONTROLs 3 and 5. "SAVED" IS A PROMISE and it may not precede the commit it describes: CONTROL 3 holds the response open for 600ms and reads in-flight off the editor while the request is still open, then saved only once the revision moved. A QUEUE THAT DOES NOT SURVIVE A RELOAD IS NOT A QUEUE: CONTROL 5 goes offline, queues an edit, RELOADS THE PAGE and finds the work still there with its idempotency key and the revision it is a change to — and asserts nothing drained itself on load. Every queued entry carries a key (a resend after a timeout that actually succeeded would otherwise duplicate a block) and an expected revision (the queue drains exactly when somebody else has been editing). The retries are capped and the queued content is recoverable, because the last honest thing an editor can do is let somebody copy what it is holding. -->
    • S7.8.g Implement operation-aware undo/redo with focus/selection preservation and no server history rewrite or side-effect replay. Evidence: undo boundary tests. <!-- evidence: undoFaults and browser CONTROLs 6 and 7. UNDO IS A NEW FORWARD OPERATION — rewriting server history says the change never happened, which is false for everybody who saw it and for every export taken since. AND IT STOPS AT AN EDIT BY SOMEBODY ELSE: the stack carries a FENCE at every remote commit, and CONTROL 7 has a second author commit for real on the server, waits for the poll to notice, then presses Ctrl+Z and asserts the undo is REFUSED WITH A REASON rather than silently reverting them. CONTROL 6 is where the browser earned its place: selecting characters 8..13, typing over them and undoing restored a caret at 9, because the entry captured its selection inside the change handler — after the browser had already replaced the text. Reading it in beforeinput is the fix, and jsdom would have agreed with the bug. A new edit clears the redo branch, and the stack is capped. -->
    • S7.8.h Display current/base revision, author/time/reason/save state and detect expected-revision conflict with compare/preserve/resolve actions. Evidence: conflict end-to-end journey. <!-- evidence: REVISION_FACTS + CONFLICT_ACTIONS + conflictJourneyFaults, and browser CONTROL 9, which is the journey the item asks for: type a headline, have a second author commit for real, save, take the 409, and walk compare -> resolve. A CONFLICT MUST PRESERVE THE LOCAL WORK — every other outcome here is arguable and this one is not — so the control asserts the typed text is in the box at every step, including after the resolution. The comparison shows BOTH sides (showing only what is stored makes the decision from memory), all three actions are offered, the resolution is written as a NEW revision (rev-4 -> rev-5) rather than an overwrite, and it is timestamped so a later reader can tell which edits the reconciliation had in front of it. --> <!-- S7.8 rollup: 5001 kit tests (62 new, 101st entry point) + 11 jsdom tests + apps/oshun/web/tools/block-editor-browser.mjs (10 CONTROLs, 3 FAULTs) against a real node:http server that holds a revision, normalizes what it stores, honours if-match and idempotency-key, and can be committed to by a second author. The browser found the selection-capture defect. Six kit gates, eslint and a non-vacuous tsc clean; the scanner hits on "placeholder" are the HTML attribute the rule is ABOUT and carry stub:legitimate. Registration adds the block-editor-browser target. -->
  • S7.9 Build tree and hierarchy editing with cycle prevention, keyboard reorder, and impact preview. Evidence: tree property tests and a11y. <!-- A tree is the one workbench structure whose invariants cannot be checked by looking at the thing being changed: whether a move creates a cycle is a fact about the whole graph, whether a sibling order is intact is a fact about rows nobody read, and whether a move is safe is a fact about everything hanging underneath it. So hierarchies.ts (102nd entry point, 41 tests) takes the TREE as well as the command in almost every rule, and the tests GENERATE trees rather than hand-writing them. All seven children complete. --> <!-- TWO tools, because the item asks for two kinds of evidence. libs/oshun/workbench-kit/tools/hierarchy-cycle-race.mjs races a real PostgreSQL 16.14 and produces the finding the item is named for: TWO LEGAL MOVES MAKE AN ILLEGAL TREE, at READ COMMITTED and at REPEATABLE READ alike. apps/oshun/web/tools/hierarchy-tree-browser.mjs drives the tree pattern with real keys, a real focus ring, a real scrollbar and a real modal. Verified by patching the BUILT dist twice — weakening moveWouldCycle to a one-level comparison, and silencing the cycle detector — each failing a measured control. -->

    • S7.9.a Define stable node ID/type, parent, sibling order, depth, properties, allowed parent/child relations, roots, and limits. Evidence: tree contract tests. <!-- evidence: treeContractFaults in libs/oshun/workbench-kit/src/hierarchies.ts. The two halves of the model are checked AGAINST EACH OTHER: a type declared a root whose allowed parents do not include the root means the contract disagrees with itself about where a tree starts, and a type that may sit nowhere is a promise the model cannot keep. The ceilings are refused when they are meaningless (a depth of one is a list, a fan-out of zero forbids every child) and when they are absent — an unbounded tree makes the recursive read that renders it bounded by whatever somebody imported. -->
    • S7.9.b Implement authorized add, rename/edit, duplicate, move/reparent, reorder, archive/delete, restore, and reference operations with expected revision. Evidence: command tests. <!-- evidence: TREE_OPERATIONS + SUBTREE_OPERATIONS + treeCommandFaults(command, tree). A MOVE NEEDS TWO EXPECTED REVISIONS: one for the node and one for the DESTINATION, because the parent can move, be archived, or change what it accepts between the read and the write. It also needs two authorizations — a move is a write to where it is going as much as to what is moving — and move-authorized-only-on-the- node is the one every implementation misses. Every operation that changes a subtree must be atomic over it: half a reparent is a shape nobody designed and no screen can explain. A detach to the root is called out separately, because promoting a subtree is a different act with different authorization. -->
    • S7.9.c Prevent self-parent, descendant-parent, cycle through concurrent updates, duplicate/order corruption, invalid root, and depth/fan-out overflow. Evidence: property and real-DB race tests. <!-- evidence: treeIntegrityFaults(tree, contract) + moveWouldCycle + proposedMoveFaults, driven by PROPERTY tests over generated trees and by a REAL PostgreSQL. The property tests build trees by only ever attaching to an existing node — so a cycle is impossible by construction and anything reported is the finder being wrong — then redirect every node to each of its own descendants and require a finding, then check the cheap pre-flight rule against the expensive whole-tree walk over EVERY PAIR of nodes (roughly 5,700 of them). THE DATABASE FINDING IS THE ITEM: session A moves a under b while session B moves b under a; each validates against a snapshot in which the other has not happened, each is correct, both commit, and the committed tree has a cycle NEITHER OF THEM CREATED. REPEATABLE READ does not help — a stronger snapshot changes what each transaction sees and neither was wrong about what it saw. SERIALIZABLE does, because each read the row the other wrote (unlike S6.11, where two writes were merely a dependency), and so does one lock taken on the tree BEFORE deciding. CONTROL 6 shows that locking the destination path is not the same thing: the node being moved is not on it, so the recipe either leaves the cycle or deadlocks. -->
    • S7.9.d Provide tree keyboard pattern plus explicit move before/after/ into/out, numeric position, and target picker equivalent to drag. Evidence: keyboard/screen-reader tests. <!-- evidence: treeKeyboardFaults and browser CONTROLs 1–5. THE PLACEMENTS ARE TYPED COMMANDS, not a simulated drag: before, after, into, out and a numeric position each have their own control, and the component has no drag at all — the pointer path is the same dialog, so the two cannot diverge. CONTROL 2 presses each of the seven keys the role obliges FROM A ROW WHERE IT HAS SOMEWHERE TO GO (ArrowUp and Home at the first row are no-ops by definition, which would have made two of the seven vacuous) and asserts the tree state changed. CONTROL 3 walks Tab from a control before the tree to one after it and finds at most one stop inside — the single-tab-stop rule, which is what stops a large hierarchy being a keyboard trap in practice. CONTROL 4 expands the tree past its scroller and types "zoo" to reach Zoology, which is off screen. The destination list comes from the kit's walk, so CONTROL 5 finds it excludes the node and its descendants and still offers everything else. -->
    • S7.9.e Preview affected descendants, inherited policy/taxonomy, references, permissions, gates, schedules, URLs, and destructive or cost effects. Evidence: impact fixtures. <!-- evidence: IMPACT_DIMENSIONS + impactPreviewFaults. Nine dimensions, every one of them invisible in a tree control and every one of them travelling with the node. A dimension nobody COMPUTED is a separate finding from one nobody SHOWED, because silence reads as "nothing happens". The count is refused when it comes from the rendered tree — collapsing a branch would make a move look smaller than it is — and an irreversible consequence must be stated AND must cost more than one click, because the tree is the surface people click fastest. The preview is timestamped and expires, so one computed when a branch was expanded is not shown for a move made later. -->
    • S7.9.f Handle lazy/virtualized nodes, loading/unknown child counts, remote edits, moved/deleted target, conflict, partial failure, and rollback. Evidence: state matrix. <!-- evidence: TREE_STATES + treeStateFaults. AN UNKNOWN CHILD COUNT IS NOT ZERO: a lazily loaded node whose children have not been fetched renders identically to a leaf, and somebody archives a branch believing it is empty — so the unknown state is its own presentation AND the subtree actions are withheld until the subtree is known, because the preview that would make them safe cannot be computed. A partial failure must be rolled back OR itemized per node, and the per-node outcomes must cover the request. A destination that MOVED is followable and one that was DELETED is not; rendering both as an error makes the recoverable case look permanent. -->
    • S7.9.g Provide semantic list/breadcrumb/search and preserve focus, expansion, selection, and scroll through operations. Evidence: accessible journey. <!-- evidence: treeJourneyFaults and browser CONTROLs 6–8. THE BROWSER TAUGHT THE RULE: preserving a raw scrollTop across a move is NOT preserving the reader's place, because the move inserts rows above it — restoring the number leaves them four rows from where they were, and the browser's own scroll anchoring is meanwhile trying to do the same job from the other side, so the two together land a row out. The component owns it (overflow-anchor: none) and restores by NODE and offset; CONTROL 6 measures the node the reader was looking at moving 5px while scrollTop moved 96px. Focus goes back to the moved node by identity, and it is applied AFTER the modal closes, because closing a dialog hands focus back to whatever opened it and would otherwise undo the restoration silently. CONTROL 8 reads the list view: the PATH is a column, which is what the indentation was for, and a move can be started from it. --> <!-- S7.9 rollup: 5043 kit tests (41 new, 102nd entry point) + TWO tools. hierarchy-cycle-race.mjs (6 CONTROLs, 1 FAULT) against real PostgreSQL 16.14, stable over three consecutive runs; hierarchy-tree-browser.mjs (8 CONTROLs, 2 FAULTs) in real Chromium. Verified by patching the BUILT dist twice. Six kit gates, eslint and a non-vacuous tsc clean. Registration adds the hierarchy-cycle-race and hierarchy-tree-browser targets. -->
  • S7.10 Build graph viewing with semantic list/table parity. Evidence: graph-only information is zero. <!-- The parent states its own acceptance in five words, and it is a much harder claim than "there is also a table". So graph-views.ts (103rd entry point, 110 tests) takes the PICTURE and the SEMANTIC VIEW as two separate arguments in every parity rule and compares them: a derivation cannot disagree, and finding the disagreement is the entire job. All eight children complete. --> <!-- THREE ARTIFACTS. The kit module; GraphView.tsx + GraphView.module.css rendering the same model twice from one state; and apps/oshun/web/tools/graph-view-browser.mjs, 43 controls in real Chromium. The browser is not optional here for a reason no other S7 item had: whether the layout carries meaning can ONLY be settled by laying the same graph out twice and finding that nothing the view SAYS changed, and nothing runs in jsdom. Verified by patching the BUILT dist twice — RAISING the headroom floor above what the machine measures, and silencing the label comparison — each failing a measured control. --> <!-- THE SECOND PATCH DID NOT FAIL UNTIL THE TOOL EARNED IT. Every control passed over a surface that is CORRECT, so the parity rules never reached their reporting branches and silencing one changed nothing. The tool now also runs each rule over an input broken in the way it exists to catch, built from what the browser rendered. Same lesson as S7.3, S7.4 and S7.7, and the sharpest statement of it yet: a dist patch is only detectable if some control DRIVES THE FAILING CASE, not merely if it imports the symbol. -->

    • S7.10.a Register typed node/edge/port/property/direction/version, group/hierarchy, status, source, and authorization projection. Evidence: graph view contract. <!-- evidence: graphContractFaults in libs/oshun/workbench-kit/src/graph-views.ts. The rule worth having is AN EDGE MAY NEVER BE MORE VISIBLE THAN ITS ENDPOINTS: if the node type is CONCEALED from a reader who may not see it and the edge is merely DENIED, the refusal to show the edge admits the node exists, which is exactly the disclosure the concealment was bought to prevent. Composed from S4's dispositions rather than re-decided. Alongside it: a directed edge must leave an out port and arrive at an in one or the arrowhead draws the opposite dependency; an undirected type whose two ends accept different node types is directed in disguise; and every status vocabulary must be able to say unknown, because a vocabulary that cannot spells it as absence and absence renders identically to "fine" (S7.1). -->
    • S7.10.b Render bounded pan/zoom/fit/minimap/search/filter/highlight/path/ impact layout without making spatial position the sole meaning. Evidence: visual interaction tests. <!-- evidence: viewportFaults + spatialMeaningFaults, and CONTROL 2 and CONTROL 3 of the browser tool. THE ZOOM BOUNDS ARE DERIVED, NOT CHOSEN: the floor is MIN_NODE_TARGET_PX / NODE_DIAMETER_PX = 0.25, where a node stops being a 24px pointer target, and the ceiling is integer division of the 640px narrow breakpoint by the node diameter = 6, where one node stops fitting the narrowest supported viewport. The browser clicks the controls until they stop and checks where they stopped. "Position is not the sole meaning" is settled by RUNNING THE LAYOUT TWICE under different seeds and comparing everything the surface states, character for character and in order — a listing whose ORDER came out of the simulation is its own finding, and a run in which nothing moved is reported too, because it settles nothing. -->
    • S7.10.c Expose nodes as navigable table/tree with identical labels, types, properties, status, groups, inbound/outbound counts, and details. Evidence: node parity tests. <!-- evidence: nodeParityFaults + degreeOf, and CONTROL 1 and CONTROL 7. The comparison is CHARACTER FOR CHARACTER, not "equivalent": a picture reading Ingest » v3 beside a table reading Ingest (version 3) is two names for one object and a reader moving between the surfaces cannot tell they are the same thing. A DEGREE IS THREE NUMBERS — visible, withheld, and whether it is known at all — because a count that includes an authorization-removed neighbour leaks its existence while a count that silently omits it renders a node with hidden neighbours identically to a leaf. A CONCEAL contributes to none of them and a DENY contributes to withheld, so "and three more you may not read" is sayable and "there is a fourth service" is not. An unfetched neighbourhood reads unknown rather than 0 (S7.9's unknown-child-count, one dimension up and worse: a graph loads a neighbourhood at a time, so most of the view is legitimately unfetched and every one of those nodes draws as an isolate). -->
    • S7.10.d Expose edges as navigable table/list with source, target, type, direction, version, status, and properties. Evidence: edge parity tests. <!-- evidence: edgeParityFaults. An arrowhead that disagrees with the listed direction is the sharpest one — which way a dependency runs is the single fact an edge exists to state — and it is driven in the browser by flipping the arrow on what was rendered and confirming the built kit catches it. A versioned edge type whose row carries no version says this points at the target rather than at one particular target. An edge to a node the projection does not hold is a DISCLOSURE, not a rendering bug: the line is then the only evidence the node exists, which is precisely what removing it was for. An endpoint named by id where the node table calls it something else makes the two lists impossible to cross-reference by eye. -->
    • S7.10.e Synchronize selection, focus, filters, path/neighbor expansion, inspector, and deep link bidirectionally. Evidence: synchronization tests. <!-- evidence: synchronizationFaults, and CONTROL 4 and CONTROL 5. There is no synchronization in the component at all, which is the point: ONE state and two renderers, because two states with a sync between them is a race and the reader sees whichever won. The kit rule exists to catch an implementation that did it the other way. A DEEP LINK CARRIES THE SELECTION AND NEVER THE VIEWPORT — a pan and a zoom are where one reader's simulation happened to settle, so a shared link would land the recipient on a different layout of a graph that has since moved (S7.9's "a scroll position is not a place", one dimension up). Keys are an ALLOW-LIST per S7.7. A selection the filter removed is COUNTED rather than dropped, or a bulk action reaches things nobody can see. -->
    • S7.10.f Provide keyboard/nonpointer traversal and explicit relationship operations where editable. Evidence: nonvisual journey. <!-- evidence: traversalFaults + relationshipCommandFaults, and CONTROL 6. THE KEYBOARD PATH THROUGH A GRAPH IS THE TABLE, NOT THE PICTURE, and that is a finding rather than a shortcut: there is no accessible traversal of a force-directed layout because "the node to the right" is a fact about the simulation. So the obligation is discharged over the semantic view — the browser tabs in and arrows down until every node row has been reached, presses Home and End, and selects with Enter. relationshipCommandFaults refuses an endpoint type the contract forbids, a port that does not exist, a parallel edge where the type forbids one, a versioned edge with no version and an unversioned one carrying one, a write with no expected revision, and a command that can only be AIMED rather than typed. -->
    • S7.10.g Handle huge/partial/cyclic/disconnected/permission-redacted/ stale graphs with honest counts and no layout denial of service. Evidence: resilience/performance tests. <!-- evidence: resilienceFaults + neighbourhood + componentsOf, and CONTROL 8 and CONTROL 9. A CYCLE IS A LEGITIMATE GRAPH HERE — the exact opposite of S7.9, where a cycle was the bug — so every walk terminates on one and SAYS it came back round rather than reporting a short path. "No layout denial of service" is a DEGRADATION, not a cap: past the ceiling the picture is not drawn and the tables are untouched, which is only defensible BECAUSE parity is total, so the budget and the parity claim justify each other and a degradation that also thins the table is the named failure. The browser measures the real shipped layout at the ceiling and at twice it, and the counts are read out of the rendered header: the authorized total (not what survived the layout), the component count, and the age. --> <!-- THE CROSSING POINT IS A PROPERTY OF THE MACHINE, NOT OF THE CODE. The obvious control — "the layout fits the budget at the ceiling and exceeds it at twice the ceiling" — is measurable and FALSE on a fast development box while the ceiling is still right; the crossing simply happens further out. Picking the ceiling from wherever the fastest machine crosses is picking it from the machine least in need of protection. So the ceiling is stated against two machine-independent things instead: the SHAPE of the curve (quadratic, so doubling the nodes must cost markedly more than twice as much), and the HEADROOM (LAYOUT_HEADROOM_FACTOR, how much slower a reader's machine may be and still fit). Measured 164ms at 1500 nodes here, 6.1x headroom against a declared floor of 4x, budget first exceeded at 4x the ceiling. Take the MINIMUM of several runs: scheduler noise only ever ADDS time, so the minimum estimates the capability and the mean measures the load — two of these tools at once took the same layout from 194ms to 356ms and failed the check over nothing. -->
    • S7.10.h Compare every displayed visual encoding/fact/action to the semantic view and report zero graph-only information. Evidence: parity audit. <!-- evidence: parityAudit returning graphOnlyFacts and graphOnlyActions, and CONTROL 10 — which reads the encodings the surface ACTUALLY renders out of computed style and geometry (are the fills all the same? the stroke widths? the radii?) rather than out of a list somebody maintained, because the audit is only as complete as its inventory and an encoding absent from the declaration is itself a finding. An encoding declared DECORATION whose rendering varies with the data is the worse case: information whose author denies it exists, so nobody ever gives it a column. --> <!-- POSITION IS ITS OWN CASE, AND WRITING THIS FUNCTION IS WHAT SHOWED IT. Decisions 1 and 2 read like one rule and are not. Every other encoding is either decoration (and therefore constant) or a fact (and therefore mapped to a column) — but a layout ALWAYS varies with the data, moving nodes according to the edges being the entire job, and it can never have a column because a coordinate states nothing. Run through the general rule, position makes every honest surface fail as "decoration that varies with the data"; nine of the first ten tests written here failed on exactly that. So parityAudit takes the layout-independence VERDICT as an argument, and an audit made without one reports position-not-settled-by-a-layout-run rather than passing. A position pointed at an x column is refused separately: it satisfies the parity check while stating nothing a reader can use. -->
  • S7.11 Build timeline editing with keyboard/timecode controls and semantic track list parity. Evidence: trim, move, split, and zoom journeys. <!-- THE MIRROR OF S7.10, AND THAT IS THE WHOLE DESIGN. There, spatial position was the one encoding that could never have a column, because a layout coordinate states nothing. Here the horizontal axis IS the data: a clip drawn eight pixels right of its own timecode has told the eye a different time from the one the field prints, and at 100 px/s that is a fifth of a second no amount of source-reading finds. So timelines.ts (104th entry point, 146 tests) takes the MEASURED picture and the RENDERED table as two separate arguments and compares them, and the browser tool measures the boxes rather than reading the style strings back. All eight children complete. --> <!-- THREE ARTIFACTS. The kit module; TimelineEditor.tsx + TimelineEditor.module.css rendering one state twice with 36 jsdom specs; and apps/oshun/web/tools/timeline-editor-browser.mjs, 71 controls in real Chromium against a real node:http server. Verified by patching the BUILT dist three times — silencing the position comparison, NARROWING the link-key allow-list, and RAISING the frame-target floor above what the machine renders — each failing a different measured control, and the tool green again once the dist was restored. --> <!-- THE BROWSER FOUND FOUR DEFECTS SOURCE-READING DID NOT. (1) The semantic table's nowrap timecode cells pushed the whole PAGE sideways at 360px and 640px, because a flex item defaults to min-width: auto and refuses to shrink below its content. (2) max-width: 100% resolves against the CONTENT box, so a region with a 1px border lands at 102%. (3) A VISUALLY-HIDDEN SPAN IS position: absolute AT ITS STATIC POSITION, so without a positioned ancestor it escapes the scroll container its own content lives in — the table scrolled correctly the whole time and the document scrolled anyway, by the width of a box nobody can see. (4) A state cell reading "unlocked" makes includes('locked') true, so any consumer matching on substrings reports every open clip as locked. -->

    • S7.11.a Define time base, frame/sample/beat units, frame rate/sample rate/ tempo map, rounding, origin, duration, tracks/lanes, clips/events, and stable IDs. Evidence: timeline contract. <!-- evidence: timelineContractFaults in libs/oshun/workbench-kit/src/timelines.ts. THE DECISION EVERYTHING ELSE RESTS ON: time is an exact integer in a base that divides every rate, never a float of seconds. 29.97 is not a frame rate; 30000/1001 is, and a float second loses a frame roughly every half hour of NTSC — invisible until two departments' exports disagree. The base is the FLICK, 705,600,000 per second, the unit published in 2018 for exactly this problem, because it is an exact integer multiple of every broadcast frame period INCLUDING the /1001 rates and of every standard sample period from 8 kHz to 192 kHz. Asserted: 23,543,520 flicks x 30,000 is 1001 seconds to the flick. Alongside it: the exact-arithmetic ceiling is stated and a longer document is refused rather than silently losing its last digits; a rate the base does not divide returns null rather than a rounded answer; and an identity derived from position or label is refused, because a move is not an exceptional event on this surface, it is the product. -->
    • S7.11.b Implement keyboard transport, seek, step, selection, range, zoom/fit, marker, track, clip/event traversal, and shortcut collision handling. Evidence: key map tests. <!-- evidence: keyMapFaults, resolveChord, DEFAULT_KEY_BINDINGS, and CONTROLS 4, 5, 7 and 8. A SINGLE-CHARACTER SHORTCUT IS A TYPING HAZARD, and the rule is wider than a bare letter: Shift does not rescue it, because Shift+ArrowLeft extends a text selection and Backspace is what somebody fixing a typo reaches for first. So the hazard is every chord with no Control, Alt or Meta, and each must declare itself suppressed while a field has focus. The shipped map PASSES ITS OWN CHECKER — a default map its rule rejects is a rule nothing has ever run — and it keeps J-K-L and I/O because those are the conventions of the domain. The reserved list COMPOSES S7.1's rather than restating it: a timeline shortcut that closes the browser tab is exactly as bad as one that traps focus. The zoom bounds are DERIVED — the floor is the whole timeline in the narrowest supported canvas, the ceiling is one frame filling it — and the browser clicks until they stop and checks where. -->
    • S7.11.c Provide exact start/end/in/out/duration/offset/speed/value fields with unit conversion, validation, snapping, and accessible labels. Evidence: timecode control tests. <!-- evidence: timecodeFieldFaults, parseTimecode, framesToTimecode, applySnap, and CONTROLS 9 and 15. DROP FRAME IS A LABELLING SCHEME, NOT A RATE: no frames are dropped, two frame NUMBERS are, at the top of every minute except every tenth. Two consequences the code carries — a drop-frame and a non-drop-frame timeline over the same media have IDENTICAL durations, and 00:01:00;00 is not a time but a typo a lenient parser turns into a one-frame error nobody ever finds, because the number it prints back is not the number that was typed. The SMPTE known values are pinned (frame 1800 prints 00:01:00;02, 17982 prints 00:10:00;00, 17981 prints 00:09:59;29, and the 59.94 pair) and every frame number across eleven minutes round-trips. AN END TIMECODE IS AMBIGUOUS UNTIL THE VIEW SAYS WHICH END IT MEANS — last frame or first frame after — so a field declares which, and one that declares neither is the finding. A field that snaps must SAY it snapped, so applySnap returns the target and the distance. -->
    • S7.11.d Implement move, trim, split, merge, duplicate, ripple/overwrite, reorder, enable/mute/lock, and domain plugin operations with undo. Evidence: operation known-value tests. <!-- evidence: applyOperation returning its own inverse, plus overlapFaults and desynchronisedLinks, and CONTROL 7. A CLIP HAS THREE TIMES AND A SPEED, and conflating any two of them scrubs the media: a MOVE changes start alone, a HEAD TRIM changes start, duration AND sourceIn, and the source it consumes is the timeline delta multiplied by the SPEED. Forgetting that factor is the single most common real defect in this domain and it is a known-value test — a split at four seconds into a 2x clip advances the right piece by eight seconds of source, not four. The inverse is produced by the operation rather than reconstructed by a stack, because only the operation knows what it actually did; an overwrite that destroyed material reports that it has NO single-operation inverse rather than offering a wrong one. A LINK IS BROKEN BY A RIPPLE THAT CROSSES IT — the audio under a shot the picture track rippled past comes out misaligned by exactly the rippled amount — so the result names the groups it pulled apart, compared BEFORE against AFTER so a group already out of sync is not blamed on this edit. -->
    • S7.11.e Provide track/event table/list with every identity, time/value, status, relationship, and operation available nonvisually. Evidence: semantic parity tests. <!-- evidence: timelineParityFaults, and CONTROLS 1, 2, 3 and 17. The comparison is between a MEASURED box and a PRINTED timecode: the browser reads getBoundingClientRect against the scrolling content box rather than reading the component's own style.left back, because comparing a model to itself settles nothing. Every rule then gets a NEGATIVE control over an input broken the way it exists to catch, built from what was rendered — a clip nudged 8px, a width shrunk a quarter, a label with a trailing space, a row removed, a drawn clip removed, a link the picture shows and the row omits. THE ASYMMETRY RUNS ONE WAY: the picture may show less than the table, because a clip scrolled out of the canvas is legitimately undrawn, and never more — so a row inside the VISIBLE window with nothing drawn for it is its own finding. A row that offers nothing is not a fault; a row that offers nothing AND does not say which of the three reasons it is — read-only track, locked item, not yours — renders a protected item identically to a broken control. -->
    • S7.11.f Synchronize playhead/range/selection/zoom/filter/inspector between visual and semantic views. Evidence: synchronization tests. <!-- evidence: synchronizationFaults, and CONTROLS 11 and 12. There is no synchronization in the component at all, which is the point: ONE state and two renderers, because two states with a sync between them is a race and the reader sees whichever won. The kit rule exists to catch an implementation that did it the other way. A DEEP LINK CARRIES A TIME AND NEVER A VIEWPORT — the mirror of S7.10's rule rather than a contradiction of it: a playhead and an in/out range are positions in the DATA and worth sending to somebody, while the zoom and the horizontal scroll are facts about how wide the sender's monitor was. Keys are an ALLOW-LIST per S7.7, and the browser hands the keys the page REALLY wrote to the built kit. READING THE TABLE DOES NOT SEEK: arrowing down the semantic list is how a nonvisual reader inspects a timeline, and if it also moved the playhead they could not read without editing what everybody else is watching. The filter EVALUATES the parsed tree and refuses an unknown field BY NAME — S7.10's lesson, since a filter that collects terms contributes none for a comparison and [].every() is true, so the box silently shows everything on exactly the input somebody typed most deliberately. -->
    • S7.11.g Handle variable frame rate/tempo, drop frame, long duration, overlaps, boundaries, locked/unauthorized items, remote edits, and conflicts. Evidence: boundary fixture matrix. <!-- evidence: boundaryFaults, durationPrintFaults, beatAtFlicks / flicksAtBeat / barBeatAtFlicks, and CONTROLS 10, 15, 16 and 17. ROUND FROM THE SEGMENT ORIGIN, NEVER FROM THE PREVIOUS BEAT: most tempi have no whole-flick beat, and rounding an increment and adding it drifts linearly while rounding the whole offset does not accumulate at all — asserted by showing the naive error growing tenfold with the beat count while ours stays inside half a flick. A BAR NUMBER ACCUMULATES ACROSS METER CHANGES, which the first bar test caught as a real bug: numbering from the current segment's meter alone renumbers every bar in the piece the moment somebody inserts a meter change. A DURATION IS A COUNT AND A POSITION IS A WALL CLOCK, so a span printed through the drop-frame label scheme reads ten minutes as nine fifty-nine. AN OVERLAP IS A TRACK PROPERTY, not an error — checked against the lanes the track declared. A timeline past the 24-hour wrap is refused because two instants in it print the same label. A CONCEAL is absent from the document the view is built from and a DENY is counted, so "and one you may not read" is sayable and "there is a fourth clip" is not. -->
    • S7.11.h Run trim, move, split, zoom, undo, autosave, conflict, and nonpointer journeys under supported viewports. Evidence: Playwright tests. <!-- evidence: apps/oshun/web/tools/timeline-editor-browser.mjs, 71 controls, Nx target timeline-editor-browser. The whole trim-split-undo journey runs from the harness control BEFORE the timeline with ZERO pointer events, and each step is checked in BOTH surfaces: the two halves of a split are drawn edge to edge with no gap and no overlap, and a head trim moves the drawn left edge by exactly the pixels that time maps to. THE TYPING HAZARD IS DRIVEN FOR REAL — 00:00:03:00 typed into the timecode field, then two ArrowLefts and a Backspace, and nothing split, nothing deleted, the playhead unmoved and the caret where a field owes its reader. The conflict journey uses a real node:http server the page really polls: the request count proves it is watching rather than having watched once, the remote value is one only the server knows, and undo then stops at the fence rather than re-applying a state the server never had. Measured at 360px, 640px and 1280px — where the page must not scroll sideways while the canvas and the table each must. -->
  • S7.12 Build comparison surfaces for text, metadata, images, video, audio, 3D, graphs, and gate results through typed adapters. Evidence: each adapter has a golden comparison fixture. <!-- S2.5 ALREADY DECIDED WHAT A DIFFERENCE IS, and nothing here re-decides it: every adapter emits DiffOperations, because one contract is the only thing that makes the review surface one surface. comparisons.ts (105th entry point, 99 tests) adds the eight typed adapters and the one claim that makes any of them falsifiable — a GOLDEN is a recorded input PAIR plus the recorded answer plus the adapter version it was taken under. All ten children complete. --> <!-- THREE ARTIFACTS. The kit module and its eight goldens in evidence/v1-workbenches/workbench-kit-comparison-goldens.json, recorded over fixtures that live beside the spec reproducing them; CompareSurface.tsx rendering any outcome with 26 jsdom specs; and apps/oshun/web/tools/compare-surface-browser.mjs, 44 controls in real Chromium. Verified by patching the BUILT dist twice — silencing the cursor repeat detector and emptying SURVIVING_ANCHOR_TARGETS — each failing a measured control, and the tool green again once restored. --> <!-- THE FIRST PATCH DID NOT FAIL UNTIL THE TOOL EARNED IT, the fourth sighting of S7.10's lesson. Silencing the repeat detector changed nothing because none of the six negative controls DROVE a repeat; adding one made the same patch fail. The browser also found a real focus defect and a real gap in two adapters, both described in the children below. -->

    • S7.12.a Define common compare contract for left/right/base identity and revision, adapter/version, aligned change list, selection, filters, unknown/unsupported/redacted state, and accessible summary. Evidence: compare interface. <!-- evidence: compareContractFaults, accessibleSummaryOf, goldenFaults and goldenRegistryFaults in libs/oshun/workbench-kit/src/comparisons.ts. A GOLDEN WITHOUT THE ADAPTER VERSION PINS NOTHING: the same inputs through a later adapter legitimately produce a different answer, so a golden that cannot say which version it was taken under is either failing forever or being quietly re-recorded, and both make it a test nobody can fail. A version mismatch is therefore reported as UNTAKEN rather than failed. Alongside it: TWO REVISIONS ANSWER "WHAT IS DIFFERENT" AND THREE ANSWER "WHO CHANGED IT", so a merge control offered over a two-way comparison asks for a choice without the fact that decides it; a redaction declares whether the FACT of the change is itself the secret, because usually "changed, value withheld" is the useful sentence and sometimes the one bit is what the redaction was bought to protect; and the accessible sentence is written in exactly one place, or the nine surfaces would say the same thing nine ways. -->
    • S7.12.b Implement text adapter for block/line/word/character and semantic structured changes with whitespace/locale policy. Evidence: text goldens. <!-- evidence: compareText, graphemes, prepareText, and the text/one-line-rewritten golden. A CHARACTER DIFF OPERATES ON GRAPHEME CLUSTERS, NOT CODE UNITS — a family emoji is one grapheme, five code points and eight code units, and splitting it renders as mojibake, from which the reader correctly concludes the tool is broken. NORMALIZE AND SAY THAT YOU DID: é composed and decomposed are the same text and different bytes, so comparing unnormalized reports a change nobody made while normalizing silently hides a real round-trip loss — both sides get reported. Case folding is LOCALE-DEPENDENT and asserted on the Turkish dotless i, and folding with no locale is refused BY NAME rather than picking somebody else's language. The alignment is S2.5's own longestCommonSubsequence rather than a second implementation. -->
    • S7.12.c Implement typed metadata adapter for field presence/value/unit/ enum/list/relationship changes with redaction. Evidence: metadata goldens. <!-- evidence: compareMetadata, convertQuantity, and the metadata/runtime-restated-and-rating-raised golden — which is the whole argument in one fixture: 90 minutes rewritten as 5400 seconds is ONE PRESENTATION change and the rating that really moved is one semantic one. A comparison without unit conversion makes every unit migration look like a data migration. A SET HAS NO ORDER, so reporting one reordered is reporting a change that does not exist; a SEQUENCE reordered is real, and checking membership BEFORE order is what keeps "somebody moved chapter 3" from rendering as "somebody rewrote the chapter list" — a real bug the tests caught. Absent, null and default stay three different facts (S2.5's rule, kept), and a reference RETARGETED is a different change from its target having MOVED, because the two send the reader to different places. -->
    • S7.12.d Implement image adapter for bytes/hash/dimensions/color/profile/ metadata plus aligned overlay/swipe/pixel/perceptual differences and nonvisual metrics. Evidence: image goldens. <!-- evidence: imageMetrics, meanStructuralSimilarity, compareImages and the image/one-pixel-and-a-profile golden. A PIXEL DIFF OF ZERO DOES NOT MEAN THE IMAGES ARE THE SAME: colour space, ICC profile, bit depth and EXIF orientation all survive a pixel-identical comparison and at least one of them changes what a viewer sees, so the envelope is compared separately and "identical pixels, different profile" is its own finding. EVERY VISUAL COMPARISON OWES A NUMBER — a swipe and an overlay are two ways of showing a difference to somebody who can see it, and the changed-pixel count, the peak channel delta, the PSNR and the bounding box are what a reviewer says out loud. PSNR is 10 log10(255^2/MSE) asserted at a known 28.13 dB and INFINITE for identical images rather than capped; SSIM is Wang et al. over 8x8 windows, asserted at a known 0.99548 over a uniform shift that changed every one of 64 pixels — which is the demonstration that a pixel count and a structure measure are different questions. -->
    • S7.12.e Implement video adapter for stream/codec/duration/frame rate/ tracks/metadata and time-aligned visual/audio/event differences. Evidence: video goldens. <!-- evidence: compareVideo and the video/conformed-to-25 golden. THE ALIGNMENT IS BY TIME, NEVER BY FRAME INDEX: frame 100 of a 24 fps master and frame 100 of a 25 fps conform are 4.167 s and 4.000 s, so a per-frame comparison indexed by number compares unrelated pictures and reports the whole film as changed. The same decision S7.11 made about a timeline, and the golden is the case that proves it — the same 240 frames at two rates, where the count is identical and the film is not. Duration is computed from the EXACT rational rate, so 30,000 frames at 30000/1001 is exactly 1001 seconds. A sampled frame with no counterpart at that time is named in notCompared rather than skipped, and a track set with no shared id says so rather than guessing which track became which. -->
    • S7.12.f Implement audio adapter for bytes/format/duration/channels/ loudness/peak/waveform/spectrum/events and synchronized audition with text/data alternative. Evidence: audio goldens. <!-- evidence: integratedLoudness, truePeakDbtp, rmsEnvelope, compareAudio and the audio/six-db-louder golden. Real ITU-R BS.1770-4: the two published K-weighting stages, 400 ms blocks at 75% overlap, an absolute gate at -70 LUFS and a relative gate 10 LU below what survived it. THE COEFFICIENTS ARE ASSERTED AS DATA and the RESPONSE IS MEASURED — the measured curve reproduces the published one (flat through the midband, rolling off below 100 Hz, plateauing at the +4 dB shelf), which is what catches a transposed digit that typechecks and produces a meter confidently wrong. The rest are properties of the definition rather than a calibration figure somebody half-remembered: doubling every sample moves exactly 6.02 dB, an identical second channel exactly 3.01, a surround channel exactly 10 log10(1.41), an LFE exactly nothing, and digital silence is negative infinity rather than a large negative number. THE GATE IS DEMONSTRATED, not described: two seconds of tone followed by six of silence measures within half a dB of two seconds of tone alone, where an ungated meter would be six dB out. A SAMPLE PEAK IS NOT A PEAK, so the true peak is oversampled — and the interpolation UNDERSTATES it, which is documented rather than hidden. An audition with no numeric alternative is refused: listening is the whole comparison for a hearing reader and none of it for anybody else. -->
    • S7.12.g Implement 3D adapter for stage/scene graph, geometry, topology, transforms, materials/textures, skeleton/animation, variants/refs, and OpenUSD validation. Evidence: 3D goldens. <!-- evidence: compareStages, windingReversed, COMPOSITION_ARCS and the stage-3d/parent-moved-child-did-not golden — a fixture whose entire content is the finding: the child's authored transform is identical, its world position changed because the parent moved, and attributing that to the child sends a modeller to edit a value nobody touched. The golden records ONE operation, on the parent. Alongside it: composition arc strength is USD's own LIVRPS order, and a CHANGED ARC WITH AN UNCHANGED VALUE IS A REAL CHANGE, because the same answer through a stronger arc survives edits the weaker one loses. REVERSED WINDING passes every count check and renders the mesh inside out, so it is detected face by face including rotation. A VARIANT SELECTION IS NOT A GEOMETRY EDIT — both variants still exist and somebody chose the other one. Transforms are not compared at all across differing up axes, and geometry is not compared across differing unit scales, because a point at 1.0 is a different place in each stage; both are named in notCompared rather than silently skipped. -->
    • S7.12.h Implement graph adapter for node/edge/property/group/path and layout-versus-semantic changes. Evidence: graph goldens. <!-- evidence: compareGraphs and the graph/tidied-and-one-rename golden. S7.10 established that a coordinate states nothing; this is that rule one step on. A LAYOUT CHANGE IS NOT A GRAPH CHANGE, so somebody tidying a diagram produces ONE operation, marked presentation, carrying the count of nodes that moved — rather than forty-seven entries that make the review unreadable and train everybody to skim it. layoutOnly is the answer to "is there anything here at all", and it goes false the moment one label changes. The golden holds exactly that pair: two nodes moved and one renamed, one presentation change and one real one. -->
    • S7.12.i Implement gate adapter for raw values/units, evaluator/corpus/ model, policy/threshold, verdict, freshness, evidence, waiver, and remediation. Evidence: gate goldens. <!-- evidence: compareGateRuns, verdictCauses and the gate/bar-raised-not-work-worsened golden. A VERDICT THAT FLIPPED WITHOUT THE VALUE MOVING IS A POLICY CHANGE: "the build got worse", "we raised the bar" and "we changed how we measure" are three different conversations and only one is about the work, so a changed verdict is decomposed into value-driven, threshold-driven, evaluator-driven, corpus-driven and model-driven — EVERY cause that applies, not the first, because "we raised the bar AND the number got worse" is common and reporting one of them is how half a regression goes unnoticed. Two numbers from two different evaluators are not a trend, and drawing an arrow between them is the most common way a quality dashboard lies. THE SAME VERDICT OVER DIFFERENT EVIDENCE IS A DIFFERENT CLAIM, and a surface that renders only the verdict says nothing happened. verdictCauses returning EMPTY is the answer when nothing observable changed — an earlier version had a fallback that made the list never empty, so verdict-change-with-no-cause could never fire, which is the shape of every unwired gate. -->
    • S7.12.j Provide stable difference navigation, side switching, synchronized context, comments/decision linkage, keyboard/screen-reader parity, and large-diff pagination across adapters. Evidence: common compare journeys. <!-- evidence: compareNavigationFaults, linkageFaults, differenceCursor, differenceAnnouncement, CompareSurface.tsx, and apps/oshun/web/tools/compare-surface-browser.mjs (Nx target compare-surface-browser). The word doing the work in the item is COMMON: the surface takes an outcome and knows nothing about what was compared, so the browser renders a text, a metadata and a gate comparison and asserts the navigation controls are character-for-character the same set. THE ORDER IS THE RESULT'S ORDER, ALWAYS, because every conversation about a review is conducted in ordinals, and pagination is a WINDOW onto that order — the page is DERIVED from where the cursor is, and the ordinals stay absolute so page two starts at eleven. A COMMENT HANGS ON A STABLE ANCHOR OR IT DOES NOT HANG: an operation id is built from a path, so it survives recomputing the comparison and not the next edit, and when a comparison carries no anchor the surface says so and offers no box rather than one that silently drifts. --> <!-- THE BROWSER FOUND TWO REAL DEFECTS. (1) FOCUS DIED AT EVERY PAGE BOUNDARY: .focus() called in the handler that moved the cursor targets a row that is not rendered until the page catches up, so focus fell to the document body and the next arrow key went nowhere — a keyboard reviewer stranded twelve differences into a sixty-difference review with no way back in except the mouse. Moved into an effect keyed on the page. (2) ONLY THE TEXT ADAPTER EMITTED ANCHORS, so a reviewer could comment on a sentence and not on a gate verdict; a metadata key and a gate id are both stable identity, and both adapters now anchor from the path each operation already carries, in ONE place rather than at each push. Also settled here and nowhere else: the two context panes really scroll together without the echo scrolling the first one back, the panes FOLLOW the cursor while a side switch deliberately does NOT move them, and the DOM holds one page of a review with 118 differences in it. -->
  • S7.13 Build the decision-record drawer. Evidence: selected and rejected variants remain linked to the decision. <!-- S2.7 ALREADY DECIDED WHAT A DECISION RECORD IS — its lifecycle stages, the options that lost, evidence pinned by hash, supersession that never edits. decision-drawer.ts (106th entry point, 50 tests) adds only the presentation, and the acceptance sentence is much harder than it reads: the hard word is REMAIN. A rejected option in the DOM behind a collapsed section is linked to the decision the way a footnote is linked to a book. All seven children complete. --> <!-- THREE ARTIFACTS. The kit module; DecisionDrawer.tsx rendering every option as a SIBLING with 28 jsdom specs; and apps/oshun/web/tools/decision-drawer-browser.mjs, 49 controls in real Chromium (Nx target decision-drawer-browser). Verified by patching the BUILT dist twice — silencing the reachability comparison and NARROWING OUTCOME_STATES — each failing a measured control, and green again once restored. The second patch changed nothing until a control DROVE the retained-current case; adding one earned it. Fifth sighting of S7.10's lesson. --> <!-- THE BROWSER FOUND A REAL DEFECT. S7.13.e asks for a round trip with the selection PRESERVED, and a real goBack() remounts the component: the selected option lived in React state alone, so a reader who went to check a number against its source came back to the top of the record with nothing selected. The selection now lives in the address bar (replaceState, so selecting is not a navigation) — S7.11's playhead rule, one surface on. -->

    • S7.13.a Display decision identity/type/scope/status/owner/effective version/review date and affected objects/capabilities. Evidence: header fixtures. <!-- evidence: drawerHeaderFaults in libs/oshun/workbench-kit/src/decision-drawer.ts, and CONTROL 3. The header is taken as its OWN argument rather than derived from the record, for the reason every S7 parity rule takes two: a derivation cannot disagree, and finding the disagreement is the job. AN EFFECTIVE VERSION WITH NOTHING ATTACHED IS A NUMBER — "3" answers nothing and "from version 3 of the ingest workbench" is the sentence somebody needed — so a bare one is refused separately from a wrong one. A STAGE RENDERED AS A WORD THAT IS NOT A STAGE ("live", "current", "done") leaves a reader unable to tell whether the decision binds anybody. And a header with no review statement is how a temporary exception becomes permanent without anybody deciding that: a date or a stated reason there is none, never blank. -->
    • S7.13.b List every option/variant with stable ID, summary, provenance, score/evidence, status, and compare action. Evidence: option fixtures. <!-- evidence: optionListFaults, and CONTROL 1. Each field is refused by name because each absence reads as something else: no provenance and a thing somebody proposed is indistinguishable from a thing a tool generated, which are different arguments; no score STATEMENT and a blank cell reads as a score of nothing, where "not scored" is a fine answer; no compare action and "why not this one" — the question the record exists to answer — has no control that answers it. An option the reader may not see is COUNTED on screen (S7.10's degree and S7.11's track total, third application): a silent drop renders a decision with alternatives identically to one taken without any. -->
    • S7.13.c Mark selected, rejected, retained-current, unavailable, and deferred options without hiding any considered option. Evidence: outcome state tests. <!-- evidence: DRAWER_OPTION_STATES, STATE_FOR_STATUS, outcomeStateFaults and drawerStateOf, and CONTROLS 1, 2 and 9. THE VOCABULARY MUST BE ABLE TO SAY "WE KEPT WHAT WE HAD" AND "IT WAS NOT ON OFFER", and S2.7's three statuses cannot: recording a stay-put as chosen loses that nothing changed, and recording a vendor who withdrew as rejected attributes to the team a judgement they never made. So five presentation states over three, each mapped to exactly one record status, and a rendered state that disagrees with the filed one is itself a fault — the drawer may translate and may not contradict. THE REACHABILITY RULE IS MEASURED, NOT ASSERTED: stepsToReach is counted in the browser (visible on load is zero, inside a closed disclosure is one), and a rejected option further away than the selected one is the finding. That is the difference between a record that answers the question and one that technically contains the answer. -->
    • S7.13.d Display rationale, trade-offs, consequences, risks, actors/ authority, reviews, timestamps, and immutable evidence refs. Evidence: detail tests. <!-- evidence: decisionDetailFaults. AN EMPTY RESIDUAL RISK IS A CLAIM, not an absence — it says "we say nothing is still wrong" — so rendering nothing at all lets the reader assume nobody looked, and the surface states the claim instead. A DELEGATED ACTOR IS TWO PEOPLE and the delegated AUTHORITY is a third fact: who was at the keyboard and whose authority the decision carries answer different questions, and only the second says whether it binds anybody. A REDACTED ACTOR IS NOT AN UNKNOWN ONE — somebody was there and their identity was removed for a stated cause. Every evidence reference must carry a version AND a digest, or it points at whatever that document says today, which is the failure pinning was bought to prevent. REDACTION APPLIES TO THE RATIONALE, NOT ONLY THE EVIDENCE: the reason can name the thing the evidence was redacted to protect, so redacting one and rendering the other spends the budget and keeps none of it. -->
    • S7.13.e Navigate to compare/evidence/source and back with preserved selection/focus and authorization/redaction. Evidence: drawer journey. <!-- evidence: drawerNavigationFaults, and CONTROLS 4, 5 and 8. THE DEFECT THE BROWSER FOUND: a real page.goBack() remounts the component, and the selection lived in React state alone, so the reader who went to check a number came back to the top of the record. Now in the address bar via replaceState — selecting an option is not a navigation, and a Back button that stepped through every option somebody clicked would make the real journey unreachable. A CONCEALED DESTINATION IS NOT OFFERED AT ALL: a refused link admits it exists, which is the disclosure the concealment was bought to prevent (S7.10's edge rule, applied to a link). A DENIED one is also not offered, for the different reason that a control which refuses when pressed teaches its reader to expect refusals. -->
    • S7.13.f Support propose/review/approve/reject/supersede/reverse only when authorized, with exact validation and audit. Evidence: mutation tests. <!-- evidence: decisionMutationFaults, MUTATION_FROM_STAGES, and CONTROL 7. AUTHORIZED, VALIDATED, AUDITED, IN THAT ORDER — and the stage list is part of the validation rather than a comment: supersede runs only from effective, because superseding a proposal is not a supersession, it is a withdrawal. The surface offers a control only where BOTH the authorization and the stage admit it, and the browser reads the offered set at three different stages to check. A refusal must carry the exact reason naming the stages it could have run from, because a refusal nobody can act on is a dead control with extra steps. An accepted mutation with no audit entry makes the archive unfalsifiable, and an audit entry for a REFUSED one records something that did not happen — both are named. A REVERSE THAT EDITS THE ORIGINAL is the worst of them: reversing a decision is a new decision, and editing the old one erases the reasoning of the day, which is the one thing the archive is for. -->
    • S7.13.g Render missing/inaccessible/stale evidence and supersession lineage honestly. Evidence: degraded state fixtures. <!-- evidence: degradedStateFaults, and CONTROLS 6 and 8. EVIDENCE HAS FOUR RESOLUTIONS AND ONLY ONE IS FINE: resolved, missing, inaccessible to this reader, and stale — where stale means the digest no longer matches what the author read. Rendering all four as a link tells the reader the reasoning still rests on what it rested on, which is precisely what the surface cannot know, so only resolved is an anchor and the browser checks the degraded ones are not even in the tab order. Both directions are refused: calling it resolved when the digest moved, and calling it stale when nothing changed and somebody is sent to re-read an unchanged document. A SUPERSEDED RECORD STAYS READABLE — it is not a redirect, it renders its own options and its own reasoning behind a link forward, and the link is TWO-WAY or somebody arriving at the current decision cannot find the reasoning it replaced. -->
  • S7.14 Build the gate panel with freshness, evidence, policy version, owner, remediation, waiver, and rerun state. Evidence: state matrix. <!-- evidence: libs/oshun/workbench-kit/src/gate-panel.ts (114 tests, the 107th entry point), apps/oshun/web/src/components/workbench/ GatePanel.tsx (+ module CSS, 28 jsdom tests) and apps/oshun/web/tools/gate-panel-browser.mjs (61 controls in real Chromium). S2.9/S2.10/S2.11 already decided what a gate IS, what a RUN recorded and when a waiver authorises anything; none of that is re-decided, and the whole module composes them — the fourth sighting of S7.12's rule that the kit is large now, so look for the module that already owns the vocabulary before writing one. THE PARENT ACCEPTS ON A STATE MATRIX, AND THAT IS THE RIGHT ACCEPTANCE, because every expensive failure of a gate console is one state rendered as another. Two of those get named here for the first time. STALE AND EXPIRED ARE DIFFERENT ANSWERS BECAUSE THEY ARE DIFFERENT ACTIONS: S2.10's evidenceFreshness carries both under one word with a why, and the reader's next move differs — expired is a clock, nothing is known to have changed and a rerun will very probably say the same; stale is a CHANGE, a dependency moved or the answer was given against another revision, and a rerun should be expected to say something else. Rendering them identically makes "it is a bit old" the mental model for a rights check whose consent record was withdrawn this morning. And AN EXPIRED PASS IS NOT A PASS — decision 10, the one that changes an outcome rather than a sentence. Handing S2.10's resolveEffect the VERDICT of an answer that no longer stands produces satisfied, the summary counts it, and the release ships. So a row whose answer does not stand contributes S2.8's stale NON-VALUE and the definition's own whenEvidenceUnavailable decides what that costs, which is the catalogue's already-written answer to "what do we do when this gate cannot answer" applied to the case where it once could. -->

    • S7.14.a Display gate ID/name/severity/applicability, current target revision, owner/escalation, and required/optional status. Evidence: definition fixtures. <!-- evidence: gateHeaderFaults, and CONTROL 11. SEVERITY IS PER-STAGE AND OBLIGATION IS canBlock, WHICH ARE TWO DIFFERENT QUESTIONS. Printing defaultSeverity is right most of the time and wrong in the one place somebody is making a decision, because a gate that is a warning while drafting and a blocker at publish has two correct answers. And a blocker in a SHADOW lifecycle stops nothing: a row that says "required" over it sends somebody to fix something that was never going to hold anything up, which is how a shadow rollout — whose entire purpose is to find out what a gate WOULD have done — becomes indistinguishable from an enforcement it never was. The browser drives the shadow case end to end and reads optional, a lifecycle line naming the shadow, and an effect of noted rather than blocks. Also refused in both directions: a row for a gate the catalogue does not apply here (a check the reader will wait for and nobody will run) and a gate that applies rendered as not applicable, including the case where the LEDGER says not-applicable and the CATALOGUE disagrees. -->
    • S7.14.b Display raw values/units/uncertainty, evaluator/corpus/model, policy/threshold version, verdict, reason, and completion time. Evidence: result fixtures. <!-- evidence: resultDisplayFaults, uncertaintyBand, marginWithinUncertainty, and CONTROL 9. A VERDICT INSIDE ITS OWN UNCERTAINTY IS NOT A VERDICT. S2.8 keeps the reading and its uncertainty apart precisely so the pair can be read together, and 0.82 ± 0.05 against a threshold of 0.80 is a pass whose margin is smaller than its own noise — invisible in every screenshot, survives code review because the number really is above the threshold, and it is why a gate flaps between pass and block on identical input while everybody looks for the bug in the evaluator. The band is the 95% interval of the MEAN (stddev / sqrt(n) × 1.95996), the standard ERROR and not the standard deviation: the question is how well the run pinned the value down, not how much the population varies, and using the raw stddev would widen every interval by sqrt(n) and caveat verdicts that are perfectly well separated. Strict inequality at the ends, so a band rounded to the threshold does not count as straddling it. Both directions are faults — a flat verdict over a straddling margin, and a hedge on one that clears, because a caveat on every verdict is a caveat on none. Around it: a bare number can be compared with anything, a model name without the provider that served it is a stable label over a thing that changed, and a corpus that gained rows is a different reference set. -->
    • S7.14.c Display fresh/stale/expired/missing/pending/running/unavailable/ unsupported/failed state with invalidating dependency and age. Evidence: freshness matrix. <!-- evidence: GATE_PANEL_STATES, panelStateFor, panelStateFaults, NON_VALUE_TO_PANEL_STATE, STATE_CONTRIBUTES_NON_VALUE, and CONTROLS 1 and 2. THE NINE STATES ANSWER ONE QUESTION — why is there no usable answer, and who can change that — which is what makes them a vocabulary rather than a list. THE PANEL VOCABULARY IS NARROWER THAN THE RECORD'S, SO NOTHING MAY BE DROPPED IN THE MAPPING: this is S7.13's rule from the other side, where five words covered three and the danger was contradiction; here nine cover S2.8's eight non-values plus two freshness reasons, and the danger is LOSS. insufficient-data and failed both land on failed and are two different people's work, so the reading carries sourceNonValue and the mapping is total and is data. A VERDICT MAY NEVER BE PRINTED WITHOUT ITS STANDING — an answer given against inputs that have since moved reads exactly like today's answer, which is the single most common way a gate system stops working while every light stays green — and the browser settles it as a LAYOUT fact: the two are on the same line, within 24px, because the same two strings a scroll apart are a green light and a footnote. A stale row must NAME what moved or it has silently become an expired one. AN IN-FLIGHT RUN DOES NOT RETRACT THE ANSWER ON SCREEN (a row that flipped to running would say there is nothing to read) and the answer does not hide the rerun. -->
    • S7.14.d List evidence refs with type/version/hash/access state and safe navigation. Evidence: evidence tests. <!-- evidence: evidenceListFaults, EVIDENCE_ACCESS_STATES, and CONTROL 5. THE LIST IS OVER WHAT THE DEFINITION REQUIRED, NOT OVER WHAT THE RUN RECORDED — S2.10's second decision, one surface up. A list built from the records in hand cannot distinguish "the consent record was checked" from "the consent record was never asked for", and under that list the way to make a rights gate look thorough is to collect less. digest-mismatch is the fourth access state and the one a naive surface cannot express: the reference resolves, the document is there, and it is not the document the run read — so a link to it says the verdict still rests on what it rested on, which is precisely what the panel cannot know. Only resolved may be an anchor, and the browser proves the degraded ones are not in the tab order by REAL Tab traversal rather than by an attribute (with a companion control that resolved references elsewhere ARE reachable, so the first control is not vacuous). SAFE NAVIGATION MEANS THE PANEL RESOLVES A REFERENCE THROUGH ITS OWN ROUTE: an evidence reference is a string an EVALUATOR wrote, and an anchor built from it is an open redirect with a trust badge on it. -->
    • S7.14.e Display remediation steps, affected fields/objects, documentation, appeal/review, and recompute prerequisites. Evidence: remediation tests. <!-- evidence: remediationFaults, whoToContact, contactFor, recomputePrognosis, and CONTROL 8. THE PANEL'S NINE STATES DO NOT DECIDE WHO TO CONTACT — THE UNDERLYING REASON DOES, and this is the first thing in the kit to READ four of S2.9's five owners. failed from a broken evaluator is an incident for its owner; failed from insufficient input is work for whoever submitted; unsupported is a question for whoever decided this gate applies here at all; a blocking verdict is for the remediation owner, who in practice is neither of the first two. One "contact" line for all of them pages the wrong team and puts a sentence about a platform in front of a person whose problem is editorial. THE RECOMPUTE PROGNOSIS IS THE SECOND FIRST-READER: policy.deterministic existed only to be compared between a catalogue and a contribution, and S2.9's own warning says a non-deterministic gate published as deterministic is one whose re-run is treated as a confirmation. Here that sentence becomes a control — a green DETERMINISTIC check offers no rerun because it would spend the evaluator's budget to reprint the same sentence, and the identical row on a non-deterministic gate DOES offer one. Futility is part of LEGALITY rather than advice, or the panel and its checker disagree about that exact row. Refused both ways: an offer that cannot change anything, and a withheld one where a dependency moved and the reader could have moved it themselves. -->
    • S7.14.f Display waiver/override scope/reason/approver/evidence/start/ expiry/residual risk and out-of-scope/stale status. Evidence: waiver matrix. <!-- evidence: waiverPanelFaults, and CONTROL 10. A WAIVER IS RENDERED ON THE BLOCK IT COVERS, NEVER INSTEAD OF IT. S2.10 already refuses to call waived a pass, and a surface can undo that decision simply by rendering the exception alone — after which no reader can tell a check that passed from one somebody signed for. AND A WAIVER THAT NO LONGER AUTHORISES MUST STILL RENDER (S2.11's fifth decision): one that vanishes takes with it the fact that somebody once thought this was acceptable, which is the fact an incident review is looking for. Both are faults, and so is the middle case — an expired or out-of-scope waiver renders in full, with its approver, its dates and its residual risk, and is not presented as covering anything. The browser drives ?waiver=active and ?waiver=expired over the same row: the first renders waived with the block still on screen beside it, the second renders blocks with the whole exception still readable. The scope has to be printed or an out-of-scope waiver looks like an arbitrary refusal, and a severe accepted risk with nobody acknowledging the escalation is named where S2.11 can only refuse it at write time. -->
    • S7.14.g Authorize and run/recompute/cancel/retry/appeal/waive actions through durable jobs with cost/impact confirmation. Evidence: action journeys. <!-- evidence: gateActionFaults, actionIsLegal, ACTION_FROM_STATES, and CONTROL 7. AN ACTION IS A DURABLE JOB, AND A CONTROL THAT RETURNS A VERDICT IS CLAIMING AN ANSWER IT DID NOT GET — ActionOutcome carries a verdict arm SO IT CAN BE REFUSED. A gate is evaluated by somebody else, later; a panel that flips the verdict on click has invented a synchronous evaluator and rendered a number no durable record holds. Without an idempotency key a double-click is two runs and two bills. The order is S7.13's and is not negotiable — AUTHORIZED, then LEGAL FROM THIS STATE, then confirmed proportionately, then DURABLE — and confirmation is proportionate in BOTH directions: a free action behind a typed confirmation trains everybody to dismiss the one that mattered, and a waiver, which is a signature on a risk somebody else carries, behind a single click is not a decision anybody made. run is only for a gate that has never answered and recompute only for one that has: the same job, two sentences, and a panel offering "run again" over a check that never ran once is describing a history that did not happen. THE BROWSER FOUND A REAL DEFECT HERE — see the rollup. -->
    • S7.14.h Provide keyboard/screen-reader navigation, noncolor status, compact grouping, and no pass-only aggregation hiding blockers. Evidence: accessibility tests. <!-- evidence: panelAccessibilityFaults, panelHeadline, isUnknownRow, and CONTROLS 3, 4, 5 and 6. THE HEADLINE IS BUILT FROM THE COUNTS, NOT CHECKED AGAINST THEM. A rule of the form "the headline must mention the number of blockers" is SATISFIED BY "1 of 3 checks passed" — every count is right there as a substring — and that sentence is exactly the failure this cell exists to stop. So the kit produces the line, the surface renders it, and a panel that writes its own is a disagreement rather than an argument. The order is the content: blockers, then gaps (because a gap is a decision nobody made and a waiver is one somebody signed — S2.10's ELIGIBILITY_STRICTNESS already ranks them that way), then waivers, and the pass count LAST. Zero is SPOKEN. COMPACTING IS FINE; COMPACTING A BLOCKER OUT OF SIGHT IS NOT — and collapsed is MEASURED in the browser rather than declared: the tool presses the collapse control and reads rendered heights, so the claim is about boxes rather than about a boolean, and the group's open is DERIVED so no control can force it shut. Tab order is real traversal. Non-colour status is proved under forcedColors: 'active' — a real rendering in which the browser overrides the palette — with a companion control that the colour channel really was flattened, so the proof is not vacuous. --> <!-- S7.14 rollup: 114 kit tests (107th entry point, 5565 kit tests total) + 28 jsdom tests + 61 real-Chromium controls, each rule with a NEGATIVE control driving the case it exists to catch. DIST-PATCH VERIFICATION: silencing verdict-printed-without-its-state in the BUILT dist failed its negative control, and narrowing UNKNOWN_STATES so an outage stops counting as a gap failed the summary comparison; restoring the dist returned all 61 to green. THE BROWSER FOUND A REAL DEFECT. Clicking "retry" on a row that ALREADY had an answer queued a job — and the row's state did not move, because by design an in-flight run must not retract the answer on screen. So every costly control stayed on offer and a second click was a second job and a second bill. The state was the wrong place to look: PanelReading now carries inFlightJobId SEPARATELY from the state, because they are two different facts and a row has to carry both — the answer keeps standing while a replacement is on its way. recomputePrognosis consults it first, so "wait rather than queue another" outranks every other reason. Source reading could not have found this; the two halves each looked correct. -->
  • S7.15 Build the provenance/lineage inspector. Evidence: ingredients, transformations, hashes, signatures, validation, and gaps are visible.

    • S7.15.a Display exact artifact/revision identity, byte hash/algorithm/ size/type, release refs, and manifest/ledger binding. Evidence: artifact fixtures. <!-- THE DIGEST'S ORIGIN IS THE CLAIM, NOT THE DIGEST. Four different things produce the same sixty-four characters and they render identically: bytes hashed just now, a number copied out of the manifest under examination, a checksum the store keeps for its own purposes, and a value the caller passed in. Only the first is evidence about the bytes on screen; the second is the document agreeing with itself, which is what a tampered manifest is best at, and the third is S6.12's finding — truncate an object in MinIO and it recomputes its own ETag over what it now holds. So EVIDENTIAL_DIGEST_ORIGINS has exactly one member, DIGEST_ORIGIN_LIMITS states what each of the other three cannot settle, and the header carries digestOriginRendered separately because the RECORD always knows where its digest came from and the SURFACE is where that gets dropped. A floating reference cannot head the page at all: S1.4's resolutionContract already says a signature over a channel name commits to nothing, so every panel below it would be about bytes nobody named. Both lengths are on screen because a short read is the corruption real stores produce and the one a self-consistent checksum cannot see. -->
    • S7.15.b Traverse ingredients and derivatives with relation, version/hash, source/origin, rights/consent hint, availability, and redaction. Evidence: lineage graph tests. <!-- LINEAGE IS A DAG, SO A CYCLE IS A DEFECT AND A DIAMOND IS NOT — the opposite call from S7.10, where a cycle was a legitimate graph. Two byte sequences cannot each have been made from the other, and a walk that silently absorbed one would report a finite history for a record that has none; the same photograph placed twice is ordinary and a finder that flagged it would be switched off in a week. lineageCycles is an iterative three-colour DFS so a node already FINISHED is a second path and only a node still ON THE STACK is a way back. AN ASSET HAS AT MOST ONE PARENT, because the parent is whose history this history continues. A DEGREE IS THREE NUMBERS and a concealed neighbour is in NONE of them (S7.10's rule, and S5.3 already named the classes where the existence of the row IS the disclosure) while a withheld one IS counted, because a denial is a refusal somebody can appeal. The walk carries its budget, what it stopped at and what it did not reach — an ingredient list that ran out of budget reads as the complete set of things this was made from. -->
    • S7.15.c Display ordered capture/edit/generate/transform/merge/validate/ approve/export/publish actions and parameters. Evidence: action history fixtures. <!-- THE ORDER IS THE RECORDED ORDER AND A TIMESTAMP NEVER SORTS IT. Action timestamps come from whichever machine performed the action; across a generator in one region and an approval clicked in another they are not comparable, and sorting by them shuffles somebody's history into a plausible-looking wrong order. A clock that runs backwards against the record is REPORTED — it is skew or a fabricated entry, both worth seeing — and it re-sorts nothing. A GAP IS A ROW: positions are contiguous and an absence occupies its position with its cause, because six actions with a hole where the seventh was reads as six actions. parameters distinguishes {} ("ran on its defaults") from null ("the record does not say"), which is the pair a component collapses the moment it writes parameters ?? {}. ACTION_REQUIREMENTS is data: generation needs its model AND its provider, since two deployments of one model version answer differently; a merge names what came together; an approval cannot be a service account; and a publication with its only approval AFTER it is a record of somebody approving what the world already had. -->
    • S7.15.d Display tool/app/plugin/provider/model/template/policy/environment versions and authenticated/delegated actors safely. Evidence: provenance detail tests. <!-- ALL EIGHT SLOTS ALWAYS RENDER (S7.1's rule: a vocabulary that cannot say unknown spells it as absence, and absence renders identically to fine). A panel showing five versions because three were not recorded has told the reader five things were involved. And the three absences are three different sentences: not-recorded is work somebody has to finish, not-applicable is a claim about this pipeline that must say why, and an EMPTY CELL is both at once — which is the fault that costs the most and shows the least. describeActor is the first surface reader of S1.4's refusal of display data in an identity: a signed provenance record is permanent and travels with the asset, so a name or an address in it cannot be un-published. A delegated act answers two questions and a line carrying one of them answers neither; a cross-tenant service has NO tenant rather than one nobody recorded; and a redacted actor is "identity removed (cause)", never "unknown". -->
    • S7.15.e Display signature/binding/chain/trust/revocation/timestamp/ assertion/ingredient validation independently. Evidence: validation state matrix. <!-- COMPOSE S6.7, DO NOT RE-DECIDE IT. content-credentials.ts already published the ten checks, the six outcomes and the nine states; S7.15 is the presentation, and the presentation is where the boolean comes back. EVERY CHECK RENDERS ITS OWN OUTCOME AND THE STATE NEVER RENDERS WITHOUT THE CHECK THAT DECIDED IT — CredentialStateResult has carried decidedBy since S6.7, so a surface has no excuse. A check with no result renders not-performed rather than disappearing, because an absent row reads as no problem found. Revocation and the hard binding are PRIMARY_CHECKS: an unreachable revocation responder silently becomes a trust decision for every asset validated during the outage, and that rule is held at the PRESENTATION layer too, since the honest outcome beside a sentence reading "not revoked" is the string a reader takes away. ingredient-availability: partial over forty ingredients tells nobody anything — the useful fact is which. -->
    • S7.15.f Display missing/stripped/unavailable/unsupported/untrusted/ invalid/ambiguous gaps with recovery method and limitations. Evidence: gap fixtures. <!-- RECOVERED IS NOT BOUND. When a channel has stripped the credential a perceptual match finds a record for bytes that LOOK like these, and rendering it the way a hard binding is rendered asserts the one thing a hard binding exists to establish. recoveryWorth composes S6.8's outcomes: only exact-hard-match states these bytes, and a match to a REVOKED or DELETED record is none rather than weak — a weak yes is how a withdrawn registration keeps answering. Only stripped gets a method that produces a candidate; the other six resolve to none deliberately, because "retry when the responder is back" is waiting, and calling it a recovery method makes an outage look like a procedure. stripped REQUIRES AN EXTERNAL WITNESS: missing and stripped are byte-identical from inside the asset, so the second word accuses something of removing the credential and needs a ledger record or a measured channel profile behind it. And ambiguous is two answers, not no answer — showing one of them is choosing at random. -->
    • S7.15.g Provide semantic table/tree parity, bounded graph traversal, protected-source redaction, keyboard navigation, and evidence export authorization. Evidence: accessible/security tests. <!-- THE PARITY HERE IS SYMMETRIC, WHICH INVERTS S7.10 AND S7.11. There a coordinate stated nothing and the picture was allowed to show less than the table; here the NESTING IS THE RELATION — indentation is the claim "this was made from that" — so the two views are two renderings of one statement and any difference is one of them being wrong. Both are built from ONE statement string, because two independently assembled sentences drift. The order check is the one people leave out: a table sorted by name over a tree in walk order gives two readers two histories, each internally consistent. THE EXPORT IS BUILT FROM THE READING AND AUTHORIZED SEPARATELY FROM THE VIEW: what may be read on screen inside a tenant-scoped session is not what may be handed to somebody outside it, and an export that re-queries runs with the exporter's authority over a fresh answer, so it can hold rows the reader never saw and is not covered by the redaction the surface applied. Counted at a real node:http server: pressing export made ZERO requests. And where the object is concealed there is no control at all, because a control that refuses admits the object exists. --> <!-- S7.15 rollup: 177 kit tests (108th entry point, 5743 kit tests total) + 36 jsdom tests + 96 real-Chromium controls, each rule with a NEGATIVE control driving the case it exists to catch, built from what the browser rendered. DIST-PATCH VERIFICATION both ways: widening EVIDENTIAL_DIGEST_ORIGINS in the BUILT dist so a manifest-supplied digest counts as evidence failed its control, and silencing concealed-node-rendered-as-a-row failed its; restoring the dist returned all 96 to green. THE BROWSER CORRECTED A MEASUREMENT AND FOUND A FIXTURE THAT PROVED NOTHING. The first version of the adjacency check asked whether the digest and its origin were OVERLAPPING BOXES, and at 1280px a sha256 is 71 characters that fill their own line, so the origin lands on the line below and the check called one statement two. What actually separates a statement from a footnote is the BLOCK and the DISTANCE — same dd, gap under one line — and detaching the origin into another block in the real document breaks both, which is what makes the pair load-bearing. Separately, the "an unasked check still has a row" control was answered by a fixture that answered all ten, so it was measuring nothing; it now drives a page with a check genuinely missing. Property tests over fixed-seed generated DAGs check the walk's depth against an independent shortest-path relaxation and the cycle finder against graphs that cannot contain one. -->
  • S7.16 Build job and queue views with progress, dependency graph, logs, cost, retry, cancel, and failure detail. Evidence: live fault-injection journey.

    • S7.16.a Build permission-aware queue list with job type/version, target, priority, state/stage, queued age, progress, owner, resource/provider, cost, deadline, and stable pagination/filtering. Evidence: queue tests. <!-- queued COVERS TWO THINGS ON SCREEN AND ONLY ONE OF THEM IS A QUEUE. A job whose runnableAt is in the future is waiting for a CLOCK, not for a worker, and rendering it beside the ones that are waiting for a worker is how a queue of four looks like a queue of forty and the worker pool gets the blame. So RUNNABLE_STATES is WIDER than S3.9's six (S7.13's rule) and RUNNABLE_STATE_SOURCE maps every one of its eight members back to exactly one record state, so it can say more and can never contradict. The other addition is stalled: A LEASE THAT HAS LAPSED IS NOT A WORKER, and "running on worker-7" over a job nothing is holding is the single most misleading cell a job console can draw — the state is true, the worker is true, and both are statements about the past. runnableStateOf consults S3.9's own workerHoldsLease, so the clock-skew margin comes with it. AN AGE IS READ AT COMPARISON TIME (S7.3's finding): the surface ticks its own clock, because a now captured at mount makes every age wrong by however long somebody has been looking, and the jobs queued longest are the ones a decision is about to be made about. A DEADLINE IS A STANDING, never an instant — the reason a deadline is on the row is to say whether it has passed. Keyset over (priority, enqueuedAt, jobId), because an offset over a draining queue skips and repeats exactly the rows being worked on. -->
    • S7.16.b Build job detail with immutable inputs, attempts, lease/worker, timing, checkpoints, outputs, provenance, audit/correlation, and safe support refs. Evidence: detail fixtures. <!-- THE INPUTS ARE A REFERENCE AND A HASH, NEVER A COPY. A panel holding a copy is a copy that can disagree with the reference, and it is the one the reader believes, on a page anybody with the job id can open; "immutable" is a promise until the hash beside it checks it. EVERY ATTEMPT, not the last: three timeouts then a validation error is a different story from one validation error and only the second is the submitter's. And A SUPPORT BUNDLE THAT CARRIES THE INPUTS copies a customer's data into a ticketing system with a different retention policy and a different audience, at the exact moment nobody is thinking about retention — so SupportReference carries carriesInputs and carriesSecrets as facts the checker reads rather than as promises. -->
    • S7.16.c Visualize parent/child/dependency/critical-path graph with semantic table parity and required/optional/blocking state. Evidence: graph parity tests. <!-- REQUIRED, OPTIONAL AND BLOCKING ARE THREE FACTS AND ONLY ONE IS ABOUT NOW. Required is the edge's contract, optional is what may fail without failing the parent, blocking is what is holding the parent up at this moment; a graph that draws the first and the third alike makes every ordinary dependency read as an incident. THE BROWSER SETTLED THE FOURTH: FINISHED IS NOT SATISFIED. A required child that was dead-lettered has finished — it is holding nothing up — and it has not met the requirement and never will, so drawing it satisfied says the parent is fine at the moment the parent is about to fail. SATISFYING_STATES is one member beside FINISHED_STATES, written down next to it precisely because the two are so easy to conflate. The parity here is ASYMMETRIC, unlike S7.15's lineage tree: a job graph is drawn with a layout and a coordinate states nothing (S7.10), so a bounded picture may show fewer nodes than the table and only the reverse is a fault. And the critical path is a PREDICTION from expected durations that moves as the run proceeds, so it says so or it reads as the route the work took. -->
    • S7.16.d Display completed/total/indeterminate progress by stage and aggregate without invented percentages or false completion. Evidence: progress fixtures. <!-- THE BROWSER FOUND A REAL DEFECT HERE AND jsdom AGREED WITH THE BUG. The fill element carried no inline size when the fold was indeterminate — which is correct as an intention and wrong as CSS, because a block-level child with no declared width FILLS ITS CONTAINER. The honest "we do not know how far along this is" rendered as a FULL BAR, the exact reading an unknown denominator must never produce. In jsdom the inline style is a string and the assertion that it was absent passed; in a real browser the box was 1244px wide. inline-size: 0 is now the default and the determinate case overrides it, so the only way to get a filled bar is to have a fraction to fill it with. A FULL BAR IS ALSO NOT COMPLETION: S3.9 deliberately produces no done from the progress table because completion is a STATE decided by whatever commits the last effect, and a worker that dies after its final report leaves every counter at its total — the job that never wrote its result is then one nobody is looking for and its outputs are ones nobody deletes. -->
    • S7.16.e Display privacy-safe structured logs/events with levels, timestamps, attempt/stage, filters, bounded streaming, download policy, and no content/secrets. Evidence: log security tests. <!-- A LOG LINE IS THE LEAKIEST SURFACE IN THE PRODUCT. It is written by a handler, in a hurry, about the thing that just went wrong — which is usually the input. So the console renders STRUCTURED events over a declared field ALLOW-LIST (S5.5's rule for response headers, one level up) and the renderer walks the list rather than the event, so a handler that attached the prompt gets it dropped at the surface rather than at the reviewer: measured in the browser as the string never appearing in the rendered document at all. A DROPPED LINE MUST BE VISIBLE — an unbounded tail is a leak in the browser and a load on the server, and a bounded window that discards silently renders as a shorter, calmer log, so an operator reading it during an incident concludes nothing much happened in the gap. And a download is a SECOND authorization, because a file that leaves the system is a different question from a panel that is read inside a session. -->
    • S7.16.f Display estimated/reserved/actual/refunded/variance cost and resource usage with units/rate version. Evidence: cost fixtures. <!-- A REFUND IS NOT A NEGATIVE COST AND A VARIANCE IS NOT A NUMBER ANYBODY STORES. Netting a refund against a charge loses that money moved twice — the only thing an accounts query cares about — and makes a job that was charged and refunded indistinguishable from one that was never charged. costVariance derives it, and a stored variance that disagrees with the two numbers it came from is the finding; it REFUSES across currencies and across scales rather than subtracting, because two amounts in different currencies have no difference without a rate and inventing one would put a made-up exchange rate on an accounting screen. An estimate is an INTERVAL (S5.7's rule) and a point is a claim nobody made. Every amount carries its rate version or it cannot be re-derived, and the day a price changes there is no way to tell a job that cost more from one that was priced differently. -->
    • S7.16.g Authorize cancel/retry/resume/replay/abandon with impact, idempotency, confirmation, eligible scope, and result. Evidence: control journey tests. <!-- ONE PREDICATE decides what is offered and what a press may do (S7.14's rule), so a withheld control and a checker calling it withheld cannot come from two opinions. A CONTROL PRODUCES A JOB, NEVER A STATE: a console that sets a job to queued has decided what a scheduler decides, with none of the fencing, budget or idempotency that goes with it, and the job it "retried" is now competing with whatever the last worker is still doing — measured in the browser as the row not moving and the server seeing zero requests. Impact is part of the offer, because cancelling a parent cancels its children and the confirmation has to say how many; confirmation is proportionate in BOTH directions (S7.6), since over-confirming trains everybody to dismiss the one that mattered. A bulk control freezes its scope AND the count at the freeze (S7.5). -->
    • S7.16.h Display failure class/reason/retained outputs/retryability/ remediation/dead-letter/unknown outcome without stack leakage. Evidence: fault-injection browser journey. <!-- unknown-outcome IS THE MOST IMPORTANT THING A FAILED JOB CAN SAY AND THE ONE MOST OFTEN RENDERED AS AN ORDINARY FAILURE. Every other kind tells the reader what did not happen; this one tells them nobody knows. S3.9 already carries unknownOutcome on the classification and this is its first reader — and controlIsLegal consults it, so with an at-most-once effect the retry is not offered at all, because retrying is how one payment becomes two and doing nothing is how it becomes none. The browser drives BOTH deliveries over the SAME failure and gets two different control sets, which is the whole point of the clause. A RETAINED OUTPUT is on screen and DECLARED PARTIAL, or a stage that did not finish reads as a result and nobody deletes it. And a stack frame is a leak with three payloads — a path names the deployment layout, a module frame names a version somebody can look up an advisory for, and an interpolated frame sometimes carries the value — so the reader gets a class, a reason KEY and a correlation id. --> <!-- S7.16 rollup: 142 kit tests (109th entry point, 5886 kit tests total) + 49 jsdom tests + 88 real-Chromium controls, each rule with a NEGATIVE control driving the case it exists to catch, built from what the browser rendered. DIST-PATCH VERIFICATION both ways: widening SATISFYING_STATES in the BUILT dist so a dead-lettered required child counts as met failed its control, and silencing the ready/deferred split failed its; restoring returned all 88 to green. THE BROWSER FOUND TWO REAL DEFECTS AND jsdom AGREED WITH ONE OF THEM. (1) The indeterminate progress fill had no inline size, which is correct as an intention and wrong as CSS: a block-level child with no width fills its container, so "we do not know" rendered as a full bar. jsdom asserted the inline style was absent and passed; the real box was 1244px. (2) edgeEmphasis checked FINISHED before REQUIRED, so a dead-lettered required child was drawn satisfied — the parent is about to fail and the graph said it was fine. Only success satisfies; a required edge whose child finished without succeeding falls back to required, because the requirement stands and is now unmeetable. Also corrected a measurement: a percentage inline size resolves against the CONTENT box, so comparing the fill against the bar's border box fails by exactly the border (S7.11's max-width: 100% finding from the other side). -->
  • S7.17 Build audit history with safe redaction and export. Evidence: authorization tests distinguish ordinary users and auditors.

    • S7.17.a Build immutable chronological/correlated event query by target, actor, action, outcome, revision, policy, job/release, and time with stable pagination. Evidence: query tests. <!-- THE QUERY IS ITSELF A DISCLOSURE AND ITS DIMENSIONS ARE NOT ALIKE. "What happened to this object" is ordinary; "what did this person do" is surveillance of a colleague, and it is the same screen with a different field filled in. So SENSITIVE_DIMENSIONS names it and the surface renders it as a WORD rather than a disabled control — measured in the browser as not a button and not in the tab order, because a disabled control teaches its reader that this console refuses things. THE SEQUENCE IS THE ORDER AND THE TIMESTAMP IS A FIELD (S7.16's lesson, one surface on): at comes from whichever node wrote the record. And PAGING BACK THROUGH AN APPEND-ONLY LOG IS STABLE BY CONSTRUCTION while paging forward is not — the exact opposite of S7.16's queue, where every row is moving; nothing behind the cursor can change, so only the forward direction has to say there is more. -->
    • S7.17.b Display actor/delegate, tenant-safe context, action, prior/new revision, reason, decision/policy, result, evidence refs, and correlation without copying protected content. Evidence: event fixtures. <!-- S4.12 made "without copying protected content" STRUCTURAL: a mutation carries a parameterDigest and there are no parameters to copy, and a refusal has NOWHERE to put a new revision. Both promises are broken later in exactly one place — a console that joins to the request store and offers a disclosure triangle, and a row that prints a dash where a new revision would go. So the checker looks for an expand control and the browser asserts there is no <details> element anywhere in the rendered document. A single-tenant history prints no tenant column, because a column of one repeated value trains people to stop reading the context line. -->
    • S7.17.c Group request/command/audit/event/job/provider/publish chains and preserve raw event drilldown for authorized roles. Evidence: correlation view tests. <!-- COMPOSES S5.4's THIRTEEN LINK KINDS rather than inventing a second vocabulary. purged AND missing-linked-event ARE DIFFERENT ABSENCES and on the wire they are the same shape — a link with no record behind it — separated by exactly one thing, a tombstone. absenceOf is that one thing as a function, and the two faults point opposite ways: rendering a missing link as purged turns a broken write into a retention policy, and rendering a purge as missing sends somebody to investigate a deletion that was requested. The raw event is what the projection was built by removing things FROM, so opening it is a separate question from reading the projection. -->
    • S7.17.d Apply object/property, protected-source, learner/biometric, rights, tenant, support, legal-hold, and role redaction at query/result. Evidence: ordinary-user/auditor negative tests. <!-- REDACTING AT THE RESULT ALONE IS AN ORACLE. A query that may FILTER on a field the reader may not READ answers the question with its COUNT: ask for events where the protected source is X, get seven back, and the redaction has been defeated without a single value crossing the wire. So AXIS_APPLIED_AT is data and three of the eight axes are both, and a concealed field is refused BY NAME in the same words an unknown one gets (S7.4's rule, one layer up) so the error is not a schema oracle. A LEGAL HOLD WIDENS RETENTION AND GRANTS NOTHING: it stops a purge, it says nothing about who may read what it preserved, and a console that treats "under hold" as "available to the litigation reader" has invented an authorization out of a retention rule — over exactly the records somebody has already decided are contentious. Support access is time-boxed AND itself audited, in this log. -->
    • S7.17.e Distinguish redacted, inaccessible, retained, purged, integrity- invalid, late, and missing-linked-event states. Evidence: state fixtures. <!-- SEVEN STATES, SEVEN WORDS AND SEVEN RECOURSES, all asserted DISTINCT — the vocabulary has seven members precisely because the reader's next move differs, and a surface that shares a sentence between two of them has thrown that away. redacted and inaccessible are the pair that costs the most: the first is gone for everybody and the second is gone for you, and one word for both sends somebody to request an access grant for a value that no longer exists. That check fires only when BOTH are on screen, because that is the only arrangement in which a reader could have told them apart and was not given the chance. A purge carries its tombstone or it is indistinguishable from a write that did not land; a late entry says where it belongs, because a reader scanning in order will not see it where they expect it. -->
    • S7.17.f Export a versioned checksummed snapshot with query scope, generated-by/time, schema, redaction policy, page completeness, and access controls. Evidence: export tests. <!-- PAGE COMPLETENESS IS THE FIELD EVERYBODY FORGETS AND THE ONE THE EXPORT IS FOR. A snapshot that hit a limit and stopped is a partial history in a file with a checksum on it, and everything about the way it looks says complete — S7.15's truncation rule applied to the artifact that gets filed as evidence. THE REDACTION POLICY VERSION TRAVELS WITH IT for the slower version of the same problem: the same query under a later policy returns different rows, so two exports of one unchanged history look like the history changed. Authorized separately from the view and labelled a copy, per S7.15; built from the rows on screen, proved by a real node:http server seeing ZERO requests when the button is pressed. -->
    • S7.17.g Verify audit chain/signature/hash integrity and surface failures without hiding remaining history. Evidence: tamper tests. <!-- A BROKEN CHAIN MUST NOT HIDE THE REST OF THE HISTORY. The failure this exists to prevent is a console that verifies, finds a break at entry 400 and shows nothing — it has answered "is this log intact" by making the log unreadable, at the exact moment an investigator needs the 399 entries that ARE intact and the ones after it that are still evidence of something. Measured in a real browser as every entry on both sides of the break having a non-zero box, and the ones beyond it carrying a label; the label survives forcedColors: 'active' because it is a border STYLE, with a companion control proving the colour channel really was flattened. THREE QUESTIONS, ASKED SEPARATELY — a hash that does not match, a link that does not follow and a signature nothing covers are three different incidents. And AN UNVERIFIABLE ENTRY IS NOT AN INVALID ONE: no checkpoint covers it yet, and reporting that as a failure produces an incident every time the log is read faster than it is signed. --> <!-- S7.17 rollup: 93 kit tests (110th entry point, 5980 kit tests total) + 29 jsdom tests + 69 real-Chromium controls, each rule with a NEGATIVE control driving the case it exists to catch, built from what the browser rendered. DIST-PATCH VERIFICATION both ways: emptying SENSITIVE_DIMENSIONS in the BUILT dist so an actor query needs no grant failed two controls, and weakening AXIS_APPLIED_AT so protected-source needs the result layer only failed two more; restoring returned all 69 to green. THE HARNESS TRAP WORTH RECORDING: the first break was driven at the LAST sequence in the fixture, so nothing was ever beyond it and the "entries after the break are labelled" control was answered by an empty set. A break needs entries on BOTH sides to be a break at all, and the control now asserts that first. Also hit the tsc-bypass trap: building the dist with npx tsc --outDir (Nx was broken by worktree duplicates) does not copy package.json, so packed-consumer-smoke refused the new subpath until it was copied across. -->
  • S7.18 Build bulk-action preview and result surfaces. Evidence: mixed eligibility, partial failure, cancel, and retry journeys. <!-- COMPOSE S7.5 + S4.12 + S3.9 + S7.16 + S7.6, DO NOT RE-DECIDE THEM. Seventh sighting of the S7.12 rule: S7.5 published the selection (and the rule that "everything matching" is a frozen QUERY plus the count at freeze), S4.12 the per-item result and the arithmetic that makes partial real, S3.9 the job states, failure kinds, delivery guarantees, cancellation and the noncancellable region, S7.16 which failures are the caller's to fix, S7.6 the risk table and the confirmation bands. The kit is large now; look before writing. -->

    • S7.18.a Freeze selected IDs or versioned query and show selected, currently matched, hidden/inaccessible, and changed-since-selection counts. Evidence: concurrent selection tests. <!-- A FROZEN QUERY IS A FROZEN ANSWER, NOT A SAVED SEARCH. The reason to freeze is that the preview and the execution must be about the same set, so a surface that re-runs the query when the actor presses go acts on objects created after the preview was read — never counted, never checked for eligibility, never in the confirmation. It is re-evaluated to REPORT the difference and never to widen the action, and a query freeze with no snapshot token cannot even do that, because its answer is unreproducible. "CHANGED SINCE SELECTION" IS FOUR EVENTS, and the one a query re-run can never find is the object that STILL MATCHES and has been rewritten end to end — so the detection mode is a field rather than an implementation detail. Hidden-by-your-filter and not-yours are TWO counts, because one clears and the other does not, and one line for both sends somebody to clear a filter that will never bring the rows back. AND THE SUBSETS OVERLAP, SO A PRINTED REMAINDER IS ARITHMETIC THAT DOES NOT HOLD — including when it ADDS UP. revision-moved sits inside currentlyMatched as well as changedSinceFreeze and no-longer-visible inside inaccessible, so a clean sum can be a double-count cancelled out by something the categories missed. That is why the rule has a membership half as well as an arithmetic one, and why the browser tool drives the membership half separately: the rendered counts already trip the arithmetic one, and an untested half is one a patch to the built kit walks straight through (verified: it did). -->
    • S7.18.b Display eligible/ineligible/unknown/already-satisfied items with object-specific authorization/state/dependency/policy/capacity reasons. Evidence: eligibility fixtures. <!-- A REASON BELONGS TO AN OBJECT, NOT A BUCKET. "17 ineligible: insufficient permission" is the failure this exists to prevent — the reader fixes the one reason on screen, presses go, and watches sixteen refuse for four other reasons. The five reason kinds differ in what the reader does next, so only the one whose remedy is WAITING may carry a clock: a retry-after on a policy refusal is an invitation to hammer a decision no clock changes, and on an authorization refusal it tells somebody their access is about to arrive. A CONCEALED ITEM IS NOT AN INELIGIBLE ITEM, because a refusal is a disclosure. already-satisfied is its own bucket and is not a success: nothing will change, so counting it as completed claims work nobody did, but acting anyway is not free either — the audit record, the notification and the bill are not idempotent. And unknown is an answer: the surface says which way it will be treated, and attempting the unknowns under an action that cannot be undone is the one combination that is never allowed. -->
    • S7.18.c Preview resulting state, dependencies, gates, notifications, external effects, cost, irreversibility, rollback, and confirmation requirement. Evidence: impact fixtures. <!-- NOTIFICATIONS ARE COUNTED BY RECIPIENT, NEVER BY ITEM. Five hundred objects that all notify one person is one interruption; five hundred that notify five hundred people is an event, and reporting the object count misinforms in whichever direction the fan-out runs. A ROLLBACK HAS A SCOPE AND AN EXTERNAL EFFECT IS OUTSIDE EVERY SCOPE — restoring the internal state does not unsend the mail — so a preview may only be presented as reversible when nothing in the set carries one, and irreversibility over a set is decided by the STRONGEST member. A SINGLE HEADLINE OVER TWO STARTING STATES IS TRUE OF HALF OF THEM: the other half takes a different edge with different guards, so a preview that cannot say one thing says several. And the confirmation is S7.6's own table over what will ACTUALLY BE ATTEMPTED — confirming over 500 selected when 100 are eligible over-confirms, and over 100 when the 40 unknowns will also be tried under-confirms. Measured in the browser: a twenty-thousand-object SELECTION does not escalate the band, because three objects are eligible either way. -->
    • S7.18.d Execute through durable job and display pending/running/ completed/failed/skipped/cancelled per-item and aggregate progress. Evidence: live execution journey. <!-- THE SUMMARY MAY NOT BE PRODUCED WHILE THE RUN IS STILL MOVING. S4.12's summarizeBulk is arithmetic with no notion of time: handed a run whose first object is still pending it counts zero successes and reports failed, and handed one half done it reports partial — both settled-sounding verdicts about something unfinished, and the surface would print them in large type. foldRun returns null instead, which is the whole reason to compose S4.12 through a function rather than call it. The mapping from the six presentation states back to S4.12's three is DATA, because a wider vocabulary is where the narrower one loses a word: skipped and cancelled are both refused, which is neither a success nor a failure. skipped also carries WHEN the decision was taken — skipped-at-preview was refused before anything started and the reader saw that, while skipped-at-execution stopped qualifying mid-run, and it is the only evidence anybody gets that the world moved after the preview was approved. Two origins under one sentence have been collapsed on screen however carefully the record kept them apart. And the progress denominator is the FROZEN count: a denominator discovered as the run goes makes the bar go backwards, which an operator reads — correctly — as "something is wrong", about the wrong thing. -->
    • S7.18.e Allow cancellation of unstarted work, explain in-flight behavior, and preserve completed results. Evidence: cancellation journey. <!-- CANCEL IS NOT UNDO, AND IT IS THREE ANSWERS RATHER THAN ONE. Stop what has not started, ask what is running, leave what finished alone. A control whose label or description says "undo", "revert" or "roll back" is read by somebody who needs the first two hundred objects back, and gets a stopped queue and no reversal — so the surface states that finished work stays finished, and the browser asserts those words appear nowhere in the panel. THE IN-FLIGHT ANSWER IS READ FROM S3.9's CANCELLATION POINT, NEVER GUESSED: an object inside a noncancellable region runs to completion because that is what the region is for, and one whose region has overrun its own bound is escalated and fenced because the cooperative cancel has already failed. Measured in the browser over one click: three different dispositions, the completed row not moving, and the results still on screen — a cancelled run whose results vanish is the worst of the three possible outcomes, because now nobody knows what happened. -->
    • S7.18.f Allow retry only for eligible failed items with original approved parameters/idempotency and refreshed preconditions. Evidence: retry duplication tests. <!-- THE TWO CONSTRAINTS PULL IN OPPOSITE DIRECTIONS AND BOTH ARE REQUIRED. The PARAMETERS must be the ones the actor approved, because re-reading them from the form lets an unapproved edit ride in on an old approval. The PRECONDITION must be REFRESHED, because the if-match captured before the first attempt is about a version that may no longer exist — reusing it either fails forever or, worse, succeeds against a revision the approval was never about. And the idempotency key is the SAME key, or the retry is a second action wearing the first one's name. FUTILITY IS PART OF LEGALITY (S7.14's rule): a retry over a handler-defect is a second identical failure, a second audit record and a second bill, so the surface withholds the control and says the refusal in its place rather than offering it with a warning. Retryable composes S7.16's CALLER_FIXABLE_FAILURES plus the transient kinds; unknown-outcome is decided by the DELIVERY, and only at-most-once withholds it. Proved at a real server: one request per retryable object, each carrying the key the first attempt claimed, and no request for anything that succeeded. -->
    • S7.18.g Export result/evidence and provide keyboard/screen-reader traversal of mixed status without color-only meaning. Evidence: accessibility tests. <!-- AN EXPORT THAT LISTS ONLY WHAT HAPPENED IS FILED AS WHAT THE BULK DID. A hundred completed rows is true, and the four hundred the run declined to touch are exactly the part somebody comes back for — S7.17's completeness rule pointed at the artifact that leaves the building. It is built FROM THE READING, which is a claim about what the page did NOT fetch: measured as zero requests at a real node:http server and zero at the browser. STATUS IS CARRIED BY A BORDER STYLE, NOT A COLOUR, measured under forcedColors: 'active' with the companion control that makes the proof non-vacuous — two states computing the SAME border colour, so the colour channel really was flattened, while the styles stay distinct. Every row states its word in its accessible name, has a box with a height, and is reachable by a real Tab walk; and a MIXED list offers a way to reach the failures directly, because three failures among five hundred rows are reachable the way a needle is reachable in a haystack. --> <!-- S7.18 rollup: 127 kit tests (111th entry point, 6108 kit tests total) + 46 jsdom tests + 104 real-Chromium controls, each rule with a NEGATIVE control driving the case it exists to catch, built from what the browser rendered. THE BROWSER FOUND A REAL DEFECT THE SOURCE COULD NOT SHOW: a concealed object leaked through the DRIFT list. The eligibility list filtered it out correctly and the drift row named it anyway, under the one label that makes the leak sound like a courtesy — "no longer yours to read". Decisions 4 and 5 look contradictory until you see they are about different objects: one was disclosed and then withdrawn, the other was never disclosed at all. disclosedItemIds is now an argument and the surface filters both lists through one predicate. Its COUNT had the same problem: a changed-count of 2 beside a list of 1 says "and one more you may not see", so the panel renders the disclosed count. AND THE TOOL FOUND TWO DEFECTS IN ITSELF, both of which had been reporting green. React renders <button data-harness-after> as data-harness-after="true" and never as the empty string, so the tab walk compared against '' , walked the whole document, wrapped, and reported that the order never leaves the surface. And the 360px negative control measured the ROOT while the overflow is taken by the panel's own overflow-x: auto — restated as three measurements (the container took it, the document did not move, and the container SCROLLS rather than clipping) it is S7.1's finding as a measurement. DIST-PATCH VERIFICATION, THREE WAYS: emptying DRIFT_KINDS_ALSO_COUNTED_ELSEWHERE failed one control (but only AFTER a control was added that drives the membership half — the first attempt stayed green because the rendered counts already tripped the arithmetic half, which is the S7.10 lesson landing for a sixth time); widening RETRYABLE_FAILURES with handler-defect failed one; emptying TEMPORAL_INELIGIBILITY_REASONS failed two. Restoring returned all 104 to green. -->
  • S7.19 Build notification center and preference entry points. Evidence: unread counts reconcile with server state. <!-- THE FIRST DECISION IS WHAT THIS MODULE IS NOT. libs/veritas/notifications already owns the delivery system — the channels, providers, categories, urgencies, quiet hours, digest frequencies and delivery statuses. A second copy of any of those lists in the kit would compile, would be convenient, and would be wrong within a release, so the taxonomy ARRIVES as a NotificationTaxonomy, exactly as S3.2's role model arrives in actors.ts. The kit's contribution is the DRIFT, not the list. (It is also what the buildable-package rootDir constraint requires: this package cannot import that one.) -->

    • S7.19.a Build permission-aware paginated notification list with event type, urgency, domain/workbench, actor-safe summary, target, time, read/archived state, and action. Evidence: list tests. <!-- THE SUMMARY IS THE LEAK, NOT THE LINK. A notification summary is the one string in the product that also lands in a push payload, an email preview and a lock screen, so it is a TEMPLATE plus FIELDS with a disposition each, and renderSummary is the only place substitution happens — a surface that builds the sentence itself has already had the chance to put the value in. AND THE DOOR AND THE NAMEPLATE ARE DECIDED SEPARATELY. Whether the reader may OPEN the target is a different question from whether they may know what it is CALLED, and conflating them is how a summary that carefully withholds a document title sits beside a target line that prints it. The jsdom spec found exactly that: every surface that renders one fact twice checks one of them. targetLabelDisposition is the second question and withheld-value-repeated-outside-the-summary is the cross-check. READ STATE IS NOT A SORT KEY. Sorting unread first means marking one read makes it jump out of the page and pulls an unseen one in, so the reader loses their place by doing the one thing the list is for. -->
    • S7.19.b Group/deduplicate event storms while preserving distinct assignments, decisions, blockers, failures, mentions, and security/safety alerts. Evidence: grouping fixtures. <!-- SEVEN KINDS MAY NEVER BE COLLAPSED, and it is data rather than a threshold, because the argument is never about how many — it is about which. Each is a notice whose whole value is that one person sees one of them, and "and 47 others" is exactly how the one that mattered disappears. A group also states its distinct-TARGET count, since fifty notices about one object and fifty about fifty objects are different situations that one number cannot tell apart. DEDUPLICATION IS NOT GROUPING. A second delivery of one event is not a second thing that happened, and an event every delivery of which was suppressed was never counted at all — the invisible half of the at-least-once problem S3.7 exists for, and the half that has no symptom. A group's size is therefore over DISTINCT EVENT IDS. -->
    • S7.19.c Implement mark read/unread, mark group/all, archive/unarchive, delete/retention, and target navigation with optimistic rollback. Evidence: lifecycle tests. <!-- BOTH HALVES OF A ROLLBACK ARE EASY TO WRITE AND ONE IS EASY TO FORGET. A failed mark-read that restores the row and not the badge leaves two numbers on one screen disagreeing forever, which the reader resolves by ignoring the badge — and the badge was the only thing that would have brought them back. Faults in both directions, plus the refusal that was kept, plus a rollback nothing explained (a row that silently springs back reads as a bug in the page rather than a refusal). A MANDATORY NOTICE MAY BE READ AND NOT LOST: archive and delete are refused, mark-read is not. A delete names the retention rule that permitted it, or nobody can say whether it should still have been there. And a sweep must state its SCOPE: "mark all read" over a filtered list that clears the whole inbox destroys, in one press, exactly the notices the filter was hiding — the set the reader was not looking at. -->
    • S7.19.d Reconcile unread counts across tabs/devices/reconnect/duplicate delivery and never infer read state only from opening the center. Evidence: real-time count tests. <!-- OPENING THE CENTRE IS NOT READING. It is the most common defect in this surface and the most damaging, because the badge is the only thing that would have brought the reader back to the one that mattered. Measured in a real browser as a real document load: the badge survives the mount, every unread row is still unread, and nothing is written to the server just for looking. THE COUNT IS RECONCILED, NEVER ACCUMULATED. A reconnect resyncs rather than replaying a delta — a delta across a gap counts whatever arrived twice and misses whatever did not, and the whole point of a reconnection is that the client does not know which. The server wins, and a disagreement is REPORTED rather than silently overwritten, because it means a local action was lost and overwriting it quietly destroys the only evidence of that. Two real browser tabs are opened and compared. -->
    • S7.19.e Expose preference entry by event/category, urgency, domain/ project, role, channel, locale/time zone, quiet hours/digest, and nonmutable mandatory notices. Evidence: preference tests. <!-- A PREFERENCE FOR A CHANNEL THE DELIVERY SYSTEM DOES NOT HAVE IS A SWITCH THAT DOES NOTHING, and this is the drift check the taxonomy argument exists for. A switch labelled "SMS" over a deployment with no SMS provider reads as a promise, and the reader who turns it ON is now expecting messages that will never arrive — which is worse than the reader who turns it off. A MANDATORY NOTICE IS A WORD, NOT A DISABLED SWITCH. A control that silently refuses is worse than none, because the reader has been told they turned it off; measured in the browser as not a button, not an input, and not in the tab order. QUIET HOURS DELAY, THEY DO NOT SUPPRESS — dropping is a data-loss bug wearing a preference's clothes, and the person who set it asked for later, not never. The mandatory ones pierce them, which is what quiet hours are FOR: the things that can wait only get to wait if the things that cannot, do not. And the window is read in the READER'S time zone; the server's zone is the setting nobody notices is wrong until they are woken up. -->
    • S7.19.f Show notification capability/degraded/delivery failure state and protect source/content/learner/biometric/private target detail. Evidence: state/security tests. <!-- A DELIVERY FAILURE IS ABOUT ONE CHANNEL AND MUST NAME IT. The banner reading "notifications are degraded" while only SMS is down is read by everybody, is true for almost nobody, and is dismissed — and so is the next one, on the day the whole thing is down. A global banner is honest only when every channel is impaired, which the browser drives both ways. Naming a channel that is working is the other half of the same failure. The protection half composes S5.3 directly: a field in one of the PRESENCE-DISCLOSING classes is withheld from a summary even under an allow disposition, because naming it at all answers the question — and a notification is the least controllable place it could be named. -->
    • S7.19.g Meet keyboard, screen-reader, focus, live announcement, narrow layout, and long-locale requirements. Evidence: accessibility journey. <!-- ANNOUNCING EVERY ARRIVAL IS A DENIAL OF SERVICE ON A SCREEN READER. An assertive live region per arrival interrupts once per notification, and forty in a storm is forty interruptions in the middle of whatever was being read; the page is then turned off, which costs them the urgent one as well. Only the urgencies at or above a threshold TAKEN FROM THE TAXONOMY interrupt — a threshold over an unordered vocabulary is a guess — and everything else moves a polite count. Both regions stay in the accessibility tree with nothing to say. The announcement is the SAME string the row shows, which is the kit's, so a value withheld from the screen cannot be spoken — an announcement is read out in a room, which is the one audience nobody models. Read state survives forcedColors: 'active' because it is a border STYLE, with the companion control proving the colour channel really was flattened. Real Tab traversal reaches every notice and leaves the surface, and the 360px check is load-bearing: forcing the rows to nowrap overflows the LIST, the document still does not move, and the list SCROLLS rather than clipping. --> <!-- S7.19 rollup: 89 kit tests (112th entry point, 6198 kit tests total) + 28 jsdom tests + 80 real-Chromium controls, each rule with a NEGATIVE control driving the case it exists to catch. THE JSDOM SPEC FOUND A REAL LEAK: the summary correctly withheld a document title and the target line beside it printed the same string. I had conflated "may you open it" with "may you know its name"; they are two dispositions, and the cross-check withheld-value-repeated-outside-the-summary catches the general shape — a surface rendering one fact in two places checks one of them. THE BROWSER TOOL'S OWN RECONSTRUCTION FAILED TWICE FIRST, and both are worth recording. An empty summaryFields list makes renderSummary return the template unsubstituted, so every row was reported as a sentence built somewhere else; and reading the target line back as the NAME accuses the surface of the leak it had just prevented, because under a refusal that line holds the class word. DIST-PATCH VERIFICATION, THREE WAYS: dropping deny from VALUE_WITHHOLDING_DISPOSITIONS failed five controls, dropping mention from NEVER_COLLAPSED_KINDS failed two, and making urgencyAtLeast always true failed two. Restoring returned all 80 to green. -->
  • S7.20 Build contextual help that describes real capability and failure semantics. Evidence: help links are versioned with the surfaced feature. <!-- THE ACCEPTANCE IS THE WHOLE SUBJECT. A help link that resolves to "latest" is not a stale document; it is a CONFIDENT WRONG ANSWER, because after one release it describes a different product to somebody reading a screen that has not changed — which is worse than no link at all. -->

    • S7.20.a Register help by capability/route/operation/field/block/gate/ error/state/contract version and locale. Evidence: help registry. <!-- HELP IS KEYED BY THE OWNING REGISTRY'S OWN IDENTIFIER, never a slug invented for the docs, which is what turns the drift check into a set comparison rather than a judgement. The anchor carries the CONTRACT VERSION of the thing and the locale, so a key names one document about one version of one feature in one language. -->
    • S7.20.b Write utility guidance for purpose, prerequisites, authorized audience, inputs/outputs, lifecycle, limitations, cost, safety/rights, failure/recovery, and support. Evidence: content completeness review. <!-- AN EMPTY SLOT IS A CLAIM, SO THERE ARE THREE STATES AND NOT TWO. "No limitations" and "nobody has written the limitations" are different sentences and a string | null can only say one of them. And WHICH slots are mandatory is DERIVED from the thing's own declared effects rather than remembered: an operation carrying a payment must say what it costs, one publishing irreversibly must say so under safety and rights. The purpose is the one slot never optional and the one held to a single sentence, because it is read in a tooltip beside a control somebody is already halfway through using — a paragraph there is a paragraph nobody reads, and everything below it then goes unread too. -->
    • S7.20.c Link help at labels, states, disabled reasons, errors, remediation, inspectors, command palette, and documentation entry points without clutter. Evidence: placement review. <!-- THE MOST-ASKED QUESTION IS "WHY CAN'T I PRESS THIS", AND THE ANSWER ALREADY EXISTS — as the guard's remediation, the capability's state or the authorization disposition. Help that restates it in its own words creates a SECOND answer, and the paraphrase is the one that goes stale without anybody noticing, because both read plausibly. So the panel renders the reason verbatim and has no prop that could carry a second wording of it. AND CLUTTER IS MEASURABLE. A help affordance beside every label is not help; it is noise that trains people to ignore the one beside the field that needed it. One affordance per element, and a ceiling on the share of controls carrying one — counted off the rendered document, with a companion load that exceeds the ceiling so the measurement is load-bearing rather than true of any fixture. -->
    • S7.20.d Resolve the exact feature/version and show stale/missing/ incompatible/unauthorized help honestly. Evidence: version matching tests. <!-- FOUR NON-ANSWERS, FOUR DIFFERENT SENTENCES, MEASURED ON ONE SCREEN. Missing reads as "there is nothing to know", unauthorized as "there is nothing here", stale as current, and incompatible as either — and each sends the reader somewhere different. It is not enough for them to be four values in a type; the browser collects the four rendered strings from four affordances in one document and asserts they are distinct. STALE AND INCOMPATIBLE ARE SEPARATED BY THE READER'S NEXT MOVE. Stale help was written for THIS feature and has not caught up, so it is worth reading with care and its body is shown; incompatible help is about a version this reader does not have. Help written for a LATER contract is incompatible and never stale, for the same reason. UNAUTHORIZED IS ANSWERED BEFORE THE LOOKUP when the thing is concealed — an entry found and refused is an entry whose existence has been disclosed. And the link carries the version of the ENTRY that was rendered, not of the surface and not "latest", which the browser checks by resolving a v2 entry on a v4 surface and reading /v2/ out of the href. -->
    • S7.20.e Generate links from the same registry/contracts as capability and errors and detect drift/broken anchors. Evidence: drift tests. <!-- BOTH DIRECTIONS, BECAUSE THEY ROT DIFFERENTLY AND ONLY ONE OF THEM IS EVER NOTICED. Help for something no registry declares is found by the first reader who follows it. A declared feature with NO help is found by nobody, because there is no broken link to notice — nobody reports the absence of a link they never saw. A BROKEN ANCHOR IS WORSE THAN NO LINK: it lands the reader at the top of a four-thousand-word page, which they read as "the answer is not here", and then they stop asking. Anchors are checked against what the document declares it contains, and a document nothing indexes is its own fault. A retired anchor still serving help and help left behind a contract bump are the two remaining directions. -->
    • S7.20.f Validate concise product-language copy, keyboard/focus behavior, screen reader, locale, narrow layout, and no secret/protected details. Evidence: content accessibility tests. <!-- THE LEAK SCAN IS S1.10's OWN, NOT A SECOND COPY. That module already decided what a credential, a connection string, a stack frame and a cross-tenant identifier look like, and help text is the LAST place anybody thinks to check — which is exactly why it is where a support engineer pastes a working command with the token still in it. THE FOCUS ROUND TRIP IS MEASURED: opening moves focus INTO the panel and closing puts it back on the trigger, against document.activeElement after real clicks with the panel actually laid out — the case a useLayoutEffect restoration silently loses. AND THE BROWSER FOUND A REAL CSS DEFECT: a WIDTH bounded by the viewport is not a RIGHT EDGE bounded by the viewport. min(320px, 100vw - 32px) is the obvious rule and it is wrong, because the panel's ORIGIN is the trigger rather than the viewport — a 320px panel hanging off a control 115px in ends at 435px and pushes a 360px page sideways. CSS cannot express "as wide as the space to my right" for an anchored element, so below 480px the panel stops being anchored and becomes a sheet held against the viewport, which is the only position CSS alone can guarantee. A locale whose words are three times as long is loaded as a second viewport case. --> <!-- S7.20 rollup: 62 kit tests (113th entry point, 6261 kit tests total) + 16 jsdom tests + 51 real-Chromium controls. Every one of the 37 declared fault kinds is both raised in the source and asserted in the spec — checked mechanically, because a declared-and-never-fired rule is the shape of every unwired gate. THE NEGATIVE CONTROL HAD TO BE REWRITTEN ONCE, and the reason is worth keeping: after the sheet fix, widening the panel proved nothing, because a position: fixed element pinned to both edges does not move the document however wide it is told to be. The control now RE-ANCHORS the panel to its trigger, which reproduces exactly the defect the browser found. A negative control has to break the thing that is doing the work, not a thing that used to. DIST-PATCH VERIFICATION both ways: dropping incompatible from NON_ANSWER_RESOLUTIONS failed the four-sentences control, and dropping cost from the payment requirement in EFFECT_REQUIRES_SLOT failed two. Restoring returned all 51 to green. -->
  • S7.21 Add design tokens and primitives to Storybook or the repository equivalent. Evidence: automated interaction, a11y, theme, and visual regression suites. <!-- Closed by S7.21.a-h: typed tokens with external constraints rather than agreed values, nineteen primitives with no app vocabulary, the state matrix split by owner, fourteen contexts and the four families the contract cannot name, the interaction suite where a dispatched event is not a gesture, the accessibility coverage that leads with the thirty-two criteria a checker cannot reach, the pinned baselines with a fourth cause, and the ship boundary where three leaks in four are invisible in the bytes. Six browser tools, one bundler tool and seven generators, each with a --check in verify:inventory:v1-workbenches. -->

    • S7.21.a Publish typed tokens for typography, spacing, sizing, layout, density, color/accent/semantic state, border/radius, elevation, motion, z-index, and data visualization. Evidence: token docs. <!-- tokens.ts already carried six of the twelve — colour, spacing, radius, typography, motion, elevation. The other six are in scale-tokens.ts, kept separate because they share one property the first six do not: EVERY ONE OF THEM HAS AN EXTERNAL CONSTRAINT THAT CAN BE CHECKED RATHER THAN AGREED. A spacing ramp is a house style and a test pinning md === 16 only records that somebody typed 16. A target size is not: WCAG 2.5.5 says 24 CSS pixels and 2.5.8 says 44, and the touch density takes the AAA figure because a touch density that does not is a density with no reason to exist. So the spec asserts the CONSTRAINT — the tokens can be retuned and cannot be retuned past the rule. CONTRAST IS THE WRONG METRIC FOR A CATEGORICAL PALETTE, and writing the test is what showed it. A contrast ratio is a ratio of LUMINANCE, and a palette separated by hue has near-identical luminances on purpose — that is what makes it unordered. Requiring 3:1 between every pair would have forced it into a ramp and destroyed the property it exists for, while two colours at delta-E 60 sit at contrast 1.1. So separability is measured in CIELAB with CIE76 delta-E (a documented approximation: CIEDE2000 is more accurate and forty lines longer, and at the 25 threshold they differ by units rather than by verdict), and contrast is kept for the question it really answers — whether a mark is readable against the surface behind it. There is a test that DEMONSTRATES the point: the palette is separable and would fail a contrast rule. Both transforms are checked against published values rather than against themselves — L* 100 for white, 53.585 for #808080, and 53.24/80.09/67.20 for sRGB red, which a transposed matrix row still misses while giving a plausible-looking number. White lands at 100.0000039 rather than 100 because the published D65 white point is rounded to five places, and the test says so instead of nudging the white point until it passes, which is how a colour pipeline ends up half a standard. The other constraints: a sequential ramp STRICTLY MONOTONIC in luminance (a ramp that is not stops encoding order the moment it is printed, and the failure is invisible to whoever picked it); a diverging ramp lightest at its midpoint and odd-length, or it has no midpoint to be neutral at; a z-index scale whose members are DISTINCT (or the stacking is decided by document order) with gaps of 100 so a layer can be inserted without renumbering, and a stated narrative the tests hold to; a focus ring as its OWN token at the WCAG 2.4.13 2px floor, so retuning the border scale cannot thin it by accident; breakpoints strictly ascending from 360, which is the width every browser tool in this repository measures; and a reading measure inside the 45-75 character range. 52 tests. -->

    • S7.21.b Publish base controls, forms, status, navigation, overlay, workspace/inspector, table/list/tree, feedback, and accessibility primitives without app-specific vocabulary. Evidence: primitive index. <!-- Nineteen primitives across the nine groups, in apps/oshun/web/src/components/workbench/primitives, with their rules in @oshun/workbench-kit/primitives and their index generated from the source by generate-workbench-primitive-index.mts. A PRIMITIVE IS EXACTLY THE DECISIONS THE APP MAY NO LONGER MAKE, and it fails in two opposite directions that are each invisible from inside the other. Too specific is a domain word in a shared file. Too generic is a div with props — it moved nothing, so every consumer still decides the focus ring, the target size and the disabled semantics, and the "system" is a naming convention. primitive-owns-no-invariant is the direction nobody reports, because a component that decides nothing never breaks. THE CONTRACT ALREADY EXISTED AND HAD NO CONSUMER. ADR-S4 declared nine categories of --wk-* role names scoped to [data-workbench-theme] at S0.9, and before this item the string --wk- appeared in four places — the ADR, the generator that emits its prototype stylesheet, that stylesheet, and its README — and in no component. What shipped instead was a second undeclared palette: twenty --wb-* properties defined INSIDE one component's CSS module, plus raw hexes that near-miss those values (#16181d beside rgb(16, 24, 40), which is #101828). The primitive layer is that contract's first reader. AND THE CONTRACT'S OWN GATE RAN IN ONE DIRECTION. It checked that every declared role appears in tokens.css and never the reverse, so four -fg roles the stylesheet has emitted since S0.9 were declared by no category. A missing role is found by the first host that sets it; an undeclared one is found by nobody, because there is no broken reference to notice. Both roles and the reverse check are now in the generator. APP VOCABULARY LEAKS AT FOUR LAYERS AND ONLY THE FIRST IS OBVIOUS: the NAME, the PROP NAMES, the ENUMERATED MEMBERS, and the DEFAULT TEXT. The last is the one that hides — a component can pass the other three and still carry a domain sentence in a default argument, which is the string nobody reads at the call site. So a primitive renders no text it was not given, and the tone glyphs are allowed because a SYMBOL IS A CHANNEL AND A WORD IS VOCABULARY: ✓ says the same thing in every language on the roadmap and "Passed" says it in one. THE FILTER THAT MAKES A LEXICON A LEXICON IS A CITATION, NOT A STATISTIC, and measuring is what showed it. A lexicon derived from 104 kit vocabularies is a dictionary — it convicts Table, TextInput and every honest primitive. The obvious repair is frequency, and it runs EXACTLY BACKWARDS: stale appears in 23 modules here, concealed in 11 and quarantined in 8, while modal, toggle, orientation and density appear in ONE each, because a domain reuses its own words hardest and reaches for a platform word about once. So the filter is PLATFORM_VOCABULARY — the words of HTML, ARIA, CSS and the DOM, which are the words a primitive is written in — and running it renamed TreeView to Tree and InspectorPane to DetailPane rather than being tuned around them. A DESIGN SYSTEM ALREADY EXISTED AND MOST OF IT IS THE APP. src/design-system/components holds 63 components, and beside Button and Divider sit GroundingStateBadge, ProvenanceBadge and HabitCheckbox. classifyComponents publishes the separation rather than fixing it — one product's system carrying that product's words is CORRECT, and the fault would be shipping them in the layer seven workbenches share. Calibrating it was the work: the first rule was "a vocabulary that is not presentational", which needed nothing external and convicted Card of elevated|glass, Avatar of xs|sm|md and a polymorphic row of div|span|article. It had discovered that the exemption list is not "the presentational vocabulary" but "the words these nineteen primitives happen to use" — A CLOSED VOCABULARY IS A PROMISE THE PRIMITIVE LAYER MAKES ABOUT ITSELF, not a test that can be applied to somebody else's component. Recalibrated it finds 55 of 63, including Spinner, whose name is as generic as exists and which ships the word "Loading". THE BROWSER FOUND FOUR THINGS SOURCE-READING COULD NOT. (1) THE ROVING TABINDEX WAS HALF-IMPLEMENTED: an arrow press moved the selection and left focus on a tab whose tabindex had just gone to -1, so the next press came from nowhere. Focus moves in an EFFECT keyed on the selection, not in the handler, where the new tab has not rendered yet. (2) NO box-sizing: border-box, so the detail pane's own border put it at 102% and overflowed a 360px container by two pixels. (3) THE WCAG FLOOR NEVER BOUND: with fixed padding a control's intrinsic text height already clears 24px at every density, so the max() was guarding nothing and a control asserted at the shipped densities passes with or without it. Padding now scales with the density too, and the tool drives the density to 0.1 to put the floor in the regime it exists for. (4) THE FIRST VERSION OF THAT CONTROL PASSED AT A DENSITY THAT WAS NEVER APPLIED, because the override went on the theme scope and the Region's own data-density rule is more specific. AN UNRESOLVED CUSTOM PROPERTY IS INVISIBLE, which is why the token proof is a COMPUTED value taken twice — inside the scope and outside it. A var() naming an undefined property is invalid at computed-value time: an inherited property takes its parent's value and a non-inherited one takes its initial value, nothing throws, and color in particular lands on something that looks deliberate. The outside copy is what catches a hardcoded value or a var(--x, #fallback), and the two halves are proved SEPARATELY — one control drives a hardcode (both fire) and one drives a fallback that survives with no theme (only the second fires). Also repaired here because it guards this registration: the S1.2 export map's test had pinned 46 subpaths by hand and the kit passed 100, so the gate had been red since 2026-08-03 and every S7 item since had widened the gap unnoticed — the exact failure the generator's own "the approved set is DERIVED, never authored" comment exists to prevent. Five assertions now derive what they used to restate, and a bogus entry point injected into the generator still fails two of them. 44 browser controls, 63 kit tests, 43 jsdom tests, 13 generator tests. DIST-PATCH VERIFICATION three ways: silencing token-not-load-bearing failed two controls, widening PRIMITIVE_CONSUMED_CATEGORIES with data-visualization failed two, and silencing primitive-owns-no-invariant failed one. Restoring returned 44/44. -->

    • S7.21.c Add default, hover, active, focus, selected, disabled, read-only, loading, empty, error, stale, fixture, degraded, conflict, and destructive stories. Evidence: state coverage. <!-- 57 stories in 20 CSF files beside the primitives, the matrix in @oshun/workbench-kit/primitive-states, and evidence/v1-workbenches/workbench-primitive-state-coverage.json. A STATE MATRIX WITH ONE AXIS HAS ALREADY LOST, BECAUSE THE FIFTEEN STATES ARE OWNED BY THREE DIFFERENT PARTIES. The BROWSER owns hover, active and focus: nobody sets them, no argument produces them, and a story that RENDERS one is a picture of a state the product does not have — it reviews well, it passes a visual baseline, and the real hover can be doing nothing at all. The COMPONENT owns whichever of the rest it has an argument for, which is a FACT read out of S7.21.b's parsed index rather than a decision. The COMPOSITION owns the five no primitive has an argument for anywhere — stale, fixture, degraded, conflict, destructive — which are exactly the words S7.21.b forbids a primitive to know. AND A COMPOSITION STATE IS A CLAIM ABOUT THE LAYER, NOT ABOUT EACH COMPONENT. The first version of the module asked it per primitive: 5 states × 19 components = 95 cells, of which "a stale VisuallyHidden" is representative — not a gap in the coverage, a sentence with no meaning. Worse, it would have been answered by somebody adding tone="stale", which is how the vocabulary rule gets undone: not by anybody arguing with it, but by a story that was awkward to write. NEVER_A_COMPONENT_STATE refuses to call one of those a component state however the API changes. INTERACTIVITY IS MEASURED, NOT DERIVED. Reading it off the API looked possible and gets Field wrong (its children is a render prop, so it parses as a handler) and NavList wrong (it takes no handler and is a list of links). Whether something can be hovered is a fact about what it renders, so the browser reports it and the kit takes the report — which also means the coverage generator REFUSES to run without the measurement rather than quietly marking those cells inapplicable, because a coverage report that improves when the browser tool stops being run is worse than none. THE BROWSER FOUND THREE THINGS. (1) THE LAYER HAD NO POINTER STATES AT ALL — hover, active and rest computed the same canonical string on every control, so nothing in it answered a pointer. Two rules fixed it, drawn from contract roles, with opacity for the press because it survives forced colours where every colour collapses. (2) A TREE ITEM TOOK A CLICK AND SAID SO NOWHERE: no cursor: pointer, so hover measured as INAPPLICABLE to the one collection primitive that is operated. (3) THE RULE ABOUT FOCUS RINGS WAS WRONG FOR TEXT FIELDS, and the browser is the authority: :focus-visible MATCHES a clicked input, because somebody who clicked into a field is about to type and needs to see where. A text field also has no press state at all — a pointer down places a caret rather than taking an action — so one measured acceptsText flag exempts it from both the ring rule and the press comparison. The measurement also settled two story sets against me and was right both times: a CLOSED ModalDialog and a Popover panel whose own box is tabindex="-1" have nothing to hover, so their interaction stories were removed rather than the check being softened. 16 browser controls, 34 kit tests, 57 of 57 applicable cells told (21 browser-owned, 31 component-owned, 5 layer compositions). -->

    • S7.21.d Add comfortable/compact/touch, theme/forced-colors, narrow/wide, zoom, reduced-motion, RTL/bidi, long-string, and font-fallback stories. Evidence: cross-context coverage. <!-- Fourteen contexts across the parent's eight families, in @oshun/workbench-kit/primitive-contexts, with nineteen stories beside the primitives and tools/primitive-contexts-browser.mjs driving every one of them in a real browser. THE CONTRACT'S CONDITIONS ARE THE ONES A VALUE CAN ANSWER, AND THE ONES THAT BREAK A LAYOUT ARE THE OTHER FOUR. ADR-S4's a11y matrix names five — theme, density, zoom, reduced motion, forced colours — and S7.21.b made every primitive declare which of them it answers. Every one of the five is answered by SUBSTITUTING A VALUE: a colour role changes, the density multiplier changes, a duration goes to zero. The four families the contract does not name are exactly the four no substitution can answer: no token makes a control fit in less room, no token mirrors an inset box-shadow because box-shadow has no logical form at all, no token makes a sixty-four character digest break, and no token gives a substituted face the metrics of the one that was asked for. So a primitive can declare all five, pass every check the index makes, and be broken in half the families — measured here at four to seven unclaimed families for every visible component, and the contract's own matrix makes that invisible because it looks complete. A CONTEXT IS A PROPERTY OF THE DOCUMENT, NOT AN ARGUMENT TO A COMPONENT. You cannot put one button into forced colours while the button beside it stays as it was, so the unit is ONE story per context whose subject is every primitive that context reaches. A per-(context, primitive) grid is 266 pictures of something that cannot happen, and — S7.21.c's lesson, learned the same way — it would have been answered by somebody adding a forcedColors prop. The gallery is what makes the subject derivable rather than remembered: a twentieth primitive fails every context story until it is in there. EVERY CONTEXT NEEDS TWO MEASUREMENTS AND THE FIRST IS THE ONE EVERYBODY FORGETS — was it APPLIED, and did the layer SURVIVE it. Three prior sightings in this repository of the first going missing (S7.14's forced-colours companion, S7.21.b's density applied to a scope a more specific rule defeated, S7.10's gate answering an empty list), so each context declares a proof of application and the invariants are SKIPPED when it fails. And application is a property of the DOCUMENT while answering is a property of the PRIMITIVE: a density that moved twelve primitives and not the thirteenth arrived perfectly well, and the thirteenth is S7.5's checkbox. Collapsing the two files a real defect as a broken harness. WHICH COMBINATIONS MATTER IS DERIVED, NOT CHOSEN. Fourteen contexts is 16,384 combinations. Each declares the BUDGET it spends and the SIGN of its effect, and two must be told together exactly when one raises demand on a budget the other lowers the supply of, or when both substitute the same channel — fourteen pairs, of which five need a story because both halves are the story's to render. reduced-motion pairs with NOTHING, because nothing else spends the motion budget, and that absence is a result rather than an omission. TWO SUBSTITUTIONS ON ONE CHANNEL DO NOT BOTH APPLY, AND THE OBVIOUS WAY TO MEASURE THE PRECEDENCE IS WRONG. "The reader's palette beats the host's" is not "the rendering is identical to forced colours alone": the user agent forces the USED colour and leaves the declaration alone, so a host that named LinkText survives forced colours legitimately while a host that named its own blue is flattened to CanvasText — two different renderings, both correct. Asserting equality reported five primitives as leaking. What a host may never do is get a colour of its own onto the screen, so the rule is that every used colour is one the reader's palette supplied, with the palette RESOLVED from the browser rather than listed. ZOOM IS TWO MECHANISMS AND THEY FAIL IN OPPOSITE DIRECTIONS. The user agent's text zoom scales the root font size, so every rem grows — while a media query's rem does not, because font-relative units there resolve against the INITIAL font size, and a breakpoint therefore slides away from the panel it guards. The contract's own --wk-text-scale is the reverse: only the declarations that read it grow, so type gets bigger inside boxes that did not. One name for both would have tested whichever one somebody implemented. THE BROWSER FOUND SEVEN REAL DEFECTS, none of them visible in source. (1) NavList and Tree marked the current and selected rows with an INSET BOX-SHADOW, which has no logical form — so the marker drew on the physical left in a right-to-left layout, on the trailing edge of every row — and which forced colours forces to none, so the one channel the marker had disappeared for exactly the reader who set the palette. One border-inline-start fixes both. (2) NavList, Tabs and Tree are interactive targets that answered NO DENSITY at all: their padding read --wk-space-* directly, so a host setting [data-density] got bigger buttons beside navigation that did not move. (3) --wk-font-sans and --wk-text-scale had NO BASE READER — ten of the seventeen visible primitives did not move when either was retuned, and every primitive that declared no font-size rendered in the surrounding application's default. (4) InlineMessage distinguished its six tones with a border COLOUR alone, so two tones computed the same appearance under forced colours; it now carries StatusMark's glyph table rather than a second one. (5) At 1600px the region and the detail pane set prose at 129 and 177 characters: S7.21.a published the 45–75 measure as a constraint and nothing read it, so the contract gains a --wk-measure role and the text — not the box — is capped. (6) A sixty-four character digest at 360px pushed the DOCUMENT sideways from a table cell, a status mark and a button. (7) An <input> sizes itself from size, which is a width in ems, so a reader who doubles the text size at 360px gets a field 31px wider than the document — and NEITHER context alone shows it. Harness lessons worth keeping: THE REFERENCE FOR ONE HALF OF A PAIR IS THE OTHER HALF ALONE, because measured against the plain baseline "did the layout mirror" is asked of a document that also doubled its type and every mirrored surface comes back broken; the pair's scope is the INTERSECTION, or a detail pane gets asked about a touch density; the mirror anchor is the first LINE BOX rather than the first element child, which changes with the content; a form control's hidden overflow is reachable with the caret and is not clipped content; and the reading measure is the longest rendered LINE, because a seventeen-character label in a 1600px box is not a 180-character line. 42 browser controls, 56 kit tests, 32 jsdom tests, 181 of 181 in-scope cells measured. DIST-PATCH VERIFICATION three ways: emptying SUBSTITUTION_BUDGETS failed the pair control, silencing the target-floor comparison failed its negative control, and NARROWING viewport-narrow's required feature to motion changed nothing at all until a control was added that the viewport family reaches every primitive the browser saw on the screen — S7.10's asymmetry again, and the fourth sighting of a scope that can only be narrowed. Restoring returned 42/42. --> <!-- S7.21.f/g/h remain: automated WCAG checks, visual baselines, and the bundle scan. -->

    • S7.21.e Add play-function interaction/keyboard/focus/announcement/ pointer-alternative tests. Evidence: passing interaction suite. <!-- Fourteen play functions in primitives/interactions.stories.tsx, the model in @oshun/workbench-kit/primitive-interactions, the measurement in tools/primitive-interactions-browser.mjs (75 controls), and the coverage in evidence/v1-workbenches/workbench-primitive-interaction-coverage.json at 38 of 38 cells. THE PARENT'S FIVE WORDS ARE NOT FIVE WAYS OF SAYING THE SAME THING, and only the first is about the person who performed the gesture. Interaction asks whether it produced the result it names; keyboard asks the same question of somebody who cannot point, focus of somebody who cannot see where the page went, announcement of somebody who cannot see the page at all, and pointer-alternative of somebody who cannot execute a path. The four that follow are ONE QUESTION ASKED OF FOUR OTHER READERS, and a component can answer the first perfectly while failing all four — which is what this layer was doing in four separate places, having passed every check S7.21.b, S7.21.c and S7.21.d make. EVERY OPERATION NEEDS BOTH PATHS AND THE TWO GAPS ARE INVISIBLE TO OPPOSITE PEOPLE. A pointer-only operation is invisible to whoever built it with a keyboard and a keyboard-only one to whoever built it with a mouse; nobody drives both, so the check is a MATRIX of operation by path. And the halves are not equally defended: WCAG 2.1.1 requires the keyboard path and NO success criterion anywhere requires the pointer one. That is why Tree shipped with a disclosure a mouse could not reach — clicking selected, there was no affordance that toggled, and ArrowRight was the only way into a branch — and why Popover shipped able to be dismissed with Escape and by nothing a pointer could do. Both passed every automated accessibility checker ever written, because none of them looks. THE OTHER TWO DEFECTS. Tree had exactly ONE tab stop and no arrow navigation at all, so a keyboard reader arrived at the first node and that was the entire tree — while the component's own invariant said it was "walked with the arrow keys as WAI-ARIA's pattern requires". ModalDialog called its caller TWICE for one Escape, because the press fires cancel and then close and both were wired; setOpen(false) twice survives that and a handler that pops a stack does not, and both halves read correctly in source. A DISPATCHED EVENT IS NOT A GESTURE, WHICH IS THE TRAP IN PLAY FUNCTIONS SPECIFICALLY. userEvent.click constructs an event sequence and hands it to a node; a real pointer is hit-tested first. Measured here: with a modal <dialog> open, a dispatched click on the button behind it ran the button's handler and a real click at the same coordinates did not. A play function is therefore capable of asserting a press no user can make, in green. The harness then produced the same finding again by accident — the obvious way to write "dismiss the popover by pressing outside" pressed a control the 320px panel was drawn straight over, and only the real pointer noticed. So the play functions are written against an input INTERFACE, the runner decides which implementation they get, every observation records which one it was, and the kit refuses to settle keyboard, focus or pointer-alternative from a constructed event. Two of the fourteen plays pass under real input and fail under userEvent, which is the evidence that the distinction is observable. BEING ASSOCIATED IS NOT BEING ANNOUNCED. There are two ways a change reaches somebody who is not looking — focus lands on the thing that changed, or a live region says so — and aria-describedby is neither, because a description is read when focus ARRIVES and a validation error exists precisely because focus is already there. Field rendered its error, associated it correctly, satisfied every check of its own wiring, and said it to nobody. The fix is a live region that is OLDER than its message, deliberately not hidden while empty, since display: none takes an element out of the accessibility tree and a region that hides itself re-enters it together with its first message. RUNNING THE RULE CORRECTED IT THREE TIMES, each worth keeping. open on a <dialog> is the platform's visibility mechanism rather than a state anybody is told, so leaving it among the announced attributes convicted a dialog of closing correctly. A change inside an aria-hidden subtree is not an announcement question at all, which is what the disclosure glyph is. And a set that permits exactly ONE holder says everything about the losers by naming the winner — announcing the tab that just went unselected is "Overview, not selected" on every arrow press, so implied-by-the-single-selection is a third announcing route, narrowed to aria-selected/aria-current/aria-checked because two branches can both be open and two filters can both be on. reached IS DERIVED FROM THE CANONICAL STATE AND NEVER REPORTED BY THE HARNESS: a tool that hands over its own yes-that-worked boolean can hand it over for a gesture that did nothing, which is the shape of every unwired gate this phase has found. Which also caught a story of its own — a popover whose canonical state was a boolean is back where it started by the time the tool reads it, because the play has to open the panel before it can close it. A FEW OPERATIONS GENUINELY HAVE ONE PATH AND THE EXEMPTION IS CHECKED RATHER THAN BELIEVED. Typing has no pointer gesture — the pointer path for text entry is an on-screen keyboard, which is the platform's — and a modal's dismissal control is the caller's content, because a shared layer drawing its own "Close" would be putting this file's wording on every product's screen. Both are sentences rather than flags, both leave no cell in the denominator, and an exemption for a path the measurement finds anyway is reported as path-exempted-and-then-found. S7.21.C'S INTERACTIVITY MEASUREMENT IS THE WRONG EVIDENCE HERE and the first full run proved it: "an operation on a primitive with nothing operable in it" convicted Popover and ModalDialog, because that measurement is taken over S7.21.c's RESTING stories, where an overlay is closed and has no box. The parsed index answers the useful question instead, in the direction nobody reports — WHICH HANDLER PROPS NOTHING DRIVES. And a focus outcome had to become PER PATH, because Escape says only "take this away" while a press outside also says "I am over here now"; one answer for both convicts whichever half it was not written for. A CONTAINER'S CENTRE IS NOT A POINT ON THE CONTAINER, which the parent's own S7.21.c tool found out the moment an open branch drew its children under its row rather than beside it: the geometric centre of the branch landed on a child, the press focused the child, and three controls reported a working component as having no pointer states. Every pointer gesture there now aims at the element's own first row. 75 browser controls, 59 kit tests, 18 jsdom tests, 38 of 38 cells. DIST-PATCH VERIFICATION three ways: emptying SINGLE_HOLDER_ATTRIBUTES failed one control, making reached always false failed twelve, and silencing focus-left-outside-the-tab-order failed its negative control. Restoring returned 75/75. --> <!-- S7.21.f/g/h remain. -->

    • S7.21.f Run automated WCAG checks and real semantic assertions for complex widgets. Evidence: passing accessibility suite. <!-- primitive-accessibility.ts (the standard, the dispositions and the widget contracts), tools/primitive-accessibility-browser.mjs (axe-core over every subject plus the computed accessibility tree read through CDP), __fixtures__/primitiveAccessibilityHost.tsx and primitiveStoryKinds.ts. 44 browser controls, 68 kit tests, 39 fault kinds, 55 criteria each with exactly one derived disposition. A PASSING AUTOMATED SUITE IS QUOTED AS CONFORMANCE, AND THE COVERAGE IS THEREFORE THE DELIVERABLE. Measured against the installed axe-core rather than argued from a blog post: 104 rules, 74 carrying a WCAG success-criterion tag, and between them they touch 23 of the 55 Level A and AA criteria in WCAG 2.2. So the very best an automated pass can mean is that a fifth of the standard was checked and nothing whatever was said about the other thirty-two, while the artifact everybody forwards reads "0 violations". suiteHeadline is BUILT from the counts (S7.14's rule) and the run reports 30 of 55 answered by a machine — 8 by the checker, 22 by controls here — with 11 the host's, 14 not exercised, and nothing left unanswered. inapplicable AND incomplete BOTH READ AS GREEN AND NEITHER IS AN ANSWER. A rule that matched no node is the empty-argument-answering-no- problems shape this repository has now found in six modules; incomplete is the checker saying it could not tell, dropped by almost every integration because the line everybody prints is violations.length. Only passed and violated are evidence. Measured here, color-contrast comes back INCOMPLETE over the tone glyphs — "element content contains only non-text characters", a correct refusal about a real question, since the glyph is the second channel S7.21.d added so a tone survives forced colours. THE CHECKER'S THIRD ANSWER IS WHERE THE WORK IS, so a resolution names a control here, the control measures the glyph at the 3:1 non-text floor, and resolution-for-a-rule-that-decided-on-its-own catches the other direction. THE ATTRIBUTE IS THE INPUT AND THE ACCESSIBLE NAME IS THE OUTPUT. Everything S7.21.b through S7.21.e asserts is markup, and a screen reader reads none of it. Reading the COMPUTED tree found two live defects in the first pass: Meter's aria-labelledby named an element no story rendered, so a progress bar was called nothing while every check of its markup passed; and every Tabs in the layer declared aria-controls pointing at a panel that was not in the document — from a component whose own invariant said the panel was "reachable from the tab rather than only by sight". The strip also wrote the CALLER'S tab id straight into the document's global id namespace, so two strips sharing a vocabulary collided. Both are the arrangement Field exists to prevent — three identifiers that have to agree, written by hand in two files — so the panels are the component's now, aria-controls appears only when there is a panel to point at, and the ids are namespaced by useId. The interactions file had the same defect independently: it rendered only the SELECTED panel, so the other tab pointed at nothing. THE INSTRUMENT HAS A RANGE AND IT IS PART OF THE COVERAGE. The protocol that reports the tree carries no aria-rowcount, aria-rowindex, aria-posinset or aria-setsize — exactly the properties a windowed table's first invariant is about. Asserting the DOM attribute quietly instead would claim tree-level evidence the suite does not have, and dropping the check would report a green table. So the observable set is MEASURED from a reference document, a contract property outside it is answered at the DOM with the caveat recorded, and settled-from-the-dom-with-no-caveat catches the arrangement where somebody forgets. A gap in the reference document showed up as a caveat rather than a silent pass the first time it ran, which is the arrangement working — the gap was the reference's. WRITING THAT CONTRACT DOWN FOUND A REAL DEFECT IN THE TABLE. WAI-ARIA counts every row a table draws, header included; its own example is a thousand records at aria-rowcount="1001" with the header at index 1. This component draws a header the caller never mentions and was neither numbering nor counting it, so a caller passing the count of its records and the absolute position of each described a table one row shorter than the one on the screen with every position one out. The fixture that showed it had a header at row 1 and a first record also at row 1. The offset is the component's, because a caller counts the rows it HAS and this counts the rows it DRAWS. AND THE ONE WORTH THE ITEM: A LAYER CAN MAKE A CRITERION IMPOSSIBLE FOR EVERY HOST BUILT ON IT AND NOTHING ANYWHERE WILL SAY SO. WCAG 1.3.5 asks a field collecting information about the person using it to declare that purpose. TextInput owns the <input> and had no autocomplete argument, so no product on the layer could ever comply — and the checker reports the rule INAPPLICABLE, because the attribute that would have been wrong is absent. Silence from every direction about a criterion the layer had removed from the table. So a criterion may be delegated to a host only by naming the HANDLE the host reaches for, the handle is checked against S7.21.b's parsed index, and delegation-whose-handle-the-layer-does-not- offer is the fault. Its negative control removes autocomplete from the index and watches it fire. A FRAGMENT IS NOT A SMALLER USAGE, AND A CHECKER RUN OVER ONE DAMAGES THE LAYER. TextInput takes its name from the Field around it by design, so a TextInput fragment has no name and cannot be given one; run the naming rules over it and the only way to green is an aria-label on a component that deliberately has none, which is how a design system ends up with aria-label="Text input" everywhere and a label rule true nowhere. The default is usage so a new story is held to the full rule until somebody writes down why not, the obligation MOVES rather than vanishing (every bound control needs a complete usage somewhere), and a fragment may omit a name while never pointing at one that is missing. A HOST RULE IS A REDIRECTION, NOT AN EXCUSE. A gallery of nineteen primitives has five <nav>s with one name between them and no <main>, and a checker is right about all of it and none of it is a component's. Twenty-eight rules are classified as the document's WITH A REASON EACH, a host classification naming a rule the checker does not have is a fault, and a criterion whose only automated evidence is a host rule is reported as the host's rather than as passing. The off-by-default set is READ OUT OF THE CHECKER and it is two mechanisms: an enabled flag on eight rules, and a separate run-time exclusion of the seven tagged experimental — reading only the first left seven rules looking unclassified. EVERY EXCLUSION CARRIES WHAT WOULD REFUTE IT, over four channels: an element the browser would have to find, a rule that would have to fire, a stylesheet declaration, and the layer's own source — the last because WCAG 2.5.4 is about an API that leaves no element and no declaration behind it, and an exclusion nothing can contradict is a shrug. RUNNING THE RULES CORRECTED TWO OF THEM. aria-controls on the unselected tab of a WAI-ARIA tab strip names a panel that is hidden, exactly as the pattern prescribes, so the user agent drops the relation from the tree; treating that as a dropped property convicted the pattern of following itself, and relations moved out of the state table because a relation's presence is a fact about its TARGET. And the layer was reported as reordering its own content on the strength of a border radius, because order: is a substring of border: and direction: of flex-direction: — a property name is matched as a property. THREE HARNESS DEFECTS WORTH KEEPING. The focus-ring control resolved two custom properties through a probe span and reported 1.00:1, because --wk-focus-ring is a whole outline shorthand rather than a colour and the surface role it was compared against is a name the contract does not have — two unresolvable properties compute the same inherited colour, so the control was comparing a value with itself and calling the answer a failure. getComputedStyle returns color(srgb 0.92 0.92 0.92) for anything that came through a color-mix, which a parser that only knew rgb() read as null. And the three WCAG 1.4.13 questions cannot be asked of a story whose onClose is () => undefined: Escape "failed" to close a panel that had no close handler, which is a harness reporting its own fixture as a defect. SPECIFIED_VOCABULARIES IS A THIRD ANSWER TO "WHOSE WORD IS THIS". S7.21.b had two — presentational, or a domain word one workbench chose — and email, postal-code and given-name are neither: they are the HTML autofill tokens WCAG 1.3.5 names, implemented by the browser, and a vocabulary this layer may not extend or cut down. Parking thirty-nine of them in PRESENTATIONAL_VOCABULARY would claim bday-month is a way of drawing something and would exempt the word everywhere. The exemption names one prop on one primitive and substitutes a STRICTER test, checked both ways: a member the standard lacks is a fault, and a token the prop refuses is a purpose no host can state. DIST-PATCH VERIFICATION three ways: silencing computed-name-is-empty failed its negative control, emptying INCOMPLETE_RESOLUTIONS failed four controls and dropped 1.4.3 from automated to unanswered, and dropping expanded from STATE_PROPERTY_SOURCES failed the dropped-property control. Restoring returned 44/44. --> <!-- S7.21.g/h remain: visual baselines, and the bundle scan. -->

    • S7.21.g Pin visual baselines for intended compositions and fail on accidental token/layout/state changes across themes/viewports. Evidence: visual regression report. <!-- primitive-baselines.ts + baselines.stories.tsx (four intended compositions) + baselineCompositions.ts (the declaration module both sides read) + primitiveBaselinesHost.tsx + primitive-baselines-browser .mjs (37 controls) + generate-workbench-primitive-baseline-coverage.mts (the report, with --check in verify:inventory:v1-workbenches), over a committed recording in workbench-primitive-baseline-pin.json and a hand-edited approval ledger beside it. THE PARENT'S THREE CAUSES ARE THE THREE THINGS THAT CHANGE AROUND THE LAYER AND NOT THE ONE THAT CHANGES INSIDE IT, and that is the sentence the whole item turns on. Ask where each of the three comes from: a TOKEN change is the host handing the layer different values, a LAYOUT change is the window handing it a different amount of room, a STATE change is the data handing it a different condition. All three are INPUTS. The change a visual regression suite is actually installed to catch is the stylesheet edit somebody made to the layer itself — a border that became a shadow, an element that stopped reading --wk-color-muted and started reading --wk-color-ink — and it is in none of them. A vocabulary of only the parent's three files every such edit under one of them, and the report then says "token change" for a change no token made, which is worse than silence because it sends the reader to the theme file. So there are FOUR causes, CAUSES_THE_PARENT_NAMES keeps the addition visible, and declaration is the fourth. A SCREENSHOT IS A BASELINE WITH NO VOCABULARY. A pixel comparison answers "1,240 pixels differ", which names none of the parent's causes and therefore cannot tell an intended change from an accident; it is a function of the machine that recorded it; and it is BLIND to a third of what was asked for, because aria-busy arriving on a table is a state change with no pixels whatever. So the PIN is a structured signature with one channel per cause and the raster is measured INSIDE a run, where the machine is constant. Measured: of four deliberate changes the recording saw all four, ONE of them leaving no pixels at all, and none visible only as pixels. WHAT THE RECORDING CANNOT SEE IS MEASURED, NOT ASSERTED — the S7.21.f lesson applied to an instrument rather than to a standard. A catalogue of real changes, one per cause plus a null change, is applied to a live document and each is recorded against BOTH instruments, so a blind spot is a number in the report rather than a silence. It also calibrates the one tolerance in the design from both sides at once: the geometry is quantised into units of the cell's own inline size (a pinned pixel is a fact about the recorder's font build), and the quantum is proved fine enough by a layout mutation that must move it and coarse enough by a two-load comparison that must not. ACCIDENTAL IS THE HARD WORD, AND AN APPROVAL IS WHERE EVERY VISUAL SUITE DIES — never on the first red run, but on the wholesale re-record three weeks later where forty images are accepted because somebody made one of them. So a difference is permitted only by a declared intent whose CAUSE and SCOPE must both match, which is what stops a token approval from laundering a layout regression sitting beside it (driven as a control: two differences, one intent, exactly one still accidental and it is the geometry one), and which must name the PIN REVISION it was written against — S7.12's golden rule applied to the approval, so an approval cannot outlive the recording it was written for. AND COVERAGE CAN SHRINK TO ZERO WITHOUT A SINGLE FAILURE: delete the stories, delete the baselines, and every visual suite ever written reports success. A pinned cell that no longer renders is therefore a FAULT, and a primitive in no intended composition is named — checked against S7.21.b's PARSED INDEX rather than against the composition list, because a list of subjects somebody remembered makes the coverage claim circular. All 19 reached by 4 compositions, none a subset of another. THE BROWSER FOUND THREE DEFECTS, TWO OF THEM IN THIS ITEM'S OWN INSTRUMENT, AND THE THIRD WAS THE INSTRUMENT MEASURING ITSELF: (1) The reference channel first asked "which role has this value", and a value-to-role lookup OVER-ATTRIBUTES: in the neutral theme --wk-color-ink is CanvasText, the same black an element that sets no colour at all inherits from the user agent, so every such element was recorded as reading the ink role — and retuning ink then presented as a DECLARATION change at a dozen places. The catalogue's retheme caught it on the first run. Provenance is now MEASURED: every role is substituted with a value unique to it, all at once, and what the engine computes is a fingerprint of the DECLARATION with the host's palette factored out by construction. Inheritance is handled exactly, composed values come back as compositions of sentinels, one style recalculation for the whole cell. (2) A PROPERTY WITH A TRANSITION DOES NOT JUMP TO ITS NEW VALUE — it animates, so getComputedStyle hands back the OLD one for as long as the transition runs. Read straight after the substitution, Button's background and ToggleButton's border came back as the host's real accent and rule colours in BOTH themes, which is exactly the theme-dependence the channel exists to remove, and each cell's answer looked perfectly plausible on its own — only a control comparing two palettes could see it. Worse, the motion roles are substituted too, so the sentinel durations made the wrong answer last a second. transition-property: none stops the animating without touching transition-duration, which this recording reads: turning the durations off would have silently emptied one of the properties in the channel. (3) A HARNESS THAT MEASURES ITS OWN SIDE EFFECTS. Three loads of the modal composition compared in isolation were pixel-identical and the same three inside the tool were not — because signing a cell puts the roles back afterwards, and putting them back starts a transition on every property that declares one, so the screenshot caught a button mid-fade at whatever moment the call returned. The fix is ordering, and it is the rule: the raster is of the document as RENDERED, never of the document as PROBED. Also corrected during the build: the report first rebuilt its repeat pairs as the same signature twice, which satisfies the stability rule with data incapable of failing it — the empty-argument-answering-no-problems shape arriving through an ARTIFACT rather than through a function, eighth sighting. The measurement now records the second signature exactly when it disagreed. DIST-PATCH VERIFICATION four ways: mapping the reference channel onto the token cause failed the catalogue (a declaration change filed under a cause an approval would authorise), dropping the CAUSE half of intentExplains failed the laundering control, widening scopeCovers to always true failed the unused-approval control, and narrowing PINNED_APPLICANTS to drop the viewport contexts failed both the derived context list and pinned-cell-no-longer-rendered over twelve real cells. Restoring returned 37/37, and a second full run reproduced the recording byte for byte. -->

    • S7.21.h Prevent stories/fixtures/test providers from production exports and bundles. Evidence: bundle scan. <!-- ship-boundary.ts (72 tests, 25 fault kinds) + apps/oshun/web/tools/ship-boundary-scan.mjs (24 controls, real esbuild at production settings) + generate-workbench-ship-boundary.mts (the report, with --check in verify:inventory:v1-workbenches) + a no-restricted-imports ban in the app's eslint config, over a committed measurement in evidence/v1-workbenches/workbench-ship-boundary.json. A PRODUCTION BUNDLER IS AN OPTIMISER, AND AN OPTIMISER ERASES THE EVIDENCE WHILE LEAVING THE LIABILITY. The parent names a BUNDLE SCAN, which read plainly is a search of the emitted output, and that is the one instrument incapable of settling it. Four leaks were planted into this layer's own barrel, one per class, and each was built by a real esbuild at production settings. Against a 1,151-byte reference: a re-exported story emitted 1,151 bytes with its title string nowhere in the output, a fixture emitted 1,151, and a test provider reaching a development-only package emitted 1,654 with its own code shaken away while the bundler still resolved fifteen files out of that package. Only the harness survived, at 17,358, and it survived because it MOUNTS A ROOT AT MODULE SCOPE and a side effect is the one thing tree shaking may not remove. Of four real leaks the served bytes showed ONE; the import was still in the source in all four cases. So the load-bearing channel is REACHABILITY, taken from the bundler's own metafile before any optimisation, and the emitted bytes are corroboration rather than the finding. Measured over the finished instrument: reachability 4/4, emitted bytes 1/4, resolved dependencies 1/4. AND THE BYTE ANSWER IS A PROPERTY OF SOMEBODY ELSE'S CONFIGURATION, WHICH IS DRIVEN TWICE RATHER THAN ARGUED. The same planted story emitted 1,841 bytes with tree shaking and 11,442 without, over an identical source tree. The same four leaks were seen by the bytes 1/4 through a consumer that imports one export and 4/4 through one that re-exports the barrel — so the scan blinds itself exactly as the consumer becomes more selective, which is the direction real pages go, and neither knob is in this layer. THE PARENT NAMES THREE INERT THINGS AND THE FOURTH ONE RUNS. A story is a catalogue, a fixture is data and a test provider is an implementation somebody asked for; all three sit there until something renders them. The harness that mounts the layer into a document is in none of the parent's words, and it is the only one of the four that executes on import, the only one an optimiser cannot remove and by fifteen times the largest. TEST_ONLY_CLASSES has four members, CLASSES_THE_PARENT_NAMES keeps the addition visible, and harness is the fourth. A CLASS IS ALSO NOT A CONSEQUENCE: what a file is belongs to the test apparatus and what it costs in production is derived, one file carries several at once, and fabricated-content is in that list because a fixture a production module reads is not weight in a bundle — it is a surface showing data nobody has. ONE OF THE FOUR CANNOT BE CONFIRMED FROM ITS CONTENTS AND SAYING SO IS THE POINT. A harness works at import, a story carries a story-format default export and a test provider imports something the install does not have: each announces itself in its own syntax and a declaration that disagrees is a finding. A fixture is a table of ids indistinguishable from a table of ids somebody ships, so any rule invented to tell them apart gets tuned until it agrees. CLASS_WITH_NO_SYNTACTIC_EVIDENCE records that the residual class's only evidence is where somebody put it, which means a fixture leaving the fixtures directory stops being classifiable and the reachability rule is all that is left. THE ENTRY SET IS THE ARGUMENT. This layer has no route consumer at all, so an audit rooted at the product's pages would traverse everything, never arrive, and report zero leaks in the layer it was written for — the empty-argument-answering-no-problems shape, NINTH sighting, arriving through the ROOT SET rather than through a function argument or an artifact. no-production-entry-point is a fault and the headline says out loud that no application file imports the layer yet. THE SCAN FOUND A REAL DEFECT ON ITS FIRST RUN THAT HAD NOTHING TO DO WITH STORIES: @oshun/workbench-kit is imported by fifty shipped files in apps/oshun/web/src and appears in neither dependencies nor devDependencies. It builds because Next resolves it through tsconfig.base.json; any resolution that goes through the manifest does not have it. Fixed by declaring it. It also produced shipped-file-no-entry-reaches for fifteen components, because the barrel named five of twenty — a gate whose coverage is a quarter of the directory reports a clean layer either way — so the barrel now names every shipped component. THREE DEFECTS IN THIS ITEM'S OWN INSTRUMENT, EACH FOUND BY RUNNING IT: (1) the module-scope-effect walk exempted its own root, so a top-level const readQueue = () => { window.localStorage... } read as work performed at import and convicted two shipped components — a function initializer performs no work at import whatever is inside it. (2) It counted every occurrence of the text document, including parameter names and property accesses, which is ordinary vocabulary in an editor component. (3) The token corpus for the byte channel was assembled from parsed string LITERALS while the search over a minified output can only be a SUBSTRING test — asymmetric, and not theoretically: Hierarchy survived as a supposed discriminator for a story and then matched HierarchyTree in the bytes. Both sides are now includes over raw text, and the corpus is derived from the bundler's own shipped inputs rather than from a remembered list of directories, which had left verified, completed, sequence and onSelect looking distinctive. Excluding tokens that appear in the CLEAN BUNDLE instead would be circular: a token is in that bundle either because something that ships says it or because the file leaked, and the second case is the finding. THE BAN IS SCOPED, NOT THE IGNORE LIST. The lint rule fired on twenty-three harnesses doing the correct thing until __fixtures__ was exempted from THAT RULE rather than added to the config's global ignore list, because these files mount real components into real documents and adding a directory there switches off forty rules instead of one. DIST-PATCH VERIFICATION three ways: silencing the module-scope arm of evidencedDisposition failed four controls, pointing REQUIRED_CHANNEL at the emitted bytes failed five, and dropping harness from TEST_ONLY_CLASSES failed two. Restoring returned 24/24 each time. -->

  • S7.22 Enforce no unnecessary card nesting, no app heroes, no decorative gradient chrome, and no competing accent colors. Evidence: design conformance review.

    • S7.22.a Inventory every container/card/panel and require that cards represent a selectable, draggable, comparable, previewable, or otherwise meaningful object interaction. Evidence: card-use review. <!-- libs/oshun/workbench-kit/src/card-use.ts (67 tests, 18 fault kinds) + apps/oshun/web/tools/card-use-browser.mjs (77 controls, 21 fixture hosts, real Chromium). Evidence: evidence/v1-workbenches/workbench-card-use.json — 998 containers, 22 cards, every card names an interaction and every one of them was DRIVEN, every component the directory ships reached. S7.1.g's RULE WAS CORRECT AND HAD NEVER BEEN SHOWN THE SET IT GOVERNS. workbench-shell-browser.mjs collects a container only where the element is a BUTTON that draws a border and reads the interaction off an attribute the component author wrote, so both halves of the argument were declarations and the decorative wrapper the rule exists to catch — a div with a border, a radius and sixteen pixels of padding — was never in the list. Empty-argument shape, tenth sighting, arriving through the INVENTORY. Measured: of the sixteen data-card-interaction attributes in the layer, NOT ONE was on an element this review derives as a card, and twelve of them said movable, a word nobody had defined — the shipped rule is interaction !== 'none', which every undefined word satisfies. So the inventory is DERIVED FROM WHAT THE ENGINE PAINTED and the declaration is the second opinion. RUNNING IT CORRECTED THE MODEL FIVE TIMES, each correction a class of finding against surfaces doing the right thing, and the first run filed four hundred of them. (1) A LINE IS NOT A BOX HOWEVER MUCH SPACE IS AROUND IT: reading inner space as box-hood called forty list items with a four-pixel left rule cards, and a list with rules between its items is precisely what S7.22.b recommends flattening a card mosaic INTO, so the rule asked people to flatten what they had already flattened. Three strokes close a shape. (2) A GRID IS ONE ENCLOSURE DRAWN N BY M TIMES, NOT N BY M CONTAINERS — the line between two cells belongs to neither, and 578 table cells came back as decorative cards that ought to be draggable. (3) A LANDMARK CANNOT BE A COLLECTION MEMBER however many identically-drawn siblings it has: a header and a footer sharing a border style with a third box repeats perfectly well and is not three of anything. (4) FOR A CONTROL THE REPEAT IS NOT ENOUGH — a row of filter buttons repeats, is separately named and encloses its content exactly as a list of objects does, and only an announcement that it stands for an object tells them apart. (5) AN INERT BOX OF ONE TEXT RUN IS A MARK: eighteen integrity chips carrying a real non-colour signal in their border width and style read as eighteen decorative cards, and a box of nothing but controls is a control group, not a card. THE MARKUP'S OWN ANNOUNCEMENT COUNTS AND ITS ABSENCE IS A FAULT: aria-selected says a thing can be chosen whether or not it is chosen now, and S7.1.g reached for a private attribute because inferring from aria-current read a STATE as a capability — but it is the attribute's PRESENCE that is the capability and its VALUE that is the state. Three of the seven words have a standard announcement and four have none, and saying which is which is the point. The shell fixture then failed its own new rule: it wrote aria-current={selected ? true : undefined}, and React's undefined idiom is exactly what removes the presence. THE VOCABULARY REFUSES FOUR WORDS BY NAME with a reason each — actionable is a border drawn around a button, hoverable is no interaction for a keyboard or a finger, clickable names the input device rather than what happens to the object, readable justifies every border anybody draws — and ADDS THREE the parent does not name (navigable, expandable, dismissable: the ones that do not leave the object where it was, and the most ordinary real card there is opens the thing it stands for). A CLICK IS NOT A DRAG: driving every word with click() reported twelve shipped stage-board cards as inert, <details> discloses through its open property and never writes aria-expanded, and a card's identity has to survive the move it is asked to perform — the board prints its column's label inside the card, so identifying it by its text made a successful move look like no move at all. EVERY OTHER GROUP IS TRIED, NOT THE FIRST: draggable claims the card can be moved, not that it can be moved anywhere, and where it may go is the graph's business. REMEDIATION: movable -> draggable (12); the job console's <tr> dropped its attribute, because a table row is drawn with a rule and aria-selected already announces it; and 31 boxes across seven components — audit entries, drift rows, notice rows, gate rows, decision sections, the bulk panel's seven stacked tiles, a refusal paragraph — became lists and sections, each keeping the non-colour signal on the channel its own browser proof measures. All nine affected browser tools and 256 jsdom tests re-run green. A PRE-EXISTING RED GATE FOUND AND FIXED: workbench-shell-browser.mjs had a hand-written esbuild alias whose bare entry SHADOWED every subpath, so it had crashed on startup since browserHost.tsx began importing @oshun/workbench-kit/identity. DIST-PATCH VERIFICATION four ways. Raising MIN_CLOSED_EDGES failed 17 controls, silencing the descendant-seat arm failed 1, dropping the mark branch failed the layer verdict — and EMPTYING REFUSED_CARD_INTERACTIONS CHANGED NOTHING, because the four refusal controls are generated by iterating that very list: no words, no controls, no failures, a green run over a rule that had been deleted. The empty-argument shape inside the harness itself, eleventh sighting. A control loop fed by the thing it verifies now states how many controls it expects to produce, and the same patch then failed. Restoring returned 77/77. -->

    • S7.22.b Flatten nested cards/panels into workspace regions, sections, columns, dividers, lists, or semantic groups where borders/elevation add no meaning. Evidence: before/after composition review. <!-- libs/oshun/workbench-kit/src/composition.ts (102 tests, 20 fault kinds) + apps/oshun/web/tools/composition-browser.mjs (60 controls, 21 fixture hosts, real Chromium). Evidence: evidence/v1-workbenches/workbench-composition.json — 998 containers, 520 nested boundaries, every one of them warranted, no box drawing a channel it cannot justify, deepest run of parallel edges 3 of 3. A NESTED BOUNDARY IS NEVER JUDGED ON ITS OWN, which is what makes this a different question from S7.22.a rather than the same one again: a border that would be the whole point of a box standing on the page says nothing at all when it is drawn four pixels inside another border that has already gathered the same content. So this module composes card-use.ts outright — derivedContainerRole says what a box is, drawsAsABox says whether a stroke closed a shape — because two measurements of what a container is would drift within a release and the second one would be the one nobody notices is wrong. THE CHANNEL MAKES THE CLAIM, NOT THE BOX: a border says an edge is here, an elevation says this is above the surface, a fill says this is a different surface, an inset rule says a line marks a state, so a box drawing two of them can be right about one and the remedy for a warranted border under an unwarranted fill is to drop the fill rather than the box. A WARRANT IS MEASURED AND SEVEN OF THE TEN ARE THE ROLE — and THREE ROLES BRING NO WARRANT AT ALL. A section, a panel and a well are exactly what the parent's word "container" hides, and they are this item's whole subject, which the table says rather than a predicate. "IT SEPARATES IT FROM ITS SIBLINGS" IS DELIBERATELY NOT A WARRANT, because it is the argument every card mosaic makes for itself and it is true: a rule separates them too, with one stroke instead of four, which is precisely why dividers and lists are on the parent's list of destinations. Separation is not a warrant AND it is something the boundary carries, and that pair is the item — a destination that cannot separate may not be prescribed for a box that has a neighbour, or the cheapest answer is always to run every stack of panels together into one column of text. THE DESTINATION IS CONSTRAINED BY THE ARRANGEMENT, WHICH IS GEOMETRY: a horizontal rule between two boxes side by side separates nothing and a wrapped set has no line through it at all, so the sibling rectangles decide which of the parent's six replacements are available, and unwrap is the seventh the parent does not name. A DESTINATION IS LEGAL ONLY IF IT CAN CARRY WHAT THE BOUNDARY WAS CARRYING — the rule S7.22.a's own remediation was already obeying by hand — so a divider carries a state on its own stroke and no name, a list carries membership, a group carries a name and separates nothing, and removing a border that was the only thing announcing a group is a deletion with a tidier diff. THE BEFORE/AFTER IS DRIVEN, NOT RECORDED: a pair of snapshots from two commits is a golden that pins nothing once somebody re-records it, so every prescription is PERFORMED in the live document and the built kit compares what the two documents state in order, the named groups in the accessibility tree, the reading order and the state signals — additions allowed anywhere, nothing dropped, shared statements never shuffled. RUNNING IT CORRECTED THE MODEL SIX TIMES and the first run filed sixty- nine findings against surfaces doing the right thing. (1) A ROLE THAT EXPLAINS THE BOX EXPLAINS EVERY CHANNEL IT DRAWS: the warrants were written per channel, so a table cell's border was justified by its cell-hood and its wash was not, and a hundred and forty-four filled cells came back as findings. (2) The colour-only rule fired on forty-seven buttons whose border is drawn because they are the thing you press; it is now asked only of a channel the colour is the ONLY candidate justification for. (3) A LIST IS A RUN, and membership is a claim about one: a gate row's result and remediation blocks are drawn alike and separated by the row's own field table, so gathering them into one list moved the fields out from between them — the before/after reported the reordering and the prescription was wrong, which is the transition doing the job it exists for. (4) A BOX WITH NOTHING IN IT IS NOT A CONTAINER, IT IS A DRAWN INDICATION: the timeline's range band is a translucent wash with no children and no text whose whole job is to say WHERE the range is, and there is nothing inside it for a boundary to be about. (5) The channel rules were escaping the item's scope onto boxes standing on the page, whose channels are S7.22.a's question. (6) A probe holder that is itself a box contains every probe's text, so selecting probes by text picked the HOLDER and eight controls ran against a box none of their rules were about. REMEDIATION: the gate panel's result, remediation and waiver blocks became sections — a rule above the block plus a heading that names it — and the waiver kept its own signal on a NON-COLOUR channel (a double rule rather than an amber wash), because colour alone is a state nobody reading in one colour can see. The gate panel's 61 browser controls, card-use's 77 and its 28 jsdom tests all re-run green. NO RULE GOES UNEXERCISED: the layer is now clean, so compositionFaults and flatteningFaults over the real observations both answer an argument nothing in them can fail — every one of the twenty faults therefore has to have been produced by a control in the same run, over a document the engine painted, and the four ways a flattening actually goes wrong (removing the box with its contents, gathering non-adjacent boxes, deepening the composition, doing nothing at all) are driven rather than described. DIST-PATCH VERIFICATION three ways: dropping the divider's warrant failed 5 controls, lowering the parallel- edge ceiling to 2 failed 2, and silencing the named-group arm failed 2 — including the coverage check, which is what catches a rule that has been deleted rather than merely never reached. Restoring returned 60/60. -->

    • S7.22.c Remove marketing hero, campaign headline, stat strip, logo cloud, and aspirational banner patterns from operator/authoring routes. Evidence: route screenshot audit. <!-- libs/oshun/workbench-kit/src/first-screen.ts (62 tests, 13 fault kinds) + apps/oshun/web/tools/first-screen-browser.mjs (44 controls, real Chromium, real PNGs). Evidence: evidence/v1-workbenches/workbench-first-screen.json and twenty captures under evidence/v1-workbenches/first-screen/ — 20 surfaces (5 routes x 2 archetypes x 2 viewports), every audience derived, none of the five patterns found, worst first screen 1% empty of a 17% budget, every surface captured. A HERO IS AN ALLOCATION, NOT A STYLE. Two blocks can be pixel-identical — same height, same type scale, same full-bleed width — and only one of them is a hero, because the difference is what the block OFFERS the person reading it: a release console's masthead carrying the release name, its gate verdict and a publish button occupies exactly the space a marketing hero occupies and is the most useful thing on the page. So nothing here measures a gradient or a font size on its own; every block on the first screen is asked what a reader can DO with it, and the five named patterns are five shapes of the answer "nothing". THE AUDIENCE IS THE ARGUMENT AND GETTING IT WRONG FAILS BOTH WAYS: the rule is addressed to operator and authoring routes, so auditing a landing page as an operator surface files every finding against a page doing its job, and classifying an operator surface as public lands the exemption on exactly the page the rule exists for. It is derived from S1.8's descriptor, which already knew and had never been asked — scope, roles and deepLink.shareSafety, the last of which has said public / tenant-only / never-share since S1.8 and is almost literally the question — and a route none of the three settles is reported NOT DERIVABLE rather than assumed. THE FIRST RUN FOUND EXACTLY THAT: four of the layer's five routes declared platform scope and no deep link at all, so sixteen of the twenty surfaces could not be placed. That is inaccurate as well as unauditable — a brief is a workspace object and its link is not for a stranger — and correcting the descriptors made S1.8's own parser refuse a tenant-only link with no expiry, which is a URL that keeps working after the person it was sent to has left. PRIME SPACE IS VERTICAL SPACE ABOVE THE FOLD, NOT THE TOP OF THE DOCUMENT, so the unit is the INTERSECTION of a block with the first viewport in pixels, and the audit runs at 1440x900 and at 390x844 because everything a hero displaces at handset width goes further down rather than to one side. The blocks are FOUND rather than named: descend through wrappers that have one child covering their own area, stop at the first level that partitions into more than one part, and KEEP DESCENDING WHILE A PART STILL HOLDS THE WORK to a bound of three. One level was not enough and the first run showed why — the shell's outer band held the work, its region grid held the work, and the audit stopped with a single block covering the entire screen and offering the-work, which is exactly where a hero inside a workspace would have hidden. AN OFFER IS ONE OF A CLOSED LIST AND what-you-may-not-do IS ON IT DELIBERATELY: a refusal at the top of a surface is the most useful thing prime space can carry when it is true, and a rule that scored it as decoration would ask every surface to bury the one thing its reader needs first. A HEADING IS NOT A HERO — an operator surface names itself and that is orientation — so a campaign headline is measured as a heading ABOVE the surface's own name or drawn LARGER than it. THE FIRST RUN REPORTED A SHIPPED SHELL AS CARRYING ONE, because the surface named itself with an aria-label and the rule compared a 13px section title against a name that has no size, no position and nothing a sighted reader can see at all. Both arms now require the name to be a HEADING, and a working surface that names itself only in the accessibility tree is a finding of its own. AND A NUMBER YOU CANNOT ACT ON IS A STAT STRIP: not a number, since a job console's counts are the work when pressing one filters the list, and every clause of "three or more identically drawn tiles, each a numeral and a caption, none focusable, selectable or linking anywhere" is load-bearing. THE SCREENSHOT IS THE ARTIFACT AND THE MEASUREMENT IS THE EVIDENCE. A picture is the one form of evidence nothing can re-check — it cannot fail, it cannot be compared, and a year later nobody can tell what viewport it was taken at — so a capture and a measurement are taken from the same page at the same viewport and a pair whose two halves disagree about that viewport is refused rather than filed. RANGE, MEASURED RATHER THAN ASSERTED: the product ships 720 page files under apps/oshun/web/src/app and not one resolves to an S1.8 route descriptor, so this audit can derive an audience for none of them and does not claim to have reviewed them. The count is in the evidence. The model and the tool generalise to those routes the moment they carry a descriptor; what is missing is the argument, not the rule. A production next build of the app was run to establish that this is a descriptor gap rather than a rendering one — it succeeds, with a 10GB heap. NO RULE GOES UNEXERCISED: the layer is clean, so firstScreenFaults over the twenty real surfaces answers a question nothing in them can fail, and every one of the thirteen faults has to have been produced by a control in the same run over a document the engine painted. Harness trap: a probe planted inside a holder div is measured at the holder's full width, so a 60vw block planted to drive the rule that catches what the five patterns MISS came back as a marketing hero — the harness must not be the thing under test (S7.2). DIST-PATCH VERIFICATION three ways: tightening the empty-space budget past what the layer draws failed 1 control, narrowing AT_WORK_SCOPES to drop workspace failed 2, and silencing isStatStrip failed 2 including the coverage check. Restoring returned 44/44. -->

    • S7.22.d Remove decorative gradients, glows, glass effects, ornamental icons, thick repeated borders, and shadows that do not convey state/ hierarchy. Evidence: chrome audit. <!-- libs/oshun/workbench-kit/src/chrome.ts (40 tests, 8 fault kinds) + apps/oshun/web/tools/chrome-browser.mjs (35 controls, 21 fixture hosts, real Chromium). Evidence: evidence/v1-workbenches/workbench-chrome.json — 54 effects, every one of them conveying something, 50 by state and 19 by hierarchy, 3 of the 6 effect kinds present in the layer at all. THE QUALIFIER IS THE RULE. The item lists six CSS effects and then qualifies the whole list with five words — "that do not convey state/hierarchy" — and read without them it is a ban on six properties: a layer obeying that would delete the inset rule marking a blocked gate and the elevation lifting a dialog off the page. Read with them, the qualifier names TWO justifications and they are TWO COMPARISONS AGAINST TWO DIFFERENT SETS. An effect conveys a STATE when it distinguishes this thing from the things BESIDE it; it conveys HIERARCHY when it distinguishes this thing from the thing AROUND it. Sideways and outwards. An effect the same as its neighbours' and the same as its container's orders nothing and marks nothing — it is a texture, and that is the finding. THE SIBLINGS ARE THE COMPARISON SET FOR A STATE, NOT THE CLASS LIST. S7.22.b keyed the instances of a part on the class list, which is right for asking whether a component part varies and wrong here: a modifier class makes row unread and row read two parts and each of them is perfectly constant, while the reader is comparing a row to the row under it. AN ELEVATION WITH NO OFFSET IS A GLOW, and the difference is where the light is. A shadow offset downward claims a surface above another under light from above, which a reader can check against everything else on the screen; a zero-offset shadow with a wide blur claims a thing lit from nowhere. And a zero-offset shadow with almost NO blur is a focus RING, which is the most important mark on a keyboard surface — so the blur is the discriminator rather than the offset alone, and both ends of that threshold are defended. GLASS IS A BACKDROP FILTER AND NOTHING ELSE: a translucent background is a colour with an alpha, this layer has several that are ordinary control fills, and treating alpha as glass would file findings against every rgba() in it. AN ICON THAT NAMES ITSELF IS NOT ORNAMENTAL, and state and hierarchy are deliberately NOT on its warrant list — a mark does not become meaningful by differing from the mark beside it, so it either says what it means, sits inside a control whose target it is part of, or draws a value. ONE THICK BORDER IS EMPHASIS AND THREE IDENTICAL ONES ARE A TEXTURE: the item says thick REPEATED borders and the repetition is doing the work, since a four-pixel leading rule on the one blocked gate in a list is the most useful mark on the surface. THE ONE REAL FINDING WAS A DEFECT IN THE MEASUREMENT, AND IT IS S7.22.a's SOLE-ELEMENT-CHILD FINDING IN ANOTHER GUISE. The run reported the navigation's current-item marker — a three-pixel leading rule whose whole purpose is to mark one entry among several — as a border conveying nothing, because the marker is drawn on the anchor and the anchor is the only element inside its list item, so it had no sibling at all. The reader is comparing list ITEMS. The comparison set now walks up through sole-child steps first and reads the effect off anywhere inside each sibling, and the marker earns itself by state as it always did. AN EFFECT WITH NEITHER A SIBLING NOR A CONTAINER IS UNTAKEN, NOT FAILED: both comparisons come back false over an empty argument, and that reads exactly like an effect that conveyed nothing. NO RULE GOES UNEXERCISED, and here it matters more than anywhere: the layer paints NO gradient, NO glass and NO glow at all, so three of the six rules answer an empty argument over the real surfaces and come back clean — indistinguishable from three rules that have been deleted. Each of the six is therefore planted into a live document TWICE, once in the shape the rule exists to catch and once in the shape that earns its place, and every fault has to have been produced by a control in the same run. DIST-PATCH VERIFICATION three ways: lowering THICK_BORDER_PX to 1 made every hairline a thick border and failed the layer verdict, silencing the sideways comparison failed it again over the 50 effects it justifies, and raising GLOW_MIN_BLUR_PX past every real shadow failed 3 controls including both coverage checks. Restoring returned 35/35. -->

    • S7.22.e Enforce one primary accent and separate accessible semantic statuses with noncolor cues. Evidence: token/contrast audit. <!-- libs/oshun/workbench-kit/src/signals.ts (82 tests, 11 fault kinds) + apps/oshun/web/tools/signals-browser.mjs (33 controls, 21 hosts, real Chromium). Evidence: evidence/v1-workbenches/workbench-signals.json — ONE accent hue family, 190 of 190 paints governed by a role, 10 statuses across 2 vocabularies with NONE carried by colour alone, and 154 of 154 measured pairs clearing WCAG 2.1 AA. AN ACCENT IS A HUE FAMILY, NOT A HEX. A surface painting one blue at three lightnesses has ONE accent and a rule counting distinct values reports three; a surface painting a blue primary button beside a teal go link has two and that same rule reports both cases identically. The unit is therefore the hue family in OKLCh, which was fitted to perceptual data so one angular tolerance means the same thing everywhere on the wheel — HSL's would not, since it puts pure blue and pure yellow at the same lightness. HUE CANNOT SEPARATE A COLOUR FROM A GREY AND ONLY CHROMA CAN, which the measurement found rather than an argument: every neutral in this layer is a COOL grey sitting within a couple of degrees of the accent's own hue, so grouping by hue alone reports every divider on every surface as a second accent painted in the accent's colour. CHROMATIC_CHROMA_FLOOR is therefore the whole of the measurement, placed at a twenty-percent tint of a saturated accent over white — the strongest wash that still reads as a tinted surface rather than as the colour itself. ONE ACCENT MEASURED IS NOT ONE ACCENT ENFORCED, AND THE ITEM'S VERB IS ENFORCE. The first run counted one hue family and found the count was a COINCIDENCE: ADR-S4's token contract resolved in ONE of twenty-one documents and governed 17 of 184 paints, because the only element in the product carrying data-workbench-theme was a fixture inside the primitive catalogue. The shell — the thing every workbench surface renders inside — never turned the contract on, so "one accent" held because sixteen authors had reached for the same blue that week. WorkbenchShell now carries the attribute and publishes the nine roles from the --wb-* values it already declared (not one painted colour moves), the fixture hosts mount through mountThemed so the harness matches production, and forty-five literals across eight stylesheets resolve to roles: 21 of 21 documents, 190 of 190 paints. SEPARATE IS ABOUT THE ACCENT AND NOT ABOUT THE OTHER STATUSES, AND THE ITEM'S TWO CLAUSES ARE WHAT SETTLE IT. The first run also convicted the gate panel for drawing three blocked gates in one red and separating them with three different words — which is correct, and the rule that convicts it demands a distinct hue per member of every vocabulary, which is the FIRST clause of this same item inverted: manufacturing a colour per status is precisely how a surface acquires the competing accents "one primary accent" exists to prevent. A shared colour is a channel DECLINING TO CARRY, so it is reported as a fact and the fault is reserved for the case where nothing separates two members at all. A NON-COLOUR CUE MUST DISTINGUISH RATHER THAN MERELY EXIST, so the comparison set for a cue is the other members of its own vocabulary — S7.22.d's sideways comparison moved off the document and onto the vocabulary, because the reader is not comparing this row to the row under it, they are comparing this status to the one it could be confused with. AND THERE ARE TWO READERS REACHED BY DIFFERENT CHANNELS. An aria-label is a complete answer for somebody using a screen reader and nothing at all for the reader 1.4.1 was written for, who can see the screen and cannot separate its hues; a dashed border is the exact mirror; and the glyph this layer's status primitive draws is aria-hidden ON PURPOSE, which makes it a sighted channel and not a screen-reader one. A single has-a-cue answer passes a surface unreadable to precisely the person the criterion protects. The first run's six colour-alone findings were all one missing channel: THE ACCESSIBILITY TREE CARRIES A CHANNEL OF ITS OWN, and announced-role is the counterpart of border-style — the inline message renders role="alert" for one tone and role="status" for the other, stating the distinction in the one place that reader is listening. THE PAIR'S KIND DECIDES THE CONTRAST FLOOR AND THE KIND IS MEASURED. WCAG 2.1 has several numbers and which applies is a property of what the pair IS; the floors are read from docs/releases/p2/accessibility-baseline.md rather than chosen here, because §STD.1 has not yet recorded the 2.2 target and 2.2 changed neither 1.4.3 nor 1.4.11. Measured on this layer's own pixels a disabled toolbar button sits at 3.06:1 — its criterion met and the wrong criterion failed from one pair of colours. THIRTY REAL CONTRAST DEFECTS, ALL FIXED. A control's boundary is what identifies it as a control, so 1.4.11 asks 3:1 of it against the surrounding fill, and the layer drew fourteen of them in the DIVIDER colour at 1.24:1 — a boundary a reader can find only because they already knew the control was there. --wb-control-line is that budget written down. The block editor's toolbar left its buttons entirely to the user agent, which supplied a #d6d6d6 boundary at 1.45:1 and a #b4b4b4 disabled label at 1.99:1, invisible to review because nothing in the file put them there. And three of the contract's own four state roles were CSS named colours that clear the 3:1 a badge is judged at and fail the 4.5:1 the words on it need: SeaGreen 4.25, DarkGoldenrod 3.25, SteelBlue 4.11, each replaced by the same hue one step darker. FOUR DEFECTS IN THE MEASUREMENT ITSELF, each of which passed a plausible run first. (1) AN IDENTITY IS NOT A VOCABULARY AND EVERY PAINT IS A FUNCTION OF AN IDENTITY: the derivation reported the gate panel's rows as a status vocabulary whose members were the gate IDS, because a map from distinct keys is a function whatever it returns, so a status must PARTITION its group where an identity ENUMERATES it. (2) A GLYPH IS A MARK, NOT A CHARACTER IN A SENTENCE — scanning every non-letter in an element's text reported the gate rows as carrying an identical glyph, which was the middle dot in "BLOCK · fresh". (3) WHICH SIDE OF AN ELEMENT THE BACKGROUND IS ON DEPENDS ON WHAT IS PAINTED, and it was exactly backwards: text sits ON its own fill, a border sits at the EDGE, so a primary button's border was measured against the button's own accent fill and came back 1.00:1. (4) THE CANVAS IS A SYSTEM COLOUR, SO MEASURE IT — nothing declares a background on the root, getComputedStyle reports transparent, and both defaulting to white and refusing to answer are wrong; color: Canvas resolves what the agent is actually painting. THE HARNESS MUST NOT BE THE THING UNDER TEST, twice: the primitive catalogue plants a subtree with no contract above it in colours its own comment calls "nothing any palette would choose", and measured as part of the layer that probe is two competing accents and two unreadable pairs. DIST-PATCH VERIFICATION three ways: widening ACCENT_HUE_TOLERANCE_DEG to 180 so one family swallows every hue failed 4 controls, dropping CHROMATIC_CHROMA_FLOOR to 0.001 so every grey became a colour failed 4, and silencing namesTheStatus — the only channel reaching both readers — failed 5 including both coverage checks. Restoring returned 33/33. --> <!-- OPEN, and recorded rather than resolved: the layer's two ratified palettes disagree about warning. The shell's is 13.5° from its failure and the components' is 41°, so unifying on the shell's palette would DEGRADE the separation this item asks for. Both now resolve through the contract, so the choice is one host block rather than forty-five literals — but which values a host publishes is a product decision this audit deliberately does not make. -->

    • S7.22.f Verify each route opens on the actual working surface with one dominant workspace, readable dense information, concise utility copy, and contextual inspector/status. Evidence: hierarchy review. <!-- libs/oshun/workbench-kit/src/working-surface.ts (74 tests, 26 fault kinds) + apps/oshun/web/tools/working-surface-browser.mjs (60 controls, 20 surfaces, real Chromium). Evidence: evidence/v1-workbenches/workbench-working-surface.json — 20 surfaces, every one measured, every route opening on its own working surface, 80 regions ranked, narrowest workspace 57% of the 50% a majority requires. THE VEIN, AND IT WAS IN THE INSTRUMENT BEFORE IT WAS IN THE LAYER. The shell fixture read its route with useState(currentRouteId), which takes an initial value and nothing else, so a host re-rendering it in place — which is what every browser tool here does, because remounting would throw away the selection and focus those tools measure — got the route it asked for once and the first route forever after. S7.22.c mounted five routes and was handed one, five times: 16 of the 20 captures under evidence/v1-workbenches/first-screen/ were byte-identical duplicates of four images filed under twenty route names, and not one check in that run could have noticed, because nothing compared the surface against the route it was asked for. surface-does-not-name-the-route-it-opened is that missing join, asked before every other rule, since a run that measures the wrong page perfectly reports a clean layer. After the fix the captures are 18 distinct images of 20, and the two that still match are the detail route in both archetypes, which is what the archetype axis means here. A ROUTE WAS NOT A PROP EITHER, SO FOUR OF FIVE ROUTES SHOWED A FIFTH ROUTE'S OBJECTS. The fixture rendered one list everywhere and swapped its rows on the ARCHETYPE, so the claim this item makes could not be failed or passed. Each route now renders what it is named after and the archetype lands on the axis where it is genuinely a choice: an editor is read closely and a console is scanned, so S7.1.b's two archetypes take S7.5's two row heights. "OPENS ON" IS A CLAIM ABOUT A STATE NOBODY HAD PRODUCED. Every reading in this layer was taken after arriving, which is the same moment — but the only way to be wrong about opens is to measure after touching something, so a landing reading is taken with nothing touched and a second after exactly one interaction. The two together separate a surface that is empty from one that hides its work behind a choice, which a single reading cannot. Driving it also exposed a harness defect of the same family: re-rendering between surfaces kept the inspector's state, so the "arrival" reading of the second route showed the first route's object open in the panel. The page is loaded again for every surface (S7.2's same-document-navigation lesson, one surface on). DOMINANCE IS A MAJORITY AND A MAJORITY MAKES "ONE" A THEOREM. S7.1 published DOMINANT_WIDTH_SHARE = 0.5 against "the arrangement's content width" — a denominator never defined, worth twenty points of verdict depending on whether the rails are inside it — and the argument it judges was DECLARED by its caller and had never been measured off a painted document. Over the sum of what the shell paints, a majority is held by at most one region, so the item's word "one" needs no second rule that could disagree with the first. The number is taken from S7.1 rather than repeated, so the two cannot drift. THE BROWSER FOUND THREE DEFECTS SOURCE-READING DID NOT. (1) THE INSPECTOR RESIZE CONTROL RESIZED NOTHING: --wb-inspector-width was declared on the inspector and the rule that reads it is grid-template-columns on its PARENT, and custom properties inherit downwards only, so the region announced aria-valuenow moving 320 to 720 while its box stayed at 320px for every keystroke and every drag. (2) Once it worked, the shell's own flat 720px maximum put it into an arrangement its own thesis forbids: 1440 less a 56px rail, a 220px section rail and a 720px inspector leaves the workspace 444px, which is less than the inspector beside it. inspectorCeilingPx derives the ceiling the dominance claim implies from geometry the browser measures — 417px at the reference viewport — and the shell now clamps to it. (3) The fixture's heading used the shell's .workspaceTitle, which carries flex: 1 1 auto for a horizontal header; in a column-flex surface it grew along the column and measured 820x267, a heading taking a third of an editor. "READABLE DENSE" IS A TENSION AND NEITHER HALF IS A FONT SIZE. They pull on one knob in opposite directions, so a rule written against the knob is answered by the wrong half every time. Density is INFORMATION STATED per screen, its floor derived from two constants this layer already ships (S7.5's 44px comfortable row and S7.6's 220px column minimum give 9,680 square pixels as the most generous room one fact needs; four times that is 25 facts per megapixel). Readability is ROBUSTNESS — the criteria a crammed layout is the thing most likely to break — plus the 75-character measure S7.21.d already published. They are not opposites on every axis: a denser arrangement puts text in narrower columns and IMPROVES the measure, which is why one number would have been wrong in both directions at once. S7.22.e's VEIN, ONE TOKEN OVER. ADR-S4 publishes --wk-text-scale so a reader who needs larger text can have it, and primitives.module.css was the only file in the layer that consulted it: doubling it moved nine declarations, left sixty-seven fixed, and 23 of 58 text runs on a shell surface did not move at all. Sixty-seven font-size declarations across seventeen stylesheets now resolve through the token, so at the default scale of 1 not one rendered pixel moves and the control finally reaches the reader who needs it. MOST TEXT ON A WORKING SURFACE IS NOT COPY, so a word count over a rendered document measures the tenant's titles and reports them as the product being verbose. Every run carries a PROVENANCE and a run attributable to neither side is REPORTED rather than counted — the audit found three of those on the first run and they were real omissions. The budget is then per ROLE, because one budget over every string convicts the refusal that has to explain itself and acquits the button labelled "Go"; and the ordering is proved rather than asserted, with one planted sentence driving both arms (unattributable, so the register rule stays silent; the same run attributed to the product, so it fires). A DENSITY FLOOR NEEDS ONE EXEMPTION AND IT MUST BE MEASURED. An editor's information is what is IN it, not the labels around it, so a rule counting stated facts would ask every editor in the product to become a form. The exemption is not "this is an editor" but a measurement that the editable region really is the majority of what the surface drew — which is why the case it must not cover, a sparse page with a note box on it, stays convicted. AND CONTEXTUAL IS A RELATION, SO IT IS DRIVEN: two selections, joined to what the harness selected rather than to what the panel says it selected, plus the mirror — a panel showing a subject nobody chose is a second workspace wearing an inspector's label. HARNESS TRAP: browserHost prepends a button outside the shell for another tool's focus proof, and twenty pixels of furniture above a 100vh shell pushes the status strip off the bottom of every first screen. Removed for this run, and CONTROL 6 puts it back to prove that removing it was necessary rather than convenient. DIST-PATCH VERIFICATION three ways: raising DOMINANT_AREA_SHARE to 0.9 failed 6 controls, silencing marketingMarksIn failed 1, and flattening inspectorCeilingPx to the shipped 720 failed 2. Restoring returned 60/60. --> <!-- Re-ran after the fixture and stylesheet changes, all clean: workbench-shell-browser (every control and fault as declared), card-use-browser (1037 containers, 19 cards, every one driven), chrome-browser (54 effects, all conveying), composition-browser (520 nested boundaries, all warranted), signals-browser (1 accent family, 189/189 paints, 153/153 pairs at AA), first-screen-browser (20 surfaces, no pattern found), ship-boundary-scan (25/25) and packed-consumer-smoke (127 subpaths). The primitive chain is untouched: the text-scale migration deliberately skipped primitives/, which already consulted the token. --> <!-- OPEN, and recorded rather than resolved: the density floor is stated per SCREEN, so a surface listing a genuinely short collection can sit under it for a reason belonging to the record rather than to the layout. The run reports the measured figure for every surface rather than exempting the case, and the fixture's release list was lengthened to the length of its other lists rather than the floor being lowered — but a product surface over a four-record tenancy will read as sparse and that is a product decision this audit does not make. -->

    • S7.22.g Review all supported themes/densities/viewports and record approved exceptions with product/design/accessibility rationale. Evidence: signed design conformance report. <!-- NOT LOCALLY MARKABLE, AND THE REASON IS THE ITEM'S OWN EVIDENCE CLAUSE. §2.1: human governance approvals remain unchecked until performed by the named human, and an agent may prepare an engineering-readiness packet but may not fabricate approval. An EXCEPTION is not a failure somebody explained, it is one somebody with authority ACCEPTED — a rationale with no approver is a note — and the item names three rationales that are not interchangeable, because a product rationale ("this surface is not offered on handsets") cannot excuse an accessibility non-conformance: the reader that criterion protects is not a market segment. So the signature is the evidence, and nothing in this session may produce it. WHAT ALREADY EXISTS AS MEASURED EVIDENCE toward the packet, all of it re-runnable: the fourteen render contexts (three densities, the alternate theme, forced colours, both viewport ends, text zoom, type scale, reduced motion, RTL, bidi, long strings and font fallback) are declared in libs/oshun/workbench-kit/src/primitive-contexts.ts and measured over the primitive catalogue by primitive-contexts-browser.mjs; signals-browser.mjs measures every colour pair against the recorded WCAG 2.1 AA floors; working-surface-browser.mjs measures the layer at two viewport ends including the reflow width and the reader's text scale. WHAT THE PACKET STILL NEEDS: the cross-product run over the workbench SURFACES rather than the primitive catalogue, an exception register keyed to the failing cells and carrying an approver per rationale kind, and three signatures. Left unchecked deliberately; under-reporting is safer than a fabricated approval. -->

S8 — Collaboration, review, notifications, and activity#

  • S8.1 Extract approval workflows with role, quorum, sequence, separation-of-duties, and conflict-of-interest rules. Evidence: known-value policy tests. <!-- libs/oshun/workbench-kit/src/approval-workflows.ts (61 tests, 30 fault kinds), subpath ./approval-workflows. Rollup of S8.1.a-g below; every child is a section of the one module because the seven are one policy — an eligibility rule that could not see the quorum, or a quorum that could not see a recusal, would be seven policies that agree by accident. COMPOSE S2.13, DO NOT RE-DECIDE IT — eighth sighting of the S7.12 rule. signoff.ts had already settled what a SIGNATURE is: automation may schedule, remind, escalate, display and record a human gate and may never produce one; a service actor with the right role claim satisfies every role check ever written and is refused on its KIND before its role; a delegated actor may read everything and sign nothing; a step-up reused from login is a property of the session rather than of the judgement; and a rejection STANDS. None of it is repeated here, and stageStanding is the seam that hands the signature half back to evaluateQuorum. Gates run: packed-consumer-smoke (128 subpaths, no import-time side effects), router-vocabulary-scan (119 modules), the export-map generator (128 kit-owned entry points, 0 violations) and the API reference (19 pages). -->

    • S8.1.a Define workflow/version, target/revision, approval stages, verdicts, actors/eligible pools, deadlines, and terminal outcomes. Evidence: approval workflow schema. <!-- STAGE_KINDS (5), STAGE_VERDICTS (3), WORKFLOW_OUTCOMES (6), WORKFLOW_ROLES (4), StageDefinition, WorkflowDefinition, ApprovalTarget. TWO OUTCOMES ARE NOT VERDICTS: withdrawn is the proposer taking it back, and recording it as a rejection attributes a judgement to reviewers who never made it (S7.13's rule about the word a vocabulary cannot say); expired is terminal because a workflow that quietly stays open is one nobody is accountable for. A workflow with no stage is REFUSED rather than passed — it approves everything handed to it and reads in an audit as a workflow that ran. policyVersion is on the definition rather than beside it, because S8.1.g needs it pinned at the moment of signing. -->
    • S8.1.b Define role requirements and eligibility by tenant, object, lifecycle, assignment, specialist qualification, and delegated scope. Evidence: role policy tests. <!-- eligibility() over eight ELIGIBILITY_REASONS, one of which is a yes. EVERY ANSWER IS A REASON RATHER THAN A BOOLEAN: a check that says no without saying why produces a workflow nobody can unblock, and the person who has to unblock it is by construction not the person the rule was protecting against. The ACTOR KIND IS ASKED FIRST, for S2.13's reason — it is the only disqualification no configuration may switch off. redacted is on that list for a different reason from the other three: an erased identity cannot be shown to be independent of anybody. A DELEGATION IS SCOPED TO STAGES: somebody handed the legal review while a colleague is away has not been handed the release approval, and a deputy whose delegator is not in the pool is not in the pool either. -->
    • S8.1.c Define quorum math for eligible, completed, approving, rejecting, abstaining, recused, replaced, unavailable, and delegated reviewers. Evidence: quorum truth tables. <!-- POPULATION_RULES is the truth table written out as data and asserted row by row, over three columns: counts toward the quorum, stays in the pool, lowers the bar. THE DENOMINATOR IS THE WHOLE OF QUORUM MATH, AND THE TWO THAT LOOK ALIKE ARE OPPOSITES. A RECUSAL removes somebody from the pool AND lowers the bar, because it says this person MAY NOT judge and a quorum that outlives its own pool deadlocks with everybody behaving correctly. An UNAVAILABILITY removes them and lowers NOTHING, because it says they are NOT HERE TODAY, and a bar that falls for that is one anybody can lower with a calendar — which is how a two-of-three becomes a one-of-one in August. Known values: two-of-four with one recusal requires one and is satisfied; the same stage with one absence still requires two and is not. A DELEGATION MOVES AN AUTHORITY AND DOES NOT ADD A HEAD, so the count is over distinct authorities and two seats held by one authority are one approval with a count beside it. THE ADVERSARIAL PASS CAUGHT A DEAD RULE IN THIS MODULE'S OWN CODE, which is the shape it was written to hunt. unavailability-lowered-the-bar first compared readQuorum's reduction against readQuorum's own arithmetic — required = max(1, quorum - recusals) against quorum - recusals — and could never fire. THE CHECKED PARTY MUST NOT SUPPLY BOTH SIDES: a deployment that lowered its bar for absences did so before handing anything here, so appliedQuorum arrives as data and the rule compares it to what the populations justify. Split into two faults, both driven: the named case where absences account for the difference, and applied-quorum-below-what-the-populations-justify where nothing does. -->
    • S8.1.d Define sequential, parallel, conditional, repeat-on-change, escalation, and veto stages. Evidence: sequence tests. <!-- A VETO IS NOT A REJECTION WITH MORE WEIGHT. A rejection is counted and can in principle be answered by a revision; a veto ends the workflow whatever the count says, which is what makes it worth having and what makes handing it out a decision rather than a default — so a veto stage that names no holder is refused, a veto held by somebody outside the pool is refused (an authority to end a review held by somebody who cannot take part in it is one nobody in the room can see), and a veto the count outlived is reported. A CONDITIONAL STAGE THAT STATES NO CONDITION runs on everything, and a legal review that runs on every change is a legal review nobody reads. AN ESCALATION REACHED BEFORE THE STAGE IT RESCUES is the ordinary path with a different name, and A DEADLINE WITH NOTHING ON THE OTHER SIDE OF IT is a note in a ticket. Two stages may share a position only where both are parallel; a sequence in which two stages are both next has no order to be out of. deriveOutcome asks the veto first and the rejection second, and the outcome is DERIVED so a stored one that disagrees with its own stages is reportable. -->
    • S8.1.e Enforce proposer/reviewer/approver/publisher and other separation-of-duties constraints. Evidence: self-approval negative tests. <!-- THE TYPE CHECKER FOUND THE FINDING, AND IT INVERTED WHERE THE RULE BELONGS. The first draft read the authority off the signature, and a SignoffRecord CANNOT BE DELEGATED at all — admitSigner refuses acting-on-behalf-of-somebody-else and by is a HumanPresence no delegated actor can produce. So a separation check written over signatures is safe BY CONSTRUCTION and reads exactly like a check that had done its job. THE GUARANTEE IS ONE LEVEL DOWN AND THE HOLE IS ONE LEVEL UP: the workflow SEAT carries the authority it draws on separately from the person filling it, which is how a proposer's authority reaches an approval without their name appearing on it. The rule reads seats and joins signatures to them. AND IT IS OVER THE WHOLE HISTORY OF THE OBJECT, NOT OVER THIS REVISION. Somebody who proposed revision three and approves revision seven is still the proposer, and a check against the current revision's proposer says yes — which lets one person hold every role over an object's life, one revision at a time. publisher is in the vocabulary because the separation people forget is the last one: the person who approved a release pressing publish is the same person deciding twice, by which point everybody has stopped watching. -->
    • S8.1.f Define conflict disclosure, recusal, reassignment, confidential detail, override authority, and audit. Evidence: COI lifecycle tests. <!-- A CONFLICT'S DETAIL IS OFTEN CONFIDENTIAL AND ITS EXISTENCE NEVER IS. The quorum has to know two reviewers share a conflict without anybody reading what it is, which is why the shared thing is an identifier — S2.13's conflictId composed rather than re-decided — and a deployment that can only see the sharing by opening the detail is reported. A DISCLOSURE THAT CHANGES NOTHING is a paragraph nobody has to act on, so a disclosure with neither a recusal nor an override is a fault. AN OVERRIDE IS ITSELF A JUDGEMENT and needs its own separation: the conflicted party overriding their own recusal clears their own conflict, and the proposer overriding it is the person who wants it approved choosing who may judge it. Both are reported and an override by anybody else stands. -->
    • S8.1.g Preserve exact evidence/policy/target versions and invalidate stale approvals. Evidence: dependency change tests. <!-- S2.13 already invalidates on a moved TARGET and moved EVIDENCE; the third one is THE RULE ITSELF. An approval collected under a two-of-three does not satisfy a three-of-three, and nothing in the record can say which was in force unless the policy version is pinned at the moment of signing — so signedUnderPolicyVersion is recorded per signature and compared against the definition's current policyVersion. REPEAT-ON-CHANGE IS WHAT MAKES AN APPROVAL ABOUT A REVISION rather than about an object, and an object is whatever it was last edited into. The stage DECLARES whether a material change reopens it, because a stage that says no is saying the judgement survives an edit — a real and rare thing, a licence check over a fixed dependency set — and it has to be said rather than defaulted. A record added after a terminal outcome is reported: a decision that keeps accepting signatures is one that can be changed after it was announced. -->
  • S8.2 Extract anchored comment threads with stable target adapters. Evidence: anchors survive unrelated revisions or are explicitly orphaned. <!-- libs/oshun/workbench-kit/src/comment-threads.ts (101 tests, 66 fault kinds), subpath ./comment-threads. Rollup of S8.2.a-g below. THE VEIN: S1.4.e WROTE THE ANSWER DOWN IN FULL AND NOTHING HAS EVER PRODUCED ONE. AnchorResolution — three statuses, five orphan causes, a rebase confidence, and a sentence instructing callers to treat a low-confidence rebase as ambiguous — is a complete and correct description of the only question an anchor exists to be asked, and a repository-wide search finds it referenced in exactly two places, both of them its own declaration. StableAnchor has producers (S2.5 hangs two on every diff operation, S7.9 asks a block type for its own); the RESOLUTION had no constructor and no reader anywhere. The judge was written and the case never came. This module is the producer. AND WRITING THE PRODUCER FOUND THE CASE THE TYPE WAS SHORT, which is the whole argument for writing it: all five shipped causes are statements about the CONTENT, and none of them can say the thing that actually happens in a deployment, which is that this build has no adapter for this anchor kind and did not look. unsupported is now the sixth. A type nobody constructs is a type nobody discovers is missing a case. THE EVIDENCE CLAUSE IS A TAUTOLOGY OVER THE TYPE AND THAT IS THE POINT: "survive or be explicitly orphaned" is checkable only because nothing here can return a fourth thing, and the number that makes the summary worth reading is unexamined — how many anchors this build did not look at. COMPOSE, DO NOT RE-DECIDE — ninth sighting of the S7.12 rule, and it reached six modules this time: S1.4.e owns the anchor kinds and the resolution vocabulary, S2.5 owns the change types and the semantic / presentation split, S7.6 owns flicks and frame rates, S7.8 owns the three identity sources and which one survives an edit, S3.13 owns the sensitivity classes and which of them make existence the disclosure, S4.3 owns conceal outranking deny. Nothing on that list is restated here. Verification: npx tsc -p tsconfig.lib.json and -p tsconfig.spec.json clean, 101 tests, the full kit suite at 7280 across 125 files, eslint and prettier clean, packed-consumer-smoke at 129 subpaths, the router vocabulary scan clean, the stub-indicator scan at zero actionable hits, the NUL scan clean, both kit generators regenerated with the export map at zero violations and the API reference violation count unmoved from its committed 73. Five SOURCE INJECTIONS proved the spec is not vacuous — see S8.2.d. -->

    • S8.2.a Define thread/comment identity, target adapter/type, stable anchor, base revision, author/delegate, visibility, state, timestamps, and parent/reply. Evidence: comment schema. <!-- threadSchemaFaults over ThreadRecord/CommentRecord, 22 fault kinds. THE ANCHOR BELONGS TO THE THREAD, NOT TO THE COMMENT — a comment carrying its own would let a reply re-point the conversation, so a reader scrolling to the third one lands somewhere the first two were not talking about, and "resolve this thread" stops being a sentence that means anything. The record makes it unexpressible and the hole it leaves is reply-to-a-comment-in-another-thread, which is that defect arriving through a parent id. A COMMENT IS NOT A SIGNATURE AND THE DIFFERENCE RUNS THE OTHER WAY FROM S8.1's: S2.13 refuses a delegated SIGNER because a signature is a judgement nobody may hold on somebody's behalf; a comment is a STATEMENT and a deputy may perfectly well make one. What may not happen is that it READS as the principal's. And since ActorRef carries both identities by construction, a check over author alone is safe for free and reads exactly like a check that had done its job — the S8.1 shape again — so the record carries attributedTo, the identity the SURFACE displays, and the rule is there. presentedAsAutomated is the same argument for a gate bot: it may comment, and a machine remark rendered like a colleague's is answered like a colleague's. VISIBILITY IS A FLOOR ON WHO MAY BE TOLD, NEVER A GRANT: the vocabulary is ordered by breadth so a target's ceiling is arithmetic rather than a table somebody keeps in step, and reviewers sits between tenant and participants because a review note about a contributor's work is read by more people than the conversation has and fewer than the object has. ARCHIVE IS TERMINAL AND RESOLVE IS NOT — an answer can be wrong, so a resolution is reversible; an archived thread is off the object. --> <!-- THE ADVERSARIAL PASS OVER MY OWN MODULE FOUND THE EMPTY-ARGUMENT SHAPE, eleventh sighting: every rule that consults the target's policy was written policy !== undefined && …, so a thread whose target has no policy row had its visibility ceiling, its classification and its who-closed-it rule all silently not run and came back with no findings — indistinguishable from a thread that passed. An absent ceiling is not a ceiling of unlimited; it is a check that did not happen, and thread-on-a-target-with-no-declared-policy says so. --> <!-- SEPARATION OF DUTIES REACHES THE COMMENT SURFACE TOO: thread-resolved-by-the-reviewed-party-with-no-answer is S8.1's rule one floor down — the person whose work is being reviewed closing the reviewer's mark with nothing said back is one person deciding twice, and the mark disappears with no record that anybody agreed. Reported rather than refused, because a typo the author simply fixes is the ordinary case; a reply of any kind clears it. -->
    • S8.2.b Register entity/field/block/range/temporal/graph/canvas/hierarchy anchor resolve/rebase/context adapters. Evidence: adapter contract tests. <!-- buildAnchorRegistry over AnchorAdapterDeclaration, 8 fault kinds, all eight of S1.4.e's anchor kinds required. THE REGISTRY IS BUILT BY THE DEPLOYMENT RATHER THAN FROZEN INTO THE KIT. A kit shipping it as a constant is one in which no deployment can ever add the temporal adapter, so every temporal anchor everywhere is permanently unsupported with nothing anybody can do about it — the frozen-capability-registry shape, caught before it was written. AN ADAPTER'S IDENTITY SOURCE IS THE CEILING ON EVERY ANSWER IT CAN GIVE, and this is the finding that makes the declaration load-bearing rather than decorative. S7.8 already found the three ways a thing can be named and that only an assigned id survives an edit; an adapter addressing by POSITION or by CONTENT HASH cannot distinguish "it survived" from "something else is in that slot now", so over any structural edit its best honest answer is ambiguous whatever its rebase arithmetic believes. resolveAnchor applies it before dispatching to any adapter, and the test shows the same edit rebasing cleanly on stable ids and orphaning on positional ones. A CONTEXT KIND A COORDINATE SPACE CANNOT PRODUCE IS A PROMISE THAT DELIVERS WHATEVER WAS NEAREST: an entity anchor has no excerpt, there is no passage, only the whole object; a temporal anchor has no structural path. CONTEXT_KINDS_FOR_ANCHOR is the map and an adapter claiming outside it is reported. And an adapter that REBASES what it cannot RESOLVE is moving an anchor to a place nobody checked. --> <!-- The version field is S2.5's, for S2.5's reason: a cached resolution computed under an older rebase rule is not a resolution, it is last week's answer to a question whose rules moved. -->
    • S8.2.c Preserve original excerpt/thumbnail/structural context and exact revision without leaking protected content. Evidence: context security tests. <!-- contextFaults over AnchorContext, 9 fault kinds. A CONTEXT SNAPSHOT IS A COPY THAT OUTLIVES ITS ACCESS CONTROL. The excerpt the item asks for is the sentence the comment is about, stored in the COMMENT table — a different table, with different permissions, that no retention sweep over the document will ever visit. So it is bounded against a kit-owned ceiling (past some length it stops being a pointer and becomes a second copy of the document), it carries the class of what it was taken from, it is re-checked against the target's CURRENT protection rather than the one at capture, and when the source is erased it goes too: erasing a document and keeping the passage in a comment is not erasure, it is a copy nobody looked for. AND THE MIRROR DEFECT IS JUST AS REAL AND POINTS THE OPPOSITE WAY. A context RE-READ from the current revision makes the comment look like it was about text that did not exist when it was written. Never refreshed, sometimes withheld — and the two do not conflict, because withholding is not refreshing. The tell for a refresh needs no ordering and none is available (a revision id is not a number): the captured revision being the one BEING READ, while the anchor was taken at another, is the whole of it. A WITHHELD CONTEXT MUST SAY SO. An excerpt silently blanked renders as a comment about nothing rather than as a refusal, which is S1.6.e's rule about fabricated fallbacks reaching a review surface. FOR A PRESENCE-DISCLOSING CLASS THERE IS NO SAFE EXCERPT AT ALL — S3.13's protected-source, health, biometric, where the existence of the row IS the disclosure — so a stored context answers exactly the question the object's own access control refuses, sealed or not. --> <!-- The budget is counted in CODE POINTS rather than UTF-16 units, and the test pins it: ten astral characters are twenty units, so a unit-counting budget convicts every emoji and lets twice as much Latin text through as it meant to. --> <!-- NOT MEASURED HERE, said rather than implied: nothing in this module reads a document. Whether the excerpt a host captured is really the passage under the anchor is the host's to prove; what is checked is the REVISION it was captured at, the CLASS it was taken from, its SIZE, and whether the adapter that owns the anchor preserves that kind at all. -->
    • S8.2.d Rebase anchors across unrelated edits, moves, split/merge, timing conversion, and schema migration where unambiguous. Evidence: rebasing fixtures. <!-- resolveAnchor + eight per-kind resolvers + applyTextEdit + applyConfidenceThreshold, 33 known-value rebasing tests. THE EDITS ARE S2.5's VOCABULARY, NOT A SECOND ONE — ChangeType and DiffSignificance are imported, and the coordinate SPACES are deliberately fewer than the anchor kinds: text covers both block and range, because a model with one arm per kind makes the caller describe the same deletion twice and lets the two descriptions disagree. A RANGE ANCHOR IS ABOUT ITS CONTENT, NOT ABOUT ITS COORDINATES, which settles the boundary case — and the boundary case is the COMMON case, somebody typing at the caret where a highlight begins. THE TEST FOUND WHAT THE OBVIOUS SHAPE CANNOT SAY: the first draft shared one shiftOffset(o) between the two ends, which is what everybody writes, and it treats a range as two independent points. A RANGE'S TWO ENDS HAVE OPPOSITE GRAVITY: text arriving AT the start pushes the start right, text arriving AT the end does not push the end right, because a range is the content between its ends rather than the ends. Under one map an insertion at the start either extends the selection at both ends or at neither, and the comment ends up about a sentence the author never highlighted. Known values: insert 4 at 10 into 10..20 gives 14..24 (ten code points, the same ten); insert 4 at 20 gives 10..20 unchanged. THE TYPE ALREADY REFUSED THE ROUNDED ANSWER AND NOBODY HAD EVER MADE IT SAY SO. parseStableAnchor rejects a fractional frame with the words "rounding it silently moves the anchor" — a rule written for S1.4.e and never exercised, because nothing had ever converted an anchor. So a note on frame 100 conformed from 24fps to 25fps has NO REBASE IT IS ALLOWED TO RETURN: 2 940 000 000 flicks over a 28 224 000-flick frame is 104.1666…, and it orphans as ambiguous for a human to pick. The same conversion at frame 24 lands on 25 exactly and rebases at confidence 0.96, which is how much of the original frame the new one covers. AN ANCHOR MUST NOT MOVE FOR A CHANGE THAT WAS NOT A CHANGE TO THE WORK. S2.5 separated presentation from semantic so a review would not report forty-seven changes for an afternoon of tidying a graph; the same operations reach here, and a comment reporting itself rebased because somebody dragged a node teaches its reader that "rebased" means nothing. A SPLIT IS THE OPERATION THAT DECIDES THE WHOLE SECTION'S CHARACTER: a BLOCK anchor over a split paragraph orphans, because no arithmetic picks a half; a RANGE anchor over a split OUTSIDE its selection follows the half that holds it with the offsets adjusted; and a range split THROUGH its selection orphans, because the range now spans two blocks and a range anchor cannot say that. The item's word is orphan rather than move, and coarsening a canvas part up to its object is a silent move dressed as a rescue. A SCHEMA THAT MOVED WITH NO DECLARED MAPPING IS migrated, derived from RevisionRef.schemaVersion rather than from a flag: the field may still be spelled the same and mean something else, and no evidence available here can say otherwise. THE CONFIDENCE IS AN UPPER BOUND AND IT IS MONOTONE — the smallest the range has ever been is the most of it that can have survived, because an insertion cannot restore what a deletion took. Deleting 15..20 out of 10..20 gives 10..15 at confidence 0.5. --> <!-- applyConfidenceThreshold is S1.4.e's own sentence — "below a caller's threshold, treat as ambiguous" — carried out for the first time, and it is a separate function rather than a parameter because the threshold is a PRODUCT decision: a legal review wants every doubt in front of a human, a copy edit does not, and one resolution serves both. --> <!-- A programming error is thrown rather than orphaned: S1.4.e says an anchor from another entity is a wiring bug, and returning orphaned would hide a bug behind a product state the UI already knows how to render. --> <!-- VERIFIED BY FIVE SOURCE INJECTIONS, since a headless module has no dist-patch ritual: sharing one gravity between the two range ends failed 3 tests, rounding the timing conversion failed 1, reporting a missing adapter as deleted failed 3, dropping the identity-source ceiling failed 1, and letting presentation edits through failed 1. -->
    • S8.2.e Mark deleted/replaced/ambiguous/permission-hidden/unsupported anchors orphaned with reason and original context rather than moving silently. Evidence: orphan tests. <!-- orphanFaults over OrphanPresentation, 7 fault kinds, plus the addition to S1.4.e that writing the producer forced. THE ITEM NAMES FIVE CAUSES AND THE TYPE SHIPPED FIVE AND THEY WERE NOT THE SAME FIVE. AnchorResolution had deleted / replaced / migrated / permission-hidden / ambiguous; the item asks for deleted / replaced / ambiguous / permission-hidden / UNSUPPORTED. All of the shipped five are statements about the CONTENT. unsupported is a statement about the DEPLOYMENT: this build has no adapter for this kind of anchor and DID NOT LOOK. Reporting that as deleted tells a reviewer their clip is gone when what is missing is the code that would have gone to find it, which is the fabricated answer S1.6.e forbids — so ANCHOR_ORPHAN_CAUSES is now six, and unsupported-reported-as-a-fact-about-the-content is the fault that catches a surface collapsing them. A TYPE NOBODY CONSTRUCTS IS A TYPE NOBODY DISCOVERS IS SHORT A CASE. The gap had been sitting in the file since S1.4.e and is only visible from the producer's side. AN ORPHAN IS NOT A DELETION: the thread stays, readable, and it must keep the ORIGINAL anchor and the ORIGINAL context or no human can put it back — a comment reading "this sentence is wrong" with no sentence is a remark about a document nobody can reconstruct. The one cause exempted from the context requirement is permission-hidden, where there was never anything this reader could be shown. A REFUSAL THAT EXPLAINS ITSELF PRECISELY IS A DISCLOSURE. Rendering "there is a comment here you may not see" over a protected-source target answers exactly the question the class exists to refuse, so for S3.13's presence-disclosing classes S4.3's conceal outranks deny reaches the comment layer: the thread is ABSENT rather than explained. AND THE MIRROR IS A FAULT TOO — a surviving anchor advertised as broken sends a reviewer looking for a deletion that never happened. --> <!-- The spec proves all six causes are REACHABLE from real inputs rather than merely declared: a removed block, a replaced one, a moved schema with no mapping, a reader who may not see the target, a split paragraph, and a registry with the canvas adapter left out. --> <!-- summarizeResolutions builds its per-cause counters FROM the vocabulary rather than writing them out, so a seventh cause added to S1.4.e cannot land here as an undefined + 1 that turns a count into NaN and a summary into a blank. unexamined is surfaced separately because it is the one number an operator needs: how many anchors this build did not look at. -->
    • S8.2.f Implement create/reply/edit-window/resolve/reopen/archive/link/ navigate with authorization, expected revision, notifications, and audit. Evidence: lifecycle tests. <!-- lifecycleFaults + editWindowState over OperationAttempt, 14 fault kinds, all eight operations. AN EDIT WINDOW IS ABOUT THE READER, NOT THE WRITER, and every implementation that is only a clock has the rule backwards. The reason a window exists at all is that somebody may already have ANSWERED; once they have, editing rewrites the thing they answered and their reply becomes a response to words nobody wrote. So it closes on the first reply as well as on the clock, whichever comes first — five minutes in, well inside any clock, an answered comment is closed — and the two closures are reported separately, because a product offering "edit anyway, marked as edited" needs to know which one it is arguing with. A WITHDRAWN reply does not close it, since there is no longer an answer to invalidate. ARCHIVE IS TERMINAL FOR BOTH DIRECTIONS: a reply arriving after it and a reopen of it are both faults, which is S8.1's rule about a record added after a terminal outcome — a decision that keeps accepting input is one that can be changed after it was announced. NAVIGATION IS THE OPERATION WHOSE FAILURE IS INVISIBLE. Asked to go to an orphaned anchor a surface either does not move or moves to whatever is now at those coordinates, and BOTH look like the product working, so the refusal has to be shown. It is also the one operation that must not be written to the mutation audit: a read filed as a change makes every "who changed this" query answer with the people who looked. A LINK THAT PINS NO REVISION opens against head and shows different text than the notice that carried it described. THE NOTICE IS THE LEAK, NOT THE LINK — S7.19's first decision, and a comment body is the one string in this product that lands on a lock screen, so a notification quoting it is reported whatever the thread's visibility says. --> <!-- Every mutation carries an expected revision, which is S2.3's rule rather than a new one: there is no nullable precondition anywhere in this kit because a caller who does not know what they are anchoring to must not be able to say so. navigate and link are exempt because they change nothing. --> <!-- THE ADVERSARIAL PASS FOUND THE EMPTY-ARGUMENT SHAPE HERE TOO: an edit attempt naming no comment in the thread has no window to be inside, so the window rule never ran and the attempt reported clean — the same defect as the missing policy row in S8.2.a, in a different section of the same file. edit-that-names-no-comment closes it. -->
    • S8.2.g Provide keyboard/screen-reader bidirectional navigation between anchor and thread. Evidence: accessible review journey. <!-- reviewJourneyFaults over ReviewJourneyObservation, 8 fault kinds, the shape S7.4's treeJourneyFaults established for a declared journey. BIDIRECTIONAL MEANS THE ROUND TRIP ENDS WHERE IT STARTED. Two one-way links are two one-way links, and the rule that checks only for their existence passes the failure that actually happens: a "back to the text" that returns focus to the top of the document has ended the review for everybody not using a mouse, and both links work perfectly. So the journey is measured as a round trip and the position before is compared against the position after. THE ANNOUNCEMENT MUST SAY WHAT THE COMMENT IS ATTACHED TO, because the sighted reader is told that by the highlight and this reader is told it by nothing — "comment, 2 replies" is a different amount of information depending on who is receiving it. And an ORPHANED thread announced as though it still points somewhere is the silent move of S8.2.e reaching the one reader who cannot see that it did not move. A MARK THAT IS ONLY A COLOUR is S7.22.e's second reader arriving here: somebody who can see the screen and cannot separate its hues. AND THE FOCUS ORDER FOLLOWS THE DOCUMENT RATHER THAN THE THREAD LIST — a review read in creation order asks its reader to hold the document in their head while the sighted reader reads it off the page. --> <!-- Headless, and deliberately: the parent's evidence clause is about anchors surviving revisions, and S7's browser-tool recipe belongs to the items whose evidence is a rendered surface. What is checked here is that a host which DECLARES a journey has declared a coherent one; driving a real review surface belongs to the workbench that ships one. -->
  • S8.3 Add suggestions distinct from comments and direct edits. Evidence: accept/reject creates attributed revisions. <!-- libs/oshun/workbench-kit/src/suggestions.ts (101 tests, 113 fault kinds), subpath ./suggestions. Rollup of S8.3.a-g below; every child is a section of the one module, because a proposal whose schema could not see its own acceptance rule would be two policies agreeing by accident. THE VEIN — AN ACCEPTED SUGGESTION HAS TWO PARTIES AND THE REVISION RECORD HAD ONE SLOT. A comment says a change is needed, an edit makes it and takes the responsibility, and a suggestion is the third thing: somebody who can see what the change should be, writing it out, without the standing to commit it. It is the only one of the three with two people in it. RevisionRecord carries one author, which was enough for exactly as long as nothing could put somebody else's content into a revision — and then one field has to answer both "who changed this?", whose answer is the person accountable for the object being in this state, and "whose words are these?", whose answer is the person who wrote them. AND THE LOSS IS NOT THE CREDIT, IT IS A CONTROL. S8.1 checks separation of duties over the WHOLE HISTORY of the object, against roleHistory.proposer, documented as "every authority that has ever held each role on this object, all revisions" — and a host can build that list from nowhere but the revision history. So the laundering path S8.1 closed at the SEAT reopens one level down at the CONTRIBUTION: write the change as a proposal, have a colleague admit it, then approve it yourself. Every name on the way is a different name and the words are yours. revisions.ts now carries RevisionContribution, checkHistoryIntegrity gained the two ways it can be a fiction, and proposingAuthorities() is the seam that hands S8.1 the list it was always documented to receive — with a negative control asserting that the same history WITHOUT a contribution hands back only the accepter. ActorRef CANNOT EXPRESS IT AND STATES THE INVERSE: delegated's onBehalfOf is somebody else's AUTHORITY over your act, and this is your own authority over somebody else's words. The two are not the same relation pointed the other way; they are different relations, and the union has one of them. COMPOSE, DO NOT RE-DECIDE (ninth sighting of the S7.12 rule): S2.5 owns the change vocabulary and the semantic/presentation split, S1.4.e owns the anchor and S8.2 owns its resolution, S2.4 owns lineage and immutability, S2.10 owns submit-versus-apply, S4.3 owns the object decision, S7.8 owns dirty state, S2.13 owns what automation may never do. Verification: 9 SOURCE INJECTIONS, each caught, against a control arm that passes — a null lineage base (5 fails), a contribution never written (3), a rebase trusted without the threshold (3), the proposer-authorization check guarded on its own presence (2), the dirty-comparison move unchecked (2), a withdrawal compared at the actor id rather than the authority (1), a partial acceptance left live (1), the path collision underived (2), and automation allowed to admit (3). Plus a coverage test asserting all 113 declared fault kinds are reachable — which is what found that this module's OWN acceptance-whose-revision-is-authored-by-the-proposer was guarded by the condition that would have made it fire, so it could never fire and read exactly like a rule that had passed. The check written to hunt that shape, built with it. -->

    • S8.3.a Define suggestion identity, typed operation/patch, target/base revision/anchor, proposer, rationale, evidence, state, and expiry. Evidence: suggestion schema. <!-- SuggestionRecord + ProposedOperation + SuggestionEvidence, suggestionSchemaFaults over 27 kinds. THE PATCH IS S2.5's VOCABULARY RATHER THAN A SECOND ONE — the change types, the semantic/presentation split and the DiffValue refusals are the ones a diff already speaks, so a client that can render a review can render a proposal without a second renderer that disagrees with the first about what moved means. And from is the value AS THE PROPOSER SAW IT, which is the field that makes a stale proposal detectable at all: if the target no longer holds it, somebody has been here since, and admitting to silently discards their work. SEVEN STATES, AND accepted IS NOT partially-accepted — the first says the document now contains everything the proposer asked for and the second says it contains some of it, which is a different thing to have agreed to and a different thing to be told your proposal received. EVIDENCE IS A POINTER AND A RATIONALE IS AN ARGUMENT. Every evidence kind names something outside the proposal that the reader can go and check, so an entry with an empty reference is a second rationale wearing evidence's name — and a reviewer told a proposal is evidenced has been told something that is not true. The note may be empty; the reference may not. AN EXPIRY IS A PROMISE TO THE PROPOSER, so one that precedes the proposal means it was never open for anybody to read, and an expired record with no declared expiry means something closed it on a clock nobody published. ACCEPTANCE AND ENDING ARE TWO NULLABLE FIELDS RATHER THAN ONE OUTCOME UNION, because they are not alternatives of equal weight: an acceptance produces a revision and has to point at it, and the other four produce nothing and must be provably unable to. Their coherence with state is checked rather than assumed, which is the only way a stored row saying rejected while carrying an acceptance is ever found. -->
    • S8.3.b Render proposed change and semantic diff without applying it to authored truth or dirty state. Evidence: proposal isolation tests. <!-- PreviewObservation + isolationFaults, 14 kinds. THE SECOND HALF OF THE ITEM IS THE ONE THAT BITES. Nobody writes a revision to render a preview — that fault is real and rare. Everybody renders the preview by putting the proposed text into the editor's buffer, and from that moment it is indistinguishable from the author's own unsaved work: the next save, or the autosave nobody pressed, commits somebody else's proposal under the author's name with no acceptance anywhere in the record. S7.8 HAD ALREADY CHOSEN THE PROPERTY THAT SURVIVES A PROJECTION and did not know it. Dirty as a COMPARISON AGAINST THE BASE can ignore something that was never in the base; dirty as a COUNT OF KEYSTROKES cannot tell the author's typing from the projection's insertion, so it either loses their work or attributes the proposal to them. The counted-keystrokes source is a fault here for a reason S7.8 never had to state. THE MIRROR OF THE DIRTY FLAG IS AS REAL AS THE FLAG: a projection that makes a clean document dirty warns the author about unsaved changes they did not make, and the one prompt that could have saved them is now the one they have learned to dismiss. A document that was ALREADY dirty is left alone, or the finding trains its reader to ignore it. A PROPOSAL IS NOT A LOCK. Proposals are the cheapest thing in the system to create, so a preview that takes the document away from its author lets anybody who may propose stop anybody who may write. AND THE SEMANTIC DIFF FAILS BOTH WAYS: counting a presentation part reports tidying as change (S2.5's whole reason), and omitting a semantic one means the reader who reads the summary instead of the document is deciding without the change in front of them. -->
    • S8.3.c Authorize creation independently from direct edit and approval rights. Evidence: permission matrix. <!-- PermissionMatrix + CapabilityRule + permissionMatrixFaults, 14 kinds. "INDEPENDENTLY" FAILS IN BOTH DIRECTIONS AND ONLY ONE OF THEM LOOKS LIKE A BUG. A deployment that CONFERS propose on everybody holding edit-directly has renamed a right rather than created one, and the population the feature exists for — the people who can see what the change should be and may not make it — still has nothing. A deployment that REQUIRES edit-directly in order to propose has made the feature unavailable to exactly that population, so nothing is ever proposed. That is why the rule carries a RELATION rather than being a graph edge: confers and requires are not inverses and they break the item's claim in opposite directions. THE SHARP EDGE IS THAT ACCEPTANCE MUST BE A SUBSET OF DIRECT EDIT. An accepter who may not edit is an escalation channel, not a safeguard: a proposer puts any change at all in front of them, and admitting it writes content into the document that nobody trusted to write it ever agreed to. The reading that feels safer — "they can only accept, not author" — is the one that is wrong. A GRANT SAYS WHETHER SOMEBODY MAY EVER, AND AN OBJECT DECISION SAYS WHETHER THEY MAY HERE. So an exercise carries the S4.3 disposition, and one with null is the empty-argument shape made a fault rather than a silent pass (twelfth sighting): a system holding only the matrix authorizes the reader of one tenant's proposal against another tenant's rules. Both disagreements are reported, and the one that cannot be safe is named as such. -->
    • S8.3.d Revalidate target revision, anchor, schema, authorization, validation, and dependencies at acceptance. Evidence: stale suggestion tests. <!-- revalidate → RevalidationReport, the item's six checks by name, 14 fault kinds. THE AUTHORIZATION REVALIDATED AT ACCEPTANCE IS THE PROPOSER'S, NOT THE ACCEPTER'S — and this is the check nobody writes, because at the moment of acceptance the proposer is not the one making a request. The accepter's rights get checked because they are acting; the content is coming from somebody who may have left the team, had a grant revoked, or been moved out of the workspace, and admitting it then is admitting content from outside the trust boundary through a door that has closed. A MOVED DOCUMENT IS A FACT AND NOT A REFUSAL, which is why the verdicts are four-valued. "The document has moved under this proposal AND the anchor still resolves" is exactly the sentence the accepter needs and exactly the one a boolean cannot say. not-checked is the fourth, because a build that did not look must not report what a build that looked and was satisfied reports. ACCEPTANCE IS S8.2's applyConfidenceThreshold FINDING ITS SECOND CALLER, at a higher floor, and the asymmetry is the argument: a comment shown in the wrong place is read by somebody who can see it is the wrong place, and a patch applied in the wrong place is a change to the document that looks exactly like a change somebody meant. A SCHEMA MIGRATION IS REFUSED RATHER THAN REPORTED, because the paths in the patch are the old shape's paths and a migration is precisely the event that changes what one of them means. Nothing revalidates that except a person. A DECLARED DEPENDENCY WITH NO STATE IS THE EMPTY-ARGUMENT SHAPE AGAIN and is reported rather than skipped — a patch whose citation was deleted last week still applies cleanly and says something nobody can stand behind. AND THE WHOLE THING IS TAKEN AT APPLY, not at submit: everything between the two is a change nobody revalidated against, and the window is exactly as long as the queue is deep. -->
    • S8.3.e Accept whole or supported atomic parts by creating an attributed new revision linked to proposer and accepting actor. Evidence: acceptance lineage tests. <!-- acceptSuggestion + partSelectionRefusals + acceptanceFaults + proposingAuthorities, 20 kinds. THE VEIN, and see the parent. revisions.ts gained RevisionContribution (contributor, proposalId, amended), contribution joined IMMUTABLE_FIELDS — an attribution that can be rewritten is one that can be removed by whoever the history was going to embarrass — and checkHistoryIntegrity gained two kinds. A revision that credits its OWN AUTHOR manufactures a second party, and every count taken over that history gets a head there is nobody to fill; a contribution naming no proposal has lost the only thing that makes the credit checkable. amended IS ON THE RECORD BECAUSE AN AMENDED PROPOSAL IS STILL THE PROPOSER'S CONTRIBUTION AND IS NO LONGER THEIR TEXT, and a reader told only "contributed by X" will quote them on words they did not write. THE LINEAGE FIELD FOR THIS WAS ALREADY THERE. Lineage.linear separates parent — where the content is going — from base — what the change was expressed against — and a proposal admitted after the document moved on is that separation's own case. An acceptance writing base: null throws away the only record of what the proposer could see, and after that revision nothing anywhere else has it. "SUPPORTED" IS LOAD-BEARING IN THE ITEM'S OWN SENTENCE. Two couplings and only one is declarable: a DECLARED requirement is a MEANING — this sentence only reads correctly after the one before it was rewritten — and nothing but the proposer can know it, while a DERIVED one is a collision the paths can see. Admitting one of a coupled pair produces a document neither party proposed, signed by both. A PARTIAL ACCEPTANCE IS TERMINAL, because the parts left behind were judged against a document the acceptance itself replaced; leaving them live invites a second accepter to admit a change revalidated against nothing. SELF-ACCEPTANCE IS COMPARED AT THE AUTHORITY, NOT THE ACTOR ID — a deputy admitting the proposal of the principal they act for is two names and one judgement, which is S8.1's seat walk-around arriving at the contribution. Where a deployment permits it, the result is produced as a DIRECT EDIT with no contribution written, because "X's proposal was accepted by X" is the sentence that puts the phantom head into the count. AND AUTOMATION MAY PROPOSE AND MAY NOT ADMIT (S2.13 composed, not re-decided): a linter that can write the fix it found is the best case for this whole feature, and deciding that somebody else's words belong in the document is a judgement that needs somebody who can be asked about it. -->
    • S8.3.f Reject/withdraw/expire/supersede with reason while retaining immutable proposal/audit and no content mutation. Evidence: rejection tests. <!-- SuggestionEnding + endingFaults, 15 kinds. FOUR ENDINGS, FOUR DIFFERENT FACTS, AND ONE OF THEM HAS NO ACTOR. The item says "with reason" and the type deliberately does not give expired one: nobody decided an expiry, the clock did, so a shape demanding a human reason there is answered by whatever string the sweep job was written with, and every expired proposal in the system carries the same invented sentence in the field a reader goes to for a person's thinking. THE S7.13 RULE, A THIRD TIME, AND IT IS CHECKABLE IN BOTH DIRECTIONS. A rejection recorded on an ending the PROPOSER made is a withdrawal filed as a refusal — it puts a rejection on the record of somebody who was never refused. A withdrawal recorded on an ending SOMEBODY ELSE made files a reviewer's judgement as the proposer changing their mind, and the reviewer who made it cannot be asked why. Compared at the authority, so a deputy withdrawing on the proposer's behalf is correctly the proposer. A SUPERSESSION IS THE ONLY ENDING THAT POINTS SOMEWHERE, so it is the only one that can point wrong: at nothing, at another object, at a proposal that PREDATES it (the older thinking presented as the newer), or at one that dropped the discussion — and that last is the expensive one, because every objection raised against the first proposal has to be raised again by somebody who remembers it. DECLINING A CHANGE IS THE ONE OPERATION THAT MUST LEAVE THE DOCUMENT EXACTLY AS IT FOUND IT, so the content hash and the head are both compared across the ending, and the proposal's own operations, rationale, evidence, proposer, base, anchor and timestamp are frozen: the words that were declined are the evidence for the decision, and a proposal that can be tidied after it is judged is one nobody can be held to. -->
    • S8.3.g Keep comment discussion and suggestion state linked but semantically distinct. Evidence: thread/suggestion tests. <!-- LinkObservation + linkFaults, 9 kinds. THE COUPLING FAILS IN BOTH DIRECTIONS AND BOTH ARE SHIPPED. Resolving the discussion must not admit the change — people agree in words all the time without anybody pressing the button, and a system that presses it for them has recorded an acceptance nobody made. Admitting the change must not settle the discussion — the change going in is not an answer to the argument about it, and the objection that was open is now filed as settled. AND THE FAILURE IS NOT ONLY COUPLING. An archived thread with a live proposal hanging off it takes a pending change off every surface the discussion appears on while it is still waiting for somebody to decide it; one control standing for both states leaves a reader unable to tell an unanswered question from an unadmitted change, which need different people to act. S8.2's EIGHTH DECISION REACHES HERE UNCHANGED — two one-way links are two one-way links — and the two directions are separate faults because they are separate fixes: a reviewer at the change with no way to the argument has lost the half they need, and everybody arguing about the change without the change in front of them has lost the other. A LINK NAMING A THREAD WITH NO STATE IS THE EMPTY-ARGUMENT SHAPE, and the test asserts it is the ONLY finding returned: every rule below reads that state, so guarded on its presence they would all silently not run and the observation would come back clean. A proposal argued NOWHERE is not a fault — a one-line fix needs no discussion — and inventing one would make every trivial proposal a finding. -->
  • S8.4 Add mentions with tenant-scoped identity resolution. Evidence: unauthorized identities cannot be discovered. <!-- libs/oshun/workbench-kit/src/mentions.ts (104 tests, 66 fault kinds), subpath ./mentions. Rollup of S8.4.a-f below; every child is a section of the one module, because a picker whose scope rule could not see the resolution that happens at submission would be a control on a door beside which there is no wall. THE HOLE HAS TWO FACES AND THEY ARE ONE HOLE. First: THIS KIT CAN DECIDE WHETHER YOU MAY SEE AN OBJECT AND HAS NEVER BEEN ASKED WHETHER YOU MAY SEE A PERSON. ObjectMetadata carries an owner and a list of assignees and ENTITY_KINDS lists twelve things a row can be, none of which is a human being — people appear in this kit only as attributes of things. So every visibility decision ever written here has the shape "given this object, may this actor read it", and a mention search asks the one question that shape cannot hold, which is who is there. THAT IS WHY THE ITEM'S EVIDENCE CLAUSE SAYS DISCOVERED RATHER THAN READ: the harm is enumeration, and an object model answers reads. authorization.ts gets halfway, since conceal and presence-is-the-disclosure already exist so that a 403 does not appear where a 404 belongs — but A CONCEALMENT PROTECTS A NAMED SUBJECT AND A SEARCH HAS NO NAMED SUBJECT. Nobody asks whether they may see one person; they type three letters and receive a set, and what leaks is not a refusal but the difference between two answers that both succeeded. Second: A MENTION IS THE FIRST IDENTITY REFERENCE IN THIS KIT THAT LIVES IN CONTENT. Every identity mechanism here operates on FIELDS — ActorRef is a column on a revision, parseActorRef refuses display data by name so identity cannot drift into a display string, and deletion.ts's residueIn hunts survivors by walking Object.entries(record). A mention is a substring in a body, so the sweep that redacts every actor field on every row will not touch it, and after an erasure the thread still says the name in prose. The redacted ActorRef variant — which exists so history can still say a mutation had an actor — has no counterpart anywhere in content, and identityResidue is that counterpart. COMPOSE, DO NOT RE-DECIDE (tenth sighting of the S7.12 rule): S4.3 owns conceal versus deny, S3.2 owns the tenant and workspace scoping, S4.2 owns the principal mapping and the rule that the kit holds no roster, S8.2 owns the thread a naming lands in and the finding that a context snapshot outlives its access control, S7.19 owns the notification taxonomy and keeps mention among the kinds a storm collapse may never swallow, S7.13 owns withdrawal, and identity.ts owns what an id may not be made of. Verification: 14 SOURCE INJECTIONS, each caught, against a control arm that passes; npx tsc -p tsconfig.lib.json and -p tsconfig.spec.json clean; the full kit suite green at 127 files / 7492 tests; export map 0 violations and 131 entry points; API reference 21 -> 22 pages at the unchanged 73-violation baseline; packed-consumer smoke and router-vocabulary scan both OK; stub-indicator scan and NUL scan clean. -->

    • S8.4.a Define mention token/reference independent from display name and support human/team/role/on-call aliases as policy permits. Evidence: mention schema. <!-- MentionReference / MentionOccurrence / mentionSchemaFaults, 12 fault kinds. THE FOUR ALIAS KINDS DIFFER IN WHAT THEIR RESOLUTION IS A FUNCTION OF, and RESOLUTION_DEPENDS_ON writes the axis down rather than leaving it in four branches: nothing, membership, the object, the instant. A human names a person. A team names a set that changes. A role names the holders OF THIS OBJECT, so the same token in two threads names different people. And on-call names a rota position whose entire purpose is that it moves — re-derived in August, March's "@on-call" names somebody who has never seen the thread, and a reader chasing why nobody answered asks the wrong person. So every group kind PINS the set it was derived over with the instant it was derived at, and the pin is the population while the delivery is what happened to each of them. "Independent from display name" is enforced by shape rather than by assertion: a reference that IS the typed text, or that carries an address, a handle or a written-out name, is the parseActorRef argument arriving where identity lives in content — a name changes and a handle is reassigned, and either used as identity makes every mention of somebody a mention of nobody the day they marry. resolvableKinds is separate from permittedKinds for the reason S8.2 needed unsupported: "policy forbids it" and "this build cannot do it" are different facts and a reader told the wrong one debugs the wrong system. -->
    • S8.4.b Search only identities visible to actor in current tenant/ workspace/object and authorized mention scope. Evidence: identity search negative tests. <!-- isDiscoverable / searchFaults / DirectoryEntry, 7 fault kinds. The predicate is ONE named function because S8.4.d needs the same one at submission and the two paths must not be able to drift; everything else in the module calls it. The negative tests cover the five ways a served row should not have been there plus the two nobody writes: the SEARCHER'S own workspace (the candidate belongs to the partition and the person asking does not, so the picker is answering for a partition they are outside — one of the fourteen injections, and the suite catches it), and serving anybody at all to a searcher who cannot read the target, since a mention picker is a directory query wearing a document's clothes and the document is the only thing that scoped it. A CANDIDATE WHO CANNOT READ THE TARGET IS A POLICY DECISION RATHER THAN A BUG, because mentioning somebody into a thread is how they get pulled into it — but where the deployment does not permit it, offering them makes the picker a membership oracle for every object in the tenant. Suspended is deliberately NOT hidden: they are still here, they still get named, and it is the delivery that waits. -->
    • S8.4.c Prevent enumeration through query timing, result counts, avatars, errors, autocomplete, copied markup, or guessed stable IDs. Evidence: discovery abuse tests. <!-- searchIdentities / responseDifferences / probeForOracle / discoveryFaults / AvatarSession, 12 fault kinds. THE SEVEN CHANNELS DIVIDE IN TWO AND THE DIVISION IS THE HONEST PART. Four are IN THE RESPONSE, so a real function can be held to them — and the property is not "the hidden rows are withheld", which is true of every implementation with the bug, but that THE RESPONSE IS A FUNCTION ONLY OF THE VISIBLE SET. The suite proves it by running the real search over two directories that differ by five people the actor may not see and asserting the responses are identical, then by probing a hidden colleague against somebody who does not exist and asserting no channel separates them. AN ORACLE DOES NOT NEED A POSITIVE ANSWER: two queries that both return nothing are still a directory if one of them says so differently. Three channels are properties of the SYSTEM AROUND the response and no returned value can measure a database's plan or intercept a paste, so those are declared and the declaration is checked — and a filter applied AFTER retrieval is a timing oracle whatever the milliseconds say, because the work done before the answer is proportional to what the actor may not see. AVATARS ARE ISSUED, NOT DERIVED: a url that is a function of the subject id is a directory lookup with an image content type, so an attacker cannot construct the avatar of somebody they have not been shown; the cost is that a cross-session cache needs exactly the function that is the oracle, and it is taken deliberately. The prefix minimum is recorded as a COST rather than a control, and claiming it as one is its own fault kind, because it is the reason nobody builds the control that would work. -->
    • S8.4.d Resolve at submission, validate current eligibility, and preserve redacted historical reference if identity later changes/leaves. Evidence: lifecycle tests. <!-- resolveMention / renderMention / identityResidue / lifecycleFaults, 11 fault kinds. THE PICKER IS NOT THE ONLY PRODUCER, SO THE SEARCH RULE IS NOT THE CONTROL: a mention token is TEXT and arrives by paste, by quotation, by typed markup, by an API client that never opened a picker, which is the only reading of "resolve at submission" that makes it a security property rather than a caching decision. AN UNAUTHORIZED REFERENCE THEREFORE FAILS EXACTLY AS A TYPO FAILS — ActorId segments are lowercase alphanumerics with hyphens, so actor:ama-mensah is a name away from being guessed, and the defence cannot be that ids are hard to guess but that HOLDING ONE CONFERS NOTHING. The test asserts deep equality across a typo, a guessed id of a hidden colleague, a guessed id from another workspace, a reference pasted from another tenant, an erased identity and nothing claimed at all: six failures that differ by a field are six answers. The three refusals that ARE distinguishable are facts about the deployment rather than about a person. AND redacted CANNOT CARRY A DEPARTURE: its three causes are erasure-request, retention-expiry and legal-hold, every one of them a statement that DATA WAS REMOVED, and somebody leaving removes nothing — rendering a departed colleague as redacted tells every reader the record was destroyed when it was not, which is a worse lie than the one it was avoiding. So a render answers TWO questions with two answers: what the text meant when it was written, which does not change, and what this reader may learn from it now, which is re-derived every time. For an ERASURE the typed text IS the name, so the render is a fixed label and identityResidue is the content-level counterpart to residueIn's field walk. -->
    • S8.4.e Create idempotent notification/activity refs with source context, urgency, privacy/redaction, locale, and preference policy. Evidence: mention delivery tests. <!-- mentionDeliveryKey / planDeliveries / sentDeliveryFaults / acknowledgeToAuthor, 9 delivery and 5 acknowledgement fault kinds. WHAT IS NOT IN THE KEY IS THE DECISION: no offset, so one person named three times in a comment is told once; no revision, so a re-save is not a second naming; no occurrence id, so removing a mention and adding it back does not ring twice. planDeliveries takes a COMMENT rather than an occurrence for that reason — given one at a time each call sees a clean slate and the duplicate rule becomes a property of the caller's iteration order. THE SOURCE CONTEXT IS THE PAYLOAD THAT CROSSES THE BOUNDARY, and S8.2's finding that a context snapshot outlives its access control arrives here on the other axis: a mention of somebody who cannot read the thread is an ACCESS REQUEST, and "Ama named you: '...the settlement figure is 4.2m...'" hands them the document through the notification system, which has no idea it did. So the naming travels and the excerpt does not, as two dispositions rather than one boolean, and the withholding is conceal rather than deny because a notice saying "there is an excerpt you may not see" has disclosed the excerpt. THE ACKNOWLEDGEMENT TO THE AUTHOR IS ABOUT THE RECORD AND NEVER ABOUT THE DELIVERY: every honest sentence about a delivery is a fact about the recipient, and "3 of 12 could not be reached" measures a team roster the author can already see against an access control list they cannot — repeat it with a second team and the list is enumerated with no refusal anywhere. The audience note is a CONSTANT, because the moment it takes a parameter somebody puts a count in it. What the author IS told is what resolved and what did not, both facts about their own text. two-deliveries-under-one-key is checked over the OUTBOX rather than over the planner's output, because the planner cannot emit it and a fault kind only its own producer can never produce is a hole with a name — a retry or a second worker is where the invariant is worth having. -->
    • S8.4.f Handle self, duplicate, nested quote, edit-add/remove, deleted content, suspended user, and inaccessible target. Evidence: edge-case tests. <!-- classifyOccurrence / occurrenceFaults / OCCURRENCE_RULE_ORDER, 6 dispositions, 13 reasons and 10 fault kinds. ONE FUNCTION RATHER THAN EIGHT, because several of the cases are true at once in the situations that actually go wrong — a suspended colleague named twice inside a quotation of a deleted comment is what a thread looks like after a week — and eight independent checks agreeing is eight chances for two of them to disagree. A MENTION INSIDE A QUOTATION IS A REPLAY AND NOT A NAMING: identical token, different act, and nothing in the token can say which, so resolution is POSITIONAL and quoted regions are a structural fact the host declares. AN EDIT CANNOT UN-SEND, so a removed naming is RETRACTED rather than deleted (S7.13 arriving a fourth time) or somebody who read the notification goes looking for why they were named and finds a thread that never names them. hold and record-only are the pair that must not merge — not now against never — or a suspended colleague's mentions are lost forever or every self-mention re-sends on the next sweep. THE PRECEDENCE TEST HAD TO BE MADE CUMULATIVE TO BE A TEST AT ALL: with thirteen contexts each carrying one wrong fact, every rule fires correctly in isolation and the suite passes against ANY ordering, which the injection that moved the duplicate check above the quotation check proved by escaping. The descent now starts with every condition true at once and turns off only the one above at each step, so a rule that moves up is caught by the step whose answer it steals. occurrenceFaults compares a host's actual handling against the rule in ONE direction, ranked by DISPOSITION_PERMISSIVENESS: a deployment that held something this module would have sent has been careful and its users will tell it; one that sent something this module would have held has already published, and nobody tells it anything. --> <!-- S8.4 marked after reading src/mentions.ts and src/mentions.spec.ts in full in this session: 104 tests, every one of the 66 declared fault kinds proved reachable by the test that runs last, 14 source injections each caught against a control arm that passes, and the whole kit suite green. -->
  • S8.5 Add assignments, due dates, escalation, reassignment, and workload views. Evidence: state-machine tests. <!-- src/assignments.ts + src/assignments.spec.ts, 159 tests, 95 fault kinds, every one proved reachable by the test that runs last. THE VEIN: AN ASSIGNMENT IN THIS KIT IS A COLUMN ON THE THING BEING WORKED ON, AND S8.5 ASKS FOR A THING OF ITS OWN. indexes.ts declares the physical model and workbench_entity carries the whole of assignment in four columns — assigneeId nullable, reviewerId nullable, dueAt nullable and openState with a declared distinctValues: 2 against this item's TEN named states — while no assignment entity appears anywhere in the statistics. So an assignment has no id, no assigner, no reason, no priority, no SLA class and NO HISTORY, and UPDATE workbench_entity SET assigneeId = 'kofi' does not end the previous assignment, it makes it never have existed. S8.5.a asks first for identity and S8.5.e asks for preserved history: a state machine over a column overwrite is a state machine with one state. THE MODEL ALREADY ADMITTED ITS CARDINALITY WAS WRONG — reviewerId is a SECOND singular column added because a second person needed to hold one object at one time in a different role, an error paid for one role at a time, and there is no third column for the person who has been OFFERED it. THE GREP: take an item that asks for a NOUN and look for a TABLE with that name in the storage model; where the concept exists only as columns on somebody else's row it has no identity, no history and no state machine. Corroborating tell: two singular columns for one relation in different roles. -->

    • S8.5.a Define assignment identity/type, target/revision, assignee/ eligible pool, assigner, role, priority, state, reason, and timestamps. Evidence: assignment schema. <!-- AssignmentRecord + assignmentSchemaFaults, 14 fault kinds. THREE DECLARATIONS OF ASSIGNMENT EXIST A LIBRARY APART AND DISAGREE: authorization.ts's plural assignees, functions.ts's plural assignedTo, and indexes.ts's one nullable assigneeId with a partial index defined by IS NULL. THE RESOLUTION IS NOT THAT ONE IS WRONG — they answer different questions and nothing ever said so: the plural list is who may TOUCH it, the singular column is whose PLATE it is on, and an OBSERVER proves they are different by belonging in the first and in neither column. Both are derived from one ledger by derivedAcl and derivedWorkColumns and checked in BOTH DIRECTIONS by projectionFaults, because they fail differently and both silently: a column ahead of the ledger authorises somebody no assignment supports, a ledger ahead of the column leaves the person doing the work locked out and drops the row into wbe_team_unassigned, which is partial on assigneeId IS NULL, so it appears in the UNCLAIMED queue while somebody is already doing it. everAssigned is the producer functions.ts never had: not-the-assignee reads ActorHistory.assignedTo and gives the right reason for reading the object rather than the request, then stops one step short, because the object's assignee column is also something the actor can have changed — take yourself off the assignment and the rule meant to catch you finds an empty list, which is S8.1's laundering at the SEAT and S8.3's at the CONTRIBUTION arriving a third time at the ASSIGNMENT. Negative control: an offer that was never accepted is NOT a past assignment, or a declined offer disqualifies people from work they refused. -->
    • S8.5.b Define offered/accepted/active/blocked/completed/declined/ cancelled/overdue/escalated/reassigned transitions and authorization. Evidence: state machine tests. <!-- ASSIGNMENT_STATES (EIGHT of the item's ten), TRANSITIONS (15 edges), transitionFaults (13 kinds), hostStateModelFaults (4). THREE OF THE TEN ARE ON DIFFERENT AXES AND THE EXTENSION IS THE PROOF: if overdue is a STATE then extending a deadline has to decide which state to go back to, and the record no longer says whether the person had accepted it or was blocked on somebody — a state machine with a transition that time makes and an extension unmakes is not one. If escalated is a state then an escalated assignment is not blocked and the cause has been erased by the response to it. So both live on their own axis (dueStanding, EscalationRecord) and hostStateModelFaults is how a deployment that put them in its state column finds out — it also catches the storage model as declared, a two-valued flag. THE TWO ABSENT EDGES ARE THE CONTENT: no offered -> active, because work cannot start before somebody agrees and an accepted state nobody accepted is a consent that never happened; and no blocked -> completed, because finishing out of a block leaves nothing saying whether the blocker was overcome or ignored. COMPLETE IS THE ASSIGNEE'S AND CANCEL IS THE ASSIGNER'S: an assigner marking work complete is one person deciding another person did something, indistinguishable in the record from the work having been done — with an EXEMPTION for self-assigned work, where there is only one person and the rule is about the second. An offer LAPSING has no actor (S8.3's expiry rule), and the sweep job's principal in the field a reader goes to for a person's decision is the invented sentence. THE PRECEDENCE DESCENT IS CUMULATIVE — every condition wrong at once, one turned off per step — because with one wrong fact per case every ordering passes. -->
    • S8.5.c Define due instant, originating time zone, reminder schedule, extension, SLA class, pause behavior, and DST handling. Evidence: clock-controlled tests. <!-- instantOfWallClock / commitDue / extendByCalendarDays / scheduleReminders / slaStanding / pauseFaults, 13 clock faults, and REAL arithmetic over the platform's own zone database rather than declarations about it. A DUE DATE IS A PROMISE MADE IN A PLACE AND Instant REFUSES TO CARRY ONE: parseInstant takes Z only and says why, that string order is time order, which is right for a log and wrong for a commitment — and the entity's own statistics agree, 900 distinct dueAt values across 400,000 rows being the signature of dates rather than moments. So a commitment stores the WALL CLOCK, the ZONE and the INSTANT, and DST proves the three are not redundant. KNOWN VALUES: 17:00 New York on 2026-03-07 is 22:00Z and on 03-08 is 21:00Z, so "one more day" moves the instant TWENTY-THREE HOURS and +24h delivers at 18:00 local — an hour late, on the day the deadline was already tight; 25 hours over the autumn transition; exactly 24 in Accra, which has no summer time. 02:30 on 2026-03-08 denotes NO instant in New York and 01:30 on 11-01 denotes TWO an hour apart, both reported rather than silently resolved. Auckland goes back in April and forward in September, which is the control that nothing assumes March means forward. A reminder authored as "24 hours before" fires at 16:00 local and one authored as "17:00 the day before" fires at 17:00 — different rules, not two spellings. A PAUSE IS ON THE CLOCK THAT MEASURES, NOT ON THE PROMISE THAT WAS MADE: stopping a measurement is something anybody may do and moving a deadline is a renegotiation, and one control for both lets a stopwatch rewrite a promise. Genuinely clock-controlled: the module reads no ambient clock at all — zero Date.now(), every instant an argument. -->
    • S8.5.d Escalate overdue/declined/blocked/unavailable/policy-triggered work to configured actor/role/on-call with deduplication. Evidence: escalation tests. <!-- ESCALATION_CAUSES (all five), EscalationTarget, escalationKey, escalationFaults (11 kinds). ESCALATION IS AN AUTHORIZATION EVENT WITH NO HUMAN IN IT: it creates an assignment, and an assignment grants access, so A CLOCK CANNOT GRANT ACCESS and the target must be independently eligible — decided by rules that already exist rather than by the deadline passing. S8.4's rota lesson arrives with a bigger consequence than a misdirected notification, because a rota re-derived at read time hands August's on-call the object March's escalation opened; the axis is IMPORTED from RESOLUTION_DEPENDS_ON rather than restated, and the spec asserts the two agree on every kind. THE DEDUP KEY IS (assignment, cause, GENERATION) and the third part earns its place: an escalation fires, somebody extends the deadline, the deadline passes again — that is a second situation, and a key over (assignment, cause) alone suppresses it forever. Escalating is asking for help and is NOT a renegotiation, so it must not move the deadline; an unavailability escalation to an unavailable target has moved work from one absence to another; and an escalation that resolves back to the person who is stuck reminds somebody who already knows while nobody with the authority to help hears anything. -->
    • S8.5.e Reassign with reason, preserved history/comments/work, old/new notifications, authority, and due/SLA policy. Evidence: reassignment tests. <!-- Reassignment / reassignmentFaults (11 kinds) / handoverNotices / handoverNoticeFaults (2). The predecessor goes TERMINAL in reassigned and the successor names it, which is what makes "preserved history" a row somebody can find rather than a value that was overwritten. THE DUE POLICY IS A VALUE AND NOT A NULL: inherit keeps a promise made outside the team, reset starts the clock for somebody beginning from nothing, and renegotiate says the deadline is not yet agreed — which is the honest answer more often than either, and a null due date and a due date nobody has agreed are different things to a queue. ACCESS THAT OUTLIVED THE WORK is decision 2's consequence nobody plans for: on a protected object, taking somebody off the assignment locks out the person best placed to explain it, and the answer is not to leave them on the list — that grant never expires, because no future event is about them — but an explicit handover grant WITH AN END. And S8.4's acknowledgement rule arrives where it is easiest to get wrong: the reason somebody came off a job is about that person, so "reassigned from Ama (falling behind)" is a performance assessment delivered to a colleague by a workflow engine. Two audiences, two payloads, and the incoming notice has no parameter to put a reason in. -->
    • S8.5.f Build authorization-aware workload counts/queues by assignee, role, priority, state, due/SLA, domain, capacity, and locale/time zone. Evidence: workload query tests. <!-- WorkloadQuery / workloadFaults (9 kinds) / countWorkload / calendarBand / weekShapeFor. THIS ITEM BREAKS A DECLARED AUTHORIZATION IN S3.5. indexes.ts permits inbox.assignee-open to drop its workspace predicate and writes down what stands in for it — "the assignee predicate is the authorization: the caller reads their own assignments and no query shape lets them ask for anybody else's" — and S8.5.f IS that query shape, so the moment a manager binds another person's id the thing authorizing the widening is gone and the scope has to come back. The count is then mentions.ts's count oracle moved from people to rows: "Ama has 14 open items" told to a reader who may see 9 discloses the 5, so the total is taken AFTER the reader's filter and a filtered figure says that it is one. Capacity is the derived form of the same oracle. LOCALE IS NOT DECORATION — weekShapeFor reads real week data, and the known value is that ONE Sunday deadline is this-week for a reader in London and later for one in New York, while a Saturday is this-week in the United States and later in Egypt, whose week starts on Saturday and whose weekend is Friday and Saturday. Work with no SLA class is counted as unmeasured rather than within, because reporting the absence of a promise as the keeping of one is how a queue says everything is fine. -->
    • S8.5.g Reconcile assignment state with review/gate/target changes and avoid ghost/duplicate work. Evidence: dependency tests. <!-- assignmentStanding / reconciliationFaults (7 kinds) / duplicateWorkFaults. GHOST WORK IS AN ASSIGNMENT WHOSE SUBJECT MOVED, and approval-workflows.ts already owns the question one level up: repeatOnChange says whether a judgement survives an edit, so it is COMPOSED rather than re-decided. An assignment has a third answer an approval never needs, because an approval is asked about a revision and an assignment is asked about a JOB, and a job can be called off — VOID. The descent puts the three void causes first, because work nobody wants is not work somebody is behind on; a DEPARTURE ORPHANS rather than voids, because the obligation was the organisation's and somebody leaving does not discharge it, so closing it removes what was actually needed from the queue. THE DUPLICATE IS THE FAULT THE COLUMN MODEL COULD NOT REPRESENT: with one assigneeId per row a second simultaneous assignment for one (target, ask, role) is an OVERWRITE rather than a collision, so nothing has ever detected it — and it is the ordinary result of an escalation that assigned somebody without terminating the original. Keyed on the ask because two reviewers reading one document for different things are two jobs; one AUTHORITY holding two records is bookkeeping, not two people doing one job. --> <!-- S8.5 marked after writing and reading src/assignments.ts and src/assignments.spec.ts in full in this session: 159 tests, every one of the 95 declared fault kinds proved reachable by the test that runs last, 17 source injections each caught against a control arm that passes, and the whole kit suite green at 7652 tests. -->
  • S8.6 Add presence and active-selection awareness with privacy controls. Evidence: reconnect and stale-presence cleanup tests. <!-- libs/oshun/workbench-kit/src/collaborator-presence.ts (136 tests, 58 fault kinds), subpath ./collaborator-presence. Rollup of S8.6.a-f below; every child is a section of the one module because the six are one question — an expiry rule that could not see the publication, or a publication that could not see the lease, would be six policies agreeing by accident. THE HOLE: PRESENCE IS THE FIRST THING IN THIS KIT THAT MUST STOP BEING TRUE ON ITS OWN, AND EVERY DURABILITY RULE THE KIT HAS SAYS THAT IS A BUG. indexes.ts declares EIGHT entities — workbench_entity, gate_run, waiver_grant, entity_taxonomy_link, taxonomy_term, editorial_calendar_slot, audit_event, recompute_job — and every one is a durable row that exists until something deletes it; before this module a grep for ttl, heartbeat or ephemeral across all 124 shared modules returned NOTHING. The discipline on top of that model is explicit and load-bearing: release-locks.ts decision 5 is "A LOCK THAT EXPIRES QUIETLY IS NOT A LOCK. Expiry is an event with an outcome and an actor, not the absence of one", and audit.ts decision 1 is that a decision surface with no event is "a thing the platform does and does not record". S8.6.d asks for the inverse of the first IN ITS OWN TEXT — "without durable online claims" — and S8.6.f asks for the inverse of the second. THE RESOLUTION IS NOT THAT ONE OF THEM IS WRONG. A DECISION IS DURABLE AND ITS END NEEDS AN ACTOR; AN OBSERVATION IS PERISHABLE AND ITS END NEEDS ONLY A CLOCK. A lock is something the system CLAIMED on somebody's behalf and a claim that evaporates leaves nobody able to say who decided; presence claims nothing, so the honest end of a report whose subject stopped reporting is that it CEASES. Demanding an actor there is suggestions.ts's invented sentence with the volume up — a laptop lid closing is not an act of the platform. So the module checks for a record CROSSING the line, because presence does not leak as a table marked presence: it leaks as a lastSeenAt column on workbench_entity and as a metric labelled by actor id. -->

    • S8.6.a Define ephemeral presence session ID, actor-safe identity, workbench/object, permitted sublocation/selection, device/tab, state, heartbeat, and expiry. Evidence: presence schema. <!-- THREE IDENTITY LEVELS, because RECONNECT AND A SECOND TAB ARE DIFFERENT EVENTS AND ONE IDENTIFIER CANNOT TELL THEM APART: key presence on the actor and a reconnect looks like a second person, key it on the connection and closing one tab looks like leaving while another is open. So ACTOR / ENDPOINT (device+tab, survives a reconnect) / CONNECTION (ephemeral, does not). WHAT THE CLIENT CLAIMS IS SPLIT FROM WHAT THE SERVER CONCLUDES — REPORTED_ACTIVITIES against PRESENCE_STANDINGS — because a single status field lets a client declare itself live; every instant on the record is server-stamped except reportedAt, which is labelled and read by nothing. ROOM_ACTOR_KINDS IS NARROWER THAN ActorRef: a retention sweep reading a document is not a person in the room, and an erased identity has no present tense. And a session id DERIVED from the actor is a name badge — release-locks.ts decision 1 again, with a sharper consequence, since streams are addressed by session id and a viewer reduced to pseudonymous de-anonymises the room by reading the identifier the transport handed them; the check covers plain text, hex and base64url, and the base64 arm is case-SENSITIVE because its alphabet carries bits in its case. -->
    • S8.6.b Authorize publish/subscribe by tenant/object/property/protected context and minimize precision/content exposed. Evidence: privacy tests. <!-- DISCOVERABILITY IS NOT TRACKABILITY, AND THE KIT ONLY HAS THE FIRST. S8.4 established that ENTITY_KINDS names twelve things a row can be and none is a human being, so every authorization here reads "given this OBJECT, may this ACTOR read it"; presence points the arrow the other way and discloses a fact about a PERSON to an object's viewers. isDiscoverable answers whether you may FIND somebody and NOTHING answers whether you may WATCH them — a directory lookup you are entitled to perform is not a trace you are entitled to keep. So a publication is the JOIN of two permissions and exposureFor refuses to derive the second from the first; the one-way implication (not discoverable => never named) is asserted in the spec against the REAL isDiscoverable over a 32-case cross-product, and the reverse is shown NOT to hold. PRESENCE IS ALSO THE FIRST SIGNAL GENERATED BY LOOKING, and looking is the one act that has never disclosed itself: a comment or an assignment is a deliberate act its author made public, while authorization.ts already carries support-may-not-see-presence and an auditor, a legal-hold reviewer and an approver checking a conflict are the same shape. THAT IS WHY THE LADDER HAS FOUR RUNGS AND NOT THREE — counted still says somebody is here, and "and 1 other" told to the subject of an investigation is the disclosure, so an observer contributes nothing, not even a change in a number. A SELECTION IS CONTENT: a range is a block id and two offsets, end states the block is at least that long, and an EMPTY RANGE IS A CARET, which is not where somebody is reading but where they are about to type — so precision DEGRADES down a ladder per viewer rather than vanishing, and the ceiling is a function of the reader, applied on the way out. -->
    • S8.6.c Handle multiple tabs/devices, reconnect/session replacement, network partition, browser sleep, logout, permission loss, and object deletion. Evidence: lifecycle tests. <!-- AN AUTHORIZATION DECIDED ONCE AND NEVER RE-DECIDED IS A SUBSCRIPTION: every check in this kit runs at request time and a presence stream is a request that stays open for an afternoon, so "logout, permission loss, object deletion" is a list of things that must reach INSIDE an authorization that already returned allow. The roster counts distinct PEOPLE, because "3 people are here" over three of one person's tabs is a false statement about a room, and the fault that catches a reducer keyed on the ACTOR cannot be seen by comparing rosters — after such a reducer runs the person is correctly absent from both sides — so it compares the sibling connections by id. A LATE HEARTBEAT DOES NOT REVIVE AN EXPIRED CONNECTION (found by chasing an unread field on the observation): a socket unreachable for longer than its lease has been reported gone to everybody, and extending it when the packet lands makes somebody reappear who was announced as having left. AND A DEPARTURE DISCLOSES IN BOTH DIRECTIONS: identical to the ROOM for every cause, asserted exhaustively over the union rather than by eye, because a channel distinguishing "closed the tab" from "lost access" from "left the company" is an HR feed; and DISTINCT to the SUBJECT, because actors.ts decision 2 already settled that a rejection says which kind and "you were disconnected" and "you may no longer open this" need opposite responses from the person they happened to. -->
    • S8.6.d Expire stale presence by server clock/lease and reconcile missed disconnects without durable online claims. Evidence: cleanup tests. <!-- THE VEIN, MADE CHECKABLE. A HEARTBEAT IS THE CHECKED PARTY SUPPLYING ITS OWN LIVENESS (S8.1's rule, in its most direct form): a client that supplies the TIME of its own heartbeat holds a room open forever with a clock set forward and places itself in a document at an hour it was not there — so the server stamps, and the spec asserts the standing does not move as reportedAt sweeps a decade in both directions. The lease is read from the SESSION, not from the policy in force today, so a policy change cannot retroactively expire a connection granted something longer; and a lease shorter than TWO heartbeats is a fault, because one dropped packet then empties the room, which on a train is every few minutes. ExpiryReconciliation HAS NOWHERE TO WRITE A DEPARTURE INSTANT OR AN ACTOR and the spec asserts that on the field list: a sweep knows ONE thing, the last instant a heartbeat arrived, and "Ama left at 16:42" is a fabrication built from it — she left at some point in the lease window and the sweep observed its own timer. DeclaredExpiryRecord is the shape a host produces when it models this the way the kit models the end of a lock, and every field it fills in is a fault. -->
    • S8.6.e Render collaborator presence/selection with accessible identity and noncolor cues without obscuring authoring content. Evidence: UI accessibility tests. <!-- COMPOSED FROM signals.ts RATHER THAN RE-DECIDED, and headless, because what the clause actually names is measurable without a browser. COLOUR IS THE ONLY CHANNEL BY WHICH COLLABORATOR IDENTITY IS CONVENTIONALLY CARRIED, which is precisely the S7.22.e failure, so the comparison set for a cue is the OTHER PEOPLE IN THE ROOM — a vocabulary whose members arrive at runtime — and cues that all land on one reader are reported too, since an assistive name is nothing on a greyscale screen and a glyph is nothing in the accessibility tree. "WITHOUT OBSCURING AUTHORING CONTENT" IS A CONTRAST MEASUREMENT AND THE OBVIOUS WAY TO TAKE IT IS WRONG: a translucent presence wash sits BETWEEN the ink and the page, so the number that reaches the reader is the composite, and an audit measuring ink against the page passes while the selected paragraph is the one nobody can read. Known values from the WCAG definition computed independently of signals.ts: black on white is 21:1, and blue at 85% drops it to 2.8078:1. Also the badge drawn over the insertion point, and a live region wired to the heartbeat, which interrupts a screen reader every few seconds and makes the document unusable. -->
    • S8.6.f Keep presence out of immutable authored state/audit by default while emitting bounded operational telemetry. Evidence: persistence/ telemetry tests. <!-- THE TELEMETRY IS THE BACK DOOR THROUGH WHICH THE EPHEMERAL BECOMES DURABLE, and it is the half of this item that a persistence check alone would miss. The durable surfaces are named individually because they are reached by different mistakes — authored state by a convenience column (workbench_entity.lastViewedBy, added to save a join), the audit by uncoveredSurfaces pulling presence in as a decision surface, the search index by a denormalising writer, the backup by being in any of them. Then the word doing the work in "bounded operational telemetry" is BOUNDED: a counter labelled by actor id is a durable online claim wearing a metrics costume, kept by the time series store long after the lease expired, and it answers "when was Ama at work" better than the table we declined to write. So labels identifying a PERSON, a SESSION or a single OBJECT are refused OUTRIGHT rather than budgeted — the spec proves a per-person metric sits well inside the series budget and is still refused — and everything else is bounded by a cardinality computed as the product of the declared label domains, plus a retention window that answers "is the socket layer healthy" and not "who was working late in March". --> <!-- S8.6 marked after writing and reading src/collaborator-presence.ts and src/collaborator-presence.spec.ts in full in this session: 136 tests, every one of the 58 declared fault kinds proved reachable by the test that runs last, 25 source injections each caught against a control arm that passes, and the whole kit suite green at 7789 tests. -->
  • S8.7 Add notification routing by event, role, urgency, locale, channel, quiet hours, and frequency caps. Evidence: known-value routing suite. <!-- libs/oshun/workbench-kit/src/notification-routing.ts (138 tests, 98 fault kinds), subpath ./notification-routing. Rollup of S8.7.a-g below; every child is a section of the one module because the seven are one question — an audience that could not see the preferences, or a preference that could not see the transport, would be seven policies agreeing by accident. THE HOLE: THERE IS A READER AND THERE IS A SENDER AND NOTHING TURNS AN EVENT INTO A RECIPIENT. S7.19 (notification-center.ts) is the reader, and its header names what it does not do — "It does not deliver a notification, hold a device token, evaluate a quiet-hours window against a clock, or decide what a category means. libs/veritas/notifications owns all of that." That library is real and it is not small: roughly seven thousand lines of FCM, APNs and Web Push; SendGrid, SES and SMTP; Twilio, Africa's Talking and Hubtel; Twilio and Meta WhatsApp. Its entry point is send(request: NotificationRequest), and NotificationRequest carries userId: string — ONE person — beside channels already chosen, title and body already rendered, and category and urgency already decided. Every question this item asks is answered BEFORE that call is made. THE CENSUS CLOSES THE GAP: nothing in the monorepo imports that library. tsconfig.base.json carried no path mapping for @veritas/notifications, so nothing could; its createNotificationService is called from exactly one place, which is its own service.ts; and the only references to it outside its own tree are three doc comments and two hand-copied fixtures — the kit's own spec transcribes its vocabulary under the note "copied as a fixture rather than imported". (Every same-named NotificationService elsewhere in apps/ belongs to a different domain.) The delivery system has no caller, and the step that would produce its argument is this item. THE GREP THAT FINDS THIS VEIN: TAKE THE SENTENCE IN A MODULE THAT SAYS WHO OWNS THE REST OF THE PROBLEM, AND GO AND COUNT THAT OWNER'S CALLERS. A deferral reads as diligence and is only true if somebody makes the call. It is S8.2's producer-less judge turned around — not a judge nothing feeds but a SENDER NOTHING CALLS, which is harder to see because the sender is the part that obviously works. The corroborating tell is one line in the kit's own idiom: mentions.ts declares RecipientContext.channel, documented as the weakest urgency wanted "ON THIS CHANNEL", and planDeliveries reads the threshold, the locale, the standing and the mute and never reads channel. S8.6's rule — an unused field on an input type is a missing rule — and the missing rule is the whole of S8.7.e. Twelve decisions, of which the load-bearing ones: the fan-out is where authorization happens and it happens ONCE PER PERSON, so the unit is the triple (event, recipient, channel); a person reached by four routes is one person and FOUR GROUNDS and the strongest wins rather than the first, which is the only defect here that leaves both audiences identical; escalation is the one ground that reaches OUTSIDE the audience, so its notice is composed from classes rather than values; mandatory pierces preference and timing and NEVER pierces authorization; quiet hours are a wall clock, start <= t && t < end is unsatisfiable for every wrapped window, and the release is itself a wall clock that may denote no instant; a cap is a cap on INTERRUPTION and not on record, and counted over its own suppressions it never reopens; fallback must not increase exposure; an unverified address is not a weak address, it is somebody else's; a bounce is evidence about the ADDRESS and a failure about the PROVIDER and only one may be retried; and in one state machine acknowledged needs an actor while expired needs only a clock — S8.6's boundary with both halves side by side. Verified: 44 SOURCE INJECTIONS, all caught, against a control arm that passes. The first 36 found FOUR places where the spec asserted a constant AGAINST ITSELF (looping over DISPOSITIONS_THAT_STOP_A_NOTICE rather than naming its members, TERMINAL_DELIVERY_STATES.every(isTerminalDelivery), and two more) — removing conceal from the authorization gate and bounced from the terminal states both passed. Every ordered or meaning-bearing vocabulary is now written out in the spec rather than read back from the source, and the last 8 injections delete or reorder those lists to prove it. The per-fault reachability test forced two fixes in the module itself: mandatory-notice-silenced-by-a- preference could never fire because the floor always won the comparison it was checking, and quiet-window-evaluated-as-a-simple-range compared the module against its own answer. Writing the spec found two more: the escalation renderer demanded two different strings for one notice, and planNotifications reported required-notice-muted-rather-than-moved and then obeyed the mute anyway — the constant encoding that rule, SILENCEABLE_OBLIGATIONS, had been declared and never used. Clock known values computed from UTC offsets independently of both assignments.ts and this module: a 22:00-07:00 New York window is 480 minutes the night the clocks go forward and 600 the night they go back, never 540; a release at 02:30 on 2026-03-08 is skipped and resolves to 2026-03-08T07:30:00Z; Auckland goes BACK on 2026-04-05, so the same 02:30 is repeated there — the control that nothing assumes March or forward. -->

    • S8.7.a Register notification event types with source, recipients/ audience rules, default urgency, required/optional delivery, sensitivity, template, dedupe key, and expiry. Evidence: event registry. <!-- EventRegistration + eventRegistryFaults, 16 fault kinds. Three checks are not schema: a dedupe key built from the RENDERED summary rather than the event's fields silently un-dedupes everything in flight when somebody edits a template; an escalation window past the event's own expiry means the unanswered case is the one nobody is ever told about; and a scheduled event whose template has an {actor} slot fills it with whichever sweep job ran, which is S8.3's invented sentence. -->
    • S8.7.b Resolve explicit user, assignee, mention, owner, reviewer, role/ team, escalation, and delegated recipients at event time with tenancy/ authorization. Evidence: recipient tests. <!-- All nine grounds the item names, each carrying its own ResolutionDependency — S8.4's axis reused rather than a second one invented. foldGrounds is the one place the fold happens and takes the strongest obligation AND the strongest refusal (OUTCOME_PRECEDENCE); resolutionParityFaults compares the derivation with a host's claim in BOTH directions, because a weakened obligation leaves both lists identical. Tested with the grounds in both orders. -->
    • S8.7.c Compose event/category, urgency, domain/project, role, user, mandatory notice, and channel preferences with explicit precedence. Evidence: preference truth tables. <!-- The finding here is that a precedence over ONE list conflates two axes — WHO SET IT and HOW SPECIFIC IT IS — and both single-axis answers are wrong: ordering by specificity makes a person's own choice advisory, ordering by authorship makes a narrow control unsettable. So comparePrecedence is authority then scope. The truth table is CUMULATIVE: eleven settings present at once, one removed per step, so every step still has all the weaker ones under it. -->
    • S8.7.d Render localized templates using recipient locale/time zone, accessible/plain alternatives, safe redaction, and target link. Evidence: locale/security fixtures. <!-- Redaction composes S7.19's renderSummary rather than substituting again, because its own doc comment says it is the ONE place substitution happens. Three leaks that are not the summary: the PLAIN part (an email's text alternative is where redaction is forgotten, and it is what a screen reader reads), the LINK (/doc/Q3-payroll-restructure beside a body that withheld the title), and the transport itself — EXPOSURE_CEILING maps all nine protection.ts classes, asserted over the full cross-product against an independently transcribed table. -->
    • S8.7.e Route in-app, email, push, SMS, webhook, or other registered channel through capability/consent/verification and fallback policy. Evidence: channel tests. <!-- Capability, consent and verification are three different false answers with three different remedies and are refused separately. AN UNVERIFIED ADDRESS IS NOT A WEAK ADDRESS, IT IS SOMEBODY ELSE'S: a typo'd mailbox does not bounce, it succeeds, at a stranger. Holding a phone number is not consent, which is gated by EXPOSURE rather than by channel name. Fallback is an ordered bounded walk, not a search, so it cannot answer a push outage with SMS; and a suspended account — often a security response — is reachable in-product and nowhere else. -->
    • S8.7.f Apply quiet hours/time zone/DST, urgency bypass, digest eligibility, frequency caps, and mandatory safety/security exceptions. Evidence: clock/routing tests. <!-- Reuses assignments.ts's zone toolkit rather than a second clock. The wrap and the skipped release are both silent failures: the first makes the window unsatisfiable so nothing is ever quiet and nothing is logged, the second holds a batch until a moment that never arrives. The cap counts DELIVERIES only — counted over its own suppressions the ceiling is self-sustaining and the symptom is the absence of notifications — and the in-product centre is never capped, because what a cap rations is the interruption and not the record. -->
    • S8.7.g Persist planned/sent/delivered/failed/bounced/suppressed/deduped/ expired/acknowledged state and safe retry. Evidence: delivery lifecycle. <!-- Exactly the nine states the item names, with TRANSITION_AUTHORITIES deciding what each end needs: acknowledged is the only state the RECIPIENT authors (and opening the centre is not reading — S7.19 decision 6), expired has no actor and no reason. failed returns to sent and bounced does not, and no edge departs a terminal state at all. The key is the EVENT's (S3.6 decision 1), so replanning is byte-identical and an at-least-once dispatcher is safe. --> <!-- S8.7 marked after writing and reading src/notification-routing.ts and its spec in this session: 138 tests, 98 fault kinds all proven reachable, 44/44 source injections caught against a passing control, export-map violations 0, API pages 24 -> 25, full kit suite 7924 tests green. -->
  • S8.8 Add digest coalescing and deduplication. Evidence: event storms do not produce notification storms. <!-- libs/oshun/workbench-kit/src/notification-digests.ts (101 tests, 91 fault kinds + 7 admission refusals), subpath ./notification-digests. Rollup of S8.8.a-f below; the six are one module because they are one accumulator — a key that could not see the window, or a cap that could not see what must stay separate, would be six policies that agree by accident. THE HOLE: THERE IS A DIGEST AND THERE IS NOTHING TO PUT IN IT. @oshun/messaging-channels/digest is real and careful — eight locales, a Sunday-17:00 send window computed DST-correctly in each locale's home zone, corrections leading a deterministic section order, five items per section with an overflow line, and an empty week returning do-not-send rather than an email — and it is mounted at GET /v1/me/digest/preview. Its input arrives through DigestInputsReader, whose PRODUCTION implementation is emptyDigestInputsReader, whose get returns null for every member, always. The file's own first sentence says why: "No production producer writes a digest aggregate in this BFF." The only reader that ever returns anything is marked test-only. THE GREP: TAKE THE INTERFACE A WORKING COMPONENT READS ITS INPUT THROUGH AND GO AND LOOK AT WHAT THE PRODUCTION IMPLEMENTATION OF THAT INTERFACE RETURNS. S8.2 found a judge nothing produces for, S8.7 a sender nothing calls, and this is a CONSUMER WHOSE PRODUCTION FEED IS A CONSTANT NULL — the hardest of the three, because the null is deliberate, documented and correct. The other end of the gap is in the kit: applyTiming returns defer-to-digest, the only disposition that delays a notice without saying until when. hold-until-release carries the instant the hold lifts and reports held-past-its-own-expiry; defer-to-digest returns releaseAt: null and no finding at all, so a deferred notice is indistinguishable from a dropped one. That is structural rather than an oversight — a HOLD is a property of a NOTICE and a DIGEST is a property of a PERSON over an INTERVAL, so no function of one event could have computed it, which is why nothing accumulated. -->

    • S8.8.a Define dedupe identity by recipient, event family, target/scope, state transition, urgency, channel, and time window. Evidence: key contract. <!-- DEDUPE_COMPONENTS is exactly the seven the item names, and MISSING_COMPONENT_FAULTS gives each omission its own casualty: without the RECIPIENT the first person told is the only person told, without the TRANSITION a failure and its recovery collapse and it is the SECOND that is suppressed, without a WINDOW the suppression has no horizon and a reminder is sent once ever. deliveryKey AND THIS KEY ANSWER DIFFERENT QUESTIONS AND ONE MUST CARRY THE EVENT ID WHILE THE OTHER MUST NOT — the first asks whether this is the same DELIVERY under at-least-once dispatch, the second whether this is the same NOTICE when the world repeats itself — and a key carrying the id is decorative, which is the one everybody writes because the right one is already there. The TENANT is in the key even beside the recipient: one person in two tenants each holding a doc:atlas is a cross-tenant boundary crossed as a SILENCE. dedupeIdentityOf is what finally reads EventRegistration.dedupeKeyFields, which S8.7 declared, documented and validated and nothing ever built a key from. dedupeKeyComponents is the inverse, and it exists because at a fixed arity the separator escape ALONE is already injective — so injectivity cannot show the escape is incomplete and a round trip can. -->
    • S8.8.b Distinguish exact duplicate, update/supersession, aggregateable related events, and events that must remain separate. Evidence: classification tests. <!-- classifyPair over the item's four relations, tested first on audience, then tenant, then the same-object case, then the same-family-different-object case. COLLAPSE AND DEDUPLICATION ARE DIFFERENT OPERATIONS AND S7.19 NAMED ONLY THE FIRST: a protected kind may be exact-duplicate-suppressed — two rows for one mention is not two mentions — and may never be aggregated, so reading NEVER_COLLAPSED_KINDS as "never touch these" ships the at-least-once bug on the seven kinds that matter most. SUPERSESSION IS DECIDED BY THE EVENT'S OWN ORDER AND NEVER BY ARRIVAL, because out-of-order arrival is normal and deciding on it lets the stale event win; the symptom is a digest reporting a failure after the recovery, which somebody acts on. The host's own reading arrives as a RelationClaim so the check has something to disagree with (S8.7's reachability lesson). -->
    • S8.8.c Define digest windows/schedules by user locale/time zone, preference, urgency, quiet hours, channel, and maximum delay. Evidence: scheduling tests. <!-- nextWindowClose resolves through assignments.ts's instantOfWallClock rather than by arithmetic, so a close anchored in the spring gap denotes no instant and says so. Known values: 08:00 in New York is 13:00Z on 03-07 and 12:00Z on 03-08, so windowSpanMinutes is 1380 there and 1500 across 11-01 — never 1440. A WEEK IS WORSE: weekShapeFor makes ar-EG start on the Saturday, en-US on the Sunday and fr-FR on the Monday, so one weekly schedule resolves to three dates from one instant, and a runtime that cannot say returns null rather than guessing. The composition faults are the ones no single rule can see: a close inside quiet hours whose HOLD outruns a maximum delay the close itself respected, and a hold long enough that two windows fall due at the same instant and must merge. -->
    • S8.8.d Aggregate counts, actors, targets, status, earliest/latest, failures/blockers, and accessible links without leaking hidden items. Evidence: digest content tests. <!-- A COUNT IS A DISCLOSURE THAT CARRIES NO FIELD. S7.19's machinery governs FIELDS — a slot has a disposition and a sensitivity class — and "and 4 others" has neither while still saying how much is happening where the reader cannot look, so an inadmissible item leaves the COUNT and the remainder is returned for the caller's audit rather than for the page. A DIGEST IS THE ONLY THING IN THE SYSTEM THAT PUTS HOURS BETWEEN THE DECISION AND THE DELIVERY, so it is the only place authorization can go stale while a notice waits, and it is re-asked at composition against the audience as it stands now. composeDigestContent re-authorizes BEFORE it counts, because counting first and filtering the rows afterwards is the version that ships — the rows visibly disappear and the number visibly does not. A failure or blocker is NAMED rather than counted; a target whose own summary withheld what it is called is counted rather than named; and a link carrying a credential onto a surface nobody authenticated to is refused outright. -->
    • S8.8.e Atomically claim/send digest items with idempotency and handle concurrent events, retry, partial channel failure, and late arrivals. Evidence: race/fault tests. <!-- A DIGEST'S IDEMPOTENCY KEY IS A FUNCTION OF THE WINDOW AND NEVER OF ITS MEMBERS: digestIdempotencyKey(tenant, recipient, channel, windowId). A key hashed over member ids is stable right up until a late arrival lands, at which point it matches no previous send and the whole digest goes out again with one extra line in it — and the only way to SEE that is to ask the host for its key twice over the same window with different contents, which hostKeyAfterALateArrival does. A late arrival therefore belongs to the NEXT window, which is why there must always be one to belong to. The lease, the fencing token and the holder are all checked, a claim taken before the close fixes nothing because the set it was taken over is not the set that is sent, a partial channel failure is a fact about ONE channel (S7.19 decision 11) and a retry that reopens the window re-sends the channel that worked, and a window that gives up while still holding notices is a drop with a state name in front of it. An empty digest and a digest of one are both refused. -->
    • S8.8.f Enforce per-recipient/type/channel caps and escalate/drop/summarize according to policy rather than silently discard critical events. Evidence: storm load tests. <!-- THE DIGEST IS THE BOTTOM OF THE LADDER, SO ITS ONLY LAWFUL OVERFLOW IS A SUMMARY. Quiet hours delay rather than suppress (S7.19 decision 10); a frequency cap displaces into a digest rather than into nothing (S8.7 decision 7); there is nothing underneath a digest, so a drop there is the same data loss reached through the last door left open. escalate is not a third option at the same level — it is the admission gate's answer for something that should never have been digested, and used as an overflow it turns the cap into a delay before the storm. AND THERE ARE TWO CEILINGS RATHER THAN ONE: a limit on how many digests with none on their size is one envelope with four hundred lines in it, and a limit on size with none on number mints another every twenty items, so a deployment that sets one has capped nothing. The evidence clause is a NUMBER: coalesceDigest takes 603 arrivals — 500 distinct comments, 100 at-least-once repeats, and a mention, a blocker and a required assignment buried in them — and produces 6 notifications, with all three criticals still their own notification, 440 members summarised rather than discarded, and digestCapFaults's independently computed bound agreeing at 6. --> <!-- S8.8 marked after writing and reading src/notification-digests.ts and its spec in this session: 103 tests, 91 fault kinds and 7 refusals all proven reachable by the suite's own sweep, 64/64 source injections caught against a passing control arm, export-map violations 0, API pages 25 -> 26, full kit suite 8032 tests green. Named dedupeIdentityKey and not dedupeKey, because envelopes.ts already exports the latter for the TRANSPORT's redelivery key (subject#sequence) — decision 2's confusion, which the kit's own namespace must not carry. THE ADMISSION GATE also closes a hole in S8.7: planNotifications marks a delivery defer-to-digest whenever preference.outcome.decision === 'digest', consulting neither the obligation nor registration.digestEligible, and preferenceParityFaults — the function that refuses a digest choice on a notice above optional — is not one the plan calls. admitToDigest refuses it downstream and returns send-now, because declining to DELAY something is not a decision to withhold it. -->
  • S8.9 Add immutable activity-feed events with safe summaries. Evidence: rebuild from audit produces the same ordering.

    • S8.9.a Define activity event ID/version/type, target/revision, actor-safe reference, outcome, time, causation/correlation, source audit/event, and visibility. Evidence: activity schema.

    • S8.9.b Generate events from immutable audit/domain events rather than UI guesses and prevent update/delete outside retention/redaction policy. Evidence: source linkage tests.

    • S8.9.c Generate structured localized safe summaries with no raw content/secrets/protected identity/learner/biometric/unreleased asset. Evidence: canary-data tests.

    • S8.9.d Apply tenant/object/property/audience authorization to events, counts, actors, targets, linked refs, and export. Evidence: disclosure tests.

    • S8.9.e Order deterministically by authoritative time/sequence and stable tie-breaker; handle late/out-of-order/replayed events. Evidence: ordering tests.

    • S8.9.f Rebuild from empty feed using audit/event history and compare event IDs/order/content/visibility to stored projection. Evidence: replay parity test. <!-- S8.9 marked after writing and reading src/activity-feed.ts and its spec in this session: 169 tests, 130 fault kinds and 4 projector refusals all proven reachable by the suite's own sweep, 95/95 source injections caught against a passing control arm, export-map violations 0, entry points 135 -> 136, API pages 26 -> 27, full kit suite 8033 -> 8202 tests green. THE HOLE: the activity feed is not a projection of anything. GET /v1/activity is mounted, authorized, cached, paginated and rendered by a 668-line component, and every row in it is composed at read time from the CURRENT value of a mutable entity — notebook.updatedAt, habit.lastKeptAtIso, session.lastActivityAt. Its ids are the ENTITY's (nisaba-notebook-${notebookId}, and state.ts writes notebook.updatedAt = annotation.updatedAt in place); one row is stamped occurredAt: fallbackTimestamp, which is generatedAt, which is new Date().toISOString() at the top of the handler; and the sort has no tie-breaker, so rows at one instant keep the arrival order of the per-domain batches of Promise.all(authorizedDomains.map(...)) — an array derived from the VIEWER'S SCOPES. No table anywhere stores an activity event, so S8.9.f had neither of its two operands, which is why this item needed a PROJECTOR (projectActivityFeed) and not only a comparator. The kit-side tell is audit.ts, whose replay/parityFaults look like they cover this and are per-stream: a gapless sequence is a total order inside one stream and no order at all across the streams a feed merges. Two escapes in the injection harness were real gaps in my own spec rather than in the module — the arity check in activityEventIdComponents is load-bearing only for an OVER-LONG id, and the stream component of the order key is invisible to any test whose tied rows share a sequence number.

      2026-08-26 — THE ROUTE WAS FIXED, and the module is the reason it could be. GET /v1/activity now carries timeSource on every row (recorded | current-state | observed), so a reader can tell an append-only record from a rendering of a mutable entity from a row with no source time at all; sorts through a TOTAL comparator built to compareFeedOrder's rule minus the stream sequence, which none of the tables behind this route carries; REFUSES a row whose source time will not parse instead of stamping it with generatedAt and floating it to the top; drops occurredAt: fallbackTimestamp in favour of an observed row that sorts BELOW every timed row; carries the derived-from revision in the ids of the two rows read off mutable entities (nisaba-notebook-${id}@${updatedAt}, metis-tutoring-${id}@${updatedAt}); and reads arete check-ins from habitCheckIns — the append-only log the same write path was already filling, one row per (habit, user, day), which this route had never opened. It had been reading habit.todayKept and habit.lastKeptAtIso, two columns updateStreak overwrites, so one habit had exactly one check-in row for its whole lifetime: today's. 8 known-value cases in src/routes/activity-ordering.spec.ts. What is still open is projectActivityFeed itself: adopting ActivityEvent needs a per-stream sequence, and v1_arete_check_in, v1_nyx_logged_observation and v1_tara_sitting_completion_event have no such column — that is a migration, not a route change.

      The route's own suite was 5-of-10 RED before this pass and is now 10/10. Every one of the five failed on its FIXTURE, not its subject: they named veritas and metis, which V1_SCOPED_DOMAIN_IDS deferred to V1.2. That deferral is a deliberate single-point cut in libs/oshun/navigation/src/release-scope.ts — "their libraries, contracts, routes, and BFF services all stay in the tree" — so the route's veritas and metis branches are RETAINED CODE, not dead code, and lifting the deferral is one edit there. Fixtures moved onto V1.0 rooms the way assistant-session-store.spec.ts already records doing; the partial-failure test now fails nyx instead of veritas; and the metis test became a lock on the cut itself — it still seeds a real ingest job and now asserts the job reaches no feed and a deferred-only grant is refused. A DEFERRED ROOM'S STATUS IS ABSENT, NOT forbidden: out of release scope is a different claim from withheld from this member. -->

  • S8.10 Add review inbox and saved queues. Evidence: server-side filters, assignment, SLA, and pagination are tested.

    • S8.10.a Define inbox item projection for review/assignment/gate/signoff/ escalation target, revision, role, priority, state, due/SLA, freshness, blockers, and safe summary. Evidence: projection schema.
    • S8.10.b Resolve eligible items from current assignment/role/delegation/ quorum/stage/authorization and remove stale or completed work. Evidence: eligibility tests.
    • S8.10.c Implement server-side full-text/structured filters for domain, workbench, type, state, stage, assignee/role, priority, due/SLA, freshness, gate, and target. Evidence: query tests.
    • S8.10.d Implement deterministic sort/cursor pagination and live count/ status reconciliation under concurrent assignment or decision changes. Evidence: pagination tests.
    • S8.10.e Persist private/team saved queues with query, columns, grouping, sort, density, sharing, schema migration, and authorization. Evidence: saved queue tests.
    • S8.10.f Support accept/start/block/complete/decline/reassign/escalate/ bulk actions with expected revision, impact, audit, and partial result. Evidence: inbox action tests.
    • S8.10.g Provide keyboard/screen-reader/virtualization/deep-link and narrow review experience. Evidence: accessible journey. <!-- S8.10 marked after writing and reading src/review-inbox.ts and its spec in this session: 221 tests, 84 fault kinds plus the admission, action, saved-queue and cursor refusal vocabularies and all five eligibility grounds proven reachable by the suite's own sweep, 61/61 source injections caught against a passing control arm, export-map violations 0, entry points 136 -> 137, API pages 27 -> 28, full kit suite 8207 -> 8428 tests green. THE HOLE: the narrowing happens where the rows are DRAWN and not where they LIVE. AdminWorkspaceStateStore .getInboxWorkspace folds six per-workspace queues into one array, ships the whole thing, and says so in its own comment — "Filtering / sorting happen client-side (the contract owns the predicates)" — which UnifiedInboxPanel.tsx repeats: "The panel is fully client-side: filters and sort live in local state". matchesAdminInboxFilter and sortAdminInboxItems in @oshun/contracts are correct, tested, and on the wrong side of the wire, so every row the filter removes has already crossed it — including the ones a free-text search reaches by concatenating title and subtitle, where subtitle is a customer display name for support and a region for a privacy request. Four faults are visible in that same source: the count is items.length over an array that still holds every terminal row (buildInboxItemFromReview computes terminal only to set slaStatus: 'paused' and pushes anyway), so "open items" counts finished work; deriveInboxSlaStatus returns no_sla both for a surface with no clock and for a due date that failed to parse, so a breach and a promise nobody made are one value; normalizeInboxPriority (null) puts every model row in a band called unknown that PRIORITY_RANK then ORDERS; and sortAdminInboxItems ends at Date.parse(b.lastUpdatedAt) - Date.parse(a.lastUpdatedAt) with no unique backstop, which is both not total — so indexes.ts's own rule says it cannot be paged over, and it is not — and NaN for an unparseable time, which every JavaScript sort reads as equal. Saved queues repeat the shape one level up: saved-views-store.ts stores isShared and lists only users.get(input.userId), so a queue marked shared is visible to exactly one person and wears a badge saying otherwise, and importState rehydrates a saved filter verbatim with no version and no migration. THE GREP: the parent's evidence clause begins with the words "server-side" — take the surface it names and go and look at which side of the wire its predicates run on. Three faults I declared were unreachable until the registry sweep found them: count-includes-what-the-list-excludes had no producer, the four cursor-not-bound-to-* kinds had none either, and off-axis-segment-moves-with-the-direction was UNSATISFIABLE BY CONSTRUCTION because it asked orderInbox whether it agreed with itself. One injection escape was a real point about the module rather than the spec: the grounds were collected in the same order as their strengths, so the sort that decides between them was dead code that looked load-bearing — collecting cheapest-lookup-first makes the rule the only thing that answers. -->
  • S8.11 Add reviewer handoff and temporary delegation. Evidence: expired delegations stop authorizing actions.

    • S8.11.a Define handoff for selected assignment/review target with source/ destination reviewer, reason, context, unresolved work, due/SLA, and effective time. Evidence: handoff schema.
    • S8.11.b Validate destination eligibility, role/qualification, tenant/ object access, workload/capacity, COI, and availability. Evidence: eligibility negative tests.
    • S8.11.c Transfer ownership atomically, preserve prior reviewer history, comments/evidence, due/SLA policy, notifications, and audit. Evidence: handoff integration tests.
    • S8.11.d Define delegation scope by action/role/domain/workbench/object/ target, delegator/delegate, start/expiry, reason, and nondelegable actions. Evidence: delegation schema.
    • S8.11.e Require delegator authority and delegate eligibility, prohibit privilege amplification/transitive redelegation unless explicit, and enforce separation/COI. Evidence: authorization tests.
    • S8.11.f Activate/revoke/expire delegation by server time and re-evaluate pending work/approvals without retroactively altering completed actions. Evidence: clock-controlled tests.
    • S8.11.g Display effective delegated actor on actions/audit and notify affected parties. Evidence: attribution tests. <!-- S8.11 marked after writing and reading src/reviewer-handoff.ts and its spec in this session: 141 tests, 107 fault kinds plus the six grant standings, the eight authorization refusals, the eight sweep outcomes, the six attribution shapes and the six destination axes all proven reachable by the suite's own sweep, 141/141 source injections caught against a passing control arm, export-map violations 0, entry points 137 -> 138, API pages 28 -> 29, full kit suite 8429 -> 8570 tests green. THE HOLE: the parent's evidence clause is "expired delegations stop authorizing actions", and before writing anything the question worth asking is what would make that sentence FALSE rather than merely untested. It takes a delegation that cannot expire, and that is what was here. CanonicalClaims (S4.2.a) carries two grants on adjacent lines: supportGrant is a SupportGrantClaim of { grantId, expiresAt, reason } which resolveSession turns into supportGrantLive = expiresAt > now, which authorization.ts threads through RequestContext, tenancy.ts refuses on and router.ts carries to the edge — five modules of lifecycle for one grant kind — while delegationGrant is string | null, validated for spelling and for being present whenever onBehalfOf is, then copied verbatim into ActorRef.delegated.grant. There is no delegationGrantLive; the identifier does not occur anywhere in this repository. Fourteen sites read that string and two of them AUTHORIZE on it — restore.ts refuses unless request.actor.grant === policy.restoreGrant, stages.ts unless command.actor.grant === transition.requiresScope — so authority is a string comparison against a name, and a string does not expire. parseActorRef even states the reason the field must exist, "or a revoked delegation cannot be traced to what it authorised", and the register it would be traced to was never built. The product side matches: AdminWorkspaceStateStore.delegateReview takes { reviewId, delegateToReviewerId, note }, overwrites review.assignedReviewerId, increments a scalar delegationCount and returns, with no window, no scope, no revocation route and no second actor, so "temporary delegation" is a permanent transfer of ownership and no delegation in the product has an end that could lapse. Four of AdminReviewerRecord's fields — active, role, specialties, availabilityStatus — are read by nothing on that path, which is S8.11.b's five eligibility axes with none of them enforced; the capacity counter is Math.min(dailyCapacity, currentAssigned + 1), so overload is not merely unrefused but unrepresentable and the >= dailyCapacity test after it can only ever be an equality; the predecessor survives only inside the free text of "Delegated from Ama to Kofi: <note>"; and the audit event is written with actorId: userId and actorLabel: target.displayName, so the row records the giver's id under the receiver's name, which reads correctly to a person and wrongly to a query and is unrecoverable afterwards. THE GREP: when the parent's evidence clause names a LIFECYCLE, find the field that carries the thing whose lifecycle it is and ask whether it has a shape a lifecycle could be stored on — then find its SIBLING that does, because the sibling proves the omission was not ignorance. So: A HANDOFF MOVES THE WORK AND A DELEGATION LENDS THE AUTHORITY, the first permanent and the second nothing without an end; A GRANT IS A ROW IN A REGISTER AND NOT A TOKEN IN A REQUEST, so resolveGrant can FAIL and a work row whose only ground names a grant the register cannot resolve is a fault rather than an absence; EXPIRY IS EVALUATED AGAINST A SERVER CLOCK AND THE CLOCK'S PROVENANCE IS PART OF THE ANSWER; AN EXPIRY THAT IS ONLY CHECKED ON WRITE IS NOT AN EXPIRY, so the lifecycle has a SWEEP whose last run is part of the observation and a register whose ended grants are older than its last sweep is refused; A REVOCATION IS NOT AN EARLY EXPIRY, so the grant carries issuedNotAfter beside notAfter and an end that moved with no revocation is a fault; WORK ALREADY DONE UNDER A LIVE GRANT STAYS DONE while work STARTED and not finished is withdrawn rather than grandfathered, because half an approval is not an approval; WITHDRAWN WORK HAS TO GO SOMEWHERE, since a row belonging to nobody looks on every surface in this kit exactly like a row that has been dealt with; A GRANT CANNOT OUTLIVE THE AUTHORITY IT WAS CUT FROM, which is why the delegator's CURRENT standing is an input and lapsed-with-its-source is its own standing — S5.12's frozen delegatorMaxBreadth is the right cap in the wrong tense and no window check will ever find it; UNSTATED SCOPE IS NOT UNLIMITED SCOPE across all six dimensions, so all is a variant carrying a written justification and an empty these selects nothing; NONDELEGABLE ACTIONS ARE DERIVED from ACTIONS_BY_GROUND.delegation rather than restated, and the one rule that is not derivable is a PAIR, completing a sign-off, because S2.13 is about what is attested to and not about the verb; ELIGIBILITY IS EVALUATED AT THE INSTANT THE HANDOFF TAKES EFFECT and not at the instant it is asked for; and A DELEGATED ACTION HAS TWO ACTORS AND BOTH GO IN THE RECORD, where the failure worth its own fault kind is neither omission but the CROSS. -->
  • S8.12 Add redaction rules for protected sources, learner records, biometric data, and unreleased assets. Evidence: notification and activity snapshots contain no protected payloads.

    • S8.12.a Define protected-source identity/contact/document/relationship/ location fields and allowed pseudonym/role summaries per audience. Evidence: source redaction policy.

    • S8.12.b Define learner identity/enrollment/activity/assessment/ accommodation/integrity/credential fields and educational-role access. Evidence: learner redaction policy.

    • S8.12.c Define face/voice/body/motion/health/embedding/template/derived biometric fields and consent/purpose/role access. Evidence: biometric redaction policy.

    • S8.12.d Define unreleased content/title/media/script/build/research/ model/product/embargo metadata and audience/embargo access. Evidence: unreleased asset policy.

    • S8.12.e Apply structured allowlist/redaction before template rendering, summary generation, channel routing, activity projection, search/index, export, logs, and support tools. Evidence: boundary coverage tests.

    • S8.12.f Prevent inference through counts, actor names/avatars, target titles/URLs, snippets, grouping, timing, errors, and deep links. Evidence: side-channel tests.

    • S8.12.g Use canary values across notifications/activity/digests/emails/ push/webhooks/exports/snapshots and fail on any payload leak. Evidence: automated leakage suite. <!-- S8.12 marked after writing and reading src/redaction.ts and its spec in this session: 119 tests, 75 fault kinds across nine vocabularies, and the four embargo standings and six encoding forms all proven reachable by the suite's own sweep. 85/85 source injections caught against a passing control arm; the two that escaped the first pass were real spec gaps and are recorded below. Export-map violations 0, entry points 138 -> 139, API pages 29 -> 30, packed-consumer smoke OK at 139 subpaths, full kit suite 8688 -> 8690 tests green.

      THE HOLE: TWO OF THE FOUR CATEGORIES THIS ITEM NAMES COULD NOT BE STATED ON THE FIELD TYPE THAT DECIDES WHETHER A VALUE LEAVES THE PRODUCT. activity-feed.ts has carried all four in ACTIVITY_CLASSIFICATIONS since S8.9, with an audience ceiling for each and a rule that an authored title may not appear on a row carrying one — its own comment even says "these are the classes S8.12 will write rules for". protection.ts's SensitivityClass, which is the vocabulary a SummaryField is classified in and therefore the one EXPOSURE_CEILING is keyed on, named two. There was no member for a learner record or an unreleased asset, so both could only be declared sensitivity: null — and exposurePermits answers true for a null class at EVERY rung, including carrier-plaintext and third-party-endpoint. The activity feed refused to carry the event and the notification renderer, which is the surface that actually leaves the building, waved the same fact through to an SMS and to a webhook. Both classes are added here with ceilings (unreleased-asset in-product, since an embargo is broken by a mail preview; learner-record account-mailbox, since that is where an institution has always written to a student and the next rung is a lock screen somebody else is reading), and the regression is tested as a PAIR through exposurePermits — the refusal at carrier-plaintext beside the permission for null — because the first case could not previously be written at all.

      THE PRODUCT SIDE HAS THE SHAPE THAT GAP PRODUCES. In @oshun/messaging-channels, decideMemoryIngestion returns a MemoryIngestionDecision whose redactionRules is a list of English verb phrases — strip-channel-handles, strip-phone-numbers, strip-initdata. The identifier redactionRules occurs THREE times in this repository: the interface, that producer, and one test asserting the array .toContain('strip-initdata'). Nothing interprets a rule and no production call site calls the function. Its two neighbours in the same file are the same shape: consentRevocationCascade returns strings like revoke-binding:u1:telegram-bot that no executor consumes, and crisisCrossChannelSuppression returns [...new Set(input.channelIds)] — every channel it was handed, deduplicated — from a function whose name is suppression. Downstream, buildPushPayload puts deep-link into data even on a SILENT push, and /v1/notifications composes a flat { title, body } per row with no field structure and no class, so there is nowhere on the wire for any of this to be said. THAT REMEDIATION IS NOT IN THIS ITEM and is recorded here as found.

      THE GREP THAT FINDS THIS VEIN: take the vocabulary the parent phase names its protected classes in, find the module that already carries all of them, then check the vocabulary the module that decides TRAVEL is keyed on. The surface that refuses to RENDER is never the surface that puts a value on a wire, and an unstatable class defaults to unlimited.

      TWELVE DECISIONS. REDACTION IS AN ALLOWLIST AND THE DEFAULT FOR AN UNKNOWN FIELD IS TO WITHHOLD IT, because a denylist is a list of the leaks somebody has already had — the projection returns released, withheld and unregistered as three outcomes, and a denylist implementation passes every other test in the spec and fails exactly that one. DISCLOSURE IS A LADDER AND NOT A BOOLEAN (none < existence < value < basis), because an instructor must know a student HAS an accommodation, since they have to implement it, and must not know what it is FOR. PRESENCE-DISCLOSURE IS A PROPERTY OF THE FIELD AND NOT OF THE CATEGORY: enrollment existing is what being a student means and accommodation existing is a disability disclosure, in one category on one row, so protection.ts's class-level list stays a FLOOR this refines rather than a duplicate. THE AUDIENCE LADDER MEASURES BREADTH AND NOT ENTITLEMENT, so the substitute map is deliberately not monotone in it — auditors ranks wider than workbench-operators and receives MORE, and the pair is asserted, because a map sorted by rank is what somebody writes when they tidy this up. A PSEUDONYM IS ASSIGNED AND NEVER DERIVED, and ONE CODENAME PER SOURCE PER STORY, because a codename reused across two stories publishes that they share a source: the redaction is intact in both documents and the leak is in the PAIR. A ROLE SUMMARY IS A REDACTION ONLY ABOVE A COHORT FLOOR, and an unmeasured cohort is refused separately from a measured small one. A DERIVED BIOMETRIC IS THE BIOMETRIC — an embedding re-identifies exactly as well as the image, survives every deletion aimed at the image, and is the artefact a withdrawal fails to reach, so a withdrawal that has not reached the derivatives is its own fault and not a variety of "withdrawn". CONSENT IS PER PURPOSE AND THREE PURPOSES MAY NEVER BE INFERRED FROM ANOTHER (model-training, voice-synthesis, research), because each makes something durable out of a capture the subject agreed to for something that ends. AN EMBARGO WITH NO LIFT INSTANT IS NOT AN EMBARGO, and an ABSENT embargo record is a fault rather than a permission, since defaulting it to release makes forgetting to write one the fastest route to publication. THE PROJECTION HAPPENS ONCE UPSTREAM OF ALL EIGHT BOUNDARIES and each boundary declares which projection it consumed, because eight redactors is eight places to forget and the one that gets forgotten is the log — the fault that matters is boundary-re-read-the-record-after-the-projection, since enrichment downstream of a redaction reads as a feature in the diff. THE SIDE CHANNEL IS THE PAYLOAD: a small count, an avatar URL that is a hash of an address, a deep link whose path carries the record id, a GROUPING key the reader may not see, and an emission time equal to the event time each disclose with every field withheld — and a small count is SUPPRESSED rather than rounded, because a rounded count still moves when one row is added, so two readings difference back to the row. AND A CANARY IS FOLLOWED THROUGH THE ENCODINGS THE CHANNELS APPLY, because a scan that only calls includes misses the value percent-encoded into a link, base64'd into a push data field, backslash-escaped by a serializer or entity-escaped into a mail part — which is five of the six forms.

      TRAPS. THE INJECTION HARNESS FOUND TWO REAL SPEC GAPS, both of which were holes in my own tests rather than in the module, and both are now closed: the base64 corpus mapped entirely into [A-Za-z0-9], so swapping the alphabet for the base64url one — the mistake somebody makes while "fixing" a URL — passed every sample; and the token floor in the codename check was untested from below. THE TOKEN FLOOR IS TWO AND NOT THREE: Li, Wu, Ho, Ng and Yi are whole surnames, so three is a check that works for European names and not for Chinese, Korean or Vietnamese ones, and it is now pinned from both sides by two cases that each fail against exactly one of the wrong answers. A RAW NUL BYTE IN A SOURCE FILE MAKES IT INVISIBLE TO GREP: one reached this module as a composite-key separator and the mandatory stub-indicator scan came back clean because grep had skipped the file entirely. It is written as \u0000 now. SIX OTHER PRODUCTION MODULES IN THIS PACKAGE CARRY RAW NULs and are invisible the same way — collaborator-presence, determinism, diff, errors, retries, router — so every grep-based signoff over them has been vacuous; re-scanned with grep -a here, they hide nothing actionable, but the hazard is real and unfixed. FINALLY, pnpm nx production-scan is RED and was red before this change: gate-definitions.ts (S2.9, 2026-08-02) exports CONFORMANCE_FAULTS/ConformanceFault, and src/testing/model-provider-conformance.ts (S5.1, 2026-08-04) later declared the same names, so the scan's name-derived token list flags the production module. All three inputs are byte-identical to HEAD. -->

S9 — Taxonomy, localization, calendar, catalog, and discovery#

  • S9.1 Extract tighten-only registry merge with code floors and provenance. Evidence: every attempted loosening fails. <!-- S9.1 marked after writing and reading libs/oshun/workbench-kit/src/ registry-merge.ts and its spec (94 tests, 100% of lines and functions; the residual uncovered branches are the per-case shape guards that narrow the union for the compiler and two sort tie-breaks). Entry point @oshun/workbench-kit/registry-merge added to the DERIVED export map (generate-workbench-kit-export-map.mjs, 140 entry points, no drift, no violations), package.json, and the tsconfig.base.json alias table; public-api.spec.ts resolves it (149 tests) and tools/production-scan.mjs is GREEN over the emitted graph.

    THE VEIN: A TIGHTEN-ONLY RULE STATED OVER THE LABEL OF A CHANGE INSTEAD OF
    OVER ITS VALUE CANNOT REFUSE ANYTHING. The grep is the word the rule is
    written in — `tighten` — followed by, for every comment promising that a
    validator rejects relaxations, looking at the RECORD the validator is
    handed and asking whether both sides of the comparison are on it. Three
    answers in this tree. `libs/metis/learning/src/mastery/tenant-override.ts`
    is honest and narrow: `enforceTighten(candidate, baseline, 'higher' |
    'lower')` throws, over five fields named one by one, two layers, numbers
    only, no record, and a silent fall-through to the baseline for anything
    that is not a number. `policy-plugins.ts` (S5.9.e) has the general shape —
    `weakeningFaults` checks a requested value against a code floor AND a
    domain floor and carries the observation this module is built on, that a
    domain floor beneath the code floor "was accepted once and has been
    quietly ineffective ever since; whoever set it believes it is in force" —
    but `weakeningFaults`, `EvaluationRecord` and `PolicyOverride` each occur
    exactly three times in the repository: declaration, barrel, own spec.
    NOTHING PRODUCES AN `EvaluationRecord`, so it is a correct checker with no
    merge in front of it. And the live one, fixed in this commit:
    `@oshun/persona-registry`'s `launch-persona-config.ts` promised above
    `PersonaPolicyOverride` that "the validator rejects relaxations", the
    override was `{ ruleId, action, rationale }` with NO VALUE ON IT, and the
    check was `tighteningActions.includes(override.action)` over the same
    three strings the `action` union already permits — true for every
    inhabitant of the type. Eleven shipped in the launch catalog, including
    `{ ruleId: 'tone.pace', action: 'raise-floor' }`. Raise it to what?
    
    NINE DECISIONS, each with a test that fails without it. PRECEDENCE NEVER
    PICKS THE VALUE — the truth table's `stricter` and `looser` rows are
    identical whichever layer is broader, and rank appears only in the two
    `equal` rows, where it decides ATTRIBUTION and nothing moves. THE TIE GOES
    TO THE BROADER SOURCE and the narrower is reported redundant, because the
    broader one is what still holds when the narrower is deleted. A FLOOR IS
    NOT A DEFAULT AND A BELOW-FLOOR CONTRIBUTION IS REFUSED RATHER THAN
    CLAMPED, since clamping succeeds, discards the author's value and leaves
    them believing it is in force — which is why one refusal fails the WHOLE
    merge. THE CODE LAYER CANNOT BE CONTRIBUTED, because a source that can
    name itself `code` is a source that can raise its own authority. THE FLOOR
    TABLE IS AN ALLOWLIST AND AN UNDECLARED KEY IS REFUSED, because a merge
    that passes unknown keys through is last-write-wins for exactly the keys
    nobody has thought about. INCOMPARABLE IS A REFUSAL AND NOT AN INVITATION
    TO COMBINE: intersecting two audiences yields a value stricter than both
    that NOBODY WROTE. A PROHIBITION IS NOT A FLOOR AND STRICTER IS NOT SAFER
    (`retentionDays: 0` is a lower ceiling than 30 and destroys the audit
    trail). DELETION IS A LOOSENING. AND A CONTRIBUTION IS ADMITTED BY ITS
    PROVENANCE BEFORE ITS VALUE IS READ — expiry, staleness, scope, signature
    and a content hash recomputed against the value — because the strictest
    number in the file is worth nothing if the file was signed before the
    number was edited. -->
    
    • S9.1.a Define registry entry identity/version, owner, value schema, scope, status, source, authority, effective period, and provenance. Evidence: registry contract. <!-- evidence: RegistryEntry carries all nine, and authority is 'proposal' | 'binding-floor' rather than an issuer name, because the question a merge has to answer about a contribution is whether it COMPETES on strictness or BOUNDS everything narrower than it — the generalisation of the domainFloor policy-plugins.ts hard-codes. REGISTRY_SOURCE_LAYERS orders the seven from broadest to narrowest, and CONTRIBUTABLE_LAYERS excludes code. -->
    • S9.1.b Define immutable code floors for required entries, minimum severity/constraint, prohibited values, trust/safety requirements, and schema/version bounds. Evidence: floor schema. <!-- evidence: RegistryCodeFloor + defineCodeFloors, which deep- freezes and fingerprints the table (order-independent) and REFUSES ten contract mistakes that would otherwise surface as a merge behaving strangely: a floor on its own prohibited list (the key then has no lawful value at all, including the one it falls back to), an ordinal floor outside its own vocabulary, a severity field with no vocabulary, a vocabulary on a field that never reads one, two floors for one key, an alias cycle, an alias to nothing, an alias shadowing a real key, a floor whose shape disagrees with its field kind, and a key the reading code requires that no floor declares. -->
    • S9.1.c Define source precedence for code, signed policy, tenant, project, domain, environment, and runtime overrides without last-write-wins. Evidence: precedence truth table. <!-- evidence: PRECEDENCE_TRUTH_TABLE, eight rows over relation x breadth, total and duplicate-free by test. The without last-write-wins clause is discharged by two assertions rather than by prose: precedenceOutcome('stricter', true) === precedenceOutcome('stricter', false) and the same for looser. Rank is read in exactly two rows and both are equal, where it decides which source is CREDITED and the value does not move either way. -->
    • S9.1.d Implement field-aware merge that permits additions and equal/ stricter values while rejecting deletion, weakening, wider audience, lower threshold/severity, longer expiry, or unknown comparison. Evidence: tighten-only property tests. <!-- evidence: eight declared orderings (REGISTRY_FIELD_KINDS) are the eight this clause names — a lower threshold and a lower severity are the same move in two vocabularies, a wider audience is a superset, a longer expiry is a larger duration, an unknown comparison is opaque or an ordinal outside its vocabulary, and a deletion is the removal of any of them. every field kind refuses a loosening of its own ordering walks all eight, so a comparison written the wrong way round fails there even though every hand-written example still passes. A field kind is DECLARED and never inferred: the spec's gate.minSeverityToBlock runs its vocabulary critical -> info, because a gate that blocks at low blocks on strictly more, and the same five strings in the natural order mean the opposite. -->
    • S9.1.e Preserve contributing source/version/hash/actor/decision/time and rejected override reason for every effective field. Evidence: field provenance tests. <!-- evidence: FieldContribution carries all six plus role (code-floor / binding-floor / proposal) and disposition (effective / redundant-equal / superseded-by-stricter) with a sentence for each non-effective one. The code floor appears AS a contribution, so a field held only by the floor still has a row. explainField renders it in the terms support is actually asked in: "your project set 90 days and the tenant requires 30" is the question, and a merge returning only the winning value leaves everyone staring at a config file that says 90. -->
    • S9.1.f Reject conflicts, ambiguous partial order, incompatible schema, cycles/aliases, unsigned required inputs, and stale sources. Evidence: negative fixture matrix. <!-- evidence: 37 fault kinds in REGISTRY_MERGE_FAULTS, and the matrix is STRUCTURAL rather than exemplary — every fault kind is reachable iterates the union and fails if any member has no fixture that produces it, so a fault added without a way to trigger it is a test failure rather than a comment. Aliases are resolved BEFORE anything else is checked, which is what makes the old and the new name compete for one field instead of the looser one being in force under a name nothing reads. -->
    • S9.1.g Generate deterministic effective registry and diff/explanation. Evidence: repeat generation and loosening mutation tests. <!-- evidence: three input permutations produce deep-equal registries and an identical digest; re-stamping provenance does not move the digest and any effective value does. strictestOf computes MAXIMAL ELEMENTS rather than folding a running maximum, because a fold's answer depends on visit order for an unordered set — [{a,b},{b,c},{a,b,c}] left to right reports the first pair and right to left never compares them — and the spec asserts three orderings agree. The loosening mutation test walks every entry, moves that one a step toward looser in its own ordering, and requires the merge to refuse. diffEffective adds reattributed: a value that did not move but is now held by a different source is one deletion away from moving, and calling that unchanged is true about the value and wrong about the system. -->
  • S9.2 Add registry draft, review, publish, deprecate, alias, and migration lifecycle. Evidence: published artifacts remain resolvable. <!-- S9.2 marked after writing and reading libs/oshun/workbench-kit/src/ registry-lifecycle.ts and its spec (87 tests, 100% of statements, lines and functions; the residual uncovered branches are shape guards and ?? defaults). Entry point @oshun/workbench-kit/registry-lifecycle added to the DERIVED export map (141 entry points, no drift, no violations), package.json and the tsconfig.base.json alias table; public-api.spec.ts resolves it (150 tests), the whole kit suite is green (137 files, 8873 tests) and tools/production-scan.mjs is GREEN over the emitted graph (132 production entry points, 264 files).

    THE VEIN: A LIFECYCLE STATE IS A STATEMENT ABOUT WHAT MAY BE WRITTEN NEXT,
    AND IT HAD BEEN INSTALLED ON THE PATH THAT READS WHAT WAS ALREADY WRITTEN.
    The grep is not a word. It is COUNTING RESOLVERS: one resolver means the
    read path and the write path are the same path, and every rule written for
    one of them is in force over the other. `libs/shared/event-bus/src/
    topic-registry.ts` is the live instance and it is a real registry with a
    real lifecycle — 65 topics, schema versions carrying `lifecycle: 'active' |
    'deprecated' | 'retired'`, a `defaultSchemaVersion`, a declared payload
    shape, and enforcement that throws `Schema version 1.0 for event topic X is
    retired`. It has exactly one resolver and it is called `resolveForPublish`.
    `EventBus.publish` stamps `eventSchemaVersion` onto every envelope and
    persists it under a TTL; `EventBus.subscribe` hands that envelope to a
    handler and NEVER RESOLVES THE VERSION IT NAMES. So the schema lookup that
    exists is the one the author performs, the lifecycle gate sits on it, and
    the reader of an already-written artifact has one way through:
    `allowRetired: true`, a flag whose name says the caller is doing something
    exceptional at the moment they are doing the only thing left. Retiring a
    version makes every event ever published under it unreadable through the
    registry, and that is the moment the archive of them is at its largest.
    
    The same file carries S9.1's vein in its own dialect, which is why it is
    worth naming twice. Each schema version declares `payloadSchema.fields` and
    `payloadSchema.required`, and the only check over them,
    `validateSchemaVersion`, asserts that every member of `required` appears in
    `fields` — THE DECLARATION COMPARED WITH ITSELF. 64 of the 65 topics ship
    no `validate` function and `validatePayload` answered
    `{ valid: true, issues: [] }` when there is none, so a payload missing every
    field the registry calls required was reported valid by the registry that
    calls them required. And `deprecates?: string` occurred twice in the whole
    repository: the declaration and one spec fixture. Nothing set it in
    production and nothing read it.
    
    BOTH ARE FIXED IN THIS COMMIT, not merely described.
    `EventTopicRegistry.resolveForConsume` is the read path: it never branches
    on a lifecycle state, reports it as advisory text instead, and returns
    nothing only for reasons about the registry (topic not registered, version
    not declared). `validateAgainstDescriptor` makes `required` a requirement
    and every result now carries `checkedBy`, so "nothing validated this" is no
    longer spelled the same way as "this is valid"; `EventBus` runs it, refusing
    in `enforced` mode and reporting in `advisory`, which is the dial that
    already existed for exactly this and keeps a check that was never on from
    refusing traffic on the day it is switched on. A `deprecates` naming a
    version its topic does not declare is now a registration error, and
    `resolveForConsume` reads it. A publish against a non-active version is
    logged rather than passing silently under `allowDeprecated: true`. The
    event-bus suite is green (42 tests, 5 new) and so is every dependent:
    veritas-cascade-worker, trust-safety, deletion-fanout, isis/event-publisher,
    service-discovery, domain-veritas (331).
    
    TEN DECISIONS in the kit module, each with a test that fails without it.
    THERE ARE TWO RESOLVERS AND THEY ANSWER DIFFERENT QUESTIONS —
    `resolveCurrent` answers what to write against, consults the lifecycle and
    names the replacement when it refuses; `resolveHistorical` answers what a
    reference meant and does not read the state to decide, only to report. A
    VERSION RESOLVES THROUGH ITS OWN ALIAS TABLE AND NOT TODAY'S, because an
    alias is a claim with a date and resolving a 2024 reference through the 2026
    table is how a document silently acquires a meaning nobody gave it — the
    reference did not move, the dictionary did. IDENTITY IS MINTED ONCE AND THE
    SEQUENCE POSITION IS ASSIGNED LATER, so an abandoned draft leaves no gap and
    nothing sorts by an id issued before the order was known. A REVIEW DECISION
    IS BOUND TO THE CONTENT IT REVIEWED, which is the only thing separating
    "three people approved this" from "three people approved something, and then
    it was edited". PUBLISH DECIDES AND DOES NOT WRITE, returning the artifact
    with every cache key to invalidate and notification to send, because a
    function that both computes and performs them is one where the decision can
    be skipped by calling the performer. AN EFFECTIVE TIME MAY BE IN THE FUTURE
    AND NEVER IN THE PAST. A LOOSENING ACROSS VERSIONS IS LAWFUL AND MUST BE
    NAMED BY KEY — S9.1 refuses a loosening between contributions competing in
    ONE merge, and across versions a control is sometimes deliberately relaxed,
    so refusing outright would be wrong and allowing it silently is how a
    control evaporates over three releases that each looked small. AN ALIAS IS
    AN INPUT VOCABULARY. A SPLIT IS NOT A FUNCTION. AND ABSENCE OF MEASUREMENT
    IS NOT ABSENCE OF USE. -->
    
    • S9.2.a Define immutable registry version identity and draft/review/ approved/published/deprecated/retired state machine. Evidence: lifecycle tests. <!-- evidence: RegistryVersionIdentity carries versionId (minted at draft open, never reused, opaque so nothing can sort by it) and publishedOrdinal (assigned at publish, dense — the abandoned-draft test opens v2, never publishes it, and requires v3 to take ordinal 2). LIFECYCLE_TRANSITIONS is data, successorStates is DERIVED from it rather than written beside it, and the authority lives on the EDGE not the state, which is the only way in-review -> draft can be both an author withdrawing and a reviewer returning. There is no edge out of retired: retirement is taken on evidence about a closed interval, which cannot expire. reaches every state from draft and names every authority on at least one edge are structural. -->
    • S9.2.b Create/edit drafts against an explicit base with validation, semantic diff, conflict, ownership, reason, and provenance. Evidence: draft tests. <!-- evidence: a base is REQUIRED whenever one exists, because an implicit base is whatever was head when the draft happened to be read — which makes the draft's meaning depend on a moment nobody recorded and makes the base-moved check at publish impossible to state. Ownership is longest-prefix so gate. and gate.minSeverityToBlock can belong to different teams, and a key with NO owning prefix is not a fault: an ownership table that has not reached a new area should slow nobody down. contentHashOf covers content and nothing else — the same entries opened by a different author at a different instant hash identically, which is what stops an edit to the reason invalidating every approval. every semantic change kind is produced iterates SEMANTIC_CHANGE_KINDS, so a diff that can name a change it cannot detect is a test failure. -->
    • S9.2.c Define review assignments, required roles/quorum, comments, decisions, tighten-only/floor checks, and stale-review invalidation. Evidence: review tests. <!-- evidence: quorum counts DISTINCT reviewers whose approval carries the draft's current content hash, so one reviewer approving twice is one approval and an edit drops every tally to zero while still REPORTING the stale ones (a reviewer whose approval silently disappeared concludes the system lost it). A self-approval is a fault rather than an ignored row, for the same reason. One request-changes against the current content blocks however many approvals stand beside it, because a quorum that can outvote an objection is a headcount; a block with an empty comment is itself refused, since it cannot be actioned. An abstain is recorded and never counted, because "nobody looked" and "somebody looked and would not decide" are different facts. The tighten-only and floor checks are not reimplemented here: publish runs S9.1's mergeRegistry over the draft once per scope the version names. -->
    • S9.2.d Publish atomically with immutable artifact/hash, effective time, compatibility, notifications, cache/index invalidation, and audit. Evidence: publish tests. <!-- evidence: publishVersion returns the artifact, the next catalog, the semantic diff, the invalidations, the notifications and the audit — and PERFORMS none of them, which is stages.ts's rule for transitions applied to publication. The input catalog is asserted unchanged. Order matters: identity, base, review, content, then effect, because checking the effective registry first spends the expensive work on a change whose base moved under it. base-moved is the one that catches the real failure — an approved draft based on a version that is no longer head was reviewed against a registry that no longer exists, and publishing it reverts somebody else's version without anybody deciding to. A surface that declares NO key dependency is invalidated by everything, because a projection whose inputs nobody wrote down is always potentially stale. Compatibility is about RESOLVABILITY, not agreement: migrated is the distinction worth having, a change that would break references and shipped its bridge in the same version. Two bugs were found by coverage rather than by reading: a multi-scope version could never publish (S9.1's merge refuses an entry addressed elsewhere, correctly — the per-scope merge has to be handed the entries for that scope), and a rename read as a deletion until the base was aligned through the NEW version's alias table, which made every bridged rename report a regression. -->
    • S9.2.e Deprecate with replacement, migration guide, warning, supported- through/sunset, owner, and usage telemetry. Evidence: deprecation tests. <!-- evidence: the artifact is untouched and the notice sits beside it, which is decision 1 seen from the storage side. withdrawn is a first-class replacement with a signed rationale rather than a null successor, because "there is no successor" is a decision somebody made and a missing field is a decision nobody made. A sunset not after the deprecation is refused: a sunset in the past is a retirement wearing a warning. deprecationWarning reads nothing but the notice, so two consumers receive byte-identical text — the form in which "without exposing other consumers" is checkable rather than promised. THE RULE THAT MATTERS IS THE THIRD RETIREMENT BLOCKER: zero rows in a telemetry table is not zero usage, it is zero measurement, and the two are indistinguishable to anything that only sums. The observation windows are unioned and compared against the whole interval from sunset to now, and each gap is reported AS AN INTERVAL so somebody can go and find out what was running then. The spec asserts the measured-zero and the unmeasured zero are the same number and only one of them is evidence. An amendment may extend a sunset and may not take one back. -->
    • S9.2.f Add alias with stable source/target, scope, noncollision, noncycle, display/search behavior, and deprecation status. Evidence: alias tests. <!-- evidence: eight refusals — collision with a live key, missing target, duplicate in one scope, cycle, a target with no entry in a scope that covers the alias, displayed-but-unsearchable, deprecated-and- displayed, and self-reference. searchable and displayed are separate fields because they are separate decisions with OPPOSITE defaults: you may find a record by the name it was filed under, and rendering that name back to a person is how the rename fails to finish, since the retired spelling re-enters new documents by copy-paste. A cycle returns null rather than the last name before the loop, because a resolver that answers from inside a cycle gives a different answer depending on where it entered; resolveHistorical reports that as a key absence rather than resolving it. -->
    • S9.2.g Define migration mappings for removed/merged/split/moved/renamed entries with loss/conflict report and resumable backfill. Evidence: migration tests. <!-- evidence: A SPLIT IS NOT A FUNCTION. One key becoming three has no mapping unless a discriminator picks the target, so the plan reports the row ambiguous rather than taking the first target — a backfill that silently picked is indistinguishable afterwards from one that knew. A merge is lossy and a removal is dropped, both named in migrationLossReport. The plan is deterministic (a shuffled mapping list produces an identical hash and an identical step order) and the cursor carries the plan hash: RESUMING A CHANGED PLAN FROM AN INDEX SKIPS ROWS THE NEW PLAN PUT BEFORE IT THAT WERE NEVER PROCESSED, and the run then reports completion over a set it never covered. a backfill resumed at any point covers exactly what one pass covers splits the run at EVERY index rather than one, because an off-by-one survives a single hand-picked break point half the time. -->
    • S9.2.h Resolve exact historical published versions and aliases after later updates/deprecation/retirement. Evidence: historical resolution tests. <!-- evidence: every published state still resolves history iterates PUBLISHED_STATES, drives one catalog into each, and requires an answer in every one — the parent's evidence clause quantified over the state vocabulary instead of sampled at the state somebody remembered to test. A lifecycle gate added to the read path fails there, which is exactly what topic-registry.ts did and what no test of it caught. Nothing in resolveHistorical branches on the state: it reads it, puts it on the result and renders it as advisory, and fails only for reasons about the artifact iterates HISTORICAL_RESOLUTION_FAILURES to show all four are statements about the artifact. Resolution goes through the version's OWN alias table: v1 reads gate.minSeverityToBlock directly as medium, v2 reads the same name through an alias onto the renamed key as low, the two answers differ, both are correct, and mutating v2's alias table does not move v1's answer. resolutionHistory shows the whole sequence, because "this document says X and the console says Y" is the question support is actually asked. -->
  • S9.3 Add taxonomy impact preview. Evidence: affected queries, artifacts, filters, exports, and URLs are enumerated before change. <!-- S9.3 marked after writing and reading libs/oshun/workbench-kit/src/ taxonomy-impact.ts (+ spec, 123 tests, 100% of statements, branches, functions and lines, 37 fault kinds) and after FIXING the shipped preview the item is about. THE VEIN IS VISIBLE IN A RETURN VALUE WITHOUT READING THE BODY: libs/oshun/studio-authoring/src/taxonomy-curation/curation.ts accumulated affectedArtifactIds into a Set and affectedRecommendationSlots / affectedSearchResults with += IN THE SAME LOOP, so one population came back under two cardinality rules at once — the union on one row and the multiset sum on the next — and the operation whose entire purpose is the overlap is the merge. THE GREP IS NOT A WORD, IT IS TWO CARDINALITY RULES INSIDE ONE RETURN VALUE; where a preview both unions and sums, one of them is wrong, and the wrong one is whichever number somebody later divides by to size a maintenance window. Its own test is what kept it: it merges tag-a and tag-b, puts art-2 on both, and asserts 3, then 30, then 300 — three is the union, thirty and three hundred are 12+18 and 100+200 — and a test that agrees with itself twice in opposite directions reads as confirmation. Three more faults sat beside it, all fixed in the same commit rather than described: it read request.targetIds and nothing else, so a reparent of a node with four hundred descendants previewed the epicentre rather than the radius; every lookup was map.get(id) ?? 0, so an index that was down, a tag nobody scanned and a tag genuinely used by nothing produced the same zero, AND A ZERO IS THE ONE ANSWER THAT MAKES SOMEBODY CLICK APPLY; and ONTOLOGY_ROLLOUT_STAGES opened with a stage named preview whose transition function never mentioned one — requestId sat on both types, nothing joined them, and preview -> staged -> rolled-out advanced whether or not anybody had computed anything. A stage name is not a gate. The fix carries populations rather than numbers (a NUMBER CANNOT BE DEDUPLICATED, which is exactly why the artifact count was right and the other two were wrong), walks the closure, reports unmeasuredTagIds per dimension, and makes tryAdvanceOntologyRollout refuse preview-missing, preview-for-another-request and preview-incomplete. -->

    • S9.3.a Accept a versioned proposed add/rename/move/merge/split/alias/ deprecate/retire change against exact published base. Evidence: preview input contract. <!-- evidence: acceptProposal binds the proposal to the exact BYTES of the base, not to its name — base-moved is checked before anything is enumerated, because a base that merely matches by id is one that moved while the form was open and every count over it is a count of a population that no longer exists. All eight kinds are declared in CHANGE_SHAPES as DATA, and refuses the wrong arity for every kind walks that table rather than sampling the kind somebody remembered, so a ninth kind added without an arity rule fails instead of being accepted in any shape. refuses every pre-publication state quantifies over the state vocabulary the same way. The conflict rule reads mutatedKeys, not the keys a change NAMES: two moves onto one parent redefine nothing about that parent and their order changes nothing, so they are accepted, while two changes rewriting one declaration are still refused. And a split with no discriminator is deliberately NOT refused — telling the proposer that twelve thousand assignments have nowhere to go is the most useful thing the module does, and refusing would withhold exactly that. -->
    • S9.3.b Enumerate affected stored entity assignments, inherited classifications, translations, rules, gates, permissions, and dependent registry entries. Evidence: data impact rows. <!-- evidence: the seven kinds the clause names are the seven members of DEPENDENT_KINDS that SURFACE_OF_KIND maps to data, asserted as a count rather than by spot-check. The rows are the correction of the shipped defect: counts a record that carries both merged keys ONCE rebuilds curation.ts's own fixture and gets 3 on EVERY dimension, and partitions the affected records proves no id appears under two rows, so a surface total is the size of a union. A translation on both merged keys is a meaning-collision rather than a retarget, because two labels for one concept is a question ("which wording survives"), not a migration. -->
    • S9.3.c Enumerate affected search queries/index docs/facets/suggestions, saved views/queues, filters, reports, and dashboards. Evidence: discovery impact rows. <!-- evidence: nine kinds on the discovery surface, and the surface where WHO HOLDS THE REFERENCE stops being a detail: an index document is one we rewrite and a saved view is one a person holds, so the same rename is retargeted for both and only the second puts its owner on owners with mustBeNotified. A report and a dashboard carry a time series, so a merge, a split or a move gives them series-discontinuity and a history-annotation decision — the chart that silently spans the change is a chart nobody can read. an add widens what its parent contains, which nobody edited is the S9.3 case the shipped preview could not express at all: a saved view reading "everything under X" returns more rows tomorrow and no query changed. -->
    • S9.3.d Enumerate affected bundle/export/import/interchange values, consumer contracts, external IDs, catalog/collection rules, and published artifacts. Evidence: integration impact rows. <!-- evidence: eight kinds, and the one place REACH_OF_KIND had to agree with S9.2 rather than with intuition — published-artifact is outside, because S9.2.d fixes its bytes forever and a preview that listed it among the things we rewrite would be promising to edit an immutable artifact. Export, interchange value, consumer contract and external id are outside for the ordinary reason: they are already in somebody else's system. An alias carries them and a rename without one BREAKS them, and no alias can carry a SPLIT at all, an alias being a function and one name to three targets not being one. -->
    • S9.3.e Enumerate canonical/deep URLs, navigation, redirects, bookmarks, caches, and share links. Evidence: route impact rows. <!-- evidence: seven kinds, and the surface that reads a DIFFERENT mitigation from the other three. an alias does not move a URL — a redirect does: the registry vocabulary and the route table are separate machines, so effectOn reads mitigations.redirectsInstalled for route and change.aliasInstalled everywhere else, and a proposal that installs one and not the other gets rows saying so. A share link already sent and a bookmark in somebody's browser are the references a rename cannot reach, which is why they are broken rather than orphaned — those are different remedies, one a redirect and one a migration. -->
    • S9.3.f Calculate counts, samples, ambiguous mappings, required decisions, migrations/backfills/reindex, cost/time, owners, and rollback. Evidence: impact plan. <!-- evidence: ImpactCount is the spine — exact, at-least, estimated (carrying its observed rows AND its fraction) or unknown (carrying a reason) — and there is exactly one path to a zero in the module: a kind covered by a COMPLETE source with no matches. An unscanned kind contributes an unknown to its surface total, so the ignorance reaches the number an operator reads instead of being absorbed at the leaf. combineAll is a symmetric function of the whole list because folding two at a time is NOT associative once ignorance is in the lattice, and is NOT the repeated application of a binary operator demonstrates it: the same three counts answer at least 14 one way and about 1004 the other. Cost is an ImpactCount for the same reason — a cost derived from a floor is a floor, and that is the number somebody schedules a maintenance window against. A required decision is ONE question about ONE change with every table aggregated behind it, kept out of the handled column because twelve thousand undecidable rows printed beside the retargeted ones read as twelve thousand handled. Rollback is DERIVED from what the change destroys, never asserted: a merge without a captured pre-image is irreversible and says which keys need one, and a reference outside our reach that will be dead keeps it irreversible even with a snapshot. -->
    • S9.3.g Keep preview read-only and revalidate base/proposal/authorization when applying through a separate audited job. Evidence: no-side-effect and stale-preview tests. <!-- evidence: does not touch a single input deep-freezes the catalog, proposal, sources and ownership, runs the preview and compares the JSON before and after — a frozen input mutated under ESM strict mode throws, so the test fails loudly rather than by diff. authorizeApply does not read the plan's conclusions: it RECOMPUTES the plan from the catalog, proposal and inventories as they are now and compares hashes, so a moved base, an edited proposal, a re-read inventory and an altered plan are each refused BY NAME with the hash comparison as the backstop. Staleness is measured from the OLDEST contributing observation rather than from when the plan was assembled, because a plan built a second ago out of a four-hour-old index is four hours old. The strongest rule is preview-insufficient: a preview that could not see part of the system authorizes NOTHING, which is precisely the gate ?? 0 exists to slip past. The job carries the plan hash, the proposal hash, the base content hash, the authorizing actor, the revalidation instant and the audit reference. --> <!-- Registering the entry point was the usual four edits — the KIT_OWNED_ENTRY_POINTS row, package.json exports, the tsconfig.base.json alias, then regenerate — and the generator NAMES the two you forget (unaliased-approved-entry-point, unexported-approved-entry-point) rather than making you find them. 142 entry points, no drift, no violations. --> <!-- NOT closed by this work and deliberately left for a host: the module takes the dependent inventories as input and does not go and read them. That is the correct seam — the stores, indexes, route tables and clocks are the host's — but it means a deployment gets kind-unscanned for every class it has no source for until somebody writes those sources, and completeness: 'insufficient' is the honest answer until they do. -->
  • S9.4 Add localized label and description support without localizing stable IDs. Evidence: locale fallback tests. <!-- S9.4 marked after writing and reading libs/oshun/workbench-kit/src/ localized-labels.ts (+ spec, 179 tests, 100% of statements, branches, functions and lines) and after FIXING the shipped localization workflow the item is about. THE VEIN: A LOCALE-DEPENDENT JUDGEMENT WRITTEN WITH A LOCALE-INDEPENDENT PRIMITIVE ANSWERS THE SAME THING EVERYWHERE, AND THE PLACE IT ANSWERS WRONG IS ALWAYS THE LOCALE THE FEATURE WAS ADDED FOR. THE GREP IS NOT A WORD: find every comparison the localization path makes and ask which of startsWith on a tag, \W, .length, toLowerCase, localeCompare and a hand-written script range it is built out of, because each of those is a decision about language wearing the clothes of a string operation. Three measured instances shipped in libs/oshun/studio-authoring/ src/localization-workflow/, all fixed in this commit rather than described. (1) fuzzyMatchTranslationMemory scored candidates with a Jaccard similarity over text.toLowerCase().split(/\W+/), which is ASCII, so for ja/zh/ko/ar/he/th/hi both token sets came back empty and its own guard if (aTokens.size === 0 && bTokens.size === 0) return 1 reported a PERFECT MATCH BETWEEN TWO UNRELATED SENTENCES — measured at exactly 1 for the Japanese, Arabic and Hebrew pairs and 0 for the English control, so the one language the translation memory worked for was the one nobody translates into; it then returned the entry's stored score instead of the similarity it computed. (2) The RTL check was targetLocale.startsWith('ar-') || startsWith('he-'), and libs/oshun/i18n's OSHUN_LAUNCH_LOCALES ships 'ar' and 'he' WITHOUT A REGION — so the two right-to-left locales the product launches with were the two it could not see, along with fa/ur/ps/sd/ug/dv/yi/ckb; listing tags would not have fixed it, because DIRECTION IS A PROPERTY OF THE SCRIPT (ku is Latin and ckb is Arabic). (3) detectHonorificGap used /[さま|さん|くん|ちゃん|殿|様]/u, which is a CHARACTER CLASS and not an alternation, so it matched any text containing さ, ま or ん — and its own fixture 田中が到着しました contains ま, so the assertion in a test named "flags ja-JP target missing -san / -sama" had only ever been satisfied by the generic info fallback at the bottom of the function. Two more sat beside them: detectStaleTranslations answered {stale:false} when the source hash was ABSENT from the index it was handed, so a verified match and a failed lookup were one answer; and surfaceCustomerStalenessBanners — THE ONLY CUSTOMER-VISIBLE STRING THE LOCALIZATION SYSTEM ITSELF PRODUCES — was a hardcoded English sentence labelled with a locale read from a map named localeByDomain indexed by segmentId and defaulted to 'en-US': a fallback persisted as a translation, committed by the localizer. prioritizeLocaleQueue also ignored the nowUnixSeconds it was handed, so overdue work never rose. -->

    • S9.4.a Keep stable registry/taxonomy ID, code, alias target, parent/path, version, and interchange value locale-independent. Evidence: contract tests. <!-- S9.4.a: LOCALIZED TEXT IS A SIDECAR KEYED BY THE STABLE ID AND NEVER A FIELD ON THE ENTITY — S9.2 fixes a published version's bytes forever, so a French typo correction must not change a content hash, invalidate an alias resolution or need a version whose diff a reviewer reads. LOCALE_INDEPENDENT_FIELDS names the seven; admitStableIdentity refuses anything but lowercase ASCII (NFC and NFD are two rows to ===) and refuses a path that does not END at parentId, because a path and a parent pointer are two spellings of one fact. identityInvariantFaults is the contract test itself: the same entry rendered in eight locales including ar, he and ja-JP must carry one fingerprint, and the spec also feeds it a Turkish-folded code and a collator-reordered alias list to prove it can fail. Ordering of ids is compareCodepoints, never localeCompare, which consults the RUNTIME's default locale. -->
    • S9.4.b Store localized label/short/long description/help/keywords with canonical locale, source locale, status, author/translator, revision, provenance, and updated time. Evidence: localized value schema. <!-- S9.4.b: LocalizedTextValue / LocalizedKeywordsValue carry every field the clause names, and admitLocalizedValue reports every fault at once rather than the first. The refusals worth naming are the ones a SCREEN causes rather than a programmer: a machine provenance must name its engine (an unattributed machine translation cannot be re-run or recalled when the engine is found wrong) and a human one must not; a human translation must name its translator; an authored value is its own source and may not carry one; and a human-translated, machine-translated or post-edited value whose source locale IS its own locale is refused, because that row is one the coverage report counts and the translator never saw. Keywords are a list and never a sentence: they are joined against a query, and a surface that displayed them would be showing a reader the terms somebody guessed they might type. -->
    • S9.4.c Define exact locale, language, configured fallback chain, source/default, and missing behavior without silently persisting a fallback as translation. Evidence: fallback tests. <!-- S9.4.c: parseLocale canonicalises through Intl.getCanonicalLocales, so en-us, EN-US and the 1989 spelling iw are not three rows, and the und refusal is checked on the canonical STRING because new Intl.Locale('und').language is undefined and a check made after construction reports "no primary language subtag" about a tag whose primary language subtag is right there. admitLocaleCatalog refuses chain cycles at configuration time — a loop found at request time is a hang on a page load — and buildFallbackChain walks exact, configured, RFC 4647 truncation, source, default, listing a locale ONCE at its first position. A FALLBACK IS RENDERED AND NEVER PERSISTED: resolved-fallback is a provenance the read path produces and admitLocalizedValue refuses BY NAME, and a target byte-identical to its source is refused unless declared a source copy, because a brand name legitimately does not change and a fallback somebody pressed save on looks exactly the same. -->
    • S9.4.d Distinguish absent, draft, machine-suggested, human-reviewed, approved, stale, rejected, obsolete, and inaccessible values. Evidence: state fixtures. <!-- S9.4.d: nine states, and TWO OF THEM CANNOT BE STORED — absent and inaccessible are answers about a LOOKUP, and a stored row saying "there is no row" is a contradiction that also hides itself from every coverage count; admitLocalizedValue refuses them as resolution-only-status. They stay in one enumeration because a caller must tell them apart: absent means translate this, inaccessible means somebody already did and this actor may not read it, and collapsing them either leaks the existence of embargoed copy or commissions a second translation of something already paid for. DISPLAY_TABLE crosses status with audience and has THREE dispositions, because two are not enough: a stale approved translation beats no label and pretending it is current is what the notice prevents. machine-suggested reaches an end user under no configuration. Staleness is three-valued for the same reason the shipped detector was wrong — unverifiable keeps its stored status and carries a notice, so an index outage does not read as a translation regression. -->
    • S9.4.e Validate script/direction, length/markup/placeholders/links, terminology, uniqueness where required, and unsupported locale. Evidence: validation tests. <!-- S9.4.e: seventeen fault kinds, and the measurement is the point. Length is GRAPHEMES ('a👨‍👩‍👧b'.length is 10 and its grapheme count is 3, so a budget in code units is a different budget in every script). Script conformance is a measured RATIO over letters using Script_Extensions rather than Script — U+30FC, the Japanese prolonged sound mark, is Script=Common, so a check built on Script calls an ordinary ポートレート mixed-script — and UNICODE_SCRIPTS_OF_ISO15924 maps the composite codes CLDR returns, because /\p{Script=Jpan}/u is a SyntaxError and a validator built on them fails to LOAD rather than failing open. Placeholders and link targets are compared as sets in BOTH directions (a dropped {count} and an invented {userName} are different bugs; a translated href is an open redirect that arrived through a spreadsheet), the placeholder pattern uses a LOOKAHEAD so {count, plural, one {#} other {#}} is not missed, and U+202A–U+202E embedding/override controls are refused while the isolates are permitted, because an override makes the label a person reads differ from the one that sorts and exports. Uniqueness is per (scope, field, locale) under one declared fold that uses toLowerCase, never toLocaleLowerCase. -->
    • S9.4.f Search/sort/display by requested locale while retaining stable ID for API/filter/URL/export and exposing fallback/source state. Evidence: end-to-end locale tests. <!-- S9.4.f: resolveLocalizedText NEVER returns a bare string — a function returning string has already discarded whether the caller is looking at a French label, an English one or the key, and the caller renders all three identically; every read answers with the resolved locale, the chain step, the effective status and the notices. Direction comes from the locale the TEXT came from, not the one requested, so an English fallback is not laid out right-to-left. renderEntryList orders by an Intl.Collator for the requested locale and BREAKS TIES ON THE STABLE ID, because a collator calls many pairs equal and a list whose equal-label rows come back differently each call is one where a pagination cursor skips a row; the collation locale is read off the constructed collator's resolvedOptions() rather than from supportedLocalesOf, which hands the tag straight back. An unlabelled row sorts by its id rather than as an empty string at the top of the list. searchLocalizedEntries matches a stable id exactly, outranking every text match and unaffected by locale. And buildInterchangeRow TAKES NO LOCALE PARAMETER AT ALL — the spec asserts it is byte-identical for ar and ja-JP — and carries only approved labels, because an export is read by a system that cannot see a status. -->
  • S9.5 Add editorial calendar entities for tentative and committed slots. Evidence: committed capacity cannot be double-booked. <!-- S9.5 marked after writing and reading libs/oshun/workbench-kit/src/ editorial-calendar.ts (+ spec, 103 tests, 100% of statements, branches, functions and lines) and tools/double-booking-race.mjs (run against the live PostgreSQL 16.14), and after FIXING both shipped calendars. THE VEIN: FIND THE STRING OR THE TYPE NAMING THE FAULT THE MODULE EXISTS TO PREVENT AND THEN FIND THE WRITE THAT COULD HAVE REFUSED IT; WHERE THE FAULT IS A NOUN IN A REPORT AND NOT A REFUSAL IN A COMMAND, THE SYSTEM DETECTS WHAT IT WAS BUILT TO PREVENT. libs/aphrodite/talent/src/calendar/availability-calendar.ts knows exactly what a double booking is — CalendarConflict with conflictType: 'double-booking', found by pairwise slotsOverlap over slots that are ALREADY BOOKED — and the only path into that status was updateSlotStatus(slotId, status), whose entire body was const updated = { ...slot, status }; this.slots.set(slotId, updated). It consulted slotsOverlap, exported from the same file, not at all; it consulted the BookingBlock prep/recovery window its own blockForBooking stored, not at all, because nothing read bookingBlocks but a getter; it consulted the blackout dates not at all. libs/aphrodite/venue-management/src/scheduling/venue-scheduler.ts is the same shape with a sharper edge, because the check it needed is CALLED TWICE, three lines apart, for the two things that are not the booking: createSlot walks every blackout asking timesOverlap, then every maintenance window asking timesOverlap, and never once asks it about another SLOT — so two overlapping AVAILABLE slots at one venue were created and booked without complaint — while venueCapacities, the only number saying how much of the venue there is, is read in exactly one place, matchPerformersToSlots, to SCORE a suggestion. And holdSlot(slotId, notes?) set a status and wrote a note: no holder, no granted time, no expiry, no clock, with bookSlot accepting a HELD slot from any caller — a hold that never expires and that anybody may book over is a label, and its test asserted the label. All three fixed in this commit: bookSlot is now a guarded write in both files (returning a named refusal in the talent engine, throwing in the venue scheduler), updateSlotStatus REFUSES BOOKED and points at it, holds carry a holder and an expiry swept by expireHolds(now), and detectConflicts is documented as what it can now honestly be: the detector for conflicts that ARRIVED from an import or a raised minRestHours, because a detector runs over a snapshot and a refusal has to be part of the write. -->

    • S9.5.a Define calendar/resource/slot IDs, owner/scope, title-safe ref, start/end/time zone, status, priority, target/type, capacity demand, dependency, revision, and audit. Evidence: slot schema. <!-- S9.5.a: CalendarSlot carries all seventeen, and admitSlot reports every fault at once. Two checks earn their place. AN INTERVAL IS HALF-OPEN — [start, end) — so back-to-back slots do not conflict and nobody subtracts a second to work around it, and a zero-length interval is refused because it occupies nothing and therefore passes every exclusion check ever written. And THE OCCUPIED INTERVAL MUST CONTAIN THE BOOKED ONE: setup and teardown extend it outwards, so a swapped pair makes the exclusion narrower than what the calendar displays, which is a double booking that looks correct on every screen. titleSafeRef is an opaque handle and never the title, because the title of an unannounced production is what an embargo exists to protect and a field that SOMETIMES holds it leaks it through the one endpoint nobody thought about. The time zone is checked against Intl.DateTimeFormat rather than a pattern, because Continent/City is a shape and Africa/Atlantis matches it. -->
    • S9.5.b Define proposed/held/tentative/committed/in-progress/completed/ cancelled/expired/conflicted transitions and authorization. Evidence: slot state tests. <!-- S9.5.b: TRANSITION_TABLE and TRANSITION_ROLES are data. commit may be reached from proposed, held and tentative and NOT from conflicted — the world already moved under something, and committing on top of it is how the conflict acquires a second occupant. NO HUMAN ROLE MAY expire: the clock performs it, because an expiry a person performs is a cancellation nobody has to sign, and authorizeTransition refuses both directions of that (clock-only-transition). The refusals are separate kinds because they are separate remedies — cross-tenant is not a permission to grant but a different calendar and short-circuits alone, role-may-not-perform means ask somebody else, not-the-holder means the hold is somebody's, wrong-source-state means the world moved — and a single boolean makes all four "no" and the third unfixable by whoever hit it. Property writes are authorized SEPARATELY, because a scheduler may commit a slot and still not be allowed to change its priority. -->
    • S9.5.c Define tentative hold duration/expiry, renewal, overbooking visibility, ranking, confirmation deadline, and no guaranteed-capacity claim. Evidence: hold clock tests. <!-- S9.5.c: A HOLD IS A LEASE OR IT IS A LABEL — holder, granted instant, expiry, bounded renewals, confirmation deadline that the lease must outlive. Expiry is a SWEEP against a clock and never a read-side check, because a hold that is only expired when somebody looks is holding the interval in every query that does not look; and an OVERDUE confirmation is reported separately and releases nothing, because automatic release means an editor who was in a meeting loses a studio. Renewal is measured from NOW rather than added to the old expiry, so a script cannot hold an interval indefinitely by renewing early, and the bound exists because AN UNBOUNDED RENEWAL IS A COMMITMENT MADE IN INSTALMENTS THAT NOBODY SIGNED. A HOLD NEVER CLAIMS CAPACITY: holds are granted while the resource is fully committed — that IS overbooking visibility — maxConcurrentHolds bounds the QUEUE and not the capacity, and describeHoldPosition returns guaranteesCapacity: false as a LITERAL TYPE so no surface rendering it can describe it as anything else. -->
    • S9.5.d Define committed reservation transaction, capacity unit/ availability, exclusivity/share rules, expected revision, and idempotency. Evidence: reservation contract. <!-- S9.5.d: AN AVAILABILITY CHECK IS NOT A RESERVATION. evaluateReservation decides against a snapshot and, on acceptance, returns a ReservationCommand carrying the exclusion key, both expected revisions and the idempotency key — because a detector can run over a snapshot and a refusal cannot; it has to be part of the write. The exclusion key is tenant and resource and NOT the interval: an interval in the key makes every booking unique and the constraint decorative. CAPACITY IS A QUANTITY WITH A UNIT AND TWO UNITS DO NOT ADD — refused, never converted, the factor being a domain fact this contract does not have. EXCLUSIVE AND SHARED ARE DIFFERENT ARITHMETIC AND NOT DIFFERENT AMOUNTS: modelling exclusivity as "demand equals capacity" makes an exclusive booking silently shareable the moment somebody buys a second encoder. And capacity is drawn down AT AN INSTANT, so the check is the PEAK concurrent demand over the overlap and never the sum over the window, which refuses bookings the resource could have taken. Idempotency is about the request: same key and digest replays, same key and a different digest is refused rather than becoming a second booking. -->
    • S9.5.e Enforce exclusion/capacity constraints under concurrent commit, resize, move, cancel, resource change, and recurrence materialization. Evidence: real-DB double-booking tests. <!-- S9.5.e: libs/oshun/workbench-kit/tools/double-booking-race.mjs, target double-booking-race, run against the live PostgreSQL 16.14 on this box. It FAILS rather than skips with no database. THE CLAIM: four schedulers that all read "free" — the situation evaluateReservation answering ok to four callers at once MEANS — produce exactly ONE committed booking under EXCLUDE USING gist (exclusion_key WITH =, occupied WITH &&) WHERE (state is capacity-occupying), measured. CONTROL 1 drops the constraint, changes nothing else, and the same code leaves FOUR overlapping committed bookings, because A READ DOES NOT CONFLICT WITH ANOTHER TRANSACTION'S UNCOMMITTED WRITE — which is the shipped read-then-write shape and the reason the contract hands back a command rather than an answer. Five more controls: a RESIZE and a MOVE into an occupied interval (refused by the same constraint, accepted without it — the exclusion covers UPDATE, which an insert-time check does not); a CANCEL carrying a stale revision matching zero rows where the unguarded form cancels a booking whose interval it never saw; SHARED capacity, which no exclusion constraint expresses because a sum over concurrent rows is not a pairwise predicate, reaching peak 2 under a FOR UPDATE on the resource row and peak 3 without it; and a RECURRENCE whose fourth occurrence collides, leaving 0 rows materialised in one transaction and 3 orphans when each occurrence autocommits. A rendezvous makes the interleaving decided rather than hoped for — without it control 1 was flaky in the direction that matters. And the deadlock the constraint sometimes throws instead of queueing (40P01) is classified as a REFUSAL and named, because nothing was booked and the caller must retry, which is what the idempotency key on the command is for. -->
    • S9.5.f Reconcile resource/target/project dependencies, conflicts, notifications, publication/release state, and calendar projections. Evidence: integration tests. <!-- S9.5.f: reconcileCalendar reports six kinds — capacity lowered, resource withdrawn, blackout introduced, dependency cancelled, dependency moved past its dependant, publication window missed — each with the transition it IMPLIES and the roles to notify, and it PERFORMS NONE OF THEM. THE WORLD CHANGING UNDER A COMMITMENT PRODUCES conflicted, WHICH STILL OCCUPIES CAPACITY: a resource withdrawn does not cancel the reservations it breaks, because a production silently losing a booking is the failure this whole module is downstream of, and a state that stopped occupying would hand the interval to the next requester while the production that was already booked still believes it has it. Capacity is not compared across mismatched units here either. -->
    • S9.5.g Apply tenant/object/property authorization and protect embargoed/ confidential titles in free-busy views. Evidence: disclosure tests. <!-- S9.5.g: FREE-BUSY CANNOT CARRY A TITLE AND THE TYPE IS WHAT GUARANTEES IT — FreeBusyInterval has no field a title fits in, and a redaction applied at render time is one the next endpoint forgets; the endpoint that forgets it is always the one added for a partner integration. Another TENANT's slots are omitted ENTIRELY rather than shown as opaque busy blocks, because the pattern of when a shared studio is busy is itself information and a competitor reads a shoot schedule out of the shape of it. Within the tenant the disclosure class decides: public shows the ref, internal shows it to participants, schedulers and owners, opaque shows a busy interval and not even the resource to anybody who is not a participant — which is what an embargoed title needs, the ref being the handle a partner looks the title up by. And mergeFreeBusy merges ADJACENT blocks as well as overlapping ones, on purpose: three back-to-back half-hour reviews and one ninety-minute shoot must read the same, and leaving a zero-length gap between them is the same disclosure the merge exists to remove. -->
  • S9.6 Add recurrence, time-zone, daylight-saving, embargo, blackout, and dependency semantics. Evidence: calendar edge-case suite. <!-- S9.6 marked after writing and reading libs/oshun/workbench-kit/src/ calendar-recurrence.ts (+ spec, 106 tests, 100% of statements, branches, functions and lines) and after FIXING all FOUR shipped local-time converters. THE VEIN: A LOCAL TIME IS NOT AN INSTANT, AND A FUNCTION FROM (DATE, TIME, ZONE) TO A SINGLE INSTANT WITH NO FAILURE MODE IS ANSWERING QUESTIONS THAT HAVE NO ANSWER AND QUESTIONS THAT HAVE TWO. THE GREP IS NOT A WORD, IT IS A RETURN TYPE: find every function whose result is one instant and whose input is a wall clock, and ask what it returns for the two hours a year the answer is not one instant. Those are the two hours a year when every alarm in the system is wrong, and they are the hours nobody writes a fixture for. MEASURED against tzdata 2026a. libs/oshun/domain-tara/src/ ritual-scheduling.ts guessed the instant by reading the local time as UTC, formatting the guess back and subtracting the difference TWICE, with no convergence check and no failure mode — and it was wrong in BOTH DIRECTIONS. Asked for 2026-03-08 02:30 in America/New_York, a local time the clock SKIPS, it answered 06:30Z, which reads 01:30 local: an hour early, silently. Asked for 2026-10-04 02:15 in Australia/Lord_Howe, whose spring transition is THIRTY MINUTES, it answered an instant reading 02:45: half an hour late, the opposite direction, for the same reason. Asked for 2026-11-01 01:30 in America/New_York, a local time that happens TWICE, it picked one of the two by accident of its initial guess and the caller could not request the other. The same shape shipped three more times, each subtly different, which is what makes it a vein rather than a bug. libs/oshun/domain-veritas/src/retraction-ux/notification.ts ran the identical loop FOUR times instead of two — more iterations do not make a gap have an answer. apps/veritas/video/src/services/podcast/shows.ts did ONE offset lookup and subtracted, which is wrong exactly on the day the offset changes, which is the day a weekly production schedule is computed across: measured, its answer for 2026-03-08 03:30 in America/New_York read 04:30. And libs/arete/ai-coach/src/weekly-review.ts took toUtcFromLocal(localDateTime, timezoneOffsetMinutes) — AN OFFSET IS NOT A ZONE. A stored offset is wrong for half the year in every zone that observes daylight saving, and the function had no way to find out, because the fact that would have told it was discarded at the boundary; its parameter is now an IANA zone and its result carries which one. All four now use the two-candidate method and report unique, ambiguity-resolved or gap-shifted. -->

    • S9.6.a Define supported recurrence rule/version, start, count/until, intervals, weekdays/month rules, inclusions/exclusions, exception edits, and series/occurrence identity. Evidence: recurrence contract. <!-- S9.6.a: RecurrenceRule carries the version, frequency, interval, start, count-or-until, weekdays, month days, an explicit month-overflow policy, RDATEs and EXDATEs. admitRecurrenceRule REFUSES AN UNBOUNDED SERIES: it is legal in iCalendar and every consumer of one has to invent its own horizon, and no two of them invent the same one — a caller that genuinely wants forever says so with an until somebody had to type. It refuses a series bounded twice, and an unsupported rule VERSION rather than expanding it to the wrong dates. monthOverflow is on the rule because neither answer is right for every series: skip is iCalendar's and it is the one that surprises people — monthly-on-the-31st has SEVEN occurrences in 2026, not twelve — while clamp moves to the last day. AN OCCURRENCE'S IDENTITY IS ITS ORIGINAL START AND NEVER ITS CURRENT ONE: an exception keyed on where the occurrence IS cannot be found after it moves, so the move is applied again on the next expansion, and again, and the rehearsal walks forward an hour a day until somebody notices. The spec asserts the expansion is idempotent under a move, which is what that rule buys. -->
    • S9.6.b Store source local date/time and IANA zone plus resolved instant/ offset and tzdata version without relying on server locale. Evidence: temporal schema tests. <!-- S9.6.b: ZonedDateTime is {localDate, localTime, zone} — the SOURCE, what a person entered, and the only form that survives a tzdata update. ResolvedInstant is derived and carries the offset, the resolution kind and the TZDATA VERSION. That version is an INPUT: this module never reads process.versions.tz, because it is a contract, it runs in a browser, and a version read on the server is a claim about the wrong machine. staleResolutions re-resolves the source against the data in force now and reports the DRIFT — a zone whose rules changed did not change any stored instant, and the difference between the instant we stored and the instant the wall time now means is exactly the meeting somebody misses. partsInZone uses hourCycle: 'h23' rather than hour12: false, because the latter yields hour 24 for midnight in several locales and a 24 silently becomes the next day when fed back into Date.UTC; and the BC era is converted astronomically (1 - year, since there is no year zero) rather than by a sign flip that would be off by one for every date before the common era. -->
    • S9.6.c Define nonexistent/ambiguous DST time choice, repeated/skipped occurrence, zone rule change, leap/day boundary, and all-day behavior. Evidence: DST edge fixtures. <!-- S9.6.c: possibleInstantsFor returns ZERO, ONE or TWO instants, and the size is the answer. The method: take the offsets a day either side, form one candidate per offset, and KEEP ONLY THE CANDIDATES THAT FORMAT BACK to the wall clock asked for — the step a fixed-point loop cannot take, because it has converged to something and has nothing to compare the something against. Fixtures: 0 for the New York hour gap AND for the thirty-minute Lord Howe gap; 2 for the New York repeated hour, for the thirty-minute Lord Howe one, and for 2018-11-04 00:30 in America/Havana, a zone that repeats MIDNIGHT. A NON-EXISTENT CALENDAR DATE IS A DIFFERENT FAULT FROM A NON-EXISTENT LOCAL TIME — 2026-02-29 is a typo and 2026-03-08 02:30 is a policy question, and answering "no instants" to both makes the wrong person responsible. AN ALL-DAY EVENT IS A DATE RANGE AND NEVER A MIDNIGHT INSTANT, because Asia/Beirut skipped midnight on 2023-03-26 and America/Havana has repeated it: the projection walks forward in quarter-hour steps (a transition can be thirty minutes, so an hourly probe steps over the gap it should report walking out of) and reports the wall time it actually used — 01:00 for Beirut. -->
    • S9.6.d Define embargo start/end/zone, allowed prepublication actors, early release, extension/lift, clock source, target/channel/territory, and fail-closed state. Evidence: embargo tests. <!-- S9.6.d: AN EMBARGO FAILS CLOSED. No clock, an unresolvable lift time, a request from a territory nobody classified — all indeterminate with mayPublish: false, because the failure mode of the other choice is publishing an unannounced title and it is not recoverable. The lift resolves with onAmbiguous: 'later' and onNonexistent: 'shift-forward' and neither is arbitrary: on an ambiguous lift the embargo holds through BOTH readings of the repeated hour, and on a nonexistent one it holds until after the gap — AN EMBARGO RESOLVES TOWARDS STAYING IN FORCE. An early release is an explicit, ATTRIBUTED act checked before the clock, so a lift time that resolved oddly can never produce one; an extension moves the effective lift and keeps the original on the record. -->
    • S9.6.e Define resource/tenant/global blackout type, interval/recurrence, capacity effect, priority, override authority, and conflict. Evidence: blackout tests. <!-- S9.6.e: A BLACKOUT SCOPE IS A LATTICE AND THE NARROWER SCOPE CANNOT WIDEN IT. An override is authorised against the scope that IMPOSED the blackout, so a resource owner declaring an override on their own resource cannot book through a platform maintenance window — and the refused override is REPORTED rather than silently ignored, because a booking refused for a reason the person believes they already handled is worse than one refused plainly. Blocking blackouts come back broadest scope first, reduce-capacity takes the LOWEST remaining capacity of every applicable reduction, and a blackout recurs on the same rule engine as everything else. The expansion window is WIDENED to the left by the blackout's duration: an occurrence that starts before the window and runs into it covers the window just as much as one starting inside it, and an expansion over the bare window never returns it — a maintenance shift that began an hour ago, invisible to the booking made now. -->
    • S9.6.f Define finish/start, release/gate, asset/source, resource, approval, and cross-domain dependencies with lag/lead and unavailable state. Evidence: dependency scheduling tests. <!-- S9.6.f: four dependency types, seven subjects, lag and lead as one signed minute count. THE EARLIEST START DERIVED FROM AN UNSCHEDULED PREDECESSOR IS null AND NOT ZERO: a planner that reads unknown as zero takes the maximum over a set containing zero, gets the other constraints, and schedules a chain whose first link nobody has planned — today, in front of everything. An UNAVAILABLE predecessor (a withdrawn asset, a revoked approval) is unschedulable rather than late, because it does not become available by waiting. A violated constraint still reports the bound it violated, so the fix is visible. -->
    • S9.6.g Materialize/query bounded windows deterministically and detect conflicts without infinite expansion or stale tzdata. Evidence: performance/property tests. <!-- S9.6.g: expandSeries takes a window AND a cap on both occurrences and rule STEPS — the step cap separately, because a rule whose every occurrence is excluded still has to stop and a cap counted on emitted occurrences would let it run forever. It reports stoppedBecause and truncated, because A SILENTLY CAPPED LIST READS AS A COMPLETE ONE, and that is how a conflict check passes over occurrences it never saw. THE RULE WALKS THE CALENDAR, NOT THE CLOCK: adding 24 hours to an instant is not tomorrow at the same time, it is the next day at a different time twice a year, so each occurrence's wall clock is resolved on its own and the spec pins the offset changing from -300 to -240 across the transition while the local time stays 09:00. COUNT COUNTS WHAT THE RULE PRODUCED, before the window and before the exclusions — counting after exclusions would REFILL them, so a series somebody shortened by one would end a day later than it did before, the opposite of what they asked for. Determinism is asserted directly (two expansions agree, and the order is the exported compareOccurrences), and staleness is staleResolutions from S9.6.b. -->
  • S9.7 Add drag and non-drag scheduling with server conflict checks. Evidence: concurrent reservation test. <!-- S9.7 marked after writing and reading libs/oshun/workbench-kit/src/ scheduling-board.ts (+ spec, 193 tests, 100% of statements, branches, functions and lines), apps/oshun/web/src/components/workbench/ SchedulingBoard.tsx (+ module CSS, + 25 jsdom tests) and apps/oshun/web/tools/scheduling-board-browser.mjs (54 controls in a real Chromium), and after fixing the shipped scheduler the item is about.

    THE VEIN: FIND THE GESTURE HANDLER THAT COMPUTES THE NEW VALUE, THEN FIND
    THE REQUEST IT MAKES. WHERE THE HANDLER'S OWN CALLBACK IS THE WRITE, THE
    DROP WAS THE DECISION AND THE SERVER WAS NEVER ASKED. The grep is `onDrop`
    / `onDragEnd` / `endDrag`, and the second grep — in the same file — is
    `revision`, `idempot`, `fetch`, `await`. Four hits and zero hits is the
    shape, and `libs/yemaya/ui/src/components/SchedulingEngine/
    DragDropSchedule.tsx` is exactly it: `endDrag` computed new dates and
    called `onTaskMove?.(id, start, end)`, with no request, no expected
    revision, no idempotency key and no question asked of anything that knows
    what else is booked. Its one validator, `validateDrop`, asked whether the
    GESTURE was enabled and then `return true`, so it was a deny-list of four
    cases over an open world.
    
    Reading the rest of that file with the question "what does this gesture
    actually produce" found six more, which is what makes it a vein rather
    than a bug. A RESIZE WAS COMPUTED AS A MOVE, because `calculateNewDates`
    never read `dragData.operation` — it returned the drop date plus the
    original duration for every one of them, so `onTaskResize` was handed a
    task of UNCHANGED duration at a new start, and the one thing a resize
    exists to do was the one thing that path could not do. THE POINTER WAS NOT
    AN INPUT: that function took `_pixelsPerDay` and never read it, its one
    caller passed two arguments, and `calculateDateOffset` — the only function
    that turns pixels into time, and the only one with a snapping test — was
    called by nothing, so how far somebody dragged had no effect on the result
    at all. EVERY SUB-DAY BOOKING BECAME A DAY, because the duration was
    `Math.max(1, Math.ceil(ms / 86400000))` DAYS re-applied through `setDate`:
    a sixty-minute review dragged once was a one-day booking of the room, and
    the test pinned the whole-day case by re-deriving the duration with the
    same `Math.ceil`, so it passed straight over the rounding. `setDate` WALKS
    THE LOCAL CALENDAR, so the same arithmetic yielded 23 or 25 hours across a
    transition — S9.6's vein one surface up, with the duration preserved in
    neither wall clock nor elapsed time because nobody had chosen. THE
    PROVIDER WAS DECORATIVE AND THE DROP WAS A NO-OP: `DraggableTask`,
    `DropZone`, `DragPreview` and `SortableTaskList` each called
    `useDragDrop()` rather than `useDragDropContext()`, so each got its own
    `useState` and every drop zone's `dragData` was permanently null —
    `isValidDropTarget` answered false for every target and rendered EVERY
    zone in the invalid style during a drag, `DragPreview` returned null, and
    `onDrop`'s `setTimeout(() => endDrag(), 0)` ran an `endDrag` closed over
    the pre-drop state that hit its own early return having called
    `onDragEnd(null)`. And `dataTransfer.getData()` RETURNS THE EMPTY STRING
    DURING `dragover` in every browser, so the validation inside `onDragOver`
    never ran and `dropEffect` was never set. There was also no keyboard path
    of any kind.
    
    All eight are fixed in place, and the module is pointed at as canonical
    (`scope:yemaya` may not depend on `scope:oshun`, so it cannot compose it
    — same boundary S9.6 hit for `scope:arete`). The drop is now a PROPOSAL:
    `onProposeMove` returns a decision and the callbacks run only after one is
    accepted; with nothing wired to decide, the drop is REFUSED with
    `NO_DECIDER_REASON` rather than performed. Four of that file's own tests
    asserted the defect and are rewritten.
    
    FIFTEEN DECISIONS, and the three the browser settled are worth the space.
    A GESTURE IS A PROPOSAL, and `committed` is the literal type `false` while
    the local read is `authority: 'local-snapshot'`, because the client's
    vocabulary has no success word — a local evaluation is `refused` or
    `no-local-objection`, never `accepted`. A DURATION IS PRESERVED IN ONE OF
    TWO INVARIANTS AND THE CALLER MUST SAY WHICH: measured on tzdata 2026a, a
    three-hour booking moved onto `2026-03-08` in `America/New_York` ends at
    04:00 under `wall-clock` and 05:00 under `elapsed`, and a system with only
    one is wrong for half its bookings. POSITION IS ARITHMETIC IN THE BOARD'S
    ZONE AND DURATION IS ARITHMETIC IN THE SLOT'S, because the person pointed
    at a grid drawn in one zone and the booking belongs to a resource living
    in another. SNAPPING BELONGS TO THE POINTER PARSER AND NOT TO THE
    PLACEMENT, so a keyboard command means exactly what it says and both paths
    reach one `proposePlacement` through one door — which makes non-drag
    parity a theorem rather than two tests that agree today; and it is
    measured from LOCAL MIDNIGHT, since a fifteen-minute grid computed on
    epoch milliseconds lands off the hour in every half-hour zone. A SNAP CAN
    LAND IN A GAP AND THE ONLY HONEST POLICY FOR A GESTURE IS `reject`. A
    MOVE IS A RE-COMMITMENT and is evaluated from the state it will return to,
    while `conflicted` and `in-progress` stay refused for S9.5's reasons. A
    MOVE BETWEEN RESOURCES IS TWO CAPACITY CHANGES, so both revisions travel.
    A CONFLICT IS PROJECTED THROUGH THE READER'S DISCLOSURE CLASS AND THE
    COUNT IS PART OF THE PROJECTION. AN UNKNOWN OUTCOME IS RESOLVED BY
    REPLAYING THE SAME KEY, NEVER BY READING, and may not be rendered as a
    failure. AND UNDO IS A COMPENSATION AND SOMETIMES A REQUEST.
    
    THE BROWSER FOUND A REAL DEFECT IN THE KIT, and it is the sharpest thing
    here: `mintIdempotencyKey` carried the INTENT KIND, and a 15px drag and
    one ArrowRight land on the identical interval on the identical lane — so
    they minted `k=move-to` and `k=move-by`, two keys, and a double-tap or a
    slow drag followed by a nudge was two bookings of one interval with
    idempotency protecting neither. THE KEY NAMES THE RESULT BEING ASKED FOR
    AND NEVER THE GESTURE THAT ASKED; it is keyed on the placement (booked AND
    occupied, since a changeover change is a different write) plus the
    expected revisions, which is also what makes the NEXT request distinct
    without a counter. Nothing in jsdom could have found it: the two paths
    have to reach the same pixel for the keys to be comparable at all.
    
    Both dist-patch directions verified: silencing
    `gesture-without-a-keyboard-command` and silencing the move arm of
    `move-changed-the-duration` in the BUILT `dist/.../scheduling-board.js`
    each failed a control; restoring returned 54/54.
    
    TRAPS. The gate `packed-consumer-smoke` was RED on twelve problems before
    this change and is red on six after: five declared subpaths did not import
    because the dist `package.json` was stale (a manual `tsc --outDir` bypass
    does not copy it — fixed by copying it), and `Object.freeze` AT MODULE
    SCOPE is work a bundler may drop under `"sideEffects": false`. This module
    writes its constant tables as literals with `readonly` types instead, so
    it is off that list; `calendar-recurrence`, `editorial-calendar`,
    `localized-labels`, `registry-lifecycle`, `registry-merge` and
    `taxonomy-impact` are still on it and several of them build `Set`s and
    `Map`s at module scope, which needs restructuring rather than unwrapping.
    `production-scan` is GREEN again (it was red on a name collision through
    2026-08-10). A v8 two-element `Array.sort` calls its comparator ONCE, so
    the `||` tie-break in ten inline comparators was unreachable — exported as
    `compareFindings` and tested directly, which is also the right shape for a
    surface merging findings from several finders. And the board never moves a
    booking itself, so repeated key presses do NOT accumulate: every press is
    computed from the same unchanged record, which is the claim under test and
    also the reason the browser fixture has to place its New York booking at
    01:45 to reach 02:00 in one step. -->
    
    • S9.7.a Load resource/calendar capacity, tentative/committed slots, dependencies, blackouts, embargo, permissions, version, and current time from server. Evidence: scheduling view tests. <!-- S9.7.a: THE LOAD-BEARING FIELD IS THE ORIGIN OF THE CLAIM, NOT THE NUMBER (S7.3's shape, applied to a clock and to a permission). SchedulingView carries clockOrigin, of which exactly two of four are in AUTHORITATIVE_CLOCK_ORIGINS, and permissions.origin, of which derived-from-payload-shape is a fault — which fields an endpoint chose to project is not a statement about what this person may do. viewClock advances the server's reading by the time the client has observed since receipt and NEVER runs backwards past it, because a machine that resumed from sleep is wrong by hours in the direction that shows a lapsed hold as live. ViewCoverage is per class of thing, because a board drawn over an interval wider than what was fetched renders the remainder as FREE, and free is the one thing an unanswered interval is not. loadedStates makes the same point about the other axis: a board that draws only commitments shows held space as available and sends every requester into a refusal. An absent embargo on a protected slot is a fault rather than a permission (S9.6.d). And gesture-only-capability enforces S9.7.c at the point the permissions ARRIVE rather than at the point somebody notices they cannot Tab to anything. -->

    • S9.7.b Implement pointer/touch drag/resize preview with snapping, time zone/DST, target resource, duration, dependency, and impact; do not commit optimistically. Evidence: drag tests. <!-- S9.7.b: the preview's committed is the literal type false and the impact's authority is the literal 'local-snapshot', so no surface can widen either; the React component renders both into the DOM and the browser tool reads them back. previewGesture is proposePlacement(gestureToIntent(...)) — the snap lives in the parser and the placement never rounds. A resize moves exactly one edge STRUCTURALLY (the other is copied from the original, not recomputed), and gestureAuditFaults is the finder that catches the other arrangement over a recorded before/after — handed what calculateNewDates produces it reports resize-changed-both-edges, move-changed-the-duration and setup-teardown-lost. Setup and teardown keep their original lengths through a move. A snapped edge that names no instant is REFUSED, never shifted, and the browser drives that case for real on 2026-03-08 in New York and spends no request on it. The real drag is page.mouse.down/move/up over a captured pointer, and the interval the command carries is recomputed by the BUILT kit from the pixel distance rather than copied into the assertion. -->

    • S9.7.c Provide keyboard move/resize/target controls and explicit edit dialog with exact date/time/zone/duration/resource as non-drag parity. Evidence: accessible alternatives. <!-- S9.7.c: parity is STRUCTURAL — one proposePlacement, reached by the pointer through gestureToIntent and by the keyboard through an exact intent — and the spec pins it by serializing two placements made under different modalities and comparing them character for character. nonDragParityFaults checks over EDGES rather than names, because a command list that can nudge a start and an end but never move the whole booking has the same gap as one missing a key and comparing labels would not find it. zone is on EDIT_DIALOG_FIELDS because a date and a time without one are not an instant, and a dialog that omits it silently means "the board's zone", which is not where most bookings live. A field the principal may not write renders as a DIFFERENT ELEMENT rather than a disabled input (S7.8), asserted in jsdom by tag name. The browser walks Tab through every booking and every resize handle and out the far side. -->

    • S9.7.d Send expected calendar/resource/slot revisions and idempotency key to an atomic server capacity/conflict check and reservation. Evidence: command contract. <!-- S9.7.d: buildSchedulingCommand composes S9.5's evaluateReservation rather than restating it, and uses the result in ONE direction — to refuse. A local pass is no-local-objection, never accepted, because the snapshot is old by the time it is read and the world can only have got more booked since. A move between lanes carries BOTH resource revisions, because the source frees capacity and the target consumes it and one expected revision covers one of them. The key is minted per REQUEST and contains the expected revisions, which is what makes a retry free and the next request distinct without a counter — and a real browser proved it must not contain the intent kind. attempt is deliberately outside the key; retryCommand advances it and attemptTranscriptFaults audits an observed sequence for a retry that minted a new key, a board that moved before an answer, and a read issued after a timeout. -->

    • S9.7.e Return conflicting slots/resources/dependencies/blackouts/ embargo/capacity with protected safe detail and alternatives where permitted. Evidence: conflict fixtures. <!-- S9.7.e: THE COUNT IS PART OF THE PROJECTION. "Three conflicts" over an opaque calendar measures how busy somebody else is, and two such readings difference back to a single row — so an opaque conflict set is ONE statement with no count and no ref, an internal one is a count with no ref, and only a wholly public one names titleSafeRefs. A conflicting slot the board never loaded is treated as opaque rather than as absent, and one opaque member makes a mixed set opaque. Every ReservationFaultKind maps to a conflict source and a reader-facing noun built here rather than written by a surface. An alternative is an OFFER, so suggestAlternatives searches only lanes the reader may BOOK rather than merely see, inside the window the board fetched, clear of the blackouts it just drew; alternativeFaults audits offers whoever produced them, and the browser drives the see-only case through the built kit as a negative control. -->

    • S9.7.f Handle concurrent accept/reject, stale display, tentative expiry, permission change, partial bulk, offline, retry, and unknown outcome. Evidence: race/fault tests. <!-- S9.7.f: OUTCOME_PRECEDENCE puts a refusal nobody can fix by editing first, then the ones a re-read changes, then the ones an edit changes — S7.2 reached the same order for route outcomes for the same reason, and strongestRefusal is one function so a single item and a batch agree. unknown is the one every scheduler gets wrong: the reservation may exist, so mayReportFailure is false, the next step is replay-the-same-key, and there is no branch anywhere that reads to find out. The browser drives a thrown transport, checks the retry carries the same key with the next attempt number, and counts ZERO network requests in between. bulkSummary is BUILT rather than checked (S7.14): "1 of 3 moved, 2 did not" names both numbers, and an unknown item is spoken as "may or may not have". An offline entry carries its expected revisions and an expiry derived from the server's own staleness budget — not for safety, which the revisions already give, but because a queue that fires two hundred refusals when a laptop wakes up has taught nobody anything. -->

    • S9.7.g Restore focus/selection and announce accepted/refused/conflict result with undo/compensating behavior where supported. Evidence: accessible journey. <!-- S9.7.g: focus is restored in an EFFECT keyed on the submission rather than in the handler that produced it, because the element the id names is re-rendered by the same commit and focus() on a node not yet in the document moves focus to the body and reports success (S7.12); the browser asserts document.activeElement is the live booking and not the body, after a dismissal and after an answer. restorationFaults takes the LIVE element and slot ids as arguments so the same finder works over a browser transcript and over component state, and it does not blame a refusal for dropping a slot that left the board. THE ANNOUNCEMENT NAMES THE RESULTING TIME rather than saying "moved" — a reader who cannot see the board learns nothing from "moved", and that is the whole reason the live region exists; a refusal is assertive because the reader is about to act on a board that did not change. UNDO IS A COMPENSATION AND SOMETIMES A REQUEST: giving something back always works and taking something back does not, so undoing a move is re-request-the-original with guaranteed: false and a label that says somebody may have taken it, rather than an inverse the system cannot perform. Work that has started is not undoable at all. -->

      <!-- S9.7 FRONTIER NOTE. The next unchecked parent is S9.8 (capacity resources and reservation units as domain plugins), whose evidence clause is GPU, reviewer, DCC host, editorial and distribution examples sharing one reservation engine without shared code. -->

  • S9.8 Add capacity resources and reservation units as domain plugins. Evidence: GPU, reviewer, DCC host, editorial, and distribution examples. <!-- S9.8: shipped as libs/oshun/workbench-kit/src/capacity-plugins.ts (+ spec, 183 tests, 100% of statements/branches/functions/lines) and libs/oshun/workbench-kit/tools/co-allocation-race.mjs (nx target co-allocation-race, real PostgreSQL). 147 kit entry points.

    THE VEIN, STATED SO IT TRANSFERS: FIND THE NUMBER THE REQUEST DECLARES AND
    THE LIMIT THE RESOURCE DECLARES, THEN FIND THE SUBTRACTION BETWEEN THEM.
    WHERE THE LIMIT IS APPLIED TO THE RECORD INSTEAD OF TO THE DECISION -- A
    CLAMP, A SCORE, A REPORT -- THE RESOURCE IS OVER-COMMITTED IN EXACTLY THAT
    UNIT, AND THE DETECTOR THAT WOULD HAVE CAUGHT IT IS OFTEN UNREACHABLE BY
    CONSTRUCTION. The grep is mechanical: take the requirement type field by
    field, and for each field find the member of the LOAD type that spends it.
    A field with no member is a unit the scheduler cannot run out of.
    
    `libs/metis/multimedia/src/video/gpu-worker-pool.ts` is the sharpest
    instance, because the clamp and the detector are eleven lines apart. A job
    carries `estimatedMemoryMB` and the pool carries `memoryLimitMB`;
    `processNext` assigned any job to any idle worker without asking about
    memory at all, and then stored
    `Math.min(this.config.memoryLimitMB, Math.max(0, Math.round(value)))`. The
    limit was applied to the number being STORED, so `healthCheck`'s
    `worker.memoryUsageMB > this.config.memoryLimitMB` -- the only comparison
    of the two in the file -- COULD NEVER BE TRUE for any value that reached
    it. A 1024 MB composition on a 256 MB pool was admitted, recorded as 256
    and reported healthy, and the file's own test pinned it:
    `it('clamps an over-budget estimate to the memory limit')`. Four more
    followed from one root cause, that the pool never retained the reservation
    it granted: `processNext` returned the in-flight job and kept nothing, so
    `stopWorker` searched the queue and the completed list, found the running
    job in neither and DROPPED it; `completeJob` and `restartWorker` rebuilt a
    replacement with `compositionId: ''`, a fresh `createdAt` and a priority of
    1, so what came back was not the job that ran and a restarted job lost its
    place in the queue; and `getJob` could not find a rendering job at all.
    
    `libs/oshun/render-farm` is the same shape in the field a scheduler is
    most likely to declare and least likely to spend.
    `RenderResourceRequirements` carries `licenses?: Record<string, number>`
    and `RenderWorkerLoad` has no member for seats, so `capabilitiesSatisfy`
    asked about the renderer, the cores, the memory, the storage, the features
    and the GPU and never about a licence: a task declaring `{ arnold: 1 }`
    was dispatched whatever the site's licence server had left. The ledger
    exists one file over, and `advanced.ts::selectPlacement` consulted it with
    `licenses.available(feature) < seats` and never called `acquire` -- S9.5's
    vein one layer up, an availability check that is not a reservation. Its
    `gpuCompatible` never subtracted `worker.load.gpuCount` although the
    sibling `gpuSatisfies` did, so placement put work on machines whose cards
    were all busy. Its ranking added a region flag worth 100, the card's VRAM
    in gibibytes and a task count times twenty into one `score`, so a bigger
    card outvoted the requester's own region and did so even when the request
    needed no GPU. `evaluateLimit` projected `concurrentRunningTasks` as the
    CURRENT value while `maxQueuedJobs` two branches below projected
    `current + 1`, so the one limit that exists to bound concurrency could
    never be exceeded by the submission it was evaluating. And a fractional
    GPU count reached `normalizePositiveInteger`, where `Math.trunc(0.5)` is 0
    and the guard threw "Expected a positive integer" from whichever load
    calculation ran first.
    
    All eleven are fixed in place with tests, and every fix was
    mutation-tested: reverting the admission check, the licence acquisition,
    the concurrency projection, the rider draw-down and the staleness
    degradation each turned a test red.
    
    THE DESIGN, IN THIRTEEN DECISIONS. A RESOURCE HAS SEVERAL CAPACITY
    DIMENSIONS AND EXACTLY ONE OF THEM BEARS THE BOOKING: the bearing
    dimension is the calendar row and carries the exclusion key, the expected
    revisions, the hold and the idempotency key, and every other dimension is
    a RIDER on that row -- VRAM, licence seats, focus minutes, egress. Riders
    are refusals rather than warnings, and they are exactly what the shipped
    schedulers forget, because the bearing one is the only dimension with a
    state machine to remind them. Sharing is a property of the ROW and never
    of a rider, which a synthetic sixth plugin found during the build.
    A DIMENSION THE RESOURCE DOES NOT DECLARE IS A REFUSAL AND NEVER A ZERO.
    THE ID SAYS WHICH POOL, THE UNIT SAYS WHAT THE NUMBER MEANS, AND BOTH
    TRAVEL, because the factor between gib and gb is 1.073741824 and the
    factor between a claim and a focus-minute does not exist. A LEVEL IS DRAWN
    DOWN AT AN INSTANT AND A RATE IS DRAWN DOWN OVER A WINDOW: levels reuse
    S9.5's peak-concurrent sweep, rates use the exact maximum of a sliding
    integral over a piecewise-constant density, and a rate does not convert
    across periods either, since a daily budget can be spent before lunch.
    A FRACTION IS EXPRESSIBLE OR IT IS REFUSED AND NEVER ROUNDED. HEALTH IS
    OBSERVED RATHER THAN DECLARED AND A DEAD RESOURCE LOOKS EXACTLY LIKE AN
    IDLE ONE. AN INTERVAL THE SNAPSHOT DOES NOT COVER IS UNKNOWN, NOT FREE. A
    LIFECYCLE STATE SAYS WHAT MAY BE BOOKED NEXT AND NEVER WHAT IS ALREADY
    BOOKED. A QUOTA IS CHARGED TO THE REQUESTER AND CAPACITY TO THE RESOURCE,
    so they are different refusals with different remedies, and every fault
    carries a remediation class because a refusal without one is a dead end
    whatever its prose says. MONEY IS AN INTEGER IN MINOR UNITS AND A CURRENCY.
    PREEMPTION IS A PLAN AND NOT AN ACT. A RESERVATION THAT SPANS RESOURCES IS
    ALL-OR-NOTHING AND ORDERED. AND THE ENGINE HAS NO BRANCH ON A PLUGIN. -->
    
    • S9.8.a Define plugin-owned resource type/ID/version, tenant/scope, capabilities/constraints, calendar/availability, capacity units, location/zone, cost, health, and lifecycle. Evidence: resource contract. <!-- S9.8.a: PluginResource carries all ten, and the two that decide everything else are the ones a declaration is most likely to get wrong. admitPluginResource refuses a resource with NO bearing dimension (no calendar row, so nothing carries its exclusion key or its revision) and one with TWO (two rows for one booking, and the second is the one nobody cancels). Capabilities are TAGGED values rather than strings, because the three comparisons a scheduler makes are three different orderings: 550.54.14 and 550.9 order one way as versions, another as text and a third as floats, and only one of those admits a host that can run the job. A number carries its unit and gib-against-gb is attribute-unit-mismatch rather than a 7.4 per cent error. An attribute the resource does not DECLARE is unsatisfied, never compatible, because absence is the state a fleet acquires the moment somebody adds a field to the inventory. observedHealth degrades a stale healthy to unknown against the clock, and a reading from the future is unknown too -- a host that stops answering has no running jobs, no memory in use and no errors, so every metric says idle. LIFECYCLE_ADMITS_NEW_WORK is a total table, and draining is the row worth stating twice: it accepts nothing new and keeps everything it has, and folding it into retired cancels the work draining exists to let finish. -->

    • S9.8.b Define reservation request units, amount/rate, interval, exclusivity/share, setup/teardown, priority, preemption, quota, cost, and result. Evidence: reservation contract. <!-- S9.8.b: PluginReservationRequest names all ten and the interval it occupies is the booked one extended OUTWARDS by setup and teardown (S9.5 decision 7), which is also what cost is charged over: a changeover is time the resource cannot sell to anybody else, and a farm billing only the booked interval gives it away. peakWindowedDemand is the rate arithmetic and it is exact rather than sampled -- the integral of a piecewise-constant density over a sliding window is piecewise LINEAR in the window's start, with breakpoints only at the density's own and at those shifted back by one period, so the maximum is attained at one of those or at an end of the range and the function enumerates exactly that set. A quota is the same arithmetic charged to the REQUESTER instead of the resource, which is why it is a different fault with a different remedy: waiting fixes a full resource and never fixes an exhausted allocation. estimateReservationCost returns one total PER CURRENCY and there is no field a cross-currency sum could be put in, and every line is an integer of minor units, because money held as a float acquires a rounding error per line and reconciles against nothing. -->

    • S9.8.c Define availability/capacity calculation, tentative hold, commitment, release, resize, move, expiry, failure, and reconciliation interfaces. Evidence: plugin lifecycle tests. <!-- S9.8.c: calculateAvailability answers open, closed or unknown rather than a number, because the failure it exists to stop is a surface drawing free space over an interval nobody answered for: a window past the snapshot's coverage is unknown, a stale health reading is unknown, and only open may be drawn as space. planRelease returns capacity from NOW rather than from the booked start, since a job cancelled halfway through has already consumed the first half, and a release at or after the end frees nothing and says so instead of reporting a zero-length interval that every overlap check passes. planResize moves exactly one edge STRUCTURALLY -- the other is copied -- and produces a REQUEST rather than a mutation, because on a rate dimension the same amount in a shorter interval is a HIGHER draw, so even a shrink has to be asked. planMove carries both revisions and refuses a target that cannot account for a rider rather than dropping it. planFailure splits three facts everybody folds together: a run that failed is over and its capacity returns; a resource that went unreachable has said nothing about the work, so the capacity stays claimed and somebody reconciles; and an unknown outcome may not be reported as a failure at all, because the reservation may exist and the next step is to replay the same key rather than to read and guess (S9.7.f). reconcileResource reports eight kinds of drift and performs none of them. -->

    • S9.8.d Implement GPU example with model/VRAM/compute/region/driver, fractional/exclusive capacity, quota, health, and cost. Evidence: GPU fixtures. <!-- S9.8.d: three pools, because a card is three things at once -- a DEVICE somebody may hold exclusively, seven COMPUTE SLICES and eighty gibibytes of VRAM. A scheduler with only the first oversubscribes memory the moment two jobs share a card, which is what RenderWorkerLoad (a gpuCount and nothing else) does. FRACTIONAL GPU IS NOT AN ARBITRARY FRACTION: MIG_PROFILES is the A100-80GB table, seven compute slices and eight ten-gibibyte memory slices, and the spec checks the arithmetic invariant rather than the names (no profile's slices times its instance count exceeds 7, none's memory exceeds 80). Asking for 0.5 of a card is not a smaller request, it is one the device cannot express, and a MIG instance takes a WHOLE device from the device pool plus part of that card's slices and memory -- charging it a fraction of a device would let two instances of two profiles be placed as though they were on different cards. CUDA_MINIMUM_DRIVER makes "newer driver" a comparison that can be made, and the spec pins that 550.54.14 > 550.9 fails under both text and float comparison. -->

    • S9.8.e Implement reviewer example with qualification/role/working hours/ load/COI, count/effort units, delegation, and privacy. Evidence: reviewer fixtures. <!-- S9.8.e: TWO pools, because a queue that counts only claims treats a four-hour appeal and a five-minute triage as the same load -- libs/contracts/concordia's maxConcurrentClaims is the first pool and there is no second one. reviewer.effort-minute is a RATE over a day, because attention is spent and does not come back when the item closes, and the spec drives the case a claim count cannot see: a second four-hour appeal that fits three concurrent claims and does not fit a 300-minute day. qualificationsAt is a function of the CLOCK, because an expired certification is not a qualification and a resource whose attributes were built once at registration is a reviewer who stays qualified forever. withinWorkingHours walks the interval in the REVIEWER's zone rather than sampling its ends, and the spec drives a 16:30-18:30 booking against a window that closes at 17:00. A LAPSED CONFLICT IS STILL A CONFLICT INSIDE THE COOLING-OFF PERIOD, so filtering on endedAtMs === null clears somebody who left the company last week. DELEGATION MOVES THE CLAIM AND NEVER THE QUALIFICATION: the delegate's own qualifications and own conflicts are re-checked, because the alternative is a decision signed by somebody who was not allowed to make it, discovered at audit. And WHY a reviewer is unavailable is projected through the reader's class: "at capacity" is ordinary, "conflict of interest" says this person has a financial or family relationship with the party under review, and that is employment data wearing a scheduling word. -->

    • S9.8.f Implement DCC host example with application/version/plugin/ license/GPU/session/exclusive slot, liveness, and maintenance. Evidence: DCC fixtures. <!-- S9.8.f: THE LICENCE IS NOT A PROPERTY OF THE HOST. A floating seat lives on the licence server, is shared by every host in the building, and is therefore its own resource with its own revision -- modelling it as a field on the worker is exactly what lets two hosts hand out one seat, and it is why coAllocate exists. dcc.gpu-device is deliberately not gpu.device: the same card counted in two pools is a card that can be handed out twice. The session is exclusive and the host offers one, because an artist opening Maya on a box somebody else is using is not a capacity question. sessionLiveness never releases on silence, in any of its three verdicts: an artist on a flaky VPN, a host whose agent crashed while Maya kept running, and a machine that was unplugged look identical from here, and two of the three still have somebody's unsaved work in them. Maintenance is a blackout with a window rather than a deletion, and a window declared over existing commitments is a reconciliation finding. -->

    • S9.8.g Implement editorial and distribution examples with slot/ throughput/channel/territory/embargo/deadline/blackout semantics. Evidence: domain fixtures. <!-- S9.8.g: the editorial desk separates what it can HOLD from what it can PUT OUT -- six pieces in production and four a day through subbing and legal are different numbers, and a model with only the first schedules a week's output for Tuesday. A DEADLINE IS A PROPERTY OF THE REQUEST AND NOT A REFUSAL BY THE RESOURCE: the desk has not run out of anything, so reporting it as a capacity fault sends the requester to look for another desk, and deadlineFinding is separate from the decision with a change-the-request remedy. Distribution needs three pools, and egress shows why granularity matters -- 0.1 Mbps steps, and an offer that is not a whole number of them is capacity the channel advertises and cannot hand out. distributionGate composes S9.6's decideEmbargo rather than restating it, so the fail-closed behaviour survives: an unclassified territory and a lift time that will not resolve are both indeterminate, and indeterminate is never a yes. The blackout is checked SECOND and independently, because a released title still may not go out during a territory blackout and folding the two answers together loses which one refused. -->

    • S9.8.h Prove plugins share reservation engine without shared code assuming one unit or silently converting incompatible capacities. Evidence: cross-resource isolation tests. <!-- S9.8.h: proved four ways. STRUCTURALLY, every filter in the module is by (resourceId, dimensionId), and the spec asserts that a GPU decision is byte-identical whether or not a reviewer's effort and a desk's throughput are sitting in the same draw array. pluginRegistryFaults catches the three ways two plugins can collide, and the subtlest is unit-used-with-two-meanings: two plugins may both count in minute, but one metering it as a rate over a day and another as a level makes the same word mean "how many at once" in one place and "how many per day" in another, and the first person to write a shared report will add them. templateConformanceFaults keeps a resource from renaming its own units out of comparability with the rest of its type. And the proof that the ENGINE has no branch on a plugin is a SIXTH plugin that exists only in the spec -- a ceramics kiln with chambers and fuel therms -- decided end to end by the same functions, and refused by them for the same reasons and in its own unit. The fifth way is the one no in-process suite can do: tools/co-allocation-race.mjs races four contenders for one DCC session AND one licence seat against a real PostgreSQL 16.14, and its five controls each reproduce a shipped fault before the claim is measured against them -- THE RIDER IS NOT THE ROW (with only the bearing exclusion, four sessions commit against a pool of one seat and the seat pool is never touched), A CLAMP IS NOT A REFUSAL (a demand of 8 against an offer of 1 is admitted and stored as exactly 1, so the health check that compares them can never fire), A PARTIAL ACQUISITION IS WORSE THAN A REFUSAL (per-leg transactions leave three contenders holding hosts they cannot use), ORDER DECIDES DEADLOCK, and UNITS DO NOT ADD (a seat-hour draw against a seat pool is refused by the storage layer, because a schema that stores a bare number is one that adds them). Both dist-patch directions verified: moving the DCC bearing flag in the BUILT dist/.../capacity-plugins.js turned it red on two findings, and restoring it returned it to green. -->

      <!-- S9.9 FRONTIER NOTE. The next unchecked parent is S9.10 (prohibit direct production-catalog writes), whose evidence clause is that all changes pass through reviewed versioned bundles -- which is S9.9's bundle made mandatory, so the first question is which write paths exist that do not go through one. -->

  • S9.9 Add collection, relation, feature, rank, eligibility, and audience curation primitives. Evidence: deterministic curation bundle. <!-- S9.9: shipped as libs/oshun/workbench-kit/src/curation-bundles.ts (+ spec, 192 tests, 100% of statements/branches/functions) and libs/oshun/workbench-kit/tools/curation-determinism.mjs (nx target curation-determinism, child processes under five locale/zone combinations). 148 kit entry points.

    THE VEIN, STATED SO IT TRANSFERS: A CURATED ORDER IS A DECISION ABOUT
    MEMBERSHIP AND NOT DECORATION. WHEREVER A SORT'S COMPARATOR CAN RETURN
    ZERO FOR TWO DISTINCT ITEMS AND A LIMIT IS APPLIED DOWNSTREAM OF IT, THE
    SET OF THINGS PEOPLE SEE IS DECIDED BY THE ORDER THE ROWS HAPPENED TO
    ARRIVE IN. The grep is mechanical and it is two greps. First: every
    `.sort(` whose comparator has no total tie break, then look downstream for
    `.slice(`, `limit`, `[0]` or `break`. Second: every ordering primitive
    that reads ambient host state rather than its input -- `localeCompare`
    with no locale argument, iteration over a `Map` or `Set`, and `+=` over
    floats.
    
    `libs/oshun/customer-curation/src/collections/smart-collections.ts` is the
    sharpest instance, because it is LIVE -- `apps/oshun/bff` depends on it --
    and because the failure is not a reordering. `evaluateSmartCollection`
    sorts `b.createdAtUnixSeconds - a.createdAtUnixSeconds` and then
    `sorted.slice(0, rules.limit)`, so three artifacts created in the same
    second with a limit of two return `[a, b]` from one input permutation and
    `[c, b]` from another. Not a different order -- a different SET, with no
    way for anybody to say why `c` is not on the shelf. Measured against the
    shipped function before it was touched. Its `alphabetical` arm is
    `a.title.localeCompare(b.title)`, which orders `["ärlig","i","I","z"]`
    under en-US, `["ärlig","I","i","z"]` under tr-TR and
    `["i","I","z","ärlig"]` under sv-SE -- three orders from one input, and
    `LC_ALL` is not a field in any bundle. And `smartRulesEqual`, whose stated
    job is "to dedupe smart collections within an owner's library and to
    short-circuit re-evaluation when nothing changed", canonicalises with
    `JSON.stringify(pred)`, which preserves key INSERTION order: two identical
    rule sets built by two code paths compared unequal, so the short circuit
    never fired and the re-evaluation ran every time.
    
    `libs/oshun/search-discovery` has the same shape three more times.
    `candidates/generators.ts::findObjectsMatchingInterests` walks the catalog
    as `for (const [objectId, object] of objectsById)` and `break`s at twenty,
    so which twenty objects a cohort's interests match is decided by the order
    the catalog `Map` was built in. `ranker/features.ts::scoreRankerCandidate`
    accumulates `total += value * weight` over `input.features`, a `Map`, so
    the same eight readings inserted in reverse produce composites differing
    by 2.22e-16 -- which is not nothing, because it is the sign of the
    comparator that ranks them. The same function reads
    `weights.weights.get(featureClass) ?? 0` per PRESENT feature, so a feature
    class nobody measured contributes exactly what a feature class measured at
    zero contributes: both score 0.435 on the same candidate, and a collector
    that has gone down is indistinguishable from a signal that came back
    empty. `applyCoherenceConstraints` then applies greedy `maxPerSource` and
    `maxPerDomain` caps -- and when `enforceTaraCenteredHome` is false it
    applies them to `input.ranked` WITHOUT SORTING IT AT ALL, so the caps keep
    whichever items came first; the sorted branch had no tie break either.
    Every one of its ten exclusion paths was a bare `continue`, so an item
    dropped for a takedown, for rights, for consent or for residency left the
    same trace as an item that was never a candidate, which is none.
    `enforceAntiMonoculture` stops at one remaining item and returns a set
    that still violates the ratio it was asked for, reported as success.
    
    Three more from asking what each input MEANS rather than how it is
    ordered. `catalog/object-classes.ts::isObjectFresh` is
    `nowUnixSeconds - reviewedAt <= f.windowSeconds`, a ONE-SIDED window that
    every negative age satisfies, so an object published in 2033 is fresh in
    2023 and gets fresher the further ahead it is dated.
    `aggregation/decay.ts::decayWeight` clamps age with
    `Math.max(0, now - capturedAt)`, so a signal timestamped ten years in the
    future keeps weight 1.0 and never decays -- the one field a client
    supplies is the field that turns the decay off. And `timeOfDayBucket`
    computes `Math.floor((unixSeconds % 86400) / 3600)`, which is the UTC
    hour: eight in the evening in Los Angeles is 04:00Z and buckets as
    `night`, under a dimension that exists to say when in somebody's day a
    thing happened and beneath a reason string that reads "because it's
    evening".
    
    `objectIsEntitled` is the only one that is a leak rather than a wobble.
    `creator-only` was decided as `viewer.entitlement === 'creator-only' ||
    viewer.entitlement === 'admin-only'` -- a comparison of TIERS -- so any
    viewer holding the creator tier was entitled to every other creator's
    private objects. It was written that way because it could not be written
    any other way: `SearchableObject` carried `personaBinding` and `tenantId`
    and had no field naming the creator at all, so the rule named a
    RELATIONSHIP the schema could not express and a tier comparison was what
    was left. A TIER IS NOT AN IDENTITY, and the fix is a field.
    
    All eleven are fixed in place with tests, and every fix was
    mutation-tested: reverting the tie break, the code-point sort, the
    canonical JSON, the canonical feature order, the coherence ordering, the
    two-sided freshness window, the identity comparison, the future-capture
    refusal and the zone-aware bucket each turned a test red.
    
    FOURTEEN DECISIONS. A COLLECTION IS ENUMERATED OR DERIVED AND THE TWO HAVE
    DIFFERENT IDENTITY, because a derived one's members are a FUNCTION of a
    corpus and a version naming only the rules denotes a different member set
    every time the corpus moves. AN ORDER KEY LIST IS PART OF THE MEMBERSHIP
    DECISION AND MUST BE TOTAL BEFORE A LIMIT MAY BE APPLIED, tested the way
    `indexes.ts::sortIsTotal` tests a database query: the sorted columns plus
    the fixed ones must contain a declared unique key. TEXT IS COMPARED BY
    CODE POINT, OR BY A COLLATION THE BUNDLE NAMES AND HASHES -- and the quiet
    offender is JavaScript's own `<`, which orders UTF-16 CODE UNITS and so
    puts an emoji-led title before U+FFFD where UTF-8 byte order and
    `COLLATE "C"` put it after, which is a page ordered in the process
    disagreeing with the same page ordered in the index. A SIGNAL CARRIES ITS
    UNIT, SCALE, SOURCE, SOURCE VERSION AND OBSERVATION TIME, AND POLICY NEVER
    TRAVELS WITH IT: the signal type has no weight field and the term type has
    no value field, because a weight applied at capture is a number nobody can
    re-weight and nobody can audit. AN ABSENT SIGNAL IS NOT A ZERO -- a term
    refuses the item or takes a DECLARED default that is written into its
    reasons, and there is no third behaviour, because contributing nothing
    silently is a demotion applied to whichever items a broken collector
    happened to cover. THE COMPOSITE IS EXACT INTEGER ARITHMETIC SUMMED IN THE
    POLICY'S DECLARED TERM ORDER, in basis points over readings held as
    integers at a declared decimal scale, and a policy whose worst case could
    leave the exact range is refused. EVERY EXCLUSION CARRIES ITS REASONS, ALL
    OF THEM, because "why isn't my thing on the shelf" is the only question
    anybody ever asks and fixing the first of four rejections and being
    refused again is how a queue stops being used. AN INPUT THE SNAPSHOT DOES
    NOT COVER IS INDETERMINATE, AND INDETERMINATE IS NEVER ELIGIBLE. A TIER IS
    NOT AN IDENTITY. A SEGMENT BELOW ITS COHORT FLOOR IS REFUSED AT EVALUATION
    AND NOT SUPPRESSED AT REPORT, because a suppressed count still answers the
    question when the person asking chose the predicate -- and DIFFERENCING
    makes the floor a property of every PAIR rather than of one segment, since
    a permitted segment of 40 and a permitted nested one of 38 identify two
    people by subtraction. A SEGMENT MAY NOT BE DEFINED ON A SPECIAL-CATEGORY
    ATTRIBUTE NOR ON A DECLARED PROXY FOR ONE, because a wellness affinity
    inferred from which health articles somebody read is health data with a
    different name and a derivation does not launder the category. A RELATION
    IS DIRECTIONAL AND ITS CONSTRAINTS BELONG TO ITS TYPE, so acyclicity is
    walked per type -- a mixed walk finds cycles that are cycles in neither
    graph. A FEATURE PLACEMENT IS A CLAIM ON A SLOT OVER AN INTERVAL, two
    overlapping exclusive claims are decided by declared priority and a
    priority tie is a fault rather than a coin toss, and a slot that resolves
    to nothing says WHICH of the four reasons it was, because an empty shelf
    and a shelf whose only candidate was refused for rights look identical to
    a reader and must not look identical to an operator. AND THE HASH COVERS
    THE INPUTS, THE POLICY AND THE AMBIENT DECLARATION RATHER THAN THE OUTPUT,
    because an output hash agrees exactly when two builds made the same
    mistake. -->
    
    • S9.9.a Define versioned collection identity/type/owner/scope/title refs/ lifecycle and ordered or rule-driven membership. Evidence: collection schema. <!-- S9.9.a: CurationCollection carries all seven, and the field that decides the rest is membership. AN ENUMERATED COLLECTION'S MEMBERS ARE ITS CONTENT and change only by an edit that moves a revision; A DERIVED COLLECTION'S MEMBERS ARE A FUNCTION OF A CORPUS AT AN INSTANT, so corpusVersion is part of its identity and collectionFaults refuses a derived membership without one -- a rule collection versioned by its rules alone is a bundle whose members change underneath a frozen id, which is exactly what a reader who pins a version is trying to prevent. The title is a REF into the localized-label registry rather than a string, because S9.4 already decided that a label is localized and an id is not, and a collection carrying a raw title is one whose name cannot be translated without changing the collection. The lifecycle is asserted EQUAL to registry-lifecycle.ts's six states in the spec rather than imported, so the two cannot drift; SERVABLE_LIFECYCLE_STATES is deliberately a strict subset of that module's PUBLISHED_STATES, which also contains retired, because a retired ENTRY must keep resolving for the history that named it and a retired SHELF must stop being drawn. A duplicate ordinal is a fault rather than a tolerance, because a tie in an enumerated collection is the arrival-order membership this module exists to refuse, and a blank tenant is separated from a null one because null says unscoped and "" says somebody built an id from an empty field. -->

    • S9.9.b Define typed directional relation identity, source/target, role/strength/order, validity period, provenance, and constraints. Evidence: relation schema. <!-- S9.9.b: CurationRelation names all seven and the constraints are properties of the TYPE rather than of the graph. Acyclicity is walked per relation type, and the spec drives the case that makes it matter: an a -> b prerequisite and a b -> a related-to form a cycle in the union and in neither graph, so a single walk over all edges reports a cycle nobody has, and a system that then refuses the edit refuses a correct one. Only edges IN FORCE at the instant are walked, because a prerequisite that lapsed last year does not make a cycle today, and the same restriction governs functionality and out-degree. An empty half-open validity interval is its own fault, because a relation valid from 200 to 200 holds at no instant and is visible in the store and in no result -- the shape somebody stares at for an afternoon. STRENGTH IS A READING AND MAY BE unjudged, for the same reason a rank signal may be absent: an editor who has not judged how strong a link is has not judged it to be zero. The cycle walk is iterative rather than recursive (a citation graph is as deep as its corpus, and a stack overflow in a validator reads as a crash rather than as a finding) and visits nodes in code-point order, so a graph with several cycles reports the same one on every run. -->

    • S9.9.c Define feature placement/slot, target, creative/metadata ref, start/end, priority, exclusivity, fallback, and target/channel context. Evidence: feature schema. <!-- S9.9.c: a placement is a CLAIM ON A SLOT OVER AN INTERVAL, which is S9.5's reservation composed rather than restated. Two exclusive claims that overlap are a conflict decided by declared priority, and an equal priority is a separate fault rather than a coin toss -- the shipped alternative is whichever row the store returned first. The overlap check is per CONTEXT-INTERSECTING pair, so one placement for GB and one for GH on the same slot at the same time are not reported: they can never be shown to the same reader, and a conflict report full of those stops being read. peakOverlap sweeps endpoints rather than sampling, and sorts ends BEFORE starts at an equal instant, because an interval is half-open and a sweep that opened first would report a phantom overlap on every back-to-back campaign. resolveSlot answers placed, fallback or empty and populates reasons on ALL THREE: an empty shelf and a shelf whose only candidate was refused for rights look identical to a reader, and the first is a scheduling gap while the second is a legal question. Its four empty-reasons are distinct -- nothing targets the slot, everything is outside its interval, everything excludes this channel/territory/locale, or a named placement was refused with its dimension -- and the winner is chosen by priority, then later start, then id in code-point order, which is what makes the choice total. -->

    • S9.9.d Define raw rank signals/units/source/version/time separately from rank policy/weights/tie-break/diversity and final position. Evidence: deterministic ranking tests. <!-- S9.9.d: the separation is structural and the spec asserts it as structure: RankSignal has no weight key and RankTerm has no value key. A pipeline that stored 0.35 * dwell has thrown away the dwell, and no later policy can recover it. Readings are INTEGERS at a declared decimal scale and weights are integer basis points, so the composite is summed exactly and in the POLICY's declared term order -- the caller's map order cannot reach it, which is the whole of scoreRankerCandidate's 2.22e-16. A rescale that would round is refused rather than performed, because a rank that rounds its inputs is a rank two pipelines can disagree about, and both directions are driven: 3000 at 1e-3 becomes 3 exactly and 3500 does not become 3.5. MAX_TERM_MAGNITUDE is a billion, which is where the two constraints meet -- it leaves room for ninety full-weight terms before the worst case reaches 2^53-1, and the first version had it at a trillion, where the module refused its own example policy. The spec caught that, which is what a spec is for. policy-tie-break-not-total is the finding that matters: two items with equal composites keep the order they arrived in, and a limit downstream turns that into membership. Items refused for a missing required signal, cut by the limit, or demoted by a diversity cap are all CARRIED in the result with their reasons and their contributions, next to the ones that ranked. -->

    • S9.9.e Define eligibility rules for lifecycle/gates/freshness/rights/ consent/embargo/audience/locale/territory/channel/capacity and exact refusal reasons. Evidence: eligibility truth tables. <!-- S9.9.e: eleven dimensions, each a separate refusal with a separate remediation class, because a reader whose item was refused for rights and one whose item was refused for freshness have nothing to do with each other and a single boolean sends both to the same page. ELIGIBILITY_COMBINATION is the truth table as data, and its asymmetry is the decision: ineligible absorbs indeterminate and indeterminate absorbs eligible, so a known refusal stays a fact when something else times out and an unknown never becomes a yes. A DIMENSION NOBODY READ IS indeterminate RATHER THAN ABSENT -- a pipeline that forgot to check rights has not established that rights are fine -- and the spec drives exactly that by removing one dimension from an otherwise all-eligible set. Every reason is kept rather than the first, because a queue that answers one rejection at a time over a review cycle that takes a day is a queue people stop putting things into. freshnessAt has THREE states because an item that is not yet current is neither fresh nor stale, and merging it into either loses the difference between an embargoed piece and an expired one; the shipped one-sided window is the reason. And evaluateRelationshipScope is the creator-only fix: it resolves a REF against the party the item names, and an item that records no holder is indeterminate rather than eligible -- which is precisely the state that made the shipped rule compare tiers. -->

    • S9.9.f Define audience segments by stable governed refs, allowed attributes/purpose/consent, version, minimum cohort/privacy, and noninference rules. Evidence: audience security tests. <!-- S9.9.f: a segment is defined over GOVERNED ATTRIBUTE REFS, and the registry's classification is what decides whether it may be targeted on. special-category-derived is the class that matters: an attribute INFERRED from a special category is still that category, so a "wellness affinity" computed from which health articles somebody read is health data with a different name, and a registry that classes it as ordinary launders it. Proxy detection is structural rather than statistical, because the registry already knows what each attribute was derived from. A SEGMENT BELOW ITS FLOOR IS REFUSED AT EVALUATION AND NOT SUPPRESSED AT REPORT: suppressing the number still answers the question when the person asking chose the predicate, since "fewer than five" for country = GH AND role = admin AND salaryBand = 9 has already said there is at least one and the asker supplied every term. The floor is also absolute rather than self-declared, because a minimum a segment sets for itself can be set to one. And differencingFaults is the disclosure a per-segment check cannot see: two segments of 40 and 38 where the second is contained in the first each pass their own floor, and the two people in the difference are identified by subtraction -- so the floor is a property of every PAIR, and the intersection case is the same statement for overlapping rather than nested segments. -->

    • S9.9.g Compose/validate curation with no duplicates/cycles/invalid refs, deterministic ordering, preview/diff, review, and expected revision. Evidence: composition tests. <!-- S9.9.g: composeCuration validates every part TOGETHER rather than in sequence, because the faults that matter are between the parts -- a placement pointing at a slot that does not exist, a collection member whose ref does not resolve, a policy whose tie break is not total over the items it is about to order. A ref carries the VERSION it was resolved at, and ref-resolves-to-a-different-version is separate from ref-does-not-resolve because the remedies differ: the first is a bundle that needs rebuilding against the corpus it ran on and the second is a typo or a deletion. diffCuration reports member-moved as its own kind rather than as a removal plus an addition, because a reviewer approving a reorder and one approving a swap are approving different things and a diff that renders both as -a +a asks them to spot it; and ambient-changed exists so that a bundle whose members moved because somebody's container changed time zone says so instead of presenting it as an editorial decision. requiredReview takes the STRICTEST any single change requires -- an eligibility change is rights-and-owner because it is the one that can put something in front of somebody not permitted to see it -- and an ambient change raises nothing at all, because asking an owner to approve a container move hides the container move. -->

    • S9.9.h Build a versioned bundle whose identical inputs produce identical members/ranks/eligibility/reasons/hash. Evidence: reproducibility test. <!-- S9.9.h: THE HASH COVERS THE INPUTS, THE POLICY AND THE AMBIENT DECLARATION -- NOT THE OUTPUT. Hashing the output makes two bundles agree exactly when nothing interesting is true: a build that dropped every member because a collector was down hashes identically to another that did the same, and the pair reads as a reproduction. The spec pins the consequence: hand-editing a bundle's members and verdicts does NOT move the hash, and it is the per-axis findings that catch it and name which axis moved. Every ambient input the computation could read is declared in AmbientDeclaration and hashed with it, and a build whose declaration is incomplete is REFUSED rather than defaulted, because a bundle that reads Intl.DateTimeFormat().resolvedOptions().timeZone is a property of the container it ran in and container images change without anybody deciding to change a ranking. canonicalize is composed from revisions.ts rather than written again, which is why building returns a ParseResult: a bundle carrying a NaN would otherwise hash the same as one carrying Infinity, both having serialized to null.

      THE PROOF IS `tools/curation-determinism.mjs`, and it is the half no
      in-process suite can settle: a single process has one locale, one time
      zone, one ICU build and one heap layout, so running a build twice in it
      proves the function is not calling `Math.random` and cannot prove it is
      not calling the container. The tool builds the SAME bundle in child
      processes under `LC_ALL` of en_US, tr_TR, sv_SE, C and unset and `TZ`
      fourteen hours apart across the date line, times four input
      permutations, and requires one members list, one excluded list, one
      verdict set, one input digest and one hash across all twenty -- after
      first checking that the children really did resolve to four different
      locales, so the claim is not one container measured twenty times. Six
      controls each reproduce a shipped defect before the claim is measured
      against them: A LOCALE SORT IS A CONTAINER SORT (three orders from one
      input), A LIMIT OVER A TIED ORDER CHOOSES MEMBERS (four different member
      SETS from four permutations), FLOAT ADDITION IS NOT ASSOCIATIVE
      (0.6800000000000002 against 0.6799999999999999), A MAP WALK WITH A BREAK
      IS AN ARRIVAL-ORDER FILTER, A CLOCK READ IS AN UNDECLARED INPUT, and
      UTF-16 ORDER IS NOT UTF-8 ORDER. A control that stops reproducing its
      defect is reported as a BROKEN CONTROL rather than as a pass.
      
      AND IT EARNED ITS PLACE ON ITS FIRST RUN. It reported four hashes where
      the members, the excluded list and the verdicts were all identical: the
      enumerated membership was being canonicalised in whatever array order
      the caller built it, although the ordinal is the content and the array
      order is presentation. A real defect in this module, invisible to 189
      in-process tests, found by the one thing that varies what a process
      cannot vary about itself. Both halves are now pinned in the spec -- the
      same membership in two orders digests the same, and a membership whose
      ORDINALS moved digests differently, because a reorder is an editorial
      change that must not ship under a pinned version. -->
      
  • S9.10 Prohibit direct production-catalog writes. Evidence: all changes pass through reviewed versioned bundles.

    • S9.10.a Inventory database/admin/API/job/import/script/provider paths capable of changing production catalog, search, feature, rank, collection, or audience state. Evidence: write-path register.

    • S9.10.b Define draft proposal/bundle, semantic diff/impact, validation/ gates, required reviews/approvals, schedule/embargo, publish, verify, and rollback lifecycle. Evidence: catalog workflow contract.

    • S9.10.c Restrict production catalog credentials and write functions to the publisher identity and deny humans/hosts/domain services direct access. Evidence: authorization tests.

    • S9.10.d Bind publisher execution to immutable approved bundle/hash, target/environment, expected production version, idempotency, and step-up. Evidence: command tests.

    • S9.10.e Prevent alternate database, search index, cache, compatibility, migration, admin, or provider API writes from bypassing the workflow. Evidence: bypass tests.

    • S9.10.f Read back and verify production catalog, preserve immutable release/audit, and reconcile partial/unknown/rollback outcomes. Evidence: end-to-end publish tests.

    • S9.10.g Add CI/runtime detection for unregistered catalog mutation code/ credentials. Evidence: deliberate bypass failure. <!-- S9.10: shipped as libs/oshun/workbench-kit/src/catalog-write-paths.ts (+ spec, 109 tests, 100% of statements, branches and functions) and libs/oshun/workbench-kit/tools/catalog-bypass-scan.mjs with its register tools/catalog-write-paths.json (nx target catalog-bypass-scan). 149 kit entry points.

      This item is the word ALL in S9.9's evidence clause, and ALL is a claim
      about a closed world: false the moment one door exists that the bundle
      does not lead through, and invisible to the bundle, because that door is
      precisely the thing that never produces one.
      
      THE VEIN, STATED SO IT TRANSFERS: AN APPROVAL THAT TRAVELS ON THE THING
      BEING APPROVED IS NOT AN APPROVAL, AND A CHECK THAT QUANTIFIES OVER THE
      LIST THE WRITER PASSED CANNOT REFUSE WHAT THE WRITER LEFT OUT. The grep
      is mechanical and it is three greps. In every admission, advance or gate
      function: (1) which fields of the verdict are read off the SUBJECT --
      `edge.reviewState`, `gate.satisfied`, `gate.required`; (2) which are
      taken from a CALLER-SUPPLIED BOOLEAN -- `reviewerIsExpert`, `isAdmin`;
      (3) which JOIN THEIR EVIDENCE BY IDENTIFIER where the evidence is about
      content -- `preview.requestId === rollout.requestId`. All three are
      shipped here, and every claim below was measured against the live
      function before it was touched.
      
      `search-discovery/src/concept-graph/schema.ts::admitConceptEdge` is the
      admission control for the concept graph and its high-stakes rule reads
      `edge.reviewState !== 'approved' || edge.approvedByExpertId === null` --
      both fields of the edge the caller is submitting. A caller who writes
      `reviewState: 'approved', approvedByExpertId: 'me'` is admitted and no
      reviewer ever existed: the gate asks the thing being admitted whether it
      was approved. Worse in the other direction: for the nine kinds that are
      NOT prerequisite or contradicts there was no review rule at all, so an
      edge whose own state read `rejected` -- a reviewer looked at it and said
      no -- entered the live graph, and `listNodeNeighbors`, in the module
      whose sibling `queries.ts` filters all eight of its traversals through
      `approvedEdges` with a comment saying why, reported it as a fact about
      the concept.
      
      `concept-graph/curation.ts::reviewConceptEdgeProposal` took
      `reviewerIsExpert: boolean` as an INPUT, so the expertise gating every
      high-stakes edge in the system was whatever the caller passed, and it
      let `reviewerId` equal the proposal's `proposerId`: alice proposed,
      alice approved, alice certified alice, and the edge went into the graph
      with a provenance record naming alice as its reviewer.
      `withdrawConceptEdgeProposal` took no actor at all -- its own doc says
      "proposer cancels their own proposal" -- so anybody could withdraw
      anybody's, and the record it wrote set `reviewedAtUnixSeconds` while
      leaving `reviewerId` null: a decision by nobody, at a time.
      `listPendingProposals` applied its age filter only when BOTH
      `olderThanSeconds` and `nowUnixSeconds` were supplied, so
      `{ olderThanSeconds: 604800 }` returned every pending proposal -- the
      widest possible answer to a request to narrow -- and
      `summariseCurationQueue` folded negative ages into a maximum starting at
      zero, so a queue of future-stamped work reported no old work.
      
      `studio-authoring/src/editorial-lifecycle/lifecycle.ts::tryEditorialAdvance`
      was one line: `input.gates.filter((g) => g.required && !g.satisfied)`.
      `input.gates` is the caller's array and both booleans are the caller's,
      so the function had no idea which gates exist and `approved ->
      published` with `gates: []` succeeded. The producer of the right list
      was one file away -- `checklists.ts::resolveEditorialChecklist` resolves
      the required gates per content type and state and its own doc says the
      result "can be handed straight to `tryEditorialAdvance`" -- and nothing
      checked that it had been, while `apps/oshun/bff/.../editorial-publisher.ts`
      takes `gatesFor` from the deployment and two of its four call sites
      passed `gates: []` outright. The kit had already decided the general
      form in S2.9, where `UNAVAILABLE_BEHAVIOURS` has no `pass` member "not
      as a rule -- as a type"; a decision that stays inside the kit is a
      decision the library that needed it never got. The same fix found the
      gate pointing backwards: the non-publish arm of the checklist returned
      the authoring-evaluation gate for EVERY state, so a takedown was gated
      on a check about the quality of the writing.
      
      `studio-authoring/src/taxonomy-curation/curation.ts::tryAdvanceOntologyRollout`
      is the third grep. A previous pass fixed the missing preview and wrote
      "a stage name is not a gate" above the fix; the join it added is
      `preview.requestId !== rollout.requestId`. The preview is a computation
      over `request.targetIds`, the request is a mutable record, and the join
      is on its id. Measured: a preview whose `affectedTagIds` is
      `["tag:obscure"]` admitted a request whose targets were
      `["tag:every-artifact-we-have"]`, reporting `complete: true` and a blast
      radius of one tag. The same function rolled out from `staged` with no
      preview, no approval, no reviewer, a year-old preview, and a request
      whose own `contestStatus` read `contested`.
      
      And two silent-drop writes in the index. `catalog/index-engine.ts::indexObject`
      answered `SearchIndexEngine` for both outcomes, so an object of another
      class was dropped and the caller got back a value indistinguishable from
      a completed write; `indexObjectInRegistry` did the same for a class the
      manifest has no engine for, which is a deployment fault. Its embedding
      arm read `engine.embedding && payload.embedding ? indexed :
      engine.embedding`, so re-indexing an object whose embedding was
      withdrawn KEPT THE PREVIOUS VECTOR: the lexical document said one thing
      and the vector said what the document used to say, and a semantic query
      still returned the object on the strength of text that had been edited
      or taken down.
      
      Sixteen in-place fixes across the three libraries and the BFF, and all
      thirty-two mutations of them were caught: reverting the approval record,
      the self-review refusal, the roster resolution, the two-sided
      qualification window, the withdrawal actor, the future-dated queue
      count, the rejected-edge neighbour filter, the four-field provenance
      predicate, the requirement loop, the deployment-added gate, the
      ungated removal, the canonical-form binding, the preview window, the
      approval requirement, the contested refusal, the class-mismatch refusal
      and the stale-vector removal each turns a test red.
      
      TEN DECISIONS. A PROHIBITION IS A CLOSED WORLD AND A REGISTER IS ONLY A
      CLAIM ABOUT ONE, so `registerFaults` judges the entries it has and the
      register is worth what `reconcileRegister` says when it is set against
      an INDEPENDENT enumeration -- and a reconciliation against zero
      observations reports zero unregistered paths, which reads identically to
      compliance, so an empty or partial scan is itself a finding. AN OMITTED
      GATE IS AN UNMET GATE: `changeAdmissionFaults` iterates the CONTRACT's
      requirements and looks each up in what was presented, and never the
      reverse, because the two loops are one keystroke apart and only one of
      them can refuse an omission. AN APPROVAL BINDS TO A DIGEST AND NOT TO AN
      IDENTIFIER -- the id says WHICH change, the digest says WHICH BYTES, and
      evidence whose `subjectDigest` is null is refused rather than warned
      about. A CHANNEL IS CLOSED BY A MECHANISM OR IT IS OPEN, so
      `documented-convention` and `code-review` are in the vocabulary
      precisely so a register can SAY them and both report as open; the one
      self-serving closure, `writes-no-shared-state`, carries its criterion in
      the type's own comment. THE PUBLISHER IS AN IDENTITY AND NOT A ROLE
      PEOPLE CAN ASSUME, so `assumableBy` is read as well as `principal` and a
      break-glass grant must expire and must name where its use is recorded.
      THE EXPECTED PRODUCTION VERSION IS PART OF THE COMMAND, because
      publishing onto whatever is there now is a lost update with a receipt.
      AN INDEX, A CACHE AND A PROJECTION ARE PRODUCTION CATALOG STATE, and two
      writers into one derived class is its own finding even when both are
      legitimate, because each reads back exactly what it wrote. AN UNKNOWN
      OUTCOME IS NOT A FAILED ONE, so `reconcilePublication` answers with five
      outcomes and names `reconcile-before-any-retry` rather than retrying.
      THE VERIFICATION READS THE TARGET AND NOT THE RECEIPT: an observation
      sourced from the command's own reply is a fault AND is excluded from the
      read count, so a publication verified entirely against its own answer
      reconciles to `unknown`. AND EVERY REFUSAL NAMES THE ROUTE IT WOULD HAVE
      TAKEN INSTEAD, because a prohibition with nowhere to go is one people
      route around, and the route they find is the one nobody registered.
      
      `tools/catalog-bypass-scan.mjs` is the half no in-process suite can
      settle. It reads 1395 real source files under four roots, strips
      comments and string literals first -- every file that DESCRIBES the scan
      mentions every pattern it looks for, so a scanner that greps raw text
      reports its own documentation as a fleet of bypasses and is switched off
      within a day -- and finds every SQL DML against a catalog table, every
      search-index write, every taxonomy admission, every editorial advance
      and every catalog credential read, anchoring each on its enclosing
      exported symbol rather than on a line number that moves. It found
      fourteen real sites, every one of which is now classified in the
      register with an owner and a closure. SIX CONTROLS, each of which must
      reproduce its defect first: an unregistered SQL write, an unregistered
      index write, an unregistered credential, a MENTION that must produce
      nothing, an EMPTY SCAN that must not read as a clean one, and a stale
      register entry. A deliberate `indexObjectInRegistry` call added to
      `libs/oshun/customer-curation` fails it with the file, the symbol and
      the matched line, and removing one entry from the register fails it the
      other way; exit 1 both times, exit 0 clean.
      
      Scope stated: `schedule/embargo` in S9.10.b is the `scheduled` stage in
      the contract's stage list -- the embargo predicate itself is S9.6's
      `lifecycle.ts::isEmbargoed` and is not restated here -- and S9.10.f's
      "end-to-end publish tests" are exhaustive over the reconciliation of a
      read-back, since V1 has no live production catalog service to publish
      into; the live end-to-end equivalent is the BFF editorial publisher
      walking the real state machine under the checklist's required gate ids.
      -->
      
  • S9.11 Add search indexing events with replay and dead-letter handling. Evidence: full rebuild matches incremental index.

    • S9.11.a Define versioned index event types for upsert/delete/tombstone/ permission/taxonomy/localization/rights/gate/catalog/visibility changes. Evidence: event schemas.

    • S9.11.b Emit authoritative event with tenant/object/revision/sequence, changed field classes, causation/correlation, and projection version via outbox. Evidence: atomicity tests.

    • S9.11.c Consume idempotently with per-object/version ordering, late/ duplicate/out-of-order/gap/deletion handling, and property redaction. Evidence: delivery tests.

    • S9.11.d Build documents through versioned projection adapter and perform atomic replace/delete without partial searchable state. Evidence: indexing integration tests.

    • S9.11.e Retry transient failures and dead-letter malformed/unsupported/ deterministic failures with owner, safe cause, source event, and replay. Evidence: dead-letter tests.

    • S9.11.f Implement checkpointed full rebuild from authoritative data/ events into new index version with validation and atomic alias/cutover. Evidence: rebuild rehearsal.

    • S9.11.g Compare incremental and rebuild docs/counts/hashes/queries/ permissions and repair divergence. Evidence: parity report. <!-- S9.11: shipped as libs/oshun/workbench-kit/src/search-index-events.ts (+ spec, 98 tests, 23 of 23 mutations of the shipped decisions caught), entry point ./search-index-events (150 approved, no drift), plus in-place fixes in search-discovery, domain-tara, offline and apps/oshun/mobile.

      The evidence clause -- "full rebuild matches incremental index" -- is the
      only evidence in the phase that cannot be produced by the thing it
      judges. An index is a DERIVED store, and every other check on it reads
      the derived store and asks the derived store whether it is right, which
      is blind to exactly the two failures that leave nothing in it to read:
      the object that never arrived, and the object that should have left and
      never did.
      
      THE VEIN, STATED SO IT TRANSFERS: A DERIVED STORE IS ONLY AS TRUE AS THE
      ENUMERATION THAT MAINTAINS IT, AND AN ENUMERATION OVER THE DERIVED STORE
      CANNOT SEE WHAT NEVER ARRIVED. The grep is one question asked of every
      maintenance loop: WHAT DOES IT ITERATE? A loop over the derived store can
      refresh and can evict and can never ADD. A loop over the source can add
      and can update and can never DELETE. Only a loop over the UNION is a
      reconciliation -- and the moment one side's membership is read off the
      other, the comparison agrees with itself for every input.
      
      Four instances, measured before anything was written, all the same loop.
      
      `search-discovery/src/catalog/index-engine.ts::freshenIndexEngine` walked
      `lexical.docs`, `embedding.entries` and `facets.indexedAtByObject` --
      three enumerations of the INDEX -- and refreshed or evicted whatever was
      past the class budget. An object present in `liveObjects` and absent from
      every sub-index was never looked at, so a document whose indexing event
      was lost was invisible to the pipeline whose NAME is freshness. Worse in
      the other direction: an object deleted from the catalog while its entry
      was still fresh was not stale, so it was not in the stale set, so it was
      not evicted -- and `defaultFreshnessTargetSecondsFor` answers
      `365 * 86400` for every class outside veritas, nyx, metis and arete, so a
      takedown that reached the catalog and not the index stayed searchable for
      up to a year while the report called the class clean. It now walks the
      union, and a snapshot must DECLARE whether it is the complete class,
      because a partial one may not evict anything.
      
      `domain-tara/src/favorites-sync.ts` called
      `getFavorites({ limit: 500 })` and handed the result over as the
      authoritative remote set. A PAGE IS NOT A SET: the five hundred and first
      favourite was "not remote", so it was counted as a remote deletion and
      dropped from the device. The same function called an item `unchanged`
      when its id appeared on both sides, so a meditation renamed,
      recategorised or re-timed reconciled as unchanged and the stale copy was
      the one kept. It now pages to a short page, degrades to `partial` rather
      than deleting when an account outruns the budget, dedupes rows an offset
      walk repeats, and counts `updated` apart from `unchanged`.
      
      `domain-tara/src/ritual-completion-events.ts` builds the only search
      document any V1 library produces from an event, and keyed it
      `${event.eventId}-search` where the event id ended in
      `Math.random().toString(36).slice(2, 8)`. A DOCUMENT IS KEYED BY THE
      OBJECT IT DESCRIBES AND NEVER BY THE EVENT THAT ANNOUNCED IT: with a
      random component in the key, one completion delivered twice was two home
      signals, two activity rows, two search documents and six
      recommendations, none carrying a field naming what they duplicated, and
      a rebuild -- which reads completions, not deliveries -- could never
      produce the same set. Every projection is now keyed on
      `taraRitualCompletionObjectId`, and the test that asserted
      `secondEvent.eventId).not.toBe(firstEvent.eventId)` was pinning the
      defect.
      
      `offline/src/queue.ts::process` was the dead letter that is not one. On
      the attempt reaching `policy.maxAttempts` it pushed an item result and
      `continue`d WITHOUT pushing the item onto `nextQueue` -- and `nextQueue`
      is what `saveQueue` persists -- so the operation was deleted from durable
      storage and survived only in the array returned to the caller. Its one
      caller, `apps/oshun/mobile/.../offlineQueueController.ts`, kept those in
      a field, in memory, on a mobile client: the process that died between the
      failure and the retry lost the user's queued write with no record
      anywhere that it had existed. And every failure was transient, so a
      payload the server would never accept was retried four times on a metered
      radio and then discarded. There is now a persisted dead-letter store, a
      caller-supplied failure classifier, replay under the ORIGINAL id (the
      previous `enqueue(type, payload)` minted a new one, so a retry was a
      second write), and per-object head-of-line ordering so a queued edit
      cannot be overtaken by the next edit to the same object.
      
      TEN DECISIONS. A DELETE AND A TOMBSTONE ARE DIFFERENT EVENTS AND THE
      REBUILD IS WHERE THE DIFFERENCE BITES -- a tombstone is the one fact that
      is true of the index and false of the authoritative store, because the
      row is usually still there, so a plan whose `tombstoneSource` is `none`
      republishes every takedown it ever applied, and the parity report then
      calls those documents `only-in-rebuild`, which is the shape of a MISSING
      document; repairing it the obvious way puts them back, which is why
      `repairPlan` consults the tombstone list before anything else. A REVISION
      ORDERS CONTENT AND A SEQUENCE ORDERS EVENTS, AND ONLY ONE OF THEM MOVES
      WHEN A PERMISSION CHANGES, so deduplicating on the revision drops every
      envelope-only change -- precisely the class whose staleness is a
      disclosure. THE STALENESS BUDGET IS A PROPERTY OF THE FIELD CLASS AND NOT
      OF THE PIPELINE, since one "index lag" number treats a corrected title
      and a revoked ACL alike and is always chosen for the title. A GAP IS NOT
      A REASON TO STOP AND IT IS NOT A REASON TO PROCEED, so it is held for a
      window and then answered with `request-replay` naming the missing range.
      THE DISPOSITION OF A FAILURE IS A PROPERTY OF THE FAILURE AND NOT OF THE
      ATTEMPT COUNT -- only `unknown` consults the count, exactly as S3.6's
      `decideDelivery` already decided -- and `hold-for-upgrade` is the third
      answer for an event emitted for a projection this worker does not have.
      A DEAD LETTER THAT DOES NOT CARRY ITS SOURCE EVENT IS A LOG LINE; it
      names an owner, and its cause is CHECKED against the values the redaction
      policy protects rather than trusted, because an operational store is
      where careful redaction goes to be undone. A DOCUMENT REPLACE IS ONE
      OPERATION OR THE INDEX IS QUERYABLE IN A STATE NOTHING PRODUCED, across
      sub-indexes too: a write that fills the lexical document and leaves the
      previous vector is the defect this repository already fixed once in
      `indexObject`, and it is a plan-level fault here so the next pipeline
      cannot reach it independently. AN ANALYZER CHANGE IS AN INDEX VERSION
      CHANGE -- nothing in a document changes when the tokenizer does, so the
      old documents are simply unfindable by queries the new analyzer parses,
      which no document comparison can see and only a QUERY probe can. A
      REBUILD PINS A WATERMARK BEFORE IT READS ANYTHING, or every change made
      while it ran is silently rolled back by the cutover, and it checkpoints
      against its own plan digest, because a checkpoint is an index into an
      ORDER and resuming a changed plan reports completion over a set it never
      covered. AND A PARITY REPORT WHOSE TWO SIDES SHARE AN ENUMERATION PROVES
      NOTHING, WHILE ONE OVER AN EMPTY CORPUS PROVES LESS: both produce zero
      divergences, which reads exactly like parity, so `conclusive` is a field
      and `rebuildMatchesIncremental` requires it.
      
      Scope stated: S9.11.d's "indexing integration tests" and S9.11.f's
      "rebuild rehearsal" are exercised against the real in-process index
      engine and the real offline queue over real storage, since V1 has no live
      Elasticsearch or OpenSearch deployment to cut an alias over; the kit
      owns the plan, the cutover refusal and the parity arithmetic as data, and
      `freshenIndexEngine`'s union loop is the live reconciliation those rules
      describe. Emission composes S3.6's `outbox.ts` -- the transaction-bound
      row, per-subject order and the `retryable | permanent | unknown`
      classification are not re-decided here, only specialised to what a search
      event has to carry.
      -->
      <!-- S9.11 marked after writing and reading
      `libs/oshun/workbench-kit/src/search-index-events.ts` end to end, running
      its 98 tests, mutating 23 of its decisions and confirming each turns a
      test red, and after running the four repaired libraries' own suites:
      search-discovery 166, domain-tara 318, offline 13, mobile jest 68. -->
      
  • S9.12 Add permission-aware search projection. Evidence: hidden titles and counts do not leak.

    • S9.12.a Define searchable/displayable/facetable/sortable/suggestible/ exportable fields per property authorization, audience, lifecycle, classification, and tenant scope. Evidence: projection policy.

    • S9.12.b Project tenant/workspace/object ACL or authorization filter version plus protected-source/learner/biometric/unreleased redaction without raw credentials/roles. Evidence: index document tests.

    • S9.12.c Apply authorization before query, total/count, facet, group, suggestion/autocomplete, highlight/snippet, ranking, and pagination. Evidence: disclosure negative tests.

    • S9.12.d Revalidate current object/property authorization at result use to cover stale index and emergency revocation. Evidence: stale ACL tests.

    • S9.12.e Update/delete index immediately on permission/ownership/ classification/embargo/rights/consent/lifecycle change. Evidence: invalidation latency tests.

    • S9.12.f Prevent hidden existence inference through timing, score, spelling suggestion, related item, empty-versus-forbidden errors, cached result, and cursor. Evidence: side-channel tests.

    • S9.12.g Test two tenants with colliding IDs/titles/content/taxonomy and ordinary/assigned/reviewer/auditor/support roles. Evidence: isolation matrix. <!-- S9.12: shipped as libs/oshun/workbench-kit/src/search-permission-projection.ts (+ spec, 95 tests, 54 of 54 mutations of the shipped decisions caught), entry point ./search-permission-projection (151 approved, no drift), plus in-place repairs in libs/contracts, apps/oshun/bff, search-discovery and domain-tara.

      The evidence clause is "hidden titles and counts do not leak", and the
      second noun does the work. Filtering the ROWS is true of every
      implementation that has this defect, because a hidden row is always
      withheld; what escapes is the total, the page fill, the facet count, the
      suggestion, the score of the rows that remain, and the id in the cursor.
      
      THE VEIN, STATED SO IT TRANSFERS: A SEARCH SURFACE IS A SECOND READ PATH
      TO EVERY ROW IT COVERS, AND ITS AUTHORIZATION IS ASKED ABOUT THE QUERY
      RATHER THAN ABOUT THE ROW. The direct route asks "may you read THIS
      object". The search route asks "may you use THIS facet", "may you see
      THIS domain", "may you reach THIS index" -- questions about the request,
      answered before a row is loaded -- and then reads whatever the collector
      reads. Two greps find it everywhere: which noun the permission check
      takes as its SUBJECT, and which values in the response were computed from
      rows the check never saw.
      
      Four instances, measured against the running code before anything was
      written.
      
      `apps/oshun/bff/src/admin/admin-search.ts::runAdminUniversalSearch` gates
      on `facetsVisibleTo(scopes)`, and the `libs/contracts` table unlocked the
      `cases` facet for any ONE of `admin:workspace:support`, `:privacy`,
      `:rights` or `:incident` while `collectCaseHits` read all four
      workspaces. THE FACET IS A UNION OF SOURCES AND THE REQUIREMENT WAS AN
      ANY-OF OVER THAT UNION, so any-of collapsed to all-of. Measured against
      the booted BFF with a token carrying only `admin:workspace:incident`:
      `GET /v1/admin/workspaces/privacy` answered
      `403 admin_scope_missing` and
      `GET /v1/admin/search?q=privacy&facets=cases` answered `200` with three
      privacy requests and a rights request. `artifacts` collapsed the same
      way -- a `:persona`-only operator was served `policy-privacy-dsar` from a
      workspace whose own route refuses them. And the needle was scored against
      `item.subjectId`, which no hit renders: `q=user-privacy-02` returned one
      export request and never printed the id, so the query box answered a
      question about a data subject that the answer was careful not to state.
      The requirement now lives on the SOURCE, the facet's requirement is
      DERIVED from its sources, and a partially served facet reports
      `skippedSources` rather than looking like an empty workspace.
      
      `search-discovery/src/catalog/index-engine.ts` ran
      `searchLexicalIndex(index, query, topK)` to `hits.slice(0, topK)` and
      offered `filterVisibleHits` as the visibility step AFTERWARDS. Two
      defects in one line: the caller who asked for twenty got three, and the
      seventeen missing rows were the count of the matches that viewer may not
      see. The scores were worse -- BM25 reads `docs.size`,
      `documentFrequency` and `averageLength`, three statistics over the WHOLE
      corpus, so the rank of a document a viewer can see moved when a document
      they cannot see was added, which is a cross-tenant existence oracle
      anybody who can write a document can operate. `aggregateFacetCounts`
      counted over whatever scope it was handed and `filterByFacets` with no
      filters handed back every indexed object. All three now take a required
      `AuthorizedObjectIds`, the statistics are recomputed over it, and
      `revalidateHits` separates a hit whose object LEFT THE CATALOG (the index
      is behind) from one whose object refuses this viewer (the query was
      answered over too wide a set) -- the previous `if (!object) return false`
      reported neither.
      
      `search-discovery/src/candidates/generators.ts` builds the related rails
      and every one of them cut its top-K before entitlement, which
      `applyCoherenceConstraints` applied afterwards. A hidden object consumed
      the slot and the rail came back short: four slots rendering two tells the
      reader that two objects exist which are more similar to what they are
      reading than anything they may open, and which ones is often obvious from
      the seed. Measured: with the two nearest neighbours hidden and
      `maxPerSeed: 2`, the rail returned NOTHING.
      
      `domain-tara/src/ritual-completion-events.ts` builds the only search
      document any V1 library produces, and it carried a title, a summary, a
      body, tokens and a route -- a member's practice record -- and no owner,
      no tenant, no audience and no filter version. A DOCUMENT WITH NO SUBJECT
      CANNOT BE FILTERED TO ONE, however carefully the query is written, and
      the one place the member appeared was inside `documentId`, which is the
      field that travels furthest. It now carries `subjectId`, `audience:
      'subject-only'` and `authorizationFilterVersion`; the key is left alone,
      because S9.11 keyed every projection on the completion deliberately.
      
      TEN DECISIONS. THE UNIT OF AUTHORIZATION IS THE ROW AND A FACET IS NOT A
      ROW -- a facet, a tab, a domain filter and an index name are groupings
      the UI chose, and the moment one spans two sources with different owners
      an any-of requirement is read by the collector as all-of. A FIELD'S
      PROJECTION IS SIX PERMISSIONS AND THE MATCH IS ONE OF THEM: searching
      discloses EXISTENCE, faceting the value AND a count, sorting ORDER,
      suggesting the value verbatim -- so a searchable-and-not-displayable
      field is a match oracle, licensed only by a named human acceptance
      carrying its reason, and never for a class where presence IS the
      disclosure. THE DOCUMENT HOLDS THE UNION AND THE ANSWER HOLDS THE
      INTERSECTION, which is why the authorization filter is a required field
      of the document rather than a lookup beside it. THE AUTHORIZED CORPUS IS
      A TYPE AND NOT A STEP -- every stage after authorization takes
      `AuthorizedCorpus`, which `authorizeCorpus` is the only function able to
      produce, so a stage cannot read the wider set because it holds no
      reference to it; a stage ordering written down as a list of names is a
      comment, and this is the same ordering held by the compiler. THE SCORE IS
      A CORPUS STATISTIC SO THE CORPUS HAS TO BE THE VIEWER'S. TRUNCATION AFTER
      AUTHORIZATION, NEVER BEFORE. AN EMPTY ANSWER AND A FORBIDDEN ANSWER ARE
      THE SAME ANSWER, so there is no forbidden code and no field able to carry
      one -- S4.3's conceal-outranks-deny at the level of a result set. A
      CURSOR IS BOUND TO THE VISIBILITY IT WAS MINTED UNDER, since S4.9 already
      binds tenant, actor and query and none of those move when an ACL does --
      and the digest covers the INPUTS to visibility and never the outcome, or
      indexing an ordinary document would invalidate every outstanding token
      and the digest would itself be a value moving with the corpus.
      REVALIDATION IS A BACKSTOP AND NOT A CONTROL: it shrinks the rows, cannot
      repair the aggregates without the corpus, says so in the report, treats
      could-not-tell as a drop that is NOT evidence of a stale index, and
      discards whole any answer older than the revocation epoch. AND A
      NARROWING IS A DELETE WHILE A WIDENING IS AN UPDATE, AND ONLY ONE OF THEM
      HAS A BUDGET -- a reshape is a narrowing, because taking the safe half
      first costs nothing.
      
      Scope stated: S9.12.e's "invalidation latency tests" are exercised
      against observations of a pipeline rather than against a live indexing
      cluster, since V1 has no Elasticsearch or OpenSearch deployment to
      measure a queue on; the kit owns the action, the budget per change class
      and the breach arithmetic as data, and the four repaired libraries are
      where the decisions are applied. S9.12.b's redaction composes S3.12's
      `STRATEGY_CAPABILITIES` and S3.13's `PRESENCE_DISCLOSING_CLASSES` rather
      than re-deciding which operations a protected field can support.
      -->
      <!-- S9.12 marked after writing and reading
      `libs/oshun/workbench-kit/src/search-permission-projection.ts` end to
      end, running its 95 tests, mutating 54 of its decisions and confirming
      each turns a test red, and after running the repaired libraries' own
      suites: workbench-kit 10264, search-discovery 179, domain-tara 321,
      contracts admin-universal-search 22, BFF admin-search route 15 plus the
      OpenAPI coverage contract. -->
      
  • S9.13 Add localization jobs, translation memory refs, glossary refs, review, and parity gates. Evidence: locale bundle round-trip.

    • S9.13.a Define localization job source revision/locale, target locale/ script/region, field/unit list, provider/method, references, policy, budget, and lifecycle. Evidence: job schema.

    • S9.13.b Segment structured content without localizing IDs/code/URLs/ placeholders/standards and preserve stable source anchors/order/context. Evidence: segmentation tests.

    • S9.13.c Reference authorized versioned translation memory matches with source/target locale, segment/hash, score, rights/confidentiality, and provenance. Evidence: TM tests.

    • S9.13.d Reference versioned glossary terms, morphology/context, required/ forbidden translations, case, locale scope, and precedence. Evidence: glossary tests.

    • S9.13.e Store generated/human translation as proposal with provider/ model/template/run/cost or translator identity and no auto-approval. Evidence: proposal lifecycle tests.

    • S9.13.f Implement linguistic/domain/legal/safety/accessibility review, suggestions, comments, required roles/quorum, and approval per locale. Evidence: review workflow tests.

    • S9.13.g Gate completeness, placeholder/markup/reference parity, terminology, numbers/units/dates, links, rights, layout/overflow, accessibility, and source freshness. Evidence: parity gate tests.

    • S9.13.h Build/export/import locale bundle with source/target revisions, stable IDs, statuses, refs, provenance, hashes, and loss report; require round-trip identity. Evidence: bundle tests. <!-- S9.13: shipped as libs/oshun/workbench-kit/src/localization-jobs.ts + .spec.ts (153 tests), entry point ./localization-jobs (152 approved, no drift), plus in-place repairs in libs/oshun/studio-authoring/src/localization-workflow/.

      THE VEIN: A TRANSLATION IS A REWRITE PERFORMED BY SOMEBODY WHO CANNOT
      READ THE PARTS THAT MUST NOT CHANGE. Every localization judgement is
      about a SPAN, and it is one of two exactly opposite judgements -- this
      span must survive byte for byte, or this span must be REBUILT to the
      target language's own grammar -- and every check in this repository was
      asked about the WHOLE STRING, which can express neither. Three greps:
      which checks take the entire sourceText/targetText as their subject;
      which check that guards VERBATIM survival normalises first (case-folds,
      trims, de-duplicates), because that one has erased the property it was
      asked to certify; and which check requires the target to MATCH THE
      SOURCE'S STRUCTURE, for the spans where matching the source is the
      defect.
      
      Six instances, all measured before anything was written.
      `studio-authoring/.../glossary-editor.ts::enforceGlossaryInSegment` was
      three defects in five lines, all the first grep: it decided a
      do-not-translate term survived with
      `targetText.toLowerCase().includes(term.toLowerCase())`, so `OSHUN`
      against a German target reading `oshun ist bereit` answered `[]`; it
      matched as a substring with no boundary, so the term `Tara` reported
      `do-not-translate-violated` against the source `A tarantula appeared`;
      and it was count-blind, so `OSHUN, OSHUN, and OSHUN again` translated as
      `OSHUN und nichts weiter` also answered `[]`.
      `translation.ts::evaluateLocaleQA` measured expansion over the whole
      segment: `Save` to `Speichern unter` correctly warned at 3.75x, and
      appending the same 96-character URL to BOTH sides made the identical
      defect produce no finding at all. `locale-formats.ts` scanned the whole
      target, so `Bestellnummer 1,234,567 wurde storniert` was reported as a
      number-format leak whose repair rewrites an order number, and
      `https://example.com/docs/3/4/2026` as a retained US date whose repair
      breaks a URL. `translation.ts` also had no identity where a reference
      belongs: `lookupGlossary` took no scope, so a lineage-scoped term and a
      global one were told apart by ARRAY ORDER (measured: the same query
      answered `exercice` and `pratique contemplative` depending on which was
      pushed first); `TranslationMemoryEntry` had four fields, so a match
      returned another customer's sentence at similarity 1 with nothing saying
      whose it was; and a segment built `status: 'approved'` with
      `reviewedByCreatorId: null` scored `launchable: true, coverageRatio: 1`.
      `workbench-kit/src/localized-labels.ts` (S9.4) compares
      `new Set(extractPlaceholders(source))` against the target's set, and a
      set is not parity: the `#` inside a plural branch is invisible, so
      `{count, plural, one {# item selected} other {# items selected}}` and a
      German translation with no `#` at all both extract to exactly
      `["count"]`; the Set destroys multiplicity, so two `{city}` against one
      compare equal; and it strips `{{`/`}}` as the escape, which is Rust's
      and .NET's, while ICU quotes with an apostrophe. The third grep is asked
      nowhere: `Intl.PluralRules` gives the eight `libs/oshun/i18n` launch
      locales two categories for en-US and de-DE, three for fr-FR, pt-BR and
      he, one for ja-JP and SIX for ar, so parity with the source is exactly
      the rule that keeps an Arabic string wrong for 2, 3, 11 and 100 items --
      and the only parameterised customer strings the product produces are
      undeclared, `privacy/src/receipts-feed/receipts-feed.ts` emitting 23 copy
      keys across 18 parameterised messages with 26 parameter names of which
      the catalogue holds ZERO, while the catalogue's own 28 keys carry no
      substitution at all.
      
      TWELVE DECISIONS. A SEGMENT IS A SEQUENCE OF SPANS AND EACH SPAN HAS A
      DISPOSITION (verbatim, rebuilt, translated). PARITY IS A SEQUENCE OF
      KEYS WITH THEIR MULTIPLICITIES, NEVER A SET, with `#` counted per BRANCH
      and order compared only for the kinds consumed positionally, since
      demanding a named argument keep its position would forbid translating
      into Japanese. THE TARGET'S PLURAL CATEGORIES COME FROM THE TARGET, and
      an exact `=1` selector never stands in for the `one` category, which in
      Russian also matches 21. A VERBATIM CHECK NEVER NORMALISES, so a
      do-not-translate term is `exact` by construction and a term declaring
      otherwise is refused. A TERM OCCURS AT A PLACE, counted on
      `Intl.Segmenter` word boundaries so it works for Japanese, and never
      inside a protected span. PRECEDENCE IS A TOTAL ORDER OVER SIX SCOPES AND
      A TIE IS A REFUSAL. A MEMORY MATCH IS SOMEBODY ELSE'S SENTENCE AND
      ARRIVES WITH ITS OWNER ATTACHED, `unknown` rights being a refusal rather
      than a permission and the locale pair compared on the resolved SCRIPT so
      `zh-Hans` does not serve `zh-Hant`. A PROPOSAL CANNOT BE BORN APPROVED:
      the status is not an input and no transition reaches `approved` except
      from `in-review`. REVIEW IS PER LOCALE AND PER DIMENSION, and a
      credential for a LANGUAGE is not a credential for a LOCALE unless the
      policy says so in a field. A GATE ANSWERS PASS, FAIL OR UNEVALUABLE AND
      SILENCE IS NEVER A PASS. EXPANSION IS MEASURED OVER THE SKELETON. AND A
      LOSSY EXPORT AND A DEFAULTING IMPORT ARE ONE BUG SEEN TWICE -- the
      importer defaults nothing, an absent field is a fault, and a bundle with
      a non-empty loss report may be read and may not be imported back over
      what it came from.
      
      Scope stated: S9.13.a's "field/unit list" is the segmented unit list
      rather than a second field vocabulary, because S9.4's `LOCALIZED_FIELDS`
      already names the fields and S9.13.b's segmentation is what turns them
      into units. S9.13.c's "segment/hash" is the SKELETON digest, so a match
      implies the same protected spans and can actually be spliced; the raw
      text digest would agree for two texts whose URLs differ. S9.13.d's
      "morphology" is the declared surface-form list rather than a stemmer,
      because a stemmer that guesses is wrong in the language the term was
      added for. S9.13.g's "rights" gate re-reads the memory entry statuses at
      gate time and reports an unresolvable reference as a BLOCK rather than a
      pass; propagating a withdrawal to the derived work is S9.14's
      invalidation and is not claimed here. The four repaired
      `studio-authoring` functions duplicate the minimum span primitive the
      `layer:shared-logic` boundary forces (that package may not import
      `layer:workbench-kit`), the same way S9.4's RTL script set is duplicated
      there.
      -->
      <!-- S9.13 marked after writing and reading
      `libs/oshun/workbench-kit/src/localization-jobs.ts` end to end, running
      its 153 tests, mutating 128 of its decisions and confirming each turns a
      test red, and after running the repaired library's own suite
      (studio-authoring 192, of which 10 pinned the defects above and were
      rewritten, plus 11 new ones) and the whole workbench-kit suite (10418).
      -->
      
  • S9.14 Add cross-locale dependency invalidation. Evidence: source changes mark translated dependents stale.

    • S9.14.a Record dependency from each translated field/segment/bundle/ review/gate/release to exact source field/anchor/revision/hash and translation inputs. Evidence: dependency schema.

    • S9.14.b Classify source changes as content, structure, identifier, placeholder/markup, link, media, rights/consent, glossary/TM, policy, metadata, or nonsemantic. Evidence: change classifier tests.

    • S9.14.c Determine affected locales/segments/references/layout/parity gates/catalog/search/releases transitively without invalidating unrelated translations. Evidence: blast-radius tests.

    • S9.14.d Mark dependencies stale transactionally with source old/new revision, change reason/type, time, and event. Evidence: staleness tests.

    • S9.14.e Queue idempotent retranslation/review/parity/reindex/rebundle or require human action according to policy and cost. Evidence: workflow routing tests.

    • S9.14.f Block publication/use where stale translated content violates target policy and surface exact affected locales/segments. Evidence: eligibility tests.

    • S9.14.g Reconcile missed invalidation by comparing stored dependency versions to authoritative sources. Evidence: repair tests. <!-- S9.14: shipped as libs/oshun/workbench-kit/src/cross-locale-invalidation.ts (entry point ./cross-locale-invalidation, contract runtime, 153 approved entry points and no drift), with in-place repairs to libs/oshun/workbench-kit/src/localization-jobs.ts and to four shipped functions in libs/oshun/studio-authoring.

      THE VEIN. A DERIVED WORK IS A FUNCTION OF EVERY INPUT THAT WENT INTO IT,
      AND EVERY INVALIDATOR IN THIS REPOSITORY WATCHES THE ONE INPUT THAT
      SHARES ITS NAME. A translation is made from a source span, and it is
      also made from the glossary decisions that constrained its wording, the
      memory entries its sentences were assembled from, the rights those
      entries carry, the policy that judged it, the model version that
      produced it and the asset its alternative text describes. Staleness is
      computed against the source text; the other six are recorded on the
      derived work in full and read by nothing, so a withdrawn sentence stays
      published and certified. And the one input that IS watched is watched as
      a WHOLE, through a single digest, so the answer is the same boolean for
      every dependent: the two failures are opposite and simultaneous, a
      trailing space costing eight locales a full retranslation while a rights
      withdrawal costs nothing. Three greps find it. What is the SUBJECT of
      the staleness comparison -- is it one scalar standing for an entire
      upstream? List the inputs the derived record NAMES, then grep each name
      inside the staleness and gate code; the gap between the two lists is the
      set of ways the artifact can go wrong without anybody hearing. And which
      invalidator can be asked AGAIN -- every one of them is a push at the
      moment of a change, so a miss is permanent and invisible.
      
      SEVEN MEASURED INSTANCES, all run before anything was written.
      `studio-authoring/src/localization-workflow/translation.ts::detectStaleTranslations`
      over eight locale rows of one unit, with the source moved by a single
      trailing space, answers `stale count: 8 of 8`, every reason
      `source-text-changed-since-translation`, and its verdict has four fields
      -- `segmentId`, `state`, `stale`, `reason` -- of which none can say WHAT
      changed, so a corrected typo and a rewritten section cost the same eight
      retranslations. The same function over a unit whose memory entry was
      withdrawn that morning answers `{state:'fresh', stale:false,
      reason:null}` and `buildLocaleLaunchScorecard` downstream answers
      `launchable:true, coverageRatio:1`; `TranslationSegment` has nine fields
      and names no input except a digest of the source text, so there is no
      field on the row a withdrawal could ever have been matched against.
      `editorial-lifecycle/staleness.ts::detectStaleArtifacts` reads
      `sourceLastChangedAtUnixSeconds !== undefined`, so an artifact whose
      drift signal was never wired reports exactly like one whose source has
      not moved: measured with two published articles reviewed ten seconds ago
      -- one with no signal, one whose source changed in the review's own
      second, since the comparison is `>` -- the sweep returns `[]`, and
      `StaleArtifact` carries a reason named `source-drift` and no field
      naming which source. `asset-metadata/metadata.ts::buildAssetReplacementCascade`
      returns four fields whose invalidation is `input.asset.usedByArtifactIds`,
      with no old or new revision, no reason, no classification and no locale,
      so replacing a photograph with a re-encode of the same photograph and
      replacing it with a different photograph produce byte-identical
      cascades. `asset-metadata/approval-queue.ts::surfaceBrokenLinks` guards
      its rights alarm with `asset.usedByArtifactIds.length > 0`: measured, an
      asset whose licence expired four months ago raises one signal when a
      usage row happens to be present and ZERO when the list is empty -- and
      the list is empty for every asset in the product, because
      `usedByArtifactIds: []` in `asset-metadata/bulk-upload.ts` is the only
      assignment to that field in the repository, so a condition that reads as
      a filter is an off switch on the one class of change that cannot be
      deferred. `workbench-kit/localization-jobs.ts::runLocalizationGates`
      does re-read memory entry status at gate time and compares `status`
      only, while the `entryVersion` carried on the `MemoryMatchRef` beside it
      is compared to nothing anywhere in the module, so an entry that was
      CORRECTED -- same id, new version, still `active` -- passes the rights
      gate. And `glossaryRefs` is written by `admitTranslationProposal`,
      carried through `buildLocaleBundle`, validated by `importLocaleBundle`,
      and appears ZERO times inside `runLocalizationGates`.
      
      ELEVEN DECISIONS. A DEPENDENCY NAMES A PART OF A NAMED REVISION OF A
      NAMED OBJECT AND EVERY INPUT CLASS IS A DEPENDENCY -- seven input kinds,
      each carrying an identity and a VERSION, because a reference with no
      version can never be shown to have moved. AN ANCHOR THAT MEANS A
      POSITION IS NOT AN ANCHOR: `paragraph-2`, `3`, `node-1` are refused at
      admission, since inserting a unit above one of them rebinds it to
      somebody else's sentence without moving any digest it owns. A CHANGE IS
      A SET OF FACTS ABOUT PARTS, NEVER A LABEL ON A WHOLE, each fact naming
      its anchor, and one edit routinely produces several of different
      classes. `nonsemantic` IS A CLAIM ABOUT THE RENDERED RESULT UNDER A
      DECLARED NORMALIZATION AND THE ROLE DECIDES WHICH NORMALIZATIONS APPLY,
      so a whitespace fold is invisible in a paragraph and CHANGES THE
      MEASUREMENT of a capped `button-label`, `unicode-nfc` changes what a
      `keyword` matches, and a change touching a protected span is never
      nonsemantic however the prose compares. SILENCE IS NEVER A PASS, SO
      THERE IS A TWELFTH CLASS: `unclassified` is what a digest disagreement
      nothing else explains becomes, and it routes to a human and blocks under
      every tolerance, because a change nobody could describe is not one
      anybody can price. A BLAST RADIUS IS COMPUTED FROM THE PART AND CROSSES
      A LOCALE BOUNDARY ONLY WHERE THE TARGET SAYS IT DOES -- a fact reaches a
      dependency when the dependency's input covers the fact's anchor, and the
      same argument-reference edit is a parity re-check for `ja-JP`, which
      owes one plural branch, and a retranslation for `ar`, which owes six.
      THE ACTION IS PART OF THE FINDING AND THE CLASSES DO NOT SHARE ONE:
      retranslate, resegment, rebind, re-splice, re-check parity, re-check
      terminology, replace, re-gate, reindex, rebundle, re-review the
      alternative text, triage. AN IGNORED CHANGE STILL MOVES THE WATERMARK,
      with the reason recorded beside it, because a reconciler that keeps
      rediscovering an edit somebody already decided to ignore is a reconciler
      that gets switched off. AN IDEMPOTENCY KEY NAMES THE RESULT AND NEVER
      THE EVENT: automatic execution is a permission the policy grants per
      action and per locale, a locale with no policy has granted nothing, an
      unpriceable action is a human decision rather than a cheap one, and a
      rights change is never queued at all. AN ELIGIBILITY ANSWER NAMES THE
      UNITS, per locale, in three verdicts, with the blocked ones carrying
      exact unit ids rather than a count and `undecidable` reserved for a unit
      in scope that no dependency covers. AND A PUSH INVALIDATOR'S MISSES ARE
      INVISIBLE, SO THE ONLY HONEST ONE CAN BE ASKED AGAIN --
      `reconcileDependencies` compares stored input versions against the
      authoritative ones and reports a missed invalidation, an unresolvable
      input, a mark queued against a revision already superseded, an orphan
      mark and a consumed input no dependency names, and reports the COVERAGE
      of the snapshot it was given, because a sweep over half the inputs that
      finds nothing has found nothing about the other half.
      
      REPAIRS IN PLACE. `detectStaleTranslations` now takes the recorded
      non-source inputs and an input index and answers over both dimensions,
      with every finding naming its input and the two versions;
      `TranslationSegment.inputs` is `null` when nobody recorded them, which
      is `unverifiable` rather than `fresh`, and the index's `coverage` field
      is what tells an input that is GONE from one the index could not answer
      about. `detectStaleArtifacts` takes a list of sources rather than an
      optional timestamp, reports `drift-unverifiable` for an unwired
      artifact, compares with `>=` so an edit in the review's own second is
      flagged, and names the drifted sources. `buildAssetReplacementCascade`
      classifies the replacement against the perceptual hash
      (`identical-bytes` / `re-encode` / `different-media` / `unclassifiable`,
      the last because guessing cheap would waive an accessibility re-review
      nobody performed), carries both versions and a reason, and takes its
      dependents from an enumeration it was GIVEN with the coverage that
      enumeration can honestly claim. `surfaceBrokenLinks` no longer gates on
      the producerless usage list: an expired licence raises its signal and
      stays quiet only when a COMPLETE enumeration says the asset is unused.
      And the rights gate compares `entryVersion`, blocking when an active
      entry's version cannot be read and warning when it has moved.
      
      Scope stated: S9.14.b's `placeholder/markup` class is named
      `argument-markup`, keeping S9.13's vocabulary -- an argument reference
      is what the syntax holds -- and because the repository's own stub scan
      greps the other word. S9.14.c's `catalog/search` propagation is the
      `reindex` action on the `metadata` class rather than a second index
      model, because S9.11 already owns the index event and its projection.
      S9.14.e's cost is a per-action unit price against a per-locale budget
      rather than a scheduler, since the queue is the host's. S9.14.g's repair
      marks only the MISSED invalidations: an unresolvable input and an
      unrecorded dependency both need somebody to decide what the dependent
      should have pointed at, and inventing a mark for either would be the
      invalidator asserting a fact it does not have. The `TranslationInputRef`
      kinds duplicated in `studio-authoring` are the minimum the
      `layer:shared-logic` boundary forces (that package may not import
      `layer:workbench-kit`), the same way S9.4's RTL script set and S9.13's
      span primitive are duplicated there.
      
      Pre-existing and untouched: `generate-workbench-kit-api-reference.mjs`
      was already red at HEAD with 86 `undocumented-module` violations, one
      for every S9.x kit module since S9.4, and this module is the 87th of
      that same class; `check-conventions-ratchet.mjs` was already red at HEAD
      on `test-dot-test-suffix` for six committed `*.test.ts(x)` files, none
      of them touched here.
      -->
      <!-- S9.14 marked after writing and reading
      `libs/oshun/workbench-kit/src/cross-locale-invalidation.ts` end to end,
      running its 133 tests, mutating 138 of its decisions and confirming each
      one turns a test red, and after mutating the five repaired shipped
      functions 41 more times with the same result -- 179 mutations, 179
      caught. Suites run: workbench-kit 10554 across 149 files (of which
      localization-jobs 155, two of them new), studio-authoring 211 (of which
      16 new, six rewritten where they pinned the defects above, and the rest
      given the two fields the records grew).
      Gates run by hand: the stub-indicator scan over every changed file
      (zero hits), eslint, prettier, `tsc --noEmit` on both `tsconfig.lib.json`
      and `tsconfig.spec.json` in both libraries proved with a deliberate-error
      canary, and `generate-workbench-kit-export-map.mjs --check` (no drift,
      153 entry points, no violations).
      -->
      
  • S9.15 Add accessible calendar, catalog, and taxonomy UI primitives. Evidence: keyboard-only full journeys.

    • S9.15.a Build calendar agenda/list/grid views with semantic date/time/ zone/resource/slot/status/capacity/conflict data and synchronized selection. Evidence: calendar parity tests.

    • S9.15.b Provide keyboard create/navigate/select/move/resize/recurrence/ exception/resource/commit/cancel and exact form alternatives to drag. Evidence: keyboard calendar journey.

    • S9.15.c Build catalog table/list/collection/feature/rank/eligibility/ audience views with preview/diff/review/publish state. Evidence: catalog semantic tests.

    • S9.15.d Provide keyboard curation, ordering, targeting, eligibility inspection, preview, review, and bundle publication journey. Evidence: keyboard catalog journey.

    • S9.15.e Build taxonomy tree/list/search/detail/impact/diff/lifecycle/ translation views with nonvisual hierarchy and counts. Evidence: taxonomy semantic tests.

    • S9.15.f Provide keyboard add/rename/move/merge/split/alias/deprecate/ migrate/review/publish journey with cycle and impact feedback. Evidence: keyboard taxonomy journey.

    • S9.15.g Implement focus restoration, announcements, noncolor status, zoom/reflow, RTL/time-zone/locale, virtualization, conflict/offline/ degraded states across all three. Evidence: cross-primitive a11y matrix. <!-- S9.15: shipped as libs/oshun/workbench-kit/src/domain-views.ts (entry point ./domain-views, 154 approved, no drift), apps/oshun/web/src/components/workbench/{CalendarViews,CatalogViews, TaxonomyViews}.tsx + module CSS, and apps/oshun/web/tools/domain-views-browser.mjs.

      THE VEIN: A VIEW IS A SAMPLE AND EVERY SENTENCE IT WRITES IS ABOUT THE
      POPULATION. Three greps -- find the count and ask whose set it is over;
      find the CONTROL whose options are derived from the rows it happens to
      hold; find the field that NAMES the omission and grep it in the render.
      Eight measured instances across four shipped surfaces, all repaired in
      place. `StudioTaraWorkbenchCalendar.tsx` declared `ProgramEntry.clipped`
      and read it NOWHERE while its lane summary counted the truncated slot
      list and printed it as a number; declared `reservations` and read that
      nowhere either, so a whole lane of locked editorial commitments was
      fetched and never drawn by a component whose own comment said "every
      dated item"; bucketed days with `toISOString().slice(0, 10)`, which is
      the UTC day and is the reader's day in exactly one zone, and took its
      window from `getUTCFullYear()` on a LOCAL `new Date()`, so "This month"
      was the wrong month on the last day of one in a positive offset; and
      `reservedByCycle.get(cycle) ?? 0` printed "reserves 0 commitments" for a
      cycle the planner never answered about.
      `StudioTaraWorkbenchCatalog.tsx` built its category, teacher and series
      dropdowns from the rows the SERVER had already filtered, so choosing a
      category collapsed that control to one option and the door locked behind
      the reader, while a browser-side teacher filter could stay ON with no
      control anywhere showing it; and it printed `{filtered.length} of
      {state.total} items`, two denominators in one sentence.
      `StudioTaraWorkbenchCollectionEditor.tsx` wrote `position: index` to the
      server for a `curated` collection, drew the rank as `{index + 1}.` in a
      span, and offered NO way to change it -- not a keyboard way, no way.
      `StudioTaraWorkbenchCategories.tsx` opened its editor as a form appended
      below the whole list, moved focus nowhere and announced nothing.
      
      Decisions: EVERY COUNT IS AN S9.3 `ImpactCount` AND NEVER A NUMBER,
      rendered through `formatCount`. A VIEW'S POPULATION IS RETURNED WITH ITS
      ROWS, computed by the same pass that placed them, so the count cannot
      disagree with the drawing, and the seven `POPULATION_LIMITS` are derived
      from the refusals rather than declared. A SHORTFALL THE READER CAN UNDO
      IS AN OFFER AND ONE THEY CANNOT IS A DISCLOSURE. A FACET BUILT FROM THE
      RESULT AND APPLIED AT THE SERVER IS A ONE-WAY STREET, and a selected
      value absent from its own option list is a filter that is on and
      invisible. PARITY IS NOT "THE SAME ROWS APPEAR": every entry leaves a
      projection through one of two doors, and an entry every view refuses is a
      commitment no reader can reach. AN INSTANT IS NOT A ZERO-WIDTH INTERVAL
      -- the half-open overlap test compares a point against itself, so a due
      date sitting exactly on the window's start is the one item the range says
      it does not contain. A ROW IS NOT AN ENTRY. AN UNDATED ROW HAS NO DAY. A
      DAY IS READ IN A NAMED ZONE AND WALKED AS A CALENDAR DATE. A SELECTION
      THAT SURVIVES A VIEW SWITCH IS THE DANGEROUS HALF, so a member the target
      view cannot address is WITHHELD -- kept, counted in the action scope and
      disclosed -- and DROPPED only when the row is gone. A CATALOG ROW UNDER
      REVIEW IS TWO ROWS. AN ORDER OPERATION IN A FILTERED VIEW IS AMBIGUOUS,
      so a relative move is refused with the absolute alternative named.
      `aria-setsize` HAS ITS OWN UNKNOWN AND IT IS -1. A SEARCH VIEW DROPS THE
      HIERARCHY. A CYCLE REFUSAL NAMES THE ANCESTOR and a committed key-meaning
      change names its blast radius, both against S8.1's sixty-second ceiling.
      AND THE MATRIX IS ELEVEN CONDITIONS TIMES THREE FAMILIES, thirty-three
      cells, where the eight a browser can settle may not be answered by a
      declaration.
      
      THE BROWSER FOUND FOUR THINGS NOTHING ELSE COULD. A ROVING TABINDEX
      WITHOUT A COMPOSITE ROLE is a set of rows most of which are not in the
      tab order with nothing telling a reader to press an arrow -- measured as
      one reachable entry out of five, and fixed by making the panel a
      multi-select listbox whose reachability proof is the ARROW walk. A KEY
      THAT IS A SPACE CANNOT SURVIVE A SPACE-JOINED ATTRIBUTE, so `data-keys`
      is JSON. `inline-size: 100%` ON A PADDED, BORDERED ROW OVERFLOWS ITS
      PARENT BY EXACTLY THE PADDING PLUS THE BORDER -- 326 against 304 at 320px
      and 1286 against 1264 under `dir="rtl"`, the same 22px both times. And
      the field a reader sees and the value a form submits came from two
      places, so the absolute-position control submitted an empty string.
      
      Scope stated: the calendar family models a `CalendarEntry` rather than
      requiring S9.5's `CalendarSlot`, because an agenda shows program windows,
      dossier due dates and agent cycles beside bookings; `entryFromSlot` is
      the composition point so a slot's decisions are not made twice. S9.15.b's
      move and resize reach S9.7's `proposePlacement` through the host rather
      than being re-decided here. S8.1's `treeKeyboardFaults` already owns the
      generic tree keyboard pattern, so the taxonomy family adds only the
      operations S9.15.f names that a generic tree does not have. -->
      <!-- S9.15 marked after writing and reading
      `libs/oshun/workbench-kit/src/domain-views.ts` end to end, running its
      224 tests, and mutating 203 of its decisions -- 203 caught, nineteen of
      them only after the spec gaps the first pass exposed were closed and four
      equivalent mutants were re-aimed. The fourteen in-place repairs to the
      four shipped studio surfaces were mutated back to what shipped: 14 of 14
      caught. 217 mutations, 217 caught.
      Suites run: workbench-kit 10779 across 150 files (of which
      `domain-views` 224), and 55 jsdom tests across
      `DomainViews.spec.tsx`, `StudioTaraWorkbenchCurationRepairs.spec.tsx` and
      `StudioTaraWorkbenchCalendar.spec.tsx`.
      Real Chromium: `apps/oshun/web/tools/domain-views-browser.mjs`, 69
      controls -- three keyboard-only journeys plus the 33-cell matrix, every
      cell measured -- with both dist-patch directions verified (patching
      `commandFaults` in `dist` turned three verdicts red, patching
      `reachabilityFaults` turned another red, and restoring returned all 69).
      Gates run by hand: the stub-indicator scan over every changed file (zero
      hits beyond pre-existing `// stub:legitimate` input placeholders),
      eslint, prettier, `tsc --noEmit` on `tsconfig.lib.json`,
      `tsconfig.spec.json` and `apps/oshun/web/tsconfig.wbcheck.json` -- all
      three proved with a deliberate-error canary, and the four repaired studio
      components ADDED to `tsconfig.wbcheck.json`, which is the first time they
      have been typechecked at all -- and
      `generate-workbench-kit-export-map.mjs --check` (no drift, 154 entry
      points, no violations).
      Pre-existing and untouched: `generate-workbench-kit-api-reference.mjs`
      was already red at HEAD with 87 `undocumented-module` violations, one for
      every S9.x kit module since S9.4, and this module is the 88th of that
      same class; `tools/conventions/check-conventions-ratchet.mjs` was already
      red at HEAD on `test-dot-test-suffix` for six committed `*.test.ts(x)`
      files, none of them touched here, and every file added by this item is
      `.spec.ts(x)`. -->
      
  • S9.16 Add exportable calendar and catalog audit snapshots. Evidence: stable schema and authorization tests.

    • S9.16.a Define versioned snapshot envelope with type/scope/query/time window/target, source versions, generated actor/time, schema, redaction policy, counts, pagination completeness, and hash. Evidence: envelope schema.

    • S9.16.b Export calendar resources/capacity, slots/recurrence/exceptions, states, dependencies, blackouts/embargo, conflicts, actors, changes, and audit refs. Evidence: calendar snapshot fixtures.

    • S9.16.c Export catalog/collection/relation/feature/rank/eligibility/ audience, bundle/release, target verification, correction/withdrawal, actors, changes, and audit refs. Evidence: catalog snapshot fixtures.

    • S9.16.d Apply tenant/object/property/protected/embargo/audience role and purpose authorization at query and serialization. Evidence: ordinary/ auditor cross-tenant negative tests.

    • S9.16.e Redact or pseudonymize content/identity according to snapshot purpose while preserving stable internal refs only for authorized exports. Evidence: canary data tests.

    • S9.16.f Generate deterministically with stable order/format/time semantics and sign/checksum immutable output. Evidence: reproducibility tests.

    • S9.16.g Verify import/read tooling, schema compatibility, complete page capture, retention/legal hold, and audit of export/download. Evidence: snapshot lifecycle tests. <!-- S9.16: shipped as libs/oshun/workbench-kit/src/audit-snapshots.ts (entry point ./audit-snapshots, 155 approved, no drift), two committed goldens with pinned canonical digests under libs/oshun/workbench-kit/schemas/audit-snapshot-{calendar,catalog}.golden.json, and in-place repairs to apps/oshun/bff/src/routes/admin-audit-log.ts, apps/oshun/bff/src/admin/admin-audit-events-store.ts, apps/oshun/bff/src/routes/tenant-console-reads.ts, libs/oshun/tenant-console/src/bulk-ops/bulk-export.ts, apps/oshun/bff/src/routes/tara-workbench.ts and apps/oshun/web/src/components/studio/StudioTaraWorkbenchCalendar.tsx.

      THE VEIN: A VIEW IS READ BY SOMEBODY WHO CAN ASK AGAIN AND A SNAPSHOT IS
      READ BY SOMEBODY WHO CANNOT. S9.15's reader is present at the moment of
      the read and can widen the filter, press page two or notice the spinner;
      this reader is a regulator, an auditor or the same operator in eighteen
      months, and the bytes are the only witness they will ever have. So every
      question they will ask has to be answerable FROM THE DOCUMENT, including
      the two a live view never has to answer: is this all of it, and what was
      the question. Three greps -- find every LIMIT on the export path and ask
      what in the output discloses it; find every predicate that decided
      membership and ask whether the document repeats it; find every number,
      hash and instant and ask whether a reader can re-derive it from the
      bytes. Ten measured instances across six shipped surfaces.
      `admin-audit-log.ts`'s investigation export called
      `listAcrossOperators({ ..., limit: 500 })` and wrote
      `## Matched events (${events.length})` into the Markdown an investigator
      downloads, so the size of the SAMPLE was printed as the size of the
      population with no hasMore, no total, no second page and no parameter
      that would fetch one -- and the audit event the export itself writes
      recorded `matchedEventCount: matched.length`, which is 500 again. It
      truncated from the WRONG END, because the store sorts
      reverse-chronologically before slicing, so what fell off was the
      beginning of the incident. It resolved bookmarks with
      `.map(readEvent).filter(e => e !== null)` and rendered the bookmark
      section only `if (bookmarkedEvents.length > 0)`, so an investigation
      whose ten bookmarks had aged out of the ring buffer exported a document
      in which "there were none" and "there were ten and they are gone" were
      the same bytes. `recordInvestigationAuditEvent` put the DECLARED
      bookmark count in its base payload and the export's extraPayload
      overwrote it with the RESOLVED count, so one field name carried two
      populations depending on which action wrote the row -- and a shipped
      test asserted `bookmarkedEventCount: 0` for an investigation with one
      bookmark, locking that in. The Markdown carried the title, two dates and
      a list of events and NOT THE FILTER, so the one artefact that leaves the
      building could be neither reproduced nor challenged; it carried no
      generated instant while the JSON did, and neither carried a checksum.
      `listAcrossOperators` ordered by
      `b.occurredAt.localeCompare(a.occurredAt)` before slicing, which is both
      S6.5's locale-dependent comparator and S9.9's
      `limit-over-non-total-order` -- a limit over a tied order does not
      truncate a list, it CHOOSES between the tied rows by arrival order.
      `tenant-console-reads.ts` read `listAcrossOperators({ limit: 5000 })`,
      filtered the result to this tenant and sliced again: the limit clamps to
      500, so the filter ran over the 500 most recent events across EVERY
      operator, and a tenant with hundreds of events saw an empty cockpit
      that is indistinguishable from having none -- A FILTER APPLIED TO A PAGE
      IS A FILTER OVER THE WRONG POPULATION, so the tenant predicate is now a
      field on the store's query.
      `@oshun/tenant-console`'s `bulk-export.ts` sorted the records inside
      every exported file and the file list inside the manifest with
      `localeCompare`, and both feed a sha256, so the integrity manifest of a
      data export was a function of the exporting process's LANG; and its
      `verifyManifest` walked `manifest.files` checking each was present with
      the right bytes, so a file in the download and absent from the manifest
      verified clean, while `totalRecordCount` was never compared against the
      sum of the per-file counts -- the chunks were never required to TILE the
      population. `tara-workbench.ts`'s calendar feed enumerated agent cycles
      with `for (let i = 0; i < 366 && cursor <= toUnix; i += 1)` and said
      nothing when the cap bit, and `StudioTaraWorkbenchCalendar.tsx` printed
      that lane's size as an `exact` count.
      
      Decisions: A COUNT IS A TRIPLE AND NEVER A NUMBER -- matched as an S9.3
      `ImpactCount` because a truncated read honestly knows only `at-least`,
      emitted as what is in the bytes, withheld as the shape of the hole
      itemised by reason -- and `countAccounts` is the invariant tying them
      together. COMPLETENESS IS A PROOF AND NOT A FLAG: the envelope carries
      the PAGES with their row counts and their start and end cursors, and the
      claim is re-derived rather than trusted, so a page fetched from
      somewhere the previous page did not end is a torn read whose missing
      rows are in no page at all. THE WHOLE DOCUMENT IS READ AS OF ONE MOMENT
      OR IT IS NOT A SNAPSHOT, which is the other side of S4.9's admission
      that a keyset traversal without one may miss and repeat rows and of
      S1.5's note that offset paging is correct only over a frozen snapshot.
      AN OMITTED SECTION AND AN EMPTY SECTION ARE DIFFERENT FACTS, so absence
      is a value with a reason. A SNAPSHOT IS A CLOSED WORLD: an id it
      mentions and does not contain is a question its reader cannot go and
      ask, because the system has moved on. AUTHORIZATION IS ASKED TWICE AND
      THE SECOND TIME IS THE ONE PEOPLE SKIP, because a snapshot is
      DENORMALISED and every embedded actor, conflicting slot and audience
      definition is a second read the query's admission check never saw. THE
      TARGET IS NOT THE SCOPE -- the scope decides which rows are selected and
      the target decides how much of each may be shown, whether two exports
      may be joined, and who has to be named when the document turns up
      somewhere it should not be. A PSEUDONYM IS A JOIN KEY, so the salt is
      per-snapshot unless a named authority approved linkage; and one subject
      under two codenames splits a person while one codename over two subjects
      MERGES them, which is not a loss of resolution but a false statement
      about conduct. THE HASH COVERS THE DOCUMENT AND NOT A HAND-WRITTEN LIST
      OF ITS FIELDS. AND A HOLD OUTRANKS AN EXPIRY WHILE A DOWNLOAD IS ITS OWN
      EVENT: a retention schedule that wins over a legal hold destroys
      evidence on time, an operator who can release the hold on the evidence
      about them has a delete button with extra steps, and generating a
      document is a decision while handing over the bytes is the disclosure.
      
      Scope stated: the module is `contract`-class, so the stores, clocks,
      hash functions and readers it is decided against are the host's; S9.9's
      `orderFaults`/`orderIsTotal`/`compareCodePoints`, S8.12's embargo
      standing, exposure and audience ceilings, S9.5's `admitInterval` and
      S9.3's `ImpactCount` are composed rather than re-decided. The audience
      ceiling needed one named exception: `learner-record` tops out at
      `object-collaborators` and `auditors` ranks above it on a ladder that
      counts PEOPLE, so `regulatory-audit` and `legal-discovery` license that
      one band and nothing wider. -->
      

S10 — Observability, reliability, privacy, and cost#

  • S10.1 Define shared OpenTelemetry resource and span attributes. Evidence: schema lint rejects unbounded content attributes.

    • S10.1.a Define resource attributes for service/library, version, domain/ workbench, host/runtime, environment, region, deployment, worker/ provider class, and tenant-safe scope. Evidence: resource schema.

    • S10.1.b Define span naming and attributes for UI navigation/action, HTTP/ RPC, command/query, database, outbox/inbox/event, job, provider/model, blob/bundle, review/gate, publish/verify, and bridge. Evidence: span schema.

    • S10.1.c Define status/error/retry/cancel/timeout/partial/unknown outcome, correlation, version, target, usage, and cost attributes with units. Evidence: semantic convention tests.

    • S10.1.d Allowlist bounded values and explicitly prohibit raw IDs, content, prompts, URLs/paths, filenames, stack/error messages, PII, protected identity, biometric/learner data, and secrets. Evidence: attribute policy.

    • S10.1.e Set cardinality and length budgets, unknown/other behavior, and schema version/deprecation ownership. Evidence: lint configuration.

    • S10.1.f Add compile/runtime lint and canary-data tests that reject unregistered or unbounded attributes. Evidence: deliberate violation test. <!-- S10.1: shipped as libs/oshun/workbench-kit/src/telemetry-attributes.ts (entry point @oshun/workbench-kit/telemetry-attributes, contract-class, 156 approved entry points, no drift) with a 193-test spec, plus repairs to @oshun/tracing and a 14-test regression spec (libs/shared/tracing/src/attribute-hygiene.spec.ts).

      THE VEIN: WHAT AN ATTRIBUTE COSTS AND WHAT IT DISCLOSES ARE BOTH DECIDED
      BY ITS VALUE, AND EVERY CHECK IN THIS REPOSITORY WAS WRITTEN ABOUT ITS
      NAME. A name is what the author chose; a value is what the world
      supplies. And the same value costs different amounts depending on where
      it lands: on a span attribute a distinct value is an ADDEND -- one field
      on one record, dropped when the span ages out -- while on a metric
      dimension it is a MULTIPLIER, a new time series against every other
      dimension, retained for the metric's whole retention and paid for at
      every collection interval whether or not anything happened; on a
      resource it multiplies every signal the process emits. The API is
      `setAttribute(key, value)` in all three places and the economics are
      three orders of magnitude apart, so "is this attribute bounded" is not a
      question about the attribute, it is a question about the PAIR. Telemetry
      is also the only output in this system with no boundary check -- a
      response has a schema, a write has a column type, S9.16's export has an
      envelope, and a span has the backend creating a column for whatever
      arrives -- so the producer is the only enforcement point that exists,
      and the producer is every file in the repository.
      
      Twelve measured instances across five shipped surfaces.
      `libs/shared/tracing/src/otlp-metrics.ts` opens with "Real OTLP/HTTP
      metric export for BOUNDED, metadata-only metric points" and then appends
      `'oshun.observed_at_unix_ms': new Date(point.observedAt).getTime()` to
      every point's dimension set -- a distinct value per observation, so
      every point exported creates its own time series, forever, added by the
      export path where no caller asked for it and no caller can turn it off.
      A histogram with a unique dimension per observation is a log with a
      worse query language and a longer retention. The same file's
      `validatePoint` is the repository's ONLY content prohibition and tests
      `/(conversation|transcript|...|prompt|source\.body)/iu` against
      `Object.keys(point.attributes)`, never the value, so `{ q: <the text the
      user typed> }` is admitted and `{ 'queue.message_count': 5 }` is refused
      -- unsound in one direction, incomplete in the other, and defeated by
      renaming in either. That prohibition is never applied to the RESOURCE,
      which is the set attached to EVERY point, and `...input.resourceAttributes`
      is spread LAST after `service.name`, so a caller can relabel its
      telemetry as another service's -- identity is the one thing a caller
      must not assert about itself. The same file keys its instrument cache on
      `${point.name}:${point.unit}`, so one metric recorded once as `ms` and
      once as `s` does not fail but FORKS into two instruments with one name
      whose values differ by a thousand; and it calls `meter.createHistogram`
      for every point because `OtlpMetricPoint` has no field in which any
      other instrument kind could be named -- the kind is not chosen wrongly,
      it is UNSTATABLE, which is the S8 family again.
      `libs/shared/tracing/src/middleware/hono.ts`, on the real BFF request
      path, sets `'http.url': c.req.url` -- the full URL INCLUDING THE QUERY
      STRING, so every search term, filter, reset link and signature ever
      passed as a query parameter is copied verbatim to a vendor that retains
      it on a schedule nobody here controls and indexes it for staff who were
      never granted the tenant. It sets `'net.peer.ip'` from
      `x-forwarded-for`, which is personal data in its own right and
      attacker-supplied besides, alongside `http.user_agent` from an
      eight-kilobyte header and `request.id` from a CLIENT-SUPPLIED
      `x-request-id`, so the length of that attribute is chosen by whoever is
      calling. Its `generateSpanName` falls back to
      `${c.req.method} ${url.pathname}`, putting a concrete path into the
      primary grouping key of every APM product -- one operation per object --
      at the exact place a template was in scope. That middleware also
      declares `recordRequestBody`, `recordResponseBody` and `maxBodySize`
      ("default: 10KB") and reads NONE of the three: the bodies were never
      recorded, which is the safe outcome arrived at by accident, and an
      operator who wrote `recordRequestBody: false` had turned off something
      that was never on while believing they had made a decision. Raw error
      text reached attributes in six places (`hono.ts` status message,
      `tracer.ts` twice including `exception.stacktrace`, `decorators.ts`
      twice, `middleware/index.ts` once), and an exception message is a string
      assembled at the failure site out of exactly the values that were in
      scope when things went wrong. And
      `libs/oshun/analytics/src/tracing-manifest.ts` says downstream SDKs
      "compile this manifest into their span builders", that
      `resolveSpanAttributes` is "used by tracers to validate emitted
      attribute sets at boot time" and that `isPii` is "used by sanitizers",
      while NOTHING outside the file imports any of them -- six hundred lines
      of convention with no reader is not a schema, it is a document that
      happens to compile; and its `type: 'string'` vocabulary has no field in
      which an enumeration could be written, so BOUNDED -- the one word in
      this item's evidence line -- is inexpressible in it even if somebody did
      read it.
      
      Decisions. ADMISSION IS ASKED ABOUT A PAIR AND NEVER ABOUT AN ATTRIBUTE,
      because `tenant.id` is correct on a span, ruinous as a dimension and
      false on a resource shared by many tenants. BOUNDED IS A STATEMENT ABOUT
      A SET, SO THE REGISTRY HAS TO BE ABLE TO WRITE THE SET DOWN: enum,
      boolean, quantity with a unit, bucketing with its boundaries, template
      with its parameters, pattern, opaque identifier, pseudonym or free text
      -- and only the ones with a computable cardinality may be a dimension.
      CARDINALITY IS A PRODUCT AND BUDGETS ARE SPENT BY THE METRIC, since four
      dimensions of five, eight, three and twelve are 1,440 series and each
      looked modest in review. A DISTINCT-VALUE COUNT IS AN S9.3
      `ImpactCount`, because coercing an unknown to a number is the arithmetic
      that keeps an estimate small right up until the bill arrives. A RESOURCE
      IS BOUNDED BY THE FLEET AND A DIMENSION BY ITS DOMAIN -- which is why
      `service.instance.id` is correct on one and ruinous on the other -- so
      the resource's real question is not "is this small" but "is this fixed
      for the lifetime of the process". FOLDING TO `_other_` IS A TRUNCATION
      AND AN UNDISCLOSED TRUNCATION IS A FALSE STATEMENT ABOUT THE POPULATION
      (S9.15, S9.16), so folds are COUNTED -- and the ledger that counts them
      is itself capped, because a ledger keeping every distinct folded value
      in order to count them grows exactly as fast as the cardinality it
      exists to prevent, which is why its distinct count is `at-least`. THE
      SPAN NAME IS AN ATTRIBUTE WITH NO KEY, which is how a concrete path
      survives in a library that has a prohibition list. AN IDENTIFIER IS A
      TOKEN AND A SENTENCE IS NOT A TOKEN, because an opaque-id slot accepts
      any string by construction and is otherwise the widest hole in the
      schema -- found by the regression tests, not by design. AND A
      PROHIBITION IS TESTED AGAINST VALUES WITH CANARIES, POSITIVE AND
      NEGATIVE, since a corpus of nothing but known-bad strings is passed
      perfectly by a detector that refuses everything, and such a detector is
      indistinguishable from a working one until it has refused enough
      legitimate traffic that somebody switches the lint off.
      
      Repairs shipped in `@oshun/tracing`: the per-observation timestamp
      dimension deleted; the content prohibition extended to VALUES and to the
      resource; identity sealed by spreading caller attributes first; the
      instrument cache keyed on the name alone with unit and kind
      disagreements refused as conflicts; `OtlpInstrumentKind` made statable;
      `http.url` and `net.peer.ip` removed entirely and replaced by the route
      SHAPE via a new `templatizePath`; user agent truncated and
      `x-request-id` accepted only as a bounded token; the three
      never-read body controls replaced by one that is read; and every raw
      error message and stack trace replaced by `errorClassName`. One shipped
      test asserted the defect (`status.message` `toBe('Test error')`) and now
      asserts the class.
      
      Verification: 193 kit tests, 88 tracing tests, 0 typecheck errors, no
      export-map drift. A 22-mutation sweep over the kit module caught 19; the
      three survivors were each proven EQUIVALENT rather than untested by
      mutating the redundant partner (removing the sealed-key rejection is
      caught; making the shared-regex state observable by breaking the scan
      loop early is caught; the FNV per-iteration `>>> 0` is equivalent by
      congruence mod 2^32, since every subsequent `^` and `<<` re-coerces).
      
      Scope stated: the module is `contract`-class, so the tracers, meters,
      clocks and exporters it is decided against are the host's; S9.3's
      `ImpactCount` and S9.9's `compareCodePoints` are composed rather than
      re-decided. The `@oshun/tracing` repairs are self-contained rather than
      importing the kit, because `libs/shared` is the lower layer and must not
      depend on `libs/oshun`. -->
      
  • S10.2 Instrument browser-to-worker trace propagation. Evidence: one synthetic journey has an unbroken trace.

    • S10.2.a Create/continue W3C trace context at trusted browser/desktop ingress and validate sampling/baggage propagation without accepting spoofed internal identity. Evidence: ingress tests.

    • S10.2.b Propagate through web navigation/action, BFF request, actor/auth/ authorization, domain service, database/blob/cache/search, and response. Evidence: synchronous trace tests.

    • S10.2.c Persist trace/correlation in outbox/event/inbox/job parent-child/ attempt and restore context in worker. Evidence: asynchronous tests.

    • S10.2.d Propagate allowed context to provider/model/tool, publisher, callback/webhook, DCC/engine/bridge, and read-back verification. Evidence: external boundary tests.

    • S10.2.e Link audit, job, provider, artifact, bundle, release, and external refs as span attributes/events without high-cardinality metrics. Evidence: trace evidence tests.

    • S10.2.f Preserve parent/link semantics across retry, fan-out/fan-in, queue delay, batch, cancellation, unknown outcome, and compensation. Evidence: topology tests.

    • S10.2.g Run a synthetic browser-to-worker-to-publisher journey and assert no broken segment, wrong tenant context, or sensitive baggage. Evidence: captured trace. <!-- S10.2: shipped as libs/oshun/workbench-kit/src/trace-continuity.ts (entry point @oshun/workbench-kit/trace-continuity, contract-class) with a 131-test spec, plus repairs to @oshun/tracing, @oshun/gpu-dispatcher, @oshun/bff-kit and @lilith/service-lib and a 19-test regression spec (libs/shared/tracing/src/propagation-continuity.spec.ts).

      THE VEIN: A TRACE IS A CLAIM ABOUT CAUSATION AND IT IS MADE OF EDGES,
      NOT SPANS, AND EVERY CHECK IN THIS REPOSITORY IS ABOUT A SPAN. Every
      span can be perfectly formed -- correct name, bounded attributes, all of
      S10.1 satisfied -- and the trace still be false, because what makes two
      spans one trace is the EDGE between them, and an edge exists only where
      the context survived a medium that does not natively carry it: an HTTP
      header, a queue message attribute, a database column, a webhook body, a
      retry, a four-hour delay. In process the edge is free and invisible;
      across any of those it is a field somebody had to write and somebody
      else had to read. AND A MISSING EDGE RAISES NO ERROR -- it does not look
      like a failure, it looks like two well-formed traces. Nothing throws,
      nothing is refused, no dashboard turns red, and the only symptom is that
      during an incident the worker's work is not under the request that asked
      for it, discovered by the person who needed it at the moment they needed
      it. So the grep is not "is there a tracer"; it is: for every
      asynchronous hop, NAME THE FIELD THE CONTEXT RIDES IN, then show that
      field has both a writer and a reader.
      
      Ten measured instances. `EventContext.traceParent` (`outbox.ts`,
      `quarantine.ts`) is documented "W3C traceparent, or null when the
      mutation had no trace" and has two declarations, ZERO production writers
      and ZERO production readers; twelve of its fourteen occurrences are in
      specs and seven of those set it `null`, and `quarantine.spec.ts` puts
      `'00-trace-span-01'` in it -- which is not a traceparent, and nothing
      objected because nothing parses it. `@oshun/gpu-dispatcher`'s
      `tracePoll` carries the comment "Try to link to parent trace if
      available" and then writes `'trace.linked_trace_id'` as a string
      ATTRIBUTE, starting the span with neither parent nor link, so the poll
      is the root of a NEW trace and the connection exists only as text --
      THE EDGE WAS DESCRIBED RATHER THAN MADE -- while `createJobTraceLink`,
      twelve lines further down the same file, already built the correct
      structure and had no caller. The same library's
      `injectTraceContextToJob` returns the job input UNCHANGED when there is
      no active span, so the enqueue succeeds, the job runs, and the edge is
      absent with no marker anywhere.
      `LightweightTracer.extractContext(_carrier)` returned
      `context.active()` and never looked at its argument, while implementing
      `Tracer`, being exported from the package root, and being what the Hono
      middleware calls on the real BFF request path -- two implementations of
      one interface disagreeing about whether propagation happens, with the
      type system calling them substitutable; and its `injectContext` wrote
      `00-${traceId}-${spanId}-01`, hardcoding the sampled flag (a trace
      sampled OUT telling every downstream service it was sampled IN) and
      erasing the seven reserved bits, and never wrote `tracestate` or
      `baggage` at all. `@oshun/bff-kit`'s `residency-guard` wrote
      `traceId: extractTraceId(request)` where `extractTraceId` returned the
      WHOLE `traceparent` header into a field called `traceId`, so every audit
      row it wrote carried a superstring of the key everything else joins on
      -- the one record whose entire purpose is to be found later joined to no
      trace at all -- with `req_${Date.now()}` as the fallback.
      `@lilith/service-lib`'s `health-response` returned
      `ctx.spanId ?? traceHeader.slice(0, 16)` as `x-span-id`: a span id no
      span ever had, derived deterministically from the trace id so every
      request in that trace lacking one gets the SAME value, two lines below
      `randomBytes(16)` minting a trace id at RESPONSE time. `parseTraceState`
      enforced no member cap, no byte cap and no grammar while
      `updateTraceState` prepended and kept everything (one member per hop,
      forever); `parseBaggage` enforced none of W3C's 180/8192/4096 limits and
      called `decodeURIComponent` unguarded, so `baggage: %=1` threw URIError
      inside the propagation path. And `tracing-manifest.ts` declares a
      `propagation` channel list per span family -- exactly the right subject
      -- which nothing outside that file imports.
      
      DECISIONS. THE UNIT IS THE EDGE AND EVERY EDGE NAMES ITS CARRIER, so
      "we propagate context" cannot be written down without saying in what.
      ABSENCE IS ONLY EXPRESSIBLE AGAINST A DECLARATION -- a verifier over the
      edges it was given can report that they are malformed and can never
      report that one is missing, because absence has no representative in the
      data -- so the journey is DECLARED first and `verifyJourney` reports
      declared legs with no observed edge, which is the difference between
      "we found no broken edges" and "we found no edges". THE FIRST EDGE IS
      CHOSEN BY A STRANGER, SO IT IS CONTINUED AND DISTRUSTED AT THE SAME
      TIME: the trace id is honoured (S4.13's asymmetry, composed rather than
      re-decided), the SAMPLED FLAG is not, because it is not an identifier
      but a SPEND INSTRUCTION, and when refused the decision falls to
      consistent trace-id-ratio sampling so every service computes the same
      answer without communicating. A REFUSAL TO CONTINUE IS ITSELF AN EDGE --
      a trace already recorded under another tenant restarts AND LINKS, since
      continuing would join two tenants' journeys and in a trace UI a join is
      a disclosure. THE DEFERRED EDGE INVERTS PARENTHOOD: a job enqueued now
      and run in four hours is not a CHILD of the enqueue, because the parent
      ended hours ago and a child outliving its parent makes the parent's
      duration -- the number every latency dashboard is built on --
      meaningless; the answer is a LINK, and the same holds for retry
      (attempts link to the intent, they do not nest) and fan-in (a child has
      exactly one parent and a join has several). A FABRICATED JOIN KEY IS
      WORSE THAN A MISSING ONE, because a null is a visible hole and a minted
      one is a hole that reports itself as full. BAGGAGE IS THE ONLY CONTEXT
      THAT REACHES EVERY DOWNSTREAM SERVICE INCLUDING ONES WE DO NOT OWN, so
      its audience ceiling is per-key and enforced per-HOP (the tenant
      pseudonym is fine on the way to our worker and a disclosure on the way
      to a model provider, which one decision at ingress cannot express) --
      and WHO MAY SET A KEY IS A DIFFERENT QUESTION FROM HOW FAR IT MAY
      TRAVEL, which is where S10.2.a's "without accepting spoofed internal
      identity" lives. CAPS ARE PART OF THE FORMAT, not an optimisation.
      
      Repairs shipped: `LightweightTracer.extractContext` now extracts;
      `injectContext` writes the real flags and forwards tracestate; the
      lightweight span inherits its parent's sampling instead of asserting
      `traceFlags: 1`; tracestate and baggage are capped to the specification
      with the largest member shed first and a non-throwing decode;
      `tracePoll` passes a real span LINK and `createJobTraceLink` carries the
      parent's real flags instead of `1 // Sampled`; a missing context at
      enqueue is now WRITTEN DOWN (`trace.unpropagated`) instead of returning
      unchanged; the residency guard parses the trace id out of the
      traceparent, validates the `x-trace-id` fallback, drops the clock-minted
      one and fills in the span id it was passing as `null` from the header it
      was already reading; health responses ECHO OR OMIT a join key and never
      invent one. S10.1's `oshun.target.ref` was widened from `SPAN_ONLY` to
      `EVERY_RECORD`, which its own description had said since 1.0.0, and its
      metric budget now spends the `ImpactCount` FLOOR (a pre-existing
      typecheck error, invisible because the library's `tsconfig.json` has an
      empty `include` and compiles nothing). Four shipped tests asserted the
      fabricated ids and now assert the repair.
      
      Verification: 131 kit tests, 107 tracing tests, 248 gpu-dispatcher, 26
      lilith, 3 bff-kit; 0 typecheck errors across all five projects; no
      export-map drift (148/148 modules). A 37-mutation sweep caught 35; the
      two survivors were proven EQUIVALENT by mutating the redundant partner
      (`fan-in` is both `many` and `deferred` so either branch answers `link`
      -- flipping its timing IS caught; `updateTraceState`'s own-key filter is
      redundant with `parseTraceState`'s first-wins dedupe -- removing the
      dedupe IS caught). Two survivors were real gaps and were closed: a
      boundary test that derived its input from `CLOCK_SKEW_TOLERANCE_MS`
      passed for every value of it, and every malformed trace id the obvious
      tests reached happened to land on `false` anyway, so the sampler's
      validation could be deleted unnoticed until a canary whose digits
      would have sampled IN was added.
      
      Scope stated: the module is `contract`-class, so the tracers, clocks and
      exporters it is decided against are the host's; S4.13's
      `parseTraceparent` and journey graph (`correlation.ts`), S10.1's
      `admitAttribute`/`classifyValue` and S9.3's `ImpactCount` are composed
      rather than re-decided. The `libs/shared` repairs are self-contained
      rather than importing the kit, because `libs/shared` is the lower layer.
      The canonical audit schema still requires a non-empty `traceId`
      (`z.string().min(1)`, documented "always present for correlation"),
      which is the pressure that produced the invented one; widening it
      touches every audit producer and consumer and was left alone. -->
      
  • S10.3 Instrument stage, gate, review, job, publish, and curation metrics. Evidence: metric names and units pass semantic review.

    • S10.3.a Define stage transition requests/refusals/completions/duration/ current distribution with bounded domain/stage/action/outcome labels. Evidence: stage metrics.

    • S10.3.b Define gate runs/duration/verdict/freshness age/invalidation/ recomputation/waiver/unavailable metrics with evaluator/policy families. Evidence: gate metrics.

    • S10.3.c Define review assignments/queue age/SLA/verdict/quorum/escalation/ delegation/COI/approval duration metrics without identity labels. Evidence: review metrics.

    • S10.3.d Define job queue/execution/attempt/retry/cancel/dead-letter/ resource/usage/cost/outcome metrics by bounded job/provider/resource class. Evidence: job metrics.

    • S10.3.e Define publish plan/eligibility/refusal/execute/read-back/ mismatch/compensation/target latency and outcome metrics. Evidence: publish metrics.

    • S10.3.f Define curation proposal/review/bundle/publish, eligible/ ineligible, collection/feature size/rank drift, catalog verify/correction metrics. Evidence: curation metrics.

    • S10.3.g Verify instrument type, unit, boundaries, aggregation, cardinality, privacy, owner, dashboard, and alert use. Evidence: semantic review. <!-- S10.3 shipped 2026-08-13.

      THE VEIN: A METRIC IS A CLAIM ABOUT A POPULATION AND AN INSTRUMENT ONLY
      EVER SEES WHAT REACHED IT, AND EVERY CHECK IN THIS REPOSITORY IS ABOUT
      THE POINT THAT ARRIVED. There are four ways not to arrive, all of which
      still return a number, and this repository shipped all four.
      
      NO DENOMINATOR. `libs/shared/gpu-dispatcher/src/metrics.ts` declared
      `gpu_jobs_total{job_type, priority}`,
      `gpu_jobs_completed_total{job_type, endpoint}`,
      `gpu_jobs_failed_total{job_type, endpoint, error_type}` and
      `gpu_jobs_cancelled_total{job_type}` -- four names, four label sets, one
      question. `job_type` was the only key common to all four, so "what
      fraction of high-priority jobs succeeded" had no answer: the outcomes
      were never keyed on priority and no query could recover it.
      
      NO WRITER. `gpu_jobs_running` was declared in the interface, registered
      with prom-client, given a help string and two label names, and
      incremented by nothing. Its only test was
      `expect(metrics.jobsRunning).toBeDefined()` -- an assertion that the
      instrument exists and never that it moves -- while the collector beside
      it maintained the same quantity in memory and put it in the snapshot. A
      panel over an empty series draws a flat line, and flat reads as healthy.
      
      NO RESOLUTION. `QUEUE_WAIT_BUCKETS` ended at 300s while
      `GPU_DURATION_BUCKETS` beside it ended at 3600 -- and queue wait is the
      one of the two that grows without bound, because that is what a backlog
      IS. `histogram_quantile` answers a quantile falling in `+Inf` with the
      highest FINITE boundary, so p99 queue wait reported 300 and stopped
      moving: the series went FLAT during exactly the incident it exists to
      show. `libs/shared/metrics` made it worse by defaulting every histogram
      with no explicit buckets to `HISTOGRAM_BUCKETS.HTTP_DURATION`, 5ms to
      10s, a hypothesis about a different subject applied platform-wide.
      
      NO RETURN PATH. `recordJobCancellation` incremented a counter and
      touched no gauge, so `gpu_jobs_pending` rose by one per cancellation and
      never came back -- forever -- while the collector's in-memory
      `jobStats.pending--` got it right. Two representations of one quantity,
      disagreeing by exactly the cancellation count, and the wrong one was the
      one that left the process. Because the instrument is a level nobody
      rates it, so the drift had no slope to notice.
      
      Plus the value-side twin of S10.1: AN INSTRUMENT TYPE LICENSES
      ARITHMETIC AND THE VALUE HAS TO SUPPORT IT. `updateEndpointHealth`
      stored `healthy -> 1, degraded -> 0.5, unhealthy -> 0` and
      `updateCircuitBreakerState` stored `closed -> 0, half-open -> 0.5,
      open -> 1` -- enumerations wearing numbers, running in OPPOSITE
      directions in one file, so 1 meant "good" in one gauge and "bad" in the
      other. `avg` over endpoints returns 0.5 both when every endpoint is
      degraded and when half are perfect and half are dead.
      
      Three more in the substrate, `libs/shared/metrics/src/registry.ts`: the
      instrument cache was keyed on the NAME ALONE and was a get-or-create, so
      a second `histogram({name, buckets})` returned the first instrument and
      silently discarded the second declaration's buckets and label names
      (S10.1 fixed exactly this in the OTLP path with
      `assertNoInstrumentConflicts`; the Prometheus path, which the job and
      cache metrics actually use, did not have it); `PromGauge.get()` ended
      `?? 0`, so a health check querying a misspelled label read a permanently
      healthy zero; and `setDefaultLabels` assigned rather than merged, so a
      later caller adding `region` dropped the `service` set at construction
      and the process exported under no service at all.
      
      SHIPPED. `libs/oshun/workbench-kit/src/workflow-metrics.ts`
      (contract-class, entry point `@oshun/workbench-kit/workflow-metrics`):
      38 metric declarations across the six families (a-f), each with
      instrument, UCUM unit, value semantics, bounded label domains, explicit
      histogram boundaries with a declared operating range, owner and
      escalation, recording site, and named dashboard/alert/objective readers;
      seven lifecycle machines the levels and outcome partitions are checked
      against; and `reviewWorkflowMetricCatalog` (g), which reports across the
      nine axes S10.3.g names plus three the vein adds -- name, denominator,
      return-path. 20,069 series against a 100,000 budget. 157 tests.
      
      The checks that are algorithms rather than assertions: a threshold must
      be EXACTLY a boundary, because the fraction under T is
      `bucket{le=T}/count` and there is no such bucket otherwise; consecutive
      boundaries are capped at a ratio of 4, because a quantile inside a
      bucket is interpolated across it and the shipped 300 -> 3600 step is a
      twelvefold error bar; a denominator's labels must be a SUBSET of the
      numerator's, which is the gpu-dispatcher defect stated as a rule; a
      level's `leavesOn` must cover every transition out of the counted states,
      derived from the machine rather than declared, which is the cancellation
      leak stated as a rule; and k-anonymity by PIGEONHOLE -- `population /
      cells` below the floor means at least one cell is, whatever the
      distribution -- which is what makes S10.3.c's "without identity labels"
      achievable, since `{domain, gate, verdict, shift}` contains no
      identifier and still names one person when one person reviews that gate
      on that shift.
      
      REPAIRS. `libs/shared/metrics`: histogram buckets are now required with
      no fallback; one declaration ledger across all four instrument kinds
      that refuses a disagreeing re-declaration and rolls back on construction
      failure; boundaries validated strictly increasing (prom-client silently
      SORTS what it is given); `Gauge.get` widened to `number | undefined`;
      queries on undeclared label keys refused; `Counter.get()` sums instead
      of returning the first series as the whole; `set(labels)`/`observe(labels)`
      with no value throw instead of silently dropping the write;
      `setDefaultLabels` merges and refuses to rebind an identity key already
      set; `cache_keys_total` renamed (a gauge wearing a counter's suffix) and
      `queue_jobs_completed`/`_failed` given theirs; `BACKLOG_AGE_SECONDS` and
      `LONG_JOB_SECONDS` presets added because nothing there could express a
      backlog. `libs/shared/gpu-dispatcher`: one `gpu_job_outcomes_total`
      counter carrying every denominator label plus an exhaustive four-value
      outcome partition, replacing three counters with three label sets;
      `gpu_jobs_running` given real edges (and `gpu_jobs_dispatched` retired,
      since `sum by (endpoint)` recovers it); `recordJobCancellation` given a
      required `phase` with no default, because only the caller knows which
      level the job was in and a default would be a guess that drives the
      other one negative; `recordJobDeadLettered` added for the fourth
      terminal outcome; queue-wait buckets extended to the 24h dead-letter
      deadline and the 300 -> 3600 duration gap filled; both enum gauges
      converted to state sets with one polarity; and `GpuMetricsOptions.prefix`
      deleted -- it was documented "(default: 'gpu_dispatcher')" and read by
      nothing.
      
      The compiled `.js`/`.d.ts` shadows tracked next to
      `libs/shared/metrics/src/*.ts` were patched in lockstep: vitest resolves
      `./types` to the stale `types.js`, so editing only the `.ts` changed
      nothing and the test kept asserting the old name. Regenerating them with
      plain `tsc` was tried and backed out -- the tracked artifacts carry
      `./registry.js` import specifiers that a plain emit does not produce.
      
      Verification: 11,510 workbench-kit tests, 71 metrics, 271
      gpu-dispatcher; 0 typecheck errors across all three projects under both
      their lib and spec configs; export map regenerated clean at 158 entry
      points and 0 violations -- which also retires two pre-existing
      violations, since S10.2 added `./trace-continuity` to package.json and
      tsconfig.base.json without a ratifying entry and `--check` had been RED
      ever since. A 61-mutation sweep caught 60. The two survivors were both
      real gaps and both closed: `coversFloor` could be pinned to true because
      every assertion that touched it went through `boundaryFaults`, which
      refuses the same condition by a different route; and the quantile
      comparison at the overflow boundary turned on floating point, since
      `1 - 0.95` is `0.050000000000000044` and a declared p95 against a
      declared 0.05 overflow share -- the exact case where the quantile sits
      ON the boundary and is unresolvable -- came out resolvable by one part
      in 10^17. A `SHARE_TOLERANCE` of 1e-12 fixes it and makes the strict/
      non-strict distinction a provably equivalent mutant: removing the
      tolerance IS caught.
      
      Four targeted mutations confirmed the gpu-dispatcher spec earns its
      keep: reinstating the cancellation leak fails 4 tests, the writer-less
      running gauge 5, the 300s queue-wait ceiling 3, and the enum-in-value
      health gauge 3.
      
      Scope stated: the module is `contract`-class, so the meter, clock and
      exporter it is decided against are the host's; S10.1's `Unit`,
      `ValueDomain`, `domainCardinality` and S9.3's `ImpactCount` are composed
      rather than re-decided, and S10.1's `UNITS` table gained the four
      workflow annotations (`{job}`, `{run}`, `{review}`, `{transition}`)
      rather than the catalogue forking its own. The declared figures a review
      cannot compute -- a subject population, an overflow share -- are carried
      with a required source field rather than estimated. `libs/shared` was
      repaired self-contained rather than importing the kit, because
      `libs/shared` is the lower layer. Pre-existing and left alone: 16
      typecheck errors in `libs/shared/gpu-dispatcher/src/validation.spec.ts`
      about a required `GpuJobResult.metrics`, unrelated to this work and
      present before it. -->
      
  • S10.4 Define SLOs and error-budget policies for interactive and asynchronous paths. Evidence: owners approve measurable objectives.

    • S10.4.a Inventory critical interactive read/search/navigation/save/ mutation/review/gate/publish-control journeys and user-visible success. Evidence: interactive SLI register.

    • S10.4.b Inventory critical queued job/gate/provider/render/import/export/ publish/notification/index/localization workflows and terminal success. Evidence: async SLI register.

    • S10.4.c Define availability/correctness, latency/deadline/queue-age, durability/freshness, and partial/degraded success semantics with qualifying events and exclusions. Evidence: SLI specifications.

    • S10.4.d Set target/window/error budget by domain/risk/priority and dependency contribution without hiding missing telemetry. Evidence: SLO table.

    • S10.4.e Define multiwindow burn alerts, ownership/escalation, freeze/ rollback/capacity actions, and release/feature-work error-budget policy. Evidence: policy.

    • S10.4.f Validate SLIs against synthetic success/failure/degraded/missing data and historical or load evidence. Evidence: SLI tests.

    • S10.4.g Obtain named service/domain/product/operations owner approval and quarterly review. Evidence: approval records. <!-- S10.4: shipped as libs/oshun/workbench-kit/src/service-objectives.ts (contract-class, entry point ./service-objectives, 159 approved entry points and no violations) with a 125-test spec, plus in-place repairs to libs/shakti/deployment/src/monitoring-observability.ts -- the only shipped SLO register in this repository -- and two small extensions to S10.3's workflow-metrics.ts enums.

      THE VEIN: AN OBJECTIVE IS A RATIO AND THE FAILING SYSTEM COMPUTES ITS
      OWN DENOMINATOR. S10.3 was about the observation that never reached the
      instrument; this is one layer up and worse, because here the
      observations DID arrive and the ratio still lies. `good / valid` has two
      sides and the second one is never a fact about the world -- it is a set
      the system chooses. Which events qualify, which are excused, which have
      not finished yet and which are simply missing are all decided by code
      that the outage is happening to, and a number between zero and one comes
      out regardless, onto a dashboard with a green threshold line. Five ways
      the denominator moves, all five measured in that one shakti file.
      
      1. THE DENOMINATOR IS THE COMPLETIONS. `slo-media-transcoding-throughput`
      promised "95% of transcoding jobs complete within their expected
      duration" and computed
      `completed_total{within_sla="true"} / completed_total`. Both sides count
      completions, so a job that has not completed is in neither, and the one
      failure that matters -- the queue stops draining -- moves the ratio not
      at all. `completionsOnlyValue` is shipped beside `evaluateSli` so the
      spec can put both numbers on one set of observations: 950 successes and
      50 permanently stalled jobs read 0.95 honestly and 1.0 on that query.
      The fix is not a different query, because "everything admitted" is also
      wrong -- it is depressed by work still legitimately in flight, in
      proportion to the ARRIVAL RATE, so a traffic spike is indistinguishable
      from an outage, people learn the alert is noise, and the next honest
      attempt reaches for the completions denominator again. The correct
      denominator is admissions DUE BY NOW, which makes the deadline a field.
      
      2. THE EXCLUSION IS DECIDED BY THE THING THAT FAILED.
      `slo-notification-delivery` divided an unfiltered `delivered_total` by
      `dispatched_total{recipient_valid="true"}` -- two metric names, one
      filter, on one side -- so the numerator counted deliveries whose
      dispatches were never counted and the ratio could exceed one, which is
      an error budget that refills. Underneath the arithmetic: an exclusion
      whose predicate is computed FROM THE OUTCOME is evaluated by the code
      that produced the outcome. "Exclude client errors" removes a validator
      regression, a bad schema migration and a clock skew that rejects every
      token from the denominator, and the graph improves while it happens.
      `Exclusion` therefore carries `decidedBy` and `evidenceTiming`, and
      subject-system + from-outcome is refused; the same predicate is
      admissible when the fact predates the attempt, because a recipient
      marked invalid at enrolment cannot be manufactured by today's incident.
      
      3. THE QUERY MATCHES NO SERIES AT ALL. That module declared seven
      `PrometheusMetricDefinition`s and FIVE of its seven objectives were
      computed from names that were not among them --
      `shakti_workout_session_saves_total`,
      `shakti_media_transcode_completed_total`,
      `shakti_auth_request_duration_seconds`,
      `shakti_notification_delivered_total`/`_dispatched_total` and
      `shakti_biometrics_resource_utilization_ratio`. In PromQL a name that
      matches nothing is an empty vector, `empty / empty` is empty, and
      `expr > threshold` over an empty vector is never true: AN SLI THAT DOES
      NOT EXIST IS NEVER BREACHED. The register and the metric store sat in
      one file four hundred lines apart with nothing joining them.
      
      4. THE BURN RATE HAS NO PERIOD. The fraction of budget an alert lets you
      spend before firing is exactly `rate x window / period`, and the burn
      rate CANCELS out of that expression -- derived in
      `budgetFractionAtDetection` and asserted for a 20x and a 100x burn on
      the same alert. So `{24h, 3.0}` is a 10% alert on 30 days and a 43%
      alert on 7. Those rungs were transplanted: `slo-biometrics-ingestion`
      ran a 7-day window with `{24h, 4.0}`, which fires only after 57.1% of
      the week's budget is gone, and `{6h, 8.0}` at 28.6%;
      `slo-media-transcoding-throughput` spent 28.6% on its own 24h rung. And
      every one of the seven had the other hole: an alert never fires below
      its own rate, the lowest rate anywhere in the file was 2.0, so a steady
      1.5x burn empties a 30-day budget in twenty days and trips nothing ever.
      `burnLadder` derives the ladder instead -- four budget-consumption
      decisions (2%, 5%, 10%, 10%) at rates 14.4/6/3/1, with `W = f x P / R`
      rounded DOWN so rounding can only spend less. On a 30-day period that
      reproduces the familiar 1h/6h/24h/72h table exactly, which is the point:
      those are not four remembered constants.
      
      5. NO DATA IS NOT GOOD DATA. `0/0` during an exporter outage, a
      collector restart or a regional blackout makes every comparison against
      the result false, so the alert is silent for exactly the duration of the
      blackout and the compliance figure is an average over the intervals that
      reported. `evaluateSli` returns `null` and never a number for an empty
      valid set; `NO_DATA_POLICIES` refuses `good` outright and requires
      `unknown` to carry a coverage objective whose own policy is `bad` --
      a recursion that terminates, because if the coverage series is also
      missing, that IS the outage.
      
      Other decisions worth keeping. AN OBSERVATION IS A RECORD WRITTEN BY A
      COMPONENT, SO IT CANNOT WITNESS THE FAILURE OF ANYTHING UPSTREAM OF
      ITSELF: a server-handler SLI is blind to the request that never reached
      the handler, and -- the case that surprises people -- a client-observed
      availability SLI is blind to `client` and `network` despite sitting in
      front of both, because a browser that never loaded sends no beacon and
      "no beacons" is the same bytes as "no traffic". PARTIAL CREDIT NEEDS A
      FULL-SUCCESS FLOOR, because 30% of events at 0.5 credit reads 85% while
      nearly a third of the population had a bad time. A DEPENDENCY WITH NO
      PUBLISHED NUMBER CANNOT BE MULTIPLIED IN, so `achievableCeiling` returns
      an UPPER bound and names the unmeasured, and an objective carrying one
      must disclose it -- `slo:publish-target-ack` does, because an app store
      publishes no availability to us and writing 0.999 because it feels about
      right is the hidden telemetry S10.4.d refuses. A STANDBY IS A CLAIM
      ABOUT WHAT HAPPENS DURING A FAILURE, so an unexercised redundancy group
      contributes its primary alone. THE SHORT WINDOW IS WHAT MAKES AN ALERT
      STOP, not a second alert -- shakti's `{5m, 14.4}` beside its
      `{1h, 14.4}` was a standalone page on 0.17% of a budget next to a rung
      that could only ever fire later. A FREEZE THAT BLOCKS THE ROLLBACK
      BLOCKS THE REMEDY, and a freeze the frozen owner may lift ends whenever
      it becomes inconvenient. AND AN APPROVAL ATTESTS TO A VERSION, NOT A
      NAME: `objectiveFingerprint` covers the target, window, grades,
      exclusions, dependencies, alerts and policies and deliberately excludes
      prose, so rewording a rationale does not invalidate a signature and
      changing a target does.
      
      The bidirectional closure is the reason this module is where it is.
      Five S10.3 metric declarations carried a `consumers` entry of kind `slo`
      -- `slo:stage-transition-interactive`, `slo:gate-run-completion`,
      `slo:review-sla-48h`, `slo:job-success-rate`, `slo:publish-target-ack`
      -- and until now NOTHING defined them: the metric knew an objective was
      reading it and the objective's target, window, budget, owner and policy
      existed nowhere. `reviewObjectiveRegister` refuses both directions.
      
      Shipped register: 8 interactive journeys covering all eight kinds S10.4.a
      names, 11 async workflows covering all ten S10.4.b names plus `review`
      (a stated scope decision -- a review assignment is a queue with a human
      worker, and leaving it out would put the one workflow whose deadline is
      48 HOURS in the register that measures milliseconds), and 20 objectives
      with four named owner roles, four-rung ladders and four-policy sets each.
      Two S10.3 enum extensions rather than forks: `localization` joined
      `JOB_CLASSES` and `notification` joined `PUBLISH_TARGET_CLASSES`, since
      a notification is a publish -- something leaves the building for a target
      that acknowledges it or does not, and the read-back check is exactly
      "did the person get told".
      
      Shakti repairs: the seven missing metrics declared (with 0.2 present as
      an explicit boundary of the new auth histogram, because a threshold
      between boundaries silently measures the fraction under 0.25 -- a 25%
      looser objective than the one approved); the notification ratio made
      symmetric and moved onto an outcome partition with `expired` written by
      a sweep; the media denominator moved from completions to admissions with
      `abandoned` for the stall, and its `sliBasis` corrected from
      `window-based`, which contradicted its own counter-ratio query; all seven
      ladders redivided for their own periods with a 1x rung added and the
      5m/1h duplicates folded; and `validateSloDefinitions` added so none of it
      can regress -- it resolves every query against the metric register,
      checks `le=` against declared buckets, checks ratio symmetry, checks the
      query range against the claimed window, and checks the ladder arithmetic.
      It refuses a query it cannot parse rather than passing it silently,
      because a parser that returns success without checking is the same defect
      as the empty vector.
      
      Verification: 11,638 workbench-kit tests (125 new), 2,311 shakti
      deployment tests (19 new); 0 typecheck errors across both projects under
      both lib and spec configs; export map regenerated clean at 159 entry
      points, 0 violations. A 75-mutation sweep caught 69 on the first pass;
      all six survivors were real gaps and all six were closed (the exact
      `admittedAt + deadline` boundary; the `budgetRemaining` clamp at one; the
      long:short ratio band on both validators; the workflow-admission check,
      which was masked because the partition check refused the same condition
      by a different route -- S10.3's lesson, met again; and the blind-spot
      check, which nothing exercised through `reviewObjective` because no
      shipped journey has a blind spot). Re-run: 75/75. One defect the sweep
      found in my own code first: `budgetRemaining(0.99, 0.01)` returned
      8.88e-16 rather than 0, because `1 - 0.99` is `0.010000000000000009` --
      so the one moment a `budget-exhausted` freeze exists for was the moment
      it would not have fired, by one part in 10^15.
      
      Scope stated: the module is `contract`-class, so the clock, the query
      engine, the alert manager and the release gate it is decided against are
      the host's. S10.1's `fnv1a32` and `admitSpanName`, S10.3's lifecycles,
      catalogue and `resolvesThreshold`, and S9.3's `ImpactCount` are composed
      rather than re-decided -- the event-denominated budget is an
      `ImpactCount` because it is a target multiplied by a volume forecast, and
      printing that as a bare integer is how a forecast becomes a fact between
      one slide and the next. -->
      
  • S10.5 Add dependency health, saturation, queue age, and circuit-breaker dashboards. Evidence: fault injection changes the expected panels.

    • S10.5.a Register databases, blobs, queues/workers, caches/search, providers/models/tools, signing/trust, publishers/channels, bridges/ hosts, identity, and telemetry dependencies with owner/SLO. Evidence: dependency register.

    • S10.5.b Display configured/not-configured/healthy/degraded/unavailable/ failed/stale-probe, latency/error rate, last success/failure, and affected capability. Evidence: health panels.

    • S10.5.c Display CPU/memory/GPU/storage/connection/thread/license/token/ quota/provider capacity versus limits and trend. Evidence: saturation panels.

    • S10.5.d Display queue depth/age by priority/type/tenant-safe cohort, throughput, lease loss, retries, dead letters, worker availability, and oldest work. Evidence: queue panels.

    • S10.5.e Display circuit closed/open/half-open, reason, threshold, failures, opened/until, probes, affected calls, fallback/degradation, and authorized controls. Evidence: circuit panels.

    • S10.5.f Link SLO/burn, traces/logs/jobs/incidents/runbooks/owners and apply bounded labels/access/redaction. Evidence: dashboard review.

    • S10.5.g Inject timeout/throttle/saturation/queue stall/circuit/dead-letter faults and assert exact panel/alert/recovery changes. Evidence: dashboard fault tests. <!-- S10.5: shipped as libs/oshun/workbench-kit/src/dependency-dashboards.ts (contract-class, entry point ./dependency-dashboards, 160 approved entry points and no violations) with a 110-test spec, plus in-place repairs to libs/lilith/service-lib/circuit-breaker.ts (the live breaker where the vein is measured), libs/freya/db/src/observability.ts (ten breaker configurations) and libs/shakti/deployment/src/monitoring-observability.ts (the dashboard register), and six dependency contributions added to S10.4's objectives.

      THE VEIN: THE INSTRUMENT IS DOWNSTREAM OF THE MITIGATION, SO EVERY
      PROTECTION IMPROVES THE PANEL WHILE THE OUTAGE CONTINUES. S10.3 was the
      observation that never reached the instrument and S10.4 the ratio whose
      denominator the failing system chooses. This is the layer a person reads
      at 3am, and it has a defect the other two do not: in front of every
      dependency sits a timeout, a retry, a breaker, a throttle, a bulkhead, a
      fallback, a dead-letter queue or a cache, each of which exists to REMOVE
      traffic, each of which therefore changes the population the panel is
      computed over, and each of which engages exactly during the incident. The
      direction is never random -- a mitigation removes the bad events, because
      that is what a mitigation IS -- so a panel over the survivors moves the
      reassuring way at the moment the system is worst.
      
      Measured rather than argued. `libs/lilith/service-lib/circuit-breaker.ts`
      is a live breaker with a `getStatus()` built for a panel. Its `execute()`
      increments `metrics.totalRequests` for every call INCLUDING the
      short-circuited ones and returns without touching
      `metrics.totalFailures`; `failureRate` is the second over the first.
      Driven by hand against a dependency 100% down for the whole observation,
      that panel reads 100.00%, then 33.33%, 4.35%, 0.45%, 0.04% as traffic
      continues. The error rate does not stop rising -- it FALLS, monotonically,
      toward zero, for as long as the dependency stays broken, and falls faster
      the harder the caller retries. Nothing is miscounted; every increment is
      correct on its own terms. The population changed underneath the ratio.
      
      Five more in that one file, all repaired. THE SUPPRESSED EVENTS ARE
      COUNTED NOWHERE -- `grep shortCircuit` returned nothing, so the number
      that replaced the errors was not merely off the panel, it was not
      recorded. THE MITIGATION'S OWN REJECTION IS BLAMED ON THE DEPENDENCY:
      the concurrency limiter called `recordFailure('overload')`, so a burst of
      local concurrency opened the circuit against a perfectly healthy
      dependency and the panel attributed our saturation to them. THE FALLBACK
      IS COUNTED TWO WAYS FOR ONE OUTCOME -- unrecorded in the open branch,
      a failure in the catch branch, so whether a degraded answer was a failure
      depended on which side of the breaker state the caller arrived on.
      HALF-OPEN HAD NO PROBE BUDGET and closed on ONE success, so recovery was
      a thundering herd authorised by a single lucky probe. And THE MANUAL
      CONTROL ERASED ITS OWN EVIDENCE: `reset()` assigned `this.state = CLOSED`
      and THEN called `setState(CLOSED)`, so the recorded transition was
      `closed -> closed` with the fallthrough reason `State change` and the
      `open -> closed` edge never entered `stateChanges` at all -- the one
      action an operator takes during an incident was the one action the
      history could not show, and a post-incident review read the override as
      a recovery.
      
      The theorem worth keeping: A LATENCY PERCENTILE ABOVE THE CENSORED SHARE
      IS THE TIMEOUT, NOT A MEASUREMENT. A timeout at T censors exactly the
      attempts slower than T, so the recorded sample is precisely the fastest
      `1-c` of attempts in order and an observed q-quantile is the population's
      `q(1-c)` quantile -- exact, not approximate. Under 2% timeouts a panel
      labelled p99 is the population's p97.02, and every quantile at or above
      `1-c` is bounded below by T and above by nothing, because the attempts
      that make it up were stopped and how long they had left is not in the
      data. The panel goes FLAT at T as the dependency degrades, and flat reads
      as stable. `identifiedQuantile` and `quantileIsIdentified` ship it.
      
      Other decisions worth keeping. AN ABSOLUTE FAILURE THRESHOLD IS A RATE
      THRESHOLD DIVIDED BY A TRAFFIC LEVEL NOBODY WROTE DOWN, and it is wrong
      in both directions at once: freya's `paystack` opens on five failures in
      sixty seconds, which is a 0.083% error rate at a hundred calls a second
      -- tripping on a dependency performing far better than its objective
      requires -- and at one call a minute the window does not hold five calls,
      so nothing opens it however completely the dependency has failed. A
      BREAKER CANNOT OPEN BEFORE ONE FAILURE HAS BEEN RECOGNISED AS ONE, so the
      request timeout is a hard floor on detection and `ai_vision`'s sixty
      seconds means every caller in the first minute waits the full sixty. A
      LEVEL WITH NO LIMIT IS NOT A SATURATION -- `shakti_biometrics_resource_
      utilization_ratio` is named a ratio, typed a ratio and divided by nothing
      anybody declared. THE BINDING LIMIT IS THE SMALLEST IN A CHAIN NOBODY
      DREW, and scaling out makes a shared limit bind harder while every
      per-instance panel improves. A MEAN ACROSS INSTANCES HIDES THE ONE
      REPLICA THAT IS FULL (one at 100% and nine at 10% averages 19% while a
      tenth of all traffic is refused). DEPTH COUNTS THE READY SET AND A LEASE
      HOLDS WORK OUT OF IT, so work held by a dead worker is in no terminal
      state and no ready state -- the depth gauge reads empty, the age gauge
      reads empty, and detection has a floor of one lease duration that nothing
      on the board discloses. A DRAIN ESTIMATE THAT IGNORES ARRIVALS
      under-reports without bound and prints a finite number for a queue that
      never empties. AN EMPTY QUEUE AND A BROKEN PRODUCER DRAW THE SAME LINE,
      so `starved` needs a declared expected arrival rate to exist as a verdict
      at all. A CUMULATIVE DEAD-LETTER TOTAL IS MONOTONE, so an alert on it
      fires once and is silenced permanently, and the next batch of abandoned
      work arrives into a muted alert. A TRACE LINK IS EITHER AN EXEMPLAR OR AN
      UNBOUNDED LABEL. And `not-configured` IS A CONFIGURATION FACT, NOT A
      HEALTH FACT -- three axes collapsed into one enum is what lets a
      dependency nobody configured render as healthy and an hour-old cached
      answer render as current, so `panelStateOf` is a total join over the
      three and the rendered value is compared against it rather than trusted.
      
      The evidence clause, run rather than asserted. Six fault transforms MOVE
      work between census fields instead of inventing it, and the partition
      identities are checked after every one at seven magnitudes, so a delta is
      a real change and not an accounting artefact. Against the board shape
      this repository ships -- an error rate over CALLS, a depth over the READY
      set, no witness panels -- three of the item's six faults are
      undetectable, and two of those three move it the REASSURING way:
      injecting `circuit-open` improves the error rate and the latency
      quantile, injecting `dead-letter` improves the queue depth, and injecting
      a 50% `throttle` moves nothing at all. The board does not go quiet, it
      goes green, and the reader acquires positive evidence to look elsewhere.
      With the witness panels the same six are all detected -- and the
      reassuring movement does not go away, it gets accompanied, which is the
      whole design.
      
      The bidirectional closure, both ways. Ten `dep:` ids appeared in S10.4
      objective contribution chains and NOTHING defined any of them: no kind,
      no owner, no affected capability, no panel, while each carried a declared
      availability being multiplied into a ceiling. And six dependencies a
      workbench actually calls -- the worker pool every asynchronous objective
      runs on, the signing service and trust store a publish needs, the channel
      a notification leaves over, the cache the interactive latency objective
      depends on, the provider gateway -- were in no chain at all, so
      `achievableCeiling` multiplied them at 1.0 and every one of those
      ceilings was optimistic. Added, with all twenty objectives re-checked:
      every target still sits under its corrected ceiling. The three remaining
      uncovered dependencies are all `supporting`-tier and are reported at
      `repair` severity rather than refused, because a supporting dependency
      with no objective is a stated product position and the register records
      it instead of manufacturing a target.
      
      Shipped register: 19 dependencies covering all 17 kinds and all ten
      groups the item names, one board per dependency with its readings DERIVED
      from the protections in front of it (so adding a `throttle` adds a
      `throttle-rate` panel and forgetting to build it is a finding rather than
      a silence), and a nine-entry suppression table binding each mitigation to
      the ONE reading that witnesses what it removed.
      
      Verification: 11,766 workbench-kit tests (127 new); 47 lilith
      service-lib circuit-breaker tests (11 new, all 36 prior still passing);
      114 freya db tests (7 new); 2,319 shakti deployment tests (7 new);
      0 typecheck errors across all four projects under both lib and spec
      configs; export map regenerated clean at 160 entry points, 0 violations.
      A 149-mutation sweep caught 129 on the first pass; all 15 real survivors
      were genuine gaps and all 15 were closed (a timed-out attempt counted as
      an error; the identified-quantile clamp; not-configured beating an
      expired probe; a resource already past its limit; arrivals exactly
      meeting completions; an ESTIMATED population used at its estimate rather
      than its observed floor for k-anonymity; forced-closed asked for a
      short-circuit count; the supporting-tier severity split, in both
      directions; a dependency an objective names and the register omits; the
      comparison tolerance on a delta; the circuit and queue reading sets; and
      `reviewDashboard` admitting a board that draws everything and detects
      nothing). Re-run: 149/149. The sweep also found a defect in this module
      itself: `injectFault` carried a cumulative total forward across recovery
      so that a monotone panel would show as latched, and NO reading kind read
      a cumulative total, so the mechanism could not fire. `DashboardSpec`
      gained `deadLetterCumulative` -- the same shape of choice as
      `depthCoversUnsettled` and `errorRateDenominator`, and the only one
      whose consequence appears on the way OUT of an incident.
      
      Scope stated: the module is `contract`-class, so the clock, the query
      engine, the dashboard server and the fault injector it is decided against
      are the host's. S10.4's owners, objectives and ceilings, S10.3's
      k-anonymity, aggregations and label privacy, S10.1's bounded domains and
      S9.3's `ImpactCount` are composed rather than re-decided. The seven
      original shakti dashboards were NOT given invented panel queries: their
      contents live in a Grafana this repository does not contain, so they
      report `no-declared-panels` and say why. -->
      
  • S10.6 Add privacy-safe structured logging. Evidence: automated log scan catches content, tokens, secrets, PII, and protected identifiers.

    • S10.6.a Define versioned event names and allowlisted typed fields for lifecycle/state/error-class/correlation/trace/operation/version/ duration/usage/cost and bounded context. Evidence: logging schema.

    • S10.6.b Define severity and required/optional sampling for validation, auth, dependency, retry, job, provider, publish, audit linkage, degradation, and incident events. Evidence: event catalog.

    • S10.6.c Exclude raw authored/source content, prompts/responses, file/URL/ query/path, tokens/credentials, stack/SQL, PII, protected identity, learner/biometric, and unreleased asset detail by default. Evidence: field prohibition policy.

    • S10.6.d Redact/tokenize/pseudonymize only approved fields before log emission and treat redaction failure as drop/quarantine, not raw fallback. Evidence: redaction tests.

    • S10.6.e Configure transport encryption, regional storage, access, retention/deletion/legal hold, export, sampling/rate limits, and unavailable exporter behavior. Evidence: log lifecycle tests.

    • S10.6.f Scan source/config/generated logs and runtime output with canary content/secrets/PII/protected IDs; minimize retained failures. Evidence: automated leakage report. <!-- S10.6: shipped as libs/oshun/workbench-kit/src/log-privacy.ts (contract-class, entry point ./log-privacy, 161 approved entry points and no violations) with a 239-test spec, plus the harness tools/log-leak-scan.mjs (nx target log-leak-scan) and in-place repairs to libs/shared/logging -- the new src/redaction.ts, and changes to logger.ts, types.ts, middleware/request-logger.ts and transports/elasticsearch.ts -- with a 93-test redaction.spec.ts.

      THE VEIN: A RECORD FANS OUT TO EVERY SINK AND THE REDACTION WAS
      ATTACHED TO ONE OF THEM -- AND THE ONE IT WAS ATTACHED TO IS THE COPY
      NOBODY KEEPS. S10.3 was the observation that never reached the
      instrument, S10.4 the ratio whose denominator the failing system
      chooses, S10.5 the panel computed over what a mitigation had already
      removed. This is the only one of the four where the defect is not a
      wrong number: a log record is not aggregated, sampled or divided by
      anything. It is COPIED, verbatim, to every destination somebody
      attached, and a safety property attached to one destination is not a
      property of the record.
      
      Measured rather than argued. `OshunLogger.log()` called
      `this.pino[level](logData, message)` and then, separately,
      `createLogEntry(...)` followed by `writeToTransports(entry)`.
      `redact.paths` is a pino option, so pino censored what pino serialized,
      while `createLogEntry` re-read the ORIGINAL `data` and `error` objects
      and handed them to the Elasticsearch, HTTP, TCP and file transports.
      Driven by hand through the real logger with nine canaries, three were
      stopped on stdout and NONE on the transports: `password`, `apiKey` and
      `nested.secret` were redacted for the console and shipped raw to the
      index, while `deep.a.token`, an axios error's
      `config.headers.Authorization`, its `config.data` prompt, a learner
      email, a reset token in the message and a pg `internalQuery` leaked on
      both. Not one canary was stopped on the path that leaves the box.
      
      The direction is what makes it a vein rather than a bug. The sink that
      redacted is a terminal a developer reads for ten seconds and discards;
      the sinks that did not are the ones with a retention measured in months
      and a search box on top. So an engineer who tests redaction the obvious
      way -- log a secret, look at the console, see `[REDACTED]` -- has run
      the experiment on the only copy that is not kept, and has acquired
      positive evidence for a property the retained copies do not have.
      `redactionCoverage` names that shape specifically:
      `only-the-ephemeral-sink-is-sanitized` fires when the covered sink has
      the SHORTEST retention, and it is distinct from
      `redaction-attached-to-sink-not-record`, because partial coverage is
      merely bad and partial coverage of the throwaway copy is misleading.
      
      Six more in shipped code, all repaired. A DENYLIST OVER AN OPEN WORLD OF
      FIELD NAMES HAS UNBOUNDED FALSE NEGATIVES: `DEFAULT_REDACT_PATHS` was
      sixteen entries and pino's `*` matches exactly ONE level, so
      `deep.a.token` at depth three was unmatched -- the list was not short
      through carelessness but because the set of names that can carry a
      secret is not enumerable in advance. AN ERROR IS AN OPEN MAP AND THE
      LOGGER SPREAD IT: `serializeError` walked
      `Object.getOwnPropertyNames`, skipped five names and copied the rest, so
      axios brought `config.headers.Authorization` and node-postgres brought
      `internalQuery` and `detail`, with `stack` copied unconditionally while
      S10.6.c prohibits stack and SQL by name. A PATH-BASED REDACTOR CANNOT
      SEE INTO A STRING: `createMessage` returned
      `${req.method} ${req.url} ...` with `req.url` being `originalUrl`, query
      string included, and the same file put `data.query` and `data.ip` on
      every record UNCONDITIONALLY -- `logHeaders` gated headers, `logBody`
      gated bodies, and nothing at all gated the query string or the client
      IP. THE FIELD PROHIBITION POLICY WAS A TYPE, AND A TYPE FORBIDS NOTHING
      AT RUNTIME: `PrivacyConfig` declared six controls, was exported from
      `index.ts`, was not a member of `LoggerConfig`, was accepted by no
      parameter and read by no code, while the repository's only spelling of
      `redactFields` was a bare literal in `FREYA_LOG_CONFIG` passed to no
      logger, whose one test asserts
      `expect(FREYA_LOG_CONFIG.redactFields).toContain('password')` -- reading
      the array back out of the object that declares it. REDACTION FAILURE
      FELL BACK TO RAW in two places, `writeToTransports` catching into
      `console.error(name, error)` and the Elasticsearch transport's
      `sendBatch` doing `buffer.unshift(...entries)`, an unbounded re-queue,
      so the failure path of a sink that could not accept a record was a
      second sink that agreed to no policy. And pino's `base` was
      `this.context`, merged into every line without passing through
      anything, so the base context reached stdout unredacted AND duplicated
      on every record.
      
      The finding that was not in the item's text and changes what the
      evidence means: THE TEST SUITE WAS EXERCISING A STALE BUILD ARTIFACT.
      `libs/shared/logging/src` shipped TRACKED `logger.js`, `index.js`,
      `types.js`, `events.js` and their `.d.ts` and `.map` twins beside the
      `.ts`, committed in July while the `.ts` were last touched in April and
      May. Vitest resolved `./logger` to `logger.js`; tsx resolved it to
      `logger.ts`. Proved by driving both: the repair produced
      `droppedProperties` on the transport entry under tsx and a full stack
      under vitest, from the same source line. So every one of the 52
      pre-existing tests had been passing against a compiled copy of code
      nobody had edited in three months. All twelve artifacts deleted; the 52
      tests still pass, now against the source. This is also exactly what
      S10.6.f's word "generated" is for -- a committed `.js` is invisible to
      any scan written `--include="*.ts"` -- so `generatedShadows()` in the
      harness is a standing check for it. NOT FIXED, and reported rather than
      silently left: seven other libraries under `libs/shared` have the same
      tracked build output (41 `.js` and 77 `.d.ts` across config, database,
      http-client, metrics, security, storage and types). Each needs its own
      test verification, since a suite may currently pass BECAUSE of the stale
      copy, and that is outside what S10.6 asks for.
      
      What the repair can and cannot reach, stated rather than glossed. Re-run
      after the fix, the same nine canaries go from 3/9 stopped on one sink and
      0/9 on the other to 7/9 stopped on BOTH, with perfect sink parity on all
      nine. The two that remain are both carried IN THE MESSAGE STRING, and
      they are unfixable in a general-purpose logger by construction: the
      interpolation happened at the call site before the logger was called, so
      the logger receives one opaque sentence and no path-based rule can reach
      into it. That is why the kit's contract makes the message a
      `templateId` plus bounded parameters and `sanitize` refuses a free-text
      message by name -- the structural fix belongs in the contract a caller
      opts into, and the general logger gets the fan-out fix.
      
      Decisions worth keeping. AN ALLOWLIST PROJECTION HAS NO DEPTH TO BE
      DEFEATED AT -- `project` emits the declared fields and drops the rest,
      so an undeclared field is ABSENT rather than unmatched, which is the
      only answer to a denylist over an open world. A FIELD IS CLASSIFIED BY
      WHAT IT CARRIES AND NOT BY WHAT IT IS NAMED, so every surviving string
      value is re-scanned with S10.1's detectors and a `safe` declaration
      contradicted by its value is quarantined -- a declaration is a claim by
      whoever wrote the schema about what call sites will put there. THE
      WEAKEST SINK DEFINES THE RECORD'S EXPOSURE, so coverage is per-sink with
      no averaging, and egress dominates retention when choosing the weakest
      because a record we can still purge is a different problem from one a
      vendor has indexed. A COUNTED EVENT MAY NOT BE SAMPLED, and rate
      limiting is WORSE than sampling for a count rather than better: a
      sampled count divides back out by a known rate, and a rate-limited one
      is censored at a ceiling nobody records, so it reads FLAT at the limit
      however far the true rate rises -- S10.5's censored quantile arriving
      through the logging pipe. HARM DECIDES THE RESPONSE AND THE RESPONSES
      DIFFER: a credential is closed by rotation, personal data cannot be
      un-disclosed, and an embargo is the one harm that EXPIRES, which is why
      `unreleased-asset` gets relief on a long retention and nothing else
      does. A TOKEN OVER A LOW-CARDINALITY DOMAIN IS A RENAME, so
      pseudonymisation has a declared floor. AND A LEAK REPORT THAT QUOTES THE
      LEAK IS A SECOND LEAK -- S10.6.f's "minimize retained failures" is the
      only clause in the item about the SCANNER rather than the logger, and
      every finding carries a fingerprint and a length, composing the decision
      S10.1's `HazardFinding` already made rather than restating it.
      
      Composition rather than re-decision. S10.1's thirteen hazard classes are
      imported and extended by exactly four -- `sql`, `query-string`,
      `model-output` and `unreleased-asset` -- being the classes S10.6.c names
      that span attributes never carry; restating the thirteen would create
      two lists that drift, and a value classified `pii` by one and not the
      other passes whichever check runs second. `classifyValue`, `hazardsIn`,
      `isBounded`, `templateFaults`, `OTHER_SENTINEL` and `fnv1a32` are
      S10.1's; `ImpactCount` is S9.3's, and a sampled count comes back
      `estimated` carrying both its observed value and its fraction. One rule
      was corrected by the register catching it: requiring `isBounded` of
      every field domain imported a METRICS constraint -- a dimension costs
      one series per distinct value -- into a context where a field is stored
      once per occurrence, and it forbade the `template` domain that a route
      needs. Scoped to `enum`; boundedness is still required of context keys,
      which is where S10.6.a's "bounded context" actually lives. The route
      field is `safe` rather than `url` because S10.1's detectors return no
      hazard for `/studio/{projectId}/assets` and `url` for a real one, and
      what keeps that honest is `templateFaults` refusing a value whose
      segment still looks like an identifier -- "the template was concatenated
      rather than substituted", which is the shipped middleware's `req.url` in
      one sentence.
      
      The evidence clause, run rather than asserted. `tools/log-leak-scan.mjs`
      drives the REAL logger with ten canaries and FOUR controls, because a
      harness with no negative control proves that it ran and not that it can
      fail: a positive control planted in a field nothing matches MUST reach
      both sinks or the capture rig is not observing what it claims; the
      hazard canaries must be stopped on BOTH; and SINK PARITY is checked over
      every canary rather than asserted for the ones somebody remembered. All
      three of its failure modes have been observed failing -- the static pass
      reported 12 findings before the fixture allowances, and the
      generated-shadow check would have reported 4 before the deletion. Its
      allowances are RFC 2606 reserved domains and TLDs, RFC 5737 and RFC 3849
      documentation IP ranges, the published test card numbers and one
      explicit `NOTREAL-` marker, tested against the MATCHED SPAN rather than
      the line -- an allowance evaluated against the line would let one
      `example.com` in a fixture hide a real key three columns to the right.
      
      One defect this item's own change introduced, caught by the pre-commit
      ratchet and worth keeping: A SHARED LIBRARY MAY NOT DISCRIMINATE A UNION
      ON A BOOLEAN, BECAUSE ITS CONSUMERS CHOOSE `strict`. `redactEntrySafely`
      first returned `{ ok: true; entry } | { ok: false; quarantined }`, which
      typechecks cleanly under `libs/shared/logging`'s own config and FAILS in
      `libs/yemaya/agents`, whose `tsconfig.json` sets `"strict": false` --
      without `strictNullChecks` the `true`/`false` literal types widen to
      `boolean`, the union stops discriminating, and every variant-specific
      property access is an error. The library does not control the settings
      it is compiled under, so it has to pick the narrowing that survives
      them: the result is now discriminated by a STRING, and the three
      outcomes are genuinely three (`redacted`/`dropped`/`quarantined`) rather
      than two with `quarantined: null` standing in for the third.
      
      The commit's typecheck ran 399 projects because `tsconfig.base.json` is
      staged. Twelve fail, and all twelve are red for reasons that predate
      this change -- `@hathor/llm-npc` on a non-exhaustive switch (TS2366),
      `@hathor/event-handlers` on a stale composite build-info (TS6305),
      `@yemaya/agents` and the `@lilith/*` set on the `libs/contracts` refined
      -zod incompatibility. Only two of the twelve import `@oshun/logging` at
      all, and both report zero logging errors after the fix above, so the
      remaining failures are not this change's. Committed with
      `SKIP_TYPECHECK=1`, which is the escape the hook itself prints for this
      case; the two projects this item owns typecheck clean under both their
      lib and spec configs.
      
      Verification: 12,006 workbench-kit tests (239 new), 157 files, all
      passing; 145 shared/logging tests (93 new, all 52 prior still passing);
      0 typecheck errors in both projects under both lib and spec configs
      (including 12 pre-existing `void`-return errors in the shared/logging
      specs, fixed); export map regenerated clean at 161 entry points, 0
      violations; `log-leak-scan` exits 0 with the positive control escaping
      and all nine hazard canaries contained on both sinks.
      
      Scope stated: the module is `contract`-class, so the clock, the
      transports, the log store and the CI runner it is decided against are
      the host's. Every function is a decision over data. -->
      
      NOTE (S10.6, for S10.7): `evaluateDeletion` over the shipped sink
      register answers `partial` -- only Elasticsearch can delete by query,
      `http` is append-only-immutable and the rest expire by partition. S10.7
      asks for retention, legal hold and integrity over AUDIT records and will
      need that answer, along with `evaluateLegalHold`'s
      `hold-cannot-override-partition-expiry`, which is the case where a hold
      is recorded and the data goes anyway.
      
  • S10.7 Add audit retention, export, legal hold, and integrity verification. Evidence: tamper test and retention test.

    • S10.7.a Classify audit event types/data fields by retention source, minimum/maximum, tenant/regulatory/contractual policy, and purge rules. Evidence: audit retention matrix.

    • S10.7.b Implement immutable append-only storage/partitioning with hash/signature chain or equivalent integrity, trusted timestamp, backup, and restricted writer. Evidence: integrity design tests.

    • S10.7.c Apply retention/expiry/archive/purge while preserving required tombstone/integrity linkage and deleting prohibited content. Evidence: clock-controlled retention tests.

    • S10.7.d Implement legal hold by scope/reason/authority/start/release, prevent held purge, restrict visibility, and audit every action. Evidence: hold tests.

    • S10.7.e Export authorized query/time/tenant/object/event scope in stable schema with redaction policy, completeness, hashes/signature, chain verification, and export audit. Evidence: export tests.

    • S10.7.f Continuously/sample-verify sequence/hash/signature/timestamp/ storage/backup consistency and surface missing/reordered/changed events. Evidence: verification job.

    • S10.7.g Tamper, delete, reorder, duplicate, corrupt, expire, hold, export, and restore audit fixtures. Evidence: adversarial lifecycle report.

      <!-- S10.7: shipped as `libs/oshun/workbench-kit/src/audit-integrity.ts`
      (contract-class, entry point `./audit-integrity`, 162 entry points 0
      violations) + 101-test spec; `tools/audit-integrity-scan.mjs` (nx target
      `audit-integrity-scan`, run under tsx because `@oshun/audit-platform`'s
      `main` is `./src/index.ts`); and repairs to
      `libs/shared/audit-platform`'s `retention.ts`, `hash-chain.ts` and
      `evidence-export.ts`.
      
      THE VEIN: AN INTEGRITY CHECK IS A COMPARISON, AND BOTH SIDES OF THIS
      ONE CAME OUT OF THE SAME BOX. A hash recomputed from the document that
      stores it proves the document is internally consistent -- a property
      every competently forged document has. Detection needs ONE value the
      checker holds and the writer could not rewrite; if none can be named,
      what ran was a consistency check and it has to SAY so rather than
      return ok.
      
      Measured before writing anything, by driving the real services through
      sixteen probes: ten were undetected. `verifyChain()` seeds its expected
      previous-hash from `this.chain.range(start-1,start)[0].chainHash` --
      the same sidecar under attack -- and had no anchor parameter at all, so
      editing an event and recomputing the sidecar forward returned `ok:true`
      with 10/10 verified; truncating the tail of log AND chain returned
      `ok:true`; restoring from a five-event-old backup returned `ok:true`
      with the tip going backwards recorded nowhere. `getChainTip()` was
      commented "used by external anchors" and nothing supplied one.
      `verifyExport()` recomputed `hashBundle(bundle.events)` and compared it
      to `bundle.custody.bundleHash`, both operands inside the document, and
      NEVER READ `bundle.signature` -- so a forged signature verified. The
      seal covered the ANSWER and not the QUESTION: `filterApplied`, `reason`
      and `generatedBy` were all rewritable over the same bytes, making a
      single-tenant subset re-presentable as a complete export of everything.
      And the check was only runnable by the party being audited, since the
      chain arm needs `options.chainedStore`, i.e. the live store.
      
      Retention was the same failure pointed at time. `ageInDays` returned
      NaN for an unparseable `occurredAt`, NaN loses all four `<` comparisons
      in the tier ladder, and the record fell through to the final `else`:
      `occurredAt: 'not-a-timestamp'` came back `tier: 'past_retention',
      ageDays: NaN` and landed on `getPurgeCandidates` on its first sweep. A
      far-future date came back `ageDays: 0` and was retained forever. The
      age was keyed on `occurredAt`, which the writer chooses and the ingest
      service copies unbounded beside the `ingestedAt` it stamps itself.
      `redactEvent` returned a redacted COPY and deleted nothing;
      `getPurgeCandidates` returned a list; and `CanonicalAuditEventStore` has
      no delete method -- so "no delete method is the strongest enforcement
      of immutability" is really the absence of a governed deletion path,
      which forces every real deletion through an ungoverned one, while
      `InMemoryCanonicalAuditEventStore.resetForTests()` clears every event on
      the production type behind a comment claiming it "only runs in test
      setup". A legal hold was `RetentionPolicy.legalHold`, one boolean on a
      config row: `update({...policy, active:false})` reported `legalHold:
      false` for a record reported `true` a moment earlier because the
      inactive branch returned before the hold branch, and `unregister(tag)`
      removed the hold outright -- neither leaving a record, naming an
      authority, or being a release. And `redactEvent`'s `'pii'` mode
      replaced `actorId` with `sha256(actorId).slice(0,32)` "so identity
      correlation remains possible", which reproduced
      `sha256('ada@example.com')` truncated to 32 hex EXACTLY -- S3.3.g's
      `content-digest` residue written down as a feature -- while leaving
      `reason` and `metadata`, where the address survived verbatim.
      
      SEVEN DECISIONS. Verification TAKES an anchor and refuses without one,
      reporting `unverifiable` rather than `verified` for every sequence no
      anchor covers, which makes the exposure window a real quantity. A
      witness that is the writer is NOT a witness, so the six witness kinds
      are graded and only three count as coverage. Retention and integrity
      are the same mechanism pointed in opposite directions, so expiry
      PRODUCES a tombstone. The digest that keeps the chain and the digest
      that re-identifies the subject are the same bytes and only a KEY
      separates them: the commitment is keyed with a per-record nonce the
      purge destroys, so the link survives and the lookup key does not, and
      `tombstoneFaults` asks about the nonce by name. The seal covers the
      question and the answer, with completeness four-valued because
      `duplicated` was measured rather than imagined -- the store's cursor
      silently restarts at the head when it cannot find the cursor row, and
      the real `EvidenceExportService` produced a bundle reporting
      `eventCount: 7` over six distinct events with `verifyExport` returning
      `ok: true`, the duplication sealed in and certified. A break stops the
      claim and not the report, so the run emits a row for EVERY entry, three
      checks each, three-valued -- the `IntegrityReport` S7.17 already
      renders and which nothing in this repository could produce. And a
      deletion's record must outlive what it deleted, the one retention rule
      specific to an audit log: if the class recording a purge expires before
      the class it purged, "lawfully deleted" and "never here" become the
      same bytes.
      
      Repairs to the shipped platform, all measured: `verifyChain` now takes
      anchors, walks every entry instead of returning at the first break,
      emits three-valued per-entry rows, and gives `order-violation` its
      first producer; `checkTipAgainstAnchors` catches truncation and
      rollback, which no per-entry walk can. `verifyExport` reads the
      signature, verifies it under a key the CALLER already trusted, seals
      the question in a `custodyHash`, reports distinct-vs-emitted counts and
      a completeness verdict, and returns `unverifiable` rather than `ok`
      when nothing outside the bundle was supplied. Retention refuses to date
      an undatable record instead of purging it, measures age on the
      platform-stamped clock and reports which one it used, decides the hold
      before the policy's own state, and requires a keyed salt for
      pseudonymisation -- dropping the identifier rather than emitting a
      lookup key when none is configured.
      
      Verification: 12,108 workbench-kit tests (101 new), 158 files, all
      passing; 1,000 audit-platform tests (28 new, every prior one still
      passing), 49 files; 0 typecheck errors in both projects under lib and
      spec configs; export map regenerated clean at 162 entry points, 0
      violations; `audit-integrity-scan` exits 0 with all three controls
      passing. The scan reports three columns -- shipped-with-nothing,
      shipped-given-an-anchor, kit -- and the gap between the first two IS
      the cell: with nothing from outside the shipped verifiers miss 6 of 11
      lifecycle actions, and handing the SAME code an anchor or a trusted key
      recovers 4 of them. CONTROL 3 is the strong result: `reviewLifecycle`,
      handed the platform's signal inventory, predicts the measured miss set
      EXACTLY in both columns (6/6 and 2/2, no false positives or negatives).
      The two it still misses under either column -- `duplicate` and `hold` --
      are design gaps rather than missing calls: there is no per-object
      sequence for a replay to repeat, and a hold that IS a config row is
      deleted when the row is.
      
      CONTROL 3 earned its place three times. Its first version predicted
      misses from `needsExternalValue` alone and was WRONG about three
      attacks, because a verifier misses an attack for two different reasons
      -- no witness, or no reading at all -- and that column only covers the
      first. Its second version went stale the moment the platform was
      repaired and caught that too. Its third arm asks the same question of
      THIS MODULE, and the first answer was a finding: `VERIFICATION_SIGNALS`
      named `tip-below-anchor`, `LIFECYCLE_BINDINGS` bound both `truncate`
      and `restore` to it, and `runVerification` never emitted it -- the
      `signal-with-no-producer` fault this module defines, in this module,
      with the scan's kit column quietly computing the detection in the
      harness instead. `runVerification` now produces it (and
      `verificationHeadline` leads with it, because `exposureWindow` counts
      entries ABOVE the anchor and a truncated log has none, so the headline
      would otherwise have claimed full coverage of a log missing its tail).
      All three arms now predict `none`.
      
      Scope stated: the module is `contract`-class, so the digest, the key,
      the timestamp authority, the clock and the store it is decided against
      are the host's. S3.3's retention resolution, hold record and purge
      steps, S7.17's report shape, S10.6's field classes and S9.3's
      `ImpactCount` are composed rather than re-decided.
      
      TRAP FOUND, FIXED IN THIS LIBRARY AND OPEN ELSEWHERE: a raw NUL byte in
      a source file makes it BINARY to grep, so the repository's own
      mandatory stub-indicator scan and `production-scan` read nothing from
      it and report zero hits for the same reason a clean file does. This
      module was written with two (separator bytes in `sealPreimage`) and the
      first adversarial scan over it was therefore meaningless. Fixed here by
      length-framing each member instead of joining on a separator -- which
      also removes a real ambiguity the module's own text forbids -- and in
      `log-privacy.ts` (S10.6) and `trace-continuity.ts` (S10.2) by writing
      the separator as `\u0000`. **33 files repo-wide still have this and are
      invisible to every grep-based gate**, across contracts, euterpe, metis,
      isis, yemaya, sophia, aphrodite and the V-series. -->
      
      NOTE (S10.7, for S10.8): `AUDIT_RECORD_CLASSES` includes
      `billing-event`, and `classRetention` resolves its bounds but nothing
      yet meters it. S10.8 asks for cost attribution with sampled invoice
      reconciliation, and a cost record is an audit record with a number on
      it: its retention row, its `attestsAbout` edge and the
      `attestation-expires-before-its-subject` rule all apply, and a
      reconciliation that outlives the records it reconciled is the same
      defect this cell named.
      
  • S10.8 Add provider, storage, compute, egress, and human-review cost attribution. Evidence: sampled invoice reconciliation.

    • S10.8.a Maintain versioned rate cards and units for provider/model/tool, database/blob/cache/search/queue/log, CPU/GPU/memory, transfer/egress, DCC/license, and reviewer role/time. Evidence: rate registry.

    • S10.8.b Capture metered quantity, unit, source, time window, region/ tier, attempt, job/run/artifact/release, actor/project/tenant/provider attribution, and confidence. Evidence: usage schema.

    • S10.8.c Calculate estimated/reserved/actual/refunded/unallocated cost with currency/rate version and no fabricated zero on missing data. Evidence: calculation tests.

    • S10.8.d Allocate shared storage/compute/egress/observability/license/ reviewer cost through approved reproducible rules and retain raw/derived distinction. Evidence: allocation tests.

    • S10.8.e Reconcile provider invoices, cloud billing exports, storage/ egress meters, license records, and review time samples to internal usage. Evidence: sampled reconciliation.

    • S10.8.f Record variance/missing/late/duplicate/disputed/refund/credit and correction without rewriting prior periods. Evidence: variance tests.

    • S10.8.g Expose permission-aware cost by domain/workbench/project/tenant/ capability/provider/resource/release with privacy and bounded labels. Evidence: cost projection tests.

      <!-- S10.8: shipped as `libs/oshun/workbench-kit/src/cost-attribution.ts`
      (contract-class, entry point `./cost-attribution`, 163 entry points 0
      violations) + a 107-test spec, `tools/cost-attribution-scan.mjs` (nx
      target `cost-attribution-scan`, under tsx), and repairs to
      `libs/shared/ai/src/usage/index.ts`,
      `libs/shared/gpu-dispatcher/src/cost-tracker.ts` and
      `libs/oshun/creative-orchestrator/src/usage-accounting.ts`.
      
      THE VEIN: A COST IS A MEASUREMENT TIMES A PRICE, AND EVERY MISSING
      FACTOR RESOLVES TO A NUMBER INSTEAD OF A REFUSAL. Thirty probes over
      the three shipped meters, twenty-five findings, five green controls.
      Four ways it fails, plus a fifth with no code to point at:
      
      1. AN INVENTED PRICE. `UsageTracker.getPricing` ended with
         `// Default to a reasonable estimate` and `{ 1.0, 3.0 }` per million.
         Twenty opus-class calls through a `maxCost: 1.0` gate settle at
         $3.20 and the gate fires on that, against ~$60.00 of real spend.
      2. A SUBSTITUTED PRICE. `CostTracker.calculateCost` priced an unknown
         GPU at the DEFAULT type's rate: 60s of `'H200'` recorded 2.34c at
         RTX_4090's rate, 7.1x under H100, while the record still said H200.
         The number and the label came from different GPUs.
      3. AN ABSENT PRICE. One line down, `return 0` — an hour of H100 against
         an empty card recorded $0.00 and summed in as real. Its neighbours:
         an untimed job took `calculateDuration` -> 0 -> free, `enabled:false`
         handed the caller a $0.00 record to persist, and cents rounded to
         1/100c made every T4 job under 0.18s free ($28.00 lost per million,
         job count exactly right).
      4. AN ABSENT QUANTITY, in the module that gets everything else right.
         `priceUsage` THROWS when a unit has no rate — the honest seam — but
         the loop opens `units[unit] ?? 0; if (quantity === 0) continue`, so
         THE REFUSAL WAS GUARDED BY THE PRESENCE OF THE VERY DATA WHOSE
         ABSENCE IT EXISTS TO CATCH. Report the unpriced unit and it refuses;
         omit it and the same receipt prices clean. Measured: `priceUsage(…,
         {})` returns `costUsdMicros: 0`, zero line items, basis
         `'metered-rate-table'`; `{}` and `{ cachedInputTokens: 0 }` return
         byte-identical results.
      5. And an unmetered class contributes the most confident zero of all,
         because no branch was taken to reach it. `storageByteSeconds` and
         `egressBytes` are in the unit vocabulary with no producer anywhere,
         and reviewer time has no rate in the repository at all.
      
      THE WITNESS IS THE STATEMENT AND IT IS ONLY A WITNESS OVER THE PART
      THAT DID NOT COME FROM IT. `MeasuredUsageReceipt.costBasis` already
      recorded, per receipt, whether the figure came from our rate table, the
      provider's billing, or an unsettled reservation ceiling —
      `reconcileProviderInvoice` never read it. So a 100% provider-billed
      total reconciled `matched` with drift 0 against that provider's own
      invoice; one 900,000-micro ceiling against a correct 15-micro invoice
      reported drift-alert at -899,985, reading as the PROVIDER under-billing
      us by $0.90; an invoice of 0 with zero receipts reported `matched`.
      
      ALSO MEASURED AND REPAIRED: `cacheReadInputTokens` and
      `cacheCreationInputTokens` arrive on the response, are stored on the
      record, and were read by NOTHING — 100k cache-read tokens on an Opus
      model recorded $0.0525 against a real $0.2025, and 100k cache-WRITE
      tokens $0.0525 against $1.9275 (36.7x under, on the most expensive
      token class this repository's own workload generates).
      
      TRAP, IN MY OWN REPAIR: `ceil(0.028 * (100/1000) * 10_000)` returns 29
      rather than 28, because `0.028 * 0.1` is `0.0028000000000000004` and
      the ceiling amplifies the last bit into a whole unit. The rate is now
      rounded to an integer of micro-cents per second ONCE, where the
      intended value is unambiguous, so every later step is exact. And the
      first pass at the same repair fixed `totalCostCents` while leaving
      `byJobType`/`byGpuType` doing `+= costCents ?? 0` — the same defect one
      field over, caught by the mandatory adversarial scan over my own diff.
      
      The scan reports THREE columns — legacy (the pre-S10.8 behaviour,
      reproduced from the original source lines), shipped, kit — and the gap
      between the first two IS the cell: legacy 0/12 distinguishable, shipped
      12/12, kit 12/12, with a control proving the legacy arm misses
      everything so the harness can be seen to detect rather than only ever
      print clean. -->
      
      NOTE (S10.8, for S10.9): `reviewCostSystem` predicts ten defect classes
      and this harness has scenarios for six. The four unexercised —
      `unversioned-rate`, `unmetered-class-as-zero`, `prior-period-rewrite`,
      `residual-disclosure` — are real and named rather than quietly dropped.
      S10.9 asks for spend and unallocated-cost anomaly alerts, and an alert
      computed over a total that cannot say which of its parts were priced
      inherits every one of them: a burn-rate forecast over a partial
      denominator under-predicts exhaustion by exactly the share nobody
      metered, and `costComplete` is the field that says so.
      
  • S10.9 Add quota and anomaly alerts before hard exhaustion. Evidence: threshold and recovery tests.

    • S10.9.a Register rate/concurrency/storage/compute/GPU/token/provider/ license/queue/spend/reviewer capacity quotas by scope, window, owner, and hard limit. Evidence: quota registry.

    • S10.9.b Define warning/critical forecast thresholds by absolute use, percentage, remaining time, burn rate, queue growth, and provider/reset delay. Evidence: threshold policy.

    • S10.9.c Detect volume/rate/variance/retry/failure/resource/provider/cost/ unallocated and behavioral anomalies with versioned model/rule and uncertainty. Evidence: anomaly fixtures.

    • S10.9.d Alert before exhaustion with affected capability/scope, current/ forecast/limit/reset, owner, dashboard/runbook, mitigation, and safe context. Evidence: alert payload tests.

    • S10.9.e Deduplicate correlated quota/anomaly/dependency symptoms and route/escalate by ownership and severity. Evidence: event storm tests.

    • S10.9.f Test recovery/reset/quota increase/demand reduction/provider recovery and clear only after sustained safe state. Evidence: clock-controlled recovery tests.

    • S10.9.g Verify hard limits still fail closed if warning/alert pipeline is unavailable. Evidence: alert outage test.

      <!-- S10.9: shipped as `libs/oshun/workbench-kit/src/exhaustion-alerts.ts`
      (contract-class, entry point `./exhaustion-alerts`, 164 entry points /
      0 violations) with a 108-test spec, plus
      `libs/oshun/workbench-kit/tools/exhaustion-alert-scan.mjs` (nx target
      `exhaustion-alert-scan`).
      
      THE VEIN: AN ALERT BEFORE EXHAUSTION IS A CLAIM ABOUT HOW MUCH TIME IS
      LEFT, AND EVERY THRESHOLD IN THIS REPOSITORY IS SET ON A LEVEL. "80% of
      the cap" answers how much is gone; the alert exists to answer how long
      you have got, and those coincide only when the burn rate is constant and
      known -- and the burn rate is the one quantity nobody measured. The rule:
      FOR EVERY THRESHOLD, STATE THE LEAD TIME IT BUYS AND THE MITIGATION IT
      HAS TO OUTRUN.
      
      Measured against the shipped code on 2026-08-14, all reproduced:
      
      1. `checkBudget` warns at `utilization >= warnAtFraction`. Against a
         1000-unit cap at 500 units a minute the first `warn` lands at t=95s
         and `exceeded` at t=119s -- twenty-four seconds. `BudgetCheck` was
         `{ verdict, utilization }`, so a caller could not tell twenty-four
         seconds from twenty-four hours.
      2. And at a coarse grain the warning state is UNREACHABLE: from used=700
         of 1000, a request of 50 is `ok` (0.750) and a request of 301 is
         `exceeded` (1.001).
      3. THE DEADLINE WAS ANCHORED TO THE EVENT IT WAS SUPPOSED TO OUTLIVE,
         and it is four words of source: `elapsedSinceCap = nowUnixSeconds -
         meter.lastUsedAtUnixSeconds`. The name says "since cap", the value
         says "since last use", and `consumeBudget` rewrites it on every
         consumption -- a timer that measures from the most recent occurrence
         of the thing it is timing out never expires while that thing keeps
         happening. Measured: 100 units/60s against a 1000-unit cap with a
         300s grace sits at `grace` after 24h at 143.4x over, with
         `buildUsageMeterCard` carrying the identical anchor so the customer
         still reads "Grace window active -- please wrap up in-progress work".
         The only exit is to stop, that is, to have already stopped.
      4. COMPOSED, 2 AND 3 GIVE UNBOUNDED OVERSPEND WITH ZERO WARNINGS. Over a
         30-day period: warn=0, exceeded=0, grace=41,667, final usage 4,168x
         the cap -- and the state it rests in is the one meaning "handled".
      5. A GRACE WINDOW IS A LOAN AGAINST A WARNING, and nothing asked whether
         the warning was delivered. That is S10.9.g from the other side: the
         hard limit must not depend on the alert pipeline, and the soft slack
         must.
      6. `zScore` returned `value === mean ? 0 : Infinity` for a zero-variance
         baseline, so against a baseline flat at 100 both 99 and 0 scored
         +Infinity -- two windows of an agent getting CHEAPER proposed a
         quarantine. `@metis/cost-tracking`'s `detectOutliers` and
         `@lakshmi/transactions` get the identical case right, so it is a
         defect and not a house convention.
      7. And +/-Infinity does not survive the wire: `JSON.stringify` turns it
         into `null`, so the proposal's whole evidence arrived reading as "no
         score computed".
      8. `AnomalyBaseline.sampleCount` was stored and read by NOTHING:
         baselines of n=2 and n=10,000 produced byte-identical verdicts,
         quarantine included, with no sample size, interval or rule version on
         any of them.
      9. `QuotaTracker.TenantState.warned` was declared, initialised `false`,
         reset `false` on window roll, and assigned `true` nowhere -- 100
         requests against a 100-request quota at an 80% warn point produced 21
         identical warnings with `warned` still `false`.
      10. A DEDUP MARK MUST NOT OUTLIVE THE WINDOW IT DEDUPLICATES WITHIN, and
         the ratchet that works is worse than the one that does not.
         `CostCalculator.lastAlertThreshold` suppresses repeats correctly and
         is cleared only by `clear()`/`resetTracking()`, while the window it
         guards is a ROLLING THIRTY DAYS. Measured on the real object: alert at
         95% of a 10,000c budget, take the S10.9.f recovery path and raise it
         to 100,000c, burn 95.5% of the new one -- ZERO alerts. Cross 100% once
         and no budget alert ever fires again for that calculator's lifetime.
         Recovery is the thing that silences the next exhaustion.
      11. `checkBudgetAlert` documents "the highest threshold that has been
         crossed" and assigned unconditionally in array order, so at 95% used
         `[0.5,0.75,0.9]` returned 0.9 and `[0.9,0.75,0.5]` returned 0.5.
      12. Five of the eleven classes -- concurrency, provider, license, queue,
         reviewer -- had no registry entry anywhere, and an unregistered quota
         contributes the most confident silence of all.
      
      TRAPS, BOTH IN MY OWN DIFF, both caught by the mandatory adversarial
      pass: `forecastExhaustion` reported `burnPerMs: 0`, `method:
      'endpoint-exact'` and `resetDetected: false` whenever there was no burn
      estimate -- an unmeasured rate reported as flat, a method that never ran,
      and a negative finding from an examination that never happened, which is
      S10.8's vein reproduced inside S10.9's own repair. And
      `const signed = z ?? 0` in the anomaly repair was a fabricated reading
      one guard away from a decision. Also: `precision: 'exact'` was claimed
      over a fitted float slope, and `dispatch` folded two UNRELATED quotas
      into one page because any non-root alert was suppressed as
      `derived-from-root` -- correlation becoming censorship.
      
      The scan reports THREE columns -- legacy (the pre-S10.9 behaviour,
      reproduced from the original source lines), shipped, kit -- over
      seventeen conditions: legacy 0/17, shipped 12/17, kit 17/17, and it exits
      non-zero if the legacy arm ever scores as well as the kit, which would
      mean the harness had stopped measuring. The five the shipped arm does not
      reach are the ones with no in-tree seam to reach them with: forecasting,
      pipeline-conditional grace, the enforcement split, the unregistered-class
      list and the invented reset instant all live in the kit by design. -->
      
      NOTE (S10.9, for S10.10): the enforcement split this section ships --
      `hardLimitReached` computed before the pipeline is read, soft slack
      withdrawn when it cannot deliver -- is one dependency's degradation plan
      written out by hand. S10.10 asks for that plan for EVERY dependency, and
      the field it will need is the one `QUOTA_WINDOW_KINDS` already forced:
      `concurrent` and `non-renewing` have no reset instant, so "degraded until
      it recovers" has no recovery time to quote and a mode matrix that assumes
      one will invent it.
      
  • S10.10 Add graceful degradation plans for each dependency. Evidence: dependency outage matrix names what remains read-only or unavailable.

    • S10.10.a Inventory each dependency and capability/operation/read/write/ job/gate/publish/release path that uses it. Evidence: dependency- capability matrix.

    • S10.10.b Define trigger/classification for not-configured, slow, partial, stale, rate-limited, unavailable, corrupt, incompatible, and recovery state. Evidence: trigger tests.

    • S10.10.c Decide per capability what remains live read, cached/stale read, local/offline edit, queued operation, read-only, limited, manually handled, or unavailable. Evidence: mode matrix.

    • S10.10.d Define data/gate freshness, authorization, trust, rights/ consent, cost, capacity, and safety constraints that prohibit fallback or fail-open. Evidence: invariants.

    • S10.10.e Implement circuit/timeout/retry/cache/queue/provider alternative and explicit user/API state without fabricated output. Evidence: implementation tests.

    • S10.10.f Define recovery probe, reconciliation/replay/recompute/ revalidation, stale state clearing, user notification, and SLO. Evidence: recovery tests.

    • S10.10.g Fault each dependency and verify exact retained/unavailable behavior across UI/API/job/audit/telemetry/runbook. Evidence: outage report.

      <!-- S10.10: shipped as `libs/oshun/workbench-kit/src/degradation-plans.ts`
      (contract-class, entry point `./degradation-plans`, 165 entry points / 0
      violations) with a 113-test spec, plus
      `libs/oshun/workbench-kit/tools/degradation-plan-scan.mjs` (nx target
      `degradation-plan-scan`), and repairs to
      `libs/shared/rate-limit/src/limiters/graceful.ts` (+ a 20-test spec),
      `libs/shared/rate-limit/src/types.ts`,
      `libs/shared/rate-limit/src/middleware/rate-limit.ts`,
      `libs/shared/http-client/src/circuit-breaker.ts`,
      `libs/shared/http-client/src/types.ts` and
      `libs/shared/resilience/src/circuit-breaker.ts`.
      
      THE VEIN: A DEGRADED ANSWER IS STILL AN ANSWER, AND EVERY ONE OF THEM
      WAS RETURNED IN THE TYPE THAT MEANS "WE ASKED". A degradation plan is a
      sentence about a dependency and the caller never sees the sentence -- the
      caller sees a value with the same type it had when everything worked, so
      the field that meant "the policy permits this" now means "we could not
      reach the thing that decides", and nothing on the value says which. The
      rule: FOR EVERY FALLBACK, NAME THE FIELD ON THE RETURNED VALUE THAT SAYS
      IT IS ONE, AND THE CONSUMER THAT BRANCHES ON THAT FIELD.
      
      The subject is `@oshun/rate-limit`'s `GracefulRateLimiter` -- this
      repository's only component whose name is the thing this cell asks for --
      plus both circuit breakers. Measured on 2026-08-14, all reproduced:
      
      I. THE VERDICT TRAVELS WITHOUT THE MODE.
      1. `RateLimitResult` is `{ allowed, remaining, limit, resetAt,
         retryAfter }` and not one field is provenance. `isDegraded()` is on
         the limiter; the result is what reaches `createRateLimitMiddleware`,
         which writes it into headers and a body. The two part company at the
         return statement.
      2. AND THE ONE NUMBER THAT COULD HAVE HINTED AT IT IS UNREPRESENTABLE IN
         BOTH TRANSPORTS. `createAllowResult()` returned `limit: Infinity`; the
         middleware does `result.limit.toString()`, so the header was literally
         `X-RateLimit-Limit: Infinity` (parses to NaN) and `JSON.stringify`
         rendered the same field `null`. Measured: `{"allowed":true,
         "remaining":null,"limit":null,...}`. S10.9's seventh finding on the
         degradation path.
      3. AND THE OTHER DEFAULT LIES ABOUT WHOSE FAULT IT IS. `deny-all` against
         a dead limiter produced HTTP **429** with `X-RateLimit-Limit: 0` -- the
         user told their quota is zero when the truth is that the quota is
         unknown. Wrong status class, wrong party, and "wait and retry" is the
         mitigation that keeps the dependency down.
      
      II. THE FALLBACK ANSWERS A DIFFERENT QUESTION.
      4. `fallbackStrategy: 'in-memory'` constructed `new InMemoryRateLimiter()`
         = 100/60s, with no field on the config through which the primary's
         policy could be passed. A deployment whose policy is 1000 reported
         `limit: 100`, in the same field, unmarked -- and per process, so the
         effective global limit was 100 x replicas while each client was told
         100. `status()` during an outage returned `{allowed:true, remaining:10,
         limit:10}` for a user who had spent their quota, from a store that had
         never seen the key.
      
      III. A TIMEOUT IS NOT A FAILURE TO ACT.
      5. `withTimeout` neither cancelled nor reconciled. Measured with a 20ms
         budget against a limiter that consumes and then answers in 200ms: the
         primary recorded one token, the fallback recorded another. Two tokens
         for one request.
      6. AND THE TIMEOUT BUDGET WAS DESTROYED BY WRITING THE FIELD. The
         constructor was `{ ...defaults, ...config }` and `operationTimeout` is
         optional, so `operationTimeout: undefined` -- which the type permits
         and every options-forwarding wrapper produces -- reached
         `setTimeout(fn, undefined)`, which fires after **1ms**. Measured
         against a HEALTHY limiter at a realistic 3ms round trip: 20 of 20
         answered by the fallback, `degraded=true`, primary consuming them all
         as well. The same spread disabled the breaker: `failureThreshold:
         undefined` makes `n >= undefined` always false, measured
         `circuit=closed` after 20 consecutive failures.
      
      IV. RECOVERY. `degraded` cleared on ONE success in the closed state. The
      substitute was never cleared or reconciled: six consumes during an
      outage, then a SUCCESSFUL `reset('user:jane')` against the recovered
      primary, and the fallback still held `remaining: 4`. `resetCircuit()`
      assigned `circuitState = 'closed'` and THEN read `previousState` from it,
      so the one event recording a manual intervention always said
      `{previousState:'closed', newState:'closed'}`. And `getCircuitState()`
      never evaluated the reset timeout, so an idle limiter read `open` forever.
      
      V. THE SAME SHAPE TWO LIBRARIES OVER. `@oshun/http-client`'s breaker
      returned `true` from `canExecute()` for every caller in half-open --
      measured **500 of 500** admitted on the first tick after `resetTimeout`,
      so the recovery probe was the entire production load aimed at the
      dependency that had just failed -- and `getMetrics().successes` was the
      half-open probe counter, measured **0** after 100 successful requests
      beside `totalRequests: 100`. `@oshun/resilience`'s breaker handed the
      caller `resetAt = openedAt + resetTimeoutMs` when refusing a half-open
      probe, which is by definition in the past: measured **Retry-After
      -0.5s**, i.e. "retry now", during the one state that exists to admit one
      probe at a time.
      
      VI. AND THE PLAN WITH NO CALLER. `GracefulRateLimiter` is exported from
      `@oshun/rate-limit`'s public barrel and constructed nowhere in this
      repository, so every defect above was latent on an exported API rather
      than live -- which is its own finding and the reason this cell ships the
      matrix rather than a runtime.
      
      VII. AND TWO IN THE KIT'S OWN S10.5 REGISTER, found by turning the rule
      on the prior cell. `dep:trust-store` declares `mitigations: ['timeout',
      'cache']` against `'verify a signature'`, and `dep:identity` declares
      `cache` against `'authorize any request at all'`. Revocation and
      withdrawal are NEGATIVE facts a cached copy can only be missing, so a
      cache in front of either fails open by construction, and neither
      registration carried a staleness bound or a direction. The rule those two
      needed: A CACHE MAY MAKE A DECISION STRICTER, NEVER LAXER.
      `cachedDecisionRisks()` reports both as a function rather than as a
      paragraph, so it stays true as the register changes.
      
      DECISIONS. A MODE IS SERVED BY DEPENDENCIES AND THE GRAPH DECIDES
      WHETHER IT CAN SERVE: `DEPENDENCY_SUBSTRATE` is new here because S10.5's
      `observedThrough` answers "can this dashboard report its own failure" and
      not "does this fallback run through the thing that failed".
      `dep:primary-store-scan` is a real standby for `dep:search-index` and is
      not one for `dep:primary-store`. READ-ONLY IS A PROPERTY OF THE SET AND
      NOT OF ANY PATH -- every effectful path stopped or queued and at least one
      read still answering -- computed by `productMode()` from the rows rather
      than declared beside them; `dep:primary-store` unavailable comes out
      `read-only` with 7 of 7 read paths answering and none of its 8 effectful
      ones, and `corrupt` comes out `unavailable`. UNKNOWN IS A CONDITION: a
      classifier handed no probe answers `unknown`, never `healthy`. CORRUPT
      OUTRANKS UNAVAILABLE and may not be replaced by a copy of itself.
      
      TRAPS, ALL FOUR IN MY OWN DIFF, all caught by the mandatory adversarial
      pass or by the module's own review of its own register:
      (1) the first `not-configured-treated-as-an-outage` check read the
      `because` prose for incident words and REFUSED the sentence "this is a
      decision about what was bought and not an incident" -- a word-level check
      cannot see a negation, so the state is now a `presentedAs` field;
      (2) `path:assisted-action`'s `read-only` fallback was served by
      `dep:model-provider-primary`, which rests on the gateway it was meant to
      survive -- found by my own substrate check, fixed by subtracting the
      closure;
      (3) the repaired `createAllowResult` fell back to `limit: 0` when no
      policy was declared, which is S10.8's vein reproduced inside S10.10's own
      fix -- a missing factor resolving to a number instead of a refusal; the
      policy is now required and the constructor throws without it;
      (4) `recordSuccess` read `consecutiveSuccesses` AFTER
      `transitionTo('closed')` zeroed it, so the counter that gates recovery
      was reset by the transition that indicates recovery -- the run length is
      now captured first.
      Plus one in the shipped register rather than the code: the recovery plan
      was derived from the path's KIND and the matrix from the dependency's
      ALTERNATIVES, so a save path whose only degraded mode is `unavailable`
      declared a replay of work it never accepted. `recoveryFor` now reads the
      modes the matrix actually contains.
      
      The shipped matrix is 41 capability paths over the 19 registered
      dependencies, 82 plans, 9 impairment rows each, reviewed clean:
      `reviewDegradationPlans` returns 0 refusals and 0 repairs.
      
      The scan reports THREE columns -- legacy (the pre-S10.10 behaviour,
      reproduced from the original source lines), shipped, kit -- over twenty
      conditions: legacy **0/20**, shipped **16/20**, kit **20/20**, and it
      exits non-zero if the legacy arm ever scores as well as the kit. The four
      the shipped arm does not reach are the ones with no in-tree seam: the
      substrate graph, the direction a cache may move a decision, the outage
      report's unobserved-versus-unaffected distinction and the
      configuration-versus-incident split all live in the kit by design. -->
      
      NOTE (S10.10, for S10.11): `productMode` calls `dep:primary-store`
      unavailable `read-only` because seven read paths still answer -- from
      `dep:read-cache`, whose contents were written by the store that is now
      down. S10.11 asks for a cross-store recovery point, and the same question
      arrives one layer lower: a backup of the datastore and a backup of the
      cache, the blob store and the search index are four snapshots taken at
      four instants, and `DEPENDENCY_SUBSTRATE` already says which of them is
      derived from which. A restore that reconciles rows against blobs without
      asking which side was authoritative at the recovery point is the same
      defect as a fallback that reports a substitute's number in the field that
      means the policy.
      
  • S10.11 Add backup consistency checks across relational, blob, search, and provenance stores. Evidence: restore drill.

    • S10.11.a Define a cross-store recovery point using database transaction/ log, blob inventory/version, search projection/checkpoint, provenance/ ledger manifest, event/outbox, and key/config markers. Evidence: consistency contract.

    • S10.11.b Verify backup catalogs, objects, checksums, encryption/ decryptability, schema/version, retention, immutability, and required external refs before restore. Evidence: backup preflight tests.

    • S10.11.c Restore relational metadata/revisions/events/audit, blob bytes, provenance manifests/ledger, and rebuild or restore search into isolation. Evidence: restore commands/log.

    • S10.11.d Reconcile rows↔blobs, revisions↔hashes, provenance↔ingredients, search docs↔authoritative projections, tombstones/holds, and tenant counts. Evidence: integrity queries.

    • S10.11.e Rehash stored bytes and validate read/search/history/audit/ lineage/gate/bundle/publish-preparation journeys. Evidence: restored system tests.

    • S10.11.f Measure achieved point/time, missing/inconsistent/orphan/ protected findings, remediation, and next drill. Evidence: signed drill report.

      <!-- S10.11: shipped as `libs/oshun/workbench-kit/src/restore-consistency.ts`
      (contract-class, entry point `./restore-consistency`, 166 entry points /
      0 violations) with a 62-test spec, plus
      `libs/oshun/workbench-kit/tools/restore-consistency-scan.mjs` (nx target
      `restore-consistency-scan`), and repairs to
      `scripts/operations/v1-restore-reconcile.mjs` (+8 tests, 14 total).
      
      THE VEIN: A RESTORE IS A CLAIM ABOUT WHAT CAME BACK, AND EVERY CHECK
      ASKED WHETHER THE SERVICE ANSWERS. A drill proves the estate can be
      rebuilt; what the checks establish is that a process started. Redis
      answers PING, a directory exists, a mirror exits zero, a reconciler finds
      no contradiction between two lists it was handed -- and every one of
      those passes over an EMPTY restore, which is the failure the drill exists
      to catch. The rule: FOR EVERY RESTORE CHECK, NAME THE EXPECTED COUNT IT
      IS COMPARED AGAINST, AND THE INSTANT THAT EXPECTATION WAS RECORDED AT. A
      check with no expectation cannot tell "nothing was lost" from "nothing
      was examined"; an expectation with no instant cannot tell a lost row from
      one written after the backup boundary.
      
      Measured on 2026-08-14, all reproduced:
      
      I. AN EMPTY RESTORE PASSES.
      1. `reconcileRestore` returned `safeToResumeAutonomy: true, blockers: []`
         over SEVEN EMPTY INVENTORIES -- the same verdict a full estate
         produces. `validateRecordArray` is satisfied by an empty array, and no
         field on the result was a count.
      2. TWO DIFFERENT CLAIMS, ONE DENOMINATOR. `safeToResumeAutonomy` and
         `noCompletedMutatingEffectWillRepeat` were both literally
         `blockers.length === 0`, so an EXPIRED APPROVAL turned both false and
         an estate with no provider effects turned both true -- the second
         being true because nothing was examined.
      3. `redisRestoreVerified: true` IS ASSERTED FROM A PING. The box-loss
         drill's report is a jq template with seven literal `true`s; the only
         preceding Redis step is `redis-cli PING`, which an empty Redis answers.
         `objectRestoreVerified`, `fileStateRestoreVerified` and
         `configurationRestoreVerified` rest on a `mc mirror` that succeeds over
         an empty export, `[ -d "${path}" ]` on directories that may be empty,
         and `[ -f stack.env ]` which a zero-byte file satisfies. Only the
         PostgreSQL path has a floor (`[ "${checksum_count}" -gt 0 ]` with a
         matching `database_count` equality): one class of six is counted. And
         `keyRecoveryVerified: true` has no step at all.
      
      II. THE RECOVERY POINT IS ONE INSTANT AND EVERY STORE WAS CAPTURED AT ITS
      OWN.
      4. NOT ONE OF THE SIX MARKERS S10.11.a NAMES EXISTS. `db-backup.sh` sets
         one `TIMESTAMP` and writes `created=$TIMESTAMP` into a manifest
         covering N databases dumped SEQUENTIALLY, each `pg_dump` in its own
         transaction snapshot; neither backup script passes `--snapshot=`, calls
         `pg_export_snapshot()`, or records `txid_current()` or
         `pg_current_wal_lsn()`. `pg-backup.sh`'s per-database `createdAt` is
         `date -u` evaluated AFTER the dump finished -- a wall clock, and a wall
         clock cannot say which committed transactions are inside.
      5. THE ONE NUMBER THAT MEASURES THE SKEW IS WRITTEN AND READ BY NOTHING.
         `offbox-backup.sh` records `latestPostgresAgeSeconds` -- an explicit
         acknowledgement that the relational side is older than the Redis
         checkpoint and MinIO mirror taken in the same run -- and a grep for a
         reader returns the two lines that write it.
         `OSHUN_OFFBOX_HEALTH_MAX_AGE_SECONDS` bounds `now - generated`: the age
         of the RECEIPT, not the skew between the stores it describes.
      6. AND TWO OF THE FOUR STORE CLASSES ARE NOT IN THE BACKUP SET. The
         snapshot declares `postgresql`, `redis`, `minio`,
         `content-service-runs`, `computer-use-snapshots`,
         `configuration-and-secrets`. Search is real (`elasticsearch:8.11.0`,
         `qdrant`) with no marker, no backup and no reconciliation collection.
         Provenance is the sharpest omission: `isis_provenance_ledger` is a hash
         chain (`sequence`, `chain_fingerprint`,
         `parent_chain_fingerprint REFERENCES ...`), so its last
         `(sequence, chain_fingerprint)` IS a recovery-point marker sitting
         unused in the schema -- the one store in the estate that can prove its
         own completeness, and nothing asks it to.
      
      III. AND THE COMPARISON THAT DOES RUN COMPARES TWO COPIES OF THE SAME
      NUMBER. `reconcileExactCollection('artifacts', 'digestSha256', ...)`
      matched restored artifacts to external ones on the digest RECORDED IN THE
      ROW on each side, so an artifact whose bytes no longer hash to their
      recorded digest reconciled clean. S10.7's vein on the restore path.
      
      DECISIONS. A MARKER RECORDS WHAT IS INSIDE AND AN INSTANT ORDERS THE
      CAPTURES, and a recovery point needs both, because ordering is what
      decides which inconsistencies are possible. A WALL CLOCK IS ADMISSIBLE
      FOR NOTHING BUT A CACHE. THE REFERENCED STORE MUST BE CAPTURED NO EARLIER
      THAN THE STORE THAT REFERENCES IT -- which makes inserts produce harmless
      orphans instead of dangling references -- and the safe order is NECESSARY
      AND NOT SUFFICIENT, because a delete in the skew window still strands the
      reference unless tombstone retention is at least as long as the skew. AN
      UNMATCHED ROW ON THE AUTHORITATIVE SIDE IS DATA LOSS AND ON THE DERIVED
      SIDE IS GARBAGE. AGREEMENT BETWEEN TWO DERIVED SIDES IS NOT EVIDENCE
      ABOUT THE RECORD. AN EXPECTATION READ OUT OF THE RESTORED COPY HAS ITS
      CONCLUSION IN ITS PREMISES. RECOVERY TIME IS MEASURED FROM DETECTION and
      the achieved recovery point from the EARLIEST capture, because work
      committed after the oldest store's capture is not uniformly recoverable.
      
      REPAIRS. `v1-restore-reconcile.mjs` now takes an optional `expectations`
      block -- an old input still PARSES and no longer PASSES, which is the
      fail-closed direction -- refuses a count read from the restored copy or
      carrying no instant, reports an `examined` block so a clean verdict reads
      as coverage, splits `noCompletedMutatingEffectWillRepeat` from
      `safeToResumeAutonomy` so an expired approval no longer refutes it and an
      unexamined effect set no longer confirms it, and refuses an artifact whose
      stored bytes were never rehashed. Result schema bumped to `.v2`.
      
      TRAP, IN MY OWN MODULE, caught by the adversarial pass: every check in
      `recoveryPointFaults` is a loop over markers or over required classes, so
      a recovery point with NO markers and a caller who required NO classes
      passed all of them -- this cell's own defect, in this cell. There is now
      an unconditional `recovery-point-with-no-markers` refusal ahead of the
      loops.
      
      The scan reports THREE columns -- legacy (the pre-S10.11 reconciler,
      reproduced from the original source lines), shipped, kit -- over sixteen
      conditions: legacy **0/16**, shipped **9/16**, kit **16/16**, and it exits
      non-zero if the legacy arm ever scores as well as the kit. The seven the
      shipped arm does not reach need a per-store capture marker, a reference
      graph, a refutability declaration and a detection instant, none of which
      the shell tooling has a seam for; they live in the kit by design. -->
      
      NOTE (S10.11, for S10.12): this cell measures a drill that an operator
      chose to run, on a host they prepared, at an instant they picked. S10.12
      asks for game-day scenarios with a measured DETECTION time, and
      `DrillTiming` already refuses to compute a recovery time without one --
      the field is there and the estate has no producer for it. The scenario
      register S10.12.a wants is the same shape as S10.10's mode matrix one
      layer up: a scenario that no capability path traverses is a rehearsal for
      an incident nobody can have.
      
  • S10.12 Add disaster-recovery runbooks and game-day scenarios. Evidence: measured RPO/RTO and follow-up issues.

    • S10.12.a Define disaster scenarios for regional/service/data/key/queue/ provider/search/blob/database/bridge compromise or outage and per- capability RPO/RTO. Evidence: DR scenario register.

      <!-- S10.12.a: shipped as `libs/oshun/workbench-kit/src/game-day.ts`
      (contract-class, entry point `./game-day`, 167 entry points / 0
      violations) with a 118-test spec, plus
      `libs/oshun/workbench-kit/tools/game-day-scan.mjs` (nx target
      `game-day-scan`, legacy **0/17**, shipped **5/17**, kit **17/17**), and
      repairs to `scripts/isis/run_runpod_disaster_recovery_gameday.sh`,
      `run_provider_outage_drill.sh` and
      `run_endpoint_region_failover_drill.sh` with 12 tests in
      `scripts/operations/v1-game-day.test.mjs` (wired into
      `verify:operations:v1-disaster-recovery`).
      
      THE VEIN: A GAME DAY IS AN EXPERIMENT, AND EVERY ONE OF THESE REPORTED A
      RESULT WITHOUT PERFORMING IT. The rule: for every scenario, name the
      capability path its fault interrupts and the observation that would have
      refuted its pass.
      
      The register itself is `WORKBENCH_GAME_DAY_SCENARIOS` -- ten scenarios,
      one per class S10.12.a names, every `targets` entry a dependency id in
      the S10.5 register -- and `WORKBENCH_RECOVERY_OBJECTIVES`, a recovery
      point and recovery time for each of the 41 capabilities in S10.10's path
      register. Which capabilities a scenario interrupts is DERIVED through the
      substrate closure rather than declared, so a scenario cannot claim an
      impact the dependency graph does not support, and an optional use is
      reported as degradation rather than interruption.
      
      MEASURED, BEFORE THE REPAIRS (all reproduced by running the code):
      
      1. The invocation `docs/reference/ci-cd.md` line 248 gives --
      `STRICT=false DRY_RUN=true RUN_PROVIDER_FAILURE_SCENARIO=false
      RUN_REGION_FAILURE_SCENARIO=false` -- emitted `"summary": {"total": 2,
      "pass": 1, "fail": 0, "skip": 1}`, `"overall": {"pass": true}` and logged
      `all game-day checks passed`, having injected no fault and executed no
      scenario.
      
      2. The predicate is the defect and it is the same in all three scripts:
      `pass: (all(.[]; .status != "fail"))` reads the absence of a failure as
      the presence of a success, so a run in which everything was skipped is
      the strongest pass the report can express.
      
      3. `DRY_RUN` defaults to `true` in all three, and in dry run the
      mutations are not applied AND every verification records `skip`.
      
      4. NEITHER DRILL COULD HAVE INJECTED ANYTHING IF IT HAD TRIED: both gate
      on `jq -r '.isReadOnly // true'`, and jq's alternative operator treats a
      literal `false` as absent, so the expression returns `true` for EVERY
      input. `echo '{"isReadOnly":false}' | jq -r '.isReadOnly // true'` prints
      `true`. Both drills therefore always took the `Environment is read-only
      for feature flags` branch. Reproduced against a stub admin API: with the
      gate repaired the three flag PATCHes go out and `drill.simulate_outage`
      passes; as shipped, none do.
      
      5. Nothing schedules any of them -- no workflow, no package.json script,
      no cron -- so S10.12.e's ROTATION has no second run to rotate to.
      
      6. `libs/neith/gameday`'s `SCENARIO_LIBRARY` targets `api`, `db`, `web`,
      `cache`, `auth`, `payments`, `edge`: not one is a dependency id.
      Four of the ten classes (queue, search, blob, bridge) have no scenario;
      the objectives are prose graded by a facilitator typing five numbers into
      `scoreDrill`; and the library has NO CALLER anywhere in the repo.
      
      7. Neither mutating drill had a `trap`, so any failure between injection
      and the linear restore step near the end exited with the fault applied.
      
      8. All ten rollback rehearsal records under
      `docs/releases/v1/verification/rollback-plan/2026-05-15/` carry
      `rehearsed-at` and `observed-RTO` and no detection instant -- S10.11's
      finding, restated -- and there is no fencing anywhere in the estate: a
      repo-wide search for `split-brain`, `fence`, `fencing`, `STONITH` returns
      GDPR subject fences and geofencing.
      
      REPAIRED IN THIS CELL, WITH TESTS: the game day's verdict is now
      three-valued (`passed` / `failed` / `inconclusive`) and a skip is not a
      pass; a scenario whose own drill reported `inconclusive` is not counted
      as executed; a drill that ran and FAILED is reported as a failure rather
      than as a missing artifact (`jq -e '.overall.pass'` exits non-zero on a
      literal `false`); the read-only gate honours an explicit `false`; both
      mutating drills reverse their injection from an EXIT/INT/TERM trap, and
      the handler claims itself BEFORE restoring, because the restore issues
      its own requests and a second signal arriving during them re-entered it
      without bound (found by the test, in my own code); and the report now
      carries detection latency, decision latency and a recovery time measured
      from the fault when the operator supplies the instants, and refuses to
      call an injected run conclusive without a detection instant.
      
      KEY DECISIONS worth reusing. NO CAPABILITY CAN BE PROMISED A RECOVERY
      POINT TIGHTER THAN THE CAPTURE SKEW, whatever tier it is in, because
      S10.11 established that the honest achieved recovery point is the
      EARLIEST store capture -- so every objective in the register carries the
      same recovery point and the tiering lives in the recovery TIME, and a
      register with no measured skew is refused outright rather than graded.
      THE RESTORE ORDER IS THE REVERSE OF S10.11'S CAPTURE ORDER AND NOT THE
      SAME RULE: capture the referring store first so inserts leave orphans,
      restore the referenced store first so no row is exposed pointing at bytes
      that are not back. A GAME DAY KNOWS WHEN THE FAULT STARTED BECAUSE IT
      CAUSED IT, so its recovery time runs from injection, where S10.11's runs
      from detection -- detection is the earliest instant a drill can honestly
      start from when nobody observed the onset, and using it here would
      discount exactly the interval the drill exists to measure.
      
      TRAP, AGAIN IN MY OWN MODULE: the adversarial pass found four checks that
      could not fire on the input whose defect they exist to catch -- a plan
      section silent when a scenario had no plan at all, an observer with no
      declared route skipped rather than reported, a duplicate recovery
      objective silently overwriting, and a `recoveredTo` after the fault
      producing a negative recovery point nothing refused. All four are now
      faults with tests.
      
      A SECOND PASS OVER MY OWN EXPORTS found three more of the same shape:
      `GameDayReviewInput.objectiveRegister` was a field `reviewGameDay` never
      read and `GAME_DAY_OBJECTIVE_REGISTER` a re-export nothing consumed
      (both deleted -- a declared input nothing reads is the same claim as a
      check that never runs); `RUNBOOK_SECTIONS` was a closed list nothing
      iterated, now wired through `missingRunbookSections`, which asks which of
      S10.12.b's eleven sections a runbook is silent on rather than what is
      wrong with the ones it has. Writing that function exposed a real gap:
      `recoveryInfrastructureDependencyId: null` means "the recovery runs in
      place", and nothing refused running in place for a REGIONAL scenario,
      where in place is exactly what is gone. It does now, and a `dep:`-prefixed
      value the substrate does not contain is refused as well, because a typo
      in that field wears the shape of an answer. -->
      
    • S10.12.b Document detection/declaration authority, communications, dependency assumptions, recovery infrastructure/secrets, restore order, fencing, validation, failover, and rollback/failback. Evidence: versioned runbooks.

      <!-- S10.12.b: shipped as nine new runbooks under `docs/runbooks/`
      (`v1-dr-signing-service-outage`, `v1-dr-relational-corruption`,
      `v1-dr-signing-key-compromise`, `v1-dr-queue-loss`,
      `v1-dr-model-provider-outage`, `v1-dr-search-index-loss`,
      `v1-dr-blob-store-loss`, `v1-dr-database-loss`,
      `v1-dr-bridge-host-loss`), `v1-disaster-recovery.md` raised to version 2,
      `libs/oshun/workbench-kit/tools/runbook-scan.mjs` (nx target
      `runbook-scan`, `pnpm run verify:operations:v1-runbooks`), a rewritten
      16-test `scripts/operations/v1-dr-runbook.test.mjs`, and a repair to
      `WORKBENCH_GAME_DAY_SCENARIOS` with two tests. All ten registered
      scenarios now have a runbook; `runbookCoverageFaults` returns [].
      
      THE VEIN: A CONTRACT THAT GRADES ONLY ITS OWN FIXTURES SAYS NOTHING
      ABOUT THE ESTATE. S10.12.a shipped `Runbook`, `runbookFaults`,
      `missingRunbookSections` and `runbookCoverageFaults`, and every caller
      outside the module was `game-day.spec.ts` building `runbookOf()` by
      hand. Twenty-one fault kinds, none of them falsifiable, because no real
      runbook was ever passed in. The rule: a check whose only input is a
      fixture is the same claim as a check that never runs.
      
      THE BINDING. Each runbook publishes its structured statement as a fenced
      JSON block after an `<!-- oshun:runbook -->` marker, IN the document
      rather than beside it, so the statement the checker grades and the
      statement the operator follows are one object. `runbook-scan.mjs` parses
      it, grades it with the shipped contract, and refuses an undeclared
      finding, a DECLARED finding that stopped firing, a scenario with no
      runbook, two runbooks for one scenario, a silent section, and a
      dependency id the prose never mentions. Eight negative controls mutate
      the REAL statements and the scan exits 2 rather than reporting if any
      goes undetected -- including a guard that refuses a control whose fault
      is already present unmutated, since that control proves nothing.
      
      THE REGISTER WAS WRONG, AND THE RUNBOOKS ARE HOW IT SHOWED.
      `gd:regional-loss` listed four targets. `docker/docker-compose.yml`
      declares postgres, redis, minio, elasticsearch, kafka, prometheus,
      grafana, loki, tempo, jaeger and every API in ONE file with no placement
      constraint, which is one host. The omission was not a smaller claim, it
      was a WRONG one in the direction that flatters the estate: without
      `dep:identity` the model said a total box loss leaves SIGN-IN WORKING,
      and without `dep:notification-channel` it said the messages announcing
      the outage travel over a sender that is inside it -- so
      `communication-over-the-failed-dependency` and
      `authority-reachable-only-over-the-failed-dependency` could not fire on
      the one scenario they exist for. Four targets added, each citing what
      puts it on the box. The repair immediately produced three findings the
      fixture had been hiding: an observer blinded by the fault it watches
      (the panel reports over `dep:telemetry-pipeline`, which is now inside
      the blast radius -- which is exactly why every core dependency in the
      S10.5 register carries an `outOfBandCheck` naming a prober outside the
      collector failure domain), and two recovery-time breaches, because a box
      loss now interrupts "sign in", whose promise is fifteen minutes.
      
      MEASURED, AND EACH IS A RUNBOOK SECTION SOMEBODY WOULD OTHERWISE HAVE
      WRITTEN WRONG:
      
      1. THE PAGE THAT WOULD ANNOUNCE A BOX LOSS IS EVALUATED ON THE BOX.
      `docs/operations/alert-routing.md` routes `Prometheus up == 0 for any
      production target` to PagerDuty `oshun-platform` at SEV-1, and Prometheus
      is a service in the same compose file. The highest-severity route in the
      estate is the one `gd:regional-loss` removes first. Its `CloudWatch RDS`
      rows describe an AWS deployment; V1's PostgreSQL runs in docker.
      
      2. THE C2PA VERIFIER READS THE PUBLIC KEY OUT OF THE MANIFEST IT IS
      VERIFYING. `verifyAudioC2paManifest`
      (`apps/oshun/bff/src/generation/audio-c2pa-signer.ts`) calls
      `createPublicKey(manifest.publicKeyPem)`. Nothing compares `signerKeyId`
      or `publicKeyPem` against an allowlist and `dep:trust-store` has no
      implementation, so ROTATION DOES NOT REVOKE: manifests signed under a
      disclosed key keep verifying, and the disclosed key mints new ones that
      verify. S10.7's vein, in the key scenario.
      
      3. THERE IS A STANDBY-PROMOTION SCRIPT AND NOTHING TO PROMOTE.
      `infra/yemaya/postgres-ha/failover.sh` runs `pg_ctl promote`; the V1
      compose declares one PostgreSQL service and no standby, replica or
      pooler, and nothing outside `infra/yemaya/postgres-ha/` references that
      directory. `check_standby` refuses, correctly. Two of its steps would
      mislead if it did run: `update_pgbouncer()` logs "PgBouncer update
      requires manual configuration or automation" and RETURNS 0, so
      `do_failover` proceeds as though the pooler were repointed; and
      `promote_standby` calls `read -p` on lag > 60s, so it blocks on stdin
      under automation. Its split-brain control is a human typing `yes`.
      
      4. THE V1 JOB QUEUE IS NOT A BROKER. `jobs-route.ts` holds jobs in an
      in-process map and persists the whole set through `createSnapshotSink`
      into the `admin_store_snapshot` table -- so the durable record is in
      `dep:primary-store`, and a queue loss is a reload rather than a replay.
      
      5. `docs/operations/on-call.md` EXISTS, and an earlier draft of these
      runbooks said it did not. It describes a weekly primary/secondary
      rotation and an escalation ladder, names NO person, and holds no rotation
      config while instructing the reader to "open a PR against the rotation
      config in this doc". None of the register's four `rota:` identifiers
      appears in it or in `alert-routing.md`. The claim in every runbook is
      therefore the narrow, checkable one -- the name is registered, the route
      from it to whoever carries the pager is not -- and the test asserts THAT,
      not the absence of on-call documentation.
      
      THE ABSENCE REGISTER is the unusual half. Eight claims each runbook makes
      about what this repository does NOT have (no reindex entry point, the
      search index in no backup, no queue drain or replay script, no Postgres
      standby, the pgbouncer no-op, the self-attesting verifier, the fail-closed
      signer, the unused PITR config) are predicates in
      `v1-dr-runbook.test.mjs`, and each FAILS the day its absence stops being
      true. A document describing a gap the estate has closed misleads in the
      direction of caution, which is still misleading.
      
      TRAPS IN MY OWN WORK, both found by testing rather than by reading. The
      prose-vs-statement cross-check compared the statement against the whole
      document -- which CONTAINS the statement -- so every dependency id
      trivially "appeared in the prose" and the check could never fail; fixed
      to compare against the text before the marker, whereupon it immediately
      caught that `v1-disaster-recovery.md` §6 named three stores in prose and
      ten in its restore order. And `RUNBOOKS[].prose` was a field nothing read
      until that fix, which is the S10.12.a trap repeated: a declared input
      nothing reads is a check that never runs. Also corrected: a claim that
      `providerEffects` is the reconciler's first collection (it is the second
      of seven), and a sentence asserting "zero matches for replica" in the
      compose file, which matches `KAFKA_OFFSETS_TOPIC_REPLICATION_FACTOR` --
      an answer about the wrong subject in the right shape, in a paragraph
      warning about exactly that.
      
      AND THE LARGEST TRAP WAS IN THE CLAIM I INHERITED AND REPEATED. Three
      runbooks said "there is no fencing anywhere in this estate", which is
      S10.12.a's finding 12 restated. It is a claim about the WHOLE REPOSITORY
      supported by a grep over THREE DIRECTORIES for FOUR WORDS
      (`split-brain|fence|fencing|STONITH` across `docs/`, `infra/`,
      `scripts/`), and the estate's one real fencing token is in neither the
      directories nor the vocabulary: `live_media_pipeline_job_attempts`
      `.lease_token`
      (`libs/shared/live-media/src/sql-media-pipeline-job-store.ts`) carries
      `AND lease_token = $4 AND status = 'leased'` in the `UPDATE WHERE` of
      every heartbeat, lease expiry and completion, so a worker holding a
      superseded lease updates ZERO ROWS. That is textbook lease fencing and it
      works. What is absent is the STORE-LEVEL equivalent -- nothing rejects
      the writes of a host that returns after its replacement was promoted --
      which is what `FencingStep` asks for. All four fencing sections now scope
      the absence and NAME the token, because "the estate has never done this"
      and "the estate has not done this here" are different sentences and only
      the second is true; the second also tells the reader which pattern to
      copy. Finding 12 in the module header is narrowed in place, and the
      token's continued existence is a predicate in the absence register, so
      the narrowed claim is refutable in both directions. Same shape as the
      Yemaya vein: A CLAIM ABOUT THE WHOLE REPOSITORY HAS TO BE CHECKED
      AGAINST THE WHOLE REPOSITORY. -->
      
      NOTE (S10.12.b, on the authority sections): every runbook names its
      authority from the S10.5 dependency register (`workbench-platform
      on-call`, `delivery pipeline on-call`, `production operations duty
      officer`) rather than declaring it UNASSIGNED as S10.12.a's runbook did.
      That is a change of position and it was deliberate: the register already
      carries an owner with a role, a name and a `rota:` contact for every
      dependency, so writing UNASSIGNED was under-reporting a fact the estate
      had. What it does NOT have is a resolution from those identifiers to a
      person, and that is stated in every runbook and machine-checked. Naming
      the registered rota is reading the register; inventing a person would
      have been the defect S10.12.a warned about.
      
    • S10.12.c Document reconciliation/replay for jobs, leases, outbox/inbox/ events, provider outcomes, pending uploads, gates, and releases. Evidence: state-specific procedures.

      <!-- S10.12.c: shipped as `docs/runbooks/v1-reconciliation-and-replay.md`
      (nine state-specific procedures, one per subject, publishing a
      machine-checkable statement after an `<!-- oshun:replay -->` marker),
      `scripts/operations/v1-replay-inventory.mjs` (the missing PRODUCER, run
      against the live database), `libs/oshun/workbench-kit/tools/replay-scan.mjs`
      (nx target `replay-scan`, `pnpm run verify:operations:v1-replay`, 10
      negative controls), a 23-test `scripts/operations/v1-replay-procedure.test.mjs`
      wired into `verify:operations:v1-disaster-recovery` (66 tests pass), and
      a repair to `replayFaults` with 4 new spec tests (129 pass).
      
      THE VEIN: A STATE CLASS IS NOT A TABLE NAME. Every claim that this
      estate lacks a state class was a search over `CREATE TABLE` names, and
      V1's durable state lives in ONE key-value table -- `admin_store_snapshot`,
      measured at 415 rows under 52 store keys -- where a second job store,
      thirty-eight crypto invoices carrying on-chain transaction ids and the
      editorial release streams all sit invisible to that search. The rule:
      for every state class, enumerate the STORE KEYS as well as the tables,
      and name the query that produces the inventory.
      
      AND THE SECOND HALF, WHICH IS WHY THE CELL NEEDED A PRODUCER RATHER THAN
      ONLY A DOCUMENT: `scripts/operations/v1-restore-reconcile.mjs` is named
      as `nextRequiredControl` by `box-loss-restore-drill.sh` and by every
      S10.12.b runbook, and NOTHING IN THIS REPOSITORY PRODUCES ITS INPUT. A
      grep for `oshun.v1-restore-reconciliation-input.v1` returns the
      validator, its own unit test, and the kit scan that builds a synthetic
      one. The operator following the runbook reaches the required control and
      the step has no first action. IT HAS NO PRODUCER FOR A REASON THAT
      DECIDES WHAT TO BUILD: its seven collections match on `externalRef` or
      `idempotencyKey`, and the V1 schema's 146 tables contain NO COLUMN of
      either name (`external_refs`, plural and JSONB, is a different thing on
      two tables); four of the seven have no table at all; and
      `artifact_detection_report`, the nearest name match for `artifacts`, is
      about visual artifacts in generated media. The input describes an estate
      this is not, and writing a producer for it would mean inventing the
      columns to fill it. So the producer that shipped reads the NINE state
      classes instead, and the document says plainly that it is not a drop-in
      for the reconciler.
      
      MEASURED AGAINST THE RUNNING DATABASE on 2026-08-14 (docker
      `oshun-dev-postgres`, `oshun_dev`), all reproduced by running the code:
      
      1. jobs: 63 records over two store keys, 24 IN FLIGHT -- `generation-jobs`
      8 (7 failed, 1 queued) and `metis-ingest-jobs` 55 (32 completed, 23
      `awaiting_source_content`). Only the first has a boot recovery.
      
      2. provider-outcomes: 38 crypto invoices -- 20 expired, 10 PENDING, 8
      confirmed and all 8 carrying a `txId`. This is the one subject in the
      estate whose external side genuinely outlives a restore of ours and can
      be queried independently, and nothing reconciles against it.
      
      3. leases: the estate's one real fence (`lease_token`, S10.12.b's
      finding) is correct, and NOTHING RUNS IT. `recoverExpiredLeases` is
      reachable only through `MediaPipelineWorker` -> `createLiveMediaPipelineRuntime`
      -> `apps/aphrodite/streaming`, which appears in no compose file, no
      infrastructure manifest and no workflow; `docker/docker-compose.yml`
      declares no such service; and the relation exists in no database on the
      box, because the compose declares an `aphrodite` database the running
      server does not have. The table is created by `libs/aphrodite/database`,
      not by V1's schema -- which the scan caught in my own statement.
      
      4. releases were recorded as absent by the S10.12.a groundwork on the
      strength of "no `CREATE TABLE` matching `release` anywhere". That is
      true of the table names and FALSE ABOUT THE ESTATE:
      `v1_metis_publication_package` carries `release_id` and `release_history`
      as COLUMNS, and `admin-editorial-release-streams` is a store key. The
      same error as S10.12.b's fencing claim, one level down. Corrected in
      place, and both halves are now predicates.
      
      5. jobs' boot recovery reconciles the JOB and not the two Redis ledgers
      the job wrote before dispatch: `activeQuotaGate.admit()` and
      `activeBudgetGate.reserve()` both run pre-dispatch and nothing releases
      either, so a job held for review after a restore keeps its quota slot
      and its estimated spend.
      
      THE CONTRACT REPAIR. `externallyVisible` was read on exactly ONE branch
      of `replayFaults` -- the `kind: 'none'` branch -- and ignored on the
      other four, so an effect the outside world saw and a purely internal one
      were graded identically the moment a procedure declared any
      reconciliation at all: a replay that verifies itself by counting our own
      rows passed every check. Added `externalAuthority` and
      `externally-visible-effect-verified-only-from-our-own-records`. Two of
      the nine subjects now DECLARE that fault rather than silence it (jobs has
      no provider-side receipt; releases have no readback from
      `dep:external-publish-target`), and the declaration fails in both
      directions.
      
      TRAPS IN MY OWN WORK, all found by the adversarial pass rather than by
      reading. (a) `inFlight: measured.reduce((sum, m) => sum + (m.inFlight ?? 0), 0)`
      -- THE `?? 0` VEIN, IN THE PRODUCER SHIPPED BY THE CELL THAT QUOTES IT:
      a subject whose stores carry no state field reported `inFlight: 0`,
      which says none was in flight where the truth is that the store cannot
      answer. Now null, with `inFlightMeasuredBy` naming which stores counted.
      (b) The producer answered "is this class absent" out of ITS OWN
      REGISTER; it now queries the live schema and reports
      `declared-absent-and-contradicted`. (c) `compareInventories` accepted a
      restored side with MORE records than the source as `reconciled` --
      duplication is the failure every idempotency key here exists to prevent,
      and it was the one outcome the comparison could not report. (d) In the
      scan, the dependency-id prose check sat AFTER the absent branch's
      `continue`, so for three of the nine subjects it could not fire; moved,
      with its own control. (e) Two bindings for one subject silently kept the
      last. (f) I wrote that `admin_store_snapshot.version` is "the one
      per-store marker the estate has, and S10.11 recorded it had none" --
      an overreach: S10.11's subject was which committed transactions are
      inside a dump, which a per-document counter cannot answer. Narrowed in
      place to what it does do (order two readings of one store key).
      
      AND THE VEIN CAUGHT THE DOCUMENT THAT STATES IT. §§3-4 shipped in the
      first commit saying "there is no outbox in V1" and "there is no
      consumer-side deduplication table", on the strength of no `CREATE TABLE`
      matching `outbox` or `inbox` -- the exact error §0 of the same document
      warns about, made two sections later. **V1 HAS BOTH AND THEY ARE
      COMPLETE.** `apps/oshun/bff/src/metis/academic-integrity-appeal-store.ts`
      (store key `metis-academic-records`) keeps `operations`, an append-only
      log carrying a `revision`, and derives `notificationIntents` from it by
      REPLAY, so a verdict transition and the notification intent it produces
      are one durable operation -- transactional publication, without a table
      named for it. `deliveredNotificationIds` is the acknowledgement set and
      `listPendingNotifications()` is literally the intents minus it, which is
      the inbox pattern. `reconcilePendingAcademicIntegrityNotifications()` is
      called from `server.ts` at boot, deliberately after both durable stores
      hydrate, and returns `{ attempted, pending }` with `pending` RECOMPUTED
      after the drain rather than assumed. The idempotency key is the stable
      `messageId`, and the module says so: "a crash after message commit but
      before outbox acknowledgement is retried as a harmless duplicate on
      restart". What it lacks against `study.domain_event_ack` is `attempts`
      and `dead_letter`, so a poison intent is retried at every boot forever
      with nothing counting the failures. Corrected in the document, the
      register, the statement and the tests, and the correction is left
      visible in §0 rather than rewritten away -- it is the strongest
      argument for the rule that this cell produced. Only `pending-uploads`
      remains declared absent, re-checked with a second vocabulary
      (`multipart`, `presigned`, `uploadId`) rather than the first.
      
      THE SAME SELF-REFERENCE BIT THREE TIMES IN ONE PREDICATE. "Nothing
      produces the reconciler's input" is a claim about the repository, and
      the claim's own artifacts kept entering its result set: first the TODOS
      note quoting the schema string, then the new producer's header comments,
      then the test's own search pattern. Narrowed to the ASSIGNMENT form and
      scoped away from markdown and from itself, because naming a thing is not
      producing it.
      
      AND A CHECK IN S10.12.b'S OWN TEST WAS WRONG, exposed by this document
      being written: `no other runbook still claims to cover the V1 recovery
      path` asserted an exclusion BY NAME while its comment stated a
      REACHABILITY requirement. Under the name rule the only way to satisfy it
      was to leave the reconciliation handoff undocumented, which is the
      condition that whole file exists to end. Repaired to assert what it
      meant -- a document naming the recovery path must be LINKED from a
      scenario runbook -- which is the stronger check, and
      `v1-disaster-recovery.md` §10 now links here. -->
      
    • S10.12.d Prepare synthetic/production-like data, fault controls, observers, timers, safety boundaries, abort, evidence capture, and success criteria. Evidence: game-day plan.

      <!-- S10.12.d: shipped as `docs/runbooks/v1-game-day-plan.md` (ten
      plans, one per registered scenario, publishing a machine-checkable
      statement after an `<!-- oshun:game-day-plan -->` marker),
      `scripts/operations/v1-fault-control.mjs` (the missing INJECTOR: six
      controls, each applied/confirmed/reversed on this box, with a durable
      reversal ledger and a `--self-test` that stands up its own containers),
      `libs/oshun/workbench-kit/tools/plan-scan.mjs` (nx target `plan-scan`,
      `pnpm run verify:operations:v1-game-day-plan`, 13 negative + 4 positive
      controls), a 14-test `scripts/operations/v1-game-day-plan.test.mjs`
      wired into `verify:operations:v1-disaster-recovery` (86 tests pass), and
      contract repairs in `game-day.ts` with 15 new spec tests (144 pass).
      
      THE VEIN: A FAULT CONTROL IS A CLAIM THAT SOMETHING BROKE, AND EVERY
      FIELD IN THE CONTRACT IS A SENTENCE ABOUT THE COMMAND RATHER THAN A
      READING FROM THE ESTATE. The rule: for every fault control, name the
      observation that distinguishes the injected estate from a healthy one,
      and confirm the injection by TAKING it. S10.12.a's rule was about the
      scenario's pass; this is the same rule one level down, about the
      injection itself.
      
      MEASURED ON 2026-08-14 against throwaway containers on this box, all
      reproduced by `--self-test`:
      
      1. `docker exec <c> kill -9 1` EXITS 0 AND KILLS NOTHING. The kernel
      discards SIGKILL sent to a PID namespace's init from inside it, so the
      process the drill meant to crash keeps serving. `kill -TERM 1` DOES
      work, because postgres installs a handler. The strongest signal is the
      one that does nothing and it is the one an operator reaches for.
      
      2. THE OBVIOUS CONTROLS DO NOT PRODUCE THE SAME FAULT, and there are
      THREE client-visible shapes, not one: `docker stop` -> `could not
      translate host name` in 539ms; `docker network disconnect` -> the same
      in 359ms; `docker pause` -> `timeout expired` in 5407ms; `iptables
      -j DROP` -> the same in 5370ms; `iptables -j REJECT --reject-with
      tcp-reset` -> `Connection refused` in 316ms. A host that is really gone
      KEEPS ITS NAME, so `docker stop` rehearses a graceful deregistration
      and measures a recovery time with every caller's connect timeout
      removed from it. `name-unresolvable` matches no scenario in the
      register; the control is registered anyway, so a plan naming it is
      REFUSED rather than left undescribed.
      
      3. ZEROING ONE 8KiB HEAP PAGE TOOK 185 OF 2000 ROWS OUT OF
      `select count(*)` WITH NO ERROR, `max(id)` still 2000 and the
      connection healthy. It is silent because `data_checksums` is `off` --
      which is how the RUNNING estate is configured -- so no drill on this
      estate can be graded by "did the store stay up".
      
      4. A TRAP IS A PROMISE A PROCESS MAKES ABOUT ITS OWN DEATH. S10.12.a's
      own repair added EXIT/INT/TERM traps to the two mutating drills; a
      drill that paused a container and was then SIGKILLed left it `paused`
      with nothing to unpause it. SIGKILL, OOM and reboot are exactly the
      deaths a trap cannot honour, and `reversesOnAbort: true` is satisfied
      by a trap. The driver writes the reversal record BEFORE injecting, so
      `--reverse-all` drains it from an unrelated process.
      
      5. THE ESTATE DEPLOYS NO TELEMETRY PIPELINE AT ALL.
      `infra/hetzner/docker-compose.yml` -- the compose the deploy workflow
      and every `v1-dr-*` runbook name -- has ZERO mentions of Prometheus,
      Grafana, Loki, Tempo, Jaeger, OTLP or Sentry. The S10.12.a register put
      `dep:telemetry-pipeline` in the regional blast radius citing
      `docker/docker-compose.yml`, which is the DEVELOPMENT compose (44
      mentions) and describes a different estate. The conclusion survives and
      the reason is worse: AN OBSERVER INSIDE THE BLAST RADIUS AND AN
      OBSERVER THAT WAS NEVER DEPLOYED PRODUCE THE SAME SILENCE, and only the
      first has a name in the contract.
      
      6. All 14 docker healthchecks on the estate run against `localhost`
      INSIDE the container they check, so a paused container cannot run its
      own healthcheck and a box loss takes every one. Following the
      contract's refusal of a blinded observer is what produced the two
      observers the plans actually use.
      
      THE CONTRACT REPAIRS, each forced by a reading rather than argued.
      `FaultControl` gained `presentsAs` (checked against a
      `PRESENTATIONS_BY_IMPAIRMENT` table, so an `unavailable` scenario
      cannot be rehearsed with a control that deregisters a name),
      `confirmedBy` and `movesTheRecoveryPoint`; `Observer` gained `detects`;
      `GameDayPlan` gained `unrehearsableBecause`. THE RECOVERY-POINT
      REQUIREMENT IS NOW CONDITIONAL AND THE RECOVERY-TIME ONE IS NOT: as
      shipped, `no-criterion-compares-the-recovery-point` fired on every plan
      without an RPO criterion, but most controls cannot cause data loss at
      all (a postgres crash replays its log -- measured), so it was demanding
      a criterion that passes whatever the estate does, which is the
      `aspect-passed-with-no-evidence` shape the same module refuses two
      sections down. AND `observer-blinded-by-the-injected-fault` TREATED
      EVERY IMPAIRMENT AS IF IT WERE `unavailable`: a `corrupt` store ANSWERS,
      so the check refused the only observer a corruption drill can have.
      Blindness now follows `silencesObservers(impairment)`.
      
      THE THIRD OPTION. Five of the ten scenarios cannot be injected on this
      estate -- no signing service is deployed, the model provider needs the
      undeployed ISIS admin API, the search index has no process of its own
      (it IS the relational store), and the bridge host is somebody else's
      workstation. Without `unrehearsableBecause` the author must either
      invent a control (the defect this cell warns about) or omit the plan,
      and an omitted plan is `scenario-with-no-plan`, which reads as an
      oversight rather than a measured impossibility. The declaration fails in
      BOTH directions: `planFaults` refuses one that also ships a control, and
      the scan re-checks each against the deployed compose, with four positive
      controls proving those patterns CAN match -- because a declaration
      guarded by a regex that cannot fire is an absence nothing re-checks.
      
      TRAPS IN MY OWN WORK, found by the adversarial pass. (a) THE
      SELF-REFERENCE, exactly as S10.12.c recorded it one section up: "does
      the document state this target has no process" was
      `markdown.includes(target)`, and the statement block IS part of the
      markdown, so every target the statement named satisfied the question the
      statement raised. The check could not fail for any input; it now reads
      the prose BEFORE the marker, with a negative control using
      `dep:trust-store`. (b) MY OWN PROBE WAS BLIND TO THE FAULT THAT HIDES
      BEST: `presentationOf` settled `healthy` on the socket's `connect`
      event, and a container frozen with `docker pause` reported HEALTHY --
      the kernel completes the handshake from the listen backlog whether or
      not any process will ever read from it. Found by the driver's own
      self-test reporting `injected: false` on a demonstrably paused
      container. It now requires an ANSWER, and counts a close-without-bytes
      as one, because only a live process hangs up. (c) The destructive
      confirmation asked whether the census was ZERO, which is right only for
      a control that destroys everything -- against the corruption control it
      would have reported the estate healthy for the one fault a liveness
      check already misses. It is a before/after comparison now. (d)
      `--self-test` exited 0 when docker was absent and `--reverse-all` exited
      0 with a destructive fault still applied: S10.12.a's finding 2 rebuilt
      twice inside the driver written to end it. Now 3 and non-zero. -->
      
      FINDINGS FOR S10.12.g, from this cell: F-1 no telemetry pipeline is
      deployed (high); F-2 all 14 healthchecks are inside their own container
      (high); F-3 the S10.12.a register cites the development compose for the
      estate's blast radius (medium); F-4 `data_checksums` is off, so page
      corruption is silent (high); F-5 a trap-based reversal does not survive
      SIGKILL (medium); F-6 five of ten scenarios are unrehearsable here
      (low).
      
    • S10.12.e Execute rotating scenarios end-to-end without undocumented access/dependency and measure actual detection, decision, recovery point, recovery time, and correctness. Evidence: timed report.

      <!-- S10.12.e: shipped as `scripts/operations/v1-game-day-execute.mjs`
      (five drills, each standing up a disposable estate, seeding it, backing
      it up, injecting, POLLING for detection, restoring and comparing a
      census), the timed report at
      `docs/releases/v1/verification/game-day/2026-08-14/execution-report.json`,
      `libs/oshun/workbench-kit/tools/execution-scan.mjs` (nx target
      `execution-scan`, `pnpm run verify:operations:v1-game-day-execution`),
      a `rotate` job on `.github/workflows/v1-disaster-recovery.yml` with a
      WEEKLY `schedule:` trigger, and a 15th test in
      `v1-game-day-plan.test.mjs`.
      
      THE PREMISE CHANGED, NOT THE CONCLUSION. The S10.12.a note recorded this
      cell as NOT LOCALLY ACTIONABLE because "executing rotating scenarios
      end-to-end needs a live ISIS API and a real host", and that was true of
      the estate as it stood: every fault control in the repository injected
      by `PATCH`ing feature flags against an admin API nobody deploys here.
      S10.12.d shipped an injector that needs nothing but docker, so the
      missing piece was never the host. RE-READ A "NOT ACTIONABLE" NOTE
      AGAINST THE ESTATE THAT EXISTS NOW, NOT THE ONE IT WAS WRITTEN ABOUT.
      
      MEASURED, all five quantities read from the clock at the step that
      produced them, none supplied and none derived from another:
      
      | scenario                   | detection | recovery time | recovery point |
      | -------------------------- | --------- | ------------- | -------------- |
      | `gd:database-loss`         | 1520ms    | 1669ms        | 0ms            |
      | `gd:queue-loss`            | 1521ms    | 1572ms        | 0ms            |
      | `gd:blob-store-loss`       | 1520ms    | 1571ms        | 0ms            |
      | `gd:regional-loss`         | 1553ms    | 1907ms        | 0ms            |
      | `gd:relational-corruption` | 385ms     | 631ms         | **1433ms**     |
      
      THE ONLY NON-ZERO RECOVERY POINT IS THE ONLY DRILL THAT CAN DESTROY
      DATA, which is S10.12.d's finding arriving as a measurement rather than
      an argument: the corruption drill writes 100 rows AFTER the backup and
      before the fault, and those rows are genuinely unrecoverable. The
      restore returning exactly 2000 is the correctness check -- MORE would
      mean the backup restored was not the one taken.
      
      WHAT IT DOES NOT CLAIM. `estate: "disposable"` is a field in every
      execution record, not a caveat in prose. The runs stand up their own
      postgres/redis/minio because the development box's are shared with other
      worktrees, which is the `production-like` dataset §1 of the plan
      prescribes and the only one it allows. This establishes the procedure
      works and the measurements are real; it does not establish the
      production estate's recovery time.
      
      AND THE ROTATION IS NOW A ROTATION. S10.12.a's finding 5 -- "nothing
      schedules any of them ... there is no rotation because there is no
      second run" -- was still true after the first pass, and one pass is not
      a rotation however many scenarios it covers. The `rotate` job runs
      weekly and, in order: grades the plans, proves every fault control still
      injects AND reverses, executes the rotation, grades the timed report,
      keeps the report as an artifact, and drains the reversal ledger under
      `if: always()` -- because a trap does not survive the death that needs
      one. Five of the ten scenario classes stay absent from the rotation and
      are DECLARED, read from the plan document rather than restated, and
      asserted in both directions: a class this estate can inject that was
      left out fails, and a class the plan calls unrehearsable that the
      rotation exercised fails too.
      
      THE SCAN REFUSES THREE THINGS THE CONTRACT CANNOT SEE: a detection
      latency of zero (the shape S10.11 found in all ten rollback records -- a
      number written rather than measured; verified to fire by stamping both
      instants together), a restore that finished before it started, and a
      report that does not say what estate it ran against.
      
      TRAP, IN THIS FILE'S OWN HEADER: the comment documenting what the scan
      grades contained the glob `game-day/*/execution-report.json`, and `*/`
      CLOSED THE BLOCK COMMENT -- the parser then read the prose as code and
      failed at EOF, 200 lines later, with an error pointing nowhere near it.
      `node --check` names the real line where a bundler's scanner does not. -->
      
    • S10.12.f Test split-brain prevention, degraded operation, queued work, user communication, monitoring/alerts, backup of recovered primary, and failback. Evidence: acceptance matrix.

      <!-- S10.12.f: shipped as `scripts/operations/v1-acceptance-drill.mjs`
      (ONE incident on a disposable estate built in this runbook's own
      topology: a box carrying the relational store, the queue, the blob store
      and the surface; an off-box repository; a clean recovery host), the
      matrix at
      `docs/releases/v1/verification/game-day/2026-08-14/acceptance-matrix.json`,
      `libs/oshun/workbench-kit/tools/acceptance-scan.mjs` (nx target
      `acceptance-scan`, `pnpm run verify:operations:v1-acceptance-matrix`, 29
      negative controls), the operator-facing
      `docs/runbooks/v1-acceptance-matrix.md`, 11 tests in
      `scripts/operations/v1-acceptance-matrix.test.mjs`, and an `acceptance`
      job on `.github/workflows/v1-disaster-recovery.yml`.
      
      THE PREMISE CHANGED AGAIN, EXACTLY AS IT DID FOR (e). The S10.12.a note
      left this cell unchecked because "an acceptance matrix is filled by an
      executed drill; a matrix written without one is exactly the
      `aspect-passed-with-no-evidence` the module refuses" -- true of the
      estate as it stood, and untrue of the estate S10.12.d and S10.12.e left
      behind. The missing piece was never the aspects, it was a drill, and
      there is now an injector that needs nothing but docker and a harness
      that stands up an estate and measures what happened to it.
      
      A ROW THAT PASSES IS A SENTENCE SOMEBODY WILL TRUST ABOUT THE ESTATE,
      AND THAT DECIDED THE HARDEST VERDICT IN THE MATRIX. The drill deploys a
      real Prometheus, a real blackbox exporter and a real rule, and the alert
      fires 9.1s after the box dies -- an honest measurement of a pipeline
      this drill stood up. `infra/hetzner/docker-compose.yml` contains NO
      prometheus, grafana, loki, tempo, jaeger, OTLP collector, blackbox
      exporter or alertmanager, so on the deployed estate nothing would have
      fired at all. The rule applied throughout: AN ASPECT PASSES WHEN THE
      PROPERTY WAS MEASURED ON THE ESTATE'S OWN MECHANISM, fails when the
      drill measured that the mechanism is absent or broken, and the
      measurement that would flatter it goes in the evidence instead of in the
      verdict.
      
      MEASURED, one incident, seven facets, every outcome computed from a
      comparison rather than written down:
      
      | aspect                            | outcome    | what decided it                                                                  |
      | --------------------------------- | ---------- | -------------------------------------------------------------------------------- |
      | `split-brain-prevention`          | **failed** | the returning box and the estate that replaced it both wrote id 541, two values   |
      | `degraded-operation`              | passed     | correct answers from part of the dependency set on 3 of 5 surveys, 0 empty-success |
      | `queued-work`                     | **failed** | the verified checkpoint restores 0 of 25 entries into a redis that answers PONG   |
      | `user-communication`              | passed     | the off-box notice read back 200; the on-box control did not answer in 2000ms     |
      | `monitoring-and-alerts`           | **failed** | the rule fires in 9088ms and 0 services that could run it are deployed            |
      | `backup-of-the-recovered-primary` | passed     | 10027 bytes captured between the restore and the failback, restoring to 501 rows  |
      | `failback`                        | passed     | the rebuilt box came back with 501 rows, 22 drained jobs, 25 entries, 12 objects  |
      
      THE FINDING THIS CELL WOULD KEEP IF IT KEPT ONLY ONE. `offbox-backup.sh`
      forces a `BGSAVE` and refuses to finish without `redis/dump.rdb` -- and
      the restore never reads that file. `redis-server --appendonly yes`, as
      the deployed compose starts it, creates a new empty AOF on start and
      ignores the RDB entirely: measured at 0 of 25 entries, answering PONG.
      `box-loss-restore-drill.sh` writes `redisRestoreVerified: true` on the
      strength of exactly that PING, which is the mechanism behind S10.11's
      finding rather than a restatement of it. The archive as a whole does
      carry the queue (the AOF travels with the data directory, and the
      directory restore came back with all 25 entries and the two-entry
      pending list intact), so what is broken is the artifact the backup
      VERIFIES: the queue's recovery point is not the instant the backup
      script measured, and a restore that took the checkpoint at its word
      would come up empty and pass its own check.
      
      AND THE COST OF THE MISSING FENCE, IN ROWS. §8 of the runbook says
      `Fencing -- ABSENT`, and this is what that sentence buys: the box came
      back after the clean host was serving, both accepted an insert on one
      primary key, and the failback -- which restores the capture -- discarded
      the row the old box accepted. Two writable estates for one key space,
      and everything the returned box wrote is thrown away by the procedure
      that repairs it. The same drill shows the estate's ONE real fence
      working one layer down: a stale `lease_token` matches 0 rows while the
      current holder's write lands, so the hole is specifically the promotion
      path.
      
      FOUR INSTRUMENT BUGS, ALL OF THEM IN THE DIRECTION THAT FLATTERS THE
      ESTATE, all found by running the drill rather than reading it:
      1. psql prints the command tag as well as the RETURNING row, so
         comparing stdout to the key reported BOTH writes refused while both
         rows were demonstrably there. The first matrix therefore recorded
         `split-brain-prevention: passed`. WHAT A STORE ACCEPTED IS A QUESTION
         FOR THE STORE, and the check now reads the value back.
      2. the alert poll was an attempt count, not a deadline: 120 iterations
         in under two seconds, faster than the rule's own `for:` window could
         elapse, recording `alertFiringAtMs: null` -- which reads exactly like
         a pipeline that never fires. AN INSTRUMENT THAT STOPS WATCHING BEFORE
         THE THING IT WATCHES FOR CAN HAPPEN IS NOT MEASURING.
      3. XAUTOCLAIM's reply begins with the next CURSOR, which is also
         `digits-dash-digits`, and the drain applied it as a unit of work: 23
         rows for 22 entries, the extra one keyed by a position in a scan.
      4. the on-box negative control fell back to `127.0.0.1` when the dead
         container had no address -- a question about the wrong machine, and
         one a busy development box could answer yes to. The addresses are now
         read before the kill.
      
      A NEGATIVE CONTROL FOR THE OBSERVER ITSELF. After the failback the drill
      FREEZES the same store with `docker pause` for 9s: the TCP-connect probe
      still reports `probe_success=1` and no alert fires, while the probe that
      requires an ANSWER reports a timeout. The pipeline the estate does not
      deploy would be blind to a wedged dependency even once it exists, which
      is S10.12.d's finding about `presentationOf` arriving as a property of a
      real alert rule.
      
      THE DOCUMENT AND THE REPORT ARE TWO ARTIFACTS AND ARE GRADED AGAINST
      EACH OTHER. S10.12.b's prose cross-check compared a statement to the
      document CONTAINING it, so it could never fail; here
      `v1-acceptance-matrix.md` is written by hand and `acceptance-matrix.json`
      by a drill, and the scan fails if an outcome stated in one differs from
      the one measured in the other, if a finding is missing from the
      document, or if the document cites a path nobody can open. The scan also
      re-derives every outcome from the report's own measurements (a row
      flipped without touching a number fails), grades the post-recovery
      capture with S10.11's `recoveryPointFaults`, times this runbook's own
      sentence about capturing before failing back, and re-reads every claim
      the matrix makes about the repository -- so the day somebody deploys
      Prometheus, writes a fencing step or stops verifying a checkpoint
      nothing reads, the matrix is describing an estate that no longer exists
      and the scan says so.
      
      WHAT IT DOES NOT CLAIM. `estate: "disposable"` is a field in the report.
      It does not establish the production estate's recovery time; it does not
      speak for the other nine registered scenarios, which have plans and
      runbooks and no acceptance matrix; and every step was performed by a
      program that already knew the order.
      
      FINDINGS FOR S10.12.g, emitted by the drill with the measurement that
      produced each and in the severities `FollowUp` admits: F-f1 a returning
      box writes into the key space of the estate that replaced it (critical);
      F-f2 the backup verifies a redis checkpoint the restore never reads
      (critical); F-f3 no telemetry service is deployed, so no alert has a
      producer (major); F-f4 no degraded mode exists during a regional loss --
      every fallback is co-located with what it backs up (major); F-f5 work
      written after the capture is lost and still sits on the returned box
      (minor); F-f6 a connect-only probe is blind to a frozen dependency
      (minor). -->
      
    • S10.12.g Assign severity/owner/deadline to every finding, link backlog, retest critical fixes, and update runbook. Evidence: closed follow-up register.

      <!-- S10.12.g: shipped as
      `docs/releases/v1/verification/game-day/2026-08-14/follow-up-register.json`
      (THIRTY findings, every one this section measured), the operator-facing
      `docs/runbooks/v1-follow-up-register.md`,
      `scripts/operations/v1-follow-up-retest.mjs` (four retests against real
      containers, `pnpm run run:operations:v1-follow-up-retest`),
      `libs/oshun/workbench-kit/tools/follow-up-scan.mjs` (nx target
      `follow-up-scan`, `pnpm run verify:operations:v1-follow-ups`, 27 negative
      controls), `infra/hetzner/backup/verify-restored-state.sh`, 10 tests in
      `scripts/operations/v1-follow-up-register.test.mjs`, and a `follow-ups`
      job on `.github/workflows/v1-disaster-recovery.yml`.
      
      A CLOSURE IS A CLAIM THAT THE FIX WORKS, AND THE ONLY EVIDENCE FOR IT IS
      RUNNING THE THING AGAIN. `followUpFaults` shipped in S10.12.a refusing
      `critical-follow-up-closed-without-a-retest`, and its only callers were
      spec fixtures -- the shape every other half of this section was found in.
      The register it now grades closes ten of thirty findings, and the four
      retests behind those closures were executed twice on this box with
      identical verdicts: `retest:F-f2` and `retest:key-recovery` and
      `retest:drill-verdict` came back `fixed`, and `retest:D-F4` came back
      `remedy-does-not-fix-it`.
      
      THE RETEST THAT REFUTED ITS OWN FINDING, WHICH IS WHY THE VOCABULARY HAS
      A WORD FOR IT. S10.12.d raised `D-F4` as "`data_checksums` is off, so
      page corruption is silent", and both the game-day plan §2.1 and the fault
      control's own recorded reading stated that causally. Four arms, same
      table, same injection:
      
      | page data | `data_checksums` | what `count(*)` returned          |
      | --------- | ---------------- | --------------------------------- |
      | random    | off              | `ERROR: invalid page in block 1`  |
      | random    | on               | the same ERROR, plus the checksum |
      | zeros     | off              | **1815 of 2000 rows, no error**   |
      | zeros     | on               | **1815 of 2000 rows, no error**   |
      
      AN ALL-ZERO PAGE IS A VALID UNUSED PAGE TO POSTGRESQL BY DESIGN, so the
      silence was never the missing checksum; a garbled page was already
      refused by the page-header check before any checksum was involved.
      `POSTGRES_INITDB_ARGS: --data-checksums` is set anyway -- it buys the
      diagnosis on the garbled arms, costs nothing at initdb, and applies only
      to clusters initialised after it, which means every clean-host restore
      and NOT the running production volume. `D-F4` stays OPEN with its
      statement corrected in all three places that carried the wrong one
      (§2.1, the §10 findings table, `v1-fault-control.mjs`), and the S10.12.d
      test that guarded the refuted claim was rewritten rather than deleted:
      the finding is live, so what it guards now is the corrected sentence and
      the detection path that does work.
      
      THE OWNER FIELD IS READ, NOT ASSIGNED, WHICH IS THE DEFECT S10.12.a
      WARNED THIS CELL ABOUT. The contract refuses a team noun; it cannot tell
      whether the person named exists. Every owner here must be a contact in
      the SHIPPED dependency register and, where the entry names a dependency,
      one of THAT dependency's owners. Deadlines are arithmetic over a policy
      stated once (critical 14 days, major 30, minor 90), so a row cannot be
      quietly given six more weeks. Closures must match a record in
      `follow-up-retests.json` whose verdict is `fixed` and whose FINISHING
      INSTANT is the register's `retestedAtMs` -- a closure backed by a run
      rather than by a merged diff. And every finding the acceptance matrix
      emitted must appear, under its own id or an alias (`F-f3` is absorbed by
      `D-F1`, `A-12` by `F-f1`, `B-3` by `F-f2`).
      
      TWO REPAIRS TO THE RESTORE PATH, BOTH RETESTED. `F-f2`: `offbox-backup.sh`
      forced a `BGSAVE` and refused to finish without `redis/dump.rdb`, and the
      restore never read that file -- `redis-server --appendonly yes` creates a
      fresh AOF on start and ignores the RDB, so a host handed the verified
      checkpoint comes up with 0 of 25 entries answering PONG, and the drill
      wrote `redisRestoreVerified: true` on the strength of that PING. The
      capture now records `redisKeyCount`, `redisPersistentKeyCount` and
      `minioObjectCount` where the restore can read them, and
      `verify-restored-state.sh` compares them: two numbers, because keys with
      a TTL may legitimately expire and keys without one cannot, so only the
      persistent count is a floor. `B-2`: `keyRecoveryVerified: true` was a
      literal with nothing above it, and the required set is now DERIVED from
      every `${VAR:?}` in the deployed compose -- the set the stack refuses to
      start without -- so a new required secret cannot be forgotten.
      
      THE SEAM NEITHER SIDE'S TESTS COVERED, found by reading the two suites
      against each other rather than either on its own. The producer's test
      asserted the counts were COMPUTED (its `jq` stub printed every
      `--argjson` as one flat object and threw the filter away); the consumer's
      test asserted its reads against a manifest written BY HAND in its own
      fixture. So `inventory.redisKeyCount` could have moved to top level, been
      renamed or been mistyped, every restore would have refused every
      snapshot, and both suites would have stayed green. The stub now EVALUATES
      the object construction, and a new test extracts the paths
      `box-loss-restore-drill.sh` reads out of the manifest and resolves each
      against the manifest `offbox-backup.sh` just wrote. Verified by breaking
      it on purpose: moving the field out of `inventory` fails both, naming the
      path and both scripts.
      
      THE DOCUMENT AND THE REGISTER ARE GRADED CELL BY CELL, not just on the
      word `closed`. S10.12.b's prose cross-check compared a statement to the
      document CONTAINING it and so could never fail; S10.12.f fixed that for
      one field. Here state, severity, owner, deadline AND the tally sentence
      under the table are each compared, because the reader acts on the
      severity and the date as much as on the state, and a row that agrees on
      one column and lies on another is not a smaller defect. Five of the 27
      negative controls mutate the DOCUMENT rather than the register -- without
      those, the comparison has a side nothing can move, which is the same
      defect one level up.
      
      WHAT WAS BROKEN IN THE IN-FLIGHT WORK AND IS FIXED. The scan CRASHED --
      `context.document` was read and never populated, so the central artifact
      of this cell did not run at all. The overdue branch was unreachable on a
      wall clock (every deadline is in the future), and an unreachable check
      reports exactly what an absent one does, so the clock is injectable
      (`OSHUN_FOLLOW_UP_NOW_MS`) and is exercised at 2027-01-01, where all
      twenty open findings report late. `retest:key-recovery`'s third arm
      hardcoded `met: true` and fed it into the verdict -- a sentence dressed
      as a measurement -- and now measures the artifact, because the pre-fix
      behaviour was the ABSENCE of a check and running nothing against a bad
      input accepts it by definition.
      
      A SCAN WHOSE INPUTS ARE MISSING MUST NOT LOOK LIKE A CLEAN ONE, which is
      `all(.[]; .status != "fail")` over an empty array one level up. The scan
      refuses an absent register, an absent retest report and an absent
      published document; the test file exercises all three against a synthetic
      estate built at the same depth, and the complement -- that the real one
      grades clean -- so the refusals are not unconditional. The same register
      copied into an estate where its backlog links do not resolve exits 2, not
      1: with the unmutated register already carrying a fault a control exists
      to provoke, that control proves nothing, and the scan refuses to report
      from checks it can no longer vouch for.
      
      THE SCAN FAILS IN TWO WAYS AND THEY MEAN DIFFERENT THINGS. A MALFORMED
      register is a defect in the artifacts; an OVERDUE one is well-formed and
      the work is late, which is what a deadline is for. Both exit non-zero and
      the output separates them, because an operator sent looking for a broken
      document when a critical finding is simply past its date has been
      misdirected by the instrument.
      
      WHAT IT DOES NOT CLAIM. Twenty findings are open, and the register
      records who is accountable and by when -- not that anybody is working on
      them. `docs/operations/on-call.md` names no person and holds no rotation
      config, so the rota identifiers are real register entries that resolve to
      nobody carrying a pager; the owner column says which rota is accountable,
      not who will answer, and that gap is itself `A-11`'s neighbour. Nobody
      has agreed to the deadline policy (`agreedBy: null`); the windows are the
      ones this cell proposes, enforced arithmetically so that changing one is
      a commit somebody reviews. And the register covers what the game-day
      programme found -- not the estate's whole defect surface. -->
      
      NOTE (S10.12.a, on b-g): b through g are LEFT UNCHECKED DELIBERATELY, and
      the contract each of them needs is already shipped in
      `libs/oshun/workbench-kit/src/game-day.ts` -- what is missing is the
      artifact, not the definition. `Runbook` + `runbookFaults` name the eleven
      sections S10.12.b lists and enforce them (authority is a named person or
      rota, communications must not travel over the failed dependency, the
      recovery must not assume it, secrets must be recoverable out of band, the
      restore order must not contradict the reference graph, a promotion needs
      a fencing token applied BEFORE it, and there must be a failback);
      `ReplayProcedure` + `replayFaults` name S10.12.c's nine subjects and
      refuse a replay window bounded by a wall clock instead of by the recovery
      point; `GameDayPlan` + `planFaults` name S10.12.d's fault controls,
      observers, timers, abort, evidence and success criteria; `AspectResult` +
      `acceptanceFaults` name S10.12.f's seven aspects; `FollowUp` +
      `followUpFaults` name S10.12.g's fields and refuse a critical finding
      closed without a retest.
      
      Each remains unchecked for its own reason. (b) asks for VERSIONED
      RUNBOOKS. CORRECTION, measured after the S10.12.a note was first written:
      the estate has 27 runbooks under `docs/runbooks/`, several of them
      substantive and dated (`database-backups.md`, `object-storage-backups.md`,
      `provider-failover.md`, `shell-outage.md`, each with an `Owner:` and a
      `Last reviewed:` line). The gap is narrower and sharper than "there are
      none": **NOT ONE RUNBOOK IN THE REPO MENTIONS THE V1 RECOVERY PATH.** A
      grep across `docs/runbooks/`, `docs/reference/runbooks/` and
      `V3/runbooks/` for `box-loss-restore-drill`, `v1-restore-reconcile`,
      `offbox-backup` or `restic` returns NOTHING. `database-backups.md` -- the
      runbook an operator reaches for -- documents `scripts/db-backup.sh`, an
      RDS/S3/cron cadence and a `db-restore.sh` whole-database restore, and
      never mentions that a V1 restore is not finished when it finishes:
      `box-loss-restore-drill.sh` ends by writing
      `status: "restore-pass-reconciliation-required"`,
      `safeToResumeAutonomy: false` and
      `nextRequiredControl: "scripts/operations/v1-restore-reconcile.mjs"`, and
      no runbook carried that step. Its `Owner:` is `platform team`, which the
      module refuses as a team noun.
      
      PARTIALLY CLOSED IN THE SAME PASS: `docs/runbooks/v1-disaster-recovery.md`
      now exists -- all eleven sections, every step citing a script that is
      really there, machine-checked by
      `scripts/operations/v1-dr-runbook.test.mjs` (10 tests, in
      `verify:operations:v1-disaster-recovery`). The checks are the ones a
      reader cannot perform: every cited repository path exists, every `§N`
      cross-reference resolves, no authority line names a collective, the
      restore order is stated in the REFERENCE direction (blob before
      relational -- the reverse of S10.11's capture order), the reconciler is
      named along with the proof fields that demand it, and the two sections
      that cannot be written from the repository are DECLARED (`Owner:
      UNASSIGNED`, `## 8. Fencing — ABSENT`) rather than omitted, because an
      absent section and a satisfied one look identical in a document. It also
      states the gaps rather than papering over them: `keyRecoveryVerified` is
      unproven, and Redis/MinIO/file-state/configuration are asserted from
      steps an EMPTY store also passes.
      
      S10.12.b STAYS UNCHECKED, and the criterion for marking it is now
      precise: ONE of the ten registered scenarios has a runbook.
      `runbookCoverageFaults(WORKBENCH_GAME_DAY_SCENARIOS, [...])` reports the
      other nine, and each needs the same treatment -- written from scripts
      that exist, or declared absent where they do not. (c) the replay procedures: the outbox, lease and
      provider-outcome reconciliations were not read in this cell and must not
      be invented. **(c) IS NOW DONE** -- they were read, against the running
      database as well as the source, and the answer to two of the three was
      that the store is not where a table-name search looks: the outbox is
      genuinely absent, the lease fence is real and its reclaim loop is
      deployed nowhere, and the provider outcomes are thirty-eight crypto
      invoices under a snapshot store key. See the S10.12.c note above. (d) a game-day plan needs the fault controls that actually
      exist on this estate, which is a Hetzner box plus docker compose, and
      naming a control that has not been tried is a plan for a rehearsal nobody
      can run. (e) executing rotating scenarios end-to-end needs a live ISIS
      API and a real host: NOT LOCALLY ACTIONABLE, and the harness now refuses
      to call an un-injected run a pass rather than pretending otherwise. (f)
      an acceptance matrix is filled by an executed drill; a matrix written
      without one is exactly the `aspect-passed-with-no-evidence` the module
      refuses. **(f) IS NOW DONE** -- the drill exists, and the matrix it
      filled fails three of the seven aspects with a measurement behind each.
      See the S10.12.f note above. (g) a follow-up register needs owners and backlog references,
      and inventing a name to satisfy a field is the same defect as inventing a
      result.
      
      The thirteen findings this cell measured ARE the raw material for (g);
      they are enumerated in the S10.12.a note above and in the module header,
      with the four that were repaired marked as such.
      
  • S10.13 Add data-subject export and deletion orchestration hooks. Evidence: protected retention exceptions and deletion receipts are tested.

    <!-- S10.13 PARENT: all eight children are marked, and the parent's own
    evidence clause is the one S10.13.f and S10.13.g deliver — the retention
    exception is tested (`subject-retention-exception.spec.ts`, 34 tests) and
    the deletion receipt is tested (`subject-receipt.spec.ts`, 30 tests).
    
    WHAT BACKS THIS MARK. Eight kit entry points, 279 tests and eight scans
    carrying 115 negative controls between them, all passing against the
    source:
    
    | cell | module                     | tests | controls |
    | ---- | -------------------------- | ----- | -------- |
    | a    | subject-requests           | 25    | 9        |
    | b    | subject-data-map           | 17    | 14       |
    | c    | subject-orchestration      | 68    | 21       |
    | d    | subject-export             | 42    | 22       |
    | e    | subject-erasure            | 40    | 15       |
    | f    | subject-retention-exception| 34    | 12       |
    | g    | subject-receipt            | 30    | 11       |
    | h    | subject-lifecycle-matrix   | 23    | 11       |
    
    a and b were shipped in an earlier session; both modules were read in full
    in the session that marked this parent, and both scans pass.
    
    THE THROUGH-LINE, which only appears once the eight are read together: THIS
    ESTATE CONTAINS ITS OWN ANSWER AND APPLIES IT ON ONE PATH ONLY. The tip
    ledger pseudonymises a counterparty on erasure and hands them over whole on
    export. 23 stores fence on a digest and 35 on the raw account id. The
    deletion path signs every attestation with Ed25519 and the export's digest
    is self-computed. The legal hold is category-blind and covers all 28 purge
    partitions; the retention exception is category-scoped into a vocabulary
    that reaches 12 of them. The one orchestration with a plan executes
    nothing; the two that execute have no plan. In every pair the estate has
    already built the right thing once.
    
    AND THE ONE FACT THAT RUNS THROUGH ALL EIGHT: the machinery is real —
    signed attestations, fail-closed omission, honest `not_found`, a fence in
    58 stores, an object-locked archive — and what is missing is almost always
    a SENTENCE rather than a mechanism: what the subject was told, why the data
    is held, which tenant a row belongs to, what must happen first. -->
    
    • S10.13.a Define verified request identity/authority, subject identifiers, tenant/scope, jurisdiction/purpose, request type, deadline, state, communication, and audit. Evidence: request contract.

      <!-- S10.13.a: shipped as
      `libs/oshun/workbench-kit/src/subject-requests.ts` (`SubjectRequest`,
      `subjectRequestFaults` with 29 fault kinds, `ESTATE_REQUEST_SHAPES`,
      `estateCoverageFaults`), 25 tests in `subject-requests.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-request-scan.mjs` (nx target
      `subject-request-scan`, `pnpm run verify:operations:v1-subject-requests`,
      9 negative controls), and the three registrations a kit entry point needs
      (package.json exports, `tsconfig.base.json` alias, `KIT_OWNED_ENTRY_POINTS`).
      
      AN OPERATOR'S AUTHORIZATION IS NOT THE SUBJECT'S REQUEST. They are
      different claims, made by different people, about different things.
      `POST /v1/admin/privacy/dsar/erase` gates on
      `admin:workspace:privacy` -- authorization to work the machinery -- and
      NOTHING on the type it accepts records that the subject asked or that
      anybody checked who they are. Both directions of that are dangerous in a
      way ordinary work is not: an erasure performed for the wrong person
      destroys a real account, and an access export performed for the wrong
      person IS the breach it was meant to prevent.
      
      THREE TYPES MODEL A SUBJECT REQUEST IN THIS REPOSITORY, THEY DISAGREE,
      AND THE RICHEST OF THEM IS THE ONE NOTHING CALLS. Measured, not asserted
      -- the scan reads each interface out of its source and re-derives the
      caller claim from a repository-wide search:
      
      | type                              | elements | reached by                          |
      | --------------------------------- | -------- | ----------------------------------- |
      | `privacy.DsarRequest`             | 6 of 12  | **nothing outside its own spec**    |
      | `privacy.DeletionRequest`         | 5 of 12  | the body of the erasure route       |
      | `persistence.DsarDeletionRequest` | 3 of 12  | the route, and the cascade that erases |
      
      FIVE OF THE TWELVE ELEMENTS ARE REACHABLE AT ALL (subject identifiers,
      tenant, scope, deadline, state). FOUR ARE CARRIED BY NO REQUEST TYPE IN
      THE REPOSITORY: authority, purpose, communication, audit. Three more --
      verified identity, jurisdiction, request type -- exist only on the type
      nothing reaches, which means the running estate cannot ask about them.
      
      THE COMPOUND ELEMENTS ARE SPLIT, AND THAT IS WHERE TWO OF THE FINDINGS
      WERE HIDING. S10.13.a names nine items and four are compounds. Graded as
      compounds, `DsarRequest` would score "jurisdiction/purpose" on a
      jurisdiction field alone and "identity/authority" on an identity check
      alone -- and it is precisely the missing halves that matter.
      `requestingActorKind` ADMITS `'authorized-representative'` and the type
      carries no evidence of that representative's authority, so
      `checkEligibility` treats a verified representative exactly like a
      verified subject. VERIFYING A REPRESENTATIVE IS NOT VERIFYING THEIR
      AUTHORITY: a solicitor's own identity checked to the highest strength
      available establishes nothing about their standing to act for this
      subject, and those are two evidence chains of which only the second
      decides whether the request is honoured.
      
      A DEADLINE IS AN OBLIGATION AND AN ETA IS A FORECAST. Neither reachable
      type carries a date the operator is held to: `executionEtaUnixSeconds` is
      a forecast of when work will finish and `hardDeleteAtUnixSeconds` is when
      the erasure executes. The contract anchors the deadline to RECEIPT
      specifically, because anchoring it to a step the operator controls --
      identity verification, say -- hands an unbounded extension to whoever is
      behind: don't take the step, and the clock never starts. An extension is
      a recorded act with a reason and a grantor, and a due date beyond the
      furthest recorded one is an extension nobody granted.
      
      THE ONLY SUBJECT-COMMUNICATION FIELD IN THE WHOLE PRIVACY LIBRARY IS ON
      THE OPERATOR-ACCESS TYPE. `OperatorAccessRequest.notificationToSubjectByUnixSeconds`
      and `OperatorAccessRequest.auditId` are the estate's only instances of
      two of the nine elements -- so the path where staff read a subject's data
      owes them a sentence and carries an audit anchor, and the path where the
      subject asserts a right does neither. Erasure is also the one operation
      that removes the rows which would show what was done, so a request with
      no audit id leaves a destructive act recorded only by its own absence.
      
      NINE TYPES IN THIS REPOSITORY ARE CALLED `DeletionRequest`, AND
      `DsarRequest` IS DECLARED TWICE. Demeter's portability lifecycle, two
      Iris privacy surfaces, two Aphrodite sovereignty models, Aphrodite's GDPR
      types, Kuanyin's autonomy analytics and Yemaya's account-deletion service
      all declare their own. THE SCAN FOUND THIS BY CONTRADICTING ME: my first
      caller check was a name search, it reported callers for a type that has
      none, and the negative controls refused to report rather than pass. A
      file declaring its own type of that name is now a HOMONYM rather than a
      caller, every one is recorded, and an undeclared one fails -- because
      "nothing calls this" is a claim about the whole repository and a name
      search answers it about all nine. The same pass caught the scan counting
      its OWN register file as a caller.
      
      A DECLARED ABSENCE IS A TWO-WAY CLAIM, which is S10.12.f's lesson applied
      to a type instead of a runbook. Every element the register says a type
      LACKS is searched for in that type's source by an indicator pattern, so
      the day somebody adds `auditId` to `DeletionRequest` the scan fails and
      the register has to be updated. Without that half, a gap statement
      outlives the fix and the estate looks worse than it is -- the mirror of
      the failure mode where it looks better.
      
      WHAT IT DOES NOT CLAIM. This defines the contract and measures the gap;
      it does not close it. No route was changed, no type was migrated, and the
      five reachable elements are still five. `SubjectRequest` has no producer
      yet -- which is the shape this section keeps finding, and is why the cell
      ships the estate register and the scan rather than the type alone. The
      remedy belongs to S10.13.b onward: the data map, the orchestration plan,
      and the receipt that reconciles them. -->
      
    • S10.13.b Register domain/store/provider/processor data locations, identity mapping, export format, deletion/anonymization action, retention/hold exceptions, owner, and verification. Evidence: data map.

      <!-- S10.13.b: shipped as
      `libs/oshun/workbench-kit/src/subject-data-map.ts` (`DataLocation`,
      `dataMapFaults` with 20 fault kinds, `dataMapCoverageFaults`, and
      `ESTATE_DATA_MAP` -- 24 locations), 17 tests in
      `subject-data-map.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-data-map-scan.mjs` (nx target
      `subject-data-map-scan`,
      `pnpm run verify:operations:v1-subject-data-map`, 14 negative controls),
      and the three registrations a kit entry point needs.
      
      A STORE THAT CANNOT FIND THE SUBJECT CANNOT DELETE THEM, AND SAYING IT
      DELETES THEM IS THE WORST ROW IN THE MAP. A shared catalog with no
      per-subject column, a store whose rows are enumerated by whoever calls
      the eraser, an opaque volume: each can be given a deletion action in a
      spreadsheet and none can honour it. The identity mapping is therefore the
      FIRST column, and an unselectable mapping with a deletion action is
      refused twice -- once against the row, and once against the factory
      `buildBffDeletionErasers` really uses.
      
      FOUR VOCABULARIES DESCRIBE THE SAME SUBJECT'S DATA AND NONE OF THEM IS A
      MAP. Measured:
      
      | list                       | size | what it says                                       |
      | -------------------------- | ---- | -------------------------------------------------- |
      | `OSHUN_ARTIFACT_SCOPES`    | 28   | what an artifact IS -- never where it lives         |
      | `DELETION_SERVICE_IDS`     | 16   | what the fan-out erases; all 16 ARE artifact types  |
      | `EXPORT_SCOPES`            | 11   | **nine are not artifact type names**                |
      | backup `stateClasses`      | 6    | the only list about WHERE, and it is about volumes  |
      
      AGAINST THE SEVEN ATTRIBUTES THIS CELL ASKS FOR, the artifact scopes
      already carry export format and deletion action, carry retention only as
      a post-deletion window with no hold exceptions, and carry data LOCATION,
      IDENTITY MAPPING, OWNER and VERIFICATION not at all. Those four are what
      makes a map a map, and they are what this register adds.
      
      ONE PRECISE GAP THE FOUR VOCABULARIES HID. Twelve artifact types have no
      deletion service. Eleven are honestly `redaction_only` or
      `not_deletable`. The twelfth is **`persona`** -- `deletable: true`,
      `soft_delete_then_purge`, 90 days -- and no service erases it under that
      name. The BFF's `personalization_vector` eraser deletes an
      `activePersona` partition, so the data is probably covered under another
      service's name, AND NOTHING RECORDS THAT CORRESPONDENCE. A comparison of
      the two lists cannot see it, which is precisely the question a data map
      exists to answer.
      
      TWO CAPTURED VOLUMES ARE NAMED BY NO ERASER. `content-service-runs` and
      `computer-use-snapshots` are in the off-box snapshot and mentioned
      nowhere in `data-deletion`, `deletion-fanout` or `privacy`. Whether a
      subject is in them is UNDETERMINED, and the map records that as `opaque`
      rather than leaving them out -- a screen capture of a session is exactly
      what a subject means by "everything you have about me".
      
      THE ARCHIVE IS A LOCATION, AND IT IS THE ONE THAT UNDOES EVERY ERASURE
      ABOVE. 30 days, `provider-object-lock`, so a snapshot cannot be edited to
      remove a subject even if somebody wanted to; expiry is the only
      mechanism, and a clean-host restore replays every state class wholesale.
      **THIRTY DAYS IS THE ANSWER TO "HOW LONG AFTER AN ERASURE CAN THIS
      SUBJECT COME BACK", and until this row nothing in the repository said
      so.** The scan binds that number to `RETENTION_DAYS` in
      `offbox-backup.sh`, so changing the backup makes the map wrong out loud.
      
      WHAT THE DELETION SYSTEM GETS RIGHT, because a cell that only reports
      gaps is not measuring. Sixteen services with real erasers wired at the
      BFF boundary, Ed25519 attestations per service, fail-closed omission (an
      unconfigured category is left out and a deletion requesting it fails
      rather than silently passing), bytes purged before the metadata that
      maps to them, sessions erased before the profile so refresh credentials
      cannot outlive the account, and `voice_profile` honestly reporting
      `not_found` for a shared catalog instead of claiming an erasure it
      cannot perform. That last one is the opposite of a stub, and the map
      records it as `not-subject-keyed` + `none`.
      
      THE SCAN FAILS ON DRIFT AND REPORTS COVERAGE. Malformation and drift
      from source are failures: a service the map does not place, an eraser
      factory it describes wrongly, an archive window that no longer matches
      the backup, an unregistered dependency or artifact type. The ten coverage
      findings are REPORTED and do not fail, because they are the estate's gaps
      rather than the map's, and a map that correctly records a gap is doing
      its job -- the same split S10.12.f's acceptance matrix ships under. The
      state classes are read out of a jq filter inside a shell script, which no
      TypeScript import can reach and which is exactly where a list nobody
      maintains would hide.
      
      WHAT IT DOES NOT CLAIM. It does not establish that a subject IS or is not
      in the two opaque volumes -- it records that nobody has established it.
      It does not close the export vocabulary mismatch, and `buildExportBundle`
      still has no production caller while the real exporter at
      `apps/oshun/bff/src/data-export/` runs on artifact types instead. Owners
      are rotas from the S10.5 register, which S10.12.g already recorded
      resolve to nobody carrying a pager. -->
      
    • S10.13.c Plan authorized search/export/delete with impact, dependencies, derived/backup/log/audit/provenance/search/cache/provider data, exclusions, and approvals. Evidence: orchestration plan tests.

      <!-- S10.13.c: shipped as
      `libs/oshun/workbench-kit/src/subject-orchestration.ts`
      (`OrchestrationPlan`, `orchestrationPlanFaults` with 44 fault kinds,
      `planCoverageFaults`, `planExecutionOrder`, `canonicalPlanStatement`,
      `ESTATE_ORCHESTRATIONS` and `buildEstateDeletionPlan`), 68 tests in
      `subject-orchestration.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-orchestration-scan.mjs` (nx
      target `subject-orchestration-scan`,
      `pnpm run verify:operations:v1-subject-orchestration`, 21 negative
      controls), and the registrations a kit entry point needs.
      
      A DEPENDENCY IS A CLAIM ABOUT WHAT MUST HAPPEN FIRST, AND A POSITION IN A
      LIST IS NOT ONE. This estate states exactly one ordering requirement, it
      is a safety property, and it is carried by an array index:
      `buildBffDeletionErasers` pushes `session` before `profile` and says why
      -- "so active refresh credentials cannot outlive account removal DURING
      THE SEQUENTIAL FANOUT". The sequential `executeDeletionFanout` that would
      honour that array has **no caller outside `*.test.ts`**. The production
      path is `buildDeletionRunner`, which registers every eraser as an
      independent bus consumer and dispatches ONE
      `subject.deletion.requested` event whose replies are collected in arrival
      order. The comment names the mechanism that would deliver the guarantee
      and the mechanism is not the one that runs.
      
      FOUR ORCHESTRATIONS, AND THE ONLY ONE WITH A PLAN EXECUTES NOTHING.
      Measured, each read out of the file that runs it:
      
      | orchestration                 | named steps | executor | outcomes  |
      | ----------------------------- | ----------- | -------- | --------- |
      | admin deletion workflow (010) | **8**       | **none** | typed     |
      | customer deletion fan-out     | 0           | real     | measured  |
      | admin DSAR erasure cascade    | 0           | real     | measured  |
      | customer data export          | 20          | real     | measured  |
      
      `ADMIN_DELETION_JOB_STAGES` is `discovery, redaction, content_removal,
      backup_purge, analytics_scrub, external_integrations, verification,
      closed` -- almost exactly the class list this cell asks for -- and the
      only writers of a stage transition are an admin route, its state store
      and an operator's React panel. `recordCount` and `failedCount` arrive in
      the request body, so a completed stage is a claim. `finalize_job` refuses
      only while a stage is `running` or `pending`, and a stage exists only
      once `start_stage` created it, so **seven of the eight can be absent and
      the job finalizes `succeeded`**: an unstarted stage and an inapplicable
      one are the same record.
      
      `backup_purge` IS A STAGE THE ESTATE'S OWN BACKUP MAKES IMPOSSIBLE.
      `offbox-backup.sh` accepts only `provider-object-lock`,
      `append-only-rest-server` and `provider-retention-lock`, under all of
      which a snapshot cannot be edited to remove a subject -- S10.13.b's
      thirty days, from the other side. An operator can still report the stage
      succeeded with a number they typed.
      
      THE HOLD IS CHECKED AFTER THE ERASURE, AND IT IS NOT GIVEN ANY HOLDS.
      `executeDsarErasureForSubject` runs `executeDsarDeletionCascade` -- real
      `deleteMany`s -- and only then calls `advanceToHardDelete`, whose result
      type can return `blocked-by-hold` about data that is already gone. And
      the gate is `input.holds !== undefined` while the runtime calls it with
      `{ request, nowUnixSeconds, shred }`, so an ABSENT ARGUMENT MEANS "NO
      HOLDS EXIST" and the check is skipped outright. Holds are evaluated once,
      in `createDeletionRequest`, which stamps `blocked-by-hold` or `pending`
      at filing -- so a hold placed during the soft-delete window, which is
      when holds are placed, is never consulted by the acts it exists to stop.
      The one orchestration that DOES check a hold before every stage is the
      one that erases nothing.
      
      A FORMAT IS A CHOICE ABOUT RENDERING AND SILENTLY A CHOICE ABOUT SCOPE.
      `resolveArtifactTypes` drops any artifact type whose `exportFormats`
      exclude the requested format: of the 20 exportable customer-visible
      types a subject receives **20 as json, 7 as csv, 5 as zip and 3 as
      markdown** -- and a dropped type gets NO section in the bundle, not the
      `available: false` with a reason the assembler produces for a type it
      cannot collect. The artifact is identical to one where those scopes held
      nothing. Two types are exportable and unreachable -- `persona` and
      `provenance_bundle`, both `exportable: true, customerVisible: false` --
      and `bundle-builder.ts` ships a real `persona` collector no customer
      request can reach.
      
      `redaction_only` IS A MODE NOTHING PERFORMS. Six artifact types declare
      it, the contract defines it as "retains the artifact with PII scrubbed",
      `resolveArtifacts` drops those types on `isArtifactDeletable` instead of
      scrubbing them, and NOTHING in the repository branches on
      `deletionMode ===` at all. The `redaction` stage is the only place the
      word appears in an executable position, in the orchestration with no
      executor.
      
      THE PLAN THE ESTATE RUNS, TRANSCRIBED AND GRADED BY THE SHIPPED
      CONTRACT. `buildEstateDeletionPlan` writes down what an `account_full`
      deletion really is -- fifteen erasures, one verification, `voice_profile`
      excluded because the builder constructs it through
      `createNoSubjectDataEraser` -- and the contract reports 18 findings on
      it. **No step states an impact**, and the estate has none to state:
      `rowsRemoved` exists only in the attestation a service signs AFTER
      erasing, and no search precedes the run. Against S10.13.b's 24-location
      map it reaches or excludes 16 and leaves **eight in silence**: the
      PostgreSQL cluster, Redis, MinIO, the two opaque volumes, the
      configuration snapshot, the 30-day archive that restores all of them, and
      `loc:persona`. **Two of the nine data classes are accounted for**;
      backup, log, audit, provenance, search, cache and provider are questions
      the plan cannot be asked. Twelve artifact types are dropped before a plan
      is built with no record that they were dropped.
      
      AND NOTHING IN THIS ESTATE SEARCHES. The cascade takes the records the
      caller enumerated -- its header says the subject-to-records mapping is
      the caller's responsibility by design -- the fan-out takes the artifact
      types the contract lists, and `discovery` is a stage in the orchestration
      that executes nothing. Impact cannot be stated by a plan that never
      looked, which is why the contract refuses an estimate credited to a
      search the plan does not contain.
      
      DECISIONS. AN OMISSION IS NOT AN EXCLUSION, because a location in no step
      and no exclusion arrives at the subject as silence, and silence is
      indistinguishable from a store that held nothing. AN EXCLUSION IS A FACT
      ABOUT THE MOMENT OF EXECUTION, not of filing, and a lapsed one is refused
      rather than honoured. IMPACT IS A MEASUREMENT OR A NUMBER SOMEBODY TYPED,
      so a declared count on an irreversible step is refused. AUTHORIZATION IS
      PER OPERATION, because searching, disclosing and destroying are three
      acts and one approval over "the request" collapses them in the dangerous
      direction -- and S10.13.a's rule is carried in, so a step that discloses
      to the requester on an operator scope alone is refused. AN APPROVAL BINDS
      A REVISION, and a subject asking for their own erasure IS the authority
      rather than a bypass of it, so only the other authorities owe a second
      signature. A STEP CANNOT RUN AFTER THE THING THAT LETS IT FIND THE
      SUBJECT IS GONE: afterwards it is not slow, it is blind -- it finds
      nothing and reports it.
      
      THE SCAN FAILS ON DRIFT AND REPORTS THE ESTATE. It re-derives the stage
      list, the writers of a stage transition, the caller set of the sequential
      fan-out, the ORDER of the two calls inside `executeDsarErasureForSubject`
      and the argument the second is not given, both scope filters and what
      each drops, and the accepted archive immutability modes. The 47 findings
      about the estate are REPORTED and do not fail, the same split S10.13.b
      ships under. Two of its own defects were caught by its negative controls:
      it counted its own register file as a stage writer -- the trap S10.13.b
      recorded -- and it counted `evidence_pack` as a store holding nothing
      because the builder names it under both factories.
      
      WHAT IT DOES NOT CLAIM. It does not fix any of this. The four
      orchestrations still do not know about each other; `AdminDeletionJob`'s
      `requestId` is documented as a UI deep-link and nothing writes a stage
      from a real erasure. It does not establish what is in the two opaque
      volumes, which S10.13.b already recorded as undetermined. -->
      
      <!-- Frontier note: S10.13.d onward inherit this vocabulary. The export
      isolation tests (d) have a measured starting point in the format gate
      above; the deletion tests (e) have the rehydration pair S10.13.b named
      and the ordering finding here; the exception record (f) has the hold
      checkpoints; the receipt (g) has the 47 findings this cell reports as the
      list of things a receipt would have to be honest about. -->
      
    • S10.13.d Export access-controlled, portable, understandable, checksummed/encrypted data with source/category/purpose/time and no other subject/tenant data. Evidence: export isolation tests.

      <!-- S10.13.d: shipped as
      `libs/oshun/workbench-kit/src/subject-export.ts` (`ExportBundle`,
      `exportBundleFaults` with 25 fault kinds, `exportIsolationFaults`,
      `twoPartyExportFaults`, `exportElementFaults`, `ESTATE_TWO_PARTY_STORES`,
      `ESTATE_EXPORT_SCOPES`, `buildEstateExportBundle`), 42 tests in
      `subject-export.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-export-scan.mjs` (nx target
      `subject-export-scan`, `pnpm run verify:operations:v1-subject-export`,
      22 negative controls), and the registrations a kit entry point needs.
      
      A RECORD ABOUT TWO PEOPLE IS NOT THE FIRST ONE'S TO RECEIVE -- AND THE
      ERASURE IN THE SAME FILE ALREADY KNOWS IT. This is not a principle
      imported from outside. `tip-ledger-store.ts` erases a subject by
      replacing THEIR field -- `viewerId` if they were the viewer, `creatorId`
      if they were the creator -- with `redacted:<digest>` and leaving the
      counterparty's field intact, because the counterparty's record of that
      transaction is the counterparty's data. **A hundred lines up in the same
      file, `getExportRecord` spreads the whole entry** and hands both ids to
      whichever of the two asked first. Measured across four stores:
      
      | store         | erasure does            | export does | counterparty |
      | ------------- | ----------------------- | ----------- | ------------ |
      | tip-ledger    | pseudonymises one field | copies whole| every tipper |
      | group-session | prunes one roster entry | copies whole| every member |
      | follow        | removes own edges only  | copies whole| every follower |
      | appreciation  | removes own rows        | own rows    | **none**     |
      
      `appreciation` is the control: keyed by the subject, naming nobody else,
      and it produces no finding -- a register in which every row is a finding
      is measuring its own appetite. `group-session` is the sharpest, because
      the store ALREADY separates a private leave audit from a shared one, so
      it is not that nobody thought about who sees what: `exportSession` strips
      `inviteTokenSha256` and nothing else, and a participant receives the
      host's id, the whole roster and the shared leave log.
      
      A PORTABLE FORMAT IS A PROMISE ABOUT THE BYTES, AND THIS ONE IS A FIELD
      INSIDE THEM. `assembleExportBundle` always `JSON.stringify`s, the route
      hardcodes `contentType: 'application/json'`, the download is
      `<exportId>.bundle.json`, and **there is no csv, markdown or zip renderer
      anywhere under `data-export/`** -- while the bundle records
      `format: 'csv'` when csv was asked for. So the format parameter does not
      do the one thing it is for, and does do the one thing it must not: it
      drops scopes (S10.13.c measured 20 json / 7 csv / 5 zip / 3 markdown of
      20). A subject who asks for markdown receives three scopes, in JSON, in a
      file that says it is markdown.
      
      AN EXPORT IS A NEW COPY, AND WHERE IT LANDS DECIDES HOW LONG IT OUTLIVES
      AN ERASURE. The assembled bytes are held `bytesBase64` -- an encoding,
      not a protection -- in `admin_store_snapshot`; nothing in the durable
      snapshot store encrypts; and that is the `postgresql` state class the
      off-box archive captures under object lock for **30 days**, the number
      S10.13.b bound to the backup script. So an access export taken shortly
      before an erasure puts a complete unencrypted copy of the subject beyond
      that erasure's reach for a month. The LIVE bundle is genuinely erased,
      through the `customer_export_bundles` partition on the
      `generated_artifact` eraser -- which is what keeps this a finding about
      the ARCHIVE rather than about the store.
      
      THE TENANT IS SELECTED AND NEVER FILTERED ON, AND THE FIRST VERSION OF
      THIS REGISTER GOT IT WRONG. It said `tenantId` did not appear in the
      export surface at all; a negative control falsified that. The truth is
      worse: the canonical V1 readers `select` `tenantId: true` on every row
      and type their predicate as `where: { readonly userId: string }`, so a
      tenant filter is **unexpressible** rather than merely unused. A subject
      present in two tenants receives both, each row stamped with the tenant it
      came from, inside a bundle that has no tenant of its own. The evidence of
      the leak is printed in the payload.
      
      THE ELEVEN ELEMENTS, WITH BOTH COMPOUNDS SPLIT. Graded as
      "checksummed/encrypted" this export scores on a sha256 it really
      computes; graded as two, it is checksummed and stores a plaintext copy of
      the subject in a captured table. Four elements are carried by nothing --
      **understandability, encryption, purpose and tenant isolation**. Purpose
      has ZERO occurrences anywhere under `data-export/`, so a subject is told
      what is held and never why, which is the half an objection is made
      against. Source is a habit rather than a field: `BundleScopeSection` has
      none, `storagePartition` appears inside items where a collector happened
      to add one, and `coverageGaps` names a partition as a first-class field
      only when the data is ABSENT -- the estate is more rigorous about where
      data is not than about where it is. Category is
      `appliesToCategories[0]` and five scopes roll up to two, every second one
      `audit_history`, so **a subject who asks for their audit history receives
      sections labelled `billing_receipts`, `consent_history` and
      `support_history`** and nothing says those are what they asked for. Time
      is the instant of assembly, not the period covered, so an empty section
      cannot be read.
      
      ACCESS CONTROL IS THE ONE ELEMENT WITH NO CAVEAT, and a cell that only
      reports gaps is not measuring: a customer session, a 403 on
      `bundle.userId !== auth.userId`, a 404 unless the request is `completed`,
      an expiry purge before every read, `cache-control: no-store`, and an
      `ExportBundleSubjectDeletedError` fence. The collection is real too --
      each section hashes what it collected, the bundle sha256 is over the
      bytes the customer downloads, and a type with no in-BFF store is recorded
      `available: false` with a reason rather than fabricated. Four of the
      twenty scopes have no collector: `voice_profile`, `avatar_pack`,
      `audit_event`, `analytics_event` -- and the last two are the audit and
      derived-analytics classes S10.13.c found no deletion step for, so the
      same two classes are absent from both operations.
      
      THE DESTRUCTIVE OPERATION IS SIGNED AND THE DISCLOSING ONE IS NOT.
      Deletion produces Ed25519 attestations the orchestrator verifies against
      a public key; the export's digest is computed and served by the same
      surface, so it detects corruption and not substitution. That is recorded
      as a `repair` rather than a refusal, because a digest that only detects
      corruption is doing a real job badly named.
      
      WHAT IT DOES NOT CLAIM. The transcribed estate bundle carries NO items,
      on purpose: the estate's sections are assembled from per-store records
      whose shapes differ store by store, and inventing item rows to grade
      would be grading this file's imagination. So `exportIsolationFaults` is
      VACUOUS over it and the scan says so in as many words rather than letting
      a zero read as clean -- the item-level answer is the four-store two-party
      register. It does not fix any of this, and it does not establish that the
      three disclosing stores are the only ones; it establishes that they are
      the ones whose erasure already implements the rule their export does
      not. -->
      
    • S10.13.e Delete/anonymize/withdraw in dependency-safe idempotent jobs, preserve only lawful held/minimal audit data, and prevent rehydration from cache/index/event/backup replay. Evidence: deletion tests.

      <!-- S10.13.e: shipped as
      `libs/oshun/workbench-kit/src/subject-erasure.ts` (`ErasureRun`,
      `erasureFaults` with 23 fault kinds, `executionOrder`,
      `estateFenceFaults`, `ESTATE_FENCE_CONVENTIONS`,
      `ESTATE_REHYDRATION_GUARDS`, `ESTATE_WITHDRAWAL`,
      `buildEstateErasureRun`), 40 tests in `subject-erasure.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-erasure-scan.mjs` (nx target
      `subject-erasure-scan`, `pnpm run verify:operations:v1-subject-erasure`,
      15 negative controls), and the registrations a kit entry point needs.
      
      A TOMBSTONE THAT LIVES INSIDE THE THING BEING RESTORED CANNOT SURVIVE THE
      RESTORE. Fifty-eight stores carry a deleted-subject fence and it is a
      REAL one: a second erasure is a no-op returning zero, and **118 non-test
      BFF files assert an active subject before a write**, so a replayed
      event's handler is refused. But the fence is a field of the same snapshot
      document as the rows, and every restore does
      `this.deletedSubjectIds = new Set(snapshot.deletedSubjectIds)` -- a
      replacement, not a union. **A restore from before the erasure returns the
      subject's rows AND the older fence set that does not contain them.**
      S10.13.b measured that archive at 30 days under `provider-object-lock`,
      so that is the window in which the estate's only anti-rehydration
      mechanism is absent from the thing being restored. This is a claim about
      a SHAPE rather than a symbol, so the scan reads the restore assignments
      themselves and fails the day one becomes a union.
      
      THREE OF THE FOUR SOURCES ARE GENUINELY GUARDED, which is what makes the
      fourth a finding rather than a mood:
      
      | source       | guard                          | survives its replay |
      | ------------ | ------------------------------ | ------------------- |
      | cache        | the owning store's fence       | yes                 |
      | search-index | **not captured at all**        | n/a                 |
      | event-replay | `assertSubjectActive` on write | yes                 |
      | backup       | the same fence                 | **no**              |
      
      The deployed compose runs no elasticsearch, opensearch, qdrant or
      meilisearch, so "not captured" is the honest answer and the scan fails
      the day a search profile is deployed. Four persistent buses call
      `replayUnacked()` at boot -- conversation, assignment, session,
      content-drop -- and their payloads are subject data, so an event unacked
      at the moment of an erasure IS redelivered afterwards and IS refused. The
      deletion bus alone is built `persistence: false`, deliberately: its
      signed attestations are the audit trail rather than a stream.
      
      THE PROOF OF AN ERASURE IS KEPT TWO WAYS IN ONE ESTATE. **35 stores keep
      the subject's own account id** in `deletedSubjectIds`; **23 keep a
      digest** in `deletedSubjectDigests`/`redactedSubjectDigests`. The split
      is clean -- no store carries both -- and nothing in the repository
      chooses. So nearly half the estate already demonstrates the minimal
      record of an erasure and the other half retains the person it erased. The
      interesting finding is not that a raw id is worse than a digest; it is
      that both are here, in stores that sit next to each other.
      
      AND THE RECORD OF THE ERASURE IS FILED UNDER THE ACCOUNT IT ERASED.
      `apps/oshun/bff/src/data-deletion/state.ts` has no `deleteForSubject` and
      no fence of its own, and its state map is keyed by `userId` -- so the
      whole deletion request, with its artifacts, derived jobs, tombstones and
      their signed attestations, the confirmation token, the reason and every
      timestamp, outlives the subject under their own key with no expiry. The
      tombstones' own stored shape carries service, status, rows, detail,
      signature, key id and instant and **no subject id**, so the contract
      correctly does not flag them: the residue that identifies the subject is
      the key it is filed under, not the payload.
      
      A WITHDRAWAL NAMES THE TWO THINGS THAT MUST STOP AND NOTHING ASKS EITHER
      OF THEM. `buildWithdrawalCascade` sets `storageBlockedAt` and
      `recallBlockedAt` to the withdrawal instant and builds up to **six** jobs
      -- a silent `.slice(0, 6)` -- with an eta 24 hours out. The two
      timestamps are read by exactly two things: the serializer that echoes
      them back to the customer, and the auth store's type declaration. No
      storage path asks whether storage is blocked; no recall path asks whether
      recall is; nothing anywhere advances a cascade job's status, so the eta
      passes unobserved. The subject is shown a block that is a record of an
      intention published as an effect -- which is the fourth jobs list in this
      section with no executor, after S10.13.c's three.
      
      DECISIONS. AN IDEMPOTENCY FENCE IS A RECORD OF A PERSON, necessarily, so
      the question is which record and not whether. A GUARD IS SCOPED BY WHAT
      IT SURVIVES: a fence that survives a re-run and not a restore is not a
      weaker fence, it is a guard against a different event, so the combination
      is refused rather than counted. A FENCE KEPT IN THE STORE IT GUARDS is
      only as durable as the thing it exists to outlive. A DELETE NEEDS NO
      FENCE AND AN ANONYMISE OR A WITHDRAW DOES, because repeating those
      changes the result -- a second anonymise pseudonymises the pseudonym, a
      second withdrawal re-dates a block the subject already has. ANONYMISATION
      IS AN EFFECT AND NOT A SOFTER DELETE, so it names the fields it replaces
      and what they become, and a replacement that still names the person is
      refused. AND AN HONEST "NOT CAPTURED" IS THE STRONGEST ANSWER THERE IS,
      so nothing further is asked of a source that holds no data.
      
      WHAT IT DOES NOT CLAIM. The transcribed run declares no dependencies
      between its fifteen jobs, because the estate declares none -- S10.13.c
      measured that -- and filling one in here would be transcribing a plan
      that is not run. It does not establish that the 35 raw-id stores are
      wrong and the 23 digest stores right; it establishes that the estate has
      not decided, and that it already owns the smaller answer. -->
      
    • S10.13.f Record protected retention exception with authority/legal basis, scope, fields, reason-safe communication, expiry/review, access, and later deletion. Evidence: exception tests.

      <!-- S10.13.f: shipped as
      `libs/oshun/workbench-kit/src/subject-retention-exception.ts`
      (`RetentionException`, `retentionExceptionFaults` with 24 fault kinds,
      `exceptionReachFaults`, `exceptionElementFaults`,
      `ESTATE_PURGE_PARTITIONS`, `buildEstateRetentionException`), 34 tests in
      `subject-retention-exception.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-retention-exception-scan.mjs` (nx
      target `subject-retention-exception-scan`,
      `pnpm run verify:operations:v1-subject-retention-exception`, 12 negative
      controls), and the registrations a kit entry point needs.
      
      AN EXCEPTION THAT CANNOT NAME WHAT IT PROTECTS PROTECTS NOTHING. An
      operator granting a retention exception selects from **23** privacy data
      categories; `runComplianceRetention` evaluates
      `exception.dataCategories.includes(partition.dataCategory)` against
      **10**; **six are the same word**. The bridge between the two,
      `ComplianceRetentionAdminWorkspace`, types the field
      `readonly string[]`, so nothing in the compiler will ever notice them
      diverging.
      
      SIXTEEN OF THE TWENTY-EIGHT WIRED PARTITIONS ARE PURGED UNDER A CATEGORY
      NO EXCEPTION CAN NAME. `purgeExpiredComplianceRecords` wires 28
      partitions -- 5 `billing_history`, 5 `support_case`, 1
      `data_export_request`, 1 `consent_record`, **15 `review_package` and 1
      `evidence_pack`** -- and the last two are not in the 23. So an operator
      who must preserve DSA moderation reviews for a regulatory inquiry,
      operator incidents for a safety investigation, or Metis academic evidence
      for a dispute **has no category to say so**, and the purge removes them
      on schedule while the exception reads `active`. The other direction is
      the same defect: **17 of the 23 selectable categories** are asked about
      by no partition at all, so choosing `biometric` or `location` protects
      nothing and looks like it does.
      
      THE BLUNT INSTRUMENT BESIDE IT WORKS PRECISELY BECAUSE IT IS BLUNT. The
      legal-hold filter in the same function is by subject and status alone, it
      names no category, and it therefore covers every partition. The estate's
      precise instrument protects 12 of 28 and its imprecise one protects all
      28.
      
      AN UNREADABLE EXPIRY BECOMES AN ETERNAL ONE.
      `buildComplianceRetentionAuthority` resolves an `effectiveUntil` that
      does not parse to `null`, and the filter reads
      `effectiveUntilUnixSeconds === null` as unbounded. Neither half looks
      wrong alone, which is why the scan checks them together: a typo in a date
      is a permanent hold. This is S10.8's missing-factor defect -- every
      absent value resolving to a number instead of a refusal -- in a date.
      
      FOUR OF THE TEN ELEMENTS ARE CARRIED AND CAREFULLY CHECKED, and a cell
      that only reports gaps is not measuring: `AdminRetentionExceptionSchema`
      refuses an `effectiveUntil` at or before `grantedAt`, refuses a
      revocation with no revoker and no reason, and refuses
      `justification: 'legal_hold'` with no `legalHoldId`. Authority, legal
      basis, expiry and review are all genuinely there.
      
      FOUR ARE CARRIED BY NOTHING: **fields, subject-notification,
      reason-disclosure and access.** The record has no field-level scope, so
      "keep the billing record" is the only sentence it can say and never "keep
      the tax total and not the phone number". It says nothing about what the
      subject was told -- which S10.13.c and S10.13.d both refuse -- and
      nothing about whether telling them the reason is safe, **while three of
      the eight justifications (`fraud_investigation`, `safety_investigation`,
      `regulatory_inquiry`) are ones where it is not.** And nothing limits who
      may read data that survived an erasure for one named purpose.
      
      TWO OF THE FOUR STATUSES ARE ASSIGNED BY NOTHING. The mutator sets
      `active` and `revoked`; `expiring_soon` and `expired` are set nowhere,
      and the mutator even branches on `expiring_soon` -- code reading a state
      nothing can enter. This is MITIGATED, and the mitigation is the estate
      getting something right: the executor filters on `effectiveUntil`
      regardless of status, so the date is honoured even though the label
      never changes.
      
      DECISIONS. TELLING A SUBJECT THAT DATA IS HELD IS NOT TELLING THEM WHY,
      AND ONLY ONE OF THOSE IS ALWAYS SAFE -- so "reason-safe communication" is
      split into two fields and the refusals run in both directions: an
      exception nobody was told about, and a reason disclosed where disclosing
      it defeats the purpose of holding the data. A REASON WITHHELD IS WITHHELD
      UNTIL A DATE, because an order that forbids telling the subject does not
      forbid it for ever. SCOPE HAS A GRAIN, and a category-grained exception
      always keeps whichever of the obligation and the category is larger. AN
      UNREADABLE EXPIRY IS NOT AN ABSENT ONE and is refused separately. AND AN
      EXCEPTION THAT NEVER RELEASES IS A RETENTION POLICY.
      
      WHAT IT DOES NOT CLAIM. It does not establish that the 16 unprotectable
      partitions are wrong to be purged -- their retention windows may be
      right. It establishes that no exception can change them, which is a
      different and worse fact, because the surface that exists to change them
      reports success. -->
      
    • S10.13.g Reconcile each processor/store outcome and issue a signed receipt of completed/pending/refused/retained items without leaking hidden data. Evidence: receipt fixtures.

      <!-- S10.13.g: shipped as
      `libs/oshun/workbench-kit/src/subject-receipt.ts` (`Receipt`,
      `receiptFaults` with 25 fault kinds, `estateReceiptFaults`,
      `ESTATE_RECEIPTS`, `buildEstateSubjectReceipt`), 30 tests in
      `subject-receipt.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-receipt-scan.mjs` (nx target
      `subject-receipt-scan`, `pnpm run verify:operations:v1-subject-receipt`,
      11 negative controls), and the registrations a kit entry point needs.
      
      A RECEIPT IS THE ONE ARTIFACT THAT DESCRIBES THE SUBJECT'S DATA TO THE
      SUBJECT, SO EVERY WORD IN IT IS BOTH AN ANSWER AND A DISCLOSURE.
      "Retained under a legal hold" answers the question and tells them an
      investigation exists; "erased from `metis_assessment_evidence`" answers
      it and hands them an internal store name. So a receipt is graded twice:
      for what it fails to say, and for what it says too loudly.
      
      THREE RECEIPT-SHAPED THINGS DIVIDE THE JOB SO THAT NO SINGLE ONE DOES IT:
      
      | artifact                     | reaches      | signed | can say "kept" |
      | ---------------------------- | ------------ | ------ | -------------- |
      | `DsarDeletionReceipt`        | an operator  | **no** | no             |
      | `DeletionAttestation`        | the customer | per item | no           |
      | `ComplianceRetentionReceipt` | **a log line** | no   | **yes**        |
      
      **THE ONLY RECEIPT WITH A WORD FOR DATA DELIBERATELY KEPT IS THE ONE
      NOBODY RECEIVES, AND THE ONLY ONE THAT IS SIGNED HAS NO SUCH WORD.**
      `blocked_by_legal_hold` and `blocked_by_retention_exception` exist in
      exactly one type, and that type is handed to `app.log.info`. So the
      entire retention-exception apparatus S10.13.f measured produces outcomes
      no subject can ever be shown.
      
      `refused` HAS NO WORD ANYWHERE AT ALL. `failed` is an error, not a
      refusal, and the difference is the whole of a subject's ability to
      appeal. `not-found` conflates three facts -- already gone, never held,
      and a store that cannot select the subject at all, which is S10.13.b's
      `voice_profile`.
      
      THE SIGNATURE IS PER ITEM AND NEVER OVER THE SET. `DeletionAttestation`
      is the strongest artifact in this whole section: a real Ed25519 signature
      over service, subject, deletion id, outcome and instant, verified against
      a public key before a deletion may complete. And nothing signs the
      COLLECTION, so an item can be dropped and every surviving signature still
      verifies -- while the claim a subject most needs is that these are ALL
      the items.
      
      NO RECEIPT DISTINGUISHES A PROCESSOR. Exactly one is in the deletion
      path: the wearable bridge, whose external delete throws
      (`WearableHealthExternalDeletionError`) and correctly fails the receipt
      closed. It produces no item of its own, so one `rowsRemoved` covers both
      sides of that boundary and **a subject cannot tell from any receipt that
      a third party ever held their data.**
      
      DECISIONS. FOUR STATES, AND `retained` IS THE ONE THAT MATTERS, because
      it is the only one a subject cannot infer from silence -- a receipt
      without it reports a partial erasure as a complete one by omission. A
      PENDING ITEM IS PENDING UNTIL A DATE, or "pending" and "we have not got
      to it" are the same word. RECONCILIATION IS PER PARTY AND A PROCESSOR IS
      A PARTY. AN INTERNAL IDENTIFIER IS NOT AN ANSWER -- which is why the
      audience is a field, and why the same wording is refused on a subject's
      receipt and allowed on an operator's. AND A SIGNATURE IS WHAT MAKES A
      RECEIPT EVIDENCE rather than a claim by the party with the motive to make
      it.
      
      WHAT IT DOES NOT CLAIM. The transcribed subject receipt adds no
      `retained` item, because the estate cannot produce one: the tombstone
      shape is `{ tombstoneId, artifactType, createdAt, expiresAt,
      attestation? }` and the attestation's vocabulary is
      `erased | not_found | failed`. Inventing one would be transcribing a
      receipt the estate cannot issue. -->
      
    • S10.13.h Test cancellation, duplicate request, identity mismatch, provider unavailable, partial failure, hold release, backup restore, and deadline escalation. Evidence: lifecycle matrix.

      <!-- S10.13.h: shipped as
      `libs/oshun/workbench-kit/src/subject-lifecycle-matrix.ts`
      (`LifecycleRow`, `lifecycleMatrixFaults` with 14 fault kinds,
      `ESTATE_LIFECYCLE_MATRIX` — the eight scenarios over eleven paths), 23
      tests in `subject-lifecycle-matrix.spec.ts`,
      `libs/oshun/workbench-kit/tools/subject-lifecycle-matrix-scan.mjs` (nx
      target `subject-lifecycle-matrix-scan`,
      `pnpm run verify:operations:v1-subject-lifecycle-matrix`, 11 negative
      controls), and the registrations a kit entry point needs.
      
      A SCENARIO YOU CANNOT REACH IS NOT A SCENARIO YOU PASS. A coverage report
      has two colours and this needs three: "tested and works", "reachable and
      untested", and "there is nothing to test" produce the same empty test
      file, and only the middle one is fixed by writing a test.
      
      | scenario             | reachable | driven | outcome        |
      | -------------------- | --------- | ------ | -------------- |
      | cancellation         | yes       | yes    | no-op          |
      | duplicate-request    | yes       | yes    | refused        |
      | identity-mismatch    | **split** | export only | refused   |
      | provider-unavailable | yes       | yes    | failed-closed  |
      | partial-failure      | yes       | yes    | failed-closed  |
      | hold-release         | **no**    | —      | —              |
      | backup-restore       | yes       | **no** | refused        |
      | deadline-escalation  | **no**    | —      | —              |
      
      FIVE OF THE EIGHT ARE HANDLED AND DRIVEN AGAINST THE REAL MECHANISM, and
      a matrix in which everything is broken is measuring its own appetite:
      `runDeletion` re-reads the request after the fan-out returns and refuses
      to finalize one cancelled mid-run; a duplicate is 409
      `active_deletion_exists` over a fence in 58 stores; the export download
      is 403 `bundle_owner_mismatch`; a provider outage throws
      `WearableHealthExternalDeletionError` and fails closed; a partial failure
      finalizes `deletion_incomplete` naming the unsatisfied services.
      
      HOLD RELEASE IS UNREACHABLE BECAUSE NOBODY CAN OPERATE A LEGAL HOLD. The
      hold is the one instrument in this estate that genuinely blocks a
      deletion -- the admin workflow refuses `start_stage` for a held subject
      and the compliance purge skips every partition for one, category-blind
      where S10.13.f's exception reaches only 12 of 28. And **`legalHolds` is
      READ in two places and MUTATED in none**: no action route, no
      `legalHolds.push` or `.find` anywhere in the admin state store, and the
      only released hold in the repository is a seed literal called
      `lh-released-matter-closed`. Nobody can raise one when a matter opens or
      clear one when it closes.
      
      DEADLINE ESCALATION IS UNREACHABLE BECAUSE THERE IS NO DEADLINE.
      S10.13.a measured `deadline` as carried only by `DsarRequest`, which
      nothing outside its own spec file calls; the two dates the estate keeps
      are an execution schedule and a forecast; and a search of the whole
      subject-request path for an escalation returns nothing. **There is
      nothing to escalate from**, so the row is unreachable rather than
      untested -- and that is why it is the one row with no mechanism to name,
      which the contract refuses out loud.
      
      IDENTITY MISMATCH IS OPPOSITE ON ITS TWO PATHS: exercised and refused on
      the disclosing one, and unreachable on the destroying one, because the
      erase route gates on `admin:workspace:privacy` and its request type has
      no identity to disagree with. One row per scenario would have averaged
      those into "partly covered", which is the answer that stops anybody
      looking -- so reachability is recorded per path.
      
      BACKUP RESTORE IS THE ONE REACHABLE ROW NOTHING DRIVES, and writing the
      test would fail: S10.13.e measured that every restore replaces the fence
      set rather than unioning it. The existing durability specs restore the
      snapshot that CONTAINS the deletion, which is a different question.
      
      DECISIONS. REACHABILITY COMES FIRST, so a row that is unreachable AND
      exercised is a contradiction the matrix refuses rather than a strong
      result -- whatever that test drove, it was not the estate. AN
      UNREACHABLE ROW OWES A REASON AND THE REASON IS A CLAIM ABOUT THE SOURCE,
      because "no mechanism" and "a mechanism nobody can operate" are different
      repairs: one is a feature, the other a route. A TEST AGAINST A DOUBLE IS
      NOT NOTHING AND IT IS NOT THE ESTATE. AND THE EXPECTED OUTCOME IS PART OF
      THE ROW, because a scenario handled without one is a scenario somebody
      watched happen.
      
      THE SCAN OPENS EVERY CITED TEST FILE, because a matrix that names a test
      nobody can open is the defect it exists to prevent; all seven are
      present. One of its negative controls caught a search that was reading a
      generated Prisma client -- `escalationRules` in another domain's column
      list -- and the search was narrowed to the subject-request path file by
      file. -->
      
  • S10.14 Add runtime feature-flag inventory and expiry alerts. Evidence: stale flags fail the release-readiness check.

    <!-- S10.14 PARENT: a-f all shipped and each re-read in this session. Six
    kit entry points, 246 tests, six scans with 76 negative controls between
    them.
    
    THE THREAD THAT RUNS THROUGH ALL SIX IS ONE ABSENCE. S10.14.a found that
    nothing in this estate emits a flag evaluation. .b that no declaration site
    can hold an expiry, and that the six sites which declare flags are not the
    two that answer them. .c that four of the ten integrity conditions are
    therefore undetectable rather than clean. .d that the only durations in the
    estate resolve to no instant, and that the one runbook step about flags is
    false in all three of its clauses. .e that the release gate blocks on two
    real conditions and cannot test three. And .f that the removal lifecycle's
    own precondition -- a verified cohort observation -- can never be met, for
    the reason .a wrote down first.
    
    **THE ESTATE'S FLAG MACHINERY IS NOT ABSENT. IT IS UNOBSERVED.** Every
    piece is real: sha256 bucketing, tenant allow/denylists, a kill switch,
    `unknown_flag` with `partial: true`, a waiver chain that sums its
    instalments, a gate policy type that makes a fabricated pass untypeable.
    What is missing throughout is the one thing that would let any of it be
    checked -- a producer -- and its absence is what turns each careful
    mechanism into a claim nobody can grade.
    
    The evidence line reads "stale flags fail the release-readiness check".
    They now do: `feature-flag-release-readiness-scan` runs the gate over a
    deliberate stale flag every time and exits non-zero if it passes. And the
    real estate's verdict is `block`, on six posture-changing environment gates
    with no owner and one flag read in production that no site declares. -->
    
    • S10.14.a Discover declared flags from registry/code/config/provider and runtime-evaluated flags from telemetry without treating dynamic values as source truth. Evidence: flag inventory.

      <!-- S10.14.a: shipped as
      `libs/oshun/workbench-kit/src/feature-flag-inventory.ts`
      (`FlagInventory`, `flagInventoryFaults` with 17 fault kinds,
      `duplicateDeclarationFaults`, `ESTATE_FLAG_SOURCES`,
      `ESTATE_FLAG_INVENTORY` — 24 flags), 28 tests in
      `feature-flag-inventory.spec.ts`,
      `libs/oshun/workbench-kit/tools/feature-flag-inventory-scan.mjs` (nx
      target `feature-flag-inventory-scan`,
      `pnpm run verify:operations:v1-feature-flag-inventory`, 10 negative
      controls), and the registrations a kit entry point needs.
      
      AN INVENTORY IS A CLAIM ABOUT WHAT EXISTS, AND THIS ONE ONLY LOOKS WHERE
      IT ALREADY KNOWS. Of the five sources S10.14.a names, the estate reads
      one:
      
      | source    | producer                       | reader | yields |
      | --------- | ------------------------------ | ------ | ------ |
      | registry  | **two** literals, no shared source | yes | 11     |
      | code      | the two literals themselves    | **no** | —      |
      | config    | the deployment environment     | **no** | 13     |
      | provider  | **none**                       | no     | 0      |
      | telemetry | **none**                       | no     | 0      |
      
      NOTHING EMITS A FLAG EVALUATION. `routes/feature-flags.ts` evaluates and
      returns; it logs nothing, records nothing and reports nothing to
      analytics. So **the one source that could find a flag nobody declared has
      no producer**, and its silence is indistinguishable from an estate with
      no undeclared flags. There is no flag provider either — LaunchDarkly
      appears in this repository only in Shakti prose about deployment
      strategies and in a `provider?: 'local' | 'launchdarkly' | 'unleash'`
      field on a built type.
      
      AND CONFIG HOLDS THIRTEEN GATES NO REGISTRY LISTS AND NO OWNER CLAIMS.
      `OSHUN_BFF_ENABLE_TEST_HARNESSES` exposes harness routes on a running
      service; `OSHUN_DEV_AUTO_VERIFY_EMAIL` treats a signup's email as
      verified without the subject proving it; `OSHUN_DEV_SEED_ARETE_HABITS`
      writes seed data into a running estate; `OSHUN_TELEMETRY_SILENT` silences
      the instrument every other cell of S10 depends on; four adapters choose
      which implementation of a domain loads; and `OSHUN_ENV` is read in
      eighteen non-test files, which makes it the most consulted flag in the
      estate. **The registry knows what the home screen looks like; nothing
      knows whether the service is accepting unverified signups.**
      
      TWO OF THE THIRTEEN HAVE NO PRODUCTION READER AT ALL.
      `OSHUN_BFF_ALLOW_DEV_TOKENS` and `OSHUN_BFF_FORCE_SIGNED_TOKENS` — both
      permission gates, one of them named for the safety property rather than
      the relaxation — are read only by the auth route test that deletes and
      restores them around a case. Whatever they once allowed, the only thing
      that still remembers them is the test that guards them.
      
      TWO REGISTRIES, ELEVEN KEYS EACH, NO SHARED SOURCE. They agree today;
      nothing makes them agree tomorrow, and their FIELD sets already differ —
      only the mobile one carries a `description`, only the BFF one carries an
      `experiment` — so a reader of either gets a partial answer and cannot
      tell which half is missing.
      
      WHAT THE ESTATE GETS EXACTLY RIGHT, and it is this cell's own rule: **a
      dynamic value cannot define a flag.** An override is looked up by a
      DECLARED definition's key, so an override naming an unknown flag is
      ignored, and a requested key nothing declares comes back `unknown_flag`
      with `partial: true`. The registry is careful too — eleven keys, each
      with an owner, a default and tags, five carrying an experiment with a
      stable sha256 bucketing function.
      
      DECISIONS. A SOURCE EITHER DECLARES OR WITNESSES, AND TELEMETRY ONLY
      WITNESSES — that is what "without treating dynamic values as source
      truth" is as a check rather than a sentence. A FLAG WITNESSED AND NEVER
      DECLARED IS THE FINDING, NOT AN ERROR, because it is the only original
      result an inventory can have. A SOURCE WITH NO PRODUCER IS REFUSED ONCE,
      in the source register, rather than once per flag — which is why a
      declared flag nothing witnessed is only a `repair` here. CONFIG IS A
      DECLARING SOURCE, because an environment variable a deploy sets and the
      code branches on is a flag by every property that matters, and excluding
      it for having no `key` field is how the dangerous ones stay uncounted.
      
      WHAT IT DOES NOT CLAIM. It does not establish that the 13 config gates
      SHOULD be in the registry — some may be deploy-time build switches that
      never vary in production. It establishes that nothing has decided, and
      that the surface which reviews flags cannot be asked about them. -->
      
    • S10.14.b Reconcile key/type/owner/purpose/capability/audience/default/ environments/targeting/start/review/expiry/rollback/removal task and code/config references. Evidence: completeness report.

      <!-- S10.14.b: shipped as
      `libs/oshun/workbench-kit/src/feature-flag-completeness.ts`
      (`FLAG_ATTRIBUTES` — the fifteen the cell names plus the two reference
      classes, `reconcileAttribute` with 9 verdicts, `flagCompletenessReport`,
      `flagCompletenessScore`, `flagCompletenessFaults` with 19 fault kinds,
      `ESTATE_DECLARATION_SITES`, `ESTATE_FLAG_COMPLETENESS`), 59 tests in
      `feature-flag-completeness.spec.ts`,
      `libs/oshun/workbench-kit/tools/feature-flag-completeness-scan.mjs` (nx
      target `feature-flag-completeness-scan`,
      `pnpm run verify:operations:v1-feature-flag-completeness`, 15 negative
      controls), and the four registrations a kit entry point needs.
      
      **THE FLAG THAT DECIDES AND THE FLAG THAT IS DESCRIBED ARE NOT THE SAME
      FLAG.** S10.14.a found two registries. There are six sites, and the two
      that a running reader actually consults are the two with the fewest
      fields:
      
      | site                                                     | shape                        | keys | reaches a rendering decision |
      | -------------------------------------------------------- | ---------------------------- | ---- | ---------------------------- |
      | `apps/oshun/bff/.../tenant-feature-flags.ts`             | `Record<string, Definition>` | 11   | no                           |
      | `apps/oshun/mobile/src/experimentation/feature-flags.ts` | `Definition[]`               | 11   | no                           |
      | `apps/oshun/web/src/experimentation/feature-flags.ts`    | `Definition[]`               | 11   | no                           |
      | `apps/oshun/mobile/src/config/runtime.ts`                | `Record<string, boolean>`    | 7    | **yes — 4 call sites**       |
      | `apps/oshun/web/src/config/runtime-context.tsx`          | `Record<string, boolean>`    | 0    | **yes — 1 call site**        |
      | `libs/oshun/shell-core/src/feature-experiments.ts`       | `Definition[]`               | 6    | one internal caller          |
      
      THREE EXPORTED `useFeatureFlag(key: string, fallback = false): boolean`
      FUNCTIONS EXIST AND THE IMPORT PATH IS THE ONLY THING THAT SAYS WHICH
      SOURCE ANSWERS. Two are built over the registries — the mobile
      `FeatureFlagService` on a five-minute TTL and the web `hooks.ts` — and
      **neither has a single non-test importer**. The one the product imports
      is `@/config/runtime-context`, over the env map. So the whole pipeline
      the registry serves (tenant overrides, allow/denylists, kill switch,
      sha256 bucketing, the evaluation route) reaches nothing: an operator
      setting a tenant override for `quickActionsTray` changes the BFF's answer
      and does not change the home screen.
      
      AND THE SECOND-ORDER RULE THAT MAKES IT A CHECK: **AN ABSENT VALUE AND AN
      ABSENT FIELD ARE THE SAME WORD AND DIFFERENT REPAIRS.** "This flag has no
      expiry" and "no flag can have an expiry" both print as `expiry: missing`,
      and one is an owner's minute while the other cannot be done at all until
      a type changes. **Ten of the sixteen attributes are unholdable by every
      site in the estate** — type, capability, environments, start, review,
      expiry, rollback, removal-task, code-references, config-references —
      searched for under every plausible field name across all six files, zero
      hits. **All five lifecycle attributes are among them, so S10.14.d
      (alert before review/expiry), .e (fail a release on expired) and .f
      (close after cleanup) each have no field to read.**
      
      MEASUREMENTS THAT CORRECT S10.14.a. Its note says the two registries
      "agree today". They do not, and the reason is the cell's own lesson: it
      compared KEY SETS, and **the key is the one attribute records cannot
      disagree on — a difference in it produces two flags rather than one
      conflict**. Re-read attribute by attribute across all three registries:
      `activityAchievements` is owned by `activity-team` in the BFF and the web
      registry and by `shell-team` in the mobile one, so **the answer to "who
      turns this off in an incident" depends on which file the operator
      opens**; `quickActionsTray` is tagged `navigation` in two and `home` in
      the third.
      
      PURPOSE IS PRESENT TWICE AND CANNOT BE RECONCILED BY COMPARISON. Nine of
      the eleven shared flags carry two differently worded descriptions, and at
      least one pair states different behaviour: `notificationsDeduplicationHints`
      is "De-duplicate similar notification titles in the inbox" on web
      and "Show deduplication hints in notification center" on mobile — a
      content change and a UI affordance. Differing prose may describe the same
      thing and identical prose is a copy, so equality answers neither
      question; the verdict says so instead of picking one. **And the one
      registry that evaluates has no `description` field at all.**
      
      WHAT THE READERS NAME AND NOTHING DESCRIBES. `mocapCaptureStudio` and
      `livingSceneComposer` are read on the home tab and the mocap screen and
      exist only in the env map, whose type holds a key and a default — so
      thirteen of their attributes are not blank, they are unrecordable.
      `studio.study` is read by `isStudioStudyNavVisible` as
      `featureFlags['studio.study'] === false`, which makes its default ON, and
      **no site in any namespace declares it**. Meanwhile four of the five
      experiments — every one at `rolloutPercentage: 0` — are described in
      three registries and read by nothing.
      
      BOTH RUNTIME SITES ACCEPT A KEY FROM OUTSIDE. `EXPO_PUBLIC_FF_<ANYTHING>`
      becomes a flag with no membership test, and `toFeatureFlagKey`
      lowercases before camel-casing, so `EXPO_PUBLIC_FF_homeLayoutV2=true`
      creates `homelayoutv2` and the real flag keeps its default. **The estate
      contains the correct rule and applies it on one path only**: the BFF
      returns `unknown_flag` with `partial: true` for exactly this case.
      
      THE SCORE REPORTS TWO NUMBERS AND NEVER SUMS THEM, because one total
      lets a schema gap be closed by filling a field in a file no reader opens.
      
      TRAP, caught by a negative control firing on the UNMUTATED run: the scan
      searching for its own subjects found `feature-flag-inventory.ts` and
      `feature-flag-completeness.ts` and reported four experiments as having
      readers. **A register that counts its own family invents a defect** — the
      fix is a path prefix, and the control is what made it visible. -->
      
    • S10.14.c Detect undeclared, duplicate/conflicting, unused, always-on/off, unknown state, expired, missing owner/purpose/removal, unsafe default, and contradictory nested flags. Evidence: integrity tests.

      <!-- S10.14.c: shipped as
      `libs/oshun/workbench-kit/src/feature-flag-integrity.ts`
      (`INTEGRITY_CONDITIONS` — the ten the cell names, split to thirteen where
      it compounds them, `CONDITION_EVIDENCE` classifying each by the evidence
      it needs, `integrityRunFaults` with 14 fault kinds, `gateDefaultFaults`,
      `invariantFlagFaults`, `discardedStateFaults`, `grantsWhenUnset`,
      `withholdingValues`, `synonymsOf`, `ESTATE_BOOLEAN_GATES`,
      `ESTATE_INTEGRITY_RUN`), 43 tests in `feature-flag-integrity.spec.ts`,
      `libs/oshun/workbench-kit/tools/feature-flag-integrity-scan.mjs` (nx
      target `feature-flag-integrity-scan`,
      `pnpm run verify:operations:v1-feature-flag-integrity`, 16 negative
      controls), and the four registrations a kit entry point needs.
      
      **A CONDITION ABOUT A RUN CANNOT BE DETECTED IN A DECLARATION, AND THE
      TWO HAVE THE SAME NAME.** The cell lists ten conditions as if they were
      one kind of thing. Eight are properties of a RECORD and you find them by
      reading a file. Four — unused, always-on, always-off, unknown-state — are
      properties of a RUN, and S10.14.a established that nothing in this estate
      emits a flag evaluation. One, expired, needs a date S10.14.b measured no
      declaration site can hold. **So a detector that reads
      `defaultValue: true, active: true, no experiment` and reports "always-on"
      has found something real and named it wrongly**: nothing in the
      DEFINITION can make it vary is a different claim from it has been on in
      the live system, and the first is silenced by adding an experiment block
      while the flag stays on. Every run condition here therefore carries a
      DECLARATION SIBLING under its own name —
      `invariant-by-construction`, `no-reading-expression`,
      `state-discarded-at-the-client` — and the run condition returns
      `undetectable`, not `clean`. **A DETECTOR WITH NO OBSERVATION SOURCE
      REPORTS EXACTLY WHAT A CLEAN ESTATE REPORTS**, which is this cell's
      single easiest wrong answer.
      
      SEVENTEEN BOOLEAN COMPARISONS OVER FOURTEEN KEYS, AND SEVEN OF THE KEYS
      ARE IN NO INVENTORY. S10.14.a counted thirteen config gates; a sweep
      finds seven more capability gates, four of which expose a test harness
      (`OSHUN_ENABLE_TEST_HARNESSES`, one word from the BFF gate and on the
      other tier), a proxy bypass (`…_STUDIO_ISIS_LANE_E2E_BYPASS`), an
      operator preview surface, and a header-policy override. **The condition
      S10.14.a said only telemetry could find — a flag nobody declared — is
      found by reading the source, seven times.**
      
      POLARITY ALONE DECIDES NOTHING; POLARITY TIMES WHAT IS GRANTED DECIDES
      EVERYTHING. `OSHUN_AUTONOMY_AUTOCYCLE !== '0'` starts an autonomous cycle
      driver unless the variable is exactly `'0'`.
      `OSHUN_DEV_SEED_SESSION_ENABLED === 'false'` gates a route that signs up
      or logs in a fixed account and sets both session cookies, and it is
      enabled unless the variable is exactly `'false'`.
      `OSHUN_TELEMETRY_SILENT !== 'true'` has the **same** polarity and is
      entirely safe, because what an unset variable grants there is an
      instrument. The check carries both halves or it is a style rule.
      
      AND EACH OFF-SWITCH ACCEPTS ONE SPELLING, AND THEY ARE DIFFERENT ONES.
      `AUTOCYCLE` accepts `'0'`; `SEED_SESSION_ENABLED` accepts `'false'`. An
      operator who learned `=0` from the first leaves the second granting.
      This estate has a parser that accepts `false`, `0`, `no` and `off` —
      `parseBooleanString` in the mobile runtime — and **not one environment
      gate uses it**.
      
      NESTING IS AN `&&`, SO NOBODY CAN BE TOLD ABOUT IT.
      `OSHUN_TELEGRAM_GROUNDER_TEST_SCENARIOS` does nothing unless
      `OSHUN_BFF_ENABLE_TEST_HARNESSES` is also on; the dependency lives in a
      conjunction in `telegram/webhook.ts` and in no registry. The estate has
      the right shape and uses it in exactly one place — shell-core experiments
      carry `requiredFlags` and `isExperimentEligible` refuses when one is off
      — except that `assignConsumerShellExperiment` checks the override
      **before** eligibility, so an override assigns a variant for an
      experiment whose prerequisites are off, under a `reason` that names where
      the value came from and not the constraint it broke.
      
      ONE KEY, SIX BINDINGS, AND ITS NAME IS NONE OF THEM.
      `EXPO_PUBLIC_E2E_DISABLE_AUTOFILL` is bound to `E2E_AUTOMATION_ENABLED`,
      `E2E_HOME_TEST_HOOKS_ENABLED`, `E2E_PROFILE_TEST_HOOKS_ENABLED`,
      `E2E_TARA_TEST_HOOKS_ENABLED`, `E2E_ACTIVITY_TEST_HOOKS_ENABLED` and
      `E2E_SHOW_PASSWORD_TEXT` — the last renders the sign-in password as text.
      A seventh reader is the only one that honours the name, and it binds no
      identifier, which is why a binding census is not a reader census. **The
      disagreement is between READERS and a detector that compares declarations
      cannot see it.**
      
      UNKNOWN STATE IS NOT A GAP IN WHAT THE ESTATE KNOWS. The BFF returns a
      per-flag `source`, a `reason` from nine values, the experiment
      assignment, an `errors` array and `partial`. **Both clients parse
      `{ flags }` and drop the rest**, and each one's QA surface,
      `getAllFlags()`, returns `{ value, overridden }` with no source. A
      requested key the registry does not declare comes back as an error the
      client discards and the screen shows the default.
      
      TWO TRAPS, BOTH CAUGHT BY A NEGATIVE CONTROL FIRING ON THE UNMUTATED RUN.
      **An operator is not a polarity**: the first version read `!==` as
      "grants when unset", and `if (X !== 'true') return null;` is an `equals`
      gate written inside-out — the polarity of a call site is the operator
      combined with what the branch does, which is the same shape as the
      unsafe-default rule itself. And **a register that searches for its own
      subjects finds itself**, this time because `grep -h` had stripped the
      path the workbench-kit filter needed, so the register's own `source:`
      literals came back as call sites and gave two gates an invented second
      polarity. Keep the path; filter on it.
      
      AND A THIRD OF THE SAME SHAPE, from the toolchain rather than a control:
      **a register of source text is indistinguishable, to a text-matching
      gate, from source that does the thing.** Recording each gate's
      comparison verbatim made this module read as one that consults the
      environment directly, and the repository's `direct-process-env`
      conventions ratchet refused the commit -- correctly, on the evidence
      available to it. The `process.` prefix carries nothing the `at` field
      does not already give, so the register drops it. -->
      
    • S10.14.d Alert owner/backups before review/expiry and escalate expired flags with affected routes/capabilities/cohorts and removal runbook. Evidence: clock/routing tests.

      <!-- S10.14.d: shipped as
      `libs/oshun/workbench-kit/src/feature-flag-expiry-alerts.ts`
      (`resolveDeadline`, `alertsDue`, `isReachable`, `escalationFaults` with
      12 fault kinds, `runbookStepFaults`, `ESTATE_EXPERIMENT_DEADLINES`,
      `ESTATE_STOPPING_CONDITIONS`, `ESTATE_ESCALATION`,
      `ESTATE_KILL_SWITCH_STEP`), 40 tests in
      `feature-flag-expiry-alerts.spec.ts`,
      `libs/oshun/workbench-kit/tools/feature-flag-expiry-alert-scan.mjs` (nx
      target `feature-flag-expiry-alert-scan`,
      `pnpm run verify:operations:v1-feature-flag-expiry-alerts`, 14 negative
      controls), and the four registrations a kit entry point needs.
      
      **A DURATION IS NOT A DEADLINE UNTIL SOMETHING STARTS IT.**
      `apps/oshun/mobile/src/experimentation/ab-framework.ts` -- a site
      S10.14.b's six did not include, because it declares EXPERIMENTS rather
      than flags -- carries `startDate: string | null`,
      `endDate: string | null` and `maxDurationDays: number`. All three
      experiments have `startDate: null`, `endDate: null`, and a real,
      deliberately differentiated limit: **28, 21 and 14 days from nothing**.
      `resolveDeadline` refuses to substitute `now` for the missing origin.
      That substitution is available, type-checks, and makes every duration
      resolve -- to a deadline always `durationDays` in the future however late
      you ask, so **the alert never fires and the flag never looks late**. Same
      shape as S10.9's grace window anchored to the event it was supposed to
      outlive: the quantity is real, the anchor is wrong, and the check goes
      unreachable rather than failing. The clock is a parameter throughout.
      
      AND THE MEASUREMENT THAT MAKES IT BITE: **THE STOPPING CONDITION THAT
      COUNTS PEOPLE IS ENFORCED AND THE ONE THAT COUNTS DAYS IS NOT.**
      `generateAnalysisReport` takes `durationDays` and `minSampleSize` as
      ADJACENT PARAMETERS. One becomes
      `hasMinSample = variants.every((v) => v.sampleSize >= minSampleSize)`.
      The other is copied into the report twice and interpolated into
      `**Duration**: N days`, and compared to nothing. `maxDurationDays` is
      read by nothing at all -- its interface, three literals, and no other
      line in the repository. So the two ways an experiment could end sit three
      lines apart in one object and only one is a check; and with every rollout
      at zero (S10.14.c) nobody is bucketed, so the enforced one can never be
      met either. **An experiment that can neither reach its sample size nor
      run out of time does not end.**
      
      A LABEL IS NOT AN ADDRESS. `owner: 'design-team'` resolves to nothing:
      the five files that mention a team string are the flag and experiment
      literals themselves, and no directory, rota, channel map or contact list
      in the repository maps one to a destination. **The scan checks this by
      looking for a SIXTH file** -- which is what a directory would be -- so it
      is a live check and not one that can only pass. And BACKUPS, S10.14.d's
      second recipient class, have no field at any site, so the alert would go
      to one unreachable name and one absence would silence it.
      
      THE PAYLOAD. Of routes/capabilities/cohorts, two are `unexpressible` on
      S10.14.b's measurement -- nothing links a flag key to the routes that
      branch on it, and `domainScope` names domains rather than capabilities --
      so no owner can supply them, only a type change. Cohorts exist.
      
      **A RUNBOOK STEP IS A SET OF CLAIMS ABOUT THE ESTATE, AND EVERY CLAUSE OF
      THE ONLY ONE THERE IS IS FALSE.** `docs/runbooks/shell-outage.md:88`
      reads: *"Kill-switch a recent flag: if a feature flag in
      `release-taxonomy.ts` was advanced in the last hour, set its cohort to 0%
      via the flag-admin API and observe bootstrap recovery."* Three claims,
      each checked by the method its kind implies:
      
      - A LOCATION. `release-taxonomy.ts` declares **no feature flags**; its
        four `flagged` occurrences are ProvenanceBundle moderation states and a
        `flag_reason` telemetry attribute.
      - AN OPERATION. The flag-admin API is
        `/v1/admin/studio/experimentation-flags` and its only two routes are a
        `GET` list and a `POST .../evaluate` -- **both reads**. The sole way to
        move a cohort is `OSHUN_TENANT_EXPERIMENT_SCOPES`, an environment
        variable, which is a redeploy.
      - A PRECONDITION. **Nothing in the estate records when a flag changed**,
        so the operator cannot evaluate the clause the step opens with.
      
      An operator following this step during an incident cannot complete its
      first clause. And no runbook in the estate covers RETIRING a flag at all
      -- the one flag step there is covers killing one mid-incident, which is
      the opposite operation.
      
      THE SCAN QUOTES, SO THE SCAN CHECKS THE QUOTATION. A register whose
      content is three refutations rots differently from one whose content is
      counts: the file stays, the step keeps its title, the words change, and
      the register goes on refuting a sentence nobody wrote. Each quoted clause
      is searched for in the runbook, and each verdict is re-derived rather
      than trusted -- the location by reading the file, the operation by
      enumerating the route's verbs, the precondition by searching for a
      changed-at field. The scan also prints a POSITIVE CONTROL: given an
      origin, the same 28 days yields an instant and two due alerts. **The
      arithmetic is not what is missing.** -->
      
    • S10.14.e Fail release readiness on expired or critical incomplete flags and prevent waived expiry from becoming indefinite. Evidence: deliberate stale flag test.

      <!-- S10.14.e: shipped as
      `libs/oshun/workbench-kit/src/feature-flag-release-readiness.ts`
      (`FLAG_READINESS_CONDITIONS`, `flagReadinessVerdict`,
      `flagReadinessFaults` with 10 fault kinds, `DELIBERATE_STALE_FLAG`,
      `specimenFaults`, `FLAG_WAIVER_REQUIREMENTS`, `waiverRequirementFaults`,
      `ESTATE_READINESS_VERDICT`), 35 tests in
      `feature-flag-release-readiness.spec.ts`,
      `libs/oshun/workbench-kit/tools/feature-flag-release-readiness-scan.mjs`
      (nx target `feature-flag-release-readiness-scan`,
      `pnpm run verify:operations:v1-feature-flag-release-readiness`, 11
      negative controls), and the four registrations a kit entry point needs.
      
      **BOTH HALVES OF THIS CELL WERE ALREADY SOLVED IN THIS KIT, FOR A
      DIFFERENT SUBJECT.** `plugins.ts` types `GatePolicy.onStale`,
      `onNotConfigured` and `onFailed` as `NonPassingVerdict` — and says why:
      *"`pass` and `warn` are absent by construction. A policy that could map
      an unconfigured gate to `pass` is the fabricated success the
      honest-capability invariant exists to prevent, and the only reliable
      place to stop it is the type of the field somebody would have to set."*
      **A green gate that had no measurement is untypeable.** And
      `gate-waivers.ts` already walks a supersession chain with `renewalChain`,
      summing the instalments against the gate's own limit, because *"eight
      consecutive forty-eight-hour waivers are a two-week exemption that no
      single approval ever authorised"* — which is exactly "prevent waived
      expiry from becoming indefinite", already written and tested.
      
      **AND NEITHER HAS EVER BEEN POINTED AT A FLAG.** `gate-definitions.ts`
      contains the word "flag" zero times; no `GATE_TYPES` member names one; no
      gate in the registry takes a flag as its subject. So S10.14's own
      evidence line — "stale flags fail the release-readiness check" — names
      the one subject the release machinery has never been asked about, and the
      check passes every release by never having been given anything to fail
      on. This cell therefore BINDS rather than reimplements: it records the
      four requirements standing between a flag expiry and the waiver
      machinery, naming the shipped mechanism for each, because **a second
      implementation of a rule is a second place for it to differ.**
      
      **A GATE'S PASS MEANS NOTHING UNTIL SOMETHING HAS FAILED IT**, which is
      why the cell's evidence line is a METHOD rather than a behaviour. A
      readiness check that has never refused a release is indistinguishable
      from one whose predicate is `true`, and reading it cannot tell them
      apart. So `DELIBERATE_STALE_FLAG` ships as a specimen — eight days past
      an expiry, with no owner — and **the scan runs the gate over it every
      time and exits non-zero if it passes.** It carries TWO defects on
      purpose, so that a gate which quietly lost either check still fails on
      the other and looks fine only to a test that asserts both.
      
      THE SPECIMEN HAD TO BE MANUFACTURED, AND THAT IS THE SECOND FINDING. The
      estate cannot exhibit an expired flag — not because its flags are fresh
      but because S10.14.b measured that none can carry a date and S10.14.d
      that the only durations there are resolve to no instant. **"No expired
      flags" is true here the way "no unicorns escaped" is true.** Three of
      the five conditions come back `indeterminate` — a word this estate's own
      vocabulary already has — and reporting them as `pass` is precisely the
      substitution `NonPassingVerdict` makes untypeable one file over.
      
      THE VERDICT ON THE REAL ESTATE IS `block`, AND NOT VACUOUSLY. Two
      conditions the estate CAN exhibit are dirty, both established by the
      cells before this one: **six posture-changing environment gates with no
      owner** (S10.14.c measured that no environment gate names one anywhere,
      and these are the ones whose value changes what the service will accept
      — including the autonomy driver and the session seeder), and
      **`studio.study`, read by a navigation predicate and declared by no site
      in any namespace** (S10.14.b). The scan recounts the first list from
      S10.14.c's register rather than restating it, so the two cannot drift.
      
      **A CHECK THAT CANNOT FIRE IS THE DEFECT THIS CELL IS ABOUT, AND I WROTE
      ONE.** The first draft of `release-passed-on-an-unexhibitable-condition`
      asked whether the overall verdict was `pass` while something was
      unexhibitable — and every member of `NON_PASSING_VERDICTS` outranks
      `pass`, so one unexhibitable condition makes an overall `pass`
      structurally impossible. It was not a strict check; it was a check that
      reports nothing. Replaced by
      `release-clean-only-where-it-could-be-dirty`, which asks the reachable
      question one repair away: every condition the estate CAN exhibit is
      clean, and the rest could not be tested. A test now proves the old form
      was unreachable, so the reasoning survives the fix. -->
      
    • S10.14.f Track code/data/schema/contract/test/docs/telemetry cleanup and close/remove registry entry after verified cohort observation. Evidence: removal lifecycle tests.

      <!-- S10.14.f: shipped as
      `libs/oshun/workbench-kit/src/feature-flag-removal.ts`
      (`REMOVAL_CLASSES`, `REMOVAL_STATES`, `cohortWasObserved`,
      `nextRemovalState`, `removalPlanFaults` with 11 fault kinds,
      `unenforcedSchemaFaults`, `ESTATE_PERSISTED_OVERRIDE`,
      `ESTATE_REMOVAL_PLAN`, `ESTATE_SCHEMA_ENFORCEMENT`), 41 tests in
      `feature-flag-removal.spec.ts`,
      `libs/oshun/workbench-kit/tools/feature-flag-removal-scan.mjs` (nx target
      `feature-flag-removal-scan`,
      `pnpm run verify:operations:v1-feature-flag-removal`, 10 negative
      controls), and the four registrations a kit entry point needs.
      
      **A REMOVAL IS A CLAIM THAT NOTHING IS LEFT, AND THE ONE COPY YOU CANNOT
      DELETE IS ON A DEVICE YOU DO NOT OWN.** The cell names seven cleanup
      classes and the estate has an eighth. `apps/oshun/web/src/experimentation/feature-flags.ts`
      persists its overrides to `localStorage` under `oshun-ff-overrides`;
      `loadOverrides()` reads the blob back as
      `new Map(Object.entries(parsed))` and **checks no key against any
      registry**; and `resolve(key)` consults the overrides BEFORE the remote
      value and before the default. So a flag's override outlives its
      declaration, on hardware the estate cannot reach, at the top of the
      precedence order — and if the key is ever reused, a decision somebody
      made about the old feature silently applies to the new one, for that
      person, forever. The mobile service holds its overrides in memory and has
      a `clearOverrides()`, which is what makes the web copy the finding rather
      than a generality.
      
      **A BLAST RADIUS COMPUTED BY NAME IS NOT A BLAST RADIUS.** Measured:
      `deepLinkHydration` appears in **nine** non-generated files and is the
      flag in **three** — the other six are about deep-link hydration the
      FEATURE. `crashReporting` appears across five unrelated product domains
      as a user preference. A removal plan built from a search sends its reader
      to edit files that have nothing to do with the flag, so every site
      carries HOW it was established and a name-matched one is refused.
      
      **EXACTLY ONE OF THE SEVEN CLASSES IS ENFORCED BY A MACHINE, AND ONLY FOR
      SIX KEYS.** `ConsumerShellFeatureFlagId` is a union type, so removing
      `deepLinkHydration` fails to compile at
      `requiredFlags: ['deepLinkHydration', 'currentDomainPersistence']` and
      narrows five `Record<ConsumerShellFeatureFlagId, boolean>` positions with
      it. The eleven registry flags are `key: string` and the seven env-map
      flags are `Record<string, boolean>` keys: removing one of those compiles
      everywhere, and `isRuntimeFeatureEnabled` returns the caller's fallback
      rather than saying it is gone.
      
      **AND THE FAMILY CLOSES WHERE IT OPENED.** S10.14.f's own precondition is
      "after verified cohort observation", and S10.14.a established that
      nothing in this estate emits a flag evaluation. So no cohort has been
      watched, `cohortWasObserved` is false for every flag, and **by the
      estate's own rule no flag may ever be removed**. `nextRemovalState`
      exists precisely to report that: a plan correctly stopped at
      `cleanup-verified` produces NO fault — it has claimed nothing it cannot
      support — so a reader of the fault list sees a clean removal going
      nowhere. Only something that tries the door finds it has no handle. One
      producer opens it and nothing else has to change; a test asserts that
      too.
      
      The plan shipped is for `homeLayoutV2`, the easiest case in the estate:
      three registries declare it and S10.14.b measured that no production
      expression reads it. It still cannot be removed, and its single refusal
      is the copy on somebody else's device.
      
      TRAP, caught by the scan's own control on its first run: the plan's three
      registry citations were tidy one-line SUMMARIES of multi-line literals —
      quotations that could not be found in the files they cited. **An ellipsis
      is a summary wearing a quotation's clothes.** All quotes are now verbatim,
      the scan re-reads every one of them in the file it names, and a test
      refuses an ellipsis in any of them. -->
      

S11 — Shared verification and release harness#

  • S11.1 Parameterize the Tara route-integrity test. Evidence: missing, duplicate, forbidden, and unowned routes fail. <!-- S11.1 CLOSED with a–g. Eight kit entry points, 409 tests, 51 negative controls across seven scans, and one result that the parent's evidence clause did not anticipate: on the host that mounts the workbench, none of the four defects it names can be detected today — not because the detectors are missing, but because the inputs are.

    **A ROUTE IS TRUE AT FOUR ALTITUDES — DESCRIBED, MOUNTED, REACHABLE,
    ADVERTISED — AND THEY DO NOT NEST** (a). Each adjacent pair is witnessed
    coming apart in `apps/oshun/web`: `/domains/tara` is mounted and undescribed,
    the 87 `/studio/isis` pages are described and unreachable behind one empty
    array, and `studio.study` hides a nav entry while the route answers. The
    test being parameterized compares two DESCRIBED facts, so it has no
    evidence side at all.
    
    Then: **A ROUTE TREE IS NOT A TREE OF ROUTES** (b) — proven by
    reproducing the framework's own normalization exactly, 772 patterns
    against 772, which took one correction no reading of the tree suggests.
    **A MISSING ROUTE IS A REQUEST THAT NOW GETS SOMETHING ELSE** (c) — five
    absorbing patterns in three classes, one of them unbounded. **TWO ROUTES
    COLLIDE WHEN A REQUEST COULD GO TO EITHER, AND THE COORDINATE IS USUALLY
    THE STAGE** (d) — `/metis/upload` answered three times and reachable once.
    **A BOUNDARY IS A CONJUNCTION AND ONLY THE TERMS THE CALLER CANNOT SET ARE
    THE BOUNDARY** (e) — one guard resting on one of four terms, one "gate"
    resting on none. **A ROUTE'S METADATA SLOT IS FULL AND NONE OF IT IS ABOUT
    US** (f) — 651 of 720 pages carry metadata and all seven of its fields are
    a crawler's. And **THE TWO SYNTHETIC DOMAINS CARRY THE METADATA THE REAL
    ONE DOES NOT** (g).
    
    Three findings in the estate worth acting on independently of the harness:
    **two shipped Tara surfaces** (`programs/[programId]`,
    `sources/[sourceId]`) are absent from the route map its own IA, hub and
    command palette read; **six of the nineteen `OSHUN_APP_LINK_PATHS` claims
    resolve to nothing**, `/home` at any depth, while both tests guarding that
    register compare it to a copy of itself; and the estate's one unbounded
    absorber, `/workspace/[[...path]]`, is empty today and that emptiness is
    the only reason a missing-route check means anything under it. -->
    
    • S11.1.a Define test inputs for app/source roots, framework/proxy/ middleware routes, route descriptors, host, audience, owner, capability, lifecycle, allowed/forbidden patterns, redirects, and feature flags. Evidence: harness contract.

      <!-- S11.1.a: shipped as
      `libs/oshun/workbench-kit/src/route-integrity.ts` (`ROUTE_ALTITUDES`,
      `ROUTE_INPUT_PROVENANCES` + `ALTITUDES_ESTABLISHED_BY`,
      `ROUTE_INPUT_CLASSES` + `ALLOWED_ALTITUDES_BY_CLASS`,
      `ROUTE_INPUT_DISCRIMINATIONS`, `EVIDENCE_COVERAGE`,
      `parseRouteInputSupply`, `ROUTE_INTEGRITY_CHECKS`,
      `resolveCheckReadiness`/`resolveAllCheckReadiness`,
      `routeInputSupplyFindings` with 6 fault kinds,
      `separationSupportFindings`, `OSHUN_WEB_ALTITUDE_SEPARATIONS`,
      `OSHUN_WEB_ROUTE_INPUT_SUPPLY`, `TARA_ROUTE_MAP_SPEC_SUPPLY`), 52 tests
      in `route-integrity.spec.ts`,
      `libs/oshun/workbench-kit/tools/route-integrity-inputs-scan.mjs` (nx
      target `route-integrity-inputs-scan`,
      `pnpm run verify:operations:v1-route-integrity-inputs`, 8 negative
      controls), and the registrations a kit entry point needs.
      
      **A ROUTE IS TRUE AT FOUR ALTITUDES — DESCRIBED, MOUNTED, REACHABLE,
      ADVERTISED — AND THEY DO NOT NEST.** The tempting reading is a ladder,
      each altitude contained in the one below, so that an input at any of them
      is a weaker version of the same fact. `apps/oshun/web` refutes it at
      every adjacent pair, in the direction that makes containment false, and
      each refutation is a live surface:
      
      - **Described ⊅ mounted.** `apps/oshun/web/src/app/domains/tara/` is a
        directory with seven children and **no `page.tsx` of its own**. No
        source file describes `/domains/tara`; the framework routes it through
        the `[domainId]` sibling — absent from the 726 static routes, matched
        by `^/domains/([^/]+?)(?:/)?$`. **Eleven of the proxy's eighteen legacy
        redirect targets are in this class**, so a source-tree check reports
        eleven broken redirects that all work.
      - **Mounted ⊅ reachable.** `apps/oshun/web/src/app/studio/isis/` holds
        **87 first-level segments, every one with its own `page.tsx`**.
        `isAaaOnlyRoute` is a length check plus a negated `includes` against
        `STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS`, which is `Object.freeze([])`, so
        every non-empty segment is AAA-only and `resolveStudioBoundary` returns
        `hard-block-404` for an ordinary session. **87 pages, zero surfaces,
        and the difference is one empty array.**
      - **Reachable ⊅ advertised.** `studio.study` is the estate's only
        route-adjacent flag and `isStudioStudyNavVisible` uses it to remove a
        **navigation entry**. The function's own doc says what that does not
        do: "Route resolution … and breadcrumbs stay aware of `/studio/study`
        regardless of this decision, so a deep link still renders." Correct
        behaviour, and it means a flag input says what is offered, never what
        is open.
      
      **WHAT THAT MAKES OF THE TEST BEING PARAMETERIZED.**
      `apps/oshun/web/src/components/studio/StudioTaraWorkbenchRouteMap.spec.ts`
      compares `entry.implemented` — a hand-maintained boolean — against
      `existsSync(pageFileFor(entry.path))`. **Both sides are DESCRIBED.** It
      never reaches the framework, the proxy, or the flags, so both witnesses
      are invisible to it: `/domains/tara` would grade missing and is not, any
      of the 87 Isis pages would grade present and is not reachable. Graded by
      the shipped contract, `TARA_ROUTE_MAP_SPEC_SUPPLY` comes out
      `source-only-supply` and all four checks come out non-runnable — so
      parameterizing it is not the substitution of a constant; the check has no
      evidence side, and eleven of the twelve input names exist to give it one.
      
      **AN INPUT THAT CANNOT TAKE TWO VALUES CANNOT MAKE A CHECK FAIL.** The
      parent evidence clause promises that unowned routes fail. Ownership here
      is `domains.json` — 58 domains, each with an `owner`, **all 58
      `@GreyChimp`** — generating a CODEOWNERS that opens `* @GreyChimp`. The
      check is unfalsifiable twice over, so the contract reports it `vacuous`
      rather than green.
      
      **AGAINST THIS ESTATE ALL FOUR CHECKS ARE UNSUPPLIED, EACH FOR A
      DIFFERENT MISSING CLASS.** `missing` and `duplicate` want a route
      descriptor: `RouteDescriptor` carries seven of the twelve names at once
      and **`parseRouteDescriptor` has no production caller** — every one of
      its nine call sites is a spec, an archetype, a `__fixtures__` module or a
      tool. `forbidden` wants a host and an audience: **no production file
      under `apps/oshun/web/src` imports either enum**. `unowned` wants a
      capability and a lifecycle: the 6714-row capability inventory has **zero
      rows under `apps/oshun/web`**, because the inventory's `workspaceRoots`
      are the seven initiative domains and its manifest calls this app the
      `externalProxyHost`. The same scoping leaves the committed
      `routes.json` with 125 rows across eight hosts and **none for the host
      that mounts the workbench** — whose only complete enumeration is the
      gitignored `.next/app-path-routes-manifest.json`, which is why the
      contract has an `evidence-not-regenerable` verdict at all.
      
      **AND THE COVERAGE AXIS THE ESTATE FORCES.** Reachability here is real
      and small: `middleware-studio-boundary.test.ts` imports the exported
      `proxy` and calls it with a `NextRequest`, over **four** `/studio/isis`
      segments (one deliberately not on disk) of the 87. Discriminating and
      sampled are orthogonal, so the register carries both, and `forbidden`
      grades `evidence-sampled` once its absent inputs are supplied — routes
      nobody looked at must not pass by not being looked at.
      
      DECISIONS. **A DESCRIPTOR IS WHAT YOU CHECK, NOT WHAT YOU CHECK WITH**:
      `declared`, `source-tree` and `source-literal` all top out at
      `described`, enforced in the parser, so no amount of reading source
      produces an evidence side. Two of S11.1.a's twelve names are compounds
      and both are split — framework/proxy/middleware because the framework
      enumerates, the proxy decides and the matcher decides whether the proxy
      is consulted; allowed/forbidden because the allow-list is empty and
      enforcing while the 31-name catalog enforces nothing and misses 56 of the
      87 segments on disk. **A CLASS REGISTERED AS UNPRODUCED IS NOT A CLASS
      THE REGISTER FORGOT**, so `unproduced-class` and `unregistered-class` are
      separate faults with separate repairs. Readiness is ordered by REPAIR —
      an absent input before a misplaced altitude before a blunt one — and the
      two silent verdicts, `altitude-collapsed` and `vacuous`, are named
      because a check in either state runs, reports green, and has refuted
      nothing.
      
      TRAPS, both caught by the module's own machinery on its first runs. The
      parser refused **my own first register**: it placed the matcher and both
      pattern lists at `reachable` on the reasoning that they decide
      reachability, and reading a frozen array is not making a request. Then
      the scan refuted **my own "nothing imports the proxy"** — a grep run from
      the wrong working directory — when
      `apps/oshun/web/src/__tests__/middleware-studio-boundary.test.ts` does
      exactly that, which upgraded the estate's reachability supply from a
      runtime observation to a proxy evaluation and changed the sample from 2
      to 4. And a third: **the register's own prose tripped its own absence
      check** — writing "`parseRouteDescriptor` has no production caller" into
      the export-map rationale made a bare-name `git grep` report that sentence
      as the producer it denies, so the caller pattern now requires the open
      paren, because a file that NAMES a symbol is not a file that calls it.
      
      Also repaired here, because it blocked this entry point: the export-map
      register in `scripts/v1-workbenches/generate-workbench-kit-export-map.mjs`
      had gone five entry points stale (S10.14.b–f), so
      `evidence/v1-workbenches/workbench-kit-export-map.json` carried **10
      violations** and `public-api.spec.ts` was red on this branch. All six
      rows are now registered and the map regenerates to **182 entry points, 0
      violations**. -->
      
    • S11.1.b Extract route discovery and normalization without Tara path/ role/capability constants. Evidence: Tara and synthetic domain snapshots.

      <!-- S11.1.b: shipped as
      `libs/oshun/workbench-kit/src/route-discovery.ts` (`SEGMENT_KINDS` +
      `SEGMENT_ARITY`, `classifySegment`, `RouteConvention` with
      `NEXT_APP_ROUTER`/`NEXT_PAGES_ROUTER`, `MetadataRouteFile`,
      `normalizeRoutePath`, `matchRoutePath`, `discoverRoutes`,
      `collisionFindings` with 6 fault kinds, `routeSetDifference`,
      `shapeOfDeclaredPath`, `parseRouteSnapshot`, and four snapshots —
      `TARA_ROUTE_SNAPSHOT` plus three synthetics), 42 tests in
      `route-discovery.spec.ts`,
      `libs/oshun/workbench-kit/tools/route-discovery-scan.mjs` (nx target
      `route-discovery-scan`,
      `pnpm run verify:operations:v1-route-discovery`, 9 negative controls),
      and the registrations a kit entry point needs.
      
      **A ROUTE TREE IS NOT A TREE OF ROUTES.** The Tara original is four
      lines — `routePath.replace('/studio/', '')` joined to `page.tsx` — and
      every one of the three things wrong with it survives the removal of the
      constants, because none of them is about Tara:
      
      - **A FILE THAT IS NOT A ROUTE.** 845 directories under
        `apps/oshun/web/src/app` yield 720 pages and 39 handlers; 19 files are
        `layout`/`loading`/`error`/`not-found`/`template` and eleven
        directories are private. The sharpest case is not in that tree at all:
        **six apps here have a `src/pages` directory that looks exactly like a
        Pages Router tree and is not one** — arete/web, demeter/web,
        hathor/workbench, isis/web, sophia/workbench, yemaya/studio-web, **256
        route-shaped files**, and the only thing that says so is the ABSENCE
        OF A `next.config` IN AN ANCESTOR, a file that is not in the tree being
        walked. BASE.1.d already paid for this: 329 routes before the gate,
        125 after.
      - **A SEGMENT THAT IS NOT IN THE URL.** `(workspace)` is a route group
        the framework deletes, and its own manifest records the mapping:
        `/(workspace)/workspace/[[...path]]/page` → `/workspace/[[...path]]`.
      - **A URL THAT IS NOT IN THE TREE.** Three optional catch-alls make
        whole subtrees infinite, so **the mounted set is unbounded and only
        the PATTERN set is finite** — "is this path mounted" is a match, never
        a lookup, and no `existsSync` can answer it.
      
      **A ROUTE SET HAS TWO ENDS, AND A CHECK THAT WALKS ONE OF THEM CANNOT BE
      SURPRISED.** The Tara spec iterates the map and asks the filesystem about
      each entry; it never iterates the filesystem. Measured 2026-08-15:
      **thirteen `page.tsx` files ship under `/studio/tara-workbench` and the
      route map declares eleven.** The two it does not know about are
      `programs/[programId]` and `sources/[sourceId]` — and the sibling
      `concepts/[conceptId]` IS declared, so this is not a convention against
      detail routes, it is two of three collections missing theirs. That map's
      own header calls it the single source of truth that "navigation, the
      studio hub, and (P12.2) the command palette all read", so two shipped
      surfaces are absent from the IA, the hub, and the palette.
      
      **THE MEASUREMENT THAT SETTLES WHETHER THE NORMALIZATION IS RIGHT.**
      There is a second implementation of it in this repository — the
      framework's — and it writes its answer down. Run over the real
      `apps/oshun/web/src/app` tree, **901 files across 845 directories, this
      module produces 772 URL patterns and `app-path-routes-manifest.json`
      contains the same 772**, with nothing in either direction. Getting there
      took one correction no amount of reading the tree would suggest: the
      first run agreed on 759 and missed thirteen, all of them files that are
      routes while sitting BESIDE `page.tsx` rather than instead of it — ten
      `opengraph-image`/`twitter-image` files, `sitemap.ts` served at
      `/sitemap.xml`, and the two reserved internal routes. That is
      `MetadataRouteFile`, and it is the first rule one level further down: a
      file that is not a route, until it is.
      
      DECISIONS. **A SLOT PAGE IS NOT A ROUTE**, unlike a group's child — the
      first draft gave `parallel-slot` a group's arity and
      `stories/[storyId]/@sidebar/page.tsx` was then reported as a duplicate
      URL, an invented collision. **TWO KEYS PER ROUTE**, because `/x/[id]` and
      `/x/[slug]` are different URL patterns and the same route: keeping only
      the pattern invents a collision, keeping only the shape cannot say which
      name to fix. **A RENAMED PARAMETER IS A NAMING FAULT AND NOT A MISSING
      ROUTE**, so `routeSetDifference` compares shapes — reporting it as both
      missing and extra sends a reader to create a file that exists. **THE
      HOST SUPPLIES THE FILE LIST** rather than the module reading a disk,
      which is what lets a synthetic domain be a literal instead of a fixture
      directory, and what makes "Evidence: Tara and synthetic domain
      snapshots" one code path over four inputs. Tara's tree uses only two of
      the eight segment kinds, so the three synthetics carry the rest: a group,
      a slot, an optional catch-all, a private directory, a non-route file and
      a foreign extension; one snapshot per collision class; and a Vite
      lookalike that reproduces the ancestor miscount.
      
      TRAP, and it is the one this repository has recorded before: the first
      draft used a **raw NUL byte** as the collision-key separator, which made
      the entire source file BINARY to `grep` — the mandatory adversarial stub
      scan then returned zero hits for exactly the same reason a clean file
      does. It is now `KEY_SEPARATOR = '\u0000'`, written as an escape, with
      the reason recorded beside it.
      
      **AND THE SAME BYTE HAD ALREADY EATEN THIS LEDGER.** Checking for it
      found a raw NUL inside the S10.2 note two thousand lines above -- the
      one that recorded the problem and said "33 files repo-wide still have
      this and are invisible to grep" -- so this 60,000-line file has itself
      been binary to `grep` ever since. Measured now: **43 tracked
      `.ts`/`.tsx`/`.mjs` files carrying 79 raw NUL bytes**, every occurrence
      a separator inside a string or template literal. All 79 are replaced
      with the `\u0000` escape, which is byte-identical at runtime and
      restores those files to the repo-wide searches the quality standards
      depend on. Every rewritten file was re-parsed with the TypeScript
      compiler -- 43 files, 0 syntax errors -- and the affected suites re-run.
      `libs/oshun/product-graph` has two failures (a manifest staleness hash
      and a signal count) which reproduce identically against the pre-change
      files and are unrelated to this work. -->
      
    • S11.1.c Detect source-described versus effectively mounted missing routes. Evidence: deliberate missing-route test.

      <!-- S11.1.c: shipped as
      `libs/oshun/workbench-kit/src/route-mounting.ts` (`ABSORPTION_BOUNDS`,
      `MOUNT_VERDICTS`, `isAbsorbing`, `exemplarPath`, `mountFindings`,
      `PATH_VERDICTS` + `pathFindings`, `absorptionZones`,
      `undetectableRegions`, `withRouteRemoved`, `parseAbsorberRecord`,
      `OSHUN_WEB_ABSORBERS`, `OSHUN_WEB_REGISTER_PATHS`,
      `OSHUN_WEB_APP_LINK_CLAIMS` + `deadAppLinkClaims`), 24 tests in
      `route-mounting.spec.ts`,
      `libs/oshun/workbench-kit/tools/route-mounting-scan.mjs` (nx target
      `route-mounting-scan`, `pnpm run verify:operations:v1-route-mounting`,
      7 negative controls), and the registrations a kit entry point needs.
      
      **A MISSING ROUTE IS A REQUEST THAT NOW GETS SOMETHING ELSE, AND ONLY
      SOMETIMES NOTHING.** Deleting a page has two outcomes and a detector that
      asks "does this path still resolve" scores them backwards. If nothing
      else matches, the request 404s: loud, monitored, reported. If a dynamic
      pattern above it matches, the request returns **200 with a different
      component**, and the surface is gone without a single error. The second
      is worse and it is the one a naive check calls healthy — so the question
      is not whether a pattern exists but **whether it refuses**.
      
      **FIVE PATTERNS IN `apps/oshun/web` ANSWER FOR MORE THAN ONE PATH, IN
      THREE CLASSES.** `/domains/[domainId]`, `/app/[surface]/[[...path]]` and
      `/d/[domain]/[[...path]]` each check their parameter against a named
      allow-list and call `notFound()` — **bounded-locally**, so a deletion
      under them stays visible. `/api/study/[...path]` validates nothing and
      returns the upstream's status — **bounded-elsewhere**, a real third class
      because the refusal exists and is not this repository's.
      `/workspace/[[...path]]` is **unbounded**: thirty-six lines, no
      validation, no `notFound`, any suffix joined into `resumePath` and the
      same `MultiPanelWorkspace` rendered, so `/workspace/anything/at/all` is a
      200. It is the one region of this estate where a deleted route cannot be
      detected as missing — and it contains **zero static routes today**, which
      is the only reason the check means anything there. Reported as a ZONE
      rather than as a clean bill of health for its contents, because the next
      page added under `/workspace/` is born inside it.
      
      **AT THE PATTERN LEVEL THIS ESTATE IS CLEAN AND THE WHOLE GAP LIVES AT
      THE PATH LEVEL.** 772 described against 772 mounted, nothing either way.
      Across every concrete path the app's registers name — the proxy's
      `PUBLIC_PATHS`, its eighteen `LEGACY_REDIRECTS` targets, and the
      navigation constants — 51 distinct paths: **44 answered by their own
      route, 4 absorbed by `/domains/[domainId]`, 3 that are prefix constants
      and not routes at all**. The four absorbed ones are `/domains/tara`,
      `/domains/veritas`, `/domains/nyx` and `/domains/arete`, and the
      absorption is **deliberate** — `DOMAIN_META` names all six navigable
      domains and an unknown id is `notFound()`. That is the whole reason the
      classes exist: a designed landing page and a swallowed deletion have the
      identical shape from outside, so a detector cannot tell them apart
      without being told which absorbers are allowed.
      
      **AND THE REGISTER WHOSE ERRORS LEAVE THE REPOSITORY.**
      `OSHUN_APP_LINK_PATHS` declares which URLs the native app claims, served
      to Apple and Google by the two `.well-known` handlers. Nineteen patterns,
      and **six have nothing behind them**: `/home`, `/share`, `/app`,
      `/domains` and `/v3` have no exact route, and `/v6/*` is a wildcard over
      a subtree with zero children. `/home` is the unambiguous one — no exact
      route, no children at any depth — and `WEB_SHELL_ROUTE_PATHS.home` is
      `'/'`, so the claim names a path the shell register itself spells
      differently. It is not entirely unhandled: `useDeepLinkHydration` maps
      `'/home'` to `'/'` in a **client-side** table beside `/explore`,
      `/activity`, `/library` and `/profile` — the only one of the five with no
      server route. So a visitor with the app installed gets a screen and a
      visitor without it gets a 404 served before any client code runs, which
      is S11.1.a's altitude lesson costing something outside the repository.
      
      **BOTH TESTS THAT GUARD THAT REGISTER COMPARE IT TO A COPY OF ITSELF.**
      `deep-link-associations.ts:111` builds the document as
      `paths: [...OSHUN_APP_LINK_PATHS]`;
      `deep-link-association-files.test.ts:41` asserts
      `expect(new Set(detail?.paths ?? [])).toEqual(new Set(OSHUN_APP_LINK_PATHS))`
      — `[...X]` equals `X` — and line 23 asserts the constant contains four
      of its own members. `deep-link-association-routes.test.ts` asserts the
      served document contains `['/d/*', '/domains/*', '/assistant/*', '/v3/*',
      '/v6/*']`, one of which covers an empty subtree. The register has exactly
      two consumers: the builder that copies it and the tests that check the
      copy. **Nothing anywhere asks whether any of the nineteen URLs resolves.**
      Same shape as the Tara route-integrity spec S11.1.b measured, in a second
      place, guarding a claim made to two app stores.
      
      DECISIONS. **THE DELIBERATE REMOVAL IS A FUNCTION**, not a hand-edited
      fixture: `withRouteRemoved` is called by the tests and by the scan, so
      the control is the same code path as the claim, and it **throws when the
      pattern was not there** — a control that silently removed nothing would
      pass for the same reason a working detector does. There are **two**
      missing-route tests because one is not enough: the same deletion must
      come out `described-not-mounted` where nothing absorbs and
      `described-but-absorbed` where something does, and a detector that passes
      the first and fails the second scores the silent case as healthy. **AN
      EXEMPLAR PATH IS `__param__`**, deliberately ugly, so an absorber that
      matches it matches because of its shape and not a coincidence of naming;
      an optional catch-all contributes nothing to its exemplar, which is the
      shortest path it accepts and the one most likely to collide with a static
      route above it. **THE ZONE IS REPORTED EVEN WHEN EMPTY**, because an
      empty unbounded zone is a property of the estate that the next commit can
      change. -->
      
    • S11.1.d Detect duplicate effective method/path/host, alias collision, case/normalization collision, and redirect cycle. Evidence: duplicate fixture tests.

      <!-- S11.1.d: shipped as
      `libs/oshun/workbench-kit/src/route-collisions.ts` (`ROUTE_COORDINATES`,
      `EffectiveRoute` + `RouteOutcome`, `COLLISION_KINDS`,
      `collisionFindings`, `aliasCollisionFindings`, `followRedirects` +
      `chainFindings`, `parseProxyStages`, `OSHUN_WEB_PROXY_STAGES`,
      `OSHUN_WEB_METIS_UPLOAD_RULES`, `OSHUN_WEB_COLLAPSED_CHAINS`), 28 tests
      in `route-collisions.spec.ts`,
      `libs/oshun/workbench-kit/tools/route-collisions-scan.mjs` (nx target
      `route-collisions-scan`,
      `pnpm run verify:operations:v1-route-collisions`, 8 negative controls),
      and the registrations a kit entry point needs. The case/normalization
      collision detector is S11.1.b's `collisionFindings`, which already
      reports `case-collision`, `duplicate-url` and `parameter-name-conflict`
      over a discovered tree; this cell adds the three the file tree cannot
      produce on its own.
      
      **TWO ROUTES COLLIDE WHEN A REQUEST COULD GO TO EITHER, AND THE
      COORDINATE THAT DECIDES IT IS USUALLY THE STAGE.** A middleware is not a
      redirect table. `apps/oshun/web/src/proxy.ts` is **seven ordered stages**
      over one request — rate limit, release scope, legacy redirects, canonical
      shell alias, authentication, the §24.11 studio boundary, the Metis upload
      alias — and each may answer and return. Two rules that would both fire do
      not both fire: the earlier wins and the later is **unreachable**, which
      is different from unused because nothing can test it.
      
      **THE ESTATE'S ONE ALIAS COLLISION IS THREE DEEP.** `/metis/upload` has
      three implementations of what happens to it:
      
      1. **Stage 2, release scope** — `resolveDeferredDomainFromRoute` matches
         on the prefixes `['', '/domains', '/d', '/api']`, so the **bare
         prefix** makes `/metis/upload` itself match the root `/metis`, and the
         gate returns a **307 to `/release-scope/metis`**.
      2. **Stage 7, `metisUploadAliasResponse`** — a **308 to `/metis/byom`**.
      3. **The page.** `apps/oshun/web/src/app/metis/upload/page.tsx` is a real
         mounted route whose whole body is `permanentRedirect('/metis/byom')`.
      
      **Only the first runs.** Metis is one of the two deferred domains and the
      gate is first on purpose — its own doc says why: "whether a room ships in
      this release is not a permission question, and answering it first keeps
      the refusal identical for everyone." So the other two are **dormant**
      rather than dead: the deferral is temporary by construction ("their route
      trees, components, and services all stay in the tree"), and both wake on
      the same day. They agree on `/metis/byom` today. **Nothing keeps that
      true, because no request reaches either of them.**
      
      **AND THE CHAIN THE ESTATE COLLAPSES ON PURPOSE.** `/news` and `/reader`
      are legacy aliases for `/domains/veritas`, and `veritas` is the other
      deferred domain — so read off the tables the chain is two hops,
      `/news` → `/domains/veritas` → `/release-scope/veritas`. The proxy
      resolves the **target's** deferral inside the redirect branch and emits
      one 307 with the reason written down: a 301 "would bake a deferred URL
      into a browser's cache". **A cycle detector over the redirect TABLE would
      report `/news` clean and would also be blind to the two hops it
      collapses**, because the graph that matters is the graph of stages.
      
      What the four detectors find on this estate: duplicate effective
      method/path/host **0** (19 redirect rules over 726 static routes, and no
      two handlers share a path — 39 real handlers, 15 GET-only, 21 POST-only,
      3 multi-method); alias collision **1**; case/normalization collision
      **0**; redirect cycle **0**.
      
      DECISIONS. **A DUPLICATE AND A SHADOW ARE DIFFERENT FINDINGS**, because
      the repairs differ — a duplicate is deleted, a shadowed rule is deleted
      OR moved earlier — and the finding is reported against the LATER rule,
      the one a reader has to decide about. **METHOD AND HOST ARE
      COORDINATES**: a GET and a POST at one path are two routes, and one path
      on two hosts is two routes unless an edge merges them (this app's one
      merge, the `/v1/:path*` rewrite to the BFF, has no local shadow — there
      is no `/v1` tree in the web app). **A CYCLE IS REPORTED BY THE REPEAT AND
      NOT BY EXHAUSTION**, so a long legitimate chain and a loop are told
      apart. **AND THE WALKER SORTS BY STAGE**, which is what makes it agree
      with the middleware instead of with the table. -->
      
    • S11.1.e Detect host/audience/boundary forbidden routes including source- mounted but middleware-blocked or inadvertently enabled surfaces. Evidence: forbidden fixture tests.

      <!-- S11.1.e: shipped as
      `libs/oshun/workbench-kit/src/route-boundaries.ts` (`TERM_CONTROL` +
      `CALLER_REACHABLE`, `BoundaryTerm`/`BoundaryGuard`, `loadBearingTerms`,
      `BOUNDARY_FAULTS` + `boundaryFindings`, `selfDeclaredGateFindings`,
      `SURFACE_VERDICTS` + `surfaceFindings`, `parseBoundaryGuard`,
      `OSHUN_WEB_BOUNDARY_GUARDS`), 23 tests in `route-boundaries.spec.ts`,
      `libs/oshun/workbench-kit/tools/route-boundaries-scan.mjs` (nx target
      `route-boundaries-scan`,
      `pnpm run verify:operations:v1-route-boundaries`, 8 negative controls),
      and the registrations a kit entry point needs.
      
      **A BOUNDARY IS A CONJUNCTION, AND ONLY THE TERMS THE CALLER CANNOT SET
      ARE THE BOUNDARY.** A guard reads as strong in proportion to how many
      conditions it names, and that is exactly backwards: every conjunct the
      requester controls is a condition they satisfy on the way in.
      
      **THE HOST AND AUDIENCE HALF HAS NO INPUT.** S11.1.a measured `HOSTS` and
      `AUDIENCES` as naming six each with no production importer under
      `apps/oshun/web/src`, so **zero** routes in this app declare either. The
      boundary half is all that can run.
      
      **ONE LOAD-BEARING TERM OUT OF FOUR.**
      `shouldBypassStudioIsisBoundaryForE2E` turns off the §24.11 hard block —
      the only thing between an ordinary session and 87 AAA-only pages — and it
      names four conditions:
      
      | term | who sets it |
      | ---- | ----------- |
      | `pathname.startsWith(STUDIO_ISIS_PREFIX)` | the caller |
      | `searchParams.get('__oshunStudioIsisLaneE2E') === '1'` | the caller |
      | `process.env.OSHUN_ENABLE_STUDIO_ISIS_LANE_E2E_BYPASS === 'true'` | **the deploy** |
      | `isLocalhostRequest(request)` | derived from the Host header |
      
      The fourth reads `request.nextUrl.hostname`, which a Next middleware
      derives from the request's Host header, and `next.config.mjs` pins no
      host — its only host-shaped setting is `allowedDevOrigins` for the dev
      server. So **three of the four arrive with the request and one does
      not.** The bypass is closed in production, and it is closed by one term.
      That is not a criticism of the guard; it is the number a reader needs,
      because four conditions look like four independent reasons and there is
      one. S10.14.c had already recorded this flag as granting a privilege and
      absent from the inventory; what is new here is which of its terms a
      caller controls.
      
      **AND ONE "GATE" WITH NO UNCONTROLLED TERM AT ALL.**
      `apps/oshun/web/src/app/studio/generation/living-scene/page.tsx` is in
      the proxy's `PUBLIC_PATHS`, so it is reached without a session. It
      renders `data-living-scene-route-entitlement-gate="route"` — the route
      asserting that it **is** the gate — and builds its whole access context
      from six `x-oshun-*` request headers (`tier`, `living-scenes-enabled`,
      `living-scene-templates`, `living-scene-frozen-workflow-classes`,
      `living-scene-locales`, `locale`). The proxy copies inbound headers
      verbatim — `new Headers(request.headers)` plus a request id and a locale
      — and a search of its 666 lines for `x-oshun` returns **nothing**: none
      of the six is set, stripped, or validated on the way in. It is the only
      route in the app that reads an `x-oshun-*` header.
      
      **THE SAME HEADER NAME IS DISTRUSTED ONE LAYER DOWN.**
      `libs/shared/bff-kit/src/entitlements.ts` reads `x-oshun-tier` and gates
      it on `!isProductionRuntime()`, with the reason written beside it: "in
      production a client can no longer self-assert 'pro'." Two layers read one
      header and one of them believes it. **What that buys a caller is
      bounded** — the template list is still intersected with
      `APPROVED_LIVING_SCENE_TEMPLATE_IDS`, `x-oshun-living-scenes-enabled`
      defaults to enabled so it can only close the surface, and the generation
      itself goes through the BFF where the header is production-gated. So this
      is **a gate that is not one, rather than a demonstrated escalation**, and
      the register says so in those words.
      
      Coverage, measured: **503 mounted `/studio/*` routes, 88 under the
      boundary prefix, 415 the boundary says nothing about**; and zero of the
      31 catalogued AAA capability names appear at any path outside
      `/studio/isis/`, so there is no name-shaped leak.
      
      DECISIONS. **A HEADER-DERIVED TERM IS ITS OWN CONTROL**, separate from
      `caller-set`, because it READS as server state and is not:
      `isLocalhostRequest` looks like a fact about where the process runs and
      is a fact about the request being examined, and filing it under
      `server-state` would count the bypass as resting on two terms instead of
      one. **`single-load-bearing-term` IS A FINDING AND NOT AN ERROR**,
      because a guard resting on one deployment-set term is a normal correct
      shape and what is not normal is believing it rests on four. **A
      SELF-DECLARED GATE IS A SEPARATE SUBJECT** from a guard in a pipeline,
      because the assertion is the reason to look. **AND THIS MODULE DOES NOT
      DECIDE WHETHER A BOUNDED GUARD IS STRONG ENOUGH** — that is a judgement
      about threat and it is the host's; it reports how many of the reasons are
      reasons. -->
      
    • S11.1.f Detect missing/unknown owner, capability, contract/version, audience, lifecycle, and deprecation/migration metadata. Evidence: ownership fixture tests.

      <!-- S11.1.f: shipped as
      `libs/oshun/workbench-kit/src/route-metadata.ts`
      (`ROUTE_METADATA_CLASSES` — eight, both compounds split —
      `METADATA_RESOLUTIONS` + `CAN_REPORT_UNKNOWN`, `MetadataSupply`,
      `METADATA_FAULTS` + `metadataFindings`, `MetadataFieldCensus` +
      `slotFindings`, `parseMetadataSupply`, `OSHUN_WEB_ROUTE_METADATA`,
      `OSHUN_WEB_METADATA_FIELDS`, `OSHUN_WEB_CANONICAL_MISMATCHES`), 22 tests
      in `route-metadata.spec.ts`,
      `libs/oshun/workbench-kit/tools/route-metadata-scan.mjs` (nx target
      `route-metadata-scan`, `pnpm run verify:operations:v1-route-metadata`,
      7 negative controls), and the registrations a kit entry point needs.
      
      **A ROUTE'S METADATA SLOT IS FULL, AND EVERY FIELD IN IT IS ADDRESSED TO
      A MACHINE THAT IS NOT US.** The framework gives every page a metadata
      export and this estate fills it: **651 of 720 pages export a `metadata`
      object**, six more compute one with `generateMetadata`. Its fields,
      counted exhaustively: `title` 651, `description` 648, `alternates` 620,
      `robots` 35, `openGraph` 15, `twitter` 4, `keywords` 1 — a crawler's
      brief and a link preview's, without exception. Not one of the eight
      classes appears in any of them, and **no page exports a home-grown route
      constant either**: a sweep for a screaming-case export over every page
      returns nothing, and the only route segment config in use is `dynamic`,
      on four pages. The nearest thing to an audience declaration is
      `robots: { index: false }`, on **36** pages — the only per-route
      statement about who may see a surface, and it is addressed to a search
      engine.
      
      That pairing is the finding. Eight zeroes on their own would leave a
      reader unable to tell "nobody writes route metadata" from "everybody
      writes route metadata and none of it is this", and the repairs differ.
      
      **A DEFAULT IS NOT AN ANSWER; IT IS THE ABSENCE OF ONE WEARING AN
      ANSWER'S CLOTHES.** Owner is the single class that resolves for every
      route, and it resolves because `.github/CODEOWNERS` opens with
      `* @GreyChimp` — a catch-all over every path. A checker that scored
      resolutions reports owner **720 of 720** and is measuring a wildcard.
      **"Unknown owner" is not a state this estate can be in**, which is why
      `MetadataResolution` separates `defaulted` from `declared` and why the
      finding is `resolution-is-a-default` rather than a pass.
      
      Of the eight: **one resolves by default and cannot say "nobody"; one —
      capability — has a 6714-row register with zero rows under
      `apps/oshun/web`; and six have no producer at all.** The types exist in
      the kit (`HOSTS`, `AUDIENCES`, `LIFECYCLES`, `SemanticVersion`,
      `redirectTo`) and `parseRouteDescriptor`, which is where a route would
      carry them, has no production caller.
      
      DECISIONS. **`contract` AND `contract-version` ARE SEPARATE** because the
      descriptor carries the second and names the first nowhere — the estate
      can state a version of a thing it never names. **`deprecation-notice` AND
      `migration-target` ARE SEPARATE** because `lifecycle: 'deprecated'` and
      `redirectTo` are independent, and the announcement without the exit is a
      real state; the estate does move callers — eighteen legacy redirects and
      two canonical prefixes — and it does so in the proxy, where the route
      being left behind says nothing about where its callers went.
      **`alternates.canonical` IS THE ONE FIELD ANOTHER ROUTE COULD
      CONTRADICT**: 614 declare one and **exactly one differs from its own
      path** — `/workspace/[[...path]]` declaring `/workspace`, which is
      correct, because an optional catch-all renders one page for an unbounded
      set of URLs and saying so is what the field is for. A second
      disagreement would be a route calling itself a duplicate of another, and
      nothing else in this estate can express that.
      
      TRAP, caught by the scan on its own first run: this register's census
      counted `robots` with a whole-file grep (**39**) where a census counts a
      top-level metadata field (**35**), and `index: false` counts **36** —
      three different questions collapsed into one number. All three are now
      recorded separately and the scan recounts each from the tree. -->
      
    • S11.1.g Produce source/actionable failure diff and run Tara plus two archetype fixtures. Evidence: parameterized suite.

      <!-- S11.1.g: shipped as
      `libs/oshun/workbench-kit/src/route-suite.ts` (`DIFF_ACTIONS`,
      `DiffEntry` with its `blocks` leverage count, `parseDiffEntry` +
      `diffFindings`, `SuiteDomain`/`SuiteCell`, `runSuite`, `suiteDiff`,
      `checkCoverage`, and the three domains `TARA_DOMAIN`,
      `STORY_DESK_DOMAIN`, `INCIDENT_CONSOLE_DOMAIN`), 21 tests in
      `route-suite.spec.ts`,
      `libs/oshun/workbench-kit/tools/route-suite-scan.mjs` (nx target
      `route-suite-scan`, `pnpm run verify:operations:v1-route-suite`, 6
      negative controls), and the registrations a kit entry point needs.
      
      **THE TWO SYNTHETIC DOMAINS CARRY THE METADATA THE REAL ONE DOES NOT.**
      `STORY_DESK_ROUTES` and `INCIDENT_CONSOLE_ROUTES` are the only route
      descriptors in this repository — **seven between them** — and every one
      states an owning capability, a contract version, a lifecycle, an audience
      set, a host, a scope binding, a data classification, a redirect target
      and a flag list, each built through `parseRouteDescriptor` under
      `expectOk`, which is what makes them descriptors rather than object
      literals shaped like descriptors. Tara — the quality exemplar, the real
      domain, the one this whole ledger is measured against — states a path, a
      purpose, a surface, an `ownerRole` and a boolean. **Run the same four
      checks over all three and the fixtures are the only ones that can
      answer.**
      
      That is not a defect in the fixtures and not a scolding of Tara. It is
      what a parameterized suite is FOR: the archetypes are what a supplied
      domain looks like, and running them beside the real one is the only way
      to see that the real one is not supplied. A suite over Tara alone reports
      four unsupplied checks with nothing to compare them against.
      
      **THE MATRIX, MEASURED.** No check is runnable on any of the three, and
      the domains block on **different** classes — which is the whole reason
      the cell asks for three:
      
      | | missing | duplicate | forbidden | unowned |
      | --- | --- | --- | --- | --- |
      | tara | `framework-route` | `framework-route`, `redirect` | + `host`, `audience` | + `capability`, `lifecycle`, `owner` |
      | story-desk | + `app-source-root` | `framework-route` | (no host/audience gap) | `owner` only |
      | incident-console | + `app-source-root` | `framework-route` | (no host/audience gap) | `owner` only |
      
      Tara has a file tree and a register and no metadata; the archetypes have
      complete metadata and no tree. The mirror image, exactly. And the two
      archetypes block identically, which is the control: they share no route
      path, so a difference between them would be a harness bug rather than a
      fact about either.
      
      **A FAILURE DIFF IS ACTIONABLE ONLY IF IT NAMES SOMETHING SOMEBODY CAN
      OPEN.** "Route X is missing" is a fact; "X is declared at
      `StudioTaraWorkbenchRouteMap.ts` and no file under
      `app/studio/tara-workbench` produces it" is a repair, because it names
      both ends and the reader knows which one to change. Every entry carries a
      SOURCE that must look like a repo-relative path — `parseDiffEntry`
      refuses "the route map" or "the proxy", a name for a thing rather than a
      place — and a REMEDY that must describe a change. Over the three domains
      the diff has **25 entries**, and the three at the top are the same class
      for each domain: `framework-route`, **unblocking 4 of 4 checks apiece**,
      against twenty-two that unblock one.
      
      DECISIONS. **THE OUTPUT IS A MATRIX AND NOT A VERDICT**, because a
      boolean reports the same green for a domain that passes every check and
      one that can run none. **ONE ENTRY PER DOMAIN PER BLOCKING CLASS**, not
      per failing cell, because a domain that supplies a route descriptor
      unblocks two checks at once and a diff listing checks would send its
      reader round one repair twice. **THE LEVERAGE COUNT IS A FACT ABOUT THE
      WHOLE MATRIX** and is computed before any entry is built. -->
      
  • S11.2 Parameterize a live-data/full-journey Playwright scaffold. Evidence: fixtures are opt-in and visibly labeled.

    • S11.2.a Define workbench fixture for host/base URL, actor/tenant/roles, capability probes, data reset/seed/read-back, critical routes/journeys, state selectors, and evidence output. Evidence: scaffold interface.

      <!-- S11.2.a: shipped as
      `libs/oshun/workbench-kit/src/journey-scaffold.ts` (`SCAFFOLD_SIDES`,
      `SCAFFOLD_SLOTS` + `SCAFFOLD_SLOT_SPECS`, `INSTALLATION_SHAPES` +
      `SHAPE_DECLINE_GRAIN` + `DECLINABLE_SHAPES`, `INTERPOSITION_SITES` +
      `SITE_APP_OBSERVABILITY` + `APP_INVISIBLE_SITES`, `LABEL_CHANNELS` +
      `CHANNEL_REQUIRES_APP_KNOWLEDGE` + `labelChannelsAvailableFor`,
      `READ_BACK_PATHS`, `ABSENCE_POLICIES`, `parseScaffoldSlotSupply`,
      `SCAFFOLD_FAULTS` with 8 kinds, `scaffoldSupplyFindings`,
      `scaffoldSupplyCensus`, `SCAFFOLD_JOURNEYS` +
      `resolveJourneyReadiness`/`resolveAllJourneyReadiness`,
      `TARA_E2E_SCAFFOLD_SUPPLY`, `CENTRAL_E2E_SCAFFOLD_SUPPLY`,
      `OSHUN_FIXTURE_LABEL_REACH`), 114 tests in `journey-scaffold.spec.ts`,
      `libs/oshun/workbench-kit/tools/journey-scaffold-scan.mjs` (nx target
      `journey-scaffold-scan`, `pnpm run verify:operations:v1-journey-scaffold`,
      12 negative controls), and the five registrations a kit entry point
      needs.
      
      **A SUBSTITUTE IS INSTALLED ON ONE SIDE OF THE GLASS AND LABELLED ON THE
      OTHER.** The parent's evidence clause — "fixtures are opt-in and visibly
      labeled" — reads as one demand with two adjectives. It is two demands on
      two different parties, and each fails here for a structural reason rather
      than by oversight:
      
      - **Opt-in is a property of the INSTALLATION SHAPE, not of the intent.**
        Playwright has exactly one shape a test can decline: the option fixture
        `[default, { option: true }]`, settable by `test.use()`. A `page`
        override, an `auto` fixture, an `afterEach`, a project `use:` block, a
        `globalSetup` and a `webServer` env var are installed for every test in
        their scope and no test can refuse them.
        `apps/tara/web/e2e/fixtures.ts` contains **both shapes within thirty
        lines of each other**: `skipAxe` arrives from
        `libs/shared/testing/src/playwright/axe.ts` as
        `[false, { option: true }]` and any spec may switch the accessibility
        scan off, while the **seven `context.route` handlers** that decide what
        the product is shown are installed by overriding `page` itself, so all
        **ten spec files** under `e2e/tests` receive them and none can refuse.
        The declinable one is the accessibility check; the undeclinable one is
        the fixture data.
      - **Visible labelling is a property of the INTERPOSITION SITE, not of the
        product's willingness.** `route.fulfill` replaces the response BENEATH
        the application, which has no fact available to it saying otherwise.
        This estate's badge — `FixtureBadge` in `CapabilityBanner.tsx`,
        rendering `data-fixture-badge` / `data-fixture-source` /
        `data-fixture-version` — is driven by a `fixtureSource` **prop**, so it
        can only ever announce a substitution the application performed itself,
        and a scaffold's substitutions are precisely the ones it cannot see.
        `labelChannelsAvailableFor` DERIVES that from the observability table
        and the app-knowledge table rather than asserting it.
      
      **AND THE ESTATE PROVES IT FROM THE OTHER SIDE**, which is why the
      conclusion is structural and not a complaint. There IS a working fixture
      label in this product, on a mounted route: `/operator/metis` asks the BFF
      for OneRoster with `timeoutMs: 3000`, returns `getMetisOperatorFixture()`
      when nothing comes back, and renders
      `data-operator-metis-oneroster-source`, typed `'fixture' | 'bff'` and set
      to the first by the fixture and the second by `adaptOneRosterStatus`. It
      works because the substitution is the page's OWN. **The badge built for
      the general case is unreachable** — `apps/oshun/web/src/components/
      workbench/` has zero importers outside its own directory — **and the
      annotation built for one specific fallback is live.**
      
      **A READ-BACK THROUGH THE CHANNEL THAT WROTE IT CONFIRMS THE WRITE AND
      NOTHING ELSE.** `apps/tara/web/e2e/auth.setup.ts` plants a session with
      `page.addInitScript` into `localStorage`, then asserts
      `expect.poll(() => localStorage.getItem(key)).toBe(JSON.stringify(
      persistedAuthState))` — the exact value it just wrote, out of the exact
      store it wrote it to. The suite's only real login,
      `'authenticate as free user'`, is `setup.skip(true, ...)`.
      
      Measured 2026-08-15, **the two exemplar scaffolds fail oppositely**,
      which is why both are registered. Tara substitutes the actor, the roles
      and the seed — all three invisible to the app, none labelled — so
      `authenticated-navigation` grades **fabricated**. `testing/e2e`
      substitutes nothing and grades **unverified**, because an absent
      `TEST_USER_EMAIL` makes its setup log "creating empty auth state", write
      the storage state and return **green**, and its `globalSetup`
      `console.warn`s a failed health check and continues. **Neither suite
      fills tenant, capability-probe, data-reset or data-read-back**, so
      neither of S11.2.g's two exemplar journeys is runnable on either — and
      **neither has an opt-in switch at all**: no `PLAYWRIGHT_LIVE`, `E2E_LIVE`,
      `USE_FIXTURES` or any sibling appears in either tree, so "opt-in" has no
      mechanism to be true of yet. -->
      
    • S11.2.b Provision unique nonfixture tenant/account/data through real service boundaries and verify source values before UI use. Evidence: setup/teardown tests.

      <!-- S11.2.b: shipped as
      `libs/oshun/workbench-kit/src/journey-provisioning.ts`
      (`PROVISIONED_SUBJECTS` × `PROVISIONING_ACTS`, `BOUNDARY_KINDS`,
      `AUTHORITY_KINDS` + `AUTHORITY_ORIGINS`, `IDENTITY_SOURCES`,
      `READ_BACK_MODES`, `RESIDUE_KINDS`, `parseProvisioningStep`,
      `resolveProvisioningChain`/`resolveAllProvisioningChains` over five
      `CHAIN_VERDICTS`, `PROVISIONING_FAULTS` with 9 kinds,
      `provisioningFindings`, `cleanupPathFindings`, `provisioningCensus`,
      `authorityAsymmetries`, `OSHUN_PROVISIONING_REGISTER`,
      `OSHUN_AUTHORITY_SUPPLY`, `OSHUN_CROSS_PROCESS_CONTROLS`,
      `OSHUN_CLEANUP_PATHS`), 109 tests in `journey-provisioning.spec.ts`
      organised as the SETUP / VERIFY / TEARDOWN questions the evidence clause
      asks, `libs/oshun/workbench-kit/tools/journey-provisioning-scan.mjs` (nx
      target `journey-provisioning-scan`,
      `pnpm run verify:operations:v1-journey-provisioning`, 14 negative
      controls), and the five registrations a kit entry point needs.
      
      **EVERY PROVISIONING ACT IS DECIDED IN THE SERVER'S PROCESS, AND THE
      HARNESS'S ONLY LEVER IS A REQUEST.**
      
      **A GRANT IS ISSUED, NOT REQUESTED — AND THIS ESTATE HAS TWO REAL
      BOUNDARIES WITH NO LINK BETWEEN THEM.** `POST /v1/tenants/onboard` is a
      genuine tenant provisioner: `originGuard`, `csrfGuard`,
      `collectionsAuthGuard`, contract validation, and
      `tenantOnboardingStore.onboardDurably`. It admits an actor only through
      `hasTenantOperatorScope` — `admin:*` or `admin:workspace:tenant-console`.
      And `buildAuthSession`, the single function through which every signup
      and every login issues a token, writes `scopes: ['domain:*']` as a
      literal taking no input — **the only `scopes:` literal in
      `customer-auth-store.ts`**. So no account this estate can create can
      create a tenant, and the gap is not a missing permission model but a
      missing edge.
      
      **AN AUTHORITY YOU CAN MINT IS NOT AN AUTHORITY YOU WERE GRANTED.**
      `createDevAuthToken` is `` `dev.${toBase64Url(JSON.stringify(payload))}` ``
      — unsigned, so any scope set can be written by hand. The estate closes
      that correctly: `areDevTokensAllowed` returns `false` whenever
      `NODE_ENV`, `OSHUN_ENV` or `RUNTIME_ENV` is `production`, refusing the
      token `401 dev_tokens_disabled`. Which is exactly why a provisioning
      helper built on one is a trap — it works on a developer's machine and the
      first environment that matters refuses it with a status the journey reads
      as a product defect.
      
      **A SETUP THAT FALLS BACK IS NOT A SETUP.**
      `apps/oshun/web/src/app/api/dev/seed-session/route.ts` is the only
      account provisioner a test can call and it is real: it posts
      `/v1/auth/signup` to the BFF over HTTP and sets both session cookies. Its
      email defaults to the constant `'e2e@oshun.local'` and a `409` falls back
      to `/v1/auth/login` with the same fixed password — which its own doc
      calls idempotent, and which makes a **unique** account unobtainable
      through it.
      
      **FOUR CONTROLS ARE SET ON THE WRONG SIDE.** (1) The seeder assigns
      `process.env.OSHUN_DEV_AUTO_VERIFY_EMAIL = 'true'` in the **Next**
      process and `customer-auth-store.ts` reads it in the **BFF** process,
      across the `fetch` in `forwardJsonRequest` — the line is inert, and the
      behaviour it claims to cause belongs to `apps/oshun/bff/project.json`,
      whose dev target sets the variable on the server's command line. (2) The
      plan is `resolveDevSignupPlan()` reading `OSHUN_DEV_SIGNUP_PLAN` from the
      BFF's environment, so a harness needing a `premium` actor cannot ask for
      one. (3) Whether an immediate deletion erases is decided at boot by a
      Redis URL **and** an attestation signer; without both the request is
      honestly left `scheduled` for `GRACE_PERIOD_DAYS = 30`. (4)
      `allowFixtures()` is `!isProductionRuntime()`, so outside production the
      tenant roster concatenates three fixture tenants onto the onboarded ones
      **with no field telling them apart**.
      
      THE MATRIX, computed: **three of five subjects are provisionable end to
      end** (account, session, domain-record). `tenant/create` is
      `authority-self-asserted` and `tenant/destroy` is `no-boundary`, because
      `TenantOnboardingStore` has `onboardDurably` and **no removal of any
      kind** — its only deletion vocabulary is `deletedActorDigests`, a DSAR
      fence on the operator who onboarded. `authority/create` is `no-boundary`.
      **No subject but `session` can be reset**, and the harness's own cleanup
      — the `CLEANUP_TEST_DATA` branch of
      `testing/e2e/setup/global-teardown.ts` — announces
      `'Cleaning up test data...'` over a body that is one comment. -->
      
    • S11.2.c Keep fixture providers/data/intercepts opt-in per test with persistent visible badge and metadata. Evidence: fixture mode tests.

      <!-- S11.2.c: shipped as
      `libs/oshun/workbench-kit/src/fixture-mode.ts` (`FIXTURE_CLASSES` +
      `CLASS_INSTALLED_IN`, `FIXTURE_GATES` + `GATE_DECIDER` + `GATE_GRAIN` +
      `GATE_MAY_FIRE_IN_PRODUCTION` + `TEST_SETTABLE_GATES`, `ANNOUNCEMENTS` +
      `ANNOUNCEMENT_REACHES_RENDERER` + `canFeedLabelContract` +
      `labelContractFeeds`, `parseFixtureGateSupply`, `FIXTURE_MODE_FAULTS`
      with 8 kinds, `fixtureModeFindings`, `fixtureModeCensus`,
      `resolveOptInReadiness`/`resolveAllOptInReadiness` over six
      `OPT_IN_READINESS` states, `OSHUN_FIXTURE_GATE_SUPPLY`,
      `OSHUN_LABEL_CONTRACT_FEEDS`, `OSHUN_REQUEST_CHANNEL`), 109 tests in
      `fixture-mode.spec.ts`,
      `libs/oshun/workbench-kit/tools/fixture-mode-scan.mjs` (nx target
      `fixture-mode-scan`, `pnpm run verify:operations:v1-fixture-mode`, 14
      negative controls), and the five registrations a kit entry point needs.
      
      **THE LABEL CONTRACT IS COMPLETE, AND ITS FIRST FIELD IS THE ONE NOBODY
      IN THIS ESTATE CAN SUPPLY.** Half of this cell was already built and
      built well: S7.3.c's `FixtureLabelling` carries seven fault kinds,
      `REQUIRED_LABEL_SURFACES` insists the badge leaves the banner for the
      working context, and it refuses a label without a source, without a
      version, dismissible, or signalled by colour alone. It begins
      `readonly fixtureActive: boolean` — and **nothing outside
      `capability-banner.spec.ts` has ever constructed one**, so not one of the
      seven has ever been asked about a real surface. The estate says why in
      its own words: `useCapabilityProbe` sets the origin from
      `transport.token === null ? 'unauthenticated-probe-response' :
      'authenticated-probe-response'` under the comment *"The origin is a fact
      about the REQUEST, decided here and not inferred from the answer: a 200
      says nothing about whether a credential was sent."* Exactly so — and the
      same sentence forbids the client ever naming `fixture-file`,
      `demo-dataset` or `developer-override`. **Seven of the nine
      `LIVE_CLAIM_ORIGINS` have no producer, and the three that name a
      substitution are among them.**
      
      **"OPT-IN" PRESUPPOSES ONE SWITCH, AND THIS ESTATE HAS FIVE GATES ON ONE
      CLASS OF THING.** Seventeen `*_FIXTURE` constants in `domain-stubs.ts`:
      **10** through `guardedFixtureRoute` (`503 not_configured` in
      production); **2** through `jsonRoute`, which has no condition at all and
      therefore answers in production; **2** behind `allowFixtures()` =
      `!isProductionRuntime()`, both of which **concatenate** onto live rows
      with no field telling them apart; **1** behind
      `input.includeFixture ? ARETE_OFFERINGS_FIXTURE.offerings : []` — the
      only gate *shaped* to take an outside answer, handed
      `includeFixture: allowFixtures()`; and **2** behind `if (!pg)`, which
      serve whenever Postgres is unreachable, in any environment, precisely
      when a reader most needs telling. Five gates, four deciders — the
      deployment, nobody, a dependency's health, and one line of in-process
      code passing the deployment's answer along — and **not one of them is the
      request**.
      
      **THE ONLY THING THAT CROSSES ALL THREE PROCESSES IS THE REQUEST.** A
      provider is chosen in the service, data in the renderer's layer, an
      intercept in the Playwright process. The channel already exists and is
      used heavily — **the BFF reads 45 distinct `x-oshun-*` request
      headers** — and not one is about fixtures.
      
      WHAT IS ALREADY RIGHT, AND PROVES THE REST BUILDABLE: `/operator/metis`
      falls back to `getMetisOperatorFixture()` when the BFF returns nothing
      and renders `data-operator-metis-oneroster-source`, typed
      `'fixture' | 'bff'`. It works because the **renderer** made the choice.
      And **its forty-four siblings are the same shape** — `getAreteHabits`
      asks `/v1/arete/habits` and returns `getAreteHabitsFixture()` exactly as
      `getMetisOperator` does, with no field to say so. The difference is not a
      mechanism, a boundary or a permission: one of them wrote the fact down.
      
      Counted by how each gate is **wired** rather than by its shape (the one
      `caller-parameter` gate is production-safe because its argument is
      `allowFixtures()`), **49 of the estate's 69 substitutable subjects can
      answer a production request, and none of the 49 leaves a trace.** All
      three classes grade short of opt-in per test: provider
      `deployment-scoped`, data `undeclinable`, intercept `run-scoped`. -->
      
    • S11.2.d Fail when interception, hard-coded fallback, demo seed, empty success, or build-time guess masquerades as live. Evidence: anti-mock negative tests.

      <!-- S11.2.d: shipped as `libs/oshun/workbench-kit/src/anti-mock.ts`
      (`MASQUERADE_CLASSES` + `ABSENCE_CLAIMED`, `VANTAGE_POINTS`,
      `EVIDENCE_KINDS` + `EVIDENCE_VANTAGE` + `EVIDENCE_IS_SHAPE_ONLY` +
      `POSITIVE_TRACE_EVIDENCE`, `ANTI_MOCK_CHECKS` + `antiMockCheck`,
      `parseEvidenceSupply`, `resolveAntiMockReadiness`/
      `resolveAllAntiMockReadiness` over five `ANTI_MOCK_READINESS` states,
      `ANTI_MOCK_FAULTS` with 6 kinds, `antiMockFindings`, `antiMockCensus`,
      `OSHUN_ANTI_MOCK_EVIDENCE`, `OSHUN_INVERTED_INSTRUMENTS`), 76 tests in
      `anti-mock.spec.ts`,
      `libs/oshun/workbench-kit/tools/anti-mock-scan.mjs` (nx target
      `anti-mock-scan`, `pnpm run verify:operations:v1-anti-mock`, 13 negative
      controls), and the five registrations a kit entry point needs.
      
      **EVERY ONE OF THE FIVE MASQUERADES IS A CLAIM ABOUT WHAT DID NOT
      HAPPEN, AND A RESPONSE ONLY RECORDS WHAT DID.** An interception means no
      request left the browser; a hard-coded fallback means no query ran; a
      demo seed means nobody created that row in this run; an empty success
      means nothing was asked; a build-time guess means nothing has been
      observed since the build. Not one leaves a positive mark in the artifact
      a test looks at — the 200 is a real 200, the body is well-formed, the
      schema validates, the screenshot is correct. So an anti-mock check cannot
      be a predicate over the response; it has to be a predicate over the
      response **and a record of what should have happened**, which is exactly
      what S11.2.b's read-back and S11.2.c's announcement exist to produce.
      
      **THE SAME OBSERVATION IS A DEFECT AND A CORRECT ANSWER, AND ONLY WHAT
      YOU PUT THERE TELLS THEM APART.** This ledger asks for both directions
      inside one section: S11.2.d wants an empty success masquerading as live
      to FAIL, and **BASE.7.d** wants a legitimate empty result exercised with
      the UI verified not to substitute sample content — so an empty answer
      must also stand. The two responses are byte for byte the same. A check
      that fires on emptiness fails the second requirement to satisfy the
      first, so `empty-success` is refutable by `provisioned-read-back` and by
      nothing else.
      
      FOUR OF THE SIX INSTRUMENTS EXIST HERE, AND EVERY ONE IS POINTED
      ELSEWHERE. `response.serverAddr()` returns `null` for a fulfilled route
      and an address for a served one — the exact positive trace an
      interception erases — and is **called nowhere**, while
      `fromServiceWorker()` is called **twelve times** across ten files under
      `apps/oshun/web/e2e` and every one asserts `toBe(true)`: the correct
      assertion for an offline-shell test and the exact opposite of an
      anti-mock one. `waitForRequest`/`page.on('request')` appear in
      **fourteen** files there and **one** across the two exemplar journey
      suites. S1.13.d's whole runtime fixture boundary —
      `SYNTHETIC_SEGMENT_PREFIX`, `findSyntheticValues` with declared depth and
      node bounds, `assertNoSyntheticValues` throwing `SyntheticValueError`,
      and no `stripSynthetic` because a half-fixture value is one nobody can
      reason about — has **no consumer under `apps/` at all**, and not one of
      the sixty-nine substitutable subjects S11.2.c counted carries the prefix.
      The response annotation has one producer, `/operator/metis`, which
      neither exemplar suite visits.
      
      AND THE TWO IT DOES NOT HAVE: a **provisioned read-back** before UI use
      (S11.2.b measured zero in both suites, and it is the only instrument that
      can grade `empty-success`), and a **build stamp** — no build id, commit
      sha or bundle timestamp reaches any response from `apps/oshun/web` or
      `apps/oshun/bff`.
      
      SO OF THE FIVE, **exactly one is refutable in a journey suite today** —
      `interception`, and on the weaker of its two instruments: `waitForRequest`
      proves a request left, `serverAddr()` would prove who answered, and the
      suites produce the first. `hard-coded-fallback` and `build-time-guess`
      grade `instrument-elsewhere`; `demo-seed` and `empty-success` grade
      `instrument-unused`.
      
      TRAP, caught by the scan on its first run: this register said ten
      `fromServiceWorker()` call sites because I had read a `head -10`. The
      real count is twelve, across ten files. **A count taken from a truncated
      listing is a count of the listing.** -->
      
    • S11.2.e Provide reusable authenticated navigation, direct-link, network/ console capture, save/read-back, job wait, gate/review/publish, failure, and cleanup helpers. Evidence: example journeys.

      <!-- S11.2.e: shipped as
      `libs/oshun/workbench-kit/src/journey-helpers.ts` (`HELPER_FAMILIES` +
      `HELPER_PROMISE` + `ACCEPTABLE_SIGNALS`, `SIGNAL_KINDS` +
      `SIGNAL_ESTABLISHES_A_FACT` + `ABSENCE_SIGNALS`,
      `TIMEOUT_DISPOSITIONS`, `parseHelperSupply`, `resolveHelperReadiness`/
      `resolveAllHelperReadiness` over five `HELPER_READINESS` states,
      `HELPER_FAULTS` with 6 kinds, `helperFindings`, `helperCensus`,
      `OSHUN_JOURNEY_HELPER_SUPPLY`, `OSHUN_SHARED_WAITS`), 66 tests in
      `journey-helpers.spec.ts`,
      `libs/oshun/workbench-kit/tools/journey-helpers-scan.mjs` (nx target
      `journey-helpers-scan`, `pnpm run verify:operations:v1-journey-helpers`,
      13 negative controls), and the five registrations a kit entry point
      needs.
      
      Read the parent's list once and every entry is the same shape: do a
      thing, then wait until it has happened. **So the reusable part is not the
      doing. It is the waiting, and a helper is only as good as the signal it
      stops on.**
      
      **A SIGNAL WHOSE TRUTH CONDITION IS AN ABSENCE IS SATISFIED SOONEST BY
      THE THING THAT NEVER HAPPENED.** `waitForLoadState('networkidle')`
      resolves when there has been no network activity for a moment — a fact
      about the page having stopped asking, not about it being ready. A page
      whose requests were all fulfilled from the test process settles sooner
      than one that made round trips, and a page that asked for nothing is idle
      immediately, so **the estate's default readiness wait is fastest exactly
      when S11.2.d's masquerades are present.** It cannot tell "everything
      arrived" from "nothing was requested", because those are the same
      observation. It is reached from **23 call sites** across the Tara suite's
      page objects, **eight of them in `auth.page.ts`**.
      
      **A WAIT THAT SWALLOWS ITS TIMEOUT IS A SLEEP**, and
      `apps/tara/web/e2e/pages/base.page.ts` states the distinction twice in
      one file: `isVisible()` wraps a `waitFor` in a `try` and returns `false`
      on throw — **correct**, it is a predicate and both answers are answers —
      while `waitForLoadingComplete()` does
      `loadingSpinner.waitFor({ state: 'hidden', timeout: 30000 }).catch(() => {})`
      under the comment "Loading spinner might not exist, which is fine", so it
      returns whether the spinner went away or thirty seconds elapsed with it
      still turning; `waitForToastDismiss()` is the same shape at ten. The
      author knew the distinction — the predicate proves it — and two of the
      three waits are on the wrong side of it.
      
      **A RECORDER THAT WRITES NOWHERE OBSERVED AND DID NOT RECORD.** Both
      capture call sites subscribe with `page.on(...)` and hold what they hear
      in a local, and S11.2.a measured that neither suite calls
      `testInfo.attach` at all.
      
      WHAT THE TWO EXEMPLAR SUITES ACTUALLY WAIT ON, counted: `toBeVisible`
      **683**, `toHaveURL` **61**, `networkidle` **26**, and one call site
      apiece for `page.on('console')`, `page.on('pageerror')`,
      `page.on('request')` and `waitForResponse`. `expect.poll` and
      `waitForRequest`: **zero**. **Four call sites in total observe anything
      the network did**, against 744 that watch the DOM and the address bar.
      
      THE MATRIX: **two of nine families establish their promise** —
      `direct-link` (and only because `page.goto`'s own default load event does
      the work, not because anything chose it) and `failure-surface`. One waits
      on an absence, three are weaker than their promise, and **three —
      `job-wait`, `gate-publish` and `cleanup` — are served by nothing at
      all**, which is consistent with S11.2.b: no subject but a session can be
      removed, so a cleanup helper would have nothing to call. -->
      
    • S11.2.f Capture screenshots, traces, console, requests, test data refs, correlation IDs, state, and exact command per step/failure. Evidence: evidence bundle.

      <!-- S11.2.f: shipped as
      `libs/oshun/workbench-kit/src/evidence-bundle.ts` (`EVIDENCE_ARTIFACTS` +
      `ARTIFACT_ANSWERS`, `CAPTURE_OCCASIONS` + `CAPTURED_ATTEMPTS` +
      `capturedAttemptFor` + `ATTEMPT_DESCRIBES_THE_FAILURE`, `ARTIFACT_SINKS`
      + `SINK_SURVIVES_THE_RUN` + `SINK_IS_UNIQUE_PER_ATTEMPT`,
      `parseCapturePolicy`, `resolveEvidenceReadiness`/
      `resolveAllEvidenceReadiness` over five `EVIDENCE_READINESS` states,
      `EVIDENCE_FAULTS` with 7 kinds, `evidenceFindings`, `evidenceCensus`,
      `OSHUN_RETRY_BUDGETS`, `TARA_CAPTURE_POLICY`, `CENTRAL_CAPTURE_POLICY`,
      `OSHUN_OUT_OF_BAND_ARTIFACTS`), 81 tests in `evidence-bundle.spec.ts`,
      `libs/oshun/workbench-kit/tools/evidence-bundle-scan.mjs` (nx target
      `evidence-bundle-scan`, `pnpm run verify:operations:v1-evidence-bundle`,
      14 negative controls), and the five registrations a kit entry point
      needs.
      
      Eight artifacts, and the temptation is to grade them present or absent.
      Presence is the easy half and it is not the half that goes wrong.
      
      **AN ARTIFACT OF THE RETRY IS NOT AN ARTIFACT OF THE FAILURE.** Both
      exemplar suites configure `trace: 'on-first-retry'` beside
      `screenshot: 'only-on-failure'`, and those two settings capture two
      different executions of the same test — the screenshot during the attempt
      that failed, the trace during the attempt **after** it, with a fresh
      browser and a fresh clock. When the failure is flaky the trace is of a run
      that passed; when it is deterministic it is of a second, identical-looking
      failure that is nonetheless a different one. **Nothing in either bundle
      says which artifact came from which attempt**, so a reader lines up a
      screenshot and a trace and reasons about a run that never happened.
      
      **AND THE SAME SETTING CAPTURES NOTHING WHERE MOST DEBUGGING HAPPENS.**
      Both configs declare `retries` as a ternary on the CI environment
      variable — two when set, zero when not. Outside CI there is no first
      retry, so `on-first-retry` never fires: on a developer's machine the Tara
      suite produces **no trace and no video**, by configuration, and **exactly
      one of the eight artifacts describes the failure** — the screenshot.
      `capturedAttemptFor` derives that from the occasion and the budget rather
      than asserting it, and `on-retry` is the only occasion whose answer the
      budget changes.
      
      **AN IDENTIFIER YOU DID NOT CHOOSE ARRIVES TOO LATE TO NAME THE THING
      THAT FAILED.** The BFF reads `x-correlation-id` and `x-trace-id`, so the
      join exists. Four files under `testing/e2e/tests/api` read one back off a
      **response**, and **not one test in the estate sets one on a request** —
      so a navigation that timed out, a route fulfilled from the test process,
      and every failure before the answer have no id at all. And the four
      readers are all `if (requestId) { expect(typeof requestId).toBe('string'); }`
      — **guarded by the value they assert about**, so a server sending no
      header passes exactly as one sending it does.
      
      **A FIXED PATH IS NOT AN ADDRESS WHEN SEVEN PROJECTS SHARE IT.**
      `base.page.ts` exposes `screenshot(name)` writing
      `test-results/screenshots/${name}.png`, outside `testInfo` so the report
      never sees it, at a path derived from the caller's label alone — while the
      Tara config declares **seven projects** over `fullyParallel: true`, five
      running the same specs on different browsers.
      
      THE MATRIX, in CI: Tara captures **2 of 8** artifacts and `testing/e2e`
      **5 of 8**, and in both suites **exactly one describes the failure**.
      Console, request, test-data-ref, application-state and
      reproduction-command are uncaptured in Tara; in `testing/e2e` the three
      that are captured are all `lost-with-the-run`. Neither suite reads
      `testInfo.status`, so nothing anywhere runs differently when a test fails.
      
      TRAP, caught by the scan on its first run: the check for a request-side
      correlation header matched the header NAME, which found the four readers
      and reported the repair as already done. **A setter appears as an object
      key and a reader as a subscript; the colon is the whole difference.** -->
      
    • S11.2.g Run content-authoring and operations-console exemplar journeys with fixtures disabled. Evidence: two passing live suites.

      <!-- S11.2.g LEFT UNCHECKED ON PURPOSE, with the gate that says why:
      `libs/oshun/workbench-kit/src/journey-readiness.ts` (`RUN_ENVIRONMENTS` +
      `AUTHORITY_ENVIRONMENTS`, `fixtureFreeEnvironments`,
      `fixtureFreeEnvironmentsIfUnconditionalGatesWereFixed`,
      `tenantReachableEnvironments`, `runnableEnvironments`,
      `liveRunContradiction`, `LIVE_PRECONDITIONS` + `PRECONDITION_OWNER` +
      `PRECONDITION_ASKS`, `resolvePrecondition`, `gateVerdict`,
      `resolveExemplarReadiness`/`resolveAllExemplarReadiness`,
      `readinessCensus`, `repairLeverage`), 47 tests in
      `journey-readiness.spec.ts`,
      `libs/oshun/workbench-kit/tools/journey-readiness-scan.mjs` (nx target
      `journey-readiness-scan`,
      `pnpm run verify:operations:v1-journey-readiness`, 13 negative
      controls), and the five registrations a kit entry point needs.
      
      **The two suites cannot be written, and the reason is not a missing
      feature.** Nothing here fabricates them: the deliverable is the gate that
      computes their readiness from the six preceding cells' own registers, and
      it refuses.
      
      **"WITH FIXTURES DISABLED" AND "PROVISIONED THROUGH REAL BOUNDARIES" ARE
      SATISFIABLE IN DISJOINT ENVIRONMENTS.** Both halves are derived:
      
      - `fixtureFreeEnvironments()` reads S11.2.c's gate register and returns
        **nothing**. Of the estate's sixty-nine substitutable subjects,
        **forty-nine serve in production** — two through `jsonRoute`, which
        evaluates no condition at all, and forty-seven behind a dependency's
        silence, which no environment suppresses. "Fixtures disabled" is not a
        state a test cannot enter; it is a state this estate has no environment
        for.
      - `tenantReachableEnvironments()` reads S11.2.b's register and returns
        **non-production only**: creating a tenant needs `admin:*` or
        `admin:workspace:tenant-console`, the only route to either is an
        unsigned `dev.` token, and `areDevTokensAllowed` refuses those wherever
        the runtime is production.
      
      Both exemplar journeys require a tenant, so the intersection is empty —
      **the closer the estate is brought to "fixtures disabled", the further it
      is from "provisioned through real boundaries."** And the counterfactual
      is computed too: repair the unconditional gates and the fixture-free
      environment becomes **production**, which is exactly where the tenant
      provisioner closes. **The second blocker survives the first repair.**
      
      THE GATE, run: **0 of 2 journeys runnable, 0 of 7 preconditions met, and
      all seven cells own a blocker** —
      `an-environment-satisfies-both` (S11.2.g),
      `fixtures-declinable-per-test` (S11.2.c: provider deployment-scoped, data
      undeclinable, intercept run-scoped), `subjects-provisionable` (S11.2.b:
      tenant authority-self-asserted), `no-unlabelled-substitution` (S11.2.a),
      `masquerades-refutable` (S11.2.d: four of five unrefutable),
      `helpers-establish-their-promise` (S11.2.e: job-wait, gate-publish and
      cleanup absent), and `failure-leaves-evidence` (S11.2.f: 1 of 8 artifacts
      outside CI). `repairLeverage` reports every cell unblocking exactly two,
      which is itself the finding: **S11.2.g is not one repair away from
      anything.**
      
      The scan re-reads both halves of the contradiction from source every run,
      and **fails if any suite ever claims to be a live exemplar journey while
      the gate still refuses** — a file claiming live in an estate that cannot
      be live is S11.2.d's masquerade at the scale of a whole suite.
      
      TRAP, caught by the gate on its first run: `resolveProvisioningChain`
      names its answer `verdict`, and reading `.readiness` returned `undefined`
      for every subject — filtering nothing out and listing four blockers where
      there is one. **The right conclusion for the wrong reasons, which is the
      failure mode a composed gate is most prone to.** -->
      
  • S11.3 Upgrade the axe harness and rule policy to WCAG 2.2 AA. Evidence: tests cover the new 2.2 criteria applicable to workbenches.

    <!-- S11.3 a–f ARE ALL CLOSED. THE PARENT IS LEFT UNCHECKED ON PURPOSE,
    because its evidence line is a measurable claim and the measurement is
    1 of 6.
    
    Six kit entry points, 249 tests, 83 negative controls:
    `axe-policy` (a), `wcag22-coverage` (b), `scan-matrix` (c),
    `a11y-findings` (d), `a11y-waivers` (e), `semantic-assertions` (f), each
    with a spec, a scan, an nx target, a `verify:operations:*` script and the
    five registrations.
    
    WHY THE PARENT DOES NOT CLOSE. Of the six criteria WCAG 2.2 adds at A and
    AA, S11.3.b's map reports **1 covered** — 2.5.8 Target Size, the only one a
    scanner over a still document can reach, which is exactly and
    independently why it is the only one axe has a rule for. **2 have the
    technique aimed elsewhere** (2.4.11 and 2.5.7: the estate drives a real
    browser, opens a Popover, compares bounding boxes and walks operations for
    a non-drag path — against a gallery of nineteen primitives). **3 have no
    instrument of their shape at all** (3.2.6, 3.3.7, 3.3.8 are quantified over
    more than one document, and every accessibility instrument in the estate
    takes one page and returns a verdict about it).
    
    AND THE UPGRADE ITSELF IS WORTH ONE RULE. S11.3.a measured the six declared
    tags at 56, 3, **0**, 2, **0** and 1 rules: the 2.1 upgrade bought two
    rules, the 2.2 upgrade bought one, and `wcag21a` and `wcag22a` name nothing
    at all. The one rule 2.2 buys ships with its enabled flag off, so naming
    the tag is what runs it — and the site that asks for everything is the only
    one that misses it.
    
    WHAT WOULD CLOSE IT, IN ORDER. (1) Aim the S7.21.f driven controls at the
    workbench applications: that moves 2.4.11 to covered with **no new
    technique**, and the counterfactual is computed rather than asserted.
    (2) Declare the three registers nobody produces — which elements are help
    mechanisms, which flows are multi-step processes, which steps authenticate,
    and which operations are performed by dragging. Those are product
    decisions and not repository edits, which is why S11.3.b files them as
    `input-has-no-producer` rather than as work. (3) Build the multi-document
    instrument 3.2.6 needs; the estate already visits the pages and compares
    none of them.
    
    THREE LIVE DEFECTS THE FAMILY MEASURED, none of which is about 2.2:
    **1.4.2 Audio Control and 3.3.2 Labels or Instructions are Level A, are
    scanned, are reported, and cannot fail any gate in the estate** — their
    only rules top out at `moderate` and every gate fails on `serious`
    (S11.3.d). **The predicate that runs after 620 tests has never been shown
    to reject anything**: `libs/shared/testing/src/playwright/` holds `axe.ts`,
    `index.ts` and no spec. **The product suites read the accessibility tree
    zero times** while making 29 assertions about the attributes that feed it
    (S11.3.f). -->
    
    • S11.3.a Pin axe/browser versions, WCAG 2.0/2.1/2.2 A/AA tags, rules, experimental/manual flags, impact policy, and allowed exclusions. Evidence: harness configuration review.

      <!-- S11.3.a: shipped as `libs/oshun/workbench-kit/src/axe-policy.ts`
      (`POLICY_DIMENSIONS` + `DIMENSION_ASKS`, `PIN_STRENGTHS` +
      `PIN_STRENGTH_RANK` + `isReproducible`, `OSHUN_ENGINE_BINDINGS`,
      `AxeCatalogue`, `rulesSelectedByTags`, `RuleSelection` +
      `rulesSelectedBy` + `declaredTags`, `criteriaLostToFlags`, `tagYields`,
      `MEASURED_SELECTION`, `OSHUN_AXE_HARNESSES`, `POLICY_FAULTS` +
      `policyFaults`, `gateCoverageFaults`, `dimensionCensus`,
      `policyHeadline`), 55 tests in `axe-policy.spec.ts`,
      `libs/oshun/workbench-kit/tools/axe-policy-scan.mjs` (nx target
      `axe-policy-scan`, `pnpm run verify:operations:v1-axe-policy`, 17
      negative controls), and the five registrations a kit entry point needs.
      
      THE VEIN: **A TAG IS A QUERY, NOT A POLICY.** The estate upgraded to
      WCAG 2.2 by adding two tags to two helpers, which reads like a policy
      change and is a change of search terms against a table shipped by
      somebody else. Run the query against the installed engine and the six
      declared tags are worth **56, 3, ZERO, 2, ZERO and 1** rules —
      fifty-nine of the sixty-two come from the WCAG 2.0 terms alone, the 2.1
      upgrade is worth two rules and the 2.2 upgrade one, and **two of the six
      terms name nothing at all**. `wcag21a` is the sharper half because every
      site declares it: its one rule is `label-content-name-mismatch` for
      **2.5.3 Label in Name, Level A**, and axe ships it `experimental`, which
      axe's own default `tagExclude` drops.
      
      **AND THE ONE RULE 2.2 BUYS SHIPS SWITCHED OFF.** `target-size` carries
      `enabled: false`, and axe consults that flag ONLY when the include list
      is empty, so naming the tag is what turns it on. Which gives the result
      that decides the cell: **the site asking for EVERYTHING is the only one
      that misses the 2.2 rule.** `apps/oshun/tenant-admin/e2e/fixtures.ts`
      calls `.analyze()` with no tags, gets 89 rules including 30
      best-practice ones, and does not run `target-size`, because with an
      empty include list the enabled flag applies again. **Asking for more got
      less.**
      
      `rulesSelectedByTags` reproduces axe's `matchTags` and was checked
      against **five selections measured in a real Chromium — 61, 62, 1, 0, 89
      — and matched 5/5**; the scan re-derives all five every run, so an
      engine bump re-opens the measurement instead of quietly falsifying it.
      It also **refuses an empty tag list because the engine does**
      (`runOnly.values must be a non-empty array`, thrown inside the page) —
      the first version of the suite asserted the opposite, which read as
      obvious and was false.
      
      SECOND: **THE GATE AND THE HELPER ARE NOT THE SAME HARNESS.** 539 of 560
      web specs and 81 of 82 admin specs end in the automatic `afterEach` from
      `libs/shared/testing/src/playwright/axe.ts`, which declares the 2.1 set;
      the 2.2 set is reached by **54 explicit calls in 22 files**. "The
      harness is WCAG 2.2 AA" is a true sentence about a helper and a false
      one about the gate.
      
      THIRD: **THE DIMENSION NOBODY STATES HAS A PRICE PAYABLE IN CRITERIA.**
      One of eight sites states a flag policy and it is the ledger rather than
      a gate; what the other seven inherit removes the ONLY rule the engine
      has for three live A/AA criteria — **1.2.1** (deprecated
      `audio-caption`), **1.3.4** (experimental `css-orientation-lock`) and
      **2.5.3** — and all three then come back `inapplicable`, the same answer
      a scan gives for a criterion nothing on the page could have violated.
      
      FOURTH: **AN EXCLUSION WITH NO TERMS IS THE MOST USED CONTROL HERE** —
      `test.use({ skipAxe: true })` **247 times across 241 files**, a boolean
      with no owner, reason or expiry; beside it `TEMPORARILY_DISABLED_RULES`
      (`color-contrast` 1.4.3 AA, `scrollable-region-focusable` 2.1.1 A) with
      the word "temporarily" and no date.
      
      FIFTH, which the word "pin" is about: **A PIN IS A CLAIM ABOUT A RUN AND
      NO RUN HERE RECORDS WHAT IT USED.** Every result carries
      `testEngine.version` and a `testEnvironment` and all eight sites discard
      both. **Three ranges decide the engine** — catalogue `^4.11.4`,
      `@axe-core/playwright@4.11.3 → ~4.11.4`, `axe-playwright@2.2.2 →
      ^4.10.1` — and two live in packages this repository does not edit. The
      browser is worse: `devices['Desktop Chrome']` sets a user agent and a
      viewport, not a build, and this machine has two Chromium builds
      installed. **0 of 7 dimensions are exact at every site.**
      
      DECISIONS. SELECTION IS THREE MODES AND NOT A NULLABLE TAG LIST, because
      axe has `tag`, `rule` and absent and they are not variations on one
      another — modelling `withRules` as no-tags reported the estate's
      one-rule contrast probe as running 89. THE LOST-CRITERIA LIST TAKES THE
      SELECTIONS IN FORCE AS AN ARGUMENT, so 2.5.8 is not counted as lost when
      the two helpers recover it by naming its tag, and scope is S7.21.f's A/AA
      list so four AAA criteria do not inflate it. AND SILENCE ABOUT THE
      BROWSER IS NOT A FAULT AGAINST A HARNESS, because no harness in any
      estate states a browser build; it is answered by `run-not-recorded`,
      which is the repair that exists.
      
      TRAPS, both caught by the scan on its own register: the admin helper's
      reach read "2 files" and is 3, and the story lane is configured in TWO
      files — tags in `.storybook/preview.ts`, invocation and report policy in
      `.storybook/test-runner.ts` — so a row naming one leaves the other
      unowned and `sources` is plural. -->
      
    • S11.3.b Map WCAG 2.2 focus-not-obscured, dragging movements, target size, consistent help, redundant entry, and accessible authentication to automated/manual/journey checks. Evidence: criterion coverage map.

      <!-- S11.3.b: shipped as
      `libs/oshun/workbench-kit/src/wcag22-coverage.ts` (`NEW_AT_2_2` +
      `newCriteria`, `EvidenceShape` + `CRITERION_SHAPES`, `CHECK_METHODS` +
      `methodsFor`, `CRITERION_INPUTS`, `SUBJECT_LAYERS` + `PRODUCT_LAYER`,
      `NON_AXE_INSTRUMENTS` + `instrumentsFromHarnesses` + `canExpress`,
      `COVERAGE_VERDICTS` + `coverageOf` + `coverageMap`, `COVERAGE_FAULTS` +
      `coverageFaults`, `automatedReach`, `coverageCensus`,
      `coverageHeadline`), 41 tests in `wcag22-coverage.spec.ts`,
      `libs/oshun/workbench-kit/tools/wcag22-coverage-scan.mjs` (nx target
      `wcag22-coverage-scan`, `pnpm run verify:operations:v1-wcag22-coverage`,
      14 negative controls), and the five registrations.
      
      **THE METHOD FOLLOWS FROM THE SHAPE OF THE EVIDENCE, NOT FROM HOW HARD
      THE CRITERION FEELS.** Two questions settle it: how much has to be in
      evidence at once — one document in one state, one across several, or
      more than one document — which separates automated from journey; and
      whether the instrument can produce a definite PASS or only a definite
      FAIL with a residue, which is what `manual` is actually for. Nothing in
      the file assigns a method by hand. The six split **1 automated / 5
      journey, 4 also manual**, and the automated one is **2.5.8** — which is
      the one axe has a rule for, arrived at by a second derivation that never
      consults axe. The scan fails if the two ever disagree.
      
      THE VEIN: **THE ESTATE ALREADY OWNS THE TECHNIQUE FOR THE HARDEST THREE
      AND POINTS IT AT THE LAYER WITH NO PRODUCTS IN IT.** S7.21.f drives a
      real browser, opens a Popover and compares bounding boxes to answer
      2.4.11 (`focus:not-obscured-by-an-overlay`); walks every operation for a
      non-drag path to answer 2.5.7; measures a target floor for 2.5.8. All
      three are real, driven, and run against a bundled gallery of nineteen
      primitives at `?subject=…`. The workbench applications — where the
      drawer that covers the focused field actually lives — get a static
      scanner whose rule set contains one of the three. **The instrument is
      not missing; it is aimed one layer below the products**, and the
      counterfactual is computed: aim it at the applications and 2.4.11
      becomes `covered` with no new technique, while **2.5.7 does not move**,
      because its input has no producer — a second obstacle and a different
      job.
      
      SECOND: **ITERATING IS NOT RELATING.** Six web routes and seventeen
      admin workspaces are visited under an accessibility test, and both are
      `for (const … ) { test(…) }` — N verdicts about N pages. 3.2.6 is a
      claim about a **pair**, and no instrument here produces a verdict whose
      subject is more than one document. The pages are already being visited;
      nothing compares two of them, and no rule set ever will.
      
      THIRD: **A DELEGATION NOBODY CATCHES IS AN ABOLITION.** S7.21.f
      dispositions 3.2.6 as delegated-to-the-host ("a property of the site,
      not of any one page") and 3.3.7/3.3.8 as not-exercised ("no notion of a
      step", "no authentication step"). **This is the layer they were handed
      to**, and across the whole V1 estate those three criterion numbers
      appear in exactly one module and its spec — the one that handed them up.
      S7.21.f wrote that sentence about props; it arrives here one layer later
      about criteria.
      
      FOURTH: **THE INPUT IS PART OF THE MAP.** Four of six need the estate to
      declare something first — which elements are help mechanisms, which
      flows are multi-step, which steps authenticate, which operations are
      performed by dragging — and **none of the four has a producer**. S7.20's
      help registry looks exactly like the producer for the first and is not:
      it keys help CONTENT by capability, route, operation, field, block,
      gate, error, state and contract, and 3.2.6 is about placement.
      
      AND THE ONE THAT IS COVERED IS COVERED ONLY FOR WHAT THE ENGINE COULD
      DECIDE: 2.5.8's inline/essential/user-agent exceptions leave an
      `incomplete` residue, and S11.3.a measured that seven of eight sites
      read no incomplete results at all — so `residueUnread` is a column
      beside the verdict rather than folded into it.
      
      DECISIONS. THE VERDICT REPORTS THE FIRST OBSTACLE RATHER THAN A SCORE,
      because "no instrument exists", "one exists and is aimed elsewhere" and
      "one is aimed right with nothing to hand it" are three different jobs.
      MANUAL IS ADDED ON TOP OF THE MACHINE AND NEVER INSTEAD.
      
      TRAP, caught by the scan's own controls: **a criterion number is a
      substring of a version number** — searching for `3.2.6` finds
      `Phase 69.3.2.6`, and `\b` does not help because the preceding character
      is a dot. **And the fix is not symmetric**: a dot to the right is usually
      a full stop, so guarding both sides alike rejected `see 2.5.8.`, the
      commonest way a reference is written. Two controls, one per side. -->
      
    • S11.3.c Scan route/state/theme/density/viewport/locale matrix after stable load and each relevant overlay/conflict/error/destructive state. Evidence: scan runner tests.

      <!-- S11.3.c: shipped as `libs/oshun/workbench-kit/src/scan-matrix.ts`
      (`SCAN_AXES`, `SCAN_STATES`, `AxisSupply` + `OSHUN_AXIS_SUPPLY`,
      `STATE_RELEVANCE_PRODUCER`, `CONFOUNDS`, `INTERACTING_PAIRS`,
      `planCells` + `fullFactorialSize` + `pairCovered`, `ScanRecord` +
      `coverageOf`, `SuiteRows`, `MATRIX_FAULTS` + `matrixFaults`,
      `matrixHeadline`), 39 tests in `scan-matrix.spec.ts`,
      `libs/oshun/workbench-kit/tools/scan-matrix-scan.mjs` (nx target
      `scan-matrix-scan`, `pnpm run verify:operations:v1-scan-matrix`, 18
      negative controls), and the five registrations.
      
      THE VEIN: **EVERY AXIS IS SETTABLE FROM A TEST AND FIVE OF THE SIX ARE
      NEVER SET.** Counted across every accessibility-named spec in web and
      admin: **zero** set a theme, **zero** set a locale, **zero** set a
      density, **zero** carry the `@mobile` tag the Playwright config uses to
      reach the second viewport — and that config's own comment says each spec
      executes on exactly one project. Route is varied and the rest sit at
      whatever the default is, so the scan set is a **line**. Nothing is out of
      reach: theme and locale are `localStorage`, density is an attribute,
      viewport is a runner project. **No missing capability, no blocked
      environment — the matrix is a matrix nobody enumerated.**
      
      WHAT THAT COSTS IN THE ESTATE'S OWN DECLARED VALUES: the design tokens
      ship **four themes** (`light`, `dark`, `highContrastLight`,
      `highContrastDark`) — two built for the users most likely to need a
      contrast check — and no accessibility scan runs under any but the
      default. The launch locales are **eight** and **two are right-to-left**;
      `applyLocaleToDocument` sets `documentElement.dir` from the locale, so an
      RTL rendering is one init script away, and **no accessibility scan has
      ever run right-to-left**. Against the estate: **16 routes, 3264 cells at
      full factorial, 57 in the covering plan, 0 located by evidence.**
      
      SECOND: **A CELL IS COVERED BY A SCAN WHOSE COORDINATES ARE KNOWN, AND NO
      SCAN HERE RECORDS ITS OWN.** The sharpest line of it — the web signoff
      suite scans `/` **twice**, once at load and once with the notifications
      drawer clicked open and `aria-modal` asserted, two genuinely different
      states, and **the two records are byte-identical**, because a route
      (recoverable only from the test title) is all either carries.
      
      THIRD: **A CONSTANT YOU DID NOT CHOOSE IS A CONFOUND.**
      `apps/oshun/web/e2e/fixtures.ts` emulates `reducedMotion: 'reduce'`
      suite-wide for a good and documented reason — entrance fades sampled
      mid-transition give false contrast failures — and the correct repair
      **pinned an axis** at the value most users are not in, across 539 specs.
      The admin fixtures emulate nothing, so the two suites scan **opposite
      values of one condition** and neither declares it.
      
      FOURTH: **ROWS ARE NOT CELLS.** The admin key-workspaces table is
      **17 rows over 10 distinct paths**, and `/analytics` is the subject of
      **seven** — rows differing only in which test id they wait for before
      scanning. `runAdminAccessibilityScan` is called with no `include`, so it
      reads the whole document every time and the seven are one scan behind
      seven different waits. `scopedToTheRow` is the field that decides it, and
      it is why the same shape is **not** a fault in the web suite, which does
      scope (`include: ['[data-shell-root]']`).
      
      THE REDUCTION, which is the part a runner lives or dies on: the full
      product is thousands of scans, so `planCells` covers every axis value
      once and every **interacting pair** once, and each pair carries the
      **mechanism** by which its two conditions can disagree. A pair with no
      mechanism is not in the plan — that is the difference between a covering
      array and a bill somebody quietly stops paying — and `theme x locale` is
      left out with the reason written down. `route x state` is outside the
      reduction, because a state that exists and is never entered is the gap
      the cell is about. **Which states a route has is an input with no
      producer**, so "this route has no error state" and "nobody looked at its
      error state" are currently the same reading.
      
      TRAPS: a check for a recorded coordinate matched two attachments in
      `search-scan-density-accessibility.spec.ts` whose "density" is the
      density of **search results**, not the UI axis — **the name argument is
      not a coordinate**, so the attach call's name is stripped and only the
      payload is searched. And a control asserting the state pairs were
      uncovered went stale the moment the register recorded the one overlay the
      estate really does scan; it now asserts the opposite, with a second
      control showing they close only because ONE route declares ONE state —
      which separates a weak plan from a missing input. -->
      
    • S11.3.d Fail on new unwaived AA violations and require rule/selector/ state/screenshot/DOM snippet/source owner in findings. Evidence: deliberate regression.

      <!-- S11.3.d: shipped as `libs/oshun/workbench-kit/src/a11y-findings.ts`
      (`FINDING_FIELDS` + `FIELD_PURPOSE` + `FIELD_PRESENCE` +
      `OSHUN_FINDING_FIELDS`, `GatePredicate` + `OSHUN_GATES` + `gateFails` +
      `gateThreshold`, `severityTable`, `gateAdmits`,
      `unenforceableCriteria`, `OSHUN_DELIBERATE_REGRESSIONS` +
      `STRICT_GATE` + `runRegression`, `FINDING_FAULTS` + `findingFaults`,
      `findingsHeadline`), 44 tests in `a11y-findings.spec.ts`,
      `libs/oshun/workbench-kit/tools/a11y-findings-scan.mjs` (nx target
      `a11y-findings-scan`, `pnpm run verify:operations:v1-a11y-findings`, 12
      negative controls), and the five registrations.
      
      THE VEIN: **AN IMPACT IS THE CHECKER'S SEVERITY JUDGEMENT AND A LEVEL IS
      THE STANDARD'S, AND EVERY GATE HERE IS KEYED ON THE FIRST AND NAMED FOR
      THE SECOND.** Every gate fails on `serious`+`critical` and every one is
      called a WCAG 2.1/2.2 AA gate. Against axe's own rule→check→impact
      chain, **5 of the 62 rules the conformance selection runs can never
      produce a violation above `moderate`** — `aria-deprecated-role`,
      `form-field-multiple-labels`, `html-xml-lang-mismatch`,
      `no-autoplay-audio`, `server-side-image-map` — so a real, detected,
      printed violation of any of them **passes every gate in the estate.**
      
      AND FOR TWO LEVEL A CRITERIA THAT IS THE ONLY RULE THERE IS: **3.3.2
      Labels or Instructions** and **1.4.2 Audio Control**. A page that
      autoplays audio with no way to stop it is scanned, found, printed and
      passes — which S11.3.b's coverage map reports as `covered`, because
      **coverage and enforceability are different questions.** The severity
      table is re-derived from the installed engine every run and the scan
      fails if the answer moves.
      
      SECOND: **"NEW" NEEDS A REFERENT AND AN ABSOLUTE GATE IS NOT ONE — BUT
      IT IS NOT NOTHING EITHER.** Nothing baselines an accessibility result.
      Inside the fail set that is *stricter* than a baseline: zero is the
      strictest referent there is. **Below** the threshold there is neither a
      gate nor a baseline, so a new AA violation at `moderate` is admitted once
      and admitted forever.
      
      THIRD, the proof the cell asks for: **A GATE THAT HAS NEVER FAILED IS A
      GATE NOBODY HAS TESTED.** `libs/shared/testing/src/playwright/` contains
      `axe.ts` and `index.ts` and **no spec** — the predicate that runs after
      620 tests has never been shown to reject anything. So the six
      **deliberate regressions** here are executable against a reproduction of
      that predicate, `?? 'moderate'` and all; two controls prove it can
      reject, and **three document a defect by passing a Level A violation on
      purpose**. The scan fails if any stops behaving as recorded — including
      if somebody repairs the gate, which is the outcome that should force this
      register rewritten.
      
      FOURTH: **A FIELD THE ENGINE ALREADY RETURNS IS NOT A MISSING FIELD, IT
      IS A DISCARDED ONE.** axe returns `node.html` on every violating node and
      both formatters drop it; it survives only where seven specs attach the
      whole result as JSON through a helper each defines separately. **2 of the
      6 fields reach the reader**, 2 are in artifacts, and `state` and
      `source-owner` are nowhere — the second because S11.1.a measured the
      ownership file as a single catch-all, so a lookup would answer, and
      answer the same thing every time.
      
      DECISIONS. A REGRESSION RECORDS WHAT THE GATE DOES AND NOT WHAT IT
      SHOULD, which is the only way one can document a defect rather than
      assert it away. THE UNGRADED VIOLATION IS POSED TO BOTH THE ESTATE'S GATE
      AND A STRICT ONE, so the estate's answer reads as a choice. AND A
      CRITERION IS UNENFORCEABLE ONLY WHEN EVERY ONE OF ITS RULES IS ADMITTED —
      4.1.2 has an admitted rule and a critical one, so it is not on the list.
      
      TRAP: the scan first built the severity table from the estate's *widest*
      selection, which is the untagged 89-rule default, and reported 24
      admitted rules — most of them best-practice with no criterion at all.
      **"The landmark helper cannot fail the build" is not the same finding as
      "WCAG 1.4.2 cannot fail the build"**, and only the second is a
      conformance claim, so the table is built from the widest *tag-mode*
      selection. -->
      
    • S11.3.e Define waiver schema with false-positive proof or temporary exception, scope, owner, approver, expiry, remediation, and retest. Evidence: expiry tests.

      <!-- S11.3.e: shipped as `libs/oshun/workbench-kit/src/a11y-waivers.ts`
      (`APPLICATION_POINTS` + `waivable`, `WAIVER_FIELDS` + `FIELD_ASKS`,
      `EXCLUSION_KINDS` + `KIND_CLAIMS` + `KIND_OBLIGATIONS` +
      `NEVER_EXPIRES`, `FindingScope` + `A11yWaiver`, `standingAt` +
      `authorisesOn`, `GATE_WAIVER_FIT`, `OSHUN_EXCLUSIONS`, `WAIVER_FAULTS` +
      `waiverFaults`, `waiverCensus`, `waiverHeadline`), 36 tests in
      `a11y-waivers.spec.ts`,
      `libs/oshun/workbench-kit/tools/a11y-waivers-scan.mjs` (nx target
      `a11y-waivers-scan`, `pnpm run verify:operations:v1-a11y-waivers`, 11
      negative controls), and the five registrations.
      
      THE VEIN: **A WAIVER IS A DECISION ABOUT A FINDING, AND EVERY EXCLUSION
      IN THIS ESTATE IS APPLIED BEFORE THERE IS ONE.** `skipAxe` prevents the
      scan; `disableRules` stops the rule running; the story parameter skips
      the story. Across **247 opt-outs in 241 files** there is nothing to
      attach a waiver TO, because nothing was computed — and a proof, an owner
      and a date are all statements *about a finding nobody has*. **The repair
      is not a longer `skipAxe`**: it is to move the exclusion to the other
      side of the scan — run it always, then suppress a NAMED finding on terms
      — with the consequence nobody will enjoy: the day those 247 become
      suppressions, the estate learns what they were hiding. `waivable` is one
      line and **0 of the estate's 4 mechanisms pass it.**
      
      SECOND: **THE SCHEMA ALREADY EXISTS IN THIS LIBRARY AND IT DOES NOT FIT,
      FOR THE REASON THAT MAKES IT GOOD.** S2.11's `gate-waivers.ts` has
      request-vs-grant, scope, digested evidence, risk, a validity window, a
      review date, revocation and renewal — and its **first** requirement is a
      `runId`, under a comment saying *"a waiver against a gate rather than a
      result is an exemption, and it will outlive every reason it was granted
      for."* That is exactly what an accessibility exclusion is. It also wants
      an `inputDigest` (the input is a rendered page nobody hashes) and a
      tenant-shaped scope (a finding is rule × selector × route × state).
      **5 of its 9 fields fit and 4 do not**, named one by one, so "reuse
      S2.11" is a plan with a list rather than a gesture.
      
      THIRD, and why there are **three** kinds and not two: **A CLAIM ABOUT THE
      INSTRUMENT IS NOT A WAIVER.** The estate's one well-formed exclusion,
      `JSDOM_UNEVALUABLE_RULES`, says jsdom resolves no computed colours and
      names the Playwright lane as what settles contrast instead. That must
      never acquire an expiry — **a deadline on a fact about a runtime is a
      date that arrives and changes nothing** — and, equally, classing a real
      debt as an instrument limit is how one stops expiring.
      
      FOURTH: **AN EXPIRY IS A DATE AND A WORD IS NOT ONE.**
      `TEMPORARILY_DISABLED_RULES` turns off `color-contrast` (1.4.3 AA) and
      `scrollable-region-focusable` (2.1.1 A), and "temporarily" is the entire
      temporal content.
      
      DECISIONS. THE CLOCK IS AN ARGUMENT, so the cell's named evidence —
      expiry tests — is a test and not a wait; the boundary is inclusive on the
      day, because a date is a day and not an instant. AN UNREADABLE DATE
      RETURNS `unbound` AND NOT `active`. THE STATUS VOCABULARY IS S2.11's,
      imported rather than restated. AND THE MISSING-EXPIRY FAULT IS DRIVEN BY
      THE OBLIGATIONS, not by a separate rule — the first version asked every
      kind that is not an instrument limit for an expiry, which flagged a
      false-positive claim for a date it does not owe.
      
      TRAP, caught by this scan against its own sibling: **A MENTION IN PROSE
      IS NOT A USE.** Counting lines gave **248** `skipAxe` opt-outs, and one
      was a sentence in `studio-route-render-smoke.spec.ts`'s doc comment
      explaining that the file uses one. S11.3.a's register said 248 and now
      says **247**; both cells now filter comment lines and report the same
      number. -->
      
    • S11.3.f Add custom semantic assertions where axe cannot prove names/ states/relationships, announcements, complex widget patterns, or visual obstruction. Evidence: custom test index.

      <!-- S11.3.f: shipped as
      `libs/oshun/workbench-kit/src/semantic-assertions.ts`
      (`ASSERTION_SUBJECTS` + `SUBJECT_ASKS`, `WHY_AXE_CANNOT` +
      `WHY_READS_AS` + `WHY_EXPLAINED`, `OBSERVATION_CHANNELS` +
      `SETTLEMENTS` + `CHANNEL_SETTLEMENT` + `settles`, `CustomAssertion` +
      `OSHUN_CUSTOM_ASSERTIONS`, `ASSERTION_FAULTS` + `assertionFaults`,
      `assertionCensus`, `assertionHeadline`), 34 tests in
      `semantic-assertions.spec.ts`,
      `libs/oshun/workbench-kit/tools/semantic-assertions-scan.mjs` (nx target
      `semantic-assertions-scan`,
      `pnpm run verify:operations:v1-semantic-assertions`, 11 negative
      controls), and the five registrations.
      
      THE VEIN: **A CUSTOM SEMANTIC ASSERTION THAT READS THE DOM IS CHECKING
      THE INPUT, AND EVERY ONE OF THE ESTATE'S PRODUCT-LEVEL ASSERTIONS DOES.**
      Counted across every accessibility-named spec in web and admin: **29**
      `toHaveAttribute('aria-…')` assertions, **50** role queries, and **zero**
      reads of the accessibility tree. An `aria-labelledby` is what the author
      wrote; the accessible name is what the user agent computed from it, and
      one pointing at an element not in the document produces a widget with
      **no name at all** while the attribute assertion passes.
      
      ONE LAYER DOWN THE ESTATE ALREADY OWNS THE INSTRUMENT: S7.21.f opens a
      CDP session, calls `Accessibility.getFullAXTree`, and carries a fault
      named **`property-written-in-the-dom-that-the-tree-does-not-carry`** —
      precisely the case the twenty-nine cannot see, *by construction*,
      because they read the side of the seam the fault is about. It contracts
      **13** widget patterns, observes announcements through a live-region
      probe, and compares bounding boxes for obstruction — all against a
      bundled gallery. Same shape as S11.3.b; the contribution here is
      sharper: **a named fault the estate can detect, has an instrument for,
      and cannot detect at the layer the fault is about.**
      
      SECOND: **A CUSTOM ASSERTION IS NEEDED MOST WHERE AXE REPORTS SUCCESS.**
      Three reasons the checker cannot prove something, and only one is
      visible. `no-rule` is honest. `could-not-decide` is axe's third outcome,
      and **37 of the 62 running rules have a check that can return it** — so
      part of this very index is published once per run and S11.3.a measured
      that **7 of 8 sites discard it**. `inapplicable-by-construction` is the
      worst: the rule exists and the attribute it would have judged was never
      rendered, so it reports the same word it uses for a page that could not
      have violated the criterion. **Two of the three read as green, and those
      are the two a custom assertion is FOR.**
      
      DECISIONS. THE SETTLEMENT TABLE IS THE ARGUMENT — a claim per
      (channel, subject) with a reason, because unexplained verdicts grade by
      assertion rather than by channel. **A CAVEAT IS NOT A PASS**: the 50 role
      queries compute a name in the test process rather than the browser and
      reach no relationship, so they are their own row and not coverage; and
      **no channel strictly settles a widget pattern**, because the protocol
      carries neither `aria-posinset` nor `aria-setsize` — a limit of the
      instrument, not of the estate. AND AN ENTRY'S SUBJECT IS NULLABLE,
      because of one row: **the estate's only product-level geometry calls use
      the channel that settles obstruction and measure a TOUCH TARGET**, so
      filing them under `visual-obstruction` because the channel matches would
      have reported the subject as covered by an assertion that never asks it.
      
      AGAINST THE ESTATE: **0 of 4 subjects settled at the product layer**, one
      reached only with a caveat, and three techniques that settle theirs
      aimed at the gallery. A control confirms that aiming the tree instrument
      at the products would settle one, with no new technique. -->
      
  • S11.4 Add keyboard-only journey helpers. Evidence: no pointer calls in the critical keyboard suite.

    <!-- S11.4 LEFT UNCHECKED ON PURPOSE. Six children shipped; a–d and f are
    `[x]`, e is `[ ]` with its own gate. The parent is unchecked because its
    evidence line **has no subject and would not be sufficient if it had
    one**, and both halves of that were measured by its own children.
    
    **NO SUBJECT.** S11.4.d built the guard the line describes and then went
    looking for what to run it over: **nothing in this repository declares a
    keyboard suite** — no Playwright tag, no project, no grep-able marker.
    `keyboard-only` appears in three test titles, one journey-registry entry,
    some typed text and some prose, and **a title is not a selector**. So
    "the critical keyboard suite" names nothing that can be enumerated, and a
    claim about the pointer calls in it is a claim about an empty set.
    
    **NOT SUFFICIENT IF IT HAD ONE.** S11.4.e found the nearest thing to that
    scope — the estate's critical-journey register, eight distinct tests named
    by file and title — and ran S11.4.d's guard over it: **four of the eight
    do not hold.** One opens the palette with a chord and then **clicks five
    palette items**; `home-keyboard-focus-path` **presses no key at all**,
    asserting `tabindex` is `-1` and then `0`, which reads the attribute that
    decides the tab order instead of walking it; and two set focus by method
    call. The one test titled `keyboard-only` **passes the parent's evidence
    line exactly** — zero pointer calls, zero forced clicks, zero test ids —
    and calls `.focus()` once, which reaches anything at `tabindex="-1"`, in a
    closed subtree, or skipped by the tab order. **A test that satisfies this
    line can still have assumed the thing it exists to prove.**
    
    **AND THE HELPERS THE CELL ASKS FOR CANNOT ALL BE WRITTEN**, for reasons
    that are product gaps rather than test gaps: S11.4.c found **no keyboard
    path at all** for reorder, move or range selection, with five drag
    surfaces carrying no keyboard affordance anywhere in their file; S11.4.b
    found the one key family that matters for eight launch locales —
    composition — reachable by **neither** of the runner's text APIs, only by
    a CDP `Input.imeSetComposition`, which appears nowhere here; and S11.4.a
    found **one of seven** focus claims established. The gate in S11.4.e
    computes each of those from the owning cell's own register, so **it moves
    when a blocker clears** rather than when somebody edits it: today it
    reports **0 of 9 critical acts runnable keyboard-only**.
    
    What would close this parent is a declared suite, a keyboard path for the
    three interactions that have none, and a composition mechanism — none of
    which is a test somebody forgot to write. -->
    
    • S11.4.a Provide focus-current/order/visible/not-obscured/within-region/ restored assertions and focus history evidence. Evidence: helper tests.

      <!-- S11.4.a: shipped as
      `libs/oshun/workbench-kit/src/focus-assertions.ts` (`FOCUS_CLAIMS` +
      `CLAIM_ASKS`, `EVIDENCE_NEEDS` + `CLAIM_SHAPES`,
      `OBSERVATION_PATTERNS` + `PATTERN_DESCRIBED` + `establishes`,
      `CHANNEL_FOR` + `obstructionChannelSettles`, `OSHUN_FOCUS_EVIDENCE`,
      `FOCUS_ORDER_RESOLVER`, `FOCUS_FAULTS` + `focusFaults`, `focusCensus`,
      `focusHeadline`), 31 tests in `focus-assertions.spec.ts`,
      `libs/oshun/workbench-kit/tools/focus-assertions-scan.mjs` (nx target
      `focus-assertions-scan`,
      `pnpm run verify:operations:v1-focus-assertions`, 12 negative controls),
      and the five registrations.
      
      THE SEVEN SPLIT THREE WAYS ONCE YOU ASK WHAT EACH NEEDS OBSERVED: three
      are about **one instant** (current, visible, not-obscured), one relates
      **two** (restored), and three need **every step** (order, within-region,
      history).
      
      THE VEIN: **A SEARCH FOR AN ELEMENT IS NOT AN ASSERTION ABOUT AN ORDER.**
      The estate's strongest keyboard-reachability check presses Tab **up to
      eighty times** until the assistant composer reports itself focused, then
      asserts that it was reached. That establishes reachability — **2.1.1
      Keyboard** — and says nothing about **2.4.3 Focus Order**, which is a
      claim about the sequence. And the eighty is read out of the loop by the
      scan rather than remembered, because the point is that **nothing on the
      page produced it**: a surface with a longer focus order fails, and the
      message says the element was not reached rather than that the budget ran
      out.
      
      SECOND: **TAB IS A MOVE AND AN ASSERTION IS A CLAIM — 14 of the estate's
      19 Tab presses are moves**, followed by `type(…)` rather than
      `expect(…)`. If focus lands elsewhere the text arrives in the wrong field
      and the failure names neither focus nor order.
      
      THIRD, and it stings: **THE CORRECT RESOLVER EXISTS IN THIS REPOSITORY
      AND IS SERVED TO THE PRODUCT.**
      `apps/oshun/bff/src/studio/focus-order-store.ts` computes the real Tab
      traversal — positive `tabIndex` first ordered by index then document
      position, then `tabIndex` 0 in document order, negatives
      programmatic-only, disabled and hidden skipped, the positive-index
      anti-pattern flagged. Deterministic, tested, and exposed at
      **`/v1/admin/studio/focus-order`** for a designer to inspect a component
      with. `resolveFocusOrder` has **exactly two importers** — the route that
      serves it and the store that defines it — and **no suite uses it to check
      the product's own focus order.**
      
      FOURTH: **FOCUS HISTORY IS WHAT MAKES A FAILURE READABLE.** `toBeFocused`
      fails with the state at the end; an order fails because of the path. The
      estate records none.
      
      AGAINST THE ESTATE: 65 `toBeFocused()`, 19 Tab presses (5 asserted), 20
      `document.activeElement` reads, 3 `aria-modal` assertions, 2 hand-written
      returns, and **zero** checks of the focus indicator. **4 of 7 claims
      observed, ONE established, 0 behind a shared helper.**
      
      DECISIONS. The two claims that need a painted page are refused to every
      keyboard pattern and reported under **their own fault kind**, because
      "walk it better" and "read the painted page at all" are different
      repairs. A SINGLE ASSERTION CANNOT ESTABLISH CONTAINMENT — focus inside a
      region at two moments is consistent with focus having left and come back.
      AND A RETURN NEEDS A BEFORE: an assertion made only afterwards is
      asserting a constant somebody typed, which is what both of the estate's
      focus-restored assertions do.
      
      TRAP: the register said 4 `aria-modal` assertions and 21
      `activeElement` reads; the scan's prose filter counted 3 and 20 — one of
      each was a comment. Same class as S11.3.e's 248→247. -->
      
    • S11.4.b Provide Tab/Shift-Tab, arrows, Home/End/Page, Enter/Space, Escape, type-ahead, application shortcuts, and composition-safe helpers. Evidence: key helper tests.

      <!-- S11.4.b: shipped as `libs/oshun/workbench-kit/src/key-helpers.ts`
      (`KEY_FAMILIES` + `FAMILY_ASKS` + `FAMILY_REQUIRES`,
      `INPUT_MECHANISMS` + `MEASURED_MECHANISMS` + `MEASURED_ON`,
      `canExercise` + `mechanismsFor`, `OSHUN_KEY_EVIDENCE`,
      `OSHUN_COMPOSITION_EXPOSURE`, `OSHUN_HELPER_BUDGETS`, `KEY_FAULTS` +
      `keyFaults`, `keyCensus`, `keyHeadline`), 27 tests in
      `key-helpers.spec.ts`,
      `libs/oshun/workbench-kit/tools/key-helpers-scan.mjs` (nx target
      `key-helpers-scan`, `pnpm run verify:operations:v1-key-helpers`, 12
      negative controls), and the five registrations.
      
      Seven of the eight families are `keyboard.press` with a different
      argument. **The eighth is not an argument at all.**
      
      THE VEIN: **A HELPER FAMILY IS DEFINED BY THE INPUT IT CAN PRODUCE, AND
      NEITHER OF THE RUNNER'S TEXT APIS CAN PRODUCE THE ONE THAT MATTERS.**
      Measured in a real Chromium on an input with every listener attached:
      
      - `keyboard.type('a')` → `press('Enter')`: `isComposing` **false**
        throughout, no composition event at all.
      - `keyboard.insertText('あ')` → `press('Enter')`: an `input` with the
        text and **no keydown for it whatsoever**, then `isComposing` false.
      - CDP `Input.imeSetComposition` → `press('Enter')`: `compositionstart`,
        `compositionupdate`, `input` with `isComposing: true`, and then
        **`keydown Enter` carrying `isComposing: true`.**
      
      Only the third produces the input a composition-safe helper exists to
      test. **`imeSetComposition` and `CompositionEvent` appear nowhere in this
      repository** — and the capability is not missing: the estate already
      opens CDP sessions in four places.
      
      WHAT THAT COSTS, COUNTED: **29 text-entry elements** (`<input>`,
      `<textarea>`) in web and admin carry an inline `onKeyDown` acting on
      Enter — among them the shell's universal search, the assistant composer,
      and Arete's coach, goals and journal inputs — and **`isComposing`,
      `keyCode === 229` and `compositionstart` appear ZERO times in product
      code.** For a user typing Japanese — **one of the eight launch locales** —
      pressing Enter to *confirm an IME candidate* fires every one of them: the
      search submits, the message sends, the half-composed tag is added. **The
      count is the EXPOSURE and not the defect count**, and the difference is
      the point: without the mechanism, a safe handler and an unsafe one are
      the same observation.
      
      SECOND: **THE ESTATE'S ONLY KEY HELPER IS A SEARCH, AND THERE ARE TWO OF
      THEM WITH TWO BUDGETS** — a local `tabTo({ maxTabs = 24 })` inside one
      spec, and an inline loop pressing Tab up to **80** times in another. Both
      are read out of their loops by the scan; **neither number came from a
      page.** 24 against 80 is not a disagreement about two surfaces — it is
      two people guessing.
      
      THIRD: **A TEXT API THAT SKIPS THE KEY EVENT CANNOT EXERCISE A KEY
      HANDLER**, which makes `insertText` exactly right for setting a field up
      and exactly wrong for testing anything listening for a key.
      
      AGAINST THE ESTATE: **6 of 8 families exercised** (tab 19, arrows 22,
      paging 6 with no Page key pressed anywhere, activate 148, dismiss 32,
      shortcuts 19), **1 behind any helper at all**, and that one lives inside
      a spec file. Type-ahead and composition are unexercised.
      
      DECISIONS. THE NARROWING IS BY OPENING TAG, not by file — a file
      containing an Enter handler and an input *somewhere* gives 55, and the
      first of those is a `<li role="button">` with no text entry and so no
      composition to be unsafe about. AND TYPE-AHEAD'S USE COUNT IS `null`
      RATHER THAN 0, because a call site doing it by hand is not separable from
      ordinary typing by any search — which is itself the argument for a
      helper. -->
      
    • S11.4.c Provide semantic selection/range/additive, reorder/move/resize/ trim/pan/zoom/connect, compare, command palette, review, and confirmation helpers without pointer APIs. Evidence: complex interaction examples.

      <!-- S11.4.c: shipped as
      `libs/oshun/workbench-kit/src/interaction-helpers.ts`
      (`COMPLEX_INTERACTIONS` + `KEYBOARD_IDIOM` + `SURFACE_REGISTER` +
      `enumerable`, `WORD_COUNTS_THAT_ARE_NOT_SURFACE_COUNTS`,
      `DRAG_VERDICTS` + `dragVerdict` + `OSHUN_DRAG_REGISTER` + `dragCensus`,
      `PATH_STATES` + `OSHUN_INTERACTION_EVIDENCE`, `INTERACTION_FAULTS` +
      `interactionFaults`, `interactionCensus`, `interactionHeadline`), 26
      tests in `interaction-helpers.spec.ts`,
      `libs/oshun/workbench-kit/tools/interaction-helpers-scan.mjs` (nx target
      `interaction-helpers-scan`,
      `pnpm run verify:operations:v1-interaction-helpers`, 11 negative
      controls), and the five registrations.
      
      THE VEIN: **YOU CANNOT WRITE A HELPER FOR A SET YOU CANNOT ENUMERATE.**
      For exactly one of the twelve — dragging — this cell builds the register,
      and it takes a JSX parser: walk every opening tag, track brace depth so a
      handler body full of `>` does not end it early, and ask whether *that*
      tag carries the affordance. For the other eleven **the word is not the
      thing**: "review" appears **9,779** times across the two products and
      "compare" **1,384**, and neither is a count of an affordance. **11 of 12
      interactions have no register of their surfaces** — the same shape
      S11.3.b found for 2.5.7, 3.2.6, 3.3.7 and 3.3.8, arriving one cell later
      with a name.
      
      AND THE ONE REGISTER EARNS ITS KEEP IMMEDIATELY. **S11.3.b filed 2.5.7's
      input — "which operations are performed by dragging" — as having no
      producer. `OSHUN_DRAG_REGISTER` is that input.** Nine product elements
      carry `draggable` or `onDragStart`: **1** has an `onKeyDown` on the same
      element, **3** have an arrow handler or a move control elsewhere in their
      file, and **5 have none of the three anywhere in the file they live in** —
      two in `WidgetSidebar`, one each in `TaraSceneEditor`,
      `AreteSevenHabitsSystem` and `AreteVisionBoard`.
      
      **THOSE FIVE ARE CANDIDATES AND NOT VERDICTS, AND THE DIFFERENCE IS THE
      POINT.** An alternative can live in a sibling component and a drag can be
      essential, so the verdict is **`needs-exercising`** — source has run out,
      not the question. What settles them is exercising the surface, which is
      the helper this cell asks for driven by the instrument S11.3.b measured
      as aimed at a gallery. **Before this register, a 2.5.7 review had nowhere
      to start.**
      
      SECOND: **A HELPER WITHOUT A PATH IS NOT A TEST GAP, IT IS A PRODUCT
      GAP.** The command palette has a path and no helper — Control+K, arrows
      and Enter, 9 keyboard drives written out at each site. A `WidgetSidebar`
      reorder has no path at all, and filing that as a missing helper would put
      it in the wrong queue.
      
      THIRD, measured and small: **range selection is never driven by a
      keyboard here.** `press('Shift+Arrow…')` appears **zero** times in the V1
      suites while two products declare `aria-multiselectable`, and additive
      and range selection are the two the cell names first.
      
      AGAINST THE ESTATE: **12 interactions, 2 with an enumerable surface set
      (both through the one register), 2 with a keyboard path, 0 with a
      helper.** The scan rebuilds the drag register from the products every run
      and fails if it and the module disagree. -->
      
    • S11.4.d Detect pointer/touch/mouse calls, coordinate clicks, force-click, direct DOM evaluation, inaccessible test IDs, and accessibility-tree bypass in keyboard suites. Evidence: static/runtime guard.

      <!-- S11.4.d: shipped as
      `libs/oshun/workbench-kit/src/keyboard-guard.ts` (`BYPASS_KINDS` +
      `BYPASS_ASKS`, `DECIDABILITIES` + `DECIDABILITY` + `WHY_NOT_STATIC`,
      `MEASURED_CLICK_BEHAVIOUR`, `KEYBOARD_SUITE_DECLARATION` +
      `KEYBOARD_ONLY_MENTIONS`, `OSHUN_BYPASS_CENSUS`,
      `THE_ONE_KEYBOARD_ONLY_TEST`, `GUARD_FAULTS` + `guardFaults`,
      `guardCensus`, `guardHeadline`), 25 tests in `keyboard-guard.spec.ts`,
      and **`libs/oshun/workbench-kit/tools/keyboard-guard-scan.mjs`, which IS
      the static half of the guard** (nx target `keyboard-guard-scan`,
      `pnpm run verify:operations:v1-keyboard-guard`, 11 negative controls),
      plus the five registrations.
      
      THE VEIN: **`force: true` DOES NOT MAKE THE CLICK LAND — IT MAKES THE
      FAILURE SILENT.** Measured in a real Chromium on a button under a
      full-page overlay:
      
      - `locator.click()` **refused**, timing out on the receives-events
        actionability check; the target's handler never ran and the call threw.
      - `locator.click({ force: true })` dispatched a real mouse event at the
        target's coordinates and **the COVERING element received it**; the
        target's handler never ran and **the call resolved successfully**.
      
      So a forced click does not click the thing you named — it sends a real
      click at that thing's coordinates and whatever is on top gets it, then
      reports success. Each of the estate's **39** forced clicks is a call that
      **cannot fail for the one reason it was almost certainly added**, and can
      silently activate a different control. And that actionability check is
      **the estate's only automatic detector of an obscured control**, since
      S11.3.f measured 2.4.11 as having no instrument aimed at the products.
      
      SECOND: **IN THIS RUNNER, A POINTER CALL AND A COORDINATE CLICK ARE THE
      SAME CALL.** Measured: a click on an element at (50,60) sized 100×40
      dispatched a *trusted* event at exactly **(100,80)** — the centre of its
      box. There is no non-coordinate click in the API, so the cell's two items
      collapse into one.
      
      THIRD, about the evidence line itself: **"NO POINTER CALLS" IS NECESSARY
      AND NOT SUFFICIENT.** The one test in the estate titled *"keyboard-only
      members can reach the prompt input, send button, and voice toggle"* has
      **0 clicks, 0 forced clicks, 0 hovers, 0 mouse calls, 0 test ids** — it
      passes the parent's evidence line exactly — and calls **`.focus()`
      once**. A method call reaches anything at `tabindex="-1"`, anything in a
      closed subtree, anything the tab order skips: a keyboard-only test that
      sets focus that way **has assumed the thing it exists to prove**.
      
      FOURTH, which decides the guard's design: **THREE OF THE SIX ARE
      STATICALLY DECIDABLE AND THREE ARE NOT.** The guard runs three and
      deliberately **does not** report the 759 `evaluate` calls (most read
      `document.activeElement` or seed a store — a rule that reports them all
      is a rule somebody turns off) or call the 1,345 test ids inaccessible
      (that is a question about the element each resolves to). Both are
      reported as needing a runtime fixture, which is what "static/runtime
      guard" means.
      
      AND THE GUARD HAS NO SCOPE. **Nothing declares a keyboard suite** — no
      tag, no project, no marker. "keyboard-only" appears in three test titles,
      one journey-registry entry, some typed text and some prose, and a title
      is not a selector. **The parent's evidence line has no subject.**
      
      COUNTS: pointer 3554, coordinate 3546, force 39, evaluate 759, test id
      1345, tree bypass 20341.
      
      TRAPS, both caught here and fixed across all ten scans in this family:
      **(1) `execFileSync`'s default 1 MB buffer overflows on a common token —
      `locator(` returns 20,341 lines — and a `catch` that returns `[]` records
      the throw as a count of ZERO.** Every scan now sets a 64 MB buffer and
      rethrows anything but git-grep's exit-1 "no matches". **(2) A SCAN THAT
      SEARCHES THE REPOSITORY COUNTS ITS OWN PROSE** — fourth sighting: the
      register names the symbols it measures, so its own module, scan, nx
      target, ledger entry and generated export map all match. Every repo-wide
      search now excludes the cell's own artifacts by path. Both were found by
      controls, not by review. -->
      
    • S11.4.e Run critical create/edit/save/conflict/review/gate/bulk/publish/ recover journeys from fresh page through only keyboard input. Evidence: journey suite.

      <!-- S11.4.e LEFT UNCHECKED ON PURPOSE, with the gate that says why:
      `libs/oshun/workbench-kit/src/keyboard-journeys.ts` (`CRITICAL_ACTS` +
      `ACT_NEEDS`, `OSHUN_JOURNEY_SURFACES` + `registerNames`,
      `CITATION_VERDICTS` + `citationVerdict` + `OSHUN_CITED_TESTS`,
      `PRECONDITIONS` + `PRECONDITION_OWNER` + `resolvePrecondition` +
      `resolveActReadiness` + `resolveAllActReadiness`, `JOURNEY_FAULTS` +
      `journeyFaults`, `journeyCensus`, `journeyHeadline`), 33 tests in
      `keyboard-journeys.spec.ts`,
      `libs/oshun/workbench-kit/tools/keyboard-journeys-scan.mjs` (nx target
      `keyboard-journeys-scan`,
      `pnpm run verify:operations:v1-keyboard-journeys`, 10 negative
      controls), and the five registrations.
      
      **The nine journeys cannot be written, and the reason is not a missing
      test.** THE GATE, run: **0 of 9 acts runnable, every one blocked by 4 or
      5 preconditions, and every precondition owned by a sibling cell.**
      
      THE VEIN: **A JOURNEY IS SOMETHING THE USER DOES AND THE REGISTER
      ENUMERATES PLACES THEY GO.** The estate keeps a *good* critical-journey
      register with a `keyboardOnly` coverage dimension —
      `apps/oshun/web/e2e/support/critical-web-journeys.ts` — **14 entries,
      every one carrying a keyboard-only citation** with a file, a title
      fragment and a sentence of evidence. It reads as complete. And its 14
      entries are **surfaces**: home, explore, library, activity, search, the
      assistant dock, notifications, profile, six domains. **Not one of the
      nine acts appears in it.** Coverage of every place is not coverage of any
      act, and a create that cannot be done by keyboard is invisible to a
      register organised by route.
      
      SECOND, and this is what the register makes possible: **A CITATION IS A
      CLAIM THAT CAN BE CHECKED.** S11.4.d built a guard and reported that
      nothing declared a keyboard suite for it to run over. **This register IS
      that scope** — 8 distinct tests, named by file and title. Run the guard
      over them and **4 of the 8 do not hold**:
      
      - `shell-keyboard-workflows.spec.ts` "Control+K command palette supports
        search, creation, agent, and recent workflows" opens the palette with a
        chord and then **clicks five palette items**
        (`locator('[data-item-id="…"]').click()`).
      - `home-keyboard-focus-path.spec.ts` "hidden shell controls stay out of
        the tab order until visible" — cited for the Home journey — **presses
        no key at all.** It asserts `tabindex` is `-1`, then `0`. That is
        S7.21.f's third trap in the keyboard family: `tabindex="0"` is what the
        author wrote; a control with it inside an `inert` subtree, inside
        `display:none`, or inside a closed dialog is reached by nothing.
      - `assistant-accessibility.spec.ts` (×1) and
        `tara-session-player-accessibility.spec.ts` (×6) **set focus by method
        call.**
      
      THIRD: **THE GATE REFUSES FOR REASONS THE SIBLINGS ALREADY MEASURED, AND
      NOT ONE IS ABOUT WRITING A TEST.** No helper for any of the 12 complex
      interactions; **no keyboard path at all for reorder, move or range
      selection** (5 drag surfaces with no keyboard affordance anywhere in
      their file); 1 of 7 focus claims established; the only key helper a Tab
      search local to one spec; and no scope for the guard. **A keyboard-only
      `reorder` journey is not unwritten work — it is work that cannot start
      until a product offers a path.**
      
      Every precondition is resolved by calling the owning cell's own register,
      so the gate moves when a blocker clears rather than when somebody edits
      this file. The scan re-parses the estate's register every run — including
      its **shared citation constants**, which a naive parse misses entirely,
      reporting every journey as uncited — and opens every cited test.
      
      TRAP, committed by this cell and caught by its own controls:
      `registerNames` used `includes` and reported the register as naming the
      **`save`** act, because a surface is called
      `library-**save**d-continuity`. **A substring is not a name** — the
      cell's own vein turned on the cell. -->
      
    • S11.4.f Exercise supported browser/OS/editor/assistive shortcut collisions and alternate/remapped path. Evidence: collision tests. <!-- S11.4.f: shipped as libs/oshun/workbench-kit/src/shortcut-collisions.ts (+ .spec.ts, 37 tests) with tools/shortcut-collisions-scan.mjs (15 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/shortcut-collisions, verify:operations:v1-shortcut-collisions, and the shortcut-collisions-scan Nx target. Entry point 207 of 207, 0 export-map violations.

      **THE VEIN: THIS ESTATE MAKES A SHORTCUT CLAIM THREE DIFFERENT WAYS, IN
      ONE PRODUCT.** `AccessibilityShell.tsx` **never calls
      `preventDefault`** — `scheduleShellShortcut` defers a tick, checks
      `event.defaultPrevented` and yields if anything else claimed the key, so
      the browser's action and the shell's both happen.
      `CommandPalette.tsx` **calls `preventDefault` immediately** on
      Cmd/Ctrl+K, which is where Chrome and Firefox focus the address bar for
      a search. `TaraSitPlayer.tsx` binds **bare `k`, bare Space and bare
      arrows on `window`**. Three policies, three owners, seven branches — and
      **nothing anywhere lets a user change or disable any of them**, which is
      the cell's second ask answered `REMAP_PATH = null`.
      
      **THE SECOND RULE, AND IT IS WHERE THE COLLISION ACTUALLY IS: A MODIFIER
      GUARD IS PART OF THE CHORD, AND TWO OF THE SHELL'S FOUR BRANCHES DO NOT
      WRITE ONE.** `Alt+<digit>` and `Alt+A` are both
      `event.altKey && !event.ctrlKey && !event.metaKey && !event.shiftKey`.
      `Alt+/` is `event.altKey && event.key === '/'`, and the fourth branch is
      `event.key === 'Escape'`. **`Ctrl+Alt` is not a spare combination**: on
      Windows it is **AltGr**, which four of the eight launch locales
      (S11.3.c's) hold down to type `@`, `\`, `€` and `~`; on macOS it is
      **VoiceOver's own modifier**, held before every VO command. The
      combination the two guarded branches exclude, the other two admit.
      
      **THE THIRD, AND IT IS THE SAME SHAPE ONE LEVEL DOWN: A GUARD PROTECTS
      THE BRANCH IT IS WRITTEN INSIDE.** The player's handler is four
      branches and the `isInteractiveKeyboardTarget(e.target)` early return is
      written **inside the first one** (`TaraSitPlayer.tsx:419-423`).
      `ArrowLeft` and `ArrowRight` are tested five lines below it and
      **outside it** (`:424-429`), so an arrow pressed while the caret is in a
      text field **moves the caret and scrubs the audio**. The estate has
      **exactly one** text-target guard, it is **correct**, and it covers two
      of that handler's four keys — which is why `actsOn` takes the guard per
      **branch**: a register that took it per file would clear the two
      branches it does not cover.
      
      **AND THE SHELL'S FOURTH BRANCH IS BARE `Escape` CALLING `blur()` ON
      `document.activeElement`**, with no modifier guard and no target guard.
      Escape is how a user dismisses an IME candidate window; **S11.4.b
      measured** that the keydown which does it carries `isComposing: true`
      and `keyCode` 229, and that no product handler reads either. A Japanese,
      Korean or Chinese user backing out of a candidate list also loses the
      field.
      
      **DECISIONS.** **THE ARITHMETIC SAYS A GUARD ADMITS A COMBINATION, NOT
      THAT A USER PRODUCES ONE.** Whether `AltGr+K` types a character on any
      particular layout is a fact about that layout's third level and this
      register does not hold one — so a modifier is either `held-across-keys`,
      where admitting the combination is admitting the key, or
      `a-named-chord`, which collides only with a branch bound to that key.
      No finding claims a keycap the register cannot see. **WCAG 2.1.4 IS
      SCOPED BY `event.key` LENGTH**, because the criterion covers letters,
      punctuation, numbers and symbols and a `key` of length 1 is exactly one
      of those: three bare bindings produce **one** 2.1.4 fault, and the
      arrows and Escape are reported as collisions rather than as failures of
      a criterion that does not cover them.
      
      **THE SCAN COPIES NOTHING.** It cuts each branch out of its source by
      matching brackets, derives the required and excluded flags from the
      condition, derives the keys from the key comparisons — and the digit
      branch's five from the **length of its own `routes` array**, the thing
      that would change if a sixth tab were added — and reads the text guard
      from the branch **body**.
      
      **THE READING:** 22 faults; 5 of 7 branches collide; all four sources
      the cell names are reached; 1 text guard; no remap path in either
      product. **WHAT IS RIGHT AND WORTH SAYING:**
      `resolveShellRouteShortcutIndex` reads `event.code` before `event.key`,
      so `Alt+3` is the third **physical** digit rather than whatever
      character a layout puts there. -->
      
  • S11.5 Add screen-reader announcement assertions. Evidence: async status and error messages are covered.

    <!-- S11.5 LEFT UNCHECKED ON PURPOSE. Five children shipped `[x]` (a–e);
    .f is `[ ]` because its verb is "run" and no reader can be driven here.
    The parent is unchecked because **its evidence line is a coverage claim
    and both halves of it were measured false by its own children.**
    
    **"ASYNC STATUS … COVERED" — S11.5.c: 11 of the 22 matrix cells are
    progress claims, and a progress claim is about what happened BEFORE the
    result replaced it.** No read of the final DOM contains one. The estate's
    shape is a ternary whose branches are two different nodes, so a progress
    message and its result are two insertions rather than one region speaking
    twice, and **whether the first was ever spoken is unobservable without the
    change observer S11.5.a found nobody aims at a live region.**
    
    **"ERROR MESSAGES … COVERED" — S11.5.b: 1 of 5 validation fixtures can
    run, and it is the one nobody has written.** There is **no validation
    summary in either product** against 507 submit handlers; a control error
    needs the computed description, which only the AX tree carries and no
    suite asks for; a correction is a SUPPRESSION and a resolution is a
    DEPARTURE, and **a live region is deaf to departures by default**
    (`aria-relevant` = "additions text").
    
    **AND THE THROUGH-LINE OF THE WHOLE FAMILY: THE ESTATE ALREADY CONTAINS
    SEVEN CORRECT MECHANISMS, EVERY ONE DOCUMENTED WITH ITS INVARIANTS, AND
    ALMOST NOTHING CALLS THEM.** `LiveRegion` (0 shipping callers, and its doc
    states the announcement rule verbatim), `useLiveAnnouncer` (1),
    `useErrorAssociation` (0), `useFocusTrap` (0), `useRovingTabIndex` (0),
    the shell's `announce` (0 importers, mounted app-wide, fails silent),
    `Popover` and `ModalDialog` (0). Meanwhile **808 hand-rolled live
    elements**, **54 hand-rolled dialogs**, and **five separate announcer
    implementations answering one question three ways**.
    
    What would close this parent is not a test: it is routing the product
    through the mechanisms it already has, a validation summary, and one
    change observer aimed at a live region. Where the estate IS good — 175 of
    180 failure regions announced assertively, two of three virtualized
    surfaces answering position by two different mechanisms, and a route
    announcer the framework provides free — the children say so. -->
    
    • S11.5.a Instrument accessible name/description/error/message, live region role/politeness/atomic/relevant, focus, DOM order, language, and visibility without replacing real AT smoke tests. Evidence: assertion library. <!-- S11.5.a: shipped as libs/oshun/workbench-kit/src/announcement-instruments.ts (+ .spec.ts, 32 tests) with tools/announcement-instruments-scan.mjs (12 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/announcement-instruments, verify:operations:v1-announcement-instruments, and the announcement-instruments-scan Nx target. Entry point 208 of 208, 0 export-map violations.

      **THE VEIN: AN ANNOUNCEMENT IS A CHANGE OBSERVED, AND EVERY ASSERTION IN
      THIS ESTATE READS A STATE AFTER THE FACT.** A live region is announced
      because it changed **while something was watching**. A region inserted
      into the document together with its first message was not there to
      change; on most browser and reader pairings a polite one is silent, and
      **every test that checks the text is present passes**. So the fact
      register carries a **thirteenth** entry the cell does not name —
      `present-before-the-change` — needing a capability
      (`change-observation`) that no attribute read provides.
      
      **THE ESTATE ALREADY WROTE THAT RULE DOWN.**
      `apps/oshun/web/src/components/workbench/primitives/accessibility.tsx`
      exports a `LiveRegion` whose doc says it in as many words — "THE
      CONTAINER EXISTS BEFORE THE MESSAGE DOES, and getting that backwards is
      the single most common way a live region announces nothing" — rendering
      unconditionally with `aria-atomic="true"` and declaring both invariants
      on the component. **It has three usages: a fixture host and two
      Storybook stories. None of them ships.** Beside it, walked **tag by
      tag** with brace-depth tracking rather than counted by file, the two
      products carry **808 hand-rolled live elements**, of which **25** set
      `aria-atomic` and **464 are gated by a conditional in the expression
      they sit directly inside** — the exact shape the primitive prevents.
      
      **AND THE 464 ARE NOT 464 DEFECTS**, which is the other half of the
      cell. **313 are `role="alert"` or `aria-live="assertive"`**, and an
      alert added to the document is announced by the major readers — the
      standard toast pattern, not a bug. **151 are polite**, and those are the
      silent case. They are filed **needs-exercising, not fails**, because
      what settles one is an observation this module does not make.
      
      **WHAT THE SUITES ACTUALLY POINT AT A REGION**, over 5,264 test files:
      **112 role queries**, **23 politeness assertions**, **3 atomic
      assertions**, **0 relevant assertions**, and **four `MutationObserver`s,
      not one of them on a live region** — one samples chip paint, one hides
      the Next dev overlay, one times a search cold start, and the fourth is a
      comment in `panel-timeline.ts` explaining that it samples *instead of*
      observing. The instrument is in the estate's hands four times, aimed
      elsewhere.
      
      **THE SECOND RULE, AND IT IS WHY THIS IS A LIBRARY AND NOT A LINT: A
      COMPUTED PROPERTY IS NOT AN ATTRIBUTE, AND HALF OF THESE FACTS ARE
      COMPUTED.** `role="status"` **is** polite and atomic, and
      `getAttribute('aria-live')` returns `null` for every one of the
      estate's **331** of them. An accessible name is computed from several
      sources in order, so `aria-label` is an input and not the answer; a
      language is inherited; a visibility is resolved style. So each fact
      declares the **capability** it needs and each instrument what it
      provides, and `settles` is that one comparison rather than a table.
      
      **THE THIRD — THE CELL'S OWN CAVEAT MADE INTO DATA: NO FACT HERE IS THE
      UTTERANCE.** `WHAT_ONLY_AN_UTTERANCE_SETTLES` lists four questions the
      library refuses, each with its owner: whether a reader speaks an
      inserted polite region and whether it honours
      `aria-relevant="removals"` (**S11.5.f**), and what happens when two
      regions change in one tick and what is said when a message is replaced
      mid-utterance (**S11.5.d**).
      
      **DECISIONS.** The gate test is **narrow and says so**: it asks whether
      the tag's existence is gated by the expression it sits *directly
      inside*, not whether the element can ever be absent — an early
      `return null` three functions up also decides that — so it under-counts
      on purpose. And `grader-asked-for-a-value` is asked of what the suites
      **aim**, never of the instrument list, where a measurer always exists
      beside axe: **a check that can never fire is not a check** (S10.14's
      trap, caught here before shipping).
      
      **THE READING:** 4 of 13 facts have an instrument the suites aim at a
      region; 1 of 6 instruments is aimed at one at all; 12 faults. **The scan
      caught its own register**: the 113th role query is a comment about
      Next's route announcer, and **a mention in prose is not a use** →
      112. -->
      
    • S11.5.b Assert validation summary/control error, focus movement, correction, and resolved error announcements. Evidence: validation fixtures. <!-- S11.5.b: shipped as libs/oshun/workbench-kit/src/validation-announcements.ts (+ .spec.ts, 29 tests) with tools/validation-announcements-scan.mjs (11 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/validation-announcements, verify:operations:v1-validation-announcements, and the validation-announcements-scan Nx target. Entry point 209 of 209, 0 export-map violations.

      **THE VEIN: A LIVE REGION ANNOUNCES ARRIVALS. `aria-relevant` DEFAULTS
      TO "additions text", SO ONE OF THE CELL'S FIVE EVENTS IS SILENT BY
      CONSTRUCTION AND ANOTHER CANNOT BE OBSERVED AT ALL.** Sorted by what has
      to happen rather than by what they are called, the five are not five of
      a kind: **summary** and **control-error** are ARRIVALS; **focus-
      movement** is a MOVE, the only one the platform announces for free (a
      reader speaks whatever took focus — no region involved); **resolution**
      is a DEPARTURE — removing the error node is a removal, the default
      `relevant` does not include removals, so **the error disappearing
      announces nothing**, and the only carrier is a positive message pushed
      into a region that was already there, which turns it back into an
      arrival; and **correction** is a SUPPRESSION — the requirement is that
      nothing is re-announced on every keystroke, and **an absence cannot be
      read out of a final state**.
      
      **THE SECOND RULE: A SUMMARY IS A PATTERN, NOT A WORD.** Searched as the
      word, `ErrorSummary` appears in five files and **every hit is a job
      record's `lastErrorSummary` column** or a paragraph in a studio
      workspace. Searched as the pattern the cell means — a form-level error
      list whose entries link to their fields — the two products have
      **zero**, against **507 `onSubmit` handlers** and 21 `<form>` tags. The
      cell's first item has nothing to assert.
      
      **THE THIRD, AND IT IS S11.5.a's SHAPE ONE FLOOR DOWN: THE ESTATE
      CONTAINS THE ANSWER AND ALMOST NOTHING CALLS IT.**
      `apps/oshun/web/src/components/studio/a11y/accessibilityHooks.tsx`
      exports **`useErrorAssociation`** — `aria-invalid` +
      `aria-describedby` + `role="alert"` on the message so it **is announced
      when it appears**, with the association dropped entirely when there is
      no error so nothing points at a missing element — **zero callers**.
      **`useLiveAnnouncer`** keeps stable polite and assertive regions, both
      atomic, clearing then re-setting on a tick so an identical repeated
      message still reads as a change (its own comment: "a remounted aria-live
      region often fails to announce") — **one caller**. `useFocusTrap` and
      `useRovingTabIndex`, same file, **none**.
      
      **WHAT THE SUITES ASSERT:** 11 mentions of `aria-invalid`, 3
      `aria-describedby` assertions, **0 `aria-relevant` anywhere**, and — the
      number that decides the cell — **not one test that both touches an
      invalid control and asserts what has focus**. 71 `toBeFocused`
      assertions, none after a failed submit.
      
      **THE READING: 1 of 5 fixtures runnable, and it is `focus-movement` —
      the only one that could run today is the only one nobody has written.**
      The module specifies each fixture and does **not** write them, because
      two of the five have no subject in this estate and a fixture asserting
      them would be asserting its own scaffolding; `fixtureReadiness` computes
      every blocker from **S11.5.a's** registers, so the gate moves when the
      instrument or the product changes rather than when somebody edits a
      table. 11 faults. -->
      
    • S11.5.c Assert save/offline/retry/conflict/recovery and job/gate/review/ comment/publish/destructive progress/result announcements. Evidence: async state matrix. <!-- S11.5.c: shipped as libs/oshun/workbench-kit/src/async-announcements.ts (+ .spec.ts, 22 tests) with tools/async-announcements-scan.mjs (10 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/async-announcements, verify:operations:v1-async-announcements, and the async-announcements-scan Nx target. Entry point 210 of 210, 0 export-map violations. The matrix is the cell's eleven subjects × two phases = 22 cells, each saying what the announcement has to carry (conflict/result: which version won and what the user has to do about it; gate/result: the verdict and the reason, not just the verdict; destructive/result: what was deleted and whether it can be undone).

      **THE VEIN: A PROGRESS ANNOUNCEMENT AND ITS RESULT ARE WRITTEN AS TWO
      BRANCHES OF ONE CONDITIONAL, SO THEY ARE TWO NODES AND NOT ONE REGION
      SPEAKING TWICE.** `state === 'busy' ? <span role="status">{busyLabel}
      </span> : state === 'error' ? <span role="status">…</span> : null` is the
      estate's shape and it is **two insertions** — S11.5.a's thirteenth fact
      applies to each separately. **Whether the progress message was ever
      spoken before the result replaced it is an ordering claim, and no read
      of the final DOM contains one**, so **11 of the 22 cells are marked not
      fillable from source** and handed to the change observer.
      
      **THE SECOND RULE, AND IT IS THE DECIDABLE ONE: TONE IS DECIDED BY THE
      OUTCOME, NOT BY THE SUBJECT.** Progress is polite (interrupting a reader
      mid-sentence to say "Saving…" is worse than waiting); success is polite;
      a failure the user's own action produced is **assertive**, because a
      polite failure queues behind whatever is being read and lands after they
      have moved on.
      
      **AND THE ESTATE IS GOOD AT IT.** Of **464 conditionally-rendered live
      regions**, **180 are gated on a failure condition** and **175 are
      `role="alert"` or assertive** — the strongest accessibility discipline
      measured anywhere in this family. So the deliverable is **the five that
      are not**, named with the gate that renders each:
      `AtelierPlateActions.tsx:218` (`state === 'error'`),
      `SupportContinuityDiagnosticsPanel.tsx:417` (`error &&`),
      `MetisAssessmentEditor.tsx:416` (`submitState === 'error'`),
      `IntegrityAppealWorkspace.tsx:271` (`load.kind === 'load-error'`), and
      `NotificationsCenterPanel.tsx:1691` (`feedState === 'unavailable'`).
      
      **ONE OF THE FIVE IS RECORDED AS DEFENSIBLE AND FOUR ARE NOT — a
      distinction worth making rather than a number worth inflating.** The
      notifications one announces a **disclosed outage**, an ambient page
      state, which is exactly what polite is for; the other four are results
      of something the user just did.
      
      **THE THIRD: `aria-busy` IS NOT AN ANNOUNCEMENT.** It tells a reader to
      hold off while a subtree updates and says nothing itself. The products
      carry **216 `aria-busy` tags and exactly one is on a live region** —
      `StudyRouteBoundary.tsx:197`, set only while loading with its politeness
      keyed to its own role, and **it is right**. A register counting
      `aria-busy` as progress coverage would report 216 announcements that do
      not exist.
      
      **DECISIONS.** The literal-text window (260 chars) is **stated in the
      register**, because the number depends on it — a message beginning 300
      characters in reads as computed at 260 — and the scan uses the same
      window, so the two agree by construction rather than by luck. And the
      scan **fails both ways** on the five: a named site the walk cannot find
      is a repair the register should stop naming; an unnamed one it finds is
      a new defect. 6 faults. -->
      
    • S11.5.d Assert debouncing/deduplication, priority/interruption, no stale queued message, and no announcement flood under rapid events. Evidence: event storm tests. <!-- S11.5.d: shipped as libs/oshun/workbench-kit/src/announcement-storms.ts (+ .spec.ts, 29 tests) with tools/announcement-storms-scan.mjs (12 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/announcement-storms, verify:operations:v1-announcement-storms, and the announcement-storms-scan Nx target. Entry point 211 of 211, 0 export-map violations.

      **THE VEIN: A STORM POLICY THAT GROUPS WHAT YOU WILL READ LATER SAYS
      NOTHING ABOUT WHAT IS SPOKEN NOW.** **S7.19.b already solved this** —
      groups and deduplicates event storms, sizes a group over **distinct
      event ids**, states its distinct-target count, names seven kinds that
      may never be collapsed. It operates on the **notification feed**, a list
      somebody opens and reads. An announcement is not a list; it is a moment.
      **None of that policy is on the announcement path.**
      
      **THE SECOND RULE, WITH THE ESTATE'S OWN CODE AS EVIDENCE: DEDUPLICATION
      AND RE-ANNOUNCEMENT ARE THE SAME MECHANISM RUN BACKWARDS, AND FIVE
      ANNOUNCERS ANSWER ONE QUESTION THREE WAYS.** **Three say yes on
      purpose** — `useLiveAnnouncer` clears then re-sets after a tick, the
      shell's `announce` clears then writes on the next frame ("so repeated
      identical messages still trigger"), `KeyboardTriageReviewSurface` bumps
      an id. **One says no by accident** — `StudyWorkbench`, the most-called
      announcer at **ten sites** with every overlay handler routed through it,
      is `setAnnouncement(message)` with no id and no timer, so **React bails
      out on an equal string**: deduplication as a consequence of `setState`
      rather than a decision, and the estate's only implementation of it. And
      the `LiveRegion` primitive is a container, not a queue. Repeating is
      right when the same thing happened again and wrong when it is still
      true, and **nothing distinguishes those two cases**.
      
      **THE THIRD: NO ANNOUNCER HAS A RATE LIMIT**, and the two most-called
      are on repeat paths — `KeyboardTriageReviewSurface` announces the
      artifact ref, priority and stage on **every arrow move** through a
      review queue.
      
      **THE FOURTH: THE MOST GLOBALLY AVAILABLE ANNOUNCER IS CALLED BY NOBODY
      AND WOULD FAIL SILENT IF IT WERE.** `AccessibilityShell` exports
      `announce(message)` and mounts its region from `layout.tsx` on every
      page; the body is `announceRef?.(message)` on a nullable module
      singleton, **returning `undefined` whether or not anything was said**,
      with **zero importers**.
      
      **A WORD IS NOT THE THING, FOR THE FOURTH TIME IN THIS FAMILY.**
      `announceRecordsChanged` is a DOM event bus for stale data and reaches
      no reader. And the estate's **event storm tests number zero**: `storm`,
      `rapid`, `burst`, `flood` across 5,264 suite files yield a **comment
      about announcement tone**, a **mass-share 429 guard**, and a character
      named **Kaelen Stormweaver** — all three recorded, because a zero
      nobody can reproduce is worth nothing.
      
      **DECISIONS.** **A call site is not a shipping call site**: the
      primitive has three and none ships, so both numbers are carried rather
      than one quietly filtered into the other — **the scan derived three
      where the register said zero**, which is how the distinction got made.
      And each announcer's calls are counted **the way its own callers reach
      it**: a bare `announce(` belongs to the module function only in a file
      that imports it, without which the count attributed ten of
      `StudyWorkbench`'s calls to a function nothing imports. **Four of the
      five requirements are producer-side and one is not** — which of two
      regions is spoken first is the reader's queue, already filed by S11.5.a
      as S11.5.f's. 7 faults. -->
      
    • S11.5.e Assert dynamic counts, selection, virtualized position, sort/filter changes, dialog/drawer, route transition, and focus restoration context. Evidence: navigation fixtures. <!-- S11.5.e: shipped as libs/oshun/workbench-kit/src/navigation-announcements.ts (+ .spec.ts, 23 tests) with tools/navigation-announcements-scan.mjs (10 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/navigation-announcements, verify:operations:v1-navigation-announcements, and the navigation-announcements-scan Nx target. Entry point 212 of 212, 0 export-map violations.

      **THE VEIN: THE SAME FACT HAS A DIFFERENT MECHANISM IN EVERY PATTERN,
      AND A REGISTER THAT LOOKS FOR ONE OF THEM REPORTS THE OTHERS AS
      ABSENT.** "Where am I in this list" is `aria-setsize`/`aria-posinset` in
      a listbox or tree, `aria-rowcount`/`aria-rowindex` in a grid, and a live
      region stating the visible range in a plain scrolling list — **all three
      correct**. This cell was **one grep from a false finding**: `aria-setsize`
      appears in six files and **none of them virtualizes**, which reads as
      "every virtualized surface is silent" until you look. Of the estate's
      **three real virtualized surfaces, two answer position and they use two
      different mechanisms** — `VirtualList` renders an `srOnly` polite region
      saying "Showing items N to M of TOTAL"; `DataTable` sets `aria-rowcount`
      to the total and `aria-rowindex` to the **absolute** index, with a
      comment saying why. The one that answers by none is the compare matrix.
      
      **AND `overscan` IS NOT A VIRTUALIZATION MARKER EITHER.** Seven files
      match it: three virtualize, two are non-JSX helpers, and **two are other
      trades entirely** — a cinematography workspace's **broadcast overscan
      safety** and a drone panel's **frustum overscan cushion**. A word is not
      the thing, for the fifth time in this family.
      
      **THE SECOND: ONE OF THE SEVEN IS THE FRAMEWORK'S, AND THE ESTATE'S ONLY
      INTERACTION WITH IT IS AVOIDING IT.** Next renders `<div role="alert"
      id="__next-route-announcer__">` and speaks the new page's title on every
      client navigation — **already there, on every route, for free**. The
      estate's only two mentions are comments in one spec explaining that a
      locator had to be scoped to `p[role="alert"]` **so as not to collide
      with it**. Recorded as `framework-provided` rather than present or
      missing, because both of those words would be wrong.
      
      **THE THIRD — THIS FAMILY'S RUNNING TOTAL: `overlay.tsx` EXPORTS A
      CORRECT `Popover` AND `ModalDialog` AND NOTHING THAT SHIPS USES
      EITHER.** The popover reads `document.activeElement` **at open time
      rather than assuming**, and restores only if the saved element
      `isConnected` — because `focus()` on a detached element does nothing and
      says nothing about it. The dialog uses `showModal()`, "so the inertness
      is the platform's and covers what a JS trap cannot reach". **Eight call
      sites, all specs and stories, zero shipping** — against **54 shipping
      files that hand-roll `role="dialog"`** and **five that restore focus at
      all**. Across S11.5 that makes **seven correct mechanisms with almost no
      adoption**: `LiveRegion` 0, `useLiveAnnouncer` 1, `useErrorAssociation`
      0, `useFocusTrap` 0, `useRovingTabIndex` 0, the shell's `announce` 0
      importers, `Popover`/`ModalDialog` 0.
      
      **TRAP CAUGHT BY THE SCAN:** the register quoted `ModalDialog`'s
      invariant **without its backticks** around `showModal()`, and the
      verbatim check failed. **A quotation missing one character is not a
      quotation** — same family as S10.14's ellipsis. The scan also tightened
      four counts my shell greps had inflated by not excluding stories,
      fixtures and prose (90→76, 202→193, 55→54). 6 faults. -->
      
    • S11.5.f Run representative NVDA/JAWS/VoiceOver or supported real screen- reader speech-log journeys and reconcile differences from DOM-level assertions. Evidence: AT matrix. <!-- S11.5.f LEFT UNCHECKED ON PURPOSE, with the matrix shipped empty: libs/oshun/workbench-kit/src/at-reconciliation.ts (+ .spec.ts, 27 tests) with tools/at-reconciliation-scan.mjs (10 negative controls), registered as @oshun/workbench-kit/at-reconciliation, verify:operations:v1-at-reconciliation, and the at-reconciliation-scan Nx target. Entry point 213 of 213.

      **THE CELL'S VERB IS "RUN", AND NO SCREEN READER CAN BE DRIVEN FROM THIS
      HOST.** NVDA, JAWS and Narrator are Windows; VoiceOver is Apple's;
      TalkBack is Android's; Orca needs a desktop session with a running
      speech service. And — measured rather than assumed — **nothing in this
      repository drives one**: `speechLog` and `speech-log` appear nowhere
      outside this cell's own family, and the estate's only register of
      readers is `platformScreenReaderApis` in
      `libs/iris/accessibility/src/index.ts`, a **six-name array inside a
      bridge policy**, which is a declaration and not a driver.
      
      **SO ALL 36 CELLS ARE `unobserved` AND NOT ONE IS FILLED IN FROM
      MEMORY.** What a reader does with an inserted polite region, or with
      `aria-relevant="removals"`, or with two regions changing in one tick, is
      a fact about a program nobody here has run. **Writing those verdicts
      down because they are widely believed would be exactly the thing the
      rest of S11.5 exists to catch.**
      
      **WHAT IS SHIPPED, AND IT IS THE CELL'S SECOND CLAUSE.** **The
      questions**: six, none invented here — four imported from S11.5.a's own
      list, one S11.5.b's departure, one S11.5.d's queue — each carrying **who
      raised it and who owns answering it**, which are *not* the same cell
      (S11.5.a raised all four of its own; two are owned by S11.5.d because
      they are about a queue, not a reader). **The pairs**: a reader is never
      tested alone but against a browser, because the tree that reaches it is
      the browser's — so the matrix is 6 pairings × 6 questions. **The
      reconciliation**, which is not "did they match": a DOM assertion passing
      while the reader is silent has **three** meanings — the assertion was
      wrong, the reader does not support the pattern, or the pattern is
      ambiguous and both are defensible — and collapsing them into "fail"
      tells nobody what to change. The reverse — the reader speaks where the
      DOM said nothing — is **always** the assertion being wrong, because an
      utterance is the ground truth this family has been approximating.
      
      **AND WHAT A SPEECH LOG HAS TO BE:** not a transcript but a
      **time-ordered list of utterances on a clock alignable with the page's
      own events** — without it, "the save message was spoken" cannot be told
      from "the save message was spoken after the user had already navigated
      away", which is the whole reason S11.5.d exists. The register states the
      requirement and deliberately **does not name a vendor facility**, since
      which switch produces such a log on a reader nobody here has run is one
      more claim about an unobserved system.
      
      **WHAT WOULD CLOSE IT:** a Windows host with NVDA or JAWS, an Apple host
      with VoiceOver, or a desktop Linux session with Orca; a speech-log
      capture meeting the four stated requirements; and a product decision
      about which pairs are supported. None of those is a test somebody forgot
      to write. -->
      
  • S11.6 Add responsive and theme visual-regression baselines. Evidence: approved baseline matrix.

    <!-- S11.6 LEFT UNCHECKED ON PURPOSE. All seven children are `[x]`. The
    parent is unchecked because **its evidence line asks for an APPROVED
    baseline matrix, and S11.6.g measured that nothing approves anything.**
    
    **THE MATRIX EXISTS.** `libs/oshun/design-language/src/visual-qa.ts` holds
    nineteen claims, eighteen of them `blocking: true`, each naming its spec,
    test title, state, viewports, evidence kind and capture artifact — and its
    own doc says the mappings exist so they can be "required at signoff time".
    
    **NOTHING APPROVES IT.** No hook and no workflow notices a baseline
    change. The one CI job named Visual Regression is inert **five independent
    ways** — `|| true`, a target declared in zero `project.json` files, six
    specs that are not behind that target, an `if: failure()` upload on a job
    that cannot fail, and globs that name `__snapshots__` where Playwright
    writes `<spec>.spec.ts-snapshots`. **Three of the five would each stop it
    alone.**
    
    **AND THE TWO AXES THE PARENT NAMES ARE THE TWO LEAST COVERED.**
    *Responsive*: 3 of 7 viewport cases declared, 3 exercised, **and not the
    same three** — zoom/reflow declared and never driven, orientation driven
    only by the mobile Detox suite, and persistent scrollbar, software
    keyboard and panel arrangement neither. *Theme*: the product declares
    three Lilith variants and **every harness lists two — the missing one is
    `cream-hc`, the high-contrast theme**, so a contrast regression in it
    changes no committed pixel.
    
    **AND WHAT THE MATRIX WOULD BE APPROVING.** 4 of 19 citations resolve as
    written (12 more on a dash); 7 of 32 state-index cells are filled and 69
    of 74 baselines are a surface at rest; 1 of 6 content shapes is captured,
    with a 505-line bidi library and a 404-line pseudo-localiser rendering
    into none of them; 91 pixel tolerances carry seven unnamed values; and a
    signature over the pixels covers **2 of the 8 review dimensions**, because
    six are decided by geometry, an engine rule, a focus assertion or a
    semantic read — and one of those, semantic state, **no threshold can
    reach**.
    
    What would close this parent is a gate: notice a baseline change, require
    a reason and a linked implementation, and fail without them — three of
    those four decidable from a diff and a commit message. Where the estate IS
    good, the children say so: the masks are named and justified region by
    region, `unstableRegions` is headed "Regions whose CHANGE is not a
    regression", the theme helpers refuse to plant an attribute the product
    never sets, and `tables-and-details` pins six controls and needs one
    tolerance. -->
    
    • S11.6.a Define supported narrow/medium/wide, zoom/reflow, orientation, DPR, persistent scrollbar, software keyboard, and panel arrangement viewport cases. Evidence: viewport manifest. <!-- S11.6.a: shipped as libs/oshun/workbench-kit/src/viewport-manifest.ts (+ .spec.ts, 20 tests) with tools/viewport-manifest-scan.mjs (11 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/viewport-manifest, verify:operations:v1-viewport-manifest, and the viewport-manifest-scan Nx target. Entry point 214 of 214, 0 export-map violations.

      **THE VEIN: A CITATION IS A STRING MATCH, AND A DASH IS NOT A DASH.**
      `libs/oshun/design-language/src/visual-qa.ts` is a **good** register —
      nineteen claims, each naming the shell track, surface, UX claim,
      **Playwright spec**, **test title**, state id, viewports, evidence kind,
      capture artifact and whether it blocks; **18 of 19 blocking**; its own
      doc says the mappings exist so they can be "queried, validated, and
      required at signoff time". **All 19 name a spec file that exists. Four
      name a test title that does.** **Twelve more resolve the moment ` - ` is
      read as ` — `** — the register writes a hyphen-minus, the specs write an
      em dash (44 of 44 titles). **The test is there and the citation cannot
      find it**, so a validator taking the register at its word reports twelve
      missing tests that are not missing. **Three are genuinely absent**
      (`UX-SHELL-01`, `UX-ADMIN-SHELL-01`, `UX-ADMIN-QUEUE-01`).
      
      **THE SECOND: A BASELINE CARRIES THE PLATFORM THAT MADE IT.** Playwright
      stamps `-{project}-{platform}` onto every snapshot name, and the
      committed baselines are **67 `chromium-darwin` and 16
      `chromium-linux`**. A run on a platform with no baseline **does not fail
      the comparison — there is nothing to compare** — so the same suite is a
      regression test on a Mac and a **baseline writer** everywhere else, and
      which one is decided by the runner's OS rather than by the suite.
      
      **THE THIRD: THREE MANIFESTS, AND THE SIX VISUAL-REGRESSION SPECS USE
      NONE OF THEM.** `OSHUN_QA_VIEWPORTS` (7 sizes), `e2e-inspect`'s
      `VIEWPORTS` (3 lenses **with scale factors and a 200% zoom**), and
      S11.3.c's `OSHUN_AXIS_SUPPLY` (2). The specs declare their sizes inline
      — **49 literal sizes in one file, 30 in another** — and `mobile-large`
      is named by the manifest and claimed by nothing.
      
      **THREE OF SEVEN CASES ARE DECLARED AND THREE ARE EXERCISED, AND THEY
      ARE NOT THE SAME THREE:** `zoom-reflow` declared in the inspect harness
      and driven nowhere; `orientation` driven by the mobile Detox suite and
      declared nowhere; **persistent scrollbar, software keyboard and panel
      arrangement neither** — each recorded with what it costs (a software
      keyboard shrinks the **visual** viewport without changing the layout
      viewport; a persistent scrollbar moves a layout across a breakpoint a
      headless run never crosses).
      
      **WHAT IS RIGHT AND WORTH SAYING:** `lenses.ts` refuses to plant the
      theme attribute from an init script — the default theme is the
      attribute's **absence**, so forcing it is a state the product never has
      — and ships `expectThemeApplied` because "without this check a 'both
      themes' pass can silently inspect the same theme twice". **A positive
      control, arrived at independently.**
      
      **The scan expands the `<viewport>` template and strips the platform
      stamp before looking for an artifact** — a check that skipped either
      would report 13 missing baselines that are committed. 11 faults. -->
      
    • S11.6.b Define light/dark/high-contrast/forced-colors/branded themes, comfortable/compact/touch densities, reduced motion, and color scheme. Evidence: context manifest. <!-- S11.6.b: shipped as libs/oshun/workbench-kit/src/context-manifest.ts (+ .spec.ts, 21 tests) with tools/context-manifest-scan.mjs (11 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/context-manifest, verify:operations:v1-context-manifest, and the context-manifest-scan Nx target. Entry point 215 of 215, 0 export-map violations.

      **THE VEIN: THE ONE THEME THE CELL NAMES AS ITS OWN CASE IS THE ONE
      EVERY HARNESS LIST OMITS.** The product declares **three** Lilith
      variants — `cream`, `cream-hc`, `dusk` — as a union type with a matching
      guard in `LilithThemeBootstrap.tsx`, and **`cream-hc` is the
      high-contrast one**, with its own block in `globals-v2.css`. The inspect
      harness lists two. The assistant visual spec carries its **own copy** of
      the same `useTheme` helper and lists the same two. **The theme whose
      entire purpose is contrast is the one no visual baseline is taken in**,
      so a contrast regression in it changes no committed pixel.
      
      **THE SECOND — S11.3.f ARRIVING FOR A SECOND CONTROL FAMILY: THE
      INSTRUMENT EXISTS, IT IS GOOD, AND IT IS AIMED AT THE GALLERY.**
      `forced-colors` is answered by **13 shipping stylesheets** and **eleven
      `*-browser.mjs` harnesses emulate `forcedColors: 'active'`**, their docs
      explaining what must stay distinct under it — and **every one mounts a
      fixture host**. The six visual-regression specs that capture a product
      page emulate it **zero** times. `prefers-reduced-motion`: 45 shipping
      files, three harnesses, **one** visual spec (for the good reason that
      freezing animation makes a capture stable — which also means
      `no-preference` is the rendering never captured).
      `prefers-color-scheme`: four files, **no** visual spec. **A stylesheet
      that answers a media query no capture emulates is a branch no baseline
      has ever rendered.**
      
      **THE THIRD: A DENSITY REGISTER EXISTS AND IT IS A FIXTURE'S.** The
      cell's three densities are exactly the union in
      `primitiveBaselinesHost.tsx`; the products use `data-density` in 16
      files; and the word `density` carries **three other vocabularies** in
      this estate (`expanded`, `balanced`, `standard`, `spacious`).
      
      **WHAT IS RIGHT, TWICE:** both theme helpers refuse to plant the
      attribute from an init script — the default theme is the attribute's
      **absence** — and both verify it took, because "without this check a
      'both themes' pass can silently inspect the same theme twice". **A
      positive control, arrived at independently, and the second time by
      copying the first.**
      
      **CORRECTED BY THE SCAN:** I first wrote that `forcedColors` appears
      nowhere in the repository. It appears in eleven harnesses. The register
      now counts harness uses **separately** from spec uses, because folding
      them together would let eleven harnesses answer for a product none of
      them loads. Two file counts also tightened (21→13, 54→45) once
      spec/test files were excluded. 9 faults. -->
      
    • S11.6.c Define representative Latin/long pseudo/RTL/mixed-bidi/CJK/ fallback-font content and deterministic data/clock/assets. Evidence: locale fixture manifest. <!-- S11.6.c: shipped as libs/oshun/workbench-kit/src/locale-fixtures.ts (+ .spec.ts, 20 tests) with tools/locale-fixtures-scan.mjs (11 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/locale-fixtures, verify:operations:v1-locale-fixtures, and the locale-fixtures-scan Nx target. Entry point 216 of 216, 0 export-map violations.

      **THE VEIN: A 505-LINE BIDI LIBRARY WITH NEUTRAL PROTECTION, AND NO
      BASELINE RENDERS ONE RIGHT-TO-LEFT PIXEL.**
      `studio/i18n/rtlLayout.ts` names **ten surfaces and ten direction
      policies**, maps physical properties to logical ones, and carries
      `assertNoPartialMirroring`, `findUnprotectedNeutrals` and
      `protectNeutralIn` — **the mixed-bidi mechanism itself**, the thing that
      stops a neutral character between an English run and an Arabic one
      landing on the wrong side. Pure, with a React binding and its own spec.
      **Two product files import it. Zero baselines are captured in
      `dir="rtl"`.**
      
      **THE SECOND — THE SAME SHAPE AGAIN: THE PSEUDO-LOCALISER EXISTS TOO
      AND NOTHING IMPORTS IT.** `i18n/localizationChecks.ts` is **404 more
      lines** — `pseudoLocalize`, `isPseudoLocalized`, expansion bands, a
      per-language expansion table, first-strong isolates — with a conformance
      spec driving it and **zero product importers**. **Two libraries, 909
      lines, one spec each, two product files between them.**
      
      **A NOTE ON HOW THAT WAS FOUND, BECAUSE IT IS THIS CELL'S OWN LESSON.**
      The register first said pseudo-localisation **did not exist**: a grep for
      `pseudo` that excluded CSS pseudo-elements returned the BFF's redaction
      **pseudonyms**, and reading the head of that list was enough to conclude
      there was nothing there. **The scan's positive control — "the word is
      present, so the zero is not a broken query" — caught it.** *A zero from
      a query nobody positive-controlled is not a measurement.*
      
      **THE THIRD — DETERMINISM: THERE IS ONE CLOCK FREEZE IN THE ESTATE AND
      IT IS A PRIVATE FUNCTION IN ONE SPEC.** `freezeTime` is declared at line
      308 of `shell-first-impression-visual.spec.ts`, called once, **exported
      nowhere**; the other five visual specs do not freeze time. Data is
      stubbed however each spec felt like (7 route stubs in one, none in
      another); **nothing pins a font or an image**. A baseline is a claim
      that the same inputs produce the same pixels, and each spec makes that
      claim with its own arrangement.
      
      **THE READING:** 1 of 6 content shapes captured, **6 of 6 supported**;
      2 of 3 determinism controls arranged and **0 shared**. 8 faults. -->
      
    • S11.6.d Capture shell/workspace/inspector/status plus loading/empty/ error/degraded/conflict/destructive/complex visualization states. Evidence: state baseline index. <!-- S11.6.d: shipped as libs/oshun/workbench-kit/src/state-baselines.ts (+ .spec.ts, 20 tests) with tools/state-baselines-scan.mjs (8 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/state-baselines, verify:operations:v1-state-baselines, and the state-baselines-scan Nx target. Entry point 217 of 217, 0 export-map violations.

      **THE VEIN: A BASELINE SUITE PHOTOGRAPHS THE STATE THE PRODUCT IS IN
      WHEN NOTHING IS WRONG.** 74 unique baselines; reading their own
      filenames — **the only place a capture says what it captured** — **five
      name a state** (`page-404-desktop`, `page-error-desktop`,
      `page-home-loading-desktop`, `page-loading-skeleton-desktop`,
      `page-explore-offline-mobile`). The other **69 are a surface at rest**.
      Those are worth having, and they are **the state a regression is least
      likely to reach first**.
      
      **THE SECOND: ONE OF THE FOUR SURFACES THE CELL NAMES HAS NO BASELINE AT
      ALL.** shell 11, workspace 2, inspector 1, **`status` none** — and the
      cell asks for it because a status region is where a degraded or
      conflicted system says so.
      
      **THE THIRD — WHY THIS IS AN INDEX AND NOT A COUNT: THE MISSING STATES
      ARE THE DEAR ONES.** Loading, error and degraded come from a **route
      stub**, which these specs already write. Empty needs a deliberately
      empty fixture. **Conflict needs two writers and a version that lost.**
      **Destructive needs an irreversible act that must not happen.** A
      complex visualisation is only complex with enough data to be. **Every
      missing state needs a built fixture; every present state is one a route
      stub produces** — a fact about the cost of the capture, not about
      anybody's diligence.
      
      **THE READING: 7 of 32 index cells filled (3 of 4 surfaces, 5 of 8
      states), 0 unfilled states a route stub could produce.** The scan
      rebuilds the index from the committed filenames, because a title can say
      "loading state" and photograph a settled page. 5 faults. -->
      
    • S11.6.e Mask only proven nondeterministic data and retain layout/content areas so regressions cannot be hidden. Evidence: mask review. <!-- S11.6.e: shipped as libs/oshun/workbench-kit/src/mask-review.ts (+ .spec.ts, 19 tests) with tools/mask-review-scan.mjs (9 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/mask-review, verify:operations:v1-mask-review, and the mask-review-scan Nx target. Entry point 218 of 218, 0 export-map violations.

      **THE VEIN: THE MASK IS THE PART THAT IS REVIEWED AND THE TOLERANCE IS
      THE PART THAT IS NOT.** The estate masks, and it masks **well**: two of
      six specs pass a `mask`, both through a named helper, both saying what
      they cover and why. `unstableRegions` is headed **"Regions whose CHANGE
      is not a regression"** with a comment per selector — the greeting says
      "Good morning" or "Good evening" without anything having regressed; the
      diagnostics blocks are ids and counters *put behind a disclosure
      precisely because they are not member copy*; the account chip carries a
      seeded name encoding a timestamp. That is exactly the cell's ask, and a
      mask **keeps the layout**.
      
      **AND THEN THERE ARE 91 PIXEL TOLERANCES CARRYING SEVEN DIFFERENT
      NUMBERS, NONE OF THEM NAMED** — 0.02 (35×), 0.03 (18×), 0.01 (17×),
      0.001 (16×), 0.004, 0.015, 0.05. A tolerance is **unbounded in
      location**: it forgives a difference of any nature anywhere in the
      capture as long as it is small, so it **cannot be reviewed region by
      region**. The estate *has* a named tier register — `VISUAL_THRESHOLD`,
      three tiers with a comment each — as a `const` inside one spec **of
      another domain**, reachable by no V1 spec.
      
      **THE SECOND, WHICH THE MATRIX MAKES VISIBLE: THE TOLERANCE IS WHAT EACH
      SPEC PAYS FOR THE CONTROLS IT DOES NOT HAVE.** No two of the six control
      the same things. `tables-and-details` pins viewport, locale, `timezoneId:
      'UTC'` and reduced motion, disables animations, **awaits
      `document.fonts.ready` so baseline and replay paint with the same
      font**, and masks a volatile-selector register — and needs **one**
      tolerance. `design-system` controls **nothing** and carries **20**.
      `visual-regression-pages` controls one thing and carries **47**. Not
      diligence — arithmetic: every control removed is bought back as slack,
      and **slack is the suppression nobody can review**.
      
      **TRAP, RECORDED IN THE SCAN'S OWN HEADER:** `git grep --
      'apps/oshun/*/e2e'` **matches nothing** in this repository and returns
      **zero rather than an error**. An earlier draft of this register
      concluded from those zeros that the estate masked nothing, set no
      tolerance and disabled no animation — **the exact opposite of the
      truth**. Every path in the scan is now a literal directory, and the
      controls positive-check each count. 4 faults. -->
      
    • S11.6.f Review at intended pixel/structural thresholds for overflow, clipping, overlap, focus, contrast, semantic state, card/chrome drift, and font/icon change. Evidence: signed baseline matrix. <!-- S11.6.f: shipped as libs/oshun/workbench-kit/src/review-dimensions.ts (+ .spec.ts, 20 tests) with tools/review-dimensions-scan.mjs (8 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/review-dimensions, verify:operations:v1-review-dimensions, and the review-dimensions-scan Nx target. Entry point 219 of 219, 0 export-map violations.

      **THE VEIN: A PIXEL DIFF NOTICES A DIFFERENCE AND CANNOT NAME ITS
      KIND.** The eight are not eight of a kind:
      
      **Two are a screenshot's own** — card/chrome drift and font/icon change
      have **no structural signature**: same box, same roles, same geometry,
      different picture. The 74 baselines are exactly the right instrument.
      
      **Three are geometry** — overflow is `scrollWidth > clientWidth`,
      clipping is a box outside its container, overlap is two boxes
      intersecting. Each is a **decision that arrives with the number that
      made it**, which a diff image never has. **The estate agrees**: its
      visual-QA register carries an `evidenceKind` field and **both claims it
      marks `playwright-layout-assertion` are about horizontal spill**, with a
      comment saying to use them "only for viewport-fit claims whose primary
      evidence is measured geometry and overflow checks". The answer exists
      and is applied to **one geometric dimension of three**, while **397
      suite files already touch a geometry API** — so choosing a picture over
      a measurement is a choice, not a capability gap.
      
      **Two have a better instrument elsewhere, pointed away** — contrast is
      axe's `color-contrast` (S11.3.f: aimed at the primitive gallery); focus
      is S11.4.a's (**one of seven claims established**).
      
      **And one is invisible to a screenshot entirely** — `aria-pressed`
      flipping changes **no pixel** unless the styling depends on it, and a
      semantic state that no longer matches the visual one is exactly the
      defect worth catching. **No threshold, however low, brings it into a
      picture.**
      
      **WHAT THAT MEANS FOR THE SIGNATURE:** a signed baseline matrix claims
      somebody looked at eight things; six are decided by an instrument other
      than the picture, four of those exist and point elsewhere, and **a
      signature over the pixels alone covers two**. 7 faults. -->
      
    • S11.6.g Require named review/update reason and linked implementation for baseline changes. Evidence: visual approval gate. <!-- S11.6.g: shipped as libs/oshun/workbench-kit/src/visual-approval-gate.ts (+ .spec.ts, 19 tests) with tools/visual-approval-gate-scan.mjs (9 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/visual-approval-gate, verify:operations:v1-visual-approval-gate, and the visual-approval-gate-scan Nx target. Entry point 220 of 220, 0 export-map violations.

      **THE VEIN: A GATE THAT CANNOT FAIL IS NOT A GATE.** **Nothing in this
      repository gates a baseline change** — not the pre-commit hook (which
      locks V2 binary assets and ratchets six conventions and never mentions a
      `.png`), not any workflow. And the one CI job named **Visual
      Regression** is inert **five times over**, each confirmed against the
      file that carries it:
      
      1. **`|| true`** — `pnpm nx run-many --target=test:visual --all || true`
         discards the exit code, so a failing comparison passes.
      2. **The target does not exist** — `test:visual` is declared in **zero**
         `project.json` files; the one `package.json` with it belongs to
         another domain's admin app.
      3. **None of the six V1 visual specs is behind it** — they run under the
         `e2e` targets, so even a real `test:visual` would not reach them.
      4. **The artifact upload is `if: failure()`** — a job that cannot fail
         never uploads, so **a human cannot review what a human is not
         shown**.
      5. **The globs do not match the baselines** — it collects
         `__snapshots__` and `visual-diff`; Playwright writes
         `<spec>.spec.ts-snapshots`. The **34 `__snapshots__` directories that
         do exist hold vitest text snapshots**, and `visual-diff` exists
         nowhere.
      
      **Three of the five would each stop the job alone, and that count is the
      deliverable:** a reader who sees the `|| true` and deletes it will
      conclude the gate now works, and it will still be running a target
      nothing declares over specs that are not behind it.
      
      **AND THE SAME JOB IS WHY S11.6.a's PLATFORM SPLIT MATTERS:** it runs on
      `ubuntu-latest`, and **16 of 83** committed baselines carry that stamp —
      so even a working comparison would find no baseline for most captures
      and **write one rather than fail**.
      
      **WHAT A GATE WOULD HAVE TO DO** is recorded rather than implemented:
      notice a baseline change, require a reason, require a linked
      implementation, and **fail** when either is missing — **three of the
      four decidable from a diff and a commit message**, and the fourth not a
      measurement at all but a decision to exit non-zero, which is exactly the
      one the existing job discards. 4 faults. -->
      
  • S11.7 Add deterministic clock, UUID, provider, queue, and storage test adapters. Evidence: no production adapter imports test fixtures.

    <!-- S11.7 LEFT UNCHECKED ON PURPOSE. All seven children are `[x]`, and
    unlike its neighbours this family **built things**: five working adapters,
    not five registers — `clock-adapter`, `id-adapter`, `provider-adapter`,
    `queue-adapter`, `storage-adapter`, each with its own suite driving
    behaviour, plus `adapter-contracts` (the measurement) and
    `test-only-boundary` (the enforcement).
    
    **The parent is unchecked because its evidence line is a claim nothing
    can currently make.** "No production adapter imports test fixtures" is
    S11.7.g's subject, and S11.7.g measured: **six libraries publish a
    `./testing` subpath and one of them is banned**; the estate's rule is
    exemplary and **scoped to one package**; and the only bundle check
    **weighs** the artifact rather than reading it, so **a budget in kilobytes
    is asked to answer "did the forgery kit ship"**. Five of the six
    publishers could be imported from production today and nothing would
    say so.
    
    **AND THE ADAPTERS ARE BUILT, NOT WIRED.** They are contract subpaths of
    the kit; the products still take `now` in **twelve incompatible shapes**
    (784 declarations), still call `randomUUID` **179 times** in the BFF with
    no seam, and still have no queue port. The one subject that already had a
    real port — object storage, with `AssetObjectStore`, a factory and
    `assetResolverFrom(download)` — is the shape the other four lack, and it
    is the estate's own work rather than this family's.
    
    What would close this parent is a wiring and a rule: route the products'
    `now` and id calls through the ports, extend the S1.13.e ban to the other
    five `./testing` publishers, and add a scan that **reads** a built
    artifact rather than weighing it. Three of those are mechanical; the
    fourth — deciding which shape of `now` every call site should take — is
    the one that needs a person. -->
    
    • S11.7.a Define production interfaces and explicit test-only entry points for clock/time zone/timer, stable/random ID, provider, queue/worker, and blob/object storage. Evidence: adapter contracts. <!-- S11.7.a: shipped as libs/oshun/workbench-kit/src/adapter-contracts.ts (+ .spec.ts, 20 tests) with tools/adapter-contracts-scan.mjs (10 negative controls, exits 2 if any goes undetected), registered as @oshun/workbench-kit/adapter-contracts, verify:operations:v1-adapter-contracts, and the adapter-contracts-scan Nx target. Entry point 221 of 221, 0 export-map violations.

      **THE VEIN: THE SEAM IS EVERYWHERE AND THE CONTRACT IS NOWHERE.** The
      estate injects `now` **784 times** across the oshun and shared code, in
      **twelve distinct shapes** — three value types (`string`, `Date`,
      `number`) × two forms (a value or a getter) × two arities (required or
      optional), and **every one of the twelve is in use**: `now: string` 123,
      `now: Date` 105, `now?: () => Date` 102, `now: () => Date` 102, down to
      `now?: number` 4. **Not a missing seam — a seam at almost every call
      site with no contract behind any of them**, so a controllable clock has
      to be written once per shape, and the cell's evidence line ("no
      production adapter imports test fixtures") **cannot even be asked**,
      because there is no adapter to ask it about.
      
      **AND THERE ARE NAMES, PLURAL, FOR THE SAME THING:** `ai-platform`
      exports an `interface Clock`; `creative-autonomy` exports **`type Clock
      = () => number` AND `type NowFn = () => number` in the same domain**.
      
      **THE SECOND HALF IS ALREADY SOLVED ELSEWHERE:** **six libraries declare
      a `./testing` subpath** in their own `package.json` — exactly the
      explicit test-only entry point the cell asks for, enforceable by the
      module boundaries `eslint.config.js` already declares. **None of the
      five subjects has one.**
      
      **THE THIRD — THE ESTATE'S OWN ANSWER AGAIN: ONE OF THE FIVE ALREADY HAS
      THE PORT.** `apps/oshun/bff/src/assets/asset-store.ts` declares
      `AssetResolver`, an `AssetObjectStore extends AssetResolver`, a
      `resolveAssetObjectStore` factory, and — the part that matters —
      **`assetResolverFrom(download)`, a constructor that takes the one
      function a double has to supply**; a `GeneratedArtifactObjectStoreFactory`
      and an integration test sit beside it. Built, **for one subject of
      five**, while the BFF calls `randomUUID` **179 times** with no seam at
      all.
      
      **A NOTE ON HOW THAT WAS FOUND, BECAUSE IT IS THE SAME LESSON TWICE:**
      this register first said the BFF touched **no** object storage, because
      a shell loop handed an escaped-alternation pattern to `git grep -lE` and
      got zero. **The scan's positive control caught it** — exactly as it
      caught the pseudo-localiser in S11.6.c. *A zero from a query nobody
      positive-controlled is not a measurement.* 10 faults. -->
      
    • S11.7.b Implement controllable wall/monotonic clock, timers, DST/time zone, advance/run-pending, and timeout/backoff semantics. Evidence: clock adapter tests. <!-- S11.7.b: shipped as libs/oshun/workbench-kit/src/clock-adapter.ts — a working implementation, not a register — with .spec.ts (22 tests driving behaviour) and tools/clock-adapter-scan.mjs (10 negative controls), registered as @oshun/workbench-kit/clock-adapter, verify:operations:v1-clock-adapter, and the clock-adapter-scan Nx target. Entry point 222 of 222, 0 export-map violations.

      **THE RULE THAT DECIDES THE DESIGN: A WALL CLOCK AND A MONOTONIC CLOCK
      ARE TWO CLOCKS, AND A TIMEOUT MEASURED ON THE WRONG ONE IS NOT A
      TIMEOUT.** Wall time is what a user sees and **can move backwards** — an
      NTP correction, a manual change, a VM resuming from a snapshot;
      monotonic time cannot. So a **duration** is measured on the monotonic
      clock, a **deadline** is displayed on the wall clock, `setWallClock`
      moves wall time alone and `advance` moves both — **so a test can produce
      the NTP step that breaks a naive timeout**, which is the only way anyone
      finds one. Driven end to end: after a backwards step of an hour and six
      seconds of real time, the monotonic deadline **has** expired and the
      wall comparison still has **3,599,000 ms** to go.
      
      **THE SECOND: ADDING 24 HOURS IS NOT ADDING A DAY.** A physical day is
      23 or 25 hours twice a year, so `localParts` answers local wall time in
      a named IANA zone through `Intl.DateTimeFormat` — the platform's own
      tables — and the suite advances across the **real Europe/London
      boundary** (2026-03-29) reading the offset go 0 → 60. **A scan control
      fails if this platform's tables do not put a change there**, because a
      DST test on a stale ICU build asserts nothing while passing.
      
      **THE THIRD: `advance` MUST BE EQUIVALENT TO TIME PASSING.** A timer
      scheduled by a callback running during the advance is due inside the
      same window if its deadline falls there; timers run in due order, ties
      by insertion; and **the clock stops at each timer's own due time** so a
      callback reading the clock sees the instant it was scheduled for, not
      the destination.
      
      **DECISIONS.** **Jitter is a parameter, not a call to a random source** —
      a backoff schedule a test cannot reproduce is one nobody can assert
      about — and a ratio outside `[0,1)` is **refused rather than clamped**.
      **Nothing patches a global**: a port is taken as an argument, not
      installed, and the scan checks that the module touches neither
      `setTimeout` nor `Date`. -->
      
    • S11.7.c Implement seeded/scripted UUID/ID sequences, collision/exhaustion/ invalid generation, and no production predictability. Evidence: ID adapter tests. <!-- S11.7.c: shipped as libs/oshun/workbench-kit/src/id-adapter.ts — a working implementation — with .spec.ts (20 tests) and tools/id-adapter-scan.mjs (11 negative controls), registered as @oshun/workbench-kit/id-adapter, verify:operations:v1-id-adapter, and the id-adapter-scan Nx target. Entry point 223 of 223, 0 export-map violations.

      **THE RULE: A PREDICTABLE IDENTIFIER THAT REACHES PRODUCTION MUST SAY
      SO.** A seeded id generator is a security defect wearing a test double's
      clothes — a guessable session id, invite token or object key is the whole
      game — and "it only exists in tests" is a claim about a wiring nobody
      checks. So it is answered twice: **every identifier is marked** (they
      begin `7e57`, hex for "test"), with `assertNotATestId` as a fail-loud
      seam for the paths that **mint** an id; and **the odds of a real UUID
      carrying the marker by chance are stated — 1 in 65,536** — because a
      marker treated as a proof is a false negative waiting to happen.
      
      **AND THE MARKED IDS ARE STILL VALID UUIDs** — version nibble 4, variant
      bits `10xx`, the right dashes. A double producing `id-1` would pass
      every test and fail against a `uuid` column, a parser or a router,
      **hiding exactly the bugs it exists to find**.
      
      **A SCRIPT THAT RUNS OUT HAS TO SAY SO** — exhaustion throws rather than
      wrapping (a silent collision) or returning `undefined` (a failure three
      layers away). **And a collision is something a test wants**: a script
      may repeat, so `scriptedCollisions` **reports** the repeats for the
      author rather than refusing them.
      
      **DETERMINISM WITHOUT `Math.random`:** SplitMix64 over a 64-bit state,
      written out in eight lines of `BigInt` so the sequence is **pinned to
      this file** and cannot move under a dependency bump — a seeded generator
      whose output moves is worse than none.
      
      The scan pins a **golden vector** for seed 42, runs **10,000
      identifiers** checking distinctness, shape and marker, checks all four
      legal variant nibbles appear (so the shape is not one lucky path), and
      confirms the module reaches for no real randomness — **a check that
      caught its own comment-stripper reporting a single-line JSDoc that
      *names* `crypto.randomUUID` in prose**. -->
      
    • S11.7.d Implement provider scripts for chunks/results/usage/cost/safety/ errors/delays/cancel/unknown outcome with exact request capture. Evidence: provider adapter tests. <!-- S11.7.d: shipped as libs/oshun/workbench-kit/src/provider-adapter.ts — a working implementation — with .spec.ts (17 tests) and tools/provider-adapter-scan.mjs (12 negative controls), registered as @oshun/workbench-kit/provider-adapter, verify:operations:v1-provider-adapter, and the provider-adapter-scan Nx target. Entry point 224 of 224, 0 export-map violations.

      **THE RULE: AN UNKNOWN OUTCOME IS A THIRD RESULT, NOT AN ERROR.** A call
      can succeed, fail visibly, or **have the connection drop after the
      request was accepted** — the work may have happened, been billed and
      been recorded upstream, and this side cannot tell. A double that models
      that as an error teaches every consumer to treat it as "it did not
      happen", **and the consumer then retries a generation that already
      completed**.
      
      **WHAT THE ESTATE ALREADY HAS, AND WHERE IT STOPS.**
      `libs/iris/agents/core` declares a real port — `AgenticProvider` with
      `complete(request, signal?: AbortSignal)` — so **cancellation is already
      in the contract**, and the assistant factory takes an env documented
      "injectable for testing". Two things it cannot express: its taxonomy is
      **`PROVIDER_ERROR | UNKNOWN_TOOL | DUPLICATE_TOOL | INVALID_REQUEST`**
      and **none of them is the ambiguous outcome**, so a dropped connection
      arrives as a plain provider error and **reads as safe to retry — the bug
      arriving through the type system**; and `Promise<AgenticCompletion>` is
      all-or-nothing, with no way to say "four chunks arrived and then the
      connection died", so **the partial output somebody may already have
      shown a user is not in the result at all**.
      
      **THE SECOND: A COST IS A MEASUREMENT TIMES A PRICE, AND A MISSING PRICE
      IS A REFUSAL.** `costOf` throws rather than returning zero — *a cost of
      zero is a number a dashboard will happily add up, and the sum of a
      hundred unpriced calls is a confident nought*. `totalCost` refuses a
      call with no usage rather than skipping it. And **an unknown outcome
      reports no usage by construction**, so it cannot be priced at all.
      
      **THE THIRD: EXACT REQUEST CAPTURE MEANS EXACT** — recorded unmodified
      and undefaulted, so a test can assert on the field somebody added by
      accident and the prompt that grew a trailing newline.
      
      **Delays are advanced on S11.7.b's clock** rather than slept, and the
      capture keeps **how many chunks were delivered before a cancel**. -->
      
    • S11.7.e Implement queue scripts for delivery/order/duplicate/lease/ heartbeat/retry/cancel/dead-letter/worker loss and observable state. Evidence: queue adapter tests. <!-- S11.7.e: shipped as libs/oshun/workbench-kit/src/queue-adapter.ts — a working implementation — with .spec.ts (16 tests) and tools/queue-adapter-scan.mjs (7 negative controls), registered as @oshun/workbench-kit/queue-adapter, verify:operations:v1-queue-adapter, and the queue-adapter-scan Nx target. Entry point 225 of 225, 0 export-map violations.

      **THE RULE: A LEASE IS A CLAIM WITH AN EXPIRY, AND A QUEUE CANNOT TELL A
      DEAD WORKER FROM A SLOW ONE.** Everything difficult follows from it: when
      a lease runs out the queue does not know whether the worker crashed or
      is merely slower than somebody guessed, so it **redelivers** — same
      message id, incremented attempt, **and nothing said to the worker that
      had it**. That is at-least-once delivery; a double that cannot produce
      it cannot test the only part of the consumer that matters.
      
      **THE SECOND, AND IT IS THE ONE THAT BITES: A DEAD-LETTER POLICY COUNTS
      DELIVERIES, NOT FAILURES.** A message whose worker dies five times **was
      never nacked once** — nobody reported anything — so a policy keyed on
      failures never fires and the message circulates for ever while every
      dashboard shows zero errors. The scan drives that sequence end to end:
      **3 deliveries, 0 rejections, 1 dead-letter.**
      
      **THE THIRD: A HEARTBEAT EXTENDS A CLAIM, IT DOES NOT DISCHARGE ONE** —
      and it is capped from the moment the claim was granted, because without
      the cap **the only difference between a stuck worker and a careful one
      is patience**.
      
      **AND ORDER IS PER PARTITION, NEVER GLOBAL** — a partition with a message
      in flight offers nothing, so its order holds across a redelivery while
      different partitions interleave; the scan's second sequence produces
      exactly the reordering a global-order consumer would get wrong.
      
      Time is S11.7.b's clock; an acknowledgement from an expired lease
      **throws** rather than silently succeeding. -->
      
    • S11.7.f Implement storage scripts for bytes/version/hash/range/latency/ partial/corruption/eventual consistency/permission/failure. Evidence: storage adapter tests. <!-- S11.7.f: shipped as libs/oshun/workbench-kit/src/storage-adapter.ts — a working implementation — with .spec.ts (18 tests) and tools/storage-adapter-scan.mjs (9 negative controls), registered as @oshun/workbench-kit/storage-adapter, verify:operations:v1-storage-adapter, and the storage-adapter-scan Nx target. Entry point 226 of 226, 0 export-map violations.

      **THE RULE: A WRITE THAT SUCCEEDED IS NOT A WRITE YOU CAN READ.** An
      eventually consistent store answers a `put` with a success and then
      answers a `get` with the **previous** version — or nothing, if the object
      is new. Every consumer that writes and immediately reads back is either
      correct about that or **has a bug that appears under load and nowhere
      else**. `setConsistency` opens the window, `advanceReplication` closes
      it; the scan drives a consumer straight into it.
      
      **THE SECOND: A STORE THAT CANNOT LIE CANNOT TEST A CHECKSUM.**
      `corrupt()` returns **altered bytes with the original hash** — exactly
      what a silently damaged object looks like, and exactly what a consumer
      comparing them will catch. A double that always returns what it was
      given lets a consumer skip verification for ever and pass every test.
      
      **THE THIRD: PARTIAL AND RANGE ARE DIFFERENT THINGS AND ONLY ONE IS AN
      ERROR.** A **range** read is a slice the caller asked for; a **partial**
      read is a **prefix of the truth with nothing saying so**. Modelling them
      alike teaches a consumer to treat a truncated body as a successful short
      read — *which is how a half-written file gets committed*.
      
      **THE HASH IS FNV-1a OVER 64 BITS AND IS NOT CRYPTOGRAPHIC** — written
      out so its values are pinned to this file, detecting the accidental
      corruption a store produces and nothing an adversary produces. The
      module says so and **the scan checks that it still does**, because a
      test-double digest quietly used as an integrity guarantee is a defect
      with no symptom. Latency is advanced on S11.7.b's clock. -->
      
    • S11.7.g Enforce dependency rules and production bundle scan so no app/ production adapter imports test builders/fixtures. Evidence: deliberate invalid import and bundle tests. <!-- S11.7.g: shipped as libs/oshun/workbench-kit/src/test-only-boundary.ts (+ .spec.ts, 17 tests) with tools/test-only-boundary-scan.mjs (9 negative controls), registered as @oshun/workbench-kit/test-only-boundary, verify:operations:v1-test-only-boundary, and the test-only-boundary-scan Nx target. Entry point 227 of 227, 0 export-map violations.

      **THE VEIN: A DEPENDENCY RULE READS WHAT SOMEBODY WROTE, AND A BUNDLE
      SCAN READS WHAT ARRIVED.** A `no-restricted-imports` rule matches the
      **specifier at an import site** — exactly right for a developer typing a
      test-only path into production, and **blind to the chain where nobody
      types it**: a production module imports a barrel, the barrel re-exports
      a helper, the helper imports the builders, and every site is spelled
      legally.
      
      **WHAT THE ESTATE ALREADY HAS IS EXEMPLARY.** S1.13.e bans
      `@oshun/workbench-kit/testing*` repo-wide, adds a **second rule for the
      relative `../testing/` spelling** the specifier pattern cannot see,
      explains why it uses the `@typescript-eslint` rule rather than the core
      one (flat config resolves by last match), and — the part worth copying —
      **names what it does not cover**: spec files are globally unlinted, so
      the ban rests on "every linted file is production source" rather than on
      a carve-out. **And it is proved**:
      `tools/test-only-import-fixtures.mjs` writes a real TypeScript file,
      runs the repository's own ESLint, and checks **both halves** — *"a rule
      that also refuses the allowed spelling is not a rule"*. That **is** the
      deliberate invalid import this cell asks for.
      
      **WHERE IT STOPS IS THE MEASUREMENT: SIX libraries publish a `./testing`
      subpath and ONE is banned.** `shared/live-media`, `calliope/stage`,
      `isis/comfyui-sdk`, `yemaya/agents` and `yemaya/study-workspace` each
      publish a test-only entry point with nothing stopping a production
      import. The rule is not weak — it is **scoped to one package while the
      convention is used by six**.
      
      **AND THE BUNDLE HALF DOES NOT EXIST.** `oshun-web-bundle-monitor.yml`
      **weighs** the bundle — size analysis, a budget, a failure if over — and
      **a budget is not a content scan**: a test builder that adds four
      kilobytes passes comfortably, and *"did the forgery kit ship"* is not a
      question a number in kilobytes answers.
      
      **DECISION:** this scan deliberately **does not re-run the existing
      prover** — a second tool making the same deliberate invalid import would
      be **a copy claiming to be a check**. 7 faults. -->
      
  • S11.8 Add real-database isolation and migration harnesses. Evidence: multi-tenant round-trip tests. <!-- S11.8 parent — a–g all closed. Seven kit modules, seven scans, 160 unit tests, five registrations each → 234 entry points, 0 violations. Every scan verified green in one sweep on 2026-08-15, and every one of them follows the same rule: NO SERVER IS EXIT 1, because an unmeasurable fact and an absent one produce the same silence.

THE PARENT'S EVIDENCE — a multi-tenant round-trip against a real database — EXISTS AND RAN. tools/isolation-paths-scan.mjs applies the 35 migrations to a fresh PostgreSQL 16.14 database, puts ONE natural key under TWO tenants, and drives write → read → search → history → event through the real ContractPersistenceService. Four legs read tenant-scoped. A deliberately tenant-blind reader over the same rows reads leaks, so the fixture is known to be capable of failing.

AND THE WRITE LEG DOES NOT COMPLETE, WHICH IS THE HARNESS WORKING. put() looks the record up by @@unique([tenantId, contractSchema, sourceRecordId]) and then creates the row with id taken from the CONTRACT PAYLOAD, and id String @id is a primary key that is not tenant-scoped — so the second tenant is refused with P2002, and the scan had to build the collision underneath the repository to continue. DurationBucketSchema.id is a five-value enum, so the first tenant to persist a duration bucket blocks every other tenant for ever. 91 of 147 models carry the shape. That defect is the reason this section asks for these harnesses and it was invisible to every single-tenant fixture in the estate.

WHAT THE SEVEN CELLS ESTABLISHED, each measured rather than argued:

.a the production bootstrap leaves 24 tables, 24 keys, 110 indexes and ZERO foreign keys, check constraints, policies and triggers .b the 35 migrations do not build the datamodel the client is generated from (drift exit 2, 73 statements, 17 of 147 tables), one model no migration creates, and all three migrate diff invocations pass flags Prisma 7 removed .c no fixture in the estate has ever put one natural key under two tenants, and 99 tenant-scoped compound keys make it legal by design .d 5 of the 9 paths scope by tenant; the write path leaks; and 5 of the 8 database-backed integration suites FAIL the moment a database is present .e eight concurrent save() calls ended at version 1 — seven lost updates, and the version says one save happened .f the migration files hold 0 destructive statements and the script the deploy applies holds 16, in 0 transactions, with no down migration .g 294 live EXPLAIN checks, taken with sequential scans switched off; the narrowing is real and — measured — currently costs nothing

FOUR TIMES A SCAN REFUTED ITS OWN REGISTER and the register was corrected: the 11 double-produced tables AGREE exactly; readEvent is NOT an escalation; the destructive count is 16 and not 18; and the planner settings change no plan today. A register nothing can contradict is not a measurement.

TRAP, twice in one file: git grep READS THE INDEX, so this cell's own module, spec, scan AND the export-map register — which is a register OF SOURCE TEXT and matches every term any entry discusses — all began matching pg_locks the moment they were committed, taking the count 1 → 4 → 2. A REGISTER THAT COUNTS ITS OWN FAMILY INVENTS A DEFECT. All four are excluded by path, and the absent-term control is now assembled at runtime so its literal never exists in the file that searches for it. -->

  • S11.8.a Provision supported database version/extensions in isolated schema/database/container with production constraints/policies/triggers and deterministic cleanup. Evidence: harness setup tests. <!-- S11.8.a — libs/oshun/workbench-kit/src/database-harness.ts (+ .spec.ts, 14 tests) and tools/database-harness-scan.mjs (7 negative controls; exit 2 if any goes undetected; NO SERVER IS EXIT 1, NEVER A PASS).

THE RULE: A HARNESS THAT PROVISIONS LESS THAN PRODUCTION IS TESTING SOMETHING ELSE, AND ONE THAT CLAIMS MORE IS WORSE.

The cell asks for "production constraints/policies/triggers". A harness can only supply those if production HAS them, so the deliverable is not a list of things to create — it is an introspection of what the production bootstrap actually leaves behind. The scan creates oshun_harness_<run>, points the BFF's own tryConnectPostgres() at it via OSHUN_STUBS_DATABASE_URL so the PRODUCTION bootstrapSchema() runs — the DDL itself, not a copy of it — reads pg_catalog, and drops the database.

THE READING, taken live from PostgreSQL 16.14 on 2026-08-15:

tables 24 · primary keys 24 · indexes 110 · unique constraints 1
FOREIGN KEYS 0 · CHECK CONSTRAINTS 0 · POLICIES 0 · TRIGGERS 0
tables with row security 0 · extensions: plpgsql

So a harness that faithfully provisions production's constraints, policies and triggers provisions ALMOST NONE OF THEM — the cell's requirement is satisfied by supplying nothing, and that is a fact about the estate rather than about the harness. Concretely: goal3_stub_habit_entries.habit_id is TEXT NOT NULL with no REFERENCES, so a child row can name a parent that never existed and the server will store it; and with zero row policies, tenant isolation — the evidence S11.8's parent asks for — is enforced by whatever each query remembers to write, nowhere else. Four findings recorded: no-foreign-keys, no-row-policies, no-triggers, no-check-constraints.

THE SECOND: ISOLATION IS A PROPERTY OF THE CONTAINER, NOT OF THE CLEANUP. A harness that truncates tables at the end shares a namespace with everything else on the server: a parallel run sees half-deleted rows, and a failed run leaves them for the next one. A separate DATABASE makes the isolation structural and the cleanup one statement that cannot partially succeed. assertDroppable() throws on any name outside oshun_harness_* — the cleanup path of a test harness is the one place a DROP DATABASE gets written with a variable in it, and this server also holds oshun_dev and nine databases belonging to other work.

THE THIRD: the kit takes a PORT, not a driver. Sql is one function — send this text, get rows — so the 14 unit tests drive a recorded double (no server needed to check the naming, the guard, the catalogue queries and the faults) and the scan drives the real thing. AND AN ABSENT SERVER AND A CLEAN SCHEMA PRODUCE THE SAME SILENCE, which is why no server is a failure and not a pass.

Registered five ways (exports, tsconfig path, root script verify:operations:v1-database-harness, Nx target database-harness-scan, export-map register) → 228 entry points, 0 violations; public-api.spec.ts 237 passing. -->

  • S11.8.b Apply migrations from empty, prior supported versions, and production-like snapshots; verify generated clients/schema metadata. Evidence: migration paths. <!-- S11.8.b — libs/oshun/workbench-kit/src/migration-paths.ts (+ .spec.ts, 27 tests) and tools/migration-paths-scan.mjs (8 negative controls; re-measures the drift against a live server; NO SERVER IS EXIT 1).

THE RULE: A MIGRATION PATH IS A MECHANISM PLUS A STARTING STATE, AND TWO PATHS THAT END AT THE SAME SCHEMA FROM DIFFERENT STARTS ARE NOT ONE PATH.

Four paths exist and they are different machines:

ci-migrate-deploy oshun-ci.yml:335 ordered files FROM EMPTY records deploy-migrate-diff deploy.sh:166 structural diff FROM LIVE no record database-yml-drift database.yml:177 structural diff FROM EMPTY ✗ not V1 bff-bootstrap domain-stubs-pg.ts:394 if-not-exists FROM LIVE no record

So the 35 migration files are exercised in CI and NEVER EXECUTED IN PRODUCTION: production's schema is whatever migrate diff synthesised, and no _prisma_migrations row exists there to say which version a database is at. The deploy's own comment says why — the datamodel declares no datasource url — which makes it a considered decision and does not change what it costs. Of the three starting states the cell names, ONE (empty) is exercised by anything.

MEASURED, not asserted. Applying all 35 migrations to a fresh PostgreSQL 16.14 database (35 recorded, 147 tables for 147 models) and diffing the result against schema.prisma:

exit 2 · 73 statements · 17 of 147 tables · 25 SET DATA TYPE
2 foreign keys dropped and NEVER RECREATED
  (decision_supersedes_fkey, ori_presence_inference_usage_fkey)
1 CREATE TABLE — v1_cross_cutting_gold_set_entry

The SET DATA TYPE block is real drift, not a Prisma-version artifact: schema.prisma carries zero @db.Timestamptz annotations, four migrations write TIMESTAMPTZ, and the live column is timestamp with time zone.

v1_cross_cutting_gold_set_entry is a model at schema.prisma:4584 that no migration mentions and that is absent from a fully migrated database. The generated client has a delegate for it. Its only callers are three test files, two of which skip unless OSHUN_V1_DATABASE_URL is set — and their own comment says "normal CI has no DB", so the absence is undetectable there by construction.

THE CHECK THAT WOULD HAVE CAUGHT IT EXISTS AND DOES NOT COVER THIS SCHEMA. database.yml globs the libs/<domain>/database directories in three places (destructive-op scan :104, prisma validate :151, migrate deploy + migrate diff --exit-code :164–180). Twelve schemas match. V1's is at libs/oshun/persistence — and persistence is not database.

AND ALL THREE migrate diff INVOCATIONS PASS FLAGS THE PINNED PRISMA REMOVED. The catalogue pins @prisma/client ^7.8.0; --from-url and --to-schema-datamodel are absent from both 7.0.0 and 7.8.0 (verified by asking the published CLI), and --from-migrations now demands a datasource.shadowDatabaseUrl no prisma.config.ts in this repository sets. The deploy's copy dies under set -euo pipefail — fail-closed, correct direction, and the migration step cannot complete. database.yml's copy is caught by || { echo "Schema drift detected in $schema"; exit 1; }, and an argument error exits 1 exactly as an error does: A CHECK THAT NEVER RAN REPORTS THE DEFECT IT WAS NEVER ABLE TO TEST FOR.

WHAT I EXPECTED AND DID NOT FIND. Eleven of the 24 tables the BFF bootstraps are also Prisma models, and CREATE TABLE IF NOT EXISTS is a no-op against a table Prisma already made — so I compared every column and type of all eleven between a BFF-only database and a fully migrated one. THEY AGREE EXACTLY. Recorded as two producers with no check, not as a defect. A related one that is real: the bootstrap adds 13 tables the datamodel does not know, so a production-shaped diff emits DROP TABLE for all 13 into psql -v ON_ERROR_STOP=1 — which has not happened only because the diff command errors first. Two defects holding each other up is not a safeguard.

Registered five ways → 229 entry points, 0 violations; public-api.spec.ts 238 passing. Every harness database dropped; the other 21 on the server untouched. -->

  • S11.8.c Provide transactional/nontransactional seed builders for two tenants with colliding IDs/data and role/ownership/protected cases. Evidence: fixture setup. <!-- S11.8.c — libs/oshun/workbench-kit/src/tenant-seeds.ts (+ .spec.ts, 36 tests) and tools/tenant-seeds-scan.mjs (9 negative controls; drives the instrument against a live server; NO SERVER IS EXIT 1).

THE RULE: A COLLIDING ID IS THE ONLY FIXTURE THAT CAN CATCH A QUERY THAT FORGOT ITS TENANT.

Give tenant A a row with id a1 and tenant B a row with id b1, and a query reading WHERE id = 'a1' with no AND tenant_id = $1 STILL RETURNS EXACTLY A'S ROW. The assertion passes. The isolation was never tested. Only when both tenants hold a row under the same natural key does the missing predicate show.

And the schema permits it BY DESIGN: 99 of the datamodel's compound keys are @@unique([tenantId, …]) or @@id([tenantId, …]), and the BFF's own goal3_stub_scim_user is PRIMARY KEY (tenant_id, external_id). A collision is the legal state the key was shaped to allow — not a violation to engineer around.

THE SECOND: TRANSACTIONAL AND NONTRANSACTIONAL ARE NOT TWO STYLES OF THE SAME BUILDER. THEY DECIDE WHO CAN SEE THE DATA. A transactional seed is private to its connection until it commits, so a subject that takes its own connection from a pool — an HTTP handler, a worker, a listener — sees an EMPTY DATABASE and passes because nothing was found, which is the same shape of pass as isolation working. canSeedFor() refuses that combination outright.

THE THIRD: the cases fail in different places, so they are separate builders. role breaks a permission check that resolves a role without the tenant (and every audit line still names the right person); ownership breaks an owner_user_id match wherever user ids also collide; protected destroys a row under legal hold AND REPORTS SUCCESS, because it did delete something.

THE INSTRUMENT REFUSES TO ANSWER FROM A NON-COLLIDING SEED. isolationVerdict returns inconclusive, with the reason, when the plan holds no collision or when the reader found nothing — because a reader returning one row proves nothing unless the other tenant had a row it could have wrongly returned.

THE SCAN PRODUCES BOTH VERDICTS AGAINST A REAL DATABASE. It seeds a real colliding pair into goal3_stub_scim_user, then runs a tenant-scoped reader and a tenant-BLIND one and requires tenant-scoped and leaks respectively. Live result: 2 rows seeded, scoped → tenant-scoped, blind → leaks, 0 rows after cleanup. If the blind reader did not leak, the fixture did not collide and every other reading would be worthless — so that is a hard failure, not a note.

THE ESTATE, AND THE SCAN CORRECTED ME TWICE. Eight database-backed suites in libs/oshun/persistence: seven name one tenant, ONE names two, ONE uses a transaction, 26 deleteMany cleanups. I had written 0 and 0.

The two-tenant suite comes within one line of this fixture and stops. contract-record-repository.integration.test.ts declares BUCKET_TENANT and SITTING_TENANT, and they are NAMESPACES FOR TWO UNRELATED SUBJECTS: taraDurationBucket sees only the first, taraSittingCompletionEvent only the second. One table sees both — oshunV1TombstoneAuditEvent, in a deleteMany({ tenantId: { in: [a, b] } }) and two per-tenant count calls — and the rows the two tenants put there are audit records OF DIFFERENT THINGS, sharing no natural key. So the register gained a second measure: delegatesWrittenUnderTwoTenants, which is 0. NAMING TWO TENANTS IS NOT PUTTING ROWS OF BOTH IN ONE TABLE.

Registered five ways → 230 entry points, 0 violations; public-api.spec.ts 239 passing. Harness database dropped. -->

  • S11.8.d Run read/write/search/history/blob ref/event/job/export/audit tenant isolation through real repositories/services. Evidence: multi-tenant tests. <!-- S11.8.d — libs/oshun/workbench-kit/src/isolation-paths.ts (+ .spec.ts, 20 tests) and tools/isolation-paths-scan.mjs (11 negative controls; drives the REAL ContractPersistenceService against a REAL database; NO SERVER IS EXIT 1).

THE RULE: A PATH WITH NO TENANT BOUNDARY CANNOT PASS A TENANT-ISOLATION TEST, AND REPORTING IT GREEN IS THE DEFECT.

Of the nine paths the cell names: 7 are served, 5 scope by tenant, and the other four are not gaps —

blob-ref, job no repository exists in the V1 estate at all audit AdminAuditEvent has NO tenantId; it keys on userId and a nullable workspaceId, and listAcrossOperators turns an empty filter into null — no restriction, by design export a subject-access bundle assembled per USER; a tenant probe against it would be measuring the wrong boundary

path-scoped-by-something-else and probeable-but-unprobed are separate fault kinds on purpose: folding them together lets an unrun probe shelter behind a legitimate design decision.

A REQUIRED PARAMETER IS NOT AN APPLIED PREDICATE. That ContractListQuery declares readonly tenantId: string says the caller must pass one and nothing about the WHERE clause. So the verdicts come from driving the real service against a real PostgreSQL 16.14 database with ONE NATURAL KEY HELD BY TWO TENANTS.

LIVE DEFECT — THE WRITE PATH LEAKS, AND NO SINGLE-TENANT FIXTURE COULD EVER SEE IT. ContractRecordRepository.put() looks the record up by @@unique([tenantId, contractSchema, sourceRecordId]) — correctly tenant-scoped — and then creates the row with id taken from the CONTRACT PAYLOAD, and id String @id is the primary key and is NOT tenant-scoped. The second tenant to write a natural key the first one used is refused with P2002. And DurationBucketSchema.id is a five-value enum (micro, short, standard, long, retreat), so this is not an unlucky coincidence:

THE FIRST TENANT TO PERSIST A DURATION BUCKET BLOCKS EVERY OTHER TENANT
FROM EVER HAVING ONE.

91 of the 147 models carry the same shape — a bare id String @id beside a tenant-scoped compound unique — and 99 contracts are registered through this repository.

The other four paths HOLD: read, search, history and event all read tenant-scoped, proved against a real collision the scan had to build UNDERNEATH the repository, because the repository itself cannot make one. A deliberately tenant-blind reader over the same rows reads leaks, so the fixture is known to be capable of failing.

A SECOND LIVE FINDING, MEASURED THE SAME WAY: FIVE OF THE EIGHT DATABASE-BACKED INTEGRATION SUITES FAIL THE MOMENT A DATABASE IS PRESENT. They skip unless OSHUN_V1_DATABASE_URL is set and their own comment says "normal CI has no DB", so nobody has run them. Against a fully migrated database:

contract-record-repository          1 failed 1 passed  ZodError — the
                                    DurationBucket contract gained
                                    defaultSessionMinutes,
                                    minimumHoursBetweenRecommendations and
                                    cadenceRationale; the fixture still has
                                    five fields
dsar-deletion-cascade               1 failed           same put path
dsar-erasure-runtime                2 failed           same put path
gold-set-entry-persistence          crashes in setup   table does not exist
gold-set-entry-store                crashes in setup   table does not exist
durable-admin-audit-events-store    1 passed
durable-memory-store                1 passed
durable-snapshot-store              5 passed

The two crashes confirm S11.8.b's v1_cross_cutting_gold_set_entry operationally: that model is not a latent gap, it breaks two suites on contact.

WHAT I LOOKED FOR AND DID NOT FIND. readEvent(eventId) is the one method in the audit store with no scope predicate — every sibling filters by userId — and it is reached over HTTP from the investigation export, resolving bookmark ids the caller chose. It looks exactly like a scope bypass. It is not one: the same caller can set an empty filter and read every operator's events through the ordinary list path, so the unscoped read grants no authority the scoped one withholds.

The scan compares the write verdict to the register in BOTH directions, so a fix trips it too and the register cannot silently go stale.

Registered five ways → 231 entry points, 0 violations; public-api.spec.ts 240 passing. Every harness database dropped. -->

  • S11.8.e Run unique/foreign/check/concurrency/idempotency/outbox/inbox/ lease/retention/legal-hold constraints under contention. Evidence: constraint tests. <!-- S11.8.e — libs/oshun/workbench-kit/src/contention-constraints.ts (+ .spec.ts, 23 tests) and tools/contention-constraints-scan.mjs (8 negative controls; RACES the real store; NO SERVER IS EXIT 1).

THE RULE: A CONSTRAINT UNDER CONTENTION IS ENFORCED BY WHOEVER HOLDS THE LOCK, AND FIVE OF THESE TEN HAVE NO LOCK BECAUSE THEY HAVE NO ROW.

unique       103 @@unique  → real PostgreSQL unique indexes
foreign       19 @relation → real foreign keys
check          NONE — Prisma cannot express one, live schema has zero
concurrency   version Int ×4, revision Int ×10, etag ×1  → read-then-write
idempotency   Redis; production sets OSHUN_BFF_IDEMPOTENCY_REDIS_URL
lease         lease Json? on work_item, guarded by a ROW LOCK
outbox        no model
inbox         no model
legal-hold    no model — and 78 TypeScript files that mention one
retention     19 models carry retention fields

You cannot contend for what has no row, so the honest first output is which of the ten can be contended at all (6), not a pass rate.

THE FINDING: THE ESTATE CONTAINS ITS OWN ANSWER AND APPLIES IT ON ONE PATH ONLY.

apps/oshun/bff/src/workbench/intent-store.ts:550 does it exactly right — opens a transaction, takes SELECT … FROM work_item WHERE id = $1 FOR UPDATE, and checks its compare-and-set INSIDE that transaction, with a comment saying why: "so a racer that already released and re-leased the item cannot have its fresh lease stolen". Eight FOR UPDATE sites across four files.

libs/oshun/persistence/src/durable-snapshot-store.ts:104 does not, and it holds V1's admin state (admin_store_snapshot, 52 store keys per S10.12.c):

const existing = await …findUnique({ where: { id: key } });
const version  = (existing?.version ?? 0) + 1;
await …upsert({ where: { id: key }, create: {…}, update: { document, version, updatedAt } });

A VERSION COLUMN THAT IS NOT IN THE WHERE CLAUSE IS A COUNTER, NOT A LOCK — and it is worse than no version at all, because the number looks like optimistic concurrency, so a reader comparing versions concludes nothing was lost. The doc comment above save() states the invariant this breaks: the new version "is the partition's current version + 1, so it strictly increases across saves".

AND THE PORT'S TYPE FORBIDS THE FIX. AdminStoreSnapshotDelegate.upsert declares where: { id: string } and there is no updateMany. A conditional update cannot be expressed through this interface at all — a shape the seam was cut to, not a slip at one call site.

MEASURED, NOT ARGUED. Eight concurrent save() calls at one partition:

8 of 8 succeeded · the row ended at VERSION 1 · 7 LOST UPDATES
the surviving document is writer 6; writers 0–5 and 7 are gone
and the version says one save happened

A ONE-WRITER RACE IS NOT A RACE. lostWrites of a single writer is 0 and so is lostWrites of a perfectly locked eight-writer race — the same number for "proved it holds" and "did not try" — so raceIsConclusive refuses a reading with fewer than two writers or two successes, and the scan fails rather than reports if the race comes out inconclusive.

CREDITED, NOT FLAGGED: the idempotency fallback logs "idempotency store: in-memory … replay safety is per-process only" and production configures Redis — an honest fail-loud seam.

Registered five ways → 232 entry points, 0 violations; public-api.spec.ts 241 passing. Harness database dropped. -->

  • S11.8.f Rehearse expand/backfill/verify/contract with old/new binaries, live writes, interruption/resume, rollback/compensation, and production- size data. Evidence: migration report. <!-- S11.8.f — libs/oshun/workbench-kit/src/expand-contract-rehearsal.ts (+ .spec.ts, 24 tests) and tools/expand-contract-rehearsal-scan.mjs (7 negative controls; rebuilds a production-shaped database and re-measures the applied script; NO SERVER IS EXIT 1).

THE RULE: A MIGRATION THAT DROPS NOTHING AND AN APPLY THAT DROPS SIXTEEN THINGS ARE THE SAME DEPLOY.

the 35 migration FILES     0 destructive · 0 data statements · 0 down
what the deploy APPLIES   86 statements · 16 DESTRUCTIVE · 0 transactions

Every migration file is an expand — no DROP TABLE, DROP COLUMN, SET NOT NULL or RENAME between them. That is the property a reviewer relies on, and it is a property of THE FILES. Production applies a computed diff instead, and against a production-shaped database (the migrated schema AND the BFF's bootstrap in one database, which production guarantees because OSHUN_STUBS_DATABASE_URL is deliberately unset — the terraform says so) that script drops 13 tables, 1 index and 2 constraints.

So the contract phase is NEVER AUTHORED and is PERFORMED ON EVERY DEPLOY, on tables nobody listed. And the destructive-operation guard cannot see it twice over: it globs the libs/<domain>/database directories, which do not include this schema — and even correctly aimed it reads the MIGRATION FILES, where there is genuinely nothing to find. A GUARD THAT READS WHAT WAS WRITTEN CANNOT SEE A DESTRUCTION THAT IS COMPUTED.

THERE IS NO RESUME POINT AND NO REVERSAL, AND BOTH ARE STRUCTURAL. The generated script has no BEGIN and no COMMIT, so psql applies 86 statements in autocommit and an interruption at statement 40 leaves 39 applied and 47 not; migrate diff never writes a _prisma_migrations row, so nothing records where it stopped. Zero down migrations exist — Prisma has no down concept, and the deploy's head -n -20 retention of applied scripts is a record of what happened, not a way to undo it. Two of the 35 files DO write their own BEGIN/COMMIT, so the estate knows the technique; the path that runs in production is the one without it.

BACKFILL AND VERIFY ARE NOT SLOW, THEY ARE ABSENT. Not one of the 35 migrations contains an UPDATE, INSERT or DELETE, and an expand/contract cycle exists to move DATA. 1 of 4 phases is performed deliberately.

0 of 5 conditions are rehearsed, and 2 of them cannot be: interruption-and-resume and rollback-or-compensation are not missing tests — the mechanism has to change first. condition-unrehearsed and condition-impossible are separate fault kinds for exactly that reason.

CORRECTED BY MY OWN SCAN: I first counted 18 destructive statements from a grep whose pattern also matched the ALTER TABLE line that merely OPENS a multi-line statement, so two statements were counted twice. The scan classifies whole statements and reports 16. A count off a line prefix is a count of lines.

Registered five ways → 233 entry points, 0 violations; public-api.spec.ts 242 passing. Harness database dropped. -->

  • S11.8.g Capture slow queries/locks/plans and fail on unsafe sequential scans or incompatible schema readiness. Evidence: database verification output. <!-- S11.8.g — libs/oshun/workbench-kit/src/plan-capture.ts (+ .spec.ts, 23 tests) and tools/plan-capture-scan.mjs (10 negative controls; runs the estate's own EXPLAIN checks BOTH ways against a live database; NO SERVER IS EXIT 1).

THE RULE: A PLAN TAKEN WITH SEQUENTIAL SCANS DISABLED CANNOT REPORT A SEQUENTIAL SCAN — AND WHETHER THAT IS COSTING ANYTHING TODAY IS A SEPARATE QUESTION THAT HAS TO BE MEASURED.

validate-explain-checks.ts is real and thorough: 294 EXPLAIN checks across 98 tables, derived from the 99 registered contracts in three query patterns (identity_lookup, live_contract_scan, payload_hash_drift), run LIVE in CI against a postgres:16-alpine service. More plan coverage than most estates have.

And before planning anything it issues, with no comment:

SET plan_cache_mode = 'force_custom_plan'
SET enable_seqscan = off
SET enable_sort = off

The first is right. The second is a real narrowing: the script asserts that the REQUIRED INDEX NAME appears in the plan, and enable_seqscan = off guarantees an index path if any index can serve the query — so a query that would sequentially scan in production is forced onto an index and the check passes.

an index exists that can serve this query          ← what it proves
production will not sequentially scan this table   ← what it cannot

WHAT I ASSUMED AND MY OWN SCAN REFUTED. I wrote that the setting is necessary because an empty table always seq-scans regardless of its indexes. Running all three query patterns against a freshly migrated database with the planner left alone produces ZERO sequential scans, and the plans are IDENTICAL with and without the settings — these predicates are selective equalities on the exact indexed columns and PostgreSQL takes the index anyway.

So the narrowing is REAL AND CURRENTLY COSTS NOTHING. Both numbers are recorded rather than one verdict, and the scan fails if they ever diverge in either direction — a regression and a fix both trip it.

THE ESTATE CONTAINS ITS OWN ANSWER AND APPLIES IT ON ONE PATH ONLY — twice more:

slow queries infra/terraform-v1/modules/rds/main.tf:131 sets shared_preload_libraries = pg_stat_statements,vector and log_min_duration_statement = 500; the deployed Hetzner postgres command sets shared_buffers, effective_cache_size, max_connections and neither of those locks a Grafana panel plots sum by (mode) (pg_locks_count) from Prometheus, and the deployed compose runs no prometheus, no grafana and no postgres exporter — neither deploy.sh nor deploy-hetzner.yml mentions docker/observability at all. A READER WITH NO PRODUCER.

(My first pass said pg_locks appears in 0 files. It appears in 1 — the scan's git grep found the dashboard my extension-filtered grep had excluded.)

SCHEMA READINESS HAS NOTHING TO COMPARE AGAINST: a readiness gate compares the schema a binary expects with the one the database has, and migrate diff never writes a _prisma_migrations row (S11.8.b), so nothing states what production is at.

1 of 4 captures is taken where it runs.

Registered five ways → 234 entry points, 0 violations; public-api.spec.ts 243 passing. Harness database dropped. -->

  • S11.9 Add contract fuzzing for envelopes, manifests, and remote payloads. Evidence: minimized failure corpus is retained. <!-- S11.9 parent — a–g all closed. Seven kit modules, seven scans, 183 unit tests, two committed corpora, one new CI workflow. 241 entry points, 0 violations; the kit's 15,257 tests pass and all 83 file-only scans pass.

THE PARENT'S EVIDENCE IS RETAINED AND REPLAYS. evidence/v1-workbenches/workbench-kit-minimized-corpus.jsonl holds 13 reproducers, every one carrying seed, schemaId, version, expectedResult, the rule that computed the signature, and the minimized text; every one replays against the parser it was minimized from; and shrinking-scan rebuilds, rewrites and re-replays the whole file on every run. Beside it, workbench-kit-negative-envelope-corpus.jsonl holds the 18 wire-text cases.

ONE RULE RUNS THROUGH ALL SEVEN CELLS: DECLARE WHAT YOU EXPECT TO HAPPEN, NOT MERELY THAT SOMETHING DID.

.a a generated envelope invalid for TWO reasons tests neither, so every mutation names the one ParseErrorCode it provokes .b a negative corpus stored as OBJECTS cannot hold half the cases it names — {"kind":"a","kind":"b"} is one key after JSON.parse .c a relationship is broken by changing ONE side, and a builder that regenerates both cannot produce one .d a polyglot is a fact about a SET and the decider takes a SCALAR .e an invariant you cannot OBSERVE cannot be asserted — four verdicts, not two, because unobservable and vacuous are not held .f a shrink is only valid if the smaller case fails for the SAME reason .g a check that nothing SCHEDULES is a check nobody runs

WHAT THE SECTION FOUND IN THE ESTATE, all measured:

· all four encoding mutations are ACCEPTED — a lone surrogate, a NUL, a bidi override and an NFD/NFC pair, each a string in a string-typed field · a duplicate tenant is ACCEPTED with the LAST declaration winning, so a signature over one and an authorization over the other disagree invisibly · a duplicate kind silently RE-ROUTES the envelope to a different parser before any validation runs · a large integer is accepted with a DIFFERENT VALUE than was sent · three of the ten manifest relationships are NEVER COMPARED — a manifest may credit anybody as its producer and the assembly passes · the sniffer's four facts cannot reach the content decision's one string, so two of three collapses ACCEPT a polyglot and none can quarantine it · 74 of 82 verification scripts were run by nothing, and the first honest run of them found nine failing scans — three real estate drift, five instances of one systemic trap, and one real crash in shipped kit code

AND THE METHOD CORRECTED ME REPEATEDLY, which is the point of building it this way. My survivesAsAnObject compared an expression to ITSELF and could never fail (2 lost cases where the truth is 7). I claimed the sniffer would miss the canonical polyglot; it does not. I claimed { ...parsed } drops __proto__; only Object.assign does, and it changes the prototype too. I tiered the schedule on 30 database scans when the number is 8. Every one was found by running the thing rather than reading it.

THE TRAP OF THE SECTION, worth carrying forward: A REGISTER THAT COUNTS ITS OWN FAMILY INVENTS A DEFECT, AND THE FAMILY HAS FIVE MEMBERS — module, spec, scan, the Nx doc in project.json, and the export-map reason — four of them carrying prose about the terms the scan searches for, and one living OUTSIDE libs/oshun/workbench-kit. It fired four separate times in S11.9.g alone. -->

  • S11.9.a Generate schema-valid, boundary-valid, and invalid command/query/ event/job/probe envelopes across kind/version/field/size/order/encoding. Evidence: envelope property suite. <!-- S11.9.a — libs/oshun/workbench-kit/src/envelope-generators.ts (+ .spec.ts, 28 tests incl. a fast-check property over the real parsers) and tools/envelope-generators-scan.mjs (9 negative controls). fast-check added to the kit as a catalog devDependency.

THE RULE: A GENERATED ENVELOPE THAT IS INVALID FOR TWO REASONS TESTS NEITHER.

Take a valid command, set schemaVersion to 99 to exercise the version gate, and — because the mutation also dropped idempotencyKey — the parser answers missing-field. The assertion "it was rejected" passes. The version gate never ran. So every mutation declares the ONE ParseErrorCode it exists to provoke, and outcomeVerdict returns rejected-for-the-wrong-reason when the parser answers with a different one — a failure, not a pass.

THE SECOND: BOUNDARY-VALID IS THE LARGEST INPUT THAT MUST SUCCEED, NOT THE SMALLEST THAT MUST FAIL. A generator making small valid envelopes and enormous invalid ones never evaluates the comparison operator — > and >= behave identically on every input it produces — so the boundary class emits envelopes of exactly MAX_ENVELOPE_BYTES that must parse.

WHAT THE ESTATE HAS, MEASURED. workbench-kit-envelope-goldens.json holds 57 hand-authored cases, 22 accept and 35 reject, each carrying its expected code. Sorted by the six dimensions this cell names:

field 18 · version 4 · kind 3 · size 3 · order 3 · ENCODING 0

and it provokes 7 of the 14 declared ParseErrorCode values — the seven it misses are exactly the identity parser's own rejections. A hand-authored corpus covers the cases somebody thought of. That is not a criticism of it; it is why a generator is a different artifact rather than a bigger version of the same one.

THE FINDING: all four encoding mutations are ACCEPTED today. A lone UTF-16 surrogate, a NUL byte, a right-to-left override and an NFD/NFC pair all pass parseCommandEnvelope, because each is a string in a string-typed field and a type check cannot see the difference. The scan reports them as accepted-but-should-not-be — 20 of 245 applications — rather than failing, because that is this cell's finding and not its bug.

THE THIRD, WHICH THE GENERATOR TAUGHT ME: THERE ARE TWO ENTRY POINTS AND ONLY ONE WEIGHS THE ENVELOPE. The five parse<Kind>Envelope functions take an already-parsed object and enforce shape; parseEnvelopeJson takes the wire TEXT and checks the bytes BEFORE JSON.parse, which is the right order and its own comment says why. My first pass asserted the size limit at the wrong door.

AND THE CELL'S OWN RULE CAUGHT ME TWICE MORE. The size mutation first padded payload — which job-update and capability-probe do not declare — so a mutation about SIZE silently became one about SHAPE on two of the five kinds and came back unknown-field. Its replacement padded extensions under x-pad, which parseExtensions refuses because keys must be <vendor>.<name>. Both were found by RUNNING the mutations, not by reading them. And field/undefined- not-null targeted missing-field and the parser answers not-a-string, because it distinguishes an absent key from a present one holding the wrong type — a sharper answer than the register assumed.

Registered five ways → 235 entry points, 0 violations; public-api.spec.ts 244 passing. -->

  • S11.9.b Fuzz duplicate/unknown/missing fields, discriminators, enums, IDs, numbers, dates, Unicode, nesting, arrays, aliases, and older/newer/unsupported versions. Evidence: negative envelope corpus. <!-- S11.9.b — libs/oshun/workbench-kit/src/negative-corpus.ts (+ .spec.ts, 26 tests) and tools/negative-corpus-scan.mjs (8 negative controls), emitting evidence/v1-workbenches/workbench-kit-negative-envelope-corpus.jsonl — 18 cases over all 13 categories.

THE RULE: A NEGATIVE CORPUS STORED AS OBJECTS CANNOT HOLD HALF THE CASES IT NAMES.

{"kind":"command","kind":"event"}  →  JSON.parse  →  { kind: 'event' }

One key. JSON.parse resolves a repeated key by last-wins, silently, before any parser runs — so a corpus entry written as a JSON object, which is how all 56 envelopes in workbench-kit-envelope-goldens.json are stored, has had the defect removed from it before it reached disk. The same is true of precision above 2^53, of -0 (parses as -0, serializes as 0), of NaN and a leading BOM (not JSON at all), and of escaped key forms.

5 of the 13 categories are text-only, and 7 of the 18 cases do not survive object storage. So the corpus is JSON Lines with the envelope held as a STRING, and the scan PROVES the format matters rather than asserting it: it writes the corpus, reads it back byte-identical, then stores the same cases as objects, reads THAT back, and requires exactly 7 to have changed. If none had, the text format would be unnecessary and the register wrong.

FINDINGS, all measured against the real parseEnvelopeJson:

· the duplicate TENANT is ACCEPTED and the last declaration wins, so a signature computed over one and an authorization computed over the other disagree and nothing downstream can see there were two · the duplicate KIND silently RE-ROUTES the envelope to a different parser before any validation runs — the rejection that follows is correct and its reason is not the one a reader would guess · a large integer inside a legal extension is ACCEPTED with a different value than was sent · an escaped lone surrogate is ACCEPTED, confirming S11.9.a's finding at the wire level

__proto__ ARRIVES AS AN ORDINARY OWN KEY AND THE TWO COPY IDIOMS DISAGREE: { ...parsed } keeps the key and leaves the prototype alone; Object.assign({}, parsed) DROPS the key and CHANGES THE PROTOTYPE, because spread defines own properties and assign triggers the setter. I first wrote that both did the second; only one does.

OLDER, NEWER AND UNSUPPORTED ARE THREE ANSWERS, NOT TWO. A newer envelope that still parses must be TOLERATED — that is what a closed shape with an extension namespace is for — and a corpus that only knows supported from unsupported cannot express forward compatibility at all.

AND THE MEASURE ITSELF WAS WRONG FIRST. survivesAsAnObject compared JSON.stringify(JSON.parse(text)) with JSON.stringify(JSON.parse(text)) — the same expression twice, a check that could never return false — and reported 2 lost cases where the truth is 7. Replaced with a comparison against the ORIGINAL minified text, with the minified precondition asserted rather than assumed.

Registered five ways → 236 entry points, 0 violations; public-api.spec.ts 245 passing. -->

  • S11.9.c Generate profile-valid and invalid bundle manifests with path/ ref/hash/size/media/schema/dependency/provenance/version/lifecycle relationships. Evidence: manifest property suite. <!-- S11.9.c — libs/oshun/workbench-kit/src/manifest-relations.ts (+ .spec.ts, 23 tests against the real bundleAssemblyFaults) and tools/manifest-relations-scan.mjs (7 negative controls).

THE RULE: A RELATIONSHIP IS BROKEN BY CHANGING ONE SIDE, AND A BUILDER THAT REGENERATES BOTH CANNOT PRODUCE ONE.

bundleAssemblyFaults takes FIVE inputs — profile, request, layout, observed, manifest — and every one of its thirteen faults is a DISAGREEMENT between two of them. A generator that derives the manifest from the layout can never emit one that disagrees with it: the derivation is what maintains the invariant. Feed a mutated artifact through the builder and the digest is recomputed, the path re-templated, the bytes re-counted, and the five inputs come out consistent again — a different VALID bundle. So consistentTuple builds the valid tuple (the thing the estate has only inside bundles.spec.ts) and each break reaches in and changes exactly one side of one edge, which the spec and scan both verify by comparing object identity across all five inputs.

BREAKING ONE EDGE MUST PRODUCE EXACTLY ONE FAULT. also-broke-something-else is a distinct verdict from wrong-fault because they call for different repairs: the first means the mutation is too broad, the second means the register is wrong. And expectedFaults is a SET — renaming a manifest entry's path is inherently TWO facts (an entry the layout lacks, and a layout entry the manifest omits), and a register demanding one would be wrong about the checker rather than about the bundle.

WHAT THE ESTATE HAS. MANIFEST_FIXTURE_CORPUS holds 16 fixtures across five classes, 2 valid and 14 invalid — and every one is a LONE MANIFEST: { id, fixtureClass, expect, because, manifest }, with no profile, layout or observed entries beside it. A single document can be malformed, out of bounds or version-skewed; it cannot DISAGREE WITH ANYTHING, because there is nothing in the fixture for it to disagree with. That corpus tests the schema validator.

THE FINDINGS: THREE OF THE TEN RELATIONSHIPS ARE NEVER COMPARED.

media entries are matched BY PATH and only the DIGEST is compared, so the manifest's own kind, slotId and byteLength are never checked against the layout's — a manifest can call a document a signature and the assembly passes ref the external-reference branch reads the LAYOUT's externalRef; the manifest's own field is never read by the check at all provenance producedBy and producedAt are never read, so a manifest may credit anybody and the assembly passes — and the manifest is the only record an auditor will ever see

All three are declared as expectedFaults: [], so the register STATES the gap rather than tripping over it, and the scan fails in both directions if any of them starts firing.

A FOURTH CORRECTION WAS MINE: I expected changing bytesRead to raise the byte-length fault. It raises digest-not-verified-from-the-bytes instead, which is the checker being right — a file shorter than expected fails to verify before anybody counts it — and the byte-length fault is about the DECLARED length. And the media break first wrote 'signature' unconditionally, which is a no-op on the layout's first entry, so it now chooses relative to what is there — the same lesson as S11.9.a's kind mutation.

Registered five ways → 237 entry points, 0 violations; public-api.spec.ts 246 passing. -->

  • S11.9.d Fuzz raw remote JSON/XML/form/multipart/archive/media/stream/ provider/tool/webhook payloads for malformed/truncated/oversized/ ambiguous/polyglot/version-skew input. Evidence: remote payload suite. <!-- S11.9.d — libs/oshun/workbench-kit/src/remote-payloads.ts (+ .spec.ts, 24 tests) and tools/remote-payloads-scan.mjs (7 negative controls). 11 payloads over all 10 kinds and all 6 defect classes.

THE RULE: A POLYGLOT IS A FACT ABOUT A SET, AND THE DECIDER TAKES A SCALAR.

BOTH HALVES OF THE ESTATE ARE RIGHT, AND I CHECKED RATHER THAN ASSUMED. sniffMediaType(bytes) returns FOUR facts — every signature matched, the leading one, polyglot, executable — and its comment explains why it returns a LIST: "a GIF header with a zip's central directory at the end is a valid image to an image decoder and a valid archive to an unzipper". I expected the 15-signature table to miss that case. It does not:

GIF89a + padding + PK\x05\x06
  → matches ['application/zip','image/gif'], leading 'image/gif', polyglot true

thanks to a scanning trailing-zip signature at offset −1 written for exactly this. And decideContent checks the SNIFFED type against the policy — noting that "the DECLARED type is the far side's claim and is not what is checked" — and QUARANTINES rather than refuses on a mismatch.

AND THEY CANNOT BE JOINED WITHOUT LOSS. ResponseObservation carries sniffedMediaType: string. One string. No field for the match list, none for polyglot, none for executable. Measured across all three available collapses, against a policy accepting image/gif:

leading         → collapsed image/gif        → ACCEPTED
first-match     → collapsed application/zip  → refused
trust-declared  → collapsed image/gif        → ACCEPTED
                   …and all three discard the polyglot flag

None can quarantine, because the decision never receives it. Two of the three accept. trust-declared disables the check entirely: sniffed equals declared, so nothing ever mismatches and the accepted-media-type test becomes a test of the far side's own claim.

AND THE TWO HALVES ARE JOINED NOWHERE. sniffedMediaType has 0 producers and decideContent 0 callers outside a spec, so the collapse has never been written down and the first host to write it will choose one of the three by instinct. decideWithSniff is that join, written once, in the open, taking the collapse as an EXPLICIT argument because there is no correct default — a function that picked one silently would be making the decision this module exists to surface.

THE SECOND: THREE OF THE SIX DEFECT CLASSES ARE NOT ABOUT THE PAYLOAD. Malformed, truncated and oversized are properties of the bytes — you can hold the payload alone and say whether it has them. Ambiguous, polyglot and version-skew are properties of a RELATIONSHIP between the bytes and whoever decides what they are: the same payload is unambiguous to a reader that sniffs and ambiguous to one that trusts a header. A corpus filing all six under "bad input" loses that.

The scan re-derives the signature table from source, re-counts the producers, and re-runs the collapse experiment — failing in BOTH directions, so a change that let the decision carry the polyglot trips it exactly as a regression would.

Registered five ways → 238 entry points, 0 violations; public-api.spec.ts 247 passing. -->

  • S11.9.e Assert bounded time/memory/depth/output/logging, no crash/hang/ traversal/SSRF/leak, stable typed rejection, and no partial side effect. Evidence: invariants. <!-- S11.9.e — libs/oshun/workbench-kit/src/fuzz-invariants.ts (+ .spec.ts, 27 tests) and tools/fuzz-invariants-scan.mjs (7 negative controls), run over the S11.9.b and S11.9.d corpora against the real parseEnvelopeJson.

THE RULE: AN INVARIANT YOU CANNOT OBSERVE CANNOT BE ASSERTED, AND ASSERTING IT ANYWAY IS THE DEFECT.

"No partial side effect" is a claim about what happened somewhere else. "No leak" is about what appeared in an output. "Bounded logging" is about what was written. A fuzz run that feeds a malformed envelope to a parser, gets a clean typed rejection, and concludes that nothing was logged, written or leaked has concluded three things from a return value that carries none of them.

So each invariant records THE OBSERVATION IT REQUIRES — a clock, a recorder, a mutation log, an egress log, a process probe, a second run, or the value itself — and checkInvariants returns FOUR verdicts:

held          observed, and it holds
violated      observed, and it does not
unobservable  the seam that would show it is not wired
vacuous       the subject cannot violate it

The last two are the point. unobservable reported as held is the unrunnable-lane defect. vacuous reported as held is how a suite claims coverage it does not have: a parser is pure, so "no partial side effect" is TRUE of it and says nothing about the writer three layers down.

MEASURED over 29 inputs:

6 HELD · 0 violated · 3 UNOBSERVABLE · 3 VACUOUS
unobservable: bounded-memory, bounded-depth, bounded-logging
vacuous:      no-ssrf, no-traversal, no-partial-side-effect

SIX OF TWELVE WERE ACTUALLY CHECKED, and the harness reports mostly-unchecked rather than twelve passes.

THE SECOND: A STABLE TYPED REJECTION IS A CLAIM ABOUT TWO RUNS, NOT ONE. Every other invariant can be judged from a single execution, so runTwice is built into the recorder rather than left to a caller who might skip it when the suite is slow — a minimized reproducer that does not reproduce is worse than none.

THE THIRD: A CRASH AND A HANG ARE THE SAME OBSERVATION FROM DIFFERENT SIDES. A throw is loud and immediate; a hang is silent and indefinite, and a harness with no deadline turns the second into a process killed by something outside the test whose report says nothing at all. The deadline is part of the observation.

AND PURITY BOUNDS EFFECTS, NOT WORK: bounded-time and bounded-memory are deliberately NOT vacuous for a pure subject, because a pure function can still take quadratic time and allocate a gigabyte on an adversarial input.

Registered five ways → 239 entry points, 0 violations; public-api.spec.ts 248 passing. -->

  • S11.9.f Seed runs deterministically, shrink failures, retain smallest reproducer with seed/schema/version/expected result, and promote to regression fixture. Evidence: minimized corpus workflow. <!-- S11.9.f — libs/oshun/workbench-kit/src/shrinking.ts (+ .spec.ts, 22 tests) and tools/shrinking-scan.mjs (7 negative controls), emitting evidence/v1-workbenches/workbench-kit-minimized-corpus.jsonl — 13 reproducers, 6170 bytes saved, every one replaying.

THE RULE: A SHRINK IS ONLY VALID IF THE SMALLER CASE FAILS FOR THE SAME REASON.

Measured on the estate's own versions/below-minimum case, the same input shrunk twice:

still-fails     548B → **2B** `{}`                  fails **unknown-kind**
same-signature  548B → 18B  `{"kind":"command"}`  fails invalid-field

The first is a reproducer for a DIFFERENT DEFECT, reported with total confidence — whoever picks it up fixes the wrong thing. So shrink takes a predicate returning a FAILURE SIGNATURE, not a boolean, and keeps a candidate only when the signature matches; three candidates were refused for drift on this one case. That is S11.9.a's discipline applied where "still fails" is most tempting.

A REPRODUCER IS A CLAIM SOMEBODY ELSE CAN RUN, AND FOUR OF ITS FIVE FIELDS ARE ABOUT THE RUNNER: the schema says what it was validated against, the version which build, the expectedResult what "reproduces" means, and the seed how to reach the generator state. A corpus of texts alone is a corpus of things that USED to fail.

THE THIRD, AND IT LIMITS THE WHOLE TECHNIQUE: A SHRINK CAN BE NO MORE PRECISE THAN THE FAILURE THE SUBJECT REPORTS. That 18-byte reproducer is {"kind":"command"} — which carries NO schemaVersion rather than a zero one. parseCommon answers a missing and a zero schemaVersion with a byte-identical invalid-field | schemaVersion must be an integer >= 1, verified here. A finer signature does not help: code-only and code-plus-message reach the same 18 bytes. The fix is a subject that distinguishes, not a predicate that squints harder — and the same parser DOES distinguish elsewhere (not-a-string vs missing-field for envelopeId, S11.9.a). Two levels of precision in one file.

And deletion is CHEAPER than substitution for any shrinker, so wherever a subject conflates "absent" with "wrong value", the minimized case will always be about absence. Worth knowing before reading a reproducer as a description of the bug.

A PROMOTED FIXTURE THAT STOPS FAILING MUST SAY SO, NOT PASS. replay returns no-longer-reproduces — a fix is good news that must be written down, because the expectation is now wrong and the fixture covers nothing until somebody says what it should cover instead. A regression corpus whose entries can quietly become vacuous grows without ever telling you what it still covers.

The scan runs the whole loop every time — shrink, build, write, read back, replay — and re-runs the two-predicate comparison, failing if the loose predicate ever stops drifting.

Registered five ways → 240 entry points, 0 violations; public-api.spec.ts 249 passing. -->

  • S11.9.g Run fast mutation corpus in PR and broader scheduled fuzzing with resource limits and owner routing. Evidence: CI schedule. <!-- S11.9.g — libs/oshun/workbench-kit/src/fuzz-schedule.ts (+ .spec.ts, 22 tests), tools/fuzz-schedule-scan.mjs (6 negative controls), and the deliverable itself: .github/workflows/v1-workbench-kit-verification.yml.

THE RULE: A CHECK THAT NOTHING SCHEDULES IS A CHECK NOBODY RUNS, AND A REPOSITORY FULL OF THEM LOOKS EXACTLY LIKE A REPOSITORY THAT VERIFIES ITSELF.

package.json declared 82 verify:operations:* scripts and exactly 8 were named by any workflow, all in v1-disaster-recovery.yml. The other 74 worked and nothing ran them. That is not a criticism of the scans; it is the shape of defect hardest to see from inside — a directory of ninety scanners is indistinguishable at a glance from a repository that checks ninety things.

A TIER IS DECIDED BY WHAT A CHECK NEEDS, NOT BY HOW IMPORTANT IT IS. 8 of the 90 scans open a PostgreSQL connection and cannot run in a job that has none, however important; the other 82 read files and finish in ~1.7s each, however minor. (tools/ holds 135 .mjs files — 90 scans and 45 race probes the scans drive. 31 of the 135 open a database and only 8 of them are scans, a distinction my first tiering got wrong and this cell's own scan caught.)

OWNER ROUTING WITH ONE OWNER IS NOT ROUTING: .github/CODEOWNERS carries 94 rules and one owner, catch-all included. Recorded as single-owner — a real state whose consequence is that the routing cannot go wrong and cannot help.

THE WORKFLOW partitions the scans by path at run time, so a scan added tomorrow is scheduled without anybody filing it; runs the files-only tier on every PR touching the kit and the database tier nightly against a postgres service; refuses an empty partition (a pattern matching nothing would leave the job green having run nothing); reports exit 2 as undetected negative control distinctly from a failure; and carries no continue-on-error and no || true — six other workflows in this repository already do.

AND THE FIRST HONEST RUN FOUND NINE FAILING SCANS.

· three were genuine estate drift from commits that landed on main — a register said 759 .evaluate( calls in the e2e suites and there are 760; verified at the register's own authoring commit, where it was 759 · five were one systemic trap. A REGISTER THAT COUNTS ITS OWN FAMILY INVENTS A DEFECT, and THE FAMILY HAS FIVE MEMBERS — module, spec, scan, the Nx doc in project.json, and the export-map reason — four of which carry PROSE about the very terms the scan searches for, and the fifth lives OUTSIDE libs/oshun/workbench-kit, so every exclusion written as a startsWith on that directory misses it. One scan searched for a literal its own source contains, inside the regex doing the searching. · one was a real crash in shipped kit code. game-day.ts gained announcedToActors, unrehearsableBecause, presentsAs, confirmedBy, movesTheRecoveryPoint and detects after its .mjs probe was written. The absent fields read as undefined, and undefined slipped past every === null guard and threw on .length. A .MJS PROBE AGAINST A TYPED MODULE HAS NO COMPILER BETWEEN THEM, so a field added to the type is not added to the fixture and nothing says so until the probe runs — which nothing did. Four of seventeen conditions were left indistinguishable; the kit arm now scores 17/17.

All nine repaired. All 83 file-only scans pass and the kit's 15,257 tests pass. Registered five ways → 241 entry points, 0 violations. -->

  • S11.10 Add malicious upload and remote-fetch fixtures. Evidence: archive, media, path, metadata, and SSRF cases. <!-- S11.10 CLOSED — a through g. Seven cells, seven corpora, and one shape that turned up so often it stopped being a coincidence.

THE SHAPE: A HAZARD IS A PROPERTY OF A PAIR, AND EVERY CORPUS THAT HOLDS ONLY ONE HALF OF THE PAIR IS ANSWERING A QUESTION NOBODY ASKED.

.a a collision needs two names, and it is a collision only on a filesystem that folds — so a name check reading one path cannot hold one. .b 'executes' is not a property of a file, it is a property of what OPENS it. .c the same pair again, and the estate refuted three of my four claims before the fourth survived. .d a metadata value is dangerous in one SINK and inert in another; '=1+1' is four characters in HTML and a live formula in a CSV export. .e a spelling is a bypass only against a guard that SEES it, and 8 of the estate's 21 SSRF targets reach a post-parse guard respelt. .f a transport hazard is a transfer that never ends, and every observation the estate can make is a record of one that DID. .g and then: five of the six closing expectations cannot be violated by a fixture at all, because the subject was never the fixture.

Every one of those needed a register that returns BOTH lists — where a case bites and where it is inert — because a register that recorded only the first makes "harmless everywhere" and "never tested" identical. Three cells needed a THIRD value on top of that: S11.9.e's unobservable, S11.10.f's from-neither, S11.10.g's intentionally-persistent. In each case a boolean would have reported a defect against something correct, or a pass against something nobody checked.

WHAT THE ESTATE ALREADY HAD, because measuring first was the discipline that made this section work. bundle-paths modelled all three filesystems and refuted three of .a's claims. The media sniffer caught the canonical GIF+zip polyglot I predicted it would miss. fetching.ts covered ten of .e's eleven categories and all of .f's machinery. Five of the seven cells found MORE credit than gap, and saying so is the honest half of each entry.

LIVE DEFECTS FOUND AND FIXED: game-day-scan leaking 4 temp directories a run (88 accumulated, 2.4 MB) and asTheResolverSeesIt quoting an attacker-chosen hostname into a diagnostic. LIVE DEFECTS FOUND AND RECORDED: three of FetchBounds' five fields have no reader, so a five-gigabyte compressed response that decompresses to four kilobytes is ACCEPTED; timeout was a declared refusal nothing in the repository could produce; and the SSRF corpus's own prose names a trailing-dot bypass the corpus does not contain.

AND THE TRAP THAT KEPT FIRING: a register that counts its own family invents a fact. Five separate instances across this section, the last two in .g alone — a network census that found a scan SEARCHING for a fetch call, and a writer census that found the pattern naming the writing functions inside its own source. The answer is never a longer exclusion list: it is to ask what a file IMPORTS rather than what it mentions, or to assemble the search string at run time so the searcher cannot match itself.

7 modules, 7 specs, 7 scans, 5 registrations each. Kit suite 244 files / 15,435 tests green; 248 entry points, 0 violations. -->

  • S11.10.a Add archive fixtures for traversal/absolute/symlink/hardlink/ device/duplicate/case/Unicode/reserved paths, excessive entries/depth, nested archives, and decompression bombs. Evidence: archive corpus. <!-- S11.10.a — libs/oshun/workbench-kit/src/archive-fixtures.ts (+ .spec.ts, 27 tests against the estate's real checkers) and tools/archive-fixtures-scan.mjs (8 negative controls). 15 recipes over all 13 hazards.

WHAT I CLAIMED AND THE ESTATE REFUTED — THREE TIMES. I set out to report that bundle-paths.ts reads only names, so an entry's KIND, a collision's DESTINATION and a resolved traversal are all invisible to it:

· specialFileFaults checks ADMISSIBLE_FILE_KINDS — symlink, hardlink, device and fifo are refused BY KIND already · collisionFaults(paths, target) TAKES the destination filesystem, and all three behaviours are modelled: POSIX_EXT4, WINDOWS_NTFS and APPLE_APFS with unicodeForm: 'nfd'. NFC/NFD of one word → unicode-fold-collision on APFS and nothing on ext4; README.md/readme.md → case-fold-collision on NTFS · a name resolving INSIDE the root (a/b/../c/thing.txt) is reported as dot-dot-segment and not as escapes-its-own-root

That last one caught my scan, not the estate: I treated any finding as a refusal, which would have reported the estate as rejecting a legitimate entry. AN OBSERVATION AND A REFUSAL ARE DIFFERENT FINDINGS, AND A READER THAT COUNTS THEM TOGETHER INVENTS A DEFECT.

THE RULE THAT SURVIVED: A KIND CHECK AND A TARGET CHECK ARE DIFFERENT CHECKS, AND ONLY ONE CAN BE ASKED HERE. ObservedFileEntry is { path, kind, linkCount } — there is no field for where a link points, and nothing in the module names one. So the only expressible policy is "no links at all", which is sound. The moment a deployment admits symlinks, because a bundle legitimately carries one, nothing can ask where one goes: a link named docs/readme is confined, normalized, in NFC and innocent by every check in the file, and its target is ../../../../etc/passwd. The scan reads the field list out of the source and fails in both directions — the day a target field appears, the register is wrong and says so.

THE CHECKERS EXIST AND NOTHING FED THEM THE HOSTILE INPUT — the same shape as S11.9.a, five parsers and no generator. The corpus is the deliverable.

AND AN ARCHIVE FIXTURE THAT IS SAFE TO COMMIT IS NOT THE ARCHIVE YOU NEED. A bomb is 42 KB expanding to petabytes; committing one puts a live hazard in every clone, flags every content scanner, and is one double-click from a full disk in any tool that previews archives helpfully. So the corpus holds recipes — deterministic builders with expected refusals — and the bytes exist only inside the run. 4 of the 15 must never be files: the bomb, a million entries, a 2000-deep path and a device node. The million-entry recipe builds and counts without being walked.

Registered five ways → 242 entry points, 0 violations; public-api.spec.ts 251 passing. -->

  • S11.10.b Add image/audio/video/document/3D fixtures for wrong declared/ actual type, polyglot, malformed headers/chunks, huge dimensions/ duration/frames, parser bugs, embedded active content, and corrupted metadata. Evidence: media corpus. <!-- S11.10.b — libs/oshun/workbench-kit/src/media-fixtures.ts (+ .spec.ts, 23 tests against the estate's own sniffer) and tools/media-fixtures-scan.mjs (8 negative controls). 11 cases over all 8 hazards.

THE RULE: "EXECUTES" IS NOT A PROPERTY OF A FILE, IT IS A PROPERTY OF WHAT OPENS IT.

EXECUTABLE_MEDIA_TYPES lists four formats — ELF, Mach-O, a Windows PE, a shell script — with a comment saying, correctly, "not a policy: a fact about the format". It is a fact about the format when the opener is an operating system. A media pipeline's opener is a browser, a PDF viewer, an image library. Measured against the estate's own sniffMediaType:

an SVG with a <script> element   →  matches [], executable FALSE
an HTML file with a <script>     →  matches [], executable FALSE
a PDF carrying /JavaScript       →  matches ['application/pdf'], executable FALSE

A user asked for an image uploads a program, and every check in the pipeline agrees it is not executable.

AND THE CURRENT SAFETY IS ACCIDENTAL. Two of the three sniff as no format at all, so a policy with a non-empty permittedMediaTypes refuses them as declared-type-unrecognisable. That works — because the sniffer does not know what an SVG is. The day somebody adds an SVG signature, which is the obvious improvement to make when SVG uploads are wanted, the accidental refusal becomes an acceptance and executable: false is what decides it. The scan fails if a signature for SVG or HTML appears, so the improvement cannot land silently.

THE SECOND: A DIMENSION IS A NUMBER IN A HEADER, AND THE ALLOCATION HAPPENS BEFORE ANYTHING VERIFIES IT.

a 33-byte PNG header declaring 65535 × 65535
  = 4,294,836,225 pixels = 17.18 GB of RGBA
  and maxByteSize bounds 33

Nothing in the kit reads a header past its magic number, so nothing could bound it. declaredPixels computes the claim so a caller can refuse on the claim rather than on the size — and returns null rather than zero for anything that is not a PNG header, because a zero from an unparsed header and a zero from a real one are the same number and not the same fact. The ordinary 1920×1080 control lives in the same 33 bytes, so the check is shown to discriminate rather than to refuse everything.

THE THIRD: A CHUNK LENGTH IS ALSO A NUMBER IN A HEADER. A chunk claiming 4 GB inside a 20-byte file is a refusal or a read past the end depending entirely on whether the parser believes the number or the buffer.

Registered five ways → 243 entry points, 0 violations; public-api.spec.ts 252 passing. -->

  • S11.10.c Add filenames/paths/IDs for Unicode normalization/confusables/ bidi/invisible/control, encoded traversal, drive/UNC/device, trailing dots/spaces, separators, and extreme length. Evidence: path corpus. <!-- S11.10.c — two vectors added to PATH_VECTOR_CORPUS (27 → 29) and libs/oshun/workbench-kit/src/path-pairs.ts (+ .spec.ts, 19 tests) with tools/path-pairs-scan.mjs (7 negative controls).

MOST OF THIS CELL WAS ALREADY DONE. PATH_VECTOR_CORPUS carried 27 vectors covering 9 of the 10 categories — normalization, bidi, invisible, control, encoded traversal, drive/UNC/device, separators, reserved names, and confusables through a real per-component script-mixing check with Latin, Cyrillic and Greek ranges. Each vector records the DECODE STAGE at which it becomes visible. Saying so is the honest first half of this entry.

THE ONE GAP WAS LENGTH — AND A CORPUS OF HOSTILE CHARACTERS HAS NO ENTRY FOR "TOO MANY ORDINARY ONES." Every existing vector holds something wrong IN THE STRING: a control character, a bidi override, a Cyrillic lookalike, a percent-encoded separator. The two added here hold nothing wrong at all — three hundred ordinary letters, forty ordinary directory components — and the fault is the COUNT. component-longer-than-the-limit and deeper-than-the-limit both already fired; the corpus's longest path was 32 characters.

THE SECOND: A HAZARD THAT NEEDS TWO INPUTS CANNOT LIVE IN A CORPUS OF ONE. Third time this section has met that shape — S11.9.b (a corpus of objects cannot hold a duplicate key), S11.9.c (a lone manifest cannot disagree with anything), and now this. bundlePathFaults('story/body.json.') returns [], correctly: there is nothing wrong with it. It becomes a hazard only beside story/body.json, and only where a filesystem trims. The estate files it exactly there, as trailing-dot-or-space-collision.

Measured against the estate's three modelled destinations:

body.json / body.json.     NTFS only
README.md / readme.md      NTFS and APFS
café NFC / café NFD        APFS only

Three pairs, three answers, and no pair is a hazard everywhere — one that were would be a property of the names and would belong in the single-path corpus. So "does this bundle contain a collision?" has no answer until the destination is named: a bundle built on Linux, scanned on Linux and extracted on a developer's Mac passes every check and overwrites a file.

readPair returns BOTH the dangerous and the harmless destinations, because a register recording only where a pair is dangerous would make "harmless everywhere" and "never tested" look the same — the same defect as reporting an unobservable invariant as held (S11.9.e).

AND MY OWN FREEZE CAUGHT ME: .sort() mutates in place, and readPair returns frozen arrays, so the spec's assertions threw until they sorted a copy.

Registered five ways → 244 entry points, 0 violations; public-api.spec.ts and bundle-paths.spec.ts 297 passing together. -->

  • S11.10.d Add EXIF/XMP/ID3/container/document/archive/3D/filename/title/ caption/comment metadata containing injection, secrets, protected data, URLs, scripts, and oversized nested values. Evidence: metadata corpus. <!-- S11.10.d — libs/oshun/workbench-kit/src/metadata-fixtures.ts (+ .spec.ts, 20 tests) and tools/metadata-fixtures-scan.mjs (7 negative controls). 11 payloads over 6/6 classes, 8/8 fields and 8/8 sinks.

THE RULE: A FIELD CALLED "COMMENT" IS INPUT, AND ITS NAME IS THE WHOLE REASON NOBODY TREATS IT AS ONE. An EXIF UserComment, an XMP dc:title, an ID3 artist tag, a PDF /Title, a filename — every one is bytes the uploader chose, arriving labelled as a description rather than a request, and then rendered in an admin panel, written into a log line, interpolated into an export, used to name a download.

THE SECOND, AND IT IS THE SHAPE OF THIS WHOLE SECTION: THE HAZARD IS IN THE PAIR (BYTES, CONSUMER), NOT IN THE BYTES. Fourth time S11.9–S11.10 has met it:

S11.9.d   a payload has no format; a READER has a format
S11.10.b  "executes" is a property of what OPENS it
S11.10.c  a collision is a property of the DESTINATION
S11.10.d  a metadata value is a hazard in one SINK and inert in another

=1+1 is four harmless characters in an HTML page and a live formula the moment the same value reaches a spreadsheet export — an HTML escaper does nothing to it and is not wrong to. A newline is nothing in a caption and a forged log record in a log line. ${…} is inert in a renderer that escapes and a read of the environment in one that interpolates, and which of those a codebase uses is not a property of the value.

So every payload records the sinks it is dangerous in AND the sinks it is inert in, and the two lists must PARTITION the sink set exactly. A sink named in neither is one nobody has considered, and a register permitting that state would make "inert everywhere" and "never thought about" look the same — the same defect as reporting an unobservable invariant as held (S11.9.e) and as recording only where a path pair is dangerous (S11.10.c). No payload is dangerous everywhere (that would be a property of the bytes) and none is dangerous nowhere (that demonstrates nothing).

A NUL is filed as a WRITE THAT CANNOT HAPPEN rather than an injection: PostgreSQL refuses it outright, so the upload succeeds and the row does not.

THE THIRD: THE SIZE OF A METADATA FIELD IS NOT BOUNDED BY THE SIZE OF THE FILE. XMP is XML inside an image and XML nests, so a 2 kB JPEG can carry two thousand levels of nesting and exhaust a recursive walker from a file that passed every size limit. The oversize is in a part nothing weighs separately.

AND THE NUL IS BUILT WITH AN ESCAPE. A literal NUL byte in a source file makes it binary to grep, so a scan over it reports zero hits for the same reason a clean file does — a trap this repository has been caught by before. The first draft of the module carried a raw one; it was replaced, and the scan now checks this cell's own three files for raw NULs on every run, with a positive control that the files are non-empty so "no NULs" is not the silence of an unread file.

Registered five ways → 245 entry points, 0 violations; public-api.spec.ts 254 passing. -->

  • S11.10.e Add URLs/redirect/DNS fixtures for loopback/private/link-local/ reserved/metadata/internal, userinfo, IDNA, IP encodings, rebinding, redirect escape, alternate ports/schemes, and credential leakage. Evidence: SSRF corpus. <!-- S11.10.e — MOST OF THIS CELL WAS ALREADY BUILT, AND THE ONE THING MISSING WAS NOT A CATEGORY.

MEASURED FIRST, as S11.10.a and .c taught. libs/oshun/workbench-kit/src/ fetching.ts already carries: SSRF_CORPUS, 21 targets over 9 classes covering loopback, private, link-local, reserved, metadata and encoded IP forms; REDIRECT_POLICIES with validate-every-hop and a maximumRedirects bound (redirect escape); connectionWasPinned comparing validatedAddress against connectedAddress, which is the whole of DNS rebinding; withoutUserinfo (userinfo); redactUrl, looksSigned and SIGNED_URL_MARKERS (credential leakage); allowedSchemes and allowedPorts on RemoteFetchPolicy with scheme-outside-http and unbounded-port faults (alternate ports/schemes); and DEPLOYMENT_SUPPLIED_CLASSES marking internal-service as the deployment's to supply. Ten of the cell's categories were done. Only IDNA had zero cases — a grep for punycode, idna and xn-- across fetching.ts returns 0.

But IDNA turned out to be the symptom, not the finding.

THE RULE: A SPELLING IS A BYPASS ONLY AGAINST A GUARD THAT SEES IT.

Third consecutive cell to land on this shape, so it is a shape and not a coincidence. S11.10.c: whether a bundle contains a collision has no answer until the destination filesystem is named. S11.10.d: whether '=1+1' is dangerous has no answer until the sink is named. Here: whether 2130706433 is a bypass has no answer until somebody says WHERE THE GUARD STANDS.

runCorpus(refuses, corpus) hands the guard entry.target — the string somebody typed into the corpus. A deployment whose guard reads the raw user string and one whose guard reads new URL(input).hostname are two different systems, they fail to different inputs, and both call runCorpus and record the same green result.

WHAT WAS MEASURED — all 21 targets through Node's own WHATWG URL parser. EIGHT reach a post-parse guard as a string the corpus does not contain, three ways:

127.1 2130706433 0177.0.0.1 -> REWRITTEN to 127.0.0.1. Against a post-parse guard these three test nothing. They were real bypasses, against the inet_aton-era checkers they were written for, which read the string. ::ffff:127.0.0.1 ::ffff:169.254.169.254 -> REWRITTEN into hex: [::ffff:7f00:1] and [::ffff:a9fe:a9fe]. A post-parse guard comparing against the corpus string never matches. THE ONE DIRECTION WHERE STANDING AFTER THE PARSER IS THE WEAKER POSITION. ::1 fe80::1 fc00::1 -> BRACKETED. Raw new URL THROWS on a bare IPv6 literal — it is not a URL host. asTheResolverSeesIt brackets rather than reporting it unparseable, because refusing to answer would hide the address rather than reveal it. Same address, different string: a membership test against ::1 does not match [::1].

Not one of those eight is wrong. Each is wrong ONLY at an unstated vantage point, which is the whole point.

THE DEMONSTRATION, and it is decisive. Build the most literal guard a deployment could write from that corpus — membership in SSRF_CORPUS — and run the corpus against that ONE guard twice:

as-written   0 got through
as-resolved  6 got through: [::1] [::ffff:7f00:1] [::ffff:a9fe:a9fe]
             [fe80::1] [fc00::1] example.com

Five are forbidden addresses in another spelling. The sixth is the credentialed entry, whose target is a userinfo string rather than a destination; a guard SHOULD let example.com through and that entry's hazard is leakage, not where the socket goes — recorded in the spec rather than counted as a miss. Nothing about the corpus or the guard changed between those two runs. runCorpus reports the clean answer, correctly, for the only question it can ask.

THE PARSER CREATES SPELLINGS TOO. Circled digits are IDNA compatibility variants of the digits, so a name containing no digit and no dot resolves to 169.254.169.254 — the target that makes an SSRF a credential theft rather than a port scan. A tab inside a host is STRIPPED by the parser, joining two labels a reader sees as separate. Userinfo puts a friendly name where every reader, every log line and every indexOf check looks, and the metadata endpoint where the socket goes; a hash turns a trailing at-sign into decoration, so somebody who has learnt the userinfo trick reads that one backwards and calls it safe. A Cyrillic lookalike does NOT reach the name it resembles, which is why it is included: a denylist comparison fails and so does the attack, while an ALLOWLIST comparison also fails and refuses a legitimate name that renders identically.

AND ONE THAT NORMALIZING DOES NOT FIX. 127 written with U+3002 IDEOGRAPHIC FULL STOP separators is unchanged by NFKC — U+3002 is not a compatibility variant of the full stop — so a guard that normalizes and then checks still sees one dotless label while fetch sees 127.0.0.1. IDNA's own label-separator rule is what turns it into an address and the only thing in reach implementing that rule is the URL parser. Hence asTheResolverSeesIt: NORMALIZE BY PARSING, NOT BY NORMALIZING.

WHAT I GOT WRONG, TWICE, AND THE REGISTER CAUGHT BOTH. The reach-not-as-declared check parses every spelling's written and compares it to the declared reaches. On its first run it reported that idna/circled-digits-loopback reaches 128.0.0.1 and idna/circled-digits-metadata reaches 179.254.179.254 — I had typed a circled 8 for a 7 and a circled 17 for a 16. That is the entire reason the field is computed rather than believed. I also wrote in the module header that the parser THROWS on all five bare IPv6 targets, which is true of raw new URL and FALSE of asTheResolverSeesIt, which brackets them; the header now states all three outcomes separately and the census reports them as three numbers (5 rewritten, 3 bracketed, 0 refused) rather than one.

ALSO FOUND: the corpus prose names 'a trailing dot that changes DNS resolution but not string comparison' among its bypasses and the corpus contains no target ending in a dot. Supplied here as parser-rewrite/trailing-dot, and the scan re-reads that prose each run so the finding disappears the day somebody adds a case for it.

SHIPPED. libs/oshun/workbench-kit/src/url-spellings.ts — VANTAGE_POINTS with what each one sees, asTheResolverSeesIt, changesUnderParsing, 12 SPELLINGS over 5 classes (idna, numeric-form, deceptive-authority, parser-rewrite, confusable), readCorpusAtBothVantagePoints, runCorpusAt(point, refuses, corpus), 6 fault kinds, census and headline. Every spelling names the vantage points it fools AND the ones where it is inert, and the two must PARTITION the vantage-point set — same discipline as S11.10.d's sink partition and S11.10.c's returning both destination lists, so a vantage point nobody considered cannot look like one where the spelling is known to be harmless. A target the parser refuses is reported unreachable, not as a pass.

VERIFIED. 31 tests in src/url-spellings.spec.ts; kit suite 242 files / 15,383 tests green. tools/url-spellings-scan.mjs re-derives all four corpus counts by re-parsing every target, recomputes the two-vantage-point demonstration from scratch, verifies all 12 declared reaches against the parser, and checks this cell's own three files for raw NUL bytes; 8 negative controls, exits 2 if any goes undetected. Five registrations done — kit package.json exports, tsconfig.base.json path, root script verify:operations:v1-url-spellings, Nx target url-spellings-scan, export-map register; generator run: 246 entry points, 0 violations. -->

  • S11.10.f Add slow headers/body, chunked lie, no length, endless stream, timeout/retry, TLS/host mismatch, oversized compressed response, and malicious content-type fixtures. Evidence: transport corpus. <!-- S11.10.f — THE RULE: A TRANSPORT HAZARD IS A TRANSFER THAT NEVER ENDS, AND EVERY OBSERVATION THE ESTATE CAN MAKE IS A RECORD OF ONE THAT DID.

Six of the eight hazards this cell names are not responses at all. A slow-loris header dribble, a body arriving one byte a minute, an endless stream, a chunked encoding that never sends its terminating zero — each is a socket still open, still making progress by every measure the far side controls, and never finishing. Only two — an oversized compressed response and a lying content-type — can be held in your hand and judged.

MEASURED FIRST. fetching.ts is a serious piece of work and almost all of the machinery is there. What is missing is the thing that reads it:

FetchTimeouts 6 fields. ALL SIX READ, ALL SIX INSIDE validateFetchPolicy, WHICH COMPARES THEM TO EACH OTHER (phase <= total). Not one is ever compared to an elapsed measurement, because no function in the estate takes one. A policy may declare headersMs: 5000 and nothing can be handed "96 header bytes, 90 seconds". FetchBounds 5 fields. THREE HAVE NO READER AT ALL — compressedBytes, archiveEntries, archiveDepth. Verified by calling the estate's own decideContent: five gigabytes on the wire that decompress to four kilobytes comes back ACCEPTED, because the decompressed bound is satisfied, the ratio is below one, and the bound written for exactly this case is never read. FETCH_REFUSALS 11 declared, 4 returned by fetching.ts (all in decideContent), 6 left to the host through GuardedFetchSeam — legitimate for a seam — and timeout PRODUCED BY NOTHING IN THE REPOSITORY. No host could produce it either: nothing gives one an elapsed time. FetchOutcome fetched | refused. No third state. A transfer that is still going has neither, and the type cannot say so. ResponseObservation 4 fields, every one a TOTAL. A chunked lie is a disagreement between a declared length and what arrived, and there is no field for a declared length.

SHIPPED. libs/oshun/workbench-kit/src/transport-fixtures.ts — TransferProgress, the observation the estate cannot make (phase, elapsed total, elapsed in phase, time since the last byte, header and body bytes, what the far side claimed, and whether a chunked response is still unterminated); decideInFlight, THE FIRST PRODUCER OF timeout IN THIS REPOSITORY, phase-aware because ninety seconds in headers and ninety in body are refused by different clauses of one policy; TRANSFER_PHASES mapped 1:1 onto the FetchTimeouts fields, with totalMs the one deliberately unmapped because it bounds all of them; 9 cases over 9 hazards; 7 fault kinds; census and headline.

A null declaredLength is kept distinct from a zero one, because "would not say" and "said it would send nothing" are different claims and only the second makes one further byte an overrun. The chunked lie abandons with a NULL refusal, because FETCH_REFUSALS has no kind for a sender that exceeded its own declaration and borrowing one that means something else would file it under a wrong cause forever.

A THIRD DECIDABILITY, WHICH THIS CORPUS FOUND BY BEING WRONG. The TLS host mismatch was first filed as in-flight, and the register reported it not-abandoned: decideInFlight returns keep-waiting for it. THAT ANSWER IS CORRECT. A handshake that completes quickly to the wrong name is not slow, not large and not malformed; every bound the policy states is satisfied, and the fact that decides it — the name on the certificate — is in no type here. So it is filed from-neither, because "nothing in reach can decide this" is a different statement from "it passed". A corpus that had quietly retuned a threshold until the case abandoned would have shipped a working-looking check over a fact it never read. Same discipline as S11.9.e's fourth verdict and S11.10.c returning both destination lists.

AND A CHECK THE CORPUS NEEDED: caught-by-a-different-bound. THREE cases were written with numbers that abandoned for a reason other than the bound their own prose named — the retry case at nine seconds in connect against a connectMs of three, and the two slow cases pushed past the total — and all three ABANDONED, so the suite was green and each demonstrated something weaker than it claimed. A NUMBER THAT AGREES WITH A VERDICT IS NOT A NUMBER THAT AGREES WITH THE REASON. The check reads the decider's own detail line rather than re-deriving the decision, because a second implementation would agree with itself and not with the decider. All three cases were rebuilt to sit inside every other bound, so each now shows why its own bound exists.

TWO TRAPS, BOTH IN THE SCAN, BOTH ALREADY KNOWN TO THIS SECTION. git grep READS THE INDEX, so the scan's first run reported that its own module does not exist — twice — because the files were not staged; fixed with --untracked rather than by staging, so the answer no longer depends on whether somebody ran git add. And THE SCAN'S OWN QUERY CONTAINED THE LITERAL IT SEARCHED FOR: grepping for the timeout-refusal string found the grep. The pattern is now assembled at run time from two halves, which is better than an exclusion list because an exclusion list can also hide a real hit; a positive control confirms the query still matches a refusal known to be produced.

VERIFIED. 27 tests in src/transport-fixtures.spec.ts; kit suite 243 files / 15,410 tests green. tools/transport-fixtures-scan.mjs re-parses fetching.ts on every run to recount its timeout fields, its bound fields, which bounds have no reader and which refusals are returned — nothing is quoted from the register — and it re-derives validateFetchPolicy's line range from the source rather than hardcoding it. 9 negative controls, exits 2 if any goes undetected. Five registrations done — kit package.json exports, tsconfig.base.json path, root script verify:operations:v1-transport-fixtures, Nx target transport-fixtures-scan, export-map register; generator run: 247 entry points, 0 violations. -->

  • S11.10.g Assert quarantine/no parse before scans, resource bounds, stable refusal, redacted diagnostics, cleanup, and zero outbound/ persistent side effects. Evidence: fixture expectation matrix. <!-- S11.10.g — THE RULE: FIVE OF THESE SIX EXPECTATIONS CANNOT BE VIOLATED BY A FIXTURE, AND ALL SIX CAN BE VIOLATED BY WHAT RUNS IT.

Written as a matrix of fixtures × expectations this grades six greens and says nothing. MEASURED: all 239 modules under libs/oshun/workbench-kit/src — every one, not a sample — import ZERO capabilities. No node:fs, no node:child_process, no node:net, no database client. A frozen array cannot fail to clean up, cannot open a socket, cannot exceed a bound it never allocates. Grading it on those is not a weak test, IT IS NOT A TEST: the green is a property of the KIND of thing being graded. So the subject is wrong, and naming the right one is the cell.

WHAT ACTUALLY RUNS THE FIXTURES: 97 scan tools — 89 read the filesystem, 68 spawn a subprocess, 9 open a database connection, 0 make an outbound call, 4 WRITE.

TWO LIVE DEFECTS, BOTH FOUND HERE AND BOTH FIXED IN THIS COMMIT.

FIRST — game-day-scan.mjs called mkdtempSync on every invocation, four invocations a run, and there was NO rmSync anywhere in the file. Measured by counting the OS temp directory before and after a run: four more every time, 88 accumulated at 2.4 MB. It now removes the tree in a finally, because the interesting runs of that scan are the ones that throw. The 88 leftovers were swept. And this cell's scan RE-RUNS IT and re-counts rather than grepping for rmSync — a grep would pass on a file that calls it in a branch nothing reaches.

SECOND — seven diagnostics across the six corpora are built from their own input. Six use a lookup key or a hand-written corpus id: bytes a developer typed into this repository, and quoting them is how a caller finds their typo. ONE received a hostname supplied by whoever is being defended against — because that is the entire purpose of the function it lives in — and QUOTED it into its refusal message. asTheResolverSeesIt now reports the length and refuses to carry the bytes.

WHETHER ECHOING BYTES INTO A DIAGNOSTIC IS A DEFECT DEPENDS ON WHO CHOSE THE BYTES, AND A CHECK THAT ASKS "DOES THE MESSAGE CONTAIN THE INPUT" CANNOT TELL. Applied to all seven, a redaction rule strips the identifier out of every useful error in the kit. Applied to none, it misses the one that matters.

AND PROVENANCE IS NOT CARRIAGE, established by being wrong. The first register had ONE field for both, and the spec caught it: after the fix landed it went on reporting a leak, because the site still RECEIVES an attacker-chosen string — permanently, that is what the function is for — and no longer QUOTES it. bytesChosenBy is a fact about the data flow; quotesThem is a fact about the message. One field has to either forget where the bytes came from or keep reporting a leak that has been closed.

PERSISTENCE HAS THREE VALUES, NOT A BOOLEAN — removed, left-behind, intentionally-persistent — because a cleansUp flag marks the two tracked evidence files this section was ASKED to produce as failures, and would call deleting them a fix. Same shape as S11.10.f's third decidability and S11.10.c returning both destination lists.

TRAPS — THE SAME ONE TWICE, FROM OPPOSITE SIDES. A text search for outbound network access found a scan that SEARCHES for a fetch call and holds it as a string, so the searcher looked like a caller. And the writer census flagged THIS CELL'S OWN SCAN, because the pattern naming the writing functions appears as a regex literal inside it. Fifth time this section has been caught by a register that counts its own family. Both censuses are now decided by what a file IMPORTS, which is a capability, rather than by what it mentions, which is not.

SHIPPED. libs/oshun/workbench-kit/src/fixture-expectations.ts — EXPECTATIONS (6) with the SUBJECTS each binds to and why the others cannot violate it, vacuousForFixtures, DIAGNOSTIC_ECHOES (7) with bytesChosenBy and quotesThem, hostileInputSites and leakingEchoes, RUNNER_WRITES (4) over three persistence values, 5 fault kinds, census and headline.

VERIFIED. 22 tests in src/fixture-expectations.spec.ts; kit suite 244 files / 15,435 tests green. tools/fixture-expectations-scan.mjs recomputes the capability census by reading every src module and every scan tool, re-derives the echo list BY CALLING each diagnostic with a marked string, and re-measures cleanup by running game-day-scan and counting the temp directory before and after; a positive control confirms the capability query matches where a capability IS present. 8 negative controls, exits 2 if any goes undetected. Five registrations done — kit package.json exports, tsconfig.base.json path, root script verify:operations:v1-fixture-expectations, Nx target fixture-expectations-scan, export-map register; generator run: 248 entry points, 0 violations. -->

  • S11.11 Add provider fault-injection fixtures for timeout, throttle, malformed output, partial output, safety refusal, and cost overrun. Evidence: exact UI and API results. <!-- S11.11 CLOSED — a through g. Seven cells against a provider surface that was already substantial, and one shape underneath all of them.

THE SHAPE: THE ESTATE DECIDES WELL AND RECORDS THE DECISION IN A FIELD NOTHING FILLS IN, OR ON A KEY ONLY ONE SIDE CAN SUPPLY.

.a unknown is what the CALLER writes down when it stops waiting, and ProviderScriptStep lets the PROVIDER declare it — putting the caller's ignorance in the provider's mouth. Nothing compares elapsed time to a deadline, so a scripted hour still returns ok. .b the provider sent a Retry-After and RetryContext has eight fields, none of them it — so a caller burns all five attempts in 31.0s against a 60s window and providerRetryDecision answers with 999ms. .c six validators exist and correctionAllowed takes a family array NOTHING PRODUCES; the join is missing and the order is the whole thing. .d Usage is two required numbers, so a half-reported meter is rounded to zero and prices identically to a call that produced nothing; and safety defaults to allowed when the provider said nothing. .e refusalFaults is keyed on faultClass, which only a PROVIDER can populate, so working around a LOCAL refusal is invisible — and humanInitiated, the one legitimate way past a refusal, is read once and SET BY NOTHING. .f reserve compares a bare number to budgets that each carry a unit, and settle is a report that releases nothing, so an unknown outcome holds its room for ever. .g and FAULT_TABLE answers what to do next in six fields, exactly one of which is something a caller is told.

WHAT THE ESTATE ALREADY HAD, because crediting it first was the discipline that made every finding above precise. Full-jitter backoff with the randomness supplied. A retry-share ceiling with the best note in the library on why per-request bounds turn a partial outage into a full one. costOf throwing on a missing price and totalCost throwing on a null usage. A 1,471-line structured output validator with eighteen lexical faults and a semantic pass that refuses javascript: links. And the sentence this section kept coming back to: "an automatic re-ask with different wording after a safety refusal is not error handling: it is the system looking for a phrasing that gets the answer that was declined, run at machine speed with nobody in the sequence."

THE METHOD THAT KEPT WORKING: measure by CALLING, not by reading. Every cell's central claim is re-established on every scan run by driving the estate's own functions, so if somebody fixes one of these — widens Usage, types the estimate, writes the appeal path — the scan fails and the register is rewritten rather than going quietly stale.

A CORRECTION THE SECTION OWES: S11.11.c first reported SIX blocked corrections and the number was an artefact of its own context naming a trust zone that does not exist. The real number is ONE. The scan now checks the baseline first.

AND A DEFECT I SHIPPED AND CAUGHT FOUR TIMES: a fault kind whose condition is the negation of the branch it guards. .a had one, .b had two, .d had three, .f had one — seven in all, every one unsatisfiable by construction and every one sitting in a list looking like coverage. They are now swept invariants with positive controls. A FAULT KIND THAT NOTHING CAN PRODUCE IS THE SAME DEFECT AS A GATE THAT HAS NEVER FAILED.

7 modules, 7 specs, 7 scans, 5 registrations each. Kit suite 251 files / 15,625 tests green; 255 entry points, 0 violations. -->

  • S11.11.a Inject connect/first-byte/idle/total/tool/stage timeout before acceptance, during stream, after provider success, and with unknown outcome. Evidence: timeout fixtures. <!-- S11.11.a — THE RULE: unknown IS NOT SOMETHING A PROVIDER RETURNS. IT IS WHAT THE CALLER WRITES DOWN WHEN IT STOPS WAITING.

The estate says this itself and says it well. OUTCOME_MEANS.unknown in provider-adapter.ts: "the connection ended after the request was accepted and before the answer arrived; the work MAY have happened, been billed and been recorded upstream, and this side cannot tell." That is a statement about what THIS SIDE cannot tell — and ProviderScriptStep lets a fixture author write outcome: 'unknown', which puts the caller's ignorance in the provider's mouth.

MEASURED BY DRIVING createScriptedProvider, not by reading it:

  1. delayMs is advanced ONCE, BEFORE the chunk loop. Three chunks after a 30s delay all arrive at monotonic 30000. Every scripted delay is structurally a FIRST-BYTE delay, and a gap BETWEEN chunks — the whole of an idle timeout — cannot be expressed.
  2. Nothing anywhere compares elapsed time to a deadline. A step with delayMs: 3_600_000 and outcome: 'ok' returns ok, with its text and its usage, after a scripted hour. So "the provider succeeded and the caller had already given up" — one of this cell's four moments — has no way to occur.
  3. A scripted unknown returns usage: null. CORRECT for the caller's record, and it is the whole problem: the PROVIDER's ledger, which does know what it billed, has no representation. The fixture cannot say "5,000 output tokens were billed that this side will never see" — the fact that makes an unknown outcome expensive rather than merely annoying.

SHIPPED. libs/oshun/workbench-kit/src/timeout-fixtures.ts — ProviderLedger (what the provider did, with per-chunk arrival times and what it will bill regardless of what reached the caller) and CallerRecord (EVERY FIELD DERIVED by observe() from those arrivals against the caller's own deadlines). Nothing declares the caller's outcome, because the moment a fixture can, the cases where the two ledgers disagree stop being distinguishable from the cases where they agree.

The six points are six reductions of ONE timeline. connect and first-byte both reduce to "time before the first chunk", so a scalar delay expresses exactly those two AND CONFUSES THEM — and only one of them permits a free retry. 8 fixtures, 6/6 points, 4/4 moments, 18,980 billed output tokens no caller recorded.

FOUR AGREEMENT CLASSES, GRADED OVER THE CORPUS RATHER THAN PER FIXTURE — and this axis exists BECAUSE THE FIRST VERSION WAS WRONG. It graded every fixture on whether the ledgers DISAGREE and reported four defects against four CORRECT fixtures: a caller that stops before acceptance knows the work did not start, and a provider whose own connection dropped does not know either, so both sides writing the same thing is right in both cases. Disagreement is not a quality bar a fixture must clear; it is a third state. The fallback then swallowed a clean success and filed it as shared ignorance, which is why agreed-it-succeeded is a fourth class and is the POSITIVE CONTROL: without one in the corpus, an observer that answered unknown to everything would look correct.

AND A FAULT KIND THAT COULD NEVER FIRE. "The caller wrote failed after acceptance" had as its condition the exact negation of the one branch in observe() that writes failed, so it was unsatisfiable by construction and would have sat in the fault list looking like coverage forever. It is now an invariant about the OBSERVER, swept over 375 ledgers in both the spec and the scan — where the sweep IMMEDIATELY REFUTED IT AS FIRST STATED, because a provider that ends a call with an explicit failure inside every deadline HAS told the caller the work did not happen, which is the one case the caller can see. The invariant is about TIMEOUTS, and it now says so.

VERIFIED. 29 tests in src/timeout-fixtures.spec.ts; kit suite 245 files / 15,465 tests green. tools/timeout-fixtures-scan.mjs RE-DRIVES createScriptedProvider on every run for all three limits — so if somebody improves the provider double, which would be good news, this scan fails and the register is rewritten rather than going quietly stale — recomputes both corpus totals rather than reading the census, and re-sweeps the observer with a positive control (90 of 375 ledgers finish inside every bound and keep the provider's outcome, so an observer that answered unknown to everything would fail). 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 249 entry points, 0 violations. -->

  • S11.11.b Inject throttle/capacity with retry-after, quota reset, transient/terminal variants, repeated exhaustion, and circuit transition. Evidence: throttle fixtures. <!-- S11.11.b — THE RULE: A BACKOFF COMPUTED FROM THE ATTEMPT NUMBER IS A GUESS, AND THE PROVIDER SENT THE ANSWER.

CREDIT FIRST. provider-lifecycle.ts has a genuinely good retry surface: full-jitter backoff with the randomness supplied as a parameter so the schedule stays deterministic, a per-window retryShareCeiling with an excellent note on why per-request bounds are what turn a partial outage into a full one, ten refusals ordered by certainty, and throttle and quota as distinct classes.

MEASURED BY CALLING IT. RetryContext has EIGHT fields — faultClass, attempt, requestAgeMs, cancelled, idempotencyKey, windowRetries, windowCalls, reconciliation — and NOT ONE IS WHAT THE PROVIDER SAID. backoffDelayMs(attempt, budget, jitter) takes three arguments and none is a Retry-After. So a 429 that says "come back in 60 seconds" and a 429 with no header at all get the SAME delay, because the header has nowhere to go.

five-attempt budget, 1s base, 30s cap, maximum jitter: 999 + 1998 + 3996 + 7992 + 15984 = 31.0s across ALL FIVE attempts

Against a Retry-After: 60, every one is refused again before the window opens. The budget is spent, the load is added at exactly the moment the provider is shedding it, and the caller NEVER ONCE WAITED LONG ENOUGH. providerRetryDecision answers that sixty-second instruction with 999 MILLISECONDS.

THE SECOND, AND IT IS WHY ONE retryAfterMs FIELD WOULD NOT BE ENOUGH: A RETRY-AFTER IS A DURATION AND A QUOTA RESET IS AN INSTANT, AND JITTER MEANS THE OPPOSITE THING TO EACH.

A duration is measured from each caller's own response, so a thousand callers throttled across a second are already spread across a second when they return — jitter helps and is not load-bearing. "Your quota resets at 00:00 UTC" is the SAME MILLISECOND for every caller on the plan: they arrive together, and the first thing the provider does after restoring capacity is receive everything it refused while it was down. There jitter is the whole mechanism, and it must land AFTER the instant, because ARRIVING EARLY AT A RESET IS NOT EARLY — it is a refused request added to the moment the provider is least able to absorb one. The SAME HEADER in its HTTP-date form behaves like the reset rather than like its own duration form, which is why the signal KIND and not the header name decides whether spread is needed.

AND A CIRCUIT IS NOT A BUDGET. circuit appears nowhere in provider-lifecycle.ts, and retryShareCeiling is not one: a ceiling bounds how much of a window may be retries; a circuit stops asking altogether and then asks ONE question. That half-open probe EXPECTS to fail, so charging it to a caller's attempt budget spends a real user's retry on a health check — and a failed probe restarts the open window from NOW rather than from the original opening, because otherwise each window is shorter than the last and a failing provider is probed ever faster.

SHIPPED. libs/oshun/workbench-kit/src/throttle-fixtures.ts — SIGNAL_KINDS (5) with a perCaller shape each, ThrottleSignal carrying a duration OR an instant, waitFor() honouring what was said and spreading only shared instants, waitAsTheEstateDoes() for the comparison, attemptsWastedByGuessing(), CIRCUIT_STATES with circuitStep(), 8 fixtures over 6/6 aspects, 5 fault kinds, census and headline. 21 attempts wasted by guessing across the corpus.

TWO FAULT KINDS THAT COULD NEVER FIRE — "a per-caller duration got spread" and "spread landed before the instant" — were the negations of what waitFor does, the same defect S11.11.a found and the same fix: they are now invariants about the FUNCTION, swept over 312 waits at every jitter including out-of-range ones, with a positive control confirming spread IS applied somewhere (a waitFor that never spread anything would satisfy the second one trivially).

VERIFIED. 28 tests in src/throttle-fixtures.spec.ts; kit suite 246 files / 15,494 tests green. tools/throttle-fixtures-scan.mjs re-reads provider-lifecycle.ts for RetryContext's and RetryBudget's fields (with a positive control that the field queries find the fields that ARE there), re-calls backoffDelayMs and providerRetryDecision, recomputes the wasted-attempt total rather than reading the census, re-sweeps both invariants and walks the circuit through all three states. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 250 entry points, 0 violations. -->

  • S11.11.c Inject malformed syntax, wrong schema/version/type/unit/ref, duplicate/extra fields, unsafe URLs/tool calls, and semantic invariant failure. Evidence: malformed fixtures. <!-- S11.11.c — THE RULE: THE SIX FAULT FAMILIES ARE A SEQUENCE, NOT A SET, AND correctionAllowed TAKES A SET.

CREDIT FIRST, AND IT IS MOST OF THE CELL. structured-output.ts is 1,471 lines and covers nearly everything named here: scanJsonSource (18 lexical faults, including duplicate-key, which I expected it to miss); envelopeFaults for schema, version, discriminator, required and extra fields; resultSemanticFaults for units, citations, entity scope, cross-field invariants, links to undeclared hosts and links whose scheme is javascript:/data:/vbscript:/file:; preParseFaults for truncation and content-filter refusal; and a repair layer with MEANING_CHANGING_REPAIRS and a provenance check that will not let a repaired answer pass as an original.

CORRECTABLE_FAMILIES is ['lexical','envelope'], with a rationale worth quoting: "a model told 'that citation did not resolve' returns one that does, and the second answer is worse than the first because it now looks correct."

WHAT IS MISSING IS THE JOIN. MEASURED: outside index.ts — the barrel, not a caller — scanJsonSource, preParseFaults and resultSemanticFaults have ZERO callers in this repository. correctionAllowed(families) takes an array of families and NOTHING ANYWHERE PRODUCES ONE.

AND THE ORDER IS THE WHOLE THING. A document that did not parse has no envelope to check; a wrong envelope has no coherent result for the semantic pass. A classifier that unions all six reports facts DERIVED FROM A DOCUMENT THAT WAS NEVER VALID — and hands them to the function that decides whether to retry.

correctionAllowed(['lexical'])             -> allowed: TRUE
correctionAllowed(['lexical','semantic'])  -> allowed: FALSE, blockedBy: [semantic]

TWO COSTS, AND ONLY ONE IS A CHANGED VERDICT.

VERDICT (1 case). A trailing comma AND a bad citation: ordered stops at lexical and permits a retry; the union salvages the comma, reads the citation out of the salvage, adds semantic, refuses. Whether that citation survives a retry is exactly what nobody can know until the document parses.

REASON (3 cases). A truncated body is ONE fault ordered and THREE unioned, two of them read from a fragment. The verdict agrees and the record does not, and the record is what somebody reads at three in the morning.

A CORRECTION I OWE THE READER. The first version of this cell reported SIX blocked corrections. That was an artefact of its own ResultSemanticContext setting readmittedAs: 'model-generated' — NOT a TrustZoneKind; the real one is prior-generated-output. Every case was firing output-readmitted-above-a-generated-zone, and the "blocking" was my broken baseline agreeing with itself. With the zone corrected the count is ONE, and it took a deliberately compound case to get that one. The effect is real and much smaller than I first reported. The scan now checks the baseline FIRST.

A LIVE FINDING. EnvelopeContract lists requiredFields and allowedFields and NO TYPES, so envelopeFaults({kind:'finding',schemaVersion:2,title:42}, CONTRACT) returns an EMPTY LIST — verified by calling it, with a positive control that the same checker does refuse a wrong discriminator. A required field present with the wrong type passes every check in the estate and a renderer turns it into the string "42".

TOOL CALLS ARE THE ONE GENUINE GAP. The word tool appears once in structured-output.ts. A tool call is not a field in a result — it is a REQUEST, and its arguments are the model's text arriving at something that will act on them. toolCallFaults refuses a tool that was not offered, arguments that do not parse, a path escaping the workspace, the same non-location scheme set the estate refuses in a result link, and the identical call twice in one turn.

SHIPPED. libs/oshun/workbench-kit/src/malformed-fixtures.ts — CLASSIFICATION_ORDER with ORDER_REASONS saying what each rung needs the previous one to have established, classify() stopping at the first family that fails and reporting BOTH what it found and what it never reached, classifyEverything() for the comparison, toolCallFaults(), 15 cases held as TEXT (S11.9.b's lesson) over 11/11 kinds, 5 fault kinds, census and headline.

TRAPS, BOTH KNOWN TO THIS SECTION. The scan's family-array query found ITSELF — the SIXTH instance of a register counting its own family — fixed by assembling the pattern at run time rather than by an exclusion, because an exclusion can also hide a real hit. And A SPACE TYPED INSIDE A TEMPLATE LITERAL LANDED IN THE MODULE AS A RAW NUL BYTE, which makes the file binary to grep so every scan over it reports zero hits for the same reason a clean file does. This cell's own NUL check caught it; the separator is now written as an escape.

VERIFIED. 30 tests in src/malformed-fixtures.spec.ts; kit suite 247 files / 15,525 tests green. tools/malformed-fixtures-scan.mjs re-greps for callers of each checker on every run (with a positive control that the query finds callers where they DO exist), re-derives both union costs per case rather than reading the census, and checks the baseline before anything else. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 251 entry points, 0 violations. -->

  • S11.11.d Inject partial stream/output/usage/safety metadata, truncated finish, connection loss, cancellation, and inconsistent provider status. Evidence: partial fixtures. <!-- S11.11.d — THE RULE: A PARTIAL FACT MUST BE ROUNDED TO A WHOLE ONE IN EITHER DIRECTION, AND BOTH DIRECTIONS ARE WRONG.

CREDIT FIRST. The estate is careful about WHOLLY absent facts: costOf throws on a missing price ("a cost is a measurement times a price and a missing price is a refusal, not a zero") and totalCost throws on usage: null ("a partial sum wearing a total's clothes"). Both right.

MEASURED BY CALLING IT. Usage is {inputTokens: number; outputTokens: number} — two REQUIRED numbers, no nulls. A provider that reports its input count and dies before its output count, which is what a cut stream looks like, has exactly two encodings and both destroy information:

outputTokens: 0   -> costOf(...) = 0.003, and a call that GENUINELY produced
                     no output ALSO costs 0.003. Understated, and reported
                     as measured.
usage: null       -> totalCost throws, correctly, and the input count the
                     provider DID send is discarded with it. A partial
                     measurement becomes no measurement.

The honest answer is neither: AT LEAST 0.003, UPPER BOUND UNKNOWN. costRange returns that — a floor, a ceiling that may be absent, and measured kept separate from the bounds so a dashboard can add up what it knows and say how much of the total is a floor rather than a figure. It composes: a total over ranges is a range.

THE SECOND: THE DEFAULT ASSERTS A VERDICT NOBODY GAVE. createScriptedProvider writes safety: step.safety ?? 'allowed', so a provider that returned no safety verdict is recorded as having ALLOWED the content, and SAFETY_VERDICTS has no value for "it did not say". readSafety adds not-reported; permitsUse refuses it.

THE THIRD: AN INCONSISTENT STATUS IS NOT A BAD VALUE. No single field can be inconsistent with itself — it is FOUR CHANNELS disagreeing (transport, body, finish-reason, stream-terminator), and the only way to see it is to read all four and KEEP the disagreement. Collapsing to one boolean is what makes it invisible, and a 200 carrying an error body is the case every HTTP client reports as a success.

A FOURTH CHANNEL VERDICT THIS CORPUS FOUND BY BEING WRONG. finishReason: 'length' was first classed as a channel saying the call FAILED, so every truncated-but-successful response came back INCONSISTENT — while the fixture's own prose already said it "dissents about COMPLETENESS rather than about success". A call cut at the token limit ran, billed and terminated cleanly. says-it-was-cut-short is that distinction; consistency is now a question about SUCCESS and incompleteness is a second question, reported separately.

SHIPPED. libs/oshun/workbench-kit/src/partial-fixtures.ts — PartialUsage, costRange, costAsTheEstateWouldRound for the comparison, SAFETY_READINGS with a fourth value, readSafety/permitsUse, STATUS_CHANNELS with channelVerdict and readStatus, 9 fixtures over 8/8 aspects, 3 fault kinds, census and headline. 5 disagree about success, 2 report an incomplete answer, 5 carry no safety verdict, and 5 cannot be priced to a figure — the estate would report 0.0105 for them AS MEASURED. The corpus holds a REAL zero as well as an absence so the two are exercised apart.

THREE FAULT KINDS THAT COULD NEVER FIRE — each the negation of what costRange does — became swept invariants over 25 usages with a positive control, the same defect and fix as S11.11.a and .b. The third is load-bearing: the rounded cost MUST equal the honest floor, because the finding is that the number a caller gets today IS the floor presented as the figure.

VERIFIED. 26 tests in src/partial-fixtures.spec.ts; kit suite 248 files / 15,551 tests green. tools/partial-fixtures-scan.mjs re-reads Usage's field types from the source, re-calls costOf and the scripted provider for both central claims, recomputes every corpus total rather than reading the census, and re-sweeps the three invariants. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 252 entry points, 0 violations. -->

  • S11.11.e Inject provider and local safety/policy refusal classes with nonretryable behavior, appeal/review, and no prompt-around fallback. Evidence: refusal fixtures. <!-- S11.11.e — THE RULE: THE DETECTOR IS KEYED ON WHO REFUSED, AND WORKING AROUND A REFUSAL IS THE SAME ACT WHOEVER MADE IT.

CREDIT FIRST, AND IT IS SUBSTANTIAL. provider-lifecycle.ts has the best piece of writing in this library on the subject: "an automatic re-ask with different wording after a safety refusal is not error handling: it is the system looking for a phrasing that gets the answer that was declined, run at machine speed with nobody in the sequence — and it is reached by ordinary-looking code that regenerates on failure." The machinery under it is real: AttemptRecord carries a promptDigest precisely so the comparison is possible, refusalFaults reports four kinds, and FAULT_TABLE marks the three refusal classes retryability: 'no' because a refusal is a decision, not a failure.

TWO MEASUREMENTS, AND THEY FIT TOGETHER BADLY.

FIRST — PROVIDER_FAULT_CLASSES is eleven values and every one is something a PROVIDER does. A refusal made by THIS SYSTEM — an entitlement check, a withdrawn consent, a rights holder's terms, a local policy engine — has no class, so the attempt following it carries faultClass: null and refusalFaults SKIPS IT ON ITS FIRST LINE. Verified by calling it with the identical attempt pair twice:

provider safety refusal, then a reworded prompt
    -> ['prompt-changed-after-a-refusal']
LOCAL refusal, then the same rewording
    -> []

The same act. The second is arguably worse, because it is the refusal this system is answerable for and nobody outside ever sees the loop.

SECOND — humanInitiated is the ONE exemption the rule grants: "a person who reads a refusal and edits their request is doing something the system should support; the fault is the loop that does it without them." It is declared on AttemptRecord, READ ONCE in refusalFaults, and SET BY NOTHING in this repository — checked with a positive control that the same query finds writers of a field that has one.

SO THE ONLY LEGITIMATE WAY PAST A REFUSAL HAS NO PRODUCER, AND THE FORBIDDEN WAY IS UNDETECTABLE WHENEVER THE REFUSAL WAS LOCAL. What is reachable is exactly the thing the rule was written to stop.

SHIPPED. libs/oshun/workbench-kit/src/refusal-fixtures.ts — REFUSAL_ORIGINS, LOCAL_REFUSAL_CLASSES kept deliberately OUT of ProviderFaultClass (adding them would say a provider can refuse on our entitlement rules), an Attempt extending the estate's record with a refusedBy either side can populate, an origin-blind worksAround that keeps the humanInitiated exemption exactly as written, asTheEstateSees for the comparison, APPEAL_ROUTES with per-route producers, 7 fixtures, 5 fault kinds, census and headline.

The routes are honest about their limits: a provider's safety refusal is not ours to overturn, so the only route against it is an out-of-band support ticket measured in days — which is precisely why an automated loop is so tempting and so wrong. And a withdrawn consent is the one refusal in the corpus with NO route at all, because there is nobody here who may overturn it.

THE NUMBERS. The origin-blind detector catches 4 and the estate's catches 1, so THREE WORKAROUNDS ARE INVISIBLE TODAY, and 3 of 3 appeal routes have no producer. Those three faults are reported on every run and they ARE the finding rather than a defect in the corpus — the spec asserts exactly three, all of them the appeal gap. An ordinary retry after a timeout is the positive control, without which a detector that reported everything would look correct on all six other cases.

VERIFIED. 20 tests in src/refusal-fixtures.spec.ts; kit suite 249 files / 15,573 tests green. tools/refusal-fixtures-scan.mjs re-calls refusalFaults with both attempt pairs on every run and re-greps for a writer of the exemption with a positive control — so if somebody adds a local refusal class or writes the appeal path, either of which would be good news, this scan fails and the register is rewritten. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 253 entry points, 0 violations. -->

  • S11.11.f Inject estimate-underrun, reservation expiry, per-attempt overrun, missing metering, currency/rate mismatch, budget exhaustion, and kill switch. Evidence: cost fixtures. <!-- S11.11.f — TWO RULES, AND THEY ARE INDEPENDENT.

THE FIRST: A BUDGET CARRIES ITS UNIT AND THE ESTIMATE COMPARED AGAINST IT DOES NOT.

SpendBudget has a unit: string. reserve(budgets, estimate, ...) takes ONE BARE NUMBER and compares it to every scope's remaining room. MEASURED, with a tenant budget in US cents and a provider budget in vendor credits:

reserve([tenant 10,000 usd-cents, provider 50 vendor-credits], 40)
    -> RESERVED. Forty US cents now charged against a credit budget.
reserve([the same two], 100)
    -> REFUSED, "the provider budget has 50 vendor-credits left" —
       while the 10,000-cent tenant budget sits untouched.

Both answers are wrong, in opposite directions. And THE REFUSAL MESSAGE PRINTS THE UNIT: the code knows units exist at the exact moment it declines to compare them.

THE SECOND: A RESERVATION IS A PROMISE TO SPEND, AND EVERY PATH THAT ENDS WITHOUT A SETTLEMENT HOLDS IT FOR EVER.

settle(reservationId, estimated, actual) takes three arguments — no budget, no time — and returns a REPORT. It is a good report and its overran comment is right. But NOTHING IN limits.ts ASSIGNS TO A BUDGET'S reserved FIELD (0 assignments, measured), and a reservation has no expiry: the exception has an expiresAt and the reservation does not.

So S11.11.a's case arrives here. A call whose outcome is unknown has no actual and never will, so settle can never be called, and the room is held for ever — the budget shrinks by one estimate every time a provider call ends ambiguously, which is exactly the failure mode a retry budget exists to survive.

SHIPPED. libs/oshun/workbench-kit/src/cost-fixtures.ts — Money (an amount WITH its unit), reserveInUnits as a PRECONDITION rather than a replacement (once the units agree it calls the estate's own reserve), incomparable as a fourth outcome because "refused" says waiting will help and "incomparable" says nobody can tell, a Reservation with holdsUntil, and resolveReservation with four endings — settled / released / expired / still-held, the last REPORTED rather than treated as an ending because an unresolved reservation is the state this module exists to make visible. A settlement whose actual is in a different unit does not settle either. A kill switch is kept distinct from an empty budget, because zeroing a ceiling tells somebody their budget ran out when it is full. 10 fixtures, 7/7 aspects, 4/4 endings, 4 fault kinds.

ONE FAULT KIND WAS REMOVED for being unreachable AND malformed — it asked whether an EMPTY ledger reports room, a question with one answer — and became three swept ledger invariants. The third is this cell's own opening defect arriving at the other end of the module: a held-room total across two units names NO unit rather than adding cents to credits.

VERIFIED. 25 tests in src/cost-fixtures.spec.ts; kit suite 250 files / 15,599 tests green. tools/cost-fixtures-scan.mjs re-reads reserve's and settle's SIGNATURES from limits.ts, re-runs both mixed-unit reservations, recounts the assignments to reserved, and re-sweeps the ledger invariants — where the third one caught a bug in the scan itself: the reservation it built to make a two-unit ledger expired immediately, so heldRoom filtered it out and the ledger looked single-unit. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 254 entry points, 0 violations. -->

  • S11.11.g Assert exact typed API status/code/state/provenance/correlation/ retry metadata and exact UI utility state/actions/no fabricated result. Evidence: API/UI snapshot matrix. <!-- S11.11.g — THE RULE: THE TABLE ANSWERS WHAT TO DO NEXT AND THE MATRIX ASKS WHAT TO SAY.

MEASURED. FAULT_TABLE classifies all eleven provider fault classes and its entries carry exactly SIX fields:

faultClass  retryability  family  outcomeUnknown  mayHaveBeenCharged  isRefusal

Every one is a decision THIS SYSTEM makes — retry or not, is the outcome ambiguous, might we have been billed, is this a decision rather than a failure. NOT ONE IS SOMETHING A CALLER IS TOLD.

Of this cell's six API fields, exactly ONE is a lookup: retry metadata, through retryability, which is a Retryability imported from errors.ts so the two vocabularies already agree. The other five are DECIDED here. RpcStatus lives in errors.ts and the table names an ErrorFamily instead — provider covers a throttle, a timeout and a safety refusal, which are resource-exhausted, deadline-exceeded and failed-precondition. There is no error code. And there cannot be a correlation id: it belongs to the REQUEST, not the fault, so it can never come from a table keyed by fault class — which is exactly why it must be carried explicitly and exactly what gets dropped.

None of that is a defect in the table. A table that decided what to render would be the wrong table. It is a statement about where the matrix comes from, and this cell is where those decisions are written down in the open rather than by whichever component renders first.

THE ONE RULE THAT IS NOT A CHOICE: NO FABRICATED RESULT. outcomeUnknown makes it load-bearing — a call whose outcome is unknown MAY have produced a result and MAY have been billed, so a screen that renders one is showing something nobody computed. Every row carries what the UI shows AND whether that presentation can contain a result. A partial result is permitted exactly when it is MARKED partial. No refusal row offers a retry, because a retry button on a refusal is S11.11.e's prompt-around loop with a person clicking it. And a screen that knows an instant does not show a spinner.

AND THE STATUS CANNOT DECIDE THE SCREEN, which is why this is a matrix and not a lookup table:

internal            -> retrying | failed | result        (3 screens)
resource-exhausted  -> waiting-until | blocked-needs-a-person

A throttle is waited out in seconds; a budget is not waited out at all, somebody raises it. malformed/lexical and malformed/semantic share a fault class AND a status and differ in code and screen, because a model told its citation did not resolve returns one that does. A local refusal takes permission-denied where a provider's takes failed-precondition, and its only action is a reviewer.

SHIPPED. libs/oshun/workbench-kit/src/api-ui-matrix.ts — API_FIELDS with FIELD_SOURCES recording per field whether it is a lookup or a decision, UI_STATES with STATES_THAT_SHOW_A_RESULT, UI_ACTIONS, 12 rows over all 6 families S11.11.a–.f injected, 7 fault kinds, census and headline. Every status, retryability, state, action and fault class is taken from the estate's own vocabularies and checked against them on every scan run.

VERIFIED. 25 tests in src/api-ui-matrix.spec.ts; kit suite 251 files / 15,625 tests green. tools/api-ui-matrix-scan.mjs reads FAULT_TABLE's ENTRY SHAPE on every run and recomputes which of the six fields it can answer — so if somebody adds a status or a code to the table, which would be good news, the "one of six" becomes wrong out loud rather than quietly. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 255 entry points, 0 violations. -->

  • S11.12 Add queue fault-injection for lease loss, duplicate delivery, out-of-order events, retry exhaustion, and cancellation race. Evidence: no duplicated side effects. <!-- S11.12 CLOSED — a through g, against the best-argued code in this library.

THE SHAPE: EVERY FINDING HERE IS ABOUT WHAT A CORRECT DESIGN CANNOT EXPRESS, NOT ABOUT IT BEING WRONG. jobs.ts, idempotency.ts, inbox.ts and retries.ts are excellent, and this section quotes them at length because the quotations ARE the argument. Seven times the answer was: the mechanism is right, the vocabulary it decides over is one thing short.

.a CHECKPOINT_ORDERINGS deliberately omits checkpoint-first, so the silent failure the module names CANNOT BE WRITTEN DOWN by a fixture. A type that makes a bad state unrepresentable makes it uninjectable. .b decideClaim returns claimed to BOTH racers, correctly — and the estate names the layer that decides that race for revisions and not for idempotency keys, while scopeKey already produces the six-tuple. .c OrderingRequest's seven fields are all one subject's one sequence, so a cross-subject dependent event is APPLIED. Order is a fact about a sequence; a dependency is a fact about two. .d classifyJobFailure takes two arguments and cannot count, so "transient repeats" and "permanent" are the same answer — and defer protects the queue from its jobs while nothing protects a dependency from the queue. .e JobRecord has one coarse state, so all four positions inside running get one answer and cost four different things. A cancel after the effect loses the RECORD, not the effect. .f four of ten invariants are evaluable from a corpus of faults and six are not, and saying which is which is the whole value of the suite. .g and a replay is authorized by a capability and made safe by a reconciliation, and only the first is checked — narrowed, because planDeadLetterReplay preserves the idempotency key and catches every replay whose first attempt left a record.

WHAT THE ESTATE ALREADY HAD, quoted because crediting it first is what made each finding precise: "expiry protects the QUEUE; only a monotonically increasing token, checked at the WRITE, protects the DATA" · "a stored record in completed with a different request digest is not a replay" · "a gap this new is almost always an event still in flight" · "a deferred job must not burn an attempt" · "collapsing the two tells the operator the effects have stopped at the exact moment they have not" · "this is the cell people fill in with retry because the alternative needs a person, and it is the cell that produces the duplicate charge".

MY OWN DEFECT, ELEVEN TIMES, AND NOW A RULE. A fault kind whose condition no input can reach — because it checks a module's own constants rather than the input. Seven in S11.11, four more here. All are swept invariants now, and the rule is stated where the eleventh was found: A FAULT KIND MUST RANGE OVER THE INPUT; A STATEMENT ABOUT A MODULE'S OWN CONSTANTS IS AN INVARIANT.

AND A FAMILY OF QUERY BUGS, five in this section, every one producing a wrong answer for a structural reason rather than a factual one: a JSON string's character count compared to a column count; a wrapped block comment whose sentence no flat regex matches; a guessed function name whose empty match passed a negative check; a non-greedy brace match that stopped at an inline parameter type; and a slice that ran into the next section's banner. A ZERO FROM A QUERY NOBODY POSITIVE-CONTROLLED IS NOT A MEASUREMENT, AND NEITHER IS A ONE.

7 modules, 7 specs, 7 scans, 5 registrations each. Kit suite 258 files / 15,789 tests green; 262 entry points, 0 violations. -->

  • S11.12.a Inject lease expiry before/during/after checkpoint or external effect, failed heartbeat, fencing violation, worker pause/death, and new worker takeover. Evidence: lease fixtures. <!-- S11.12.a — THE RULE: A TYPE THAT MAKES A BAD STATE UNREPRESENTABLE ALSO MAKES IT UNINJECTABLE, AND FAULT INJECTION IS THE ONE PLACE THAT MATTERS.

CREDIT FIRST, AND IT IS THE WHOLE SETUP. jobs.ts is 1,929 lines and the best-argued module in this library. Decision one is this cell's subject: "A LEASE IS NOT A LOCK, AND WHAT MAKES IT SAFE IS THE FENCING TOKEN. A worker whose lease expired does not know its lease expired: it was paused — a collector, a stalled syscall, a partition — the coordinator handed the job to somebody else, and then it wakes up and writes. Expiry protects the QUEUE; only a monotonically increasing token, checked at the WRITE, protects the DATA."

Decision three creates the problem: "effect-then-checkpoint REPEATS the effect on resume, which idempotence answers; checkpoint-then-effect SKIPS it, and nothing anywhere reports a stage that was never run. CHECKPOINT_ORDERINGS has no checkpoint-first member."

MEASURED: exactly two members, and the dangerous third absent by design.

THE CONSEQUENCE. That is right for production and it means the silent failure CANNOT BE WRITTEN DOWN. S11.12 asks for FIXTURES that inject a lease fault, and a CheckpointOrdering cannot hold checkpoint-first — so no fixture can put a system into the state whose whole danger is that nothing reports it.

A CORPUS THEREFORE NEEDS ITS OWN VOCABULARY, DELIBERATELY WIDER, AND HAS TO SAY SO. INJECTABLE_ORDERINGS is production's plus the forbidden member; productionCannotExpress marks every use; the two are never merged; and the scan checks on every run that NOTHING outside this cell imports the wider list — because the day it becomes an importable production type is the day the constraint stops holding.

THE SECOND MEASUREMENT: claimJob has ZERO callers outside a spec. The lease machinery is complete, argued and driven by nothing — the same shape S11.10.a found for the archive checkers and S11.11.c for the output validators.

SHIPPED. libs/oshun/workbench-kit/src/lease-fixtures.ts — POSITION_COSTS, a table over expiry position × ordering whose argument IS the table: the same position costs a REPEAT under one ordering and a SKIP under the other, and only the repeat is visible. wouldBeAccepted restates fencing so a fixture can assert on it, refusing a token EQUAL to the highest seen as firmly as a lower one, because two workers holding the same token is the state fencing exists to prevent. 9 fixtures over 6/6 events, 4/4 positions, 4/4 loss kinds; 7 fault kinds. ONE fixture uses the forbidden ordering and it is the SAME ONE whose loss nothing reports.

THREE DELIBERATE CHOICES. The dead-worker fixture records in its own reason that it proves NOTHING alone — a dead worker never writes, so a system with no fencing passes it — which is why the paused-worker case sits beside it. The clock-skew fixture puts the worker 45s behind so its lease looks live to it and expired to the coordinator, and the verdict still turns on tokens rather than on either clock. And the one ordering with NO bad position is in the corpus so the others are not made to look inevitable.

VERIFIED. 22 tests in src/lease-fixtures.spec.ts; kit suite 252 files / 15,648 tests green. tools/lease-fixtures-scan.mjs re-reads CHECKPOINT_ORDERINGS on every run (so adding checkpoint-first fails loudly), greps for importers of the wider vocabulary, and sweeps 25 token pairs with a positive control — a check that refused everything would satisfy every corpus case. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 256 entry points, 0 violations. -->

  • S11.12.b Deliver same command/event/job concurrently, after ack loss, after success, after retry, and with altered payload under same ID. Evidence: duplicate fixtures. <!-- S11.12.b — THE RULE: decideClaim RETURNS claimed TO BOTH RACERS, AND THE ESTATE NAMES THE LAYER THAT DECIDES THAT RACE FOR REVISIONS AND NOT FOR IDEMPOTENCY KEYS.

CREDIT FIRST, AND IT IS MOST OF THE CELL. IdempotencyScope is five fields plus a target with a note on what dropping each one costs. And decideClaim COMPARES THE DIGEST BEFORE THE STATE: "a stored record in completed with a different request digest is not a replay — it is a caller reusing a key for something else, and returning the stored result reports success for something that never ran, to somebody who will never look again." So altered-payload-under-the-same- ID, this cell's sharpest item, is covered — and inbox.ts answers it a second time by comparing payloadHash before admitting an event.

WHAT IS NOT ANSWERED IS CONCURRENCY, AND THE MODULE IS RIGHT NOT TO ANSWER IT.

MEASURED by calling it: decideClaim(null, digest, at) returns claimed, so two callers racing before either has written both see nothing and BOTH are told to proceed. That is correct. A pure function over "whatever record already exists" cannot decide a race, and pretending it could would be worse than not trying.

The race is decided by a uniqueness constraint at the storage layer — and the estate says so IN THIS VERY MODULE, about a different constraint:

unique-sequence-per-entity, layer: 'table'
"UNIQUE (entity_id, sequence); this is the constraint that decides an
 append-only race, and a SELECT-then-INSERT does not"

CONSTRAINT_PLACEMENT holds EIGHT entries and every one is a REVISION constraint. MEASURED: not one names the idempotency scope. And scopeKey already produces the exact six-tuple such an index would be over — tenant, client, operation, operation version, target, key — so the thing to place is written down and its placement is not.

SHIPPED. libs/oshun/workbench-kit/src/duplicate-fixtures.ts — SCOPE_PLACEMENT, stating that layer in the same shape CONSTRAINT_PLACEMENT uses, with the estate's own sentence and none of the words changed; ARRIVAL_KINDS, PAIR_OUTCOMES, and 6 delivery pairs over 5/5 kinds producing 5/5 outcomes, of which exactly ONE — the concurrent one — leaves both callers claimed while every other is settled by the decision function alone. 5 fault kinds.

A SEVENTH FAULT KIND THAT NO INPUT COULD REACH was removed and asserted directly instead — it compared two of this module's own constants, which makes it a consistency check rather than a finding. AND IT CAUGHT A REAL BUG ON THE WAY IN, worth recording because it is the shape this section keeps finding elsewhere: the first version compared scopeKey(scope).length to the column count, and scopeKey returns a JSON STRING — so it was comparing twenty-four CHARACTERS to six COLUMNS. Two quantities that are not the same KIND of thing, compared because both are numbers.

VERIFIED. 21 tests in src/duplicate-fixtures.spec.ts; kit suite 253 files / 15,670 tests green. tools/duplicate-fixtures-scan.mjs re-calls decideClaim for both racers AND for a caller who arrives after the record exists — the positive control, since a function that always said claimed would satisfy the first line — and re-reads CONSTRAINT_PLACEMENT on every run, so adding the scope placement fails loudly. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 257 entry points, 0 violations. -->

  • S11.12.c Deliver sequence gaps, late predecessor, reordered independent/ dependent events, stale version, replayed tombstone, and partition reorder. Evidence: ordering fixtures. <!-- S11.12.c — THE RULE: ORDER IS A FACT ABOUT A SEQUENCE, AND A DEPENDENCY IS A FACT ABOUT TWO.

CREDIT FIRST, AND IT IS FOUR OF THE SEVEN ITEMS. inbox.ts's decideOrdering has five dispositions and the note that carries the whole design: "A GAP IS NOT A LOSS UNTIL IT IS OLD. Position 5 arriving with 4 unseen is almost always 4 still in flight, so 5 is held; treating the gap as a loss immediately turns every reordering into an incident. Past the tolerance it stops being a wait and becomes a fact somebody has to be told."

So a sequence gap is hold-for-gap; a late predecessor is the same thing named from the other end; a reordered INDEPENDENT event is apply-out-of-order when its kind declares commutes WITH A REASON (the estate requires one, because "an unexplained true is what somebody set to stop the queue filling up"); and a stale version is ignore-superseded.

MEASURED, AND THIS IS THE GAP. OrderingRequest has SEVEN fields — appliedThrough, held, sequence, ordering, now, gapFirstSeenAt, gapToleranceSeconds — and every one is about ONE SUBJECT'S ONE SEQUENCE. Called with a dependent event that is position 1 on its own subject and whose predecessor lives on a DIFFERENT subject:

{ disposition: 'apply', detail: 'the next position in the subject' }

AND IT IS RIGHT. On its own sequence it IS the next position, and there is no field in the request that could mention the other subject, so the question was never asked.

FOURTH TIME THIS SECTION HAS MET THIS SHAPE, so it is a shape. S11.9.b: a corpus of JSON objects cannot hold a duplicate key. S11.9.c: a corpus of lone manifests cannot hold a relationship fault. S11.10.c: a corpus of single paths cannot hold a collision. And here: a request carrying one sequence cannot hold a cross-sequence dependency.

SHIPPED. libs/oshun/workbench-kit/src/ordering-fixtures.ts — Dependency, AppliedThroughBySubject (a MAP, because the point is that there is more than one), and decideWithDependencies, which asks the ESTATE'S question first and the cross-sequence one second. That order is deliberate: A DEPENDENCY DOES NOT OVERRIDE A GAP. An event whose own predecessor is missing holds for that regardless, and a superseded event stays superseded, so the dependency check only ever turns an apply into a hold. Unmet dependencies are reported even when the within-sequence answer already holds, because the two release at DIFFERENT MOMENTS and a consumer knowing only the first would release too early. 8 fixtures, 7/7 hazards, 6/6 dispositions against the estate's 5, 6 fault kinds.

THREE ARE DECIDED DIFFERENTLY once a second sequence is in scope — the dependent event, the replayed tombstone and the partition reorder — and ALL THREE get apply from the estate. The tombstone is the sharpest: on its own stream it is position 1 and perfectly in order, and what makes it wrong is the state of a subject it names. A partition reorder is cross-sequence by construction, since each partition is internally ordered and the relationship between them is what no single sequence carries.

A fixture with a SATISFIED dependency is the positive control, without which a decider that held everything carrying a dependency would look correct on every other case. And every dependency must say why, because one that does not holds the queue for ever and nobody can tell whether removing it is safe.

VERIFIED. 22 tests in src/ordering-fixtures.spec.ts; kit suite 254 files / 15,693 tests green. tools/ordering-fixtures-scan.mjs re-reads OrderingRequest's FIELD LIST from a fixture rather than restating it (so widening the request fails loudly), re-calls decideOrdering for the dependent case, and recomputes crossesSequences per fixture by comparing the two decisions rather than reading the flag. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 258 entry points, 0 violations. -->

  • S11.12.d Exhaust retry by transient repeats, permanent reclassification, max attempts/age/cost, poison payload, dead-letter/replay, and circuit open. Evidence: retry fixtures. <!-- S11.12.d — TWO RULES, AND THEY ARE THE SAME MISSING INPUT SEEN TWICE: HISTORY.

THE FIRST: A CLASSIFICATION THAT DOES NOT COUNT CANNOT RECLASSIFY.

classifyJobFailure(failure, delivery) is a pure function of the failure KIND and the effect delivery, and it is a good one — its unknown-outcome rule is the best thing in the module: "with an idempotent effect an unknown outcome is just a retry, and with an at-most-once effect it is a question for a person, because retrying is how one payment becomes two and doing nothing is how it becomes none."

MEASURED by calling it: dependency-unavailable classifies after-backoff on attempt 1 and on attempt 50, and the function takes TWO arguments so there is no third that could carry a count. THIS CELL'S FIRST TWO ITEMS — "transient repeats" and "permanent reclassification" — ARE THE SAME ANSWER FROM IT, and a failure that has now happened fifty times is still filed as the kind of thing that goes away on its own.

That is not a defect. A classifier over one failure cannot count, and one that guessed from a single failure would be guessing. What is missing is a caller that keeps the count.

THE SECOND: defer PROTECTS THE QUEUE FROM ITS OWN JOBS, AND NOTHING PROTECTS A DEPENDENCY FROM THE QUEUE.

decideJobRetry's third answer is excellent: "A queue at capacity has not refused the attempt — it has not offered one — so a deferred job must not burn an attempt or reserve budget for one. Folding capacity into retry is how a busy hour quietly consumes every job's attempt allowance and leaves a queue full of jobs that are out of attempts and were never actually run."

MEASURED: JobRetryContext carries queueRunning and queueConcurrency — the QUEUE'S OWN capacity — and nothing about the health of what the job calls. Five hundred jobs failing on one broken dependency each back off on their own schedule and each return, and that sentence applies to the dependency with NONE OF THE WORDS CHANGED.

SHIPPED. libs/oshun/workbench-kit/src/retry-exhaustion-fixtures.ts — FailureHistory; reclassify, which calls the classifier FIRST and keeps its answer as classified because the point is the DIFFERENCE, hardens only after-backoff and only in one direction, and files a repeated failure with the SAME payload as after-change rather than after-remediation (nobody fixes the dependency; somebody changes the input); and decideAgainstDependency, reusing capability-state.ts's three circuit positions rather than inventing a parallel set, where a defer NEVER burns an attempt and the half-open PROBE does — because unlike a capacity defer it is a real request that really ran, and a probe that cost nothing would let a job probe for ever. 11 fixtures, 8/8 hazards, 3/3 reclassifications, 5 fault kinds; 3 changed by history, 2 deferred for the dependency, 4 past a ceiling. Attempts, age and cost are kept as three separate ceilings because hitting each is a different sentence to write in a log.

The estate's best rule is kept intact and tested: an unknown outcome with an at-most-once effect is after-remediation AND is NOT deferred by an open circuit — because the question is not whether the dependency is well, it is whether the effect happened.

TWO MORE FAULT KINDS THAT COULD NEVER FIRE became invariants swept over 192 combinations with positive controls (a decider that never deferred and a reclassifier that never changed anything would satisfy both trivially). Eighth and ninth instances in this section.

A TRAP: the scan's check that the estate's note exists found NOTHING on its first run, because the note is a WRAPPED block comment and the sentence is split by newlines and leading asterisks. It flattens the comment now — a zero from a query nobody wrap-proofed is not a measurement.

VERIFIED. 25 tests in src/retry-exhaustion-fixtures.spec.ts; kit suite 255 files / 15,719 tests green. tools/retry-exhaustion-fixtures-scan.mjs re-calls classifyJobFailure (with a positive control that it DOES discriminate between kinds), re-reads JobRetryContext's field list from jobs.ts, and re-sweeps both invariants. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 259 entry points, 0 violations. -->

  • S11.12.e Race cancellation before lease, during stage/checkpoint/ external effect, after success/before commit, with retry/parent-child, and with compensation. Evidence: cancellation fixtures. <!-- S11.12.e — THE RULE: A CANCEL AFTER THE EFFECT AND BEFORE THE COMMIT DOES NOT STOP ANYTHING — IT LOSES THE RECORD OF SOMETHING THAT ALREADY HAPPENED.

CREDIT FIRST. requestCancel refuses a cancel with no reason ("indistinguishable from a bug in the caller, and the job it killed cannot be told apart from one that failed"), and keeps cancelling as a state with a real duration: "collapsing the two — flipping a running job straight to cancelled — tells the operator the effects have stopped at the exact moment they have not, and they act on that." regionFaults carries the best fault in the module: a noncancellable region longer than the lease "turns a safety mechanism into the failure it was protecting against."

MEASURED: JobRecord has NINETEEN fields and exactly ONE is about how far the worker has got — a job-level state — and requestCancel takes only the job and the request. So every position this cell names lives INSIDE one state, and the four that matter get ONE answer from the state machine and cost FOUR different things:

during a stage             — nothing has escaped
during a checkpoint        — stopping leaves it half-written
inside an external effect  — the request has left and cannot be recalled
after success, before commit — THE EFFECT ALREADY HAPPENED

That last one is worth the rule. Cancelling there stops nothing — the work is done, the money moved, the file was written — and what the cancel destroys is the RECORD. The job then looks cancelled to everybody while its effect stands. The reverse of every other cancellation hazard, and why stop cannot be the answer everywhere.

SHIPPED. libs/oshun/workbench-kit/src/cancellation-fixtures.ts — WORKER_POSITIONS and POSITION_RULES, mapping each to one of the estate's OWN three cancellation-point actions rather than inventing new ones. A checkpoint write is CONTINUED rather than stopped, because a half-written checkpoint is read on resume and believed; the two positions where stopping is wrong get escalate-and-fence, which is what stops a second worker writing a conflicting record while somebody decides. cancelAt asks requestCancel FIRST — verified at ALL SIX positions: a reasonless cancel is refused wherever the worker is, and a terminal job is terminal whatever the worker thinks it is doing. 10 fixtures, 8/8 races, 6/6 positions, 3/3 actions, 6 fault kinds; 4 settled by state, 6 by position, 3 compensations owed.

A parent source and an actor source get the same answer at the same position, and a deadline cancel owes the same compensation as a person's — because what is owed depends on the POSITION and not on who asked.

TWO COSTS CANNOT ARRIVE THROUGH A POSITIONED DECISION, because their positions only occur with a job state the estate settles alone (queued and terminal). The module records WHICH positions and WHY rather than carrying an exclusion list with no reason attached — the next reader could not otherwise tell it from a case somebody forgot.

AND A TENTH UNREACHABLE FAULT KIND was removed here, which is enough instances to state the pattern as a rule: A FAULT KIND MUST RANGE OVER THE INPUT. A statement about a module's own constants is an INVARIANT, and it belongs in a sweep where it can fail.

VERIFIED. 23 tests in src/cancellation-fixtures.spec.ts; kit suite 256 files / 15,743 tests green. tools/cancellation-fixtures-scan.mjs re-reads JobRecord's FIELD LIST on every run (so adding a stage or phase fails loudly) and re-calls requestCancel at all six positions for both the reasonless and the terminal case. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 260 entry points, 0 violations. -->

  • S11.12.f Assert idempotent authoritative and external side effects, monotonic state/progress, correct result/audit/cost, preserved checkpoint, and no false success. Evidence: invariant suite. <!-- S11.12.f — THE RULE: AN INVARIANT SUITE OVER FIXTURES CHECKS THE DECISIONS, NOT THE SYSTEM — AND SAYING WHICH IS WHICH IS THE WHOLE VALUE OF IT.

The cell asks for nine assertions. A tenth is added — decision-agreement — because it is the structural property the whole of S11.12 rests on: every cell added a second decision beside an estate one and every one was built to ask the estate FIRST, and nothing until now checked that the ordering actually holds.

S11.12.a–.e built five corpora of FAULTS, and a fault fixture is a DESCRIPTION of a situation rather than a RUN of a system. So FOUR of the ten are evaluable here and SIX are not, and the difference is not a matter of effort: a fixture saying "the lease expired after the effect and before the checkpoint" carries no effect that could be applied twice, no audit row, and no money. Asserting idempotence over it would be asserting something about a system nobody ran.

REPORTING THAT IS THE POINT, and it is why this reuses S11.9.e's four verdicts — held, violated, unobservable, vacuous. An invariant that could not be evaluated must not look like one that was and passed, and A SUITE OF TEN GREENS OVER FOUR EVALUABLE PROPERTIES IS THE SHAPE OF A CHECK NOBODY HAS EVER SEEN FAIL.

OBSERVABLE                      examined
  monotonic-state                  5    every stale token refused, every
                                        terminal job answered as terminal
  preserved-checkpoint             9    every skipped effect silent, every
                                        repeat loud
  no-false-success                24    across three corpora
  decision-agreement              29    every added decision asked first
                                 ---
                                  67

UNOBSERVABLE — each recorded with what it would need
  idempotent-authoritative-effect  an effect that was actually applied
  idempotent-external-effect       a flight that actually left
  monotonic-progress               a progress reading no corpus carries
  correct-result                   work that finished
  correct-audit                    a row somebody wrote
  correct-cost                     a measured cost — and S11.11.f showed
                                   that is a floor whenever a counter is
                                   missing

SHIPPED. libs/oshun/workbench-kit/src/queue-invariants.ts — QUEUE_INVARIANTS, INVARIANT_NEEDS stating per invariant what a fixture would have to carry AND whether any corpus carries it, runSuite() over the five corpora, 4 fault kinds, census and headline.

AN ELEVENTH UNREACHABLE FAULT KIND was removed here under the rule S11.12.e stated after the tenth: a fault kind must range over the input, and a statement about a module's own constants is an invariant.

VERIFIED. 20 tests in src/queue-invariants.spec.ts; kit suite 257 files / 15,764 tests green. tools/queue-invariants-scan.mjs re-runs the suite and RECOMPUTES every examined count FROM THE CORPORA rather than reading it off the results — so a suite that quietly stopped looking at a corpus would fail — and checks that no unobservable invariant reports examined cases and no invariant returns a verdict without looking at anything. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 261 entry points, 0 violations. -->

  • S11.12.g Reconcile unknown outcomes and prove authorized replay does not duplicate effect. Evidence: recovery journey. <!-- S11.12.g — THE RULE: A REPLAY IS AUTHORIZED BY A CAPABILITY AND MADE SAFE BY A RECONCILIATION, AND ONLY THE FIRST IS CHECKED.

Both halves already exist in the estate, in different files, and nothing joins them.

THE RECONCILIATION HALF IS EXCELLENT. mayRetryAfterUnknown's table is the best writing in the library on the subject, and its fourth row is the whole point: "reconcile, and INDETERMINATE -> refuse. This is the cell people fill in with retry because the alternative needs a person, and it is the cell that produces the duplicate charge." And ReconciliationProbe is keyed by the IDEMPOTENCY KEY rather than any local id, "because after a connection loss the local id may never have been written — which is the state being reconciled".

THE AUTHORIZATION HALF IS GOOD TOO — every dead-letter action gated on a capability, refusing not-authorized, already-resolved, inputs-purged, inputs-moved and nothing-has-changed.

AND THE REPLAY ITSELF IS CAREFUL, WHICH NARROWS THIS FINDING AND IS WORTH GETTING RIGHT. planDeadLetterReplay PRESERVES THE JOB IDENTITY AND THE IDEMPOTENCY KEY: "minting a new job id forces one of two wrong moves: reuse the key, and the enqueue is answered with the dead job's outcome; drop it, and two runs of the same work exist with nothing relating them." So a replay whose first attempt COMPLETED is caught — same key, existing record, already-completed.

WHAT REMAINS, MEASURED: · planDeadLetterReplay takes no probe and consults no reconciliation · DeadLetter carries 18 fields including failureKind — which can be unknown-outcome — and NONE that says whether the effect landed · mayRetryAfterUnknown is called ZERO times from jobs.ts, with a positive control confirming the function is really there to be called

THE GAP IS PRECISELY THE UNKNOWN OUTCOME, and it is the one place the key does not help: an attempt that died between applying its effect and writing its record leaves no record for the key to find — S11.12.b's finding arriving here, and exactly the state the probe is keyed by the idempotency key to resolve. The replay proceeds, the claim finds nothing, and the effect is applied a second time.

SHIPPED. libs/oshun/workbench-kit/src/recovery-journey.ts — RECOVERY_STEPS asking authorization, then record state, then reconciliation. THE ORDER IS THE ARGUMENT: authorization first, because telling an operator who may not replay that the effect already landed is itself a disclosure; record state second, because a resolved dead letter has nothing to replay; reconciliation last, because it is the only one that requires asking somebody else. 10 fixtures, 3/3 steps, 4/4 outcomes, 3/3 reconciliation results, 5 fault kinds; 3 refused by the estate alone, 3 allowed by it and stopped here, 0 replays of an effect that really committed. A failure whose outcome is KNOWN skips the reconciliation entirely — the positive control, since a module that refused every replay without a probe would be useless.

THREE TRAPS IN ONE SCAN, ALL THE SAME KIND. A guessed function name matched nothing and the "mentions no probe" check PASSED AGAINST AN EMPTY STRING, for the same reason a clean file does. A non-greedy match to the first newline-brace stopped at the closing brace of the INLINE PARAMETER TYPE and read 239 characters of signature, concluding the body never mentions the idempotency key. And a slice to the next export ran INTO the following section's banner comment, whose word "reconciling" reversed the answer again. A BOUNDARY THAT IS TOO LATE IS AS WRONG AS ONE THAT IS TOO EARLY; IT JUST FAILS IN THE OTHER DIRECTION.

VERIFIED. 24 tests in src/recovery-journey.spec.ts; kit suite 258 files / 15,789 tests green. tools/recovery-journey-scan.mjs re-slices planDeadLetterReplay's body between two anchors on every run, recounts the calls to mayRetryAfterUnknown, and credits the idempotency-key preservation explicitly — so if somebody adds the probe, which would be good news, this fails. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 262 entry points, 0 violations. -->

  • S11.13 Add golden bundle build/import/verify runners. Evidence: each domain profile registers a golden. <!-- a-g all closed. The evidence clause is met on the registry that exists: WORKED_BUNDLE_PROFILES has two members, bundle.story-package and bundle.dataset-release, and both register a golden -- profilesWithoutARecordedGolden() returns empty and the scan re-checks it every run. There is NO broader "domain profile" registry in the kit: the 58 entries in domains.json are Nx project domains and none of them carries a bundle profile, so "each domain profile" resolves to those two. The mechanism that keeps this true as the registry grows is the estate's own goldenCoverageFaults, which walks the SHIPPED registry rather than the coverage, so a profile added tomorrow with no golden is reported rather than invisible -- and S11.13.g runs it against WORKED_BUNDLE_PROFILES on every scan.

    WHAT THE SEVEN CELLS FOUND, IN ONE LINE EACH. .a two of my assumptions
    were wrong and checking them made the finding precise -- a real runner
    exists, and it calls 2 of the 7 checkers. .b ProfileFixture's eight fields
    carry no hash and no licence, so the registry records that a fixture
    EXISTS and never that it is the SAME one. .c a reproducibility check and a
    golden are different claims, and the estate has six controls of the first
    and none of the second. .d a parity comparison needs two independent
    reads, and the checker that says so had never been handed one; the format
    can carry four of its eight dimensions. .e the matrix had seven aspects,
    the cell names eight, and the missing one already had a checker. .f a
    rebuild from imported state must re-derive what it rebuilds from, and
    CONTROL 6 re-reads the bodies and keeps the layout. .g a golden that can
    be updated by editing it records whatever the last run did.
    
    FOUR OF THE FIVE UNFED CHECKERS NOW HAVE PRODUCERS. S11.13.a measured that
    golden-round-trip.ts has seven checkers and the runner called two.
    buildGoldenFaults (.c), importParityFaults (.d), verificationMatrixFaults
    (.e) and rebuildFaults (.f) each got their first producer in the repo;
    fixtureRegistryFaults got its first real registry in .b. All seven are now
    fed.
    
    Kit: 265 files / 15,961 tests green, 269 entry points, 0 violations, tsc
    clean and unfiltered; all seven scans re-run after their registrations. -->
    
    • S11.13.a Define runner plugin for domain profile/version, fixture set, build inputs, expected layout/manifest/archive hash, import/migration, verification, and semantic/byte identity promise. Evidence: runner API. <!-- S11.13.a — THE RULE: THE RUNNER PRODUCES BYTES AND THE CHECKERS CONSUME DESCRIPTIONS, AND FIVE OF THE SEVEN DESCRIPTIONS HAVE NO PRODUCER.

TWO ASSUMPTIONS I MADE ABOUT THIS ESTATE WERE WRONG, AND BOTH ARE RECORDED BECAUSE CHECKING THEM IS WHAT MADE THE FINDING PRECISE.

THE FIRST: that nothing performs a golden run. tools/golden-round-trip.mjs is 472 lines and performs a REAL one — it writes artefact bodies, serialises a manifest, archives with GNU tar, hashes real bytes, and repeats the whole build under shuffled input order, touched mtimes, a Turkish locale, a Kiritimati time zone, a 0077 umask and interleaved writes, comparing tar HEADERS as well as digests and asserting that localeCompare disagrees with byte order first so the locale axis is known to be live. It extracts each archive and rebuilds FROM THE EXTRACTED BYTES, byte-identical. And it reports honestly that "platform" cannot be varied on one machine, NAMING it as never perturbed rather than dropping it from the axis list. That is excellent work and S11.13.c and .f are substantially done by it.

THE SECOND: that golden-round-trip.ts is only registers. It has SEVEN checkers and they map one-to-one onto this section's seven cells.

WHAT IS ACTUALLY TRUE IS NARROWER AND CHECKABLE. The runner calls TWO of the seven — archiveReproducibilityFaults and goldenCoverageFaults. The other five take descriptions of a run:

fixtureRegistryFaults      (profiles, ProfileFixture[])
buildGoldenFaults          (BuildGolden)          ZERO producers
importParityFaults         (fixtureId, ParityComparison[])
verificationMatrixFaults   (VerificationRow[])
rebuildFaults              (RebuildComparison)    ZERO producers

The checkers are ready and nothing hands them what a real run found.

SHIPPED. libs/oshun/workbench-kit/src/golden-runner.ts — a GoldenRunner contract with ONE STAGE PER UNFED CHECKER, and each stage's return type named as the type that checker takes. THE API IS READ OFF THE CHECKERS RATHER THAN DESIGNED: if a checker's input changes, the stage that feeds it stops compiling, which is the only kind of agreement worth having between two halves of a pipeline. Every stage is required, because a checker handed nothing reports nothing — which is indistinguishable from one handed a clean run.

THE IDENTITY PROMISE IS PART OF THE CONTRACT, because this cell names it. A runner promising BYTE identity and one promising SEMANTIC identity are making different claims and failing differently, and a rebuild comparison read without knowing which was promised cannot be judged. A semantic promise with no documented difference is refused — it is a byte promise somebody was not confident enough to make — and a byte promise that documents one contradicts itself.

AND THERE IS NOTHING TO PLUG INTO. EXTENSION_POINTS offers stage-model, gate, block, preview, inspector, importer and publisher, and no golden-runner. This cell's "runner plugin" is a shape nothing can yet be registered at. That fault is reported on every run and IS the finding rather than a defect in this module, with a positive control confirming the predicate would see a point if one existed.

VERIFIED. 21 tests in src/golden-runner.spec.ts; kit suite 259 files / 15,811 tests green. tools/golden-runner-scan.mjs RECOUNTS which checkers the shipped runner calls from its own source rather than trusting the register — so wiring up a third fails loudly — enumerates golden-round-trip.ts's exported checkers to catch one added or removed, and checks every stage's produced type against that module's exports. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 263 entry points, 0 violations. -->

  • S11.13.b Register minimal/full/boundary/localized/provenance-rich/ compatibility/invalid goldens per profile with licensed source and immutable hashes. Evidence: golden manifest. <!-- S11.13.b — THE RULE: A GOLDEN IS A PROMISE THAT SOMETHING HAS NOT CHANGED, AND ProfileFixture CARRIES NO HASH.

CREDIT FIRST. GOLDEN_FIXTURE_CLASSES is exactly the seven this cell names, and fixtureRegistryFaults makes seven checks worth having — the best two being a profile with NO refused fixture (a registry where nothing is declined has never shown the validator can decline) and a fixture with no reason, because "a fixture that starts passing names the rule".

MEASURED, on ProfileFixture's EIGHT fields:

hash fields    0
licence fields 0

and golden-round-trip.ts does not use Sha256Hex anywhere in the module. Both are things this cell names — "licensed source and immutable hashes" — and both have a vocabulary elsewhere in the estate: bill-of-materials.ts exports ASSERTION_STATES = ['asserted','none','unknown','not-applicable'], which are exactly the right four for a licence, because they separate "we checked and there is none" from "nobody looked".

WHY THE HASH MATTERS MORE THAN IT LOOKS. A golden's entire job is to fail when something changed. Without one the registry records that a fixture EXISTS and never that it is the SAME fixture — so a golden edited in place goes on passing, and the run that should have caught the change is the run that was changed to accept it.

WHY THE LICENCE IS NOT A FORMALITY. A golden's content is MATERIAL, redistributed in every clone and every container image, and a fixture whose source nobody recorded cannot be shown to be redistributable at all.

SHIPPED. libs/oshun/workbench-kit/src/golden-manifest.ts — GoldenEntry with a contentHash, a recomputedHash and a LicensedSource; permitsRedistribution, which admits an asserted licence only when it names an identifier and admits not-applicable because content this repository authored has no third-party terms; 8 fault kinds; and a worked manifest of 14 entries over the 2 exported profiles and all 7 classes.

THAT MANIFEST IS HANDED TO THE ESTATE'S OWN fixtureRegistryFaults AND COMES BACK CLEAN — the first time anything has handed that checker a registry at all. The scan carries a positive control confirming the checker really does object when no fixture is expected to be refused, because an empty finding list from a checker nobody has seen refuse anything is not a measurement.

VERIFIED. 25 tests in src/golden-manifest.spec.ts; kit suite 260 files / 15,837 tests green. tools/golden-manifest-scan.mjs re-reads ProfileFixture's OWN FIELD LIST from golden-round-trip.ts on every run, so adding a hash or a licence field — either would be good news — fails loudly. 8 negative controls, exits 2 if any goes undetected. Five registrations done; generator run: 264 entry points, 0 violations. -->

  • S11.13.c Build under controlled/perturbed order/clock/locale/platform and compare layout/manifest/components/provenance/archive bytes. Evidence: build runner. <!-- A REPRODUCIBILITY CHECK AND A GOLDEN ARE DIFFERENT CLAIMS: ONE SAYS THE BUILDER AGREES WITH ITSELF, THE OTHER SAYS IT AGREES WITH A DECISION SOMEBODY WROTE DOWN, AND THE ESTATE HAS SIX CONTROLS OF THE FIRST AND NONE OF THE SECOND.

    CREDIT FIRST, AND IT IS MOST OF THIS CELL. tools/golden-round-trip.mjs
    already performs a real perturbed build: resolveLayout from the profile,
    artefact bodies written, a manifest serialised, the tree archived with
    GNU tar, bytes hashed -- then the whole build repeated under shuffled
    input order, touched mtimes, LC_ALL=tr_TR.UTF-8, TZ=Pacific/Kiritimati,
    a 0077 umask and interleaved writes, comparing the tar HEADERS as well
    as the digest. It asserts FIRST that localeCompare disagrees with byte
    order on a probe pair, so the locale axis is known live rather than
    assumed. It rebuilds from the EXTRACTED bytes, byte-identical. And it
    reports honestly that `platform` cannot be varied on one machine, naming
    it as never perturbed rather than dropping it from the axis list. Four
    of this cell's axes are driven there; the fifth is disclosed there. None
    of that is redone here.
    
    MEASURED, AND THE SCAN RETAKES THE COUNT EVERY RUN: checked-in expected
    values in that tool = ZERO. All six comparisons are sha256(perturbed
    .bytes) against sha256(baseline.bytes), both computed in one process,
    moments apart, by the same code. The estate's OWN checker names that
    shape: buildGoldenFaults calls an expectedSource of
    `regenerated-at-test-time` the case where "the expected value is
    produced by the same code that produces the actual value, so the
    comparison agrees with itself for every input and can never fail". So
    the tool detects nondeterminism -- which is what it was built for and is
    genuinely good at -- and CANNOT detect a change to the builder: rename
    every path, reorder the manifest's keys, drop provenance entirely, and
    both sides move together while all six controls pass.
    
    SHIPPED src/build-golden.ts. The four aspects BUILD_GOLDEN_ASPECTS names
    -- manifest, layout, component-inventory, provenance -- each projected
    to a canonical string, each compared against a checked-in literal, and
    the result handed to buildGoldenFaults. That makes this the FIRST
    PRODUCER of BuildGolden since that checker was written (S11.13.a
    measured it at zero). 29 spec tests; scan PASSES with 6 negative
    controls, all caught.
    
    THE RECORDED VALUE IS A READABLE STRING, NOT A DIGEST. A golden recorded
    as a digest tells whoever reviews the change that something moved and
    never what moved, which makes the review that is supposed to approve the
    new answer a rubber stamp on an opaque one. The eight literals run 145
    to 915 characters; a reviewer sees the path that changed.
    
    producedAt is EXCLUDED from the provenance aspect and the exclusion is
    NAMED in PER_RUN_PROVENANCE_FIELDS. Provenance is the one aspect not
    reproducible by construction -- it carries the instant the build ran, so
    a golden pinning it whole could never match twice, and the only way to
    make it pass is to freeze the clock, which is what the surrounding tool
    does and what a real build cannot. Pinning the stable part and naming
    the per-run part is a documented approximation; pinning nothing is what
    the estate does now. Related: `producedBy` is set on the REQUEST and is
    absent from the manifest the tool writes, so provenance never reaches
    the built bytes at all -- re-measured every scan run.
    
    ASPECT_INDEPENDENCE is five mutations that each declare which aspects
    must move AND which must not, every one verified through the estate's
    checker. If every mutation moved all four, four aspects would be
    decorative. A moved file changes layout and manifest and leaves the
    inventory alone; a changed digest leaves layout alone; a different
    producer moves provenance and nothing else -- the aspect the archive
    bytes cannot carry, because it is not in them.
    
    TWO DEFECTS THIS CELL FOUND IN ITSELF, BOTH ITS OWN SUBJECT, both kept
    as negative controls. (1) The inventory first sorted on a key of slotId
    and digest; bundle.dataset-release fills its attestation slot TWICE with
    entries agreeing on both, so the tie was broken by sort stability -- by
    the order the inputs arrived in -- INSIDE the projection meant to detect
    exactly that. It now sorts on the whole canonical element, which can
    only tie for equal elements. (2) The locale positive control first
    probed 'i' against 'I', where Turkish collation and byte order AGREE, so
    it failed; replaced with the pair the estate's own tool proved.
    
    ALSO CORRECTED HERE, and it is a process defect of mine worth recording:
    adding this cell's tsconfig path without a leading "./" raised TS5090,
    and fixing that unmasked 19 pre-existing type errors across 8 modules I
    shipped earlier this session -- invisible until now because I had been
    grepping tsc output for the current cell's filename each time. Two were
    substantive: queue-invariants filtered terminal jobs on the literal
    'failed', which is NOT a JobState, so that clause was always true and
    the set quietly lost 'dead-lettered' (now uses the estate's own
    TERMINAL_JOB_STATES); and toolCallFaults pushed five ToolCallFault kinds
    into MalformedFinding[], whose kind union declares none of them (now
    ToolCallFinding). A third: cancellation-fixtures' JobBudget carried
    maxAttempts/maxSeconds/maxCostMinorUnits, three names in no type in this
    repo, hidden by a blanket `as JobRecord`; JobBudget is unit/ceiling/
    spent. Kit: 261 files / 15,867 tests green, 265 entry points, 0
    violations, tsc clean.
    
    AND A LATENT SCAN DEFECT THE SAME PASS FOUND: malformed-fixtures' and
    lease-fixtures' scans text-searched for symbol names, so the kit's
    project.json "//" docs and the export-map generator's reason strings --
    registers of PROSE ABOUT the code, written during the five registrations
    -- counted as callers. Both scans had passed when run BEFORE those
    registrations existed, which is the order these cells are built in. Both
    now require an application or an import. RUN A CELL'S SCAN AGAIN AFTER
    ITS REGISTRATIONS. -->
    
  • S11.13.d Import/preflight/migrate/validate/persist/cleanup and compare semantic entities/artifacts/refs/lineage/loss report. Evidence: import runner. <!-- A PARITY COMPARISON NEEDS TWO INDEPENDENT READS, AND THE CHECKER THAT SAYS SO HAS NEVER BEEN HANDED ONE.

    CREDIT FIRST, AND IT IS FOUR OF THE SIX STAGES.
    tools/importer-roundtrip.mjs is 438 lines driving a REAL round trip: a
    v2 bundle built with real tar, validated against a real JSON schema with
    ajv, migrated to v4 and re-validated -- with a control asserting the v2
    manifest FAILS the v4 schema FIRST, so the migration is known to have
    moved something rather than assumed to have. It measures a real gzip
    ratio, verifies a real P-256 signature and shows it does NOT survive a
    migration, catches a naive v5 reader silently downgrading a v6 manifest,
    and its CONTROL 4 IS THE LOSS REPORT: a real migration drops a field and
    the report is the diff. It calls preflightFaults and
    migrationIntegrityFaults for real. Import, preflight, migrate, validate
    and the loss report are covered there and none of it is redone here.
    
    WHAT IS NOT DRIVEN, MEASURED BY COUNTING: `persist` and `cleanup` appear
    ZERO times in that file, and so does every one of the eight
    IMPORT_PARITY_DIMENSIONS. importParityFaults is called by nothing and
    ParityComparison has NO PRODUCER anywhere in the repo -- so
    `bothSidesRead`, the field recording whether a comparison was a
    photocopy, has never carried a value. The estate's own words for what
    that field guards: "the photocopy in the one place it is hardest to see:
    comparing the imported object to the object the importer just produced
    from the same parse tests that a function is deterministic, not that the
    bundle survived."
    
    SHIPPED src/import-parity.ts + a scan that does the real thing. A bundle
    is built, bodies written, manifest serialised, tree archived with GNU
    tar, and then EXTRACTED TO A DIFFERENT DIRECTORY, where the manifest is
    re-parsed from its file and every digest recomputed from the extracted
    bytes. `bothSidesRead` is COMPUTED from the two origins, so a comparison
    that cheated would report itself; 8 comparisons across 2 profiles, all 8
    with both sides read. compareSides refuses two different dimensions
    outright. 24 spec tests plus 4 more below; scan PASSES with 8 negative
    controls, all caught.
    
    AND THE MEASUREMENT THAT DECIDES THE CELL'S SHAPE: THE CHECKER ASKS
    ABOUT EIGHT DIMENSIONS AND THE FORMAT CAN CARRY FOUR. Across
    BundleProfile, LayoutEntry and BundleRequest -- 22 fields, re-derived by
    the scan from the interfaces themselves, with externalRef as a positive
    control so the zeros are facts and not a broken search -- there is
    nothing holding a rights position, nothing for consent and nothing for a
    gate. Those are NOT reported as mismatches: a dimension with no second
    side to read would be a defect filed against a bundle that is exactly
    what its schema says it should be. They are left out, and the estate's
    own checker reports rights-or-consent-never-compared for two of them --
    separated in its words because "a bundle that arrives with its rights
    position flattened is content being used on terms nobody agreed to, and
    it imports perfectly."
    
    CARRIAGE_STATES KEEPS A DISTINCTION THE CHECKER'S VOCABULARY CANNOT.
    provenance is `carried-but-not-written` -- producedBy and producedAt are
    on the request and neither reaches the manifest, as S11.13.c measured --
    while rights is `absent-from-the-format`. To importParityFaults both are
    simply absent from the comparison list and come back as
    dimension-never-compared; a format with no field and a field nothing
    writes are different repairs and belong in different queues.
    
    TWO DEFECTS THIS CELL FOUND IN ITSELF, both this session's recurring
    shape. (1) planWroteNothing was derived from `persisted`, which is the
    EXECUTION's writes, so every import that successfully wrote its entities
    reported that its PREVIEW had written them -- two counts of files on
    disk, related because both are counts of files on disk, answering
    different questions. The plan is now listed at its own moment and
    nothing else can stand in for it. (2) residue was every surviving file,
    so the scan reported the two entities an import had just persisted as
    residue -- filing an import's own success as a defect. A SURVIVING FILE
    IS RESIDUE ONLY RELATIVE TO WHAT THE RUN CLAIMS HAPPENED:
    bundleImportFaults reads residue only when the outcome is `failed` and
    when a rollback claims a partial import, so residueForOutcome takes the
    outcome. Both kept as tests.
    
    ALSO REPAIRED, and it is a real finding in the S11.9 fuzz-schedule cell:
    its register counted its own family, so it failed TWICE in one session
    reporting growth as drift -- every cell adds one -scan.mjs and one
    verify:operations script, which makes an equality stale by
    construction. kitScans and declaredScripts are now FLOORS, which still
    catch the direction that loses coverage (a scan deleted). In their place
    the check the count was standing in for: TEN OF THE 115 KIT SCANS HAVE
    NO verify:operations SCRIPT -- audit-integrity, catalog-bypass,
    cost-attribution, degradation-plan, exhaustion-alert, game-day,
    log-leak, production, restore-consistency, router-vocabulary. They run
    in the workflow, which globs by path, and can be invoked by name
    nowhere. That number does NOT grow with the kit, so it is checked for
    equality and a scan wired into neither trips it.
    
    Kit: 262 files / 15,896 tests green, 266 entry points, 0 violations,
    tsc clean and UNFILTERED this time. -->
    
  • S11.13.e Verify paths/schema/hash/signature/trust/dependency/profile/ target compatibility and expected invalid/refusal outcomes. Evidence: verify runner. <!-- THE MATRIX HAD SEVEN ASPECTS, THIS CELL NAMES EIGHT, AND THE MISSING ONE IS THE ONE THE ESTATE ALREADY HAD A CHECKER FOR.

    VERIFICATION_ASPECTS was schema, paths, hashes, signature, trust,
    profile-semantics and target-semantics. This cell asks for
    paths/schema/hash/signature/trust/dependency/profile/target, and
    golden-round-trip.ts did not contain the string "dependenc" anywhere.
    
    It was not missing for want of an implementation. bill-of-materials.ts
    has EXTERNAL_COMPONENT_KINDS -- url, package, model, provider-endpoint,
    resolver, tool, plugin, font, license-document, dataset -- and nine
    external-component faults with externalComponentFaults deciding them,
    including unpinned-remote-required-at-runtime and
    pinned-remote-with-no-digest, which is exactly what verifying a
    dependency means. A whole checker, and no aspect to put its answer in.
    `dependency` is added to the estate's list here and wired to it.
    
    SHIPPED src/verify-runner.ts + scan. SIXTEEN RUNS: eight aspects in BOTH
    directions, every verdict from a real estate function or real crypto --
    validatorFaults, resolveLayout, stampFaults, bundleProfileFaults,
    targetAccepts, externalComponentFaults, a real sha256 and a real P-256
    key pair -- never from a table of expected answers. 8 refusals, each
    tripping the fault its binding names. 21 spec tests; scan PASSES with 8
    negative controls, all caught. verificationMatrixFaults had NO producer
    before this: VerificationRow appears in golden-runner.ts as a stage's
    return type, which names it without making one.
    
    THE THIRD OBSERVED VALUE IS LOAD-BEARING. The estate keeps
    aspect-that-did-not-run apart from a mismatch and says why: "an
    unreachable trust store and a valid chain both leave the signature
    aspect with nothing to report, and only one of them is fine." A runner
    that could only say accepted or refused renders a verification that
    never executed identically to one that passed.
    
    BOTH DIRECTIONS, because matrix-with-no-expected-refusals exists for a
    matrix of green rows nobody has ever seen say no, and
    expected-refusal-that-was-accepted is, in the estate's words, "the
    finding this whole matrix exists for. Everything else says the pipeline
    still handles what it handled; only an invalid fixture that got through
    says a check stopped running -- and a check that stopped running
    produces no error anywhere else, ever."
    
    FOUR DEFECTS THIS CELL FOUND IN ITSELF, AND ONE CHECK FOUND THREE OF
    THEM. The module's header promised that a refusal must name the finding
    it was refused FOR -- a refusal for the wrong reason is an acceptance
    wearing the right answer -- while the scan asserted only that a refusal
    happened. Adding the check the prose already demanded immediately
    showed: (1) the dependency refusal fixture used reachability 'public',
    which is in no ReachabilityState, so it was ACCEPTED, and only a
    negative control noticed; (2) the schema ACCEPTED case was refused,
    because the digest a validator was compiled from lives on ValidatorBuild
    and I had left it undefined; (3) two expected fault names were guesses
    no estate list contains -- the real ones are
    schema-changed-since-the-validator-was-compiled and
    target-requiring-a-slot-the-profile-does-not-declare. THE LESSON: A
    CHECK THAT SOMETHING WAS REFUSED IS NOT A CHECK THAT IT WAS REFUSED FOR
    THE REASON YOU MEANT, AND THE PROSE HAD SAID SO ALREADY.
    
    (4) The matrix was judged one direction at a time, and the accepted half
    came back matrix-with-no-expected-refusals -- correctly, because a half
    containing only acceptances is exactly what that fault is for. Splitting
    the matrix defeats the check that it has ever been seen to say no; the
    checker takes aspects as a Set for coverage, so two rows per aspect is
    what it is built for. Judged as one 16-row matrix.
    
    Kit: 263 files / 15,918 tests green, 267 entry points, 0 violations, tsc
    clean and unfiltered. Every scan re-run AFTER the five registrations. -->
    
  • S11.13.f Rebuild from imported state and compare byte or documented semantic identity. Evidence: round-trip runner. <!-- A REBUILD FROM IMPORTED STATE MUST RE-DERIVE WHAT IT REBUILDS FROM, AND CONTROL 6 RE-READS THE BODIES AND KEEPS THE LAYOUT.

    CREDIT, AND IT IS A REAL REBUILD. tools/golden-round-trip.mjs CONTROL 6
    extracts each archive to a fresh directory, reads the manifest back off
    it, checks that manifest names the profile it should, re-reads every
    artifact body from the extracted tree, rebuilds, re-archives and asserts
    the two archives hash the same. Through real tar, and it passes.
    
    WHAT IT KEEPS, AND WHY IT MATTERS. The rebuild is assembled as
    `{ ...input, body: readFileSync(...) }` over the ORIGINAL inputs, and
    handed the ORIGINAL `entries` -- the resolved layout. So one field comes
    back off the disk and everything else is the object that was already
    there: the paths, the slot each artifact belongs to, its media type, its
    ordinal, its declared digest. Those are exactly the things an import has
    to reconstruct, and keeping them means the rebuild CANNOT FAIL on them.
    A manifest that omitted a path, recorded a slot wrongly or lost an
    ordinal would still rebuild byte-identically, because nothing ever asked
    the manifest for any of it. This is S11.13.d's one-side-derived-from-the-
    other arriving in the rebuild, and it is the difference between the
    "extracted state" the tool uses and the "imported state" this cell asks
    for: extracting is untarring, importing is reconstructing the entities.
    
    SHIPPED src/rebuild-identity.ts + scan. ReconstructedState is a type
    with NO field through which an original object can arrive, and
    rebuildFrom takes one and no second argument, so a rebuild written
    against it cannot quietly keep the layout it was supposed to re-derive.
    18 spec tests; scan PASSES with 9 negative controls, all caught.
    
    THE DEMONSTRATION IS THE WHOLE CELL IN TWO ARCHIVES. A bundle carrying
    one file that no manifest entry names, rebuilt twice: from the ORIGINAL
    layout it comes out byte-identical and reports nothing; from the
    MANIFEST it loses the file, and reconstructionFaults names it as
    file-in-the-archive-with-no-manifest-entry. Both halves asserted, so
    neither can quietly stop being true.
    
    AND THE PROMISE NOBODY MADE. rebuildFaults reports
    profile-that-promises-no-identity-at-all when the promise is null,
    because "a rebuild comparison against no promise passes whatever comes
    out". BundleProfile is profileId, profileVersion, manifestPath,
    manifestSchemaId, manifestSchemaVersion, slots, targets and limits --
    there is NO field for an identity promise, so every profile in the
    estate promises nothing and that fault fires for all of them.
    PROFILE_PROMISES is the declaration the profile cannot carry;
    promiseFor deliberately returns null for anything undeclared so the
    fault can still fire, because defaulting it would hide the one finding
    that says a rebuild test proves nothing. Both worked profiles promise
    byte identity, and the estate already demonstrates they can keep it.
    
    TWO DEFECTS OF MINE, FIXED HERE. (1) golden-runner.ts (S11.13.a)
    declared its own IDENTITY_PROMISES = ['byte', 'semantic'] beside the
    estate's ['byte-identical', 'semantically-identical'] -- the same two
    concepts under two spellings in one library, with RebuildComparison
    typed against one of them. It now re-exports the estate's. The estate
    warns about exactly this shape a few hundred lines from where it
    happened: "a second list of the things that vary between two machines
    would drift from the first, and the one that drifted would be the one
    nobody was reading." (2) RECONSTRUCTION_FAULTS shipped with
    manifest-that-names-another-profile emitted by NOTHING, and I wrote a
    comment in the spec excusing it -- the unreachable-fault-kind defect
    this section keeps finding in other people's code, THIRTEENTH INSTANCE,
    mine. reconstructionFaults now takes the profile the archive was opened
    as, and the kind is reachable and controlled.
    
    Kit: 264 files / 15,937 tests green, 268 entry points, 0 violations,
    tsc clean and unfiltered; every scan re-run after the registrations. -->
    
  • S11.13.g Fail registry/CI when a profile/version lacks current golden or a golden update lacks reviewed source/semantic rationale. Evidence: coverage gate. <!-- A GOLDEN THAT CAN BE UPDATED BY EDITING IT IS A GOLDEN THAT RECORDS WHATEVER THE LAST RUN DID.

    CREDIT, AND THE FIRST HALF IS ALREADY DONE. goldenCoverageFaults walks
    the SHIPPED profile registry and asks what coverage exists for each --
    the argument order is the whole thing, in the estate's words, because a
    report built the other way round "finds every golden covered by
    definition, and a profile added tomorrow with no fixtures at all is
    precisely what it cannot see". It reports a registered profile with no
    coverage, a missing fixture class, a missing round-trip stage, coverage
    against an older profile version, and coverage for a profile nobody
    ships. tools/golden-round-trip.mjs calls it with WORKED_BUNDLE_PROFILES
    -- the real registry, not a fixture -- and CI runs that tool. "Fail when
    a profile/version lacks a current golden" is covered.
    
    THE SECOND HALF HAD NO VOCABULARY AT ALL. The words rationale,
    reviewedBy and updateReason appear NOWHERE in golden-round-trip.ts,
    golden-manifest.ts or build-golden.ts. Nothing models an UPDATE. The
    estate names the automated version --
    golden-that-rewrites-itself-on-mismatch, which "turns a checked-in
    literal into a recording of the most recent run on the first red build"
    -- and the manual version is the same thing at human speed. S11.13.c
    checked eight golden literals in precisely so a change would be visible
    in review, and VISIBLE IN REVIEW IS NOT GATED: today a red golden test
    is repaired by pasting the new value over the old one, and every check
    in this section goes green again against the new answer, with nothing
    recording that it moved or why that was allowed.
    
    SHIPPED src/golden-update-gate.ts + scan. A gate on SOURCE, RATIONALE
    and REVIEW -- all three, because any two let the third through: a
    reviewed regeneration is a reviewer approving a value the code chose, a
    rationale with no review is a note nobody read, and a review with no
    rationale is agreement to something unstated. The source vocabulary is
    IMPORTED from BuildGolden['expectedSource'] rather than restated,
    because how a value got here is the same question at record time and at
    update time, and S11.13.f had to remove one duplicate vocabulary from
    this kit already. The rationale is separate from the source because "the
    build now emits this" answers where the value came from and not why it
    is correct, and it must NAME the change it points at. reviewedBy equal
    to updatedBy is refused for the reason a signature over your own claim
    is not a second opinion. An update that moved nothing is reported rather
    than waved through, and stops there, because the remaining questions are
    about a change. 23 spec tests; scan PASSES with 10 negative controls.
    
    AND THE GATE CAN SEE AN UPDATE, which is what keeps it from being a
    shape. A gate taking updates as an argument is a gate somebody has to
    remember to feed, so the scan reads the CURRENT recorded literals out of
    build-golden.ts and the PREVIOUS ones out of `git show HEAD:`, and any
    literal that differs is a real update it has to judge. Nothing is passed
    in by hand. TWO CONTROLS GUARD THAT ZERO: the extractor's literal count
    must equal what the module exports (8 = 8), and a synthesised edit must
    be seen as exactly one update -- because a zero from a parser nobody
    proved can parse is not a measurement.
    
    Every one of the seven update fault kinds is reached from a distinct
    input, in both the spec and the scan, so none of them is the
    unreachable-fault-kind defect this section has now corrected fourteen
    times. -->
    
  • S11.14 Add provenance tamper and trust-chain fixtures. Evidence: valid, invalid, unknown, expired, and stripped states. <!-- a-g all closed. The evidence clause names five states and the corpora reach all five, verified by handing every fixture to the estate's own credentialState rather than asserting: valid-and-trusted (.a), the invalid family (.c, seven separable causes), unknown-or-unsupported (.d), expired (.e), and stripped plus partially-recoverable and missing (.a, .f).

    THE SECTION'S ONE FINDING, SEVEN TIMES. Everything a reader would act on
    differently arrives as one scalar, and each cell measured a different face
    of it: ten checks to six states (.b), seven of nine causes separable and
    only two by the state (.c), a six-month-old cache reading identically to a
    live lookup (.d), five lifecycle positions arriving as two states (.e),
    one perceptual match indistinguishable from several (.f). .g reads the
    seven dimensions side by side and names what the headline hides.
    
    AND THE SECOND SHAPE, FOUR TIMES: THE CONCEPT IS MODELLED A MODULE AWAY.
    .c's schema (manifest-schema.ts has a registered schema with a pinned ajv
    validator, for the OTHER manifest), .d's freshness (trust-results.ts
    EXPIRY_STATES), .e's consent and retention (approval-workflows.ts
    `withdrawn`, audit-snapshots.ts `retainUntilMs`), and .g's policy, which
    has no home anywhere. Recorded as modelled-elsewhere rather than absent
    throughout, because one is a wire and the other is design work.
    
    THE RESTRAINT, AND IT HELD FOUR TIMES. This section did NOT widen
    VALIDATION_CHECKS, CHECK_OUTCOMES or ExternalWitness. S11.13.e widened
    VERIFICATION_ASPECTS for the same kind of gap and that was right -- that
    list had zero consumers outside specs. These three are read by
    ProvenanceInspector.tsx, provenance-inspector.ts, suggestions.ts and a
    browser tool, so widening them changes a shipped product surface. The
    missing axis was declared beside each corpus instead, and every fixture
    still runs through the estate with the vocabulary the estate has.
    
    CREDIT, THROUGHOUT. tools/c2pa-signing.mjs is 493 lines of real OpenSSL
    certificate authorities and real P-256 signatures whose CONTROL 1 proves
    the naive "does the signature verify" returns TRUE for four of the five
    situations. content-credentials.ts is 1,293 lines that get the hard
    distinctions right -- untrusted is not invalid, an unsupported assertion
    must not invalidate, an unverified timestamp is no timestamp. And
    unavailable-revocation-treated-as-a-result reads the DETAIL STRING rather
    than the outcome, catching an outage whose message reads like an answer,
    which is the best single check in the library.
    
    Kit: 272 files / 16,120 tests green, 276 entry points, 0 violations, tsc
    clean and unfiltered; all seven scans re-run after their registrations. -->
    
    • S11.14.a Add valid embedded/sidecar/ledger fixtures with complete and partial ingredients, supported assertions, trusted chain, timestamp, and hard/soft bindings. Evidence: valid corpus. <!-- A LEDGER-RECOVERABLE CREDENTIAL IS NEVER valid-and-trusted.

      CREDIT FIRST, AND IT IS REAL CRYPTOGRAPHY. tools/c2pa-signing.mjs is 493
      lines that generate two independent root certificate authorities and
      three leaf certificates with OpenSSL, build and verify the paths with
      OpenSSL, and sign and verify over the exact bytes with node:crypto. Its
      CONTROL 1 is the argument this whole section rests on: it asserts the
      naive check -- "does the signature verify" -- returns TRUE for FOUR of
      the five situations (valid, untrusted, expired, stripped) BEFORE
      asserting credentialState returns four different answers. Three of those
      four trues are wrong, and no test that only checked the strict function
      would show it. Not re-done here.
      
      THE CELL ASKS FOR EMBEDDED, SIDECAR AND LEDGER FIXTURES, AND THE ESTATE
      IS RIGHT THAT ONLY TWO OF THOSE ARE PLACES A CREDENTIAL IS EMBEDDED.
      EMBEDDING_MECHANISMS is embedded, sidecar, embedded-and-sidecar and
      none, with no ledger -- and that is not an omission. A credential is not
      embedded in a ledger; a ledger is somewhere it can be FOUND when it is
      embedded nowhere. The estate already says so in the type system:
      ExternalWitness carries ledgerHoldsCredentialForDigest, and
      credentialState reads it only on the branch where credentialPresent is
      false.
      
      THE CONSEQUENCE, and a corpus built on the cell's wording would get it
      wrong. Measured by calling the estate: a ledger-backed absence is
      `stripped`, a soft-binding match is `partially-recoverable`, and the
      same absence with no witness is `missing`. THREE DIFFERENT ANSWERS, and
      the scan asserts all three differ, so the witness is known to decide
      something rather than assumed to. "A valid ledger fixture" is a category
      the estate deliberately does not have.
      
      SHIPPED src/credential-corpus.ts + scan. RECOVERY_SOURCES is kept as a
      SEPARATE AXIS from the embedding, because keeping them as one list is
      exactly what would make a ledger look like a fourth kind of embedding.
      Eight fixtures covering every value of all four axes, reaching 4 of the
      9 credential states. EVERY FIXTURE'S EXPECTED STATE IS RE-JUDGED ON
      EVERY RUN by handing a real ValidationReport to the estate's own
      credentialState, so the corpus cannot drift from the thing it exists to
      exercise. 22 spec tests; scan PASSES with 9 negative controls.
      
      Two details the estate gets right and this corpus records rather than
      flattens: a soft binding leaves hard-binding `not-applicable` rather
      than failed -- the outcome vocabulary has a value for exactly that --
      and a PARTIAL ingredient chain is still valid-and-trusted, because the
      estate reports the gap and does not call the credential broken.
      
      Six corpus fault kinds, every one reachable from a distinct input in
      both spec and scan. -->
      
    • S11.14.b Tamper artifact bytes, manifest/assertion/ingredient/action/ actor/tool/policy fields, signature, certificate chain, timestamp, ledger record, and binding. Evidence: tamper corpus. <!-- TEN CHECKS, SIX STATES, AND THREE CHECKS THAT FAIL WITHOUT MOVING THE STATE AT ALL. A TAMPER CORPUS THAT ASSERTS ON THE STATE IS BLIND TO THREE OF THEM AND CANNOT TELL TWO PAIRS APART.

      CREDIT: c2pa-signing.mjs CONTROL 2 already performs two REAL tamper
      routes with real cryptography -- "a flipped asset byte breaks the hard
      binding while the signature over the manifest still verifies, and a
      flipped manifest byte breaks the signature". Demonstrated on real bytes,
      not described. Not re-done here.
      
      MEASURED by failing each of the ten VALIDATION_CHECKS one at a time and
      asking credentialState:
        manifest-structure      -> unknown-or-unsupported
        hard-binding            -> tampered
        signature               -> tampered
        revocation              -> revoked
        certificate-validity    -> expired
        certificate-chain       -> cryptographically-valid-untrusted
        trust-list              -> cryptographically-valid-untrusted
        timestamp               -> valid-and-trusted   (decidedBy NULL)
        assertion-support       -> valid-and-trusted   (decidedBy NULL)
        ingredient-availability -> valid-and-trusted   (decidedBy NULL)
      
      THE THREE ARE NOT A DEFECT AND THIS CELL DOES NOT REPORT THEM AS ONE. An
      assertion the validator does not support must not invalidate a
      credential -- that is the whole point of having an assertion-support
      OUTCOME rather than a failure. An ingredient that cannot be fetched says
      nothing about the signature over what is there. And a timestamp is
      load-bearing only once the certificate has expired, which is exactly the
      case CONTROL 4 of the estate's tool covers. All three are right to leave
      the state alone.
      
      What follows is a fact about the CORPUS, not the estate: the state is a
      summary and a tamper is a fact about the REPORT. So every case declares
      the check it trips, and the scan asserts on the check and on decidedBy,
      never on the state alone -- asserting on the state would give the
      thirteen targets six answers, three of which are the answer an
      untampered credential gives.
      
      SHIPPED src/tamper-corpus.ts + scan. Thirteen targets in the cell's own
      words over FIVE distinguishable routes, with EIGHT of the thirteen --
      manifest, assertion, ingredient, action, actor, tool, policy and the
      signature itself -- landing on the signature check, because everything
      inside a signed manifest breaks the same way and the report cannot say
      which field moved. Each of those eight still records what its field
      MEANS (an action is what was done, an actor who did it, a tool what made
      it, a policy what was agreed), because the reason they are
      indistinguishable is precisely that they matter differently. 18 spec
      tests; scan PASSES with 9 negative controls.
      
      AND THE ONE TARGET OUTSIDE VALIDATION. The LEDGER RECORD lives outside
      the asset, so the signature still verifies and the binding still holds,
      and ExternalWitness is a boolean pair with no integrity of its own -- an
      edited ledger answers exactly like an untouched one. Reported as
      `outside-the-validators-reach` rather than as a passing tamper, because
      "we checked and it survived" and "nothing looked" are different
      sentences and only the second is true. (A tampered trust list is the
      same shape; S11.14.d owns that.)
      
      Every number above is RECOMPUTED by the scan from the estate rather than
      read off the register, with a positive control that at least one check
      does move the state. -->
      
    • S11.14.c Add structurally/schema/signature/binding/chain/revocation/ timestamp/assertion/ingredient invalid variants with separable causes. Evidence: invalid corpus. <!-- SEPARABILITY IS A PROPERTY OF THE REPORT, NOT OF THE STATE.

      "Separable" is the whole requirement here: nine fixtures that all come
      back "invalid" have tested that the validator says no, not that it says
      why, and the repair for a broken chain and the repair for a stale
      revocation list are done by different people on different days.
      
      COMPUTED by calling credentialState for each variant rather than
      declared. SEVEN of the nine are separable, and only TWO of those by the
      STATE alone -- chain-invalid and revocation-invalid. The other five need
      the per-check report, because S11.14.b measured that ten checks collapse
      to six states. TWO ARE NOT SEPARABLE AT ALL, AND THEY ARE A PAIR:
      structurally-invalid and schema-invalid. Inseparability is a property of
      the PAIR rather than of one member -- nothing can say which of the two
      happened, so both are unidentified. This register first said "one is
      not" and the computation corrected it.
      
      THE CAUSE. VALIDATION_CHECKS is manifest-structure, hard-binding,
      signature, certificate-chain, certificate-validity, trust-list,
      revocation, timestamp, assertion-support and ingredient-availability.
      There is NO schema check, and the whole 1,293-line module contains the
      word "schema" once. So a credential manifest that PARSES and violates
      its schema is reported as manifest-structure -- the same answer a
      corrupt one gets.
      
      Those are different failures with different repairs. A structurally
      broken manifest is corrupt or truncated: a transport problem, fixed by
      fetching the bytes again. A schema-invalid manifest parses perfectly and
      says something the specification does not allow -- a wrong field type, a
      missing required member, a version the writer and reader disagree about
      -- and no refetch will ever help.
      
      AND THE KIT ALREADY KNOWS HOW TO TELL THEM APART, FOR THE OTHER
      MANIFEST. manifest-schema.ts carries BUNDLE_MANIFEST_SCHEMA with an id,
      a version, a path, an owner, a generation command and a pinned ajv
      validator, plus schemaRegistryFaults and validatorFaults to police it.
      The bundle manifest gets a schema registry; the credential manifest gets
      one check called manifest-structure.
      
      WHAT THIS CELL DELIBERATELY DOES NOT DO. It does not add a schema member
      to VALIDATION_CHECKS. S11.13.e widened VERIFICATION_ASPECTS for the same
      kind of gap and that list had no consumers outside specs; this one is
      read by ProvenanceInspector.tsx, provenance-inspector.ts, suggestions.ts
      and a browser tool, so widening it changes a SHIPPED PRODUCT SURFACE and
      makes every existing report emit check-never-asked for a check nothing
      performs. That is a product decision, not a fixture-corpus one. The gap
      is reported instead, and the scan re-measures the absence every run so
      the day a schema check appears this register goes red.
      
      SHIPPED src/invalid-corpus.ts + scan. 19 spec tests; scan PASSES with 8
      negative controls, including the positive control that matters: breaking
      the pair makes the inseparability finding GO AWAY, so it is a fact about
      this corpus rather than something the function always says. Five fault
      kinds, every one reachable. -->
      
    • S11.14.d Add unknown signer/root/trust list/assertion/algorithm/format/ version/ledger and offline or stale trust-list variants. Evidence: unknown corpus. <!-- A TRUST LIST THAT ANSWERS FROM A CACHE IS NOT A TRUST LIST THAT ANSWERED, AND THE ESTATE HAS unavailable FOR "COULD NOT ASK" AND NOTHING FOR "ASKED SOMETHING OUT OF DATE".

      CREDIT, AND IT IS THE SHARPEST CHECK IN THE MODULE.
      unavailable-revocation-treated-as-a-result does not look at the outcome
      at all -- it reads the DETAIL STRING, and fires when a revocation the
      validator could not reach carries a message reading "not revoked",
      "good" or "valid". The estate's own words: "an outage becomes a trust
      decision, silently, for every asset validated during it." A check on
      what a report SAYS rather than on what it concluded is a rare and good
      idea, and it is exactly this cell's offline hazard. Measured both ways:
      silent on an honest detail, loud on a lying one, and both give the same
      state -- so only the report separates them.
      
      MEASURED ACROSS ALL TEN CHECKS. An `unavailable` outcome collapses the
      state to unknown-or-unsupported and produces NO report fault, except
      that one. A `not-performed` outcome gives the same state AND a report
      fault (not-performed-reported-for-a-credential-that-is-present). So the
      estate already separates "nobody asked" from "we could not reach it", in
      the report, which is the right place for it.
      
      WHAT HAS NO VOCABULARY AT ALL IS THE THIRD ONE. A stale cache does not
      fail and is not unavailable: it RETURNS AN ANSWER, so the check reports
      pass, and a pass from a list fetched six months ago is byte-identical to
      a pass from one fetched a second ago. Every certificate revoked since
      the cache was filled validates cleanly.
      
      COMPUTED, not asserted, by comparing state AND report faults across five
      trust-list conditions: live, offline-cache-fresh and offline-cache-stale
      come back as ONE answer. `unreachable` IS distinguishable, which is the
      positive control that makes that group a measurement rather than a
      broken comparison.
      
      AND THE VOCABULARY EXISTS IN THE KIT ALREADY, ONE MODULE AWAY.
      trust-results.ts exports EXPIRY_STATES = fresh, expired, never-expires
      with a temporalFaults to police them. TRUST_LIST_FRESHNESS here
      re-exports those values rather than inventing a parallel set -- the same
      shape as S11.14.c's schema gap, where the concept is modelled elsewhere
      in this library and not wired to the credential validator.
      
      As in S11.14.c this does NOT add a member to CHECK_OUTCOMES, which
      ProvenanceInspector.tsx and suggestions.ts read. The freshness lives
      beside the corpus, so a fixture can say how old its answer was even
      though a ValidationCheckResult cannot.
      
      SHIPPED src/unknown-corpus.ts + scan. Eight unknown targets over four
      states, each recorded for what makes it different: unknown-signer is
      untrusted rather than invalid; unknown-assertion must not invalidate and
      does not; unknown-algorithm can be neither verified NOR refuted, so
      `unavailable` is the only honest outcome and a `fail` would be a claim
      the validator cannot make; unknown-format and unknown-version share a
      check, the same collapse S11.14.c found. 20 spec tests; scan PASSES with
      8 negative controls. -->
      
    • S11.14.e Add expired/not-yet-valid/revoked certificate, timestamp, trust-list, assertion/policy, consent/rights, and manifest retention variants. Evidence: lifecycle corpus. <!-- FIVE LIFECYCLE POSITIONS COMPUTED, AND NOTHING JOINS THE FUNCTION THAT KNOWS TO THE FUNCTION THAT DECIDES.

      CREDIT, AND IT IS A GOOD FUNCTION. validityAtSigning takes a notBefore,
      a notAfter, an optional timestamp and a validation instant, and returns
      FIVE positions: valid-now, valid-at-signing, expired, not-yet-valid and
      unknown-without-a-timestamp. All five are reachable and all five are
      driven here. The last two show it was thought about: a certificate whose
      notBefore has not arrived is NOT expired, and an expired certificate
      with NO VERIFIED TIMESTAMP is not expired either -- it is unknown,
      because nothing establishes when the signature was made. An unverified
      timestamp gets the same answer as none at all, which is correct and
      easily missed. CONTROL 4 of c2pa-signing.mjs drives the rescue case with
      a real certificate whose real notAfter has passed.
      
      THE GAP. credentialState does not take a validityAtSigning result. It
      takes a ValidationReport, on which the whole certificate lifecycle is
      ONE CheckOutcome for certificate-validity -- so a not-yet-valid
      certificate has to be encoded as a `fail`, and S11.14.b measured that a
      failing certificate-validity produces `expired`.
      
      MEASURED: expired, not-yet-valid AND unknown-without-a-timestamp all
      arrive as the same state, and CREDENTIAL_STATES has no not-yet-valid
      member at all, so the position could not be carried even by a caller
      that computed it.
      
      The estate itself says why this matters, about a different pair: revoked
      and expired are kept apart because "a lapsed certificate is a calendar
      problem whose content is probably fine and a revoked one puts everything
      that key ever signed in doubt. The two go to different people with
      different urgency, and a system that reported them alike would route the
      second one to whoever handles renewals." A NOT-YET-VALID certificate is
      a third thing again -- usually a clock wrong somewhere, quite possibly
      the validator's own -- and reporting it as expired routes a clock
      problem to whoever handles renewals for a certificate that has not
      started yet. validityAtSigning is exported from the barrel and called by
      the signing tool: it is not dead, it is UNJOINED, which is a different
      repair.
      
      THE POSITIVE CONTROL THAT MAKES THAT A FINDING RATHER THAN AN ARTEFACT:
      the two VALID positions collapse too, and that collapse is CORRECT. So
      the comparison is known to be able to report a collapse that is not a
      defect, rather than reporting its own filter.
      
      THE OTHER FIVE SUBJECTS, and they are not equally modelled. timestamp
      has its own check and feeds validityAtSigning. trust-list has a check
      and no freshness (S11.14.d). assertion-support is a check and a lapsed
      POLICY is not -- an assertion the validator can read under a policy that
      expired is supported and out of date at once. And consent/rights and
      manifest retention have NO model here at all, while the kit carries
      `withdrawn` in approval-workflows.ts and `retainUntilMs` in
      audit-snapshots.ts. THAT IS THE THIRD TIME THIS SECTION HAS FOUND THE
      CONCEPT A MODULE AWAY AND UNWIRED, after S11.14.c's schema and
      S11.14.d's freshness, and it is recorded as
      `modelled-elsewhere-in-the-kit` rather than absent because one is a wire
      and the other is design work. The scan opens each named module and
      checks it really contains the symbol claimed.
      
      SHIPPED src/lifecycle-corpus.ts + scan. 22 spec tests; scan PASSES with
      9 negative controls. -->
      
    • S11.14.f Strip embedded metadata through realistic channel/transcode/ resize/crop/recompress/rerecord and test sidecar/soft-binding/ledger recovery, ambiguity, and failure. Evidence: stripped corpus. <!-- A SOFT BINDING IS A SEARCH AND softBindingRecovered IS A BOOLEAN.

      CREDIT: the estate's wording about a soft binding is exactly right. When
      a credential is absent and a soft binding matched, credentialState
      returns partially-recoverable and says why -- "there is a claim about
      what this probably is and NOTHING THAT BINDS IT TO THESE BYTES." That is
      the whole difference between a hash and a perceptual match in one
      sentence, and soft-binding-alone-where-a-hard-one-is-available is a
      fault kind for using the weaker route when the stronger was on offer.
      
      WHAT THE FIELD NAME ALREADY TELLS YOU, AND IT DECIDES HALF THIS CELL.
      ExternalWitness.ledgerHoldsCredentialForDigest is keyed ON THE DIGEST,
      so a ledger can only answer for bytes that still hash to what was
      registered. That splits the six channels cleanly:
      
        channel-strip   metadata removed, pixels untouched -> the digest still
                        matches, so THE LEDGER CAN ANSWER
        transcode  \
        resize      |   different bytes, so a different digest, so THE LEDGER
        crop        |   IS USELESS -- not because anything was lost, but
        recompress  |   because the recovery route is keyed on the one thing
        re-record  /    these channels change
      
      FIVE OF SIX DEFEAT THE LEDGER. A corpus that tested "stripping" with a
      single strip-only fixture would find the ledger works and stop.
      
      THE AMBIGUITY. A perceptual match over a corpus returns nothing, exactly
      one thing, or SEVERAL, and several is strictly worse than none: a wrong
      answer displays as confidently as a right one. A boolean says "something
      matched"; it cannot say "two different credentials both claim these
      bytes", which is the case where a reader is most likely to be shown
      provenance belonging to something else.
      
      COMPUTED: the one-match and several-match cases produce the same state
      AND the same witness. THE POSITIVE CONTROL is that ZERO matches IS
      distinguishable -- which is exactly what a boolean can carry, and what
      makes the collapse a measurement rather than a broken comparison.
      
      MATCH_CARDINALITY lives beside the corpus rather than widening
      ExternalWitness, which credentialState itself and ProvenanceInspector.tsx
      read. Every fixture is still run through the estate with the boolean the
      estate has -- the third time this section has reported a gap rather than
      changing a shipped surface.
      
      SHIPPED src/stripped-corpus.ts + scan. 23 spec tests; scan PASSES with 9
      negative controls. Also recorded: re-record is the one channel where
      even a perceptual match should fail, because geometry, colour and noise
      all move at once; and a SIDECAR survives a strip with the credential
      PRESENT, which is why shipping both is worth the trouble. -->
      
    • S11.14.g Assert crypto validity, trust, binding, assertion, ingredients, policy, and recovery states independently in API/UI/gates. Evidence: state matrix. <!-- THE CELL ASKS FOR SEVEN DIMENSIONS ASSERTED INDEPENDENTLY, AND EVERY MEASUREMENT IN S11.14 HAS BEEN OF THE SAME SCALAR.

        .b  ten checks reach six states, and three fail without moving it
        .c  seven of nine invalid variants separable, only two by the state
        .d  a live lookup, a fresh cache and a stale cache are one answer
        .e  five lifecycle positions arrive as two states
        .f  one perceptual match and several are one answer
      
      Each is the same shape: credentialState returns ONE CredentialState and
      everything a reader would act on differently arrives flattened. THAT IS
      NOT A DEFECT IN credentialState -- collapsing to one headline is what a
      headline is for, and `decidedBy` exists precisely because the author
      knew the headline was not enough. What was missing is the other half.
      
      CREDIT, AND THE ESTATE ALREADY NAMES THIS EXACT FAILURE.
      VALIDATION_SURFACE_FAULTS in provenance-inspector.ts (2,642 lines)
      includes one-badge-instead-of-the-matrix,
      state-rendered-without-the-check-that-decided-it,
      ingredient-availability-summarised-without-naming-the-ingredients and
      unavailable-detail-that-reads-as-an-answer. validationSurfaceFaults is
      called for real by apps/oshun/web/tools/provenance-inspector-browser.mjs,
      and that module's CREDENTIAL_CARRIAGE even includes the `ledger`
      S11.14.a found missing from EMBEDDING_MECHANISMS -- correctly, because
      carriage and embedding are different questions. RenderedCheck and
      RenderedCredentialState had NO producer anywhere in the repo.
      
      SHIPPED src/state-matrix.ts + scan. Seven dimensions read side by side,
      each cell reading only its own checks so a failing signature does not
      make the ingredients cell say anything. dimensionsTheHeadlineHides names,
      per failing check, which dimension a valid-and-trusted headline is
      covering up: MEASURED AT 3 OF 10. 28 spec tests; scan PASSES with 8
      negative controls.
      
      TWO OF THE SEVEN ARE NOT CHECKS. `policy` is answered by NOTHING -- no
      check, no outcome, no report field -- so a credential can be
      cryptographically perfect and used on terms nobody agreed to. THIRD
      SIGHTING: S11.13.d found rights and consent uncarried by the bundle
      format, S11.14.e found consent modelled in the approval workflows and
      not here. It reads `no-input-in-the-report` rather than failing, because
      a dimension nothing can answer and one answering badly are different
      repairs. `recovery` reads the ExternalWitness, and that is CORRECT:
      whether provenance can be got another way is a fact about the world
      outside the asset.
      
      AND ONE CHECK THE SEVEN HEADINGS DO NOT NAME. manifest-structure is not
      cryptography, trust, a binding, an assertion, an ingredient, a policy or
      a recovery -- it is whether the document could be read at all, upstream
      of every other question, and S11.14.c found it doing double duty for a
      schema violation too. Reported through check-no-dimension-reads rather
      than forced into a heading it does not belong to.
      
      TWO DEFECTS THIS MODULE'S OWN CHECKER FOUND IN IT ON ITS FIRST RUN.
      (1) `recovery` read `does-not-hold` for a credential that was PRESENT,
      putting a red cell beside six green ones on a perfectly good credential
      -- and worse, making every headline-hides comparison name recovery for
      every input. A DIMENSION THAT DOES NOT APPLY IS NOT A DIMENSION THAT
      FAILED; it now reads `does-not-arise`. (2) `timestamp` was read by no
      dimension until it was put where it belongs, with crypto-validity, since
      it establishes WHEN the signature was made -- which is exactly what
      validityAtSigning needs it for.
      
      The rendered matrix is handed to the estate's own validationSurfaceFaults
      and accepted, with a positive control that a NO-ROW surface is REFUSED
      with eleven findings. -->
      
  • S11.15 Add performance budgets for route load, interaction latency, save, search, compare, and virtualized scrolling. Evidence: CI trend report. <!-- a-g all closed. The evidence clause is the CI trend report, and S11.15.g measures it end to end: real repetitions asserted against the median with traces attached, and an arithmetic ceiling on what those repetitions can distinguish.

    THE SECTION'S SHAPE, SIX TIMES. Every cell found a real instrument
    measuring something adjacent to what the cell asked for, and named the
    difference rather than crediting the near-miss:
    
      .a  a real budget file CI enforces, at ONE device/network profile, and
          memory is a dimension no Lighthouse budget can carry
      .b  A POTENTIAL IS NOT A LATENCY -- max-potential-fid is measured on a
          page nobody interacted with; 14 time bounds in 176 specs, each a
          literal in the spec that asserts it
      .c  the save pipeline has STATES AND NO CLOCK, and a retry ceiling is a
          count: five attempts with no delay and five over four minutes are
          the same observation
      .d  A COMPLEXITY BUDGET BOUNDS WORK, NOT TIME -- the best-budgeted part
          of the estate, budgeting cost, freshness and backoff, and never the
          wait
      .e  every second in the compare module is a COORDINATE IN THE CONTENT,
          and there is no status for "too large"
      .f  renderedRowCount is RECORDED AND COMPARED TO NOTHING -- the number
          virtualization exists to keep small
      .g  three runs give two distinct statistics, and a threshold is not a
          trend
    
    THE DISCIPLINE THAT MADE IT HONEST: NO CELL INVENTED A NUMBER. A budget
    nobody measured is a number somebody will later read as a limit that was
    checked, which is this repository's bright line. Every register carries
    instruments and coverage; the numbers stay where the measurement is, and
    each module has a fault kind that fires if a digit appears in it. S11.15.f
    refined that guard after it fired on a cell reference and a spec filename
    -- it now requires a digit to OPEN a token, catching "200 ms" and letting
    S11.15.a through, tested both ways.
    
    AND THE VALUES THAT KEPT THE FINDINGS FAIR: a-load-time-potential apart
    from nothing (.b), an-order-or-count-check apart from nothing (.c), five
    bound kinds rather than two (.d), covered-as-correctness and
    recorded-and-unbounded as distinct states (.f), and
    runs-but-cannot-distinguish apart from not-met (.g). In every case
    collapsing them would have credited a real check as the wrong kind of
    check, or called a properly-guarded thing uncovered.
    
    Kit: 279 files / 16,275 tests green, 283 entry points, 0 violations, tsc
    clean and unfiltered; all seven scans re-run after their registrations. -->
    
    • S11.15.a Define route navigation/load/usable timing, JS/CSS/image/font/ data transfer, requests, main-thread/CPU/memory, and layout shift budgets by device/network. Evidence: route budgets. <!-- EVERY NUMBER IN THE BUDGET FILE IS ONE POINT IN A TWO-DIMENSIONAL SPACE.

      CREDIT, AND IT IS A REAL GATE THAT REALLY FAILS CI.
      apps/oshun/web/lighthouse-budget.json is not a document about budgets --
      it is the file Lighthouse CI enforces. Three path entries (/*, /search,
      /domains/nisaba), each carrying six timing budgets
      (first-contentful-paint, largest-contentful-paint, max-potential-fid,
      cumulative-layout-shift, total-blocking-time, speed-index), five
      resource-size budgets (total, script, stylesheet, image, font) and four
      request-count budgets (total, script, stylesheet, third-party). Between
      them those cover TWELVE OF THE THIRTEEN dimensions this cell names, and
      lighthouserc.cjs runs the whole thing three times per route behind a
      real authenticated session.
      
      THE GAP. The cell asks for budgets BY DEVICE AND NETWORK, and
      lighthouserc.cjs sets `preset: 'desktop'` with one throttle -- rttMs
      150, throughputKbps 1638.4, cpuSlowdownMultiplier 4 -- with no
      formFactor and no mobile config in either the web or the admin runner.
      So "by device/network" has no second value to be "by".
      
      That is NOT a criticism of the numbers, which are sensible, and the
      throttle is a real one rather than an unthrottled best case: a 4x CPU
      slowdown at roughly 1.6Mbps is a deliberately unkind desktop. It means
      something narrower and worth saying plainly -- the file states what a
      throttled desktop must achieve, and says nothing about the phone on a
      bad connection, which is the device a slow route hurts most and the one
      whose numbers would differ in KIND rather than in degree.
      
      AND ONE DIMENSION NO LIGHTHOUSE BUDGET CAN CARRY. The cell names MEMORY.
      resourceSizes is bytes over the wire, not bytes resident, and the budget
      file mentions memory zero times -- so a route that leaks 200MB into the
      heap passes every line of it. Marked needs-a-different-instrument with
      the instrument NAMED (a CDP heap sample from the existing Playwright
      inspect harness) rather than left looking covered.
      
      WHAT THIS CELL REFUSES TO DO IS INVENT NUMBERS. A budget nothing
      measures is a number somebody will later mistake for a limit that was
      checked. Every dimension is either bound to the real budget file -- with
      the value READ from it by the scan, never restated in the register -- or
      marked with the instrument it would need and NO NUMBER AT ALL, and a
      fault kind fires if a digit ever appears in the register.
      
      SHIPPED src/route-budgets.ts + scan. 20 spec tests; scan PASSES with 9
      negative controls, and re-reads both estate files every run, so adding a
      mobile profile or a memory budget -- both good news -- turns this
      register red. -->
      
    • S11.15.b Define input/selection/command/panel/table/tree/graph/timeline/ media interaction latency and dropped-frame/long-task budgets. Evidence: interaction budgets. <!-- A POTENTIAL IS NOT A LATENCY.

      CREDIT, AND THERE IS A REAL INSTRUMENT.
      apps/oshun/web/e2e-inspect/phase-7-4-long-run.spec.ts installs a genuine
      PerformanceObserver on `longtask`, drives forty frames through
      requestAnimationFrame, and bounds the elapsed time of a full-document
      scroll -- and it is honest about its own limits in a comment: "Not every
      engine exposes longtask; the wall clock still counts." That is the right
      way to carry an instrument that is not available everywhere. Long tasks
      and dropped frames ARE measurable here, on a real browser, today.
      
      THE FIRST HALF. The Lighthouse assertions cover interactivity with
      max-potential-fid, total-blocking-time and interactive, and ALL THREE
      ARE LOAD METRICS. max-potential-fid is an estimate derived from the
      longest task during page load, on a page NOBODY HAS INTERACTED WITH: it
      answers "if a user had clicked at the worst moment, how long might the
      page have taken to respond", which bounds the page's capacity to be slow
      rather than measuring any interaction that happened. Nine kinds of
      interaction are named by this cell and none of them is what that number
      is about.
      
      THE SECOND HALF, MEASURED ACROSS ALL 176 SPECS IN e2e-inspect: FOURTEEN
      toBeLessThan bounds in the entire suite. Several are not times at all --
      0.4 and 1.5 are contrast ratios, 1024 is bytes, 2 and 24 are counts --
      leaving roughly eight genuine time bounds, every one a literal written
      in the spec that asserts it, no two the same, and CAP_MS the only named
      constant among them. (A first pass counted 33 specs with a bound; the
      precise regex found 14 bounds in 12 specs, and the register is built on
      the second number.)
      
      That is NOT a criticism of those specs -- a spec bounding the thing it
      drives is exactly right. It means the estate has interaction ASSERTIONS
      and no interaction BUDGETS, and the difference matters three ways.
      Nothing says what the budget for a table interaction IS, so two specs
      may disagree and neither is wrong. Changing a budget means finding every
      literal. And a kind with no spec is INDISTINGUISHABLE from a kind whose
      spec is generous: unbudgeted and untested render identically.
      
      SHIPPED src/interaction-budgets.ts + scan. Coverage per kind across
      three instruments -- a-driven-interaction, a-load-time-potential and
      nothing -- kept as THREE because a page carrying a max-potential-fid
      budget is not unmeasured, it is measured for a different question, and
      calling that coverage is the mistake this cell exists to name. Measured:
      5 of 11 kinds driven, 3 with only a potential, 3 with nothing.
      
      The scan OPENS every spec the register names and confirms it CONTAINS a
      bound rather than merely existing, because a named file that is not the
      instrument it is named as is the same defect as no instrument at all. As
      in S11.15.a it invents no numbers, and a fault kind fires if a digit
      appears in any entry. 16 spec tests; scan PASSES with 8 negative
      controls. -->
      
    • S11.15.c Define local feedback, autosave dispatch, durable ack, offline queue, retry, and conflict timing budgets. Evidence: save budgets. <!-- THE SAVE PIPELINE IS MODELLED AS A SEQUENCE OF STATES AND CARRIES NO CLOCK.

      CREDIT, AND THE MODELLING IS EXCELLENT. block-editors.ts S7.8.f has
      AUTOSAVE_STATES of idle, pending, in-flight, saved, failed,
      queued-offline and conflicted -- every stage this cell names, under
      different words. QUEUE_DURABILITY separates in-memory and
      session-storage from local-storage and indexed-db, and DURABLE_STORES
      names which two actually survive a reload. And
      `saved-shown-before-the-commit` is the best check in the section: it
      fires when the surface says "saved" and commitAcknowledged is false,
      which is the difference between telling a writer their work is safe and
      it being safe.
      
      MEASURED ON THE MODULE: ZERO time-typed fields anywhere in it.
      AutosaveObservation's eleven fields are a state, two id lists, four
      booleans, a durability, and retryAttempts and retryCeiling -- which are
      COUNTS. All eight AUTOSAVE_FAULTS are about order, presence, durability
      or a count.
      
      So the estate checks that the states happen in the right ORDER and
      nothing checks how long any of them TAKES. A save that reaches `saved`
      thirty seconds after the keystroke, with the commit genuinely
      acknowledged and every entry carrying its idempotency key, passes all
      eight -- and it is the failure a writer actually experiences, because
      the thing they are waiting for is the acknowledgement, not its ordering.
      `retries-without-a-ceiling` is the closest the module comes, and it is a
      ceiling on the NUMBER of attempts: five with no delay and five spread
      over four minutes are the same observation.
      
      WHY THE ORDER CHECK IS STILL THE MORE IMPORTANT ONE, and this cell says
      so rather than implying otherwise: a slow save loses time, and a "saved"
      shown before the commit loses work. The estate spent its checks on the
      worse failure. What follows is that the timings are missing, not that
      the modelling is wrong.
      
      SHIPPED src/save-budgets.ts + scan. Six stages, each tied to the
      AUTOSAVE_STATE it reaches so the register is BOUND to the estate rather
      than parallel to it; each naming the estate check that covers its
      ordering (four do) or recording that nothing does (two); each naming the
      instrument that would bound its duration. Measured: 0 of 6 stages timed.
      
      The scan re-counts the absence with a POSITIVE CONTROL that the same
      duration search finds many names in jobs.ts -- because a zero from a
      search nobody proved can find anything is not a measurement -- and
      checks every state and check name against the estate's own lists. 22
      spec tests; scan PASSES with 9 negative controls. Invents no numbers,
      and a fault kind fires if a digit appears in a reason or an
      instrument. -->
      
    • S11.15.d Define search typeahead/debounce/query/server/first result/full result/facet/pagination and expensive query rejection budgets. Evidence: search budgets. <!-- THE SEARCH STACK BUDGETS EVERYTHING EXCEPT HOW LONG A SEARCH TAKES.

      CREDIT, AND THIS IS THE BEST-BUDGETED PART OF THE ESTATE. Unlike the
      save pipeline, the search stack bounds things, and hard things:
      
        complexityBudget         query-grammar.ts scores a query BEFORE
                                 running it, with per-field cost multipliers
                                 and a comment that wildcards are expensive,
                                 refusing over budget with
                                 complexity-budget-exceeded. That is this
                                 cell's expensive-query rejection,
                                 IMPLEMENTED -- and the comparison is a real
                                 comparison, not a declared field.
        staleToleranceSeconds    how out of date an index may be before a
                                 reader is told, which is the freshness
                                 question search stacks usually leave to hope.
        initial/maxDelaySeconds  a real backoff schedule for re-indexing
                                 rather than a fixed retry.
      
      MEASURED across query-grammar.ts, search-index-events.ts and
      search-permission-projection.ts: no `latency`, no responseMs, no
      queryMs, no firstResult -- nothing naming the elapsed time between a
      keystroke and a result. Every duration in those three modules is a
      FRESHNESS tolerance or a BACKOFF delay, both about the index, and the
      reader is waiting on the query.
      
      AND A COMPLEXITY BUDGET IS A BOUND ON WORK, NOT ON TIME. That is the
      whole cell. `cost > context.complexityBudget` is evaluated against a
      static score for the query SHAPE, so it promises how much work the query
      ASKS FOR -- and work is a proxy for time only when the machine is not
      busy. A query comfortably under budget, run against a shard that is
      re-indexing or serving someone else's wildcard, is slow, and nothing in
      the stack notices: the cost check passed, the index was inside its
      staleness tolerance, no retry was needed, and the reader waited anyway.
      It is a genuinely good check, and it is a check on the QUERY. The nine
      stages are about the RUN.
      
      SHIPPED src/search-budgets.ts + scan. FIVE bound kinds rather than two
      -- latency, cost, freshness, backoff, nothing -- because collapsing them
      would credit a cost check as a latency check. Measured: 0 of 9 stages
      have a latency budget, 3 are bounded by something else, 6 by nothing.
      
      first-result and full-result are separate stages because the estate's
      own TOTAL_EXACTNESS admits a `lower-bound` total, so a surface CAN show
      something early -- and whether it does is a latency trade nothing
      records.
      
      The scan re-measures the absence with a positive control in
      service-objectives.ts, and confirms every credited symbol is really in a
      search module. 21 spec tests; scan PASSES with 8 negative controls.
      Invents no numbers. -->
      
    • S11.15.e Define compare load/alignment/diff/render/navigate/filter for text/metadata/media/3D/graph/gate and large input bounds. Evidence: compare budgets. <!-- EVERY SECOND IN THE COMPARE MODULE IS A COORDINATE IN THE CONTENT, NOT A DURATION OF THE WORK.

      CREDIT: comparisons.ts covers every content kind this cell names and
      then some. COMPARE_ADAPTERS is text, metadata, image, video, audio,
      stage-3d, graph and gate -- the cell's six with media split into its
      three real cases, which is the right split, because aligning two videos
      and aligning two audio tracks are different problems. And
      COMPARISON_STATUSES separates computed, partial, unsupported, withheld
      and unavailable, so a comparison that could not run says which kind of
      could-not.
      
      MEASURED: six duration identifiers across comparisons.ts and diff.ts,
      and every one is about the material. atSeconds is documented as "Frames
      sampled at declared SECONDS, so two rates can be aligned at all" -- a
      position in a video. durationSeconds is how long the media is.
      windowSeconds is the RMS window a waveform picture is drawn from.
      measuredAtEpochMs is WHEN a measurement was taken, not how long it took,
      and that third subject is kept separate because calling a timestamp a
      timing would credit an instant as a measurement.
      
      So the module is full of time and none of it is its own. A 3D stage
      comparison that takes forty seconds to align two scenes produces exactly
      the same `computed` status as one that takes forty milliseconds, and the
      seconds in the result are the scene's.
      
      AND THERE IS NO STATUS FOR TOO LARGE. The cell asks for large-input
      bounds. COMPARISON_STATUSES has five values and none is a refusal for
      size: unsupported is about the adapter, unavailable about reachability,
      withheld about permission, and partial about some of it being computed
      -- which is what a size-limited comparison would produce BY ACCIDENT,
      reported as though the missing part were incidental rather than
      declined.
      
      That matters more here than anywhere else in S11.15, because comparison
      is the one operation whose cost grows with the SQUARE of its input in
      the general case. A diff of two large documents, an alignment of two
      long videos and a graph comparison over two big graphs are the three
      places a workbench most easily becomes unusable, and the module has no
      vocabulary for declining any of them.
      
      SHIPPED src/compare-budgets.ts + scan. The scan EXTRACTS the duration
      identifiers from the source and fails if the register does not classify
      one, so a name added later must be classified rather than assumed -- and
      its first run caught its own extractor matching a bare `Seconds` out of
      a doc comment, which is prose about the content and not a field; it now
      requires an identifier. 21 spec tests; scan PASSES with 8 negative
      controls. Invents no numbers. -->
      
    • S11.15.f Define virtualized initial/scroll/page/jump/filter/resort/ selection/focus/memory/DOM count and accessibility semantic budgets. Evidence: scrolling budgets. <!-- THE VIRTUALIZATION MODEL IS ABOUT WHAT THE WINDOW MISREPRESENTS, NOT WHAT IT COSTS.

      CREDIT, AND THE CORRECTNESS MODEL IS EXCELLENT. data-tables.ts S7.5.h
      has the right numbered rules: "a selection that only counts rendered
      rows is a lie virtualization tells"; "aria-rowcount IS THE TOTAL, NOT
      THE WINDOW"; "print and export need the nonvirtual path". And
      VIRTUALIZATION_FAULTS covers aria-rowcount-reports-the-window,
      aria-rowindex-not-absolute, dynamic-row-heights-assumed,
      focus-lost-when-a-row-is-recycled, scroll-anchored-to-a-pixel and
      print-and-export-render-the-window. That is this cell's
      accessibility-semantic budget, its focus dimension and its selection
      dimension, ALL COVERED as CORRECTNESS -- which is the right way to cover
      them and worth more than any timing. Every one of the six is about the
      gap between the WINDOW and the TOTAL.
      
      MEASURED on VirtualizationObservation: eight fields, four counts and
      four booleans, ZERO durations and ZERO memory. Both re-taken by the scan
      with positive controls (57 duration names in jobs.ts, 8 byte names in
      artifact-rehash.ts).
      
      AND THE SHARPEST INSTANCE OF THIS SECTION'S SHAPE: renderedRowCount IS
      RECORDED AND COMPARED TO NOTHING. It is the number virtualization exists
      to keep small. It sits on the observation, it is printed into a fault
      detail as renderedRowCount/totalRowCount, and no check anywhere asks
      whether it is too big -- so a table that virtualizes correctly by every
      one of the six rules, with absolute indices and a total rowcount and
      focus surviving recycling, and renders nine thousand rows into the DOM,
      passes all of them. The window is honest about being a window, and
      nothing bounds the window.
      
      Memory is the second absence and the same one: S11.15.a found the route
      budgets cannot carry memory because Lighthouse measures bytes over the
      wire, and here the observation has no field for it either -- so a
      virtualized table's whole argument, that it holds a bounded amount, is
      bounded nowhere.
      
      SHIPPED src/scrolling-budgets.ts + scan. FIVE coverage values rather
      than two, because covered-as-correctness and recorded-and-unbounded are
      real and distinct states: calling a properly-guarded dimension uncovered
      would be wrong, and calling it budgeted would be wronger. Measured: 0
      budgeted, 5 covered as correctness, 1 recorded and unbounded, 5 with
      nothing. 23 spec tests; scan PASSES with 10 negative controls.
      
      A REFINEMENT THIS CELL MADE TO ITS OWN GUARD: the no-numbers check fired
      on a cell reference and a spec filename, neither of which is a budget.
      It now requires a digit to OPEN a whitespace-delimited token, which
      catches "200 ms" and "50 rows" and lets S11.15.a through -- tested both
      ways. And its memory positive control failed on the first module it
      tried, which is a control working: it said the search was unproven, and
      it was. -->
      
    • S11.15.g Run controlled repetitions, record median/p75/p95/worst/ variance, compare baseline/trend, fail significant regression, and attach traces. Evidence: CI trend report. <!-- THREE RUNS CANNOT PRODUCE A p75, A p95 AND A WORST THAT DIFFER.

      CREDIT: the repetitions are real. lighthouserc.cjs sets numberOfRuns and
      every route is measured that many times behind a real authenticated
      session under a fixed throttle, with Lighthouse CI asserting against the
      MEDIAN rather than the best or the last -- the right choice, and the one
      that makes the repetitions worth taking. The traces are attached too, so
      a red gate arrives with its evidence.
      
      THE ARITHMETIC, and it is arithmetic rather than opinion. With three
      sorted samples and nearest-rank quantiles the median is rank 2 and every
      one of p75, p95 and worst is rank 3 -- the same observation reported
      under three names. The cell asks for five statistics; three runs yield
      TWO distinct values, and a variance estimated from three points has so
      wide an interval that it cannot separate a regression from a noisy
      morning.
      
      DERIVED, NOT ASSERTED. distinctStatisticsAt computes this for any n, and
      runsNeededForDistinctStatistics derives that TWENTY runs are needed
      before all four read different samples -- which is also where the usual
      variance rule of thumb lands, and that rule is exposed as a named
      constant a reader can disagree with rather than buried in a comparison.
      The scan READS numberOfRuns from the runner, so the arithmetic is about
      the gate that exists, and raising it turns this register red.
      
      AND `temporary-public-storage` IS NOT A BASELINE. The upload target is
      named for what it is: somewhere the reports go and do not stay. Nothing
      in the workflow keeps a previous run and nothing compares against one,
      so what the gate compares against is the budget file, which is a set of
      CONSTANTS -- and "fail a significant regression" is not what happens. A
      route that was well inside its budget and has since quadrupled passes,
      because the budget is all there is to compare with. A CONSTANT CATCHES A
      ROUTE THAT BECAME BAD; A BASELINE CATCHES A ROUTE THAT IS BECOMING BAD,
      and those are different failures caught at different times.
      
      A THRESHOLD IS A CLAIM ABOUT ONE RUN AND A TREND IS A CLAIM ABOUT A
      SEQUENCE. That is why the two halves are related: with three runs there
      is no distribution to speak of, and with temporary storage there is no
      sequence to compare against. Fixing either alone leaves the other.
      
      SHIPPED src/trend-report.ts + scan. THREE requirement states rather than
      two, because a statistic that is COMPUTED AND DEGENERATE is not a
      statistic nobody computed, and the repairs differ: one is more runs, the
      other is new machinery. Measured: 4 requirements met, 3 that run and
      cannot distinguish (p75, p95, variance), 3 not met (baseline, trend,
      fail-significant-regression). 25 spec tests; scan PASSES with 8 negative
      controls. -->
      
  • S11.16 Add high-volume API and job load profiles. Evidence: limits and backpressure activate before resource exhaustion. <!-- WHAT THE ESTATE CAN DECIDE AND WHAT IT CAN ONLY DESCRIBE, SEPARATED SEVEN TIMES.

    The section asked for load profiles and the evidence that limits activate
    before exhaustion. The limit vocabulary turned out to be complete and
    several of the limiters really run -- .d drove the rate limiter and
    watched five actors stopped by the TENANT bucket, and .f found eight
    assertion families of which most are callable. What is absent is the other
    half of every sentence: the ARRIVALS a profile describes, and the
    MEASUREMENT that would say a limit fired before something ran out.
    
    .e is the finding the rest of the section leans on: the deployed estate
    emits no telemetry, so nothing observes traffic. That is why .a could not
    state a traffic proportion, why .c has no key-skew vocabulary, and why
    .g's false-success checkers can only read descriptions of a run. A limit
    that refuses is decidable from the code; that it refused BEFORE
    exhaustion is a measurement nobody takes.
    
    NO CELL IN THIS SECTION INVENTED A NUMBER, and each module has a fault
    kind that fires if a digit opens a token in its register. The numbers that
    do appear were computed by driving the estate: .d's admission decisions,
    .f's payload verdicts, .g's scope keys and claim outcomes.
    
    Seven cells, seven scans, all passing with negative controls; the section
    closes with the kit at 290 entry points and 0 violations. -->
    
    • S11.16.a Define realistic/surge/abuse API mixes for reads/search/count/ upload/mutation/bulk/stream/health and tenant/actor/key distributions. Evidence: API profiles. <!-- THE ESTATE HAS EVERY DIMENSION A LOAD PROFILE WOULD BE BUCKETED BY, AND NO PROFILE.

      CREDIT, AND THE LIMIT VOCABULARY IS COMPLETE. limits.ts has
      RATE_DIMENSIONS -- sustained-rate, burst, concurrent-requests,
      concurrent-streams, concurrent-uploads, concurrent-jobs and
      concurrent-provider-calls -- every axis a load profile would be measured
      against, including the three concurrency kinds that behave differently
      under the same request rate. And BUCKET_KINDS is actor, tenant, route,
      client-class and trust-tier, which is this cell's tenant/actor/key
      distributions as the dimensions a limit is bucketed by.
      
      THE GAP. A rate rule is a CEILING; a mix is a DISTRIBUTION OF ARRIVALS.
      Neither derives from the other, and the measurement shows why: the eight
      operations exhaust SIX different resources first -- connection pool,
      cpu, disk, memory, network egress, nothing in particular. So a limit
      says nothing about whether the traffic underneath it is mostly reads
      with a few uploads or the reverse, and those exhaust completely
      different things at the SAME total rate. One fills a connection pool;
      the other fills a disk.
      
      A NOTE ON A WORD, KEPT FRESH BY THE SCAN. capacity-plugins.ts uses
      "distribution" twenty-four times and means CONTENT DELIVERY by it --
      distribution.channel-slots, distribution.egress-mbps,
      distribution.delivery. Searching the estate for that word and concluding
      arrival distributions are modelled would be wrong.
      
      SHIPPED src/api-profiles.ts + scan. The SHAPE of the three mixes: which
      operations each contains, which bucket dimensions each varies, what
      distinguishes it. A SURGE SHARES THE OPERATION SET OF REALISTIC, because
      a surge is a RATE and not a shape -- one that also changed the mix would
      be two experiments at once. ABUSE IS A SHAPE AND NOT A RATE: the four
      operations whose cost is set by the caller, concentrated in few actors
      so per-tenant bucketing does not spread it, and omitting health because
      an abuser has no reason to send one.
      
      AND WHAT IT REFUSES. It states NO proportions. "Seventy per cent reads"
      is a measurement of real traffic and there is no telemetry in this
      repository to take it from -- writing one down would invent the thing
      the profile exists to represent, and every later run would validate a
      guess against itself. Every mix carries proportionsFrom:
      `not-yet-observed` and the register reports it, which is the finding
      rather than a defect in the register.
      
      Health is marked as the operation that MUST NOT BE SHED: its answer
      decides whether the instance stays in rotation, so shedding it under
      load removes the instance that was coping. 22 spec tests; scan PASSES
      with 9 negative controls. -->
      
    • S11.16.b Define job mixes for short/long/GPU/provider/render/evaluator/ import/export/publish/parent-child/retry/cancel and priority/tenant distributions. Evidence: job profiles. <!-- EVERY DIMENSION OF A JOB MIX IS MODELLED SOMEWHERE, AND THE JOB'S OWN KIND IS A STRING.

      CREDIT, AND IT IS BROADER THAN IT LOOKS. jobs.ts carries
      expectedAttemptSeconds and estimatedAttemptCost per job, so short and
      long are ONE measured axis with a threshold rather than two labels --
      and the queue already reads the duration to decide whether a deadline
      can be met. job-graphs.ts is 1,076 lines of parent and child with
      FAILURE_POLICIES (fail-parent, degrade-to-partial),
      CANCELLATION_PROPAGATIONS (cancel-children, let-children-finish),
      PARENT_OUTCOMES and REUSE_DECISIONS, all closed sets -- and it IMPORTS
      JobState from jobs.ts, so the two are joined rather than parallel.
      capacity-plugins.ts carries cpuCores, memoryGb, storageGb and gpuCount
      on RenderWorkerLoad, so a GPU job's demand is a modelled quantity rather
      than an adjective. And PRIORITY_BAND is a real validated band with a job
      outside it REFUSED.
      
      THE GAP. JobRecord.type is `string` and JobRecord.queue is `string`, and
      nothing in jobs.ts compares either against anything -- measured at ZERO
      comparisons against job.type, with a POSITIVE CONTROL that the same
      shape of search finds the priority comparisons that DO exist, so the
      zero is a fact rather than a broken query.
      
      Five of the twelve kinds therefore exist as free text. Two spellings of
      one kind are two kinds. A kind that needs a GPU worker and a kind that
      needs a provider quota are the same shape to the queue. And a mix
      expressed as proportions over kinds cannot be checked against what
      actually ran, because "what kind was that" has no closed answer. The
      priority band shows what the alternative looks like: a dimension a
      profile can be written in AND verified against.
      
      WHY THIS IS NOT A COMPLAINT ABOUT `string`. A queue that shipped a
      closed union of job types would have to be edited by every team that
      adds a job, and that is a worse design -- the estate is right that the
      queue should not own the catalogue. What is missing is that NOBODY ELSE
      OWNS IT EITHER. A registry the queue validates against, held wherever
      job types are actually decided, is the shape that gives a mix something
      to be a mix OVER.
      
      SHIPPED src/job-profiles.ts + scan. modelled-in is kept SEPARATE from
      expressibility, because `provider` is modelled in jobs.ts and is still
      only a label: where a kind lives and whether a profile can be checked
      against it are two questions. Measured: 5 of 12 kinds expressible only
      as labels, 4 modelled nowhere at all. 21 spec tests; scan PASSES with 8
      negative controls. States no proportions, for the reason S11.16.a
      gives. -->
      
    • S11.16.c Define data/artifact sizes, hot/cold keys, pagination/search complexity, provider/dependency latency/error, and production-like capacity/cost constraints. Evidence: load manifest. <!-- KEY SKEW IS THE ONE DIMENSION WITH NO VOCABULARY, AND THE WORD THAT SOUNDS LIKE IT MEANS A CLOCK.

      CREDIT, AND EIGHT OF THE NINE DIMENSIONS ARE ALREADY MODELLED.
      PAYLOAD_LIMIT_KINDS covers sizes including the decompression RATIO --
      the dimension a zip bomb exploits and the one a size limit alone misses.
      COMPLEXITY_DIMENSIONS covers pagination and search with nine scored
      axes. PROVIDER_FAULTS is a closed fault set. COST_RESOURCE_CLASSES and
      COST_UNITS make spend expressible per resource.
      
      AND THE PART THAT MAKES "PRODUCTION-LIKE" REAL RATHER THAN
      ASPIRATIONAL: the deployed estate declares its own memory and cpu limits
      in infra/hetzner/docker-compose.yml -- 26 of them, mostly small
      containers -- and contention-constraints-scan.mjs ALREADY READS THAT
      FILE. The numbers are on disk and a tool already goes and gets them, so
      a manifest citing production capacity cites a fact rather than a guess.
      
      THE GAP. Nothing anywhere describes a key distribution: no hot key, no
      cold key, no partition, no shard. So a manifest cannot say that one
      tenant owns most of the traffic, or that one document is read by
      everybody -- the two shapes that turn a comfortable average into an
      outage.
      
      It matters more than it sounds, because every limit in the estate is
      BUCKETED. BUCKET_KINDS lets a limit be per-tenant or per-actor, which is
      exactly the machinery that makes skew survivable, and that machinery
      cannot be exercised by a manifest with no way to say the traffic is
      skewed. A LOAD RUN OVER UNIFORMLY DISTRIBUTED KEYS TESTS THE LIMITS AND
      NEVER TESTS THE BUCKETING.
      
      THE SECOND MISLEADING WORD IN S11.16. S11.16.a found "distribution" in
      capacity-plugins.ts means CONTENT DELIVERY; here every one of the eleven
      `skew` identifiers in the kit is a CLOCK or a VERSION skew. Both are
      recorded rather than left to be rediscovered, because a search returning
      hits for the right word with the wrong meaning is worse than one
      returning nothing: the first tells you to stop looking.
      
      SHIPPED src/load-manifest.ts + scan. The scan CLASSIFIES every skew
      identifier and checks the list is COMPLETE, so one added later must be
      READ rather than assumed. Getting there took three attempts, each
      corrected by its own control: a line-level clock word missed
      `if (skew < -1)`, where the variable IS a clock skew but the line does
      not repeat it; an identifier regex missed SkewMatrix, whose meaning
      lives in the section header "the version-skew matrix"; and the
      completeness check then found clockSkewSeconds, which nobody had
      classified. 18 spec tests; scan PASSES with 9 negative controls. -->
      
    • S11.16.d Ramp sustained, burst, spike, soak, recovery, noisy-neighbor, single-tenant abuse, dependency slowdown, worker loss, and regional capacity scenarios sequentially. Evidence: run results. <!-- A LIMITER RUN PROVES THE DECISION AND NOT THE SURVIVAL.

      This cell asks for run RESULTS, so the first question is what can
      actually be run. The answer is more than nothing and much less than
      everything, and both halves are the finding.
      
      WHAT CAN BE RUN, AND IT IS A REAL RUN. limits.ts exports consume(rules,
      request, counter), a genuine rate decision that charges EVERY bucket a
      request falls into in a fixed order -- tenant, actor, client-class,
      trust-tier, route -- first breach winning. The estate explains the order
      in its own words: "charging only the actor is the rotation bypass -- ten
      identities in one tenant get ten times the allowance -- and charging
      only the tenant lets one actor consume everybody else's", with tenant
      first "because a tenant breach is the one another caller in the same
      tenant can do nothing about".
      
      WHAT THE RUN SHOWED, every number computed by that function:
        * SINGLE-TENANT ABUSE is first refused on the TENANT bucket after ten
          requests -- five actors each under the actor limit, which would have
          passed twenty if only the actor bucket were charged. The rotation
          bypass is closed, and now DEMONSTRATED rather than described.
        * NOISY NEIGHBOUR: the quiet tenant is served in full (its counter
          reaches three, all admitted) while the loud one is refused eight
          times.
      
      AND WHAT THE RUN FOUND THAT THIS REGISTER DID NOT PREDICT.
      RateRule.windowSeconds IS TOLD TO THE CALLER AND NEVER USED TO DECIDE.
      consume reads it on exactly ONE line, to compute retryAfterSeconds; the
      decision itself is `current + cost > permitted`, and
      Counter.current(dimension, bucket, key) takes NO INSTANT -- so nothing
      in the interface lets a counter know which window a request falls in, or
      that a window has passed.
      
      That is a defensible design, since windowing belongs to a store with a
      TTL. What the signature does not do is REQUIRE it, so a counter that
      looks correct and never expires produces a limiter that never forgives,
      and no type would catch it. The consequence is demonstrated: driven
      sequentially, RECOVERY ADMITS NONE OF ITS ARRIVALS, because the count
      from the burst before it only goes up. Recovery is drivable in shape and
      degenerate in result -- the same distinction S11.15.g needed.
      
      THE OTHER FIVE are not drivable for a reason worth naming: spike and
      soak need a CLOCK, worker-loss and regional-capacity need a TOPOLOGY,
      and dependency-slowdown needs a REAL DEPENDENCY, which S11.16.c found
      the estate has no vocabulary for.
      
      AND "SEQUENTIALLY" IS A REAL CONSTRAINT, not a detail. Recovery is
      defined relative to state a previous scenario left behind, so the
      sequence is ordered with every dependant after what it follows, and the
      scan ASSERTS that ordering rather than trusting it.
      
      SHIPPED src/load-scenarios.ts + scan. 19 spec tests, six of them real
      runs; scan PASSES with 9 negative controls. -->
      
    • S11.16.e Measure latency/throughput/error, queue age/depth, saturation, DB/cache/search/blob/provider, retries/dead letters, fairness, cost, SLO, and data correctness. Evidence: dashboards/reports. <!-- EVERY MEASUREMENT IS MODELLED AND THE DEPLOYED ESTATE EMITS NOTHING.

      CREDIT, AND THE MEASUREMENT VOCABULARY IS THE BEST-DEVELOPED IN THE KIT.
      service-objectives.ts has SLI_CLASSES of availability, correctness,
      latency, deadline, queue-age, durability and freshness -- with QUEUE-AGE
      as its own class rather than folded into latency, which is right,
      because a request that waited and one that was slow are different
      failures with different repairs; and CORRECTNESS as its own, which is
      unusual and right, because a system can be fast, available and wrong and
      only that class notices. And it has OBSERVATION_POINTS:
      synthetic-probe, client-commit, edge, server-handler, worker,
      downstream-ack. Naming WHERE a measurement is taken is the thing most
      systems leave implicit, and it is the difference between a latency that
      includes the queue and one that does not.
      
      MEASURED on infra/hetzner/docker-compose.yml: 33 service blocks --
      postgres, redis, minio, the bff, web, admin, tenant-admin, a content
      service, a metis stack and nine psyche services -- and ZERO of seven
      telemetry markers (prometheus, opentelemetry, otel, statsd, metrics,
      grafana, jaeger). No exporter, no collector, no scrape target, no trace
      backend, with a positive control that the same search finds them in a
      probe that declares them.
      
      SO ALL SIX OBSERVATION POINTS ARE EQUALLY UNAVAILABLE. The vocabulary
      for saying where you would measure is complete, and there is nothing at
      any of those places to read. This is not a gap in the kit; it is a gap
      between the kit and the deployment, and the kit is the half that is
      finished.
      
      IT ALSO EXPLAINS THE REST OF S11.16. S11.16.a could not state a traffic
      proportion because nothing observes traffic. S11.16.c found no key-skew
      vocabulary, and a skew is something you notice by measuring. S11.16.d
      could drive only a pure limiter and could not show anything being
      exhausted. All three are the same absence seen from different sides:
      THE ESTATE CAN DECIDE AND CANNOT OBSERVE.
      
      And it is why the parent's evidence clause -- "limits and backpressure
      activate BEFORE resource exhaustion" -- cannot be established today. The
      first half is decidable, as S11.16.d showed by running it. The second
      needs a number from a running process, and no running process emits one.
      
      SHIPPED src/load-measurements.ts + scan. Twelve categories, each mapped
      to what describes it and to an SLI class where one fits -- with
      THROUGHPUT deliberately mapped to NONE, because an objective is about
      what a request gets and throughput is a fact about the system. 19 spec
      tests; scan PASSES with 9 negative controls. The day an exporter is
      added this register goes red and is rewritten, which is the good
      outcome. -->
      
    • S11.16.f Assert rate/concurrency/payload/query/storage/compute/spend/ queue limits, load shedding, circuit, reservation, and backpressure activate before unsafe memory/CPU/storage/provider exhaustion. Evidence: threshold assertions. <!-- "ACTIVATES BEFORE EXHAUSTION" IS A COMPARISON BETWEEN TWO NUMBERS IN DIFFERENT UNITS, AND NOTHING CONVERTS.

      This is the cell the parent's evidence clause is really about, and that
      sentence is a COMPARISON, so everything turns on what is on each side.
      
      THE LEFT SIDE IS REAL AND CHECKED, and the scan CALLS it rather than
      describing it: the rate limit admits exactly its permitted count of four
      requests, and checkPayload accepts a body inside every limit and refuses
      one over the compressed-bytes ceiling. Circuit breakers are modelled
      with states and positions, spend has reserve/settle/available over a
      budget, query cost has scoreComplexity -- most of the twelve mechanisms
      exist and several are callable, and S11.16.d already DROVE the rate
      limit through five scenarios.
      
      THE RIGHT SIDE IS REAL TOO, AND IT IS IN A DIFFERENT FILE AND A
      DIFFERENT UNIT. infra/hetzner/docker-compose.yml declares 26 memory
      limits: genuine numbers, in bytes, for the processes that would actually
      die.
      
      THE BRIDGE IS ONE QUANTITY -- how much of the resource one admitted unit
      of work consumes. The estate even has the UNIT for it: COST_UNITS
      includes memoryByteSeconds, cpuMilliseconds, gpuMilliseconds,
      storageByteSeconds and egressBytes, so "this request cost N
      memory-byte-seconds" is a sentence the kit can already say. What it
      cannot do is fill in N, because filling it in is a measurement and
      S11.16.e found nothing in the deployment emits one.
      
      So this cell's assertion is not merely unmade. It is UNMAKEABLE with
      today's inputs, and PRECISELY ONE MISSING QUANTITY BLOCKS IT -- which is
      a far more useful thing to report than a list of mechanisms.
      
      SHIPPED src/threshold-assertions.ts + scan. FOUR assertion depths rather
      than two -- nothing-to-assert, exists, activates-at-its-threshold,
      activates-before-exhaustion -- because a mechanism that activates
      correctly and cannot be related to an exhaustion point is not a
      mechanism that is missing, and the repairs differ. Measured: 0 of 12
      assertable to exhaustion, 0 of 4 bridges with a value, and TWO
      mechanisms implemented by nothing: load-shedding, and BACKPRESSURE,
      which the parent names by name. A limit REFUSES a request and
      backpressure SLOWS a producer, and nothing here slows anybody.
      
      Two are closest to complete and the register says why: SPEND, where the
      limit and the exhaustion point are already in the same unit -- money
      against a budget -- and PAYLOAD, whose threshold is already in bytes.
      PROVIDER is marked as having no local number at all: not a missing
      measurement but a missing relationship, since a provider's exhaustion
      happens where this estate cannot see.
      
      24 spec tests; scan PASSES with 10 negative controls. (The scan's own
      payload fixture was corrected by its first run: checkPayload takes
      (limits, observation), not the reverse, and the observation names
      longestArray and deepestNesting -- getting it wrong made a payload
      inside every limit come back refused on five kinds.) -->
      
    • S11.16.g Assert no cross-tenant starvation/leak, duplicate/lost effect, priority inversion, false success, or unrecovered state after load. Evidence: post-run integrity report. <!-- PRIORITY INVERSION CANNOT BE VIOLATED HERE, BECAUSE NOTHING SELECTS.

      This is the one cell in S11.16 where most of what is asked can be
      VERIFIED rather than described, because integrity properties are
      decidable by pure functions in a way resource exhaustion is not. Two of
      the six were established by CALLING the estate in this scan, and the
      answers are good ones.
      
      VERIFIED BY RUNNING IT. scopeKey puts tenantId FIRST in the tuple it
      serialises, so two tenants with otherwise identical scopes -- same
      client, same operation, same version, same target -- produce different
      keys. A cross-tenant idempotency collision is therefore structurally
      impossible rather than merely unlikely, and the scan shows it by calling
      the function on three tenants and comparing the three keys, with a
      positive control that the same probe reports allDistinct false when the
      scopes really do collide. decideClaim against a completed record returns
      already-completed rather than deciding it the way it decided the first
      attempt, so a duplicated effect is refused by a function, not by a
      convention. Cross-tenant starvation was DRIVEN in S11.16.d, where the
      quiet tenant was served in full while the loud one was refused by the
      TENANT bucket. And S11.13.d's residueForOutcome reports what survived
      relative to what a run CLAIMS happened, which is what unrecovered
      means -- an import's own successful output is not residue.
      
      THE ONE THAT CANNOT BE VIOLATED. PRIORITY_BAND is real, JobRecord
      carries a priority, and the estate validates it against the band. What
      does not exist anywhere in jobs.ts is a function that chooses which job
      runs NEXT: no dequeue, no scheduler, no ordering, no comparator;
      claimJob claims a job somebody already named. So the property is not
      violated, and it is not upheld either, because a priority inversion is
      a statement about an ORDER and there is no order to be wrong. That is a
      different report from "we checked and it holds", and collapsing the two
      would credit an absence as a guarantee -- which is the failure this
      whole ledger is about. The register carries a value for exactly that,
      unviolatable-because-unimplemented, and a fault kind that fires if
      anybody ever marks such a property complete. The scan re-measures the
      claim from jobs.ts each run rather than trusting this comment, so the
      day something starts selecting, this cell fails and has to be rewritten
      rather than quietly continuing to say an inversion is impossible.
      
      FALSE SUCCESS IS THE PARTIAL ONE, AND THE ESTATE IS GOOD AT IT.
      saved-shown-before-the-commit catches a surface that says saved before
      the commit landed, and partial-import-reported-as-a-failure catches the
      opposite -- an import that really landed reported as failed. Both are
      real checks and both read a DESCRIPTION of a run. What neither can do
      is notice a false success that nothing described, which after a load run
      is the likely kind, because the thing that would describe it is the
      telemetry S11.16.e found the deployment does not emit.
      
      Measured: 6 properties, 3 completely established (2 of those by calling
      the estate), 1 unviolatable because nothing implements what would break
      it, 3 findings all of the same kind -- property-not-completely-
      established. 22 spec tests; scan PASSES with 9 negative controls,
      including one that every fault kind is reachable from some input, after
      S11.16.b shipped a kind that could never fire. -->
      
  • S11.17 Add accessibility parity tests for canvas/table, graph/list, timeline/list, waveform/transcript, and 3D/metadata views. Evidence: semantic-equivalence assertions. <!-- a-g all closed. THE SECTION'S SHAPE: PARITY IS A TWO-SIDED CLAIM, AND THE ESTATE MAKES IT WELL FOR ONE PAIR AND ONE-SIDEDLY FOR THE REST.

    .a measured it: of fifteen facets, five support an equivalence assertion
    and two of those only for graphs, because the one complete parity audit in
    the repository takes nodeColumns and edgeColumns. Seven support presence
    and nothing stronger, and one -- range -- is described by neither
    rendering.
    
    EVERY CELL FOUND THE SAME KIND OF THING AND A DIFFERENT INSTANCE OF IT: a
    fact that exists, is correct, and does not reach the reader who cannot see
    the picture. .b: the estate's spatial doctrine is right about a computed
    layout and inverted on an authored canvas, and compareGraphs hands a
    non-visual reader "1 difference (0 to the work, 1 to its presentation)"
    for a box somebody dragged across the page. .c: pathListed and
    impactListed are BOOLEANS beside twenty-three fact-by-fact comparisons,
    and the group tree is declared, cycle-checked and flattened to one label
    on BOTH sides. .d: TimelineDocument.markers is readonly number[] while
    three keyboard commands move a reader between them -- navigable and
    unreadable -- and every "readonly value" in 3,236 lines is a boolean. .e:
    integratedLoudness computes seventeen block values and returns one number,
    so loud-then-quiet and quiet-then-loud measure 0.0002 LU apart. .f: a
    local 4x4 is correct, complete and unreadable, and variantSelections holds
    the choice and not the options.
    
    THE VOCABULARY THAT KEPT IT FAIR. Each cell needed a third or fourth
    state, and collapsing any of them would have credited an absence as a
    guarantee or called a careful thing broken: above-both-renderings (a value
    decided before either view, where a parity test would be theatre),
    as-prose-inside-a-string and outside-the-signature (present-but-unassertable,
    and a missing argument rather than a missing field), asserted-as-a-boolean
    (a list was rendered, not that it agrees), computed-and-discarded (correct
    arithmetic, missing return value), authored-and-unreadable, and
    unobserved-because-no-reader-runs-here.
    
    .g CLOSED IT BY MUTATION: 33 visual-only facts and actions removed or
    altered across all five pairs, 33 killed, over a harness that REFUSES to
    report when the honest baseline is already noisy. And it says plainly that
    the screen-reader half was not performed -- measured on the host, not read
    from a register -- rather than marking a journey nobody listened to.
    
    Seven cells, seven scans, all passing with negative controls; the section
    closes with the kit at 297 entry points and 0 violations. -->
    
    • S11.17.a Define a shared parity contract for entity IDs, values/units, relationships/order/hierarchy/time, status, selection/focus/filter/ range, operations, errors, and protected/inaccessible state. Evidence: parity schema. <!-- PARITY IS A TWO-SIDED CLAIM, AND ALMOST EVERY PIECE THE ESTATE HAS DESCRIBES ONE SIDE.

      An equivalence assertion needs both sides in one place, so a test can
      compare them and fail when they diverge. A record that describes only
      the alternative supports the weaker claim that a row EXISTS, and no
      amount of care makes it support the stronger one, because there is
      nothing to compare the row against.
      
      CREDIT FIRST, BECAUSE THREE PIECES ARE GENUINELY TWO-SIDED AND ONE IS
      EXCELLENT. graph-views.ts has a complete parity audit from S7.10.h:
      every visual encoding declares the fact it carries and the
      semanticColumn stating the same fact, every action declares its
      canvasGesture and the semanticControl that performs it, and parityAudit
      walks the encodings ACTUALLY IN USE rather than the ones somebody
      remembered to declare -- reporting graphOnlyFacts and graphOnlyActions
      with emptiness as the acceptance. That is exactly what this section
      asks for. It answers for ONE of the five pairs, and the reason is in the
      signature: two of its four required inputs are nodeColumns and
      edgeColumns, and EncodingDeclaration.encoding is a closed union of
      twelve values named for node-link diagrams, three of them declared edge
      encodings. A waveform has no edges. That is not a criticism -- it was
      written for graphs and it is right about graphs -- it is why the cell
      asks for a SHARED contract rather than pointing at the one that exists.
      
      domain-views.ts adds two more. ViewCommand carries pointerGesture and
      formAlternative in the same record, so "this can only be done by
      dragging" is decidable. And reconcileSelection compares a selection held
      in one view against the ids ANOTHER view can address, splitting them
      into carried, withheld and dropped with not-representable-in-this-view
      as a first-class reason. It is the one piece of parity machinery in the
      repository that is not shaped like the view it was written for: it takes
      ids and reasons. The scan calls it with timeline, waveform and
      scene-graph input and it answers unchanged, three times.
      
      THE ALTERNATIVE IS REQUIRED, AND IT CARRIES CONTENT RATHER THAN FACTS.
      plugins.ts makes a non-visual rendering mandatory and the comment is
      right that a projection whose only rendering is visual has no export, no
      API and no screen-reader path. But a SemanticNode is four fields over
      eleven document kinds. Order and hierarchy survive that projection
      honestly, because arrays are ordered and children nest. An entity's
      identity, a number's unit, a status and an instant all arrive as words
      inside text -- which is not absence but a third state: present, read
      aloud, and impossible to assert on, compare, export or filter. The
      estate has the most careful time model it could have, in flicks and
      rational rates and drop-frame timecode, and a SemanticNode has no time
      field, so the instant reaches the reader as characters.
      
      AND FOUR FACETS ARE OUTSIDE THE SIGNATURE. toSemanticNodes is a pure
      function of the model; selection, focus, filter and range are states of
      the VIEW. Recording those as "nothing carries it" would invite the wrong
      repair -- a field in the node tree -- when the repair is a second
      argument. TWO NEED NO PARITY AT ALL: ProjectionState is loading / ready /
      empty / unauthorized / not-configured / failed and is decided BEFORE
      either rendering runs, so there are not two descriptions to keep in step,
      and a contract demanding a test there would be asking somebody to prove
      that one value equals itself. Four sidedness values, four claims: a
      boolean would have merged the equivalence you must test with the one
      that cannot break.
      
      WHAT THIS CELL ADDS BEYOND THE REGISTER, AND THE LIVE DEFECT IT FOUND.
      SemanticNode states three rules about itself in its own comments -- text
      is never markup, text is never a translation key, a text alternative is
      required on a figure and refused elsewhere -- and NOTHING IN THE
      REPOSITORY CHECKED ANY OF THEM. semanticNodeFaults is the first checker,
      and the scan calls all four shipped implementations of
      NonvisualAlternative rather than reading them: TWO OF THE FOUR LEAK A
      TRANSLATION KEY. incident-console.archetype.ts emits
      `${entry.at} ${entry.labelKey}` as the text of a list item, so a reader
      on the alternative hears "2026-08-16T00:00:00.000Z
      timeline.incident.mitigationIneffective" where the visual timeline shows
      a local time and a translated sentence; story-desk's inspector emits
      model.headlineKey as a term. Present, required, shipped, and not
      equivalent -- which is this cell's whole point in one line.
      
      The compounds are SPLIT: the cell writes two items where there are
      eight, and they do not share an answer -- order survives a document
      projection and time does not, selection is reconciled by a real function
      and range is described nowhere. Measured: 15 facets, 5 comparable, 2
      equal by construction, 7 presence-only, 1 (range) described by neither
      side; 9 carried across all five pairs, 2 comparable for graphs alone.
      31 spec tests; scan PASSES with 12 negative controls, including one that
      the key checker stays SILENT on ordinary prose. (Its first run caught my
      own field regex counting `readonly SemanticNode[]` as a fifth field of
      SemanticNode.) -->
      
    • S11.17.b Assert canvas objects/spatial facts/selection/edit operations are present and synchronized in a navigable table/form alternative. Evidence: canvas/table fixtures. <!-- A GRAPH'S POSITION IS COMPUTED AND MEANS NOTHING. A CANVAS'S POSITION IS AUTHORED AND IS THE CONTENT. THE ESTATE HAS ONE DOCTRINE AND IT IS THE GRAPH'S.

      The kit states it three times and is right every time about the surface
      it was written for. spatialMeaningFaults reports
      position-mapped-to-a-coordinate-column when a table answers a positional
      encoding with an x column -- "it satisfies the parity check while stating
      nothing a reader can use". ComparableGraphNode.position is commented
      "Where the layout put it. Never compared". compareGraphs returns
      layoutOnly when the only thing that moved was the drawing. All three
      rest on S7.10's finding that a coordinate states nothing, because a
      layout run put it there.
      
      Every one becomes false the moment a PERSON rather than a simulation
      decides where the object goes. On an authored board, moving a box is the
      edit, overlapping two shapes is the meaning, and a table that omits the
      coordinates has dropped the content.
      
      THE SCAN SHOWS BOTH BY RUNNING THEM, each beside its graph-shaped twin,
      so the same function is seen being right and being wrong on the same
      day. The same spatialMeaningFaults call is correct for a layout and
      wrong for a board. And compareGraphs classifies a box dragged from
      (60,40) to (700,500) as significance=presentation, handing a non-visual
      reader the sentence "1 difference (0 to the work, 1 to its
      presentation)." On an authored canvas that sentence is false, and it is
      the sentence the reader who cannot see the board is given.
      
      THE REPAIR IS ONE BIT OF PROVENANCE, NOT A NEW ALGORITHM. POSITION_ORIGINS
      says where the coordinates came from, and that single fact decides which
      of the estate's existing rules applies. It sits on the OBJECT rather
      than the surface, because a real board mixes them: a person places the
      boxes and a router lays out the connectors between them, and the
      connector's coordinates state nothing while the box's are the whole
      point. The fixture is deliberately mixed for that reason, and one table
      carries both.
      
      PRESENT AND SYNCHRONIZED ARE TWO QUESTIONS, AND THE SECOND SORTS ITSELF
      BY WHERE THE STATE LIVES. An object's existence, its position, its parts
      and an applied edit are in the MODEL both renderings project, so they
      cannot disagree -- synchrony is free, and a test for it would be asking
      whether one value equals itself. A SELECTION is held by each RENDERING
      and nothing in the estate joins them, so it is the one subject that can
      drift silently, and the only one a test earns its keep on.
      reconcileSelection is exactly the missing joint and it already exists:
      the canvas selection is compared against the ROWS the table can address,
      not against the table's own selection, which would have asked a
      different question and returned an answer that looked identical.
      
      THE ANCHOR NAMES WHAT AND NEVER WHERE. StableAnchor has a canvas kind
      with objectId and an optional part, and resolveCanvas is careful with
      it: a part that dies while its object survives is deleted rather than
      coarsened up to the object, because rebasing "this handle" to "this
      shape" is a silent move dressed as a rescue. The anchor has no field for
      a place, so a comment pinned to a canvas object names the shape and can
      never name the region -- verified by parsing one and listing its fields.
      
      The table carries the numbers AND a sentence: two coordinates are the
      fact, "top left" is the fact made usable, and a table with one of them
      fails differently from a table with neither. The description is a named
      approximation -- thirds on each axis, by the object's centre -- written
      down with the numbers travelling beside it, and there is a check that
      the sentence agrees with its own coordinates.
      
      Measured: 11 parity faults over 4 subjects, 1 subject whose synchrony
      can fail. 27 spec tests; scan PASSES with 14 negative controls. (Its own
      NUL check caught 2 NUL bytes in the module I had just written -- the
      fifth occurrence of that trap this initiative -- and the RevisionRef it
      builds was refused until it carried all seven declared fields.) -->
      
    • S11.17.c Assert graph nodes/edges/properties/groups/paths/impact and relationship operations are present in list/table/tree alternatives. Evidence: graph/list fixtures. <!-- A BOOLEAN SAYS A LIST EXISTS. IT DOES NOT SAY THE LIST AGREES WITH THE PICTURE.

      THIS IS THE ONE PAIR OF THE FIVE THE ESTATE ALREADY ANSWERS, AND IT
      ANSWERS MOST OF IT VERY WELL. Saying that plainly is half this cell's
      job. nodeParityFaults compares a drawn node with its row fact by fact --
      label, type, status, group, every annotation, both degrees, the details
      target -- over twelve fault kinds, and refuses a degree that counts a
      node the view does not hold. edgeParityFaults does the same over eleven,
      including a direction the arrowhead asserts and the list contradicts,
      and an endpoint named by id where the line shows a label.
      relationshipCommandFaults covers ten. traversalFaults asks whether a
      node can be reached by keyboard at all. synchronizationFaults keeps five
      surfaces in step. The scan DRIVES the first two rather than citing them:
      silent on an honest pair, and loud when the table says blocked where the
      picture says ready, or the list contradicts an arrowhead.
      
      TWO OF THE SEVEN SUBJECTS ARE ANSWERED BY ONE BIT EACH.
      ViewportObservation has pathListed and impactListed, and viewportFaults
      reports path-shown-only-as-a-line and impact-shown-only-as-a-shade when
      either is false. Both checks are right and both are booleans: they say a
      list was RENDERED, not that the list is the path. The estate's own
      detail string names the comparison it is not making -- "the hops, their
      order and their edge types are the answer; the line is the picture of
      it". The scan puts ONE path, listed in the wrong order, to both: 0
      findings from viewportFaults with pathListed true, 4 from
      pathParityFaults, same path, same run. A path through the wrong edges
      between the right nodes passes the boolean too, and so does an impact
      set that omits the one thing that breaks.
      
      AND THE GROUP TREE IS DECLARED, CHECKED FOR CYCLES, AND FLATTENED TO ONE
      LABEL ON BOTH SIDES OF THE GLASS. GraphViewContract.groupParents is a
      genuine tree with a group-cycle fault refusing a loop. Both renderings
      then carry groupLabel: string | null, one flat string. So the picture
      and the table AGREE about the group, neither states the hierarchy, and
      group-differs-between-the-picture-and-the-table passes on two renderings
      that are equally silent -- which is exactly why the existing check
      cannot find it. Run against a three-deep tree stated as flat labels, the
      new comparison reports both nested groups. The cell asks for
      "list/table/tree alternatives" and GRAPH_SURFACES is graph, node-table,
      edge-table, inspector, deep-link. There is no tree, for the one subject
      whose shape is a tree.
      
      WHAT THIS CELL ADDS is the three comparisons the estate's own comments
      describe -- the hops and their edge types, the impact set and whether it
      is offered before the change, and the ancestry each rendering states --
      and it replaces none of the estate's checks. There is a register fault,
      deep-coverage-that-also-adds-something-here, that fires if it ever
      duplicates one, because two checkers over the same graph is how they end
      up disagreeing.
      
      THE REGISTER NAMES ITS ESTATE FAULT KINDS RATHER THAN COUNTING THEM. Its
      first scan run caught the reason: group claimed 1 and the derivation
      said 12, because a count was a number about the whole list and the claim
      was about the subject. Now every attributed kind is checked for
      membership in the estate list it came from, no kind may be claimed
      twice, and the attributions must PARTITION the node and edge lists
      exactly -- so a fault kind cannot be quietly dropped from the register.
      The same run also refused the scan's own fixture with
      unknown-neighbourhood-reported-as-zero, correctly: the nodes had no
      neighbourhoodLoaded and the rows stated a degree anyway.
      
      Measured: 7 subjects, 4 compared fact by fact by the estate, 1 as one
      label, 2 as a boolean, 3 deepened here; 12 node and 11 edge kinds
      partitioned, 10 relationship-command kinds, 2 of 10 viewport kinds.
      35 spec tests; scan PASSES with 12 negative controls, four of which
      check that the new comparisons stay SILENT on renderings that match. -->
      
    • S11.17.d Assert timeline tracks/events/start/end/duration/value/status/ overlap and trim/move/split/zoom/seek operations are present in semantic lists/forms. Evidence: timeline/list fixtures. <!-- THE TIMELINE MODELS SPANS WITH LABELS. TWO OF ITS OWN SIX TRACK KINDS ARE NEITHER A SPAN NOR A LABEL.

      CREDIT FIRST, AND IT IS OWED IN QUANTITY. timelines.ts is 3,236 lines
      and timelineParityFaults is the best parity check in the repository. It
      compares a drawn clip with its row over thirteen fault kinds, by real
      arithmetic rather than string equality: the drawn x against
      timeToPixels(start) within one device pixel, the drawn width against the
      stated duration, the end timecode against start plus duration in the
      contract's own frame rate. It knows the asymmetry runs one way -- an
      item scrolled off the canvas is legitimately undrawn, and an item INSIDE
      the visible window that the table lists and the canvas omits is
      graph-only information in the other direction. It has
      row-with-no-operations-and-no-reason, which is the disabled-field rule
      one level up: a genuinely read-only track offering nothing is fine, and
      offering nothing without saying why renders a protected item identically
      to a broken control. The scan DRIVES it: silent on an honest clip, and
      reporting drawn-position-disagrees-with-the-stated-time when the
      rectangle moves 40px from its timecode. Every fact this cell names for a
      CLIP -- track, start, end, duration, label, status, overlap, and the
      trim/move/split operations -- is already compared by something that
      runs. Zoom is bounded by clampZoom and pointerPrecision; seek is a
      playhead key in TIMELINE_LINK_KEYS.
      
      AND THEN THERE IS TRACK_KINDS, WHICH HAS SIX MEMBERS.
      
      A MARKER IS AN INSTANT WITH A MEANING, HELD AS A BARE NUMBER.
      TimelineDocument.markers is readonly number[]. Not items: numbers. So a
      marker has no id, no label, no track, no row and no operation, and the
      clip comparison -- which walks drawn clips against rows -- cannot see
      one. Meanwhile TIMELINE_COMMANDS has add-marker, next-marker and
      previous-marker with real key bindings, and SNAP_TARGETS has marker. A
      reader can add a marker, jump to the next one, and snap a clip edge to
      it, and can never be told which markers exist or what any of them means.
      Navigable and unreadable. Two markers at one instant are two equal
      numbers, so nothing can name one of them, move one of them, or say which
      one a comment is about.
      
      AN AUTOMATION POINT IS A VALUE AT AN INSTANT, AND THERE IS NO VALUE.
      automation is a declared track kind with its own native unit -- samples,
      because a volume ramp is authored against audio and not against frames,
      which the module is right about. TimelineItem carries start, duration,
      sourceIn, speed, three flags, a link group and a lane. Every "readonly
      value" in the file is the boolean on set-enabled, set-muted and
      set-locked; the scan counts them (3) and asserts none is numeric.
      TIMELINE_OPERATIONS has thirteen members and none changes a number. So
      the one thing an automation track exists to carry has no field, no
      column and no operation.
      
      WHY THIS IS NOT A COMPLAINT ABOUT A MISSING FEATURE. The cell asks for
      "value" among the facts a list must state. A span-shaped row can hold a
      marker only by inventing a duration, and an automation point only by
      dropping the number -- and BOTH would pass a check comparing what the
      picture draws with what the table lists, because neither rendering has
      the fact. That is the same shape S11.17.c found in the group tree: the
      two agree, and both are silent.
      
      Measured: 8 facts, 1 (value) with no field anywhere, 2 (track, event)
      answered for spans only, 3 shapes across 6 track kinds; 12 new fault
      kinds over markers and automation. 23 spec tests; scan PASSES with 12
      negative controls, five of which check the new comparisons stay SILENT
      on a marker or a point stated in full. (Its first run was refused by the
      estate's own comparison for a contract missing originFrames, which made
      the estate's expected timecode NaN -- the check working on the scan's
      fixture. A second run caught a value regex reading past a field, because
      the operation-request union members end in a brace, not a semicolon.) -->
      
    • S11.17.e Assert waveform/audio duration/channels/time/range/markers/ loudness/events/alerts and navigation are present in transcript/data/ event views. Evidence: waveform/transcript fixtures. <!-- A RANGE IS TWO INSTANTS, AND EVERY TIME ANCHOR IN THE ESTATE HOLDS ONE.

      StableAnchor has eight kinds and two are about position. temporal is
      {timeBase, position} -- ONE instant, in seconds or frames, with the time
      base carried beside the number so a frame position cannot be read
      against the wrong rate. range is {blockId, start, end} -- two offsets
      INSIDE A BLOCK, which is a span of text and not a span of time. So "from
      1:02 to 1:09", the selection a waveform reader makes constantly and the
      unit a transcript view exists to align with, cannot be named by any
      anchor in the repository: not commented on, not carried across a
      revision, not handed to another surface. S11.17.a recorded range as the
      one facet neither rendering describes; this is the structural reason,
      read from both types in the scan.
      
      AND THE LOUDNESS SERIES IS COMPUTED AND DISCARDED.
      
      integratedLoudness is real BS.1770-4: K-weighting through two biquads,
      400ms blocks at 75% overlap, an absolute gate at -70 LUFS and a relative
      gate at -10 LU, the channel weights and the -0.691 dB offset. It is
      careful enough to return -Infinity for digital silence, "which is the
      true value", and to flag measuredAtReferenceRate when the coefficients
      were applied at a rate they were not derived for. It computes a block
      loudness for every 400ms window, and returns blockCount, gatedBlockCount
      and one number.
      
      THE SCAN RUNS IT TWICE. Two seconds of a 1 kHz tone, loud then quiet:
      -9.7287 LUFS. The same two seconds quiet then loud: -9.7289. 0.0002 LU
      apart -- not a distinction anybody can act on. The measurement is right
      both times and it is the same measurement, because integrated loudness
      is a property of the programme rather than of any moment in it. A
      waveform shows a reader WHERE the loud part is. The seventeen blocks
      that would say so were computed inside the function and are a local
      variable. True peak is the same shape: both arrangements measure -6.02
      dBTP, so a clip at the start and a clip at the end are one reading.
      
      computed-and-discarded is therefore a state of its own in this register,
      distinct from not-modelled. The arithmetic exists and is correct; what
      is missing is a return value. That is a different repair and a much
      smaller one, and calling it "no loudness data" would hide both facts.
      momentaryLoudnessSeries supplies it, built from the estate's own
      exported kWeight, channel weights, block length and offset rather than a
      second filter, so the two cannot drift in their arithmetic. It is
      BS.1770 MOMENTARY loudness, ungated, named as such: gating is how you
      get one number for a programme, and applying it per block would discard
      exactly the quiet stretches a reader wants located. Run over the same
      fixtures it puts the loud part at 0.1s and at 1.1s respectively.
      
      CREDIT, BECAUSE THE ESTATE ALREADY NAMES THIS FAILURE. Of its eight
      media comparison faults, one is audition-with-no-text-alternative --
      "listening to two takes is the whole comparison for a hearing reader and
      none of it for" anybody else -- and another is
      channel-counts-differ-with-no-downmix-policy. Channels are declared by
      layout and weighted by name; an LFE at weight zero really is ignored,
      which the spec checks. The audio model is one of the strongest things in
      the kit. What it has no way to say is WHEN.
      
      Measured: 9 facts -- 2 computed and discarded (loudness, event), 4 not
      modelled (range, marker, alert, navigation), 1 derivable and unnamed
      (duration, a division away from fields that exist), 2 computed and
      returned. 25 spec tests; scan PASSES with 12 negative controls,
      including one that the loudness check stays SILENT when a row states the
      truth, and one that a recording shorter than a block yields no blocks
      rather than a fabricated one. -->
      
    • S11.17.f Assert 3D hierarchy/objects/transforms/geometry/materials/ animation/variants/refs/validation and selection/navigation are present in metadata/tree/table views. Evidence: 3D/metadata fixtures. <!-- A MATRIX IS SIXTEEN NUMBERS. A READER NEEDS A PLACE.

      CREDIT FIRST, AND THE MODEL IS A REAL USD STAGE. StagePrim carries a
      path, a type name, a local transform, mesh topology, a material binding,
      variant selections, references, the composition arc that supplied the
      strongest opinion, a skeleton joint count and free attributes. Stage
      carries the up axis, metres per unit and time codes per second.
      COMPOSITION_ARCS is the real strength order. And
      STAGE_COMPARISON_FAULTS is seven kinds of hard-won care: transforms
      across a differing up axis are refused rather than attempted, geometry
      across a differing unit scale likewise, a variant switch is not reported
      as a geometry change, a reference retarget is not reported as a local
      edit, and identical topology with reversed winding is caught -- the bug
      that renders as a black object and compares as no change at all.
      
      Transform.matrix is commented "Row-major 4x4. Local to the parent, never
      world", and world-transform-change-attributed-to-the-prim exists because
      the module knows exactly what that costs. It is the RIGHT choice: a
      local matrix is the authored fact, and a world matrix is a derived one
      that changes when an ancestor moves.
      
      AND IT MEANS A TABLE ROW CANNOT SAY WHERE ANYTHING IS. Sixteen numbers,
      relative to a parent the row does not name, is a complete description
      that no reader can act on -- and the scan prints one to make the point.
      The world position is a product of the chain: derivable from what the
      stage already holds, stated nowhere, and therefore recomputed by every
      surface that wants it, in whichever convention that surface guessed.
      worldTransform composes it, with the convention STATED once and PROVED
      rather than assumed -- a three-deep fixture (world at x=10, set at x=1,
      chair at y=2) whose answer a person can predict: [11, 2, 0]. It refuses
      rather than substituting an identity for a missing ancestor, because
      that would place an object where nobody authored it. readTransform gives
      a place and a size, takes scale as the length of each basis row, and
      REFUSES to name a rotation: from a general matrix it is ambiguous
      without an agreed order, and an angle under the wrong convention is
      worse than no angle. It reports orthogonal=false when shear is present,
      so a caller never has to guess whether the numbers mean what they
      usually mean.
      
      A SELECTION IS NOT A CHOICE UNTIL YOU KNOW WHAT ELSE WAS AVAILABLE.
      variantSelections is Record<string, string> -- the set name and the
      selected variant -- and nothing anywhere holds the OPTIONS. The scan
      reads StagePrim's ten fields and asserts none of them is one. So a
      metadata view can say "lod = high" and can never say "lod is one of
      high, medium, low": a reader on the table cannot tell a deliberate
      choice from the only possibility, and cannot discover that a variant
      exists to be switched. In the viewport that information is a dropdown.
      
      AND ANIMATION IS A RATE AND A JOINT COUNT. timeCodesPerSecond says how
      fast time runs and skeletonJointCount says how many joints there are;
      attributes are static strings and no field anywhere is a time sample.
      Nothing distinguishes a rigged character standing still from one that
      moves, so a metadata view has nothing to list under animation but a
      number of joints.
      
      Measured: 11 subjects -- 7 stated and readable, 1 authored and
      unreadable (transform), 2 derivable and unstated (variant, validation --
      seven fault kinds judge two stages against each other and none judges
      one on its own), 1 not held at all (animation). 28 spec tests; scan
      PASSES with 11 negative controls, three of which check the new
      comparisons stay SILENT on a view that states everything. A register of
      path faults with no producer was written and then deleted rather than
      shipped. -->
      
    • S11.17.g Mutate each visual-only fact/action and require parity test failure; test keyboard/screen-reader full journeys on alternatives. Evidence: semantic mutation report. <!-- A CHECK THAT HAS NEVER FAILED IS A CHECK NOBODY HAS TESTED.

      S11.17.b through .f each built a comparison and asserted it stays SILENT
      on an honest pair. Silence is the cheap half. This cell takes each
      visual-only fact and each visual-only action across all five view pairs,
      removes or alters it in the alternative, and requires the comparison to
      report. 33 mutations, 33 killed, printed one line each with the fault
      that caught it -- and a mutant that SURVIVES is a hole named rather than
      counted.
      
      THE GUARD THAT MAKES A KILL MEAN ANYTHING. Before any mutation runs, the
      unmutated alternative is checked and must produce nothing. Without that
      a baseline that already reports would kill every mutant for free, and
      the report would read as complete coverage of a suite that never looked
      at the mutation at all. runMutations THROWS NoisyBaselineError rather
      than returning outcomes somebody would read, and names the noisy
      findings so the fixture can be repaired. Both directions are
      positive-controlled: a deliberately blind checker produces a survivor
      with caughtBy null, and a noisy baseline is refused outright. killRate
      is exercised at 0.5 as well as 1, so a hundred percent is a measurement
      rather than the only value the function returns.
      
      Three mutation kinds are kept apart because they fail differently: a
      dropped fact is a missing row or column, an altered one is a row that
      disagrees, and a suite can catch one and not the other. All three are
      exercised, and every one of the five pairs is covered.
      
      THE KEYBOARD HALF IS DECIDABLE HERE AND IS CHECKED. An alternative that
      states a fact nobody can move focus to has not made it available; an
      operation with no keyed route is a pointer gesture wearing a button; and
      an action that does not declare where focus lands drops a keyboard
      reader at the top of the document.
      
      THE SCREEN-READER HALF WAS NOT PERFORMED, AND THAT IS THE HONEST REPORT.
      at-reconciliation.ts declares six readers and every one of them
      runnableOnThisHost: false, with reasons that are facts about platforms
      rather than excuses -- NVDA, JAWS and Narrator are Windows, VoiceOver is
      Apple and driven through its own scripting, TalkBack needs a device, and
      Orca needs a desktop session and a speech service a headless host does
      not have. The scan does not trust that register: it MEASURES the host,
      looking for five reader binaries on PATH, and fails if one is present
      while the register says none runs. Result: 0 of 5 found, 0 of 6 runnable,
      AT_MATRIX 0 of 36 cells observed, and this cell changes none of them.
      
      Recording that as unobserved rather than as a pass is the whole reason
      that module exists. Marking this cell without saying so would be the
      exact defect this section is about: a claim that a reader was served,
      made by somebody who never listened.
      
      18 spec tests; scan PASSES with 8 negative controls.
      
  • S11.18 Add an adversarial stub scan tailored to provider, score, fixture, and success-state fabrication. Evidence: reviewed zero-blocker report.

    • S11.18.a Define executable scan sources for production code/config, adapters, routes, UI, generated code, feature flags, fixtures imports, tests excluded by rule, and built bundles. Evidence: scan scope. <!-- A SCAN'S SCOPE IS THE SET OF FILES IT CAN OPEN. THIS ONE OPENS THIRTEEN SUFFIXES, AND THE LANGUAGE THE TOOLING IS WRITTEN IN IS NOT AMONG THEM.

      CREDIT FIRST, BECAUSE THIS SCAN IS REAL AND RUNS.
      scripts/stub-indicator-scan.sh is wired into the pre-commit hook, a CI
      job over the whole PR diff, a separate silent-random job, a nightly
      trend run and a per-PR comment. It has three modes, an inline allow-list
      that carries a stated reason (// stub:legitimate <why>), and it excludes
      tests BY RULE rather than by accident, with the argument written down:
      fixtures and assertions are expected to contain stub vocabulary. It even
      handles pure renames -- an R100 move carries no new markers, and without
      that carve-out a directory move re-litigates every pre-existing hit,
      which it did once and is recorded in the script. It caught this session
      twice, on the word "would" in two doc comments.
      
      THE SCOPE IS AN EXTENSION LIST, AND THAT IS THE FINDING. The filter is a
      case over suffixes: nine are languages the product is written in and
      four are V3 asset JSON. Everything else in the repository is invisible.
      Measured over the whole tracked tree -- 136,871 files -- 66,488 are in
      scope, and of the rest 5,172 carry code somebody runs: 3,638 .mjs, 526
      .js, 269 .cs, 223 .tf, 202 .sql, 197 .sh, 71 .java, 38 .cjs, 8 .go.
      
      .mjs is what every verification tool in this initiative is written in,
      every generator under scripts/, and every scanner under tools/ --
      INCLUDING THE SCANNER THAT PRINTS THAT NUMBER, which the run checks by
      name and reports out of scope. A fabricated result inside the machinery
      that produces evidence would not be seen by the scan that exists to find
      fabricated results. Configuration is unreachable for the same reason:
      only four dotted V3 suffixes reach the scan out of everything ending in
      .json, so a value fabricated in configuration is read by the product and
      by nothing else.
      
      AND THREE CLASSES ARE EXCLUDED ON PURPOSE. Generated code, built bundles
      and anything under dist/, build/, .next/ and target/ are dropped by path
      before the suffix filter. That is defensible for a pre-commit hook over
      files nobody commits, and it is not defensible for the question this
      cell asks, because the bundle is the only artifact that actually runs in
      front of a person: a stub introduced by a transform, a define or a
      bundler replacement exists in exactly one file, and it is one this scan
      never opens.
      
      excluded-by-a-stated-rule is kept distinct from
      unreachable-no-suffix-matches because an exclusion somebody wrote down
      is a decision that can be argued with, and a class the filter never
      matches is not a decision at all -- nobody has had the argument. And
      reached-only-through-another-suffix is the third: an adapter, a route
      and a feature-flag registry are covered because they happen to be
      TypeScript, so a file moved to another format takes the coverage with
      it.
      
      THE REGISTER IS CHECKED AGAINST THE SCRIPT, NOT AGAINST ITSELF. The
      suffix list and the exclusions are PARSED out of the shell source, so
      the day somebody adds .mjs this register is wrong out loud; the parser
      returns null rather than a guess when the filter changes shape, and the
      scan fails on that. A class recorded as reached must arrive as a suffix
      the script accepts, and a class recorded as excluded must arrive as a
      pattern the script really excludes. (The first version of that second
      check grepped this register's own prose for words like "deliberate" and
      failed on a perfectly good reason -- a check that reads meaning out of
      wording, which is the thing this initiative keeps finding in other
      people's registers. It is structural now.)
      
      Measured: 11 source classes, 5 reached, 3 excluded by a stated rule, 3
      unreachable. 21 spec tests; scan PASSES with 12 negative controls. -->
      
    • S11.18.b Detect fake/mock/demo/stub/in-memory/fallback provider binding, deterministic canned outputs, catch-and-success, and unconfigured-to- success paths. Evidence: provider rules/fixtures. <!-- THE SCAN LOOKS FOR THE WORD. THE DEFECT IS A SHAPE.

      scripts/stub-indicator-scan.sh is a regular expression over lines, and
      it is a good one: "in production, this", "would use", "not yet
      implemented", "placeholder", "hardcoded", "fake", "TODO:". Every one of
      those is somebody ADMITTING what they did, in English, in a comment. Its
      real power is that developers write those admissions honestly, and it
      works -- it caught this session twice.
      
      The four fabrications this cell names contain none of those words:
      
          catch { return { ok: true }; }
          if (!apiKey) return { configured: true };
          export const provider = new InMemoryProvider();
          export const score = (input) => 0.92;
      
      Four lines that fabricate a result, and a vocabulary scan reads them as
      ordinary code. THE RUN PUTS ONE FIXTURE THROUGH BOTH INSTRUMENTS, with
      the pattern PARSED out of the shell script rather than copied so the
      comparison is against what really runs: ZERO hits against FOUR findings,
      same file, same run.
      
      WHY THIS IS NOT A CRITICISM OF THAT SCAN. A grep for admissions and a
      search for shapes are different instruments and the estate should have
      both. One catches the developer who wrote down what they were doing; the
      other catches code that says nothing about itself. Neither subsumes the
      other, and the second is the half that does not exist.
      
      THE APPROXIMATION IS STATED RATHER THAN DISCOVERED LATER. No TypeScript
      parser: a lexical pre-pass blanks comments and string bodies while
      PRESERVING every offset and newline -- so a reported line is a real line
      a reader can open -- and then the detectors match on brace-balanced
      regions. The limits are written into the module, every finding carries
      its line and the text that fired, and the spec exercises the pre-pass
      against comments, strings, template literals and escaped quotes.
      
      EVERY DETECTOR IS SILENT ON THE HONEST SHAPE, which is what makes a
      finding worth reading. A catch that returns a typed failure, a default
      or an empty list produces nothing; so does one that rethrows. A
      not_configured throw, a fail-closed 503 and a configured:false produce
      nothing -- the estate's own honest seams. A zero-argument constant is a
      constant. A function that uses its argument is a function. The PARAMETER
      is the evidence for a canned output: a computation that cannot read its
      input did not happen.
      
      An inline opt-out exists (fabrication:legitimate <reason>) and REQUIRES
      a reason; a bare marker still reports. It is kept separate from the
      estate's stub:legitimate on purpose, because putting that marker on a
      line the vocabulary scan never flags would send a reader looking for a
      word that is not there.
      
      SWEPT OVER THE KIT'S OWN 302 PRODUCTION SOURCES: one finding, at
      negative-corpus.ts:446 -- `catch { return true; // not JSON at all; the
      whitespace question does not arise }`. Read by hand: the shape is real
      and the code is right, because `true` there means the question does not
      arise rather than that anything succeeded, and the reason is on the
      line. Reported rather than silenced, and the estate file was not edited
      to make this cell's number nicer.
      
      Measured: 4 fabrications, all shapes, none caught by a word. 35 spec
      tests; scan PASSES with 15 negative controls, seven of which check a
      detector stays SILENT. -->
      
    • S11.18.c Detect numeric/boolean pass/default score, random/hash/length/ keyword heuristics presented as domain evaluation, UI-computed truth, and missing-to-zero/green. Evidence: score rules/fixtures. <!-- A PROVENANCE IS A CLAIM ABOUT WHERE A NUMBER CAME FROM, AND EVERY CHECK IN THE ESTATE IS ABOUT ITS SHAPE.

      CREDIT, AND scores.ts IS ONE OF THE BEST-MODELLED THINGS IN THE KIT. A
      reading is measured or one of eight NON-VALUE STATES -- absent,
      not-applicable, insufficient-data, unsupported, unavailable, failed,
      stale, redacted -- with not-applicable and unsupported marked NOT
      COUNTED, because "this does not apply" and "we do not know" are
      different answers and averaging them together is how a dimension nobody
      could measure becomes a zero. Every reading carries a ScoreProvenance
      naming the evaluator and its version, the corpus and its version, the
      model, the provider (required whenever a model is named), the
      preprocessing, the policy the verdict would be derived under, the input
      revision, the input hash, the run and the instant. Nineteen violation
      kinds check the vector, down to an uncertainty interval that excludes
      its own value and a sample of one carrying a standard deviation.
      
      EVERY ONE OF THOSE CHECKS IS ABOUT THE RECORD. The scan builds a reading
      produced by dividing a length, gives it a provenance naming a model
      served by a provider, and puts it through the estate's OWN
      validateVector: ZERO violations out of nineteen -- the record is
      impeccable. It then reads the PRODUCER and reports the heuristic. Same
      number, same run.
      
      FOUR DETECTORS, NONE CATCHABLE BY A WORD.
      
      A judgement that DEFAULTS to a passing value: `?? 0` on a risk score is
      a claim that nothing is wrong, made by the operator that exists because
      something was missing. `?? 0` on a latency produces nothing, because the
      NAME is what separates them, and a default to a non-passing value
      produces nothing either -- defaulting to the worst case is a different
      and often correct decision.
      
      A heuristic over a length, a hash or a character code presented as a
      domain evaluation. It requires the body to read NOTHING ELSE, because a
      length among other inputs is a legitimate feature and the defect is
      length being the whole of it.
      
      A verdict computed inside the component that draws it. This is a
      fabrication rather than a style preference because such a verdict has no
      evaluator, no corpus, no policy and no run behind it: the number on the
      screen is real and its history does not exist, so it cannot be
      reproduced, compared across revisions, or explained to the person it is
      about.
      
      An absence resolved to a number. This is the one that matters most and
      the cheapest to get wrong: NON_VALUE_STATES exists, is complete, and is
      a keystroke further away than `?? 0`. Nothing in the estate makes the
      honest answer the easy one.
      
      THE SWEEP OVER 303 KIT SOURCES FOUND THREE FALSE POSITIVES AND EVERY ONE
      TIGHTENED A DETECTOR RATHER THAN AN ALLOW-LIST.
      
      `counts.set(verdict.owner, (counts.get(verdict.owner) ?? 0) + 1)` -- a
      counter, and defaulting a count to zero is correct. Only the assignment
      target and the operand's ROOT count now, walked backwards over one
      balanced group so an argument cannot supply the name.
      
      `good += creditOf(spec.grades, observation.outcome) ?? 0` -- same fix,
      same reason.
      
      And provider-registry's unprobed-provider branch, which returns
      `{ situation: 'probe-failed', status: 503, retryHelps: true, ... }` --
      code whose own detail says "an unprobed provider is not a healthy one,
      and reporting it as healthy is the failure this state exists to
      prevent". Matching ANY field with a passing value had read
      `retryHelps: true` as a claim of success. Advice about whether to retry
      is not a score, so only a field carrying the ANSWER counts now. Both
      corrections are locked in by spec tests naming the real lines.
      
      The sweep now reports zero. (Its first run also refused this scan's own
      fixture with non-finite-value, because the reading was hand-written with
      the wrong field name -- the estate's check working on my fixture. It is
      built with the estate's own `measured()` constructor now.)
      
      The lexical pre-pass, its limits and the inline opt-out are S11.18.b's,
      reused rather than reimplemented. Measured: 4 fabrications, all
      checkable only over the producer, 3 of them with an honest answer that
      already exists in the estate. 28 spec tests; scan PASSES with 13
      negative controls, six of which check a detector stays SILENT. -->
      
    • S11.18.d Detect fixture/sample/demo seed/interception/test builder in production dependency graph or unlabeled runtime output. Evidence: fixture rules/fixtures. <!-- A DEPENDENCY ON A FIXTURE IS NOT A DEFECT. A PATH FROM A NAME THAT PROMISES SOMETHING ELSE IS.

      S11.18.a recorded this class as unreachable by the estate's scan and
      gave the reason: whether production code imports a fixture is a property
      of the GRAPH and not of any line, so a grep over text cannot see it at
      all. This cell is the graph. 305 modules, 623 sibling import edges,
      walked breadth-first so the path returned is the SHORTEST one -- which
      is the edge a reader is most likely to be able to cut.
      
      WHAT IS NOT A FINDING. A kit that publishes its fixture modules as entry
      points BY NAME is being honest: a consumer who writes
      @oshun/workbench-kit/tamper-corpus has asked for fixtures, in those
      words, and got them. So the 25 declared fixture entry points are LISTED,
      not reported. Without that distinction the report is 25 findings that
      are all correct behaviour, and nobody reads the 26th. A checker whose
      subject is a corpus must import the corpus; a module about fixture MODE
      is about a mode, not a fixture.
      
      WHAT IS. queue-invariants -> lease-fixtures. A consumer importing a
      module called queue-invariants has asked for invariants; whether a
      corpus arrives with them is a fact about the graph nobody can read off
      the import statement. AND THE FINDING IS A PATH, NOT A PAIR: "X depends
      on a fixture" is unactionable, and X -> Y -> corpus names the edge
      somebody has to cut, so every finding carries the whole chain.
      
      AND THE MARK EXISTS AND NOTHING APPLIES IT. synthetic.ts ships
      SYNTHETIC_SEGMENT_PREFIX, syntheticMark() and isSyntheticId -- a real
      runtime boundary, which anti-mock.ts names as one of its strongest
      instruments, because a mark somebody deliberately put there is a
      POSITIVE trace whose presence and absence mean different things.
      Measured: 0 of 25 fixture modules import it. That is the
      "unlabeled runtime output" half of this cell, and it is not a missing
      feature -- the feature is written, exported and documented, and what is
      missing is the wire between the modules that manufacture data and the
      one module that would let a reader downstream tell that data apart.
      Whether a VALUE carries the mark is a runtime question; whether the
      module that made it can even apply one is a static edge, and that is
      what this measures.
      
      INTERCEPTION IS RECORDED AS ANSWERABLE ONLY BY A RUN. A responder
      installed in a test process is not in the import graph of the code under
      test; the estate answers it from the harness with an outbound-request
      record. There is a register fault that fires if a runtime-only source
      ever claims a static detector, because swapping the two is how a suite
      reports coverage it does not have.
      
      The specifier reader deliberately runs over the RAW source, since a
      specifier IS a string literal and S11.18.b's blanking pre-pass would
      erase exactly what this needs -- the two passes want opposite things
      from the same text. Its own spec then caught that a bare side-effect
      `import './x.js';` has no `from` and was invisible to the walk, which is
      an edge nobody could cut. The counts live in the run rather than in the
      module's prose, so they cannot go stale.
      
      Measured: 6 sources, 5 answerable by the walk and 1 only by a run. 25
      spec tests; scan PASSES with 12 negative controls, including that a
      cycle terminates, that the SHORTEST path is returned rather than the
      first found, and that renaming the marking module does not silently make
      every fixture look compliant. -->
      
    • S11.18.e Detect fabricated success/verified/published/saved/healthy/ configured/complete state from request dispatch, HTTP 2xx, empty value, build env, optimistic UI, or swallowed error. Evidence: success rules. <!-- A SUCCESS IS A CLAIM THAT SOMETHING HAPPENED. EVERY FABRICATION IN THIS CELL REPORTS A DIFFERENT, EARLIER EVENT INSTEAD.

          dispatch        we asked                      -> "it is saved"
          a 2xx status    the server answered           -> "it is published"
          an empty value  nothing came back             -> "it is healthy"
          a build value   it was true when we compiled  -> "it is configured"
          optimism        we expect it to               -> "it is complete"
          a caught error  we stopped looking            -> "it succeeded"
      
      That framing is what makes six shapes one cell rather than six unrelated
      greps. Each names a real moment that really occurred, and the defect is
      the distance between that moment and the one the word claims. It also
      gives every finding an obvious repair -- report the event you actually
      have -- which is why the substituted event travels IN the finding: "do
      not fabricate success" is advice, and "you have a dispatch and you are
      printing a save" is a change somebody can make this afternoon.
      
      TWO OF THE SIX ARE ALREADY NAMED BY THE ESTATE, AT RUNTIME. anti-mock
      has empty-success -- "a well-formed answer carrying nothing, from a
      question nobody asked" -- and build-time-guess -- "a value decided when
      the bundle was built, presented as current" -- each with the absence it
      claims written out and an instrument that refutes it. This cell asks the
      same two questions of the SOURCE, plus four the estate does not name.
      The register records which is which, with faults that fire if an
      estate-named source cites no masquerade class or a cell-only source
      cites one, and the run checks the cited classes against
      MASQUERADE_CLASSES rather than trusting them. Crediting one instrument
      with the other's reach is how a suite reports coverage it does not have.
      
      THE SUCCESS WORDS ARE THE CELL'S, NOT MINE: success, verified,
      published, saved, healthy, configured, complete. A boolean called `sent`
      set after a dispatch is accurate and produces nothing; a boolean called
      `saved` set after the same dispatch is not.
      
      AND THE DISCARD IS THE EVIDENCE. `const receipt = await send(x)` may
      well be checked, and `await send(x)` on its own line means nothing looked
      at what came back. A status branch that opens the body produces nothing.
      An empty list answered with "unknown, no check has run" produces
      nothing. A catch returning a typed failure produces nothing.
      
      THE RUN SHOWS ALL SIX ON ONE FIXTURE AND ZERO ON THE HONEST REWRITE OF
      THE SAME SIX FUNCTIONS, because a detector that has only been shown
      firing has not been shown to be quiet.
      
      Swept over the kit's production sources: one finding, at staleness.ts,
      where `truncated.length === 0` returns `complete: true`. Read by hand
      and CORRECT: there the emptiness of `truncated` IS the completeness
      rather than a substitute for it. That distinction is not one a detector
      can make in general, which is exactly why this module says a finding is
      a place to look and not a verdict. Reported rather than silenced.
      
      (Its own fixture caught that the optimistic detector saw no assertion in
      `setSaved(true)` -- a setter call is how a component says it, and the
      first version matched only fields and assignments, so the detector found
      nothing in the one shape it exists for.)
      
      Measured: 6 sources, 2 already named by the estate at runtime, 4 only
      here. 28 spec tests; scan PASSES with 14 negative controls, six of which
      check a detector stays SILENT on the honest form. -->
      
    • S11.18.f Classify blocker/review/allowed test-only/known false-positive with source, dataflow/context, owner, disposition, expiry, and linked proof; no broad ignore patterns. Evidence: triage schema. <!-- A SUPPRESSION WITH A REASON AND NO EXPIRY IS PERMANENT.

      CREDIT FIRST, AND THE ESTATE'S ALLOW-LIST IS THE RIGHT SHAPE. It is the
      narrowest one available: an end-of-line comment, // stub:legitimate
      <reason>, suppressing exactly the line it sits on. Nothing about it can
      be widened by accident, it travels with the code it excuses, and it
      requires somebody to write a sentence. Against a config file listing
      paths to ignore, that is a much better design.
      
      IT CARRIES ONE OF THE SIX FIELDS THIS CELL ASKS FOR. There is a reason.
      There is no owner, no disposition, no expiry, no linked proof, and no
      statement of the dataflow that makes the line safe. So a suppression
      written on a Tuesday to get a branch through is indistinguishable,
      forever, from one a team argued about.
      
      MEASURED OVER THE WHOLE REPOSITORY: 8,129 markers across 3,234 files.
      24 carry an owner (0.3%). 27 carry a linked proof (0.3%). TWO carry no
      reason at all. ZERO carry an expiry -- so all 8,129 are permanent, and
      nothing will ever ask about any of them again.
      
      A REASON THAT PROMISES A FUTURE IS AN EXPIRY WITH NO DATE. 206 of them
      say "until the real pose-estimation pipeline lands", "will call the
      actual model inference", "V2-deferred", "for now". Each is a claim that
      the suppression is TEMPORARY, made by somebody who did not say when --
      which is the missing field, written in prose where nothing can read it.
      That detector is a documented approximation over a small explicit word
      list, and it accepts any reason carrying a date whatever else it says:
      it cannot tell a good reason from a bad one, which nothing can by
      machine, and it can tell a reason that promises a future from one that
      does not, which is the checkable half.
      
      AND BREADTH IS A PROPERTY OF THE DIRECTIVE, NOT OF THE REASON. A
      line-scoped suppression excuses one line. A file-scoped one excuses
      every line that will ever be added to that file, including lines nobody
      has written yet, and the person who adds them never sees the comment at
      the top. The cell says "no broad ignore patterns", and the only way to
      hold that is to make breadth a field and judge it. Of 1,603
      eslint-disable directives, 763 are FILE-SCOPED block comments -- and
      they go through the SAME triageFaults as everything else, because a
      number printed beside a schema is not a number judged by it. That
      produced 3,815 findings over 763 directives and is what makes
      suppression-wider-than-a-block a finding rather than a fault kind
      nothing in this repository exhibits.
      
      THE EXPIRY CHECK TAKES THE INSTANT AS AN ARGUMENT rather than reading a
      clock -- the run uses the newest commit date -- so two runs of the same
      tree agree, and a test that asks "is this expired" does not answer
      differently tomorrow. And an allow-list may not hold a blocker, which is
      its own fault kind.
      
      Across every suppression in the repository, 7 of the 9 fault kinds are
      exhibited; record-expired and blocker-that-is-suppressed are not, which
      is reported rather than left as an unexplained gap.
      
      Measured: 6 fields, 1 carried by the marker, 1 (disposition) with
      nowhere to go at all -- so a needs-review and a known-false-positive are
      written identically and a reviewer cannot sort them -- over 4
      dispositions the marker cannot tell apart. 23 spec tests; scan PASSES
      with 14 negative controls, including that a complete record produces
      nothing and that the same promise WITH a date is accepted. -->
      
    • S11.18.g Run static patterns plus AST/import/bundle checks and targeted runtime/adversarial tests, then independently review zero-blocker report. Evidence: signed scan report. <!-- A SIGNATURE SAYS WHO WROTE A NUMBER. IT SAYS NOTHING ABOUT WHETHER THE NUMBER IS RIGHT.

      Three of this cell's four asks are ordinary work and the run does them
      for real. It EXECUTES all eight scans -- the estate's own stub scan plus
      S11.18.a through .f plus S11.17.g's mutation harness -- collects their
      exit codes, reads each one's negative-control result out of its own
      output, builds a canonical preimage, digests it, signs it with a freshly
      generated ed25519 key, verifies the signature, and shows that one
      altered field breaks verification. The preimage covers the REVIEW STATE,
      so a signature cannot be lifted from an unreviewed report onto a
      reviewed one, and it does not depend on the order the scans arrive in.
      
      THE ZERO THIS CELL ASKS FOR DOES NOT EXIST FOR THIS REPOSITORY. The
      estate's own whole-repo stub scan reports 3,341 actionable hits. The
      report states the number and names the scan that produced it, and the
      run PASSES while reporting it -- because refusing to publish an honest
      non-zero would be the exact defect this section is about, committed by
      the cell that closes it. A non-zero is a success of the instrument. The
      count is read from the scan's own output rather than collapsed from a
      non-zero exit code, because the difference between one blocker and three
      thousand is the whole content of a report.
      
      (That number is not an indictment of the estate's scan. The hook scans
      STAGED files and CI scans the PR DIFF, so the scan is a ratchet on new
      code rather than a gate on the tree -- which is a deliberate design, and
      is why the R100 rename carve-out exists. What it means is that the
      whole-repo zero this cell asks for has never been required, and this is
      the first run that asked for it.)
      
      AND INDEPENDENT IS A FACT ABOUT WHO. The fourth ask cannot be done by
      the party that produced the report -- not for want of a tool, but
      because there is one party here, and a zero produced, signed and
      reviewed by the same hand is the claim and the evidence in one hand. A
      signature over that proves only that the same hand held both. So the
      report carries unreviewed-because-the-producer-is-the-only-party as a
      VALUE rather than a footnote, and reportFaults refuses any zero-blocker
      report whose named reviewer is its own producer: the only way to get the
      strong claim is for somebody else to sign it. A report that plainly
      admits it was not reviewed produces no finding at all, because being
      unreviewed is a state and claiming otherwise is the fault.
      
      WHAT A ZERO IS WORTH WITHOUT THAT should also be said, and it is quite a
      lot: every scan in this section carries negative controls that exit 2 if
      a check cannot fail, so a zero from a scan that ran is a zero from an
      instrument that has been shown to work. What it is not is INDEPENDENT,
      and those are different claims that are easy to write as one sentence.
      
      A SCAN THAT DID NOT RUN IS NOT A SCAN THAT FOUND NOTHING, and a report
      holding one may not claim zero at all -- the same rule S11.17.g used for
      its mutation baseline, for the same reason.
      
      `bundle` is the one instrument the cell names that nothing exercises,
      for the reason S11.18.a measured: generated paths are dropped before the
      suffix filter, so no scan in this section opens a built bundle. Reported
      as an unexercised instrument rather than quietly dropped from the list.
      
      (Two of this scan's own negative controls MISSED on the first run,
      because they were derived from the live report -- which has 3,341
      blockers -- and the producer-reviewed check only fires on a ZERO-blocker
      report. A control working. They are built from an explicitly clean
      fixture now.)
      
      Measured: 8 scans, 4 of 5 instruments, 3,341 blockers, signed and
      verified, NOT independently reviewed. 26 spec tests; scan PASSES with 13
      negative controls. -->
      

S11.18 parent rollup — closed by a–g (note only; the checkbox is the S11.18 line above, §2.5.11). <!-- THE SECTION'S SHAPE: THE ESTATE HAS ONE INSTRUMENT AND IT LOOKS FOR ADMISSIONS.

  scripts/stub-indicator-scan.sh is real, runs in five places, has three
  modes, an allow-list that carries a reason, a documented test exclusion
  and a rename carve-out learned from an incident. It caught this session
  three times. What it is, is a grep for somebody ADMITTING what they did --
  "in production, this", "would use", "placeholder", "TODO:" -- and it works
  because developers write those admissions honestly.

  EVERY CELL IN THIS SECTION FOUND A DEFECT THAT SAYS NOTHING ABOUT ITSELF.
  .b: four provider fabrications with zero scan vocabulary between them, put
  through the estate's own pattern for 0 hits and through the shape
  detectors for 4. .c: a reading produced by dividing a length passes all 19
  of the estate's vector checks, because a provenance is a claim about where
  a number came from and every check is about its shape. .d: whether
  production code imports a fixture is a property of the GRAPH, invisible to
  any grep -- 305 modules and 623 edges walked to find it. .e: six ways to
  report success, each substituting a different earlier event. .f: 8,129
  suppressions, ZERO with an expiry.

  AND .a MEASURED WHY: the scan's scope is thirteen suffixes, and 5,172
  out-of-scope tracked files carry code somebody runs -- 3,638 of them .mjs,
  which is what every verification tool in this initiative is written in,
  INCLUDING THE ONE THAT PRINTS THAT NUMBER.

  THE DISCIPLINE. Every detector here is shown SILENT on the honest form as
  well as loud on the defect, because a detector only shown firing has not
  been shown to be quiet. Three sweeps over the kit's own sources found four
  false positives, and every one tightened a detector rather than an
  allow-list -- including provider-registry's unprobed-provider branch,
  whose retryHelps:true was read as a claim of success in code whose own
  comment says reporting it as healthy is the failure that state exists to
  prevent.

  .g closed it by running all eight scans, signing the report, and saying
  the two things the report cannot say: the whole-repo zero does not exist
  (3,341 hits), and nobody independent has looked.

  Seven cells, seven scans, all passing with negative controls; the section
  closes with the kit at 304 entry points and 0 violations. -->
  • S11.19 Add dependency-license, secret, SAST, and supply-chain scans to the release evidence index. Evidence: critical findings block release.
    • S11.19.a Scan direct/transitive runtime/build/dev/container/native/ model/plugin/DCC/tool dependencies for vulnerabilities, reachability, maintenance, source, hash, and advisories. Evidence: dependency report. <!-- A GATE IS THE INTERSECTION OF WHAT BLOCKS AND WHAT IS COVERED.

      CREDIT, AND THE PIPELINE IS SUBSTANTIAL. security-deps-and-secrets.yml
      runs a pnpm audit with a documented baseline and a severity threshold,
      the OSV scanner recursively over the whole tree, gitleaks and
      trufflehog, and aggregates them into a release gate.
      sbom-and-sign.yml runs syft for an SBOM, grype over it, cosign for
      signatures and attest-build-provenance. Most repositories do not have a
      supply-chain posture like this.
      
      THE GATE REQUIRES THREE OF THE FOUR SCANNERS, and says so in a comment:
      "OSV is informational on this gate; it stays in artifacts." That is
      honest labelling and deserves credit rather than a finding on its own.
      
      WHAT IT MEANS IS THE FINDING. pnpm audit reads a pnpm lockfile and
      nothing else, so OSV -- run with --recursive ./ -- is the only scanner in
      the pipeline that sees a Cargo.lock, a poetry.lock or a go.mod. And it
      is the one that cannot fail the build. Measured against the tracked
      tree: 6 of the 7 present ecosystems have NOTHING that can fail the build
      reading them -- 94 Cargo lockfiles, 34 python manifests, 3 Go modules,
      156 Dockerfiles, 102 .uplugin, 2 .csproj. Neither half of that is a
      mistake. The intersection is empty, and an intersection is what a gate
      is.
      
      Which jobs the gate REQUIRES is PARSED out of the workflow rather than
      restated here, because a register naming the blocking scanners is a
      second copy of a fact that lives in a YAML file, and the two drift. The
      run also reports when the gate runs at all: only on main and release
      branches.
      
      AND THE CELL NAMES SIX ATTRIBUTES, OF WHICH THREE ARE DATABASE
      QUESTIONS, TWO ARE LOCKFILE QUESTIONS AND ONE IS A GRAPH QUESTION.
      
      Vulnerabilities, advisories and maintenance are facts about the outside
      world that arrive from a database somebody else maintains; this
      repository can check that it ASKS, and cannot answer them itself. Source
      and hash are in the lockfile and checkable offline -- 6,502 of 6,503
      resolutions carry an integrity hash, which is a real guarantee, enforced
      by the INSTALLER rather than by the gate.
      
      REACHABILITY is neither. It is a property of the import graph,
      computable with no network and no database, and no tool in the pipeline
      computes it -- so a critical advisory against a package nothing imports
      and a critical advisory against a package on the login path arrive at
      the same severity, in the same report, with nothing to tell them apart.
      S11.18.d already built the walk; this cell records what joining it to a
      dependency name would take, and does not pretend the join exists.
      
      Measured: 6 attributes -- 1 asked by something that blocks, 3 asked only
      by something informational, 2 asked by nothing, 3 answerable inside this
      repository. 9 dependency kinds, 7 ecosystems. 15 spec tests; scan PASSES
      with 10 negative controls, including that making every scanner blocking
      reports FEWER ungated ecosystems, so the count is a measurement rather
      than a constant. -->
      
    • S11.19.b Generate license/notice/SBOM and detect prohibited/unknown/ missing/incompatible/transitive/model/data/font/media license or attribution. Evidence: license report. <!-- AN SBOM IS A BILL OF SOFTWARE. FOUR OF THE NINE THINGS THIS CELL NAMES ARE NOT SOFTWARE.

      CREDIT: sbom-and-sign.yml generates a real SBOM in both CycloneDX and
      SPDX from syft, scans it with grype, signs it with cosign and attaches a
      build-provenance attestation. Both formats carry a licence field per
      component, so for PACKAGES the question this cell asks is answered by
      machinery that exists and runs. The formats are parsed out of the
      workflow rather than asserted here.
      
      A FONT IS LICENSED. An image is licensed. An audio file is licensed. A
      model weight is licensed, usually on terms stricter than any code in the
      tree -- non-commercial, or research-only, or requiring a named
      attribution in the product. None of them is a package, so syft
      enumerates none of them, grype scans none of them, and neither SBOM has
      a component to attach a licence to.
      
      MEASURED FROM git ls-files: 545 tracked files carry terms and appear in
      no SBOM component -- 448 media (266 .png, 111 .svg, 66 .jpg, 3 .mp3, 1
      .wav, 1 .webm), 85 data (60 .csv, 25 .jsonl), 9 fonts (5 .ttf, 4
      .woff2), and 3 .onnx weights.
      
      The register's distinction is a-file-in-the-tree-with-no-manifest versus
      a-package-manifest-an-sbom-reads, because a thing with no manifest is
      not a component the scanner MISSED -- it is a component the scanner has
      no way to know about, and the repair is an inventory rather than a
      configuration change. There is a fault kind that fires if an entry ever
      claims an SBOM component for something with no manifest.
      
      AND THE ROOT OF THE TREE HAS A LICENSE AND NO NOTICE. LICENSE is
      Apache-2.0: what this repository GRANTS to whoever takes the code. A
      NOTICE points the other way -- what this repository OWES to the people
      whose work it ships -- and Apache-2.0 section 4(d) is explicit that a
      redistributed work carrying a NOTICE must carry its contents onward. So
      a tree with dependencies and no NOTICE has answered one direction only.
      Conflating the two directions is easy and expensive, so they are
      separate values in the register.
      
      WHY "TRANSITIVE" IS ITS OWN SUBJECT: a direct dependency's licence is
      something a person chose, and a transitive one arrives with it. An SBOM
      lists both identically, which is right for a bill of materials and wrong
      for a review queue -- one remedy is a decision to revisit and the other
      is a conversation with whoever owns the parent.
      
      Measured: 5 licensed subjects, 1 in the SBOM and 4 outside it over 545
      files; 1 of 2 required attribution artifacts present; 6 defect kinds the
      cell names. 16 spec tests; scan PASSES with 10 negative controls,
      including that a subject whose files are ABSENT from the tree is not
      reported, so the count is a measurement rather than a constant. -->
      
    • S11.19.c Scan source/history/config/build artifacts/images/bundles/log/ test snapshots for credentials/tokens/keys/certs/signed URLs/secret patterns and verified canaries. Evidence: secret report. <!-- A SCANNER THAT FINDS NOTHING AND A SCANNER THAT IS NOT RUNNING PRODUCE THE SAME REPORT. A CANARY IS THE ONLY THING THAT TELLS THEM APART, AND THERE IS NONE.

      CREDIT, AND BOTH SCANNERS BLOCK. gitleaks and trufflehog each run on
      every release, each with fetch-depth: 0, so both read the whole HISTORY
      and not only the working tree -- which is the half of secret scanning
      that matters most, because a key deleted in the next commit is still in
      the pack file forever. The release gate requires both. The flags and the
      checkout depths are PARSED from the workflow rather than restated here.
      
      AND TRUFFLEHOG RUNS --only-verified. That flag means a secret is
      reported only when the credential could be CONFIRMED live by calling the
      provider, and it is a deliberate, defensible answer to alert fatigue: a
      queue full of unverified findings is a queue nobody reads.
      
      It also means the class of things that do not fail the build is larger
      than it looks. 4 of the 6 kinds this cell names cannot be verified by
      anybody. A private key has no endpoint to present it to. A signed URL
      grants access by existing, and calling it would be USING it, so
      verification is not something a scanner should do. A certificate and a
      pattern match authenticate nothing. And a key rotated last week no
      longer verifies while still sitting in the history for whoever had it
      before it was rotated.
      
      Neither half is a defect on its own. Writing both down is what stops
      "the secret scan passed" being read as "there are no secrets".
      
      4 OF THE 8 SURFACES ARE NOT PRESENT WHEN THE SCAN HAPPENS. A build
      artifact, an image layer and a bundle are produced later; a runtime log
      is never on that machine at all. That is a fact about where the surface
      IS, rather than a gap in a scanner -- the same shape S11.18.a found for
      the stub scan and the bundle.
      
      AND THE CELL ASKS FOR VERIFIED CANARIES, WHICH DO NOT EXIST. Without
      one, a green secret scan and a misconfigured secret scan are
      indistinguishable, and the difference is discovered by the incident.
      
      A CANARY MUST NOT BE A WORKING CREDENTIAL is the first of the five
      requirements, and it is checked by an explicit invalidating marker
      rather than by guessing at shape: deciding "is this a real key" by
      appearance is exactly the guess this repository refuses elsewhere, and
      the cost of being wrong is a live credential in a git history. A
      proposed canary is run against all five and fails exactly one -- nobody
      has seen a scanner report it -- so it is a plan for a control rather
      than one. There are also fault kinds for a canary planted in a surface
      no scanner reaches, and for a canary of a kind --only-verified filters
      out, because either would be silent whether the scanner worked or not.
      
      Whether either scanner would find a canary is UNOBSERVED here: the run
      measures the host and finds neither binary. This cell does not claim
      otherwise.
      
      Measured: 6 kinds (2 reported under the flag, 4 not), 8 surfaces (4
      present when scanned, 4 not), 0 canaries. 15 spec tests; scan PASSES
      with 10 negative controls, including that a register WITH a canary for a
      reportable kind stops reporting the gap. -->
      
    • S11.19.d Run language/framework SAST for injection, auth, tenant, SSRF/ upload/path, crypto, deserialization, traversal, unsafe subprocess, privacy, and unsafe API use with reachability. Evidence: SAST report. <!-- CODEQL'S REACHABILITY IS REAL, AND IT STOPS AT THE LANGUAGE BOUNDARY.

      CREDIT, AND THE CONFIGURATION IS A GOOD ONE. codeql.yml runs with
      +security-extended,security-and-quality -- not the default suite, the
      wide one -- over libs, apps and lilith/services, on every push, every
      pull request and a weekly cron, excluding node_modules, dist and test
      files by path. And CodeQL is a DATAFLOW engine: the cell asks for
      reachability, and this is the one instrument in the whole pipeline that
      genuinely has it, because a taint path from a request parameter to a
      query is a far stronger claim than a pattern match.
      
      THE PATHS ARE NOT THE GAP. Measured against the tree, those three roots
      hold 50,482 of the 51,512 non-test TypeScript files -- 98.0% -- and the
      remainder is mostly generated evidence. This cell says so rather than
      reaching for the easy finding.
      
      THE LANGUAGE IS. matrix.language is ["javascript-typescript"] alone. The
      estate has 9,306 Rust files, 3,556 Python, 2,871 C/C++ and 269 C#, and
      none of them gets any security analysis. cargo clippy -D warnings is a
      correctness and style lint -- valuable, and not a taint analysis, and
      saying which it is matters.
      
      AND THE BOUNDARY IS CROSSED CONSTANTLY. 361 files hand a value to a
      subprocess and 656 reference a native binding. Every one is a place
      where a value leaves the analysed language, and an engine that is right
      about everything it can see returns nothing about the step across --
      which is exactly where an unsafe-subprocess finding would matter most.
      
      TWO OF THE ELEVEN CATEGORIES ARE NOT LANGUAGE DEFECTS. Injection, path
      traversal, SSRF, unsafe deserialization, weak crypto and unsafe API use
      are properties of code that any good engine knows about, and the wide
      suite covers them. TENANT ISOLATION and PRIVACY are not: which argument
      is a tenant, and which field is personal data, are facts about THIS
      estate that no generic query pack has ever seen. A pack cannot know that
      a column called handle is personal and one called slug is not.
      
      The estate HAS the vocabulary for the first -- a tenant id type and a
      scope key that puts the tenant first, both verified by running them in
      S11.16.g -- so a custom query is writable. There are ZERO .ql files in
      the repository. The register carries a fault that fires if any entry
      ever claims a shipped suite covers a rule only this estate knows.
      
      Measured: 11 categories (9 properties of code, 2 rules no pack knows),
      1 language analysed of 6 present, 1,017 boundary crossings, 0 custom
      queries. 11 spec tests; scan PASSES with 9 negative controls, including
      that the same boundary count with SEVERAL languages analysed is not
      reported, so the finding is about the pair rather than about the
      crossings. -->
      
    • S11.19.e Verify lockfile/integrity/signatures/provenance, pinned actions/ tools/images, build isolation, generated artifacts, package confusion, typosquat, install scripts, and signing chain. Evidence: supply-chain report. <!-- A POLICY THAT LIVES ONLY IN THE PATTERN OF PAST DECISIONS CANNOT BE VIOLATED -- ONLY DEPARTED FROM UNNOTICED.

      CREDIT, AND TWO OF THESE ARE EXCELLENT. The lockfile carries an
      integrity hash for effectively every resolution and the INSTALLER
      verifies it, so the bytes that arrive are the bytes that were locked.
      And package.json declares onlyBuiltDependencies -- an explicit
      allow-list of the 35 packages permitted to run an install script, beside
      ignore-scripts=false. That is the strongest single control in this
      section: an install script runs arbitrary code on a developer machine
      and on every CI runner, and this estate names exactly which packages may
      do it.
      
      AND THE ACTION PINNING IS A REAL POLICY THAT NOBODY WROTE DOWN. Of 796
      workflow references across 106 workflows, 11 are pinned to a commit --
      and every one is a third party that either installs a toolchain or scans
      code: dtolnay/rust-toolchain, aquasecurity/trivy-action,
      trufflesecurity/trufflehog, snyk/actions/node. That is not an accident.
      It is somebody applying a sound rule, repeatedly, on exactly the
      references where a moved tag would be worst.
      
      The pattern is DERIVED from the references rather than asserted, because
      asserting "the policy is to pin the security tools" would be a claim
      about somebody's intent, and deriving it makes the exceptions findings
      rather than opinions.
      
      THE RULE EXISTS NOWHERE. So ten other security-sensitive references are
      named by something the owner can move: sigstore/cosign-installer@v3.7.0,
      which is the signing tool the entire chain rests on;
      anchore/sbom-action@v0, a MAJOR-version tag on the action that produces
      the SBOM; anchore/scan-action/download-grype@v4;
      google/osv-scanner-action; gitleaks-action@v2; and three zaproxy
      actions.
      
      AND THE CLEAREST PROOF IS ONE ACTION TREATED TWO WAYS.
      dtolnay/rust-toolchain is SHA-pinned in one workflow and @stable -- a
      BRANCH -- in another, in the same repository, and nothing notices. The
      run reports that pair by name and fails if it ever disappears without
      the finding being rewritten.
      
      A tag is a name the owner can move; a commit is not. The gap between the
      pinned and the unpinned is not a difference in judgement -- it is the
      absence of anything that reports one.
      
      The register keeps a-pattern-nobody-wrote-down as a state distinct from
      both a mechanism and nothing, because it is strictly better than nothing
      (somebody was applying a rule) and strictly worse than a rule (nothing
      can report a departure). Three controls are held by nothing: pinned
      images (no check reads a Dockerfile FROM line), generated artifacts
      (S11.18.a measured that every scan excludes them, so what a build emits
      is checked by nothing before it is signed), and typosquat.
      
      Measured: 13 controls -- 6 enforced by a mechanism, 2 declared, 2 held
      by a pattern nobody wrote down, 3 by nothing; 626 of 796 references on a
      major-version tag. 16 spec tests; scan PASSES with 10 negative controls,
      including that the same sensitive action pinned to a COMMIT is not
      reported, so the finding is about the pinning rather than about the
      owner. -->
      
    • S11.19.f Normalize findings with scan/tool/db version, artifact/commit, severity/exploitability/scope, owner, fix/exception/expiry, retest, and raw evidence link. Evidence: evidence index rows. <!-- A LINK TO RAW EVIDENCE THAT EXPIRES IS A FINDING THAT BECOMES UNCHECKABLE ON A SCHEDULE.

      The cell requires a raw evidence link on every row, and the estate
      produces exactly that: every scanner uploads its report as a workflow
      artifact, and a finding can name the run it came from. Measured across
      the workflows: 94 uploads, 91 declaring a retention -- 4 kept for ONE
      day, 29 for seven, 36 for fourteen, 19 for thirty, 3 for ninety.
      
      So a finding raised on a Monday, with an exception granted for a
      quarter, points at a report that stopped existing before the first
      review. And the row STILL HAS A LINK IN IT, which is worse than an empty
      field, because an empty field is obviously empty.
      
      THAT IS NOT AN ARGUMENT FOR LONGER RETENTION. Storage costs money and
      old scan output ages badly. It is an argument that the retention period
      is part of the FINDING: a row whose exception outlives its evidence has
      to say so. exception-that-outlives-its-evidence names both numbers --
      "the exception runs 92 day(s) from 2026-08-01 and the evidence is kept
      for 14" -- so the repair is obvious and is a choice between two things
      somebody can actually change.
      
      AND A VERSION IS THREE VERSIONS. The cell asks for scan, tool and
      database version, and they move independently: the same tool, on the
      same commit, against a different advisory database produces a different
      answer, and only the third explains why a finding appeared with no code
      change. Recording one number called "version" loses exactly the case
      somebody will need to explain.
      
      EXPLOITABILITY IS SEPARATE FROM SEVERITY, because they answer different
      questions and a queue sorted by the wrong one wastes the week: severity
      is how bad it would be, exploitability is whether it can happen here.
      S11.19.a measured that nothing in the pipeline computes reachability, so
      that field arrives from a person -- and a critical with it unknown is a
      finding of its own, because it is the field that decides whether this
      week is spent on it.
      
      Both instants are ARGUMENTS rather than a clock read, so an expiry
      answer is reproducible: the same fix S11.18.f needed, for the same
      reason.
      
      Measured: 13 fields, 5 severities, 4 exploitability values, 3 remedies,
      12 fault kinds. 13 spec tests; scan PASSES with 11 negative controls,
      including that an exception INSIDE its evidence window is accepted and
      that the same unknown exploitability on a low severity is not a
      finding. -->
      
    • S11.19.g Block critical policy findings, expired exceptions, missing/ stale/incomplete scans, or artifact mismatch and prove with deliberate fixture. Evidence: release gate test. <!-- THE GATE ASKS WHETHER THE SCANNER SUCCEEDED. IT DOES NOT ASK WHAT THE SCANNER FOUND.

      release-gate asserts needs.<job>.result == "success" for each required
      scanner. That is a fact about a JOB. A job that ran, reported 156
      reviewed high and critical advisories, and exited zero because they are
      in a checked-in baseline is `success` -- and so is a job that found
      nothing. The gate cannot tell them apart, because the only thing it
      reads is the exit.
      
      THE BASELINE IS NOT THE DEFECT. non-v1-known-high-critical.json carries
      a schema version, the audit level it was generated at, an explicit
      scope, and a reviewedAt date of 2026-05-15; the audit tool splits
      findings by path so anything under a V1 prefix still blocks. That is a
      considered, documented, reviewable decision about which parts of a large
      monorepo gate a launch. The defect is that "the release gate passed" and
      "there are no high or critical advisories" are different statements, and
      the gate's output only supports the first.
      
      AND ONE THING THE GATE GETS RIGHT BY CONSTRUCTION, WHICH IS WORTH
      SAYING. Staleness cannot arise. The gate waits on jobs through needs:
      inside one workflow run, so every scan it depends on ran on the same
      commit, in the same run, minutes earlier. A gate that read a stored
      report would have to check how old it was; this one cannot be handed
      yesterday's answer. Recorded as impossible rather than as checked.
      
      THE CELL ASKS FOR PROOF BY DELIBERATE FIXTURE, and a gate asserted to
      block is a gate nobody has watched block -- the same argument S11.17.g
      made about mutation and S11.19.c made about canaries. So there are six
      fixtures, one per condition, each differing from a passing control in
      exactly ONE way, because a fixture that changes two things proves that
      one of them blocks and does not say which.
      
      The run drives the CONTROL first and requires it to pass. If the control
      blocks, every block below is free and proves nothing -- which is its own
      fault kind, passing-fixture-that-blocks. Then all six are driven, and
      each must block FOR ITS OWN CONDITION rather than merely block:
      fixture-that-blocks-for-another-reason is what catches a fixture that
      proves the gate says no without proving it says no for this. All six
      blocked, each for its own reason.
      
      Every reason is reported rather than returning at the first, because a
      release blocked for one reason and one blocked for four are different
      amounts of work, and a gate that stops early sends somebody back three
      times.
      
      Measured: 6 conditions -- 3 invisible to the estate gate (critical
      finding, expired exception, artifact mismatch), 1 impossible by
      construction (stale), 2 genuinely caught (missing, incomplete). 17 spec
      tests; scan PASSES with 11 negative controls, six of them one per
      fixture. -->
      

S11.19 parent rollup — closed by a–g (note only; the checkbox is the S11.19 line above, §2.5.11). <!-- THE SECTION'S SHAPE: THE INSTRUMENTS ARE REAL AND THE GATE IS AN INTERSECTION.

  This estate has a substantial security posture and this section says so
  repeatedly: pnpm audit with a reviewed baseline, OSV recursively, gitleaks
  and trufflehog over full history, CodeQL on the WIDE query suite, syft in
  two SBOM formats, grype, cosign, build-provenance attestation, an install
  script allow-list, and integrity hashes the installer enforces. Most of
  this cell's work was measuring what those instruments reach.

  AND EVERY CELL FOUND THE SAME SHAPE: AN INSTRUMENT THAT IS RIGHT ABOUT
  WHAT IT SEES, AND A GAP NOBODY DECIDED ON. .a: the only scanner that reads
  Cargo, poetry and Go is the one that cannot fail the build, so 6 of 7
  present ecosystems are ungated. .b: an SBOM is a bill of SOFTWARE, and 545
  tracked files carry licence terms with no component to attach them to. .c:
  --only-verified means 4 of 6 secret kinds cannot be reported at all, and
  there is no canary to tell a working scanner from a silent one. .d: one
  language analysed, 1,017 places where a value leaves it. .e: 11 references
  pinned to a commit and every one security-sensitive -- a real policy that
  exists nowhere, so 10 more are on movable names and ONE ACTION IS PINNED
  BOTH WAYS. .f: the raw evidence a finding links to is kept for 14 days.
  .g: the gate reads a job result, not a finding count.

  THE DISCIPLINE. Every claim about the pipeline is PARSED from the workflow
  that makes it -- the required jobs, the query suite, the scanner flags, the
  SBOM formats, the retention periods -- because a register that restates a
  YAML file is a second copy that drifts. And .e derives the pinning policy
  from the references rather than asserting it, which turns an opinion about
  intent into a measurement with exceptions.

  Seven cells, seven scans, all passing with negative controls; the section
  closes with the kit at 311 entry points and 0 violations. -->
  • S11.20 Add a release-readiness generator. Evidence: it refuses to produce a green packet when required evidence is missing or stale. <!-- THE PARENT ASKS FOR AN "IT". Seven sub-cells built the manifest (.a), the resolver (.b), the validator (.c), the freshness recomputation (.d), the truth tables (.e), the emission (.f) and the packet (.g). NONE OF THEM IS THE "IT" -- so the parent was not marked on the strength of its parts. readiness-generator.ts composes them, and composing them is where the pieces stop agreeing.

    THE ALPHABETS DO NOT MEET. .d produces four freshness verdicts. .f names
    five zero-conditions. `unknowable-no-kind-for-a-dependency` IS NOT ANY OF
    THE FIVE. Calling it fresh reports a check nothing performed; calling it
    stale claims a change nobody observed. It is filed as `blocking`, with the
    reason in a named register (UNMAPPABLE_VERDICTS) rather than picked
    quietly in a switch arm -- a mapping that chooses a bucket without saying
    so is how two honest subsystems produce one dishonest answer.
    
    A RESOLUTION FAILURE IS NOT ONE CONDITION EITHER. Five of the resolver's
    six outcomes are not `resolved`, and they map to THREE conditions because
    they need different people: resolves-to-nothing is MISSING and the
    pipeline produces it; digest-does-not-match is FAILED and somebody owns
    it; a mutable name, a cross-class substitution, and an id two requirements
    both claim are decisions, so BLOCKING.
    
    THE ORDER IS RESOLVE, THEN AGE, THEN APPROVE, stopping at the first
    answer. A requirement whose evidence resolves to nothing has no capture to
    be stale, and asking a person to approve something that does not exist
    spends the one input no pipeline can produce.
    
    THE SCAN RUNS THE THING against the parent's own sentence. A baseline that
    IS GREEN FIRST -- a generator that refuses everything satisfies the
    sentence and is useless -- then NINE WITHDRAWALS ONE AT A TIME, each
    required to refuse green FOR ITS OWN REASON, covering 5 of 5 conditions,
    with the human gates compared before and after every one. Plus a
    deterministic seal that is identical across two runs and MOVES when only
    the generator version moves while the packet contents stay byte-identical.
    And all seven sub-cell scans are INVOKED, not trusted: 7/7 green at the
    moment this claim is made.
    
    AND THE SCAN CAUGHT A DEFECT IN THE GENERATOR ITSELF. `failed` was
    unreachable, because the expected digest was read off the ref being
    checked -- both sides of the comparison out of the same box. It showed up
    as one of the five conditions never being raised by any withdrawal. The
    manifest-side digest is now a separate input, and the withdrawal battery
    covers all five.
    
    26 spec tests; scan PASSES with 12 negative controls, three of them
    silence controls (a change after the instant being judged; a change the
    estate says does not invalidate; the green baseline itself). -->
    
    • S11.20.a Define versioned readiness manifest for release/commit/artifact/ environment/target, scope/requirements, dependency graph, required evidence types, owners/approvals, freshness, waivers, and stop policy. Evidence: readiness schema. <!-- A REQUIREMENT SET RESOLVED FOR A RELEASE ID IS A SET RESOLVED FOR A NAME.

      CREDIT, AND THE SUBSTRATE IS MOSTLY THERE. gate-definitions.ts has gate
      types, audiences, per-stage severities, ownership, and a
      RequiredEvidence carrying BOTH maxAgeSeconds and invalidatedBy --
      freshness by age and by dependency change, which is the pair most
      schemes get half of. UNAVAILABLE_BEHAVIOURS has no `pass`, absent as a
      TYPE rather than as a rule, with the reason written down: a gate that
      passes when its evidence is missing passes when the thing it checks is
      broken in exactly the way that hides the evidence. gate-runs.ts gives
      every result a provenance with the target revision, target hash, gate
      version, evaluator and policy. release-locks.ts freezes ten inputs, two
      of them `gates` and `approvals`.
      
      AND THE PER-RELEASE SET ALREADY EXISTS. publish-eligibility.ts has
      RequirementResolution: a releaseId, the destination it was resolved FOR,
      the bindings it was resolved against, the requirements themselves, and
      best of all an `excused` map where every evidence kind deliberately NOT
      required carries the reason why. It even has a fault for a set that was
      resolved for the release and pointed at a destination afterwards.
      
      THIS CELL WAS WRITTEN ON THE ASSUMPTION THAT NO SUCH SET EXISTED, AND
      ITS OWN SCAN REFUTED THAT ON THE FIRST RUN -- the check "is there a
      module holding both a releaseId and a requirement list" returned
      publish-eligibility.ts and failed the run. The claim below is what
      survived, and it is sharper than the one it replaced.
      
      WHAT THE SET IS BOUND TO IS A NAME. RequirementResolution carries a
      releaseId and bindings, and NO commit and NO artifact digest -- checked
      structurally by reading its fields, so the day one is added this cell
      fails rather than going quietly stale. Two builds of the same release id
      resolve the same requirements, and nothing in the set says which bytes
      it was about. That is S11.19.g's artifact-mismatch one level down:
      there, the gate could not tell whether the artifact scanned was the
      artifact released; here, the requirement set cannot say either.
      
      AND resolvedAt IS A TIME, NOT A VERSION. Two resolutions taken a week
      apart differ by an instant, and nothing says whether the REQUIREMENTS
      changed between them. A packet assembled against a shorter list looks
      exactly as complete as one assembled against a longer one -- and the
      older one looks better, because it satisfied less.
      
      SO THIS MANIFEST ADDS FOUR THINGS TO A SHAPE THAT MOSTLY EXISTS: the
      commit and the artifact digest, a VERSION of the requirement set, a
      dependency graph between requirements (topologically ordered, returning
      null on a cycle rather than looping), and a stop policy. Everything else
      is the estate's vocabulary, arranged.
      
      requirement-class-with-no-requirement is a finding on purpose: a release
      may legitimately need no screenshots, and the manifest should have to say
      so out loud, because an absent class and a class nobody thought of look
      identical.
      
      Measured: 9 requirement classes, 4 targets, 3 stop policies, 11 fault
      kinds. 16 spec tests; scan PASSES with 10 negative controls, including
      that a dependency cycle terminates and that a missing class is reported
      BY NAME. -->
      
    • S11.20.b Resolve requirement/ledger IDs to code, tests/scans, migration, screenshots/traces, performance/SLO, runbooks, approvals, and immutable evidence hashes. Evidence: traceability resolver tests. <!-- RESOLVING TO A NAME RESOLVES TO WHATEVER IS THERE NOW.

      A requirement that resolves to evidence/scans/dependency.json is
      satisfied by whatever that path holds at the moment somebody looks. A
      requirement that resolves to a digest is satisfied by one sequence of
      bytes and by nothing else. Both look like a resolved requirement in a
      packet; only the second is still true tomorrow.
      
      This is the same rule S11.19.e found in the workflows -- a tag is a name
      the owner can move -- arriving one level up, which is why the resolution
      targets are distinguished by TYPE rather than by a convention about
      where evidence lives.
      
      MEASURED AGAINST THE ESTATE: 1,238 tracked files under evidence/, and
      ZERO of them named by a digest. Every one would be reached by a path.
      
      A RESOLVER THAT CANNOT FAIL IS A LOOKUP TABLE, so the failures are the
      content: an id that resolves to nothing, an id that resolves to evidence
      of the wrong CLASS, an id two requirements both claim, an id that
      resolves to a mutable name, and an id whose digest does not match what
      the manifest recorded. Each is a different repair and only the first is
      obvious, so the resolver reports the KIND rather than returning null.
      
      AND A CLASS MISMATCH IS THE ONE THAT LOOKS FINE. A requirement for a
      runbook, resolved to a passing test, IS resolved: there is an id, there
      is an artifact, there is a digest, and a packet counting resolved
      requirements counts it. What is missing is that a test does not tell
      anybody what to do at three in the morning. S11.20.e is where one result
      substituting for another becomes a rule; this is where it becomes
      visible.
      
      The claim map is passed IN rather than derived, because whether an id is
      claimed twice is a fact about the WHOLE set and resolving one
      requirement in isolation cannot see it.
      
      Measured: 3 resolution targets, 6 outcomes all exercised by one set, 9
      requirement classes all resolvable. 16 spec tests; scan PASSES with 10
      negative controls. -->
      
    • S11.20.c Validate artifact/commit/version/target/scope match, file/link existence, hash/signature, schema, command/result, complete pages/ matrices, owner, timestamp, expiry, and source provenance. Evidence: validator tests. <!-- A VALIDATOR THAT CHECKS THE RECORD CANNOT CHECK THE WORLD.

      The cell lists twelve checks in one sentence and they are not one kind
      of thing. Some are about the record on its own: does it have an owner,
      does it parse, does the artifact it names match the artifact the
      manifest names. Some are about correspondence with something that has to
      be PRESENT: does the file exist, does the hash match the bytes, does the
      signature verify. And one is about correspondence with something that
      only exists while it happens -- did the recorded command produce the
      recorded result.
      
      Those three fail differently and run in different places.
      
      An INTERNAL check runs anywhere, needs nothing, and is the cheapest
      thing in a release pipeline, so there is no excuse for skipping it. A
      CORRESPONDENCE check needs the artifact on the machine, which is exactly
      what S11.19.c measured is absent when the secret scan runs. A RE-RUN
      check is not a validation at all: it is another execution, and "the
      recorded command produced the recorded result" is only knowable by doing
      it again, at the cost the original run cost. It is reported as unchecked
      even when the validator is permitted to re-run, because the comparison
      belongs to whoever pays for it.
      
      THE RUN DIGESTS A REAL EVIDENCE FILE AND VALIDATES ONE ENTRY TWICE.
      With the world present: 14 of 15 passed, nothing failed, and it may be
      called validated. With the world absent: ALL ELEVEN internal checks
      still passed, one failed, three were unchecked, and it may not.
      
      A packet full of internal passes is a packet that is internally
      consistent. That is worth something, and it is not what "validated"
      sounds like: a record naming a file that was deleted, with an owner and
      a timestamp and a schema that parses, passes every internal check there
      is.
      
      EVERY CHECK RETURNS A RESULT, INCLUDING `unchecked`. A validator that
      omits the checks it could not run produces a report where absence looks
      like success -- the failure this initiative keeps finding, and the same
      distinction S11.18.g had to make about a scan that did not run. A hash
      nobody recorded, a signature nobody attempted, and bytes nobody digested
      are each unchecked rather than passed.
      
      Measured: 15 checks over 3 natures (11 internal, 3 correspondence, 1
      re-run), partitioning exactly. 16 spec tests; scan PASSES with 12
      negative controls, including a real digest mismatch against altered
      bytes. -->
      
    • S11.20.d Recompute freshness from code/contract/schema/dependency/ policy/model/corpus/fixture/environment/standard changes rather than trusting a green label. Evidence: staleness mutation tests. <!-- A LABEL SAYS WHEN SOMEBODY LOOKED. FRESHNESS IS WHETHER ANYTHING MOVED SINCE.

      CREDIT, AND staleness.ts IS ONE OF THE BEST THINGS IN THE KIT. An
      INVALIDATION_REGISTRY of 27 named mutations across 8 dependency kinds,
      each carrying `invalidates` and a REQUIRED `why` on BOTH answers -- 10
      of them saying false. The comment explains that the negative answer
      matters more: "a title correction does not invalidate a rights check" is
      a judgement somebody made, and an unexplained `invalidates: false` is a
      judgement somebody made to stop the queue filling up, and afterwards
      there is no way to tell them apart.
      
      Its doc even names the case that justifies the whole subsystem: a
      provider repoints a stable model name at new weights, nothing in this
      system changed, no revision moved, no editor did anything -- and every
      gate answered by that model is now answering a different question, which
      cannot be noticed from inside a single result.
      
      THE ALPHABET IS SHORTER THAN THE ESTATE. DEPENDENCY_KINDS has eight
      members and the cell names ten input classes. Five map: code→source,
      model→model, policy→policy, environment→target-configuration, and
      dependency is what the registry is about. FIVE HAVE NO KIND AT ALL:
      contract, schema, corpus, fixture, standard.
      
      Each of those five is load-bearing in this repository. A contract is
      GraphViewContract and TimelineContract. A schema is what the validator
      parses against. A corpus is what an evaluator scores against -- and
      ScoreProvenance already carries a corpusId AND a corpusVersion, so the
      estate VERSIONS the thing it cannot register a mutation for. Fixtures
      are the 25 modules S11.18.d walked. And a standard is WCAG 2.2,
      BS.1770-4, C2PA: all three leaned on in this section alone.
      
      A KIND THAT DOES NOT EXIST CANNOT HAVE A MUTATION REGISTERED AGAINST IT,
      so no evidence can be invalidated by one. When WCAG 2.2 becomes 2.3,
      every accessibility gate is answering a question nobody asks any more,
      and the mechanism that exists to notice exactly this cannot express it.
      
      The verdict unknowable-no-kind-for-a-dependency is what keeps the
      recomputation honest. It outranks `fresh` and is outranked by both
      staleness verdicts, because knowing something moved beats not knowing
      whether anything did -- and "fresh" for an unwatchable dependency is a
      report that nothing was watching.
      
      AND A MUTATION TEST IS THE ONLY WAY TO KNOW A RECOMPUTATION WORKS. One
      mutation per input class, each required to flip the verdict for its OWN
      class, over a baseline checked FIRST and required to be fresh -- the
      same guard S11.17.g needed.
      
      (One fault kind, mutant-that-changes-it-for-another-input, was
      UNREACHABLE as first written, because it was inferred from the verdict
      and no input could produce it. That is the defect this initiative keeps
      finding in other people's registers, so it was made meaningful instead
      of deleted: it now compares the mutant's filed input to the input its
      change is about, which catches a badly-built battery and is reachable.)
      
      Measured: 10 input classes, 5 with no kind, 4 verdicts, 27 registered
      mutations. 20 spec tests; scan PASSES with 13 negative controls. -->
      
    • S11.20.e Evaluate required pass/known limitation/waiver/N/A/human approval states and stop-the-line conditions without allowing one result to substitute for another evidence class. Evidence: truth tables. <!-- A WAIVER IS NOT A PASS. AN N/A IS NOT A PASS. A PERSON SAYING YES IS NOT EVIDENCE. AND EACH OF THEM IS EASIER TO OBTAIN THAN THE THING IT REPLACES -- which is why this cell asks for a TABLE and not a rule. A rule saying "do not let one result substitute for another" is advice. A table says, for every state a requirement can be in and every policy the release runs under, exactly what the packet may claim.

      CREDIT, BECAUSE THE ESTATE IS GOOD AT THE STATES. gate-panel.ts counts
      blocks, waived, satisfied and unknown as four separate figures and
      carries a fault named unknown-counted-as-satisfied for exactly the
      collapse this cell fears. gate-runs.ts returns 'waived' rather than
      folding it into a pass. comparisons.ts requires a waiverId before a
      verdict may become waived.
      
      THE GAP IS NOT IN THE STATES. IT IS IN THE CLASSES. Measured off the
      kit: 18 alphabets name evidence, 89 distinct members between them, and 3
      of 153 pairs share anything at all. There is no single evidence
      vocabulary -- there are many. And `RequiredEvidence.evidenceType`, the
      field a gate definition uses to say what it requires, is typed `string`.
      
      Free text, belonging to no alphabet. Any two equal strings match and
      nothing constrains what the strings are. So the rule this cell asks for
      CANNOT BE STATED at the place where requirements are defined; it can
      only be checked afterwards by somebody who already knows which
      vocabulary was meant.
      
      Alongside it: 14 distinct ways to say "we did not check", spread across
      122 of 2,499 exported alphabets, 32 of which put one of them beside a
      passing member in the SAME alphabet. That is where a substitution stops
      being a decision and becomes a matter of which comparison somebody
      happened to write.
      
      THE ANSWER IS TWO TABLES. Six states by three policies, printed by the
      scan: blocking changes the answer for exactly one state (unmet), and
      awaiting-human is the only state no evidence can resolve, because it is
      the absence of an act by a person -- a generator that resolves it has
      forged a signature, and that is a fault kind with a before/after check
      behind it rather than a value check.
      
      And a substitution matrix false off the diagonal except for two
      exceptions written down with their reasons, out of 72 off-diagonal
      pairs, asymmetric (a screenshot may stand in for a trace; a trace may
      not stand in for a screenshot), with NOTHING AT ALL permitted to satisfy
      an approval. An exception nobody stated is indistinguishable from a
      mistake.
      
      Also: advisory-only genuinely disarms a stop-the-line condition, and
      that is reported (stop-the-line-disarmed-by-the-policy) rather than
      silently allowed -- the policy is doing what it says, which is precisely
      why it belongs in the packet.
      
      37 spec tests; scan PASSES with 15 negative controls, four of them
      silence controls. -->
      
    • S11.20.f Emit green only with zero missing/stale/failed/blocking/ unapproved items; otherwise emit exact blockers and remediation while preserving unchecked human gates. Evidence: negative fixtures. <!-- NAMING WHAT IS WRONG AND SAYING WHAT TO DO ABOUT IT ARE DIFFERENT ACTS, AND ONLY ONE OF THEM HAS BEEN DONE AT SCALE.

      Counted off the kit's own source: 754 exported fault and blocker
      alphabets carrying 6,592 distinct kinds between them. This estate is
      extremely good at naming precisely what is wrong. Maps joining a fault
      kind to what somebody should do about it: ONE, with 11 rows.
      capacity-plugins.RESERVATION_FAULT_REMEDIATION. 0.17%.
      
      Widening the key side to anything at all finds a second map --
      provenance-inspector.ACTION_REQUIREMENTS, 80 rows -- and it says what a
      subject REQUIRES rather than what fixes it. So the figure is a fact
      about the estate and not about the regex that found it. (My own
      registers are part of that 6,592 and join none of it. The count includes
      my contribution to the problem.)
      
      CREDIT, AND THE TWO PLACES THAT DO IT ARE THE MODEL. capacity-plugins
      names its remedies change-the-request, wait-or-retarget,
      ask-for-more-quota, ask-somebody-else and nothing-the-requester-can-do
      -- every one says WHO acts, and the last is the honest answer nobody
      writes. authorization maps every disposition to a remediation and
      WITHHOLDS one for a concealed object, because a remediation for a
      concealed object confirms the object.
      
      THE FIVE ZERO-CONDITIONS ARE NOT ONE LIST. The cell names missing,
      stale, failed, blocking and unapproved and treats them as one set of
      counters. Only TWO of them are cleared by running something. Re-running
      a failure reproduces the failure; re-running past a block produces the
      same block; and no pipeline produces an approval.
      
      So "re-run the release pipeline" is right twice out of five and wrong
      specifically on the condition that is actually holding the release --
      which is the sentence somebody reads at three in the morning before
      running the pipeline again.
      
      A remedy therefore carries an ACTOR and an INSTRUCTION. A blocker with
      no remedy is a fault; a remedy addressed to the pipeline for a condition
      the pipeline cannot clear is a fault; and "fix it" / "investigate" /
      "TBD" are instructions that name no act by anybody.
      
      SIX NEGATIVE FIXTURES, one per condition plus a clean control -- because
      five detectors that fire on everything would pass all five of the
      others. All 6 correct, each raising its own condition and no other.
      
      AND THE HUMAN GATES COME OUT AS THEY WENT IN. Preservation is a claim
      about what the emitter did NOT do, so the check holds the BEFORE and
      compares field by field; an unchecked gate becomes an `unapproved`
      blocker rather than being resolved by something that runs.
      
      29 spec tests; scan PASSES with 16 negative controls, four of them
      silence controls. -->
      
    • S11.20.g Generate deterministic signed/checksummed packet and machine/ human summaries from the same evidence, and detect generator drift. Evidence: reproducibility test. <!-- A DIGEST ANSWERS "DID THE BYTES CHANGE". GENERATOR DRIFT IS THE CASE WHERE THE BYTES CHANGED AND THE EVIDENCE DID NOT -- WHICH ONE DIGEST CANNOT SAY.

      Two runs produce different packets. Three completely different things
      could have happened: the evidence moved, the generator was upgraded, or
      the generator is not deterministic. Only the third is a bug; the second
      is a fact somebody needs BEFORE comparing anything; the first is
      ordinary. A single checksum over the packet reports all three
      identically.
      
      So: two digests plus a generator identity, and compareRuns returns
      reproduced / generator-drift / nondeterministic-generator /
      evidence-changed. An evidence change is answered FIRST, because it
      explains any packet difference and nothing about the generator may be
      inferred from it.
      
      MEASURED, AND THE ESTATE IS EXCELLENT AT THE HARD HALF. ZERO clock
      reads, argless date constructions or random calls in executable code
      across 309 kit modules -- measured with comments stripped and a POSITIVE
      CONTROL on the stripper, because a grep that finds nothing and a grep
      that cannot see produce the same output. The 5 files the raw patterns
      match are all doc comments describing the defect. 8 modules canonicalise
      before hashing, and build-golden NAMES its per-run provenance fields
      rather than dropping them quietly, because "the golden does not cover
      this" and "the golden covers everything" are different claims.
      Determinism is not the gap.
      
      THE GAP IS THAT ALMOST NOTHING RECORDS WHAT PRODUCED IT. 69 modules
      carry a digest. 9 name a producer. 5 of 68 do both, once this register
      excludes ITSELF from its own count -- it is one of the six, and a
      register that counts its own family is measuring itself. So a checksum
      in this estate is a fact about bytes with no fact about the code that
      wrote them, and "did the generator drift" has no field to read.
      
      AND THE EXCLUSION LIST IS WHERE A GENERATOR HIDES ITS OWN INSTABILITY,
      so an excluded field that is not actually per-run is a fault, and the
      list actually used is passed in rather than read from the constant.
      
      THE CHECKSUM IS FNV-1a 64-BIT AND IT IS NOT A SIGNATURE. A real, named
      algorithm, cross-checked against sha256 over 84 comparisons of the same
      canonical bytes with 0 disagreements -- so the verdicts are a fact about
      the bytes and not about the checksum. It meets 1 of the 4 things a
      signature is asked for, and NO digest computed by the same process that
      produced the bytes can meet the other three, whatever the algorithm.
      
      That is why "stored under a name that claims a signature" is a fault
      about the field NAME rather than a keyword grep over prose: the one
      description doing the right thing is the one that says "this is not a
      signature", and a prose grep flags exactly that one.
      
      (The seal check is a PROBE, not an inspection: change only the generator
      version and require the value to move. Written as an inspection first,
      it was unreachable -- the same defect this initiative keeps finding in
      other people's registers.)
      
      36 spec tests; scan PASSES with 18 negative controls, five of them
      silence controls. -->
      
  • S11.21 Migrate Tara to the shared kit incrementally. Evidence: each slice preserves public API and behavior. <!-- DELIBERATELY LEFT UNCHECKED, with .a-.g all closed.

    The parent's evidence is "each slice preserves public API and behavior".
    That is a claim about slices that HAVE BEEN PERFORMED. None has. .a-.g
    built the preconditions -- the freeze, the ordering, the compatibility
    model, the dual-write model, the lane matrix, the rollback model and the
    deletion gate -- and every one of them is a statement about what a
    migration would have to satisfy, not a record of one that happened.
    Marking this would claim that slices preserved behaviour when no slice
    ran.
    
    AND THE SUB-CELLS' OWN MEASUREMENTS SAY THE MIGRATION CANNOT PROCEED AS
    SPECIFIED. Four independent readings, each with a positive control:
    
      S11.21.a  3 of 8 aspects cannot be frozen detectably: performance has an
                instrument and no recorded number, accessibility has features
                and no baseline, and no fixtures policy exists at all.
      S11.21.b  1 of 10 slices (branded IDs/envelopes) has NOTHING in Tara to
                migrate and 3 more are thin; the same detectors find all four
                in the kit, so the slice list describes the kit.
      S11.21.d  All 16 hand-written prisma mutations are in two seed scripts
                and there are 0 mutating route handlers, so there is no
                application call path to migrate one at a time.
      S11.21.e  4 of 7 lanes run on no Tara project, and only 2 both run here
                and produce something a before and an after can be held
                against.
    
    WHAT WOULD CLOSE IT. Slices performed against the frozen manifest, with
    each one's before/after comparison recorded on the two lanes that can
    carry one. That is an act on the estate, not an analysis of it, and it is
    the same class of thing S11.2.g was left open for.
    
    This follows the S11.20 precedent in the other direction: there the parent
    asked for an "it" that could be built and verified, so it was built. Here
    the parent asks for work to have been done. -->
    
    • S11.21.a Freeze Tara public package exports, contracts, routes, BFF/UI behavior, persistence/events/jobs, fixtures policy, performance, and accessibility baselines. Evidence: pre-migration manifest. <!-- A FREEZE IS A PROMISE THAT A VALUE WILL NOT CHANGE. YOU CANNOT MAKE THAT PROMISE ABOUT A VALUE NOBODY HAS.

      The cell names eight aspect classes and treats them as one list. They
      are FOUR KINDS, and only one of the four can show that a migration slice
      broke something:
      
        an-artifact-a-program-reads       record, migrate, re-read, compare
        an-instrument-that-records-nothing  measuring is not recording
        a-description-somebody-wrote      cannot fail; a person re-reads it
        nothing-exists
      
      READ OFF THE REAL TARA ESTATE, five of eight are readable artifacts: 31
      package exports across 9 libs, 30 route files, 24 prisma models, 106
      spec files, and exactly ONE file declaring a schema. (One. For 31
      exports. That is the contracts baseline.)
      
      THREE ARE FROZEN IN NAME ONLY.
      
      PERFORMANCE: 4 files measure -- a real WebVitalName union covering FCP,
      LCP, FID, CLS, TTFB, INP -- and no recorded number exists outside a spec
      file. "We can measure it" and "we know what it was" are different
      claims, and only the second survives a migration.
      
      ACCESSIBILITY: 5 files, 15,226 bytes, in the two accessibility
      directories. What is in them is a WebVTT caption validator and a
      SkipLink component. Those are accessibility FEATURES. A feature is not a
      baseline, and the difference is whether anything records how many
      violations there are NOW.
      
      FIXTURES POLICY: the word "fixture" appears in 25 files, 6 of them the
      Playwright test-fixture sense -- NextFixture, WorkerFixtureValue,
      TestFixtureValue. Nothing anywhere states a policy about serving
      substitute data. SAME SPELLING, DIFFERENT WORD. A grep that finds 25
      files, none of which is about the thing being asked for, is the most
      expensive kind of confirmation.
      
      EVERY ABSENCE CLAIM CARRIES A POSITIVE CONTROL. Before reporting that a
      directory holds no accessibility baseline, the run shows the directory
      exists, holds bytes, and that the SAME READER finds the 10 exports that
      ARE there. Otherwise "no baseline" and "bad path" produce one output.
      
      AND compareAspect HAS THREE ANSWERS, not two: unchanged, changed, and
      cannot-tell. A migration reporting "no differences" for an aspect with
      no baseline has reported that nothing was watching.
      
      Two controls run the comparison against the REAL baselines: dropping one
      export and dropping one prisma model each show up as `changed`.
      
      23 spec tests; scan PASSES with 13 negative controls. -->
      
    • S11.21.b Order independently reversible slices for branded IDs/ envelopes, capability/route registry, lifecycle/revisions/gates, persistence/jobs, API/security, provider/trust, bundle/publish, shell/UI, collaboration/taxonomy, and observability/testing. Evidence: migration plan. <!-- AN ORDER IS A LINE. REVERSIBILITY IS A GRAPH. AND THE GRAPH PEOPLE LOOK AT IS THE WRONG ONE.

      "Independently reversible" says that after slice B has landed, A can be
      undone without undoing B. That is false whenever A and B share anything
      -- and an import edge is only ONE way to share. Two slices that never
      import each other but both edit the same file are entangled exactly as
      thoroughly, and that entanglement is invisible to a dependency graph,
      which is the graph a migration plan is normally drawn from.
      
      So a pair is in one of THREE relations: independent, ordered-by-a-
      dependency, entangled-by-a-shared-file.
      
      MEASURED ON THE REAL ESTATE, GOOD NEWS TWICE OVER:
        12 Tara projects, 5 import edges, 5 of 132 ordered project pairs
        => 0 of 45 SLICE pairs ordered by a dependency
        12 of 45 slice pairs share at least one file (33 placements)
        25 of 566 files touched by more than one slice
      
      An import graph would call all 12 of those pairs independent. Worst:
      api-security x shell-ui share 12 files; api-security x bundle-publish 7.
      3 of 10 slices survive as reversible-alone in the cell's own order.
      
      AND THE BAD NEWS: THE SLICE LIST DESCRIBES THE KIT, NOT TARA. The same
      detectors over both estates:
      
        branded-ids-envelopes         0 in Tara,  7 in the kit
        capability-route-registry     1 in Tara, 91 in the kit
        lifecycle-revisions-gates     3 in Tara, 89 in the kit
        provider-trust                2 in Tara, 13 in the kit
      
      One absent, three thin. THE KIT COLUMN IS THE POSITIVE CONTROL: an
      absence measured with a detector that finds nothing anywhere is not an
      absence, it is a broken detector. And shell-ui reading 159 in Tara
      against 2 in the kit shows the reference is not simply the bigger
      estate -- the kit is a contract library with no React in it.
      
      YOU CANNOT ORDER A SLICE THAT HAS NOTHING TO MIGRATE. A plan whose first
      slice is "branded IDs and envelopes" is a plan for INTRODUCING them,
      which is a different job with a different rollback: there is nothing to
      roll back TO. Substance is a state (substantial / thin /
      absent-in-the-subject), not a footnote.
      
      28 spec tests; scan PASSES with 14 negative controls, three of them
      silence controls. -->
      
    • S11.21.c Add shared implementation behind existing Tara public adapters without changing consumer imports/contracts. Evidence: compatibility tests per slice. <!-- AN ADAPTER PRESERVES A SHAPE FOR SOMEBODY, AND A TEST THAT IMPORTS THE IMPLEMENTATION HAS WALKED PAST THE ADAPTER TO CHECK IT.

      Both halves of this cell turn on a question nobody asks first: WHICH
      imports, BY WHOM.
      
      FIRST, WHO CONSUMES. Measured over the WHOLE repository, because
      "nothing imports this" is false the moment one file anywhere does: 31
      published specifiers against 81,743 source files.
      
        outside-the-subject        3   (13 import sites)
        inside-the-subject-only   11
        imported-by-nothing       17
      
      @tara/features (9 outside), @tara/live-class-booking (3),
      @tara/monitoring (1). That is the entire real subject of "without
      changing consumer imports". For the other 17 -- the whole of
      @tara/analytics, all of @tara/api-client, @tara/content/api and the rest
      -- a compatibility test asserts a shape against itself, and passes for
      the same reason a deleted export would.
      
      SECOND, HOW THE TESTS REACH IT. A compatibility test has to travel the
      road the consumer travels.
      
        106 Tara spec files
        2 imports through a public @tara/ specifier
        135 through a relative path into the implementation
        => 1.5%
      
      Only 2 of 106 spec files could fail if an adapter stopped re-exporting
      something. The rest test the implementation; the public specifier could
      drop an export entirely and every one of them would still pass, for the
      same reason it passed before.
      
      THAT IS A COMPARISON WITH BOTH SIDES OUT OF THE SAME BOX, wearing
      different clothes. The point of an adapter is that the public name and
      the implementation MAY DIFFER. A test that reads the implementation has
      removed the only difference it was there to detect.
      
      14 of 31 exports need a compatibility test. 1 is backed by one that
      could fail.
      
      The claims are DERIVED from what the specs actually import, never
      declared -- a declared claim is a promise, and this asks what is true.
      
      23 spec tests; scan PASSES with 15 negative controls, two of them
      silence controls. -->
      
    • S11.21.d Migrate one call path/state/data flow at a time with expected revision/idempotency/events/audit and no dual-write divergence. Evidence: slice integration tests. <!-- A DUAL WRITE IS ONE ACT WRITTEN TWICE, AND FOUR OF THE FIVE THINGS THIS CELL ASKS TO PRESERVE COUNT ACTS.

      The state is the easy one: write the same value to both stores and they
      agree. Revision, idempotency, events and audit are LEDGERS, and a ledger
      records that something HAPPENED. Write once through two paths and:
      
        revision      each path allocates its own, so the revision stops being
                      a fact about the edit and becomes a fact about which
                      path wrote
        idempotency   the key is consumed on each side, so a retry the old
                      path correctly refuses is accepted by the new one and
                      the act happens twice
        events        every consumer receives it twice, and a consumer that is
                      itself idempotent hides this until one that is not
                      is added
        audit         the trail says the person did it twice -- the one record
                      that exists to be read by somebody reconstructing what
                      happened
      
      SO THE STATE CAN CONVERGE PERFECTLY WHILE EVERY LEDGER DIVERGES. A
      dual-write check that compares the two stores row by row and reports no
      differences is telling the truth about 1 of 5 facts, and it is the one
      that was never in danger.
      
      AND DIVERGENCE IS FOUND BY READING BOTH, WHICH IS THE ONE THING A
      CUTOVER NEVER DOES. Serve from the old side, then the new side; neither
      configuration reads both, so neither can see a difference. Detecting it
      needs a third act -- a compare read -- that serves nobody.
      
      (The per-act ledgers are different and worse: a doubled event is visible
      from a SINGLE side, because the duplicate lands on the side you are
      already reading. The reason nobody sees it is that nobody counts.)
      
      CREDIT, BECAUSE THE ESTATE ALREADY KNOWS THIS.
      protection.MIGRATION_PHASES is expand → dual-write → backfill → verify →
      read-switch → contract, with `verify` sitting exactly between the
      backfill and the read switch, and a comment explaining that separating
      the read switch from the backfill keeps the old value available while
      the new path takes production traffic. That `verify` IS the compare
      read. It also carries `backfill-without-dual-write`, the silent one.
      
      The phase list is PARSED out of protection.ts, never restated -- a
      register carrying its own copy of the estate's migration model cannot
      notice the estate changing it.
      
      MEASURED ON TARA: all 16 hand-written prisma mutations are in TWO SEED
      SCRIPTS (prisma/seed.ts 10, prisma/seed-test.ts 6), and there are ZERO
      mutating Next route handlers. So in the subject as it exists there is no
      application call path for this cell to migrate and nothing to
      dual-write. Reported with a POSITIVE CONTROL, because the reader DID
      find writes -- it found them all in the seeds.
      
      (An earlier shell grep of mine said 550 write sites. Those were
      `libs/tara/database/src/generated/client/models` and
      `node_modules/.prisma`: generated type declarations, not call sites. The
      scan excludes both, which is why it says 16.)
      
      24 spec tests; scan PASSES with 17 negative controls, three of them
      silence controls. -->
      
    • S11.21.e Run targeted unit/contract/integration/browser/performance/ accessibility/stub scans and compare behavior after every slice. Evidence: per-slice report. <!-- A SUITE GIVES YOU A VERDICT. THE CELL ASKS YOU TO COMPARE BEHAVIOUR. THOSE ARE NOT THE SAME THING.

      A suite that was green before and is green after has told you the same
      assertions still hold -- which is compatible with every behaviour the
      assertions do not mention having changed. Comparing behaviour needs a
      RECORDED OUTPUT: a golden, a report file, a measured number. 2 of the 7
      lanes (unit, integration) produce only a verdict.
      
      AND A LANE IS A TARGET ON A PROJECT, WHILE A SLICE CUTS ACROSS PROJECTS.
      7 lanes x 12 Tara projects = 84 cells. Measured off every project.json
      and the root package.json:
      
        runs on the project in question      17
        runs but on another subject          33
        right name, unreadable target        12
        does not exist                       22
      
        unit           12 of 12
        contract        0 of 12
        integration     0 of 12
        browser         2 of 12
        performance     3 of 12
        accessibility   0 of 12
        stub-scan       0 of 12
      
      FOUR OF SEVEN LANES RUN ON NO TARA PROJECT: contract, integration,
      accessibility, stub-scan.
      
      A LANE THAT EXISTS SOMEWHERE ELSE IS NOT A LANE FOR THIS SUBJECT -- the
      same shape as the fixtures trap in S11.21.a. The repository HAS
      `verify:operations:v1-a11y-findings`, `…-a11y-waivers` and
      `…-axe-policy`. Three accessibility scripts, all real, all green, and
      every one points at `libs/oshun/workbench-kit/tools`. They scan the kit.
      Running them after a Tara slice reports on a codebase the slice did not
      touch.
      
      SO COVERAGE HAS FOUR ANSWERS, and the middle two are the point.
      `runs-but-on-another-subject` is what would otherwise be counted as
      coverage. `unknown-what-it-runs-against` is for a script with the right
      name whose target could not be read from its command line --
      `contracts:check` is one. Folding that into "runs elsewhere" would
      assert something nobody measured, which is the defect this register
      exists to catch, one level up. (I wrote it the wrong way first and the
      scan's own output -- "points at null" -- is what showed it.)
      
      THE OVERLAP THAT DECIDES THE CELL: only 2 of 7 lanes both RUN on a Tara
      project AND produce something a before and an after can be held against
      -- browser and performance.
      
      25 spec tests; scan PASSES with 16 negative controls, three of them
      silence controls. -->
      
    • S11.21.f Define per-slice enablement/rollback, data compatibility/ reconciliation, old/new mixed deployment, and removal criteria. Evidence: rollback tests. <!-- A ROLLBACK GOES BACKWARDS IN CODE AND FORWARDS IN DATA.

      Deploying the previous build returns the code to what it was. It does
      NOT un-write the rows the new build wrote while it was live, and it does
      not un-tell the consumers that adopted the new shape. "Rollback" names
      four different acts and only two of them are an undo:
      
        code            returns-to-the-previous-value
        configuration   returns-to-the-previous-value
        data            MOVES FORWARD ONLY
        consumers       depends-on-somebody-else
      
      SO THE ROLLBACK TEST FOR DATA IS NOT "UNDO IT". It is: can the PREVIOUS
      code read what the new code wrote, without error and without loss. A
      compatibility property, tested by running the old reader over new rows.
      
      And consumers are outside this repository: their deploy is their
      decision and their timetable, so nothing here reverts them.
      
      THE ESTATE MAKES THIS CONCRETE. Tara's migration history is ONE
      directory -- 20260506000000_initial, 803 lines, 24 CREATE TABLE, 28,081
      bytes. Repository-wide: 189 forward .sql steps and ZERO reverse ones.
      Prisma migrate is forward-only. There is no previous schema to return to
      and no tool step that would return to it.
      
      (Reported with a POSITIVE CONTROL: the same reader found the 189 forward
      steps, so "no reverse step" is a statement about the tool and not about
      a reader that cannot see SQL.)
      
      A MIXED DEPLOYMENT IS TWO CLAIMS, NOT ONE. While both versions serve:
      old code reads rows the new code wrote, AND new code reads rows the old
      code wrote. A plan that tests one direction has tested the half that
      gets tested BY ACCIDENT -- every row that existed before the deploy was
      written by the old code. The other half is the one a rollback depends on
      and the one nobody sets up, because writing with the new code is what
      you were doing anyway.
      
      REMOVAL IS A CLAIM ABOUT TIME. The old path may go when nothing can roll
      back to it, which needs a window that has ELAPSED, not a window that
      exists -- the same distinction S10.9 needed when a grace period was
      anchored to the event it was supposed to outlive.
      
      24 spec tests; scan PASSES with 15 negative controls, three of them
      silence controls. -->
      
    • S11.21.g Migrate external consumers and delete legacy implementation/ adapters/flags only after import/usage graph and full suite prove no dependency. Evidence: deletion gate. <!-- AN IMPORT GRAPH ANSWERS "WHO IMPORTS THIS BY NAME". A DELETION ASKS "WHAT BREAKS". THOSE DIFFER BY EVERY REFERENCE THAT IS NOT A STATIC IMPORT.

      Seven reference kinds; four break, two go stale, one is unknowable:
      
        a-static-import                            breaks-at-build
        a-manifest-or-config-entry                 breaks-at-build
        a-path-mapping                             breaks-at-build
        a-dynamic-import-with-a-literal            breaks-at-runtime
        a-doc-or-a-string                          goes-stale
        a-build-output                             goes-stale
        a-dynamic-import-with-a-computed-specifier UNKNOWABLE
      
      A DOC TAG GOES STALE AND A MANIFEST ENTRY BREAKS THE BUILD, and both
      look identical to a grep over text. Every mention of @tara/content/api
      in this repository is an `@module` JSDoc tag; @tara/api-client appears
      inside a `describe()` label. A graph that counted text would call both
      load-bearing; one that counts `from` clauses calls both free. Neither is
      wrong and only one is answering the deletion question. (That also
      CONFIRMS S11.21.c: nothing imports either.)
      
      AND WHEN A SPECIFIER IS ASSEMBLED AT RUNTIME, NO STATIC GRAPH CAN ANSWER
      AT ALL. 243 such sites beside 2,671 literal ones across 81,755 files.
      Real examples: `import(`./messages/${locale}/tara.json`)` -- a Tara
      resource loaded by a name that does not exist until the locale does --
      and a verified `return import(moduleName)`.
      
      BUT ONLY THE ONES THAT COULD REACH THE SUBJECT COUNT: 225 of the 243. A
      messages path cannot resolve to a bare package specifier, and a gate
      blocking on it would block every deletion in the repository forever,
      which is a gate that has STOPPED ANSWERING rather than one being
      careful. The run compares each site's STATIC PREFIX to the subject; a
      bare identifier has no prefix and always counts.
      
      AND A GATE IS ONLY AS WIDE AS WHAT IT SEARCHED. "No dependency" is a
      claim about a repository. A claim over 5 directories from a search over
      1 is refused BEFORE anything else -- it is a different question, not a
      weaker answer.
      
      Two of the four verdicts are unknowable and both outrank `safe`: a gate
      blind to a class of reference that says safe has reported that nothing
      was looking.
      
      (The reader strips comments and requires expression position. Its first
      version counted a function signature and a doc comment as dynamic
      imports, which is what made the figure move from 476 to 243.)
      
      23 spec tests; scan PASSES with 20 negative controls, four of them
      silence controls. -->
      
  • S11.22 Run Tara's complete relevant unit, integration, contract, build, accessibility, and Playwright suite after migration. Evidence: reports are linked and green. <!-- DELIBERATELY LEFT UNCHECKED, with .a-.h all closed. Two independent reasons, either of which is sufficient.

    FIRST, "AFTER MIGRATION". No migration happened -- see the S11.21 parent,
    which is left open for the same reason. A suite run before the thing it is
    meant to verify has not verified it.
    
    SECOND, "REPORTS ARE LINKED AND GREEN" IS FALSIFIED BY .a-.h THEMSELVES.
    Each measured what the corresponding report could honestly say, and the
    answers were:
    
      .a  10 of 12 verification inputs can be pinned. The clock cannot: 491
          direct platform reads across 113 of 566 files against 5 files with
          a seam. Fixtures cannot: no policy switch exists.
      .b  0 of 106 spec files use a property-based generator; 72 are
          example-only. Credit: 2,866 cases, 4,270 assertions, 0 asserting
          nothing.
      .c  0 of 11 integration subjects are covered by a test that would FAIL
          if the dependency were absent. 0 of 106 spec files carry
          live-connection machinery.
      .d  0 of 8 contract artefacts bind both parties, so 0 of 10 suites have
          one. The generated client declares 14 paths; the provider carries
          16 route modules.
      .e  3 of 9 build gates can stop a change. 4 run on another subject.
      .f  4 of 8 accessibility rows are covered; 2 need an instrument that is
          not present on this machine at all.
      .g  1 of 3 browser artefacts describes the failing execution, in either
          environment. Fixture mode is declared nowhere.
      .h  a rerun that passes is the same observation for a fix and a flake,
          and 2 of 20 combinations are established by one.
    
    A report that is linked and green is producible from this estate today.
    It would not be false in any single line and it would not mean what the
    cell is asking it to mean, which is the failure every one of .a-.h is
    written against.
    
    WHAT WOULD CLOSE IT. The migration performed (S11.21), then the suites run
    against it with the gaps above closed or explicitly excepted -- an
    integration lane that fails without its dependency, a contract artefact
    both sides read, and gates that can stop a change. That is work on the
    estate, not analysis of it. -->
    
    • S11.22.a Pin migrated commit, dependencies, environment, database/ storage/queue/provider modes, fixtures, accounts/tenants, browser, clock, and exact commands. Evidence: verification manifest. <!-- A PIN IS A PROMISE THAT A VALUE WILL BE THE SAME NEXT TIME, AND THAT PROMISE NEEDS SOMEWHERE TO STAND.

      A commit pins itself -- git records it and nothing downstream can
      disagree. Dependencies pin through a lockfile the installer reads. A
      browser pins through a version a tool honours. Those are values somebody
      writes down and a tool then obeys.
      
      A CLOCK IS NOT LIKE THAT. Writing "clock: 2026-08-16T00:00:00Z" pins
      nothing unless every read of the current time goes through something
      that can be GIVEN that value. Code calling the platform directly reaches
      past any injection point, and the manifest line becomes a note about
      what somebody intended.
      
      6 of the 12 inputs need a seam in the code: clock, fixtures, and the
      database/storage/queue/provider modes.
      
      MEASURED WITH THE SAME DETECTOR ON TWO ESTATES, comments and quoted
      strings stripped:
      
                            Tara     kit
        Date.now()           103       0
        new Date()           358       0
        Math.random()         27       0
        performance.now()      3       0
        TOTAL                491       0
        files with >=1    113/566   0/331
      
      Five Tara files carry anything resembling an injectable clock seam,
      against 113 that go straight to the platform. One injection point beside
      hundreds of direct calls is not a seam with exceptions -- it is hundreds
      of code paths and a decoration. A fixtures policy switch: 0 files, which
      S11.21.a found by a different route.
      
      SO THE VERDICT ALPHABET HAS THREE MEMBERS. A manifest for this subject
      can honestly call 10 of 12 pinned; clock is recorded-but-not-enforced
      and fixtures is unpinnable-no-seam-exists. A manifest claiming all
      twelve produces three findings.
      
      (THE DETECTOR STRIPS QUOTED STRINGS AS WELL AS COMMENTS. Its first
      version counted one hit in the kit: a register whose prose reads "a
      constant, never `new Date()`" -- a register of source text tripping a
      detector that reads source text, which is a trap this initiative has hit
      before. Template literals are deliberately left alone, because an
      interpolated clock read executes; both directions are negative
      controls.)
      
      22 spec tests; scan PASSES with 16 negative controls, two of them
      silence controls. -->
      
    • S11.22.b Run all affected Tara/shared unit/property/known-value/ component tests with coverage mapping. Evidence: unit report. <!-- COVERAGE IS A CLAIM ABOUT WHICH LINES RAN. The cell names four kinds because each finds a class of defect the others cannot, and NO LINE NUMBER DISTINGUISHES THEM.

        example      the case somebody thought of, exactly as thought of
        property     a case nobody thought of, within reach of its generator,
                     shrunk to say which part mattered
        known-value  a disagreement with a truth from OUTSIDE this repository
                     -- the only kind that catches the whole system being
                     CONSISTENTLY wrong
        component    what renders; cannot tell a right value from a wrong one
      
      Run one example over a function and every line is covered. Run a
      thousand and the number does not move.
      
      CREDIT WHERE THE MEASUREMENT FINDS IT. Tara has 2,866 test cases and
      4,270 assertions across 106 spec files, at 1.49 assertions per case, and
      ZERO files declare a case and assert nothing. The classic way a coverage
      number rises without anybody checking anything is absent here.
      
      AND THE KINDS ARE NOT EVENLY REPRESENTED:
      
                      Tara     kit
        example         72     319
        property         0       1
        known-value      1       1
        component       33       3
        spec files     106     324
      
      NONE of Tara's 106 spec files uses a property-based generator. The kit
      column is the control: the same detector finds one there, so the zero is
      a fact about the subject rather than about a detector that finds nothing
      anywhere.
      
      A suite made entirely of examples has line coverage and NO input-space
      coverage, and that is invisible in every report this cell asks for.
      Coverage is blind to four things it is routinely read as answering,
      including which of the four kinds ran -- all four move the same number.
      
      17 spec tests; scan PASSES with 12 negative controls, three of them
      silence controls. -->
      
    • S11.22.c Run real database/storage/queue/provider/publisher/host integration, migration, tenant, fault, replay, and restore tests. Evidence: integration report. <!-- A TEST IS A REAL INTEGRATION TEST IF IT FAILS WHEN THE DEPENDENCY IS ABSENT. That is the whole definition, and it is the one nobody checks.

      "Real" is doing all the work in this cell's sentence. A suite that
      passes with nothing running has told you about the code and nothing
      about the integration, and the only way to know which one you have is to
      take the dependency away.
      
      FOUR RELATIONS, NOT TWO:
      
        connects-to-a-real-one  fails without it. This is integration.
        uses-a-double           passes either way -- CORRECT for a unit test,
                                and not integration.
        skips-when-absent       REPORTS GREEN. Nothing ran.
        never-touches-it        the code does not reach it at all.
      
      `skips-when-absent` is the one that turns an honest arrangement into a
      false report. Skipping when the database is down is sensible on a
      developer's machine and catastrophic for a release gate to count,
      because a skipped test and a passing test are the same colour in every
      summary -- the same defect as an aggregate that reports success when a
      status is merely not a failure.
      
      MEASURED:
      
        Tara spec files carrying live-connection machinery:   0 of 106
        kit  spec files carrying it, same detector:           2 of 325
      
        subject      specs  live   relation
        database         2     0   uses-a-double
        storage          5     0   uses-a-double
        queue            1     0   uses-a-double
        provider        11     0   uses-a-double
        publisher        1     0   uses-a-double
        host             0     0   never-touches-it
        migration        1     0   uses-a-double
        tenant           0     0   never-touches-it
        fault           21     0   uses-a-double
        replay           0     0   never-touches-it
        restore          1     0   uses-a-double
      
      ZERO of 11 subjects are covered by a test that would fail if the
      dependency were absent. Three -- host, tenant, replay -- are not
      mentioned by any spec at all. A report claiming all eleven produces 14
      findings.
      
      The kit column is the control: the same harness detector finds two files
      there, so the Tara zero is a fact about Tara.
      
      provesIntegration and failsWhenAbsent are deliberately the SAME
      predicate. The property that makes a test an integration test IS failing
      without the thing, and two functions that happen to agree would invite
      one of them to drift.
      
      19 spec tests; scan PASSES with 13 negative controls, one of them a
      silence control. -->
      
    • S11.22.d Run provider/consumer/schema/generated/API/event/job/bundle/ compatibility/drift contract suites. Evidence: contract report. <!-- A CONTRACT IS BETWEEN TWO PARTIES. A CONTRACT TEST THAT ONLY ONE PARTY RUNS IS A UNIT TEST WITH A LONGER NAME.

      A provider suite asserting the server's own schema passes whatever the
      client expects. A consumer suite asserting a generated client passes
      whatever the server sends. Both are green, both are honest about what
      they check, and between them they establish nothing about whether the
      two agree.
      
      FOUR STANDINGS, AND THE FOURTH IS THE ONE THIS ESTATE IS IN:
      
        shared-by-both-parties                  a real contract
        read-by-one-party                       a unit test, longer name
        read-by-nobody                          it can say anything
        generated-from-a-source-that-is-absent
      
      THAT LAST ONE IS THE INTERESTING FAILURE. A generated client is a
      snapshot of a specification at the moment somebody ran the generator. If
      the specification is not in the repository, the snapshot CANNOT BE
      REGENERATED -- so the difference between it and the current server
      cannot be computed, and "drift" is the tenth suite this cell names. The
      artefact that exists to detect drift is the one drift is invisible in.
      
      MEASURED:
      
        openapi specification documents in the repository:  21
        belonging to the subject:                            0
      
      Stated beside the reader that found the 21, because an absence measured
      by a reader that finds nothing anywhere is not an absence.
      
      THE CHEAPEST DEMONSTRATION: the generated client declares 14 paths and
      the provider carries 16 route modules. THEY DISAGREE. Two counts of one
      surface differing is the least you can do to show that no shared
      document is being enforced -- a count, deliberately, not a shape check.
      
      4 of 5 JSON schema documents (collection, course, sound, teacher) are
      mentioned by no source file anywhere in the subject, its tools or its
      scripts.
      
      0 of 8 artefacts bind both parties, so 0 of 10 suites have one. A report
      claiming all ten produces 16 findings.
      
      18 spec tests; scan PASSES with 13 negative controls, two of them
      silence controls. -->
      
    • S11.22.e Run lint/typecheck/package/bundle/production build and dependency boundary/test-fixture/stub/secret scans sequentially. Evidence: build/static reports. <!-- A GATE IS A THING THAT CAN STOP YOU. THERE ARE THREE WAYS IT CANNOT, AND ONLY ONE OF THEM LOOKS LIKE AN ABSENCE.

        does-not-exist            nothing to run
        runs-on-another-subject   green every time, about a codebase this
                                  change did not touch
        has-a-documented-bypass   it stops you unless the variable is set,
                                  and the failure message names the variable
      
      MEASURED ON THIS REPOSITORY (323 root scripts, 3 husky hooks):
      
        lint                  can-stop-you             lint
        typecheck             has-a-documented-bypass  SKIP_TYPECHECK=1, 52x
        package               does-not-exist           (nothing)
        bundle                runs-on-another-subject  -> scripts/v3/...
        production-build      can-stop-you             build
        dependency-boundary   runs-on-another-subject  -> workbench-kit/tools
        test-fixture          runs-on-another-subject  -> workbench-kit/tools
        stub-scan             can-stop-you             .husky/pre-commit
        secret-scan           runs-on-another-subject  -> workbench-kit/tools
      
      3 OF 9 CAN STOP A CHANGE. Three of the four that run elsewhere resolve
      to workbench-kit registers -- my own -- rather than to the thing being
      verified. The same trap as S11.21.e's accessibility scripts.
      
      AND THE BYPASS IS NOT HYPOTHETICAL. SKIP_TYPECHECK is named in 52 lines
      of the last 200 commit messages. This session took it on every commit,
      with the reason stated in each body. A gate that prints its own escape
      hatch when it fails has told you it is advisory; that is the honest kind
      of hole and it is still a hole.
      
      "SEQUENTIALLY" IS THE OTHER HALF OF THE CELL. Nine gates in order means
      the first failure stops the rest. A run that stops at gate two has said
      NOTHING about gates three to nine, and listing them as "not failed"
      counts seven absences as seven passes -- the skip-as-a-pass defect
      again. Three outcomes, exactly one green:
      never-ran-because-an-earlier-gate-stopped is the answer to a question
      nobody asked yet.
      
      (THE BYPASS IS ATTRIBUTED BY PROXIMITY, not co-residence. The scan's
      first run blamed SKIP_TYPECHECK for the stub scan: both live in
      .husky/pre-commit and only the typecheck is guarded. The stub scan has
      no bypass at all -- it blocked two commits this session and both were
      fixed by rewording, not by setting a variable.)
      
      19 spec tests; scan PASSES with 14 negative controls, three of them
      silence controls. -->
      
    • S11.22.f Run automated/manual/keyboard/screen-reader/semantic parity/ visual/zoom/locale accessibility matrix. Evidence: accessibility report. <!-- AN ACCESSIBILITY MATRIX HAS TWO KINDS OF ROW. A MACHINE CAN FILL ONE KIND, AND THE OTHER KIND IS WHERE THE THINGS AUTOMATION CANNOT SEE LIVE.

      5 of 8 are a machine's job. manual, screen-reader and semantic-parity
      need a person operating an assistive technology, and they are NOT the
      easy rows left over -- a control with a perfect accessible name that
      announces at the wrong moment passes every automated check ever written,
      and semantic parity compares what a person HEARS to what a person SEES,
      which is not a property of a tree.
      
      AND A ROW HAS A FURTHER STANDING: whether the instrument it needs is
      present WHERE THE RUN HAPPENS.
      
        INSTRUMENTS on this machine
          a screen reader:   ABSENT (orca/nvda/voiceover)
          at-spi2 bridge:    present
          an axe runner:     @axe-core/playwright
          a browser driver:  @playwright/test
      
        row              filler                   files   standing
        automated        a-machine                    0   fillable-and-not-filled
        manual           a-person...                  0   needs-a-person
        keyboard         a-machine                   39   filled
        screen-reader    a-person...                 17   instrument-absent-on-this-machine
        semantic-parity  a-person...                 90   instrument-absent-on-this-machine
        visual           a-machine                   34   filled
        zoom             a-machine                    5   filled
        locale           a-machine                   76   filled
      
      TWO ABSENCES, TWO COMPLETELY DIFFERENT MEANINGS. @axe-core/playwright is
      installed and 38 files in this repository use it, ZERO of them in the
      subject -- so an unfilled automated row is a statement about the
      SUBJECT. No screen reader is installed anywhere here -- so an unfilled
      screen-reader row is a statement about THIS MACHINE. A matrix that
      reports them the same way has told a reader nothing. (This is also why
      earlier sessions refused S11.5.f.)
      
      4 of 8 rows are covered. A matrix claiming all eight and calling itself
      complete produces 5 findings.
      
      (The automated row counts IMPORTS of the runner, not occurrences of the
      word. The first run called two files that merely mention "axe" filled;
      mentioning a tool and running it are different.)
      
      21 spec tests; scan PASSES with 14 negative controls, two of them
      silence controls. -->
      
    • S11.22.g Run all critical live-data and fault-state Playwright journeys with fixture mode explicit. Evidence: browser report/traces/screenshots. <!-- AN ARTEFACT OF THE RETRY IS NOT AN ARTEFACT OF THE FAILURE.

      Whether a failing browser run can be investigated depends on TWO
      settings that are read separately and mean something only together.
      PARSED from apps/tara/web/playwright.config.ts:
      
        retries: process.env.CI ? 2 : 0
        screenshot: only-on-failure
        trace: on-first-retry
        video: on-first-retry
      
      LOCALLY, retries is 0. There is never a first retry, so NO TRACE AND NO
      VIDEO ARE PRODUCED AT ALL. Both lines are individually sensible and the
      combination is a setting that does nothing. 1 of 3 artefacts describes
      the failure.
      
      IN CI, retries is 2, so both ARE produced -- from a DIFFERENT EXECUTION
      than the screenshot. The screenshot is of the run that failed; the trace
      is of a later run that may have passed. If the failure was a race, the
      trace is of the case that WORKED. One bundle, two events, presented as
      one. Still 1 of 3.
      
      So the answer has THREE members, and `a-different-execution` is the
      expensive one, because it is the one that LOOKS like evidence.
      
      AND "FIXTURE MODE EXPLICIT" IS A CLAIM ABOUT WHERE THE DECLARATION IS. A
      journey whose title says "live data" has declared it to a READER; the
      substitute is installed on the other side of the glass -- a server, a
      route handler, an environment value -- and a title cannot reach it. Only
      a declaration made where the fixture is INSTALLED can be wrong in a way
      that fails.
      
      Of the subject's 11 e2e files: 0 declare it where it is installed, 0
      where it is read. `not-declared`.
      
      This reproduces S11.2.f's finding independently, in a different
      configuration file and a different app.
      
      19 spec tests; scan PASSES with 15 negative controls, one of them a
      silence control. -->
      
    • S11.22.h Triage every failure, rerun affected/full gates, and link final current green reports with no stale result. Evidence: zero open failures. <!-- A RERUN THAT PASSES IS THE SAME OBSERVATION FOR A FIX AND FOR A FLAKE. WHAT TELLS THEM APART IS WHETHER ANYTHING CHANGED IN BETWEEN.

      A test fails, somebody looks at it, does something, runs it again, and
      it passes. Two completely different stories fit that: the change fixed
      it, or it was intermittent and this time went the other way. The passing
      run is identical in both.
      
        disposition             changed  passed  answer
        fixed                   true     true    established
        fixed                   false    true    indistinguishable-from-a-flake
        a-flake                 true     true    indistinguishable-from-a-flake
        a-flake                 false    true    established
        an-environment-problem   any     true    a-claim-a-rerun-cannot-support
        a-known-limitation       any     true    a-claim-a-rerun-cannot-support
      
      2 of 20 combinations are established by a rerun. An environment problem
      and a known limitation are claims about the world OUTSIDE the test and
      are settled by running it again in no combination at all.
      
      The three-member answer is what stops the overreach in both directions:
      "fixed" and "it was a flake" are equally unearned when nothing changed.
      
      AND CALLING SOMETHING A FLAKE IS A DECISION, NOT AN OBSERVATION. It says
      the failure was not about the code -- a claim about every future run,
      made from one. So it owes a count of how often it has been seen and
      something a person can look at. A flake with neither is a way of closing
      a ticket.
      
      "NO STALE RESULT" IS S11.20.d WEARING DIFFERENT CLOTHES. The scan drives
      it with the REAL HEAD and a REAL earlier commit rather than two invented
      strings:
      
        unit          green  from HEAD          current
        integration   green  from HEAD~5        stale-from-another-commit
        browser       green  from (none)        no-commit-recorded
      
      ALL THREE ARE GREEN and one describes the tree being released. A URL to
      a green report looks the same whichever commit produced it -- and a
      green report is the one nobody re-reads.
      
      A report claiming ZERO OPEN FAILURES over a realistic set produces 7
      findings.
      
      21 spec tests; scan PASSES with 13 negative controls, three of them
      silence controls. -->
      
  • S11.23 Compare Tara pre/post performance and accessibility baselines. Evidence: regressions are fixed or explicitly approved. <!-- DELIBERATELY LEFT UNCHECKED, with .a-.g all closed. Same shape as the S11.21 and S11.22 parents, and for a reason .a and .c each measured.

    THERE IS NO "PRE". A comparison needs two sides.
    
      S11.21.a  the accessibility baseline does not exist -- what the subject
                holds is accessibility FEATURES (a caption validator, a
                SkipLink) and nothing recording how many violations there are
                now. The performance baseline is an INSTRUMENT with no
                recorded number outside a spec file.
      S11.23.c  2 of 566 subject files record a violation count or a WCAG
                level; 2 of the 9 diff subjects need a screen reader that is
                not present on this machine at all.
    
    And there is no "post", because no migration ran (S11.21).
    
    THE PARENT'S OWN EVIDENCE PRESUPPOSES THE COMPARISON. "Regressions are
    fixed or explicitly approved" is a statement about regressions that were
    FOUND, which needs a comparison that RAN, which needs two baselines. Each
    link in that chain is missing and .a-.g say which.
    
    WHAT .a-.g DID ESTABLISH is what the comparison would have to satisfy, and
    several of those are findings in their own right: 2 of 10 equivalence
    inputs can never be equal and need a bound rather than a checkmark; at the
    3 runs this estate collects the smallest detectable difference is 4.303
    standard errors, so most differences are unreadable; a finding count going
    down is ambiguous without a denominator; only 1 of 6 delta classes is
    settled by the measurement; three named approvals are not one; and a
    published baseline moves past any unfixed regression it contains.
    
    WHAT WOULD CLOSE IT. A recorded accessibility baseline and a recorded
    performance number, taken before a migration that has happened, compared
    after it with the machinery .a-.g specifies. Two of those are acts on the
    estate and one is an act on this machine. -->
    
    • S11.23.a Verify pre/post builds use equivalent environment/data/accounts/ browser/device/network/clock/providers/routes/states and record unavoidable differences. Evidence: comparison manifest. <!-- "EQUIVALENT" IS A CLAIM ABOUT TWO RUNS. SOME OF THE TEN THINGS THIS CELL NAMES CANNOT BE EQUAL BETWEEN TWO RUNS AT ALL, AND FOR THOSE THE MANIFEST HAS TO RECORD A BOUND RATHER THAN AN ASSERTION.

        equal-by-construction        routes, states -- one tree, one value
        equal-if-pinned              environment, data, accounts, browser,
                                     device, providers
        never-equal-only-boundable   network, clock
      
      CALLING THE CLOCK AN "UNAVOIDABLE DIFFERENCE" AND RECORDING IT IS THE
      RIGHT SHAPE AND THE WRONG EMPHASIS. Two runs happen at two different
      instants BY DEFINITION. It is not an exception that arose, it is the
      guaranteed case, and a manifest listing it beside a genuinely unexpected
      difference has flattened the two.
      
      A manifest asserting all ten equal produces exactly 2 findings. An
      honest one, with bounds where a bound is the only honest record,
      produces 0.
      
      AND A BOUND WITHOUT A MEASUREMENT IS A WISH. A number, an observation
      and a unit are all required: a bound with no observation is a rule
      nobody checked, an observation with no bound is a number nobody can
      judge, and "network: comparable" is a checkmark wearing a bound's
      clothes.
      
      THE CLOCK IS RE-MEASURED HERE, not cited from S11.22.a: 461 direct
      platform reads across 110 of 566 subject files, against 0 across 339 kit
      modules with the same detector. So the clock is BOTH unpinnable in this
      subject AND guaranteed to differ. A manifest asserting clock equivalence
      would be wrong twice.
      
      20 spec tests; scan PASSES with 14 negative controls, three of them
      silence controls. (One control was mis-written first: passing a single
      entry also reports the nine that are missing, which is correct and is
      not what the control is about.) -->
      
    • S11.23.b Compare route load/transfer/request/main-thread/memory/layout, interaction/save/search/compare/virtualization, API/job/queue/provider, and cost/resource metrics statistically. Evidence: performance diff. <!-- TWO NUMBERS ARE NOT A DISTRIBUTION. One before and one after is a SUBTRACTION, and the cell asks for a statistical comparison.

      The difference between two runs is the change PLUS the noise, and
      nothing in the two numbers separates them. What separates them is the
      SPREAD. A difference smaller than the spread is not a small regression
      -- it is a measurement of the noise.
      
      So this computes the smallest difference the samples could distinguish:
      Welch's standard error sqrt(sb²/nb + sa²/na) times a two-sided 95%
      Student's t critical value, and refuses to call anything below it a
      regression.
      
      THE t VALUES MATTER AT THE n THIS ESTATE USES. apps/tara/web/
      lighthouserc.js collects `numberOfRuns: 3`. Welch's df are never below
      min(n)-1, so this takes 2, where t is 4.303 against a normal 1.96 -- AN
      INTERVAL MORE THAN TWICE AS WIDE. Using the normal value would call
      noise significant, which is the error this cell exists to prevent.
      
        n per side   df   t        smallest detectable (sd=10ms)
        1            -    -        nothing -- too few samples
        2            1    12.706   179.69ms
        3            2    4.303     40.57ms
        5            4    2.776     19.23ms
        10           9    2.262     10.66ms
        30          29    2.060      5.41ms
        100         99    1.960      2.79ms
      
      TWO COMPARISONS AT n=3, BOTH MOVING THEIR MEANS BY EXACTLY 70ms:
      
        noisy       sd 60/60   smallest detectable 210.8ms  =>
                               indistinguishable-from-noise
        repeatable  sd  2/2    smallest detectable   7.0ms  =>  regression
      
      The same difference is a regression in one and unreadable in the other.
      The number that decides is the spread, which a two-number comparison
      does not have.
      
      FOUR VERDICTS, AND TWO ARE NOT "NO CHANGE".
      `indistinguishable-from-noise` says the difference was smaller than this
      experiment could see; `too-few-samples-to-say` says the experiment could
      see nothing. A diff reporting both as "no regression" has told a reader
      a check happened.
      
      A DOCUMENTED APPROXIMATION, NAMED: the interval assumes the sample means
      are approximately normal, which at n=3 is an assumption rather than a
      fact. It is computable from what the estate collects; a bootstrap would
      need samples nobody gathers.
      
      (THE SCAN CAUGHT MY OWN FIXTURE ERROR: the first pair moved 35ms and
      70ms, so the demonstration did not demonstrate what it claimed.)
      
      27 spec tests; scan PASSES with 15 negative controls. -->
      
    • S11.23.c Compare automated WCAG findings, keyboard journeys, screen- reader announcements, focus, semantic alternatives, zoom/reflow, themes/densities, RTL/locales, and visual baselines. Evidence: accessibility diff. <!-- A FINDING COUNT IS A NUMERATOR. WITHOUT THE DENOMINATOR, A COUNT GOING DOWN IS AMBIGUOUS BETWEEN "FEWER PROBLEMS" AND "LESS WAS EXAMINED".

      Before: 40. After: 12. That looks like a very good change, and it is
      equally consistent with a page that now fails to load, a route removed
      from the crawl, a component that renders nothing, or a scan that timed
      out. Every one of those reduces the count and none of them fixes
      anything.
      
        40 -> 12, same examination   =>  fewer-problems  (an improvement)
        40 -> 12, more examined      =>  fewer-problems  (an improvement)
        40 -> 12, less examined      =>  less-was-examined
        40 -> 12, no denominator     =>  ambiguous-no-denominator
      
      2 of 4 readings of the SAME drop is an improvement. The denominator is
      checked BEFORE the numerator, so even MORE findings over LESS
      examination reads as less-was-examined -- the conservative and correct
      choice, since an improvement that only appears when less is looked at is
      not one. Same shape as S10.3.
      
      AND ONLY 2 OF THE 9 SUBJECTS ARE COUNTS AT ALL:
      
        a-count-with-a-denominator  2   automated-wcag, zoom-and-reflow
        an-ordered-sequence         2   keyboard-journeys, focus
        a-text                      2   announcements, semantic-alternatives
        a-set-of-states             2   themes/densities, rtl/locales
        an-image                    1   visual-baselines
      
      A screen-reader announcement is TEXT. Focus is an ORDER -- the same
      focusables in a different order is a different experience with an
      IDENTICAL COUNT. Subtracting two numbers is the wrong operation for 7 of
      9, and a diff reporting every subject as a delta has quietly turned
      those comparisons into counts.
      
      AND A DIFF NEEDS A BEFORE. 2 of 566 subject files record a violation
      count or a WCAG level. 2 rows need a screen reader that is not present
      on this machine. Those are two different reasons a row cannot be diffed
      and they are kept apart rather than sharing a footnote.
      
      23 spec tests; scan PASSES with 14 negative controls, two of them
      silence controls. -->
      
    • S11.23.d Distinguish regression/improvement/noise/tool/environment/ intentional semantic change with raw evidence and significance/budget. Evidence: triaged delta register. <!-- A SIGNIFICANCE TEST SAYS THE DIFFERENCE IS REAL. IT SAYS NOTHING ABOUT WHAT CAUSED IT.

      S11.23.b answers one question -- is this larger than the spread -- which
      separates `noise` from everything else and stops. The other five classes
      are ALL "the difference is real"; they differ in their CAUSE, and the
      samples record only the effect.
      
        noise                        the samples themselves
        tool                         a change in the toolchain
        environment                  a change in the machine or image
        intentional-semantic-change  a record written BEFORE the run
        regression / improvement     nothing in the run -- the RESIDUE
      
      1 of 6 is settled by the measurement. 2 need evidence the run does not
      contain. 1 needs a prior record -- a semantic change explained AFTER a
      number moves is indistinguishable from a regression with a good story,
      and the difference is entirely in WHEN it was written down.
      
      SO A REGRESSION IS WHAT IS LEFT. That is not a weakness of the method,
      it is what attribution IS. The register's job is to rule things out, in
      order, and one that jumps straight to "regression" has skipped the
      ruling-out and happened to be right some of the time.
      
      THE WHOLE SPACE, 96 combinations of the four pieces of evidence against
      the six classes:
      
        class                          supported  ruled out  unsupported
        regression                             2         14            0
        improvement                            2         14            0
        noise                                  8          0            8
        tool                                   4          8            4
        environment                            2         12            2
        intentional-semantic-change            1         14            1
      
      19 of 96 support the attribution made. Every class is supported by SOME
      combination and none by all of them -- which is what stops the table
      being a decoration. With a toolchain change present, only the tool
      attribution stands.
      
      23 spec tests; scan PASSES with 13 negative controls, two of them
      silence controls. -->
      
    • S11.23.e Fix each unapproved regression and rerun focused then complete affected comparison. Evidence: closed fix results. <!-- A FOCUSED RERUN IS A CHECK ON THE FIX. A COMPLETE RERUN IS A CHECK ON THE FIXER.

      The cell asks for focused THEN complete, and the order is the whole
      instruction:
      
        focused   did this number come back
        complete  did the change that moved this number move anything else
      
      The focused rerun is fast and is the right thing to run first. It cannot
      answer the second question -- which is the one a person under time
      pressure stops before reaching.
      
      THREE ANSWERS, NOT TWO:
      
        closed
        the-fix-worked-and-nothing-else-was-checked
        not-closed
      
      The middle one is a TRUE statement about a REAL result and it is not
      closure. Reporting it as closure is how a fix for one regression ships
      another.
      
      AND THE COMPLETE RERUN HAS TO BE THE ONE AFTER THE FIX. A green complete
      comparison from before it describes a tree that no longer exists -- the
      same staleness S11.22.h found in a linked report, through a different
      door. The scan drives that with REAL commits, not two invented strings.
      
      AND A FIX IS NOT A FIX UNTIL SOMETHING CHANGED. A regression that
      disappears with no change was not fixed, and calling it fixed removes
      the record that would have said the metric is unstable.
      
        closed-properly                 closed
        focused-only                    the-fix-worked-and-nothing-else...
        complete-from-before-the-fix    the-fix-worked-and-nothing-else...
        nothing-changed                 not-closed
      
      1 of 4 closed. A report claiming all four produces 3 findings.
      
      21 spec tests; scan PASSES with 13 negative controls, two of them
      silence controls. -->
      
    • S11.23.f For an allowed regression, record affected users/SLO/criterion, rationale, alternatives, compensating control, owner, named product/ accessibility/performance approval, expiry, and follow-up. Evidence: approved exception. <!-- AN EXCEPTION IS A DECISION TO SHIP SOMETHING WORSE. THE FIELDS MAKE IT REVIEWABLE AND THE APPROVALS MAKE IT SOMEBODY'S.

      The cell names ten things and the tenth IS NOT ONE FIELD. "Named
      product/accessibility/performance approval" is THREE approvals by three
      different people with three different jobs, and an exception carrying
      one signature has one person's judgement standing in for all three.
      
      That is not pedantry about headcount. A performance approver can say a
      200ms regression is acceptable for this feature. They cannot say it is
      acceptable for a person using a screen reader -- a different question
      about different users with a different answer.
      
      AND AN EXCEPTION WITHOUT AN EXPIRY OUTLIVES ITS REASONS. Every field
      describes a MOMENT: these users, this SLO, this alternative that was not
      available. One never revisited becomes a permanent property of the
      system that nobody chose. The instant is an ARGUMENT, never a clock
      read, so the same exception judged at two moments gives two reproducible
      answers.
      
      AND A COMPENSATING CONTROL IS A THING THAT EXISTS. "We will monitor it"
      is a sentence; a control is something that ACTS -- the same distinction
      S11.20.f drew about remediation. An approval naming a role rather than a
      person is a fault for a related reason: a role cannot be asked about it
      afterwards.
      
      9 withdrawals from a COMPLETE exception, one thing removed at a time,
      cover all 7 fault kinds -- so every finding is attributable to what was
      taken away rather than to a fixture that was never valid. The complete
      exception produces 0.
      
      20 spec tests; scan PASSES with 14 negative controls. -->
      
    • S11.23.g Publish final comparison with zero unexplained or expired regressions. Evidence: signed baseline report. <!-- "ZERO UNEXPLAINED OR EXPIRED" IS TWO CONDITIONS, AND THE SECOND ARRIVES WITHOUT ANYBODY DOING ANYTHING.

      A regression is unexplained until somebody explains it -- a state a
      person changes. An EXPIRED one is different in kind: an explained
      regression carries an exception, the exception carries an expiry, and
      the expiry passes ON ITS OWN. The report was TRUE when published and
      becomes FALSE while nobody edits it, which is the only way a document
      can go wrong with no author.
      
      THE SAME REPORT, PUBLISHED 2026-09-01, READ AT FOUR INSTANTS:
      
        2026-09-01  clean                   0 of 3 blocking
        2026-10-01  has-expired-exceptions  1 of 3 blocking
        2026-11-15  has-expired-exceptions  1 of 3 blocking
        2027-01-01  has-expired-exceptions  2 of 3 blocking
      
      NOBODY EDITED IT. So a publication check that runs once, at publication,
      has verified the only moment at which the property was guaranteed to
      hold. Findings at publication: 3. In November: 4.
      
      AND "SIGNED" IS NOT "CHECKSUMMED". 1 of 3 attestations is a signature.
      S11.20.g measured what a digest computed by the producing process can
      establish -- one of the four things a signature is asked for -- and this
      cell asks for a SIGNED report, so the register keeps the distinction
      rather than accepting a digest under a signature's name. A reader takes
      the word for the claim.
      
      AND A BASELINE REPORT IS WHAT FUTURE COMPARISONS ARE MEASURED AGAINST.
      Publishing one with an outstanding regression does not merely record it
      -- it MOVES THE BASELINE, and the next comparison finds nothing wrong,
      correctly, against a baseline that already contains it. An unfixed
      regression published as a baseline stops being a regression at all.
      
      (An exception with no expiry is left to S11.23.f rather than counted
      twice here.)
      
      24 spec tests; scan PASSES with 14 negative controls, three of them
      silence controls. -->
      
  • S11.24 Complete an adversarial code review of the extracted kit. Evidence: findings are resolved and public API is frozen for domain P1.

    • S11.24.a Freeze review commit/scope and map every public export, extension point, consumer, dependency direction, generated artifact, runtime boundary, and test evidence. Evidence: review manifest. <!-- AN ENTRY POINT IS REACHED BY AN IMPORTER OR BY A RUNNER. A REGISTER'S RUNNER IS ITS SCAN. AN ENTRY POINT WITH NEITHER IS REACHED BY NOTHING.

      This opens an adversarial review OF THIS KIT, and the first honest thing
      to say is that "nobody imports it" is the wrong test for most of it. A
      register exists to be RUN: its scan re-derives the numbers from the
      estate and reports a judgement, so a register with a working scan and no
      importer is doing exactly its job. Applying an import graph to it and
      calling the result dead code is the S11.21.g mistake made in the other
      direction.
      
      So reachability has FOUR states and three of them are fine.
      
      MEASURED ON THE KIT ITSELF, over 81,803 source files repo-wide:
      
        imported-by-a-consumer     48
        run-by-a-scan             178
        imported-and-run            0
        REACHED BY NOTHING        115   of 341 subpath exports
      
      THE ZERO IS ITSELF A FINDING. The kit has split cleanly into code
      somebody uses and registers something runs, with NO module doing both.
      
      AND 115 ARE REACHED BY NOTHING -- a public name with nothing behind it,
      indistinguishable from a working one to every check that reads the
      export map. That is the queue this review starts from. (activity-feed,
      actors, approval-workflows, artifact-rehash, assignments, audit, ...)
      
      AND THE SEVEN ASPECTS ARE NOT SEVEN LISTS: 5 distinct shapes across 7.
      A public export is a NAME; a consumer is a FILE SOMEWHERE ELSE; a
      dependency DIRECTION is a property of a PAIR; a runtime boundary is a
      claim about where code may run. A manifest rendering all seven as one
      table of names has flattened different questions into one.
      
      AND A CONSUMER COUNT IS A WHOLE-REPOSITORY CLAIM, so a manifest whose
      search was narrower than the estate it describes is refused -- exactly
      as a deletion gate is.
      
      20 spec tests; scan PASSES with 12 negative controls, three of them
      silence controls. -->
      
    • S11.24.b Review abstraction quality for Tara leakage, one-consumer extraction, generic escape hatches, unsafe defaults, hidden coupling, runtime/type mismatch, and extension invariants. Evidence: API findings. <!-- AN ABSTRACTION IS A CLAIM ABOUT MORE THAN ONE CASE, AND EVERY ONE OF THE SEVEN CONCERNS IS A WAY OF HAVING ONLY ONE.

      None of the seven is a property of a token, so a grep count answers none
      of them. Each is a RELATION and the register reads pairs: the kit against
      the estate it was cut out of, a module against the PROJECTS that import
      it, a `??` against the declared optionality of the field it is written
      on, a name against every other declaration of that name, a lookup table
      against the keys that reach it, a rule against the thing that enforces
      it.
      
      THE HEADLINE, MEASURED OVER 347 MODULES AND 81,806 SOURCE FILES
      REPO-WIDE. Of 342 subpath exports, 48 are imported at all -- and ALL 48
      BY EXACTLY ONE PROJECT. The distribution has one bucket. The whole kit is
      consumed by TWO projects, `apps/oshun/web` and `apps/oshun/bff`, and
      neither is Tara: the extraction this phase exists to serve has not been
      exercised once by a second caller.
      
      A CONSUMER IS NOT A FILE, IT IS A PROJECT. 39 of those 48 are reached
      from several files inside ONE project, and a consumer count that counted
      files reports every one of them as reuse. Five files in one app exercise
      the abstraction from the one perspective it was extracted from, which is
      why the register keeps `one-project-through-several-files` apart from
      `shared-by-more-than-one-project` and reports the measurement fault
      separately from the thing measured. (S11.24.a's "a consumer is a FILE
      SOMEWHERE ELSE" is the right shape at the wrong altitude for this
      question.)
      
      AND ONE NAME MEANS MORE THAN ONE THING 123 TIMES. Of 2,451 alphabet names
      in the kit, 125 are declared in more than one module and 123 of those
      carry DIFFERENT members. `LIFECYCLE_FAULTS` is declared in EIGHT modules
      -- audit-integrity, audit-snapshots, collaborator-presence,
      comment-threads, lifecycle-corpus, mentions, notification-center,
      response-provenance -- with eight disjoint member lists of size
      3/6/8/8/9/11/11/14.
      
      AND 44 OF THOSE NAMES ARE ALSO SERVED BY THE ROOT ENTRY POINT, so two
      legal imports of one name give two values:
      `import { LIFECYCLE_FAULTS } from '@oshun/workbench-kit'` resolves to
      response-provenance's eight and
      `from '@oshun/workbench-kit/mentions'` to a different eleven. Both doors
      are supported and THE TYPE SYSTEM CANNOT OBJECT, because each door is
      separately consistent.
      
      AND A TARA PATH IN A DATUM IS A CITATION; ONE A FUNCTION READS IS A LEAK.
      20 modules name Tara; a name-grep would report all 20. ZERO name it in a
      type or an identifier -- the leak everybody looks for is absent. 9 hold
      Tara paths in a constant NOTHING reads, which is a register saying what
      it measured, and refusing it the vocabulary would make it unable to. 7
      hold one a function READS, and those are the leakage: the module computes
      from one estate's readings, so a second consumer gets that estate's
      answer under a neutral name.
      
      AND A DEFAULT IS NOT UNSAFE FOR BEING A DEFAULT, IT IS UNSAFE FOR ITS
      DIRECTION. Of 476 `??` written on a field whose declared type admits
      absence: 260 resolve toward a stricter reading, 66 toward a refusal, and
      150 TOWARD AN ANSWER NOBODY COMPUTED. A further 129 are written on fields
      no interface in the module declares as absent, so they cannot fire at all
      and a reader cannot tell whether they are dead or whether the object is
      not the one they think.
      
      AND TWO ABSENCES IN ONE OPTIONS BAG CAN BE ANSWERED IN OPPOSITE
      DIRECTIONS -- 33 modules do it, and it is the finding a per-field rule
      cannot make, because each field alone reads as defensible. The clearest
      is `scanValue` in errors.ts: an absent `tenant` makes it report EVERY
      tenant id as foreign (stricter, correct), and an absent `canaries` makes
      it report no planted value at all (blinder). Same function, same options
      object, six lines apart, and the bag defaults to `{}` so the easiest call
      to write gets one of each.
      
      AND `unknown` IS NOT AN ESCAPE HATCH, IT IS THE OPPOSITE -- it cannot be
      used until it is narrowed, so grouping its 219 occurrences with `any`
      because both are "loose" inverts the finding. `any` is at ZERO. AND A
      HATCH IN AN IMPLEMENTATION IS A CHOICE WHILE A HATCH IN A SIGNATURE IS A
      CONTRACT, so the 32 findings are the bags and index signatures on the
      public surface; the 1,341 casts onto a kit union are not dropped, they
      are the input to the next concern.
      
      THE SHARPEST FAULT NEEDS TWO FACTS AT ONCE. `noUncheckedIndexedAccess` is
      OFF across this repo, so `Record<K, V>[k]` types as `V`. Of 79 union-keyed
      frozen lookup tables, 33 have a key manufactured somewhere by an
      assertion rather than a check, and 31 of those never check it before the
      index -- the miss types as a hit. Neither fact alone is a fault: a cast
      alone is not one, a table alone is not one, and turning the compiler
      option on would fix all 31 without touching either.
      
      AND 3 OF 4 STATED EXTENSION INVARIANTS ARE ENFORCED, which is a good
      result honestly obtained. Both directions of declaration-vs-contribution
      are real fault kinds; the seven extension points are a closed set checked
      by `includes`; measurement-vs-verdict is carried in `decidedBy` where the
      type will not let a verdict forget who made it. The fourth --
      "ABSENCE IS A VALUE, NEVER AN EXCEPTION" -- IS ANSWERED AT THE CALL SITE,
      and plugins.ts contains ZERO `try`. It declares 18 function-typed fields
      and calls one at two sites: `policy.decide(measurement)` and, as the last
      line of `activatePlugin`, `plugin.activate(context)`. That function
      returns `not-configured` and `refused` as VALUES for every absence THE
      KIT can detect, argues in its own header that a plugin signalling absence
      by throwing would be indistinguishable from one that crashed -- and then
      hands control to the plugin unguarded, which is exactly the state the
      rule forbids. The rule holds for the absences the kit finds and is
      unenforced for the one the plugin declares.
      
      MY OWN INSTRUMENT WAS WRONG TWICE AND BOTH FIXES CHANGED AN ANSWER. The
      first detector for "both directions" looked for a phrase from the doc
      comment rather than the fault kinds, and reported a rule the validator
      really does enforce as unchecked. The second matched `plugin.<field>(`
      anywhere, so it "found" contribution calls in `plugin.descriptor.x` --
      it reached the right verdict on rule 4 for a reason that was not the
      reason. A DETECTOR THAT AGREES WITH THE ANSWER YOU EXPECTED IS NOT
      THEREBY CORRECT; both now read the fault alphabet and the call sites.
      
      AND THE REGISTER CARRIES NO NUMBER IT DID NOT DERIVE -- a scan check
      fails if a two-digit number appears in it, with ledger-cell references
      stripped first so `S11.24.b` is a citation and a measurement quoted in a
      doc comment is still caught.
      
      39 spec tests; scan PASSES with 30 negative controls, ten of them silence
      controls. -->
      <!-- The five registrations: kit package.json export, root
      verify:operations script, project.json target, export-map generator, and
      the tsconfig.base.json path alias -- which the generator's own
      `unaliased-approved-entry-point` check caught me omitting. -->
      <!-- Typechecked libs/oshun/workbench-kit directly: tsconfig.lib.json is
      clean; tsconfig.spec.json has 23 errors, none of them in this cell's
      files (adapter-compatibility, duplicate-fixtures, fixture-reachability,
      graph-parity, manifest-relations). -->
      <!-- NOT DONE HERE, and named rather than folded in: the 115 entry points
      S11.24.a found reached by nothing are not re-judged here; whether the
      1,341 casts are each justified is a per-site question this cell does not
      answer; and the 44 two-door names are reported, not resolved -- renaming
      them is a public API change, which is S11.24.h. -->
      
    • S11.24.c Review identity/tenancy/object/property/function authorization, step-up, idempotency/concurrency, audit, errors, limits, fetch/upload, and secret/protected data paths adversarially. Evidence: security findings. <!-- A SECURITY DECISION IS ONLY AS BINDING AS THE WEAKEST LINK BETWEEN THE SUBJECT AND THE VERDICT.

      Reviewing fifteen dimensions one module at a time asks the same question
      fifteen times and misses the shape, because a bypass never needs the rule
      to be wrong -- it needs one of three links to be loose. HOW THE FACT
      ARRIVES: a decider that is TOLD the security-relevant fact is exactly as
      correct as whoever told it. HOW THE DECISION IS SHAPED: a verdict
      returned beside its subject can be dropped. WHAT THE VERDICT CAN SAY: a
      refusal that cannot name what disagreed sends its reader elsewhere. One
      finding under each, on 15 of 15 dimensions, every compound split.
      
      AND EACH LINK HAS A STATE THAT LOOKS LIKE THE FAULT AND IS NOT. That is
      where a naive count goes wrong, and getting the three separators right is
      most of this cell.
      
      1. TAKING THE ANSWER IS RIGHT WHEN THE DECIDER CANNOT COMPUTE THE
      QUESTION. `tenancy.ts` argues this itself, correctly and at length, for
      `stageIsImmutable`: the check "used to read `lifecycleStage ===
      'published'`, and that spelling protects a workbench whose terminal stage
      happens to be called `published` and silently protects nothing in one
      whose stage is `released`, `aired`, `filed` or `shipped`. The kit cannot
      know the word, so it takes the answer instead." That is not a loose link,
      it is the only correct design.
      
      AND THE SAME MODULE TAKES `sameTenant` THE SAME WAY. `AccessRequest`
      carries `sameTenant: boolean` and `sameWorkspace: boolean` and no tenant
      ids at all, so the module named for tenancy cannot compare tenants -- it
      can only be told. A tenant is NOT domain vocabulary: `TenantId` is one of
      17 identifier types the kit itself brands, and the decider holds both
      values. The separator is `theDeciderHoldsTheValues`, not the shape of the
      parameter, which is a boolean either way.
      
      AND THE ESTATE CONTAINS ITS OWN ANSWER, ONE MODULE AWAY. `authorization.
      ts` compares a tenant 4 times, and its header calls the re-compare after
      a scoped resolver "the single most valuable line in the module", because
      "a resolver that ignores its scope argument -- the most ordinary bug
      there is, `SELECT ... WHERE id = $1` -- is invisible to every test that
      only checks the happy path". One module distrusts its input on exactly
      this question and the neighbour accepts a claim.
      
      2. A DECISION THAT CARRIES NO SUBJECT IS NOT ADVISORY, IT IS A POLICY
      FUNCTION. A rule handed an object and returning a verdict about it has
      nothing to leak. Of 7 verdict types in the security surface carrying a
      nullable field, SIX carry a `reason`, an `error`, a `probe`, a `refusal`
      or an audit record -- all ABOUT the decision rather than subject to it,
      and reporting them would have been six false positives. The fault needs
      three facts at once: the decider LOADED the subject, it sits beside the
      verdict, and it is what the verdict is about.
      
      THE SEVENTH IS `AuthorizedLoad`, and `LoadOutcome` seventy lines above it
      is the same module getting it right. `LoadOutcome` is a SUM -- `metadata`
      exists only on the `loaded` variant, so it cannot be named without
      narrowing first. `AuthorizedLoad` is a PRODUCT: `{ outcome, metadata:
      ObjectMetadata | null, fault }`. One null check -- which every caller
      writes anyway -- yields the metadata, and the verdict beside it was never
      read.
      
      AND THE SAME FUNCTION BINDS AT ONE DISPOSITION AND NOT AT ANOTHER, which
      is the finding no per-function rule states. `authorizeObject` withholds
      the metadata at `conceal` and at `unavailable`, CORRECTLY and for the
      reason its header gives -- "a 403 is an answer: it says the object
      exists" -- and returns it on whatever the rules decided, which is allow,
      deny AND step-up. So a caller denied for not being assigned receives
      `owner`, the full `assignees` list, `team`, `classification` and the
      `sensitivity` classes: a denial that names who to ask is also a denial
      that names who to phish. Decision 2 of the module's own header bounds the
      blast radius -- `ObjectMetadata` has nowhere to put content, deliberately
      -- and decision 5 says a refusal "carries a stable reason and a
      remediation, and NEITHER CARRIES THE OBJECT". The return shape carries
      it. Binds at 2 of 3, and THE HALF THAT WORKS IS WHAT MAKES THE OTHER HALF
      HARD TO SEE.
      
      3. AND A MEMBER NAME IS NOT THE FIELD IT NAMES. Step-up is the best-built
      thing in this review and the finding is small on purpose. `Challenge`
      binds twelve fields including a `parameterDigest` -- "a different request
      is a different act" -- `CHALLENGE_BINDINGS` is DATA and `verifyChallenge`
      ITERATES it, so "a binding that stops being checked is a deletion from a
      list rather than a missing `if` nobody notices", and `bindingValue` is a
      real exhaustive switch onto real fields. Eight things really disagree and
      the alphabet names seven. The eighth is the NONCE, which has no member,
      and the caller must pass something -- so a nonce mismatch is filed as
      binding `'actor'`, which `bindingValue` maps to `actorId`. The rejection
      is CORRECT; it is reported as a thing that did not disagree.
      
      MY OWN INSTRUMENT WAS WRONG THREE TIMES AND EVERY FIX CHANGED AN ANSWER.
      The brand reader required the brand on one line, found 3 identifier types
      instead of 17, and so cleared `sameTenant` as a fact only the caller could
      know. The producer reader was a nested `indexOf` slice that never fired,
      so `AuthorizedLoad` came back as a policy function. And the vocabulary
      check compared MEMBER names against FIELD names, reporting `actorId` and
      `sessionId` as unnameable when the alphabet names them perfectly well as
      `actor` and `session` -- three unnameable things where one is real. The
      disposition reader then found nothing at all, because `search()` with a
      non-greedy `[\s\S]*?` matched from `domainPolicyRule` and stretched to
      `authorizeObject`'s return type: A NON-GREEDY SPAN FINDS THE EARLIEST
      START, NOT THE NEAREST. It now carries the producer the shape check
      already resolved, so the two readers cannot silently disagree about which
      function they are describing.
      
      AND THE REGISTER CARRIES NO NUMBER IT DID NOT DERIVE -- checked, with
      ledger-cell references stripped first.
      
      26 spec tests; scan PASSES with 20 negative controls, nine of them
      silence controls -- six of which assert that a correct design is NOT
      reported, since three of the five fault kinds exist to be silent on the
      state that resembles them. -->
      <!-- NOT DONE HERE: the fifteen dimensions are read through the three
      links, not exhausted individually -- idempotency's two-racer `claimed`
      and the reservation/settle path are named in their own cells, and this
      cell does not re-judge them. Nothing here is fixed: the shape changes are
      public API changes and belong to S11.24.g and .h. -->
      
    • S11.24.d Review persistence/migrations/events/jobs/cache/blob/provenance/ publish for atomicity, races, replay, partial/unknown outcome, retention, and recovery. Evidence: correctness findings. <!-- A DOUBT NEEDS A NAME, A PRODUCER AND A READER, AND THEY FAIL SEPARATELY.

      The cell names six properties and five are answered the same way: find
      out which state you are in, then act on it. AN UNKNOWN OUTCOME IS NOT A
      SIXTH HAZARD, IT IS THE FAILURE OF THE PRECONDITION THE OTHER FIVE ARE
      ANSWERED WITH. A retry is safe when you know the effect was partial; an
      idempotency key protects a replay when you know the key was recorded; a
      recovery needs a state known to be good. When the system cannot tell, all
      of them degrade to a guess at once -- 6 of 7 properties presuppose that
      observation, and the seventh is its loss. (`partial/unknown outcome` is
      split for exactly this reason: a partial outcome is a state the system
      OBSERVED.)
      
      So the question is whether each subject can HOLD a doubt, and that has
      four answers rather than two: no module owns the subject, no name exists
      for the doubt, a name exists with no producer, or a producer exists with
      no reader.
      
      5 OF 8 SUBJECTS HOLD A DOUBT AND ACT ON IT, and `jobs.ts` is the
      exemplar because the rule is the estate's own: it says "its external
      effect is UNKNOWN, and unknown is not failed", it keeps "never started"
      apart from "outcome unknown", and it carries `unknownOutcomeMaxAttempts`
      as a separate LOWER ceiling -- "a lower ceiling for unknown outcomes:
      each attempt may be a second effect" -- with the answer DEPENDING on
      whether the effect is idempotent. `outbox`, `blobs`, `publishers` and
      `provenance-inspector` each branch on a doubt state too; `publishers`
      carries three states with nine reading sites apiece.
      
      THE THREE THAT CANNOT, AND THE WORST IS THE QUIETEST. `cache` has NO
      MODULE in the kit at all -- no file carries a cache in its topic line,
      and crediting `freshness-recompute.ts` would be crediting S11.20.d's own
      register with a question it was not asked. And `dual-write-divergence`
      has 5 alphabets and `migration-paths` has 3, and NOT ONE MEMBER BETWEEN
      THEM names an undeterminable outcome -- in the two places where doubt is
      structural. A dual write is exactly the operation where one side lands
      and the other's outcome is unknown. A half-applied migration is none of
      `START_STATES`' four (`empty`, `prior-supported-version`,
      `production-like-snapshot`, `live-database`). THE ALPHABET CANNOT SPELL
      IT, SO A DOUBTFUL CASE MUST BE FILED UNDER A CONFIDENT ONE, and nothing
      is missing from any file.
      
      AND A DEAD MEMBER IS A FINDING WHATEVER ITS NEIGHBOURS DO. `blobs`'
      `STALENESS.unknown` is assigned at one site and read by nothing, so a
      blob whose staleness could not be determined is stored and then treated
      exactly like `current`. `provenance-inspector`'s `RIGHTS_HINTS.unknown`
      appears EXACTLY ONCE in the whole module -- in its own declaration --
      beside `not-recorded`, which is a different doubt again ("nobody wrote it
      down" is not "we looked and cannot tell"). Both sit next to a working
      sibling, so a subject-level standing alone would have hidden them: the
      review reports the subject AND the member.
      
      AND A STATE AND A FAULT KIND ARE JUDGED BY OPPOSITE TESTS. This is the
      trap the whole measurement turns on. A STATE is live when something
      BRANCHES on it, because a value stored and never consulted behaves
      identically to its confident neighbour. A FAULT KIND is live when
      something PRODUCES it, because a detector emits it and branching on one
      would be the odd thing to do.
      
      MY OWN PROBE APPLIED THE BRANCH TEST TO BOTH AND CALLED FOUR OF THE
      ESTATE'S BEST-DETECTED FAULTS INERT -- including all three of
      `PUBLISH_RECONCILIATION_FAULTS`' members named for exactly this hazard
      (`unknown-outcome-retried-before-a-probe`,
      `unknown-outcome-marked-failed-with-no-evidence`,
      `unknown-outcome-marked-succeeded-with-no-evidence`), each of which IS
      produced by a real detector. A DEFECT IN THE INSTRUMENT REPORTED AS A
      DEFECT IN THE ESTATE, and it was caught only by opening the files. An
      earlier version was worse: it tested production by deleting every
      occurrence of the literal from the file, which deletes the production
      sites too, so it could only ever answer "none".
      
      AND THE EXEMPLAR READER LOOKED IN THE STRIPPED SOURCE. Both of the
      sentences that make `jobs.ts` the exemplar live in its DOC COMMENT, and
      the reader ran on comment-stripped text -- so it answered a question
      about the stripper and reported `false` about a module that says the
      thing plainly. It now reads the raw file.
      
      AND THE REGISTER CARRIES NO NUMBER IT DID NOT DERIVE -- checked.
      
      21 spec tests; scan PASSES with 16 negative controls, four of them
      silence controls, one of which drives the state/fault-kind trap BOTH WAYS
      on the same member and facts. -->
      <!-- NOT DONE HERE: the six properties are the lens, not a 8x7 matrix
      walked cell by cell -- races and retention are read through whether the
      subject can hold the doubt they turn on, and are not separately
      exhausted. Nothing is fixed: naming a doubt in `START_STATES` or
      `dual-write-divergence` widens a public union, which is S11.24.g and .h.
      -->
      
    • S11.24.e Review provider/trust/gates/scores/fixtures/states for injection, fabricated truth/success, stale authorization, policy/version, cost, override, and human approval boundaries. Evidence: trust findings. <!-- AN ESTATE WANTS ZERO OF A HAZARD AND SOME OF A POWER, AND THE TWO ARE ANSWERED BY OPPOSITE INSTRUMENTS.

      The cell names seven slots and reading them as seven of a kind is the
      mistake. Ask of each whether ZERO IS THE GOAL. Nobody wants an injection,
      a fabricated success, or a result evaluated under a policy that is not
      the one in force -- those are HAZARDS, answered by a DETECTOR. But an
      estate that spends nothing does nothing, one that cannot override a bad
      gate cannot recover from it, and one with no human approval has removed
      the point of asking -- those are POWERS, answered by a CEILING. THE CELL
      ITSELF SAYS SO AND PUTS IT ON EXACTLY THE RIGHT THREE: it says
      "BOUNDARIES", and a boundary is what you give a thing you want some of.
      
      AND THE INSTRUMENTS DIFFER IN THE ONE PROPERTY A REVIEW DEPENDS ON. A
      DETECTOR PROVES ITSELF BY FIRING -- feed it the fault, watch it report,
      which is this phase's whole method. A CEILING CANNOT. A power exercised
      inside its limit and a power with NO limit anywhere leave THE SAME
      RECORD: work happened and nothing objected. The observation that would
      tell them apart is the one a working system never produces. So a ceiling
      is judged by its PATHS and never by its rule -- the rule is the part
      that is easy to read and the part that is never wrong.
      
      THE MEASUREMENT LANDS ON THAT LINE EXACTLY. All six hazards have a
      detector that fires: 43 of 43 fault kinds PRODUCED, across
      trust-zones.CONTROL_FAULTS, score-fabrication.SCORE_FABRICATIONS,
      provider-fabrication.PROVIDER_FABRICATIONS,
      reviewer-handoff.GRANT_AUTHORITY_FAULTS (17 alone),
      trust-results.INVALIDATION_FAULTS and .STAMP_FAULTS. And 0 OF 3 POWERS
      have a ceiling that binds every path it censused. THE ESTATE IS PERFECT
      WHERE IT COULD CHECK ITS OWN WORK AND IMPERFECT WHERE IT COULD NOT, and
      that is not a coincidence -- it is the property above.
      
      AND A FAULT KIND IS A PROPERTY OF A RECORD, SO THE WAY PAST A CEILING IS
      TO LEAVE NONE. This is why a no-record path ranks BELOW a path closed
      only by a convention: no alphabet, however long, reaches a route that
      writes nothing, because there is nothing for a fault to be ABOUT. THE
      ESTATE STATES THIS ITSELF, about its own worst case -- `a11y-waivers`
      says `skipAxe` is "a boolean fixture option ... there is no field on it
      to carry anything and no finding for a field to be about", over 247
      opt-outs, and 2 of its 3 APPLICATION_POINTS leave nothing to waive.
      `catalog-write-paths` names the write channel that "leaves no
      application record at all".
      
      AND A CLOSURE THAT NAMES AN INTENTION IS NOT ONE, discovered TWICE and
      named differently both times: `build-gates.GATE_STANDING` carries
      `has-a-documented-bypass` -- "the honest kind of hole and it is still a
      hole" -- and `catalog-write-paths` sets 2 of its 10 CLOSURES aside as
      CLOSURES_BY_CONVENTION (`documented-convention`, `code-review`). Neither
      module knows the other's word for it.
      
      AND A CENSUS IS A CLAIM ABOUT EVERY PATH. `limits` says "every one of
      these limits has a bypass, and the bypasses are more interesting than
      the limits" -- and `bypassFindings` walks RATE_DIMENSIONS, which is 2 of
      the 6 categories LIMIT_CATEGORIES_PUBLIC declares. Payload, upload,
      query-complexity and spend have no enumerated bypass.
      `cost-accounting.dispatchVerdict` is the exemplar of the other kind: it
      REFUSES at the point of use, with 5 named refusals, one of them for the
      unreserved path -- the bypass closed rather than reported.
      
      AND THE RULE AND ITS ROUTES ARE FILED APART: `override` keeps its
      ceiling and its two path censuses in THREE different modules, so a
      reviewer who opens the module deciding whether a waiver authorizes never
      meets the enumeration of the routes that never reach it.
      
      MY OWN INSTRUMENT WAS WRONG TWICE AND READING CAUGHT BOTH. A keyword
      search for path censuses found 2 and the estate has at least 5 --
      `limits`, `build-gates` and `isolation-paths` carry one under names no
      pattern matches. A KEYWORD-DERIVED ZERO IN THIS ESTATE IS ALMOST ALWAYS A
      FACT ABOUT THE KEYWORD (third recurrence). And a substring test for
      limit coverage would credit `upload` from `concurrent-uploads` -- a
      CONCURRENCY dimension, not the upload-size limit -- reporting a partial
      census as whole; the category mapping is READ, and the scan fails loudly
      on any dimension it does not classify rather than dropping it.
      
      30 spec tests; scan PASSES with 21 negative controls, four of them
      silence controls, two driving the hazard/power split BOTH WAYS on the
      same facts. -->
      <!-- NOT DONE HERE: the six named subjects are the cell's index, not the
      review's -- coverage is read where the estate files it, and every
      coverage count is reported as a FLOOR because the fault vocabulary is
      phrasal and any outside taxonomy under-counts it. Nothing is fixed:
      closing a no-record path or widening a census changes public surface,
      which is S11.24.g and .h. -->
      
    • S11.24.f Review UI/API accessibility/localization/resilience/privacy/ performance/operability and test harness ability to detect regressions. Evidence: quality findings. <!-- A QUALITY THE VALUE CARRIES TRAVELS WITH IT; A QUALITY AN INSTRUMENT VISITS IS ONLY AS REAL AS THE LAST VISIT.

      The cell names six qualities and then asks something it asks of nothing
      else in this whole review: whether the harness can DETECT A REGRESSION in
      them. THAT CLAUSE IS NOT A SEVENTH QUALITY, IT IS THE REASON THE OTHER
      SIX ARE GROUPED. Each of them fails by producing a PERFECTLY VALID VALUE
      -- an inaccessible button is a valid button, an untranslated string is a
      valid string, a leaked field is a valid field, a slow response is a valid
      response, an unoperable service is a running service, and a degraded
      answer is a valid answer, which is `degradation-plans`' own opening
      sentence. VALIDITY IS WHAT ALL SIX FAILURES HAVE, so no validator finds
      any of them and the only question left is whether anything can tell.
      
      THE KIT ANSWERS THAT IN TWO WAYS AND NEVER NAMES THE CHOICE. 5 OF 6 ARE
      ANSWERED BY MAKING THE VALUE CARRY THE FACT ABOUT ITSELF, under five
      names that are never connected: `ResolvedText` returns the text WITH the
      locale asked for, the locale it actually came from, the fallback step and
      eight notices including `fallback-used`; `publicError` and `projectFor`
      CONSTRUCT the safe shape rather than deleting fields, because "deletion
      is a blocklist, and a blocklist is wrong the first time somebody adds a
      field"; `degradation-plans` requires the returned value to NAME THE FIELD
      that says it is a fallback AND the consumer that branches on it;
      `service-objectives` requires the ratio to carry the denominator the
      failing system chose. THE ESTATE MADE THE SAME MOVE FIVE TIMES.
      
      THE SIXTH, ACCESSIBILITY, IS ANSWERED BY A VISIT -- and that is NOT a
      mistake, because no type can hold "this label is meaningful", so for part
      of the quality the visit is FORCED. But the two differ in exactly the
      property the cell's last clause asks about: a carried fact is in every
      consumer's hands on every call, and a visited fact exists only where the
      visit reached. That is why `accessibility-diff` must ask whether a
      falling count means "fewer problems" or "less was examined", and why none
      of the carried five has to.
      
      AND 1 OF THE 4 WAYS A VISIT STOPS BEING EVIDENCE HAS NO INSTRUMENT IN ANY
      ACCESSIBILITY MODULE. `not-run`, `reached-less-than-before` and
      `no-denominator-recorded` are all modelled (`a11y-waivers`,
      `accessibility-diff`). `run-on-another-subject` is modelled NOWHERE --
      and `build-gates` names that exact standing, "green every time, about a
      codebase this change did not touch", and attributes it in its own words
      to "S11.21.e's accessibility scripts". The estate has written down that
      its accessibility scan has this trap and has built nothing that detects
      it.
      
      AND THE PROTECTION IS INVERSE TO THE INVISIBILITY OF THE FAILURE. The kit
      brands 13 identifier types with a unique symbol, a strict parser and --
      its own words -- "no unsafe cast", so a mixed-up identifier CANNOT BE
      CONSTRUCTED, and that is also the failure the first test to run would
      have caught. The fields carrying the six silent qualities are 94 of 102
      bare `string`, a type that cannot distinguish a translated label from an
      untranslated one, a message key from an English sentence, or a real
      accessible name from a leftover default. The estate knows the hazard by
      name and answers it in ONE module at RUNTIME: `errors.validateMappings`
      "refuses a classifier that looks like a sentence, because that is how
      message-matching comes back after being removed" -- for a hazard the TYPE
      permits at all 24 bare `messageKey` declarations.
      
      MY OWN INSTRUMENT WAS WRONG FIRST, AND THE SCAN CAUGHT IT RATHER THAN THE
      ESTATE. Three cited sentences and one carrier name came back stale, and
      the review reported `resilience` as carried-by-nothing -- A DEFECT IN THE
      INSTRUMENT ABOUT TO BE REPORTED AS A DEFECT IN THE ESTATE. The cause was
      one thing: A PROSE CLAIM SPANS LINES AND THE COMMENT PREFIX SITS IN THE
      MIDDLE OF IT, so a regex over raw text answers a question about where the
      author happened to wrap. The reader now strips the prefix and collapses
      whitespace before matching. (S11.24.d's version of this was a reader
      searching comment-stripped source for a sentence that lives only in a
      comment; same family, opposite direction.)
      
      24 spec tests; scan PASSES with 18 negative controls, two of them silence
      controls, two driving the carried/visited split BOTH WAYS on the same
      facts. -->
      <!-- NOT DONE HERE: the six are read at the altitude the kit occupies --
      it is a library, so it can make a quality reachable or unreachable but
      cannot itself be accessible, and no browser or endpoint was driven. The
      field counts are a FLOOR over a deliberately tight field list. Nothing is
      fixed: naming a carrier or constraining `messageKey` changes public
      surface, which is S11.24.g and .h. -->
      
    • S11.24.g Classify blocker/major/minor/question, owner/fix/decision, retest, and resolve all blocker/major and every public API uncertainty. Evidence: closed finding ledger. <!-- STILL OPEN, DELIBERATELY. The ledger is BUILT, SEVEN ROWS HAVE BEEN FIXED AND RETESTED under a public-API authorisation given in the session of 2026-08-16, and it still cannot be closed here: finding-ledger.ts + .spec.ts (21 tests) + tools/finding-ledger-scan.mjs (PASSES, 18 negative controls, four of them silence controls). The evidence clause asks for a CLOSED ledger; 10 of its 13 rows must close before the freeze and 6 are still open, so the checkbox stays [ ].

      FIXED AND RETESTED (each with a citation the scan re-checks, rewritten
      from asserting the defect to asserting the fix, so a row cannot outlive
      its subject in either direction):
      
        B8  plugins.ts said "ABSENCE IS A VALUE, NEVER AN EXCEPTION" and then
            called `plugin.activate(context)` unguarded, with zero `try` in the
            file. PLUGIN_VIOLATION_KINDS gains `activation-threw` -- the only
            member of that alphabet found by RUNNING rather than by READING --
            and the throw becomes `refused`, never `not-configured`.
        C1  AuthorizedLoad was a PRODUCT in a module whose decision 5 says a
            refusal "carries a stable reason and a remediation, and NEITHER
            CARRIES THE OBJECT". It is now a SUM shaped like `LoadOutcome`
            seventy lines above it. This closes S11.24.c's second half too: the
            old shape returned the object on deny AND step-up, and A DENIAL
            THAT NAMES WHO TO ASK IS ALSO A DENIAL THAT NAMES WHO TO PHISH. The
            one consumer LOST a branch -- the state it tested for has no
            representation now.
        D3  `blobs.STALENESS.unknown` was produced and read by nothing.
            `dispositionFor` gives it a reader, and it gets its OWN disposition
            rather than either neighbour's, because a MOVED ancestor is known
            wrong and re-deriving fixes it, while an UNRESOLVED one is unknown
            and re-deriving from an ancestor you cannot read produces a second
            artifact with the same problem.
        D4  `RIGHTS_HINTS.unknown` was NEVER produced -- a caller-supplied
            field, so not D3's shape at all. TWO doubts sat side by side with
            nothing telling them apart, in the module whose stated subject is
            what a reader can TELL APART. "Nobody wrote it down" is closed by
            the holder writing it down; "we looked and cannot tell" is not.
        E1  `limits.bypassFindings` took `RateRule[]`, so a module that says
            "every one of these limits has a bypass" enumerated two of its six
            categories. It now takes a `LimitPolicy` naming all six, and both
            new members come from the module's OWN law rather than invented
            hazards: a spend scope nobody budgeted is skipped by `reserve` in
            silence, and -- since every cap is a required number, so a MISSING
            one has no representation -- the hazard the type cannot close is a
            cap that never refuses. EVERY COMPARISON WITH NaN IS FALSE.
        F1  `accessibility-subject.ts` is the instrument S11.24.f found
            missing. THE REASON IT WAS MISSING IS THE INTERESTING PART: a scan
            records what it RENDERED and a change records what somebody EDITED,
            and neither says which rendered thing an edited file is in. The
            S11.24.f scan now reports accessibility as
            `visited-with-every-loss-modelled`.
        B1  is NOT fixed and is now BOUNDED -- see below.
      
      AND THE REVIEW ITSELF WAS THE PROBLEM FIRST. Measuring B1 to decide how
      to fix it showed S11.24's own eight modules had added EIGHT name
      collisions while reporting that exact defect. `SURFACES` was declared by
      `quality-review` and `finding-ledger` -- two modules written in the same
      session -- with entirely different members, and `FACT_SOURCES` collided
      with `graph-views`, whose version the barrel serves, so the bare name and
      the subpath gave two different values. A REVIEW THAT GROWS THE DEFECT IT
      REPORTS HAS NO STANDING TO REPORT IT. All eight renamed; a ratchet
      (`two-door-names-scan`) now fails on a new one and holds the review
      modules at zero. THE FIRST RATCHET WAS BROKEN AND A POSITIVE CONTROL
      CAUGHT IT: it counted colliding NAMES, so a ninth `LIFECYCLE_FAULTS` did
      not move the number -- A RATCHET THAT COUNTS THE NAMES CANNOT SEE A NEW
      COLLISION ON AN OLD NAME. It counts duplicate DECLARATIONS now, against a
      baseline that is measured rather than guessed (the guess was 176 against
      a real 179).
      
      A FIX IS RETESTABLE AND A DECISION IS NOT, AND THE CELL'S THREE
      RESOLUTIONS SPLIT ON EXACTLY THAT LINE. A fix changes code, so running
      the test again is evidence. A decision changes a record -- an owner
      agreed the behaviour is intended, or that the cost is accepted -- and NO
      TEST PASSES BECAUSE SOMEBODY AGREED. A row claiming a retest on a
      decision is claiming evidence that cannot exist, and that is a fault kind
      here rather than a formality: it is how a ledger reports itself green.
      
      AND `question` IS NOT A SEVERITY. blocker/major/minor say how much damage
      a finding does; a question says nobody can say yet, because something has
      to be decided before the damage is knowable. It is kept OFF the ordered
      scale, so no sorting of it can make it read as at-least-major.
      
      AND WHAT MUST BE RESOLVED BEFORE A FREEZE IS DECIDED BY WHERE THE FINDING
      IS, NOT BY HOW BAD IT IS. The cell names two triggers and they select
      different rows. A major finding inside a private helper can ship and be
      fixed next week. A MINOR question about a public signature cannot,
      because S11.24.h freezes that signature and the freeze makes the answer
      permanent -- the cost of being wrong stops being the defect and becomes
      the support window. The two are OR-ed and neither implies the other.
      
      AND A ROW IS NOT EVIDENCE THAT A FINDING EXISTS. Every row cites
      something the scan RE-CHECKS against the source -- 21 distinct citations
      across 13 rows, all re-checked -- because a ledger is the one artefact in
      this review with an incentive to be wrong: a row saying "resolved" is
      cheaper to write than a fix, and a row whose subject has since changed
      reads exactly like a row that was always right.
      
      MEASURED: 13 rows over the six cells that found something -- 8 major, 3
      questions, 2 minor; 7 on the public API, 3 outside the kit, 2 on an
      instrument in this kit, 1 on a compiler setting. Exactly ONE row is
      closed, by a decision S11.24.d had already RECORDED (cache has no module,
      and crediting `freshness-recompute` would credit S11.20.d's register with
      a question it was not asked).
      
      A LEDGER CANNOT CLOSE ITSELF, and that is the finding rather than an
      excuse. Every remaining resolution is one of two things this cell cannot
      perform. Either it is a PUBLIC API CHANGE -- renaming the two-door names,
      turning `AuthorizedLoad` from a product into a sum, widening a union so a
      throwing plugin gets a typed refusal (checked: `PLUGIN_VIOLATION_KINDS`
      has no member that says a plugin threw, and all 46 describe declaration
      defects found by validation, so even the honest repair of `plugins.ts`'
      own rule 4 widens the public surface). Or it is an ACT OUTSIDE THIS KIT
      -- a revoked privilege on a write channel, a repository-wide
      `noUncheckedIndexedAccess` whose blast radius is every project, several
      hundred accessibility call sites. NO DECISION IS INVENTED HERE: an owner
      agreeing that a cost is acceptable is an act a scan cannot perform, and
      writing one down would be the fabrication this whole review exists to
      find. That is precisely the named approval S11.24.h is gated on, so .g
      and .h are not sequential in the way the list implies -- .g's remaining
      resolutions ARE .h's freeze content.
      
      WHAT REMAINS, AND WHY EACH ONE IS A DECISION RATHER THAN A FIX. Six rows,
      and not one of them is work somebody could just do.
      
        A1  whether each unreached entry point is intended surface. Only the
            freeze answers that; it IS the freeze.
        B1  which of eight `LIFECYCLE_FAULTS` is "the" one, per name, 46 times.
            Removing a name from the barrel breaks whoever imports it.
        B7  `noUncheckedIndexedAccess` is off; S11.24.b measured that turning
            it on fixes all 31 lookup-table faults without touching either the
            cast or the table. Its blast radius is every project in the
            repository, so the kit cannot take it alone.
        E2  the accessibility opt-outs: several hundred call sites in the
            subject estate, not in this kit. `a11y-waivers` already states the
            repair.
        E3  closing a write channel is an infrastructure act -- a revoked
            privilege, a read-only replica -- that no change to this kit
            performs.
        F2  whether to brand `messageKey`, parse it, or leave the runtime guard
            as the answer. S11.24.f explicitly declined to rule: "the repair is
            NOT to brand every label -- that would be the one-consumer
            abstraction S11.24.b exists to refuse." Doing it unilaterally would
            contradict this review's own finding.
      
      NO DECISION IS INVENTED HERE. An owner agreeing that a cost is acceptable
      is an act a scan cannot perform, and writing one down would be the
      fabrication this whole review exists to find. -->
      <!-- Retested together after every fix: authorization, preconditions,
      bulk-actions, graph-views, route-states, redaction, plugins, blobs,
      provenance-inspector, limits and the eight review specs; the kit
      typechecks clean (`tsc --noEmit -p tsconfig.lib.json`, exit 0) and all
      eight scans plus the ratchet exit 0. -->
      
    • S11.24.h Freeze public API/compatibility/version/support window only after clean build/contracts/Tara full suite and named architecture/ package/security/domain approval. Evidence: P1 API signoff.


Phase I — Isis governed-generation workbenches#

Isis is not one new /studio/isis-workbench page. It has three governed audiences: operator factory controls in apps/isis/web, AAA-creator controls in Yemaya Studio, and curated generation cards plus the output gallery in Oshun Studio. This phase closes the workflow across those hosts while preserving the generation-tier resolver as the single boundary authority.

I0 — Product boundary, reuse, and threat model#

  • I0.1 Reconcile every Isis route with STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS and the generation-tier resolver. Evidence: route-boundary test covers every discovered route. <!-- SUPERSEDED (2026-08-16). The old note said c-g were "runtime/design/ human and stay unchecked". Six of the seven children are now closed: .a .b .c .d .f .g. The claim that .c needed a runtime was wrong -- the generation-tier resolver is a pure function and the generator runs it.

    THE PARENT STAYS OPEN ON .e, AND ITS OWN EVIDENCE CLAUSE IS SEPARATELY
    SATISFIED. "route-boundary test covers every discovered route" is true:
    I0.1.f reports 0 uncovered register routes and 0 uncovered BFF modules
    across 1263 executed cases. But .e is a child of this parent and 192 of
    192 route rows are owed at least one cell, on two columns no scan can
    fill -- a data classification for every route, and a keep/refactor/
    redirect/retire disposition for the 104 outside the studio tree. A parent
    cannot be closed over an open child by satisfying its own sentence. -->
    
    • I0.1.a Enumerate Isis routes from apps/isis/web, Oshun Studio, Yemaya Studio, BFF/proxy, redirects, route inventories, and deep links with source/mount/host. Evidence: route register. <!-- evidence: .operatorRoutes (27 apps/isis/web react-router routes w/ permission) + .studioRoutes (88 /studio/isis/* Next.js pages) + .bffRoutes (77 admin-isis-* BFF modules), each with source/mount/host. Yemaya Studio has no separate /studio/isis mount (Studio pages are the Oshun-Studio surface). generate-isis-route-register.mjs + 6 tests, wired into inventory/verify. --> <!-- CORRECTED IN PLACE (I0.1.f pass, 2026-08-16). Two of the 27 permission labels were wrong and the test asserted the wrong ones. operatorRoutes() walked App.tsx line by line, LATCHING underAuth and currentPermission on the way into a guard and never clearing them on </Route>. That turns "the nearest ENCLOSING guard" into "the most recent guard anywhere above", which is a different thing the moment a block closes: /404 and /*, declared at the top level of <Routes> after the RequireAuth block ends, were recorded as models-gated. They are public. A LATCH IS NOT A SCOPE.

      The parser now walks the TSX syntax tree with a guard STACK pushed and
      popped on element boundaries (self-closing elements enclose nothing, so
      they never contribute a frame). Every other row is byte-identical, which
      is the check that the rewrite reproduces the parts that were right.
      
      AND THE TEST HAD ENCODED THE DEFECT: `assert.deepEqual(publicRoutes,
      ['/login','/unauthorized'])` passed against the wrong output. Rewritten
      with the fix, plus a second test that pins the nesting semantics
      directly. Downstream: I0.1.d's matrix moves 2 routes from
      live-permission-gated to live-public (see its correction), and its test
      is now pinned to exact counts rather than `>=` bounds, which could not
      tell the corrected number from the wrong one.
      
      CORRECTED AGAIN IN PLACE (I0.18 pass, 2026-08-16). THE REGISTER WAS
      MISSING A WHOLE SURFACE. It described Isis as spanning three — the
      operator SPA, /studio/isis/* and the admin-isis-* BFF modules — and
      `apps/oshun/admin/src/app/isis/*` was in none of them. That is the
      Oshun ADMIN console: eight server-rendered operator pages
      (civitai-intake, comfy-nodes, lora-training, model-merging,
      output-gallery, runpod-endpoints, voice-cloning, workflow-editor), each
      guarded by `getAdminServerSession()` and
      `sessionCanEnterWorkspace(session, 'isis')`, each calling
      `/v1/admin/isis/*`.
      
      AND IT IS THE SURFACE THE OPERATOR API EXISTS FOR. `apps/isis/web` does
      not reference the Oshun BFF anywhere — it talks to the Isis
      micro-services on ports 3000-3002 — so a register without the admin
      console held 77 operator API modules whose UI it did not know about.
      The register now enumerates four surfaces; the count moves 192 → 200.
      
      THE TEST THAT SHOULD HAVE CAUGHT IT PASSED BOTH BEFORE AND AFTER. It
      was called "enumerates all three Isis surfaces" and asserted three
      counts plus the shape of the rows it was handed, so it could not tell
      "the register enumerates every surface" from "the register enumerates
      every surface it knows about". It now pins the SET OF HOSTS, which is
      the claim the cell makes, and a second test pins each admin page's
      session and workspace guard.
      
      DOWNSTREAM, FOUR CONSUMERS READ NAMED KEYS AND NOTICED NOTHING. Adding
      `adminRoutes` to the register left `isis-effective-route-matrix`,
      `isis-tier-decision-rows`, `isis-route-boundary-cases` and
      `isis-route-dispositions` byte-identical, and every drift check stayed
      green, because each reads `operatorRoutes` / `studioRoutes` /
      `bffRoutes` by name and never asks whether it has read them all.
      ALL FOUR CONSUMERS NOW COVER THE SURFACE. I0.1.d's matrix classifies
      the admin pages under a new `live-workspace-gated` class (200 routes)
      and REFUSES a register carrying a route-bearing key it does not
      classify, with a test that proves the refusal fires. I0.1.c's tier rows
      gain a fourth decider — `oshun-admin(sessionCanEnterWorkspace)` — and
      read 800 rows over 200 entries. I0.1.e's ledger gains eight
      `oshun-admin-page` rows and the pairing they make visible. I0.1.f
      generates forty admin cases whose expectations come from the DECLARED
      `OSHUN_ADMIN_WORKSPACE_MODEL.isis.requiredScopes` table and are executed
      against the real `canEnterAdminWorkspace`, with a control proving the
      authority refuses as well as grants.
      
      AND THE I0.1.g GATE FIRED, WHICH IS THE POINT OF IT. All eight new
      routes arrived without a disposition, so the gate failed until they were
      recorded in the accepted-undecided baseline — the deferral is now
      written down (112 undecided dispositions, 196 undecided canonical hosts
      over 200 routes) rather than silent.
      
      THE PAIRING THE FOURTH SURFACE MAKES VISIBLE: 7 of the 8 admin pages
      pair by name with an `admin-isis-*` module, and 70 OF THE 77 MODULES
      HAVE NO PAGE AT ALL. `/isis/workflow-editor` is the eighth page and
      there is no `admin-isis-workflow-editor`; its capability cell is
      recorded `owed` rather than filled from a near-miss like
      `admin-isis-workflow-definitions`.
      
      This cell stays `[x]`: its own sentence — routes with source, mount and
      host — is satisfied by the corrected register. -->
      
    • I0.1.b Resolve each Oshun Studio segment against STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS, including the current empty/ deny-all behavior. Evidence: segment decision rows. <!-- evidence: .segmentDecisions — all 87 distinct /studio/isis/* segments resolved against the EMPTY allow-list ⇒ every one is aaa-only-deny (fail-closed). Cross-checked with the documented 31-segment AAA catalog: 56 segments are blocked but NOT in the catalog (doc-vs-code coverage gap). -->

    • I0.1.c Resolve every route/action through the canonical generation-tier resolver for actor, tenant, audience, asset/workflow class, and host. Evidence: tier decision rows. <!-- evidence: isis-tier-decision-rows.json — 768 rows over 192 route entries (191 distinct host+path; the operator SPA mounts / twice, once under RequireAuth and once under RequirePermission(dashboard)) x 4 tiers. generate-isis-tier-decision-rows.mts + .test.mts (14 tests), wired into inventory + verify.

      THE PRIOR NOTE HERE SAID THIS CELL WAS "runtime, not static". IT IS NOT.
      `resolveGenerationTier` is a pure function of its argument -- no clock,
      no database, no network on any path -- so the generator IMPORTS THE REAL
      MODULE AND RUNS IT. Every verdict below is a return value.
      
      THE CANONICAL RESOLVER IS CONSULTED FOR ITS LABEL AND NEVER FOR ITS
      DECISION. 88 of 191 routes reach `resolveGenerationTier`; all 88 arrive
      through `resolveStudioBoundary`, which reads `.tier` to pick between two
      denials and never touches `.surfaceAllowlist`. TIER_ALLOWLISTS -- the
      table the whole module exists to hold -- decides ZERO routes. Its one
      production reader is `GET /v1/entitlements/generation/surface/:surface`,
      where the surface is a URL parameter: THE DECISION TABLE IS REACHABLE
      ONLY THROUGH A ROUTE THAT ASKS THE CALLER WHICH QUESTION TO ANSWER. A
      test binds that claim to `studio-boundary.ts` and `AaaGate.tsx`, so
      wiring the allowlist in fails here rather than leaving a confident row
      describing code that moved underneath it.
      
      A DIMENSION YOU CANNOT VARY IS A DIMENSION NOBODY READ. Of the five the
      cell names, three are fields of `GenerationEntitlement` (actor=userId,
      tenant=tenantId, audience=persona) and two have no argument position at
      all (asset/workflow class, host). ZERO are read. Established by a
      differential probe rather than by reading the parameter list -- hold the
      tier fixed, vary one field, re-ask all 28 surfaces and all 31 AAA-only
      segments, compare the whole answer set. A DECLARED, CARRIED, DROPPED
      FIELD LOOKS EXACTLY LIKE A READ ONE UNTIL YOU VARY IT. The only field
      `resolveTierFromTags` reads is `tags`, which is not among the five.
      
      AND A PROBE THAT ONLY EVER SAYS NO IS NOT A PROBE. The criterion is
      per-tier on purpose: asking "did anything move anywhere" counts the four
      tiers' own difference as evidence about the probed field. A control runs
      the same probe over `tags` and REQUIRES a positive, plus a nonexistent
      field that must come back negative.
      
      THE ESTATE HAS FOUR AUTHORITIES AND THE RESOLVER'S DOCBLOCK CLAIMS TO BE
      THE ONE. It says "BFF, Admin, and Studio middleware all consult it
      before rendering any generation surface." Measured: the 77 admin-isis-*
      BFF modules consult `hasAdminScope` over three scopes (admin:*,
      admin:studio, admin:workspace:isis) and none imports the resolver; the
      operator SPA consults `RequirePermission` over ten permission names; the
      Studio middleware consults the boundary. Four alphabets, and 7 of the
      resolver's 28 surfaces name a route in the register at all.
      
      A REFUSAL AND A RELOCATION READ THE SAME AND ARE NOT. 14 disagreements
      between the boundary and the deny-by-default guard, ALL host-only: the
      guard says allow, the boundary says `render-yemaya-cta`. That branch is
      reached only for aaa-creator/operator-admin and redirects to a page whose
      copy says the caller is already entitled, so it AFFIRMS entitlement and
      moves the host -- the dimension neither can express. Classifying it as a
      refusal was the first version of this measurement and it hid the
      escalation: a POSITIVE CONTROL (inject `audit-trail`, an operator-admin
      -only surface, as a studio segment) passed the ratchet, because "guard
      denies + boundary sends you to the CTA" scored as agreement. Corrected;
      the control now fails two ratchets. Entitlement disagreements: 0, and
      that zero is now a gate, because `resolveStudioBoundary` treats
      aaa-creator and operator-admin identically while TIER_ALLOWLISTS
      separates them by six admin surfaces.
      
      THE GATE THAT PLUGS THE LEAK SITS BEHIND THE DOOR THAT ALREADY REFUSES
      EVERYONE. `<AaaGate>` exists because route middleware cannot see inside
      a hub that embeds an AAA workspace as a lane. All 6 embeds are mounted
      by pages under `/studio/isis/*`, and with the allow-list empty the
      boundary refuses every one of those segments for all four tiers before
      the page renders. 6 of 6 unreachable in production; the only path that
      reaches them is OSHUN_ENABLE_STUDIO_ISIS_LANE_E2E_BYPASS=true plus
      ?__oshunStudioIsisLaneE2E=1 on localhost -- the test switch that turns
      the outer guard off. The env-var name is read from `proxy.ts` by the
      generator, and a truncating character class ([A-Z_] stops at the digit
      in E2E) shipped a variable nobody could set until a test caught it.
      
      AND WHERE THEY BOTH RUN, THEY DISAGREE. 6 rows: for aaa-creator and
      operator-admin on gpu-worker / animatediff-lightning / rife-interpolation,
      `proxy.ts` redirects to /aaa-upgrade and `<AaaGate>` renders the
      workspace. Same actor, same segment, same host, opposite answers. Both
      cannot be the product rule. Every instance is currently unreachable, and
      the row records that, because a contradiction you cannot reach is a
      different severity from one you can -- a test fails the day one becomes
      reachable. -->
      <!-- Retested: 14 tests pass; `--check` reports no drift; the positive
      control (an operator-admin-only surface injected as a studio route) fails
      the two ratchets and the surface-binding known-value, then the register
      is restored byte-identical. -->
      
    • I0.1.d Classify source-mounted but unreachable, feature/provider-gated, redirected, proxied, embedded, compatibility, and live routes by effective middleware behavior. Evidence: effective route matrix. <!-- evidence: isis-effective-route-matrix.json — all 192 routes classified by effective behavior from the I0.1.a/b register + studio-boundary rule: 88 /studio/isis/* = gated-concealed (source-mounted but unreachable for the contemplative product via the EMPTY allow-list ⇒ hard-block-404), 77 live operator-api (admin-isis-* BFF), 24 live-permission-gated + 2 live-public (operator SPA). No blank cells; no redirect/proxy/compat shims in the register. generate-isis-effective-route-matrix.mjs + 5 tests. --> <!-- CORRECTED IN PLACE (I0.1.f pass, 2026-08-16). The operator SPA split is 4 live-public + 1 live-auth-gated + 22 live-permission-gated, not 2 + 24. This matrix derives its classes from the I0.1.a register, and the register mislabelled /404 and /* as models-gated because its parser latched the last guard it had seen instead of tracking enclosure. Both are public. The 88 gated-concealed and 77 live-operator-api rows are unchanged, and the finding stands.

      THE TEST COULD NOT HAVE CAUGHT THIS: it asserted `live-public >= 2` and
      `live-permission-gated >= 10`, which both wrong numbers and both right
      numbers satisfy. A DERIVED COUNT ASSERTED AS A LOWER BOUND CANNOT
      DISTINGUISH ITS SOURCE BEING WRONG FROM ITS SOURCE BEING RIGHT. Pinned
      to exact values. -->
      
    • I0.1.e Assign canonical owner/host/audience/capability, allowed controls, data classification, and keep/refactor/redirect/retire disposition. Evidence: completed route rows. <!-- STILL OPEN, DELIBERATELY. The rows are BUILT and every cell is filled: isis-route-dispositions.json, 192 routes x 8 cells, from generate-isis-route-dispositions.mts + .test.mts (14 tests), wired into inventory + verify. The evidence clause asks for COMPLETED rows and 0 of 192 are complete, because two columns cannot be assigned by anything that reads code.

      A ROW IS COMPLETE WHEN EVERY CELL EITHER CARRIES A DECISION OR NAMES WHO
      OWES ONE, and the three provenances are the whole design: `recorded`
      (the estate decided, and the cell cites where), `derived` (computed from
      a recorded fact by a rule the cell states), `owed` (nothing decides it;
      the cell names the alphabet the answer must come from and stops). A cell
      filled by this generator's opinion would be WORSE THAN AN EMPTY ONE --
      it reads like a ratified answer and nobody was ever asked. A test
      enforces that an `owed` cell cannot carry a decision and a decided cell
      cannot claim to be owed, in both directions.
      
      FILLED, WITH SOURCES: owner and owning domain on all 192 from
      `domains.json` (longest-prefix match, because `apps/oshun/` and
      `apps/oshun/bff` would otherwise race and the answer would depend on
      registry order); host-as-built on all 192; audience on all 192;
      capability on all 192 -- 140 admin-isis BFF endpoints parsed from the
      modules, the operator SPA's client operations from
      isis-page-inventory.json, the StudioIsis*Workspace components each
      studio page renders; allowed controls on all 192 (for a BFF module: its
      non-GET endpoints).
      
      THE ISIS DOMAIN OWNS 27 OF ITS 192 ROUTES. 165 live in another domain's
      applications -- 88 studio pages in `apps/oshun/web` and 77 modules in
      `apps/oshun/bff`.
      
      A RATIFIED DESTINATION IS STILL A CLAIM ABOUT A PLACE. The §24.11
      Surface Inventory is a real ratified record -- 25 surfaces, each with a
      canonical host and an entitlement -- and it assigns 16 of them to Yemaya
      Studio. Yemaya Studio's own route tables have 39 routes across BOTH
      hosts the document names (`studio-web` AND `studio-desktop`; checking one
      would be an absence claim about a place nobody looked) and NAME NONE OF
      THE 16. The claim is scoped to names on purpose and says so in the
      artifact: whether `/render-farm` is the RunPod operator dashboard under
      another label is a judgement, so the full destination list is published
      beside the number instead of being folded into it. A control proves the
      matcher can match when a route does exist, so "0 of 16" is not an
      artifact of the matcher.
      
      A KEYWORD CENSUS WOULD HAVE ANSWERED THIS WRONG. Grepping the
      destination for the surface names finds `gaussian` -- in `feGaussianBlur`,
      an SVG filter -- and `runpod` in a render-farm test. Matching is on route
      tables, not free text.
      
      THREE ARTIFACTS SPELL ONE THING THREE WAYS. Binding a route to a
      ratified row is VERBATIM only, after splitting the document's label on
      `/` (its own punctuation: "Model merger / comparison" is two surfaces on
      one row). 4 of 88 studio routes bind. The 3 surfaces that have a route
      and a plausible row but no shared name -- `3d-generation`,
      `model-comparison`, `music-generation` -- are listed by name with their
      loose candidates, unbound. A substring rule loose enough to catch "Model
      merger / comparison" from `model-comparison` also catches `comparison`
      out of any row that mentions comparing, and then the binding is the
      matcher's opinion rather than the document's.
      
      A REDIRECT IS NOT AN AUDIENCE. The first version of the audience rule
      was "tiers whose verdict is not hard-block-404", which made
      aaa-creator+operator-admin the audience of all 88 studio routes. They are
      307'd OFF the route. Corrected to "tiers whose verdict is render-as-is",
      and every studio route now reads `none-renders` with the redirect
      recorded in the source. THAT IS HOW A ROUTE WITH NO READERS REPORTS TWO.
      
      WHAT IS OWED, AND WHY NEITHER IS WORK SOMEBODY COULD JUST DO.
      
        data classification  0 of 192. The alphabet EXISTS and the estate
            already requires it: DATA_CLASSIFICATIONS (public/internal/
            confidential/restricted/personal) in workbench-kit `registry.ts`,
            and `routes.ts` makes `dataClassification` a required field of a
            registered surface. No Isis route is registered through it, and
            `libs/isis` + `apps/isis` contain zero occurrences of the field.
            Classifying 192 routes is a data-owner's judgement about what each
            one discloses; a scan can name the alphabet and cannot pick a
            member.
        disposition  88 of 192 recorded, 104 owed. The 88 studio pages carry a
            RECORDED `retire` with two citations the tests re-check -- the
            §24.11 "Routes Removed From Oshun Studio" section and the
            `STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS` comment calling the tree the
            "legacy provider-machinery tree" -- plus the named replacement
            (`/studio/generation/*`, `/studio/generation-gallery`). The 27
            operator SPA routes are owed by I0.3, which is the open cell that
            maps `apps/isis/web` pages to keep/refactor/merge/retire; assigning
            them here would pre-empt that cell with an invented decision. The
            77 BFF modules are owed by nobody yet -- nothing in the estate
            records a disposition for them, and that absence is the finding.
        canonical host  4 of 192 recorded. The ratified inventory covers
            generation SURFACES; it names no canonical host for the operator
            factory SPA, for the admin-isis-* BFF modules, or for the 84
            studio segments that are not one of its surfaces.
      
      4 rows are owed nothing but a data classification. -->
      <!-- Retested: 14 tests pass, `--check` reports no drift, and the
      adversarial stub scan over both new files returns zero hits. -->
      
    • I0.1.f Add generated route cases that prove each discovered route is allowed only for its exact tier/host and concealed/refused elsewhere. Evidence: complete route-boundary test. <!-- evidence: isis-route-boundary-cases.json — 1263 generated cases, 0 uncovered register routes, 0 uncovered BFF modules, 140 of 140 endpoints. generate-isis-route-boundary-cases.mts + .test.mts (9 tests), apps/oshun/bff/src/tests/isis-route-boundary.spec.ts (3 tests, all 840 BFF cases), and the generated sweep added to middleware-studio-boundary.test.ts (12 tests). All green.

      AN EXPECTATION COMPUTED BY RUNNING THE CODE UNDER TEST CANNOT FAIL.
      Every case's `expected` comes from a DECLARED rule and its `authority`
      is the code that will be executed: the §24.11 document for the studio
      tree, `ROUTE_PERMISSIONS`/`RequireAuth`/`RequirePermission` for the
      operator SPA, `createAuthPreHandler` + the modules' scope alphabet for
      the BFF. A control test enforces the separation — a studio case may not
      cite `studio-boundary.ts` as its expectation source.
      
      THE SPLIT: 352 studio (88 routes x 4 tiers), 71 operator SPA (each
      guarded route against its own permission, ANOTHER REAL PERMISSION FROM
      THE SAME ALPHABET rather than an invented string, and an empty set), and
      840 BFF (140 endpoints x anonymous / zero-scope / another domain's admin
      scope / each of the three Isis admin scopes).
      
      THE SWEEP REFUTED ITS OWN DECLARED RULE ON 140 ENDPOINTS. The zero-scope
      identity was written as 403 from `hasAdminScope`, and every endpoint
      answered 401. A GUARD IS A CHAIN AND THE FIRST LINK ANSWERS FIRST:
      `payloadToResolution` treats a token with an empty scope set as an
      INVALID token, so a zero-scope caller never reaches the module's scope
      check. The case is corrected to 401 with the earlier declaration cited,
      and the finding is recorded in the artifact:
      AN IDENTITY THAT IS AUTHENTICATED BUT UNPRIVILEGED HAS NO REPRESENTATION
      AT THIS BOUNDARY -- it collapses into unauthenticated. The 403 half is
      carried by `admin:workspace:aja`, a real scope from another domain, and
      it passes on all 140.
      
      A LIST OF OBSERVED ROUTES CANNOT OBSERVE A ROUTE ADDED AFTER IT WAS
      TYPED. The middleware suite's `observedLegacySegments` was 56 hand-typed
      strings tested at ONE tier (contemplative), and the `specList` beside it
      is 31 more. The generated sweep drives all 88 register segments at all
      four tiers through the real `proxy()`, so the 56-item list is removed --
      after checking that every one of its segments is in the register, because
      a hand-typed list is only safe to drop once something enumerates a
      superset of it. The doc-catalog `specList` stays: it checks the document,
      not the register, which is a different question.
      
      AND A SWEEP CAN BE REFUSED BY SOMETHING THAT IS NOT THE BOUNDARY. The
      first middleware run returned 429: `proxy.ts` rate-limits by IP before
      the §24.11 guard, so 352 requests from one address stop being boundary
      observations. Each case is a distinct visitor and now carries a distinct
      `x-forwarded-for`, and an explicit assertion fails on any 429 rather than
      letting a rate-limited request read as a pass. The BFF runner has the
      same guard for the same reason.
      
      ON "TIER/HOST": no Isis route has both a tier and a host to select
      between. The studio tree is decided by tier at one host; the operator SPA
      by permission; the BFF by scope. The cases test the authority that
      actually decides each route, and I0.1.c is the measurement of why there
      is no third dimension to test. -->
      <!-- Retested: 9 + 3 + 12 tests pass, `--check` reports no drift, the
      three new generators typecheck clean under strict tsc, and the
      adversarial stub scan returns only the documented `:param` URL
      placeholder. -->
      
    • I0.1.g Fail inventory drift when a new Isis route or allowed segment lacks a tier/host disposition. Evidence: deliberate route addition test. <!-- evidence: isis-route-disposition-gate.mts (exit 1 with named violations) + isis-route-disposition-baseline.json (104 accepted undecided dispositions, 188 undecided canonical hosts) + .test.mts (9 tests, seven of them deliberate additions). Wired into verify.

      A `--check` DRIFT GATE CANNOT ANSWER THIS. Drift-checking compares a
      committed artifact against a clean regeneration, and a regeneration that
      happily emits a new row saying `owed` is not drift -- it is AGREEMENT.
      The gate asks a different question: which routes are ALLOWED to be
      undecided, and is this one of them.
      
      AND THE ANSWER IS A SET, NOT A COUNT. A COUNT CANNOT SEE A NEW ROUTE
      ARRIVE WHILE AN OLD ONE IS DECIDED -- the total holds and the new one is
      undecided anyway. The baseline lists route keys and is checked in BOTH
      directions: a route owed a disposition that is not listed fails, and a
      listing whose route has since been decided or removed also fails, so the
      ratchet tightens instead of becoming a permanent licence a future route
      could inherit by reusing a path.
      
      AN ALLOWED SEGMENT IS HELD HIGHER THAN A BLOCKED ONE. A segment added to
      STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS renders on the contemplative product,
      so its route needs a RECORDED disposition and no baseline entry excuses
      it. A test proves the baseline cannot buy it off: adding the route to the
      accepted list silences the two deferral violations and leaves the
      reachability one standing.
      
      THE DELIBERATE ADDITION WAS ALSO RUN END TO END, not only against the
      pure function. A real `page.tsx` was written into
      `apps/oshun/web/src/app/studio/isis/gate-e2e-probe/`, the register went
      88 -> 89 studio routes, the dispositions went 192 -> 193 rows, and the
      gate exited 1 naming the new path. Then removed, and all five artifacts
      re-checked byte-identical.
      
      AND THAT ONLY WORKS BECAUSE THE CENSUS WAS FIXED FIRST. The register
      enumerated studio pages and BFF modules with `git ls-files`, which lists
      TRACKED files only -- so a page that exists on disk and has not been
      `git add`ed is invisible, and THE GATE PASSED FOR EXACTLY THE DEVELOPER
      WHO WAS ADDING THE ROUTE IT EXISTS TO CATCH. Measured on the probe:
      `git ls-files` saw 0, `--cached --others --exclude-standard` saw 1. The
      same fix removed a `catch { pages = [] }` on both enumerations, which
      made "git is broken" and "there are no Isis pages" produce the same
      register.
      
      THE GATE FIRED ON THE DIMENSION THAT WAS ACTUALLY UNDECIDED. The probe
      produced ONE violation -- no ratified canonical host -- and not a
      disposition violation, because the studio tree's `retire` is recorded for
      the tree and a new page in it inherits that. A gate that had reported
      both would have been reporting its own rules rather than the estate.
      
      AND THE GATE'S FIRST DRAFT FAILED 77 TIMES ON NOTHING. It compared the
      disposition rows' `path` (a BFF module name) with the case table's
      `route` (an endpoint) and declared all 77 modules untested. A COVERAGE
      CHECK ACROSS TWO ARTIFACTS IS A JOIN, AND A JOIN NEEDS A KEY BOTH SIDES
      CARRY. The covered set now carries both identifiers, and a control test
      asserts both are present. -->
      
  • I0.2 Ratify operator, AAA-creator, curated-creator, and contemplative capabilities separately. Evidence: signed audience-capability matrix. <!-- THE MATRIX IS BUILT AND EXECUTED; NOTHING HERE IS SIGNED, so the parent and all seven children stay [ ]. isis-audience-capability-matrix.json — 160 cells over 40 capabilities x 4 audiences, every cell a return value from the estate's own authority (checkSurfaceAccess, checkCardEntitlement, validateBulkAction, resolveGenerationCreatorTier) run over the real card catalog and the real scope declarations. generate-isis-audience-capability-matrix.mts + .test.mts (10 tests, 5 of them ratchets on findings), wired into inventory + verify.

    SIX ALPHABETS NAME THE SAME FOUR AUDIENCES AND TWO OF THEM CANNOT NAME ONE
    OF IT. `GENERATION_TIERS`, `GENERATION_CREATOR_TIERS`,
    `checkCardEntitlement.creatorTier` and `GalleryView['tier']` all carry four
    members. `CardSpec.minimumTier` and `AuthorizationCheck.tier` carry three,
    and the missing member is `contemplative` -- the contemplative product's
    own audience. A TYPE THAT CANNOT NAME AN AUDIENCE CANNOT REFUSE IT: those
    four cells read `not-expressible`, which is a different fact from `deny`
    and is recorded as one.
    
    FINDING 1 -- THE PRIMARY AUDIENCE IS ENTITLED TO SURFACES ON WHICH EVERY
    CAPABILITY REFUSES IT. `checkSurfaceAccess` ALLOWS a contemplative caller
    onto `curated-image`, `curated-living-scene` and `curated-voice-clip`.
    `checkCardEntitlement` REFUSES that caller on all 8 catalog cards, because
    every card's `minimumTier` is at least curated-creator and the type cannot
    express anything lower. The §24.11 Surface Inventory agrees with the cards
    (curated-creator+). So the resolver whose docblock calls it "the single
    source of truth" is the one out of step, and the effect today is
    fail-closed -- but `GET /v1/entitlements/generation/surface/:surface`
    answers from the resolver, so THE API THAT ANSWERS THE ENTITLEMENT
    QUESTION AND THE CODE THAT ENFORCES IT DISAGREE ABOUT THE PRODUCT'S
    PRIMARY AUDIENCE. Which of the three is wrong is a product decision.
    
    FINDING 2 -- THE SAME PREFIX MEANS "ANY ADMIN" IN ONE MODULE AND "THESE
    THREE" IN ANOTHER. Every `admin-isis-*` BFF module accepts exactly
    `admin:*`, `admin:studio`, `admin:workspace:isis`.
    `resolveGenerationCreatorTier` admits ANY scope matching
    `scope === 'admin:*' || scope.startsWith('admin:')` as the top Isis
    audience. Executed: `admin:workspace:aja` and `admin:workspace:yemaya`
    resolve to `operator-admin`, and I0.1.f already measured that both are
    refused 403 by all 140 Isis operator endpoints. ONE IDENTITY, REFUSED BY
    EVERY ISIS OPERATOR ROUTE AND ADMITTED AS ISIS'S TOP GENERATION TIER --
    which is what unlocks every curated card's `minimumTier`. Narrowing the
    rule is an authorization change with a blast radius outside this cell.
    
    FINDING 3 -- A DECLARED ENFORCEMENT LAYER WITH NO PRODUCER. The card gate's
    docblock says it "is consulted at three layers: component (UI render),
    route (Studio router), and BFF". Measured across every production source in
    `apps/oshun/{bff,web}/src` and `libs/isis`: `bff` has 3 producers,
    `component` has 1, and `route` has NONE -- only a test ever passes it. The
    authorization survives (the BFF enforces), what is missing is the
    route-level concealment, so a refused card's page renders a denial instead
    of not existing. That is the "without relying on UI hiding" question in
    .f, answered: at the route layer, the hiding is all there is.
    
    FINDING 4 -- AN AUTHORIZATION LADDER WHOSE AUDIENCE ARGUMENT IS A LITERAL.
    `validateBulkAction` distinguishes four audiences across four actions
    (re-tag and send-to-editorial need curated-creator or better; revoke-consent
    and takedown need operator-admin). All three production call sites pass
    `tier: 'operator-admin'` as a constant -- `output-gallery-store.ts:170`,
    `admin-isis-output-gallery-actions.ts:124`, `server.ts:3389` -- so the
    ladder can never refuse. The routes are admin-scope gated, which is why
    this is not a hole today; but an admin SCOPE is not an operator-admin TIER
    (finding 2 is the same seam), and three of the ladder's branches cannot
    fire.
    
    WHAT IS OWED, PER CHILD. .a-.d ask to DEFINE each audience's actors,
    operational purpose, canonical host, visibility, authority and exclusions.
    The matrix supplies every cell the estate DECIDES; it cannot supply an
    operational purpose or an exclusion nobody wrote down, and I0.1.e already
    measured that the ratified canonical host for 16 of 25 surfaces names a
    destination that holds none of them. .e asks for zero blank cells across
    eight applicability kinds (exposure/observe/control/review/approve/publish/
    recall/support) -- that alphabet does not exist anywhere in the estate, so
    there is nothing to derive it from. .f is MEASURED but not RESOLVED: its
    four findings are two authorization changes and one product decision. .g
    is a named signature from seven parties, which no scan performs.
    
    NO DECISION IS INVENTED HERE. -->
    
    • I0.2.a Define operator actors, operational purpose, canonical host, raw workflow/model/provider/resource visibility, mutation/override/emergency authority, and exclusions. Evidence: operator matrix rows.
    • I0.2.b Define AAA-creator actors, Yemaya project/asset context, advanced workflow/model/training/3D/voice/music controls, release handoff, and forbidden fleet/admin authority. Evidence: AAA rows.
    • I0.2.c Define curated-creator actors, approved task/asset cards, high- level controls, gallery/editorial handoff, and forbidden raw details. Evidence: curated rows.
    • I0.2.d Define contemplative actors, permitted narrow assistive generation, consent/safety constraints, output use, and prohibited AAA/ operator capabilities. Evidence: contemplative rows.
    • I0.2.e Decide exposure/observe/control/review/approve/publish/recall/ support applicability per canonical capability and audience. Evidence: zero blank matrix cells.
    • I0.2.f Resolve shared capabilities whose behavior/data detail differs by audience without relying on UI hiding. Evidence: contract and authorization variants.
    • I0.2.g Obtain named Isis, Yemaya, Oshun Studio, product, safety/rights, security/privacy, and architecture approval. Evidence: signed matrix.
  • I0.3 Map apps/isis/web pages to keep, refactor, merge, or retire. Evidence: every page and route has one disposition. <!-- SUPERSEDED (2026-08-16). The old note said b-g were "runtime/design/ human and stay unchecked". .b is now closed and the runtime claim was WRONG: apps/isis/web has no fixture mode at all, so fixtures-off is the only mode it has, and what a page does in it is decided by whether the service listening at the base URL the client was handed serves the endpoint the page calls. That is answered by constructing each service's real router, which needs no browser and no running stack.

    THE PARENT STAYS OPEN ON .c-.g, WHICH ARE DISPOSITIONS AND AN APPROVAL.
    Its own evidence clause -- "every page and route has one disposition" --
    is NOT yet satisfied: 0 of 26 pages carry a keep/refactor/merge/retire
    disposition. The walkthrough matrix (a behaviour class per page, with
    121 operations resolved to a service and a verdict) is the input those
    cells were missing, not a substitute for them. -->
    
    • I0.3.a Enumerate every page/layout/component route, navigation entry, deep link, API/provider/store/job dependency, role, state, and test. Evidence: source-linked page inventory. <!-- evidence: .pages — 26 apps/isis/web pages, each source-linked to its App.tsx route + permission role, with isis-client apiOperations (client.<ns>.<method>, 118 distinct), react-query query/mutation counts, localStateHooks, hooks[], hasTest. Surfaces 2 unrouted page components (JobSubmissionStudio, WorkflowControlBlendPanel) + 6 pages without a test. generate-isis-page-inventory.mjs + 7 tests, wired into inventory/verify. -->

    • I0.3.b Walk each reachable page with fixtures off and classify live, partial, fixture, not-configured, dead, duplicate, and unreachable behavior. Evidence: page walkthrough matrix. <!-- evidence: isis-page-walkthrough.json — 26 pages x a behaviour class with its reason, 121 client operations each resolved to a verb, a path, the service the SPA actually sends it to and that service's verdict. generate-isis-page-walkthrough.mts + .test.mts (22 tests, 6 of them ratchets on findings), wired into inventory + verify.

      "WITH FIXTURES OFF" NAMES A MODE THIS APP DOES NOT HAVE, AND THAT IS THE
      FIRST MEASUREMENT. `apps/isis/web` declares no fixture, mock, MSW or
      env-gated fake backend anywhere: 0 of 26 pages carry a fixture signal.
      Every page reads through one `IsisClient` built in
      `IsisClientProvider.tsx` from three base URLs, so fixtures-off is the
      only mode the app has, and a page's behaviour in it is decided by
      whether the service listening at the URL that client was handed serves
      the endpoint the page calls. That is a question about delivery, and it
      is answered by CONSTRUCTING each service's real Hono router and reading
      `app.routes` — not by a regex over route files.
      
      AN ENDPOINT THAT EXISTS AND AN ENDPOINT THAT ARRIVES ARE DIFFERENT
      CLAIMS, AND THE SPA REACHES ONE OF ITS FOUR NAMESPACES.
      
        generation  -> generationApiUrl    :3001 -> isis-workflow-registry
        models      -> generationApiUrl    :3001 -> isis-workflow-registry
        workflows   -> workflowRegistryUrl :3003 -> nothing listens
        outputs     -> outputRegistryUrl   :3002 -> isis-output-registry  OK
      
      The defaults govern, and that is checked rather than assumed: the three
      `VITE_ISIS_*_URL` names appear in `apps/isis/.env.example` only under
      its "referenced by this app but NOT in the canonical .env.example"
      list, so nothing in the estate assigns them a value. Meanwhile
      `isis-generation-api` — 59 routes, the service that owns
      `/api/v1/jobs` and `/api/v1/models` — defaults to PORT 3000, WHICH NO
      CLIENT BASE URL NAMES.
      
      86 OF 121 OPERATIONS DO NOT ARRIVE, AND ALL 121 ARE BUILT. Asking every
      service the same question, not just the intended one, is what separates
      misdelivery from absence: 0 operations are fictional. 47 reach a service
      that answers 404, 38 reach a port with no listener, 35 are served, and 1
      is served by a service it did not mean to ask.
      
      A PORT FIX WOULD NOT MAKE THE `workflows` NAMESPACE WHOLE.
      `workflows.run`, `workflows.getExecution` and
      `workflows.getExecutionHistory` are served ONLY by
      `isis-generation-api`, never by the workflow registry. The client models
      `workflows` as one host and the estate splits it across two, so no
      single base URL can be correct for all 37 of its operations.
      
      THE HEALTH PANEL IS GREEN FOR A SERVICE IT NEVER CONTACTED.
      `client.checkHealth()` is one call that fans out to `GET /health` on all
      three base URLs. `/health` exists on all three services, so the
      generation probe SUCCEEDS — answered by the workflow registry — and
      DashboardPage labels that answer "Generation API: Healthy". The
      workflows probe reports unhealthy because nothing is at :3003. An
      operator's status board is right about outputs, wrong about generation,
      and correct-by-accident about workflows.
      
      PER-PAGE: 15 not-configured, 9 live, 2 partial, and zero of fixture,
      dead, duplicate or unreachable. EVERY page that touches generation,
      models or workflows is not-configured; the only three that work against
      a backend — OutputsPage, OutputsParityPage,
      OutputComparisonWorkspacePage — are the ones that talk to the output
      registry. The remaining 6 "live" pages call no backend at all, and the
      row says so rather than borrowing the same word silently.
      
      THE TWO UNROUTED COMPONENTS ARE NOT UNREACHABLE. I0.3.a recorded
      JobSubmissionStudio and WorkflowControlBlendPanel as unrouted, which is
      true; the walkthrough adds that both are RENDERED BY a routed page
      (JobsPage and WorkflowDetailsPage), so neither is dead and neither needs
      its own disposition. `unreferenced` is 0 and so is `dead`.
      
      FOUR WAYS THE INSTRUMENT WAS WRONG BEFORE IT WAS RIGHT, ALL FIXED AND
      ALL NOW UNDER A CONTROL TEST. An unauthenticated probe returns 401 for
      EVERY path, so "absent" and "refused" are indistinguishable until auth
      is satisfied — the walkthrough forces `allowAnonymous` so the ROUTER
      answers. `client.checkHealth()` is not namespaced, and keying on
      `client.<ns>.<method>` reported DashboardPage as calling nothing.
      ApiParityMonitorPage receives the client as a `currentClient` parameter,
      and keying on the literal name `client` reported it as calling nothing
      either. Widening the receiver to any identifier then swept in
      `workflow.outputs.map(...)`, an array — so the receiver is constrained
      to bound clients plus the `*[Cc]lient` convention, and every rejected
      receiver is PRINTED IN THE ARTIFACT rather than dropped.
      
      .c-.g stay open and are not blocked on a runtime any more. They are
      Assign/Reconcile cells: keep/refactor/merge/retire is a disposition
      somebody owes, and this matrix is the input they were missing. Note that
      the wiring finding is not itself a disposition — a page is
      `not-configured` because of how the SPA is addressed, not because of
      anything the page is or does, so "retire the not-configured pages" would
      be the wrong reading of this table. -->
      <!-- Retested: 22 tests pass, `--check` reports no drift, and the
      adversarial stub scan over both new files returns zero hits. -->
      
    • I0.3.c Assign keep where audience/host/capability and live implementation are correct; record missing hardening/proof separately. Evidence: keep rows.

    • I0.3.d Assign refactor with target shared/domain boundary, contract/data/ authorization changes, migration slice, compatibility, and rollback. Evidence: refactor rows.

    • I0.3.e Assign merge with canonical destination, source-only behavior to preserve, deep-link/state migration, consumer impact, and deletion gate. Evidence: merge rows.

    • I0.3.f Assign retire with replacement/redirect/export/retention/notice/ owner/date and verify no unique capability is lost. Evidence: retirement rows.

    • I0.3.g Reconcile source routes/navigation to dispositions and obtain operator owner approval. Evidence: zero undispositioned pages/routes.

  • I0.4 Map existing Oshun Studio generation cards and gallery routes. Evidence: no planned route duplicates an existing approved surface.

    • I0.4.a Enumerate generation cards, entry actions, contexts, high-level controls, tier resolver calls, BFF/API operations, output destinations, and tests. Evidence: card inventory. <!-- evidence: isis-curated-surface-inventory.json .cards — all 8 CURATED_CARD_CATALOG entries with cardId/kind/displayName/binding/ minimumTier/tenantAllowlist/killSwitchActive/provenance, the declared high-level controls per kind parsed from the card-input interfaces (e.g. ambient-audio = durationSeconds, loopable, mood, tempoBpm), the BFF endpoints of the 3 modules that consult the card gate, the executor each route is wired to, and the tests naming each kind (every card has at least one). generate-isis-curated-surface-inventory.mts + .test.mts (13 tests), wired into inventory + verify.

      OUTPUT DESTINATION IS A SEAM, AND THAT IS THE ANSWER RATHER THAN A GAP.
      Each route takes a `*Executor` whose default throws `*_not_configured`
      and 503s; `server.ts` injects the real one. Read from the wiring:
      curated-card, music and Nyx-3D get real executors,
      `registerIsisGenerationRoute` keeps `notConfiguredProviderExecutor`. A
      default whose only behaviour is to throw is an honest fail-loud seam,
      so the surface it serves is `not-configured`, not `live`, and the row
      says which.
      
      TWO SURFACES ARE ENTITLEMENT-CHECKED AGAINST ANOTHER SURFACE'S CARD.
      There is no `music` or `nyx-3d` card kind. `music-route.ts` checks
      against `ambient-audio` and `nyx-3d-route.ts` against `illustration`.
      A KILL SWITCH IS A CLAIM ABOUT ONE SURFACE: `killSwitchActive` and
      `tenantAllowlist` on those two cards now decide for two surfaces each,
      and the borrowing surface has no control of its own. Proven by running
      the gate with the switch flipped, not by reading it. -->
      
    • I0.4.b Enumerate gallery/list/detail/compare/replay/promotion/history/ provenance/routes and deep-link state. Evidence: gallery route inventory. <!-- evidence: .gallery — ONE route, /studio/generation-gallery, carrying 7 surface families enumerated from the client's own data-* markers: branch (replay-with-tweak), compare (grid + slots + status), filter (10 facets), live (provenance inspector), output, gallery, action.

      THE CELL ASKS FOR ROUTES AND THE ESTATE HAS ONE. The compare grid, the
      branch drawer and the provenance inspector are in-page state, not
      routes, and the inventory says so rather than inventing route names for
      them.
      
      AND THE DEEP-LINK STATE IS EMPTY. Measured: `useSearchParams`,
      `searchParams`, `useRouter`, `pushState` and `window.location` are all
      absent from the gallery client. A COMPARE GRID NOBODY CAN LINK TO IS A
      COMPARE NOBODY CAN SHOW ANYONE -- ten filters, a comparison and a
      provenance selection, none of which survives a reload or can be sent to
      a colleague. That is the answer to "and deep-link state", and it is a
      finding rather than an omission in the inventory. -->
      
    • I0.4.c Classify each surface by authorized audience, asset/workflow class, live/fixture/provider state, backing contract/service/store, and current gaps. Evidence: surface matrix. <!-- evidence: the same artifact's per-route and per-card columns. Audience: tierSource per route (session / request-header / none) plus the 160-cell audience matrix from I0.2. Asset/workflow class: binding.workflowClassId and provenance.workflowClassId per card. Live/fixture/provider state: read from the executor wiring above -- 3 of 4 generation routes configured, 1 deliberately not. Backing service: the BFF module and endpoints each route reaches. Gaps: the three sections below.

      ONE CURATED ROUTE TAKES ITS AUDIENCE FROM A CALLER-SUPPLIED HEADER, ON A
      PUBLIC PATH. `/studio/generation/living-scene` is in `proxy.ts`'s
      PUBLIC_PATHS and reads six entitlement inputs straight off the request:
      `x-oshun-tier`, `x-oshun-living-scenes-enabled` (absent ⇒ ENABLED),
      `x-oshun-living-scene-templates`, `-frozen-workflow-classes`,
      `-locales`, `x-oshun-locale`. THE ESTATE ALREADY FIXED THIS ONE LAYER
      DOWN: `libs/shared/bff-kit/src/entitlements.ts` honours `x-oshun-tier`
      only outside production, precisely so "a client can no longer
      self-assert 'pro'". The page has no such guard. The BFF still refuses
      the submission, so this is what the surface OFFERS rather than what it
      grants -- which is exactly the "without relying on UI hiding" question
      I0.2.f asks. It is also the surface that advertises
      `data-living-scene-route-entitlement-gate="route"`, i.e. it claims to BE
      the route-layer gate I0.2 measured as having no producer, while taking
      its input from the caller.
      
      THE RATIFIED INVENTORY AND THE CARD ALPHABET DO NOT LINE UP. §24.11
      ratifies 9 Oshun Studio surfaces; 3 have no card kind of the same name
      ("Curated music card", "Nyx sky-event 3D briefing", "Output gallery")
      and 2 card kinds have no ratified row (`video`, `video-from-image`).
      The first two are the borrowed-card finding seen from the other side.
      
      NO SURFACE IS FIXTURE-BACKED IN THIS TREE. The seed/production split
      that exists in the admin output-gallery store has no counterpart on the
      curated routes: they are either wired to a real executor or 503. -->
      
    • I0.4.d Match every proposed curated capability/route against existing card/gallery behavior and exact reusable component/contract. Evidence: proposal-to-existing map. <!-- OPEN: .d/.e/.f all take "every PROPOSED curated capability/ route" as their input, and no proposal register exists. The proposals live dispersed through the I1-I6 prose of this file; extracting them is a reading of intent rather than a measurement, and a proposal-to-existing map built from a guess at the proposal set would be a map of the guess. .g says "after decisions", and the decisions are .d-.f.

      WHAT IS ALREADY DONE FOR THEM: the existing side of the comparison is
      complete (I0.4.a-.c), and the uniqueness and boundary tests `.g` asks
      for run today against it -- no two curated routes collide, no two share
      a source file, card ids and kinds are unique, and every curated route
      sits under one of the two prefixes `studio-boundary.ts` itself names as
      approved. -->
      

    <!-- DEDUPED (2026-08-16): I0.4.d appeared twice, the second copy verbatim and noteless. A duplicated checkbox is double-counted and only one copy can ever be marked, so the bare copy was removed and the annotated one kept. -->

    • I0.4.e Select extend/refactor/merge/redirect/retire instead of a new route whenever an approved surface owns the same audience/workflow. Evidence: zero unexplained duplicate proposals.
    • I0.4.f Record missing approved card/gallery capability as an extension task with canonical route and no raw operator/AAA control leakage. Evidence: gap rows.
    • I0.4.g Validate route/navigation/deep-link inventory after decisions. Evidence: uniqueness and boundary tests.
  • I0.5 Map Yemaya Studio generation, workflow, model, training, 3D, voice, and music controls. Evidence: deep-link and ownership matrix.

    • I0.5.a Inventory high-level and raw generation brief/run/variant/output controls and their Isis contract/provider/job paths. Evidence: generation rows. <!-- isis-generation-controls.json: derived from typed schema (GenerationType/ JobStatus/JobPriority/OutputFileType/OutputStatus enums + GenerationJob/ GeneratedOutput Prisma models + libs/isis/client resources). 15 GenerationType kinds; the GenerationJob run entity (32 fields bucketed: brief=prompt/ negativePrompt/inputUrl, run-lifecycle=status/progress/retry, job-execution= queueName/workerId/gpuType/gpuTimeSeconds/tokensUsed/costEstimate) with JobStatus(6)/JobPriority(4); the GeneratedOutput entity (29 fields) with OutputFileType(11)/OutputStatus(5); contract paths = 4 client resources (generation/models/outputs/workflows); 30 ai-provider paths. brief/run/output/ contract/provider/job all map to typed evidence; VARIANT has NO dedicated entity — variants are GeneratedOutput rows sharing a jobId (recorded honestly, not fabricated). generate-isis-generation-controls.mjs + 6 tests, wired into inventory/verify + CI. This COMPLETES the I0.5.a-d control-inventory family; I0.5.e (3D) and I0.5.f (audio) are now CLOSED and the reason given here for skipping them was wrong: "scattered across many libs (no single typed registry)" describes the shape of the work, not an obstacle to it. A family is covered by DECLARING its owning packages and folding their control-shaped exports mechanically; no single registry is needed. This COMPLETES I0.5.a-f. I0.5.g/h are Assign/Define (design). -->

    • I0.5.b Inventory workflow graph/template/node/parameter/version/import/ validation/execution controls. Evidence: workflow rows. <!-- SPEC 2026-09-11: this checkbox line had been lost in an earlier edit — only its continuation line survived and the evidence comment below was left hanging under I0.5.a. Restored as [ ]; generate-isis-workflow-controls.mjs and the comment below still exist, so re-verify per §2.1 and flip in the current conversation. --> <!-- isis-workflow-controls.json: derived from typed schema + the comfyui surface. 5 graph engines (WorkflowEngine: COMFYUI/BLENDER/UNREAL/GODOT/ CUSTOM), 12 WorkflowCategory, 4 WorkflowStatus (DRAFT/PUBLISHED/DEPRECATED/ ARCHIVED), 4 WorkflowVisibility (PRIVATE/TEAM/ORGANIZATION/PUBLIC); 4 Prisma entities (Workflow 22 fields incl. runCount/successfulRuns/avgExecutionTime, WorkflowTemplate, WorkflowVersion, WorkflowStar); 9 3d-comfyui-nodes node definitions + 25 comfyui-sdk node types (InputSlot/OutputSlot) + a node-validator (5 exports: checkTypeCompatibility/validateValue/ validateNodeSpec) + 3 execution modes (always/on_change/manual). All 8 of I0.5.b's named controls map to typed evidence, NONE absent: graph→ WorkflowEngine, template→WorkflowTemplate, node→3d-comfyui-nodes, parameter→ InputSlot/OutputSlot, version→WorkflowVersion, import→changelog, validation→ node-validator, execution→run-stats+ExecutionMode. generate-isis-workflow-controls.mjs + 5 tests, wired into inventory/verify + CI. Generation (I0.5.a) and training (I0.5.d) controls are separate. -->

    • I0.5.c Inventory model/checkpoint/adapter/LoRA/version/intake/evaluation/ selection controls. Evidence: model rows. <!-- isis-model-registry.json: derived from typed schema (the ModelType/ ModelFormat/HashAlgorithm Prisma enums + the ModelCard/ModelRegistry/ ModelVersion Prisma models). 7 model kinds (ModelType: CHECKPOINT, LORA, CONTROLNET, VAE, EMBEDDING, UPSCALER, CUSTOM) in 8 ModelFormats (SAFETENSORS/PYTORCH/ONNX/GGUF/BLEND/GLTF/FBX/OTHER), managed by 3 typed entities (52 scalar control fields parsed from *MinAggregateOutputType, each categorized by control dimension identity/classification/version/ intake/evaluation/selection). All 8 of I0.5.c's named controls map to typed evidence, NONE absent: model/checkpoint/adapter/LoRA → ModelType (adapter = CONTROLNET/LORA); version → ModelVersion + version/releaseNotes/baseModel; intake → ModelRegistry storagePath/uploaderId/sizeBytes; evaluation → status/isVerified/isActive + 4-algo hash provenance; selection → isPublic/ usageCount/downloadCount/starCount ranking. generate-isis-model-registry.mjs + 6 tests, wired into inventory/verify + CI. Generation controls (I0.5.a), workflow controls (I0.5.b), and training controls (I0.5.d) are separate. -->

    • I0.5.d Inventory dataset/training/fine-tune/checkpoint/evaluation/ promotion/resource/cost controls. Evidence: training rows. <!-- isis-training-controls.json: derived from the lora-training-surface typed contracts (libs/isis/lora-training-surface/src/training-run.ts). A 9-state TRAINING_RUN_STATES lifecycle (pending-rights-attestation → pre-flight → training → paused → evaluating → pending-promotion → promoted → rejected → aborted), 5 DatasetRightsAttestation assertions (owns-or-licensed/no-pii/ no-csam/no-likeness-without-consent/no-third-party-trademarks), 4 evaluation metrics (fid/clip-score/psnr/ssim), 38 typed control fields (rank/epochs/ learningRate/costEstimateCents/checkpointStep/datasetSizeBytes/ promotedManagedModelId/…) across 4 training libs (lora-training-surface, model-fine-tuning, training-data, gaussian-splatting) + a multi-GPU orchestration resource provider. All 8 of I0.5.d's named controls map to typed evidence, NONE absent: dataset→rights-attestation, training→lifecycle, fine-tune→LoRA-rank, checkpoint→CheckpointSample, evaluation→metrics, promotion→promoted-state, resource→multi-gpu, cost→costEstimateCents. generate-isis-training-controls.mjs + 7 tests, wired into inventory/verify + CI. (Also removed a duplicated I0.5.d anchor line.) I0.5.a (generation) and I0.5.e+ (scene/mesh/rig/DCC) remain. -->

    • I0.5.e Inventory scene/mesh/rig/material/texture/splat/animation/DCC/ engine generation and post-process controls. Evidence: 3D rows. <!-- evidence: isis-3d-controls.json — all NINE named families resolved to typed surfaces: 7,386 controls across 273 files, every family typed, nothing owed. generate-isis-3d-controls.mjs + .test.mjs (13 tests), wired into inventory + verify. Shares its extractor with I0.5.f, so both inventories read source the same way.

      Each family declares its OWNING PACKAGES and every control-shaped export
      in them is folded in mechanically. Two families share one library and
      are split by a declared file pattern: `libs/isis/universal-rigging`
      holds both the skeleton controls (rig) and the clip/blend controls
      (animation), and folding it whole into either row would misreport both.
      
      A SURFACE CAN BELONG TO TWO FAMILIES, AND THE ARTIFACT SAYS WHICH. 20
      surfaces are claimed twice and each is listed in `overlaps`. They are
      not mistakes: a Godot resource export is an interchange format AND an
      engine target, and `humanoid-animation-retargeting` is retargeting (a
      rig concern) applied to animation. A first version of the test asserted
      the rig/animation split was disjoint; the split is right and the
      assertion was wrong, so the test now requires that every shared file be
      RECORDED as an overlap rather than absent.
      
      AN ALPHABET CAN BE DECLARED AND EMPTY, AND THE FIELD IT FEEDS DECIDES
      WHAT THAT MEANS. 18 constants are `[] as const`, so every caller sees a
      named vocabulary with no word in it. Reading them without their consumer
      would report one fact where there are three. 12 are named `*_PLANNED_*`.
      Five feed `deferredTaskIds`, where empty says NOTHING IS DEFERRED — the
      opposite of a gap. The last one is the interesting one:
      `LOCAL_3D_UNIVERSAL_RIGGING_SUPPORTED_INPUT_FORMATS` is `[]`, feeds
      `supportedInputFormats`, and sits beside
      `..._PLANNED_INPUT_FORMATS = ['obj','glb','fbx','vrm']` with a test
      pinning the empty one. THE RIGGING LIBRARY SAYS PLAINLY THAT IT READS NO
      INPUT FORMAT YET AND NAMES THE FOUR IT INTENDS TO. That is an honest
      declaration, not a stub, and it is recorded as one.
      
      EXTRACTOR CORRECTIONS MADE WHILE BUILDING THIS, BOTH NOW UNDER TEST. An
      alias chain routinely passes through a MODULE-PRIVATE type
      (`export type X = UnrealImportMode` where
      `type UnrealImportMode = (typeof MODES)[number]`); refusing to follow an
      unexported link reported 34 real alphabets as having no members. And an
      interface field whose type WRAPS onto the next line
      (`readonly humanBones?:` then a wrapped `Partial<Record<…>>`) was read as
      no field at all, which made an interface with members report zero — a
      wrong answer rather than a missing one. Both fixes flow back into I0.5.f.
      
      The boundary is published: 422 files export a symbol in the 3D
      vocabulary, 273 are folded into rows, and the rest are listed by name
      with their symbols. Eight symbols are skipped and listed; each is an
      alphabet that really is empty, not an extraction failure.
      
      This CLOSES the I0.5.a-f control-inventory family. .g (Assign each
      control to an owner) and .h (Define signed deep-link/handoff context)
      are design cells and stay open. -->
      <!-- Retested: 13 tests pass, `--check` reports no drift, and the
      adversarial stub scan over both new files returns zero hits. -->
      
    • I0.5.f Inventory voice/TTS/cloning/dubbing/conversation/foley/audio and music/stem/composition/mastering generation controls. Evidence: audio/ music rows. <!-- evidence: isis-audio-controls.json — all ELEVEN named families resolved to typed surfaces: 1,947 controls across 79 files, 10 typed and 1 partial. generate-isis-audio-controls.mjs + .test.mjs (14 tests, 2 of them ratchets on findings), wired into inventory + verify.

      Method follows I0.5.a-.d: a family is covered only by a TYPE, and the row
      carries that type's actual members. Four shapes are read from source —
      interfaces with their optional flags, string/numeric unions, `as const`
      arrays, and zod `object`/`enum`. The zod reader matters: the real
      generation parameters for music, voice and SFX are module-PRIVATE consts
      in the generation-API pack schemas, and being unexported does not make a
      control surface less of one.
      
      A ROW'S COVERAGE MUST NOT DEPEND ON WHICH SYMBOLS A HUMAN LISTED. Each
      family pins the shared types it needs by name, and then its OWNING
      PACKAGES are folded in mechanically by the estate's own naming
      convention for a control (`*Request`/`*Options`/`*Config`/`*Settings`/
      `*Params`/`*Schema`/`*Kind`/`*Mode`/`*Format`/`*Strategy`/`*Profile`).
      Hand-listing had given `cloning` 5 controls; the package fold gives it
      317, because `libs/isis/voice-cloning` carries a `*Request` per
      capability — age-modified synthesis, archival restoration, multilingual
      preservation, speech-to-speech, consent rights, watermarking — and none
      of them was in anybody's list.
      
      MASTERING IS THE ONE PARTIAL, AND THE SHAPE OF THE GAP IS THE FINDING.
      Loudness is typed twice over: as a TARGET the caller sets
      (`target_lufs` on the pack schema, range -32..-6), as a MEASUREMENT
      taken afterwards (`SectionLoudnessSchema.integratedLufs` /
      `.truePeakDb`, `SongSeamMeasurement.loudnessDeltaLu`), and as a ceiling
      on the bus (`MixBus.limiterCeilingDbFs`). What has NO type anywhere is
      the processor chain that would get from one to the other — zero hits for
      a mastering chain, master bus, multiband compressor or mastering preset
      across 3,646 tracked Isis sources. THE ESTATE CAN STATE A LOUDNESS
      TARGET AND CHECK WHETHER IT WAS MET, AND TYPES NOTHING THAT WOULD MEET
      IT.
      
      AN ABSENCE IS ONLY RECORDED WHEN IT WAS LOOKED FOR, AND THE PROBE THAT
      LOOKS IS PART OF THE BUILD. Two absences claimed on a first pass were
      WRONG and the probes killed them: `composition` was called partial for
      having no section model, and `libs/isis/workflows/src/ledger/
      song-ledger.ts` types a song as eight ordered `SongSectionKind` parts
      with a `MusicalKey` and `SongLocks`; `mastering` was called absent for
      having no loudness control, and LUFS appears in 35 files. Both errors
      came from reading a `grep | head` as if it were the whole answer. The
      generator now THROWS if a negative probe starts matching, so the same
      mistake cannot be committed.
      
      WHAT THIS INVENTORY DOES NOT FOLD IN IS PUBLISHED, NOT OMITTED. The same
      audio vocabulary is swept across every tracked Isis source: 91 files
      export a matching symbol, 43 are folded into rows, and the other 48 are
      listed by name with their symbols. Most are providers, quality gates, UI
      helpers and 3D files that matched on `Mix`. One deliberate exclusion is
      recorded in the source: folding `libs/isis/workflows/src/ledger` as a
      package added 638 controls about lecture overlays and narration arcs to
      a row about musical composition, so the song ledger is pinned by file
      and the rest stays in the published gap. A ROW THAT GAINS VOLUME AND
      LOSES MEANING IS WORSE THAN A GAP SOMEBODY CAN SEE.
      
      NOT RECORDED AS A DEFECT, DELIBERATELY: the control-conditioned variant
      of all three audio packs carries the image ControlNet vocabulary
      (`pose_map`, `depth_map`, `lineart_map`, `use_canny_control`,
      `*_strength`). It looks wrong on a voice line until you find
      `control_applicability_mode: 'auto' | 'where-applicable' | 'strict'`
      beside it, which is a declared seam for exactly that. Whether a uniform
      cross-modal control contract is right is a product decision, so the
      measurement is recorded and the verdict is not. -->
      <!-- Retested: 14 tests pass, `--check` reports no drift, and the
      adversarial stub scan over both new files returns zero hits. -->
      
    • I0.5.g Assign each control to Yemaya, Isis operator, domain plugin, or excluded surface with owner, audience, capability, contract, route, and permitted detail. Evidence: ownership matrix.

    • I0.5.h Define signed deep-link/handoff context for project/asset/shot/ timeline/run/output/return path and test tenant/expiry/permission. Evidence: deep-link matrix.

  • I0.6 Map all libs/isis/* packages to operator, creator, worker, shared service, or out-of-scope. Evidence: all discovered packages classified. <!-- partial: package capability inventory (a) done in evidence/v1-workbenches/isis-package-inventory.json. b-d (assign operator/ creator/worker ownership per package) are design decisions and stay unchecked; roleSignal is provided as raw input.

                                                                                                                                                                                                                                                                    .e IS NOT PURELY A DESIGN DECISION AND THE OLD NOTE SAID IT WAS. Its
                                                                                                                                                                                                                                                                    sentence states two NECESSARY conditions, and applying them is a
                                                                                                                                                                                                                                                                    measurement: isis-shared-package-gate.json excludes 55 of 78 packages on
                                                                                                                                                                                                                                                                    those grounds alone. It stays unchecked because a necessary condition
                                                                                                                                                                                                                                                                    cannot confirm a class -- see the note on .e itself. -->
    
    • I0.6.a Enumerate every Isis library/export/consumer/dependency/test/ runtime/contract/entity/job/provider and atomic capability. Evidence: package capability inventory. <!-- evidence: .packages — all 76 libs/isis/* packages consolidated from the BASE.3 library graph (exports/inbound/outbound/disposition) + BASE.9 capabilities (496 entity+event attributed via source.file) + BASE.1 tests (1138 files). App-surface capabilities (api-operation/command from apps/isis, apps/oshun/bff) stay in I0.1/I0.3 by design. 29 no-consumer retire-candidates surfaced. generate-isis-package-inventory.mjs + 7 tests, wired into inventory/verify. -->

    • I0.6.b Classify operator-owned packages that expose fleet, provider, workflow/model registry, policy, emergency, economics, parity, or administration. Evidence: operator rows.

    • I0.6.c Classify creator-facing packages for brief/workflow/variant/ preview/compare/review/gallery or authorized project handoff. Evidence: creator rows.

    • I0.6.d Classify worker/runtime packages for generation, training, processing, validation, lineage, storage, queue, and provider adapters with operational projection. Evidence: worker rows.

    • I0.6.e Classify truly domain-neutral contract/service helpers as shared only with two real consumers and correct dependency direction. Evidence: shared rows. <!-- STILL OPEN, DELIBERATELY, AND THE GATE IS BUILT. isis-shared-package-gate.json — all 78 libs/isis/* packages put through the two conditions this cell states, with the evidence for each beside it. generate-isis-shared-package-gate.mjs + .test.mjs (12 tests), wired into inventory + verify.

      A NECESSARY CONDITION CAN EXCLUDE AND CANNOT CONFIRM. "only with two real
      consumers and correct dependency direction" is a gate on the way in, not
      a definition of the class. `truly domain-neutral` is the sufficiency test
      and this cell reserves it, so the artifact records `classifiedShared: 0`
      and never guesses. 55 packages are EXCLUDED on the cell's own grounds —
      that half is decided and is not owed to anybody. 23 are admissible and
      await one judgement each.
      
      WHY THE CONSUMER EDGES ARE RECOMPUTED RATHER THAN REUSED. The BASE.3
      library graph counts test importers deliberately — its own header says
      "application/service/worker/CLI/test/library consumers" — so a package
      imported only by somebody's spec file reads as consumed there. This cell
      says REAL consumers, so consumption is recounted at file level and a
      production import is distinguished from a test one. Result: 52 of 78
      packages do not have two production consumers, and 29 HAVE NO CONSUMER
      AT ALL — the same 29 I0.6.a's inventory reports as no-consumer
      retire-candidates, reached by a different route. Two instruments
      agreeing is worth more than either alone. The 29 are the input I0.6.f
      needs for its dead class; the disposition itself is .f's to make.
      
      DEPENDENCY DIRECTION IS ABOUT THE SCOPE, NOT THE COUNT. A shared helper
      sits BELOW every domain, so the clause is "imports no domain-scoped
      package at all". `@oshun/*` is the shared scope and does not count
      against it; `@isis/*` does, because a package that imports its own domain
      cannot move to `libs/shared` without dragging the domain with it. 11
      packages fail on that alone — `@isis/3d-semantic-editing` importing
      `@isis/universal-rigging` is the shape.
      
      TWO WAYS THE INSTRUMENT COUNTED THINGS THAT WERE NOT THERE, BOTH FIXED
      AND BOTH NOW UNDER A CONTROL TEST. A path that is not a project is not a
      domain: the first version fell back to the first two path segments for
      any importer, which made `eslint.config.js` and `scripts/v1-workbenches`
      — THIS GENERATOR'S OWN DIRECTORY — read as domains, and eight packages
      looked domain-neutral on the strength of a lint config importing them.
      And A QUOTED PACKAGE NAME IS NOT AN IMPORT: matching every
      `'@scope/name'` string made THIS GATE'S OWN TEST FILE, which compares
      against package names in its assertions, a consumer of the packages it
      tests — the artifact drifted the moment the test was written. Specifiers
      must now sit in an import position. That correction also removed `v2`
      from the consumer domains: `apps/v2/isis-*` projects NAME Isis packages
      in strings and import none of them.
      
      DOMAIN SPREAD IS REPORTED BESIDE THE VERDICT, NOT FOLDED INTO IT. Of the
      23 admissible, the ones consumed by another PRODUCT domain (yemaya,
      metis, nous, psyche, neith) are the strongest candidates; the rest are
      consumed by `isis` and by the host app `apps/oshun/*`, which is not the
      same fact as being used by two products. That distinction is
      evidence for the judgement, and making it here would BE the judgement. -->
      <!-- Retested: 13 tests pass, `--check` reports no drift, and the
      adversarial stub scan over both new files returns zero hits. -->
      
    • I0.6.f Classify compatibility/dead/experimental/vendor/test-only/ superseded/non-V1 packages with retain/retire/exclude rationale, consumers, owner, and date. Evidence: out-of-scope rows.

    • I0.6.g Resolve mixed packages at symbol/capability level rather than one broad package label. Evidence: split mapping.

    • I0.6.h Reconcile every filesystem package to exactly one or more capability rows with no unclassified library. Evidence: path check and domain-owner signoff.

  • I0.7 Decide whether any new Isis workbench domain library is needed. Evidence: ADR prefers existing @oshun/generation-control-isis/domain packages unless a concrete gap is demonstrated.

    • I0.7.a Define the exact missing capability/ownership/dependency/API/ runtime/test problem that cannot live in an existing Isis package. Evidence: source-linked gap statement.

    • I0.7.b Evaluate extending @oshun/generation-control-isis, another existing domain package, the shared workbench kit, an application-local adapter, and no-new-code options. Evidence: option matrix.

    • I0.7.c Compare public API, dependency direction, consumer count, domain-specific semantics, runtime boundary, version/release cadence, migration, and deletion cost. Evidence: trade-off table.

    • I0.7.d Prototype the smallest interface in the preferred existing boundary and record why it succeeds or fails. Evidence: compile/dependency proof.

    • I0.7.e If a new library is justified, define owner, scope/non-goals, exports, dependencies, two consumers, contracts, migration, and deletion threshold. Evidence: proposed package design.

    • I0.7.f Record the chosen decision, rejected options, consequences, compatibility, review date, and named architecture/Isis approval. Evidence: ratified ADR. <!-- MEASURED 2026-08-17 (I0.7.a–f). THE ANSWER IS NO, AND THE REASON IS MORE USEFUL THAN THE ANSWER: THE OBSTACLE IS NOT A MISSING BOUNDARY, IT IS AN UNREACHED ONE. evidence/v1-workbenches/isis-library-gap-decision.json (writer + 14 tests, wired into verify:inventory:v1-workbenches before the design doc) measures the five gaps this lane has already recorded and 0 of 5 justify a new library.

      THE GAPS ARE NOT INVENTED IN THE INSTRUMENT. Each row is read out of
      another committed artifact by dotted path, so the list is the estate's own
      record: asset-kind taxonomy (I0.8), data-class register (I0.11),
      emergency-control commands (I0.12), threat-to-test map (I0.10),
      output-format columns (I0.19). A gap whose source stops reporting it
      leaves the matrix; a new one appears in it.
      
      TWO OF THE FIVE HAVE AN EXISTING HOME THAT NOTHING ON THE ISIS PATH
      REACHES. `libs/contracts/src/common/asset.ts` already declares
      `AssetTypeSchema` (image|video|audio|model_3d|texture|material|animation|
      …) and `libs/contracts/src/common/data-rights-scopes.ts` already declares
      `OshunArtifactType`, and **zero on-path modules reach either**. So those
      two gaps are `existing-home-unreached`: no library is missing, a boundary
      is unvisited. A third (`output-format-columns`) is
      `existing-home-reached-needs-extending` — `OutputArtifactManifest` exists
      and is reached, and three columns are simply absent from it. A fourth
      (`emergency-control-commands`) must name Isis domain types, so its home is
      a `libs/isis/*` package, of which 78 already exist. The fifth is not a
      library question at all: A MAPPING BETWEEN A THREAT LIST AND A TEST ID IS
      DATA, AND NO PACKAGE CAN OWN IT.
      
      A LIBRARY IS A DEPENDENCY DIRECTION BEFORE IT IS A FOLDER, so the
      direction was measured before any option was scored: **45 `libs/isis/*`
      packages depend on `@oshun/*` and exactly ONE `libs/oshun/*` package
      depends on `@isis/*`** — `libs/oshun/privacy`, whose own header calls it
      "the account-deletion fan-out's intentional, cross-cutting reach into a
      domain data store… the one place the privacy deletion engine must touch
      domain persistence". A rule with one exception that documents itself is
      still a rule; the instrument throws if that exception moves. That
      direction is what decides the host: a capability expressible over shared
      contracts fits the preferred `@oshun/*` boundary, one that must name an
      `@isis/*` type does not.
      
      (.c) IS EIGHT MEASURED COLUMNS, NOT FOUR AND A PARAGRAPH. Every column
      the cell names maps to an OptionRow member and a test asserts the mapping
      is total, so a column cannot be answered in prose. The table is worth
      reading for the shape of the choice: the preferred library is **452
      public symbols / 24 consumers / `layer:domain` / esm, side-effect-free,
      source entry**, and the shared workbench kit is **14,298 symbols / 128
      consumers / `domain:neutral`** — so adding an Isis vocabulary there would
      change what the package IS, which its own Nx tag declares. The
      application-local adapter is the cheapest to delete (nothing outside the
      BFF can import it) and the kit the dearest.
      
      (.d) IS EXECUTED, NOT ASSERTED. The smallest interface that would give
      I0.11 its missing row — a binding from an Isis record kind to the
      canonical artifact type that already carries an export scope and a
      retention rule — was WRITTEN into
      `libs/oshun/generation-control-isis/src/`, compiled by `tsc` with the
      repo's real path mappings (**exit 0, zero diagnostics**), and DELETED in a
      `finally` whether it passed or failed: a speculative interface must never
      survive the measurement that used it. A test asserts the file is gone, and
      `.gitignore` covers it so a hard kill mid-measurement cannot leave
      something committable.
      
      THE DEPENDENCY HALF IS A SEPARATE CHECK AND THAT SEPARATION IS THE POINT.
      `tsconfig.base.json` maps **`@isis/*` → `./libs/isis/*/src/index.ts`** — a
      WILDCARD — so every Isis package resolves for every file in the repo, and
      none of them is a dependency of the preferred package. A COMPILER THAT
      RESOLVES AN IMPORT HAS NOT TOLD YOU THE IMPORT IS ALLOWED: the mapping is
      repo-wide, the dependency is per package, and only the second is a
      boundary. So the compile proof and the dependency proof are asked and
      reported separately.
      
      TRAPS. (1) `git ls-files 'dir/**/*.ts'` MATCHES NOTHING — the `**`
      pathspec is not expanded like a shell glob, and it silently reported ZERO
      exports for a 42-file package; `git ls-files 'dir'` is already recursive,
      and a control test pins the collision. (2) The first run claimed
      `apps/isis/output-registry/src/types.ts` declares `ArtifactManifest`; it
      declares `OutputArtifactManifest`, and the word-anchored check correctly
      refused to match inside it — the verification caught the author's own
      error rather than emptying a row.
      
      CELLS MARKED: .b (the option matrix covers all five options the cell
      names, each with measured columns), .c (all eight columns measured for
      every option, mapping asserted total), .d (prototype written, compiled at
      exit 0, dependency-checked and deleted). CELLS LEFT UNCHECKED: .a — "define
      the exact missing problem that cannot live in an existing package"
      presupposes such a problem, and the measured answer is that there is none
      today, which is an ongoing obligation rather than a closed one; .e is
      conditional on a new library being justified and the condition is unmet,
      so offering a design anyway would smuggle a decision in as a deliverable;
      .f needs a ratified ADR with named architecture/Isis approval, which is
      not locally actionable. The PARENT stays open on .a/.e/.f even though its
      own bar — "ADR prefers existing packages unless a concrete gap is
      demonstrated" — is now satisfied on the evidence side: no gap is
      demonstrated, and the ADR that would say so has no author. -->
      
  • I0.8 Define the asset-kind taxonomy for image, portrait, texture, video, audio, foley, voice, mesh, rig, gaussian splat, and composite scenes. Evidence: registry maps each kind to contract, preview, gates, and publisher. <!-- THERE IS NO REGISTRY. THE PARENT AND ALL THIRTEEN CHILDREN STAY [ ].

    evidence/v1-workbenches/isis-asset-kind-registry.json —
    generate-isis-asset-kind-registry.mts + .test.mts (10 tests, 4 ratchets,
    3 controls), wired into inventory + verify. The eleven kinds and the four
    columns are parsed out of THIS parent's own sentence, so its bar is the
    measurement's bar.
    
    THE CONTRACTS ARE REAL AND THEY ARE A REFERENCE. Executed against
    `evidence/v1-workbenches/asset-contracts`: all ELEVEN kinds' minimal
    fixtures parse and round-trip byte-stably through the public API, an
    unexpected field is refused with `image: unexpected field "sneakyExtra"
    (no unknown-payload bag)` rather than absorbed, and each kind declares its
    own required fields (5 for image up to 11 for scene). That is a genuine
    contract per kind — and NOTHING UNDER apps/ OR libs/ IMPORTS IT, so all
    eleven kinds have no product contract.
    
    THE ESTATE HAS SEVENTEEN ASSET-KIND ENUMERATIONS AND NONE IS THIS
    TAXONOMY. The best overlap with the eleven kinds is six, and it belongs to
    another evidence reference. On the Isis surface itself the only one is
    `unity-package-export`'s four-member `['material','model','prefab',
    'texture']`. THE LARGEST BY MEMBER COUNT IS A TRAP: `studio-authoring`
    declares eleven kinds — ambient-loop, manuscript-scan, ritual-sound-pack,
    voiceover — a DIFFERENT taxonomy that happens to be the same size, and a
    first cut reported it as the product's asset-kind enum on the strength of
    the number alone.
    
    TWO AXES THAT DO NOT MEET. The estate gates by ASSET FAMILY — 32 of them,
    each with a quality gate and a regression golden dataset — and this
    taxonomy is by ASSET KIND. NOT ONE of the 32 families declares a kind;
    they declare OUTPUT CLASSES instead (`audio_master`, `metadata_manifest`,
    `preview_image`, `project_archive`, `source_frames`, `video_master`), so
    `material-pbr-pack` — a texture family — announces itself as producing a
    `preview_image`. The gates column cannot be filled from the families
    because nothing maps a family to a kind.
    
    A MENTION IS NOT A MAPPING, and that distinction is the difference between
    this cell being nearly done and not started. Fifty-two Isis modules whose
    file name says gate or quality contain the word "image"; counting those
    reported SEVEN OF ELEVEN kinds as having all four columns. The artifact
    records the mention counts as context and answers the parent's question
    with `columnsMappedByARegistry`, which is 0 for every kind because there
    is no registry to do the mapping.
    
    AND I0.8.a'S NINE ATTRIBUTES HAVE ONE CARRIER BETWEEN THEM. Identity,
    version, parent/family, aliases, media/profile, lifecycle, owning package,
    audience/host, extension policy: a bare string-literal union carries the
    first and none of the rest. -->
    <!-- I0.8.m's validation questions — unique ids and aliases, no ambiguous
    kind resolution — are answered for the REFERENCE by the round-trip and
    required-field cases above; for the product there is no registry to
    validate. -->
    
    • I0.8.a Define taxonomy entry identity/version, parent/family, aliases, media/profile, lifecycle, owning package, audience/host, and extension policy. Evidence: asset-kind registry schema.
    • I0.8.b Define generic image kind controls/outputs, metadata, visual preview/semantic alternative, quality/safety/rights/provenance gates, and publishers. Evidence: image registry row and fixtures.
    • I0.8.c Define portrait kind identity/reference/consent controls, face preview/alternatives, impersonation/biometric/safety/quality gates, and publishers. Evidence: portrait row and fixtures.
    • I0.8.d Define texture kind tiling/channels/color/PBR target controls, material preview/metadata, seam/technical/rights/provenance gates, and publishers. Evidence: texture row and fixtures.
    • I0.8.e Define video kind duration/frame/rate/size/codec/motion/audio/ caption controls, synchronized preview/alternatives, temporal/safety/ rights/quality gates, and publishers. Evidence: video row and fixtures.
    • I0.8.f Define general audio kind duration/sample/channel/format/content controls, waveform/audition/transcript alternative, loudness/safety/ rights/quality gates, and publishers. Evidence: audio row and fixtures.
    • I0.8.g Define foley kind event/action/material/scene/time/variation/ synchronization controls, audiovisual preview, originality/rights/ technical gates, and publishers. Evidence: foley row and fixtures.
    • I0.8.h Define voice kind speaker/language/style/text/timing/delivery and consent constraints, audio/transcript preview, cloning/impersonation/ safety/rights/quality gates, and publishers. Evidence: voice row.
    • I0.8.i Define mesh kind units/axes/topology/geometry/material/UV/LOD/ collision/format controls, 3D/metadata preview, OpenUSD/geometry/rights/ provenance gates, and publishers. Evidence: mesh row and fixtures.
    • I0.8.j Define rig kind skeleton/joint/naming/orientation/constraints/ skin/controls/target controls, hierarchy/pose preview, deformation/ compatibility/rights gates, and publishers. Evidence: rig row.
    • I0.8.k Define Gaussian-splat kind coordinate/scale/SH/point-count/ cameras/compression/format controls, spatial preview/metadata, quality/privacy/rights/interchange gates, and publishers. Evidence: splat row.
    • I0.8.l Define composite-scene kind hierarchy/reference/variant/camera/ light/material/animation/audio/dependency controls, scene/semantic preview, cross-asset/rights/provenance/target gates, and publishers. Evidence: scene row.
    • I0.8.m Validate unique IDs/aliases, inheritance, audience exposure, contract/preview/gate/publisher completeness, unsupported behavior, and no ambiguous kind resolution. Evidence: registry validation tests.
  • I0.9 Define the workflow-class taxonomy and audience allowlist. Evidence: raw graph, model hash, LoRA, scheduler, and provider details cannot leak to curated payloads.

    • I0.9.a Inventory current workflow templates/graphs/nodes/providers/ models and classify generation, edit/transform, upscale/enhance, animate/video, audio/voice/music, 3D, training, analysis, and composite classes. Evidence: workflow class register. <!-- isis-workflow-class-register.json: built from CANONICAL committed artifacts (the codebase's OWN classification, not imposed judgment). The committed workflow-taxonomy-registry.json enumerates 66 workflow mappings (registry declares 64 — both recorded, discrepancy surfaced) across 11 in-code categories (sacred-art/video/fractal/flagship/audio/avatar/ composition/utility/3d/nft/texture), 37 asset families, 5 output profiles, 38 control modalities. Canonical typed vocabulary from the generated Prisma enums: WorkflowCategory(12: IMAGE_GENERATION..COMPOSITING..), GenerationType (15), WorkflowEngine(5: COMFYUI/BLENDER/UNREAL/GODOT/CUSTOM), ModelType(7)/ ModelFormat(8) = models, OutputFileType(11). Providers: 30 ai-providers + llm provider-registry. workflow-classes lib = 7 modules. The 9 proposed I0.9.a classes are covered by the canonical vocab EXCEPT "training" (honestly flagged as a real gap — no training workflow category/GenerationType exists). generate-isis-workflow-class-register.mjs + 6 tests, wired into inventory/verify + CI. The class SCHEMA (ID/version/capability/cost/lifecycle/ owner, I0.9.b) and audience allow/deny (I0.9.c) DEFINE new structure and stay unchecked. -->

    • I0.9.b Define class ID/version, supported asset kinds/operations, required inputs/outputs, parameter schema, capabilities, cost/resource class, lifecycle, and owner. Evidence: class schema.

    • I0.9.c Assign operator, AAA, curated, and contemplative allow/deny/ request/inspect/edit/execute/replay exposure per class. Evidence: audience allowlist.

    • I0.9.d Define operator payload with authorized raw graph/node/model/ adapter/LoRA/scheduler/provider/resource/debug detail. Evidence: operator contract tests.

    • I0.9.e Define AAA payload with approved advanced workflow/model/training detail but no fleet credentials, private provider internals, or operator emergency authority. Evidence: AAA contract tests.

    • I0.9.f Define curated/contemplative payload as task-specific high-level intent/constraints/reference/quality/budget/delivery choices and stable public labels. Evidence: curated contract tests.

    • I0.9.g Redact raw graph, node IDs/parameters, model/checkpoint hashes, LoRA/adapter identity, scheduler/sampler internals, provider/endpoint, secrets, pricing, and fleet topology from lower tiers. Evidence: property/bundle leak tests. <!-- THE ESTATE HAS A LEAK DETECTOR AND IT IS POINTED AT THE REQUEST. I0.9.d–g STAY [ ].

      evidence/v1-workbenches/isis-curated-payload-leak.json —
      generate-isis-curated-payload-leak.mts + .test.mts (8 tests, 3 ratchets,
      2 controls), wired into inventory + verify. The nine vectors are parsed
      out of THIS cell's own redaction list.
      
      THE DETECTOR IS DOCUMENTED OUTBOUND AND APPLIED INBOUND.
      `auditRawGraphLeak`'s own comment says it "Scans an outbound payload".
      It has ONE call site outside its module, at
      `curated-card-route.ts:150`, and it is handed
      `JSON.stringify(body.inputs)` — the REQUEST. So it stops a curated
      caller SMUGGLING a raw graph IN, which is a real and useful check, and
      it is not the check this cell asks for. NOTHING AUDITS WHAT GOES OUT.
      
      AND WHAT GOES OUT CARRIES A PROVIDER KEY. `GET
      /v1/isis/curated-cards/catalog` returns `{spec, provenance}` per card to
      any admitted caller, and `ProvenancePreview` declares `providerKey`,
      `providerRegion` and `costClass` — PROVIDER, ENDPOINT REGION AND
      PRICING, three of the nine vectors this cell asks to be redacted from
      lower tiers. Executed: the real catalog produced eight cards, each with
      a provenance object of `[costClass, providerKey, providerRegion,
      workflowClassId, workflowClassVersion]`, and the detector run over that
      exact payload returned `leakDetected=false`.
      
      THE VALUES SHIPPED TODAY ARE THE LITERAL "pending" and the catalog's own
      comment says the runtime fills them in when it resolves the tenant's
      provider. THE LEAK IS THEREFORE A PROPERTY OF THE RESPONSE TYPE, not of
      the current fixture — which is exactly what a contract-level cell asks
      about.
      
      THE DETECTOR COULD NOT HAVE OBJECTED ANYWAY. Handed one payload per
      vector it caught five — a raw ComfyUI graph, numeric node ids, a 64-hex
      model hash, a `lora-…` id and a scheduler name — and MISSED the
      provider, the secrets, the pricing and the fleet topology. Its patterns
      were written for the four kinds of thing a client might smuggle in, and
      the two vectors the response actually carries are not among them. That
      it fires at all is what makes its silence on the catalog meaningful
      rather than inconclusive.
      
      A CONTROL THE CLASSIFIER NEEDED: word-anchoring the direction vocabulary
      classified a call site handed `responsePayload` as `unknown`, because
      `\bresponse\b` does not match a camelCase compound. A response argument
      is now recognised, and a test pins it. -->
      <!-- I0.9.d–f (the operator / AAA / curated payload definitions) are
      answered from the same measurement: the curated payload is `CardSpec` +
      `ProvenancePreview`, and its fields are listed in the artifact. -->
      
    • I0.9.h Resolve workflow/tier at authoritative BFF/service boundary and reject direct IDs or parameter injection outside allowlist. Evidence: bypass negative tests. <!-- PARTIAL, READ-ONLY 2026-08-16 (I0.9.h). NO INSTRUMENT — this note records only what was read, and the cell stays unchecked because the executed negatives its Evidence line asks for were NOT built. THE ALLOWLIST THE STYLE PRESET IS CHECKED AGAINST ARRIVES IN THE SAME REQUEST AS THE STYLE PRESET. libs/isis/curated-cards/src/card-validators.ts:118 asks if (!input.supportedStylePresets.includes(i.stylePreset)), and apps/oshun/bff/src/isis/curated-card-route.ts:159 fills that array from body.supportedStylePresets ?? []. Lines 121/146 and 160 are the same shape for tenantSupportedLocales. So the second bypass class the cell names — parameter injection OUTSIDE AN ALLOWLIST — is checked against a caller-supplied allowlist, which constrains nobody. The first class is in better shape: the route resolves the workflow class from CURATED_CARD_CATALOG by cardId and never reads a body.workflowClassId, and the creator tier and tenant come from the authenticated principal (route header comment, lines 11–15). NEXT SESSION: the executed negative is cheap — validateInputs is pure — but it needs a real CuratedCardInputs for a catalogued kind; a synthetic {kind, brief, stylePreset, locale} made validateInputs return undefined for every catalogued kind. THAT IS NOT A SECOND FINDING — checked: validateInputs switches over exactly the eight kinds the catalog ships (illustration, narration, ambient-audio, explainer, caption-dub, accessibility-pass, video, video-from-image), so it is exhaustive and the undefined came from the harness, not the product. The real cause is the entry shape: listCuratedCards() returns catalogue entries DIRECTLY (no .card wrapper) and card ids are not always curated-<kind>. Build the inputs from the entry's own binding and read entry.kind off the entry. -->

    • I0.9.i Define unsupported/not-configured/degraded/version-skew and class migration behavior per audience. Evidence: capability state matrix.

  • I0.10 Threat-model brief injection, workflow injection, malicious graphs, model supply chain, unsafe media, face/voice impersonation, child safety, biometric data, training rights, and cost abuse. Evidence: mitigations map to test IDs.

    • I0.10.a Threat-model brief/prompt/reference/negative-prompt/metadata injection across audience tiers and downstream provider/tools. Evidence: brief attack tree and tests.
    • I0.10.b Threat-model workflow/template/node/parameter/import/version injection and unauthorized class/tier escalation. Evidence: workflow attack tree and tests.
    • I0.10.c Threat-model malicious graph cycles, resource bombs, unsafe nodes/scripts/custom code, filesystem/network access, deserialization, dependency escape, and hidden outputs. Evidence: graph tests.
    • I0.10.d Threat-model model/checkpoint/adapter/LoRA/plugin/container/ dependency origin, license, signature/hash, malicious serialization/ code, poisoning, backdoor, update, and compromise. Evidence: supply-chain tests.
    • I0.10.e Threat-model uploaded/fetched/generated image/video/audio/3D/ archive malformed, active, oversized, steganographic, illegal, or parser exploit content. Evidence: media threat tests.
    • I0.10.f Threat-model face and voice cloning/reenactment/dubbing/ impersonation, identity theft, fraud, consent scope, public figure, disclosure, and revocation. Evidence: impersonation tests.
    • I0.10.g Threat-model child sexual abuse/exploitation, age ambiguity, grooming/sexualization, identity/biometric processing, reporting, quarantine, and mandatory escalation. Evidence: child-safety controls.
    • I0.10.h Threat-model face/voice/body/motion embeddings/templates, inference, cross-context linking, retention, access, provider transfer, breach, and deletion. Evidence: biometric tests.
    • I0.10.i Threat-model dataset/reference/output/model training rights, license/consent, attribution, opt-out/revocation, laundering, derivative use, and territorial limits. Evidence: rights tests.
    • I0.10.j Threat-model token/GPU/storage/provider/egress/retry/batch/ training/render cost exhaustion, quota bypass, cryptomining/denial, invoice fraud, and kill-switch failure. Evidence: cost abuse tests.
    • I0.10.k Map asset/threat/actor/entry/control/detection/response/recovery/ residual risk/owner to exact automated/manual test and stop-the-line policy. Evidence: zero unmapped high-risk scenarios. <!-- MEASURED 2026-08-16 (I0.10.a–k). THERE IS NO THREAT-TO-TEST MAP, SO "ZERO UNMAPPED HIGH-RISK SCENARIOS" IS UNEVALUABLE — EVERY SCENARIO IS UNMAPPED BY CONSTRUCTION. evidence/v1-workbenches/isis-threat-mitigation-map.json (writer + 8 tests, wired into verify:inventory:v1-workbenches) parses the ten threats out of I0.10's own text and asks three SEPARATELY-FAILING questions of each: is there a control ON THE ISIS PATH, is there a TEST naming it, and is the threat MAPPED TO A TEST ID. threatMapArtifacts: 0 — no artifact anywhere in the repo maps a threat to a test id, so the third answer is false for all ten regardless of the first two. 9 of 10 are controlled on the Isis path; training rights is CONTROLLED ONLY IN ANOTHER DOMAIN (32 modules implement it, none of them on the Isis generation path) and brief injection is CONTROLLED AND NOT TESTED (4 on-path modules, 0 tests naming them). The remaining eight are controlled-and-tested-not-mapped: the control and the test both exist and nothing ties either to a scenario id, so the parent's bar ("mitigations map to test IDs") is unmet even where the mitigation is real. THE CONTROLS THAT EXIST DO REFUSE, EXECUTED, so the gap is the map and not the machinery: a ComfyUI-shaped graph inside a curated request is caught by auditRawGraphLeak (2 tokens); a below-tier caller and a kill-switched card are both denied by checkCardEntitlement; and an ordinary brief passes unflagged, so the refusals are not noise. TRAP, and the reason every vocabulary here is word-anchored: unanchored, csam matches dedTemporalSemanticSample and credited the gaussian-splatting index as a child-safety control, inflating that threat from 3 modules to 11. A control test pins the collision. Cells left unchecked: .a–.j each need their own enumeration of entry points and residual-risk owners, which this instrument does not attempt, and .k needs a map that does not exist. -->
  • I0.11 Define data classification and retention for prompts, references, biometric embeddings, model artifacts, outputs, logs, and provenance. Evidence: privacy/security approval.

    • I0.11.a Classify prompts/briefs/templates/negative prompts/system policy/ tool arguments by content, tenant, source identity, secrecy, and purpose. Evidence: prompt data rows.

    • I0.11.b Classify uploaded/fetched/link/reference media, metadata, protected identities, rights/consent, source, and derived analysis. Evidence: reference data rows.

    • I0.11.c Classify face/voice/body/motion embeddings/templates/features/ similarity scores and prohibit or constrain persistence by purpose/ consent. Evidence: biometric rows.

    • I0.11.d Classify model/checkpoint/adapter/LoRA/training dataset/ optimizer/artifact/config/evaluation/license/security metadata. Evidence: model artifact rows.

    • I0.11.e Classify generated/intermediate/partial/quarantined/rejected/ approved/published/recalled outputs, derivatives, hashes, and metadata. Evidence: output rows.

    • I0.11.f Classify provider/worker/API/security/audit/debug logs/traces/ metrics/profiles/support exports and prohibit sensitive payloads by default. Evidence: telemetry rows.

    • I0.11.g Classify provenance manifests/ledger/ingredient/action/actor/ signature/trust and protected/redacted identity fields. Evidence: provenance rows.

    • I0.11.h For each class define collection/minimization, legal basis/ consent, purpose, encryption, region/provider, access, sharing, training use, retention, hold, deletion, backup, incident, and subject rights. Evidence: complete lifecycle matrix.

    • I0.11.i Obtain named privacy/security/safety/rights/domain approval and test enforcement against stores/providers/logs/exports/deletion. Evidence: approvals and data lifecycle tests. <!-- MEASURED 2026-08-17 (I0.11.a–i). THE ESTATE HAS THE REGISTER THIS CELL ASKS FOR AND IT IS GOOD. OshunArtifactType in @oshun/contracts is a 28-value alphabet, and every value carries an export scope (OSHUN_ARTIFACT_SCOPES: exportable, formats, deletable, deletion mode, retention-after-deletion, customer visibility) plus a retention rule (OSHUN_RETENTION_RULES: default window, legal-hold extension, retention driver, purge cadence, policy slug). So the question is never whether a row SHAPE exists; it is WHICH ROW, and the answers are three. evidence/v1-workbenches/isis-data-classification-matrix.json (writer + 17 tests, wired into verify:inventory:v1-workbenches) parses the seven classes out of .a–.g's own text — each states its bar as a ROW — and asks four separately-failing questions of each.

      FIRST, FOUR OF THE SEVEN CLASSES ARE NAMED ONLY IN PART AND ONE NOT AT
      ALL. `isKnownArtifactType` returns false, EXECUTED, for `prompt`,
      `negative_prompt`, `reference_media` and `biometric_template`. So
      `GenerationJob.prompt` and `negativePrompt` — `@db.Text` columns holding
      the most sensitive free text Isis accepts — have no register row at all,
      and .b (reference media) has none of any kind: it persists as a bare
      `GenerationJob.inputUrl` and inside `Provenance.inputs`.
      `workflow_template` is a row for templates and NOT for a prompt;
      `voice_profile` is a row for a voice and NOT for a face embedding. A
      NEAR-MISS THAT FILLS A COLUMN IS WORSE THAN AN EMPTY ONE, so each row
      lists `termsWithNoRegisterRow` rather than counting itself covered.
      
      SECOND, THE ISIS PATH REACHES THE REGISTER AND BINDS THE WRONG ROW. 11
      of 68 Isis BFF modules bind an `OSHUN_RETENTION_RULES` rule and the
      histogram is `{review_package: 9, session: 1, billing_receipt: 1}` — NOT
      ONE binds `generated_artifact`, the row written for Isis output. Measured
      against the register itself: `review_package` is 2555 days on a
      `compliance_window` driver purging every 30, `generated_artifact` is 365
      on `creation_timestamp` purging every 7. A DIFFERENCE OF 2190 DAYS, so
      cost tracking and texture quality are held seven years under a
      trust-and-safety audit rationale written for appeals. A further 19
      modules declare a self-described disposition instead
      (`not_applicable` ×10, `request_only` ×4, `request_lifetime` ×2, `none`,
      `inherits_output_gallery`), which is honest where nothing durable is
      held and is not a classification.
      
      THIRD, THE CLASSIFICATION ITSELF IS ABSENT. The Isis Prisma schema
      declares NO classification, sensitivity, legal-basis or subject-rights
      column of any kind, and zero modules on the Isis path name a
      data-classification vocabulary. The positive control is NAMED rather
      than counted — a repo-wide count inside a `--check`ed artifact goes red
      on somebody else's merge — so the artifact carries six alphabets the
      estate declares elsewhere (Iris, Lilith, Lakshmi, Maat, Aphrodite,
      Yemaya) and every one is a ladder of sensitivity VALUES:
      `public|internal|confidential|restricted|top_secret`,
      `PERSONAL|SENSITIVE_PERSONAL|PSEUDONYMIZED|ANONYMIZED`. A SENSITIVITY
      LEVEL IS AN ANSWER AND THIS CELL ASKS FOR THE QUESTION — none of the six
      can name a prompt, a reference or a provenance manifest, because they are
      columns looking for rows.
      
      LIVE DEFECT, AND IT IS NOT A POLICY QUESTION: THE SUBJECT-ERASURE PATH
      DELETES FROM A TABLE NOTHING PRODUCES. `server.ts` composes the
      account-deletion fan-out's `generated_artifact` partition from
      `resolveGeneratedArtifactDeleter`, which reaches `@isis/database` and
      runs `generatedOutput.findMany`/`deleteMany` against `generated_outputs`.
      Of the 24 models the Isis schema declares, 6 have a producer among the
      Isis Prisma client's 25 consumers and ALL SIX are the `pipeline_state_*`
      family; `GeneratedOutput` is produced by NOTHING and touched only by
      `libs/oshun/privacy` — the export bundler reads it, the eraser deletes
      from it. Meanwhile the DEPLOYED `apps/isis/output-registry`
      (`infra/terraform/ecs-services/isis.tf`, built in CI) issues `CREATE
      TABLE IF NOT EXISTS outputs`, inserts into `outputs`, and every one of
      its six retention-sweep statements writes `outputs` — on the same
      `ISIS_DATABASE_URL ?? DATABASE_URL`. The eraser and the deployed service
      share ZERO tables. `if (outputs.length === 0) return 0` makes a zero-row
      erasure indistinguishable from a subject who owned nothing, so the
      fan-out signs SUCCESS and purges no bytes. THE ESTATE CONTAINS ITS OWN
      ANSWER AND THE ERASURE PATH DID NOT READ IT:
      `apps/oshun/bff/src/isis/output-lineage-store.ts` already states "No
      production generation path writes those normalized tables" and declares
      `databaseSchema: 'declared_not_producer_bound'`. NOT REPAIRED HERE ON
      PURPOSE — the fix is either repointing a GDPR deletion path or
      re-plumbing a deployed service's storage, both decisions with blast
      radius across a signed-attestation contract, and picking one silently is
      worse than reporting it. A ratchet test pins
      `tablesTheEraserAndTheServiceShare: []` so the divergence cannot close by
      accident or widen unnoticed.
      
      THE MACHINERY IS NOT THE PROBLEM AND WAS EXECUTED TO PROVE IT. The
      register resolves a real scope (`generated_artifact` → exportable,
      deletable, `soft_delete_then_purge`, 30 days after deletion; 22 of 28
      exportable, 17 deletable). The Isis-owned 3D evaluator returns a real
      `deny` with two obligations for a customer upload past its window, and
      REFUSES an `artifactKind` of `"prompt"` outright — "3D model retention
      policy does not define a rule for \"prompt\"" — so the alphabet
      genuinely cannot express the class. And `@oshun/audit-platform`'s
      retention engine (repaired under S10.7) tiers a tagged event to
      `past_retention` correctly while returning the IDENTICAL untagged
      `domain: "isis"` event, whose `newState` carries a prompt, as `tier:
      hot, policyId: null` at ageDays 1461 — four years old and still in the
      hot tier, because nothing on the Isis path writes a `retentionTag`.
      
      (.h) THEREFORE CANNOT BE STARTED FOR ONE CLASS AND IS ONE-THIRD BUILT
      FOR THE REST: of 98 lifecycle cells, 30 are defined (retention, hold,
      deletion, sharing, subject rights — each naming the register field that
      carries it), 14 are unstartable because .b has no row at all, and 54 name
      an attribute the register has no field for (collection/minimization,
      legal basis/consent, purpose, encryption, region/provider, access,
      training use, backup, incident). "Complete lifecycle matrix" is unmet.
      
      TRAPS THAT COST TWO FALSE RESULTS HERE. (1) `git grep` without `-E`
      does not honour `\b`, and that reported the wired
      `createGeneratedArtifactSubjectDeleter` as having no caller when
      `server.ts` composes it — a false absence in the direction of a
      finding, corrected before it was written down. (2) Expanding the Prisma
      consumer set by BARE BASENAME put `index` and `types` in the pattern,
      `from './index'` matched 7,077 files across every domain, and `AuditLog`
      was credited with a producer in Iris; the second hop is now
      path-qualified to the importer's own directory and a control test caps
      the set. (3) `exports` as `\bexport(s|ed)?\b` matched the TypeScript
      keyword in every module and manufactured eight enforcement tests for a
      point with none — each point token now names the SUBJECT of the point.
      
      CELLS LEFT UNCHECKED, ALL NINE. .a–.g each need rows this instrument
      measures as absent or partial, and authoring them means choosing
      retention windows and legal bases — policy, not code. .h needs the 68
      undefined cells. .i's approval half is a NAMED HUMAN SIGN-OFF and is not
      locally actionable; its enforcement half has tests at 2 of the 5 points
      it names (`logs`, `deletion`) and none at `stores`, `providers` or
      `exports`. -->
      
  • I0.12 Define the operator emergency controls for provider disable, workflow revoke, model quarantine, output recall, and spend kill switch. Evidence: authority and blast-radius decisions.

    • I0.12.a Define provider disable scope by provider/endpoint/operation/ model/class/region/tenant/global, trigger, in-flight/queued behavior, fallback prohibition/allowance, recovery, and authority. Evidence: provider control decision.

    • I0.12.b Define workflow revoke scope by template/version/class/audience/ host, future dispatch, active jobs, outputs/releases, cache, replacement, and restoration. Evidence: workflow control decision.

    • I0.12.c Define model quarantine scope by model/version/checkpoint/adapter/ LoRA, intake/evaluation/dispatch/training/outputs/derivatives, evidence, investigation, revoke/restore. Evidence: quarantine decision.

    • I0.12.d Define output recall scope by artifact/run/ingredient/model/ workflow/release/target, dependent derivatives, gallery/catalog/cache, notifications/tombstone, provider/target withdrawal, and verification. Evidence: recall decision.

    • I0.12.e Define spend kill scope by actor/project/tenant/provider/model/ operation/resource/global, reservation/new dispatch/in-flight/retry/ queued behavior, essential exceptions, and recovery. Evidence: spend control decision.

    • I0.12.f Define request/approve/activate/revoke/expire/recover workflow with least privilege, step-up, separation/quorum, emergency break-glass, immutable audit, communication, and evidence. Evidence: common control contract.

    • I0.12.g Preview direct/transitive blast radius and make all actions idempotent, rapid, fail-closed, observable, and reconcilable. Evidence: blast-radius and fault tests.

    • I0.12.h Rehearse each control and combined incidents, measure propagation and recovery, and link stop-the-line/runbooks. Evidence: game-day report. <!-- MEASURED 2026-08-17 (I0.12.a–h). A CONTROL IS A VERB AND EVERY OTHER PART OF SPEECH FAILS DIFFERENTLY. The five controls differ not in how much of each is built but in WHAT KIND OF THING each is, so kind is the first column of evidence/v1-workbenches/isis-emergency-controls.json (writer + 17 tests, wired into verify:inventory:v1-workbenches) and it is derived, never asserted: {provider disable: absent, workflow revoke: absent, model quarantine: absent, output recall: operator-command, spend kill switch: frozen-literal}. Each of .a–.e has the shape "Define <control> scope by <slash-list>, <clause>, …", so the cell hands over its own matrix — 33 scope dimensions and 23 further aspects, all parsed from the checklist.

      EXACTLY ONE OF THE FIVE IS AN OPERATOR COMMAND, AND IT IS REAL. `POST
      …/bulk-actions` in `admin-isis-output-gallery-actions.ts` is admin-scoped
      (`hasAdminScope`), takes a `reason`, is idempotent on `invocationId`, and
      has an undo endpoint. Its scope is an explicit `outputIds` LIST, not a
      query, so of the seven dimensions .d names it can express ARTIFACT and
      nothing else: `run`, `ingredient`, `release` and `target` are absent, and
      `model`/`workflow` are `filterable-only` — A FIELD THE RECORD CARRIES IS
      NOT A SCOPE THE ACTION ACCEPTS, and crediting `OutputRecord.model` would
      have credited the caller's own filtering to the control. Executed: a
      takedown sets `takenDownAtUnixSeconds` on the listed id and leaves the
      bystander `null`, and `OutputRecord` carries NO lineage field at all — so
      "dependent derivatives" is not unimplemented, it is UNEXPRESSIBLE ON THE
      SHAPE. The estate agrees with itself: `output-lineage-store.ts` declares
      `takedownCascade: 'unavailable'`. Only 2 of 33 dimensions across all five
      controls are expressible.
      
      THE SPEND KILL SWITCH IS TWO THINGS AND NEITHER IS A SWITCH. Executed:
      `evaluateBudget` at 97% projected returns `status: "kill-switch-armed",
      killSwitchArmed: true` and ZERO fields that could block, cancel or stop
      anything — the word "armed" is carrying the whole claim, and its only Isis
      consumer is `RunpodDashboardPanel.tsx`, which renders it. The curated-card
      `killSwitchActive` is worse: all EIGHT catalog cards carry `false`, every
      `spec` is `Object.freeze`d, and assigning `true` raises "Cannot assign to
      read only property 'killSwitchActive'" while the value stays `false`. A
      SWITCH THAT ONLY A DEPLOY CAN THROW IS NOT AN EMERGENCY CONTROL. The
      enforcement point is fine and was checked as a control: the gate returns
      `deny / "kill-switch active"` when the flag IS set, so the gap is the
      switch and not the gate.
      
      PROVIDER DISABLE IS AN AUTOMATIC BREAKER, WHICH IS A DIFFERENT QUESTION.
      `unified-generation-pipeline.ts` keeps a PRIVATE, in-process
      `Map<string, {open, failures}>` that trips on consecutive failures — no
      operator trigger, no authority, no route, and per-instance so it does not
      propagate to another worker. AN AUTOMATIC BREAKER ANSWERS "IS IT FAILING"
      AND AN EMERGENCY CONTROL ANSWERS "STOP USING IT": a provider that is
      healthy and unsafe passes every breaker. It is recorded as an
      `automaticSurrogate` and kept OUT of `kind`, so the row can be neither
      misread as a control nor as an empty domain. WORKFLOW REVOKE AND MODEL
      QUARANTINE DO NOT EXIST under any verb — the only `quarantine` on the
      path is 3D delivery OUTPUT quarantine, a different subject, and I1.2.d
      separately measured six of ten model-version stages, quarantine among
      them, as having no command. Their 18 dimensions are reported `no-control`
      rather than `absent`, because a zero invites the reading that somebody
      looked at the dimension.
      
      (.f) THE COMMON CONTRACT HAS NO TEETH ON ANY CONTROL. Of its seven
      properties, 2 land on a module that implements one of the five (`least
      privilege` via `hasAdminScope`, `evidence` via marketplace abuse
      reporting); `emergency break-glass` and `communication` are absent from
      the entire Isis surface; and `step-up`, `separation/quorum` and `immutable
      audit` match REAL machinery aimed at a different subject — CLI login
      reauthentication, a release gate, a provenance bundle. Every match is
      therefore partitioned by whether it lands on a control, because a
      property of the domain is not a property of the control. The closest
      common contract is 3D-model-governance exception approval
      (requested/approved/renewed/revoked/expired/evaluated): it covers four of
      .f's six verbs, misses `activate` and `recover`, and binds none of the
      five. (.g) A REVERSE DEPENDENCY INDEX EXISTS AND NO CONTROL CONSULTS IT —
      I5.13 shipped `evidence/v1-workbenches/blast-radius.json`, and zero
      control modules reference it; of .g's five action properties, 3 land on a
      control (`idempotent`, `fail-closed`, `reconcilable`), `observable` is
      real elsewhere and on no control, and `rapid` (a propagation bound) is
      absent from the whole surface. (.h) NO GAME-DAY ARTIFACT REHEARSES ANY OF
      THE FIVE, and real game-day artifacts do exist, so the zero is about the
      controls.
      
      THE HONEST SEAMS ARE RECORDED RATHER THAN COUNTED AGAINST THE ESTATE.
      `budget-management-store.ts` declares `killSwitchEnforcement:
      'unavailable'`, `mutations: 'unavailable'`, `authorityBinding:
      'not_configured'` and says returning evaluated samples "would create a
      second, fabricated authority" — a refusal to invent an authority is the
      right behaviour for a control that does not exist, and it is why these
      rows are `absent`/`frozen-literal` and not stubs.
      
      FOUR MATCHERS MANUFACTURED A POSITIVE BEFORE CORRECTION, EACH PINNED BY A
      CONTROL TEST. (1) Matching the NOUN instead of the verb made
      `apps/isis/generation-api/src/routes/jobs.ts` — a breaker consumer — a
      provider-disable route; worse, the route filter `/isis/` matched because
      THE DOMAIN NAME IS IN THE PATH, so the test is now the basename
      `admin-isis-*.ts`. (2) `killSwitchActive:` matched an interface member and
      a frozen object literal, reporting a compile-time constant as a mutation
      site; state writes are now measured at LINE level and declaration lines
      (`const`/`let`/`readonly`) are dropped, which also stopped `const
      killSwitchArmed = …` reporting the pure evaluator as mutating what it
      evaluates. (3) `recall` is never the vocabulary here — on the Isis path it
      is the ML metric (`rougeLRecall`, `minRecall`, `recallPass`,
      `totalRecall`) and it credited an evaluation suite and a grounding gate as
      recall controls; the estate's verb is `takedown`. (4) A game-day artifact
      is one whose PATH names a game day: filtering on files that merely MENTION
      one put `docs/systems/lib-v6.md` and a workbench export map in the set,
      and a phrase match inside them then read as a rehearsal.
      
      CELLS LEFT UNCHECKED, ALL EIGHT. .a, .b and .c need controls that do not
      exist; .d and .e need scopes, cascades and an authority the shapes cannot
      express; .f needs a contract that binds the controls; .g needs a preview
      that consults the index that already exists; .h needs a rehearsal of
      controls three of which cannot be rehearsed. Every cell's bar is a
      DECISION about authority and blast radius, which is a human call, so
      nothing here was authored on the instrument's behalf. -->
      
  • I0.13 Author and ratify V1/ISIS_GENERATION_WORKBENCHES_DESIGN_2026-07-23.md with current-state, audience, host, data, API, job, UI, trust, migration, rollout, and decision records. Evidence: approved doc and resolved review comments.

    • I0.13.a Document code-derived current applications/routes/components/ contracts/packages/entities/jobs/providers/tests, live/fixture/dead status, and reusable/gap disposition. Evidence: current-state section.

    • I0.13.b Document operator/AAA/curated/contemplative actors, roles, capabilities, forbidden detail/actions, approvals, and boundary matrix. Evidence: audience section.

    • I0.13.c Document canonical hosts/routes/navigation/deep links/embeds/ handoffs/proxies/redirects and migration from noncanonical surfaces. Evidence: host section.

    • I0.13.d Document asset/workflow/model/brief/plan/run/output/lineage/gate/ rights/consent/provenance/release data contracts, stores, retention, and migrations. Evidence: data section.

    • I0.13.e Document APIs/events/jobs/providers/bundles/publishers, authentication/authorization, idempotency/concurrency, limits, errors, health, and compatibility. Evidence: service sections.

    • I0.13.f Document queue/worker/resource/cost/cancel/retry/checkpoint/ fault/recovery/emergency operations. Evidence: job/operations section.

    • I0.13.g Document restrained operator/AAA/curated/gallery workspace hierarchy, responsive behavior, states, accessibility/nonvisual media/ graph/3D alternatives, localization, and utility copy. Evidence: UI section.

    • I0.13.h Document threat model, data classification, safety/child/ impersonation, rights/consent, provenance/trust, gates/overrides, incident/recall, and evidence. Evidence: trust section.

    • I0.13.i Document incremental reuse/refactor/data/route/consumer migration, compatibility, rollback, deprecation/deletion, environment/standards, verification, rollout/cohort, SLO, support, and DR. Evidence: migration/ rollout sections.

    • I0.13.j Record every product/architecture/security/privacy/safety/rights/ host/ownership/standard decision with options/evidence/rationale/ consequences/approver/review. Evidence: decision appendix.

    • I0.13.k Run source link/schema/route/inventory/requirement consistency checks, obtain named reviews, resolve every comment, and ratify exact document revision. Evidence: approval record and closed comment ledger. <!-- AUTHORED 2026-08-17 (I0.13.a–k). THE DOCUMENT NOW EXISTS: V1/ISIS_GENERATION_WORKBENCHES_DESIGN_2026-07-23.md, 343 lines, eleven sections, GENERATED by scripts/v1-workbenches/generate-isis-design-doc.mts (+ 13 tests) and --checked byte-for-byte as the LAST step of verify:inventory:v1-workbenches.

      IT IS GENERATED BECAUSE .a FIXES THE STANDARD FOR ALL ELEVEN SECTIONS:
      the current state must be CODE-DERIVED. So every number and every list in
      it is read out of a committed evidence artifact at build time — **133
      measured claims from 19 artifacts**, each read by a dotted path through a
      `fact()` accessor that THROWS when the path is absent rather than
      interpolating `undefined` into a design document, and each recorded so the
      document carries its own source ledger (§13). Every repo path it cites is
      checked to exist. A DESIGN DOCUMENT ROTS AND A MEASUREMENT DOES NOT: a
      route that appears, a package that is retired or a gate that starts
      passing now makes this document stale LOUDLY. A ratchet test also refuses
      any interpolated `undefined`, `NaN`, `[object Object]` or bare `null` —
      the first run rendered `the registry module is null`, which reads as a
      fact and is not one.
      
      WHAT THE DOCUMENT ESTABLISHES THAT WAS NOT WRITTEN DOWN ANYWHERE. §12 is
      a decision register derived from the checklist itself: every Phase I cell
      whose own bar names a decision, an approval or a ratification —
      **29 cells, all open by construction** since each is unchecked. It is
      split three ways by a PUBLISHED rule, because "decision appendix",
      "decision tests" and "threshold decision record" all contain the word
      *decision* and none of them is a decision anybody has to take: **14 are
      waiting on a named person** (including all five I0.12 control decisions
      and I0.11), 10 on a person AND on engineering work, and 5 are engineering
      work about decisions. Mixing them would make the register useless for the
      one thing it is for. `approver` and `review` are EMPTY in every row and
      the document says why — recording an approver who does not exist is the
      exact failure this lane was built to catch.
      
      TWO FINDINGS CAME OUT OF WRITING IT, BOTH CORROBORATING I0.12 FROM
      ARTIFACTS THAT WERE NOT LOOKING FOR IT. (1) The endpoint DISPATCH MODEL
      (I1.3.b–e) names an eight-state lifecycle including `disabled`, and NO
      ENUM IN SOURCE IMPLEMENTS IT — the only endpoint states the code has are
      the four health states `healthy|degraded|unhealthy|unknown`. A PROPOSED
      LIFECYCLE IS NOT A CONTROL, and a reader who took `disabled` from the
      model would conclude an operator can disable a provider. I checked this
      BEFORE trusting the I0.12 verdict, and the verdict survived. (2) The
      workflow job machine's unimplemented transitions are
      `submit|review|approve|REVOKE` — measured independently of I0.12.b, which
      found workflow revoke absent under any verb.
      
      §13.1 IS THE REASON EVERY CELL STAYS UNCHECKED, and it is a table rather
      than a paragraph: per child, which sections answer it and WHAT THE CHILD
      ASKS FOR THAT THE DOCUMENT DOES NOT CONTAIN. That column is a stated
      judgment, published so it can be argued with, because a generated
      document could claim completeness by matching words and A WORD APPEARING
      IN A SECTION IS NOT THE SAME AS THE SECTION DOCUMENTING THE SUBJECT. The
      gaps, in short: .a lacks a component-level inventory below the page; .b
      lacks the forbidden-detail lists and approval matrix (decisions, I0.2.e/g);
      .c lacks deep links, embeds, handoffs, proxies and redirects; .d lacks
      per-record migrations and the retention rules themselves (I0.11); .e lacks
      idempotency, concurrency, limits, error taxonomy, health and compatibility
      per API, and events/bundles/publishers entirely; .f documents only the
      emergency half and not queue/worker/resource/cost/cancel/retry/checkpoint;
      .g lacks workspace hierarchy, responsive behaviour, states, localization
      and utility copy; .h lacks safety/child/impersonation and rights/consent
      detail, gates and overrides, and the incident/recall runbook; .i lacks
      rollback, deprecation, cohorts, SLO, support and DR; .j lacks the approver
      and review; .k lacks named reviews, a resolved comment ledger and a
      ratified revision. A test asserts that no I0.13 cell can be checked while
      its row states a gap.
      
      CELLS LEFT UNCHECKED, ALL ELEVEN PLUS THE PARENT. The parent's bar is
      "approved doc and resolved review comments" and .k's is an approval
      record — neither is locally actionable. .a–.i each have a stated gap
      above; several of those gaps are themselves decisions listed in §12, so
      closing them is not writing prose. The document reports its own status as
      **unratified** in its first line and quantifies it in §13.2 (Phase I: 757
      of 1582 cells checked, counted between the `## Phase I` and `## Phase M`
      headers rather than by a line range — the range I first used was stale by
      the length of these notes), so a thorough current-state section cannot be
      mistaken for approval of it. -->
      
  • I0.14 Reconcile the complete ISIS_TODOS_2.md asset-family matrix: character concept, turnaround, expression, costume, prop-interaction, NPC-crowd and creature packs; environment concept, matte-painting, tileset, architecture, interior, prop, weapon, vehicle, PBR material, decal, terrain, foliage, sky, VFX, UI/HUD, storyboard, animatic, cinematic, marketing, social, audio-SFX, voice-line, and music-cue outputs. Evidence: every matrix row has an audience, authorized host, workflow/template, controls, output bundle, preview, quality profile, and release disposition. <!-- THE RECONCILIATION IS BUILT AND IT FOUND SOMETHING; THE PARENT AND ITS CHILDREN STAY [ ]. isis-asset-family-matrix.json — generate-isis-asset-family-matrix.mjs + .test.mjs (14 tests, 4 of them ratchets on findings), wired into inventory + verify.

    The family list is read from `ISIS_TODOS_2.md` itself, not from a list in
    the generator: a family the matrix names and the estate never built would
    otherwise be invisible, which is the failure mode a reconciliation exists
    to catch. 32 families, 18 MXA items each, 576 of 576 marked `[x]` in that
    document.
    
    WHAT THE ESTATE ACTUALLY HAS IS UNIFORM AND SUBSTANTIAL. All 32 families
    carry all 8 conventional artifacts. All 32 declare the SAME four workflow
    variants (baseline / image-conditioned / control-conditioned /
    consistency-locked) — one shape, no exceptions. All 32 expose five SDK
    methods (a generic dispatcher plus one per variant), a quality evaluator, a
    cost estimator and a budget guardrail, and a golden regression dataset;
    192 golden cases in total. File existence is a banned proxy, so each of
    those is the artifact's CONTENTS, read from source.
    
    SEVENTEEN OF THIRTY-TWO FAMILIES ARE GATED ON A FACE AND A WARDROBE.
    Grouping the families by their quality metric set — the one thing a
    per-family read cannot see — gives 14 distinct sets, and the largest is
    shared by 17: `face_embedding_cosine_similarity`,
    `facial_landmark_drift_px`, `garment_attribute_match_ratio`,
    `wardrobe_palette_delta_e`. Seven of those 17 are character families and
    the metrics fit. The other TEN are `architecture-kitbash-pack`,
    `environment-concept-art`, `environment-matte-painting`,
    `environment-tileset-pack`, `interior-scene-pack`, `prop-pack-general`,
    `terrain-heightmap-pack`, `vehicle-exterior-pack`,
    `vehicle-interior-pack` and `weapon-pack`. A TERRAIN HEIGHTMAP IS SCORED
    ON HOW WELL IT MATCHES A WARDROBE PALETTE.
    
    IT IS NOT INERT AND IT IS NOT ENFORCED, AND BOTH HALVES MATTER. Every one
    of the 17 gates DECLARES those metrics in its own type union, thresholds
    each from its own env key, and its golden regression cases SUPPLY VALUES
    FOR THEM — `terrain-heightmap-pack` ships a case with
    `face_embedding_cosine_similarity: 0.842`. So the metrics are live in the
    regression suite. But all 32 gates read
    `ISIS_<FAMILY>_QUALITY_GATE_ENABLED` defaulting to FALSE, so nothing is
    being wrongly rejected today. The defect is in what the estate would
    measure if it turned the gates on.
    
    THE ESTATE KNOWS HOW TO WRITE THESE, WHICH IS WHY THIS IS A COPY AND NOT A
    LIMITATION. Twelve families carry metrics written for their subject:
    `material-pbr-pack` checks seam tileability and roughness/metallic
    accuracy, `ui-icon-pack` stroke-weight consistency and contrast
    compliance, `vfx-element-pack` temporal flicker and alpha edge halo,
    `foliage-pack` canopy coverage and species mix, `voice-line-pack` phoneme
    alignment and prosody drift, `music-cue-pack` harmonic drift and cue
    timing. A second shared set is legitimate: storyboard, animatic and
    cinematic genuinely share blocking/camera/composition/style-lock.
    
    WHY THE PARENT AND ALL CHILDREN STAY OPEN. The evidence clause wants every
    row to carry audience, authorized host, workflow/template, controls,
    output bundle, preview, quality profile AND release disposition. Four of
    those are measured here. Audience and authorized host are the same
    question I0.2 measured as unratified and I0.1.e recorded as owed, and a
    release disposition is a decision. Filling them from this generator's
    opinion would read like a ratified answer nobody gave.
    
    THREE WAYS THIS INSTRUMENT UNDER-REPORTED BEFORE IT WAS RIGHT, ALL NOW
    UNDER A CONTROL TEST. A PascalCase guess at symbol names reported
    `ui-icon-pack` as having NO SDK method, NO quality evaluator and NO cost
    estimator — the estate writes acronyms in caps, so they are
    `runUIIconPack` and `evaluateUIIconPackQualityGate`; matching is now
    case-insensitive on the de-hyphenated family name. Requiring `(` straight
    after a method name missed every family's GENERIC dispatcher
    (`async runX<TWorkflowId extends …>(`), counting four SDK methods where
    there are five. And reading every `name: <number>` in a golden dataset
    swept in each case's REQUEST parameters — `cfg_scale`, `focal_length_mm`,
    `height` — and reported dozens of quality signals no gate reads; signals
    are now read only from `qualitySignals` blocks, and the true count is
    zero. -->
    
    • I0.14.a Reconcile character-concept outputs independently, including single/multi-view intent, identity/style references, anatomy/silhouette, design variation, audience/host, workflow, controls, bundle, preview, quality/rights/provenance gates, and release. Evidence: completed character-concept row.
    • I0.14.b Reconcile character-turnaround outputs independently, including canonical views, camera/pose/scale consistency, identity/wardrobe locks, sheet layout, downstream modeling, and release proof. Evidence: completed turnaround row.
    • I0.14.c Reconcile expression-sheet outputs independently, including expression taxonomy/intensity, identity/anatomy consistency, layout, rig/reference handoff, and safety/quality proof. Evidence: completed expression row.
    • I0.14.d Reconcile costume outputs independently, including garment/ material/accessory variants, identity/body/era/style locks, rights/ cultural review, turnaround/fit preview, and downstream handoff. Evidence: completed costume row.
    • I0.14.e Reconcile prop-interaction outputs independently, including character/prop identity, grasp/contact/occlusion/physics, action poses, scale, rig/animation handoff, and quality gates. Evidence: completed prop-interaction row.
    • I0.14.f Reconcile NPC-crowd pack outputs independently, including population taxonomy, identity diversity, variation rules, LOD/atlas/ instancing, bias/safety, engine bundle, and performance gates. Evidence: completed NPC-crowd row.
    • I0.14.g Reconcile creature-pack outputs independently, including species/ anatomy/size variants, turnaround/expression/motion/rig needs, safety, downstream formats, and quality. Evidence: completed creature-pack row.
    • I0.14.h Reconcile environment-concept outputs independently, including biome/location/era/mood, composition/scale/perspective, camera/lighting/ palette locks, production breakdown, and world-building handoff. Evidence: completed environment-concept row.
    • I0.14.i Reconcile matte-painting outputs independently, including plate/ camera/perspective/lighting/color/depth, alpha/layer/overscan, projection, EXR delivery, and compositing/rights/quality proof. Evidence: completed matte-painting row.
    • I0.14.j Reconcile tileset outputs independently, including grid/tile/ adjacency/rules, seamless variants, collision/metadata, atlas/engine packaging, preview map, and validation. Evidence: completed tileset row.
    • I0.14.k Reconcile architecture outputs independently, including style/ period/function, dimensions/units/modularity, interior/exterior, material/ code considerations, LOD/collision, DCC/engine handoff, and gates. Evidence: completed architecture row.
    • I0.14.l Reconcile interior outputs independently, including room program/ layout/scale, furniture/props/materials/lighting, camera/navigation, modular breakdown, DCC/engine bundle, and quality. Evidence: completed interior row.
    • I0.14.m Reconcile standalone prop outputs independently, including function/scale/material/wear/variants, turnaround, mesh/texture/LOD/ collision targets, rights, and engine/DCC release. Evidence: completed prop row.
    • I0.14.n Reconcile weapon outputs independently, including fictional/ historical context, safety policy, dimensions/mechanism/material/ variants, hands/animation, mesh/LOD/collision, rights, and audience restrictions. Evidence: completed weapon row.
    • I0.14.o Reconcile vehicle outputs independently, including class/scale/ design/variants, exterior/interior, wheel/part/rig/animation, materials/decals, LOD/collision, and simulation/engine handoff. Evidence: completed vehicle row.
    • I0.14.p Reconcile PBR-material outputs independently, including workflow/ shader model, base color/normal/roughness/metallic/AO/height/emissive/ opacity channels, tiling/scale/color space, preview, and target export. Evidence: completed PBR-material row.
    • I0.14.q Reconcile decal outputs independently, including purpose/ dimensions/alpha/channel packing/projection/atlas/variants, trademark/ rights, surface preview, DCC/engine format, and validation. Evidence: completed decal row.
    • I0.14.r Reconcile terrain outputs independently, including extent/units/ heightfield/erosion/biomes/masks/material layers/LOD/collision/streaming, GIS/source rights, preview, and engine package. Evidence: completed terrain row.
    • I0.14.s Reconcile foliage outputs independently, including species/ growth/season/variation, cards/geometry/material/wind/LOD/collision/ density, ecological/source rights, and engine bundle. Evidence: completed foliage row.
    • I0.14.t Reconcile sky outputs independently, including time/weather/ location/HDRI/cloud/sun/stars/color/exposure, environment-map format, lighting preview, rights, and engine/DCC handoff. Evidence: completed sky row.
    • I0.14.u Reconcile VFX outputs independently, including effect taxonomy/ duration/space/timing, sprites/flipbooks/particles/volumes/materials/ simulation, loop/LOD/performance, preview, and engine package. Evidence: completed VFX row.
    • I0.14.v Reconcile UI/HUD outputs independently, including platform/ resolution/state/components/icons/text/localization/accessibility/ interaction, style consistency, atlas/vector/source delivery, and rights. Evidence: completed UI/HUD row.
    • I0.14.w Reconcile storyboard outputs independently, including script/ shot/scene/panel IDs, composition/action/dialogue/camera/timing/notes, continuity, revision, accessibility, and editorial package. Evidence: completed storyboard row.
    • I0.14.x Reconcile animatic outputs independently, including ordered panels/shots, timing/camera/transitions/dialogue/music/SFX/temp media/ captions, EDL/timecode, review, and editorial export. Evidence: completed animatic row.
    • I0.14.y Reconcile cinematic outputs independently, including sequence/ shots/cameras/animation/lighting/VFX/audio/titles/captions/color/ delivery, render jobs, editorial/DCC/engine package, and release gates. Evidence: completed cinematic row.
    • I0.14.z Reconcile marketing outputs independently, including campaign/ channel/audience/territory/format/copy/brand/rights/disclosure/ accessibility/approval/variant tracking and delivery. Evidence: completed marketing row.
    • I0.14.aa Reconcile social outputs independently, including platform/ placement/dimensions/duration/safe area/caption/audio/copy/hashtags/ accessibility/moderation/rights/schedule and derivatives. Evidence: completed social row.
    • I0.14.ab Reconcile audio-SFX outputs independently, including event/ category/material/action/perspective/variation/duration/sample/channel/ loop/loudness/metadata/rights and engine package. Evidence: completed audio-SFX row.
    • I0.14.ac Reconcile voice-line outputs independently, including character/ speaker/actor consent/language/text/pronunciation/emotion/timing/takes/ processing/loudness/caption/localization and engine package. Evidence: completed voice-line row.
    • I0.14.ad Reconcile music-cue outputs independently, including brief/ scene/mood/style/key/tempo/meter/duration/structure/stems/loop/hit points/ mix/master/rights and engine/editorial package. Evidence: completed music-cue row.
    • I0.14.ae Validate that every asset-family row independently names its audience, authorized host, workflow/template, typed controls, output contract/bundle, preview/semantic alternative, quality profile, safety/ rights/provenance gates, downstream handoff, and release/exclusion. Evidence: zero incomplete matrix rows.
  • I0.15 Reconcile every consistency and conditioning mode: seed and latent reuse; identity, style, wardrobe, prop, scene, lighting, palette, lens, camera, pose, depth, temporal, and motion locks; ControlNet pose/depth/ canny/line-art/segmentation/normal; IP-Adapter style/face; InstantID; open-pose sequences; camera paths; motion vectors; masks; reference ensembles; and multi-control blends. Evidence: per-mode audience exposure, typed contract, diagnostics, gate, and unsupported-state matrix. <!-- THE RECONCILIATION IS BUILT AND EVERY ROW CAME BACK INCOMPLETE; THE PARENT AND ALL THIRTY-ONE CHILDREN STAY [ ]. evidence/v1-workbenches/isis-conditioning-mode-matrix.json — generate-isis-conditioning-mode-matrix.mjs + .test.mjs (15 tests, 6 of them ratchets on findings, 4 controls that prove each scan can come back both ways), wired into inventory + verify. The mode list is read from THIS document, not from a list in the generator; .ae is separated from the thirty by its own wording ("Validate", not "Reconcile") rather than by being the last letter.

    SIX ALPHABETS SPEAK ONE VOCABULARY AND THEY DO NOT AGREE. The canonical
    dictionary (ISIS_V2_CONTROL_TAXONOMY.json) has 38 control ids; the I2.16
    conditioning reference has 11 typed kinds; the workflow registry has 8
    consistency dimensions; workflow.schema.ts has 7 continuity locks; the
    consistency runtime has 3 domains; and the entitlement resolver — the ONLY
    mechanism in the estate that decides which audience may reach which
    capability — has 28 generation surfaces, NOT ONE of which names a mode.
    Zero exact matches, and the three token overlaps are coincidental
    (`curated-living-scene` shares "scene" with `scene-lock`). So per-mode
    audience exposure is not owed as a decision here the way I0.14's was: it
    is NOT EXPRESSIBLE, and that is a measured negative on all thirty rows.
    Three modes have no canonical control id at all — prop lock, lighting lock
    and multi-control blends — and the first two DO have typed runtime state,
    so what they lack is a dictionary entry, not an implementation.
    
    THE GRAPHS REFUTE 401 OF THE 736 CONTROL CLAIMS THE REGISTRY MAKES. A
    control names the node families that implement it, and a graph containing
    none of them cannot apply it. 23 of the 37 families the dictionary names —
    CameraControl, PromptComposer, OpenPose, InsightFace, InstantID, Canny,
    ColorPaletteGuard, KeyframeControl, TileControl among them — do not exist
    as a node class in ANY of the 66 workflow graphs. The single workflow
    declaring InstantID (`consistent-character`) has no InstantID, no
    InsightFace and no ControlNet node. All five declaring controlnet-inpaint,
    including `inpaint-edit` and `outpaint-expand`, have no inpaint node at
    all. All four declaring lens-profile-lock lack its only family.
    
    THE TEST ONLY EVER REFUTES, WHICH IS WHY IT IS TRUSTWORTHY. A family is
    classified GENERIC or DISCRIMINATING by measuring its reach across the 66
    graphs rather than by assertion, so `seed-lock` riding on KSampler is
    reported as backing that proves nothing; a templated `class_type` is
    INDETERMINATE, never absent (8 pairs).
    
    THE DICTIONARY NAMES 37 REQUIRED INPUTS AND THE WORKFLOWS ACCEPT ONE.
    `seed` is a declared parameter on 63 of 66 workflows; no workflow declares
    a parameter named after any of the other 36. Eleven controls go further —
    latent-reuse, pose-constraint, depth-constraint, palette-lock,
    identity-adapter, reference-ensemble, keyframe-anchor-lock,
    motion-vector-lock and three ControlNet arms — their required input has
    ZERO occurrences in any TypeScript file and zero in any workflow, and
    exists only in the documents that declare it. reference-ensemble is
    declared on 31 of 66 workflows and nothing in the estate can supply it.
    
    A MODALITY CONTRADICTION NAMES A PAIR, NOT A CULPRIT (76 of them). An
    audio workflow cannot wear a wardrobe, and `sfx-sound-designer`,
    `music-cue-composer`, `voice-chatterbox` and `voice-performance-synthesis`
    all declare wardrobe-lock, camera-lock, pose-constraint and
    depth-constraint. But `song-section` — the one hand-written registry row —
    declares sampler-scheduler-lock, which the dictionary omits `audio` from,
    and there it is the DICTIONARY that is short an entry. A test pins that
    counter-example so the finding cannot harden into "the registry is wrong".
    
    THE SEVEN TYPED LOCKS THAT DO EXIST ARE PARSED AND NEVER READ.
    workflow.schema.ts declares character_identity_lock, style_lock,
    scene_lock, camera_lock, wardrobe_lock, prop_lock and temporal_lock,
    defaults every one to `unlocked`, and no file under libs/isis or apps/isis
    outside a spec mentions any of them. The one rigorous typed conditioning
    contract in the estate — the I2.16 reference with its declared conventions
    and named refusals — lives under evidence/ with its own tsconfig, is
    reachable by no path alias, and is imported by zero product files. For 15
    of the 30 modes that reference IS the whole typed contract.
    
    192 CERTIFICATION REPORTS TREAT THE STRATEGY AS A LABEL, NOT A VARIABLE.
    12 consistency strategies x 16 target classes. For 15 of the 16 targets
    all 12 reports are identical in metric names, thresholds, seeds AND
    observed values — only the filename and the caseId prefix differ. 158 of
    them carry a title naming a different strategy ("Seed Lock"). The
    depth-constraint x voice-persona-pack report observes speaker identity,
    phoneme alignment, prosody, timbre and LUFS, and nothing about depth.
    scene-lock is the control that proves the fields are meant to vary: it is
    titled correctly on all 16 and is the one strategy that swaps a metric
    (sceneLayoutIoU for wardrobePaletteDeltaE), so neither 158 nor 15 is an
    artifact of a comparison that always fires.
    
    AND THE VALIDATOR THAT WOULD HAVE CAUGHT THE DRIFT IS WRITTEN, CORRECT AND
    UNCALLED. `scripts/isis/validate-v2-certification-matrix.mjs` exits 1 with
    905 errors today: the certification matrix was built over 48 workflows and
    the registry now holds 66, leaving 18 — every flagship workflow among them
    — with no certification row at all. Its only mentions in the repo are
    inside two generators that print the command into documentation.
    
    TWO REPAIRS SHIPPED WITH THIS PASS, BOTH VERIFIED. (1) The registry header
    declared `workflowCount: 64` while heading 66 mappings and no test could
    see it, because the field was read into the document type and never
    exported: every count-shaped assertion in the suite compared the mapping
    count with the mapping count. It is now exported as
    WORKFLOW_TAXONOMY_REGISTRY_DECLARED_WORKFLOW_COUNT, corrected to 66, and
    pinned by an assertion that reads the DECLARED number. (2)
    `libs/isis/workflows/src/index.js` and `index.d.ts` were committed build
    output sitting beside `index.ts`, and Vite resolved `./index` to the .js —
    which is how the new export read back `undefined` and how this was found.
    Measured before claiming: the shadow was a stale EXPORT LIST over live
    implementations, not a stale build. It re-exported from extensionless
    specifiers (`./consistency`, `./utils/registry`) that resolve to the real
    `.ts` sources, so the values the specs imported were current; what it could
    not carry was any symbol added to `index.ts` after the snapshot — 16 of the
    barrel's 191 runtime exports (the catalog renderer, the expression
    evaluator, the seed helpers) plus anything new. Only the two in-package
    specs were affected; `tsconfig.base.json` aliases `@isis/workflows` to
    `src/index.ts`, so every consumer outside the package already read source.
    Removed; all 34 spec files / 377 tests pass, and tsconfig.lib.json
    typechecks clean. 113 more tracked `.js` files still shadow a sibling `.ts`
    elsewhere (111 in `libs/isis/ai-providers`, 26 of those barrels —
    including the controlnet, ip-adapter and instantid provider barrels); that
    sweep needs its own pass.
    
    WHY EVERY CELL STAYS OPEN. Each child asks for a completed row naming
    audience exposure, diagnostics, a gate and an unsupported state for its
    mode. Every one of the thirty rows names at least one of those as missing:
    30 lack per-mode audience exposure because no mechanism can express it, 17
    have no in-product diagnostic, 3 have no canonical control id, and 3 have
    every registry declaration refuted by the graphs. I0.15.ae wants zero
    incomplete mode rows and there are thirty. -->
    
    • I0.15.a Reconcile seed reuse with provider/workflow/version scope, deterministic limits, exposure by audience, replay provenance, diagnostics, and unsupported behavior. Evidence: seed row.
    • I0.15.b Reconcile latent reuse with latent format/model/workflow/version/ storage/privacy scope, compatibility, audience restriction, diagnostics, and reproducibility. Evidence: latent row.
    • I0.15.c Reconcile identity lock with consent/biometric/reference/ embedding/strength/scope/version, drift measures, impersonation gates, and audience policy. Evidence: identity-lock row.
    • I0.15.d Reconcile style lock with reference/licensing/artist-style policy, adapter/embedding/strength/scope, consistency diagnostics, and audience exposure. Evidence: style-lock row.
    • I0.15.e Reconcile wardrobe lock with garment/accessory/material/color/ era references, character/shot scope, occlusion/fit diagnostics, and variant behavior. Evidence: wardrobe-lock row.
    • I0.15.f Reconcile prop lock with prop identity/scale/material/state/ placement/interaction references, shot scope, diagnostics, and failure behavior. Evidence: prop-lock row.
    • I0.15.g Reconcile scene lock with environment/layout/object/spatial/ weather/time references, shot/sequence scope, diagnostics, and change control. Evidence: scene-lock row.
    • I0.15.h Reconcile lighting lock with source/rig/HDRI/time/exposure/color/ shadow references, shot/sequence scope, diagnostics, and target color profile. Evidence: lighting-lock row.
    • I0.15.i Reconcile palette lock with color list/space/roles/tolerance/ source, local/global scope, accessibility/brand constraints, and drift metrics. Evidence: palette-lock row.
    • I0.15.j Reconcile lens lock with focal length/sensor/FOV/distortion/ aperture/focus and camera/model/shot scope, unit validation, and diagnostics. Evidence: lens-lock row.
    • I0.15.k Reconcile camera lock with pose/path/orientation/intrinsics/ framing/stabilization/coordinate system and shot/sequence scope. Evidence: camera-lock row.
    • I0.15.l Reconcile pose lock with skeleton/keypoints/confidence/visibility/ proportions/contact, actor/character/shot scope, diagnostics, and accessibility alternative. Evidence: pose-lock row.
    • I0.15.m Reconcile depth lock with depth type/units/range/confidence/ occlusion/resolution/camera alignment, normalization, diagnostics, and unsupported state. Evidence: depth-lock row.
    • I0.15.n Reconcile temporal lock with frame/shot identity, cadence/time base, consistency window, memory/reference state, flicker/drift metrics, and reset behavior. Evidence: temporal-lock row.
    • I0.15.o Reconcile motion lock with optical flow/trajectory/velocity/ acceleration/camera-vs-object/masks/time base, diagnostics, and failure state. Evidence: motion-lock row.
    • I0.15.p Reconcile ControlNet pose conditioning with preprocessor/model/ version, keypoint schema, strength/range, resolution, diagnostics, license, and audience. Evidence: ControlNet-pose row.
    • I0.15.q Reconcile ControlNet depth conditioning with estimator/model/ version, units/normalization/alignment, strength/range, diagnostics, and audience. Evidence: ControlNet-depth row.
    • I0.15.r Reconcile ControlNet canny conditioning with threshold/aperture/ resolution/inversion/strength/schedule, diagnostics, and typed preview. Evidence: ControlNet-canny row.
    • I0.15.s Reconcile ControlNet line-art conditioning with preprocessor/ style/model/version/resolution/strength/schedule, preview, and diagnostics. Evidence: ControlNet-line-art row.
    • I0.15.t Reconcile ControlNet segmentation conditioning with vocabulary/ class map/palette/model/version/unknown class/strength, semantic preview, and diagnostics. Evidence: ControlNet-segmentation row.
    • I0.15.u Reconcile ControlNet normal conditioning with coordinate system/ range/encoding/camera alignment/model/version/strength, 3D preview, and diagnostics. Evidence: ControlNet-normal row.
    • I0.15.v Reconcile IP-Adapter style conditioning with reference rights/ preprocessing/embedding/model/version/weight/schedule/masking, style policy, and diagnostics. Evidence: IP-Adapter-style row.
    • I0.15.w Reconcile IP-Adapter face conditioning with identity consent/ face detection/embedding/model/version/weight/masking, impersonation/ biometric gates, and diagnostics. Evidence: IP-Adapter-face row.
    • I0.15.x Reconcile InstantID with identity image/landmarks/embedding/ adapter/control versions/strength, consent/impersonation/biometric policy, and identity drift metrics. Evidence: InstantID row.
    • I0.15.y Reconcile open-pose sequences with keypoint/skeleton schema, frame rate/time base/interpolation/confidence/person tracking, sequence preview, and temporal diagnostics. Evidence: open-pose row.
    • I0.15.z Reconcile camera paths with coordinate system/keyframes/ interpolation/lens/time base/collision/constraints, 3D and timeline preview, and downstream export. Evidence: camera-path row.
    • I0.15.aa Reconcile motion vectors with encoding/coordinate space/units/ scale/direction/time base/resolution/occlusion, visualization, and validation. Evidence: motion-vector row.
    • I0.15.ab Reconcile masks with source/semantic role/format/alpha/ resolution/alignment/feather/invert/animation/lifecycle, visual preview, and privacy/safety constraints. Evidence: mask row.
    • I0.15.ac Reconcile reference ensembles with ordered refs/roles/weights/ conflicts/rights/consent/compatibility, per-reference influence/ exclusion diagnostics, and lineage. Evidence: ensemble row.
    • I0.15.ad Reconcile multi-control blends with typed control set/weights/ schedules/masks/precedence/conflict/capacity, component/aggregate diagnostics, and unsupported combinations. Evidence: multi-control row.
    • I0.15.ae Validate every mode independently for audience/host exposure, typed inputs/units/version, provider/workflow compatibility, preview/ diagnostics, raw values/gates, provenance/rights/consent, and explicit unsupported/not-configured/degraded behavior. Evidence: zero incomplete mode rows.
  • I0.16 Reconcile specialized 3D capabilities in libs/isis/*: scene from image, part-level and semantic/text mesh editing, AI texturing, PBR relighting, mesh transformers, quad meshing, model fine-tuning, video-to-mesh, score distillation, advanced/diffusion-edited Gaussian splats, gigascale scenes, asset browsing/marketplace, and universal rigging. Evidence: each capability is explicitly exposed, operationally projected, retained, retired, or excluded. <!-- THE RECONCILIATION IS BUILT AND TEN OF THE FOURTEEN CAPABILITIES ARE NONE OF THE FIVE THINGS THE EVIDENCE CLAUSE ALLOWS; THE PARENT AND ALL FIFTEEN CHILDREN STAY [ ]. evidence/v1-workbenches/isis-3d-capability-reach.json — generate-isis-3d-capability-reach.mjs + .test.mjs (12 tests, 3 ratchets, 3 controls), wired into inventory + verify. The capability list is read from THIS document; .o is separated from the fourteen by its own wording ("Validate", not "Reconcile").

    THE PARENT'S FIVE VERBS ARE ALL ABOUT REACH, SO REACH IS WHAT IS
    MEASURED, at four altitudes that do not nest: BUILT (the package holds
    exported source and tests), IMPORTED (something outside it imports it),
    APP-REACHABLE (an application arrives, directly or through a chain), and
    PROJECTED (that arrival is an API route or a UI surface).
    
    ELEVEN OF FOURTEEN HAVE NO APP-REACHABLE PACKAGE AND THEREFORE NO
    PROJECTION: .a scene-from-image, .b part-level editing, .c semantic/text
    mesh editing, .d AI texturing, .e PBR relighting, .f mesh transformers,
    .g quad meshing, .i video-to-mesh, .j score distillation, .k advanced/
    diffusion-edited splats, .l gigascale scenes. These are not empty shells —
    `scene-from-image-composer` holds 107 tracked files and 40 test files,
    `video-to-mesh` 62, `ai-texturing` 494 export statements across 40 source
    files and 35 test files. They are built, exported, tested, documented in
    the architecture deep-dive, and no application can arrive.
    
    35 OF THE 79 ISIS PACKAGES ARE NOT REACHABLE FROM ANY APPLICATION, and 14
    of the 18 packages these capabilities are built from are among them. THAT
    NUMBER IS A FLOOR, NOT AN ESTIMATE: every non-Isis library importing an
    Isis package seeds the frontier without checking whether an app reaches
    that library, and dynamic `import()` and `require()` count as edges, so
    both choices can only ADD reach.
    
    A CONSUMER THAT NOTHING REACHES IS NOT REACH. Two capabilities have
    importers and still cannot be arrived at. `ai-texturing`'s only importers
    are two files in `3d-comfyui-nodes`, and `3d-comfyui-nodes` is imported by
    nothing at all. `3d-asset-library` (705 export statements) is imported
    only from `3d-semantic-editing`, itself unreferenced. One hop of apparent
    use, terminating in silence — which a one-hop consumer count reports as
    healthy.
    
    AND A SHIPPED PAGE WITH A SIMILAR NAME IS NOT THE CAPABILITY.
    `apps/yemaya/studio-web/src/pages/RelightStudioPage` imports
    `@isis/relight-studio`, a 2,365-line module of IC-Light/GenLit FRAME
    relighting — presets, EXR passes, NLE interchange — which is the image
    relighting I0.17.c asks about, not the PBR relighting of I0.16.e. Both
    were bound to this cell in a first pass here and separated on reading the
    source; the LumiGauss relightable-Gaussian library I0.16.e does name has
    zero consumers, and a test pins `relight-studio` out of every I0.16 cell
    so the conflation cannot come back.
    
    THE GUARD AGAINST THE AUDIT CITING ITSELF IS LIVE AND CURRENTLY DECIDES
    NOTHING, and both halves are published. `scripts/v1-workbenches/**` and
    `tools/conventions/**` import three Isis packages — entitlements,
    curated-cards, output-gallery — to measure them; all three also have
    product consumers, so excluding audit paths removes nobody's only caller
    today. It is recorded because the hazard is real the moment a package's
    product consumers go away, and an audit that counted itself would report
    exactly the wrong verdict at exactly that moment.
    
    NOT ONE ROW CARRIES AN APPROVED DISPOSITION, AND THE FIELD NAME IS WHY.
    The estate holds one disposition per package and the field is called
    `proposedDisposition`, so `reuse` is a proposed reuse exactly as
    `retire-candidate` is a proposed retirement — the value does not upgrade
    the field. I0.6 produced proposals and labelled them honestly. Measured
    from the field NAMES the package inventory carries, so this stops being
    true the moment an approved one appears. I0.16.o wants zero incomplete 3D
    rows and there are fourteen; the three that fail on the disposition ALONE
    — .h model fine-tuning, .m asset browsing/marketplace, .n universal
    rigging — are the ones an application can reach and project, so the estate
    does ship some of this. -->
    
    • I0.16.a Reconcile scene-from-image generation independently: source/ camera/depth/segmentation/scale assumptions, scene graph/geometry/ material outputs, uncertainty, preview, provenance/rights, validation, target formats, audience/host, and disposition. Evidence: scene-from- image capability row.
    • I0.16.b Reconcile part-level mesh editing independently: stable part/ topology/selection identity, local operation/mask/constraint, boundary/ UV/skin/material preservation, preview/diff/undo, validation, and disposition. Evidence: part-edit row.
    • I0.16.c Reconcile semantic/text mesh editing independently: typed edit intent/target/strength/constraints, ambiguous target clarification, geometry/topology/material/rig impact, variants, provenance, quality, and disposition. Evidence: semantic mesh-edit row.
    • I0.16.d Reconcile AI texturing independently: mesh/UV/material inputs, prompt/reference/rights, texture set/channels/resolution/UDIM/color/ tiling, seam consistency, preview, PBR gates, and disposition. Evidence: AI-texturing row.
    • I0.16.e Reconcile PBR relighting independently: geometry/material/ normals/environment/light rig/camera/color inputs, relit maps/renders, energy/material consistency, preview, and disposition. Evidence: PBR-relighting row.
    • I0.16.f Reconcile mesh-transformer capabilities independently: supported model/task/version/input/output/topology/scale limits, resource/cost, quality/compatibility/rights/provenance, preview, and disposition. Evidence: mesh-transformer row.
    • I0.16.g Reconcile quad meshing independently: source mesh/point/splat, target density/flow/symmetry/boundaries/features, UV/material/rig preservation, topology metrics, preview, downstream DCC, and disposition. Evidence: quad-meshing row.
    • I0.16.h Reconcile 3D model fine-tuning independently: dataset/license/ consent/provenance, architecture/base/version, training parameters/ resource/cost, checkpoints/evaluation, promotion/quarantine, and disposition. Evidence: model fine-tuning row.
    • I0.16.i Reconcile video-to-mesh independently: video/camera/calibration/ time/mask inputs, reconstruction scope, geometry/texture/animation/ confidence outputs, identity/biometric/rights, quality, and disposition. Evidence: video-to-mesh row.
    • I0.16.j Reconcile score-distillation capabilities independently: representation/camera/render/diffusion/model/version/objective/optimizer/ seeds/resource inputs, intermediate/final artifacts, reproducibility, safety/rights, evaluation, and disposition. Evidence: score-distillation row.
    • I0.16.k Reconcile advanced and diffusion-edited Gaussian splats independently: source/camera/splat/model/edit/mask inputs, coordinate/ SH/compression outputs, temporal/spatial/artifact/privacy gates, preview/interchange, and disposition. Evidence: advanced-splat row.
    • I0.16.l Reconcile gigascale scene capabilities independently: tiling/ streaming/LOD/spatial index/resolver/coordinate/precision/distributed build, resource/cost, preview, integrity/performance, export, and disposition. Evidence: gigascale scene row.
    • I0.16.m Reconcile asset browsing/marketplace independently: catalog/ search/filter/preview/license/price/entitlement/provenance/import/ dependency/update/revocation/attribution, provider trust, audience, and disposition. Evidence: asset-marketplace row.
    • I0.16.n Reconcile universal rigging independently: source mesh/ skeleton/archetype/landmarks, auto joints/weights/controls/retarget/ constraints, target rigs/formats, deformation metrics, preview, and disposition. Evidence: universal-rigging row.
    • I0.16.o Validate every 3D capability row names owner/package/consumer, audience/host, typed contract, provider/job/resource, persistence/ lineage, UI/API/operational projection, rights/safety/privacy, quality/ compatibility/interchange, tests, and one approved disposition. Evidence: zero incomplete 3D rows.
  • I0.17 Reconcile specialized image/video/audio capabilities: portrait and face synthesis, relighting, fast generation, AnimateDiff Lightning, Florence-2 analysis, live preview, camera/motion control, AV narrative, video editing/object removal, NVIDIA RTX/RIFE/Topaz enhancement, visual dubbing, foley, TTS, text-to-audio, voice cloning, music, and conversational audio. Evidence: each capability has a host, consent/right policy, typed controls, preview, lineage, gates, and release path or explicit exclusion. <!-- THE RECONCILIATION IS BUILT AND ONE ROW OF TWENTY-ONE CARRIES THE PARENT'S SEVEN COLUMNS; THE PARENT AND ALL TWENTY-TWO CHILDREN STAY [ ]. evidence/v1-workbenches/isis-media-capability-matrix.json — generate-isis-media-capability-matrix.mjs + .test.mjs (10 tests, 3 ratchets, 2 controls), wired into inventory + verify. The capability list is read from THIS document; .v is separated by its own wording.

    THE UNIT HAD TO BE A MODULE, NOT A PACKAGE. Ten of the twenty-one do not
    live wholly in a package of their own — AnimateDiff Lightning, Florence-2,
    live preview, RTX, RIFE, Topaz, TTS, text-to-audio and conversational
    audio are directories or single files inside `@isis/ai-providers` or
    `@isis/video-enhancement`, and music spans a package AND a providers
    directory. A package-level reachability answer would call each of them
    reachable the moment anything imported the parent barrel, so reach is read
    from IMPORTS AND THEIR BINDINGS: an application reaches a capability if it
    imports a path inside the module, or imports the module's package and
    NAMES a symbol the module exports. A star or default import of the parent
    names nothing and cannot decide either way, so it is recorded as
    indeterminate rather than counted or dropped; today there are none.
    Eleven of the twenty-one have no application import at all.
    
    A COLUMN IS A MEMBER, NOT AN OCCURRENCE. Exported interface, type-literal
    AND class bodies are brace-matched and their member names are what each
    column matches, with optionality recorded, because an optional consent
    field and a required one are different promises. Class members had to be
    read too: without them a class-based module measures as empty for a reason
    about its style rather than its contract, which is how TTS first read as
    having no typed controls at all.
    
    FIVE OF THE SEVEN BIOMETRIC CAPABILITIES DECLARE NO CONSENT MEMBER
    ANYWHERE IN THEIR TYPES OR CLASSES: .a portrait synthesis, .i AV
    narrative, .o visual dubbing, .q TTS, .u conversational audio. Visual
    dubbing is the loudest — 19 source files that put one person's speech in
    another person's face, 191 exported symbols, 776 type members, and the
    strings "consent", "rights", "licence", "watermark" and "disclosure" do
    not occur in the module at all. Conversational audio carries hundreds of
    licence occurrences for its voice marketplace and zero consent. TTS is the
    sharpest: `providers/tts` declares no object type of its own and its whole
    contract is a 1,403-line `ElevenLabsProvider` class exposing `cloneVoice`,
    `cloneVoiceInstant`, `cloneVoiceProfessional` and `designVoice` — 71 class
    members, none of them consent.
    
    AND WHERE CONSENT DOES EXIST IT IS A SEPARATE REQUEST NOBODY MAKES. Voice
    cloning has 606 type members, three of which mention consent, all on one
    `VoiceConsentRights*` family in one file; face synthesis has 824 members,
    two of them on a `FaceSynthesisEthicalSafeguardsRequest`. Neither is a
    field on the generation request, so consent is a check the caller may
    decline to run — and tracing the symbols those files export shows the ONLY
    non-test callers are the package's own barrel and, for face synthesis, the
    v1-workbenches script that measures it. THIS IS THE INVERSE OF THE
    ESTATE'S OWN BEST WORK: the I2.16 conditioning reference makes
    `consentRef`, `consentExpiresAt` and `consentScopes` non-nullable on the
    face-identity layer, so a layer without consent cannot be constructed.
    
    A RIGHTS POLICY IS THE WIDEST GAP: 16 of 21 declare no licence, rights,
    royalty, attribution or ownership member of any kind. Then lineage (11),
    preview (10), release path (10), gates (6), consent (5 of the 7 biometric)
    and typed controls (2).
    
    WHY EVERY CELL STAYS OPEN. Video object removal (.k) is the one row that
    carries all seven parent columns — it ships a C2PA credential type, a mask
    preview, a provenance package and a safety context. The artifact names
    that field `carriesAllParentColumns`, not `complete`, because each child
    asks for more than the parent's seven: .k also wants a protected-evidence
    and manipulation policy and a disposition, and the only disposition the
    estate records is the `proposedDisposition` I0.16 found. I0.17.v wants
    zero incomplete media rows and there are twenty. -->
    
    • I0.17.a Reconcile portrait synthesis independently: subject/context/ style/pose/wardrobe/background references, consent/biometric/public- figure policy, identity/anatomy/quality, disclosure/provenance, host, and release/disposition. Evidence: portrait row.
    • I0.17.b Reconcile face synthesis independently: identity/attribute/ expression/age/pose/lighting controls, detection/embedding, consent/ impersonation/child safety, artifact/identity metrics, watermark/ provenance, audience, and disposition. Evidence: face row.
    • I0.17.c Reconcile image relighting independently: image/depth/normal/ mask/environment/light/camera/color inputs, shadow/specular/material/ identity preservation, diagnostics, preview, and disposition. Evidence: relighting row.
    • I0.17.d Reconcile fast-generation paths independently: model/distillation/ quantization/step/scheduler/provider/version, quality/latency/cost trade- offs, unsupported controls, tier exposure, provenance, and disposition. Evidence: fast-generation row.
    • I0.17.e Reconcile AnimateDiff Lightning independently: base model/ motion module/version/workflow/frame/rate/duration/seed/conditioning, temporal quality/flicker, license/provenance, preview, and disposition. Evidence: AnimateDiff row.
    • I0.17.f Reconcile Florence-2 analysis independently: supported caption/ detection/grounding/OCR/segmentation tasks, model/version, input/output schema/confidence, privacy/safety, evaluator limits, host, and disposition. Evidence: Florence-2 row.
    • I0.17.g Reconcile live preview independently: progressive/intermediate/ low-res/stream protocol, cancellation/backpressure/cost, unsafe/unscanned content containment, ephemeral retention, accessible alternative, and audience. Evidence: live-preview row.
    • I0.17.h Reconcile camera/motion control independently: camera path/lens/ pose/depth/flow/vector/reference/time controls, model/workflow compatibility, diagnostics/preview, and disposition. Evidence: camera-motion row.
    • I0.17.i Reconcile audiovisual narrative generation independently: script/scene/shot/character/dialogue/camera/image/video/voice/music/SFX/ captions/timecode dependencies, continuity, rights/safety, bundles, and disposition. Evidence: AV-narrative row.
    • I0.17.j Reconcile video editing independently: trim/split/reorder/ extend/inpaint/replace/style/color/frame-rate/upscale/audio/caption operations, source preservation, timeline/preview, provenance, quality, and disposition. Evidence: video-editing row.
    • I0.17.k Reconcile video object removal independently: object/mask/track/ occlusion/background/motion inputs, temporal fill consistency, protected evidence/manipulation policy, preview/diff, and disposition. Evidence: object-removal row.
    • I0.17.l Reconcile NVIDIA RTX processing independently: supported SDK/ model/operation/GPU/driver/version, inputs/outputs, latency/quality/ resource/cost, host licensing, diagnostics, and disposition. Evidence: RTX row.
    • I0.17.m Reconcile RIFE interpolation independently: implementation/ model/version/input/output frame rates/time base/scene cuts/artifacts/ audio sync, quality/latency, preview, and disposition. Evidence: RIFE row.
    • I0.17.n Reconcile Topaz enhancement independently: product/model/version/ license/host/provider, upscale/denoise/sharpen/frame operations, parameters/outputs, quality/artifact/cost, provenance, and disposition. Evidence: Topaz row.
    • I0.17.o Reconcile visual dubbing independently: source video/audio/ transcript/translation/voice/face/timing, speaker identity/consent, lip-sync/translation/audio metrics, captions/disclosure/provenance, and disposition. Evidence: visual-dubbing row.
    • I0.17.p Reconcile foley generation independently: scene/action/material/ event/time/perspective/variation prompts, audiovisual sync, source/ training rights, audio technical/quality gates, package, and disposition. Evidence: foley row.
    • I0.17.q Reconcile TTS independently: text/SSML/language/voice/style/ pronunciation/timing/output controls, voice license/consent/disclosure, safety/quality/accessibility, and disposition. Evidence: TTS row.
    • I0.17.r Reconcile text-to-audio independently: prompt/category/duration/ scene/timing/variation/loop/sample/channel controls, safety/rights/ originality, technical quality, provenance, and disposition. Evidence: text-to-audio row.
    • I0.17.s Reconcile voice cloning independently: enrollment samples/ identity verification/consent/purpose/locale/quality, embedding/model/ retention/revocation, impersonation/disclosure/watermark, access, and disposition. Evidence: voice-cloning row.
    • I0.17.t Reconcile music generation independently: brief/genre/style/ artist-policy/key/tempo/meter/structure/instrument/duration/stems/lyrics/ seed controls, originality/rights/quality, provenance, and disposition. Evidence: music row.
    • I0.17.u Reconcile conversational audio independently: participants/ voices/roles/script/turns/interruptions/emotion/timing/room/SFX/music/ captions, voice consent/safety, streaming/cancel, and disposition. Evidence: conversational-audio row.
    • I0.17.v Validate every media capability for owner/package, audience/host, typed controls/contracts, provider/workflow/job, live preview/semantic alternative, lineage/provenance, consent/rights/safety/privacy, raw quality diagnostics/gates, bundles/publish/recall, tests, and approved release or exclusion. Evidence: zero incomplete media rows.
  • I0.18 Reconcile the existing operator-only parity monitors, feature flags, audit log, RunPod topology, model/workflow parity, API parity, entitlement, anomaly, chargeback, marketplace, monetization, resource recommendation, and collection surfaces. Evidence: route-by-route keep, merge, retire, and live-data disposition. <!-- THE RECONCILIATION IS BUILT AND NOT ONE OF THE FOURTEEN FAMILIES CARRIES A PAGE, AN API, A LIVE-DATA VERDICT AND A RECORDED DISPOSITION TOGETHER; THE PARENT AND ALL FIFTEEN CHILDREN STAY [ ]. evidence/v1-workbenches/isis-operator-surface-matrix.json — generate-isis-operator-surface-matrix.mjs + .test.mjs (9 tests, 4 ratchets, 1 control), wired into inventory + verify. It joins committed evidence rather than re-deriving it: the I0.1.a register (now four surfaces), the I0.3.b page walkthrough's live-data verdicts, and the I0.1.e disposition ledger.

    THIS CELL IS WHY THE REGISTER GREW. Chasing "which routes serve the
    operator-only surfaces" found `apps/oshun/admin/src/app/isis/*` — eight
    operator pages in none of the register's three surfaces, and the only UI
    that calls the 77 admin-isis-* modules. See the I0.1.a correction; the
    register, the effective-route matrix, the tier rows, the boundary cases,
    the disposition ledger and the I0.1.g baseline were all repaired first,
    because a per-family answer built on a register missing a host would have
    been confidently wrong.
    
    FIVE FAMILIES HAVE AN OPERATOR API AND NO OPERATOR PAGE AT ALL:
    .h entitlement, .i anomaly, .j chargeback, .l monetization, .m resource
    recommendation. THREE ARE THE MIRROR IMAGE — .e feature flags and .f the
    audit log are pages in the operator SPA with no admin-isis-* module behind
    them, and .h has neither, so the surface the cell names does not exist as
    an operator surface at all.
    
    A NEAR-MISS THAT FILLS A COLUMN IS WORSE THAN AN EMPTY ONE. Three keywords
    were removed after reading what the modules actually serve:
    `admin-isis-integration-events` validates outbound integration events and
    is not the mutation/event/job/provider/release audit trail .f names;
    `admin-isis-account-protection` is abuse defence, not the
    actor/tenant/plan/licence rights of .h; and `artifact-detection` is a
    false friend, since in this estate an artifact is a visual defect in a
    generated image and not an anomaly signal. Removing them moved three
    families from "has an API" to "has none", which is the true answer, and a
    test pins them out so the coverage they manufactured cannot come back.
    
    AND WHERE THERE IS A PAGE, THE WALKTHROUGH FOUND IT NOT CONFIGURED. Six
    families — .a generation-type parity, .b workflow parity, .c model parity,
    .e feature flags, .f audit log, .g RunPod topology — have pages whose
    EVERY live-data verdict is not-configured. `/jobs/generation-parity`
    reaches 0 of its 27 client operations, the single largest unserved page in
    the estate; `/feature-flags` 0 of 3; `/audit-logs` 0 of 2. Only .d
    provider API/SDK parity has any page the walkthrough called live.
    
    THE UNCLAIMED SET IS PUBLISHED, BECAUSE IT IS THE OTHER HALF OF THE
    ANSWER: 48 of the 77 admin-isis-* modules serve surfaces this cell does
    not name at all, along with eight operator SPA routes and five admin
    console pages. Every one is an operator capability with no seat in the
    fourteen families I0.18 enumerates.
    
    WHY EVERY CELL STAYS OPEN. I0.18.o wants zero undispositioned operator
    routes. The I0.1.e ledger labels the disposition cell `owed` for every
    route in every one of these families — 112 of 200 routes are knowingly
    undecided and now recorded as such in the I0.1.g baseline — so no family
    can carry the keep/merge/retire half of its own evidence clause. -->
    
    • I0.18.a Reconcile generation-type parity monitors: authoritative type/ capability sources, expected versus implemented provider/workflow/UI/ contract coverage, divergence/freshness, owner, and disposition. Evidence: generation-type route row.
    • I0.18.b Reconcile workflow parity monitors: registry/graph/template/ version/node/provider/host expected versus deployed coverage, mismatch/ drift/unsupported state, and disposition. Evidence: workflow parity row.
    • I0.18.c Reconcile model/model-version parity monitors: registry/artifact/ provider/endpoint/evaluation/promotion/host expected versus actual versions/states and disposition. Evidence: model parity row.
    • I0.18.d Reconcile provider API/SDK/UI parity monitors: operations/ request-response/error/capability/version/control exposure and exact missing producer/consumer surface. Evidence: API parity row.
    • I0.18.e Reconcile feature-flag surfaces: registry/runtime evaluation/ owner/purpose/audience/default/expiry/rollout/rollback/removal and authorized operator control. Evidence: feature-flag row.
    • I0.18.f Reconcile audit-log surfaces: mutation/event/job/provider/ release/emergency coverage, correlation/redaction/integrity/retention/ export/access and live source. Evidence: audit row.
    • I0.18.g Reconcile RunPod topology surfaces: accounts/projects/regions/ network volumes/templates/endpoints/pods/workers/GPUs/models/queues/ health/capacity/cost/credentials with safe live probes. Evidence: topology row.
    • I0.18.h Reconcile entitlement surfaces: actor/tenant/project/plan/license/ asset/workflow/model/provider/quota/territory/time rights, decision provenance, enforcement, and appeal. Evidence: entitlement row.
    • I0.18.i Reconcile anomaly surfaces: signal/model/rule/version/baseline/ confidence/threshold/category, affected provider/job/cost/quality/ security, triage/alert/ack/resolution, and false-positive handling. Evidence: anomaly row.
    • I0.18.j Reconcile chargeback surfaces: usage/rate-card/allocation/ estimate/reservation/actual/variance/invoice/project/tenant/provider/ resource attribution, correction, export, and access. Evidence: chargeback row.
    • I0.18.k Reconcile marketplace surfaces: provider/vendor/asset/model/ workflow/license/price/entitlement/trust/security/rights/provenance/ compatibility/install/update/revoke and disposition. Evidence: marketplace row.
    • I0.18.l Reconcile monetization surfaces: offer/plan/price/currency/tax/ entitlement/quota/metering/billing/refund/revenue share/territory/ disclosure and product-boundary approval. Evidence: monetization row.
    • I0.18.m Reconcile resource-recommendation surfaces: workload/model/ quality/latency/budget/capacity inputs, candidates, raw estimates/ confidence/reasons, policy/floors, reservation, actual comparison, and no auto-dispatch. Evidence: recommendation row.
    • I0.18.n Reconcile collection surfaces: collection/member/relation/order/ taxonomy/eligibility/audience/owner/lifecycle/search/bulk/export/ publisher and operator-only versus creator exposure. Evidence: collection row.
    • I0.18.o For each route assign canonical host/owner/audience, live API/ store/provider, fixture/degraded states, keep/refactor/merge/retire/ redirect decision, migration, tests, and evidence. Evidence: zero undispositioned operator routes.
  • I0.19 Reconcile every output/interchange format and package, including PNG, JPEG, WebP, TIFF, EXR, MP4, MOV, WAV, FLAC, OBJ, FBX, glTF/GLB, USD, Alembic, Gaussian-splat/SPZ, manifests, EDL/shot packages, DCC bundles, and engine bundles. Evidence: each supported format maps contract, metadata, preview, deterministic naming, integrity, provenance, quality, downstream ingest, loss profile, and authorized host; unsupported formats are explicit. <!-- THE RECONCILIATION IS BUILT AND NOT ONE OF THE NINETEEN FORMATS CARRIES ALL TEN COLUMNS; THE PARENT AND ALL TWENTY CHILDREN STAY [ ]. evidence/v1-workbenches/isis-output-format-matrix.json — generate-isis-output-format-matrix.mjs + .test.mjs (9 tests, 3 ratchets, 2 controls), wired into inventory + verify. The format list is read from THIS document.

    THE TEMPLATE COMES FROM THE ESTATE AND THE DERIVATION REFUSED THE RICHER
    GUESS. Reading the fullest family (Exr, 16 symbols) suggests a five-symbol
    spine — ArtifactRole, DependencyManifestReference, ManifestProvenance,
    ArtifactManifest, OutputPackage. Intersecting what the families ACTUALLY
    carry gives three: ArtifactRole and ManifestProvenance are not universal.
    The generator derives the spine rather than listing it, and a test pins
    that the derivation reacts to its input instead of returning a constant.
    
    SEVENTEEN OF NINETEEN FOLLOW IT, AND THE PART THEY SHARE IS GOOD. Every
    one points its `<Fmt>ArtifactManifest` at the shared
    `OutputArtifactManifest`, which carries `filename`,
    `checksumAlgorithm`/`checksum` and `provenance` — so deterministic naming,
    integrity and provenance are answered for all seventeen at once, from one
    contract, with a dependency-manifest reference back to the checklist item
    and ticket that produced them.
    
    TWO FORMATS HAVE NO OUTPUT PACKAGE AND THEY ARE NOT THE SAME FINDING.
    GAUSSIAN SPLAT / SPZ HAS NO SYMBOL AT ALL — no package, no manifest, no
    provenance — and the format is not hypothetical:
    `libs/isis/3d-generation/src/types/3d.types.ts` carries `'spz'` in its
    format union and `libs/isis/gaussian-splatting`'s optimizer writes and
    converts it. So a splat can be produced and has no deterministic name, no
    checksum and no provenance record. DCC BUNDLES ARE THE OTHER SHAPE: they
    have `DCCBundleFileEntry`, `DCCBundleManifest` and `DCCExportBundle`, so
    the template was started and stopped before the package. Neither is
    recorded as explicitly unsupported, which is what the parent asks for when
    a format is not supported.
    
    AND THREE COLUMNS ARE CARRIED BY NO FORMAT AT ALL: quality, downstream
    ingest, and authorized host. Preview is carried by exactly one — the
    engine bundle. A loss profile by four, and only because PNG, TIFF, EXR and
    FLAC name a compression mode that happens to imply one. The four members
    every package shares are `format`, `manifest`, `schemaVersion` and
    `sourceOutputClass`. THE MATCHER IS NOT SIMPLY FAILING: preview, loss
    profile and metadata are looked for the same way and all three come back
    present somewhere, which is what makes the three zeroes mean something.
    
    WHY EVERY CELL STAYS OPEN. Each child asks for its format's own technical
    limits AND the shared columns. Seventeen rows are missing quality,
    downstream ingest and authorized host; two are missing the contract
    itself. I0.19.t wants every supported format complete and every
    unsupported one explicit, and neither half holds. -->
    
    • I0.19.a Reconcile PNG support: color type/depth/profile/alpha/metadata/ animation applicability, limits, preview, deterministic encoding/name, hash/provenance, rights/quality, ingest, and loss profile. Evidence: PNG row.
    • I0.19.b Reconcile JPEG support: profile/quality/subsampling/color/EXIF/ orientation/metadata limits, preview, deterministic encoding/name, hash/provenance, rights/quality, ingest, and lossy behavior. Evidence: JPEG row.
    • I0.19.c Reconcile WebP support: lossless/lossy/quality/alpha/animation/ color/metadata limits, preview, deterministic encoding/name, integrity/ provenance, ingest, and loss profile. Evidence: WebP row.
    • I0.19.d Reconcile TIFF support: classic/BigTIFF, pages/layers, bit depth/ sample format/compression/color/alpha/metadata limits, preview, naming, integrity/provenance, ingest, and loss. Evidence: TIFF row.
    • I0.19.e Reconcile EXR support: version/compression/channels/layers/ half-float/full-float/color/metadata/tile-scanline/deep applicability, preview/tonemap, naming, integrity/provenance, ingest, and loss. Evidence: EXR row.
    • I0.19.f Reconcile MP4 support: container/codec/profile/level/bitrate/ resolution/frame rate/color/HDR/audio/subtitles/timecode/metadata, preview, deterministic transcode/name, integrity/provenance, ingest, and loss. Evidence: MP4 row.
    • I0.19.g Reconcile MOV support: QuickTime container/codecs/profiles/ ProRes/alpha/timecode/tracks/color/HDR/audio/subtitles/metadata, preview, deterministic transcode/name, integrity/provenance, ingest, and loss. Evidence: MOV row.
    • I0.19.h Reconcile WAV support: RIFF/RF64, PCM/float, sample rate/depth/ channels/layout/BWF/timecode/metadata, audition/waveform, deterministic encoding/name, integrity/provenance, ingest, and lossless behavior. Evidence: WAV row.
    • I0.19.i Reconcile FLAC support: version/level/sample rate/depth/channels/ compression/metadata/cues/pictures, audition/waveform, deterministic encoding/name, integrity/provenance, ingest, and lossless behavior. Evidence: FLAC row.
    • I0.19.j Reconcile OBJ support: vertices/normals/UVs/faces/groups/ materials/MTL/textures/units/axes/naming/encoding/limits, preview, integrity/provenance, ingest/round-trip, and known loss. Evidence: OBJ row.
    • I0.19.k Reconcile FBX support: supported versions/binary-text, units/ axes/hierarchy/geometry/materials/textures/skeleton/skin/animation/ cameras/lights/custom props, preview, deterministic constraints, integrity/provenance, ingest, and loss. Evidence: FBX row.
    • I0.19.l Reconcile glTF/GLB support: version/extensions/buffers/images/ geometry/materials/PBR/skins/animation/cameras/lights/units/axes/ compression, preview, validation, naming, integrity/provenance, ingest, and loss. Evidence: glTF/GLB row.
    • I0.19.m Reconcile USD support: profile/version/file formats/layers/ composition/variants/payloads/refs/resolver/schemas/units/axes/materials/ skeleton/animation, preview, usdchecker, deterministic packaging, integrity/provenance, target ingest, and loss. Evidence: USD row.
    • I0.19.n Reconcile Alembic support: version/archive type/units/axes/ hierarchy/geometry/topology/UV/normals/attributes/cameras/animation/time sampling/compression, preview, integrity/provenance, ingest, and known loss. Evidence: Alembic row.
    • I0.19.o Reconcile Gaussian-splat/SPZ support: version/coordinate/ positions/scales/rotations/opacity/SH/cameras/compression/count/limits/ privacy, spatial preview, validation, naming, integrity/provenance, target ingest, and loss. Evidence: splat/SPZ row.
    • I0.19.p Reconcile manifest support: schema/version/profile/IDs/refs/ hashes/media/metadata/dependencies/rights/consent/gates/provenance/ signatures/compatibility, deterministic serialization/name, validation, import, and losslessness. Evidence: manifest row.
    • I0.19.q Reconcile EDL/shot packages: EDL/XML/AAF/OTIO or supported profile, sequence/shot/clip IDs, source refs, timecode/rate, handles/ transitions/tracks/audio/metadata/assets/manifests, validation, round-trip, and known loss. Evidence: editorial-package row.
    • I0.19.r Reconcile DCC bundles: target app/version/plugins, project/ scene/assets/refs/resolver/cache/textures/fonts/scripts policy, manifest/lock/SBOM/provenance, deterministic layout, sandbox validation, import/round-trip, and loss. Evidence: DCC-bundle row.
    • I0.19.s Reconcile engine bundles: target engine/version/platform/ renderer/plugins, assets/import settings/metadata/LODs/collision/ materials/rig/animation/audio/code policy, manifest/dependencies/ provenance, deterministic layout, validation, ingest, and loss. Evidence: engine-bundle row.
    • I0.19.t For every format/profile assign supported import/export/ preview/publish direction, asset kinds/audiences/hosts, authoritative contract/schema/validator, limits/security, naming/integrity/provenance/ quality, dependency packaging, round-trip/loss policy, version/ deprecation, and explicit unsupported behavior. Evidence: zero incomplete format rows.
  • I0.20 Crosswalk Phase I against ISIS_CHROMA_RUNPOD_MVP_TODOS_2026-09-11.md before any I6–I9 cell is started. Added 2026-09-18 by the board specification audit: that tracker has built, since 2026-09-11 and with no reference to this ledger, the RunPod endpoints, spend governor, model library and cache, Civitai intake and licence register, job/output pages in apps/isis/web, and a 3D workspace plan (T.15) — the same ground as I6.6, I7.1, I7.3, I7.5, I7.13, I7.18 and I7.19 here. Evidence: a committed crosswalk table under evidence/v1-workbenches/isis/ with one row per I6–I9 parent, in the merged / superseded / independent / split vocabulary of Isis ADR-0008.

    • I0.20.a List every checked item of the Isis tracker's C, V, A, L, E, H and T sections that ships code in apps/isis/web, apps/isis/generation-api or libs/isis/*, with the path it shipped. Evidence: shipped-surface rows keyed by tracker id.
    • I0.20.b For each I6–I9 parent, record whether an Isis-tracker item already delivers it in whole (superseded), in part (split, naming the remainder), under a different design (merged, naming which design wins and why), or not at all (independent). Evidence: zero unmapped I6–I9 parents.
    • I0.20.c Re-run the I0.3 page map over the apps/isis/web pages added after the BASE.1 inventory SHA, so keep/refactor/merge/retire covers them. Evidence: I0.3 rows cite the refreshed inventory SHA.
    • I0.20.d Record the reverse direction in the Isis tracker: every open T, E and F item that adds a page or an API to those paths names the I-cell it satisfies or states that none applies. Evidence: the Isis tracker's open dashboard items (T.15, T.22.06, E.01.02, E.02.04, F.20.03) carry the reference.

I1 — Canonical generation-control model#

  • I1.1 Verify the canonical workflow-template state machine against current contracts. Evidence: code-derived state diagram and known-value tests.

    • I1.1.a Inventory workflow/template status enums, transition commands, guards, APIs, persistence, events/jobs, operator UI, and tests from current code/contracts. Evidence: source mapping. <!-- isis-workflow-state-machine.json: source mapping of the CURRENT Isis workflow-template state machine (realized in apps/isis/workflow-registry/src/ services/workflow.service.ts). Status enums: WorkflowStatus (DRAFT/PUBLISHED/ DEPRECATED/ARCHIVED) + WorkflowVisibility (4); persistence: Workflow.status + WorkflowVersion.version/deprecated. Transition commands: 8/12 intended transitions have a code signal — the CORE lifecycle (create, edit=update, version=createVersion, activate, deprecate, archive, restore, publish) is implemented; the review sub-flow (submit/review/approve/revoke) is honestly reported ABSENT (workflow-templates have no submit→review→approve path). Guards: auth preHandlers present (routes/rbac.ts requireAuth+permission). 62 workflow tests, 16 operator-UI components, 7 event sources. Verb-synonym detection (edit→update, version→createVersion) + the git-pathspec both-globs fix (routes/.ts AND routes/**/.ts) avoid false negatives. generate-isis-workflow-state-machine.mjs + 6 tests, wired into inventory/ verify + CI. I1.1.b (diagram), I1.1.c (lifecycle reconciliation), I1.1.d-f (transition/guard/immutability verification with negative tests) build on this mapping and stay unchecked. -->

    • I1.1.b Generate the implemented state/transition/guard/actor/side-effect diagram from authoritative definitions. Evidence: code-derived diagram. <!-- evidence: isis-workflow-state-diagram.json .diagramMermaid — code-derived state diagram (DRAFT→PUBLISHED→DEPRECATED→ARCHIVED w/ create/edit/version/ publish/activate/deprecate/archive/restore) built on the I1.1.a census. generate-isis-workflow-state-diagram.mjs (--check gate), 9 tests. -->

    • I1.1.c Reconcile draft/review/approved/active/deprecated/revoked/ archived and any current states with the intended lifecycle. Evidence: state disposition table. <!-- evidence: .dispositionTable — 7 canonical states reconciled: 4 present (draft=DRAFT, active=PUBLISHED, deprecated=DEPRECATED, archived=ARCHIVED), 3 ABSENT (review/approved/revoked — workflow-templates have no review sub-flow, the I1.1.a finding). -->

    • I1.1.d Verify create/edit/submit/review/approve/activate/deprecate/revoke/ archive/restore/version transitions and exact authorization/guards. Evidence: transition matrix. <!-- evidence: .transitionMatrix — 12 canonical transitions, 8 implemented with an auth-preHandler guard (requireAuth+permission), 4 absent (submit/review/ approve/revoke). Verb-synonyms resolved (edit=update(), version=createVersion()). -->

    • I1.1.e Test invalid source/target, missing evidence, stale revision, incompatible dependency, revoked model/node, in-use template, and unauthorized/system actor. Evidence: known-value negative tests. <!-- SUPERSEDED (2026-08-16). The old note said this was "runtime: writing known-value NEGATIVE tests that execute against the running workflow service". IT IS NOT RUNTIME, in the same way I0.1.c was not. WorkflowServiceImpl picks an in-memory backend from ISIS_WORKFLOW_REGISTRY_STORAGE, so the REAL service — the same class apps/isis/workflow-registry mounts — runs in-process with no database and no HTTP, and the negative tests execute against it.

      evidence/v1-workbenches/isis-workflow-transition-probe.json —
      generate-isis-workflow-transition-probe.mts + .test.mts (6 tests, 2
      ratchets, 1 control), wired into inventory + verify. The subject is
      created with `createFromTemplate` from the service's own seeded
      template, so the definition under test is the estate's.
      
      EVERY ONE OF THE TWELVE ORDERED STATE PAIRS IS ACCEPTED. draft →
      archived, archived → published, published → draft, draft → deprecated:
      the workflow-template state machine has NO transition guard at all. The
      terminal state is not terminal, and a template can be deprecated without
      ever having been published.
      
      THE GUARDS THAT DO EXIST ARE REAL, WHICH IS WHAT MAKES THE ABSENCE MEAN
      SOMETHING. A stale `expectedUpdatedAt` is refused with
      `WorkflowUpdateConflictError`; a `createVersion` at or below the current
      semver is refused; an empty definition is refused by
      `assertDefinitionValid`; a soft-deleted workflow cannot be updated. The
      write path is capable of refusing — it simply has nothing to say about
      the state machine.
      
      A NULL RETURN IS A REFUSAL, NOT AN ACCEPTANCE. The first cut read "it
      did not throw" as "the estate performed it", which put a false
      acceptance in the artifact for the soft-deleted case: `update` returns
      `null` there because `get()` filtered the row. The refusal MODE is now
      carried beside the outcome so a throw and a null stay distinguishable.
      
      THE CELL STAYS `[ ]` ON COVERAGE, NOT ON METHOD. Three of the seven
      cases it names are tested with known values — invalid source/target,
      stale revision, unauthorized actor. FOUR ARE NOT EXPRESSIBLE AT THIS
      LAYER and are recorded as such with reasons rather than as passes: the
      workflow-template model carries no evidence record to be missing; the
      definition names custom nodes and models as strings the service resolves
      against no registry; there is no link to a model or node lifecycle; and
      nothing records that a workflow came from a template, so "in use" is not
      a state the service holds. Whether another layer can express them is the
      open question. -->
      
    • I1.1.f Verify immutable version/history/audit/event/active-run/release effects and no direct status write bypass. Evidence: integration tests. <!-- THE BYPASS HALF IS ANSWERED AND THE ANSWER IS THAT THE BYPASS IS THE ONLY PATH (2026-08-16, same probe as I1.1.e). There is no publishWorkflow, activateWorkflow, deprecateWorkflow or archiveWorkflow method on the workflow at all: update takes status from the request and writes it, and UpdateWorkflowRequestSchema validates that the value is one of the four states and nothing about the move. The named methods the I1.1.a census counted — activateVersion, deprecateVersion, restoreVersion — belong to WorkflowVersion, a different entity from the WorkflowStatus the I1.1.b diagram draws, so the census's verb synonyms attributed a version-level transition to the template.

      The immutability half — version/history/audit/event/active-run/release
      effects — is not covered by this probe and stays open. -->
      <!-- runtime: integration tests require the running service + DB. -->
      
    • I1.1.g Resolve contract/code/UI drift and fail on future state/transition additions without tests/diagram update. Evidence: drift gate. <!-- evidence: .driftGate — the generator reads the I1.1.a census, so a new WorkflowStatus value or implemented transition flows through and breaks --check + the coherence assertions (every implemented transition must have a matrix row; every code state must map to a canonical state) until the diagram is updated. -->

  • I1.2 Verify the canonical model and model-version state machines. Evidence: invalid promotion, deprecation, and archive paths fail.

    • I1.2.a Inventory model and model-version states/transitions/guards/ contracts/stores/events/jobs/UI/tests independently. Evidence: source mapping. <!-- isis-model-state-machine.json: source mapping of the CURRENT Isis model/ model-version state machine from apps/isis/generation-api/src/services/ model.service.ts (read ONCE for perf — the workflow gen's per-transition git greps timed out verify). 26 service operations classified: create/edit/delete, intake-upload (generateUploadUrl/completeUpload/cleanupAbandonedUploads), verification (verifyUploadedContent/scanUploadedContentForMalware), delivery, audit/telemetry, ops/health. VERIFICATION FINDING: the model lifecycle is WEAKLY formalized vs the workflow machine — NO typed ModelStatus enum (contrast WorkflowStatus), the intended states (draft/review/approved/active/deprecated/ revoked/archived) appear only as SCATTERED string literals across 28 files with no canonical definition, and status transitions ride on update() (no explicit approve/deprecate/revoke methods). 7 ModelType + 8 ModelFormat contracts, 39 model UI components, guards/events/stores present, but only 2 model-service test files (thin coverage). generate-isis-model-state-machine.mjs + 6 tests, wired into inventory/verify + CI. I1.2.b (identity/version boundary) and I1.2.c (lifecycle verification) build on this and stay unchecked. -->

    • I1.2.b Define model identity/ownership/license/risk/supported task/ lifecycle separately from immutable version/checkpoint/artifact/ evaluation/provider compatibility. Evidence: boundary review. <!-- evidence: isis-model-identity-boundary.json .boundary — 9 dimensions partitioned: MODEL IDENTITY (mutable) = ownership/license/risk/supported-task/ lifecycle vs IMMUTABLE VERSION (frozen) = checkpoint/artifact/evaluation/ provider-compatibility, each modeled dim mapped to real ModelCard/ModelRegistry/ ModelVersion fields (validated = drift gate). HONEST GAPS: license/risk/provider- compatibility NOT modeled today; lifecycle weak (no ModelStatus enum, I1.2.a). generate-isis-model-identity-boundary.mjs (--check gate), 9 tests. I1.2.c-g need executable tests (runtime) → parent I1.2 stays unchecked. -->

    • I1.2.c Verify model draft/review/approved/active/deprecated/revoked/ archived lifecycle and model-level default/current version semantics. Evidence: model state tests. <!-- THE WRITE-PATH CENSUS IS BUILT AND IT ANSWERS PART OF .c, .e AND .g; ALL THREE STAY [ ]. evidence/v1-workbenches/isis-model-writepath.json — generate-isis-model-writepath.mjs + .test.mjs (8 tests, 3 ratchets, 2 controls), wired into inventory + verify. It asks, for each of the 27 columns CREATE TABLE model_registry declares, WHO CAN WRITE IT: the INSERT column list, the if (updates.<field> !== undefined) SET guards, and every literal UPDATE model_registry SET … in the module — then crosses that with who READS it, who FILTERS or SORTS on it, and whether determineModelReleaseState requires it.

      `is_verified` HAS A READER, A FILTER, AN API PARAMETER, A GATE AND NO
      WRITER. Declared `BOOLEAN NOT NULL DEFAULT false`; absent from the
      INSERT column list so `create` cannot set it; no `updates.isVerified`
      guard so `update` cannot set it; no direct statement writes it. It is
      read into `ModelRegistryEntry.isVerified`, filtered on through
      `verifiedOnly` which `model.schema.ts` accepts as a coerced boolean, and
      required by `determineModelReleaseState` for `eligible`. SO
      `GET /models?verifiedOnly=true` RETURNS NOTHING, PERMANENTLY, and every
      model created through this service lands on `review` and never on
      `eligible`.
      
      `star_count` IS THE SAME SHAPE ONE STEP QUIETER: declared `INTEGER NOT
      NULL DEFAULT 0`, read into `starCount`, offered as a `sortBy` value the
      list schema accepts, and written by nothing.
      
      THE CONTROL THAT KEEPS THIS HONEST. A first cut counted only the guarded
      `update()` clauses and reported `usage_count`, `download_count` and
      `star_count` as unwritable. Two of the three ARE written — by dedicated
      increment statements — and the soft delete writes `deleted_at` and
      `is_active`. A census that misses a writer manufactures exactly the
      finding it exists to detect, so the scan now reads every literal UPDATE
      and a test pins that it sees those writers.
      
      WHY THESE CELLS STAY OPEN. .c asks for model STATE tests and there is no
      typed status enum to test (the I1.2.a finding); this census establishes
      what the fields can do, not that a lifecycle behaves. .e asks to prevent
      promotion without evidence — the evidence flag has no writer, so the
      negative test cannot even be set up on the positive side. .g asks about
      immutable artifacts and lineage as well as the bypass; only the bypass
      half is answered here, and the answer is that `is_active` rides the
      generic `update()` with no transition command. The model service is
      Postgres-only, so unlike I1.1.e there is no in-memory backend to execute
      the transitions against. -->
      
    • I1.2.d Verify version intake/quarantine/scan/evaluate/approve/promote/ active/deprecate/revoke/archive transitions and actor/guard requirements. Evidence: model-version state tests. <!-- SIX OF THE TEN STAGES HAVE NO COMMAND, NOT ONE OF THE FOUR THAT EXIST REQUIRES AN ACTOR, AND THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-model-version-lifecycle.json —
      generate-isis-model-version-lifecycle.mts + .test.mts (16 tests, 5
      ratchets, 5 controls), `--check` and the tests wired into verify. The
      ten stages are parsed out of THIS cell's own slash-list, so rewording
      the cell stops the build rather than quietly leaving a stage unmeasured.
      
      THIS ONE IS EXECUTED, WHICH I1.2.c COULD NOT BE. The note there said the
      model service is Postgres-only and so had no backend to run against; the
      local dev stack has a Postgres AND a MinIO, so the probe drives the real
      `ModelService` — the class `apps/isis/generation-api` mounts — with its
      real S3 signing client, its real hash verification and its real malware
      scanner, and records what each stage did. 16 cases, 10 findings, zero
      refusals classified `infrastructure` (see the control below). A version
      is a `model_registry` row keyed `(name, version)` and `getVersions(name)`
      selects the rows sharing a name, so the version lifecycle IS the row
      lifecycle and there is no second table to look for.
      
      THE SCAN RUNS, REPORTS `clean=false`, AND THE INTAKE TAKES THE FILE
      ANYWAY. `completeUpload` blocks on `threats.filter(t => t.type ===
      'malware')`, and `DEFAULT_MODEL_MALWARE_SCAN_ENGINE` is `'builtin'`.
      `BuiltinSecurityScanner` can emit fourteen threat types — api_key,
      command_injection, credential, encoded_payload, obfuscated_code,
      path_traversal, pii, private_key, sensitive_data, sql_injection, ssrf,
      suspicious_pattern, xss, xxe — and `malware` is not one of them; the
      only `type: 'malware'` in `scanner.ts` is inside `ClamAVScanner`.
      Executed: bytes the engine flagged came back `status="completed"
      clean=false` with 7 threats, 0 of them typed `malware`, and the upload
      completed with `malwareScan.status: 'clean'` written into the audit
      metadata. THE GATE IS NOT MISCONFIGURED, IT IS UNCLOSEABLE ON THE
      DEFAULT ENGINE, and `result.clean` — the scanner's own verdict — is
      never read.
      
      THE VERIFIED BYTES ARE NOT THE ISSUED BYTES. `POST /models/:id/upload-
      complete` takes `storagePath` from the request body; the zod schema
      checks only that it is a 1–500 character string, and nothing compares it
      to the `models/<id>/<file>` prefix `generateUploadUrl` minted. Every
      check that follows — HEAD size, download, sha256, scan — is then
      performed on whatever object the caller named, so they all pass.
      Executed: version 2.0.0's intake, naming version 1.0.0's object, was
      accepted and 2.0.0's row now points at 1.0.0's artifact. The same call
      on an already-completed version is accepted too, so the artifact binding
      is never sealed.
      
      NO STAGE REQUIRES AN ACTOR. `register(uploaderId)` is the only mutation
      whose actor has no default, and it accepts the empty string; `update`,
      `delete` and `completeUpload` all default to `'system'`, and every
      mutating route degrades an absent caller to the literal `'anonymous'`.
      Executed: a `completeUpload` with no actor recorded
      `model.upload.completed` by `system`. The 401 is real but it is in the
      chain, not the handler — `authMiddleware` on `/api/*`, mounted only when
      the app config carries `auth`, with `allowAnonymous` true under
      `NODE_ENV=development` — and `POST /models` is the one mutating route
      with no ownership comparison at all.
      
      THE SIX STAGES WITH NO COMMAND ARE quarantine, evaluate, approve,
      promote, deprecate AND revoke. Not weakly modelled — absent. Executed
      consequences: a version whose intake was refused stays live, active and
      unmarked with the rejected object still in the bucket (no quarantine);
      `update({isVerified:true})` is ACCEPTED and the flag reads back false,
      because `ModelUpdateRequestSchema` has no such field and the update is
      silently dropped (no approve); a hash-verified, malware-scanned version
      reaches release state `review` and stops there (no promote). The
      positive control separates the two possible explanations: with
      `is_verified` set by direct SQL the same predicate returns `eligible`,
      so the predicate is sound and its input has no writer — which is exactly
      what the I1.2.c census found from source and this run demonstrates at
      runtime.
      
      THE TWO STAGES THAT DO EXIST DO NOT EXIST AS COMMANDS. `active` rides
      the generic `update`, recording the audit action `model.updated`, so
      nothing distinguishes a deactivation from a description edit. `archive`
      is `delete()`, a soft delete setting `deleted_at` and `is_active`; the
      row survives and drops out of `get` and `getVersions`. And no column
      records which version is current: `getVersions` derives it as the first
      active row of an `ORDER BY created_at DESC`, so with 1.0.0 and 2.0.0
      both live it named 2.0.0, and after deactivating 2.0.0 it named 1.0.0.
      DEACTIVATION IS THE ONLY LEVER ON "CURRENT" — promoting an older version
      is not expressible.
      
      WHAT KEEPS THE PASSES HONEST. Four cases came back refused across three
      classes — `(name, version)` uniqueness, hash mismatch, size mismatch —
      and a BROKEN OBJECT STORE WOULD HAVE PRODUCED THE SAME REFUSALS. Every
      throw is therefore classified by message into a named refusal class, and
      the run aborts if any lands on `infrastructure`; all four classified as
      estate refusals. Two matchers were also narrowed after they lied: route
      handlers are bounded by the NEXT handler of any verb, because slicing
      each POST to the next POST swallowed the GET handlers between them and
      credited `POST /models` with an ownership check that belongs to
      `GET /models/:id`; and the scanner vocabulary is read per class, because
      reading the file whole puts ClamAV's `malware` in the builtin engine's
      list and silently turns the central finding into a pass.
      
      NOT WIRED INTO `inventory:v1-workbenches`. The write path needs a
      Postgres and an object store, and refuses outright without them rather
      than replacing a real recorded run with an empty one; `--check` without
      them verifies the static half and says so. Reproduce:
      `ISIS_MODEL_PROBE_DATABASE_URL=… ISIS_MODEL_PROBE_S3_ENDPOINT=…
      ISIS_MODEL_PROBE_S3_BUCKET=… ISIS_MODEL_PROBE_S3_ACCESS_KEY=…
      ISIS_MODEL_PROBE_S3_SECRET_KEY=… pnpm exec tsx
      scripts/v1-workbenches/generate-isis-model-version-lifecycle.mts`. -->
      
    • I1.2.e Prevent promotion without hash/signature/source/license/security/ safety/quality/compatibility evidence or with stale/missing results. Evidence: promotion negative tests. <!-- ONE OF THE EIGHT EVIDENCE KINDS IS GATED AND THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-model-promotion-evidence.json —
      generate-isis-model-promotion-evidence.mts + .test.mts (13 tests, 5
      ratchets, 4 controls), wired into inventory + verify. The eight kinds
      are parsed out of THIS cell's slash-list and the parser also requires
      the "stale/missing results" clause, so half an answer to half a question
      cannot read as a whole one.
      
      A NEGATIVE TEST NEEDS SOMETHING TO WITHHOLD FROM A DECISION THAT
      CONSULTS IT. There are two promotion decisions, both exported pure
      functions, so the negatives run against the real implementations with
      the estate's own builders producing the subjects. `determineModelRelease
      State` reads exactly three fields — `hash`, `isActive`, `isVerified` —
      and `buildIsisReleaseReadiness` folds the result in as one requirement,
      `model_not_verified`. Of the eight kinds the cell names, ONLY `hash` is
      an input to either. Executed: no hash → `review`, unverified → `review`,
      inactive → `blocked`, everything present → `eligible`, and a named model
      in review makes the readiness gate `block`. Those four passes are what
      make the rest mean something.
      
      THE ESTATE HAS TWO MODEL-VERSION RECORDS AND THE GATE READS THE THIN
      ONE. `model_registry` — raw SQL, written by `apps/isis/generation-api` —
      carries a hash and one opaque boolean. `model_versions` and
      `model_cards` are Prisma models carrying `licenses`, `hashes`, `safety`,
      `ethics`, `governance`, `evaluation`, `telemetry`, `lineage` and
      `source`: exactly the evidence this cell asks for. NOTHING OUTSIDE THE
      GENERATED CLIENT READS OR WRITES THEM. The seven files that mention the
      names are six generated modules and one test asserting the migration
      adds the tables, each excluded by name and reason rather than silently.
      So license, safety and quality are declared on a record no decision can
      reach; source (`base_model`, `trained_on`) and compatibility
      (`compatible_engines`) are carried on the gated record and read by no
      promotion decision; signature is not modelled anywhere; and security is
      AUDIT-TRAIL-ONLY — the malware scan runs at intake and its summary is
      written into the audit-log `metadata` JSON, where no column carries it
      and no decision can query it.
      
      THE PREDICATE'S OTHER TWO INPUTS NAME NO EVIDENCE KIND AT ALL.
      `isActive` and `isVerified` match none of the cell's eight words, and
      `isVerified` is the opaque boolean standing in for all of them — the one
      the I1.2.c write-path census found has a reader, a filter, an API
      parameter, a gate and NO WRITER. Together those two facts say the whole
      thing: seven kinds of evidence cannot reach the decision, and the eighth
      channel that could carry them cannot be written.
      
      STALENESS IS NOT A STATE THIS ESTATE CAN BE IN. Neither decision takes a
      timestamp. Executed: a record last touched in 2019 returns `eligible`,
      the same as a fresh one. Nor could a freshness bound be bolted on to
      what exists — `model_registry` has only `created_at` and `updated_at`,
      and `update()` sets `updated_at = NOW()` on any edit including a
      description change, so the row's only clock is not evidence-scoped.
      `model_versions`, which could carry per-evidence timestamps, is the
      record nothing reaches.
      
      AND THE MODEL REQUIREMENT PASSES WHEN NO MODEL IS NAMED.
      `const modelVerified = input.model ? input.model.releaseState ===
      'eligible' : true` — absence is permission. `modelId` is optional on
      `IsisControlledGenerationRequest` and `resolveGenerationContext` resolves
      it to `null` when omitted, so the omission path is reachable from the
      public request type. Executed: with no model the gate reports
      `modelVerified=true` and an empty `missingRequirements`. Combined with
      the unwritable `is_verified`, THE MODEL GATE IS UNPASSABLE WHEN A MODEL
      IS NAMED AND VACUOUS WHEN ONE IS NOT.
      
      Compatibility has one more executed wrinkle: a version with no declared
      compatible engine reaches `eligible` carrying the warning "Model has no
      declared compatible engines." A WARNING IS NOT A REFUSAL.
      
      CONTROL THAT MATTERED. An unanchored `model_versions` matches
      `athena_cad_model_versions` — a different table in a different domain —
      and would have reported a consumer for the record whose central finding
      is that it has none. The sweep is a fast `git grep -F` narrowed by
      word-anchored patterns in-process, and a test pins both that the
      collision source still exists and that the anchored pattern rejects
      it. -->
      
    • I1.2.f Prevent deprecation/archive/revoke that strands active workflows/ endpoints/runs/releases without replacement/impact/recall/migration. Evidence: dependency guard tests. <!-- A DEPENDENCY GUARD NEEDS A DEPENDENCY EDGE, AND ALL FOUR EDGES ARE STRINGS NOTHING RESOLVES. THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-model-retirement-guard.json —
      generate-isis-model-retirement-guard.mts + .test.mts (13 tests, 5
      ratchets, 5 controls), `--check` and the tests wired into verify. All
      three of the cell's lists — the transitions, the dependents and the
      remedies — are parsed out of the cell separately, and a cell that stops
      naming any one of them stops the build.
      
      TWO OF THE THREE RETIRING TRANSITIONS HAVE NO COMMAND. There is no
      `deprecate` and no `revoke` on `ModelService`; `archive` resolves to the
      soft `delete()`, whose entire body is `UPDATE model_registry SET
      deleted_at = NOW(), is_active = false WHERE id = $1 AND deleted_at IS
      NULL`. It reads no dependent table, so the count of transitions that
      consult a dependent is 0 of 3 — and two of those three could not consult
      anything because they do not exist.
      
      THE FOUR DEPENDENTS HOLD FOUR DIFFERENT SHAPES OF STRING AND NOT ONE IS
      A FOREIGN KEY. A staging recipe pins models as `model://name/variant@
      version` URIs in a JSON record (`modelAliases`); an endpoint declares
      model CLASSES (`modelClasses`), validated only for uniqueness; a run
      carries `jobs.input_model TEXT`, whatever the submitter passed; and a
      release carries `outputs.model JSONB`, a snapshot of the name, version
      and hash taken at output time. Two of the four cannot name a version at
      all. NO CODE RESOLVES ANY OF THEM back to a `model_registry` row, and no
      SQL statement anywhere in the repository names `model_registry`
      alongside `jobs`, `outputs` or a workflow table — though all of them
      share one `ISIS_DATABASE_URL`, so the join is physically available and
      simply never written.
      
      NONE OF THE FOUR REMEDIES IS A CONCEPT THE MODEL SURFACE CARRIES. No
      replacement or `replacedBy` pointer, no impact or blast-radius record,
      no recall or rollback command, no migration plan — across the service,
      its routes and its request schemas, with comments stripped so a header
      naming a remedy cannot stand in for one.
      
      EXECUTED, AGAINST THE REAL SERVICES. A run was submitted through the
      real `JobService` naming a real model registered through the real
      `ModelService`, and the model was then retired. `delete()` was accepted
      with a `pending` run naming the model; the run row still names it; and
      because a second version shares the name, the reference now denotes
      version 2.0.0 while the run was submitted against 1.0.0. The retirement
      audit entry is `model.deleted` with NO metadata and a `beforeState` of
      the model row alone — nothing anywhere records what the retirement
      affected. THE POSITIVE CONTROL SEPARATES ABSENCE FROM INCAPACITY: a
      second `delete()` of the same version WAS refused, so the write path can
      say no; it just has nothing to say about dependents.
      
      TWO MATCHERS LIED BEFORE THEY WERE FIXED, AND BOTH WOULD HAVE SOFTENED
      THE FINDING. Searching for the releases reference by its column name
      `model` returned seventy-odd files and reported the reference as
      resolved; the tokens are a conjunction now (`row.model` AND `outputs`),
      with a positive control proving the search still finds a real resolver.
      And the SQL-join sweep matched a backtick literal spanning two unrelated
      templates in `calliope-operations.ts`, where `model_registry` is the
      name of a SECRET PROVIDER, not a table; a statement must now also carry
      a SQL verb and stay under a length no real query exceeds.
      
      THE PER-REFERENCE MEASUREMENT IS A FLOOR and says so in the artifact: a
      file resolving one of these references under a different alias would not
      be counted. The join sweep is independent of aliasing and agrees.
      
      OUT OF CELL, FOUND WHILE BUILDING THIS. `JobService.initialize()` throws
      on the default configuration: `ISIS_QUEUE_PRODUCER_ENABLED` defaults to
      true and `DEFAULT_QUEUE_NAME` is `'isis:jobs:submit'`, which BullMQ
      rejects outright with "Queue name cannot contain :". Production is not
      affected — `resolveQueueProducerConfig` requires `ISIS_QUEUE_NAME` there
      — but every non-production boot that does not set it fails to start the
      job service. The probe sets `ISIS_QUEUE_PRODUCER_ENABLED=false` because
      the queue is not this cell's subject; the defect belongs to the job
      dispatch cells and is recorded here so it is not lost. -->
      
    • I1.2.g Verify immutable version artifacts/lineage/audit/events and no direct status or current-version bypass. Evidence: real integration tests. <!-- ONE OF THE FOUR SUBJECTS IS IMMUTABLE, NEITHER BYPASS IS GUARDED, AND THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-model-immutability-ledger.json —
      generate-isis-model-immutability-ledger.mts + .test.mts (12 tests, 5
      ratchets, 3 controls), `--check` and the tests wired into verify. The
      four subjects and both bypasses are parsed out of the cell.
      
      ABSENCE IS NOT IMMUTABILITY, AND THE LEDGER REFUSES TO MERGE THEM. A
      thing with no carrier cannot be mutated; scoring that as immutable would
      turn a missing record into a pass. Each subject is `immutable`,
      `mutable` or `absent`. ARTIFACTS: mutable. LINEAGE: mutable. AUDIT:
      immutable — the one real pass. EVENTS: absent.
      
      THE ARTIFACT COLUMNS ARE HELD AGAINST THE GENERIC UPDATE AND REWRITTEN
      BY THE INTAKE PATH. Executed: `update({storagePath, sizeBytes, hash})`
      left all three columns unchanged, because `ModelUpdateRequestSchema`
      has no such fields and the update silently drops them. The only writer
      is `completeUpload`, which the I1.2.d run showed accepts a second call
      on an already-completed version AND a storage path belonging to a
      different version. Those two cases are CITED BY ID out of
      `isis-model-version-lifecycle.json` and re-read at build time, so a fix
      over there changes this artifact and breaks this cell's ratchet rather
      than leaving a stale sentence behind.
      
      THE ONLY ANCESTRY A VERSION CARRIES IS TWO FREE-TEXT FIELDS THE GENERIC
      UPDATE REWRITES. Executed: `update({baseModel, trainedOn})` rewrote both
      to arbitrary strings. Neither is resolved against a registered model
      (I1.2.f measured that for every model reference in the estate), so a
      version's recorded parentage is an editable label.
      
      THE AUDIT TABLE IS APPEND-ONLY, EVERYWHERE. No `UPDATE` and no `DELETE
      FROM model_registry_audit_log` exists in any tracked `.ts`, `.tsx` or
      `.sql` file; the service only inserts and selects; and entries carry a
      before and an after state. That is a genuine pass and it is what makes
      the rest of the ledger mean something. BUT NOT EVERY WRITE IS AUDITED:
      executed, `recordUsage()` incremented `usage_count` with the audit count
      unchanged and `updated_at` NOT moving, so a column changes with no entry
      and no clock. (A first cut compared `updated_at` against the row as it
      stood at registration, when two earlier `update()` calls had already
      moved it, and reported a clock that `recordUsage` does not touch; the
      snapshot is now taken immediately before the write under test.)
      
      NO DOMAIN EVENT IS PUBLISHED ON THE MODEL PATH AT ALL. Four audit
      actions go into a table and nothing goes onto a bus. The sibling
      `JobService` in the same app holds an `eventPublisher`, so the silence
      is a choice rather than a platform limit, and the ledger records the
      contrast.
      
      NEITHER BYPASS IS GUARDED, FOR TWO DIFFERENT REASONS, AND THE VERDICTS
      KEEP THEM APART. `status`: the only status a version has is `is_active`,
      it is a field of the generic update request, and no method on the
      service names activate, deprecate, revoke or transition — so THE BYPASS
      IS THE ONLY PATH. `current-version`: the live table declares zero
      columns naming current or default, so there is NOTHING TO BYPASS —
      which is not the same as guarded, and is the other face of I1.2.d's
      finding that the current version is derived as the first active row of
      an `ORDER BY created_at DESC`. Executed against `information_schema`
      rather than the DDL text, so the claim is about the table that exists.
      
      The one thing the trail does get right: when `is_active` flips, the
      entry is the generic `model.updated`, but its `changedFields` names
      `isActive` and its before and after states do record the flip. The
      transition is unnamed, not unrecorded.
      
      Reproduce: `ISIS_MODEL_PROBE_DATABASE_URL=… pnpm exec tsx
      scripts/v1-workbenches/generate-isis-model-immutability-ledger.mts`. The
      write path refuses without a database rather than replacing a real
      recorded run with an empty one; `--check` verifies the static half and
      says so, which is why the writer is not in `inventory:v1-workbenches`. -->
      
  • I1.3 Verify provider-endpoint lifecycle and health semantics. Evidence: state/health combinations have deterministic dispatch eligibility.

    • I1.3.a Inventory endpoint configuration/lifecycle/health/circuit/ maintenance/capability/capacity states, probes, routing, APIs/stores/UI, and tests. Evidence: source mapping. <!-- isis-endpoint-state-machine.json: source mapping of the CURRENT Isis provider-endpoint state machine (libs/isis/ai-providers/src/provider-endpoints; registry/health-monitor/router/dead-letter/queue-metrics read ONCE each for perf). 4 typed health states (ISIS_ENDPOINT_HEALTH_STATES: healthy/degraded/ unhealthy/unknown), 9 provider kinds (ISIS_PROVIDER_KINDS), 7 regions, 6 typed config interfaces (IsisProviderEndpoint/CostControls/ConcurrencyControls/ RetryPolicy/SecretRotation/Registry). All 9/9 I1.3.a dimensions map to typed evidence: configuration→IsisProviderEndpoint, lifecycle→health-states, health→health-monitor probes, circuit→dead-letter, maintenance→SecretRotation, capability→provider-kinds, capacity→concurrency/cost controls, probes→ health-monitor, routing→router. POSITIVE FINDING: unlike the weakly-formalized model machine (I1.2.a), the endpoint machine is WELL-TYPED. 16 endpoint/provider UI components, 3 provider-endpoints test files (both-globs fix caught top-level specs the ** glob misses). generate-isis-endpoint-state-machine.mjs + 6 tests, wired into inventory/verify + CI. I1.3.b+ (config/lifecycle/circuit verification) build on this and stay unchecked. -->

    • I1.3.b Define configured/unconfigured/validating/ready/draining/ maintenance/disabled/retired lifecycle and authorized transitions. Evidence: lifecycle state machine. <!-- evidence: isis-endpoint-dispatch-model.json .lifecycle — 8 states (unconfigured/configured/validating/ready/draining/maintenance/disabled/ retired) + 11 authorized transitions with actor (admin/operator/system); only ready is dispatch-eligible. generate-isis-endpoint-dispatch-model.mjs (--check gate), 14 tests. -->

    • I1.3.c Define probe unknown/healthy/degraded/unhealthy/unavailable/ incompatible/stale states with checked time, latency/error/capacity/ supported operation/model detail. Evidence: health contract. <!-- evidence: .healthContract — 7 probe states (4 realized in code per I1.3.a: healthy/degraded/unhealthy/unknown; 3 contract additions: unavailable/ incompatible/stale) each with 6 detail fields (checkedAt/latencyMs/errorRate/ capacity/supportedOperations/model). 4 code states ⊆ 7 contract (validated). -->

    • I1.3.d Define circuit closed/open/half-open and throttle/quota/capacity/ regional/license/dependency state separately from health. Evidence: dispatch input schema. <!-- evidence: .dispatchInputSchema — circuit (closed/open/half-open) + throttle/quota/capacity/regional/license/dependency, defined SEPARATELY from the health contract (health is not a field of the dispatch schema). -->

    • I1.3.e Produce a deterministic eligibility/routing/refusal reason table for every lifecycle×health×circuit×maintenance×capability×capacity combination. Evidence: truth table tests. <!-- evidence: .eligibility — a DETERMINISTIC, TOTAL decision function over all 1344 combinations (8×7×3×2×2×2): 4 eligible, the rest refused with 1 of 10 priority-ordered reasons (lifecycle→maintenance→circuit→health→capability→ capacity), ALL reachable (checked by full cross-product enumeration). The "truth table tests" are pure-function known-value tests (no running service). -->

    • I1.3.f Recheck endpoint state/health/capacity/model compatibility/ entitlement/budget at dispatch reservation and handle races. Evidence: dispatch integration tests. <!-- NOT ONE OF THE SIX DIMENSIONS IS RECHECKED AT DISPATCH; THE CELL STAYS [ ]. evidence/v1-workbenches/isis-dispatch-recheck.json — generate-isis-dispatch-recheck.mjs + .test.mjs (7 tests, 2 ratchets, 2 controls), wired into inventory + verify. The six dimensions are read from THIS cell's own words and the build stops if the cell stops naming one of them.

      THE DISPATCH GUARD IS REAL, FAIL-CLOSED AND WIRED — AND IT GUARDS THE
      OUTPUT, NOT THE ENDPOINT. `evaluateIsisDispatch` denies on a block, a
      review, a gate error or any denied admission, and
      `apps/oshun/bff/src/generation/release-gate.ts` calls it. But every
      field of the `CanonicalReleaseGateMeasurement` it receives is about the
      artifact — safety-scan score, provenance manifest, watermark coverage,
      synthetic-voice disclosure, quality aggregate, policy compliance, rights
      and licence, output shape. Endpoint state, health, capacity, model
      compatibility, entitlement and budget are not among them, and the BFF
      caller names none of them either. The guard is a pure function of its
      argument, so a dimension it cannot see cannot be re-read however
      carefully it was computed at plan time.
      
      SO THE ELIGIBILITY TABLE I1.3.e BUILT IS COMPUTED ONCE AND NEVER
      RECHECKED. The plan records `providerEndpointId`, `canaryEndpointId`,
      `modelVersion` and `estimatedCostUsd`; `dispatchGuardedGeneration` then
      calls the guard and, if permitted, calls `runGeneration`. Between those
      two moments an endpoint can drain, a circuit can open, capacity can fill
      and a budget can be spent, and nothing looks again.
      
      TWO OF THE SIX ARE ABSENT FROM BOTH SIDES. Health and capacity are not
      fields of the plan either, so they are not merely un-rechecked — the
      control plane never carries them into the decision at all, though I1.3.c
      and I1.3.d defined probe states and concurrency controls for them.
      
      AND THE RACE HALF IS UNADDRESSED: no reservation, no lock, no
      compare-and-set, no idempotency key anywhere on the dispatch path. The
      estate does know how — `WorkflowServiceImpl.update` refuses a stale
      `expectedUpdatedAt` — so this is a gap, not a limitation.
      
      TWO MATCHERS HAD TO BE NARROWED BEFORE ANY OF THIS WAS TRUE. An
      unanchored `queue` matched `enqueueGenerationJob` in a header comment
      and credited the guard with a capacity recheck it does not make; an
      unanchored `lock` matched `blockedReasons`, which is on the dispatch
      path in quantity, and reported a lock where there is none. Comments are
      stripped before matching, the race vocabulary is word-anchored, and a
      control test pins that both patterns still match what they are for. -->
      
    • I1.3.g Verify drain/disable/maintenance/failover/recovery behavior for queued/in-flight/retry work and unknown external outcomes. Evidence: fault tests. <!-- THREE OF THE FIVE EVENTS ARE NOT STATES AN ENDPOINT CAN ENTER, AND THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-endpoint-fault-behaviour.json —
      generate-isis-endpoint-fault-behaviour.mts + .test.mts (12 tests, 5
      ratchets, 4 controls), wired into inventory + verify. Every endpoint
      module is pure, so the FAULT TESTS the cell asks for execute the real
      exported functions with no service and no database, and the endpoint
      fixtures are checked by the registry's own `validateIsisProviderEndpoint`
      before use.
      
      DRAIN, DISABLE AND MAINTENANCE ARE NOT IN THE CODE AT ALL. Not weakly
      modelled — absent: no field of `IsisProviderEndpoint`, no member of
      `ISIS_ENDPOINT_HEALTH_STATES`, and no identifier in any of the five
      modules names them. I1.3.b defined an eight-state lifecycle
      (unconfigured/configured/validating/ready/draining/maintenance/disabled/
      retired) as a CONTRACT; the endpoint the router reads carries `health`
      and nothing else. Executed: routing to a `degraded` endpoint is GRANTED,
      so the only lever that stops new work is `unhealthy`, which is a
      statement about how the endpoint is behaving and not a decision anyone
      made. There is no way to say "stop sending, finish what you have".
      
      WHAT DOES WORK, WORKS, AND THAT IS WHAT MAKES THE REST MEAN SOMETHING.
      Executed: an `unhealthy` endpoint is refused with
      `considered=[primary:unhealthy]`; in-flight at the concurrency ceiling
      is refused; and a failed primary routes through to its declared
      companion with reason `failover-companion-match`.
      
      QUEUE DEPTH IS MEASURED AND NEVER CONSULTED. `IsisQueueMetrics` reports
      depth, lag percentiles, throughput and success ratio per endpoint;
      `IsisRouteRequest` has no field for any of it. Executed: with a depth of
      2 recorded for the endpoint, the router granted the route.
      
      A TIMEOUT IS AN UNKNOWN OUTCOME AND THE ESTATE TREATS IT AS A FAILURE.
      `decideRetry` retries a timeout — executed, `kind="retry"` after 100ms —
      and neither the retry decision nor the route request carries an
      idempotency key, which matches I1.3.f's finding that there is none
      anywhere on the dispatch path. So a provider call that succeeded and
      whose answer was lost is retried, and the second attempt is
      indistinguishable from the first.
      
      AND THE VOCABULARY FOR "UNKNOWN" EXISTS WHERE NOTHING IS AT STAKE.
      `HEALTH_ERROR_KINDS` has seven words INCLUDING `unknown`, so a probe can
      report one. `RETRY_ERROR_KINDS` has four and `unknown` is not among
      them: executed, `decideRetry({errorKind:'unknown'})` throws
      `failure.errorKind "unknown" is not canonical`. THE ONE PLACE THAT CAN
      SAY IT IS THE ONE PLACE WHERE NO WORK IS PENDING.
      
      A LIVE DEFECT IN THE RECOVERY PATH. `evaluateEndpointHealth`'s doc
      comment says "Newest outcomes first"; its body is
      `.slice(-window.samples)`, which takes the LAST entries. Executed: the
      same eight probes — three failures then five successes — give
      `healthy` when passed oldest-first and `unhealthy` when passed
      newest-first. The one in-repo caller, `createHealthMonitorController`,
      appends to the tail and is therefore correct; ANY CALLER OBEYING THE
      DOCUMENTED CONTRACT IS JUDGED ON THE STALEST PROBES. Recovery also has
      no hysteresis and no minimum time in state: the window flips as soon as
      the ratio crosses, recorded as an observation rather than a finding
      because the cell asks what recovery does, not what it should be.
      
      AND ALMOST NOTHING CALLS ANY OF IT. Of seven entry points —
      `planIsisRoute`, `evaluateEndpointHealth`,
      `evaluateEndpointHealthBatch`, `createHealthMonitorController`,
      `decideRetry`, `buildDeadLetterEntry`, `IsisQueueMetrics` — ONE has a
      caller outside its own module and its tests, and that caller is
      `libs/isis/study-adapter/src/adapter.ts`, whose package
      `@isis/study-adapter` is imported by nothing. The second hop is
      MEASURED, not asserted, and an unmeasurable hop is flagged rather than
      reported as empty: `@isis/ai-providers` has only subpath mappings in
      `tsconfig.base.json` and no root one, so a caller living there would
      resolve to null, and the artifact carries a field that says so.
      `decideRetry` is also a name three unrelated libraries export, so a
      caller only counts if it imports the Isis package — a control pins that
      Euterpe's own `decideRetry` is not miscounted.
      
      I1.3.a called this machine WELL-TYPED and it is. The question it did not
      ask is whether anything runs it. -->
      
  • I1.4 Define a generation brief with intent, audience, asset kind, locale, references, rights assertions, safety context, quality targets, budget, and delivery target. Evidence: schema invariants. <!-- evidence: COMPILING reference evidence/v1-workbenches/generation-brief/*.ts (schema.ts 11 facet interfaces a-k + validate.ts cross-field suite l + fixtures.ts 1 valid/8 invalid + check.ts runnable), strict tsc --noEmit + invariant suite run via tsx (both gated in verify:inventory:v1-workbenches). a=BriefIdentity, b=BriefIntent(+intent fixture), c=BriefAudience(+validateAudience Entitlement), d=BriefAssetKind(+validateAssetKind, grounds to GenerationType I0.5.a), e=BriefLocale, f=BriefReference(+reference fixture), g=RightsClaim as evidence-backed claims (verified rejected without evidenceRef — NOT client truth), h=BriefSafety(+ validateSafety: minor-present⇒reviewer+consent), i=QualityTarget, j=BriefBudget (no fabricated cost — nullable ceiling), k=BriefDelivery, l=validateBrief cross-field suite. generate-generation-brief-schema.mjs (--check gate), 9 tests. -->

    • I1.4.a Define stable brief ID/version, tenant/project/parent context, author/delegate, lifecycle, base revision, title/summary, and created/updated metadata. Evidence: brief identity schema.
    • I1.4.b Define structured creative/production intent, goal/non-goals, subject/content/action/context/style/tone, required/excluded elements, and domain extension blocks. Evidence: intent fixtures.
    • I1.4.c Define audience/tier/end user/age/region/channel/purpose/ sensitivity and verify it against actor/host entitlement. Evidence: audience tests.
    • I1.4.d Define asset-kind/profile/operation/count/variant/group/ composite relationship and kind-specific typed control selection. Evidence: asset-kind invariant tests.
    • I1.4.e Define content/UI/source/target locale/language/script/direction/ cultural context and localizable versus stable identifiers. Evidence: locale tests.
    • I1.4.f Define ordered reference roles/weights/scopes/hashes/versions/ rights/consent/provenance/availability and conditioning bindings. Evidence: reference fixtures.
    • I1.4.g Define rights/license/ownership/attribution/territory/term/ derivative/training/publication assertions as claims requiring evidence, not client truth. Evidence: rights validation tests.
    • I1.4.h Define safety context for people/identity/age/biometric/voice/ sensitive content/policy/risk/reviewer/consent and required restrictions. Evidence: safety fixtures.
    • I1.4.i Define asset/delivery-specific raw quality targets/units/ thresholds/tolerance/priorities/evaluator/profile versions and unavailable behavior. Evidence: quality target tests.
    • I1.4.j Define currency/rate-card/estimate ceiling/tenant-project budget/ latency/resource/carbon or other approved constraints without fabricated cost. Evidence: budget tests.
    • I1.4.k Define delivery target/host/channel/territory/environment/format/ profile/dimensions/time/audio/3D package/naming/deadline/embargo/ publisher and compatibility. Evidence: delivery target tests.
    • I1.4.l Enforce cross-field authorization/compatibility/safety/rights/ budget/quality/delivery invariants and retain valid/invalid/full fixtures. Evidence: schema invariant suite.
  • I1.5 Define a resolved generation plan distinct from the user brief. Evidence: plan records workflow/model/provider choices and every reason. <!-- evidence: COMPILING reference evidence/v1-workbenches/generation-plan/*.ts (plan-schema.ts + resolve.ts + fixtures.ts + check.ts), imports the I1.4 brief; strict tsc --noEmit + plan suite via tsx (both gated in verify). a=PlanIdentity (immutable: planHash+version+exact briefRevision), b=resolveBriefToPlan + NormalizedIntent w/ ResolvedField.source (user/default/derived/constrained/ unsupported — over-limit count constrained to 8), c/d/e=Decision<Workflow/Model/ Provider Choice> recording chosen + rejected alternatives WITH reasons (every reason), f=ExecutionGraph, g=PlanBudget (per-stage estimate/reservation/uncertainty +stopConditions), h=validatePlan + 6 invalid fixtures, i=createReplan + ReplanLink (original kept IMMUTABLE, replanPreservesOriginal). generate-generation-plan-schema.mjs (--check gate), 8 tests. -->

    • I1.5.a Define immutable plan ID/version, originating exact brief revision, resolver/policy/catalog versions, actor/job/time, status, expiry, and plan hash. Evidence: plan identity schema.
    • I1.5.b Record normalized intent and every default/derived/unsupported/ constrained field separately from user-authored brief values. Evidence: brief-to-plan mapping tests.
    • I1.5.c Record chosen workflow/template/version/graph/class/node/ parameter profile and compatible alternatives/rejections with reasons. Evidence: workflow decision fixtures.
    • I1.5.d Record chosen model/version/checkpoint/adapter/LoRA/scheduler/ preprocessor/evaluator and alternatives/rejections with compatibility/ policy reasons. Evidence: model decision fixtures.
    • I1.5.e Record chosen provider/endpoint/region/resource/queue/failover/ priority/capacity reservation and alternatives/rejections with health/ entitlement/latency/cost reasons. Evidence: provider decision fixtures.
    • I1.5.f Record transformed/resolved inputs/references/conditioning, output kinds/counts/formats/bundles, post-processing, gates/review/ provenance/watermark/publisher steps. Evidence: execution graph.
    • I1.5.g Record per-stage estimate/reservation, resource/time/cost ranges, uncertainty, budget checks, and stop conditions. Evidence: plan budget tests.
    • I1.5.h Validate plan against current actor/tier/host, contracts/versions, compatibility, config/health, rights/consent/safety, capacity/budget, and delivery before approval/dispatch. Evidence: invalid plan tests.
    • I1.5.i Keep plans immutable and create linked replan versions when any brief/dependency/policy/provider/capacity/target input changes. Evidence: replan lineage tests.
  • I1.6 Define immutable generation execution records. Evidence: seed, parameters, inputs, policy, provider response, costs, outputs, and timings are retained. <!-- evidence: COMPILING reference evidence/v1-workbenches/execution-record/*.ts (execution-schema.ts 9 facets + record.ts validation/append-only + fixtures.ts 1 valid/7 invalid + check.ts), strict tsc --noEmit + record suite via tsx (both gated in verify). a=ExecutionIdentity (pins exact plan/brief revisions), b=Runtime Provenance, c=ParameterRecord (seed/subseeds/latents/typed params/parameterHash), d=InputArtifact (ordered+hash+role), e=PolicyProvenance (decisions+dispatchEligible; failed⇒not-eligible), f=ProviderRecord (secure REFS not inlined secrets), g=CostRecord (actual NULLABLE — no fabricated cost; variance rejected without settlement), h=OutputRecord (noOutputReason), i=TimingRecord (clock/unit), j=appendCorrection + CorrectionLink (NEW linked record, original IMMUTABLE + legal hold, correctionPreserves Original). generate-execution-record-schema.mjs (--check gate), 9 tests. -->

    • I1.6.a Define execution/run/attempt/stage IDs, exact plan/brief revisions, actor/tenant/project/host/tier, job/correlation/trace, lifecycle, and timestamps. Evidence: execution identity schema.
    • I1.6.b Record resolved workflow/graph/template/node/model/checkpoint/ adapter/LoRA/scheduler/provider/endpoint/resource/environment/container/ dependency versions. Evidence: runtime provenance tests.
    • I1.6.c Record seed/subseeds/latent refs and exact typed stage parameters/ defaults/derived values with units and immutable hashes. Evidence: parameter tests.
    • I1.6.d Record ordered immutable input/reference/conditioning/mask/ source artifact refs, hashes, roles, transformations, rights/consent, and availability. Evidence: input tests.
    • I1.6.e Record safety/rights/consent/quality/release policy, threshold, gate/evidence/waiver versions, decisions, and dispatch eligibility. Evidence: policy provenance tests.
    • I1.6.f Securely retain or reference raw provider request/response/ stream/tool events with IDs/hashes/status/finish/safety/usage/error/ cancellation and normalized adapter version. Evidence: provider record tests.
    • I1.6.g Record estimate/reservation/actual/refund/variance cost and token/ GPU/CPU/memory/storage/egress/provider/resource usage per stage/attempt. Evidence: cost tests.
    • I1.6.h Record intermediate/partial/quarantined/rejected/final output refs/hashes/metadata/state/provenance and no-output reason. Evidence: output tests.
    • I1.6.i Record queued/leased/started/first output/stage/checkpoint/ completed/cancelled/failed/retry/total timing with clock/unit semantics. Evidence: timing tests.
    • I1.6.j Enforce append-only/immutable attempts with access/retention/ deletion/legal hold and new linked record for corrections/replay. Evidence: immutability/lifecycle tests.
  • I1.7 Define output and ingredient lineage. Evidence: composites retain ordered relationships to every source and generated ingredient.

    • I1.7.a Define lineage node types for source/reference/input, intermediate, generated output, edited derivative, composite, bundle, provenance manifest, and release with exact revision/hash. Evidence: node schema. <!-- evidence (I1.7.a/b/d/e/f/g): COMPILING reference evidence/v1-workbenches/ lineage/*.ts (lineage-schema.ts 10 NodeType + 9 EdgeType w/ role/weight/mask/ time/range/part + EdgeProvenance; graph.ts traverse/computeImpact/computeRecall/ validateIntegrity/ingredientUses; fixtures.ts composite graph; check.ts). a=LineageNode, b=LineageEdge, d=ingredientUses (src used 2x at distinct layers, equal hash NOT collapsed), e=EdgeProvenance (run/stage/actor/rights per edge), f=cycle-safe bounded traversal (impact/recall), g=validateIntegrity (dangling/ cross-tenant/cross-brief/hash-mismatch rejected; inaccessible reported honestly). strict tsc --noEmit + lineage suite via tsx gated in verify. generate-lineage- schema.mjs (--check gate), 10 tests. -->

    • I1.7.b Define typed directional ordered edges for uses/conditions/ derives/transforms/composites/mixes/layers/supersedes/publishes with role/weight/mask/time/range/part metadata. Evidence: edge schema.

    • I1.7.c Create edges transactionally for ingest, planning, execution stages, post-process, manual edit, variant merge, bundle, promotion, and publication. Evidence: mutation integration tests. <!-- SUPERSEDED (2026-08-16). The old note said "runtime: requires the running service + DB". WHETHER A WRITE IS INSIDE A TRANSACTION IS A PROPERTY OF THE SOURCE. evidence/v1-workbenches/isis-lineage-write-transactionality.json — generate-isis-lineage-write-transactionality.mjs + .test.mjs (7 tests, 1 ratchet, 2 controls), wired into inventory + verify. Every statement in apps/isis/output-registry that writes a lineage-bearing column is located and scoped against the withTransaction callback bodies, BRACE-MATCHED rather than by proximity, because a write a few lines below a transaction that has already closed is not inside it.

      THE LINEAGE GRAPH HAS TWO WRITERS AND ONLY ONE OF THEM IS
      TRANSACTIONAL. `addLineageEdge` wraps its `INSERT INTO lineage_edges`
      and the accompanying audit record in one `withTransaction`, so an
      explicit edge and its audit trail land together or not at all.
      `register` — the path that creates an output WITH parents — does not: it
      inserts the output row carrying `parent_ids`, then issues a separate
      `UPDATE outputs SET child_ids = array_append(…)` against the parents,
      then writes the audit record, as THREE UN-TRANSACTED STATEMENTS.
      
      SO A FAILURE BETWEEN THEM LEAVES THE GRAPH ASYMMETRIC: a child that
      names its parents while the parents do not name the child. Both halves
      are read back — `getLineage` walks `parentIds` for ancestors and
      `childIds` for descendants — so the two directions can disagree about
      the same link, and the direction a caller asks for decides which answer
      they get.
      
      A SINGLE UN-TRANSACTED STATEMENT IS NOT THE DEFECT; A GROUP IS. The
      census reports per METHOD, and a method with one write is atomic by
      itself. `register` has three that must all land or none.
      
      AND THE CELL NAMES NINE STAGES WHILE THE ESTATE HAS TWO WRITERS.
      Ingest, planning, execution, post-process, manual edit, variant merge,
      bundle, promotion and publication are the moments an edge must be
      created transactionally; the output registry offers `register` and
      `addLineageEdge`, and nothing in it distinguishes which stage a caller
      is in. The cell stays `[ ]` on both counts. -->
      
    • I1.7.d Preserve each duplicate use and ordering/layer/track/shot/part relationship where semantically meaningful without collapsing equal hashes. Evidence: composite fixtures.

    • I1.7.e Link edge to run/stage/action/tool/model/policy/actor/time and rights/consent/license/attribution/provenance evidence. Evidence: provenance tests.

    • I1.7.f Traverse exact upstream/downstream paths and compute impact/ recall with cycle/fan-out/depth/pagination and redaction. Evidence: graph tests.

    • I1.7.g Validate no missing/dangling/cross-tenant/cross-brief-invalid/ hash-mismatched ingredient and represent inaccessible/deleted/unknown honestly. Evidence: integrity tests.

  • I1.8 Define variant groups and comparison membership. Evidence: variants cannot cross tenant, brief, or incompatible asset kinds.

    • I1.8.a Define group ID/version, tenant/project, exact brief revision, purpose/strategy, asset kind/profile, base/source, owner, lifecycle, and limits. Evidence: group schema. <!-- evidence (I1.8.a-f): COMPILING reference evidence/v1-workbenches/variant-group/*.ts (variant-schema.ts VariantGroup+VariantMember + append-only history; variant-ops.ts validateMembership/validateCompatibility/addMember/removeMember/freezeGroup/ freezeComparison/opPreservesInput; fixtures.ts + check.ts). a=VariantGroup (tenant/ project/exact-brief scoped), b=VariantMember (source provenance), c=membership rejects cross-tenant/cross-brief, d=compatibility rejects incompatible kind w/o adapter, e=IMMUTABLE lifecycle (new version + append-only history + optimistic expectedVersion, opPreserves Input), f=freezeComparison pins revision/hash + represents unavailable/stale/redacted honestly. strict tsc + variant suite via tsx gated. generate-variant-group-schema.mjs (--check), 9 tests. -->

    • I1.8.b Define member ID/order/label/source manual-or-run/plan/run/output/ revision/hash, state, provenance, created actor/time, and archive. Evidence: member schema.

    • I1.8.c Validate same tenant/project/authorized scope and exact brief or explicitly compatible fork lineage. Evidence: tenancy/brief negative tests.

    • I1.8.d Validate asset-kind/profile/semantic/time-base/dimension/channel/ 3D coordinate compatibility or require typed cross-kind adapter. Evidence: compatibility tests.

    • I1.8.e Support add/remove/reorder/rename/archive/restore/freeze/duplicate/ merge-derived member with expected revision and immutable history. Evidence: lifecycle tests.

    • I1.8.f Freeze comparison input revisions/hashes/adapters/score/gate versions and represent unavailable/stale/redacted member. Evidence: stable compare tests.

    • I1.8.g Link selection/rejection/decision/rationale/evidence and preserve all considered members. Evidence: decision integration tests. <!-- SUPERSEDED (2026-08-16). The old note said this needs "the running comparison service" and that the freeze contract "already preserves all considered members". The variant-group module is PURE, so the decision questions are function calls; and the preserve claim holds for the LIVE set, which is not the same thing.

      NONE OF THE FIVE ELEMENTS THE CELL NAMES EXISTS, AND THE CELL STAYS
      `[ ]`.
      
      evidence/v1-workbenches/isis-variant-decision-record.json —
      generate-isis-variant-decision-record.mts + .test.mts (9 tests, 3
      ratchets, 3 controls), wired into inventory + verify. Selection and
      rejection are asked for as STATES, decision as an EVENT, rationale as a
      FIELD on that event and evidence as a LINK from it — separately,
      because a schema can have one without the others and collapsing them
      would let one present element stand in for four absent ones.
      
      THE NINE OPS ARE add, archive, duplicate, freeze, merge-derived, remove,
      rename, reorder, restore. THE FIVE MEMBER STATES ARE active, archived,
      redacted, stale, unavailable. Not one names a selection or a rejection.
      A history entry carries exactly `[actorId, atVersion, memberId, op]` —
      THERE IS NOWHERE TO PUT A REASON. A frozen member carries `[adapter,
      availability, gateVersion, hash, memberId, revision, scoreVersion]`: the
      score and gate VERSIONS are pinned and no score, metric or evaluation
      reference is among them. `scoreVersion` is recorded as a NEAR-MISS
      rather than a carrier, because it is the version of a scorer and not a
      score, and "absent" should not read as "nothing resembling it exists".
      
      THE PRESERVE CLAIM, EXECUTED, IS HALF TRUE. A stale member DOES survive
      a freeze with its availability represented honestly — three members in,
      three out, `[available, available, stale]`. But `removeMember` drops the
      member from the live set and `freezeComparison` maps over the live set,
      so A MEMBER CONSIDERED AND THEN REMOVED IS ABSENT FROM THE COMPARISON:
      three members became two, and the removal survives only as a history
      entry carrying no revision, no hash and no state. "All considered
      members" and "all current members" are different sets and the freeze
      pins the second.
      
      THE ESTATE KNOWS HOW TO RECORD A DECISION. 175 non-test files carry a
      `decisionRationale` or a `rejectionReason` — persona releases, output
      spot-checks, research integrity, avatar packs. The variant comparison
      having none is a choice, not a platform limit.
      
      The positive control keeps the silences meaningful: `addMember` with a
      stale `expectedVersion` was refused with "stale group version 0 (current
      1)", so the module does say no. -->
      <!-- AND THE CONTRACT HALF IS A SPECIFICATION. I1.8.a-f cite the
      compiling reference under evidence/v1-workbenches/variant-group/, and
      the artifact measures its importers: ZERO files under apps/ or libs/
      reference it. Whatever decision record this cell wants would have to be
      built where the product can reach it. -->
      
  • I1.9 Define quality profiles per asset kind and delivery target. Evidence: thresholds are versioned configuration with code-owned floors.

    • I1.9.a Define profile ID/version/owner, asset kind/operation, delivery target/profile/environment/audience, applicability, lifecycle, and effective period. Evidence: quality profile schema. <!-- evidence (I1.9.a-f): COMPILING reference evidence/v1-workbenches/quality-profile/*.ts (quality-schema.ts QualityProfile/QualityMeasure/Threshold/CodeFloor; policy.ts applyFloor/mergePolicies/versionPolicy; fixtures.ts + check.ts). a=QualityProfile, b=QualityMeasure registry, c=Threshold fixtures, d=CODE-OWNED floors — applyFloor lets config only TIGHTEN (loosening clamped to floor + flagged, never accepted), e=merge global→…→project under TOTAL LAYER_PRECEDENCE + provenance (loosening layer overridden by code-floor), f=versionPolicy (bump + invalidate changed + retain raw). strict tsc + quality suite via tsx gated. generate-quality-profile-schema.mjs (--check), 9 tests. -->

    • I1.9.b Register raw technical/semantic/aesthetic/temporal/audio/3D/ accessibility/safety/rights/provenance measures with units/direction/ evaluator and nonvalue states. Evidence: measure registry.

    • I1.9.c Define target-specific thresholds/tolerance/severity/composition/ required evidence/freshness/remediation and human-review bands. Evidence: policy fixtures.

    • I1.9.d Define immutable code-owned minimum/maximum/fail-closed floors for safety/integrity/compatibility and only allow configuration to tighten. Evidence: loosening negative tests.

    • I1.9.e Merge global/asset-kind/target/territory/channel/tenant/project policies with field provenance and no ambiguous precedence. Evidence: merge truth tables.

    • I1.9.f Version changes, invalidate old results where required, retain raw values, and define compatibility/deprecation/migration. Evidence: policy-change tests.

    • I1.9.g Calibrate/validate profiles on representative known-value/ boundary/adversarial corpora and obtain domain/target approvals. Evidence: calibration report. <!-- THE HUMAN HALF IS REAL AND IT IS NOT THE BLOCKER. THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-quality-calibration.json —
      generate-isis-quality-calibration.mts + .test.mts (9 tests, 3 ratchets,
      2 controls), wired into inventory + verify.
      
      AN APPROVAL THAT IS UNOBTAINED AND ONE THAT IS UNRECORDABLE ARE
      DIFFERENT FINDINGS. Nobody signs off in a script, so the old note's
      "human" is fair — but `QualityProfile` declares thirteen fields
      (applicability, assetKind, audience, deliveryProfile, deliveryTarget,
      effectiveFrom, effectiveTo, environment, lifecycle, operation, ownerId,
      profileId, version) and NOT ONE of them could hold an approver, a
      signoff, a reviewer or a date. A domain or target approval obtained
      tomorrow would leave no trace. That is a schema gap, not a scheduling
      one.
      
      CALIBRATION MEANS RUNNING A MEASURE'S EVALUATOR, AND THE NAMES DO NOT
      RESOLVE. Every `QualityMeasure` names its evaluator with a bare string.
      The three declared are `laion`, `safety-classifier` and
      `tox-classifier`. `tox-classifier` appears NOWHERE ELSE IN THE
      REPOSITORY. The other two are MENTIONED — `laion` in ComfyUI checkpoint
      filenames and another reference's fixtures, `safety-classifier` in a
      Lilith test's `classifierId` and an Oshun BFF module name — and in none
      of those is the string a registry key, a `case` label or a registration
      call. A mention is not a dispatch, and the artifact records the two
      classes apart.
      
      THE CORPORA ARE REAL AND THEY ARE SOMEWHERE ELSE. `apps/isis/
      generation-api/src/routes` carries 32 regression golden datasets, one
      per asset family, each with thresholds and a quality gate. ALL 32 CARRY
      KNOWN-VALUE EXPECTATIONS. NONE CARRIES A BOUNDARY OR AN ADVERSARIAL
      CASE — two of the three kinds this cell names. And they are not attached
      to the `QualityProfile` model at all: nothing under apps/ or libs/
      imports the reference.
      
      THE PROFILE MODEL ITSELF IS SOUND, which is what makes the gaps mean
      something. Executed: `applyFloor` clamps a loosening config to the floor
      and flags `loosened=true`, while a tightening config passes through
      unflagged; `mergePolicies` returns effective thresholds each carrying
      `[floorEnforced, sourceLayer, threshold]`, so the winning layer is
      recorded.
      
      A CONTROL THAT MATTERED. Matching corpus kinds as bare words found
      "transient attack" in the audio golden set and "phrase boundaries" in
      the voice one — creative direction inside string literals — and reported
      an adversarial and a boundary corpus that do not exist. A kind now
      counts only where the term appears as an identifier before `:`/`=`/`(`
      or as a quoted value in its own right, and a test pins both prose
      strings as non-matches.
      
      AND ONE THE ARTIFACT DID TO ITSELF. The evaluator mention count sweeps
      `*.json` as well as `*.ts`, so once the artifact was committed it
      counted its OWN copy of every evaluator name and the count grew by one
      each run — `--check` could never be satisfied. The lane's own outputs
      (`evidence/v1-workbenches/<name>.json`) are excluded; the reference
      DIRECTORIES beside them are sources and stay in. -->
      <!-- Still human: the approvals themselves. Still provider-bound:
      nothing — calibration needs evaluators that resolve, which is a code
      gap, not an infrastructure one. -->
      
  • I1.10 Define release-gate policy composition. Evidence: safety, provenance, watermark, quality, policy, rights, output shape, and review requirements have deterministic precedence.

    • I1.10.a Define safety gates for content/person/identity/child/biometric/ impersonation/provider policy with nonoverrideable classes and fail- closed unavailable behavior. Evidence: safety gate policy. <!-- evidence (I1.10.a-i): COMPILING reference evidence/v1-workbenches/release-gates/*.ts (gate-schema.ts 8 GateClass policies + 9 GateOutcome states + OUTCOME_PRECEDENCE + NON_OVERRIDABLE_CLASSES; compose.ts composeRelease/enumerateTruthTable; fixtures.ts one policy per class + rule fixtures; check.ts). a=SafetyGatePolicy (nonOverridable + failClosedWhenUnavailable), b=ProvenanceGatePolicy, c=WatermarkGatePolicy, d=Quality GatePolicy (missingCountsAsPass:FALSE — no missing-to-pass), e=PolicyGateRow (version+ owner), f=RightsGatePolicy (requiresEvidence — never client truth), g=ShapeGatePolicy, h=ReviewGatePolicy (automationProhibited), i=composeRelease = DETERMINISTIC TOTAL aggregation over all 36 truth-table rows (9 outcomes x required x class) → release/ block/review-required; FAIL-CLOSED on required non-passing; safety+review never waivable; waiver w/o waiverRef blocks; optional failures warn only. strict tsc + release suite via tsx gated. generate-release-gate-composition.mjs (--check), 10 tests. -->

    • I1.10.b Define provenance gates for artifact hash, ingredient/action/ tool/actor completeness, manifest binding/signature/trust/revocation/ recovery and target requirement. Evidence: provenance policy.

    • I1.10.c Define watermark/disclosure gates for method/profile/presence/ robustness/detectability/accessibility/target/channel and stripped-state fallback. Evidence: watermark policy.

    • I1.10.d Define asset/target quality gates from raw measures/evaluator/ profile/threshold/freshness and no missing-to-pass. Evidence: quality policy.

    • I1.10.e Define platform/product/brand/editorial/marketplace/provider/ territory/channel policy gates with exact version and owner. Evidence: policy gate rows.

    • I1.10.f Define rights/consent/license/attribution/training/derivative/ identity/territory/term/revocation gates with evidence and expiry. Evidence: rights policy.

    • I1.10.g Define output-shape/format/schema/hash/size/media/technical/ bundle/dependency/target-compatibility gates. Evidence: shape policy.

    • I1.10.h Define human review/signoff roles/quorum/sequence/COI/step-up/ expiry/staleness and automation prohibition. Evidence: review policy.

    • I1.10.i Define deterministic precedence/aggregation for blocker/warning/ unavailable/stale/waived/overridden/unsupported/partial across required/ optional and target-specific gates. Evidence: exhaustive truth tables.

    • I1.10.j Bind eligibility to exact output/dependency/evaluator/policy/ evidence/waiver/signoff/target versions and audit all reasons. Evidence: composition integration tests. <!-- SUPERSEDED (2026-08-16). The old note said "runtime: requires the running service + audit store". IT DOES NOT: the binding is evaluateProofRequirements in @oshun/contracts/common, a pure function, and the evaluateReleaseGate that calls it is pure too. Both run in-process with crafted known values. evidence/v1-workbenches/isis-release-binding-probe.json — generate-isis-release-binding-probe.mts + .test.mts (7 tests, 3 ratchets, 1 control), wired into inventory + verify.

      THE BINDING THE ESTATE DOES HAVE IS STRICT AND IT WORKS. A proof
      differing from the requirement in exactly ONE way — a different subject,
      a different policy, the wrong proof type, a failed status, a revoked
      status, an expiry already passed, a verification instant in the future —
      is refused in all seven cases, each with its own reason, while the
      control proof that differs in nothing is satisfied. That control is what
      makes the seven refusals evidence about the binding rather than about
      the fixture.
      
      BUT OF THE EIGHT BINDINGS THE CELL NAMES, THE REQUIREMENT CAN DEMAND
      TWO. `ProofRequirementSchema` is `.strict()` with four fields —
      requirementId, subjectRef, requiredProofType, policyRef — so OUTPUT and
      POLICY are demandable and nothing else is. SIGNOFF is demandable only
      indirectly, by requiring the `human_approval` proof type, which cannot
      name a particular signer. DEPENDENCY, WAIVER and TARGET are expressed by
      nothing at all: no requirement field, no proof field, no proof type.
      
      AND TWO ARE CARRIED BUT NEVER COMPARED, WHICH IS THE SHARPEST CATEGORY.
      Every proof records a `producerRef` — the estate knows WHICH EVALUATOR
      issued it — and the binding never looks: the probe varies exactly that
      field and the requirement is still satisfied. Every proof carries
      `evidenceRefs` and the verdict returns them, so evidence is AUDITED; but
      no requirement can ask for particular evidence, so it is not DEMANDED.
      
      THE POLICY BINDING IS BY NAME, NOT BY VERSION. The gate builds its
      requirement with `policyRef: `isis-release:${kind}``, a constant per gate
      kind, so every evaluation of a given gate binds to the same reference
      forever and changing the policy does not invalidate a proof issued under
      the old one — which is precisely the version binding this cell asks for.
      
      The cell stays `[ ]`: two of eight demandable, one indirect, two carried
      but uncompared, three inexpressible, and no version anywhere. -->
      <!-- runtime: binding eligibility to exact live output/dependency/evaluator/
                                                                                                                                                                                                          policy/evidence/waiver/signoff/target VERSIONS and auditing every reason requires the
                                                                                                                                                                                                          running service + audit store (integration tests). The composition contract (i)
                                                                                                                                                                                                          already retains every reason. Parent I1.10 stays unchecked. -->
      
  • I1.11 Define model, adapter, LoRA, checkpoint, node, and workflow compatibility constraints. Evidence: incompatible combinations fail planning before dispatch. <!-- evidence: COMPILING reference evidence/v1-workbenches/compatibility/*.ts (constraints-schema.ts 6 component surfaces + IssueKind; solver.ts solveCompatibility; fixtures.ts 12-row MATRIX; check.ts), strict tsc --noEmit + compat suite via tsx (both gated in verify). a=ModelConstraints, b=AdapterConstraints, c=LoRAConstraints, d=CheckpointConstraints, e=NodeConstraints (socket typing), f=WorkflowConstraints, g=solveCompatibility returning EXACT conflict/unsupported/missing-metadata/version-skew w/ components+field+alternatives (NO implicit conversion; absent metadata BLOCKS; unknown extension refused not ignored), h=12-row fixture matrix (pairwise/multi-lora overlap/adapter stacking/version-skew/missing-metadata/unknown-extension/provider capability-drift/node-socket/missing-node/precision/audience/post-plan-change). PARENT satisfied: every incompatible plan has dispatchable=false — incompatible combinations FAIL PLANNING BEFORE DISPATCH. generate-compatibility-constraints.mjs (--check gate), 10 tests. -->

    • I1.11.a Define model family/architecture/task/modality/input-output/ tokenizer/encoder-decoder/precision/quantization/context/license/ provider/runtime compatibility. Evidence: model constraints.
    • I1.11.b Define adapter type/base architecture/model/version/target modules/dimensions/weights/format/precision/license and stacking/ conflict compatibility. Evidence: adapter constraints.
    • I1.11.c Define LoRA base model/version/architecture/target modules/rank/ alpha/weight format/trigger/tags/precision/license/style/identity policy and multi-LoRA compatibility. Evidence: LoRA constraints.
    • I1.11.d Define checkpoint architecture/base/version/format/hash/source/ config/tokenizer/VAE/text encoder/scheduler/precision/license/security compatibility. Evidence: checkpoint constraints.
    • I1.11.e Define node implementation/type/version/input-output socket/ schema/data type/shape/unit/asset kind/runtime/provider/resource/ determinism/security compatibility. Evidence: node constraints.
    • I1.11.f Define workflow graph/template/version required/optional nodes/ models/adapters/LoRAs/checkpoints/providers/resources/asset kinds/ audiences/outputs and target constraints. Evidence: workflow constraints.
    • I1.11.g Implement versioned constraint registry and planner solver that returns exact conflicts/alternatives/unsupported rather than dispatch or implicit conversion. Evidence: planning negative tests.
    • I1.11.h Test pairwise and multi-component combinations, version skew, missing metadata, unknown extensions, provider capability drift, and post-plan change. Evidence: compatibility fixture matrix.
  • I1.12 Define deterministic replay eligibility. Evidence: required artifacts, versions, seeds, and environment dependencies are explicit.

    • I1.12.a Define replay class: byte-deterministic, semantically reproducible-with-tolerance, provider-best-effort, or nonreplayable with reason. Evidence: replay taxonomy. <!-- evidence (I1.12.a-g): COMPILING reference evidence/v1-workbenches/replay/*.ts (replay-schema.ts ReplayClass 4-class + REQUIREDVERSION_KEYS(15)/REQUIRED_ENVIRONMENT KEYS(12) + DependencyProblemKind(6); eligibility.ts checkVersion/Artifact/Environment Completeness + classifyProviderCeiling + deriveReplayClass + checkPermission + buildReplayPreview; fixtures.ts + check.ts). a=class DERIVED from the weakest input, NEVER asserted upward (incomplete env caps below byte-deterministic even on a perfect provider; declared provider nondeterminism caps at best-effort; missing versions ⇒ nonreplayable WITH reason), b=null version is MISSING not "latest", c=artifact access/ hash/lineage, d=environment manifest, e=provider support DECLARED not assumed, f=technical reproducibility is NOT permission (reproducible-but-forbidden proven), g=preview names every problem + unknown cost stays null. strict tsc + replay suite via tsx gated. generate-replay-eligibility.mjs (--check), 11 tests. -->

    • I1.12.b Require immutable brief/plan/workflow/graph/node/model/ checkpoint/adapter/LoRA/tool/template/policy/evaluator/corpus/provider versions or hashes. Evidence: version completeness tests.

    • I1.12.c Require all source/reference/conditioning/mask/latent/ intermediate/artifact bytes and ingredient lineage with hash/access/ retention. Evidence: artifact completeness tests.

    • I1.12.d Require exact seed/subseed/random generator/algorithm/parameters/ precision/hardware/runtime/container/driver/environment/dependency/ locale/time inputs where applicable. Evidence: environment manifest.

    • I1.12.e Require provider operation/idempotency/replay support or classify provider nondeterminism/version unavailability honestly. Evidence: provider replay tests.

    • I1.12.f Validate current authorization/rights/consent/policy/budget/ safety and distinguish technical reproducibility from permitted replay. Evidence: eligibility tests.

    • I1.12.g Produce missing/inaccessible/expired/revoked/incompatible/ nondeterministic dependencies and expected tolerance/cost before replay. Evidence: replay preview.

    • I1.12.h Execute as a new linked run with same inputs/seed where eligible and compare bytes/raw metrics/semantic tolerances/provenance/cost. Evidence: replay known-value tests. <!-- EXECUTING A GENERATION DOES NEED A PROVIDER, AND THE TWO QUESTIONS BEFORE IT DO NOT. THE CELL STAYS [ ] ON BOTH.

      evidence/v1-workbenches/isis-replay-execution.json —
      generate-isis-replay-execution.mts + .test.mts (8 tests, 3 ratchets, 3
      controls), wired into inventory + verify.
      
      THERE IS NO REPLAY IN ISIS THAT PRODUCES A NEW RUN. The one operation
      Isis calls a replay is `replayDeadLetterEntry`, and its body is parsed
      rather than described: it UPDATEs the same outbox row and the same job
      row back to `pending`. No INSERT, no new id, no link. IT IS A
      RE-DISPATCH OF THE ORIGINAL, WHICH IS THE OPPOSITE OF A NEW LINKED RUN —
      and on the way it sets `attempts = 0`, `last_error = NULL` and
      `dead_lettered_at = NULL`, so the first attempt's terminal detail does
      not survive it. (`replay_count`, `last_replayed_at` and
      `last_replay_reason` do.) That is a defensible operational action; it is
      simply not what this cell asks for.
      
      A REAL LINKED REPLAY EXISTS IN A DIFFERENT DOMAIN.
      `@oshun/content-service`'s `replay(runId)` re-submits the brief and
      records `replayOf` on a separately persisted run. So the estate knows
      the shape. Two caveats measured from its own body: it re-submits with a
      MODIFIED brief id (`${briefId}__replay`), so "same inputs" is false by
      construction in one field, and it compares nothing — it returns the new
      record.
      
      THE ELIGIBILITY MODEL I1.12.a-g BUILT IS GOOD, AND IT WAS EXECUTED. A
      complete manifest on a deterministic provider classifies as
      `byte-deterministic`; a missing model version gives `nonreplayable`; a
      declared-nondeterministic provider caps at `provider-best-effort`. The
      class really is derived from the weakest input.
      
      BUT ITS PREVIEW SAYS WHETHER, NEVER WHAT. `buildReplayPreview` returns
      `[classification, eligible, estimatedCost, expectedTolerance, permitted,
      problems, technicallyReproducible]`. Two of those are THE CALLER'S OWN
      ARGUMENTS HANDED BACK — `tolerance` and `estimatedCost` are parameters —
      and none of them is a comparison plan or a list of the dimensions a
      replay would be checked against. The cell names five dimensions and the
      preview names none of them.
      
      AND A COMPARATOR IS ONLY ON THE REPLAY PATH IF THERE IS A REPLAY PATH.
      Scoped to `apps/isis`, `libs/isis` and the generation control plane,
      there are 32 exported comparator functions; matched against each
      dimension BY FUNCTION BODY, bytes has 3, raw metrics 1, provenance 1,
      cost 1 and semantic tolerances none. Not one of them can run on a replay
      of an Isis generation, because Isis produces no replay run for them to
      read.
      
      TWO NARROWINGS THE MEASUREMENT NEEDED. Matching a dimension against the
      whole FILE credited every comparator that happens to live beside the
      word "score" — 561 for one dimension, which measures nothing; the match
      is now against the function body. And `[Dd]iff` matched
      `adjustDifficulty`; it now requires the next character not to be
      lowercase. -->
      <!-- What remains genuinely provider-bound: executing a replay and
      comparing the actual bytes. That is one step, and it is the last one. -->
      
  • I1.13 Define artifact promotion across development, test, staging, and production. Evidence: no environment can consume an inadmissible state.

    • I1.13.a Define environment-specific artifact states, stores/regions/ encryption/credentials, allowed producers/consumers, and immutable identity/hash. Evidence: environment policy. <!-- evidence (I1.13.a-e/g/h): COMPILING reference evidence/v1-workbenches/promotion/*.ts (promotion-schema.ts ENVIRONMENTS(4)/ArtifactState(8)/CONSUMABLE_STATES/ CONSUMPTION_CHANNELS(9)/TEST|STAGING|PRODUCTION_PREREQUISITES(11/9/9)/LifecycleOp(7); admission.ts canConsume/promote/applyLifecycleOp; fixtures.ts + check.ts). a=EnvironmentPolicy (per-env store/region/encryption/credentialScope, never shared), b=dev artifact refused by EVERY non-dev env x EVERY channel, c/d/e=test/staging/ production gates w/ each unmet prerequisite NAMED (prod requires humanSignoff), g=all 7 lifecycle ops leave the artifact inadmissible on every channel, h=ONE admission door for all 9 channels incl. direct-id (no bypass) + concurrent source-mutation detection via observed-hash. Promotion is SEQUENTIAL, mints a NEW identity w/ lineage, and REFUSES in-place relabeling + env-skipping. strict tsc + promotion suite via tsx gated. generate-artifact-promotion.mjs (--check), 12 tests. -->

    • I1.13.b Define development intake/iteration artifacts and explicit prohibition from test/staging/production consumption without promotion. Evidence: dev isolation tests.

    • I1.13.c Define test promotion prerequisites for source/hash/schema/ malware/security/license/rights/consent/provenance and reproducible fixture/evaluation. Evidence: test gate.

    • I1.13.d Define staging promotion prerequisites for approved model/ workflow, representative quality/safety/compatibility/cost/load, target bundle, and release candidate. Evidence: staging gate.

    • I1.13.e Define production promotion prerequisites for fresh gates, human/signoff, provenance/signature, rights/consent, artifact lock, rollback/recall, and target readiness. Evidence: production gate.

    • I1.13.f Copy/replicate exact verified bytes using target-managed credentials, rehash/read-back, record source/destination identity and promotion lineage, and prohibit in-place status relabeling. Evidence: promotion integration tests. <!-- SUPERSEDED (2026-08-16). The old note said this needs "the running service + real object stores"; the dev stack HAS a real object store, and the run below used two buckets of it. It also said "the CONTRACT half is done: promote() records source/destination identity + lineage" — HALF OF THAT IS WRONG. promote() records a source artifact id and a target ENVIRONMENT. No destination store, bucket, key or region appears anywhere on the artifact it produces, and the artifact interface has no field that could hold one.

      ZERO OF THE SEVEN OBLIGATIONS ARE PERFORMED BY THE PRODUCT.
      
      evidence/v1-workbenches/isis-promotion-transfer.json —
      generate-isis-promotion-transfer.mts + .test.mts (12 tests, 4 ratchets,
      3 controls), `--check` and the tests wired into verify. Each obligation
      is anchored on a phrase that must still appear in the cell.
      
      THREE SUBJECTS, AND ONLY ONE OF THEM IS THE PRODUCT. The reference
      `evidence/v1-workbenches/promotion/*.ts` that I1.13.a–e/g/h cite; the
      estate's own `S3StorageClient`; and `apps/isis` + `libs/isis`. All three
      were measured and two were executed.
      
      THE REFERENCE IMPLEMENTS TWO OBLIGATIONS AND CANNOT EXPRESS FIVE.
      Executed: it mints a new identity carrying `sourceArtifactId`, and it
      REFUSES a promotion that reuses the same id with the message "in-place
      status relabeling is prohibited". Both real. But `promote(artifact,
      target, evidence, newArtifactId, atTime)` takes no bytes, no store and
      no credentials, so copying, rehashing, reading back and naming a
      destination are outside it by construction rather than merely undone.
      And it CARRIES THE SOURCE CONTENT HASH FORWARD UNCHANGED — the comment
      beside it reads "the SAME verified bytes", which is an assertion about
      bytes the function never sees.
      
      THE ESTATE'S CLIENT CAN CROSS BUCKETS AND DOES NOTHING ELSE THE CELL
      ASKS. Executed against MinIO with two buckets: `copy()` moved the object
      into the target bucket and returned an etag. It performs no download and
      computes no digest — the probe had to download and rehash itself to
      learn the destination bytes matched. And it issues ONE
      `CopyObjectCommand` from the destination client naming the source bucket
      in `CopySource`, so THE SAME CREDENTIAL PAIR READS THE SOURCE AND WRITES
      THE DESTINATION and a target-managed transfer is not expressible. The
      estate reads one `MODELS_S3_BUCKET` and one AWS key pair; the
      reference's `EnvironmentPolicy` declares a store and a credential scope
      per environment.
      
      NO ISIS PRODUCT FILE COPIES AN ARTIFACT BETWEEN STORES, and no product
      code imports the reference — so the two obligations that ARE implemented
      are implemented in a module nothing runs.
      
      A NEAR-MISS THAT FILLED FOUR COLUMNS. The first cut looked for `.copy(`
      and matched `buffer.copy(` across eight 3D-geometry modules, reporting
      four of the seven obligations as performed by the product — the opposite
      of the finding. A site now counts only if it issues `CopyObjectCommand`
      or calls `.copy(` in a module that holds a storage client, with a
      control pinning that the collision source is rejected and that a real
      storage copy is still found. -->
      
    • I1.13.g Define reject/quarantine/rollback/demote/revoke/recall/expire and dependent workflow/run/release/cache behavior per environment. Evidence: lifecycle tests.

    • I1.13.h Test environment credential, URL/ref, API/provider, storage, cache/search, job, and direct-ID bypass plus concurrent source mutation. Evidence: inadmissible consumption negative tests.

  • I1.14 Define provenance aggregation for multi-step workflows. Evidence: nested executions produce one verifiable release lineage.

    • I1.14.a Define workflow/parent-run/child-run/stage/action/artifact/ ingredient/transform relationships and stable ordering/branch/fan-in/ retry identity. Evidence: aggregation graph schema. <!-- evidence (I1.14.a-g): COMPILING reference evidence/v1-workbenches/provenance- aggregation/*.ts (aggregation-schema.ts RunNodeKind(5)/IngredientDisposition(9)/ AggregationIssueKind(4)/TrustStatus(4); aggregate.ts validateStageProvenance/ aggregateIngredients/composeAssertions/bindManifest/composeTrust/queryLineage; fixtures.ts + check.ts). a=RunNode w/ order+branch+ATTEMPT (retry is a distinct node), b=stage provenance validated BEFORE aggregation (unknown artifact refs, hash mismatch, missing provider/tool/model/template/policy/actor, out-of-graph run all rejected), c=NO FLATTENING (same artifact reused AND duplicated = 2 rows; all 9 dispositions survive; selected vs considered distinguishable), d=field-level sourceRunId + CONFLICT names every source (no silent winner) + redacted RECORDED, e=manifest binds final bytes + child manifests BY HASH (not inlined), f=HEADLINE: composeTrust = worst child status, parent claim preserved + parentClaimOverridden flagged (a trusted parent CANNOT hide a child failure), g=cycle-safe traversal w/ unauthorized nodes REDACTED BY NAME. strict tsc + aggregation suite via tsx gated. generate-provenance-aggregation.mjs (--check), 13 tests. -->

    • I1.14.b Ingest each child/stage raw provider/tool/model/template/policy/ actor/input/output provenance and validate hashes/references before aggregation. Evidence: stage provenance tests.

    • I1.14.c Preserve shared/reused/duplicate/optional/failed/retried/ compensated/manual edited ingredients and exact selected outputs without flattening lineage. Evidence: complex workflow fixtures.

    • I1.14.d Compose ordered action/ingredient/tool/actor/policy assertions with field-level source and identify conflicts/missing/unsupported/ redacted data. Evidence: aggregation tests.

    • I1.14.e Bind aggregate manifest/ledger record to final artifact/bundle bytes and include child manifest/hash refs rather than unbounded content. Evidence: binding tests.

    • I1.14.f Sign/validate aggregate independently while preserving invalid/ untrusted/revoked child status and preventing a trusted parent from hiding child failure. Evidence: trust composition tests.

    • I1.14.g Traverse release→final→stages→ingredients and ingredient→all derivatives/release with authorization/redaction/cycle/scale handling. Evidence: lineage query tests.

    • I1.14.h Verify rebuild/round-trip/tamper/stripped recovery and exact release eligibility based on complete current lineage. Evidence: end-to-end provenance report. <!-- SUPERSEDED (2026-08-16). The old note said a rebuild needs "the running service + real bytes". The estate's provenance ledger is a PURE hash-chained module — buildIsisProvenanceBundle, appendProvenanceLedger, verifyProvenanceLedger, buildProvenanceChainFingerprint in libs/isis/ai-providers/src/provenance/bundle.ts — so all four verbs are pure-function questions and all four were executed against the real implementation.

      THREE OF THE FOUR VERBS PASS, AND THE CELL STAYS `[ ]` ON THE FOURTH AND
      ON ELIGIBILITY.
      
      evidence/v1-workbenches/isis-provenance-recovery.json —
      generate-isis-provenance-recovery.mts + .test.mts (9 tests, 4 ratchets,
      3 controls), wired into inventory + verify.
      
      THE TAMPER DETECTION IS REAL AND SHOULD BE SAID SO. Executed: a bundle's
      fingerprint is recomputed from its fields, and one edited field is
      refused with `fingerprint does not match bundle fields`. A JSON
      round-trip of a three-entry ledger still verifies. An EDITED entry
      breaks at index 1, a TRUNCATED ledger at index 2 and a REORDERED ledger
      at index 0 — each caught, each located — while the untouched ledger
      verifies. `verifyProvenanceLedger` also enforces tenant coherence,
      one-generation-once and monotonic emission time.
      
      STRIPPED RECOVERY IS THE VERB THAT FAILS, AND IT FAILS ON A KEY. The
      persisted `isis_provenance_bundles` table declares fifteen columns and
      NOT ONE names an artifact, an output, a content hash or a storage key;
      the store offers exactly two lookups, `loadLedger(tenantId)` and
      `verifyPersistedLedger(tenantId)`. An artifact that arrives with its
      embedded provenance stripped therefore has nothing to be matched
      against: the ledger is keyed by tenant and generation id, and the
      generation id is precisely what stripping removes.
      
      AND THE VERDICT GATES NOTHING. `buildIsisReleaseReadiness` decides its
      provenance requirement with `input.expectNewOutput ? true :
      Boolean(input.outputBundle?.attachedSignals.includes('provenance'))` —
      a flag set when a provenance RECORD WAS SUPPLIED, never from whether a
      chain verified. Executed: a bundle claiming the signal is release-ready
      with an empty `missingRequirements`, and no ledger, fingerprint or
      verification verdict appears anywhere in the expression. The positive
      control matters here: removing the signal DOES block with
      `provenance_missing`, so the gate is not inert — it is reading the wrong
      thing. THE CELL ASKS FOR ELIGIBILITY BASED ON COMPLETE CURRENT LINEAGE
      AND THE GATE READS A CLAIM THAT ONE EXISTS.
      
      The one consumer of `verifyProvenanceLedger` outside its own module is
      `apps/oshun/web/src/lib/server/isis-provenance-loader.ts`, and it does
      not branch on the verdict — it spreads it into a returned object for an
      operator page to display. That is MEASURED: the binding is found by
      name and the branch is looked for on that binding. A first cut called
      any file containing a `throw` a gate, which credited the loader with
      refusing something it does not refuse.
      
      ALSO WORTH KEEPING IN VIEW: two libraries export a function called
      `buildIsisProvenanceBundle`. The Isis one builds the hash-chained
      record; the generation-control one builds the presentation summary
      carrying `attachedSignals`. The release gate reads the second. -->
      
  • I1.15 Add canonical contract fixtures for every asset kind. Evidence: fixtures parse through public contracts with no any or payload gaps. <!-- evidence: COMPILING reference evidence/v1-workbenches/asset-contracts/*.ts (contracts.ts 11-kind discriminated union + branded units Pixels/Millimeters/ Seconds/Hertz/Lufs/Degrees/Ratio01 + typed AssetRef/EvidenceRef; parse.ts public parseAsset/serializeAsset/roundTrips w/ per-kind field spec; fixtures.ts 11x5=55 fixtures; coverage.ts buildCoverageReport = the I1.15.l gate; check.ts). a=image b=portrait c=texture d=video e=audio f=foley g=voice h=mesh i=rig j=splat k=scene, each w/ minimal/full/boundary (parse + byte-stable round-trip) + invalid (refused, specific field named) + version-skew (refused as skew, NEVER migrated). PARENT BAR MET: no any and no index-signature payload bag (comment-stripped source scan); an unexpected field is REFUSED ("no unknown-payload bag") not absorbed; a missing required field is refused BY NAME not defaulted. strict tsc + fixture suite via tsx gated. generate-asset-contract-fixtures.mjs (--check), 12 tests. -->

    • I1.15.a Add minimal/full/invalid/boundary/version-skew image brief/plan/ run/output/lineage/gate/release fixtures. Evidence: public contract parse/round-trip tests.
    • I1.15.b Add portrait fixtures covering identity/reference/consent/ biometric/impersonation and missing/revoked evidence states. Evidence: portrait contract tests.
    • I1.15.c Add texture fixtures covering PBR channels/UDIM/tiling/color/ material target and unsupported channel/profile. Evidence: texture tests.
    • I1.15.d Add video fixtures covering frame/time/audio/caption/motion/ codec/delivery and partial/truncated/variable-rate states. Evidence: video tests.
    • I1.15.e Add audio fixtures covering sample/channel/duration/loudness/ transcript/codec/delivery and nonvalue/error states. Evidence: audio tests.
    • I1.15.f Add foley fixtures covering event/time/variation/perspective/ sync/loop/package and rights/quality states. Evidence: foley tests.
    • I1.15.g Add voice fixtures covering speaker/consent/language/SSML/ pronunciation/timing/disclosure/caption and revoke states. Evidence: voice tests.
    • I1.15.h Add mesh fixtures covering units/axes/geometry/topology/UV/ materials/LOD/collision/formats and invalid/unsupported cases. Evidence: mesh tests.
    • I1.15.i Add rig fixtures covering hierarchy/joints/skin/controls/ constraints/retarget/target formats and deformation failure. Evidence: rig tests.
    • I1.15.j Add Gaussian-splat fixtures covering coordinates/cameras/ positions/scales/rotations/opacity/SH/compression/point count/privacy and version failures. Evidence: splat tests.
    • I1.15.k Add composite-scene fixtures covering hierarchy/refs/variants/ cameras/lights/materials/animation/audio/dependencies/bundles and partial component states. Evidence: scene tests.
    • I1.15.l Assert every fixture uses branded typed fields, exact units/ discriminators/versions/refs, no any/unknown payload bag, public parse/ serialize/round-trip, and expected invalid refusal. Evidence: asset-kind fixture coverage gate.

I2 — Brief, prompt, workflow, and variant authoring#

  • I2.1 Implement deterministic brief-to-prompt-spec compilation. Evidence: known briefs compile to exact structured specs. <!-- evidence: COMPILING reference evidence/v1-workbenches/prompt-spec/*.ts (prompt-spec.ts versioned PromptSpec — system/policy separated, NO free-form prompt/instruction field, fenced references; compile.ts validateInput/compileBrief/ serializeSpec/recompilesIdentically/leaksRestrictedDetail; fixtures.ts 6 canonical + 10 negative briefs; goldens.ts FROZEN golden specs; check.ts). a=schema, b=validate+canonicalize (gapless ref order, 0..1 weights, approved vocab/locale/ extensions), c=approved-only resolution + full mapping trace (brief/default/derived/ omitted/unsupported — a default is visible AS a default), d=host-independent serialization (sorted keys, NFC, NO Date/Math.random/localeCompare), e=untrusted text stays DATA + tier restriction — the TRACE ITSELF redacts restricted identities (5/6 goldens unauthorized-tier carry ZERO template/model/provider strings; the 1 authorized golden legitimately does), f=3 structured error kinds, NEVER a concatenation fallback, g=6 canonical briefs → EXACT frozen goldens + byte-identical recompile (the parent's bar). strict tsc + compiler suite via tsx gated. generate-prompt-spec-compiler.mjs (--check), 12 tests. -->
    • I2.1.a Define versioned PromptSpec schema with separated system/policy, positive intent, exclusions, references/conditioning, spatial/temporal/ audio controls, output/delivery constraints, locale, and extensions. Evidence: schema fixtures.
    • I2.1.b Validate and canonicalize exact brief revision, asset/workflow/ audience profile, locale, units, ordered refs, typed controls, and policy/code floors before compile. Evidence: compiler input tests.
    • I2.1.c Resolve only versioned approved templates/vocabularies/mappings and record every source field/default/derived/omitted/unsupported decision. Evidence: mapping trace.
    • I2.1.d Serialize normalized spec in stable field/list/number/string/ Unicode order independent of host locale/time/random/process ordering. Evidence: determinism tests.
    • I2.1.e Keep untrusted user/reference content fenced as data and prohibit compiled spec from granting workflows/models/tools/provider detail outside audience. Evidence: injection/tier tests.
    • I2.1.f Return structured validation/unsupported/policy errors without falling back to generic prompt concatenation. Evidence: negative tests.
    • I2.1.g Compile canonical minimal/full/multilingual/multireference/ multimodal briefs to exact golden specs and recompile byte-identically. Evidence: golden compiler suite.
  • I2.2 Keep positive intent, exclusions, reference controls, spatial controls, temporal controls, audio controls, and delivery constraints as typed fields. Evidence: no required control is hidden in free-form text. <!-- evidence: COMPILING reference evidence/v1-workbenches/typed-controls/*.ts (controls.ts 7 typed blocks + LINTED_SURFACES(5) + HiddenControlViolation; lint.ts lintControlSet/lintFreeText/hiddenOnly/duplicatesOnly w/ 9 detectors; fixtures.ts FULLY_TYPED + AV_TYPED + 7 HIDDEN_CASES + DUPLICATE_CASE; check.ts). a=PositiveIntent (stable vocabularyVersion + namespaced domainBlocks), b=Exclusion (category/scope/SEVERITY/SOURCE — code-floor vs user distinguishable), c=ReferenceControl (order/version/hash/regionMask/influenceSchedule/preprocessing/ rights/consent/conflictBehavior), d/e/f/g=Spatial/Temporal/Audio/Delivery. PARENT BAR: prose only in notes declared NON-AUTHORITATIVE; the lint covers ALL 5 surfaces (compiler/ui/contract/import/export + notes itself) and separates a genuinely HIDDEN control (typed field EMPTY ⇒ lint FAILS) from a DUPLICATE (typed ⇒ reported, not fatal), quoting the offending text. strict tsc + controls suite via tsx gated. generate-typed-controls.mjs (--check), 11 tests. -->
    • I2.2.a Define typed positive intent for subject/content/action/context/ style/tone/composition/required elements and domain blocks with stable vocabularies/extensions. Evidence: intent field tests.
    • I2.2.b Define typed exclusions for content/objects/styles/attributes/ safety/brand/rights/technical constraints with scope/severity/source. Evidence: exclusion tests.
    • I2.2.c Define typed ordered reference roles, weights, strengths, regions/masks, influence schedule, preprocessing, versions/hashes, rights/consent, and conflict behavior. Evidence: reference control tests.
    • I2.2.d Define typed canvas/coordinate/units/aspect/resolution/crop/ layout/position/depth/pose/camera/lens/light/scene spatial controls. Evidence: spatial tests.
    • I2.2.e Define typed frame/sample/beat/timecode units, duration/rate/ sequence/shot/timing/motion/transition/loop/sync temporal controls. Evidence: temporal tests.
    • I2.2.f Define typed sample/channel/codec/loudness/peak/dialogue/voice/ pronunciation/music/SFX/foley/stem/mix/sync audio controls. Evidence: audio tests.
    • I2.2.g Define typed target/host/channel/environment/territory/locale/ format/profile/dimensions/quality/naming/package/deadline/embargo/ publisher constraints. Evidence: delivery tests.
    • I2.2.h Prevent required control names/values from appearing only in free-form prompt fields; lint compiler/UI/contracts/import/export. Evidence: deliberate hidden-control failures.
  • I2.3 Validate reference media kind, rights, checksum, dimensions, duration, codec, and moderation before planning. Evidence: invalid inputs never reach a provider. <!-- evidence: COMPILING reference evidence/v1-workbenches/reference-validation/*.ts (reference-schema.ts StoredReference w/ NO url field + IssueKind(5) + ScanKind(8) + call-counted ProviderSeam; validate.ts resolveReference/checkKind/checkRights/ checkChecksum/checkDimensions/checkTiming/checkCodec/checkScans/planWithReferences; fixtures.ts 17 bad refs + fnv1a hasher + makeProvider; check.ts). a=access/state/ retention + no mutable-URL path, b=SNIFFED kind authoritative over declared, c=rights/consent freshness+revocation, d=REAL recompute-and-compare (a flipped byte quarantines), e=dimensions, f=timing/sync, g=codec/bit-depth, h=scans must exist+PASS +be fresh+be BOUND TO THE SAME HASH. PARENT BAR (i): provider reachable ONLY via the counted seam; 17 bad refs x (alone + mixed w/ a valid ref) = 34 ZERO-PROVIDER-CALL assertions; every refusal names ref+check+kind. strict tsc + validation suite via tsx gated. generate-reference-validation.mjs (--check), 11 tests. -->
    • I2.3.a Resolve immutable authorized reference bytes/metadata and verify tenant/project/actor/property access, state, retention, and no mutable URL-only dependency. Evidence: reference resolver tests.
    • I2.3.b Sniff actual media/container/asset kind and verify it is supported by declared role, asset kind, workflow class, model/provider, and target. Evidence: kind mismatch tests.
    • I2.3.c Verify license/ownership/attribution/derivative/training/ territory/term and identity/voice/biometric consent evidence/freshness/ revocation. Evidence: rights/consent tests.
    • I2.3.d Rehash stored bytes with approved algorithm and compare metadata/ provenance/manifest; quarantine corruption. Evidence: checksum spoof tests.
    • I2.3.e Validate image/texture/3D dimensions/units/axes/resolution/point/ polygon/texture/channel limits and role alignment. Evidence: dimension boundary tests.
    • I2.3.f Validate video/audio/animation duration/frame/sample/beat/time base/rate/track/channel/synchronization limits. Evidence: duration tests.
    • I2.3.g Validate codec/container/profile/level/bit depth/color/HDR/ compression/3D format/extensions and decode/parser resource limits. Evidence: codec/format tests.
    • I2.3.h Require current malware/content/child/identity/safety/moderation/ archive/domain scans bound to the same hash. Evidence: scan freshness tests.
    • I2.3.i Stop compilation/planning/provider calls and return exact invalid/unsupported/unauthorized/stale/quarantined reason. Evidence: zero-provider-call negative assertions.
  • I2.4 Implement workflow-template selection by capability and audience. Evidence: tier-forbidden workflows never appear or dispatch.
    • I2.4.a Query approved active workflow versions by asset kind/operation/ required capabilities/input-output/profile/target and current dependency compatibility. Evidence: candidate query tests. <!-- evidence (I2.4.a-f): COMPILING reference evidence/v1-workbenches/workflow- selection/*.ts (selection-schema.ts WorkflowCandidate + CandidateSummary/Detail + RejectionCode; select.ts queryCandidates/filterByTier/filterEligibility/ rankCandidates/selectWorkflow/projectFor/dispatchWorkflow; fixtures.ts 19 deliberate rejections; check.ts). a=capability/IO/profile/target/dependency query, b=tier/host/ entitlement/tenant-policy/classification exposure, c=lifecycle+provider+resource+ budget+rights+safety (deprecated-in-window stays eligible), d=deterministic rank w/ STABLE id/version tie-break (order-independent), e=curated tiers get a Summary type that STRUCTURALLY lacks graph/nodes/model/provider (a leak = TYPE ERROR) + serialized -view secret scan, f=explicit choice runs the FULL chain so a forbidden id is REJECTED, and dispatch RE-VALIDATES (revoked-after-planning RACE refused); operator tier still allowed. strict tsc + selection suite via tsx gated. generate-workflow-selection.mjs (--check), 11 tests. -->

    • I2.4.b Filter by canonical actor tier, audience/host allowlist, entitlement, tenant/project policy, classification, and workflow exposure. Evidence: tier filter tests.

    • I2.4.c Filter revoked/deprecated-expired/quarantined/incompatible/ unavailable/not-configured/provider/resource/budget/rights/safety candidates with exact reasons. Evidence: eligibility tests.

    • I2.4.d Rank/select deterministically by policy, compatibility, quality, latency, cost, capacity, target, and explicit user choice with stable tie-break and reason. Evidence: selection fixtures.

    • I2.4.e Return allowed candidate summaries/details appropriate to audience and redact graph/node/model/provider internals from curated tiers. Evidence: property leak tests.

    • I2.4.f Revalidate exact workflow/version/audience/host/dependencies at planning and dispatch; reject direct forbidden IDs. Evidence: bypass/ race tests.

    • I2.4.g Emit capability/ineligibility state to UI/API and audit selected/ rejected workflow decisions. Evidence: end-to-end tests. <!-- PRODUCING A DECISION AND EMITTING IT ARE DIFFERENT CLAIMS. THE CELL STAYS [ ] ON BOTH HALVES.

      evidence/v1-workbenches/isis-workflow-selection-emission.json —
      generate-isis-workflow-selection-emission.mts + .test.mts (9 tests, 3
      ratchets, 3 controls), wired into inventory + verify.
      
      THE PRIOR NOTE WAS RIGHT ABOUT THE CHAIN AND THAT IS WHAT MAKES THE REST
      A FINDING. Executed: selecting for a curated actor over the reference's
      25 candidates chose one workflow and produced 20 rejections across 19
      distinct codes, and `projectFor` gave the curated actor a summary
      carrying `[assetKind, capabilities, operation, profile, version,
      workflowId]` while the operator's carries the graph, the nodes, the
      model and the provider. The state exists.
      
      NOT ONE OF THE TWENTY-ONE REJECTION CODES IS EMITTED AS A WORKFLOW
      DECISION. Three of the strings do appear on the Isis surface for other
      subjects — `unavailable` as a status across three dozen stores and
      pages, `quarantined` in the 3D quality gates, `revoked` in a compliance
      report — and none of those files carries a workflow selection anywhere.
      The other eighteen, including `tier-not-exposed`, `entitlement-missing`,
      `classification-too-high` and `rights-uncleared`, appear on the Isis
      surface not at all. No API returns a candidate list with reasons and no
      UI renders one, because there is no vocabulary in the product to render.
      
      AND THE AUDIT NAMES NO DECISION. The workflow registry writes twelve
      actions into `workflow_audit_logs`: created, deleted, updated,
      workflow.created, workflow.deleted, workflow.restored,
      workflow.template.used, workflow.updated, workflow.version.created,
      workflow.version.deleted, workflow.version.deprecated,
      workflow.version.restored. EVERY ONE IS A MUTATION OF THE WORKFLOW
      RECORD. `workflow.template.used` records that one was used, never that
      others were considered and refused — which is exactly the pair this cell
      asks to be audited.
      
      TWO NARROWINGS THE MEASUREMENT NEEDED, AND BOTH CHANGED THE ANSWER.
      Counting any file carrying the string reported TWELVE of twenty-one
      codes as emitted, on the strength of `'revoked'` in an Asase API-auth
      module and `'over-budget'` in a localisation check; a carrier must be on
      the Isis surface. Then file-level co-occurrence with selection
      vocabulary still counted `candidateVersion` in a 3D rollback policy and
      `searchCandidates` in a Civitai intake store; the workflow and the
      candidate must now appear on the SAME LINE. Both collision sources are
      pinned by tests, so the narrowing cannot silently drift back. -->
      <!-- The reference is a specification: nothing under apps/ or libs/
      imports evidence/v1-workbenches/workflow-selection/, which the artifact
      measures. -->
      
  • I2.5 Implement model selection through the governed registry. Evidence: quarantined, incompatible, unlicensed, or environment-inadmissible models are unavailable with reasons. <!-- evidence: evidence/v1-workbenches/model-selection/ (compiling reference, strict tsc + runnable suite) and model-selection.json via node scripts/v1-workbenches/generate-model-selection.mjs. All four states the parent names are distinct typed reason codes with deliberate fixtures (19 refusals); every declared code is reachable. -->
    • I2.5.a Query models/versions/checkpoints/adapters by required workflow/ node/task/modality/asset/profile/input-output/provider/runtime/hardware capabilities. Evidence: candidate query tests.
    • I2.5.b Require approved active lifecycle, verified artifact hash/source/ signature/security scans, current evaluation, and nonrevoked state. Evidence: lifecycle tests.
    • I2.5.c Require compatible architecture/version/precision/format/context/ adapter/LoRA/node/workflow/provider/endpoint/resource and target. Evidence: compatibility tests.
    • I2.5.d Require license/usage/training/derivative/attribution/territory/ commercial rights and identity/content policy for actor/audience/use. Evidence: license tests.
    • I2.5.e Require environment promotion/admission, regional availability, entitlement/quota/capacity/budget, and target data residency. Evidence: environment tests.
    • I2.5.f Rank allowed candidates deterministically by policy/quality/ latency/cost/capacity/reproducibility/target and record alternatives/ rejection reasons. Evidence: selection fixtures.
    • I2.5.g Expose audience-safe metadata and typed reasons for quarantined/ incompatible/unlicensed/inadmissible/not-configured/unavailable without leaking private model/provider detail. Evidence: UI/API tests.
    • I2.5.h Revalidate model/version/hash/licensing/evaluation/promotion/ endpoint at dispatch and block stale direct references. Evidence: race tests.
  • I2.6 Implement LoRA/modifier selection only in authorized AAA/operator hosts. Evidence: curated payload leak audit passes. <!-- evidence: evidence/v1-workbenches/modifier-selection/ (compiling reference, strict tsc + runnable suite) and modifier-selection.json via node scripts/v1-workbenches/generate-modifier-selection.mjs. The leak audit is type-level (curated contracts declare no modifier field) plus a scan with a negative control, and no curated probe can distinguish a real modifier id from a fictional one (zero registry calls). -->
    • I2.6.a Register LoRA/adapter/modifier ID/version/hash/source/base-model/ target-module/rank/trigger/strength range/license/safety/identity/style/ asset/workflow compatibility and lifecycle. Evidence: modifier schema.
    • I2.6.b Resolve operator and AAA actor/host/tier/entitlement/tenant/ project/use authorization at discovery and mutation. Evidence: authorization tests.
    • I2.6.c Filter approved/nonquarantined/compatible/licensed/environment- admitted/provider-supported modifiers and multi-modifier combinations. Evidence: eligibility tests.
    • I2.6.d Expose typed search/select/order/weight/schedule/mask/region/ enable/preset controls with ranges and effective influence diagnostics. Evidence: AAA/operator UI tests.
    • I2.6.e Revalidate exact modifier versions/hashes/combination/audience at planning/dispatch and persist effective provenance/parameters. Evidence: dispatch tests.
    • I2.6.f Omit modifier IDs/names/hashes/triggers/weights/schedules/base- model linkage and unsupported hints entirely from curated contracts/ routes/bundles/telemetry. Evidence: curated leak scan.
    • I2.6.g Reject direct API/query/body/deep-link/import attempts from curated/contemplative users without revealing existence. Evidence: negative tests.
  • I2.7 Implement sampler and advanced node controls only in authorized hosts. Evidence: route and property authorization tests. <!-- evidence: evidence/v1-workbenches/advanced-controls/ (compiling reference, strict tsc + adversarial suite) and advanced-controls.json via node scripts/v1-workbenches/generate-advanced-controls.mjs. Route authorization is the 2-host x 4-verb matrix; property authorization is the allow-list guard, which closes direct/alias/nested/unknown-field/ mass-assignment routes including a real own proto key. -->
    • I2.7.a Register sampler/scheduler/step/CFG/seed/denoise/precision/batch/ node/port/parameter/subgraph/custom-node controls with versions/types/ units/ranges/compatibility/risk/resource/cost. Evidence: control schema.
    • I2.7.b Assign operator versus AAA view/edit/import/execute rights per workflow class/node/control and prohibit curated/contemplative exposure. Evidence: permission matrix.
    • I2.7.c Build typed controls with validation/default provenance/ dependency constraints/effective values/cost-quality-latency impact and no arbitrary JSON/code. Evidence: UI/API tests.
    • I2.7.d Enforce node catalog/signature/license/security/sandbox/network/ filesystem/resource policy and graph compatibility. Evidence: custom node security tests.
    • I2.7.e Revalidate actor/host/workflow/version/node/parameter/policy/ capacity at plan/dispatch and persist effective graph/controls. Evidence: dispatch tests.
    • I2.7.f Remove/redact controls and raw values from lower-tier responses, errors, URLs, imports/exports, provenance summaries, logs, and capability hints. Evidence: leak tests.
    • I2.7.g Test route, object/property/function authorization and direct/ nested/alias/unknown-field/mass-assignment bypass. Evidence: adversarial authorization suite.
  • I2.8 Implement best-of-N variant requests with estimated cost and resource reservation. Evidence: dispatch requires explicit confirmation above the configured threshold. <!-- evidence: evidence/v1-workbenches/best-of-n/ (compiling reference, strict tsc + runnable suite) and best-of-n.json via node scripts/v1-workbenches/generate-best-of-n.mjs. The confirmation carries a digest of the estimate it approved, so a token issued for one estimate cannot authorize a changed one (count grew, or a unit price moved); all four threshold kinds are proven reachable. -->
    • I2.8.a Define requested count, candidate count, diversity strategy, selection strategy, seed policy, per-variant controls, and hard maximums as typed fields. Evidence: best-of-N contract tests.
    • I2.8.b Expand one request into deterministic candidate plans with unique run/output identities and shared request/brief lineage. Evidence: expansion fixtures.
    • I2.8.c Estimate compute, provider, storage, egress, evaluation, and post-processing cost for the complete candidate set before reservation. Evidence: estimate tests.
    • I2.8.d Reserve quota, budget, queue, GPU/CPU/VRAM, and storage for the confirmed set atomically or return an exact shortfall. Evidence: reservation contention tests.
    • I2.8.e Require typed confirmation when configured count/cost/resource/ latency thresholds are crossed and bind it to the unchanged estimate. Evidence: stale-confirmation tests.
    • I2.8.f Dispatch only confirmed/reserved candidates with bounded fan-out, cancellation, partial-failure, retry, and reservation-release behavior. Evidence: execution race tests.
    • I2.8.g Reconcile actual per-variant and aggregate usage/cost, preserving failed/cancelled candidates and selection state. Evidence: reconciliation fixtures.
  • I2.9 Implement manual and generated variant grouping. Evidence: generated scores remain provider/gate values and manual variants show score absence. <!-- evidence: evidence/v1-workbenches/variant-grouping/ (compiling reference, strict tsc + runnable suite) and variant-grouping.json via node scripts/v1-workbenches/generate-variant-grouping.mjs. ManualVariant declares no generation-provenance or score field, so a fabricated run or score is a compile error; measurements keep provider/gate attribution and absence survives aggregation. -->
    • I2.9.a Define a stable variant-set identity, ordering, labels, roles, parent/revision links, source discriminator, and membership rules. Evidence: grouping schema tests.
    • I2.9.b Add authorized manual assets without inventing generation run, model, seed, provider, prompt, or automated score provenance. Evidence: manual variant fixtures.
    • I2.9.c Add generated assets only through immutable run/output lineage and retain provider measurements and gate results as attributed values. Evidence: generated variant fixtures.
    • I2.9.d Prevent manual edits, regrouping, imports, or API payloads from rewriting source identity or converting missing scores into zero/pass. Evidence: provenance mutation tests.
    • I2.9.e Enforce asset-kind/profile/comparison compatibility while preserving mixed-source groups explicitly allowed by policy. Evidence: compatibility tests.
    • I2.9.f Record membership, reorder, remove, merge, split, and archive decisions without deleting historical group state. Evidence: audit history tests.
  • I2.10 Implement side-by-side and A/B comparison with synchronized image/video/audio/3D controls. Evidence: selection and playback state are stable and accessible. <!-- evidence (a-e, h): evidence/v1-workbenches/comparison/ (compiling reference, strict tsc + runnable suite) and comparison.json via node scripts/v1-workbenches/generate-comparison.mjs. runtime: f and g name an ACCESSIBLE audit and keyboard-focus / screen-reader behaviour on a rendered surface, which no pure model can establish; their data models (NonvisualPanel, RestorableState) are built and asserted here so the runtime audit has something to run against. The parent stays unchecked because its bar ends "and accessible". -->
    • I2.10.a Define comparison session, participant, source revision, layout, synchronized state, annotation, and decision contracts. Evidence: comparison schema tests.

    • I2.10.b Implement image pan/zoom/fit/pixel/alpha/difference controls synchronized in common coordinates with deliberate unlinking. Evidence: image interaction tests.

    • I2.10.c Implement video play/pause/scrub/frame/timecode/rate/loop/audio controls with duration/rate mismatch policies. Evidence: video sync tests.

    • I2.10.d Implement audio play/pause/scrub/loop/gain/mute/channel/waveform/ loudness controls without auto-normalizing away measured differences. Evidence: audio comparison tests.

    • I2.10.e Implement 3D camera/orbit/pan/zoom/light/background/material/ animation/LOD/wireframe controls with unit/axis/bounds reconciliation. Evidence: 3D comparison tests.

    • I2.10.f Present nonvisual metadata, provenance, raw gate values, thresholds, uncertainty, rights, and target compatibility alongside the media. Evidence: accessible comparison audit. <!-- WHAT IS BUILT IS THE PANEL, NOT THE PRESENTATION. THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-comparison-nonvisual.json —
      generate-isis-comparison-nonvisual.mts + .test.mts (7 tests, 3 ratchets,
      2 controls), wired into inventory + verify. The parent's note said no
      pure model can establish behaviour on a rendered surface. That is true
      of ACCESSIBILITY and not of PRESENCE: a page that never names
      provenance, a threshold or a licence does not render them, however it is
      rendered.
      
      THE MODEL IS GOOD AND WAS EXECUTED. The reference `NonvisualPanel`
      carries a provenance summary, gate readings, a rights summary and both
      the compatible AND the incompatible targets. Each `GateReading` carries
      the raw value, the threshold it was judged against and the uncertainty
      of the measurement together — a bare pass/fail would hide all three —
      and a missing measurement is a `null`, not a zero.
      
      FIVE OF THE SEVEN ITEMS ARE NAMED BY NO ISIS COMPARISON SURFACE. There
      are three: `OutputComparisonWorkspacePage`,
      `MeshTransformerAbComparisonPage` and
      `StudioIsisModelComparisonWorkspace`. Only two items match at all, and
      both loosely — `nonvisual metadata` on the generic words label, metadata
      and summary, and `target compatibility` on the model-comparison
      workspace. PROVENANCE, RAW GATE VALUES, THRESHOLDS, UNCERTAINTY AND
      RIGHTS APPEAR IN NONE OF THEM: the 4,969-line output comparison
      workspace mentions provenance, a threshold, uncertainty, rights and a
      licence exactly zero times each, which the test asserts word by word.
      And nothing under apps/ or libs/ imports the reference.
      
      A CONTROL THAT CHANGED THE ANSWER. Comparison pages exist all over the
      estate — Demeter compares gardens, Lilith compares research synopses,
      Hathor compares options — and auditing all 22 of them reported FIVE of
      the seven items as presented, by surfaces with nothing to do with
      generation output. The audit is scoped to the Isis surfaces, and story
      files and `__fixtures__` hosts are excluded because a fixture is not a
      surface a user reaches.
      
      NOT CLAIMED HERE: whether the panel, once rendered, is reachable by
      keyboard and announced by a screen reader. That is I2.10.g and it needs
      the browser harness. -->
      
    • I2.10.g Preserve keyboard focus, screen-reader labels, participant identity, synchronized state, and current selection across rerender, reconnect, and permitted deep link. Evidence: accessibility/state tests.

    • I2.10.h Commit A/B preference or multi-candidate selection only against the exact compared revisions and reject stale/unauthorized choices. Evidence: selection concurrency tests.

  • I2.11 Implement decision records for select, reject, park, or regenerate. Evidence: chosen and rejected output IDs remain linked. <!-- evidence: evidence/v1-workbenches/decision-records/ (compiling reference, strict tsc + runnable suite) and decision-records.json via node scripts/v1-workbenches/generate-decision-records.mjs. Linkage is enforced at write time (every compared candidate must appear as chosen or nonselected) and as a reachability invariant asserted over a four-decision select/reject/park/regenerate chain. -->
    • I2.11.a Define immutable decision identity, actor/authority, timestamp, compared set/revisions, disposition, reasons, evidence, comments, and supersession linkage. Evidence: decision schema tests.
    • I2.11.b Implement select with one or explicitly multiple chosen output revisions and retain all nonselected candidates. Evidence: selection tests.
    • I2.11.c Implement reject with typed reason codes, optional annotations, remediation, and no asset deletion. Evidence: rejection tests.
    • I2.11.d Implement park with owner, review trigger/date, hold reason, and restoration semantics. Evidence: parked-state tests.
    • I2.11.e Implement regenerate as a new request/run linked to selected source inputs, requested changes, and preserved prior output. Evidence: regeneration lineage tests.
    • I2.11.f Enforce actor/tier/project/state permissions and optimistic concurrency for each disposition. Evidence: authorization/race tests.
    • I2.11.g Surface current and superseded decisions in UI/API/export/audit without allowing later outcomes to rewrite history. Evidence: history projection tests.
  • I2.12 Implement prompt/version diff. Evidence: differences distinguish user brief, compiled spec, workflow, model, seed, and policy. <!-- evidence: evidence/v1-workbenches/version-diff/ (compiling reference, strict tsc + runnable suite) and version-diff.json via node scripts/v1-workbenches/generate-version-diff.mjs. Layer attribution is proven by single-layer fixtures asserting ONLY that layer moved; the policy-only verdict is the one certain attribution and its negative case deliberately flips a verdict AND changes the artifact. -->
    • I2.12.a Resolve two authorized immutable run/output/brief revisions and reject incomparable or redacted targets. Evidence: diff input tests.
    • I2.12.b Diff user-authored brief fields with semantic field/list/unit/ reference changes distinct from prose formatting. Evidence: brief diff fixtures.
    • I2.12.c Diff canonical compiled specs including derived/defaulted/ omitted/unsupported fields and compiler/template versions. Evidence: spec diff fixtures.
    • I2.12.d Diff workflow template/version/graph/control/dependency changes at the highest detail authorized for the audience. Evidence: workflow diff tests.
    • I2.12.e Diff model/checkpoint/adapter/modifier/provider/runtime changes with audience-appropriate redaction. Evidence: model diff tests.
    • I2.12.f Diff seed/randomness/candidate/conditioning and effective parameter changes without implying causation. Evidence: execution diff tests.
    • I2.12.g Diff policy/code/evaluator/rights/safety/target-profile versions and distinguish policy-only verdict changes. Evidence: policy diff tests.
    • I2.12.h Render added/removed/changed/unchanged/unknown/redacted states accessibly and export the same typed diff. Evidence: UI/API parity tests.
  • I2.13 Implement reusable brief presets with inheritance and code floors. Evidence: preset overrides cannot loosen safety or rights requirements. <!-- evidence: evidence/v1-workbenches/brief-presets/ (compiling reference, strict tsc + runnable suite) and brief-presets.json via node scripts/v1-workbenches/generate-brief-presets.mjs. Floors are re-applied after EVERY inheritance step and again after user overrides; a leaf that loosens all seven floor domains is clamped on every one, and each clamp records which step attempted it. -->
    • I2.13.a Define preset identity, scope, owner, audience, asset/workflow compatibility, version, lifecycle, typed values, parent, and code-floor references. Evidence: preset schema tests.
    • I2.13.b Resolve single-parent inheritance with cycle/depth detection and exact version pinning. Evidence: inheritance tests.
    • I2.13.c Merge scalar/list/map/reference/constraint fields with explicit per-field replace/append/remove/lock rules. Evidence: merge fixtures.
    • I2.13.d Enforce safety, rights, consent, privacy, retention, budget, and delivery code floors after every inheritance and user override step. Evidence: floor-bypass tests.
    • I2.13.e Preview inherited/defaulted/overridden/locked/effective values and their provenance before apply. Evidence: preset preview tests.
    • I2.13.f Version edits immutably, migrate or pin dependent briefs explicitly, and preserve reproducibility of historical runs. Evidence: version/migration tests.
    • I2.13.g Authorize create/read/update/archive/share/import/export by personal/project/tenant/system scope and audience. Evidence: scope authorization tests.
  • I2.14 Implement import/export of portable brief specs. Evidence: round-trip is versioned, validated, and strips secrets. <!-- evidence: evidence/v1-workbenches/portable-brief/ (compiling reference, strict tsc + runnable suite) and portable-brief.json via node scripts/v1-workbenches/generate-portable-brief.mjs. The portable schema has no field for any secret category; round-trip is compared against a frozen golden canonical body, and an omitted field is asserted to remain omitted rather than becoming a default. -->
    • I2.14.a Define a versioned portable envelope with manifest, canonical brief spec, referenced immutable identities, compatibility metadata, provenance summary, and integrity hash. Evidence: envelope schema.
    • I2.14.b Export only audience-authorized portable fields and omit credentials, signed URLs, provider internals, private prompts, hidden policy, tenant metadata, and nonportable local paths. Evidence: secret/ property leak tests.
    • I2.14.c Represent unavailable, redacted, external, and intentionally omitted references explicitly rather than embedding mutable links. Evidence: reference portability tests.
    • I2.14.d Parse imports with size/depth/count/format/unknown-field limits, version compatibility, signature/hash checks, and safe archive handling. Evidence: hostile import tests.
    • I2.14.e Validate imported typed values, rights/consent, asset/workflow/ model references, code floors, and target compatibility before preview. Evidence: import validation tests.
    • I2.14.f Preview migrations, substitutions, redactions, missing dependencies, conflicts, and effective values before authorized commit. Evidence: import preview tests.
    • I2.14.g Round-trip canonical supported specs byte-semantically while preserving versions and never converting omissions to defaults. Evidence: golden round-trip suite.
  • I2.15 Implement continuity-profile authoring for characters, wardrobe, props, environments, lighting, palette, lens, camera, and motion across shots or asset sets. Evidence: profile revisions, drift thresholds, diagnostics, and authorized overrides remain linked to every output. <!-- evidence (a-m): evidence/v1-workbenches/continuity/ (compiling reference, strict tsc + runnable suite) and continuity.json via node scripts/v1-workbenches/generate-continuity.mjs. The parent's own bar is k+l+m and is proven as a linkage invariant: every consumer kind pins an exact revision, overrides change status only, and an unmeasured dimension reports unavailable rather than zero. runtime: n names an ACCESSIBLE end-to-end test of a rendered authoring/ review surface, which no pure model can establish. --> <!-- SPEC 2026-09-11: parent re-opened. I2.15.n is [ ], and §2.1 forbids a checked parent while any child is unchecked. Every other child stays as it was; the parent flips again only when I2.15.n does. -->
    • I2.15.a Define versioned continuity-profile identity, scope, owners, target asset/sequence membership, reference sources, constraints, thresholds, overrides, and lifecycle. Evidence: profile schema.

    • I2.15.b Author character identity, anatomy, age presentation, facial/ body features, silhouette, scale, pose, expression, and consent-bound reference constraints. Evidence: character continuity tests.

    • I2.15.c Author wardrobe silhouette, garment layers, materials, colors, patterns, accessories, state/damage, and shot-to-shot change rules. Evidence: wardrobe tests.

    • I2.15.d Author prop identity, dimensions, materials, markings, handedness, ownership, placement, interaction, and state transitions. Evidence: prop tests.

    • I2.15.e Author environment layout, scale, geography, architecture, set-dressing anchors, weather, time, and persistent state. Evidence: environment tests.

    • I2.15.f Author lighting source/type/position/direction/color/intensity/ exposure/shadow/atmosphere continuity and intended changes. Evidence: lighting tests.

    • I2.15.g Author palette tokens, color-space values, allowed variance, semantic usage, accessibility constraints, and intentional exceptions. Evidence: palette tests.

    • I2.15.h Author lens/focal/aperture/focus/distortion/sensor/aspect/crop continuity in explicit units. Evidence: lens tests.

    • I2.15.i Author camera position/orientation/height/distance/framing/ eyeline/screen direction/axis and transition rules. Evidence: camera tests.

    • I2.15.j Author subject/object/camera motion paths, timing, velocity, acceleration, easing, cycles, and collision/contact constraints. Evidence: motion tests.

    • I2.15.k Version profile edits immutably and record which exact revision each plan, run, shot, output, and evaluation consumed. Evidence: revision-lineage tests.

    • I2.15.l Calculate typed per-dimension drift measurements, thresholds, uncertainty, unavailable status, and cross-shot/set diagnostics. Evidence: calibrated drift fixtures.

    • I2.15.m Authorize bounded, reasoned, expiring dimension-level overrides without changing raw measurements or profile history. Evidence: override authorization tests.

    • I2.15.n Present effective constraints, conflicts, drift, overrides, and remediation in authoring/comparison/review views. Evidence: accessible end-to-end continuity tests. <!-- TWO OF THE FIVE ITEMS HAVE NO WORD IN THE MODEL AND ONE OF THE THREE VIEWS DOES NOT EXIST. THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-continuity-presentation.json —
      generate-isis-continuity-presentation.mts + .test.mts (8 tests, 3
      ratchets, 2 controls), wired into inventory + verify. The parent's note
      said an accessible end-to-end test of a rendered surface is beyond a
      pure model, which is true of ACCESSIBILITY and hides two other
      questions: whether the model can express the item at all, and whether a
      view NAMES it.
      
      NOTHING IN THE CONTINUITY REFERENCE NAMES A REMEDIATION — not directly,
      and not as a next-step or suggested-fix field. NOTHING RESOLVES OR
      MERGES CONSTRAINTS INTO AN EFFECTIVE SET either: the bare word
      `constraint` is deliberately refused as a synonym, because every
      constraint interface carries it and accepting it would report the item's
      INPUTS as the item. A view cannot present a concept the model has no
      word for.
      
      A NAME IS NOT THE THING IT NAMES, IN BOTH DIRECTIONS. `conflicts` IS
      modelled — the estate calls them `AuthoringIssue`s — and reporting it
      absent because the word differs would be as wrong as crediting a
      near-miss. The row records which term matched.
      
      THERE IS NO REVIEW VIEW. Of the Isis surfaces carrying continuity state,
      two are authoring (`JobSubmissionStudio`,
      `StudioIsisAutoRiggingCorrectionStage`), one is comparison
      (`OutputComparisonWorkspacePage`) and NONE is a review, approval or
      signoff surface. One of the three places the cell names has no candidate
      at all, so nothing can be presented in all three. Three further surfaces
      carry continuity and fit no kind (`JobsPage`, `OutputsPage`,
      `WorkflowDetailsPage`); they are recorded rather than dropped.
      
      AND THE THREE MODELLED ITEMS REACH ONE VIEW KIND. Conflicts, drift and
      overrides are named by the authoring surfaces and by NEITHER the
      comparison one nor a review one — so the comparison view, which does
      exist, shows continuity without showing what drifted.
      
      The model is sound, which is what makes the gaps mean something:
      executed, `reviseProfile` produced revision key `cp-1@3` from `cp-1@2`
      rather than editing in place, and an unmeasured dimension is a distinct
      fixture from a drifted one.
      
      STILL OUT OF REACH HERE: whether these panels, once rendered, are
      keyboard-reachable and announced. That is the "accessible" half and it
      needs the browser harness. -->
      
  • I2.16 Implement authorized conditioning stacks for pose, depth, edges, line art, segmentation, normals, style, face identity, masks, camera paths, motion vectors, and multi-control strength schedules. Evidence: typed UI/API parity and effective-control provenance. <!-- evidence: evidence/v1-workbenches/conditioning/ (compiling reference, strict tsc + runnable suite) and conditioning.json via node scripts/v1-workbenches/generate-conditioning.mjs. Every convention is declared per layer; a mismatch is converted by a named transform that is recorded, or refused. UI and API project one resolved stack, and the parity fixture carries two layers of one kind so the check is not vacuous. -->
    • I2.16.a Define ordered conditioning-stack and layer contracts with kind, source/version/hash, preprocessing, spatial/temporal scope, strength schedule, blend, model/workflow compatibility, and provenance. Evidence: conditioning schema tests.
    • I2.16.b Implement pose controls for skeleton format, keypoints, confidence, people/track association, occlusion, transforms, and frame range. Evidence: pose fixtures.
    • I2.16.c Implement depth controls for units, range, inversion, clipping, normalization, confidence, camera, and resolution alignment. Evidence: depth fixtures.
    • I2.16.d Implement edge and line-art controls for detector/version, threshold, scale, polarity, cleanup, weight, and resolution. Evidence: edge/line fixtures.
    • I2.16.e Implement segmentation controls for label vocabulary/version, instance IDs, confidence, palette mapping, unknown regions, and masks. Evidence: segmentation fixtures.
    • I2.16.f Implement normal controls for tangent/object/world space, axes, handedness, range, camera, confidence, and transform. Evidence: normal fixtures.
    • I2.16.g Implement style controls with authorized reference roles, strength/schedule, region scope, protected attributes, and rights. Evidence: style conditioning tests.
    • I2.16.h Implement face-identity controls with identity subject, consent, reference quality, region/tracking, strength limits, and audience/policy. Evidence: identity safety tests.
    • I2.16.i Implement raster/vector/semantic masks with coordinate space, feather/dilate/erode/invert, frame range, version, and edit history. Evidence: mask tests.
    • I2.16.j Implement camera-path controls with units, coordinate system, transforms, lens, interpolation, timestamps, collision, and bounds. Evidence: camera-path tests.
    • I2.16.k Implement motion-vector controls with direction convention, units, scale, frame correspondence, confidence, occlusion, and bounds. Evidence: motion-vector tests.
    • I2.16.l Validate multiple controls for ordering, compatible models/ preprocessors, coordinate/time alignment, aggregate strength/resource limits, and conflicts. Evidence: multi-control tests.
    • I2.16.m Keep UI and API typed controls/defaults/errors/effective values equivalent and revalidate the exact stack at dispatch. Evidence: UI/API/dispatch parity tests.
  • I2.17 Implement shot, storyboard, animatic, and sequence authoring with shot order, duration, lens, camera motion, continuity, patch rerender, editorial timing, and render-pass requirements. Evidence: sequence rerender preserves unaffected shot identities and lineage. <!-- evidence: evidence/v1-workbenches/sequence/ (compiling reference, strict tsc + runnable suite) and sequence.json via node scripts/v1-workbenches/generate-sequence.mjs. Invalidation propagates only along declared edges, and the identity invariant compares OBJECT identity so a structurally identical re-mint is still caught. -->
    • I2.17.a Define versioned sequence, shot, board panel, animatic segment, track, transition, render requirement, and dependency contracts. Evidence: sequence schema tests.
    • I2.17.b Author stable shot identity, order, grouping, branch, dependency, status, owner, and source revision with conflict-safe reordering. Evidence: shot-order tests.
    • I2.17.c Author duration using frames/timecode/seconds with sequence rate, handles, source/record ranges, retime, and transition overlap. Evidence: timing tests.
    • I2.17.d Author per-shot lens/sensor/aperture/focus/aspect/framing values in explicit units with continuity-profile linkage. Evidence: lens tests.
    • I2.17.e Author camera pose/path/interpolation/constraints/shake/collision and preview it in shot context. Evidence: camera-motion tests.
    • I2.17.f Author storyboard panels with frame/shot linkage, composition, action/dialogue/notes, references, annotations, and revision history. Evidence: storyboard tests.
    • I2.17.g Assemble animatic picture/dialogue/SFX/music/caption tracks with sync, gain, transitions, placeholders, and missing-media states. Evidence: animatic playback tests.
    • I2.17.h Apply character/wardrobe/prop/environment/light/palette/camera/ motion continuity profiles and show shot-level conflicts/drift. Evidence: sequence continuity tests.
    • I2.17.i Define beauty/matte/depth/normal/motion/object/cryptomatte/audio/ caption and target-specific render-pass requirements. Evidence: pass manifest tests.
    • I2.17.j Plan a patch rerender from exact changed shots/frames/passes and downstream dependencies without silently widening scope. Evidence: invalidation graph tests.
    • I2.17.k Preserve unaffected shot/output identities, regenerate affected identities, and link old/new plans/runs/assets/edits. Evidence: patch lineage tests.
    • I2.17.l Export editorial interchange and review proxies with stable IDs, time bases, handles, media refs, captions, and round-trip validation. Evidence: editorial round-trip tests.
  • I2.18 Implement advanced 3D project authoring for reconstruction, scene assembly, part/semantic edits, texture/material generation, relighting, topology/LOD, rigging, splat editing, and engine/DCC packaging. Evidence: every stage is non-destructive, previewable, cancellable, and separately attributable. <!-- evidence: evidence/v1-workbenches/project3d/ (compiling reference, strict tsc + runnable suite) and project3d.json via node scripts/v1-workbenches/generate-project3d.mjs. A stage never mutates its input, so cancel/fail preserves the source and every prior stage; downstream invalidation follows declared edges only. -->
    • I2.18.a Define versioned 3D project, scene, object, component, geometry, material, rig, animation, camera, light, edit layer, dependency, and package contracts. Evidence: 3D project schemas.
    • I2.18.b Author reconstruction inputs, calibration, masks, scale, bounds, quality target, reconstruction method, uncertainty, and source linkage. Evidence: reconstruction tests.
    • I2.18.c Assemble referenced or embedded scenes with stable hierarchy, instances, transforms, units, axes, variants, dependencies, and conflict handling. Evidence: assembly tests.
    • I2.18.d Apply part-level and semantic select/mask/add/remove/replace/ transform/deform edits as reversible versioned layers. Evidence: semantic edit tests.
    • I2.18.e Generate/edit texture sets and materials with UV/UDIM mapping, channel semantics, color space, resolution, tiling, seams, baking, and source rights. Evidence: material tests.
    • I2.18.f Author relighting with environment/source lights, units, color, exposure, shadows, reflections, baking, and original-light preservation. Evidence: relighting tests.
    • I2.18.g Author topology repair/retopology/remesh/decimation/subdivision, normals, UV, manifold, and target polygon constraints non-destructively. Evidence: topology tests.
    • I2.18.h Generate and validate ordered LODs, screen thresholds, reduction settings, material/UV preservation, collision, and target budgets. Evidence: LOD tests.
    • I2.18.i Author skeleton/rig/control/skin/constraint/blend-shape and retargeting stages with bind/reference pose and deformation preview. Evidence: rigging tests.
    • I2.18.j Edit gaussian-splat crop/mask/transform/opacity/SH/compression/ density/cleanup/merge layers with renderer preview. Evidence: splat edit tests.
    • I2.18.k Package glTF/GLB/USD/USDZ/FBX and governed DCC/engine targets with dependencies, manifests, hashes, versions, licenses, and import validation. Evidence: packaging round-trip tests.
    • I2.18.l Give every stage its own plan/run/output/edit identity, preview, progress, cancellation, retry, measurements, cost, and attribution. Evidence: stage lifecycle tests.
    • I2.18.m Preserve source and prior stage revisions on failure/cancel/ replace and make downstream invalidation explicit. Evidence: destructive edit/race tests.
  • I2.19 Implement audio/voice/music project authoring for SFX, foley, dialogue, dubbing, TTS, cloned voice, conversational tracks, music cues, stems, pronunciation, timing, loudness target, and mix package. Evidence: consent, rights, multi-track preview, and output-bundle tests. <!-- evidence (a-m, o): evidence/v1-workbenches/audio/ (compiling reference, strict tsc + runnable suite) and audio.json via node scripts/v1-workbenches/generate-audio.mjs. BS.1770 gated loudness is verified against the standard's own calibration; consent is required by the type and revocation beats a live term; the bundle cannot ship an unmeasured compliance. runtime: n names playback/accessibility tests of a rendered multitrack surface, which no pure model can establish. --> <!-- SPEC 2026-09-11: parent re-opened. I2.19.n is [ ], and §2.1 forbids a checked parent while any child is unchecked. Every other child stays as it was; the parent flips again only when I2.19.n does. -->
    • I2.19.a Define versioned audio project, timeline, track, clip, take, speaker/voice, cue, stem, automation, mix, transcript, caption, and delivery contracts. Evidence: audio project schemas.

    • I2.19.b Author SFX cues with event, perspective, environment, intensity, duration, timing, variation, loop, source/generation, and rights. Evidence: SFX tests.

    • I2.19.c Author foley cues with performer/object/surface/action/contact/ perspective/timing layers and picture sync. Evidence: foley tests.

    • I2.19.d Author dialogue takes with speaker identity, script revision, language, performance direction, pronunciation, timing, room tone, channel, and consent. Evidence: dialogue tests.

    • I2.19.e Author dubbing with source segment/transcript/translation, speaker/voice mapping, lip/timing constraints, adaptation notes, and language review. Evidence: dubbing tests.

    • I2.19.f Author TTS with approved voice/version, text/SSML subset, language, pronunciation lexicon, style/rate/pitch controls, and policy. Evidence: TTS tests.

    • I2.19.g Author cloned/synthetic identity voice only with verified subject consent/scope/term/revocation, reference rights, watermark/disclosure, audience, and use policy. Evidence: cloned-voice safety tests.

    • I2.19.h Author conversational multi-speaker turns with stable speaker/ voice mapping, overlap, pauses, interruption, room, continuity, and disclosure. Evidence: conversation tests.

    • I2.19.i Author music cues with brief, tempo/meter/key, form, duration, hit points, instrumentation, mood, references/rights, loop, and ending. Evidence: music cue tests.

    • I2.19.j Define dialogue/music/SFX/foley/ambience/aux/master stems and clip/track/bus routing without destructive flattening. Evidence: stem routing tests.

    • I2.19.k Maintain versioned language/speaker pronunciation lexicons with phonemes, stress, substitutions, context, approval, and provenance. Evidence: pronunciation tests.

    • I2.19.l Author frame/timecode/sample/beat clip timing, trims, fades, crossfades, stretch, alignment, and synchronization in explicit units. Evidence: timeline edit tests.

    • I2.19.m Author target loudness/true-peak/dialogue-gated/dynamic-range/ channel/layout/sample/codec constraints and measure rather than assume compliance. Evidence: calibrated loudness tests.

    • I2.19.n Preview synchronized multitrack playback, mute/solo/gain/pan/ automation/waveform/transcript/captions accessibly and consistently. Evidence: playback/accessibility tests. <!-- THERE IS NO MULTITRACK PREVIEW ON THE ISIS SURFACE TO TEST FOR ACCESSIBILITY. THE CELL STAYS [ ].

      evidence/v1-workbenches/isis-multitrack-preview.json —
      generate-isis-multitrack-preview.mts + .test.mts (7 tests, 2 ratchets, 1
      control), wired into inventory + verify. A control a surface does not
      have cannot be tested for accessibility, so the two questions in front
      of the browser are whether the model carries each control and whether
      any Isis preview surface names it.
      
      THE MODEL CARRIES SEVEN OF THE EIGHT. Mute, solo, gain, pan, waveform,
      transcript and captions are all fields or checks of the audio reference.
      AUTOMATION IS NAMED NOWHERE IN IT — a mix with no automation curve
      cannot have one previewed, so that control is absent before any surface
      question arises.
      
      AND ONE OF THE EIGHT REACHES A SURFACE. Four Isis surfaces carry any
      playback vocabulary at all; exactly one names a waveform
      (`StudioIsisModelComparisonWorkspace`), and NONE names a mute, a solo, a
      gain, a pan, a transcript or a caption. A multitrack component DOES
      exist in the estate — `libs/yemaya/ui/src/components/MultiTrack` and the
      studio source timeline — so the absence here is a choice rather than a
      missing capability.
      
      THE AUDIO MODEL IS MEASURED, NOT ASSERTED, which is what makes the
      surface gap the finding. Executed: a BS.1770-calibrated block reads
      -3.01 LKFS, and three blocks including a silent one leave two after the
      absolute gate and one after the relative one — silence is gated out
      rather than averaged in.
      
      A CONTROL THAT CHANGED THE ANSWER. A stem rule matched `panel` and
      `panelStyle` and reported an audio pan control on an auto-rigging panel.
      Each control now declares the inflections it really has, and the test
      pins that `panel`, `console`, `against` and `commuter` match nothing. -->
      <!-- Still out of reach: whether a preview, once built, is
      keyboard-operable and announced. That needs the browser harness. -->
      
    • I2.19.o Export versioned mix/stems/captions/transcript/cue sheet/ attribution/consent/rights/measurements in a hashed output bundle. Evidence: bundle manifest/round-trip tests.

  • I2.20 Implement operator workflow graph authoring with node catalog, typed ports, subgraphs, custom-node/model dependencies, validation, version diff, dry run, benchmark, promotion, and rollback. Evidence: graph injection, incompatible node, and dependency-supply-chain tests. <!-- evidence: evidence/v1-workbenches/workflow-graph/ (compiling reference, strict tsc + runnable suite) and workflow-graph.json via node scripts/v1-workbenches/generate-workflow-graph.mjs. All three named attacks are exercised: injection (the guard iterates the catalog, not the payload), an incompatible node (ports carry their full type), and a shallow dependency lock (the closure is walked to its leaves). -->
    • I2.20.a Define immutable graph/version/node/port/edge/parameter/subgraph/ dependency/validation/benchmark/promotion contracts and canonical serialization. Evidence: workflow graph schemas.
    • I2.20.b Build an authorized searchable node catalog with signed source, version, input/output types, parameters, capabilities, compatibility, resource risk, license, lifecycle, and audience metadata. Evidence: catalog tests.
    • I2.20.c Enforce typed port modality/shape/unit/time/coordinate/color/ cardinality/optionality compatibility at connect and validate time. Evidence: port compatibility fixtures.
    • I2.20.d Author reusable version-pinned subgraphs with declared typed interfaces, parameters, dependencies, scope, and cycle/depth limits. Evidence: subgraph tests.
    • I2.20.e Resolve custom-node/model/adapter/runtime/package artifacts by immutable version/hash/signature/license/scan/promotion and complete transitive dependency lock. Evidence: supply-chain tests.
    • I2.20.f Validate required ports, graph cycles, unreachable/dead outputs, parameter ranges, resource bounds, target compatibility, audience, policy, and output contracts. Evidence: graph validation suite.
    • I2.20.g Diff canonical graph versions by node/edge/port/parameter/ dependency/interface/policy/effective behavior with stable identities. Evidence: graph diff fixtures.
    • I2.20.h Dry-run type resolution, dependency availability, authorization, resource/cost estimate, and execution plan without provider side effects. Evidence: zero-dispatch dry-run tests.
    • I2.20.i Benchmark approved fixtures with pinned inputs/environment/ dependencies and record quality/latency/cost/resource/reproducibility. Evidence: benchmark reproducibility tests.
    • I2.20.j Promote exact signed graph versions through governed environments only after required validation/evaluation/approval evidence. Evidence: promotion gate tests.
    • I2.20.k Roll back new selection to a prior admitted immutable graph without deleting runs or rewriting their consumed versions. Evidence: rollback/history tests.
    • I2.20.l Reject code/expression/template/serialized-object/unknown-field/ alias/mass-assignment graph injection and sandbox admitted execution. Evidence: adversarial graph tests.

I3 — Provider execution, jobs, and resource governance#

  • I3.1 Wire the V1 generation route to a real injected executor in an approved environment. Evidence: credentials remain external and the default stays not_configured. <!-- evidence: evidence/v1-workbenches/executor/ (compiling reference, strict tsc + runnable suite) and executor.json via node scripts/v1-workbenches/generate-executor.mjs. ExecutionRequest has no credential field; the leak scan takes the actual secret VALUES; and a call-counted executor plus a read-counted secret provider both record ZERO in every unapproved environment, including against a smuggled binding. -->
    • I3.1.a Define the executor port and typed request/result/progress/error/ cancellation contract independently of any provider SDK. Evidence: executor contract tests.
    • I3.1.b Register real executor implementations only through approved environment composition and keep the default binding explicitly not_configured. Evidence: composition tests.
    • I3.1.c Load credential references from the approved secret provider at execution time without accepting credentials in route payloads, configuration snapshots, logs, or stored plans. Evidence: secret leak tests.
    • I3.1.d Revalidate actor, project, plan, workflow/model versions, policy, reservation, and idempotency immediately before invoking the executor. Evidence: stale-plan/bypass tests.
    • I3.1.e Map executor progress, output, refusal, failure, cancellation, usage, and provenance to canonical domain events. Evidence: adapter integration tests.
    • I3.1.f Prove local/test/unapproved environments make zero provider calls and return the stable not-configured result. Evidence: environment isolation tests.
  • I3.2 Implement RunPod/ComfyUI adapter capability negotiation. Evidence: unsupported nodes/models are rejected before queueing. <!-- evidence: evidence/v1-workbenches/adapter/ (compiling reference, strict tsc + runnable suite) and adapter.json via node scripts/v1-workbenches/generate-adapter.mjs. Identity is class type + module + version and content hash, never a display name; every one of the 15 refusal paths asserts zero submissions against a call-counted queue seam. -->
    • I3.2.a Define the adapter capability document for endpoint/version, queue API, node catalog, models, adapters, formats, limits, progress, cancellation, and result transport. Evidence: capability schema tests.
    • I3.2.b Fetch and authenticate capability state with timeout, size, schema, freshness, cache, and last-known-state rules. Evidence: negotiation failure tests.
    • I3.2.c Normalize RunPod endpoint/worker/GPU and ComfyUI node/object-info data into canonical capabilities without trusting display names. Evidence: normalization fixtures.
    • I3.2.d Compare exact workflow graph, custom nodes, model hashes, adapters, controls, input/output formats, and resource requirements before queue submission. Evidence: compatibility tests.
    • I3.2.e Reject missing, stale, mismatched, quarantined, or unverified capabilities with typed reasons and no queue side effect. Evidence: zero-queue negative tests.
    • I3.2.f Detect capability changes between plan and dispatch and require replan rather than silently substituting dependencies. Evidence: race tests.
  • I3.3 Implement provider routing using health, capability, policy, residency, latency, cost, and capacity. Evidence: known routing fixtures. <!-- evidence: evidence/v1-workbenches/routing/ (compiling reference, strict tsc + runnable suite) and routing.json via node scripts/v1-workbenches/generate-routing.mjs. Endpoints differ in exactly one ranking signal so each tier is proven independently; a stale snapshot is unknown capacity, and dispatch may never substitute another region or model. -->
    • I3.3.a Build the candidate set from approved provider endpoints that match exact workflow/model/input/output capabilities. Evidence: candidate tests.
    • I3.3.b Filter endpoint lifecycle, health/circuit state, policy, tenant/ audience entitlement, regional/data-residency, rights, and security admission. Evidence: eligibility matrix.
    • I3.3.c Filter live queue/GPU/VRAM/storage/egress capacity and reservation feasibility without overcommitting stale snapshots. Evidence: capacity contention tests.
    • I3.3.d Rank eligible endpoints deterministically by policy priority, predicted latency, complete cost, capacity headroom, reliability, and stable tie-break. Evidence: routing fixtures.
    • I3.3.e Record candidates, allowed rejection reasons, selected endpoint identity/version, inputs, estimates, and routing policy version. Evidence: routing decision audit.
    • I3.3.f Revalidate the chosen endpoint and reservation at dispatch; fail/replan according to policy without unauthorized cross-region or model substitution. Evidence: dispatch race tests.
  • I3.4 Implement per-family circuit breakers and retry/backoff from the canonical policy. Evidence: clock-controlled state tests. <!-- evidence: evidence/v1-workbenches/breaker/ (compiling reference; every transition driven by an injected monotonic clock) + breaker.json via scripts/v1-workbenches/generate-breaker.mjs; 146 assertions, 25/25 mutants killed -->
    • I3.4.a Define circuit scope keys for provider, endpoint, workflow/model family, operation, region, and failure class to prevent unrelated traffic from sharing state. Evidence: scope tests.
    • I3.4.b Classify which failures count, do not count, or immediately open each circuit from the canonical versioned policy. Evidence: classifier fixtures.
    • I3.4.c Implement closed/open/half-open transitions with rolling windows, thresholds, cooldown, bounded probes, concurrency safety, and monotonic time. Evidence: fake-clock state tests.
    • I3.4.d Implement attempt limits, exponential backoff, jitter, deadline, Retry-After handling, idempotency, and retryable-stage constraints. Evidence: retry schedule tests.
    • I3.4.e Ensure cancellation, policy denial, validation failure, budget exhaustion, and non-idempotent completed stages are never retried. Evidence: forbidden retry tests.
    • I3.4.f Persist/restore sufficient breaker state across process restart and emit state/decision metrics without high-cardinality secrets. Evidence: restart/telemetry tests.
  • I3.5 Distinguish provider refusal, policy denial, capacity exhaustion, timeout, malformed response, worker loss, and internal failure. Evidence: stable failure taxonomy reaches the UI. <!-- evidence: evidence/v1-workbenches/failure-taxonomy/ (compiling reference; all 30 codes projected to a client shape with no prose field) + failure-taxonomy.json via scripts/v1-workbenches/generate-failure-taxonomy.mjs; 134 assertions, 25/25 mutants killed -->
    • I3.5.a Define stable machine codes, categories, retryability, ownership, user-safe message keys, operator detail, and terminal/nonterminal semantics for every failure class. Evidence: error schema tests.
    • I3.5.b Map provider refusal/status/body signals without exposing provider-private prompts, safety internals, or credentials. Evidence: provider refusal fixtures.
    • I3.5.c Distinguish canonical policy denial from provider refusal and retain policy version/rule/evidence appropriate to audience. Evidence: policy mapping tests.
    • I3.5.d Distinguish queue/capacity/quota/reservation exhaustion with responsible scope, recoverability, and safe retry timing. Evidence: capacity error tests.
    • I3.5.e Distinguish connect/read/overall/deadline timeouts and worker heartbeat/loss/eviction from explicit cancellation. Evidence: timeout/ worker-loss tests.
    • I3.5.f Validate provider response status, schema, content type, size, artifact manifests, hashes, and required outputs before acceptance. Evidence: malformed response tests.
    • I3.5.g Preserve unexpected internal failure correlation for operators while UI/API receive stable redacted typed errors. Evidence: projection/ leak tests.
  • I3.6 Implement durable generation jobs with stage checkpoints. Evidence: worker restart resumes safely or fails explicitly. <!-- evidence: evidence/v1-workbenches/durable-jobs/ (compiling reference; fenced leases, checkpoint provenance, replayable append-only log, real SHA-256 pinned to the FIPS vectors) + durable-jobs.json via scripts/v1-workbenches/generate-durable-jobs.mjs; 111 assertions, 28/28 mutants killed -->
    • I3.6.a Define immutable job identity, versioned plan snapshot, stage DAG, attempt, state, lease, checkpoint, output, error, usage, and event contracts. Evidence: job schemas.
    • I3.6.b Create jobs idempotently with plan/input/hash/reservation binding in one durable transaction. Evidence: duplicate submission tests.
    • I3.6.c Claim work with fenced leases, heartbeat, expiry, and compare-and- set transitions so stale workers cannot commit. Evidence: lease race tests.
    • I3.6.d Checkpoint only validated stage outputs and exact consumed input/ dependency/policy versions with integrity hashes. Evidence: checkpoint integrity tests.
    • I3.6.e On restart, resume only explicitly resumable idempotent stages; otherwise create a new attempt or terminal explicit failure. Evidence: crash/restart tests.
    • I3.6.f Handle duplicate/out-of-order progress/results and provider callbacks without regressing state or duplicating assets/cost. Evidence: event ordering tests.
    • I3.6.g Preserve complete append-only job/stage/attempt history and make current state a reproducible projection. Evidence: replay tests.
  • I3.7 Implement cancellation propagation to queued and running providers. Evidence: race tests prevent late outputs from appearing successful. <!-- evidence: evidence/v1-workbenches/cancellation/ (compiling reference; the claim/cancel race is run, and a late artifact is quarantined rather than counted) + cancellation.json via scripts/v1-workbenches/generate-cancellation.mjs; 100 assertions, 24/24 mutants killed -->
    • I3.7.a Define authorized cancellation request, reason, scope, actor, timestamp, idempotency, and queued/running/terminal semantics. Evidence: cancellation contract tests.
    • I3.7.b Atomically mark queued work cancelled and release queue/resource/ budget reservations before any claim can dispatch. Evidence: queue race tests.
    • I3.7.c Signal running workers/providers using supported cancel handles with timeout, retry, and explicit unsupported/unknown outcome. Evidence: provider cancellation tests.
    • I3.7.d Fence stage/output commits after cancellation while retaining late provider artifacts only in quarantined operator-visible evidence. Evidence: late-result race tests.
    • I3.7.e Reconcile actual usage/cost and release unused reservations even when provider cancellation is delayed or fails. Evidence: cancellation accounting tests.
    • I3.7.f Project cancelled/cancelling/cancel-failed state consistently to UI/API/events and retain immutable history. Evidence: end-to-end tests.
  • I3.8 Implement deduplication and cache eligibility. Evidence: cache keys include every output-affecting input and policy version. <!-- evidence: evidence/v1-workbenches/cache-eligibility/ (compiling reference; each of the 11 key components is changed one at a time and must move the key, and the 4 non-affecting fields must not) + cache-eligibility.json via scripts/v1-workbenches/generate-cache-eligibility.mjs; 79 assertions, 25/25 mutants killed -->
    • I3.8.a Define separately when request idempotency, in-flight coalescing, deterministic result reuse, and derived-artifact reuse are permitted. Evidence: eligibility policy tests.
    • I3.8.b Canonicalize and hash exact input bytes/revisions, prompt spec, workflow graph, models/adapters, effective controls/seed, environment, target, code/policy/evaluator versions, and tenant scope. Evidence: cache-key golden tests.
    • I3.8.c Exclude nondeterministic, private/identity-sensitive, consent- scoped, expired-rights, quarantined, mutable-dependency, or policy- forbidden runs from reuse. Evidence: exclusion tests.
    • I3.8.d Coalesce concurrent eligible requests with fenced ownership, independent authorization, cancellation/reference counts, and fair cost attribution. Evidence: concurrency tests.
    • I3.8.e Store cache metadata, artifact manifests, hashes, provenance, expiry, policy scope, and invalidation reasons without cross-tenant discovery. Evidence: isolation tests.
    • I3.8.f Revalidate authorization, rights/consent, policy, target, and artifact integrity on every lookup before returning a hit. Evidence: stale-hit tests.
  • I3.9 Verify cache hits against stored artifact hashes. Evidence: corrupt cache entries are quarantined and recomputed. <!-- evidence: evidence/v1-workbenches/cache-integrity/ (compiling reference; a corrupted stored object is quarantined by compare-and-set and replaced by a new run that supersedes the old one without rewriting it) + cache-integrity.json via scripts/v1-workbenches/generate-cache-integrity.mjs; 138 assertions, 47/47 mutants killed -->
    • I3.9.a Resolve the immutable cached artifact manifest and authorized storage objects without trusting caller-supplied locations. Evidence: resolver tests.
    • I3.9.b Stream-hash every required artifact and manifest using approved algorithms, enforcing size/count/time/resource limits. Evidence: hash verification tests.
    • I3.9.c Verify manifest/object hash, length, media signature, dependency, provenance, and encryption/key-version consistency. Evidence: corrupted cache fixtures.
    • I3.9.d Atomically quarantine the cache record and affected objects with reason/evidence, preventing races from serving the same entry. Evidence: quarantine contention tests.
    • I3.9.e Recompute through a fresh run identity/reservation and never rewrite the historical corrupt run as successful. Evidence: recompute lineage tests.
    • I3.9.f Emit bounded integrity/security telemetry and operator alerts without exposing asset content or cross-tenant keys. Evidence: telemetry leak tests.
  • I3.10 Implement queue priority with anti-starvation and tenant fairness. Evidence: load test demonstrates bounded wait. <!-- evidence: evidence/v1-workbenches/queue-priority/ (compiling reference) + queue-priority.json via scripts/v1-workbenches/generate-queue-priority.mjs, which RUNS the reference's load simulation and embeds the report it produced; 127 assertions, 32/32 mutants killed. Scope, honestly: the load run is a deterministic simulation of the SCHEDULER (throughput, isolation, wait bounds, starvation) — load testing live workers is a separate exercise and is not claimed here. -->
    • I3.10.a Define canonical base-priority classes, allowed actors/reasons, per-tenant/project/user/workflow/provider weights, and non-overridable safety/incident lanes. Evidence: priority policy tests.
    • I3.10.b Enqueue with immutable submitted time, cost/resource estimate, deadline class, tenant scope, and policy version. Evidence: enqueue schema tests.
    • I3.10.c Implement weighted fair scheduling with aging/anti-starvation, bounded priority boosts, per-scope concurrency, and deterministic tie-breaks. Evidence: scheduler model tests.
    • I3.10.d Prevent job splitting, repeated cancel/resubmit, identity/tenant switching, or direct priority fields from gaming order. Evidence: adversarial fairness tests.
    • I3.10.e Reconcile estimates against actual resource duration and feed bounded future scheduling signals without retroactive priority changes. Evidence: reconciliation tests.
    • I3.10.f Load-test mixed tenants/priorities/job sizes/failures/cancels to demonstrate throughput, isolation, configured wait bounds, and no starvation. Evidence: queue load report.
  • I3.11 Implement GPU/CPU/VRAM/storage/egress reservations. Evidence: overcommit and release-on-failure tests. <!-- evidence: evidence/v1-workbenches/reservations/ (compiling reference; an unauthorised expansion pauses the job instead of overcommitting, and all six stop paths release idempotently) + reservations.json via scripts/v1-workbenches/generate-reservations.mjs; 111 assertions, 33/33 mutants killed -->
    • I3.11.a Define resource kinds, units, scopes, capacity sources, safety headroom, reservation states, expiry, ownership, and accounting contracts. Evidence: reservation schemas.
    • I3.11.b Calculate per-stage and peak concurrent GPU count/type, CPU, RAM/VRAM, temporary/persistent storage, network, and egress needs from the exact plan. Evidence: requirement fixtures.
    • I3.11.c Atomically reserve all required scoped resources or none, with fencing token, lease, queue linkage, and exact insufficiency reasons. Evidence: multi-resource contention tests.
    • I3.11.d Revalidate and consume reservations at dispatch without allowing stale/expired/wrong-job/wrong-endpoint tokens. Evidence: dispatch race tests.
    • I3.11.e Adjust within authorized bounds for measured usage, rejecting or pausing unapproved expansion instead of silent overcommit. Evidence: resize tests.
    • I3.11.f Release unused resources idempotently on completion, failure, cancellation, timeout, lease loss, and abandoned job recovery. Evidence: release path tests.
    • I3.11.g Reconcile reservation ledger to provider/storage/network measurements and alert on leaks/negative/over-capacity invariants. Evidence: reconciliation tests.
  • I3.12 Implement pre-dispatch estimate and post-run cost reconciliation. Evidence: variance and unbilled failures are visible. <!-- evidence: evidence/v1-workbenches/cost/ (compiling reference; an unknown or a quarantined measurement marks the report INCOMPLETE rather than settling as zero, and variance carries typed causes) + cost.json via scripts/v1-workbenches/generate-cost.mjs; 109 assertions, 29/29 mutants killed -->
    • I3.12.a Define versioned price inputs and cost dimensions for provider compute/time, models/nodes, storage, egress, evaluation, post-processing, taxes/fees/credits, and internal allocation. Evidence: pricing schema.
    • I3.12.b Compute low/expected/high estimates from exact plan, candidate count, resource reservation, retry policy, target, currency, and price version. Evidence: estimate fixtures.
    • I3.12.c Bind the accepted estimate/threshold confirmation and budget reservation to the immutable plan before dispatch. Evidence: estimate binding tests.
    • I3.12.d Ingest authenticated provider usage/billing plus canonical worker/storage/egress/evaluator measurements with source timestamps and confidence. Evidence: usage ingestion tests.
    • I3.12.e Allocate actual cost by run/stage/variant/asset/project/tenant without double-counting retries, cache hits, shared jobs, or partial outputs. Evidence: allocation tests.
    • I3.12.f Represent unbilled, estimated-only, pending, credited, refunded, disputed, and unknown costs explicitly rather than as zero. Evidence: accounting state tests.
    • I3.12.g Calculate absolute/relative variance with typed causes and surface configured breach alerts and approval follow-up. Evidence: variance tests.
  • I3.13 Implement tenant, project, user, workflow, and provider quotas. Evidence: each limit is enforced server-side. <!-- evidence: evidence/v1-workbenches/quotas/ (compiling reference; all seven declared entry paths are run against an exhausted quota and refused by the one enforcement point) + quotas.json via scripts/v1-workbenches/generate-quotas.mjs; 98 assertions, 31/31 mutants killed -->
    • I3.13.a Define versioned quota dimensions, units, rolling/fixed windows, scope hierarchy, soft/hard thresholds, burst, reservation, and reset semantics. Evidence: quota schemas.
    • I3.13.b Resolve effective quotas from tenant/project/user/workflow/ provider policy without allowing a child scope to loosen a parent hard limit. Evidence: hierarchy tests.
    • I3.13.c Atomically check and reserve request/count/cost/compute/storage/ egress/concurrency quota at plan confirmation and dispatch. Evidence: quota contention tests.
    • I3.13.d Count completed, failed, cancelled, retried, cached, shared, and partial work according to explicit per-dimension rules. Evidence: quota accounting tests.
    • I3.13.e Release/refund only permitted unused reservations and retain consumed usage through retries or scope changes. Evidence: refund tests.
    • I3.13.f Return audience-safe current/limit/reset/shortfall/retry state and alert before soft/hard thresholds without exposing other tenants. Evidence: UI/API isolation tests.
    • I3.13.g Test direct route, batch, replay, agent, import, provider callback, and concurrency bypass paths server-side. Evidence: quota bypass suite.
  • I3.14 Implement worker and endpoint draining for maintenance. Evidence: no new work lands and in-flight policy is honored. <!-- evidence: evidence/v1-workbenches/draining/ (compiling reference; all three claim kinds are run against a draining target and refused by the one admission point, and the order's in-flight policy is resolved against each running stage's real capabilities) + draining.json via scripts/v1-workbenches/generate-draining.mjs; 194 assertions, 43/43 mutants -->
    • I3.14.a Define active/draining/drained/disabled state, reason, actor, effective/deadline times, in-flight policy, scope, and audit contract. Evidence: drain schema tests.
    • I3.14.b Authorize drain/restore separately for worker, pool, endpoint, provider, model/workflow family, region, and environment. Evidence: operation authorization tests.
    • I3.14.c Remove draining targets from new routing/reservation/queue claims atomically while allowing configured already-bound work. Evidence: new- work exclusion tests.
    • I3.14.d Apply finish, checkpoint-and-migrate, cancel, or deadline-kill policy to each in-flight stage with capability/idempotency validation. Evidence: in-flight matrix tests.
    • I3.14.e Reassign queued work and only migratable running work with new attempt/endpoint identity and preserved lineage/reservations. Evidence: migration tests.
    • I3.14.f Determine drained state from zero claims/in-flight callbacks/ leases/reservations and expose progress/blockers. Evidence: drain completion tests.
    • I3.14.g Restore admission explicitly after health/capability/security checks; never auto-enable expired maintenance targets. Evidence: restore tests.
  • I3.15 Implement run replay with explicit new-run identity. Evidence: historical executions remain immutable. <!-- evidence: evidence/v1-workbenches/run-replay/ (compiling reference; every run is serialized before the replay and re-compared after resolution, preview, confirmation, four recorded outcomes, a cache reuse of the source's own artifact and a second replay) + run-replay.json via scripts/v1-workbenches/generate-run-replay.mjs; 139 assertions, 45/45 mutants -->
    • I3.15.a Resolve an authorized immutable source run and enumerate exact inputs, plan, workflow/model/dependencies, controls, environment, policy, rights, and unavailable/redacted fields. Evidence: replay eligibility tests.
    • I3.15.b Distinguish exact replay, replay-under-current-policy, and fork- with-changes as separate typed operations. Evidence: replay mode tests.
    • I3.15.c Validate current artifact availability/integrity, dependency admission, rights/consent, policy, target, quota, budget, and provider capability before confirmation. Evidence: stale replay tests.
    • I3.15.d Preview substitutions, version drift, nonreproducible inputs, changed policy, expected cost, and output-identity consequences. Evidence: replay preview tests.
    • I3.15.e Create a new request/plan/run/attempt identity linked to the immutable source and record replay mode plus explicit changes. Evidence: replay lineage tests.
    • I3.15.f Never append progress/results/cost/decisions to or alter the source run, even when cache reuse is permitted. Evidence: historical immutability tests.
  • I3.16 Build brief-to-plane agent plans over operation-orchestrator, job-envelope, and token-budget. Evidence: tool grants, dependencies, stop conditions, approval points, and budget validate before dispatch. <!-- evidence: evidence/v1-workbenches/agent-plan/ (compiling reference; each of the nine plan checks is broken on its own and the resulting plan is shown to dispatch ZERO steps) + agent-plan.json via scripts/v1-workbenches/generate-agent-plan.mjs; 129 assertions, 48/48 mutants -->
    • I3.16.a Define versioned Isis agent-plan and step contracts that compose canonical operation, job-envelope, budget, asset, brief, workflow, and provenance identities. Evidence: plan schema tests.
    • I3.16.b Compile the authorized brief/spec into explicit ordered/DAG steps with typed inputs/outputs, dependencies, pre/postconditions, and completion criteria. Evidence: planning fixtures.
    • I3.16.c Resolve each tool grant to an allowlisted versioned capability, operation scope, resource bounds, data classification, and actor/audience authority. Evidence: grant tests.
    • I3.16.d Define hard stop conditions for validation/policy/rights/budget/ timeout/cancellation/dependency/evidence failures and bounded retry/ fallback behavior. Evidence: stop-condition tests.
    • I3.16.e Insert explicit human approval points only where canonical policy requires them, binding approver role and unchanged evidence/plan. Evidence: approval binding tests.
    • I3.16.f Allocate per-step and total token/tool/compute/cost/time budgets with non-transferable ceilings and reserve them before dispatch. Evidence: budget allocation tests.
    • I3.16.g Validate graph acyclicity/reachability, contract compatibility, tool availability, grants, budgets, approvals, and output destinations before any agent execution. Evidence: zero-dispatch invalid-plan tests.
  • I3.17 Build Isis-scoped agent run, queue, plan, tool-call, budget, output, and provenance views. Evidence: live traced journey and cross-tenant refusal. <!-- partial: evidence/v1-workbenches/agent-views/ (compiling reference; 134 assertions, 36/36 mutants) covers a/b/c/e/f/g/h — the projection contract, the state model, pagination and event semantics, budget/output/provenance views and the one authorizer every entry point goes through, including the cross-tenant refusal half of this parent. NOT claimed: I3.17.d (live journey tests) and I3.17.i (accessibility/performance automation) need a running surface, so they and this parent stay unchecked. See agent-views.json runtimeGated. -->
    • I3.17.a Define audience-safe query projections for agent request, plan, step, queue, run, attempt, tool call, budget, event, output, and lineage. Evidence: view contracts.
    • I3.17.b Build the run list/detail state model with filters, stable URLs, loading/empty/error/stale/cancelled/partial states, and pagination. Evidence: state tests.
    • I3.17.c Render plan DAG/step status/dependencies/approval/stop conditions and immutable version details without exposing hidden prompts/tools. Evidence: plan view tests.
    • I3.17.d Render queue position/priority class/estimate/reservation and live bounded event progress with reconnect/cursor/deduplication handling. Evidence: live journey tests.
    • I3.17.e Render each authorized tool call’s tool/version/purpose/status/ timing/budget/input-output schema summary/error and correlated step. Evidence: tool-call view tests.
    • I3.17.f Render allocated/consumed/remaining token/tool/compute/cost/time budget with unknown/pending/reconciled states. Evidence: budget view tests.
    • I3.17.g Render proposed/generated outputs, immutable asset/run lineage, provenance, quality/safety gates, decisions, and partial/missing states. Evidence: output/provenance view tests.
    • I3.17.h Enforce tenant/project/actor/audience/object/property/function authorization on queries, streams, URLs, exports, and actions. Evidence: cross-tenant/adversarial authorization tests.
    • I3.17.i Verify keyboard/screen-reader/reduced-motion/responsive behavior and performance with long plans/events/tool-call histories. Evidence: accessibility/performance automation.
  • I3.18 Intake agent results as attributed proposals or generated assets, never as automatic approvals. Evidence: agents cannot waive gates, advance human signoff, promote environments, or publish. <!-- evidence: evidence/v1-workbenches/result-intake/ (compiling reference; every privileged operation — waive-gate, advance-signoff, promote-environment, publish, apply-proposal — refuses an agent AND a service principal, including an agent holding every permission a human holds) + result-intake.json via scripts/v1-workbenches/generate-result-intake.mjs; 216 assertions, 91/91 mutants -->
    • I3.18.a Define result envelope with agent/run/step/tool/model/version, source inputs, output kind, claim/evidence, confidence, cost, and integrity binding. Evidence: result schema tests.
    • I3.18.b Validate schema, signature/hash, tenant/project scope, expected step/output contract, size/type, moderation, and provenance on intake. Evidence: hostile result tests.
    • I3.18.c Store analysis/recommendations/metadata as attributed proposals with proposed state and no mutation of canonical records. Evidence: proposal persistence tests.
    • I3.18.d Store media outputs as generated asset revisions linked to the exact plan/run/step/tool inputs and pending all applicable gates. Evidence: generated asset lineage tests.
    • I3.18.e Map no agent result field to approval, waiver, gate pass, promotion, publication, or human identity/authority. Evidence: mass- assignment/bypass tests.
    • I3.18.f Route proposals/assets into authorized human review with accept/ reject/modify/defer decisions recorded separately. Evidence: review workflow tests.
  • I3.19 Implement agent pause, redirect, cancel, budget kill, and incident controls. Evidence: race and partial-result tests preserve immutable run history. <!-- evidence: evidence/v1-workbenches/agent-controls/ (compiling reference; the race half is the compare-and-set fence a pre-pause reader is caught by, the partial half is retention + late-result quarantine, and the immutable half is the projection asserted at EVERY log prefix plus a digest of the input run taken before the suite and re-checked after it) + agent-controls.json via scripts/v1-workbenches/generate-agent-controls.mjs; 168 assertions, 96/96 mutants -->
    • I3.19.a Define separate authorized command contracts, eligible states, scope, reason, actor, idempotency, expected version, and outcome for pause/redirect/cancel/budget-kill/incident actions. Evidence: command schemas.
    • I3.19.b Pause at a safe step/tool boundary, checkpoint resumable state, stop new grants/calls, retain leases as policy permits, and expose blockers. Evidence: pause race tests.
    • I3.19.c Resume only after revalidating plan/dependencies/grants/policy/ rights/budget/approvals and create a new attempt where required. Evidence: resume validation tests.
    • I3.19.d Redirect through a new versioned plan revision with explicit changed objectives/steps/tools/budget and preserved completed history. Evidence: redirect lineage tests.
    • I3.19.e Cancel queued/running steps and downstream dependencies with provider/tool propagation, late-result fencing, and partial-result retention. Evidence: cancellation tests.
    • I3.19.f Kill on budget ceiling atomically before the next spend, handling in-flight nonrefundable usage and reconciling actual cost. Evidence: budget overshoot tests.
    • I3.19.g Quarantine/freeze affected runs, outputs, grants, dependencies, and evidence on incident command while preserving investigation access. Evidence: incident containment tests.
    • I3.19.h Keep append-only command/event/attempt/output history and never recast partial/cancelled/incident results as successful. Evidence: history projection tests.
  • I3.20 Implement multi-agent/evaluator consensus only as an inspectable decision aid with member identity/version, independence assumptions, votes/scores, disagreement, confidence, timeout/abstention, quorum, escalation, and human decision. Evidence: correlated agents cannot manufacture gate or approval authority. <!-- evidence: evidence/v1-workbenches/consensus/ (compiling reference; the panel fixture contains REAL copies of one member — same model, provider, corpus and prompt — rather than a correlation flag, and three of them carry one vote, produce no recommendation, and reach no gate) + consensus.json via scripts/v1-workbenches/generate-consensus.mjs; 141 assertions, 69/69 mutants -->
    • I3.20.a Define consensus request/member/assessment/vote/score/ abstention/disagreement/quorum/result/escalation contracts with exact versions and evidence refs. Evidence: consensus schemas.
    • I3.20.b Register each eligible agent/evaluator identity, owner, model/ code/prompt/config version, training/data/provider dependencies, scope, calibration, and lifecycle. Evidence: member registry tests.
    • I3.20.c Record declared and detected shared model/provider/data/prompt/ tool/dependency lineage as independence assumptions, never as proven independence. Evidence: correlation fixtures.
    • I3.20.d Collect assessments against the same immutable subject revision, rubric/policy/version, blinded context where required, and per-member timeout. Evidence: collection tests.
    • I3.20.e Preserve raw votes/scores/reasons/evidence/confidence plus abstention, timeout, error, conflict, and unavailable states. Evidence: assessment projection tests.
    • I3.20.f Calculate configured aggregation/quorum only from eligible members, applying correlation groups, weights, minimum independence, and deterministic tie behavior. Evidence: aggregation fixtures.
    • I3.20.g Surface disagreement dimensions, minority assessments, uncertainty, missing members, assumption violations, and escalation reason. Evidence: consensus UI/API tests.
    • I3.20.h Store consensus as an attributed advisory result and require a separate authorized human/canonical gate decision. Evidence: authority boundary tests.
  • I3.21 Implement batch LLM/media planning and processing with per-item source/prompt/model/budget, bounded fan-out, partial results, cancellation, retries, aggregation, review, and cost attribution. Evidence: no cross-item context or tenant leak and failed items cannot be hidden by aggregate success. <!-- evidence: evidence/v1-workbenches/batch/ (compiling reference; the leak half is the per-item allow-list context + tenant-and-item-namespaced credential, temp dir, cache key, log sink, callback and destination, checked by scanning one item's context for another's content; the hidden-failure half is a denominator over EVERY item plus a review that cannot read the batch's state) + batch.json via scripts/v1-workbenches/generate-batch.mjs; 139 assertions, 79/79 mutants -->
    • I3.21.a Define batch, immutable item, dependency, per-item plan/run/ output, aggregate, review, and cost contracts with maximum item/payload limits. Evidence: batch schemas.
    • I3.21.b Validate every item’s tenant/project/source revision/prompt spec/ workflow/model/rights/policy/target independently before admission. Evidence: mixed-validity intake tests.
    • I3.21.c Allocate per-item and batch token/tool/compute/cost/time/resource budgets without allowing one item to consume another’s hard ceiling. Evidence: budget isolation tests.
    • I3.21.d Schedule bounded fan-out with fair queueing, per-provider/tenant concurrency, backpressure, and no unbounded prompt/media materialization. Evidence: batch load tests.
    • I3.21.e Isolate item context, tools, credentials, temporary files, caches, logs, callbacks, and result destinations. Evidence: cross-item/ cross-tenant leak tests.
    • I3.21.f Retry only eligible failed item stages with idempotency and attempt limits; never rerun or rewrite successful siblings silently. Evidence: partial retry tests.
    • I3.21.g Cancel one item or the whole batch with explicit downstream/ in-flight semantics, late-result fencing, and reservation release. Evidence: cancellation race tests.
    • I3.21.h Aggregate counts/measurements only with success/partial/failure/ cancelled/skipped/unknown denominators and preserve each item’s state. Evidence: aggregate honesty tests.
    • I3.21.i Route each output through applicable review/gates and prohibit aggregate success from approving failed/unreviewed items. Evidence: review boundary tests.
    • I3.21.j Attribute estimates/actual cost/variance/retries/cache/shared work to each item and batch total without double counting. Evidence: batch cost reconciliation tests.

I4 — Asset-kind quality and technical validation#

  • I4.1 Define a versioned gate registry mapping every asset kind and target to mandatory, advisory, and unavailable gates. Evidence: coverage test has no unmapped publishable kind. <!-- evidence: evidence/v1-workbenches/gate-registry/ (compiling reference; coverage walks every publishable kind × target × profile and release planning refuses on a hole, so an unmapped publishable kind cannot ship) + gate-registry.json via scripts/v1-workbenches/generate-gate-registry.mjs; 82 assertions, 42/42 mutants -->

    • I4.1.a Define gate identity/version, evaluator requirements, supported asset/profile/target, applicability, severity, blocking class, inputs, outputs, units, thresholds, and lifecycle. Evidence: gate schema tests.
    • I4.1.b Register every canonical asset kind and delivery target from the source-of-truth registries without parallel aliases or free-text kinds. Evidence: registry join test.
    • I4.1.c Classify each gate per asset-kind/target/profile combination as mandatory, advisory, unavailable, or explicitly not applicable with reason. Evidence: matrix fixtures.
    • I4.1.d Validate required evaluator/model/code/calibration/dependency versions are admitted in the execution environment. Evidence: dependency admission tests.
    • I4.1.e Resolve the exact gate set deterministically and persist the registry/policy/profile versions consumed by each evaluation. Evidence: resolution tests.
    • I4.1.f Fail release planning when a publishable asset-kind/target/profile row is missing or a mandatory gate is unavailable. Evidence: coverage and fail-closed tests.
  • I4.2 Implement image dimension, color-space, alpha, compression, and artifact checks. Evidence: calibrated image fixtures. <!-- evidence: evidence/v1-workbenches/image-checks/ (compiling reference over REAL PNG chunks and JPEG segments and REAL rasters; the calibrated fixture table walks a clean hero, a mislabelled file, an undersized thumbnail and a decompression bomb end to end) + image-checks.json via scripts/v1-workbenches/generate-image-checks.mjs; 257 assertions, 99/99 mutants --> <!-- partial: the fixtures are synthesized rasters with derivable values, not photographic plates; a perceptual calibration against human-labelled images is a data-collection exercise, not code -->

    • I4.2.a Decode supported image formats with byte/pixel/dimension/time/ memory limits and verify declared format against content. Evidence: hostile decoder tests. <!-- evidence: image-decode.ts + image-bytes.ts — signature sniffing, CRC-checked PNG chunk walk and length-walked JPEG segments; all 10 refusals asserted on hostile fixtures (truncation, bad CRC, mislabelled format, a 300-byte 50000x50000 bomb, a segment flood) --> <!-- partial: the "time" budget is a deterministic STEP budget over the header walk; a wall-clock timeout belongs to the runtime that hosts it -->
    • I4.2.b Measure width, height, aspect, orientation, DPI/pixel density, and target-specific minimum/maximum/alignment constraints. Evidence: dimension boundary fixtures. <!-- evidence: image-geometry.ts — measured AS DISPLAYED (an Exif quarter-turn transposes, so a sideways banner fails on the width a viewer sees); boundaries asserted AT the bound in both directions; density read from pHYs/JFIF/Exif and honestly absent when nothing declared it -->
    • I4.2.c Detect embedded/declared/assumed color space, ICC profile, transfer/primaries, bit depth, gamut, and target conversion requirement. Evidence: color fixtures. <!-- evidence: image-color.ts — ICC colorants Bradford-adapted out of the D50 connection space then matched against the published primaries (the published sRGB, Display P3 and Adobe RGB matrix tags are the fixtures); gamut is the chromaticity-triangle area vs BT.709; an untagged file is assumed-default and a profile this layer cannot read is undecidable, never silently sRGB -->
    • I4.2.d Measure alpha presence/range/premultiplication/fringing/coverage and enforce allowed/required/forbidden alpha by target. Evidence: alpha fixtures. <!-- evidence: image-alpha.ts — premultiplication INFERRED from the pixels (a channel above its own alpha settles it; where nothing can settle it the answer is indeterminate), fringing measured as the luma delta between partial pixels and their opaque neighbours; present-but- unused is a flatten, not a rejection; all three policies asserted -->
    • I4.2.e Identify compression format/quality/subsampling/metadata and measure banding, blocking, ringing, blur, and target size budget. Evidence: compression fixtures. <!-- evidence: image-compression.ts — quality recovered by inverting the IJG scaling against the Annex K luminance table in zigzag order (tables built at 20/35/50/75/85/90 estimate back exactly; a clamped table is not-estimable and a lossless image has no quality at all); subsampling from the SOF sampling factors; banding/blocking/ringing/ blur measured on rasters whose values are derivable by hand -->
    • I4.2.f Detect configured generation/edit artifacts with raw scores, regions, uncertainty, evaluator version, and no unsupported certainty. Evidence: labeled artifact fixtures. <!-- evidence: image-artifacts.ts — every CONFIGURED detector reports exactly once; an evaluator we cannot admit reports unavailable with its version and no score (never a zero), a score outside its calibrated range is out-of-range rather than extrapolated, and a finding with no uncertainty carries its number and NO band --> <!-- partial: the detector itself is the evaluator seam — this is the plumbing that attributes, calibrates and refuses, over supplied raw findings -->
    • I4.2.g Emit raw measurements separately from target/policy verdicts and remediation. Evidence: image gate contract tests. <!-- evidence: image-report.ts — RawMeasurements carries no outcome, no violation and no remediation (asserted on the serialized record and gated on the schema's own interface bodies); every violation cites the measurement it read, the bound it compared against and both units, and every check is shown to reach the verdict -->
  • I4.3 Implement portrait identity-consistency scoring with consent-aware applicability. Evidence: positive, negative, and not-applicable fixtures. <!-- evidence: evidence/v1-workbenches/identity-consistency/ (compiling reference; the suite walks a positive match, a negative mismatch that is BANDED rather than refused, a not-applicable empty frame, a face nobody could compare, and a blocked consent) + identity-consistency.json via scripts/v1-workbenches/generate-identity-consistency.mjs; 144 assertions, 66/66 mutants --> <!-- partial: the face detector and the embedding model are the evaluator seam — similarity arrives as a probe over opaque handles, so this layer never sees a vector; a measured ROC over a labelled corpus is a data exercise, not code -->

    • I4.3.a Resolve authorized identity subject, consent scope/term/status, reference set/revisions, intended use/audience, and comparison asset. Evidence: applicability input tests. <!-- evidence: identity-consent.ts — CONSENT IS A FACT ABOUT A PERSON: resolved by subject, newest record wins (an older grant cannot resurrect a consent a newer one withdrew), scope covers use/audience/territory, and the resolved context carries the consent version, the reference revision and which references were rejected -->
    • I4.3.b Return explicit not-applicable/blocked when no identity claim, usable reference, valid consent, permitted use, or supported face region exists. Evidence: N/A/refusal fixtures. <!-- evidence: NOT APPLICABLE IS AN ANSWER, AND SO IS BLOCKED — no claim and no comparable face are not-applicable (the question does not arise); all 9 blocked reasons are separate answers, incl. absent/pending/ withdrawn/expired consent as four distinct events, with the term tested AT the instant it ends -->
    • I4.3.c Detect and associate faces across reference/output with confidence, multiple-person, occlusion, profile, age, and crop handling. Evidence: association corpus. <!-- evidence: identity-association.ts — A MATCH IS ONE-TO-ONE OR IT IS NOTHING: mutual-best with a margin taken from the TIGHTER of the two sides, a reference is consumed when claimed so a second face is reported as another person, an age-band disagreement refuses to compare at all, and all 8 ways not to associate are asserted one at a time --> <!-- partial: detection itself is the evaluator seam; this associates the detections a detector returns -->
    • I4.3.d Compute calibrated identity similarity per associated face and aggregate only according to versioned policy. Evidence: positive/ negative calibration report. <!-- evidence: identity-scoring.ts — the slice must match the pose AND age band exactly, thresholds are tested at the bound, an uncalibrated slice or an out-of-range score keeps its raw value and gets NO band; AN AGGREGATE THE POLICY DID NOT ASK FOR IS AN OPINION: single-face refuses two faces, and an unbanded face stops the aggregate rather than being skipped over by weakest-face --> <!-- partial: the calibration table is a fixture; deriving its thresholds from a labelled corpus is a measurement exercise -->
    • I4.3.e Report raw similarity, calibration band, uncertainty, association, reference quality, and failure reason without exposing biometric embeddings. Evidence: result/privacy tests. <!-- evidence: identity-report.ts — AN EMBEDDING IS THE PERSON: the reported shapes cannot carry a vector, a crop or a handle (gated on the schema's own interface bodies), and the suite serializes a report and searches for the fixture's marker; an unassociated face is reported WITH its reason rather than omitted -->
    • I4.3.f Enforce retention/access/deletion rules for derived biometric processing artifacts independently of the asset. Evidence: privacy lifecycle tests. <!-- evidence: identity-retention.ts — THE DERIVATIVE HAS ITS OWN CLOCK: access is purpose-, role-, time- and consent-bound (a withdrawal cuts it INSIDE the window), the sweep actually removes expired artifacts at the instant, erasure deletes everything biometric while the record that proves we stopped survives, and both independence directions are asserted — erasing a person leaves the pictures, deleting the asset leaves the derivatives -->
  • I4.4 Implement face/hand/anatomy artifact detection. Evidence: labeled evaluation corpus and threshold report. <!-- partial: a, c, d, e and g are done in evidence/v1-workbenches/anatomy-artifacts/ (110 assertions, 62/62 mutants). The PARENT stays open because its bar names the labelled corpus and the threshold report over it, which are I4.4.b and I4.4.f — data collection, not code -->

    • I4.4.a Define versioned artifact labels, regions, severity, applicability, evaluator outputs, and annotation guidance for faces, hands, limbs, joints, body proportions, and interactions. Evidence: taxonomy schema. <!-- evidence: evidence/v1-workbenches/anatomy-artifacts/ (compiling reference) — 18 versioned labels over 4 groups with severity, applicability, measurement kind and annotation guidance; inspectTaxonomy catches duplicate, unregistered, unguided and exemptible-structural rows, and the suite walks EVERY detection it produced to check it is reported in the unit its own taxonomy declares (which caught two real unit errors) + anatomy-artifacts.json via generate-anatomy-artifacts.mjs; 110 assertions, 62/62 mutants -->
    • I4.4.b Assemble licensed, consented, representative positive/negative/ hard-case corpus with demographic, pose, style, crop, occlusion, and multi-person coverage. Evidence: corpus manifest. <!-- blocked: assembling a licensed, consented image corpus is data collection with real people in it; no amount of code produces one -->
    • I4.4.c Detect face feature count/placement/asymmetry/merging and high-severity structural artifacts with localized confidence. Evidence: face artifact fixtures. <!-- evidence: anatomy-face.ts — real geometry over the landmark graph: feature counts, ordering down the face and eye separation as SCALE-FREE fractions of the face box (the same face at half the size is the same face), area asymmetry, and region overlap for merged features; CONFIDENCE IS PROPAGATED (a finding is as sure as its weakest landmark) and an occluded eye makes symmetry UNVERIFIABLE, not passed --> <!-- partial: the landmark estimator is the seam; this measures what it returns -->
    • I4.4.d Detect hand/finger count/shape/joint/contact/occlusion artifacts with left/right and person association. Evidence: hand fixtures. <!-- evidence: anatomy-hands.ts — finger count AND non-canonical finger names, joint bend from a real acos angle (with the domain clamp shown to matter: at one fixture's coordinates the cosine floats to 1.0000000000000004 and the unclamped angle is NaN, so the most broken finger in the frame reports nothing), distal taper, bone-vs-bone segment intersection, and A LIMB BELONGS TO SOMEBODY: two left hands on one figure is itself the artifact -->
    • I4.4.e Detect limb/joint/torso/proportion/attachment/intersection and subject-object interaction anomalies. Evidence: anatomy fixtures. <!-- evidence: anatomy-body.ts — limb counts, attachment to the joint a limb says it hangs off AND the gap to it, PROPORTION IS A RATIO (measured against the published anthropometric fractions of the figure's own height, so a smaller person is not an artifact), joint range, limb self-intersection (limbs sharing a joint touch there by construction and are not flagged), and an object declared held that touches no hand -->
    • I4.4.f Calibrate per-label/style/content thresholds and manual-review bands; report precision/recall/false-positive/false-negative slices. Evidence: threshold report.
    • I4.4.g Emit raw detections/regions/scores/uncertainty separately from policy verdict, respecting sensitive-region access. Evidence: result contract tests. <!-- evidence: anatomy-report.ts — the raw Detection carries no outcome, severity or reason (gated on the schema's own interface body); A REGION IS THE PICTURE OF A BODY, so face and hand coordinates are withheld from roles without that access — asserted on the SERIALIZED record — while the finding itself still appears; a verdict cannot come back clean while anything is unverifiable, and a structural count rejects at a confidence the ladder would only have reviewed, in a stylized profile too -->
  • I4.5 Implement texture resolution, tiling, seam, channel-packing, and PBR consistency checks. Evidence: known material fixtures. <!-- evidence: evidence/v1-workbenches/texture-checks/ (compiling reference over real texels; a material authored to tile exactly with a normal that decodes to unit length, and the same material broken one map at a time) + texture-checks.json via scripts/v1-workbenches/generate-texture-checks.mjs; 119 assertions, 58/58 mutants. Every family of check is proved to REACH the verdict, so no file could be quietly unwired -->

    • I4.5.a Resolve texture-set membership, material/shader profile, target, UV/UDIM mapping, channel semantics, color spaces, units, and expected resolution. Evidence: texture-set fixtures. <!-- evidence: texture-maps.ts resolveSet — A MAP IS A MEMBER OF A SET, NOT A FILE: a missing role, an undeclared one, two maps claiming the same role, a numeric map tagged sRGB (checked in BOTH directions) and a claimed UDIM tile with no map are all defects of the SET -->
    • I4.5.b Measure per-map dimensions, aspect, bit depth, mip presence, texel density, UDIM coverage, naming, and target budgets. Evidence: resolution fixtures. <!-- evidence: dimensions and aspect, powers of two on BOTH sides, the mip chain counted down to one texel (1024 implies eleven levels), byte budgets, the naming pattern, and texel density in texels per WORLD UNIT — so a smaller map over a smaller surface is the same density -->
    • I4.5.c Evaluate intended tile periodicity, border continuity, obvious repetition, scale, and wrap modes in UV context. Evidence: tiling fixtures. <!-- evidence: texture-tiling.ts — A TILE THAT DOES NOT MEET ITSELF IS NOT A TILE: the reading is the mean absolute difference across each wrap (both axes have their own case), repetition is the smallest shift at which the map lies on itself, A FLAT MAP DOES NOT REPEAT IT IS CONSTANT, and a texture nobody said tiles is not judged on tiling at all -->
    • I4.5.d Detect UV island/UDIM boundary seams, padding/bleed, tangent-space normal discontinuity, and channel misalignment. Evidence: seam fixtures. <!-- evidence: the boundary is measured across the island edge, padding is read from the WORST row because that is the row the mip chain will find, the normal across a seam is its own reading (a colour map has no normal to turn), and a set that declares NO background reports no padding rather than a number it could not have measured -->
    • I4.5.e Verify channel packing/order/range/inversion/defaults for roughness/metalness/AO/height/opacity/masks per target profile. Evidence: packing fixtures. <!-- evidence: texture-pbr.ts — each declared slot is read from the channel the profile names, a slot pointing past the end of the file is a blocker, a channel with one value in it is a default rather than a map, and the inversion check is LABELLED A HEURISTIC and only ever asks -->
    • I4.5.f Verify base-color nonlighting, normal convention/normalization, physical metalness/roughness/range, and cross-map PBR consistency. Evidence: PBR material corpus. <!-- evidence: A NORMAL IS A UNIT VECTOR OR IT IS NOT A NORMAL (every texel decodes to a length); the green convention is INFERRED by correlating the map against the height it was baked from — both conventions are recovered from real bakes and a flat normal infers nothing; METALNESS IS A CLASSIFICATION NOT A DIAL; base colour is bounded at both ends, so a baked shadow and an impossible highlight are both findings --> <!-- partial: the corpus is synthesized materials with derivable properties, not scanned real-world ones -->
    • I4.5.g Emit per-map/set raw measures, preview regions, verdicts, and target-specific remediation. Evidence: texture gate tests. <!-- evidence: texture-report.ts — the raw record carries no severity, bound, remediation or outcome (gated on the schema's own interface bodies and on the serialized record); every finding names its map, its measurement and a remediation, and every bound repeats its measurement's unit -->
  • I4.6 Implement video duration, frame-rate, codec, resolution, freeze, duplicate-frame, flicker, and temporal-coherence checks. Evidence: crafted clips yield known results. <!-- evidence: evidence/v1-workbenches/video-checks/ (compiling reference over a real ISO base media file built box by box and real frame planes; a clip authored to be clean, and the same clip broken one thing at a time) + video-checks.json via scripts/v1-workbenches/generate-video-checks.mjs; 200 assertions, 89/89 mutants. Every family of check is proved to REACH the verdict, so no file could be quietly unwired -->

    • I4.6.a Parse container/streams with byte/duration/frame/time/resource limits and verify declared media against decoded content. Evidence: hostile video fixtures. <!-- evidence: video-container.ts — A BOX THAT CONTAINS ITSELF IS NOT A CONTAINER: every read is bounded (a byte past the end refuses rather than returning a zero), a box that does not clear its own header is refused, a child is bounded by its PARENT rather than by the end of the file, and depth, box count, byte count, sample count and DURATION are all capped before anything is allocated against them. A DECLARATION IS NOT A MEASUREMENT: the time table, the size table and the pictures that decoded are three claims about one clip, and the disagreement is the finding -->
    • I4.6.b Measure duration, start/end/time base, frame count, constant/ variable rate, dropped/irregular timestamps, and target tolerances. Evidence: timing fixtures. <!-- evidence: A FRAME HAS A TIME, NOT AN INDEX — the rate is read from the sample deltas in the media timescale, so A MISSING FRAME IS STILL ON THE GRID (a delta of exactly twice the nominal one) while a delta between two grid points is a wobbly timestamp; the two are different departments and are measured apart from each other -->
    • I4.6.c Verify codec/profile/level/pixel format/bit depth/chroma/color/HDR/ bitrate/GOP/container and target compatibility. Evidence: codec matrix. <!-- evidence: video-codec.ts — the configuration record is walked past its parameter sets to reach chroma and depth, and AN UNDECLARED PIXEL FORMAT IS NOT A COMPLIANT ONE (a baseline record declares neither and neither is invented for it); the rate is the coded sample sizes over the media duration rather than a number in a header; the group of pictures includes the tail after the last sync sample; and an HDR transfer an SDR target cannot show is a GRADE, not a re-encode -->
    • I4.6.d Measure encoded/display resolution, aspect, rotation, crop, safe area, and target dimension/alignment constraints. Evidence: resolution fixtures. <!-- evidence: A BAR IS NOT A PICTURE — the active picture is found in the rows and columns themselves and only where they are flat AND dark on EVERY frame, so a dark first frame is a fade, a bright flat band is a slate and a clip that is dark throughout is a black clip rather than a crop it never had; a quarter turn in the track matrix swaps the sides; the aperture is bounded on all four sides; and the safe-area reading is LABELLED A HEURISTIC and only ever asks --> <!-- partial: the safe-area reading is edge detail, not recognized text -->
    • I4.6.e Detect configured freeze intervals and duplicate/near-duplicate frames while distinguishing intentional holds/loops. Evidence: freeze/ duplicate fixtures. <!-- evidence: A HOLD IS NOT A FREEZE, AND A LOOP IS NOT A DUPLICATE — a run without movement is still measured when the asset declared it held, and simply is not a finding; and A LOOP CLOSES WHEN ITS LAST STEP LOOKS LIKE ITS OTHER STEPS, so the join is held against the clip's own median step rather than against zero (a loop whose last picture equalled its first would stutter on every repeat), while a clip nobody declared a loop is not judged on its join at all -->
    • I4.6.f Detect luminance/chroma/exposure flicker with temporal windows, scene-cut awareness, localized regions, and uncertainty. Evidence: flicker fixtures. <!-- evidence: A CUT IS NOT A FLICKER — shots are found first and every window lives inside one, so two halves eighty levels apart are a cut rather than a grade going wrong; a corner that oscillates while the opposite corner goes down by the same amount is found by its TILE behind a frame mean that never moves; chroma is read from the worse of its two planes; a ramp is exposure drift and not a strobe; and a shot too short to hold one window is reported as UNCERTAIN with no number invented for it -->
    • I4.6.g Measure temporal identity/geometry/texture/background/motion coherence with scene/shot boundaries and evaluator applicability. Evidence: coherence corpus. <!-- evidence: AN EVALUATOR WITHOUT ITS SUBJECT IS NOT A PASS — identity, geometry and background need a subject track and, for the background, a camera that was not moving; without them each says which input it lacked, invents no reading, and the verdict CARRIES that it said so. The readings are taken from the pixels: appearance histogram plus chroma inside the box (a subject that keeps its brightness and changes hue is a different subject), the measured extent of the brighter half of its own box, local gradient energy, the frame outside the boxes, and the shift at which one picture best lies on the next --> <!-- partial: the corpus is synthesized clips with derivable properties, and the subject boxes are a detector's output taken as input -->
    • I4.6.h Emit timecoded raw measurements/regions separately from verdicts and remediation. Evidence: video gate contracts. <!-- evidence: video-report.ts — the raw record is timecoded throughout and carries no severity, remediation, bound or outcome (gated on the schema's own interface bodies and on the SERIALIZED record); every finding cites a measurement, repeats its unit in its bound and says when and where; a file that cannot be read is refused with its reason rather than measured as an empty one; and a file whose only problem is that it is not what it said it was still does not pass -->
  • I4.7 Implement AV-sync and caption/transcript checks where audio is present. Evidence: aligned and misaligned fixtures. <!-- evidence: evidence/v1-workbenches/av-sync-checks/ (compiling reference over a real two-track ISO base media file, speech built out of actual samples, a mouth that opens on the same syllables, and caption documents as bytes) + av-sync-checks.json via scripts/v1-workbenches/generate-av-sync-checks.mjs; 176 assertions, 91/91 mutants. The container walk is the I4.6 reference's, reused rather than written twice. Every family of check is proved to REACH the verdict -->

    • I4.7.a Detect audio presence/tracks/languages/channels and return explicit not-applicable when the profile permits silent media. Evidence: applicability tests. <!-- evidence: av-tracks.ts — the tracks are grouped out of the container walk by their own byte offsets (the only way two tracks in one file can be told apart), each carrying its handler, codec, timescale, packed ISO 639-2/T language, channels and sample rate. SILENT AND UNMEASURED ARE DIFFERENT WORDS: no audio on a target that permits silence is NOT APPLICABLE and says which of the three reasons applied, while silence nobody agreed to is a blocker. GOTCHA proved in the suite: A SAMPLE RATE IS UNSIGNED — 48000 does not fit in a signed 16.16 integer part -->
    • I4.7.b Normalize video/audio/caption time bases, edits, delays, handles, variable frame rate, and discontinuities without rewriting sources. Evidence: time-base fixtures. <!-- evidence: A CONVERSION IS NOT AN EDIT — each track keeps its own timescale, its own edit list and its own deltas, and the common millisecond timeline is a MAPPING taken from them: a declared delay pushes a track later, an edit that skips media pulls what follows earlier, a head handle sits before zero, and what the tracks have in COMMON is the only span on which they can be compared. A FRAME'S TIME IS THE SUM OF THE DELTAS BEFORE IT, shown against the index-times-nominal answer that would have been wrong -->
    • I4.7.c Measure speech-to-mouth and general event-to-sound offset/drift with speaker/event association, confidence, and timecoded regions. Evidence: aligned/misaligned AV fixtures. <!-- evidence: av-offset.ts — a real RMS envelope over the samples, correlated against the mouth track at every lag: SYNC IS A RELATION BETWEEN TWO CLOCKS, so a constant lag and a drift are separate findings with separate remediations, and A LEAD IS WORSE THAN A LAG (the two bounds are not the same number). A PEAK WITHOUT A MARGIN IS NOT A MEASUREMENT: a perfectly regular rhythm lines up equally well a syllable over, so the reading reports its own ambiguity, invents no offset, and no slope is computed out of two readings that were never firm --> <!-- partial: the mouth track is an upstream detector's output taken as input, and the corpus is synthesized speech rather than recorded -->
    • I4.7.d Parse caption/subtitle/transcript formats safely and validate encoding, language, cue order, overlap, duration, reading speed, line/ character limits, and target rules. Evidence: caption format tests. <!-- evidence: av-captions.ts — a strict UTF-8 decode written out so that an overlong form, a lone continuation byte, a surrogate half and a truncated sequence each refuse with their own reason rather than becoming a replacement character nobody notices; SubRip and WebVTT timestamps parsed rather than guessed; A CUE IS A CLAIM ABOUT A TIME, so one that ends before it starts is refused rather than measured; and READING SPEED IS A RATE, NOT A LENGTH — markup and position overrides come off before anything is counted -->
    • I4.7.e Compare transcript/caption text to speech with language/speaker/ non-speech cue handling, word timestamps, error metrics, and uncertainty. Evidence: transcript accuracy fixtures. <!-- evidence: av-transcript.ts — a word-level alignment whose edits are counted by KIND (substitutions, deletions and insertions separately). A NON-SPEECH CUE IS NOT A MISTAKE: bracketed sounds, music marks and speaker labels come out first; A DIFFERENT LANGUAGE IS NOT AN ERROR RATE (a caption in another language is unavailable, not a hundred per cent wrong); AN ERROR RATE OVER A SHORT SAMPLE IS NOISE; and a recognizer that is not sure of what it heard is not a reference to blame the captions against -->
    • I4.7.f Emit separate raw AV-sync and caption/transcript measurements, not-applicable/unavailable states, verdicts, and remediation. Evidence: result contract tests. <!-- evidence: av-report.ts — the raw record is timecoded and carries no severity, remediation, bound or outcome (gated on the schema's own interface bodies and on the SERIALIZED record); the verdict carries BOTH what did not apply and what could not be measured, which are different words for different things; a caption document that cannot be read is refused with its own reason rather than measured as an empty one; and a target that never asked how they line up is told nothing was asked instead of being handed a clean report nobody ran -->
  • I4.8 Implement audio sample-rate, channel, clipping, silence, loudness, true-peak, noise, and duration checks. Evidence: calibrated WAV fixtures. <!-- evidence: evidence/v1-workbenches/audio-checks/ (compiling reference over real RIFF/WAVE files built chunk by chunk, decoded to floats, over signals with derivable properties) + audio-checks.json via scripts/v1-workbenches/generate-audio-checks.mjs; 172 assertions, 93/93 mutants. Every family of check is proved to REACH the verdict -->

    • I4.8.a Decode supported audio/container formats with byte/sample/channel/ duration/time/memory limits and verify declared content. Evidence: hostile audio fixtures. <!-- evidence: audio-riff.ts — a bounded chunk walk (a byte past the end refuses rather than returning a zero; the pad byte after an odd chunk belongs to nobody and forgetting it lands mid-header; a chunk claiming more than the file holds is refused) with byte, frame, channel and DURATION limits applied before the arrays they would size are built. Each width is taken apart by hand and each round-trips, including a negative sample at every one of them. The header describes its samples three times over — block alignment, byte rate and chunk size — and where those disagree the disagreement is the finding -->
    • I4.8.b Measure sample rate, bit depth/format, channel count/layout/order, phase/polarity, codec/bitrate, and target compatibility. Evidence: format/channel fixtures. <!-- evidence: audio-format.ts — rate, width, format (including an extensible file whose real format is in its sub-format GUID), channel count, layout, the speaker order its channel mask names, the bits a second the samples run at, and the correlation between the channels: a pair carrying the same programme with one of them inverted reads as minus one, and cancels the moment anybody folds them to mono -->
    • I4.8.c Detect sample/sequence clipping and inter-sample true peaks per channel with exact locations and calibrated oversampling. Evidence: clipping fixtures. <!-- evidence: A SAMPLE PEAK IS NOT A PEAK — the true peak is measured on a four-times oversampled signal through a polyphase windowed-sinc built in the file (odd tap count, so the zeroth phase is the sample itself and every phase passes a constant as itself). A tone sampled either side of every crest reports a peak three decibels above any sample in it, at a position between two of them, while no single sample is at the ceiling; and the reading can never come back UNDER the samples it contains -->
    • I4.8.d Measure leading/trailing/internal silence, dropout, stuck samples, DC offset, and intended-silence exemptions. Evidence: silence/dropout fixtures. <!-- evidence: SILENCE SOMEBODY ASKED FOR IS NOT A DROPOUT — the same hole, declared, produces neither finding. Where a run sits decides what it is (leading, trailing or internal); a dropout is measured against the MEDIAN of the programme around it rather than an absolute level, so a hole that never reaches silence is still a hole; and a run of identical NON-ZERO samples is a converter that stopped, while a run of zeros is silence -->
    • I4.8.e Measure integrated/short-term/momentary loudness, loudness range, dialogue-gated loudness, and true peak to the target standard. Evidence: calibrated loudness fixtures. <!-- evidence: audio-loudness.ts — K-weighting (shelf then high pass), 400ms blocks at a 100ms hop, the absolute gate at −70 LUFS and the relative gate ten below the material itself. THE CALIBRATION IS AGAINST THEORY: the filter exposes its own transfer function, and a 1kHz tone at −20dBFS measures within a tenth of a decibel of what that function predicts — as does a 60Hz tone, which only the second stage explains. Doubling the amplitude adds exactly six decibels, a stereo pair is exactly three above the same programme on one channel, half a file of silence moves the integrated loudness by under a decibel where an ungated mean would drop three, and a file entirely below the absolute gate has no loudness at all. The range is a PERCENTILE spread rather than a span, and dialogue is the same measurement over the blocks a detector marked as speech --> <!-- partial: the weighting coefficients are the published ones for 48kHz, so a file at any other rate is reported as not calibrated rather than measured -->
    • I4.8.f Estimate noise floor/SNR/hum/buzz/hiss/clicks with speech/music/ ambience applicability, regions, and uncertainty. Evidence: noise fixtures. <!-- evidence: A NOISE FLOOR IS A QUIET PERCENTILE, NOT A MINIMUM — one ten- millisecond hole twenty decibels down is not the floor of the recording, and a recording with no quiet passage in it has no measurable floor, which is reported rather than guessed. AN EVALUATOR WITHOUT ITS CONTENT IS NOT A PASS: ambience is the floor rather than something above one, so it is not judged on its own quietness. The hum is a Goertzel bin at the mains frequency AND its harmonics (a hum that is all third harmonic is still a hum), hiss is the share of energy above four kilohertz, and a click is one sample that jumps away from BOTH its neighbours -->
    • I4.8.g Measure exact sample duration against declared/timeline/target timing and account for encoder delay/padding. Evidence: duration tests. <!-- evidence: A DECODED FILE IS LONGER THAN THE MUSIC IN IT — the encoder's priming samples and its padding come off before the length is held against a target, so the same file passes a target that asked for the programme and fails one that asked for the file; a fact chunk is a claim held against the samples rather than trusted over them; and a timeline with less room than the programme needs is its own finding -->
    • I4.8.h Emit per-channel/time-region raw results separately from verdicts and remediation. Evidence: audio gate contracts. <!-- evidence: audio-report.ts — the raw record is per channel and carries no severity, remediation, bound or outcome (gated on the schema's own interface bodies and on the SERIALIZED record); the verdict carries what was not measured and why; a file with no samples in it is refused with its reason rather than measured as a silent one; and a file whose only problem is that its header contradicts itself still does not pass -->
  • I4.9 Implement mesh topology, manifoldness, watertightness, normals, UV, scale, pivot, material, and polygon-budget checks. Evidence: golden mesh corpus. <!-- evidence: evidence/v1-workbenches/mesh-checks/ — a COMPILING strict-tsc reference (schema/math/parse/topology/manifold/normals/uv/transform/ material/budget/report + fixtures + a 274-assertion suite), gated three ways by verify:inventory:v1-workbenches (generator --check, tsc --noEmit, and the suite run under tsx). The corpus is a real glTF binary assembled chunk by chunk and a real Wavefront OBJ read line by line; the shapes in it are built so that what they are is known before anything measures them (a cube of eight points and characteristic two, a torus of forty-eight and characteristic nothing) -->

    • I4.9.a Parse supported mesh/scene formats with vertex/face/object/ material/dependency/depth/time/memory limits and content sniffing. Evidence: hostile mesh fixtures. <!-- evidence: mesh-parse.ts — THE BYTES DECIDE: the format is sniffed from the glTF magic or from a text file having both points and faces, and a file whose declared format disagrees is read as what it is with the disagreement carried. 13 refusals, each reached by its own hostile fixture (truncated, unknown-format, unsupported-version, chunk-size-invalid, chunk-overruns-file, missing-required-chunk, document-invalid, accessor-out-of-range, index-out-of-range, unsupported-component-type, node-cycle, depth-limit-exceeded, resource-limit-exceeded); byte, vertex, face, object, depth, dependency and step limits all applied before the array they would size is built, with the two paths to the shared resource refusal distinguished by their own messages. A VERTEX LIMIT IS NOT AN INDEX LIMIT, and A CORNER IS A TRIPLE: an OBJ point carrying two coordinates is two vertices, indices count from one and backwards from the end -->
    • I4.9.b Measure vertices/edges/faces/triangles/ngons, degenerate/zero- area/duplicate/isolated elements, poles, density, and topology by object. Evidence: topology fixtures. <!-- evidence: mesh-topology.ts — A SEAM IS NOT A CRACK: counted on welded points, where welding is a search across the twenty-seven cells around each candidate rather than a hash lookup, so a cube split for its normals is twenty-four vertices and eight points. A TRIANGLE COUNT IS NOT A FACE COUNT: the same cube as quads is six faces, twelve edges and twelve triangles. Degenerate, zero-area, duplicate (by point SET, so a rotated face is the same face), non-planar, isolated and redundant elements are separated from one another — a corner that meets itself makes no new edge and is reported once, as the thing it is -->
    • I4.9.c Detect nonmanifold edges/vertices, boundaries, self-intersections, holes, disconnected shells, and target-specific watertightness. Evidence: manifold/watertight fixtures. <!-- evidence: mesh-manifold.ts — A HOLE IS A LOOP, NOT A MISSING FACE: boundary edges are walked into the rings they form (an open box is one loop of four, a bowtie is two). Shells are connected components over shared edges; a pinched vertex is found by the faces around it failing to make one fan; the characteristic comes out at two for a closed box either triangulated or as quads, four for two boxes, and nothing at all for a torus whose genus is therefore one. Self-intersection is real triangle-against-triangle, and TWO TRIANGLES IN THE SAME PLANE HAVE NO SEPARATING AXIS OUT OF IT — a torus, whose rings sit in bands at the same height, is reported as a knot by any test that forgets that. WATERTIGHT IS A PROPERTY OF A TARGET: all of it is measured, and the same open box passes one target and fails another -->
    • I4.9.d Validate face/vertex/tangent normals for missing/zero/inconsistent/ inverted state, smoothing groups, hard edges, and target convention. Evidence: normal fixtures. <!-- evidence: mesh-normals.ts — A NORMAL DOES NOT RIDE THE MATRIX THE POINTS RIDE: vertex normals go through the inverse transpose, so a mirror above an object turns its surface inside out while the file is unchanged, and under a fourfold stretch the corner leans at the angle the transpose predicts. Face planes come from Newell over the whole cycle. Winding is read off the direction two faces cross the edge they share, so inverted normals (volume still positive), one flipped face (three disagreeing pairs) and a wholly inside-out solid (consistent, volume negative) are three separable findings. A hard edge is only UNSPLIT where something claims the surface is smooth across it — a smoothing group, or a shared vertex normal — so the properly split cube is clean -->
    • I4.9.e Validate UV sets, required channels, bounds, overlap, distortion, texel density, island padding, UDIM layout, and missing faces. Evidence: UV fixtures. <!-- evidence: mesh-uv.ts — A UV IS AN AREA, NOT A COORDINATE. A SHELL THAT FILLS A TILE EXACTLY IS IN THAT TILE: an upper edge landing on a border belongs to the tile it bounds, and a floor over the corners reports every well-packed island as spilling next door. A face that leaves its tile is a separate finding from an overlap; neighbours are NOT excluded from the overlap test the way they are in space, because a shell folded back over itself shares an edge AND covers the texture twice. Density is a length over a length (the same sheet on four times the surface is half the texels, not a quarter) and is carried as a distribution; padding is the gap between two ISLANDS, in texels. A face has coordinates or it has not: three corners of a quad is a face without them -->
    • I4.9.f Resolve declared/source/scene/target units and axes; measure bounds, scale, transforms, origin/pivot, grounding, and orientation. Evidence: scale/pivot fixtures. <!-- evidence: mesh-transform.ts — A NUMBER WITHOUT A UNIT IS NOT A LENGTH: what the file wrote down, what its format fixes, and what the pipeline knows are three different facts trusted in that order, and each of the three is exercised. A FILE THAT DECLARES NO UNIT HAS NO SCALE — an OBJ with nothing behind it reports null rather than guessing metres, keeps its bounds in file units, and carries the missing reading on the verdict. The hundredfold error is caught against an expected size range. A negative determinant is a mirror, grounding is measured along the axis the target calls up and in the direction it calls up, a translation moves the origin with the geometry so it does not move the pivot, and the orientation heuristic asks for review rather than failing -->
    • I4.9.g Validate material slots/assignments/names/shader compatibility/ texture references and unassigned or unused materials. Evidence: material assignment tests. <!-- evidence: mesh-material.ts — slots, assignments and the frame a normal map needs are per object; names, features, files and dependencies belong to the library. A path that names a place (rooted, drive-lettered, fetched over a network, or climbing out of the delivery) is told from one that names a file in it, while an embedded image is already here. An OBJ names its images inside a material library it did not bring, so nothing is claimed about textures that cannot be read from what was delivered -->
    • I4.9.h Enforce per-object/per-LOD/scene polygon, vertex, material, draw- call, and memory budgets for the selected target. Evidence: budget boundary tests. <!-- evidence: mesh-budget.ts — A SLOT NOBODY USES IS NOT A DRAW CALL, and a material no face wears costs no texture memory. The renderer pays for the vertices in the buffer rather than the points on the surface; indices are two bytes until the count passes sixty-five thousand five hundred and thirty-six and four bytes one vertex later (asserted at the boundary itself); a texture costs its own area plus every mip under it (a thousand-texel map is 5,592,404 bytes exactly). A level of detail is a promise to be smaller, and a level the delivery promised and did not bring is a finding of its own -->
    • I4.9.i Emit per-object and aggregate raw measurements, locations, verdicts, and repair guidance. Evidence: mesh gate contracts. <!-- evidence: mesh-report.ts — A MEASUREMENT IS NOT A VERDICT: the raw record is per object and per element and carries no severity, remediation, bound or outcome (gated on the schema's own interface bodies AND on the SERIALIZED record), while the verdict carries the object, the element, the measurement, the bound and the repair. Each item's measurement is shown to stand alone and to be carried through unaltered. Every one of the seven families of check is shown to reach the verdict from a single asset broken in all of them, so no file could be quietly unwired. AN EVALUATOR WITHOUT ITS SUBJECT IS NOT A PASS — a pair scan too large to run is carried as not measured — and a file whose only problem is that its header contradicts itself still does not pass -->
  • I4.10 Implement rig hierarchy, bind pose, skin-weight, deformation, and animation-range checks. Evidence: valid and broken rig fixtures. <!-- evidence: evidence/v1-workbenches/rig-checks/ — a COMPILING strict-tsc reference (schema/math/parse/hierarchy/bind/skin/deform/clips/report + fixtures + a 234-assertion suite), gated three ways by verify:inventory:v1-workbenches (generator --check, tsc --noEmit, and the suite run under tsx). The corpus is a real binary glTF assembled chunk by chunk — nodes, skins, inverse binds, joints, weights and animation samplers — and a real BVH written line by line with its braces, its channel lists and one motion row per frame; the skeletons in it are built so that what they are is known before anything measures them (a chain of three joints a metre apart, a limb of twelve points whose middle ring hangs half from one joint and half from the next). 61 of 61 mutations to the implementation are caught by the suite -->

    • I4.10.a Parse skeleton/skin/animation data with bone/influence/key/frame/ clip/dependency/resource limits and stable joint identities. Evidence: hostile rig fixtures. <!-- evidence: rig-parse.ts — THE BYTES DECIDE: the format is sniffed from the glTF magic or from text that declares a hierarchy, roots it and has motion under it, and a file whose delivery disagrees is read as what it is with the disagreement carried. 15 refusals, each reached by its own hostile fixture (truncated, unknown-format, unsupported-version, chunk-size-invalid, chunk-overruns-file, missing-required-chunk, document-invalid, accessor-out-of-range, joint-index-out-of-range, unsupported-component-type, joint-cycle, hierarchy-depth-exceeded, sampler-length-mismatch, frame-width-mismatch, resource-limit-exceeded); byte, joint, vertex, clip, key, depth and step limits all applied before the array they would size is built. A KEY LIMIT IS NOT A VERTEX LIMIT, and A NAME IS NOT AN IDENTITY: every joint is keyed by its path from the root, so two joints called the same thing are two joints. A cycle is refused before anything walks it, and a node carrying geometry is not a joint however deep inside the armature it hangs -->
    • I4.10.b Validate hierarchy roots, parent cycles, duplicate/missing joints, names, required bones, orientation, scale, transforms, and target convention. Evidence: hierarchy fixtures. <!-- evidence: rig-hierarchy.ts — A BONE IS THE GAP BETWEEN TWO JOINTS, NOT A JOINT: the bone a joint owns is the one below it, so three joints in a chain are two bones; A BRANCH HAS NO ONE DIRECTION, so the hips report none; a leaf has none either unless the file says where it ends, which a BVH end site does and a glTF does not; and the gap ABOVE a joint is the separate reading that catches two joints sitting on top of one another. AN AXIS IS DECLARED, NOT DISCOVERED — a humanoid is nearly as long below its root as above it, so which way up comes from the container's own specification or from the delivery, and where neither says, nothing is guessed. A mirror is a negative scale and is told from a shear, which is not a rotation and a scale at all -->
    • I4.10.c Validate bind/reference/rest pose matrices, inverse binds, skeleton/mesh transform agreement, and expected pose/profile. Evidence: bind-pose fixtures. <!-- evidence: rig-bind.ts — AN INVERSE BIND IS A CLAIM, AND A CLAIM CAN BE CHECKED: each one is put back through the skeleton and must come out as identity, a half-metre error reads as a half-metre deviation, and one that flattens a dimension is refused as something nothing can undo. A skin that declares none is read the way a renderer reads it rather than as what the exporter meant. Skinning replaces the transform on the node the mesh hangs from, so a non-identity one is reported as lost. The bind pose is classified by the arm's elevation — level is a T, forty-five degrees down is an A, and twenty degrees is NEITHER and is asked about rather than called the nearer of the two. Symmetry is a mirror of one side against the other, and a skeleton bound at a hundred times the size of the mesh is caught from both sides of the bound -->
    • I4.10.d Measure per-vertex influence count, missing/negative/NaN weights, normalization, tiny weights, joint references, and target limits. Evidence: skin-weight fixtures. <!-- evidence: rig-skin.ts — AN INFLUENCE WITH NO WEIGHT IS NOT AN INFLUENCE: the corpus writes four slots per vertex and carries weight in one or two, so a target that allows three is told nothing and a target that allows one is told about two. A WEIGHT THAT IS NOT NORMALIZED IS A SCALE, so the sum is measured per vertex; a pair summing to one with a negative half is a negative weight and not an unnormalized one; a weight too small to move anything is prunable rather than wrong, and is still an influence. A vertex with every weight at zero stays where it was bound while the rest moves, and a joint the skin lists that nothing hangs from is named -->
    • I4.10.e Exercise calibrated poses/clips and detect collapses, excessive stretch, volume loss, penetrations, detached vertices, and joint discontinuities. Evidence: deformation fixtures. <!-- evidence: rig-deform.ts — A POSE IS A VERSIONED INPUT, NOT A SCREENSHOT: linear blend skinning is actually run at named, versioned poses, and at the pose it was bound in nothing moves at all. The candy wrapper is caught for what it is — half a joint and half of the same joint turned round is a matrix that keeps one axis and flattens the other two, and the middle ring lands exactly on the bone with no area left. A CREASE IS A PROPERTY OF THE WEIGHTS, NOT OF THE POSE: the same bend folds flat under binary weights and not at all under shared ones, while the volume lost is identical in both, so creasing is the reading that tells them apart. A VERTEX THAT LEAVES ITS NEIGHBOURS BEHIND IS NOT DEFORMING — every one of its edges stretches, which is what tells it from a joint boundary doing its job. A BOUND IS TWO-SIDED (one pose stretches and only stretches, another crushes and only crushes), and a pair scan too large to run is carried as not measured rather than as nothing found -->
    • I4.10.f Validate clip names, ranges, frame/time bases, keys, interpolation, loop boundaries, root motion, events, and out-of-range tracks. Evidence: animation-range fixtures. <!-- evidence: rig-clips.ts — A ROTATION ORDER IS PART OF THE ROTATION: the BVH channel list is the order the three angles are composed in, checked against a value worked out by hand, and the same numbers in a different order are a different pose. A QUATERNION AND ITS NEGATIVE ARE THE SAME ROTATION AND NOT THE SAME KEY, which is why every pose in such a file is right and the motion between two of them goes the long way round. A MISSING FRAME IS STILL ON THE GRID, so a gap that is a whole number of frames wide is a dropped frame and one that lands between them is a retimed key. A CLIP THAT ENDS WHERE IT STARTS STUTTERS: the join is judged against the clip's own median step, so a bounce that closes on a step is clean, one whose last frame repeats its first plays that frame twice, and one that jumps nine steps pops — while A TRACK THAT NEVER MOVES CANNOT STUTTER. Root motion is measured on the ground plane only, so going up is not going anywhere -->
    • I4.10.g Emit joint/vertex/clip raw measurements and regions separately from target verdicts. Evidence: rig gate contracts. <!-- evidence: rig-report.ts — A MEASUREMENT IS NOT A VERDICT: the raw record is per joint, per vertex and per clip and carries no severity, remediation, bound or outcome (gated on the schema's own interface bodies AND on the SERIALIZED record), while the verdict carries the joint or the clip, the measurement, the bound and the repair. Every one of the five families of check is shown to reach the verdict from a single rig broken in all of them, so no file could be quietly unwired. AN EVALUATOR WITHOUT ITS SUBJECT IS NOT A PASS — a container that carries no skin, an open surface that encloses nothing, a leaf that declares no end and an arm that sits at neither pose are all carried as not measured — and a file that cannot be read is refused with its reason rather than measured as an empty skeleton, which would have no cycles, no weights and no clips, and would pass -->
  • I4.11 Implement gaussian-splat geometry, PSNR/SSIM/LPIPS, memory, and renderer-compatibility checks. Evidence: benchmark scene fixtures. <!-- evidence: evidence/v1-workbenches/splat-checks/ — a COMPILING strict-tsc reference (schema/math/parse/geometry/properties/render/metrics/resources/ compatibility/report + fixtures + a 180-assertion suite), gated three ways by verify:inventory:v1-workbenches (generator --check, tsc --noEmit, and the suite run under tsx). The corpus is a real binary PLY of the kind a gaussian-splatting trainer writes, assembled header line by header line and then row by row, and the splats in it are built so that what they are is known before anything measures them: a DC coefficient chosen to stand for exactly white, a stored scale of zero that is a splat one unit across, and a stored opacity of zero that is a splat exactly half opaque. 61 of 61 mutations to the implementation are caught by the suite -->

    • I4.11.a Parse supported splat/container versions with point/property/ coefficient/size/time/memory limits and required-field validation. Evidence: hostile splat fixtures. <!-- evidence: splat-parse.ts — THE BYTES DECIDE: a PLY says so on its first line and a delivery that calls it something else is read as what it is with the disagreement carried. 11 refusals, each reached by its own hostile fixture (truncated, unknown-format, unsupported-version, header-invalid, unsupported-encoding, unsupported-property-type, required-property-missing, sh-coefficient-count-invalid, row-count-mismatch, camera-invalid, resource-limit-exceeded); byte, point, step and pixel limits all applied before the array they would size is built. A STORED NUMBER IS NOT THE QUANTITY IT STANDS FOR — the opacity goes through the sigmoid and the scales through the exponential on the way out — A DEGREE IS A COUNT, NOT A DECLARATION, and the rest coefficients are read channel by channel rather than every third one, which is the classic way to get three colours that are each a third of three others -->
    • I4.11.b Measure point count/density/distribution/bounds/scale/orientation, invalid positions/covariances/scales/rotations/opacities, and outliers. Evidence: geometry fixtures. <!-- evidence: splat-geometry.ts — every reading is on the DECODED quantities, so a file full of negative stored scales is a file full of small splats rather than a broken one. A COVARIANCE IS A ROTATION AND THREE LENGTHS, so what can go wrong is the lengths: the needle (longest axis over shortest), the boulder, and the one that decoded to nothing at all. A cloud in a plane has no volume to be dense in and says so; outliers are found against a median absolute deviation, and A ROBUST SPREAD OF NOTHING IS NOT A SMALL SPREAD — four splats the same distance out give the method no scale, and that is reported rather than used to call the fifth one a floater -->
    • I4.11.c Validate spherical-harmonic degree/coefficient count/range, color/opacity encoding, precision, and source-format semantics. Evidence: property fixtures. <!-- evidence: splat-properties.ts — A DC COEFFICIENT IS NOT A COLOUR: the zeroth band is a constant, so the colour is C₀ times the coefficient plus a half, which is why a file of zeros is mid grey and why the coefficient that stands for white is 1.7724538509. A colour out of gamut by less than one display level is in gamut, because the coefficients come back out of four-byte floats. A coefficient below what the target's precision holds becomes a zero on arrival, and a file whose every stored opacity sits inside the unit interval has had the sigmoid applied twice -->
    • I4.11.d Render the same versioned calibrated cameras/reference frames under pinned renderer/config/color settings. Evidence: benchmark render reproducibility tests. <!-- evidence: splat-render.ts — the splats really are projected the EWA way (Σ' = J W Σ Wᵀ Jᵀ), sorted by depth and blended front to back with the transmittance walked down, so WHAT IS IN FRONT IS WHAT YOU SEE: two splats at one place on the film and different depths render as the near one, and turning the pair round renders as the other. A PICTURE IS TAKEN FROM SOMEWHERE — the camera set's version, the renderer's version and the colour space travel with every frame, the same scene renders byte for byte the same twice, and the colour space is part of the picture rather than a note about it. A DEGREE ABOVE ZERO IS A COLOUR THAT DEPENDS ON WHERE YOU STAND: one band renders the same from both cameras and two do not -->
    • I4.11.e Compute PSNR, SSIM, and LPIPS with explicit masks/color space/ crop/resolution/aggregation and not-applicable state when references are absent. Evidence: metric fixtures. <!-- evidence: splat-metrics.ts — peak signal to noise and structural similarity are arithmetic and are computed here, against values that are facts: no error at all is infinity rather than a large number, one level out over a whole picture is 48.1308 dB exactly, a picture is itself at exactly one, and two structureless pictures a tenth of a level apart come out at 22006.5025/22106.5025, which is the luminance term with its constant worked out by hand. A LEARNED METRIC WITHOUT ITS WEIGHTS IS NOT A METRIC: LPIPS is taken only from the backend the target names, a report from another backend is refused rather than used, and a target that gates on it with nothing behind it is told that its gate is not a gate. A REFERENCE THAT IS NOT THERE IS NOT A PASS -->
    • I4.11.f Measure encoded size, decoded CPU/GPU memory, peak VRAM, upload/ load time, render frame time, and target point budget. Evidence: resource benchmarks. <!-- evidence: splat-resources.ts — A BUDGET IS PAID IN BYTES SOMEBODY COUNTED: the file is the file, the decoder holds the row it read once per splat, and the card holds a position, an opacity, the six independent numbers of the covariance and every coefficient of every channel — fifty-two bytes a splat at degree zero and two hundred and thirty-two at degree three, half that at half precision. A TIME IS A MEASUREMENT: a delivery that brought no benchmark has no times, and they are left null rather than worked out from a bandwidth somebody assumed; where a peak was really measured, the bigger of it and the counted copy is what the budget has to cover -->
    • I4.11.g Test required web/mobile/native/engine renderer feature/format/ precision/SH/compression/device compatibility and degraded modes. Evidence: renderer compatibility matrix. <!-- evidence: splat-compatibility.ts — A RENDERER IS A LIST OF THINGS IT CANNOT DO: the declared maximum degree, point count, precisions, compressions and memory against what the file actually needs, per named renderer, with a renderer the delivery names and the target does not know reported as such rather than assumed to work. A FALLBACK NOBODY IMPLEMENTED IS NOT A FALLBACK — a degraded mode counts only where the renderer declares it — and the card's copy is then counted at the degree it would really carry, which is what makes a scene fit on a card the untouched file would not -->
    • I4.11.h Emit raw geometry/quality/resource/compatibility results with benchmark versions separately from verdicts. Evidence: splat gate contracts. <!-- evidence: splat-report.ts — A MEASUREMENT IS NOT A VERDICT: the raw record is per splat, per camera and per renderer and carries no severity, remediation, bound or outcome (gated on the schema's own interface bodies AND on the SERIALIZED record), while the verdict carries the thing, the measurement, the bound and the repair, and the camera set and benchmark versions travel with the readings taken under them. Every one of the six families of check is shown to reach the verdict from a single delivery broken in all of them. AN EVALUATOR WITHOUT ITS SUBJECT IS NOT A PASS — a cloud with no volume, a spread of nothing, a camera with no reference frame, a learned metric with no backend and a delivery nobody benchmarked are all carried as not measured — and a file that cannot be read is refused with its reason rather than measured as an empty cloud, which would have no outliers, no needles and no memory, and would pass -->
  • I4.12 Implement scene-package reference, dependency, unit, axis, and missing-asset checks. Evidence: broken package fixtures. <!-- evidence: evidence/v1-workbenches/package-checks/ — a COMPILING strict-tsc reference (schema/digest/paths/parse/graph/security/dependencies/ conventions/manifest/report + fixtures + a 167-assertion suite), gated three ways by verify:inventory:v1-workbenches (generator --check, tsc --noEmit, and the suite run under tsx). The corpus is a real archive: a local header in front of every member, a central directory at the end, an end-of-directory record after that, every member stored plain and padded so its data starts on a boundary a reader can map it at — and a real manifest inside it naming the entry scene, the conventions and the references. Both digests are the real ones, checked against their own published vectors (SHA-256 of "abc" and the CRC-32 check value). 61 of 61 mutations to the implementation are caught by the suite -->

    • I4.12.a Parse package/manifest/scene formats with archive/path/symlink/ entry/size/depth/reference/resource protections. Evidence: hostile package tests. <!-- evidence: package-parse.ts + package-paths.ts — A PATH IN AN ARCHIVE IS A CLAIM, NOT A LOCATION: every name is normalized before it is judged, both separators are separators, and an entry that walks out of the package, names a drive, names a server, or so much as mentions a parent (even one that resolves back inside) is refused rather than followed. THE DIRECTORY IS THE TRUTH AND THE HEADERS ARE A COPY, so the name, the method and the checksum are compared between the two and a pair that disagree is refused rather than one of them being picked — which is the whole of the parser- differential class of attack. A DECLARED SIZE IS NOT A SIZE: the budget is spent against the total the headers CLAIM before anything is unpacked. 14 refusals, each reached by its own hostile fixture, plus the alignment rule a package meant to be mapped where it lies has to keep -->
    • I4.12.b Build a canonical dependency graph for scenes, subscenes, variants, geometry, materials, textures, rigs, animation, audio, scripts, plugins, and external refs. Evidence: graph fixtures. <!-- evidence: package-graph.ts — twelve kinds of node, edges taken from the references as each file writes them, and then a breadth-first WALK from the entry scene so that every node carries the shortest way to reach it (the texture is two steps out, by way of the material). A file nothing in the manifest so much as mentions is an ORPHAN; a member the manifest declares and the entry scene cannot get to is UNREACHABLE; they are different findings about different files and the suite shows both -->
    • I4.12.c Resolve dependencies only within authorized package/registry scope, pin versions/hashes, and reject absolute/traversal/mutable remote references. Evidence: dependency security tests. <!-- evidence: package-security.ts — nothing here resolves anything on a disk; every reading is on the claim. A HASH PINS THE BYTES AND A VERSION PINS THE INTENT, which are two different promises: a remote reference carrying a version and no digest points at whatever is there tomorrow, and is refused as a reference. Absolute names, names that count their way out of the package, protocols with no registry named, and registries the target does not resolve from are each their own finding — while a reference that counts a parent and lands back INSIDE the package is what every scene format writes, and stays legal -->
    • I4.12.d Detect missing, unreadable, corrupt, hash-mismatched, duplicate, conflicting, cyclic, unsupported, or unused dependencies with paths. Evidence: broken dependency fixtures. <!-- evidence: package-dependencies.ts — AN UNUSED DEPENDENCY IS NOT A MISSING ONE: one is a file nobody asked for and the other is a request nobody shipped a file for, they are found by walking the same graph in opposite directions, and the suite shows each one NOT being reported as the other. A CYCLE IS NOT A DEEP GRAPH — the cycle finder keeps the path the walk took, and is checked on its own against a shape whose cycle is known. The archive's own checksums catch bytes that changed after they were written; the delivery's SHA-256 pins catch bytes that were swapped for other perfectly valid ones -->
    • I4.12.e Reconcile file/object/scene units, axes, handedness, up/forward, transforms, frame/sample rates, and color spaces. Evidence: convention mismatch fixtures. <!-- evidence: package-conventions.ts — AN AXIS IS DECLARED, NOT DISCOVERED, and there are three places a declaration can come from: the member, the package, or the target. A member that declares nothing INHERITS the package's, and the inheritance is RECORDED — because a model authored the other way up that says nothing looks exactly like one authored correctly that says nothing, and the only difference is that somebody has to go and look. Seven conventions per member, each resolved with its source; a rate agrees within a tolerance rather than by being equal to a number -->
    • I4.12.f Validate package manifest completeness, entry scene, target versions, licenses/attribution, provenance, and deterministic inventory. Evidence: manifest tests. <!-- evidence: package-manifest.ts — A LICENCE IS A FACT ABOUT A FILE, NOT ABOUT A PACKAGE: a package-level licence covers what its authors made, and a member that carries none is reported even inside a package that carries one. The inventory digest is SHA-256 over path, size and checksum, sorted — so the same members in a different order are the same package, and the same names over different bytes are not, both of which the suite shows -->
    • I4.12.g Emit per-dependency/convention raw state, impact graph, verdict, and remediation. Evidence: package gate contracts. <!-- evidence: package-report.ts — A MEASUREMENT IS NOT A VERDICT: the raw record is per entry, per reference and per convention and carries no severity, remediation or outcome (gated on the schema's own interface bodies AND on the SERIALIZED record), while the verdict carries the entry, the reference, the measurement and the repair. Every one of the five families of check is shown to reach the verdict from a single package broken in all of them. AN EVALUATOR WITHOUT ITS SUBJECT IS NOT A PASS — a delivery that pins nothing, a reference out to a registry, and a member that declares no conventions of its own are all carried as not measured — and a package that cannot be read is refused with its reason rather than measured as an empty one, which would have no cycles, no orphans and no missing dependencies, and would pass -->
  • I4.13 Implement delivery-target profiles for web, mobile, broadcast, game engine, DCC, and marketplace. Evidence: one asset can pass one target and fail another for known reasons. <!-- evidence: evidence/v1-workbenches/target-profiles/ — a COMPILING strict-tsc reference (units/schema/colour/budget/timecode/frames/catalogue/registry/ evaluate + six authored profile families + fixtures + a 367-assertion suite), gated three ways by verify:inventory:v1-workbenches (generator --check, tsc --noEmit, and the suite run under tsx). A TARGET DOES NOT MEASURE, IT DECIDES: nothing here reads an asset — the numbers came from the I4.5-I4.12 gates — so one record of measurements goes to every target unchanged and each reaches its own verdict over it. The hero scene package PASSES the engine and the marketplace and FAILS the web (180 MB before first paint), the handset (BC7, 4096, 85 MiB, 240k triangles and six weights a vertex) and the Maya pipeline (a Blender scene), and the European television master PASSES EBU R128 and FAILS ATSC A/85 on four counts — one of them the same true peak, −1.4 dBTP, which is inside one ceiling and two tenths of a decibel over the other. 104 of 104 mutations to the implementation are caught by the suite -->

    • I4.13.a Define immutable target-profile identity/version/lifecycle, supported asset kinds, formats, constraints, mandatory/advisory gates, packaging, metadata, and policy references. Evidence: profile schema. <!-- evidence: profile-schema.ts + profile-catalogue.ts + profile-registry.ts + profile-units.ts — identity is a target and an exact three-part version; lifecycle is draft/active/deprecated/retired and all four are published so each refusal is reachable. AN EDITED PROFILE IS A NEW PROFILE: the digest is SHA-256 over the canonicalized content and an entry that no longer hashes to the one it carries is refused rather than used. A constraint names a measurement in the 52-entry catalogue of what the earlier gates produce, one of 16 rule kinds, an obligation (mandatory by default, so nothing becomes advisory by being written carelessly) and a policy reference; publishing refuses 13 ways, including a rule asked of the wrong kind of measurement and A LIMIT IS A UNIT — a ceiling written in metres for a triangle count, or texture memory held to a number of pixels -->
    • I4.13.b Author web profiles for browser/runtime/codec/format/color/ dimensions/performance/download/accessibility/security constraints. Evidence: web profile fixtures. <!-- evidence: profiles-web.ts (16 constraints) + profile-colour.ts — codecs every evergreen browser decodes (AVC/VP9/AV1, and not HEVC), a 12 MB first-view budget, the 8192 upload limit, WebGL 2 or WebGPU, a text alternative, a caption track WHEN there is speech to caption, no embedded scripts and no external references. A COLOUR IS NOT A NAME: gamut containment is built from the primaries and the white point through XYZ, and the construction predicts the luminance weights the standards publish (0.2126/0.7152/0.0722 for Rec.709, 0.2627/0.6780/0.0593 for Rec.2020) and the published BT.2020-to-BT.709 matrix. A BOUND IS TWO-SIDED, so clipping to white and clipping to black are measured separately — which is what shows when two spaces disagree about white itself -->
    • I4.13.c Author mobile profiles for OS/device/GPU/memory/package/codec/ texture/audio/network/store and offline constraints. Evidence: mobile profile fixtures. <!-- evidence: profiles-mobile.ts (11 constraints) + profile-budget.ts — an ASTC/ETC2 sampler that has never heard of BC7, 2048 maps, a 64 MiB texture budget, four weights a vertex, stereo, a 200 MB cellular cap, an offline manifest, a store age rating, and a minimum OS read as NUMBERS so 17.10 is after 17.9. A DEVICE BUDGET IS SPENT BY THE WHOLE ASSET, NOT BY ITS BIGGEST PART: memory is summed over whole blocks and over the mip chain, which costs MORE than the textbook four thirds (699,064 bytes for a 1024 BC1 map, not 699,050.67) — and the desktop set is over the budget while no single map in it is -->
    • I4.13.d Author broadcast profiles for standard/raster/frame/interlace/ color/HDR/audio/loudness/caption/timecode/slate/container constraints. Evidence: broadcast fixtures. <!-- evidence: profiles-broadcast.ts (two profiles, 12 constraints each) + profile-timecode.ts — EBU R128 at 25 interlaced with EBU-TT-D captions, −23 ±0.5 LUFS and −1 dBTP; ATSC A/85 at 30000/1001 with CEA-708, −24 ±2 LKFS and −2 dBTP. A FRAME RATE IS A RATIONAL NUMBER: 30000/1001 is not 29.97 and the comparison is a cross-multiplication, drop-frame exists because of the difference, and the arithmetic is proven against the standard's own numbers — 17,982 frames in ten minutes, 107,892 in an hour, and every one of 28,572 labels counting back to the frame it labels. HDR delivered to an SDR service is three findings, not one: the transfer goes to the encoder and the peak (983 nits, read out of a PQ code value through ST 2084) goes to the grade -->
    • I4.13.e Author game-engine profiles for engine/version/platform/mesh/ texture/material/rig/animation/LOD/collision/performance/import rules. Evidence: engine fixtures. <!-- evidence: profiles-engine.ts (12 constraints) + profile-frames.ts — 500k triangles, three levels of detail each keeping at most three fifths of the one above, collision, BC formats, powers of two, 256 joints, eight weights, and clips at 30 or 60. A HANDEDNESS CONVERSION IS A MIRROR: the change of basis between two declared frames is computed, and getting from a right-handed authoring application to this left-handed engine has determinant −1 — which a crate survives (advisory, and it is still a hundred times too small) and a skeleton does not (refused), so the same conversion is two constraints with two obligations -->
    • I4.13.f Author DCC profiles for application/version/plugin/scene/unit/ axis/color/reference/cache/round-trip constraints. Evidence: DCC fixtures. <!-- evidence: profiles-dcc.ts (8 constraints) — which application wrote the scene, which version of it (asked ONLY when the application is the one this pipeline opens, so a Blender scene fails once rather than twice), which plugins it needs to resolve its nodes, which colour configuration and cache formats it reads, whether its references are pinned, and the axis and unit conversion into Y-up centimetres. The round trip is proven rather than guarded: both bases are orthonormal by construction, so the suite goes there and back over five frame pairs and asserts the identity and that the units come home -->
    • I4.13.g Author marketplace profiles for category/format/package/ technical/rights/metadata/preview/moderation/submission constraints. Evidence: marketplace fixtures. <!-- evidence: profiles-marketplace.ts (13 constraints) — category, a single zip with a manifest, a 2 GB intake, a triangle advisory, licences the store can sell, a description long enough to be one, three previews at 1200 on the short side, no open moderation reports and no third-party marks. The attribution constraint is the conditional one: the LICENCE decides whether the credit is even a question, so the same missing credit is a failure under CC BY and NOT APPLICABLE under CC0, with the reason naming the licence that waived it -->
    • I4.13.h Resolve profiles by exact target/version and prove the same measurements yield independently explainable pass/fail/N/A verdicts across targets. Evidence: cross-target golden suite. <!-- evidence: profile-registry.ts + profile-evaluate.ts — A PROFILE IS A VERSION, NOT A NAME: "latest", "3" and "3.1" are refused rather than rounded to whatever is newest, a retired profile refuses and names its replacement, a draft one refuses because nothing was published, and a deprecated one still decides (the deliveries planned against it were correct when they were made) with its lifecycle on the verdict. The cross-target suite runs one record past five targets and the promo past both broadcasters: every result carries its own measured value, its own limit rendered in the unit that target wrote it in, and its own reason. TWO TARGETS THAT DISAGREE MUST DIFFER IN WHAT THEY ASK is CHECKED rather than assumed — two profiles that asked the same thing under the same obligation and answered differently would be a bug in the evaluator, the detector for it is proven on a positive control, and an outcome nobody reached is not an opinion --> <!-- partial: the golden verdicts live in the suite rather than in a committed snapshot file; there is no runtime here to record one from --> <!-- partial: I4.13 is the profile LAYER — the profiles are authored against the measurements the earlier gate references produce, not against a running pipeline; wiring them to live Isis deliveries needs the runtime -->
  • I4.14 Persist raw measurements separately from policy verdicts. Evidence: threshold changes can recompute verdicts without reprocessing media. <!-- evidence: evidence/v1-workbenches/measurement-records/ — a COMPILING strict-tsc reference (schema/identity/store/privacy/evaluate/recompute + fixtures + a 154-assertion suite), gated three ways by verify:inventory:v1-workbenches. The parent's evidence is asserted directly: a threshold moves, the verdict is decided again, and every reading in the store is byte-identical afterwards. 70 of 70 mutations to the implementation are caught -->

    • I4.14.a Define immutable measurement records with subject revision, gate/evaluator/code/model/config/calibration versions, inputs, units, values, regions, uncertainty, applicability, and integrity hash. Evidence: measurement schemas. <!-- evidence: record-schema.ts + record-identity.ts + record-store.ts — AN IMMUTABLE RECORD IS ONE WHOSE IDENTITY IS ITS CONTENT: the id is a real SHA-256 over the canonical form under the name of its kind, nothing allocates an identifier, and there is no operation anywhere that edits one. AN INTEGRITY HASH THAT COVERS ONLY THE VALUE IS NOT AN INTEGRITY HASH, proven one field at a time — fifteen fields, fifteen distinct identities. A record that says two things (applied with no number, absent with one, absent with no reason) is refused rather than sealed, and the append-only store refuses a record that does not hash to what it says -->
    • I4.14.b Store measurement artifacts and sensitive derived data with classification, authorization, encryption, retention, and deletion rules. Evidence: measurement privacy tests. <!-- evidence: record-privacy.ts — SENSITIVE DERIVED DATA IS NOT THE MEASUREMENT: a loudness reading is a number and the decoded programme is a recording of somebody; a skin-tone reading is a number and the crop is a photograph of a face. Four classifications, four purposes, a key per artifact, a retention window in days, and six typed denials that tell "you may not" from "not any more" from "it is gone" — deletion outranking expiry because it is the reason. A purpose has to be one the ARTIFACT allows AND one the PRINCIPAL holds, which are two different refusals. A DELETED ARTIFACT IS NOT A MISSING MEASUREMENT: the number survives its picture, and only re-processing becomes impossible -->
    • I4.14.c Define verdict records that reference measurements plus exact gate-registry/target/policy/threshold versions and never copy ambiguous untyped values. Evidence: verdict schemas. <!-- evidence: record-schema.ts + record-evaluate.ts — A VERDICT CITES VERSIONS, NOT VALUES: it holds the identities of the measurements it read and the exact gate-registry, target-profile (name, version and digest), policy and threshold versions, and there is no field anywhere in it that could hold a number taken out of a measurement — which the suite checks on the serialized record by looking for the readings' own values in it -->
    • I4.14.d Evaluate pass/fail/advisory/manual-review/N/A/unavailable/error deterministically from immutable measurements. Evidence: verdict evaluation fixtures. <!-- evidence: record-evaluate.ts — seven outcomes, seven causes, one fixture each, and a precedence in which a verdict is the worst thing anybody said. THE UNCERTAINTY BAND CUTS BOTH WAYS: a reading whose own error bar straddles the limit cannot say which side it is on, so a pass inside the noise is as much a manual review as a failure inside it — a gate that only reviews the failures lets through everything that landed on the friendly side of noise. The edge of the band belongs to the review, because a comparison at exactly the error bar is decided by the last decimal place -->
    • I4.14.e Recompute new verdict identities after threshold/policy changes without mutating measurements or prior verdicts. Evidence: recompute history tests. <!-- evidence: record-recompute.ts + record-store.ts — A RECOMPUTE IS A NEW VERDICT, NOT A CHANGED ONE: the new one names the one it replaces, the predecessor stays byte-identical, two verdicts cannot replace the same predecessor (which would fork the history), and what a subject IS now is the one nothing superseded rather than the one stored last — asserted against a store handed the same two verdicts the other way round -->
    • I4.14.f Require media reprocessing when evaluator inputs/code/model/ calibration change and explain why verdict-only recomputation is insufficient. Evidence: invalid recompute tests. <!-- evidence: record-recompute.ts — AND AN EVALUATOR IS NOT A THRESHOLD: each of code, model, config and calibration moving is detected on its own, names the readings it invalidates, refuses the arithmetic, and explains that deciding again from them would be an opinion about a measurement nobody took. When the media those readings were taken from has been deleted or has aged out of its retention window the plan is neither recomputed nor failed but reprocess-impossible: somebody has to supply it again --> <!-- partial: the plan is over a record store rather than a running pipeline; the reprocessing it demands is somebody else's job to actually run -->
  • I4.15 Version evaluator code, model, calibration corpus, and thresholds. Evidence: every gate result is reproducible or labeled non-reproducible. <!-- evidence: evidence/v1-workbenches/evaluator-versioning/ — a COMPILING strict-tsc reference (schema/pins + four registers + binding + lifecycle + reproducibility + fixtures + a 226-assertion suite), gated three ways by verify:inventory:v1-workbenches. The parent's claim is computed rather than asserted: certifyGateResults returns everyResultAccountedFor over the reports, and the only verdict that could falsify it — a nonreproducible result naming no source — cannot be produced, because reproduce() throws instead. 38 of 38 mutations to the implementation are caught -->

    • I4.15.a Register evaluator identity/version, source/artifact hash, build/runtime/container, dependencies, configuration, supported inputs, and lifecycle. Evidence: evaluator registry schema. <!-- evidence: version-pins.ts + evaluator-registry.ts — A NAME THAT DOES NOT RESOLVE IS NOT A VERSION: source and artifact digests, compiler, flags, an image pinned by @sha256: rather than a tag, the dependency closure, a digest per configuration and the media kinds it accepts, all folded into one admission digest. A FLOATING REF IS A NAME, NOT A VERSION — latest/main/HEAD, a semver range, a wildcard segment and a bare tag are refused at REGISTRATION, not discovered at replay. A version means one thing for ever, so registering a second, different thing under it is refused while re-registering the same thing is idempotent; and standing is kept out of the digest, so retiring a build leaves what it computes untouched while a compiler bump does not. Declarations must describe the thing: a CPU evaluator may not claim a GPU's excuses, a GPU one must bind the kernel switch it admits to needing -->
    • I4.15.b Register model/checkpoint/preprocessor/tokenizer version/hash/ source/license/security/evaluation and runtime compatibility. Evidence: model dependency tests. <!-- evidence: model-registry.ts — THE PREPROCESSOR IS PART OF THE MODEL: a checkpoint is a function from tensors to tensors and the thing that makes those tensors decides which frame it saw, so a score taken under one preprocessor and pinned to another is refused rather than footnoted — that number is about a different function. Six ways a runtime can be wrong come back together instead of one at a time. A LICENCE IS A RUNTIME FACT: weights granted for evaluation may not run in a commercial gate. A PROVIDER ATTESTATION IS A PROMISE, NOT A PIN — a hosted model pins to provider:<build>, may not also claim a checkpoint digest, and must be reachable through a hosted runtime -->
    • I4.15.c Register calibration corpus version/manifest/hashes/license/ consent/coverage/labels/splits and controlled-access location. Evidence: corpus registry tests. <!-- evidence: corpus-registry.ts — A CORPUS IS PEOPLE. The manifest digest is an order-independent fold over item digests, so re-ordering the file does not restate the corpus and relabelling one photograph does. A HOLDOUT THAT SHARES AN ITEM WITH THE CALIBRATION SPLIT IS NOT A HOLDOUT, refused by item and naming both splits. COVERAGE IS DECLARED, NOT INFERRED: strata are claimed, counted in the calibration split against a floor, and the gap is reported where the corpus is actually thin — the two darkest Monk tones, one of which is not in the calibration split at all. Consent carries a purpose, an expiry and a count of who actually gave it; a location anybody may read is a redistribution whatever the ticket calls it; and no fabricated statistics — an inter-labeller agreement one person reached alone is refused -->
    • I4.15.d Register threshold/rubric/aggregation/uncertainty/manual-band/ target-profile versions with approval and effective dates. Evidence: threshold schema tests. <!-- evidence: threshold-registry.ts — AN EFFECTIVE DATE IS NOT A REGISTRATION DATE. All five — thresholds, rubric, aggregation, uncertainty policy, manual-review band — are versioned together with the target profile, an approval naming who and where it is written down, and a half-open window. A set may not take effect before it was approved, its window may not be empty, and TWO SETS IN EFFECT AT ONCE IS A REFUSAL rather than a conflict to resolve by sort order. The registry answers a profile AND a moment: the suite registers a newer, more-recently-approved set and shows the OLDER one deciding a run in June -->
    • I4.15.e Bind all exact versions and deterministic runtime controls to each measurement; record hardware/provider/environment where relevant. Evidence: execution binding tests. <!-- evidence: execution-binding.ts — the four strings on an I4.14 measurement are parsed into coordinates, resolved against all four registers, checked against what the evaluator says it is made of, and folded into one binding digest. RECORDED IS NOT BOUND: all eight environment fields are written down, only the ones the evaluator declared itself sensitive to enter the digest — so a gated sum is the same run on another chassis while a convolution is a different run after a driver upgrade. BINDING MORE THAN MATTERS IS AS BAD AS BINDING LESS: exactly the declared controls, one short and one over being the same refusal from opposite sides. The reading's own identity is deliberately NOT in the digest, so two subjects measured under one setup share a binding -->
    • I4.15.f Reproduce golden results within declared tolerances or label the measurement nonreproducible with the specific uncontrolled source. Evidence: reproducibility suite. <!-- evidence: reproducibility.ts — A TOLERANCE WIDER THAN THE DECISION BAND PROVES NOTHING, in two distinguishable ways: it reaches past the boundary nearest THIS reading, or it is wider than the whole review band and so could pin no reading's verdict. Either way the answer is INCONCLUSIVE — a third word, kept out of the pass count on purpose — while a tolerance of exactly zero stays a legitimate bit-exactness claim. A DIVERGENCE MUST NAME THE UNCONTROLLED SOURCE: it is attributed to declared sources that individually bound it, and one larger than any of them, or under a claim of none, is thrown as a false determinism claim rather than filed as a caveat — which is why nonreproducible has no path that does not name something. A comparison across two different bindings is refused as not-a-replay rather than blamed on the code -->
    • I4.15.g Prevent retired/quarantined/unlicensed/unadmitted dependencies from new runs while retaining historical resolvability. Evidence: lifecycle tests. <!-- evidence: lifecycle.ts — LIFECYCLE IS A GATE ON THE FUTURE, NOT AN ERASURE OF THE PAST: draft, retired and quarantined stop the next run; resolve and admissible are two functions rather than one with a flag, so what took a reading in 2026 still answers in 2029, down to the compiler. Deprecation gates nothing, because "there is a newer one" is a fact about the catalogue. QUARANTINE IS A DOUBT ABOUT THE PAST; RETIREMENT IS NOT — quarantine alone reaches backwards and marks the readings taken under it suspect, carrying the finding rather than a flag, and without editing records whose identity is their content. An unlicensed model and an out-of-consent corpus are inadmissible for a different reason than a withdrawn one, and say which --> <!-- partial: the register is a store rather than a running admission controller; nothing here stops a process that declines to ask it -->
  • I4.16 Add uncertainty and manual-review bands. Evidence: borderline cases never become false automatic passes. <!-- evidence: evidence/v1-workbenches/uncertainty-bands/ — a COMPILING strict-tsc reference (schema + representations + propagation + bands + routing + aggregation + presentation + fixtures + a 209-assertion suite), gated three ways by verify:inventory:v1-workbenches. The parent's evidence is the one mechanism by which a borderline reading becomes a confident one: there is no path anywhere here from an absent error bar to a number, requireHalfWidth throws instead, and the generator greps both arithmetic modules for a ?? 0. 36 of 36 mutations to the implementation are caught -->

    • I4.16.a Define permitted uncertainty representation per measurement: confidence interval, calibrated probability, variance, quality flag, unknown, or unavailable. Evidence: uncertainty schema tests. <!-- evidence: uncertainty-model.ts — UNKNOWN IS NOT ZERO. Six representations, three of which have a width and three of which do not, and the only function that returns one THROWS for those three rather than defaulting. AN UNCALIBRATED SCORE IS NOT A PROBABILITY: a probability names its calibration method and states that calibration's own error, or it belongs here as a quality flag. A VARIANCE WITHOUT A DISTRIBUTION IS NOT AN INTERVAL: the coverage factor comes from the published t table (12.706 at one degree of freedom, 2.228 at ten, 4.032 at five and 99%) and from a real inverse-normal beyond thirty, and the suite asserts that a small sample WIDENS the bar by about a third rather than narrowing it. A coverage the table does not carry is refused rather than interpolated, because interpolating it restates the coverage -->
    • I4.16.b Propagate input/reference/evaluator/aggregation uncertainty without treating missing or N/A values as zero confidence. Evidence: propagation fixtures. <!-- evidence: propagation.ts — ONE UNKNOWN POISONS THE COMBINATION, AND AN INAPPLICABLE ONE DOES NOT, which are two different absences given opposite treatment. The arithmetic is ordinary (quadrature at ρ=0, a plain sum at ρ=1, the general form between, all three asserted to order correctly); what is not ordinary is that a term with no width makes the WHOLE combination unknown and names the term, because √(0.1²+0²) is indistinguishable from having combined everything — the suite pins the exact number that must not come out. A term that did not apply is dropped and listed, since there was no value either. Bars in different units or at different coverages do not combine at all, and two readings one of which has no bar neither agree nor conflict -->
    • I4.16.c Define nonoverlapping automatic fail, manual-review, and automatic pass bands per gate/profile/version with boundary inclusivity. Evidence: boundary tests. <!-- evidence: bands.ts — THE BANDS MUST TILE THE LINE and EVERY BOUNDARY BELONGS TO EXACTLY ONE SIDE. Four distinct refusals at registration — a gap, an overlap, an edge both sides claim and an edge neither does — because they are four different mistakes, and the last is a hole exactly one value wide at exactly the value people tune to hit. A target window runs fail/review/pass/review/ fail under the same rules, so the loudness gate has two review strips. And A READING THAT SPANS A BOUNDARY CANNOT SAY WHICH SIDE IT IS ON, symmetrically: the bar reaching OUT of the pass band is as much a review as the one reaching out of the fail band. decisionGeometry computes what I4.15.f consumes -->
    • I4.16.d Route uncertainty/quality/conflict/missing-dependency conditions into manual review even when the point estimate crosses pass. Evidence: borderline fixtures. <!-- evidence: routing.ts — A CAUSE OUTRANKS A POINT ESTIMATE. Each of an unknown bar, a poor quality flag, two evaluators separated by more than their combined allowance, and a dependency that produced nothing is asserted on a reading sitting ten times its own error bar inside the pass band — which is the case a gate written around failures never checks and the case that lets everything through. A CAUSE CANNOT RESCUE A FAILURE, AND IT CANNOT CONFIRM A PASS: the asymmetry is deliberate, since doubt about a reading over the limit does not make it safe and a queue of definite failures is a queue nobody reads. A GATE THAT DID NOT APPLY HAS NO BAND, AND A GATE THAT DID NOT RUN IS NOT A PASS -->
    • I4.16.e Prohibit aggregation from averaging a manual-review/unknown mandatory gate into automatic pass. Evidence: aggregate bypass tests. <!-- evidence: aggregation.ts — A MANDATORY GATE MAY NOT BE AVERAGED. The bypass is always the same and always looks defensible: weight every gate, sum, compare. So a rule that gives a mandatory gate a weight is refused at the RULE rather than at the result, and there are two independent barriers — the advisory mean also skips mandatory gates even when handed such a rule directly. The advisory score is still computed and reported, and can only ever ask for a person, never excuse one. A gate that did not apply is excluded rather than scored 1.0, and a subject where no mandatory gate applied is not entitled to a pass at all -->
    • I4.16.f Present point value, units, uncertainty, band, threshold, applicability, evidence, and review reason. Evidence: UI/API tests. <!-- evidence: presentation.ts — PRECISION IS A CLAIM ABOUT UNCERTAINTY: the value is rounded to the place its bar reaches by the Particle Data Group's rule, all three branches asserted (0.4 keeps one figure, 0.12 keeps two, 0.0964 rounds up to 0.10 and carries the value's decimal place with it). An unknown renders as "unknown — not ±0" and the suite asserts the string never contains a rendered zero bar; the row is fed the COMBINED uncertainty rather than the reading's own, so a decision taken with no usable width cannot be shown beside a tidy ±0.05. An absent reading renders no number at all. And A REVIEW WITHOUT A REASON IS A QUEUE --> <!-- partial: the row is a view model rather than a rendered surface; no component here consumes it yet -->
  • I4.17 Add batch gate execution with per-gate progress and cancellation. Evidence: partial results are preserved but cannot authorize release. <!-- evidence: evidence/v1-workbenches/batch-gates/ — a COMPILING strict-tsc reference (schema + plan freeze + scheduler + progress + cancellation + aggregate + fixtures + a 151-assertion suite), gated three ways by verify:inventory:v1-workbenches. The parent is two claims pulling opposite ways and both are asserted: a measurement completed a second before a cancellation survives it by exact attempt identity, and neither a partial nor a COMPLETE batch can authorise anything. 29 of 29 real mutations to the implementation are caught -->

    • I4.17.a Define evaluation batch, subject, gate item, dependency, attempt, progress, measurement, verdict, cancellation, and aggregate contracts. Evidence: batch gate schemas. <!-- evidence: batch-schema.ts — eight gate states of which five are terminal, four attempt outcomes, six ordered progress stages, four resource classes, five failure classes of which three are retryable, and typed records for the frozen plan, the gate item (dependsOn/timeoutMs/retryBudget/attempts/ obligation), the attempt, the result, the reservation, the event and the aggregate. Twenty-one refusals, every one thrown somewhere and reached by the suite -->
    • I4.17.b Resolve and freeze exact subjects/revisions/gate registry/ evaluators/target profiles/policies before queueing. Evidence: batch planning tests. <!-- evidence: batch-plan.ts — A FROZEN PLAN IS THE ONLY THING THAT MAY BE QUEUED. A four-hundred-gate batch takes an hour; resolving anything lazily means the first gate and the last gate ran under different rules, invisibly, on the day somebody lands a registry change at lunchtime. So subjects, revisions, registry, evaluator builds, profile and policy are resolved once and hashed (the same canonical digest as I4.14), and queue refuses a plan that no longer hashes to what it froze rather than repairing it — repairing it would mean choosing which half of the batch to believe. Dependency cycles are found at plan time and reported as the cycle, because a deadlock in a scheduler is indistinguishable from slowness -->
    • I4.17.c Schedule bounded parallel gate execution by resource class with dependencies, fairness, retries, timeouts, and per-gate isolation. Evidence: scheduler tests. <!-- evidence: scheduler.ts — A SCHEDULER IS A PROMISE ABOUT RESOURCES, NOT ABOUT SPEED: bounds are per class, so a GPU-saturated batch still progresses on its network gates. FAIRNESS IS PER SUBJECT, NOT PER GATE — the fixture is built so a readiness-ordered queue would hand both GPU slots to the one asset holding two GPU gates, and the suite asserts they go to two DIFFERENT assets. A DEPENDENCY IS A GATE THAT MUST HAVE FINISHED, AND A FAILED ONE IS NOT A FINISHED ONE: the dependent is skipped with the reason, which the suite checks is not a failure of its own. A RETRY IS A NEW ATTEMPT with its own epoch and identity, only for transient/timeout/resource-exhausted, and the exhausted-budget message is told apart from the never-retryable one -->
    • I4.17.d Emit monotonic per-gate stage/percent/measurement-ready/verdict- ready progress with event id/cursor/deduplication. Evidence: progress stream tests. <!-- evidence: progress.ts — PROGRESS IS MONOTONIC, AND AN EVENT IS DELIVERED ONCE. Percent may not decrease, stages may not regress in the declared order, a readiness flag may not un-set, and all four are refused by name rather than clamped — clamping hides the bug that produced the regression. An identifier IS its cursor, so a reconnecting consumer asks for everything after the last it saw; re-emitting the same stage returns the SAME event rather than a second one, since a retrying emitter must not double a stream somebody is deduplicating. measurement-ready and verdict-ready are separate flags because I4.14 keeps the reading apart from the opinion about it -->
    • I4.17.e Cancel queued/running gates with evaluator propagation, late- result fencing, reservation release, and explicit cancellation outcome. Evidence: cancellation race tests. <!-- evidence: cancellation.ts — cancellation is a race by construction: at the moment somebody presses stop, evaluators are mid-decode and will produce good numbers seconds later. Both naive answers are wrong, so ONE comparison against the cancellation instant decides both — and a result from a superseded epoch is fenced the same way. Every running gate's slot is released as it is cancelled, a completed gate is NOT un-completed by the batch stopping, and a cancellation with no reason is refused because it leaves everybody guessing whether it was a person or a crash -->
    • I4.17.f Persist valid completed measurements/verdicts from partial runs, failures, and cancellations with exact attempt identities. Evidence: partial preservation tests. <!-- evidence: cancellation.ts — what finished first is kept, by attempt identity and epoch, so a measurement from attempt two is never confused with the one attempt one failed to produce. A result is written ONCE: a second, different measurement for one attempt is refused, which is what makes "preserved" mean something — a store that can be overwritten cannot promise to have kept anything. The suite cancels mid-flight and asserts the earlier measurement is still there by name while the late one never entered the store -->
    • I4.17.g Compute aggregate state with mandatory/advisory/N/A/unavailable/ error/cancelled counts and never emit release authorization. Evidence: aggregate authority tests. <!-- evidence: aggregate.ts — eleven counts and a completeness word, and NO field that could be read as an authorisation. That is checked three ways: the generator greps the interface body for ten banned names, the suite greps the SERIALIZED object for them, and carriesNoAuthorization is itself proven to catch each one by being handed an object that has it — because the way this fails is that somebody adds a convenient boolean next quarter. decisionReadiness answers the narrower honest question (is there enough for the AUTHORITY to decide) and requireReleaseAuthorization throws for anything that asks this module for a yes or no, including over a batch where every gate completed --> <!-- partial: the scheduler is a pure step function over an in-memory state rather than a running worker pool; nothing here executes an evaluator -->
  • I4.18 Add accessible quality comparison views. Evidence: raw values, thresholds, units, status, and remediation are available without color or canvas. <!-- evidence: evidence/v1-workbenches/comparison-views/ — a COMPILING strict-tsc reference (schema + status encoding + projection + summary/detail + regions + comparison + remediation + fixtures + a 124-assertion suite), gated three ways by verify:inventory:v1-workbenches. a-f are done; the PARENT stays unchecked because its evidence is about what a reader can get from a rendered surface, and I4.18.g is the item that would establish that --> <!-- partial: I4.18.g (screen reader, keyboard, zoom/reflow, contrast, reduced motion, large-set performance) needs a real browser over a real page. A view model cannot stand in for one, and the generator says so in notClaimed rather than quietly claiming the parent -->

    • I4.18.a Define a typed view model joining exact asset revision, target profile, gate registry, measurement, verdict, uncertainty, override, and evaluation run. Evidence: projection tests. <!-- evidence: projection.ts — A JOIN IS A CLAIM THAT TWO THINGS ARE ABOUT THE SAME SUBJECT. Six things join into a row and each is a place to assemble one out of parts that do not belong together — a verdict about another revision or another asset, a verdict that never cited this reading, a run under another profile or registry, an override attached to a different verdict. None of those looks wrong on screen, so each is refused rather than rendered. AN OVERRIDE IS AN OPINION WITH A NAME ON IT: it cannot edit an I4.14 record and does not try, the row carries both, an accepted failure comes out as neither that failure nor a pass, and an acceptance whose date has passed stops applying -->
    • I4.18.b Present summary counts/status with text/icon/pattern plus filter/ sort/search and explicit mandatory/advisory/N/A/unavailable distinction. Evidence: component/accessibility tests. <!-- evidence: status-encoding.ts + summary.ts — A STATUS THAT IS ONLY A COLOUR IS NOT A STATUS, and the check made here is the stronger one: the usual defence is "we also have an icon" and then five statuses share a circle, so EACH of the word, the shape and the fill pattern must separate ALL seven on its own and any collision in any channel is a refusal. No colour name appears in the module at all — colour is a fourth channel a theme may add, and nothing here depends on it. Counts are per status AND split by obligation, because "three failures" is a different sentence depending on whether they were mandatory; and the two absences I4.16 separated stay separate in every channel -->
    • I4.18.c Present every raw value, unit, evaluator/version, threshold/ band/operator, uncertainty, applicability, timestamp, and verdict without relying on color. Evidence: detail view tests. <!-- evidence: summary.ts — thirteen required labels, checked as a list, so that the detail lines read aloud in order ARE the row: gate, obligation, status, asset, metric, unit, uncertainty, applicability, evaluator build, threshold, band, timestamps and verdict, plus the acceptance and the evidence when there are any. A NUMBER WITHOUT ITS UNIT IS NOT A VALUE: localisation moves the group separator and the decimal mark across three locales and never touches the unit symbol, and a value rendered with no unit is refused -->
    • I4.18.d Provide accessible timecoded/spatial/mesh/audio regions using synchronized text tables/lists whenever visual canvas is unavailable. Evidence: noncanvas navigation tests. <!-- evidence: regions.ts — A CANVAS IS AN ENHANCEMENT, NOT A CHANNEL. A flicker at 00:04:12:07, a 288x288 patch at 480,220, vertices 1204-1336, samples 88200-92000 of the left channel: all numbers, and the waveform and the mesh viewer are ways of drawing them. Each of five kinds declares the coordinates it needs AND the ones it may not carry (a mesh region with a timecode was built from the wrong template), and tableIsComplete checks the table determines the regions on its own — two regions rendering one row is a table that has lost one -->
    • I4.18.e Compare asset revisions/targets/evaluation versions without mixing measurements and explain added/removed/changed gates. Evidence: comparison fixtures. <!-- evidence: comparison.ts — A COMPARISON MUST NOT MIX MEASUREMENTS. Revision 6 read -23.4 and revision 7 reads -23.1, and everybody goes looking at the mix — unless the meter was upgraded between the runs, in which case the 0.3 is an unknown mixture of the asset changing and the METER changing, which is the entire question anybody was asking. So a delta is computed only under a matching evaluator build, and a moved build comes back incomparable with NO number rather than a number with a warning beside it, because a number with a warning beside it is a number people quote. A gate the registry gained is added rather than a regression, and every gate on either side appears exactly once -->
    • I4.18.f Present authorized remediation, responsible owner, rerun/review action, and immutable evidence links with keyboard-safe focus behavior. Evidence: remediation journey tests. <!-- evidence: remediation.ts — REMEDIATION IS AUTHORIZED OR IT IS A SUGGESTION. Hiding the buttons a viewer may not press makes a tidy screen and makes the interface lie; the reader then asks in chat what happened to it. So the action is shown, marked unavailable, naming the capability it needs — which turns "why can't I fix this" into "who has isis.gate.rerun" — and it stays in the tab order. FOCUS ORDER IS PART OF THE CONTRACT: total, ordered, duplicate-free, with a declared return target for everything that opens, because focus that lands nowhere strands a keyboard user at the top of a four-hundred-row table. The evidence links are the row's immutable ones -->
    • I4.18.g Verify screen reader, keyboard, zoom/reflow, contrast, reduced motion, localization/number units, and large gate-set performance. Evidence: automated accessibility/performance tests.
  • I5.1 Enforce model-card and model-version license completeness before dispatch. Evidence: missing or incompatible rights fail closed. <!-- evidence: evidence/v1-workbenches/model-licensing/ (compiling strict-TS reference, 271 assertions, 101/101 mutants caught) + scripts/v1-workbenches/ generate-model-licensing.mjs (coherence + drift gate, 11 tests). Every refused path is held to the only claim that matters: the provider log is empty. Not claimed: the wiring into apps/isis/generation-api -->

    • I5.1.a Define required model-card/version license fields for source, license identity/text/version, owner, allowed/prohibited uses, commercial use, derivatives, attribution, territory, term, and evidence. Evidence: license schema tests. <!-- evidence: license-schema.ts — ABSENCE IS NOT A PERMISSION. ModelCard.licenses and ModelVersion.licenses are Json @default("[]") columns, so today a model nobody has done the paperwork for and a model whose terms forbid everything are the same value; and the nearest typed policy in the tree (ThreeDAssetResolvedLicensePolicy) spends its dimensions as allowsCommercialUse: boolean, where a licence that never mentions commercial use cannot be told from one that forbids it. Every permission here is three-valued and silence is a value the record must state out loud. AN UNRECOGNISED TERM IS NOT A PERMISSION: a clause reading "see appendix B" is refused rather than rounded towards yes. Fourteen fields, each deleted in turn by the suite; perpetual and not-revoked are positive statements with no null to be mistaken for silence -->
    • I5.1.b Validate exact model/version/adapters and all transitive licensed dependencies have complete verified nonexpired metadata. Evidence: completeness tests. <!-- evidence: completeness.ts — THE CLOSURE IS THE LICENCE. Nobody runs a checkpoint on its own and the non-commercial clause is always in the adapter somebody pulled off a hub, so the terms are resolved over the whole transitive closure and an uncatalogued dependency is refused rather than skipped. A TERM HAS TWO ENDS: not-yet-effective, expired and revoked are three refusals, and a licence that is both revoked and expired is revoked, because expiry was in the text from the beginning and revocation is the licensor changing their mind. Verification is by an approved authority within a freshness bar rather than by anybody at any time -->
    • I5.1.c Evaluate license compatibility against actor, tenant, project, workflow, input/reference rights, intended use, target, territory, and distribution terms. Evidence: compatibility matrix. <!-- evidence: compatibility.ts — RESTRICTIONS MEET AND OBLIGATIONS UNION, in opposite directions: one adapter forbidding commercial use makes the whole run non-commercial and no permissive checkpoint buys it back, while an attribution owed to one upstream is not discharged by another upstream not asking for one. A CHOICE IS BETWEEN WHOLE LICENCES, NOT BETWEEN THEIR CLAUSES, so nobody takes the commercial grant from one and the attribution waiver from the other. THE REFERENCE IMAGE IS PART OF THE LICENCE — every model can be clean and the run still refused. THE TARGET DECIDES WHETHER IT IS COMMERCIAL rather than the operator's checkbox; THE LICENSEE IS THE TENANT rather than the operator; and a licence is bound to where the work goes rather than where the operator sits. Nine dimensions, and every refusal names the node that imposed it -->
    • I5.1.d Revalidate licenses/revocations/effective dates immediately before dispatch and bind the decision/evidence versions to the run. Evidence: stale-license race tests. <!-- evidence: dispatch.ts — A DECISION IS AS OF AN INSTANT. The plan was cleared on Tuesday, the licensor pulled the model on Wednesday, the queue drains on Thursday, and every one of those is a normal day. So the plan decision is an input and never an authority: the decision is taken again against a fresh read at the dispatch instant, a read that saw different evidence supersedes rather than agrees, and the run binds the decision, the policy version and the exact evidence versions it was read against. A CLEARANCE CANNOT OUTLIVE THE LICENCE IT RESTS ON -->
    • I5.1.e Return missing/incompatible/expired/revoked/unknown as typed fail-closed outcomes with no provider call. Evidence: zero-dispatch negative tests. <!-- evidence: dispatch.ts + license-schema.ts — A GATE THAT RUNS AFTER THE CALL IS NOT A GATE. There is one path to a provider, it takes a clearance no other module can mint (a unique symbol), and six refused paths are swept with a counting provider whose log must be empty. THE OUTCOME IS A ROUTE AND THE REFUSAL IS A REASON: twenty refusals map onto five fail-closed outcomes by a table that is total by construction and can never route to cleared, and a decision takes its name from the worst thing found rather than the first thing checked -->
  • I5.2 Enforce training-data provenance requirements by workflow risk. Evidence: policy fixtures cover unknown, restricted, and approved data. <!-- evidence: evidence/v1-workbenches/training-provenance/ (compiling strict-TS reference, 236 assertions, 109/109 mutants caught) + scripts/v1-workbenches/ generate-training-provenance.mjs (coherence + drift gate, 12 tests). The fixtures cover unknown, restricted and approved data, and four more besides. Not claimed: the wiring into the live dispatch path, and the ratification of the matrix by a named owner -->

    • I5.2.a Define workflow-risk classes and the minimum dataset/source/ collection/license/consent/opt-out/geography/lineage evidence required for each. Evidence: provenance policy schema. <!-- evidence: provenance-schema.ts — THE BAR MOVES WITH THE WORKFLOW. "Is the provenance of this dataset good enough" is not a question about the dataset: the same corpus is fine for an engineer trying something on a Tuesday and nowhere near good enough for a campaign that puts a real face on a billboard, and a gate with one fixed bar either stops the Tuesday or waves through the billboard. Four rungs over the eight kinds of evidence the item lists, and THE LADDER MUST BE A LADDER — the nesting is checked, because a higher rung that quietly asks for less reads fine and inverts the policy -->
    • I5.2.b Resolve exact model/version/adapter training-data declarations, dataset versions, upstream sources, known exclusions, and evidence freshness. Evidence: provenance resolver tests. <!-- evidence: resolver.ts — A DATASET IS ITS VERSION: the lookup is by identity AND version, an unpinned "latest" is refused at the door, and an attestation scoped to another version does not count towards this one — otherwise a corpus cleaned last month certifies a model trained before the cleaning. AN OPT-OUT IS A DATE, NOT A FLAG: a sweep that ran before the snapshot leaves everything that arrived in between, while the declaration says "opt-out honoured" the whole time. Expiry, revocation and age are three separate ways for a reading to stop counting -->
    • I5.2.c Classify each dependency as approved, restricted, unknown, incomplete, expired, revoked, or conflicting without treating absence as approval. Evidence: classification fixtures. <!-- evidence: classification.ts — ABSENCE IS NOT APPROVAL: there is no path from nothing-on-file to approved. All seven classes, and TWO ANSWERS ARE WORSE THAN NO ANSWER — two live attestations that disagree about the same version make a dataset conflicting, which outranks the absence classes, because one of the two signatures is wrong and nothing on the page says which. TAINT FLOWS DOWNSTREAM AND NEVER WASHES OUT, and the ancestor that caused it is named -->
    • I5.2.d Apply intended-use/audience/identity/style/child/protected-source/ commercial risk rules to the resolved provenance graph. Evidence: risk-policy matrix. <!-- evidence: risk-policy.ts — A RULE CAN ONLY RAISE THE CLASS. All seven rules are escalations naming a rung the workflow must at least stand on, so there is no evaluation order for somebody to refactor and the answer is the maximum however they are shuffled. A RULE THAT DOES NOT NAME ITS SUBJECT IS A RUMOUR: every graph-level escalation carries the dataset that triggered it, and the request-level ones honestly carry none. THE MATRIX IS THE POLICY, AND IT MUST BE MONOTONE — 28 declared cells, approved allowed at every rung, restricted and revoked blocked at every rung, nothing softening as the rung rises -->
    • I5.2.e Block or require named review according to the ratified risk class and persist the exact decision inputs. Evidence: enforcement tests. <!-- evidence: enforcement.ts — A REVIEW WITH NO NAME ON IT IS NOT A REVIEW: the roles that may sign are named per rung, a review with an empty roster is refused, and BLOCKED IS NOT A REVIEW WITH NOBODY FREE — a block names no reviewer at all rather than one who cannot help. A DECISION YOU CANNOT REPRODUCE IS NOT A DECISION: the interesting question a year later is never what it said but why, when the registry has been rewritten twice since, so replay re-derives the action from the record's own nine kept fields with no registry in sight, and a record whose inputs no longer reach its action is refused -->
  • I5.3 Enforce reference-asset rights attestations. Evidence: expired or revoked rights invalidate downstream releases. <!-- evidence: evidence/v1-workbenches/reference-rights/ (compiling strict-TS reference, 170 assertions, 81/81 mutants caught) + scripts/v1-workbenches/ generate-reference-rights.mjs (coherence + drift gate, 13 tests). The digest is a real SHA-256 held to the NIST vectors and the signatures are a real HMAC-SHA256 held to RFC 4231, so every tamper test breaks something rather than being told it did; the asymmetric signature binds through a seam whose production implementation is @oshun/content-signing. Not claimed: the wiring into the live reference-intake path, and a named owner ratifying the recall policy -->

    • I5.3.a Define attestation subject/version/hash, owner/licensor, license/ permission basis, allowed uses, derivative/training rights, attribution, territory, term, restrictions, evidence, and signer authority. Evidence: attestation schema tests. <!-- evidence: attestation-schema.ts — AN ASSET IS ITS BYTES: the subject is an id, a version AND a content digest, and the digest is the one that decides, because a reference re-exported at another quality is a different asset wearing the same filename. A RECORD CANNOT REPORT ITS OWN REVOCATION, so the signed body has no revocation field at all — a licensor withdrawing permission is not going to come back and re-sign the record to say so. A PERMISSION HAS A ROLE: the eight ways a file is actually fed to a model are enumerated, because a grant reading "may be used as a reference" cannot tell a style transfer from a face swap -->
    • I5.3.b Verify attestation signature/authority, referenced asset hash, evidence integrity, effective dates, and current revocation state. Evidence: attestation validation tests. <!-- evidence: verification.ts — A SIGNATURE IS NOT AN AUTHORITY. Two checks that fail for different reasons: the signature verifies against the signer's key, and the SIGNER is checked against a registry of who may sign for that owner at that time. An intern's key verifies exactly as well as the rights desk's, and somebody who left in March cannot sign in April. The asset and evidence hashes are recomputed from the bytes rather than compared between two declarations, because two declarations agreeing with each other is not evidence about a file -->
    • I5.3.c Evaluate scope against the exact conditioning/reference role, workflow/model use, generated derivative, audience, project, target, territory, and publication channel. Evidence: rights matrix. <!-- evidence: scope.ts — A GRANT IS FOR A ROLE, NOT FOR A FILE, and every role a run puts the asset in is checked, including the second one. A DERIVATIVE IS NOT A USE: feeding a reference to a model and publishing what came out are two permissions, which is exactly what a reference-only licence says. A RESTRICTION IS A CLAIM ABOUT THE OUTPUT, NOT ABOUT THE INPUT — no-face-swap is tripped by what the run does, not by the file being opened -->
    • I5.3.d Bind accepted attestation versions to plan/run/output lineage and propagate required attribution/restrictions. Evidence: lineage tests. <!-- evidence: lineage.ts — A LINEAGE IS A CLAIM ABOUT THE PAST, AND THE PAST DOES NOT CHANGE: the binding records the attestation and its version, and rebinding a run to a later version is refused rather than absorbed, because otherwise every historical run appears to have been cleared under terms nobody had read at the time. AN OBLIGATION TRAVELS FURTHER THAN THE ASSET: the reference is not in the output and the credit it earned is, so attributions and restrictions accumulate down the derivation graph and a credit owed twice is paid once -->
    • I5.3.e On expiry/revocation/change, mark dependent release eligibility stale and trigger blast-radius/recall policy without rewriting prior evidence. Evidence: invalidation tests. <!-- evidence: invalidation.ts — REVOCATION REACHES BACKWARDS AND EXPIRY DOES NOT, which is the distinction the whole item turns on: a term ending was always coming, so nothing new goes out and what already shipped shipped under a live licence; a revocation is the licensor saying the permission was never what we thought, and it reaches the releases that already happened. STALE IS NOT INELIGIBLE, because a workbench that renders one as the other teaches everybody to ignore red rows. THE PAST IS APPENDED TO, NOT EDITED: each invalidation writes a new ledger entry and the earlier ones are proven byte-identical by their own digests. The blast radius is everything downstream; the recall is a policy question about the channel each release went out through -->
  • I5.4 Enforce face-synthesis subject consent with scope, purpose, territory, expiry, and revocation. Evidence: consent matrix tests. <!-- evidence: evidence/v1-workbenches/face-consent/ (compiling strict-TS reference, 265 assertions, 74/74 mutants caught) + scripts/v1-workbenches/ generate-face-consent.mjs (coherence + drift gate, 15 tests). The consent matrix is thirteen dimensions over five checkpoints, each enforced at the earliest checkpoint at which it is answerable and re-checked at every later one. The digest is imported from the I5.3 reference rather than written a second time. Verification finding recorded against the shipped libs/isis/face-synthesis/src/lib/ethical-safeguards.ts. Not claimed: the face encoder and detector are real dependencies taken through seams rather than implemented here, the wiring into the live Isis path, and the ages-of-majority table being sourced from counsel -->

    • I5.4.a Define verified subject identity, authorized representative, consent evidence/signature, synthesis operations, reference set, purpose/use, audience, project, territory, term, disclosure, retention, and revocation contract. Evidence: face-consent schema. <!-- evidence: consent-schema.ts — A FACE IS NOT A FILE: I5.3's subject is an asset identified by its bytes and this one is a person, so the reference photographs are evidence ABOUT her rather than the thing consented to. A CONSENT NAMES OPERATIONS, NOT "AI": eight operations are enumerated, because a release saying "may be used with generative tools" cannot tell a scratch removal from a puppeteering. A REVOCATION CONTRACT IS PART OF THE CONSENT — channels, effect latency and reach are required fields — and the signed body carries no withdrawal state, because a subject changing her mind will not re-sign the record to say so. 17 required fields, each deleted in turn by the suite -->
    • I5.4.b Verify subject/representative authority and liveness/identity evidence in the approved host without exposing biometric artifacts. Evidence: verification tests. <!-- evidence: verification.ts + enclave.ts — A SIGNATURE FROM THE SUBJECT IS NOT PROOF THE SUBJECT SIGNED IT (signature and identity assurance fail for different reasons; a self-asserted identity never clears the floor for face synthesis). LIVENESS IS A MOMENT, NOT A PROPERTY: a capture is bound to a challenge nonce and a window, and one that PREDATES the challenge did not answer it however recent. A TEMPLATE NEVER LEAVES THE HOST — the vectors live in a closure behind opaque handles, no exported operation returns one, and the view is assembled from an allow-list without the score, which is an oracle to hill-climb against -->
    • I5.4.c Match detected/declared subject and requested face identity use to exact consent scope before planning, dispatch, comparison, export, and release. Evidence: consent matrix. <!-- evidence: matching.ts + biometrics.ts — A CHECK AT RELEASE IS TOO LATE FOR A FACE ALREADY SENT: the matrix records the earliest checkpoint at which each of 13 dimensions is answerable and requires enforcement there, and the invariant takes the table as an argument so a deferred row can be handed to it and refused. Every gate goes through one door and every later gate asks again. A VERIFICATION THRESHOLD IS NOT AN IDENTIFICATION THRESHOLD — a 1e-4 per-comparison rate over a gallery of 10,000 false-matches 0.6321389535670295 of the time, pinned exactly; A SCORE IS NOT AN IDENTITY, so the band is inconclusive rather than rounded. Zero-dispatch proved by a counting provider on all 10 dispatch-time refusals -->
    • I5.4.d Enforce stricter rules for multiple subjects, minors, public figures, sensitive contexts, cross-purpose reuse, and synthetic subject ambiguity. Evidence: high-risk fixtures. <!-- evidence: high-risk.ts + detection.ts — AGE IS AS OF THE USE, NOT AS OF THE SIGNATURE (leap-day arithmetic exercised on both sides of 29 February) and MAJORITY IS A JURISDICTION'S ANSWER, so an unlisted territory refuses rather than defaulting to eighteen. GUARDIAN AUTHORITY ENDS AT MAJORITY: the day she turns eighteen the consent on file is her guardian's and not hers. BEING PUBLIC IS NOT CONSENTING; A SENSITIVE CONTEXT IS DECIDED BY THE OUTPUT rather than the requester's checkbox; EVERY FACE IN THE FRAME IS A SUBJECT; and AN AMBIGUOUS SUBJECT IS NOT A SYNTHETIC ONE — a generated face matching an enrolled person is that person -->
    • I5.4.e Propagate expiry/revocation to future use and governed downstream availability while preserving immutable historical proof. Evidence: revocation tests. <!-- evidence: revocation.ts — A WITHDRAWAL IS EFFECTIVE WHEN IT IS RECEIVED, NOT WHEN IT IS PROCESSED, or every hour of backlog buys an hour of use she did not agree to; and A WITHDRAWAL THROUGH THE WRONG CHANNEL IS STILL A WITHDRAWAL. AN EXPIRY IS NOT A WITHDRAWAL: a term ending stops the next release and leaves what shipped, and she changing her mind reaches it, with the action chosen per audience. DELETING THE FACE IS NOT DELETING THE RECORD THAT WE HAD IT — retention destroys the templates and photographs and keeps the digests and decisions, because the only thing that can answer "did you have permission in March" is a record made in March; the ledger is digest-chained and refuses a rewritten history -->
  • I5.5 Enforce voice-cloning consent and speaker verification in the authorized host. Evidence: impersonation and scope-escalation tests. <!-- evidence: evidence/v1-workbenches/voice-consent/ (compiling strict-TS reference, 233 assertions, 96/96 mutants caught) + scripts/v1-workbenches/ generate-voice-consent.mjs (coherence + drift gate, 16 tests). The audio arithmetic is real — durations and levels are measured off synthesised PCM and held to published values — and the digest is imported from the I5.3 reference rather than written a second time. Verification finding recorded against the shipped libs/isis/audio-generation VoiceCloneRequest and the libs/isis/music-generation keyword guardrail. Not claimed: the speaker-verification system, the spoofing and replay countermeasures, the claim classifier and the voice model are dependencies taken through seams rather than implemented here, and the wiring into the live Isis voice path is separate -->

    • I5.5.a Define speaker identity, consent grant, voice model/version, reference recordings/hashes, allowed text/use/style/language/audience/ channel/territory/term, disclosure, retention, and revocation. Evidence: voice-consent schema. <!-- evidence: voice-consent.ts — A VOICE CONSENT IS ABOUT WORDS AS WELL AS SOUND, which is the difference from I5.4's face consent and the half that gets left out: a likeness in a shot looks like her, and a voice reading a script is heard as her saying it. So the grant carries a script policy (prohibited claims, whether she speaks in the first person, a length) alongside the uses. AN ENROLLMENT IS NOT A FILE UPLOAD: every reference recording carries an ownership basis, and third-party-unverified is in the vocabulary precisely so "we are not sure where this came from" is expressible and refusable. 21 required fields, each deleted in turn by the suite, which also asserts the message names the missing field -->
    • I5.5.b Perform speaker identity/liveness/authority verification and reference quality/ownership checks only in the authorized host. Evidence: host-bound verification tests. <!-- evidence: audio.ts + speaker.ts + enrollment.ts — A DECLARED DURATION IS NOT A DURATION: the same 2,646,000-byte buffer is thirty seconds of 44.1kHz mono and 9.1875 seconds of 48kHz stereo 24-bit, so an enrollment minimum is computed from the bytes and the format. FULL SCALE IS NOT SYMMETRIC (the rail is 32767/32768, and a detector testing for 1.0 finds nothing in a take that is clipped solid). A SPEAKER VERIFIER CANNOT TELL YOU FROM YOUR OWN CLONE — the load-bearing fact: a model built from her enrollment scores high against it by construction, so verification establishes which voice and never whether a human, and A SPOOF GETS IN AT THE COUNTERMEASURE RATE ALONE (0.05, not the 5e-8 product with a 1e-6 verifier). AN EMBEDDING NEVER LEAVES THE HOST -->
    • I5.5.c Bind trained/enrolled voice artifacts to verified speaker/consent/ source versions and prohibit unauthorized export or model substitution. Evidence: voice artifact lineage tests. <!-- evidence: artifact.ts — A VOICE ARTIFACT IS ITS MODEL AND ITS VERSION, and the same name and version over a different training corpus is the substitution a version number cannot see. AN ARTIFACT IS BOUND TO THE CONSENT VERSION IT WAS CLEARED UNDER, so rebinding to a later one is refused rather than absorbed. THE SOURCE IS PART OF THE ARTIFACT: a model trained on a take the consent no longer covers is stale rather than innocent, and that is a different refusal from a substitution because the remedy is a retrain rather than a routing fix. AN EXPORT IS A COPY A WITHDRAWAL CANNOT REACH — exactly one of four destinations keeps a withdrawal enforceable and the other three need the consent to have licensed them -->
    • I5.5.d Revalidate requested script, purpose, audience, language/style, channel, territory, term, and actor at preview, generation, and release. Evidence: scope matrix. <!-- evidence: scope.ts + script.ts — A PREVIEW IS A GENERATION THAT DOES NOT SHIP, so all sixteen dimensions are checked at all three stages and the stage decides only where the audio may go and whether the label must be on it yet. REVALIDATE MEANS ASK AGAIN: the later stages trust the earlier clearance for nothing except proof that it happened. A KEYWORD LIST IS A DETECTOR, NOT A GATE — whether a consent is required follows which artifact is loaded, never the wording of the request — and AN UNASSESSED SCRIPT IS NOT A CLEAN SCRIPT, so a line the classifier cannot read refuses rather than coming back empty the way a substring scan does -->
    • I5.5.e Detect and block direct impersonation, claimed-authority, third-party enrollment, replay, cross-project, expired/revoked, and scope- escalation attempts. Evidence: adversarial voice tests. <!-- evidence: adversarial.ts — a total table, seven attacks, seven distinct refusals. AN IMPERSONATION IS DETECTED IN THE OUTPUT, NOT IN THE REQUEST: the rendered audio is scored against the registry of people this workbench can recognise, so a job that steers a stock voice until it sounds like somebody is refused having broken no field on any form. A CLAIM OF AUTHORITY IS NOT AUTHORITY. A REPLAY IS NOT A SPOOF AND NEITHER IS THE OTHER — a recording of her real voice is not synthetic, so the synthesis detector has nothing to find and a separate replay countermeasure is what closes it. AN ESCALATION IS NOT A SCOPE ERROR: one means she never agreed and needs a new consent, the other means she did and the approved job was smaller, and it needs a new preview -->
  • I5.6 Enforce child-safety restrictions and age-assurance boundaries. Evidence: policy and redaction tests. <!-- evidence: evidence/v1-workbenches/child-safety/ (twelve files, strict tsc + a 401- assertion suite run under tsx, both gated in verify:inventory:v1-workbenches) and evidence/v1-workbenches/child-safety.json, generated and drift-gated by scripts/v1-workbenches/generate-child-safety.mjs. 44 refusals, 36 rules, 30 vocabularies, 8 known answers; 113/113 mutants killed with the control surviving. The verification half is recorded against the shipped libs/isis/ai-providers/.../civitai/content-safety-provider.ts and gated so it cannot rot: (1) it conflates the viewer's age with the subject's under one word; (2) its only minor-safety signal is a CONJUNCTION over English prompt substrings (nsfwCheck.found && minorKeywords.some(...)), so dropping either word switches it off and it looks at no reference and no output; (3) verifyAge returns the same "verified" for a tick-box, a truthy documentId and a third party whose branch reads the caller's own verified:true; (4) its age is Math.floor(elapsed/365.25), which reads seventeen on the day somebody born 1 March 2000 turns eighteen — the error runs one way, denying adults rather than admitting children, and the sign is an accident of the constant rather than a decision; (5) it stamps an expiry on the one claim that cannot expire and then never reads it, so AgeGateStatus carries an 'expired' member nothing can produce. NOT CLAIMED: the estimator, the issuers, the classifier and the human review queue are seams, and wiring this gate into the live Isis generation path is separate -->

    • I5.6.a Define child-safety operation/content/identity/reference/storage/ sharing/publication restrictions by actor, audience, host, jurisdiction, and uncertainty. Evidence: child-safety policy schema. <!-- evidence: child-safety.ts + jurisdiction.ts — seven restricted subjects, each with its own refusal, conditioned on all five axes. AN AGE GATE IS ABOUT THE AUDIENCE AND A CHILD-SAFETY RULE IS ABOUT THE SUBJECT: these are two people and two questions, and the shipped Isis surface answers both under the word "age", so every rule here carries the party whose age it turns on. UNCERTAINTY IS A CONDITION, NOT A GAP IN THE TABLE — every row answers minor, adult and nobody-knows, and the parse-time invariant is that the third is never more permissive than the first. A TABLE WITH A HOLE IN IT IS A PERMISSION, so a subject with no rule refuses as unmapped rather than falling through. AN AGE IS A CALENDAR FACT, NOT A DIVISION, and THE APPLICABLE AGE DEPENDS ON THE QUESTION, NOT ONLY ON THE PLACE: majority, digital consent and age-sensitive content are three columns over six jurisdictions, WHEN MORE THAN ONE PLACE APPLIES THE STRICTEST ONE DECIDES, and AN UNLISTED TERRITORY HAS NO ANSWER -->
    • I5.6.b Treat declared/detected/uncertain minor status, age-sensitive content, and child identity references with fail-closed escalation rules. Evidence: applicability fixtures. <!-- evidence: status.ts + fixtures.ts — UNCERTAINTY IS A FIRST-CLASS STATE: three values, not a boolean whose false already decided what to do when nobody knows. The resolution is asymmetric on purpose. A DECLARATION IS NOT A DETERMINATION — the shipped verifyAge('self-declaration') returns the same "verified" a document check does, so nothing downstream can tell a tick-box from a passport; here a declared adult establishes nothing and a declared child establishes everything, because the only person with no incentive to lie in that direction is the one telling you they are a child. A DETECTION CAN RAISE PROTECTION AND NEVER LOWER IT. THE ABSENCE OF A SIGNAL IS NOT AN ADULT, and the resolution records every source it consulted so a question nobody asked is distinguishable from one that came back unsure. A SENSITIVE CONTEXT IS DECIDED BY THE OUTPUT (declared union detected), and a reference to somebody nobody can age is exactly as unusable as a reference to somebody known to be twelve -->
    • I5.6.c Accept only approved privacy-preserving age-assurance assertions with issuer, method class, confidence/threshold, scope, freshness, and no raw identity disclosure. Evidence: assertion tests. <!-- evidence: assurance.ts — AN ASSERTION IS NOT A BIRTHDATE. The parsed type has a closed key set and a payload carrying birthDate or documentId refuses for THAT before anything else is looked at, so disclosure is unrepresentable rather than forbidden — the same move I5.4 made for the face template. A METHOD CLASS IS PART OF THE CLAIM and self-declaration is not on the approved list. A CONFIDENCE BELOW ITS THRESHOLD READS UNKNOWN, NEVER ADULT: readEstimate's return type has no 'adult' in it at all. AN AGE-SENSITIVE OUTPUT RAISES THE BAR. YOU CANNOT GET YOUNGER, SO THE TWO DIRECTIONS EXPIRE DIFFERENTLY — an at-least claim is monotone and an under claim decays a year per year, which is the exact inversion of the shipped code that stamps a 365-day expiry on the adult claim and then never reads it. AN ACCURACY IS NOT A PROTECTION: with a quarter of the traffic under age and a check 98% right both ways, one admitted person in 148 is a child and one refused in seventeen is an adult, and the two move opposite ways -->
    • I5.6.d Prohibit child sexualization/exploitation, unsafe transformations, biometric/identity misuse, disallowed cloning, and prohibited retention/ sharing at input, plan, dispatch, output, and release. Evidence: end-to-end policy tests. <!-- evidence: prohibitions.ts + gate.ts — seven prohibitions over five checkpoints, one door, and a total table. A PROHIBITION CHECKED ONLY AT THE END IS AN INVITATION TO SPEND THE MONEY FIRST, so every row names the earliest checkpoint at which it can be decided and EVERY LATER CHECKPOINT ASKS AGAIN with no gaps in between. WHAT CANNOT BE SEEN YET CANNOT BE CLEARED YET: three of the seven genuinely cannot be decided at input, asking one early is itself a refusal, and saying so beats running a check that always passes and calling the row covered. A TRANSFORMATION IS A CLAIM ABOUT A PERSON in both directions — aging a child up and regressing an adult down are the same laundering move from opposite ends, so the unsafe set is keyed on the status rather than being one list. Every rule keys on "not established as an adult", so uncertainty is protected like a child, and the suite pins the dispatcher's call count at zero for every refusal that lands before dispatch — and honestly at one for the sharing rule, which cannot fire until an artifact exists -->
    • I5.6.e Redact sensitive child/age evidence from lower-tier UI/API/logs/ exports while retaining authorized safety-review access. Evidence: redaction tests. <!-- evidence: redaction.ts — A REDACTION IS NOT A DELETION: the withheld marker stays in the view carrying the kind and the tier that would be needed, because an absence teaches the operator there was nothing to see and that is how a case gets closed on an empty screen. THE MARKER SAYS WHAT IS MISSING AND WHY, NEVER WHAT IT SAID — the Withheld type has no value field, so a leaking marker does not type-check. A LOWER TIER STILL HAS TO BE ABLE TO ACT, so the refusal and the request survive every projection. A CASE HANDLE IS NOT A PERSON, and the leak scan exists precisely to catch the commonest version of this — using the identity as the handle, so it sits unredacted in every log line above a panel reading "withheld". YOU CANNOT REDACT WHAT YOU HAVE ALREADY WRITTEN: the projection happens at write time and there is deliberately no function that turns a record back. A COPY OUTLIVES THE READER — a reviewer reads the evidence on a screen and the same evidence still does not enter an export -->
    • I5.6.f Route uncertainty/reportable signals to authorized human safety review and incident/reporting procedures without automatic accusation. Evidence: escalation tests. <!-- evidence: escalation.ts — AN OBSERVATION IS NOT AN ACCUSATION. What goes to review is a measurement — which instrument looked, at what, and what it read — and there is no free-text field on an escalation at all, because the free-text field is where a sentence about somebody whose age estimator was wrong by two years gets written, and once it exists it is in a queue, a log, an export and somebody's memory. The accusatory field names refuse by key. A QUEUE NOBODY IS AUTHORIZED TO READ IS NOT A REVIEW: routing into a queue with no safety reader is the failure that looks exactly like success. AN ESCALATION IS NOT A REFUSAL AND A REFUSAL IS NOT AN ESCALATION — both have to happen, and each half alone has its own refusal. A REPORT IS A HUMAN ACT: the draft is a one-member union, the filed type is produced by exactly one function, and that function needs a named person holding the lead role, because an automated system that files reports about people will file reports about people who did nothing -->
  • I5.7 Implement input and output moderation with explicit uncertainty. Evidence: classifier version, thresholds, and human-review queue. <!-- evidence: evidence/v1-workbenches/moderation.json + moderation/ (12-file strict-tsc reference, 542 assertions, 107/107 mutants caught); generator scripts/v1-workbenches/generate-moderation.mjs (+ --check drift gate, 15 tests). Evaluator version, model digest and threshold-set version are pinned into every observation and verdict; the queue carries locks, derived priority and no path from elapsed time to permission. The classifiers themselves are taken through the RunEvaluator seam, not implemented. -->

    • I5.7.a Define moderation subjects/categories, input/output stages, evaluator versions, thresholds/bands, regions/timecodes, uncertainty, action policy, and audience projections. Evidence: moderation schemas. <!-- evidence: moderation.ts + evaluators.ts — 16 subjects over 2 stages, 11 categories (3 absorbing), 6 actions on ONE axis with humanReview as a separate boolean, half-open bands walked over all 1001 scores, 6 locus kinds with per-modality validity, 3 outcome kinds, 5 audiences. -->
    • I5.7.b Moderate brief text, structured controls, reference media, imported specs/graphs, filenames/metadata, and linked sources before planning/dispatch. Evidence: input coverage tests. <!-- evidence: input-coverage.ts — 8 input subjects, dispatch guard, graph-text extraction, unresolved-link refusal; reproduces the shipped substring matcher beside a boundary matcher and measures both over an 8-brief corpus (6 substring artifacts, 8/8 vs 4/8 tagged). -->
    • I5.7.c Moderate every generated/derived media component, transcript/ caption/metadata, preview, bundle, and variant before exposure/release. Evidence: output coverage tests. <!-- evidence: output-coverage.ts — 8 output subjects, exposure-driven coverage, evidence-stage check that stops an input pass discharging an output one, and locus mapping through crop/scale/transcode/frame-extract/ contact-sheet with restyle refusing rather than guessing. -->
    • I5.7.d Combine multiple classifier/rule results without averaging away a blocking category, and preserve conflicts/unavailable/error states. Evidence: aggregation tests. <!-- evidence: aggregation.ts — maximum within a category and strongest action across them; absorbing categories resolve by strength not declaration order; disputes, unavailability and indeterminacy are kept distinct from a clean reading; the shipped mean is reproduced only so the dilution (900 -> 500 -> 366 as clean evaluators are added) is measured. -->
    • I5.7.e Persist raw attributed scores/regions/uncertainty separately from policy verdict/action and exact evaluator/threshold versions. Evidence: moderation result tests. <!-- evidence: result.ts — ObservationRecord carries no action and no verdict; VerdictRecord pins policy/threshold-set versions; the ledger is append-only with supersession; replay re-decides from stored observations and refuses fresh readings; stored counts are recomputed from evidence. -->
    • I5.7.f Route configured uncertainty/manual bands to a prioritized human queue with subject lock, evidence, decision, and no false auto-pass. Evidence: review queue tests. <!-- evidence: queue.ts — priority derived lexicographically from four recorded facts, locks with expiry that returns the item still held and counts the abandonment, decisions requiring the lock holder and cited evidence from this verdict, senior role to release an absorbing flag, and requireNotAutoPassed refusing on both axes. --> <!-- partial: the human review rota behind the queue is an operational commitment, not a code artifact; that half stays with the operators. --> <!-- partial: wiring this gate into the live Isis generation path is separate from the reference and is not claimed here. --> <!-- I5.7 verification finding: the shipped surfaces are libs/isis/ai-providers/src/providers/civitai/content-safety-provider.ts and apps/isis/web/src/security/content-governance-policy-utils.ts. Both match keywords as substrings, so "a sensual portrait in a canteen" is classified illegal at rating X (canteen contains teen) while "a nine-year-old, unclothed" comes back safe at a reported confidence of 90 — higher than the 82 for the canteen. The reported confidence is a mean over the tags that fired and a hardcoded 90 when none did; scanOutput accepts an imageUrl and classifies output.text ?? ''; disabling output scanning returns a manufactured safe:true/confidence:100; the classification cache is keyed on content alone; checkRule returns on the first matching condition; and parseScanReport trusts stored counts while silently dropping the findings behind them. Full detail in the shippedComparison block of moderation.json. -->
  • I5.8 Implement similarity/originality screening appropriate to visual, audio, 3D, and video assets. Evidence: results are advisory or blocking only according to ratified policy. <!-- evidence: evidence/v1-workbenches/similarity.json + similarity/ (13-file strict-tsc reference, 357 assertions, 86/86 mutants caught); generator scripts/v1-workbenches/generate-similarity.mjs (+ --check drift gate, 14 tests). Organising rule: AN INVARIANCE IS A CHOICE, AND EVERY INVARIANCE IS ALSO A BLIND SPOT — every descriptor declares both lists and they are checked disjoint and non-empty at parse time. No classifier or embedding model is implemented; descriptors arrive as vectors through their seams. -->

    • I5.8.a Define protected/reference corpus scope, version, rights/access, modality-specific fingerprints/embeddings, thresholds, uncertainty, and permitted decision use. Evidence: screening policy schema. <!-- evidence: screening.ts + metrics.ts — 4 modalities over 23 channels, 5 rights classes (2 actionable), 15 adversary transforms, 5 metrics, 5 named uncertainty sources; a corpus carries id, version, coverage AND the basis for that coverage, because a not-found against 2% of the space is not an originality finding. -->
    • I5.8.b Implement visual whole-image and localized composition/content/ style similarity with crop/transform/near-duplicate handling. Evidence: visual similarity fixtures. <!-- evidence: visual.ts — tile-grid windows searched exhaustively or honestly bounded, mirror search with the flip recorded, zero-mean normalised cross-correlation (raw cosine over brightness matches every window alike), perceptual-hash Hamming contrasted against the shipped Jaccard-over-sha256 which has no intermediate value. -->
    • I5.8.c Implement audio fingerprint/melody/harmony/rhythm/timbre/voice/ sample similarity with time alignment and licensed-source exclusions. Evidence: audio similarity fixtures. <!-- evidence: audio.ts — reproduces the shipped chroma detector and measures both inversions (relative keys 1000 -> block; semitone transposition 286 -> allow); interval histograms, root motion, tempo-normalised IOI, bounded sample alignment, licensed-source exclusion before scoring. The speaker comparison is a seam to the I5.5 reference. -->
    • I5.8.d Implement 3D geometry/topology/shape/texture/material/scene similarity with units/transforms/partial components. Evidence: 3D similarity fixtures. <!-- evidence: three-d.ts — reproduces the shipped duplicate detector and measures the theft it misses (0.457, below its own 0.65 report floor); mean-normalised shape distribution invariant to rotation and scale, Euler characteristic as a filter rather than a distance, unit-normalised volume, per-component scene matching that keeps the best part rather than a mean. -->
    • I5.8.e Implement video frame/shot/temporal/audio similarity with reencoding/crop/rate/edit alignment. Evidence: video fixtures. <!-- evidence: video.ts — everything addressed in milliseconds so 24fps and 30fps are the same timeline; shot durations aligned as subsequences under an explicit tolerance, with inserted and dropped shots named; the temporal channel read against the query so a cutdown is entirely a copy. -->
    • I5.8.f Return candidate matches, protected-source identities only when authorized, raw scores/regions, uncertainty, corpus/version, and no unsupported ownership conclusion. Evidence: result/privacy tests. <!-- evidence: disclosure.ts — a stable pseudonymous handle for the unauthorised, identity for 2 of 4 audiences and scores for 3 (the identity is rationed, never the evidence); corpus/version/coverage and uncertainty on every result including negative ones; a conclusion-vocabulary gate over emitted statements naming the 3 facts a distance cannot supply. -->
    • I5.8.g Apply advisory/manual/blocking action solely from ratified modality/use/risk policy and require human rights review where specified. Evidence: policy-boundary tests. <!-- evidence: policy.ts — rows indexed by modality/channel/use with ordered thresholds and a per-row action ceiling, so a channel at its ceiling is not outvoted and style may advise and never block; an unratified policy still screens and cannot act above advisory, with the cap recorded rather than silent; a required rights review needs a named holder of the role. --> <!-- partial: the ratification itself is a human act and the rights-review rota is an operational commitment; both stay with the rights authority. --> <!-- partial: building the protected corpus, and the extractors that produce the descriptors, are separate from the reference and are not claimed here. --> <!-- I5.8 verification finding: the two shipped surfaces fail in opposite directions and both are exactly computable. libs/isis/ai-providers/.../copyrighted-style-detector.ts compares a 12-bin chroma histogram at 0.93: C major and its relative A minor are the same seven pitch classes so unrelated music scores 1 and is BLOCKED, while a semitone transposition scores exactly 2/7 and is ALLOWED — and no threshold separates them, because the separation is not in the descriptor. libs/isis/3d-marketplace-ops/duplicate- detection.ts aggregates channels by a weighted mean and drops everything under 0.65, so byte-identical textures on a 10%-decimated mesh aggregate to 0.457 and are never reported; its dominant geometry term is a step function over a sha256, its texture channel is Jaccard over digests, and its volume guard is truthiness so an absent volume and a measured zero both score 0. Neither records a corpus version, coverage, uncertainty or ratification, and both disclose the matched source identity unconditionally. Full detail in similarity.json. -->
  • I5.9 Implement visible watermark policy per audience and asset kind. Evidence: output inspection confirms coverage and legibility.

    • I5.9.a Define watermark requirements by asset kind, audience, preview/ review/release state, target/channel, disclosure text/mark, placement, persistence, and removal authority. Evidence: policy matrix.
    • I5.9.b Render image/texture watermarks with safe-area, scale, contrast, opacity, repetition, crop/resize resilience, and alpha/color handling. Evidence: image inspection fixtures.
    • I5.9.c Render video/animation watermarks with temporal coverage, movement/placement, scene/crop considerations, HDR/color, and captions/ safe areas. Evidence: video inspection fixtures.
    • I5.9.d Render audio watermarks/disclosures where policy requires with intelligibility, timing, loudness, localization, and accessibility. Evidence: audio disclosure tests.
    • I5.9.e Apply 3D/splat/package preview overlays or embedded visible disclosure assets without corrupting source geometry/package. Evidence: 3D/package tests.
    • I5.9.f Verify the final stored/delivered bytes for required legibility/ coverage after target transformation and block on missing marks. Evidence: post-write inspection tests.
    • I5.9.g Record policy/version/render config/result and govern authorized removal as a new output revision. Evidence: watermark lineage tests. <!-- evidence: evidence/v1-workbenches/watermark/ (14-file strict-tsc reference; 523 assertions run under tsx; 135/135 mutants killed with an inert control surviving) + evidence/v1-workbenches/watermark.json, generated and drift-gated by scripts/v1-workbenches/generate-watermark.mjs (+ 19 known-value tests), wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches and CI. I5.9.a is a policy matrix that is TOTAL over 8 asset kinds x 4 audiences x 3 release states x 6 channels: every one of the 576 cells resolves or refuses by name. Behind it is a 5-kind x 10-transform survival table in which three cells are 'depends-on-placement' because a crop is a geometric question no table answers, and every one of the other 47 is checked against what the delivery pipeline actually does. Disclosure is proved monotone in audience reach over the whole cross product, with the recipient trace excluded by name because you cannot trace a leak to an audience you did not name. Localisation refuses rather than falling back. The shipped leak-trace token is reproduced and its two exact collisions demonstrated; the reference replaces it with an HMAC over a length-prefixed encoding that round-trips. I5.9.b-e render and measure for real: WCAG relative luminance and contrast (white on black is exactly 2100 hundredths, #767676 on white 454), compositing in linear light, the exact minimum-opacity inverse, the crop guarantee period + side - 1 proved by exhausting offsets, the sqrt(21) ceiling over an unknown backdrop, SMPTE ST 2084 PQ against BT.2408's 203-nit reference white, coherent decibel summation, and a per-locale speaking model. I5.9.f applies the channel's transforms to real pixels and re-measures the DELIVERED frame, refusing to run at all on bytes that have not been through it. I5.9.g is an append-only revision chain in which a removal is a new revision and the AI disclosure has no removal authority at all. Records six shipped Isis surfaces as a verified structural comparison: the quality gate treats watermark as a zero-tolerance ARTIFACT while the marketplace controls REQUIRE one, and running the shipped keyword detector over the marketplace's own control string raises exactly that artifact — so the gate meant to pass the disclosed asset rejects it for carrying the disclosure. Full detail in watermark.json. -->
  • I5.10 Implement robust invisible watermarking where supported. Evidence: channel-transformation survival evaluation.

    • I5.10.a Define supported modalities/formats/targets, payload, key/version, embed/detect algorithms, strength/quality limits, disclosure, and unavailable policy. Evidence: invisible-mark schema.
    • I5.10.b Embed a minimal opaque ledger/output identifier after generation at the correct pipeline stage without encoding sensitive metadata. Evidence: embed/extract tests.
    • I5.10.c Verify the stored artifact contains the expected mark with confidence and no unacceptable quality/technical-gate regression. Evidence: post-embed validation tests.
    • I5.10.d Evaluate detection survival across configured resize/crop/ transcode/compression/color/audio resample/time-stretch/noise/screenshot/ recording transformations. Evidence: robustness benchmark.
    • I5.10.e Measure false-positive/collision/wrong-payload/tamper rates on marked and unmarked calibrated corpora. Evidence: calibration report.
    • I5.10.f Protect embed/detect keys and detailed parameters, support key rotation, and retain authorized historical verification. Evidence: key lifecycle/security tests.
    • I5.10.g Return explicit unsupported/unavailable/degraded state and never claim invisible marking from policy intent alone. Evidence: honest-state tests. <!-- evidence: evidence/v1-workbenches/invisible-mark/ (12-file strict-tsc reference; 312 assertions run under tsx; 107/107 mutants killed with an inert control surviving) + evidence/v1-workbenches/invisible-mark.json, generated and drift-gated by scripts/v1-workbenches/generate-invisible-mark.mjs (+ 15 known-value tests), wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches and CI. The embedder is REAL: quantisation index modulation with a keyed dither, so "an additive error smaller than half a step is corrected and one larger is not" is a theorem checked over every value in a period in both directions, and the payload type has one field — a 32-bit opaque ledger identifier — so the shipped forensic payload's operatorId does not typecheck into it. The detector's null distribution is computed EXACTLY with big integers (per-bit agreement of nine chips is 252/168/72/18/2 out of 512, convolved 32 times into 129 outcomes), giving 695 in a billion at 210 of 288 chips and about one at 220 — so the threshold is derived rather than chosen. The robustness benchmark applies its fourteen transformations for real: six survived, eight did not, no row the per-chip guarantee promised was lost, and one row it gave up on came back anyway. Raising the step fourfold beats the noise it could not beat and does nothing whatever for a resize, which is why availability for any resizing channel is 'unavailable' with the visible mark of I5.9 named as the fallback rather than a low score. The calibration runs a seven-file labelled corpus and reports one miss and one wrong payload beside the model's rate. Records the two shipped surfaces as a verified structural comparison: the voice module is a 446-line PLANNER that touches no audio and returns the same six stages for every input, its predicted recoverability rises by exactly 0.08 when a JSON sidecar is attached to the same audio, its survivability moves by exactly 0.12 on whether a rights form was signed, and its identifier is an unkeyed digest of the request. Full detail in invisible-mark.json. -->
  • I5.11 Implement C2PA Content Credentials with ingredient lineage. Evidence: signed artifacts validate through the pinned toolchain.

    • I5.11.a Define assertion/claim profile per asset kind/target with creator/ tool identity, actions, ingredients, AI disclosure, rights, timestamps, hashes, and privacy filtering. Evidence: manifest profile tests.
    • I5.11.b Build ingredient lineage from immutable source/reference/model/ workflow/run/output/edit identities and represent redacted/external/ missing ingredients explicitly. Evidence: lineage mapping fixtures.
    • I5.11.c Generate canonical manifests bound to final output bytes after target transformation and before registry write verification. Evidence: manifest generation tests.
    • I5.11.d Sign through approved external key custody with certificate/ timestamp/status checks, rotation, and no private-key application access. Evidence: signing integration tests.
    • I5.11.e Embed or sidecar credentials according to format/target and keep package references/hash bindings deterministic. Evidence: packaging tests.
    • I5.11.f Validate signature, trust chain, timestamp, status, asset hash, assertions, ingredients, and policy using the pinned independent toolchain. Evidence: verification fixtures.
    • I5.11.g Handle unsupported formats/signing failure/invalid credentials as explicit blocking or degraded states per target profile. Evidence: failure-mode tests. <!-- evidence: evidence/v1-workbenches/content-credentials/ (17-file strict-tsc reference; 608 assertions run under tsx; 96/97 mutants killed by the suite with an inert control surviving)
      • evidence/v1-workbenches/content-credentials.json, generated and drift-gated by scripts/v1-workbenches/generate-content-credentials.mjs (+ 15 known-value tests), wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches and CI. The cryptography is REAL and third-party checkable: RFC 8032 Ed25519 over FIPS 180-4 SHA-512, both held to published vectors in the suite and cross-checked byte-for-byte against node:crypto over 60 random hashes and 25 random keypairs while it was written, so a MAC is not quietly standing where a signature goes. Canonicalisation is RFC 8785, which is where the shipped finding lives: c2pa-metadata-embedding.ts:223 sorts assertion keys with localeCompare and claim-signing.ts:105 sorts with the default comparator, so the assertion hashes a manifest publishes are computed under one rule and the signature over it under another; the reference reproduces both functions verbatim and shows the two sha256 digests differ. The planner's serialiser also maps null, NaN, Infinity and -Infinity to one string (six colliding pairs) and renders a hole in an array as "[,1]", which is not JSON. The hard binding is a theorem, not a claim: flipping every byte of a file with a ten-byte exclusion changes the digest for all fifty-four bytes outside it and none of the ten inside, and a uniform file excluded at two different places digests differently because the exclusion list is hashed in. Five container encodings insert and recover a manifest byte-exactly with real CRC-32, and whether stripping the credential returns the original bytes is MEASURED per format: true for png/jpeg/mp4, false for wav/webp/glb, which is exactly the rewritesHeader column. Custody is a port with no method that returns key material and a signDigest that takes 32 bytes; path validation runs at the SIGNING instant, a timestamped signature survives certificate expiry and an untimestamped one does not, and a revocation's reason decides what it kills (a supersession spares everything signed before it; a compromise with no invalidity date takes everything). The validator recomputes from the bytes, has no second input for the signing certificate, refuses a manifest that grades its own homework, and treats an unknown as indeterminate rather than valid. Two guards in gate.ts are documented in the source as unreachable given the fixed manifest reservation; the suite checks the invariant they rest on rather than pretending to reach them, and the reservation's value is the one mutation the drift gate rather than the suite catches, because it is a tunable. NOT the C2PA wire format: no JUMBF, CBOR, COSE, DER, X.509 or RFC 3161, so these manifests will not open in the c2pa CLI and are not offered as though they would. Full detail in content-credentials.json. -->
  • I5.12 Implement external provenance-ledger recovery. Evidence: stripped manifest can be matched without false association.

    • I5.12.a Define a privacy-minimized ledger record containing opaque asset/ output identity, strong content hash/fingerprint versions, credential digest, issuance/status, and authorized recovery metadata. Evidence: ledger schema.
    • I5.12.b Write ledger records transactionally with release/output state and verify external acknowledgement/integrity. Evidence: write failure tests.
    • I5.12.c Recover exact-byte matches by cryptographic hash and transformed candidates by calibrated modality-specific fingerprints as distinct result classes. Evidence: recovery fixtures.
    • I5.12.d Enforce confidence/collision/ambiguity thresholds and return multiple/no-match/manual-review rather than false association. Evidence: false-match corpus.
    • I5.12.e Authorize recovery queries, rate-limit enumeration, and redact tenant/creator/source metadata until proof/role permits disclosure. Evidence: privacy/abuse tests.
    • I5.12.f Respect recall/revocation/deletion/legal-hold status while retaining minimum required provenance proof. Evidence: lifecycle tests. <!-- evidence: evidence/v1-workbenches/ledger-recovery/ (14-file strict-tsc reference; 468 assertions run under tsx; 87/87 mutants killed by the suite with an inert control surviving)
      • evidence/v1-workbenches/ledger-recovery.json, generated and drift-gated by scripts/v1-workbenches/generate-ledger-recovery.mjs (+ 14 known-value tests), wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches and CI. The threshold is DERIVED, not chosen: a real two-dimensional DCT-II produces a 63-bit code (the DC dropped because it is the brightness, which is why 63 and not 64), and exact big-integer binomial arithmetic turns a false-association budget and a ledger size into a Hamming threshold — eight bits against a thousand records, five against a million, two against a billion. Five bits is not many, and the reference says so rather than raising the threshold until the fixtures pass. Robustness is MEASURED over six rows: a 5% exposure change and a box blur move the code by zero bits, blur-then-exposure by two, heavy noise by ten, a hard grade by eighteen. The load-bearing measurement is that PLATE_B — a different picture — sits at eight bits, NEARER than the same picture under heavy noise, which is why the band between the identification and review thresholds produces a case and never a name, and why a win by under two bits is downgraded. Audio is a second algorithm rather than the same one renamed: a sub-band double difference that survives a 0.3x and a 2x gain AND a constant spectral tilt at zero bits, where a single difference would have lost ten of thirty-two. Over a labelled corpus of nine: six correct, one review, one miss, one true negative, zero false associations, and the five buckets are asserted to account for every query. The record is built from nothing upwards with a three-way exposure table (record / authorized / never), the write is transactional in both directions with an authenticated receipt and a durable sequence, disclosure needs possession of the C2PA credential of I5.11 or a role whose standing is checked against the tenant, enumeration is caught by the hit rate rather than the rate limit alone, and erasure drops the identifying half while retaining the digests, the date and the status. Two of my own defects were found and fixed while building it: a flat frame produced a 63-bit code made entirely of floating-point rounding error (now refused by a structure floor measured against the DC), and the identification and review bands were two filters that could drift apart (now one bandOf function, tested at both boundaries). NOT CLAIMED: no JPEG decode, no filterbank, and no rotation, crop or scale invariance whatever — a cropped picture will not be found, which is a real limit of a DCT construction and is stated rather than left to be discovered; and the false-match model assumes bit independence that real photographs do not have, which makes the true rate worse than the published one, so a corpus is measured beside it. Full detail in ledger-recovery.json. -->
  • I5.13 Implement consent, license, model, workflow, and source blast-radius queries. Evidence: revocation preview lists every dependent run, asset, collection, and release.

    • I5.13.a Build immutable reverse dependency edges from consent, rights/ license, source/reference, dataset/model/adapter, workflow/node, policy, run, output, asset, bundle, collection, and release versions. Evidence: dependency graph tests.
    • I5.13.b Accept only authorized exact subject/version/hash/revocation inputs and distinguish direct from transitive dependency. Evidence: query contract tests.
    • I5.13.c Traverse cycles/depth/fan-out safely with stable pagination and complete typed nodes/edges/reasons. Evidence: large graph tests.
    • I5.13.d Report affected plans/runs/outputs/assets/variants/bundles/ collections/gallery/marketplace/releases/targets and current state. Evidence: blast-radius golden fixtures.
    • I5.13.e Calculate prospective policy action, urgency, owner, replacement/ remediation, and unresolved/unknown dependencies without mutating them. Evidence: preview tests.
    • I5.13.f Enforce tenant/project/property-level redaction while preserving accurate authorized aggregate counts. Evidence: cross-tenant tests. <!-- evidence: evidence/v1-workbenches/blast-radius/ (13-file strict-tsc reference; 302 assertions run under tsx; 87/87 mutants killed by the suite with an inert control surviving)
      • evidence/v1-workbenches/blast-radius.json, generated and drift-gated by scripts/v1-workbenches/generate-blast-radius.mjs (+ 15 known-value tests), wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches and CI. A BLAST RADIUS IS A QUESTION ASKED BACKWARDS: edges are content-addressed over their own triple, so recording a dependency twice is one edge and an altered one no longer hashes to itself; every endpoint carries a version, because "this output used the kiln model" is not answerable; and the reverse index is rebuilt and checked in both directions rather than stored and trusted. The two ends of an edge fail DIFFERENTLY — a missing dependency is a broken write, a missing dependent is an unresolved one that gets counted — which was a fix, not the first design. Three properties carry the item and each is checked rather than asserted: a ten-long chain against a depth bound of six reaches six and names what it left unwalked; a key cursor and an offset cursor run over the same mid-pagination insert, and the offset returns a row the reader has already seen AND skips one they never will while the key cursor does neither; and four principals ask the same question and get four different visible sets (11, 12, 12, 0) with one identical total of 12, where a filter would have told the aggregate-only auditor zero. Urgency is DERIVED from the exposed count, which is the direct contrast with the shipped finding, and requireUrgencyDerived exists to stop it being simplified back. The preview is prospective by construction and the graph digest proves it. Two design bugs of my own were found and fixed while building: the traversal conflated MERGING a second path into a node with DOWNGRADING its relation (the merge is the ordinary case; the check now runs against the graph), and the graph builder refused the wrong endpoint. Mutation testing then found nine more holes, all one class — overlapping guards, where removing any single one left the refusal NAME unchanged and the suite silent — now isolated with per-guard cases asserting the message. SHIPPED FINDING: libs/isis/model-governance-3d/src/upstream-change-monitor.ts raises exactly this item's triggers ('license-drift' among them) and its twelve-field alert names no run, output, asset, bundle, collection or release; its severity is a literal per alert kind, so the function has nowhere for a radius to go and a drift behind one draft and one behind forty published releases are both 'critical'; and the generated LineageEdgeType's five values are all asset-to-asset, expressing 2 of the 12 kinds a revocation can start at — there is no edge type in which "this asset exists because of that person's consent" could be written down. NOT CLAIMED: the graph is an in-memory index; persistence, sharding, concurrent reverse-index consistency and cross-shard pagination are not solved, the bounds would need a resumable walk rather than a truncation in production, and the small-bucket disclosure problem (a hidden count of one is close to an identification) is named rather than solved. Full detail in blast-radius.json. -->
  • I5.14 Implement recall and quarantine propagation. Evidence: affected gallery items and releases become unavailable according to policy.

    • I5.14.a Define recall/quarantine case, trigger, subject, scope, severity, actor/authority, policy, target actions, deadlines, status, and evidence. Evidence: case schemas.
    • I5.14.b Freeze the authorized blast-radius snapshot and require explicit include/exclude reasons for any bounded scope adjustment. Evidence: scope tests.
    • I5.14.c Quarantine affected model/workflow/source/output/asset/bundle versions from new planning, reuse, review, and release atomically. Evidence: admission tests.
    • I5.14.d Unpublish/disable affected gallery, marketplace, collection, campaign, download, API, CDN, and target-distribution references according to channel policy. Evidence: propagation integration tests.
    • I5.14.e Revoke/replace signed URLs, caches, credentials/manifests where supported and verify downstream state; record irrecoverable copies. Evidence: post-action verification tests.
    • I5.14.f Notify required owners/reviewers/consumers/incidents with audience-safe reasons and track acknowledgements/escalation. Evidence: notification tests.
    • I5.14.g Restore only via a new authorized case decision after remediation and complete verification; preserve recall/quarantine history. Evidence: restoration tests. <!-- evidence: evidence/v1-workbenches/recall-propagation/ (14-file strict-tsc reference; 319 assertions run under tsx; 91/91 mutants killed by the suite with an inert control surviving)
      • evidence/v1-workbenches/recall-propagation.json, generated and drift-gated by scripts/v1-workbenches/generate-recall-propagation.mjs (+ 16 known-value tests), wired into pnpm inventory:v1-workbenches / verify:inventory:v1-workbenches and CI. A RECALL IS A DECISION SOMEBODY MADE: the case carries an actor, an authority CEILING (a reviewer decides up to routine, a rights officer up to elevated, and a court order is legal-only however small the radius) and required evidence, and nothing in it reads a clock — every instant is an argument, so a case replays identically and an escalation compares two recorded times. The I5.13 report is FROZEN and digested before anything is decided; adjustments are bounded at a quarter of the frozen set, each names an item, a direction, a reason and an actor, and inclusions and exclusions are reported separately because an exclusion is the dangerous direction. Quarantine is ONE admission set four gates consult, so a split between gates is unconstructible here — and the same check, pointed at the shipped four-boolean record, refuses it immediately. The channel table refuses to let an unverifiable channel report 'confirmed', and the hardest honesty in the item is arithmetic: the fixture propagation confirms three channels, issues three it cannot verify, and records 4139 copies in hands nobody controls, so it is NOT contained and the summary carries that number beside the successes rather than under them. Notifications are audience-filtered — a court order and a consent withdrawal produce the SAME external string, and every trigger the reference knows passes its own unsafe-word filter — with acknowledgements tracked separately from sends and escalation driven by the deadline. Restoration is a new case with its own actor, authority, remediation and verification, and it needs a prior state captured at quarantine time or it is refused rather than guessed at. SHIPPED FINDING (libs/isis/3d-quality-gates/src/delivery/output-quarantine.ts): its four reasons express NONE of the seven recall triggers (zero of seven, computed); its four independent block flags put a DISPOSED output at blocked-for-release and admitted-for-reuse, and two of the four gates this reference asks about have no flag at all; its release input has no actor, authority, remediation or verification (all four missing, computed); and release writes visibility 'internal-only' as a constant because the record never stored the prior value — structural, not behavioural, since the information is gone the moment the quarantine is applied. Its timestamps default to new Date(), so a replayed decision produces a different record. Said plainly: for what it is, a holding pen for outputs that failed this house's own gates, the module is reasonable; the finding is that it is being read as a recall mechanism. NOT CLAIMED: nothing talks to a CDN, a marketplace or a mail server — the channel table is the contract and the adapters are what a deployment writes; and making four real gates consult one store transactionally is a distributed-systems problem this reference does not solve. Full detail in recall-propagation.json. -->
  • I5.15 Protect biometric and protected-source fields in UI, logs, notifications, exports, and support tools. Evidence: redaction suite. <!-- evidence: evidence/v1-workbenches/field-protection.json + the compiling reference under evidence/v1-workbenches/field-protection/ (403 assertions, 55 named refusals, 39 rules, 144/154 mutants killed by the suite, inert control survived). Gated three ways in verify:inventory:v1-workbenches: the generator's --check drift gate, tsc --noEmit over the reference's own strict tsconfig, and a node --test suite that runs check.ts and asserts its assertion and refusal counts. The classification is grounded in eighteen real Isis properties across five shipped files, each of which the generator re-reads. SHIPPED FINDING: Isis has two redactors — apps/isis/web/src/security/redaction-utils.ts and libs/isis/llm-providers/src/client/pii-redactor.ts — and both are TEXT redactors, while every biometric property in Isis is a number vector, a buffer or a caller-supplied map. Keyword coverage of the eighteen protected properties is zero, by full path and by leaf; the only number the web one touches is tokenCount, which it turns into a string; its walk keeps one WeakSet it never clears, so an object referenced twice comes back as '[REDACTED]' the second time and an auditor cannot tell that apart from a refusal; and one URL has its query secret redacted and its identical userinfo secret preserved, under two different markers. -->

    • I5.15.a Classify face/voice embeddings, identity evidence, age assurance, consent artifacts, protected source/corpus, similarity matches, and derived identifiers at schema/property level. Evidence: data inventory.
    • I5.15.b Define per-role/purpose/host/tenant field visibility, masking, aggregation, use, export, retention, and break-glass rules. Evidence: access matrix.
    • I5.15.c Apply centralized projection/redaction before UI/API/event/ websocket/search/index/cache exposure. Evidence: property authorization tests.
    • I5.15.d Apply structured log/trace/metric/error/audit scrubbing and block sensitive values from labels, URLs, exception bodies, and debug dumps. Evidence: telemetry leak tests.
    • I5.15.e Render notifications/email/webhooks/support views with safe references and separately authorized drill-down. Evidence: notification/ support redaction tests.
    • I5.15.f Sanitize CSV/JSON/package/portable-spec/evidence exports and protect against hidden metadata/filenames/formulas. Evidence: export leak tests.
    • I5.15.g Enforce encryption/key separation/retention/deletion and audit every authorized sensitive-field access. Evidence: lifecycle/access tests.
  • I5.16 Implement authorized exception workflows without mutating original evidence. Evidence: waiver scope, expiry, approver, and audit. <!-- evidence: evidence/v1-workbenches/exception-workflow.json + the compiling reference under evidence/v1-workbenches/exception-workflow/ (15 files, 536 assertions, 66 named refusals all reached, 46 rules, 371/373 mutants killed with two survivors that are inert by construction and say so in the source). Gated three ways in verify:inventory:v1-workbenches: the generator's --check drift gate, tsc --noEmit over the reference's own strict tsconfig, and a node --test suite (18 tests) that runs check.ts and asserts its assertion and refusal counts. The gate registry is grounded in sixteen gates drawn from ten real Isis files, each of which the generator re-reads and greps for the identifier the gate was drawn from. THE EXCEPTION DOES NOT CHANGE THE MEASUREMENT, IT CHANGES WHAT WE DO ABOUT IT — the failing verdict travels through the projection byte-identically (checked by digest on both sides) and the effect is 'permitted-under-exception', never 'pass'. Nothing reads a clock: every instant is an argument, so establishing the same exception twice produces the same entry hash byte for byte. THE FLOOR IS CHECKED BEFORE THE AUTHORITY, and a compound gate inherits the floor of its parts under its own distinct refusal, because nobody asks to waive child safety — they ask to waive the release bundle that contains it. An unstated scope dimension grants NOTHING, all seven dimensions are evaluated every time (four of four mismatches reported), and the waiver is revalidated at each governed action rather than held as a token. SHIPPED FINDING: there are TWO exception workflows in this monorepo and neither knows about the other, and the one with the weaker control governs the more consequential thing. libs/contracts/src/common/entitlement-exception-policy.ts (V1-CX-003) waives an ENTITLEMENT and has an eight-rung role ladder, approver_same_as_requester, an approver-outranks-requester check, per-class duration caps and dual control. libs/isis/model-governance-3d/src/ exception-approval.ts (646 lines, exported from the library index) waives a MODEL GOVERNANCE POLICY DECISION and NEVER COMPARES THE REQUESTER TO THE APPROVER — the word "requester" does not appear in the file, which the generator asserts non-vacuously. Also computed rather than asserted: an empty scope list means EVERYTHING (all six dimensions default to [] and the badge says "Global approval"); the mismatch report short-circuits on && so an operator sees 1 of 6 wrong dimensions; effectiveAt is computed, stored, published on the type and never read by determineStatus, so a waiver marked effective in 2030 is active in 2026; three of the six decision verbs the ledger names (reject, request-changes, withdraw) cannot be written down at all; audit entry ids hash four identifiers and none of the entry's content, and caller-supplied entries are appended and re-sorted so an entry dated 1990 becomes the first thing that ever happened; an automatic expiry is recorded with actor: approver; and crypto.randomUUID() plus new Date() defaults mean the same input never produces the same record twice. Two more, from the operator's side: a PENDING exception reports blocking: false (blocking is decision === 'deny' and pending takes the warn branch), so the natural if (evaluation.blocking) reject() lets an unapproved exception through and the UI offers 'Acknowledge and renew exception review' rather than a wall — while a merely out-of-scope request DOES block, so the unapproved case is treated more leniently than the mis-scoped one; and the badge tones are inverted with respect to risk, with the scoped approval rendering 'warn' and the unscoped 'Global approval' rendering 'info'. Said plainly: within what it is the module is careful — it normalizes, dedupes, refuses an approval with no approver and a revocation before its approval, and renders an honest scoped-versus-global badge. The finding is that it is a RECORD of an exception rather than a CONTROL over one. NOT CLAIMED: nothing talks to a database, a queue, a mail server or an identity provider — the reporting line, subject owner and current digest arrive as three maps a deployment fills in. The hash chain is tamper-EVIDENT, not tamper-PROOF: defeating a rewriter who can reach every entry needs an external anchor at the seam where verifyChain is called. The floor assignment is this reference's decision, recorded rather than inferred. Full detail in exception-workflow.json. -->

    • I5.16.a Define exception request, exact subject/version/gate/policy, requested bounded waiver, rationale, evidence, risk, compensating controls, requester, approver roles, term, and status. Evidence: exception schema.
    • I5.16.b Declare nonwaivable child safety, legal/rights/consent, security, tenant isolation, human-authority, and other code-floor conditions. Evidence: nonwaivable policy tests.
    • I5.16.c Validate requester/approver separation, role/scope, conflict of interest, evidence completeness, subject freshness, and approval quorum. Evidence: authorization tests.
    • I5.16.d Record approve/reject/request-changes/withdraw/expire/revoke as immutable decisions linked to original measurements/verdict/evidence. Evidence: decision history tests.
    • I5.16.e Apply an approved exception only to exact subject revisions, gates/targets/use/audience/territory and validity window; revalidate at each governed action. Evidence: scope escalation tests.
    • I5.16.f Surface active/expired/revoked exception and compensating controls without changing the underlying failed/raw result. Evidence: projection tests.
    • I5.16.g Trigger review on subject/policy/evidence change and automatic expiry/revocation actions plus blast-radius/notification. Evidence: lifecycle tests.

I6 — Persistence, service API, and integration#

  • I6.1 Decide which existing Isis databases own briefs, plans, runs, variants, gates, rights, and releases. Evidence: ownership ADR prevents a duplicate shadow store. <!-- partial: a-d DONE. The decision is made and the prevention is live — ADR-I1 names one owner per record type and the shadow-table gate fails CI on any new duplicate. The headline: schema.prisma is NOT the whole Isis database (2 schema authorities, 39 physical tables, 8 declared by both with incompatible columns, 18/24 Prisma models with no call site, 10 connection pools so no transaction spans the split). Parent stays unchecked only because ADR-I1 is Proposed pending the named service owners' ratification (I6.1.e), matching how S0.6 was left. -->

    • I6.1.a Inventory existing stores, schemas, repositories, services, tenancy models, transactions, retention, backup, and production data for each Isis aggregate. Evidence: current-state inventory. <!-- isis-persistence-inventory.json: current-state persistence inventory. 24 Prisma aggregates over 1 schema (libs/isis/database/prisma/schema.prisma), accessed by 9 persistence services. ARCHITECTURE FINDING: 0 repository classes — Isis uses SERVICE-DIRECT Prisma access (no repository layer); 3 PrismaClient instantiations in the persistence layer — directly relevant to I6.1's goal (prevent a duplicate shadow store; the ownership ADR must reconcile these 3 clients). 4 persistence files use transactions ($transaction/ withTransaction), 7/24 aggregates are tenant-scoped (organizationId/tenantId columns: CanonicalAuditEvent/GeneratedOutput/GenerationJob/ModelRegistry/ Pipeline/...), 4 aggregates carry soft-delete (deletedAt) + a RetentionPolicy aggregate exists for retention; backup is ops/infra-level (not modeled in application code, recorded honestly). PERF: scoped greps to the DB+services layer (not the whole huge Isis tree) — 12.7s → 0.57s; both-globs fix for top-level services/*.ts; tenancy counted from MODELS (organizationId columns) not filtered files. generate-isis-persistence-inventory.mjs + 6 tests, wired into inventory/verify + CI. I6.1.b (aggregate→store map) is now DONE and supersedes this note's Prisma-only view of the database: the physical census is 39 tables under two schema authorities, not 24 aggregates. -->
    • I6.1.b Map brief, plan, run/job/stage, variant/output, gate measurement/ verdict, rights/consent/provenance, and release records to exactly one system of record. Evidence: aggregate ownership matrix. <!-- isis-aggregate-ownership.json: the matrix, computed over PHYSICAL tables rather than Prisma models — because schema.prisma is not the whole database. Isis has TWO schema authorities pointed at ONE connection string (ISIS_DATABASE_URL ?? DATABASE_URL, 10 resolvers): Prisma (24 tables) and CREATE TABLE IF NOT EXISTS run by the services (23 tables), 39 physical tables in total. 8 table names are declared by BOTH — lineage_edges, model_registry, provenance, retention_policies, workflows, workflow_versions, workflow_templates, workflow_stars — where IF NOT EXISTS is a silent no-op, so the loser queries columns that do not exist: provenance is outputId/generatedAt vs output_id/generated_at with FKs to two different tables, retention_policies is keyed on id vs name, and workflow_stars shares not one column NAME across 6 vs 4 columns. 13/23 service tables are created at runtime, not by a migration. 18/24 Prisma models have no call site at all; the only live Prisma family is pipeline_state_* — which is also the only family missing from the committed generated/schema.sql. Ownership is assigned by a recorded rule ladder (R1 live access path, R2 cascade-durability as a FILTER, R3 canonical-field coverage, R4 external, R5 unowned): 8/12 record types resolve to exactly one system of record, 0 contested, 4 are gaps (brief, variant, gate-measurement have no table anywhere; rights has three claimants and no access path). Costs are published, not hidden: the live run store (jobs) carries NO tenancy at all, losing userId/organizationId/ projectId to the declared-only rival. Consent has no Isis home — the owner is iris ConsentRecord in a different database, referenced by consent_id with no FK. generate-isis-aggregate-ownership.mjs + 38 tests, wired into inventory/verify + CI. Mutation-tested 3x (135/163 killed); it found 4 real defects in my own code — an identical-branch ternary, an unreachable R2 decision branch, an untested invariant gate, a vacuous single-database claim — all fixed. Surviving mutants are internal .sort() calls on already-ordered git output and one deliberately overlapping guard pair; each says so where it lives. I6.1.c/d and the I6.1.e ADR build on this. -->
    • I6.1.c Define read model, derived index/cache, blob, audit, outbox, and analytics ownership separately from canonical mutation ownership. Evidence: data-flow diagram. <!-- isis-derived-ownership.json: role ownership scored against the SAME census I6.1.b resolves owners from (imported, not re-derived, and asserted equal), so the two artifacts cannot drift. Per role and per owner the question asked is whether the role has its own store, is carried INSIDE the canonical row, or is absent. 3/6 are separated (audit, outbox, read-model); 2 are embedded in the canonical row — derived-index-cache (stats_runs/stats_stars/total_versions in workflows, access_count in outputs, and a searchVector) and analytics (durationMs in the stage row) — so they cannot be rebuilt or invalidated without writing to the system of record. Blob is scored pointer-vs-payload rather than lumped: storage_key in a canonical row is correct design and is NOT flagged; the one real defect is pipeline_state_phase_artifacts.inlineData holding payload in the row. Coverage is the sharp part: 7 audit stores exist yet 4/7 owners (gate-verdict, release, run, stage) have no audit trail reachable from their table, and there is exactly ONE outbox (job_outbox) covering ONE owner, so 6 systems of record mutate with no transactional publication path — I6.4 builds that, it does not extend it. Links are labelled foreign-key vs name-convention because outputs→output_audit_logs is enforced by nobody. The diagram is emitted as mermaid generated from the 12 edges. generate-isis-derived-ownership.mjs + 18 tests, wired into inventory/verify + CI. Mutation-tested: 36/51 killed, then the 9 surviving logic mutants were killed by extracting the verdict and the two invariant predicates as pure functions and exercising every branch — including pointer-only and absent, which today's schema never reaches. --> <!-- I6.1.d (consistency seams) and the I6.1.e ADR remain. -->
    • I6.1.d Resolve cross-store transaction/consistency, identity/version, tenancy, retention/deletion, migration, failure, and recovery seams. Evidence: consistency decision record. <!-- isis-consistency-seams.json: all seven named seams, each with what was observed, the conflicts actually unresolved (13), and a decision recorded against them. The decisions follow from one measured fact: Isis opens 10 independent connection pools at the one database (9 createPostgresClientFromUrl sites + 1 Prisma client), and a transaction belongs to a pool — so 10 of the owner pairs I6.1.b resolved CANNOT be written atomically together however much they share a database. Inside a pool the binding is opt-in: recordAuditLog() writes output_audit_logs through an executor DEFAULTING to this.db, so a caller who forgets the handle gets a row that survives the surrounding rollback. Identity: same physical table keyed two ways (retention_policies id vs name; workflow_stars id vs the (user_id, workflow_id) pair), and 4 owners carry no version/updated-at column so a lost update cannot be detected. Tenancy: FOUR shapes across 7 owners and 4 owners (gate-verdict, release, run, stage) with none at all. Retention: 6 owners cannot distinguish a deleted row from one that never existed, and the retention policy itself lives in a dual-authority table keyed two ways. Migration: 3/7 owners are created by a running service, 8 tables have two declaring authorities. Recovery: only stage can resume (ordinal); job_outbox's replay columns belong to the OUTBOX, not to the run, and the artifact does not credit them to it. generate-isis-consistency-seams.mjs + 21 tests, wired into inventory/verify + CI. Mutation-tested (41/77, then 5/5 and 6/10-with-3- sites-removed on targeted re-checks). It caught me FABRICATING a finding: a name filter reported rowToOutboxEntry — a row-to-object mapper — as a write escaping a transaction. A helper is now only a writer if its own body contains an INSERT/UPDATE/DELETE, and the same fix surfaced the REAL opt-in-executor hazard the name filter had missed. Also removed a dead authorityOf helper and an untestable x && seen.add(x) idiom. -->
    • I6.1.e Ratify an ADR naming owner/service/store per aggregate and explicitly prohibiting shadow canonical tables. Evidence: approved ADR and schema inventory check. <!-- partial: BOTH halves of the evidence exist; only the RATIFICATION is open. (1) The ADR is written: docs/adr/ADR-I1-isis-aggregate-ownership.md (Status: Proposed) names an owner per record type with the rule that chose it, records the derived-role split and the seam decisions, states the prohibition verbatim, and gives the four options considered with why each was rejected — every quantitative claim drawn from the committed evidence. (2) The schema inventory check is ENFORCING, not descriptive: scripts/v1-workbenches/isis-shadow-table-gate.mjs is a two-way ratchet wired into verify:inventory:v1-workbenches + CI. A NEW dual-authority table fails the build; a NEW runtime-DDL table fails; a system of record that loses its live access path fails; and PAYING DOWN a baseline entry without tightening the list fails too, so the ceiling cannot quietly become permission. The 8 dual-authority and 13 runtime-DDL tables are recorded as measured debt (I6.2 unwinds them; a gate cannot delete a table that has data in it). 11 tests, four of which force each violation class and one of which asserts the PROCESS exits non-zero — a report nothing acts on is not a gate. Unchecked because "ratify" is the named service owners' act, per the S0.6.f precedent. -->
  • I6.2 Add only missing schema through expand/backfill/verify/contract migrations. Evidence: production-size rehearsal.

    • I6.2.a Diff required contracts against ratified owner schemas and record each reusable field/table/index versus genuine gap. Evidence: schema gap report. <!-- isis-schema-gap.json: 131 fields across 14 COMMITTED contracts — the Zod payloads Isis actually PUBLISHES (libs/contracts/src/events/ isis.ts) and the brief record the I2 workbench validates against — diffed field by field against the systems of record I6.1.b resolved (imported, not restated; the suite asserts the two agree). Four buckets, and the middle two are the point: 15 reusable-exact (add nothing), 19 reusable-ELSEWHERE (the owner lacks it but a DECLARED-ONLY rival already models it, so it is a migration out of a table nothing reads, not new design), 23 genuine gaps, 74 belonging to a record type with no table at all. HEADLINE: Isis publishes projectId and userId on EVERY job event and the live owner (service:jobs) can store neither — both sit in prisma:generation_jobs, which nothing reads. Gaps carry a nearestOnOwner HINT (sizeBytes~size, filename~name) that deliberately does NOT equate them: a rename is still a column that does not exist, and the judgement belongs to whoever writes the migration. Index half is scored on the axis I6.3 makes mandatory — a tenant-LEADING index — and 4 owners are blocked-on-missing-tenancy-column, where adding an index is not the fix because the column does not exist. 4 record types need a table. generate-isis-schema-gap.mjs + 20 tests, wired into inventory/verify + CI. Mutation-tested 65/81, then 11/11 on a targeted re-check after extracting isServiceSourceFile/ownerStoreFor/isTenantLeading/indexStatus and the two meaningfulness predicates as pure functions. One survivor needed a better FIXTURE, not more code: the cross-database key had to be PRESENT in the census, or dropping the crossDatabase check still returned null through the lookup miss. --> <!-- I6.2.b-g (expand/backfill/verify/cutover/contract + a production-size rehearsal) need a live database and stay unchecked. -->
    • I6.2.b Add backward-compatible nullable/default-safe columns, tables, constraints, enum representations, and dual-read/write support without destructive rename/drop. Evidence: expand migration tests.
    • I6.2.c Backfill in resumable bounded batches with checkpoints, idempotency, throttling, tenant isolation, error quarantine, and observable progress. Evidence: backfill integration tests.
    • I6.2.d Verify row counts, null/uniqueness/FK/check invariants, sampled semantic equality, hashes, read paths, and lag before cutover. Evidence: verification report.
    • I6.2.e Switch readers/writers through a reversible compatibility phase and monitor errors/latency/data divergence. Evidence: cutover rehearsal.
    • I6.2.f Contract old schema only after rollback window and dependent version inventory prove no consumers remain. Evidence: contract gate.
    • I6.2.g Rehearse expand/backfill/verify/cutover/rollback/contract against production-size/sparsity/skew with resource and lock measurements. Evidence: migration rehearsal report.
  • I6.3 Add indexes for actor, tenant, project, workflow, model, provider, status, asset kind, dates, gate state, and lineage. Evidence: query plans.

    • I6.3.a Capture representative list/detail/search/filter/lineage/blast- radius/release query shapes, cardinalities, sort orders, and latency targets from production-scale fixtures. Evidence: workload catalog. <!-- partial: isis-workload-catalog.json covers SHAPES and SORT ORDERS (and more) but NOT cardinalities or latency targets, so the cell stays unchecked. 73 read paths across 8 service files and 18 tables in 7 shape classes, with filter columns, sort columns, pagination style and index support, extracted per SQL STATEMENT (not per method) so a method issuing two queries cannot hand both tables every predicate either uses. Index support is three-state: leading-index, partial-index-predicate (a WHERE deleted_at IS NULL on the index serves that filter), unindexed; sorts use the COMPOSITE rule (an index on (tenant_id, sequence) walks ORDER BY sequence when tenant_id is filtered). Predicates are split MANDATORY (inline in the statement, or seeded into the conditions array literal — applied by every execution) vs OPTIONAL (conditions.push, or pushed by a scope helper the method hands its array to — applied per request), because only the mandatory half belongs in a fixed index. Result: 38 filter columns, 24 led by an index, 1 partial, 13 leading none of their own, and exactly 1/55 filtered reads has NO indexed entry point: queryProvenance on provenance(model, seed, workflow), the blast-radius question, which scans. 1 unserved sort and 1 deep-page risk, both the webhook list. Cardinalities and latency targets are DECLARED ABSENT with the reason: they are properties of data and of product commitments, not of query text, and a guessed row count would justify an index nobody measured. Both need I6.3.e (a database). generate-isis-workload-catalog.mjs + 31 tests, wired into inventory/verify + CI. Mutation 55/86 then 10/11 targeted. It caught FIVE false findings in my own first draft, each removed: FROM unnest(tags) and FROM avg(x) catalogued as TABLES; a GIN index and every partial-index predicate unread, so indexed columns read as unindexed; PRIMARY KEY and UNIQUE-constraint implicit indexes unread; ORDER BY jobs.created_at recorded as a sort on a column called jobs; and (?:=|...)\b matching nothing for WHERE x = $1 — no word boundary after = — which silently dropped most predicates in the codebase. --> <!-- CORRECTED while building I6.3.c, which is what found the bugs: the catalogue went 53 -> 73 queries and 16 -> 18 tables. THREE more extraction faults, all silent, all of which made a real read path look like a table nobody queries. (1) The method-body scanner took the FIRST { after the signature, which is the one in query: Q = {} — so 19 method bodies parsed as an empty object literal and vanished, including every paginated list, the outbox claim loop claimOutboxEntries, the dead-letter feed, and BOTH audit-log readers. The body brace is now found by skipping the parameter list with balanced parens and then the return annotation, with ; meaning a declaration that must not adopt the next method's body. (2) Predicates SEEDED into the conditions array — const conditions = ['project_id = $1', 'deleted_at IS NULL'] — were invisible because only conditions.push was read, and the seeded ones are exactly the mandatory scoping predicates; that reported the webhook subscription list as an unscoped unindexed status scan when it is project-scoped and served by a partial index on that pair. (3) Conditions a method DELEGATES to a helper (this.appendAccessScopeConditions( conditions, ...)) were unread, which is the whole of tenancy on the output reads. (4) A plain WITH x AS (...) was classified lineage, filing the queue dispatcher under a shape class meant for graph walks; only WITH RECURSIVE is lineage now. (5) Found while building I6.3.d: predicates introduced by OR were invisible, because the extractor anchored on WHERE/AND only. That hid target_id in the one bidirectional lineage read — source_id = ANY(...) OR target_id = ANY(...) — so the REVERSE lineage column read as one no query ever touches. Disjuncts are now a class of their own and deliberately NOT mandatory: a composite spanning both branches of an OR serves neither, since the row may satisfy the other branch. --> <!-- I6.3.b-d are index DESIGN cells that build on this catalog; I6.3.e-f need a database and migrations. -->
    • I6.3.b Design tenant-leading composite/partial/covering indexes for actor/project/status/kind/date queue and pagination access paths. Evidence: index rationale. <!-- isis-index-rationale.json: 50 distinct access paths from the I6.3.a catalogue (one per (table, filters, sorts) triple — the same shape read from three methods is ONE index question), each with a status and a rationale that says why. Candidates are built tenancy → equality filters → sort column, because that is the order an index can be walked, and capped at 4 columns with the overflow REPORTED rather than truncated. Only MANDATORY filters enter a composite: a list endpoint that MAY filter by eight things does not want an eight-column index, so builder-supplied (per-request) predicates are excluded — that alone turned an 8-column proposal into (tenant_id, created_at). Coverage is by PREFIX not set membership, since an index on (status, tenant_id) does NOT serve a query leading with tenant_id. Result: 3 proposals, 20 already covered — most of them by a PRIMARY KEY, because a read filtering the whole PK resolves one row and a tenant-leading composite in front of that adds nothing — 19 blocked on 10 tables with NO tenancy column, where the fix is the column (I6.2), and 8 on 5 tables (model_registry, model_registry_audit_log, outputs, provenance, workflows) in a status this pass ADDED: blocked-on-untenanted-query. generate-isis-index-rationale.mjs + 22 tests, wired into inventory/verify + CI. Mutation 35/50 then 9/9 targeted after extracting preferStore / isPaginated / filtersPrimaryKey and the invariant predicates. Two of my own errors caught by reading the output: the first draft picked whichever authority came first for a dual-authority table, reporting provenance as having no tenancy column when its LIVE half has tenant_id, project_id and owner_id; and it proposed (tenant_id, id) composites in front of primary-key lookups. --> <!-- CORRECTED on the I6.3.a re-extraction, and the correction is the finding: on the richer catalogue this cell proposed NINE indexes and all nine led with a tenancy column the read never constrains — not on every execution, not per request, not even as a sort. A B-tree is entered at its leading column, so every one of them would have served its query not at all; four of that class had already shipped in the previous pass. The rule is now an invariant (everyProposalIsUsableByItsQuery) and the paths are reported blocked-on-untenanted-query: the table HAS tenancy, this read crosses tenants, and the fix is the query rather than an index that would read as a fix while changing nothing. The somethingIsProposed invariant went with them — requiring at least one proposal was an assumption about the answer, replaced by statusesDiscriminate (>=3 buckets populated), which checks the method instead. A surviving proposal now also records usableOnlyWhenScoped when its tenancy predicate is per-request: that index serves the SCOPED call, and the same endpoint called without a scope still reads globally. 1 of the 3 is in that state. -->
    • I6.3.c Design workflow/model/provider/version/status/time indexes for routing, operations, blast radius, and cost analysis. Evidence: query plan fixtures. <!-- isis-plan-fixtures.json: 30 access paths on the OPERATOR axis — one per (table, MANDATORY filters, sort) triple, since the plan depends on what every execution constrains — classified into the four purposes by the columns they constrain, never by method name. I6.3.b was the tenant-leading axis; these are asked ACROSS tenants and lead with workflow/model/provider/version/status/time. THE EVIDENCE IS A DERIVED PLAN, NOT A MEASURED ONE, and the artifact says so in every row: measured: false, no EXPLAIN text, no cost, no row estimate, no timing, and a test that fails if any of those words appear. The rule is stated once and applied everywhere: a B-tree is entered at its leading column and walked while each successive column is pinned by equality; the longest such prefix is the entry point; an ORDER BY is free only when its column follows a fully-pinned prefix; a PARTIAL index applies only when the read asserts its predicate, and the predicate it enforces filters exactly as a key column would. Where more than one index is viable, ALL of them are published with chosenBy saying the winner is a structural tie-break and that the real choice is a cost decision for I6.3.e — publishing only the winner would present a guess as a derivation. Result: 14 already-planned, 6 proposed (each with runnable CREATE INDEX CONCURRENTLY DDL, kept only when it lengthens the usable prefix or removes the sort step), 10 that no index can help, 4 sequential scans. Equality columns are ordered ALPHABETICALLY and the artifact says why: every order gives the same usable prefix, so all serve the query, and which is CHEAPEST needs selectivity nobody has measured — ordering by a guessed selectivity would publish a decision nobody made. Soft-delete flags are hoisted into the index PREDICATE rather than the key, because a read that always asserts deleted_at IS NULL gets one value on every entry it can use: it discriminates nothing there and only widens the tree. That is a structural argument, not a claim about how many rows are deleted, and it is what the webhook and workflow indexes already do. TWO HEADLINES, both about questions Isis cannot answer rather than indexes it is missing. (1) cost-analysis has ZERO read paths, and the live provenance table records cost and generation_duration next to provider, model, workflow, tenant_id and generated_at. The data is written on every generation and never queried, so there is nothing to plan and no index to design: an index serves a query, and this question has none. (2) the provider dimension has zero access paths while provenance.provider and isis_provenance_bundles.{endpoint_id, region} are written on every row — no failover query, no provider blast radius, nothing to index. Coverage is reported per purpose AND per dimension, and "recorded but never constrained" is computed per TABLE, because pooling columns across tables would report a dimension as covered on the strength of a different table using the same word. 9 of the 30 reads already cross tenants; the artifact states that the proposed indexes make an existing global read cheaper rather than opening a path that did not exist, which is the distinction I6.3.d has to keep true. The one persistence path this extraction cannot see — the Prisma-delegate pipeline_state_* repository — is DECLARED, so its absence does not read as a table nobody queries. generate-isis-plan-fixtures.mjs + 30 tests, wired into inventory/verify + CI. Building this cell is what found the four I6.3.a extraction faults and the nine unusable I6.3.b proposals recorded above. --> <!-- I6.3.e-f need a database and stay unchecked. -->
    • I6.3.d Design gate subject/profile/state/freshness and forward/reverse lineage indexes without introducing cross-tenant paths. Evidence: gate/lineage plans. <!-- isis-gate-lineage-plans.json: 22 access paths (21 gate, 1 lineage), each with a DERIVED plan (measured=false, no EXPLAIN, no cost or row estimate) and 3 proposals. The cross-tenant clause is taken literally first and then stated properly: an index never changes WHICH rows a query returns, only how they are found, so no index can introduce a cross-tenant read. What it can do is make a cross-tenant path cheap enough to be taken. So every proposal is held to two branches — lead with tenancy where the table has it (asserted by noProposalWidensScope), and where it has none, say unavoidablyCrossTenant and NAME what scopes the read instead. Four scope states, and the middle two are the point: tenant-column (7 paths), guarded-in-sql, guarded-in-application-code (1), unscoped (14). A predicate in the SQL and a filter applied after the rows come back are both "scoped" and are not equally scoped. THE LINEAGE FINDING: the cross-project lineage policy (assertCrossProjectLineageAllowed) is called from 2 sites and BOTH are write-path — the guard runs when an edge is CREATED, never when the graph is READ. On the read side getLineage walks parent/child ids with the caller's OPTIONAL scope, and the lineage_edges query itself carries no tenancy predicate at all: it reads edges across every tenant and the process discards the out-of-scope ones in JS afterwards. lineage_edges has no tenancy column, so that cannot be fixed with an index — an index there makes more out-of-scope rows arrive faster. Guard call sites are LOCATED in the source (not asserted), classified read/write-path by whether the enclosing method issues writes, and — the correction that mattered — attached only to tables that method actually reads. The first draft listed the output-lineage guards as protecting the outbox dead-letter feed, which is a safety claim about code that never runs there. THE GATE FINDING: all four gate dimensions have read paths, and that is the generous reading. 48 gate columns on live tables are constrained by NO read — including every purpose-built gate and profile column Isis has (gates, gate_ids, consistency_profile, lens_profile, asset_family, pipeline_stage, output_package_schema) and the promotion state machine's own status/from_status/to_status. The single profile column any read constrains is the staging environment. 60 indexes on tables Isis reads are entered by no catalogued read, 17 of them on gate columns — the mirror image of a missing index, write cost paid for reads that do not exist. The pipeline_state_* entries in that list are DECLARED as a limit of the extraction (they are read through Prisma delegates, which a SQL-statement catalogue cannot see) rather than left to read as a finding. A fifth I6.3.a extraction fault fell out of this cell: predicates introduced by OR were invisible, so the ONLY bidirectional lineage read in Isis — source_id = ANY(...) OR target_id = ANY(...) — was catalogued as forward-only and the REVERSE lineage column read as one no query touches. Disjuncts are now extracted into their own class and kept OUT of mandatoryFilterColumns, because a composite spanning both branches of an OR serves neither; Postgres combines separate indexes instead. generate-isis-gate-lineage-plans.mjs + 17 tests, wired into inventory/verify + CI. --> <!-- I6.3.e-f need a database and stay unchecked. -->
    • I6.3.e Measure read latency, rows scanned, index size, write amplification, lock/build impact, and redundant-index overlap at production scale. Evidence: before/after benchmark.
    • I6.3.f Create/drop indexes with online/concurrent safe migrations and regression-plan assertions. Evidence: migration/query-plan tests.
  • I6.4 Implement transaction-bound audit and outbox writes for mutations. Evidence: rollback and retry integration tests.

    • I6.4.a Define canonical mutation/audit/outbox envelope with actor/ authority, tenant/project, request/correlation/idempotency, aggregate/ revision, operation, before/after references, reason, policy, and time. Evidence: event schemas. <!-- isis-mutation-envelope.json: the ledger's own nine groups taken verbatim as the field list (15 fields, 12 required), each carrying the specific question that becomes unanswerable without it — a field nobody can name a failure for is decoration. Diffed against every persisted mutation-record surface, and the first finding is how many there are: NINE tables plus the JSON payload the outbox stores. audit_log and canonical_audit_events on the Prisma authority; workflow_audit_logs, model_registry_audit_log, output_audit_logs, generation_event_ledger, isis_provenance_ledger, workflow_promotion_transitions and job_outbox on the service-DDL one. NONE of the ten carries all 12 required fields. Coverage is four-valued and the middle two are the whole argument: carried (the column exists and is NOT NULL), optional (the same column, nullable — the schema's PERMISSION to omit it), partial (a near-miss, published with the reason it is not the thing), absent. 71 of 150 cells are absent and 26 more are near-misses, which is why the gap has stayed invisible: every uncarried field has a lookalike column somewhere, so the schemas read as though they cover it. THE FIELDS NOTHING GUARANTEES: authority appears nowhere at all except as actorRole on one surface — a role is a property of the actor, not the grant the action was taken under, so when a grant is revoked nothing identifies what it authorised. tenant exists on four surfaces and is NULLABLE ON ALL FOUR: there is no surface in Isis on which a mutation record is guaranteed to say which tenant it belongs to. policy exists on exactly one surface, nullable. revision is carried by none — the two sequence columns are per-parent ordinals that order the records without ever naming the version of the aggregate. idempotency is carried ONLY by the outbox; all eight audit surfaces have nothing but a per-row id, which is the opposite of an idempotency key, so one retried command writes one outbox row and N audit rows. AND THEY CANNOT BE REASSEMBLED: 0 of 44 pairs of distinct surfaces share a correlation column that is non-null on both sides. The three writes I6.4.b must make atomic could not afterwards be recognised as having belonged to one act even if they were. Two nominated surfaces fail the structural definition of an immutable mutation record, and they fail it differently. isis_provenance_ledger has no actor column and no operation column: it records that a bundle was appended, never who changed what. job_outbox is revised in place as delivery progresses — correct for a delivery work-item, and precisely why it cannot stand in for the audit write — and its actor and operation live inside the JSONB payload, so no SQL predicate can select the outbox by who acted or by what operation is pending. Sharpest single column: canonical_audit_events.resourceType is NOT NULL while resourceId is NULLABLE, so the canonical audit event is guaranteed to say a workflow changed and permitted to stay silent about which one. Surfaces are DISCOVERED — nominated by name, then confirmed against five structural checks — and a reverse sweep lists every table meeting the definition WITHOUT being nominated (0), so the roster cannot be quietly short. No database was read: every column, type and nullability is parsed from committed DDL, and the JSON envelope from its zod schema. generate-isis-mutation-envelope.mjs + 53 tests, wired into inventory/verify + CI. Mutation-tested (304 mutants); fixing survivors removed a VACUOUS invariant — optionalIsNeverSatisfied was state === 'optional' implies not state === 'carried', a tautology no input could falsify — replaced with one that checks each state against the nullability of the columns it names. The sweep also caught the type parser swallowing constraint text (INTEGER NOT NULL DEFAULT 0 as a type), which matters because the JSON-copy-vs-reference call is made on the type. --> <!-- Two extraction faults this cell shipped and then caught in its own spot-check, both of which had made Isis look BETTER than it is: (1) the envelope's fields are published under their full path, so metadata.userId normalised to metadatauserid and matched no vocabulary — the outbox envelope was reported as carrying no actor, no project and no correlation id when it declares all three; vocabularies now match the LEAF. (2) A field with a type column and an id column was read as carried whenever the TYPE was NOT NULL, so canonical_audit_events counted as carrying aggregate and actor on every row while both identifiers are nullable. Fields now declare which of their columns IDENTIFY, and only a required identifying column earns carried. --> <!-- I6.4.b-f are write-path and integration cells: they need the audit and outbox writes located at their call sites, and a database. -->
    • I6.4.b Write aggregate mutation, immutable audit record, and unpublished outbox row in the same owner-database transaction. Evidence: atomicity tests.
    • I6.4.c Ensure rollback leaves none of the three writes and retry with the same idempotency key creates no duplicates. Evidence: rollback/retry integration tests.
    • I6.4.d Publish outbox rows with leases/fencing, at-least-once delivery, stable event IDs/order keys, retry/backoff, and dead-letter handling. Evidence: publisher crash tests.
    • I6.4.e Make consumers idempotent against duplicate/out-of-order events and prevent stale projections from authorizing action. Evidence: consumer ordering tests.
    • I6.4.f Apply property classification/redaction to event projections and audit access while preserving authorized forensic evidence. Evidence: event leak tests.
  • I6.5 Implement brief/plan/run/variant read APIs with deterministic pagination. Evidence: contract tests.

    • I6.5.a Define versioned list/detail/filter/sort/include/projection contracts for briefs, plans, runs/stages/attempts, variant sets, outputs, decisions, and lineage. Evidence: API schemas. <!-- isis-read-contracts.json: the ledger's own seven dimensions (version, list, detail, filter, sort, include, projection) across its own nine record kinds, each dimension carrying the question that becomes unanswerable without it, diffed against 127 GET routes extracted from committed source — 51 on the Isis services (Hono) and 76 on the operator BFF (Fastify). 14 of 63 cells are served. THE HEADLINE: SIX of the nine record kinds have no GET route of any kind on either family — brief, plan, stage, attempt, variantSet, decision. Not an unwired API in front of a working store: unimplemented end to end. Only run, output and lineage are readable at all, and even there the ledger's "runs/stages/attempts" triple collapses to one flat record — the served run is a Job, renamed by toWorkflowExecution, with no stage or attempt substructure beneath it. A GET ROUTE EXISTING IS NOT A RECORD BEING READABLE, so every route is classified by what its HANDLER does, four ways: reads-store (59), static-context (40 — returns a module-level constant, the same answer for every caller), not-configured (8 — fail-closed with a machine-readable reason) and other (20). Only the first returns a record, and an invariant enforces that only a store-backed route can make a dimension served. The not-configured four are counted APART from absence deliberately: a route that refuses to fabricate a record it cannot produce is an honest seam, not a defect and not a stub, and the shape being decided while only the authority is unbound is a different amount of remaining work. Collapsing all four into "has an endpoint" is how this surface came to look finished. NO record kind serves projection — not one route declares a field selection anywhere, so every read returns the whole record and field-level authorization (I6.5.b) has nothing to hook into but post-hoc removal. Only lineage serves include, and only via the lineage read's bounded depth. And 15 of 127 routes validate a declared query contract at all; the other 112 accept whatever they are given. generate-isis-read-contracts.mjs + 42 tests, wired into inventory/verify + CI. Mutation-tested (208 mutants, 190 killed); fixing survivors added the version-dimension cases, which had no test at all. --> <!-- Cross-checked against an INDEPENDENT survey of the same surface before publication, which is what caught four silent extraction faults, every one of which had made the surface look better or larger than it is: (1) PROVIDERS.get('runpod') was extracted as a route — the GET pattern now requires a leading /, and the service route count fell to 51, matching the independent count exactly; (2) Hono sub-routers write leaf paths (outputs.get('/')) and their noun lives in the app.ts mount chain (apiV1.route('/outputs', outputs) under app.route('/api/v1', apiV1)), so before resolving mounts output had ONE route instead of eight and run two instead of sixteen; (3) the BFF's output-lineage segment matched neither output nor lineage until segments were split on their hyphens, dropping both routes of that module; (4) direction (ancestors/descendants/both) was classified a SORT key, which reported lineage as having an ordering contract when no lineage read accepts one — a generous vocabulary manufacturing the coverage it is meant to measure. --> <!-- This cell judges neither pagination determinism (I6.5.c) nor authorization (I6.5.b), and says so in notDerivableFromSource rather than letting silence read as approval. --> <!-- CORRECTED while building I6.5.b, which is what found the fault: the BFF was swept as src/routes/admin-isis-*.ts, which is a FILENAME CONVENTION and not a boundary. Six Isis-record GET routes register outside it — src/isis/*.ts and src/generation/*.ts — and the sweep counted 121 routes where there are 127. The 6 are not filler: two of them (GET /v1/generation/jobs, GET /v1/generation/jobs/:jobId) are the ONLY run reads in the repository that filter by tenant at all, so the omission made the surface look emptier than it is — the same fault as making it look fuller, in the other direction. The existing family invariant could not catch it, because it only asks that both families be reached and one glob per family satisfies that; everyRouteGlobMatchedAFile is new and asserts each declared glob individually reached a committed file. Totals moved: static-context 39→40, not-configured 4→8, other 19→20, operator-BFF 70→76. reads-store stayed at 59 and the served-cell count stayed at 14 — the six new routes serve no dimension, because none of them reads a store under this cell's definition (they read an in-process snapshot-backed collection, which I6.5.b resolves separately and publishes as a disagreement with these dispositions rather than editing them here). -->

    • I6.5.b Enforce tenant/project/actor/audience/object/property authorization before query and after joins/includes. Evidence: authorization tests. <!-- isis-read-authorization.json: the ledger's six subjects against the READ PATH of every GET route in the I6.5.a sweep — the handler slice plus every method body reachable from it through the file's own imports, to depth 3. A route is only in the matrix if its call chain terminates in a record source: 56 of the 127 swept routes do, 47 on the Isis services and 9 on the operator BFF, for 336 cells. THE HEADLINE: NOT ONE route in Isis binds tenant. Every tenant predicate on the surface is CONDITIONAL — if (scope?.tenantId) in the output registry, tenantId === null || in the operator BFF — which is isolation for callers who carry the claim and unrestricted access for callers who do not. Both read as tenant scoping in a diff, which is why this cell counts conditional apart from bound rather than with it, and why a subject can be everywhere in the code and nowhere in force. Worse than absence is caller-supplied: 2 cells where the predicate exists and its value comes from the REQUEST (if (query.organizationId), if (query.publicOnly) → is_public = true on the model list), so the caller chooses the scope and omitting the parameter WIDENS the read. 44 of the 56 record reads carry no credential-derived tenant, project, actor or audience predicate at all — including reads of run, output and lineage. The run list is SELECT * FROM jobs with a WHERE clause built only from status, type and workflowId — all supplied by the caller — and the jobs table those routes query has no tenant_id, project_id or owner_id COLUMN to bind even if one wanted to: JobListOptions, the type the query is built from, has no field for one either. The Prisma authority declares a DIFFERENT physical table, generation_jobs, WITH userId/organizationId/projectId and indexes on all three — the scoped twin is the one nothing reads, which is the I6.1 two-schema finding arriving as an authorization hole. The second enforcement point is the surface's one good result: of the 3 reads that follow edges out of the rows they selected, all 3 re-apply the scope to what they reached (the lineage BFS re-scopes per node and filters expanded parent/child ids), and 0 do not. 38 of 56 reads select every column, so property-level authorization has nothing to hook into — no route binds it, and none binds audience either. generate-isis-read-authorization.mjs + 56 tests, wired into inventory/ verify + CI. Mutation-tested: 260 mutants, 229 killed. Fixing survivors is what added the mixed-list cases this cell most needed — that ONE unconditional predicate makes a subject bound however many conditional ones sit beside it (.every, not .some), that a filter must not demote a bound predicate, that a guard-only cell says something different for the actor than for a scope, and that MIN_WHY_CHARS is an inclusive bound. The 31 survivors are equivalent mutants of three shapes, each checked by hand: ?? → || where the left operand is a regex match or a non-empty string and can never be falsy-but-defined; >= → > on caps the current data never reaches (400-line bodies, 60 bodies per route, depth 3); and > 0 → >= 0 inside finding-string ternaries whose counts are non-zero. --> <!-- Mechanism matters and is published per cell: query-predicate (the value is in the statement that selects rows), post-fetch-filter (the rows are fetched and then narrowed or 404'd — real, but it protects what is RETURNED, not what was read, counted or logged), handler-guard (decides WHETHER you may call, never WHICH rows come back), none. A guard is the right answer for actor and never for tenant. --> <!-- Building this cell found faults in its own classifiers, every one of which made the surface look BETTER than it is, and each is now a fixture test: AND ( inside a SQL string extracted as a call to a function named AND, whose five same-named "declarations" dragged a provenance WRITER onto the read path of routes that never call it; async (c) => { matched as a declaration, pulling every other handler in a route file onto every route; a bare resolveOwnerId(c); statement read as a second declaration; ): Promise<{ items: Job[] }> { rejected as not opening a body, which hid the job store and with it the largest finding; SELECT id, project_id … WHERE counted as a project predicate when the column was in the SELECT list; if (!ownerId) throw counted as a row filter rather than a credential guard; and a POST route registered between two GETs read as part of the GET above it (I6.5.a cuts handler slices at the next GET, which is right for counting routes and wrong for reading one) — that last one alone had inflated the read surface from 56 routes to 91. One fault ran the other way and is recorded because under-reporting is also wrong: this.store.get(id) was unresolved, because the receiver is a class FIELD and the probe only looked for a const. That made the Postgres job store unreachable from JobService.getStatus and dropped real unscoped reads out of the matrix entirely — 53 routes where there are 56. The fail-closed rule needed the same care in the other direction: a predicate that is conditional in the callee is unconditional at the route when the handler gets the value from a resolver that THROWS, so resolveOwnerId makes actor bound even though the store method is written defensively. Scoping that check to the subject it resolves is load-bearing — asking whether ANY body on the path throws erased every conditional on the surface, since all of them reach requireAuth. --> <!-- Two invariants were written and deleted: one asserting some cell is conditional alongside some bound, one asserting some route expands. Both are assumptions about the ANSWER — a surface where nothing is conditional is a finding, not a broken artifact. What must hold is that the classifier CAN produce each state, which is a property of the code and is tested against fixtures instead. --> <!-- 21 routes are classified differently here than by I6.5.a's handler disposition, and each is published with both verdicts rather than reconciled: I6.5.a asks what a handler AWAITS, this cell asks what its chain terminates in, so a handler awaiting a service whose method returns a constant is reads-store there and constant here. This cell judges neither pagination (I6.5.c) nor whether the value bound is the CALLER'S own — a header-supplied tenant id and an authenticated claim are indistinguishable in source — and says so in notDerivableFromSource rather than letting silence read as approval. --> <!-- The committed authentication posture of each service is published beside the matrix (allowAnonymous: NODE_ENV === 'development', requireProjectScope: true, middleware on /api/*), because a conditional cell reads differently depending on whether something in front already refuses credentials with no claim — and because that answer is not the same in every deployment. --> <!-- UNCHECKED, and it was briefly checked in error — recorded here because the reasoning is the point. This cell's verb is ENFORCE and its named evidence is AUTHORIZATION TESTS. What exists is the matrix above: a source-derived measurement of what is enforced today, which is the thing the enforcement will be written against and is NOT the enforcement. No production code changed and no authorization test was added, so the named evidence does not exist and the box does not flip. Publishing a thorough artifact is not the same as doing the task the artifact is about. What it would take, precisely: for the output registry the columns exist, so binding tenant/project unconditionally is a real but bounded change to appendAccessScopeConditions and its callers, plus a cross-tenant test per read. For the run kind it is BLOCKED — the jobs table those routes query has no tenant_id/project_id/owner_id column, so there is nothing to bind until the expand/backfill/contract work in I6.2.b-g runs, and I6.2.b-g need a database. The matrix names the 44 unscoped reads and the exact predicates a cross-tenant test should target, so the work is specified; it is not done. -->

    • I6.5.c Implement allowlisted filters/sorts with stable unique tie-break, bounded page size, opaque signed cursor, and snapshot/continuation semantics. Evidence: pagination mutation tests. <!-- @isis/read-pagination (libs/isis/read-pagination) + 54 tests, mutation-tested: 194 mutants, 174 killed (90%). Unlike I6.5.b this cell's verb is satisfiable without a database — a cursor, an allowlist, a tie-break and a page bound are pure functions — so it is IMPLEMENTED, not measured, and wired into a real read path (GET /api/v1/outputs/page, 10 route tests). The five obligations, each with the failure it prevents: ALLOWLIST — the caller names a PARAMETER and the contract decides the column, so sortBy=id;DROP TABLE outputs is refused rather than reaching the query builder. An undeclared filter is REFUSED, not ignored: ignoring it is how a caller comes to believe a narrowing happened that did not. TIE-BREAK — every ORDER BY ends in the contract's unique column. Without it two rows sharing a created_at have no defined order between them, so LIMIT/OFFSET shows one twice and skips another; the walk test paginates a 10-row table whose middle FIVE rows share one timestamp and asserts every row appears exactly once, at six page sizes, in both directions. BOUNDED PAGE — an over-large page is REFUSED, not clamped. A silently clamped page turns "there are no more rows" into a guess. SIGNED OPAQUE CURSOR — HMAC-SHA256 over a canonical body, constant-time compare, and the payload is re-validated after verification because a signature proves the bytes were not edited and not that they mean anything. FAIL-CLOSED with not_configured when no secret is bound: a default secret would issue cursors that verify in a service that cannot distinguish a forged one, which is a guarantee in name only. SNAPSHOT/CONTINUATION — the cursor pins version, resource, filters hash, sort, direction and page size, and a continuation disagreeing with any of them is refused rather than silently changing the scan. Every page, INCLUDING the first, is bounded by the snapshot the scan opened at, so a row written mid-scan cannot shift the pages; the walk test inserts one mid-scan and asserts it neither appears nor displaces anything. --> <!-- Two things are documented rather than hidden. A row BACKFILLED behind the cursor is never revisited by a forward keyset scan — that is the price of stable pages, and there is a test asserting it as expected behaviour rather than leaving it to be discovered. And a nullable sort column cannot use a row-value comparison: (a, b) < (NULL, x) is NULL, not false, so the page comes back empty forever. SortSpec.nullable is mandatory and the predicate is expanded, with the NULLS placement in the ORDER BY matching what the predicate assumed — the two disagreeing is how a keyset scan skips rows. --> <!-- Mutation survivors (20 of 194) are all equivalent and were each checked: 7 mutate TYPE annotations, erased at runtime; 6 are ?? → || on spec.allowed, where the fallback [] is truthy so both agree; 5 are || → && in guards whose second disjunct is unreachable once the first is false under the type system; 1 reorders a comparator that only has to be deterministic; 1 is a union's unreachable default branch. Fixing the REAL survivors is what added the wrong-TYPE cases — a query string can deliver an array or an object for a repeated parameter, and a guard that tests only emptiness passes it straight through — plus the falsy-value fingerprint cases, where ?? → || collapses size = 0 and size IS NULL into the same scan. --> <!-- The route is GET /outputs/page, deliberately separate from GET /outputs rather than a mode flag on it: the two make different promises about page stability, and one endpoint that sometimes guarantees it is the harder thing to reason about. The follow-through below upgrades the existing route too without pretending the two cursor formats or consistency promises are interchangeable. The standard output-registry Vitest configuration now resolves the shared pagination/config packages, and the complete registry suite passes 116/116, including the 10 snapshot route tests and OpenAPI parity. --> <!-- 2026-07-29 CONSUMER-SURFACE FOLLOW-THROUGH — source commit ab10c989ea upgrades the existing GET /api/v1/outputs contract too: its former OFFSET scan now uses the allowlisted sort column plus unique id, fetches one row beyond a bounded 1..100 page, and emits an opaque HMAC-SHA-256 continuation bound to every filter, authorization-scope component, sort key, and direction. Malformed, tampered, cross-query, cross-scope, and cross-sort continuations fail before the database is queried; missing/short deployment keys fail closed. This endpoint advertises consistency: continuation and deliberately does not claim the stronger transaction-like snapshot bound of /outputs/page. Follow-ups a8e373dc26 and 1512613f0c make the source-derived workload classifier recognize both the local builder-style cursor and the shared buildPageWhere path, then lock both production queries to keyset classification. Integration commit 5949061dec gives both endpoints the same documented ISIS_OUTPUT_CURSOR_HMAC_KEY, adds the missing OpenAPI snapshot operation and schema, and makes the full package harness resolve the shared workspaces.

      OpenAPI, the TypeScript SDK, CLI, operator Outputs page, and Chromium
      flow now consume the continuation contract. The UI keeps prior opaque
      cursors for Previous and resets continuation on any filter/sort/page-
      size change. Final focused evidence: 54 shared pagination tests, 116
      output-registry tests, 8 SDK contract tests, 72 web page tests, one live
      Chromium continuation/reset test, 3 CLI tests, and 295 deterministic
      base/Isis generator tests; lint and the production registry build pass.
      The workload census now reports three keyset reads (both output paths
      plus the pre-existing jobs path), 11 remaining limit-offset reads, and
      one remaining deep-page risk. This widens the checked cell's production
      and consumer coverage; it does not close I6.5.a's missing record kinds
      or I6.5.f's deleted/changed-row and role-tier fixture obligations. -->
      
    • I6.5.d Project explicit missing/redacted/unavailable/not-applicable/ stale states and audience-appropriate details. Evidence: projection fixtures. <!-- 2026-07-29 OUTPUT PROJECTION CLOSURE — @isis/read-projection defines and runtime-enforces the six-state discriminated field contract: present/stale fields carry values, missing/redacted/unavailable/not- applicable fields cannot, stale fields carry pinned observation and freshness timestamps, and consumer records cannot carry privileged detail. Its clock is injected, so freshness is deterministic in fixtures and remains fixed across a snapshot scan.

      `GET /api/v1/outputs/projected-page` is the first production projection:
      it accepts only the declared `projection=stateful` shape, derives
      consumer/operator/auditor from authenticated roles or permissions, and
      binds each audience to a distinct signed cursor resource. Callers cannot
      request an audience, and a privileged cursor cannot be replayed through
      another tier. Every one of the 23 declared output fields is present with
      an explicit state. Integrity material, storage locators, subject
      identifiers, and producer metadata are redacted from consumers; missing
      and unavailable explanations differ for operator and auditor needs.
      The OpenAPI union and TypeScript SDK publish the same contract.
      
      The authoritative output fixture reaches all six states from one sparse,
      stale pending document and scans serialized consumer output for protected
      hash, bucket/key, owner, metadata, and internal explanation values.
      Route fixtures additionally prove authentication, all three audiences,
      caller-audience refusal, projection allowlisting, cross-audience cursor
      refusal, scan-pinned freshness, and fail-closed cursor configuration.
      Shared-library tests pass 11/11, the output-registry suite 130/130, and
      the SDK suite 337/337; strict TypeScript, Nx boundary lint, shared/
      registry/SDK builds, OpenAPI parity, and affected inventory generators
      pass. Source-derived I6.5.a evidence now records output projection as
      served (15/63 cells overall; 16/129 GET routes validate a query schema).
      This closes the reusable state contract and production output exemplar;
      the six record kinds already reported as lacking any store-backed read
      remain implementation gaps, and I6.5.f's broader role/page fixture
      matrix remains separately open. -->
      
    • I6.5.e Bound joins/lineage expansion, batch dependent reads, avoid N+1, and verify production-scale query plans. Evidence: performance tests.

    • I6.5.f Publish OpenAPI/consumer fixtures for first/middle/final/empty pages, invalid cursors, deleted/changed rows, and role tiers. Evidence: contract suite.

  • I6.6 Implement guarded dispatch API with idempotency and budget reservation. Evidence: duplicate-submit and stale-policy tests.

    • I6.6.a Define dispatch command/result contracts with exact plan/version, expected state, estimate/confirmation, reservation, actor context, idempotency key, and policy/code versions. Evidence: API schema.

    • I6.6.b Authenticate and authorize route/function/object/property; load the canonical plan rather than trusting client-supplied workflow/model/ cost/tool details. Evidence: mass-assignment tests.

    • I6.6.c Revalidate plan inputs, rights/consent, safety, dependencies, endpoint capability, approvals, quota, budget, and freshness in one guarded transition. Evidence: guard matrix.

    • I6.6.d Atomically bind/consume budget/resource/quota reservation and create durable job/outbox/audit state. Evidence: transaction tests.

    • I6.6.e Return the original command result for identical idempotent retry and conflict for key reuse with different canonical input. Evidence: duplicate-submit tests.

    • I6.6.f Reject changed policy/plan/evidence/reservation between confirm and dispatch with no provider call. Evidence: TOCTOU tests. <!-- MEASURED 2026-08-17 (I6.6.a–f). I6.6 SAYS IMPLEMENT AND MOST OF IT ALREADY IS, which is why this reads as three specific edges rather than an absence. evidence/v1-workbenches/isis-dispatch-guard-contract.json (writer + 14 tests, wired into verify:inventory:v1-workbenches) DRIVES the real JobService with an in-memory JobStore double — the store is the dependency and the service is the thing under test, which is the only direction a double may go.

      WHAT WORKS, EXECUTED AND CREDITED FIRST. Two identical submits under one
      idempotency key produce ONE job id, ONE stored job and ONE outbox entry.
      The job and its outbox row are written inside a single
      `this.db!.withTransaction` over `jobs` and `job_outbox`. A unique-index
      race is resolved by re-reading (`idx_jobs_idempotency_key` is a real
      UNIQUE partial index, and the double raises the same `23505` so that
      branch is reachable). `evaluateIsisDispatch` is fail-closed and correct.
      
      FIRST EDGE — A REPLAY AND A CONFLICT ARE DIFFERENT ANSWERS AND THIS PATH
      GIVES ONE. Executed: the SAME key with `prompt: 'A COMPLETELY DIFFERENT
      PROMPT'` returned THE ORIGINAL job id, `status: pending`, no error — and
      the stored prompt is still the first one. A caller who reuses a key with
      changed input is told their new request was accepted and it was not. THE
      ESTATE KNOWS HOW TO TELL THEM APART AND APPLIES IT ON ONE PATH ONLY:
      `hasRecordedTransition` compares a fingerprint, for STATE TRANSITIONS,
      and on a mismatch it logs "Ignoring transition replay with conflicting
      fingerprint" and then `return true` — so a conflicting replay is DETECTED
      AND THEN TREATED AS AN IDENTICAL ONE. The same asymmetry appears in the
      route's responses: it returns 409 `JOB_RETRY_CONFLICT` and has NO
      submit-conflict code at all.
      
      SECOND EDGE — THE ROUTE AUTHENTICATES INTO ONE FIELD AND THE SERVICE READS
      ANOTHER. `withSubmissionMetadata` spreads `...request.metadata` and then
      overwrites `submittedBy` (unconditionally, from auth), and `projectId`,
      `tenantId`, `organizationId` (conditionally, when the scope supplies
      them). It never sets `userId`. `JobService.submit` reads
      `metadata.userId` as the owner, falling back to a fresh
      `crypto.randomUUID()`, and the schema declares `metadata:
      z.record(z.string(), z.unknown())`. So the AUTHENTICATED PRINCIPAL LANDS
      IN A FIELD NOTHING READS AS OWNER, and the owner comes from a free-form
      client record. Executed: `metadata.userId:
      "A-USER-THE-CALLER-NAMED"` is stored as the owner. That is the field the
      subject-erasure path filters on (see I0.11). Also executed: a
      caller-named `model` is stored as the model, because the submit path
      contains NO canonical plan load of any kind — which is precisely what .b
      forbids.
      
      THIRD EDGE — THERE IS NO CONFIRM PHASE. Zero sites for
      `confirmationToken|estimateId|confirmDispatch|/confirm`, so .f's window
      between confirm and dispatch does not exist and A TOCTOU TEST CANNOT FAIL
      AGAINST A ONE-PHASE FLOW.
      
      AND THE GUARDED TRANSITION IS NOT ONE. Of the nine revalidations .c names,
      **0 happen inside the transaction that creates the job**, and 7 are absent
      from the submit path entirely (plan inputs, rights/consent, dependencies,
      endpoint capability, approvals, quota, freshness). Budget and safety DO
      exist on the path — so the zero is about the transition and not about a
      vocabulary that finds nothing. The budget reservation is
      `activeBudgetGate.reserve(job.jobId, …)` in a DIFFERENT application, keyed
      on a job id that must already exist, so it cannot be in that transaction;
      and `audit`, one of the three kinds .d names, is not written in it at all.
      
      (.a) ONE OF SEVEN CONTRACT MEMBERS IS CARRIED. `JobSubmitRequestSchema`
      carries `idempotencyKey`; `exact plan/version`, `expected state`,
      `estimate/confirmation`, `reservation` and `policy/code versions` are
      ABSENT, and `actor context` is carried ONLY by the free-form `metadata`
      record — which is not a contract carrying it, since that record can carry
      anything and therefore pins nothing. `IsisDispatchRequest`, the one real
      guard's input, carries `{measurement, admissions}` — none of the seven.
      
      TRAPS. (1) Unanchored, `409` matches `'nvidia rtx 4090': 0.5` in the
      route's GPU price table, and the first pattern therefore reported the
      route as returning a submit conflict; the pattern is now the RESPONSE
      shape and a control test pins the collision. (2) A CONDITIONAL SPREAD IS
      AN OVERWRITE — missing `...(projectId ? { projectId } : {})` reported
      `projectId` as never set by the route, which it is, just only when the
      scope supplies a value; that case has its own verdict rather than being
      folded into either extreme.
      
      CELLS LEFT UNCHECKED, ALL SIX. .a needs five absent members and an actor
      context that is a contract rather than a free-form record; .b needs the
      route to make `userId` authoritative and the service to load a canonical
      plan; .c needs seven checks that do not exist and all nine inside one
      transition; .d needs the reservation and an audit row in the same
      transaction as the job; .e needs a canonical-input comparison and a
      conflict response on submit — the fingerprint machinery to build it with
      already exists one method away; .f needs a confirm phase before it can
      have a TOCTOU test. -->
      
  • I6.7 Implement job progress/cancel/retry/replay APIs. Evidence: state and role matrices.

    • I6.7.a Define snapshot and cursor-based event/progress contracts for job, stage, attempt, queue, reservation, cost, errors, and outputs. Evidence: progress API schemas.

    • I6.7.b Implement bounded authenticated progress reads/streams with event deduplication, reconnect, terminal closure, redaction, and no cross- tenant IDs/counts. Evidence: stream tests.

    • I6.7.c Implement cancel commands only for eligible states/roles with expected version, reason, idempotency, propagation status, and audit. Evidence: cancel state matrix.

    • I6.7.d Implement retry as a new attempt only for canonical retryable failures/stages under current policy/budget/dependency checks. Evidence: retry state matrix.

    • I6.7.e Implement replay as a new run identity with exact/current/fork mode, preview, confirmation, and immutable source linkage. Evidence: replay role/state tests.

    • I6.7.f Reject direct terminal-state rewrites, arbitrary retry-from-stage, hidden control changes, and stale expected versions. Evidence: adversarial command tests. <!-- MEASURED 2026-08-17 (I6.7.a–f). The bar is "state and role matrices". THE STATE HALF IS LARGELY BUILT AND WAS DRIVEN, NOT READ, by evidence/v1-workbenches/isis-job-command-authority.json (writer + 12 tests, wired into verify:inventory:v1-workbenches) against the real JobService with an in-memory JobStore double: cancelling a completed job is an idempotent no-op that does not touch the outbox; cancelling a running one sets status: cancelled and records cancellation.reason plus outboxEntriesCancelled: 2 — the PROPAGATION STATUS the cell asks for — and throws rather than reporting success when queue cancellation is unavailable; retry refuses a running job with JobRetryConflictError; and retry from failed produces a NEW job id carrying retry.rootJobId, parentJobId and attempt: 1, which is real immutable source linkage. Four of five probe cases are controls and all four pass.

      THE ROLE HALF IS THE FINDING AND IT IS ONE SENTENCE: THE JOB ROUTE FILE
      READS THE ACTOR 11 TIMES AND COMPARES IT 0. Every
      `auth.user?.sub ?? auth.apiKey?.ownerId` in `routes/jobs.ts` derives an id
      for ATTRIBUTION — `submittedBy`, `actorId`, `requestedBy` — and not one is
      compared to the job's owner. AN ACTOR ID YOU RECORD IS NOT AN ACTOR YOU
      CHECKED, so attribution sites and authorization sites are separate columns
      and are never summed. THE POSITIVE CONTROL IS THE NEXT FILE ALONG:
      `routes/models.ts`, same application, same helper, does
      `existing.uploaderId === userId` FIVE times. The estate knows how to
      authorize a command on a record somebody owns and does it on model routes
      and not on job routes — so any authenticated caller can cancel, retry or
      delete any job by id.
      
      THE ASYMMETRY RUNS ONE LEVEL DEEPER, INSIDE ONE SERVICE: `retry` accepts
      `requestedBy` and `authScheme`, and `cancel(id, reason?)` accepts NO ACTOR
      AT ALL. Retry is at least told who is asking and could compare; cancel
      could not enforce a role if it wanted to. Neither compares. And
      authentication itself is CONDITIONAL — `if (cfg.auth)` wrapped around
      `app.use('/api/*', authMiddleware(cfg.auth))` — so an unconfigured
      deployment authenticates nothing before any of this.
      
      (.f) IS SATISFIED BY ABSENCE RATHER THAN BY A CHECK, and that distinction
      is worth keeping: there is no `PATCH`/`PUT` on `/:id` and
      `store.update(id, Partial<Job>)` is not exposed on any route, so a direct
      terminal-state rewrite has NO PATH — a real positive, and a routing choice
      rather than a guard, since the store method exists. But nothing REJECTS a
      stale expected version because NO COMMAND ACCEPTS ONE:
      `expectedVersion|ifMatch|expectedStatus` appears nowhere in the job schema
      or routes, so optimistic concurrency is absent and .c's "expected version"
      has nothing to compare. No command writes an audit row either — the
      cancellation is recorded in the job's own metadata.
      
      (.a) THERE IS NO CURSOR CONTRACT ANYWHERE (0 sites) and four of the eight
      subjects the cell names — `stage`, `attempt`, `reservation`, `errors` —
      have no schema at all, while the other four do, so the four are not a
      pattern finding nothing. (.b) of five stream properties, two appear
      (`event deduplication`, `reconnect`). (.e) IS CITED RATHER THAN
      RE-DERIVED: I1.12.h already established that the only replay Isis has,
      `replayDeadLetterEntry`, UPDATEs the same outbox and job rows back to
      pending — so there is no new run identity for .e's exact/current/fork
      mode, preview, confirmation and source linkage to attach to. The
      instrument reads that finding out of the committed artifact by dotted path
      and throws if it moves.
      
      TRAPS. (1) The checklist wraps inside compounds and joining the lines
      leaves a space after the break: `cross-\ntenant` became "cross- tenant"
      and no token set matched it. A hyphen or slash before a space is always
      that wrap in these lists. (2) Deriving a method's parameters from the same
      regex that LOCATED its signature made the pattern consume its own closing
      paren, and `cancel(id, reason?)` came back with no parameters — which
      reads as "the service takes no arguments" rather than "the service takes
      no ACTOR". (3) The refusal control retried the job an earlier control had
      just cancelled, and A CANCELLED JOB IS RETRYABLE, so the control silently
      became an acceptance; it now has its own subject.
      
      CELLS LEFT UNCHECKED, ALL SIX. .a needs a cursor contract and four
      subject schemas; .b needs three of five stream properties; .c needs an
      expected version, a role comparison and an audit row; .d needs the
      policy/budget/dependency re-checks (the state guard is done); .e needs a
      replay that produces a new run at all; .f needs a command that accepts an
      expected version before it can reject a stale one. -->
      
  • I6.8 Implement gate run/result/rerun APIs. Evidence: stale result cannot satisfy publish eligibility.

    • I6.8.a Define gate-plan/run/item/progress/measurement/verdict/aggregate/ freshness/rerun contracts with exact subject and evaluator versions. Evidence: gate API schemas.

    • I6.8.b Resolve required gates server-side from asset kind/target/profile/ registry rather than accepting a client-declared passing set. Evidence: gate resolution tests.

    • I6.8.c Create/run/cancel gate batches with authorization, idempotency, resource limits, and immutable subject revision binding. Evidence: run command tests.

    • I6.8.d Read raw measurements/verdicts/regions/uncertainty with field- level authorization and deterministic pagination. Evidence: result API tests.

    • I6.8.e Compute freshness from subject/evaluator/gate-registry/profile/ policy/dependency versions at read and release-check time. Evidence: stale-result tests.

    • I6.8.f Rerun creates new evaluation identities, preserves old results, states reason/scope, and cannot promote release state. Evidence: rerun history tests. <!-- MEASURED 2026-08-17 (I6.8.a–f). The parent's bar is "stale result cannot satisfy publish eligibility", so the first job is finding the record that decides publishing. The estate has exactly one and it is not called a gate: ReviewPackage.releaseCheck, which contract-prisma-alignment.ts declares direct-aligned to the ISIS Prisma schema, and which the rule "Approved review packages must have release-ready, release-valid checks" turns into publish eligibility. Instrument: generate-isis-gate-result-freshness.mts (+13 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-gate-result-freshness.json.

      THE CONTRACT COMPARES INSTANTS FOURTEEN TIMES AND NEVER COMPARES THE ONE
      THAT DECIDES PUBLISHING. `compareTimestamps` runs all over
      `review-package.ts`: decisions must be chronological, audit events must
      be chronological, a delegation cannot resolve before it occurred, and —
      the same comparison this cell wants, made one field along — a stage graph
      cannot be stamped after the package was submitted. Not one of the
      fourteen sites names `lastValidatedAt`. The schema requires that field to
      EXIST when a check claims readiness and never requires it to be RECENT: A
      TIMESTAMP THAT IS ONLY CHECKED FOR EXISTENCE IS A FLAG, NOT A CLOCK.
      DRIVEN, NOT READ: the probe walked a package to `status: 'approved'`
      carrying a release check stamped `1970-01-01` against a stage graph
      stamped `2026-04-23`, and every rule passed. The parent's bar fails.
      
      WHAT IS BUILT IS CREDITED AND WAS ALSO DRIVEN. `ReleaseValidityService`
      resolves readiness server-side under 12 rules from stored state and reads
      nothing the caller supplied, separating merely-ready from strictly-valid
      with named reasons; the schema recomputes three counts and REFUSED to
      reopen an approved stage while the stored check still claimed none were
      blocking — though note which side gives way, since the new FACT is
      rejected rather than the stale CHECK invalidated; and a recompute writes
      the check and never the status, so .f's "cannot promote release state"
      holds on this path.
      
      THE ESTATE HAS ALREADY SOLVED THIS, IN ANOTHER LIBRARY, BETTER THAN THE
      CELL ASKS. `TransactionalGateRerunSchema` in `libs/contracts/run-protocol`
      carries `priorSubjectRef`, `subjectRef` and a `state: 'pending'` literal,
      and `rerunGates` re-opens every verdict whose subject was replaced as a
      NEW pending verdict derived from the NEW subject, appending rather than
      replacing and re-pointing the proof requirements — a new evaluation
      identity, the old result preserved, a rationale on the intervention, and
      no way to be born passing: four of .f's four properties, and it IS driven
      by an application (`autonomous-authoring-store.ts`). Driven here too: the
      schema accepted `state: 'pending'` and refused `state: 'pass'`. It never
      asks how OLD a verdict is because it asks what the verdict is ABOUT. A
      RESULT BOUND TO ITS SUBJECT BY REFERENCE CANNOT OUTLIVE THAT SUBJECT; A
      RESULT STAMPED WITH A TIME CAN ONLY BE CHECKED BY COMPARING THE TIME. The
      release check carries 7 members, 0 versions and NO SUBJECT REFERENCE, so
      there is nothing it can be invalidated BY.
      
      PER CELL. (.a) 0 of 9 named contracts exist on the eligibility path; 2
      (`verdict`, `rerun`) exist only in the run-protocol and 7 exist nowhere.
      (.b) the server-side resolution is real, and `UpdateReviewPackageRequest`
      ALSO accepts a client-declared `releaseCheck`, checked against three
      counts and never against the 12 rules — so the two strict gates only the
      service enforces (a populated evidence set, a governance retentionTag)
      are bypassable by declaring the passing set. (.c) 0 gate routes across
      four Isis HTTP surfaces, 0 gate aggregates among 24 Prisma models, and
      the subject revision the record does bind, `artifactVersion`, is nullable
      with no rule that ever requires it. (.d) the field-level authorization
      machinery exists and works — `@isis/read-projection`, 6 field states
      including `stale` — and projects exactly ONE resource, `output`, not a
      gate result; keyset pagination exists and paginates outputs; and all four
      things .d reads have 0 product sites against 48/116/16/290 sites in the
      reference implementation SOURCES, which is the positive control that
      keeps an all-zero column from being a broken matcher. (.e) 6 of 6 version
      inputs have no site anywhere on the eligibility path. (.f) on the
      eligibility path a recompute replaces the check in place, takes no reason
      or scope and keeps no prior result, so there is no evaluation identity for
      a rerun history to be a history OF.
      
      REACH AND WRITE PATH. 0 modules outside `@oshun/review-persistence`
      import it and its only repository implementation is in-memory. Two
      silently dropped fields were found by a RULE rather than by reading —
      every member of the package schema must be set by the assembler and every
      member of the update request merged by the store, both derived from the
      Zod schemas at runtime — and both are FIXED here: `assembleReviewPackage`
      never set `stageGraph`, so every `create()` threw and 97 of the library's
      137 tests were red; `mergeUpdate` never merged it, so no package could
      leave draft while the schema demanded a versioned binding to do so. A
      THIRD defect is REPORTED AND NOT PATCHED: `DecisionLifecycleService`
      stamps `fromState: stage.state` while the contract requires a history's
      first entry to come from `null` and forbids self-transitions, so the only
      service that records a review decision cannot record one (13 tests red).
      Both sides are coherent; choosing which is canonical changes what a review
      history MEANS in every domain on the shared contract, and that is a
      decision rather than a fix. 22 of the library's tests remain red from that
      cause and from specs that never stamp a stage graph.
      
      TRAPS. (1) The wrap has to be undone BEFORE the phrase regex, not just
      before the split: `field-\nlevel` joins to "field- level" and a pattern
      written `field-level authorization` matches nothing, reading as "the cell
      no longer lists its subjects". (2) Slash-lists and comma-lists cannot
      share a splitter — .f's "states reason/scope" is ONE property carrying a
      slash and splitting on it made four properties five. (3) An unanchored
      `gate` took `/v1/admin/studio/api-gateway-router` for a gate route twice.
      (4) `apps/oshun/bff/src/routes` is not an Isis surface — it holds every
      domain's operator API, and searching all of it reported Metis integrity
      verdicts as an Isis gate verdict contract; the surface is
      `admin-isis-*.ts`. (5) A COMMENT NAMING A THING IS NOT THE THING: three of
      four `verdict` hits were prose in a docblock, so every grep here is
      code-lines-only. (6) `id,` is shorthand for `id: id` and a matcher
      requiring the colon reported a field the assembler sets as dropped. (7)
      AN ARTIFACT THAT COUNTS ITSELF IS NOT A FIXED POINT: the .d positive
      control counted `evidence/v1-workbenches/` wholesale, and this report is
      written there, so every count grew by whatever the previous run wrote and
      `--check` failed the moment after it passed. The control reads the
      reference implementation SOURCES (`**/*.ts`), and a test pins the counts
      against the committed artifact.
      
      ALL SIX CELLS AND THE PARENT LEFT `[ ]`. .a needs the nine contracts on
      the surface that decides publishing; .b needs the declared passing set
      refused or re-derived; .c needs a gate run command and a required subject
      revision; .d needs a gate result to read before its authorization and
      pagination can apply to one; .e needs the six versions on the record, or —
      better, and already proven in this estate — a subject reference that makes
      the versions unnecessary; .f needs an evaluation identity on the
      eligibility path. -->
      
  • I6.9 Implement rights, consent, and provenance APIs with field-level authorization. Evidence: protected-property tests.

    • I6.9.a Define versioned rights/attestation/consent/subject/source/ ingredient/credential/ledger/blast-radius contracts with explicit redacted/unknown/expired/revoked states. Evidence: API schemas.

    • I6.9.b Separate metadata, evidence upload/retrieval, verification, grant/revoke, lineage query, and recovery operations by permission and host. Evidence: operation authorization matrix.

    • I6.9.c Enforce tenant/project/subject/purpose/object/property rights on every nested field, include, search, export, and error. Evidence: protected-property tests.

    • I6.9.d Accept evidence through immutable hashed storage references with content/size/scan/classification/retention validation. Evidence: hostile evidence upload tests.

    • I6.9.e Make consent/right changes versioned commands that trigger outbox/audit/blast-radius state without rewriting consumed evidence. Evidence: revocation integration tests.

    • I6.9.f Return privacy-safe counts/cursors and prevent enumeration through timing, not-found/forbidden differences, or identifiers. Evidence: enumeration tests. <!-- MEASURED 2026-08-17 (I6.9.a–f). The parent's bar is "protected-property tests", so the first job is finding the property and the record that holds it. Instrument: generate-isis-rights-consent-provenance-api.mts (+16 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-rights-consent-provenance-api.json.

      THE ESTATE HOLDS THREE PROVENANCE RECORDS AND THE TWO THAT BIND CONSENT
      PROPERLY HAVE NO ROUTE. (1) `provenance`, output-registry runtime DDL,
      request/response shapes as Zod in `output.schema.ts` and NO contract:
      consent is `consentId: z.string().min(1).optional()` — one free string,
      no referent, no state, no version. (2) `ProvenanceBundle`, the canonical
      contract, `direct`-aligned to Prisma, whose `rights` member carries
      `rightsState`, `consentRequired`, `consentRecordIds`,
      `restrictedRegions` and a training-data evidence state. (3)
      `isis_provenance_bundles` + `isis_provenance_ledger` in
      `libs/isis/ai-providers`: hash-chained, tenant-sequenced, append-only by
      trigger, with `consent_id TEXT NOT NULL CHECK (length(btrim(...)) > 0)`
      — consent as a database-enforced precondition of recording a
      generation. Records 2 and 3 have 0 routes across all four Isis HTTP
      surfaces. Record 1 has three, and a PATCH on it can clear the consent
      id. 12 routes enumerated; the lineage family is the positive control
      that the enumerator finds routes when they exist.
      
      THREE LIVE DEFECTS, ALL DRIVEN AGAINST A REAL POSTGRESQL 16.14, ALL
      FIXED. The lane opened by pointing the real `OutputServiceImpl` at a
      real database and the FIRST CALL FAILED: `INSERT has more target columns
      than expressions` (SQLSTATE 42601). `register()` named thirty columns,
      passed thirty values and wrote twenty-nine placeholders, so EVERY
      `POST /api/v1/outputs` against a database raised 42601 — and no
      provenance record can exist without an output, so this cell's entire
      subject was unreachable. The route tests mock the service and the
      service had no test that reaches a database, which is exactly how a
      statement that can never execute survives review and a type-checker. A
      RULE OUTLIVES THE BUG IT WAS WRITTEN FOR: the instrument now checks
      every fully parameterised INSERT on the Isis surfaces for
      columns == placeholders == distinct placeholders (13 statements, 0
      violations), with the original defect reconstructed as a positive
      control so a silent matcher cannot pass for a clean estate.
      
      With that fixed the probe reached the subject and found two more. A
      SUPERSEDED PROVENANCE RECORD WAS UNREADABLE AND STILL WRITABLE:
      `getProvenance` returns the record through `o.provenance_id = p.id`
      while `updateProvenance` resolved one through `p.output_id = o.id`, and
      nothing forbids a second provenance row for an output — the service DDL
      declares no unique constraint on `output_id` though the Prisma model
      does. Driven: two rows registered, the read returned only the second,
      and a PATCH cleared the FIRST row's consent id and watermark hash where
      no read path could show it. And A LEDGER THAT CANNOT BE CORRECTED MUST
      NOT BE WRITTEN FROM THE PAYLOAD: the ledger event was stamped from
      `updated.*`, the row as it stands AFTER the mutation, so a
      caller-supplied `tenantId` decided which tenant an append-only row
      belonged to — permanently, since `generation_event_ledger` refuses
      UPDATE and DELETE by trigger (the probe's UPDATE was refused). Both are
      fixed and pinned by `output.service.provenance-scope.spec.ts`, whose
      three assertions were confirmed RED against the previous code before
      being confirmed green against the new.
      
      WHAT IS BUILT IS CREDITED AND WAS ALSO DRIVEN. The lineage half of the
      same router resolves a tenant scope on all five of its routes, filters
      parent and child ids through it, refuses cross-project edges by policy
      and audits inside the write transaction. The ledger really is
      append-only, and the consent id a PATCH clears SURVIVES in the register
      event that preceded it — so evidence is not destroyed, it is unreadable,
      because no route reads the ledger. `requireOutputAdminAccess` handles
      both auth schemes. Not-found and forbidden collapse to one 404 on every
      read because the owner filter is a clause of the lookup rather than a
      check after it, which is the strongest form of .f's property and the
      routes get it for free. The BFF's `POST /v1/admin/isis/3d-provenance`
      answering 503 `..._authority_not_configured` is an honest fail-loud
      seam, not a gap.
      
      PER CELL. (.a) 9 nouns; 7 have a contract and 1 reaches an Isis API
      surface — and the noun matcher is DELIBERATELY GENEROUS (any exported
      contract symbol naming the noun as a whole word, so "source" collects
      notification sources too), which makes that 1 an UPPER BOUND. The
      mapping is derived by scanning every contract, not listed: a hand-
      written table in the first draft called `attestation`, `blast-radius`,
      `subject` and `ledger` absent and all four have contracts. `credential`
      and `ingredient` genuinely have none. All four states
      (redacted/unknown/expired/revoked) are declared in contracts and 0 are
      expressible on the provenance API. (.b) 5 of 6 operation classes are
      populated and only `evidence upload/retrieval` has NO operation on any
      Isis surface. CORRECTED IN PLACE: the first version of this note said
      `grant/revoke` was empty too, because the instrument's BFF enumeration
      NAMED one module instead of listing them —
      `admin-isis-output-gallery-actions.ts` carries `revoke-consent` and
      `takedown` bulk actions, tier-gated to `operator-admin`, reversible
      inside an undo window, the whole time. ASK EVERY SERVICE, NOT THE
      INTENDED ONE. The enumerator now walks every `admin-isis-*` module and a
      test fails if it narrows to one again. Note where the revoke lives: the
      only consent revocation in the estate is a GALLERY BULK ACTION on the
      BFF, not an operation on the record that holds the consent binding, and
      it names a `consentRecordId` the provenance API cannot resolve. 5
      distinct authorization predicates across the populated five. (.c) 1 of 6
      rights dimensions is enforced on the
      provenance routes (object identity) against 3 of 6 on lineage; tenant
      and project are 0-of-3 on provenance and 5-of-5 on lineage, which is the
      differential with its own control. `subject` and `purpose` have no
      predicate anywhere. (.d) 3 of 5 validations present, and the content
      check FAILS OPEN — unreachable storage logs a warning and proceeds with
      the client's hash, an unavailable algorithm skips the comparison
      entirely. Driven incidentally: the probe registered an output with a
      declared hash over an object that does not exist, and it succeeded.
      `scan` and `classification` are absent. (.e) the one change command
      carries no expected version, rewrites in place, writes no outbox and no
      blast-radius, and silently DROPS an `expectedVersion` a caller sends
      because the schema is non-strict. (.f) 5 probe cases, 5 accepted.
      
      CAPABILITY REACH, SPLIT RATHER THAN FOLDED. Of 6 measured capabilities,
      2 are reached by a route, 3 are built and called but NOT through HTTP
      (`persistEntry` on every job completion, `loadLedger` and
      `verifyPersistedLedger` from the Studio web server), and 1
      (`queryProvenance`) has no caller anywhere. Folding "no HTTP door" into
      "unreached" would say something false about working code: the
      hash-chained ledger is written and verified — what it has no door for is
      an operator or an auditor.
      
      DECISIONS, REPORTED AND NOT TAKEN. (1) `UpdateProvenanceRequestSchema`
      lets a caller reassign the record's OWN `tenantId`/`projectId`/
      `ownerId`. Nothing filters on them today so it grants nothing today, and
      removing accepted fields is a contract change; the ledger poisoning it
      fed is fixed. (2) Adding the `output_id` unique index the Prisma model
      already declares would remove the superseded-record class at the root,
      but it fails on any database already holding duplicates, so it needs a
      data decision first. (3) Which of the three provenance records is
      canonical, and therefore which one the API should serve, is a product
      decision this cell cannot take.
      
      TRAPS. (1) Slash-lists and comma-lists cannot share a splitter: .b's six
      classes are comma-separated and TWO CARRY A SLASH, so splitting on
      slashes reports ten. (2) A substring matcher says `ResourceSchema` names
      the noun "source"; the matcher splits PascalCase into words instead. (3)
      `three-d-provenance-store` matches a grep for `provenance-store` and is
      a different thing. (4) A hand-authored noun→contract table made four
      false absences — A ROW'S COVERAGE MUST NOT DEPEND ON WHICH FILES A HUMAN
      LISTED. (5) A perl rewrite of `codeOnly(readRepo(x))` → `readCode(x)`
      rewrote `readCode`'s OWN BODY into infinite recursion. (6) The first
      draft of the superseded-record test returned `[]` regardless of the
      statement, so it passed against the defect as happily as against the
      fix; the stub now models the row the way PostgreSQL sees it.
      
      ALL SIX CELLS AND THE PARENT LEFT `[ ]`. .a needs the nine contracts on
      the surface that serves provenance — or, better and already built, the
      surface pointed at record 2 or 3; .b needs an evidence upload and
      retrieval, and needs the revoke that exists to act on the record that
      holds the consent binding rather than on a gallery selection; .c needs
      tenant,
      project, subject and purpose resolved on the three provenance routes,
      and the `@isis/read-projection` machinery — which already projects
      outputs by audience — applied to the record carrying prompt, consent and
      watermark; .d needs a scan and a classification, and a content check
      that fails closed; .e needs a consent change to be a command with a
      version rather than a field write; .f is the closest, and needs its
      counts and cursors measured rather than inferred from the read paths'
      shape. -->
      
  • I6.10 Implement release plan/execute/verify/recall APIs. Evidence: time-of-check/time-of-use race test.

    • I6.10.a Define immutable release candidate/target plan, prerequisite snapshot, confirmation, execution step, verification, receipt, failure, rollback, and recall contracts. Evidence: release API schemas.

    • I6.10.b Plan server-side from exact asset/bundle revisions, target profiles, gates, rights/consent, provenance, approvals, destinations, capacity, cost, and policy. Evidence: planning tests.

    • I6.10.c Require authorized confirmation bound to the unchanged plan/ prerequisite hashes and idempotency key. Evidence: stale-confirmation tests.

    • I6.10.d Revalidate every prerequisite immediately before atomic release transition and each external target write. Evidence: TOCTOU tests.

    • I6.10.e Execute target steps with durable checkpoints, idempotent writes, partial-failure visibility, compensation/rollback, and immutable receipts. Evidence: failure recovery tests.

    • I6.10.f Verify final external/registry/gallery bytes, hashes, manifests, metadata, visibility, and URLs before success. Evidence: post-write verification tests.

    • I6.10.g Recall by authorized case/scope with blast-radius preview, propagation, verification, and immutable history. Evidence: recall API tests. <!-- MEASURED 2026-08-17 (I6.10.a–g). The parent's bar is a time-of-check/time-of-use race test, so the first job is finding something to race. Instrument: generate-isis-release-toctou.mts (+10 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-release-toctou.json.

      THERE IS NO ISIS RELEASE API. 0 route paths naming publish, release,
      recall or rollback across all four Isis HTTP surfaces — against a
      control of 11 provenance/lineage paths found by the same literal scan,
      so the zero is a measurement. 0 of the 9 contracts .a names exist. THE
      ESTATE HAS PLENTY OF RECEIPTS AND ROLLBACKS AND NONE OF THEM IS A
      RELEASE'S: 39 verification symbols, 27 receipts, 23 failures, 19
      rollbacks in `libs/contracts`, and not one is release-scoped. That
      unqualified column is the control for the qualified zero — and the
      qualifier is load-bearing, because dropping it reported 7 of 9 present
      when `SetPrecisionRecall` answered for "recall" and
      `AdminWebhookDeliveryReceiptOutcome` for "receipt". EVERY NOUN IN THE
      CELL IS GOVERNED BY ONE ADJECTIVE AND THE MATCHER MUST CARRY IT.
      
      FOUR VERBS LIVING IN THREE PLACES, NEVER JOINED. `plan`: nothing, so
      .b's ten planning inputs have nowhere to land. `execute`:
      `EditorialLifecyclePublisher.publish` — an autonomy binding wired into
      the plane, not a route on any Isis surface. `verify`:
      `POST /api/v1/outputs/:id/verify`, which verifies the stored BYTES
      against the recorded hash and is not the release verification .f asks
      for. `recall`: `withdraw`/`rollback` on the publisher AND a `takedown`
      bulk action on the BFF gallery — two recall paths that do not know about
      each other. 7 of 15 measured properties present.
      
      THE PARENT'S BAR, DRIVEN AGAINST THE ONE EXECUTOR THAT EXISTS, AND
      FIXED. A PREREQUISITE THAT IS NEVER RE-ASKED CANNOT BE WITHDRAWN:
      `publish` evaluated `gatesFor(item)` ONCE and walked four transitions on
      that single snapshot, committing afterwards. Driven: a rights clearance
      withdrawn between the check and the write, and the placement published
      anyway. It now re-asks the gates immediately before the commit. Two
      controls hold either side of the fix — a gate that never holds still
      refuses, and a gate that holds throughout still publishes — so the
      result is not a publisher that refuses everything.
      
      TWO MORE ON THE RECALL PATH, ALSO DRIVEN AND FIXED. A RECALL THAT
      RECORDS NEITHER WHO NOR WHY IS NOT A HISTORY: `withdraw(item,
      _actorId, _reason)` and `rollback` took both and discarded both, leaving
      a takedown transition carrying `from`, `to`, a timestamp and an empty
      gate list. Both are now stamped and survive a snapshot/restore round
      trip. And A HISTORY YOU CAN APPEND TO FROM OUTSIDE IS NOT AN IMMUTABLE
      HISTORY: `get` and `list` handed out the live record whose `history` is
      declared `readonly` and was a plain array, so a fabricated
      `takedown -> published` entry could be pushed onto a publication's
      record; both now clone, as `snapshot` already did. All three are pinned
      by `editorial-publisher-recall.spec.ts`, 7 of whose 8 assertions were
      confirmed RED against the previous code — the eighth is the control that
      must pass either way.
      
      THE ESTATE ALREADY BUILT THE BLAST RADIUS AND NOTHING CALLS IT.
      `buildTakedownCascade` in `@oshun/studio-authoring` returns the affected
      artifact ids for an entry plus its downstream, and REFUSES an empty
      reason or an empty user-visible note. 0 callers. The estate had already
      decided a takedown needs a reason, in a function no takedown path
      invokes — which is why the discard above could stand next to it.
      
      WHAT IS BUILT IS CREDITED. The publisher serializes every record
      mutation, walks the real studio-authoring state machine without skipping
      a state, restores the previous record and rethrows when a commit fails,
      short-circuits an unchanged candidate on a content digest (a real
      idempotency key), and records nothing at all for a refused publish.
      `artifactIdFor` refusing an item with no publication reference is a
      fail-closed guard, not a gap — checked before it was reported.
      
      TRAPS. (1) The checklist wraps after a SLASH in .c
      (`plan/\n prerequisite hashes`), and an `unwrap` that rejoins only
      hyphens leaves "plan/ prerequisite" — the cell then reads as though it
      no longer asks for the binding. (2) .a and .b are COMMA lists whose
      first items carry slashes (`release candidate/target plan`, `exact
      asset/bundle revisions`); a slash splitter reports eleven and twelve.
      (3) `tryEditorialAdvance` rejects an edge the state machine does not
      have BEFORE it looks at a gate, so revalidating as `to -> to` would have
      refused every publish for the wrong reason; the recheck uses the real
      edge.
      
      ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`. .a needs the nine contracts;
      .b needs a planning operation before its ten inputs have anywhere to
      land; .c needs a plan to hash before a confirmation can be bound to one;
      .d has its first half — the recheck before the atomic transition — and
      still needs revalidation before each external target write, which sits
      downstream inside `commitRecord`; .e needs a persisted receipt (one is
      computed and returned, none is stored); .f needs anything at all to
      re-read a written target before reporting success; .g needs
      `buildTakedownCascade` called by the path that performs a takedown, and
      needs the two recall paths to become one. -->
      
  • I6.11 Implement gallery promotion from Yemaya to Oshun editorial. Evidence: original AAA output and curated derivative retain lineage.

    • I6.11.a Define promotion request/candidate/decision contracts linking exact Yemaya project/output/revision and intended Oshun editorial collection/use/audience. Evidence: promotion schemas.

    • I6.11.b Authorize source project/output access, promotion authority, destination editorial scope, and protected-field projection across service boundaries. Evidence: cross-host authorization tests.

    • I6.11.c Revalidate source integrity, asset kind/target compatibility, safety/rights/consent/provenance/gates, and source lifecycle before acceptance. Evidence: eligibility tests.

    • I6.11.d Create an Oshun curated derivative/revision with new identity, bounded editorial metadata/transforms, and immutable ingredient linkage to the original AAA output. Evidence: lineage tests.

    • I6.11.e Route the derivative through Oshun editorial review/release without granting Yemaya approval/publication authority. Evidence: authority boundary tests.

    • I6.11.f Propagate source supersession/quarantine/recall/rights changes through the derivative dependency edge. Evidence: lifecycle integration tests. <!-- MEASURED 2026-08-17 (I6.11.a–f). The parent's bar is "original AAA output and curated derivative retain lineage", which presupposes a derivative — so the first job is finding the operation that makes one. Instrument: generate-isis-gallery-promotion.mts (+17 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-gallery-promotion.json.

      NO PROMOTION ROUTE, AND ONE PROMOTION VERB. 0 in-scope route paths name a
      promotion across the seven Isis/Yemaya/Oshun-editorial surfaces; the 2
      that match belong to the workflow-registry's ENVIRONMENT promotion, a
      staging recipe moving dev → staging → prod. Control: the same literal
      scan finds the 3 gallery paths, so the zero is a measurement. The
      estate's one operation whose name means "move this output into Oshun
      editorial" is the gallery bulk action `send-to-editorial`.
      
      0 OF 3 CONTRACTS, AND THE WORD IS EVERYWHERE. No promotion request,
      candidate or decision contract exists for these endpoints. THE CONTROL IS
      THE UNQUALIFIED COLUMN AND IT IS NOT EMPTY: `ModelCardPromotionDecision`
      and `ModelVersionPromotionDecision` are promotion decisions about a MODEL
      moving through a release lane. 58 exported symbols in `libs/contracts`
      name a promotion, across 13 subjects — model, model card, model version,
      policy bundle, source set, gold set, notebook, learning stage, autonomy,
      research comparative — and not one is an output or a gallery. "The
      vocabulary is missing" is refuted; the SUBJECT is missing. Request and
      candidate have no symbol under either reading: TWO SHAPES OF ABSENCE.
      
      5 OF 6 LINK FIELDS HAVE NOWHERE TO LAND. Measured against the three
      types that carry a promoted artifact end to end — `OutputRecord` (18
      fields), `EditorialRecord` (7), `ProvenanceChain` (12). Yemaya project,
      Yemaya revision, editorial collection, use and audience match no field on
      any of them; the one hit is `OutputRecord.outputId`, the gallery's own id
      rather than a reference to a Yemaya output. AN ORIGIN IS NOT A TIER:
      `OUTPUT_DOMAINS` has 7 members and no yemaya, so the seed AAA output is
      `domain: 'tara'` at the AAA tier — the gallery cannot say an output came
      from Yemaya, only that an AAA creator made it. And `ProvenanceChain`
      starts at an opportunity SIGNAL, so a promoted artifact's story would
      begin the moment Oshun noticed it.
      
      THE PARENT'S BAR, DRIVEN. `send-to-editorial` on a real output: 6 records
      before, 6 after, byte-identical — `applyEffect` has no branch for it, so
      it falls through to `return record`. 2 of the 4 bulk actions change a
      record field (takedown → `takenDownAtUnixSeconds`, re-tag → `tags`) and
      the promotion is not one of them; the 2 that do are the control. The
      store that performs the promotion has snapshot members `context` and
      `invocations` and nothing else, so there is nowhere to write the edge
      either. And the `EditorialRecord` the Oshun editorial release path writes
      has fields artifactId, categoryId, history, publicationSha256,
      reviewItemId, state, title — 0 naming a source, origin, ingredient or
      upstream, even when the review item is handed one.
      
      16 EDGE KINDS ACROSS 4 LINEAGE VOCABULARIES, 1 NAMES THE PROMOTION, AND
      IT HAS NO PRODUCER. `@isis/output-gallery` (6 kinds, nodes `outputId` +
      `job:<jobId>`), `@isis/outputs` (7, nodes keyed to an outputFileId within
      a manifestId), `@yemaya/asset-library` (3, nodes `artifactId` with an
      author and an attribution consent), and `buildTakedownCascade` (0 kinds —
      its downstream ids arrive as an ARGUMENT). Four node namespaces, none
      shared, so no single edge can span two graphs. THE SAME STRING IS A
      MEMBER OF TWO UNIONS: `send-to-editorial` is both a `LineageEdgeKind` and
      a `BulkAction` kind, 4 files construct the action and 0 construct the
      edge — a bare literal scan credits the promotion with writing its own
      lineage. 4 of the 6 gallery kinds appear nowhere in the estate at all.
      
      FIXED: A PROMOTION OF A SUBJECT THE GALLERY NEVER HELD SUCCEEDED.
      `applyBulkAction` walked the RECORDS and skipped ids it did not
      recognise, so `send-to-editorial` naming a Yemaya
      `project/output/revision` returned `ok:true` with a receipt naming it —
      and so did a takedown of outputs that do not exist. In production the
      gallery seeds EMPTY, so until a real output is cataloged every bulk
      action is in that state. `invalid-outputs` is the code the library
      defined for this and had only ever spent on the empty list. Every id must
      now resolve, the batch is refused whole rather than narrowed, and the
      check runs after the tier gate so it cannot become an existence oracle.
      4 of the 8 assertions in `output-gallery-promotion.spec.ts` were
      confirmed RED against the previous store; the other 4 are controls.
      
      FIXED: THE ONLY EDGE THE ESTATE PRODUCES WAS THE ONLY EDGE THE READER
      DROPPED. Both member-gallery readers admitted a lineage edge only when
      BOTH endpoints were visible RECORDS. Every edge the release write path
      emits ends at a `job:<jobId>` node, which is never a record — so the
      lineage panel was structurally empty for every real output. Driven: 1
      edge in the store, 0 reaching the client. Neither reader's tests noticed,
      because their fixtures used record→record edges of kinds no write path
      emits — A FIXTURE IN THE ONE SHAPE PRODUCTION CANNOT PRODUCE TESTS
      NOTHING. `buildRenderableLineage` now owns the rule for both: an edge
      must be ANCHORED in the visible set and must not REACH an output outside
      it, so a foreign tenant's output id is still withheld (pinned as a
      control at both altitudes).
      
      A DECISION, NOT A FIX: THE ARROW. `OutputLineage` names its traversals
      `parentsOf` (incoming) and `childrenOf` (outgoing), and `descendantsOf`
      walks outgoing and is documented "used by takedown cascade" — with 0
      product callers. The two producers of `derived-from` disagree about which
      way it points. `ledger-branch-bridge` writes parent → child and says so
      ("edges point parent -> child … so the takedown cascade of the parent
      reaches the retake") and has 0 product callers. `output-catalog` writes
      output → `job:<jobId>` — the plain reading of the kind's own name, the
      opposite arrow — and is the only producer that writes into a live tree.
      Driven: `descendantsOf(out_job7)` returns `[job:job7]`, so a cascade
      built on it sweeps the producing RUN as downstream, and the panel labels
      the run a Child. Both readings are defensible and the estate holds one of
      each, so settling it moves either a producer or a traversal name — plus a
      normalization for snapshots already written in the old direction. Left
      failing on purpose and recorded in the report's `edgeDirections`.
      
      MORE THAT IS MEASURED, NOT FIXED. (1) `validateBulkAction` branches on
      four tiers and the route hands it the literal `tier: 'operator-admin'`
      while never reading a tier off the auth context — which has none — so the
      gate that decides promotion authority has ONE reachable input. (2) The
      destination is a free string: `queueId` is validated against nothing, and
      no editorial queue or collection registry exists anywhere (the BFF does
      keep a moderation-queue registry, so the pattern is not foreign).
      (3) 2 of the 4 editorial transitions carry an `actorId` — the two recall
      paths I6.10 stamped — and none of the publish walk's, so a promoted
      derivative would be approved by nobody and recorded as nobody; .e's
      authority boundary has nothing to separate. (4) `takenDownAtUnixSeconds`
      is written by takedown, restored by undo, and COMPARED BY NOTHING, so a
      taken-down output can be promoted. (5) Nothing crosses a service
      boundary at all — the promotion never leaves the Oshun BFF process, so
      .b's cross-host projection has no boundary to project across.
      
      5 of 20 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`. .a
      needs the three contracts and five of its six link fields need a field to
      land in; .b needs a real tier, a resolvable destination and a boundary;
      .c needs four eligibility checks (only the precondition that the subject
      exists was added); .d needs a derivative before its identity, its bounded
      metadata or its ingredient link mean anything; .e has the review walk and
      needs the actor on the approval; .f needs `descendantsOf` bound to a
      takedown and a consent revocation that reaches a consent store. -->
      
  • I6.12 Implement signed deep links among Isis, Yemaya, Oshun Studio, and output gallery. Evidence: tenant, expiry, permission, and replay tests.

    • I6.12.a Define allowlisted link intents/hosts/routes and minimal claims for opaque subject reference, tenant/project, audience, issuer, target, issued/expiry/not-before, nonce, and version. Evidence: link schema.

    • I6.12.b Issue links only after source authorization using externally managed rotating signing keys and canonical URL encoding. Evidence: issuance tests.

    • I6.12.c Validate signature/key/status, issuer/audience/host/route, timestamps/skew, nonce/replay policy, and claim limits before routing. Evidence: validation tests.

    • I6.12.d Reauthorize the current actor against the canonical referenced object/property/function after validation; never treat possession as access. Evidence: tenant/permission tests.

    • I6.12.e Avoid credentials, private metadata, raw IDs where opaque refs suffice, return safe expired/invalid/forbidden states, and scrub referrers/logs. Evidence: link leak tests.

    • I6.12.f Test forward/back navigation, logged-out continuation, consumed/ replayed links, moved/deleted/quarantined objects, and host mismatch. Evidence: cross-host journey tests. <!-- MEASURED 2026-08-17 (I6.12.a–f). The parent's bar is tenant, expiry, permission and replay tests, so the first job is finding a signed deep link to test. Instrument: generate-isis-deep-link-tokens.mts (+11 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-deep-link-tokens.json.

      EVERY MINTED LINK IN THE FOUR HOSTS IS A PLAIN PATH CARRYING A RAW ID. 79
      distinct link templates across Isis (2,219 files), Yemaya (2,904), Oshun
      Studio (3,500) and the output gallery (12): 0 signed, 42 carrying a raw
      identifier — `/review?reviewId=rev-1`, `/models?modelId=…`,
      `/scene/<shortCode>` — 36 static, and exactly ONE opaque capability link.
      That the admin links are unsigned is not itself a defect: a navigation
      path behind a route that authorizes independently is safer than a
      capability link. It does mean .a–.c have no subject on those surfaces,
      and it makes .e's "raw IDs where opaque refs suffice" the standing state.
      
      SO THE BAR IS DRIVEN AGAINST THE ONE LINK THAT CARRIES ACCESS. A living
      scene share is a cross-host reference a person is GIVEN: a
      `/scene/<shortCode>` path plus a password, exchanged once for an
      HMAC-signed bearer token. 2 of .a's 8 claims are on it — an opaque
      subject and an expiry. Tenant/project, audience, issuer, target,
      not-before, nonce and version are not, so .c cannot check what the token
      does not carry.
      
      FIXED: A SIGNATURE OVER THE PARSED VALUE DOES NOT COVER THE TOKEN. The
      HMAC is derived from the DECODED short code and the PARSED expiry, and
      both decoders are lax — `Buffer.from(x, 'base64url')` ignores padding and
      stray bytes, `Number.parseInt` stops at the first non-digit. Driven:
      `…YQ=.…`, `…YQ==.…`, `…YQ\n.…`, `….<exp>abc.…`, `….0<exp>.…`,
      `….+<exp>.…` and `….<exp> .…` — seven distinct strings — all verified as
      the single token that had been issued, an unbounded family per issuance.
      Nothing consumes this token today, so nothing was bypassed; it matters
      because .c asks for a nonce/replay policy and A TOKEN THAT IS NOT
      CANONICAL CANNOT BE CONSUMED: a consumed-token set keyed on the string is
      defeated by re-encoding, and the same laxity breaks audit correlation.
      Both segments must now arrive in the encoding the signer emits — 7 of the
      12 assertions in `scene-unlock-token-canonical.spec.ts` were confirmed
      RED, and four controls (subject binding, expiry, wrong key, the canonical
      token itself) hold either side.
      
      FIXED: THE ONE SIGNATURE COMPARISON THAT WAS NOT CONSTANT-TIME. Of the
      seven HMAC-signed primitives in the four hosts, `signed-model-manifest`'s
      hmac-sha256 branch compared the MAC with `expected === signature`, which
      returns on the first differing character, while the ed25519 branch beside
      it went through `crypto.verify`. Replaced with a length-checked byte
      comparison; 7 of 7 primitives now compare in constant time. The spec is a
      REGRESSION GUARD, not red-to-green — a timing change alters no verdict,
      and the case it does guard is the one `timingSafeEqual` introduces:
      a wrong-length signature must be a refusal, never an exception.
      
      .b's ROTATING KEYS EXIST AND NOTHING OUTSIDE THEIR LIBRARY USES THEM.
      `@yemaya/auth` declares `KeyRotationService` with a key status, a stored
      key, `getActiveKey`, `getVerificationKey(kid)`, `getJWKS` and
      `checkAndRotate` — externally managed rotating signing keys, precisely
      what .b names. Consumers outside `libs/yemaya/auth`: 0. So the answer is
      not "the estate has no rotation"; it is that rotation is built and no
      signer reaches it. The scene token holds a single env secret and no kid;
      one primitive in the estate models key identity at all
      (`signed-model-manifest`), and it refuses on mismatch rather than trying
      a set.
      
      THE ESTATE WROTE THE DOCTRINE DOWN AND APPLIED IT TO A CURSOR.
      `libs/isis/read-pagination/src/cursor.ts` opens by naming the two
      obligations — OPAQUE so nothing downstream freezes the encoding, SIGNED
      because "the caller cannot WRITE one" — and refuses to sign without a
      secret rather than substituting a fallback. A deep link is the same
      problem shape: a caller-held reference the caller must not be able to
      write. 4 of the 7 primitives fail closed without a secret; no link does,
      because no link is signed.
      
      POSSESSION IS ACCESS, MEASURED. `VerifySceneUnlockTokenOptions` declares
      exactly `now, secret, shortCode, token` — no actor, tenant, audience or
      viewer — so verification asks nothing about who is presenting the token,
      and a verified token stands in for the password rather than being
      re-authorized against the referenced object. Driven: the same token was
      accepted on all 3 presentations inside its TTL. That is .d's whole
      subject and .c's nonce clause.
      
      .e, DRIVEN THROUGH THE REAL MINT. `librarySharePolicyStore.createShareLink`
      was called, and the deep link it returns carries 9 query parameters
      beside its opaque token — channel, collectionId, exp, flow, permission,
      recipient, role, source, target — including `recipient=…@…`, the
      recipient's email address, in a link meant to be forwarded. The token
      itself is `randomBytes(32)` with only its sha256 persisted and
      `resolveLink` reads the tier from the stored grant rather than the URL, so
      `permission` is decoration rather than an escalation — but the address is
      private metadata riding in the link.
      
      TRAPS. (1) THREE SEPARATORS IN ONE CELL FAMILY: .a's first list is
      slash-separated (`link intents/hosts/routes`), its second is comma-
      separated with slashes inside items (`tenant/project`), and .b is one
      comma-free sentence whose three requirements only separate on `using`
      and `and` — a comma splitter reports .b as a list of ONE. (2) A `**` PATH
      COMPONENT REQUIRES A DIRECTORY TO SPAN: `git ls-files
      'libs/isis/output-gallery/src/**/*.ts'` matches nothing against the flat
      `src/lineage.ts`, and the first version of this report gave the
      output-gallery host 0 files. (3) A COMPARISON NAMED `expected` IS NOT A
      DIGEST COMPARISON — `assertExactKeys` in the signup-verification store
      compares sorted OBJECT KEY NAMES, and a name-shaped test reported a token
      store as comparing a signature unsafely; the compared identifier must be
      assigned from a digest. (4) A CONSTANT IS NOT A VERIFIER:
      `SIGNUP_VERIFICATION_STORE_KEY` matches /verif/i.
      
      6 of 17 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`. .a
      needs a versioned claim set and an intent/host/route allowlist, neither of
      which exists; .b has issuance-after-authorization and an external key and
      needs the rotation the estate already built; .c has a constant-time
      signature check and needs issuer/audience/host/route, skew, a nonce
      policy and claim limits; .d needs the actor reauthorized after validation
      — today a verified token IS the authorization; .e needs opaque refs and
      the recipient's address out of the query; .f needs a journey that crosses
      two of the four hosts, and none exists. -->
      
  • I6.13 Implement capability/health APIs per host audience. Evidence: unauthorized capability details do not leak.

    • I6.13.a Define separate public/curated/AAA/operator capability and health projections with stable available/degraded/unavailable/not-configured/ unauthorized semantics. Evidence: projection schemas.

    • I6.13.b Derive effective capability from canonical audience/host policy, workflow/model/provider admission, dependencies, quota/capacity, and current health. Evidence: derivation tests.

    • I6.13.c Expose only high-level asset/operation/control availability and remediation appropriate to curated/AAA users. Evidence: lower-tier contract fixtures.

    • I6.13.d Expose bounded endpoint/worker/model/workflow dependency detail only to authorized operators with field-level controls. Evidence: operator contract tests.

    • I6.13.e Apply freshness/cache/stale/timeout/partial aggregation rules and never translate unknown/not-configured into healthy. Evidence: failure state tests.

    • I6.13.f Scan response/error/headers/telemetry for provider URLs, model/ node IDs, capacity, tenant, credentials, and hidden features by audience. Evidence: capability leak suite. <!-- MEASURED 2026-08-17 (I6.13.a–f). The parent's bar is "unauthorized capability details do not leak", so the job is to find every capability answer an unauthorized caller can get and read what is in it. Instrument: generate-isis-capability-health-api.mts (+11 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-capability-health-api.json.

      THE SKIP LIST WAS FOR A ROUTE THE MIDDLEWARE NEVER SEES. All three Isis
      services mount `authMiddleware` on `/api/*` — measured, one pattern each
      — and all three ALSO carry `skipRoutes: ['/health', '/ready']`. Both
      entries are vacuous, 6 in total: the middleware was never on those paths,
      so the skip list reads as a deliberate exemption for routes that had no
      guard to exempt them from.
      
      FIXED, AND DRIVEN ON ALL THREE. `/ready` answered with its evidence.
      generation-api handed any caller the storage bucket, its region and
      provider, the queue name, the Redis status and latency, the primary model
      provider with per-provider statuses, and the raw error string from the
      database, queue, storage and provider probes — 39 fields. output-registry
      gave the database probe latency and its raw error. workflow-registry gave
      the storage backend, the latency and its raw error, and driven, that
      error is "DATABASE_URL or ISIS_DATABASE_URL environment variable is
      required" — an anonymous caller was told which environment variable the
      deployment is missing. A READINESS VERDICT IS PUBLIC AND ITS EVIDENCE IS
      NOT: all three now gate the detail on an authenticated caller and keep
      the verdict, the status code and the named dependency ROLES public,
      because that is the contract an orchestrator needs and it names no
      resource. Driven at three levels per service — anonymous (7 fields on
      generation-api, 1 on the others, 0 identifying), operator (39, 5, 6), and
      a FORGED credential (401, not a silent downgrade to anonymous).
      
      THE GATE WAS ALREADY IN EACH SERVICE AND NOTHING USED IT. `optionalAuth`
      — `authMiddleware` with `allowAnonymous: true`, i.e. anonymous allowed
      and a presented credential verified — exists in all three middleware
      modules with zero callers. Two details it needs: its own `skipRoutes`
      must be cleared or it skips itself, and `requireProjectScope` must be off
      or an operator without a project claim is refused a health read.
      
      THE FIVE-STATE VOCABULARY EXISTS AND NO ISIS SURFACE USES IT.
      `@oshun/workbench-kit`'s `capability-state` declares configured,
      not_configured, degraded, unavailable, failed and enforces SIX invariants
      over them — a `configured` whose dependencies are not configured is
      refused ("a capability is not more available than the things it needs"),
      a `configured` that skipped one of the four verifications "is not
      configured, it is unverified", a `degraded` must name what still WORKS,
      an `unavailable` carrying a cached/fallback/stale value is refused by
      name, every detail field is scanned for credentials, and the module
      exports no boolean over the union so `degraded` cannot be rounded to a
      neighbour. 12 consumers, 0 of them on an Isis surface. The Isis health
      answers are `ok`/`error` and a boolean.
      
      MEASURED, NOT FIXED. (1) .a asks for four audience projections; there are
      now two (anonymous verdict, authenticated detail) and curated, AAA and
      operator are not distinguished from each other. (2) `/health` answers
      `status: 'ok'` as a CONSTANT whatever the dependencies say, and names the
      service and its version to an anonymous caller — liveness legitimately
      means "the process responds", so narrowing it is a decision, but the
      version disclosure is recorded. (3) .b's derivation reads database,
      queue, storage and provider liveness only — no audience/host policy, no
      workflow/model/provider admission, no quota or capacity. (4) The operator
      projection is all-or-nothing; .d's field-level controls do not exist.
      (5) .f is pinned by response-body assertions only; headers and telemetry
      are unscanned.
      
      TRAPS. (1) A COMMENT STRIPPER THAT DOES NOT KNOW ABOUT STRINGS EATS CODE:
      every service mounts `app.use('/api/*', authMiddleware(...))`, and the
      `/*` inside that PATH LITERAL opened a block comment to a regex stripper,
      which ran to the next `*/` and deleted everything between. The mount
      census came back EMPTY and its conclusion was right for the wrong reason
      — the only reason it was caught. The stripper now walks string literals,
      and the same correction was applied to the I6.11 and I6.12 instruments,
      whose numbers are unchanged by it. (2) A PHRASE THAT CROSSES A STRING
      CONCATENATION IS NOT IN THE SOURCE: two of the six kit invariants were
      matched on phrases written across a `+` join and reported as absent.
      (3) The trailing noun a slash list qualifies is not a member —
      `not-configured/unauthorized semantics` splits to "unauthorized
      semantics". (4) `apps/isis/workflow-registry`'s vitest suite OOMs at 8 GB
      on this box, at HEAD as well; its readiness fix is driven by the
      instrument's own probe instead.
      
      4 of 12 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`. .a
      needs four projections and the state vocabulary the estate already built;
      .b needs policy, admission and quota in the derivation; .c has no
      lower-tier projection at all; .d has the operator boundary and needs
      field-level controls inside it; .e needs freshness/stale/partial
      aggregation rules and a liveness answer that is not a constant; .f needs
      the scan extended past the response body. -->
      
  • I6.14 Add consumer-driven contracts for all three hosts. Evidence: incompatible API change fails CI.

    • I6.14.a Inventory Isis operator, Yemaya AAA, and Oshun curated/gallery consumers, versions, routes, events, fields, actions, and audience assumptions. Evidence: consumer inventory.

    • I6.14.b Capture provider contracts for success, empty, partial, stale, degraded, not-configured, unauthorized, validation, conflict, rate/ quota, and server failure cases. Evidence: contract fixtures.

    • I6.14.c Capture property absence/redaction, enums/discriminators, units, pagination/cursors, idempotency, versions, and error codes rather than example-only shapes. Evidence: schema assertions.

    • I6.14.d Verify each consumer contract against the real provider/service implementation and ratified source-of-truth registries in CI. Evidence: provider verification.

    • I6.14.e Verify deployed/provider version compatibility before consumer release and publish contract artifacts with ownership. Evidence: deployment gate.

    • I6.14.f Fail breaking removal/type/semantic/auth/audience changes while allowing explicitly compatible additive evolution. Evidence: deliberate breaking-change test. <!-- MEASURED 2026-08-17 (I6.14.a–f). The parent's bar is "incompatible API change fails CI", so the job is to find every check standing between an Isis provider change and the three hosts that consume it, then break the provider and watch what the checks say. Instrument: generate-isis-consumer-driven-contracts.mts (+14 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-consumer-driven-contracts.json.

      TWO ISIS API DESCRIPTIONS, AND THE BREAKING-CHANGE GATE WATCHED THE ONE NOBODY
      SERVES. The three services each generate a spec from their own route sources:
      104 paths, all `/api/v1/...`, version 0.1.0 — and 126 of their 132 operations
      answer with `JsonObject`, `additionalProperties: true`, so the payload every
      consumer reads is described as "an object". The hand-authored
      `libs/openapi/src/specs/isis/isis-api.yaml` describes 17 `/v1/...` paths at one
      origin (`localhost:3003`, which is the workflow registry's port) with 21 named
      response schemas, version 1.0.0, and an enum saying `processing` where the
      implementation says `running`. ZERO paths in common. It feeds
      `libs/openapi/src/generated/isis.ts`, which nothing imports. The CI step "Check
      OpenAPI breaking changes" (`openapi:diff:check`) read only that family.
      
      DRIVEN AT HEAD — the deliberate breaking-change test .f asks for, performed.
      (1) Renamed `jobs.get('/:id/status')` to `/:id/state` in the generation-api
      routes and regenerated the spec with the estate's own generator: the poll route
      of both consuming hosts, gone. `openapi:diff:check` reported "All specs pass";
      route parity passed; cross-service passed. One check caught it — schema
      compatibility, because the SDK still calls the literal. (2) Removed
      `GET /api/v1/jobs` the same way — the route `apps/isis/web/src/pages/JobsPage.tsx:325`
      lists jobs with. EVERY gate passed: 4 gate runs, 9 tests green. Its SDK call is
      `this.transport.get<PaginatedResponse<Job>>(…)` and the check's `(?:<[^>]+>)?`
      cannot cross the inner `>`, so it is invisible. (3) On the spec the gate DOES
      read: a required response property removed, a type changed integer→string, an
      enum narrowed 6→4 and an operation's `security` removed — "✅ No breaking
      changes", 0 detected. `schema-breaking` was declared in the change union and
      never constructed, under a docstring promising "Removed properties from required
      response schemas".
      
      FIXED. (1) The gate watches the three service specs beside the domain specs —
      10, was 7. (2) `schema-breaking` and `auth-changed` are implemented: response
      property removal, type change, enum narrowing, required→optional; request
      optional→required and enum narrowing; security added, removed or changed —
      resolved through `$ref` chains, so a change to `Job.status` is reported at every
      operation that returns a `Job`. Additive evolution (new response property,
      widened response enum, new path or operation, new optional request property) is
      reported and passes. Re-driven: mutation (2) now fails with "GET operation
      removed from /api/v1/jobs", and (3) with 16 breaking changes naming each JSON
      path. (3) `libs/openapi/scripts/diff.spec.ts` — 20 cases, one mutation per class
      plus the additive cases that must NOT fail; 9 of them go red when the new
      detection is disabled. Wired into CI's api-contracts job BEFORE the gate, so a
      hollowed-out detector fails there rather than approving a diff. (4) The AAA
      store's remote-status map is now `Record<RemoteJobStatus, IsisJobStatus>` keyed
      on the union `@isis/client` publishes: widening that union with `'timeout'` now
      fails `apps/yemaya/studio-web`'s typecheck (TS2741). It was a `switch` over
      `string` with `default: return 'pending'`, and `startPollingJob` re-arms for
      every non-terminal status — an unrecognised TERMINAL status polled the provider
      forever and displayed as pending. (5) The same store wrote
      `stage: mapRemoteProgressStage(undefined)` — the mapper's only argument, ever,
      was the literal `undefined`, so the stage was the constant `'generating'` — and
      `stageProgress: snapshot.progress % 25`. `JobStatusSnapshot` has no stage field;
      both are gone, with 10 consumer-contract tests of which 4 fail against the old
      behaviour.
      
      MEASURED, NOT FIXED. (1) .a's third host has no provider seam:
      `apps/oshun/web/src` + `apps/oshun/bff/src` is 4,640 files, 0 importing
      `@isis/client`, 0 provider origins configured — the gallery is seeded from
      `@isis/output-gallery` and a durable snapshot store, so there is nothing to
      drive a consumer contract against. The two that do consume are `apps/isis/web`
      (96 files, 117 SDK methods) and `apps/yemaya/studio-web/src/stores/isisStore.ts`
      (1 file, 4 methods). (2) 0 of the 5 CI contract specs name any consumer host;
      they are provider-authored. (3) .b: 2 of 11 cases pinned — success and stale,
      both by `apps/isis/output-registry/src/openapi.contract.spec.ts`. (4) .c: 6 of 7
      property classes asserted, five of them by that same output-registry spec,
      idempotency by none — and the payloads themselves are open objects, including
      `GET /api/v1/jobs/{id}/status`. (5) .d: 0 of 5 boot a provider app; every check
      reads a YAML document or regexes a source file. (6) .d: 21 of 112 SDK transport
      calls are invisible to the one consumer-facing check, and `models.ts` is not
      among the three resource files it reads. (7) .e: no consumer pins a provider
      version.
      
      TRAPS. (1) A NESTED GENERIC HIDES THE ROUTE: `[^>]+` stops at the first `>`, so
      every paginated SDK call is unprotected. (2) A WORD IS NOT AN ASSERTION —
      `Partial<IsisJob>` reported .b's "partial" case as pinned, and `missingFromSpec`
      reported absence/redaction as pinned by five specs; each recogniser now names
      the status code or literal state value a contract must mention, and an
      unrecognised subject throws instead of counting as unpinned. (3) EVERY ISIS
      CONTRACT SPEC CONTAINS THE WORD ISIS: a host marker built from the host slug
      reported the operator as named by artifacts carrying no consumer expectation.
      (4) `someRecord.outputs.map(…)` is not an SDK call — the receiver has to be a
      client. (5) `apps/oshun/web/.next/standalone` is TRACKED and holds committed
      copies of the app's own sources; a census under `apps/oshun/web/src` that
      forgets it double-counts every file. (6) A comment that quotes the code it
      removed matches a substring test for that code — the fix pins assert on
      assignment forms instead. (7) `apps/isis/workflow-registry`'s vitest OOMs on a
      single spec file on this box, at HEAD as well (same as I6.13); its route-parity
      spec was not run locally and nothing here touches that app.
      
      3 of 13 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`. .a needs a
      consumer inventory naming each host's routes and fields, and one of the three
      hosts has no seam to inventory; .b has 2 of 11 cases; .c cannot bind property
      assertions to 126 operations that describe no properties; .d needs a check that
      runs the provider and a comparison that can see every SDK route; .e needs a
      version pin and a published artifact that describes deployed routes; .f's
      detector is now real and proven, but the audience clause has no expression in
      these documents and the payload classes have almost nothing to bind to on the
      surface the consumers actually call. -->
      
  • I6.15 Add two-tenant integration coverage across run, blob, search, lineage, export, and recall. Evidence: no cross-tenant result or count leakage.

    • I6.15.a Provision two tenants with colliding human names, projects, filenames, tags, timestamps, model/workflow use, and distinct actors/ roles/data. Evidence: isolation fixture.

    • I6.15.b Exercise plan/dispatch/job/progress/cancel/retry/replay/output/ variant reads and direct foreign IDs/cursors/idempotency keys. Evidence: run isolation tests.

    • I6.15.c Exercise blob upload/download/signed URL/hash/cache/derivative/ temporary object and guessed-path access. Evidence: storage isolation tests.

    • I6.15.d Exercise search/filter/sort/autocomplete/facets/counts/ suggestions/index lag and foreign identifiers. Evidence: search isolation tests.

    • I6.15.e Exercise forward/reverse lineage, provenance ingredients, blast-radius, comparison, gate, audit, and aggregate counts. Evidence: graph isolation tests.

    • I6.15.f Exercise JSON/CSV/package/portable-spec/evidence/bundle exports, jobs, download URLs, filenames, metadata, and hidden sheets/entries. Evidence: export isolation tests.

    • I6.15.g Exercise quarantine/recall/unpublish/notification propagation and prove one tenant’s action cannot mutate or disclose the other. Evidence: recall isolation tests.

    • I6.15.h Inspect API/UI/error/timing/log/trace/event/cache/CDN/queue channels for result, identifier, existence, and count leakage. Evidence: cross-channel isolation report. <!-- MEASURED 2026-08-17 (I6.15.a–h). The parent's bar is "no cross-tenant result or count leakage", so the first question is not what the tests cover but whether the thing they would test exists: a read has to be able to name whose rows it is for before anyone can check that it named them right. Instrument: generate-isis-tenant-isolation.mts (+13 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-tenant-isolation.json.

      TWO OF THE THREE SERVICES CANNOT STATE A TENANT, SO THEIR READS CANNOT WITHHOLD
      ONE. The `jobs` table has no owner, tenant, organization, workspace or project
      column — the tenant a submission carries is written into `metadata JSONB` by
      `withSubmissionMetadata` and never read back — and `JobListOptions`, the only
      filter the list read accepts, offers status, type, workflowId, page, limit,
      sortBy and sortOrder. `workflows` has `owner_id` and `visibility`, and
      `WorkflowListQuery` offers neither (its `author` filter is free text a caller
      chooses). The output registry is the exception: five isolation columns, five
      indexes over them, and an `OutputAccessScope` that reaches the SQL.
      
      SO THE READ SURFACE IS UNSCOPED, NOT WEAKLY SCOPED. Of 48 GET routes across the
      three services, 10 constrain by a subject — all of them in the output registry's
      `outputs.ts` — 13 have a role gate and no subject, and 25 have neither. Named,
      because a count is not a finding: every job read
      (`GET /`, `/:id`, `/:id/status`, `/:id/output`, `/:id/audio`,
      `/:id/queue-position`, `/stats`), so the route both consuming hosts poll on every
      job and the queue counts beside it take an id from the path and hand back
      `jobService.get(id)`; every model read including `/:id/download-url`, which is
      .c's signed URL; and the workflow reads including `/:id/export`, which is .f's.
      
      FIXED, AND DRIVEN THROUGH THE REAL APP. A SCOPE THE CALLER SUPPLIES IS NOT A
      SCOPE. The output registry's `resolveTenantScope` read `X-Isis-Tenant-Id`,
      `-Organization-Id`, `-Workspace-Id` and `-Project-Id` BEFORE the credential's own
      claims and compared them to nothing, so a request header chose which tenant
      `appendAccessScopeConditions` filtered to on the list, the cursor page, the
      projected page, `/stats`, lineage, ancestors and descendants — while
      `authMiddleware` in all three services applies the opposite rule to project
      scope, refusing a header that contradicts the claim ("Project scope mismatch")
      and letting the claim win. The claim wins now, a contradicting header is a 403
      before any store call, and a header is honoured only where the credential is
      silent, where it can narrow an owner-anchored query and never widen one. Driven
      at three settings through the booted app: contradicted → 403, agreeing → past the
      rule, absent → past the rule. Blast radius before the fix was bounded, and the
      report says so: every one of those reads is also anchored to `resolveOwnerId`,
      which 401s without an owner claim, so the header mislabelled a query rather than
      opening another tenant's rows. The estate had already decided this on the
      provenance WRITE path —
      `output.service.provenance-scope.spec.ts` asserts a ledger event is attributed to
      `tenant_of_record` and `not.toContain('tenant_supplied_by_caller')`.
      
      DRIVEN, NOT FIXED. With an injected recording database, `getStats({ownerId,
      tenantId})` emits `WHERE owner_id = $1 AND tenant_id = $2`; `getStats()` emits
      `SELECT COUNT(*) as count, COALESCE(SUM(size), 0) as size FROM outputs` — no WHERE
      clause at all. `appendAccessScopeConditions` adds one condition per dimension
      PRESENT and nothing for a dimension absent, so an unstated scope is an unlimited
      one. Fixing that is a fail-closed decision with a wide blast radius, not a
      cleanup.
      
      TRAPS. (1) A CHECK CONSTRAINT CONTAINS PARENTHESES: `type TEXT NOT NULL CHECK
      (type IN (…))` closes a paren before the table does, and a non-balanced scan of
      `CREATE TABLE jobs (…)` reports two columns. (2) A COLUMN ADDED BY AN ALTER IS
      STILL A COLUMN — `outputs` is created without `tenant_id` and gains it, with
      organization, workspace and project, through `ALTER TABLE … ADD COLUMN IF NOT
      EXISTS`; a census reading only the CREATE reports the one service that HAS tenant
      columns as having none. (3) A NECESSARY CONDITION CAN EXCLUDE AND CANNOT CONFIRM:
      19 Isis specs name two distinct `tenant_*` identifiers and almost none is an
      isolation test — they are budget-scenario labels (`tenant-quota-warn`), two halves
      of one signing key (`tenant-alpha-signing-key`,
      `tenant-alpha-signing-secret-12345`) and governance fixtures that touch no read
      path; the survivors were read and the three real artifacts recorded with the
      assertion that pins each. (4) `GET /` and `GET /:id` occur in more than one
      router, so a route label without its file reports duplicates. (5) An `interface X
      extends Y` hides half the filter: `OutputAccessScope`'s four tenant dimensions are
      inherited from `TenantScope`.
      
      1 of 7 properties present — .h's scope-source clause, which the fix closed. ALL
      EIGHT CELLS AND THE PARENT LEFT `[ ]`, and the reason is not test coverage: .a
      cannot provision two tenants against `jobs` or `workflows` because neither can
      record one; .b's run reads accept no subject to compare; .c/.d/.e/.f/.g would be
      testing an isolation boundary that exists in one of three services; .h's ten
      channels have one report between them, the operator app's
      `buildOutputIsolationAuditReport`, which audits API response bodies it was handed
      (reached by `OutputsPage` and `OutputsParityPage`) and no other channel. Giving
      Isis a tenant boundary is a schema-and-product decision — a column, a migration, a
      backfill of `metadata->>'tenantId'`, a filter on every read — and it is not a
      session's cleanup. -->
      

I7 — Operator, AAA, and curated user experience#

  • I7.1 Refactor the Isis operator dashboard around queue health, provider health, release gates, spend, incidents, and actionable exceptions. Evidence: live-data Playwright journey.

    • I7.1.a Define the operator’s primary triage questions, canonical live projections, refresh/staleness semantics, permissions, and action ownership for every dashboard datum. Evidence: dashboard view contract.

    • I7.1.b Compose a restrained operations workspace with queue and incident attention first, supporting health/spend/gate context second, and no decorative hero or repeated card grid. Evidence: approved responsive layout spec.

    • I7.1.c Present queue age/depth/capacity/fairness, provider/circuit/ maintenance state, gate blockers/freshness, spend/variance, and incident severity with explicit units and timestamps. Evidence: projection tests.

    • I7.1.d Build actionable exception lists grouped by owner/severity/ deadline with stable filters, URLs, drill-down context, and bounded live updates. Evidence: exception interaction tests.

    • I7.1.e Gate pause/drain/retry/quarantine/budget/incident actions by exact object state, permission, confirmation, reason, and optimistic concurrency. Evidence: action authorization tests.

    • I7.1.f Render loading/empty/stale/degraded/partial/not-configured/error/ disconnected states without implying green health. Evidence: state matrix tests.

    • I7.1.g Verify live incident-to-detail-to-remediation-to-recovery journey, keyboard/screen reader/zoom/responsive behavior, and high-volume performance in Playwright. Evidence: dashboard E2E. <!-- MEASURED 2026-08-17 (I7.1.a–g). The parent asks the dashboard to be organised around six subjects — queue health, provider health, release gates, spend, incidents, actionable exceptions — so the measurement is which of the six it can show, which the API can feed it, and what the one datum it did show actually meant. Instrument: generate-isis-operator-dashboard.mts (+10 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-operator-dashboard.json.

      LIVENESS IS NOT READINESS. `DashboardPage.tsx` was 118 lines rendering three cards
      and three stage lanes from `client.checkHealth()`, which reads `/health` — the
      route I6.13 already recorded as answering `status: 'ok'` as a CONSTANT whatever the
      dependencies say. DRIVEN on the booted generation-api with no database configured:
      `/health` answers **200 status=ok** and `/ready` answers **503 ready=false** in the
      same second, naming all five roles it probed — jobsDatabase, modelsDatabase,
      provider, queue, storage — every one of them down. The page called that service
      `Healthy`, put `3/3` beside it and set its stage lane to the green `complete`.
      
      FIXED. `checkReadiness()` in `@isis/client` reads `/ready` per service and keeps the
      verdict when the transport throws on the 503 the route answers with, so the
      dependency map and per-check detail in that body reach the caller instead of
      becoming a bare "unreachable". The dashboard renders the readiness verdict as the
      answer, names the failing dependencies, keeps liveness as the separate smaller
      claim it is ("Processes responding"), says "Not measured" instead of `0/3` for a
      count it does not have, and derives the green stage state from a named function
      that returns it only for a `ready` verdict. 6 dashboard tests and 7 SDK tests; the
      count assertion and the latency assertion both go red against the old code.
      
      A LATENCY THAT IS NOT TIMED IS NOT A LATENCY. `checkServiceHealth` returned
      `latencyMs: 0` under the comment "Could be calculated if needed", and
      `apps/isis/cli/src/commands/health.ts:339` printed it to an operator as
      `Latency: 0ms` for every service on every run. It is timed now, on the failure path
      too.
      
      THE THREE SERVICES DO NOT ANSWER READINESS THE SAME WAY, and finding out cost a
      crash: only generation-api sends `criticalDependencies`, and it sends a
      role→verdict MAP, not the list of role names the first version of this instrument
      (and of the SDK type) assumed — `.join` threw. output-registry and
      workflow-registry send `ready` plus `checks` alone, and workflow-registry spells a
      failed check `failed` where the other two spell it `error`. The SDK now normalises
      both shapes into one `failingDependencies` list and reads both spellings. Also
      recorded: I6.13's comment said the public part of `/ready` names the dependency
      ROLES; it also carries their per-role booleans, which is more than roles and still
      names no resource.
      
      MEASURED, NOT FIXED. (1) 0 of the 6 subjects reach the page, and only 2 have a
      producer anywhere in the three services: queue health (`/api/v1/queue/stats`,
      `/jobs/stats`, `/jobs/admin/queue/health`, exposed as `getQueueStats` and
      `getQueueHealth`) and provider health (`/jobs/providers/operations`,
      `/providers/compatibility`). Release gates, spend, incidents and actionable
      exceptions have NO route at all — the dashboard could not show them however it were
      composed. (2) .e's six action verbs: the page offers none; of the six, two have a
      route (`POST /jobs/admin/queue/drain`, `POST /jobs/:id/retry`). (3) .f's eight
      states: 3 rendered (loading, empty, error); stale, degraded, partial,
      not-configured and disconnected have no representation. (4) .g: 0 of 13 operator
      Playwright specs mention the dashboard. (5) .a has no view contract and .b no
      approved layout spec.
      
      A NUMBER DERIVED FROM A HASH OF THE PROMPT IS NOT A MEASUREMENT — 21 unmeasured
      measurement-shaped fields survive on two operator pages, and both are outside this
      cell's subject but inside the quality bar:
      `apps/isis/web/src/pages/InteractiveGenerationPage.tsx:103` computes
      `latencyMs: 430 + (seed % 390)`, `consistencyScore: 86 + (seed % 11)` and
      `topologyScore: 82 + ((seed >>> 6) % 14)` from `hashText(prompt|style)` and renders
      them as a preview's latency and quality scores; and
      `apps/isis/web/src/pages/MeshTransformerAbComparisonPage.tsx` carries baked-in
      literal `topologyScore`, `manifoldScore`, `watertightScore`, `symmetryScore`,
      `faceCount`, `vertexCount`, `latencySeconds`, `peakMemoryGb` and
      `throughputMeshesPerHour` for three named 3D models and presents them as an A/B
      comparison. Both need a decision about what the page should show when no
      measurement exists; neither is a cleanup.
      
      TRAPS. (1) A VERB IN PROSE IS NOT A CONTROL: matching .e's action names against the
      whole page reported `retry` as offered, because the error state advises "Retry the
      readiness probe after network stabilization" — the census now reads button labels,
      `aria-label`s and `onClick` targets only. (2) AN ARROW FUNCTION IN AN ATTRIBUTE
      CONTAINS A `>`: `<button onClick={() => refresh()}>` ends a naive `<button[^]*?>`
      match at the arrow, and the label comes back as `refresh()}>Drain queue`; the tag
      ends at the first `>` outside braces. (3) `apps/isis/web` has a vite AND a vitest
      alias pointing `@isis/client` at the library source but no matching tsconfig path,
      so `tsc` resolves it through node_modules to `dist/index.d.ts` — a new SDK method
      is "not on type IsisClient" until the library is rebuilt (`npx tsup` from the
      library dir). Adding the tsconfig mapping instead pulls the SDK's Node-typed source
      into a browser app's check and produces 20 `Cannot find name 'process'` errors, so
      the mapping is NOT the fix.
      
      2 of 8 properties present — .c's health-datum provenance and .f's no-green-by-
      default, both closed by the fix. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`: four of
      the parent's six subjects have no producer to build against, .e's actions are
      mostly routes that do not exist, .f needs five more states, and .g needs a
      Playwright journey against a live stack. -->
      
  • I7.2 Complete the operator workflow-template registry and lifecycle UI. Evidence: draft through archive with review and audit.

    • I7.2.a Build searchable/filterable registry list and detail projections for identity/version, lifecycle, asset/operation/audience, owner, dependencies, evaluations, environments, and usage. Evidence: registry UI tests.

    • I7.2.b Build typed create/edit/version authoring for metadata, graph/ subgraph refs, controls, compatibility, rights, resource estimates, and target profiles without arbitrary JSON. Evidence: authoring tests.

    • I7.2.c Show canonical graph validation, dependency locks, security/ license scans, dry runs, benchmarks, diffs, and missing evidence before review. Evidence: validation view tests.

    • I7.2.d Implement submit/request-changes/approve/reject with reviewer role, comments, exact version/evidence binding, and immutable decisions. Evidence: review journey.

    • I7.2.e Implement promote/deprecate/revoke/quarantine/archive/restore actions with blast-radius preview, environment rules, confirmation, and audit. Evidence: lifecycle journey.

    • I7.2.f Present version history, current/admitted selections, consumers, runs, decisions, audit, and rollback options without rewriting history. Evidence: history tests.

    • I7.2.g Verify concurrency, stale evidence, direct URL/field/function authorization, accessibility, and large-registry performance. Evidence: adversarial Playwright suite. <!-- MEASURED 2026-08-18 (I7.2.a–g). The parent asks for a template registry that runs "draft through archive with review and audit", so the measurement asks what a template IS here before asking what a UI does with it. Instrument: generate-isis-template-registry.mts (+11 tests, determinism included), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-template-registry.json.

      A NOUN WITH NO LIFECYCLE COLUMN HAS NO LIFECYCLE. `workflow_templates` has 13 columns
      and none of them is a status, owner, reviewer, version or environment; the served
      `WorkflowTemplate` declares no such field; the template API is four routes —
      `GET /templates`, `GET /templates/featured`, `GET /templates/:id`,
      `POST /templates/:id/create` — and the SDK has the matching four methods. Of .a's eight
      facets, 3 are servable (identity/version, asset+operation, usage) and 5 have no field at
      all: lifecycle, owner, dependencies, evaluations, environments. Of .d's four review verbs
      and .e's six lifecycle verbs, ZERO have a template route. The estate's lifecycle
      machinery is real but one level down, on WORKFLOWS: `draft|published|deprecated|archived`
      with `POST /workflows/:id/restore`, `/versions/:version/deprecate`,
      `/versions/:version/restore`, and publish/archive as `PATCH /:id` with
      `expectedUpdatedAt`. None of it reaches templates. Meanwhile
      `libs/contracts/src/common/workflow-template.ts` (1466 lines) declares exactly the
      lifecycle the cell asks for — `draft/review/approved/published/deprecated/archived` plus
      a review state `not_required/pending/approved/rejected/changes_requested`, risk level and
      portability state — and **0 files** outside it name any of those symbols.
      
      A BAN EXPRESSED IN A VOCABULARY THE MATCHER DOES NOT SPEAK IS NOT A BAN. The one seeded
      template declares two `block`-severity graph bans, `external-code-execution` matched with
      `classTypePattern: '(?i)(exec|shell|python|subprocess)'` and `unbounded-remote-fetch`
      matched with `inputKeyPattern: '(?i)(url|uri)', requiresAllowlist: true`.
      `matchesDefinition` implemented `hasNode`, `maxSteps` and `forbiddenInputKey` — and
      nothing else. All three declared keys were unreadable, so both bans answered "no match"
      for every graph that has ever been checked. `match` is typed `Record<string, unknown>`,
      so the compiler could not see it; `validateWorkflowTemplateCoverage`, whose whole job is
      to certify that a governance block is complete, checked patternId, description and
      severity and never asked whether the ban could fire. (`(?i)` is not JS regex syntax
      either: `new RegExp('(?i)x')` throws.)
      
      FIXED. The matcher implements the declared vocabulary, translates the leading inline flag
      group, treats `requiresAllowlist` with a per-rule `allowlistedClassTypes` exception list
      (absent list = no exceptions), and reports a pattern it cannot apply as
      `DISALLOWED_PATTERN_NOT_EVALUABLE` at the pattern's own severity — fail closed, never a
      silent clean graph. Coverage rejects an unevaluable ban (`UNEVALUABLE_GRAPH_PATTERN`) and
      `summarizeWorkflowTemplateCoverage` reports the verdict per ban. Red-check: with the old
      three-key vocabulary restored, 7 tests fail, including the estate's own pre-existing
      "ships bundled templates with explicit governance coverage".
      
      AN OVERLAY THE CALLER WRITES IS NOT AN OVERLAY THE TEMPLATE APPROVED. `createFromTemplate`
      did `{...template.definition, ...request.customizations}` where `customizations` is
      `z.record(z.unknown())`, then wrote it; and `enforceTemplateGovernance` — 539 lines whose
      own doc comment says it applies the guardrails "immediately before execution" — had ZERO
      non-test callers. DRIVEN on the booted route before the fix: one request with a
      `PythonScriptExec` node running `os.system("id")`, an `HttpFetch` node pointed at
      `http://169.254.169.254/latest/meta-data/`, a `/opt/models/secret.safetensors` model
      reference, and `template_governance: {disallowedGraphPatterns: [], portabilityChecks: []}`
      returned **201 Created** with all of it stored — four `block`-severity rules violated in
      one call. FIXED: the merged graph is judged by the template's own governance before
      anything is created, the overlay may not rewrite the governance block that judges it, and
      the route answers 422 `TEMPLATE_GOVERNANCE_BLOCKED` carrying the findings. The same
      request now answers 422 naming seven findings; a clean instantiation and a benign overlay
      still answer 201; a refusal does not increment the template's usage count. Parameter
      guardrails are deliberately NOT applied at instantiation — they grade run values, which do
      not exist yet — so a definition-only entry point was added rather than misusing the
      run-time one.
      
      THE REGISTRY PAGE DROPPED THE ENTIRE GOVERNANCE BLOCK THE SERVICE COMPUTES, and the SDK
      type dropped it before that: `WorkflowTemplate` in `@isis/client` had no `governance`
      field, so the page could not have rendered it. Both carry it now, and the page shows
      template version, class, coverage verdict, each declared ban with whether the registry can
      enforce it, portability checks, guardrails, the declared version change, and the refusal
      findings when a customization is rejected. It also NAMES the facets this registry does not
      record instead of leaving blank columns to be read as "fine"; the instrument checks that
      list against the served shape.
      
      MEASURED, NOT FIXED. (1) .b: there is no create/update/version route for a template at
      all, and the only authoring affordance is the free-form JSON overlay the cell explicitly
      forbids. (2) .c: dependency locks, security/license scans, dry runs and benchmarks have no
      producer; the only "diff" is the template's own self-declared `versionDiff`. (3) .d/.e: no
      review or lifecycle verb has a route, so there is nothing for a UI to call. (4) .f: a
      template has no version history — `governance.templateVersion` is a string parsed out of
      the definition blob. (5) .g: 0 of 13 operator Playwright specs mention templates.
      (6) `enforceTemplateGovernance` — the ONLY function that applies a template's parameter
      guardrails — still has zero production call sites, so those guardrails are enforced
      nowhere.
      
      TRAPS. (1) A ROUTE MODULE CAPTURES THE SERVICE SINGLETON AT IMPORT: calling
      `resetWorkflowService()` after importing `routes/templates.js` leaves the route talking to
      one instance while the census reads another — the usage count read 0 after a successful
      create. (2) A COUNTER IS NOT DETERMINISTIC ACROSS TWO BUILDS: the drive records usage
      deltas, never absolute counts, so `build()` twice in one process serializes identically.
      (3) `vite-tsconfig-paths` in `apps/isis/workflow-registry/vitest.config.ts` OOMs this box
      on ANY spec in that project, including untouched ones, at a 6 GB heap; the suite runs in
      1.3s with an explicit alias list instead. (4) `apps/isis/web` typechecks `@isis/client`
      through `dist`, so `npx tsup` in `libs/isis/client` is required before a new SDK field
      exists for the page.
      
      3 of 12 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`: five of .a's eight
      facets have no field to serve, .b has no typed authoring path, .c's scans and dry runs have
      no producer, .d and .e have no routes to call, .f has no version history, and .g has no
      Playwright coverage. -->
      
  • I7.3 Complete the operator model/model-version intake and quarantine UI. Evidence: license, scan, evaluation, promotion, and revoke journey.

    • I7.3.a Build model/version intake for immutable source/artifact hash, architecture/format/runtime, files, dependencies, owner, license/model card, intended uses, and provider compatibility. Evidence: intake tests.

    • I7.3.b Present upload/remote retrieval progress and malware/signature/ format/safe-parser scan results with quarantined-by-default state. Evidence: scan journey.

    • I7.3.c Present license, training-data provenance, rights restrictions, model-card completeness, security risks, and exact blocking gaps. Evidence: evidence view tests.

    • I7.3.d Configure and track calibrated quality/safety/performance/ compatibility evaluations, fixtures, results, uncertainty, and failures. Evidence: evaluation journey.

    • I7.3.e Compare versions/models on authorized measurements, target fit, latency/cost/resources, rights, and risk without hiding unavailable data. Evidence: comparison tests.

    • I7.3.f Implement review/promote/deprecate/quarantine/revoke/archive with exact artifact/evidence binding, environment progression, blast-radius, reason, and confirmation. Evidence: lifecycle journey.

    • I7.3.g Show immutable provenance/decisions/audit/consumers/runs and verify stale-version, supply-chain, mass-assignment, accessibility, and performance cases. Evidence: model registry E2E. <!-- MEASURED 2026-08-18 (I7.3.a–g). The parent's evidence is a "license, scan, evaluation, promotion, and revoke journey", so the measurement asks which of those five nouns the registry can hold before asking what the console does with them. Instrument: generate-isis-model-registry.mts (+9 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-model-registry.json.

      A VERDICT FROM AN INSTRUMENT THAT CANNOT SEE THE THING IS NOT A VERDICT. Model upload
      scanning ships ENABLED, FAIL-CLOSED, on engine `builtin`. `BuiltinSecurityScanner` emits
      no `malware`-typed threat anywhere — the only `type: 'malware'` in the security library is
      inside `ClamAVScanner` — and the upload gate blocks on
      `threats.filter(t => t.type === 'malware')`, so on the default engine that predicate could
      never be true. Every upload recorded `malwareScan.status: 'clean'`. DRIVEN on the real
      scanner: a safetensors-shaped buffer (8-byte header length, JSON header, 256 KB of chained
      sha256 bytes standing in for tensors) comes back `clean: false, severity: high` with 21
      threats — 10 `sql_injection`, 10 `command_injection`, 1 `suspicious_pattern`, every one a
      false positive on binary content — while a pickle that calls `posix.system` comes back
      `clean: true` with ZERO threats. The engine is not a weak malware scanner; it is a
      source-text pattern matcher pointed at a binary, which is also why the original
      `type === 'malware'` narrowing was there: reading its verdict would refuse every real
      upload.
      
      FIXED. The gate now asks whether the configured engine can produce the verdict it is
      about to record. `MALWARE_CAPABLE_SCAN_ENGINES = {clamav}`; with `builtin` and
      fail-closed (the shipped default) the upload is REFUSED naming the engine and the two
      environment variables that resolve it, and with fail-closed off it completes recording
      `status: 'unsupported'` — never `clean`. A capable engine still blocks on a malware
      verdict. Driven: 5/5 cases hold. The two pre-existing specs that configured
      `ISIS_MODEL_MALWARE_SCAN_ENGINE: 'builtin'` while injecting a scanner returning
      ClamAV-shaped malware threats now name `clamav`: they were passing against a double that
      contradicted the engine they configured.
      
      A SCAN VERDICT LIVES IN EXACTLY ONE PLACE AND IT IS NOT THE MODEL. `model_registry` has 27
      columns and not one is a scan status, a quarantine, a license, a model card, an
      evaluation or an environment; the only dispositions are `is_active` and `is_verified`
      (and I1.2.c already recorded that `is_verified` has no writer). The verdict exists only
      inside the `model.upload.completed` audit entry's `metadata.malwareScan`. The console now
      reads it there — `ModelsPage` queries
      `getAuditLogs({modelId, action: 'model.upload.completed'})` and renders each verdict
      through `describeModelScanVerdict`, which names `skipped`, `unsupported` and `failed` as
      NON-verdicts rather than green — states plainly that the record itself carries no scan or
      quarantine state, and lists the intake facts this registry does not record.
      
      THE RICH TABLES EXIST AND NOTHING READS THEM. `libs/isis/database/prisma/schema.prisma`
      declares `model_cards` (32 fields: licenses, hashes, overview, usage, training,
      evaluation, limitations, ethics, safety, governance, citations, changelog) and
      `model_versions` (28 fields: source, lineage, files, licenses, hashes, runtime, telemetry,
      safety, governance, deployments) — exactly .a's files/license/model-card/intended-uses and
      .c's license/provenance/rights/model-card subjects. Product readers: **0 and 0**. No code
      outside the generated Prisma client calls `.modelCard.*` or `.modelVersion.*`.
      
      MEASURED, NOT FIXED. (1) .a: 5 of 8 intake subjects are servable; files, license/model
      card and intended uses have no field on the served record and live only in the unread
      tables. (2) .b: no quarantine state exists — there is no column, no route and no default;
      upload progress, signature, format and safe-parser scans have no producer at all. (3) .c:
      1 of 6 evidence subjects is servable (`trainedOn`, a free-text string). (4) .d: 0 of 3
      evaluation subjects servable; no evaluation route. (5) .e: no comparison route among the
      10 model routes. (6) .f: 0 of 6 lifecycle verbs — review, promote, deprecate, quarantine,
      revoke, archive — has a route or a control; the console offers Register, Save, Generate
      Upload URL, Complete Upload, Generate Download URL, Delete. (7) .g: 0 of 13 operator
      Playwright specs navigate to `/models`.
      
      TRAPS. (1) A WORD IS NOT A ROUTE: matching `/model/i` over the e2e specs reported 11 of 13
      as coverage, because the asset-generation specs say "model" in prose; only a `goto('/models`
      is coverage of this registry, and there are none. (2) A TEST DOUBLE THAT CONTRADICTS ITS
      OWN CONFIGURATION HIDES THE DEFECT — the malware-blocking specs configured the engine that
      cannot detect malware and injected one that can. (3) The drive must not use
      `randomBytes`: a chained-sha256 stream gives the same 21 threats every run, so the census
      is byte-stable.
      
      1 of 12 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`: the substrate has no
      column for quarantine, license, model card, evaluation or environment; the lifecycle verbs
      have no routes; and the tables that do carry those facts have no reader. -->
      
  • I7.4 Complete the provider endpoint and failover UI. Evidence: health, circuit state, maintenance, and routing reasons are live.

    • I7.4.a Build endpoint/pool/provider topology views with class, region, environment, capability versions, model/node images, GPU/resources, capacity, lifecycle, owner, and last update. Evidence: topology tests.

    • I7.4.b Present probe/queue/worker/provider health, freshness, partial/ unknown/not-configured states, incidents, and SLO history separately. Evidence: health state tests.

    • I7.4.c Present per-family circuit closed/open/half-open state, counted failures, thresholds, cooldown/probes, and recent decisions. Evidence: circuit view tests.

    • I7.4.d Explain routing candidates, filters, selected endpoint, policy, capability/residency/cost/latency/capacity reasons, and audience-safe redactions. Evidence: routing inspector tests.

    • I7.4.e Implement authorized drain/restore/disable/probe/canary/failover actions with in-flight preview, deadlines, reason, expected version, and audit. Evidence: maintenance journey.

    • I7.4.f Stream live changes with cursor/reconnect/dedup/stale handling and verify failover, race, direct-action authorization, accessibility, and large-fleet performance. Evidence: controlled failover Playwright test. <!-- MEASURED 2026-08-18 (I7.4.a–f). The parent's evidence is "health, circuit state, maintenance, and routing reasons are live", so the measurement asks for each of the four whether a producer exists before asking what the console shows. Instrument: generate-isis-provider-failover.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-provider-failover.json.

      AN ORDERING CONVENTION IS A DEFECT WAITING FOR A CALLER. `evaluateEndpointHealth`
      documented "Newest outcomes first" and implemented `.slice(-window.samples)` — the last N
      of an OLDEST-first array. DRIVEN: four recent failures and four older successes decide
      `unhealthy` when the caller hands them over newest-first and `healthy` when it hands the
      same eight over oldest-first. The one in-repo caller appends to the tail, so it was right
      by accident; anybody obeying the documentation was judging an endpoint on its stalest
      probes. FIXED: the window is the newest N by the outcomes' own `observedAtUnixSeconds`,
      with ties keeping the later array position, so the array order cannot change the verdict.
      Red-check: restoring `.slice(-window.samples)` fails the equivalence test.
      
      AN ENDPOINT MODEL WITH NO ROUTE IS NOT A TOPOLOGY. `libs/isis/ai-providers/src/
      provider-endpoints` models what .a asks for — provider, region, capabilities,
      modelClasses, cost, concurrency, retry, secretRotation, health, failoverPriority,
      failoverCompanions, lastHealthCheckAtUnixSeconds — and exports twelve entry points.
      Resolved BY IMPORT, ONE has a caller outside its own module: `planIsisRoute` in
      `libs/isis/study-adapter`, a package with ZERO importers. Of .a's ten topology subjects,
      2 are served (class, last update) and 5 more are modelled by `IsisProviderEndpoint` and
      reach no route; environment, GPU/resources and owner are modelled nowhere. The console's
      entire provider surface is one `GET /jobs/providers/operations` snapshot, polled every
      5s, with one control: Refresh snapshot.
      
      MEASURED, NOT FIXED. (1) .c: the only circuit state the API exposes is
      `RunPodCircuitBreakerState` — one provider, not "per family" — carrying state, failures,
      successes and nextAttempt; thresholds and recent decisions have no field. (2) .d: no
      routing inspector; `planIsisRoute` computes candidates, filters and a selected endpoint
      and no route serves it. (3) .e: 0 of 6 maintenance verbs — drain, restore, disable,
      probe, canary, failover — has a route or a control; the three provider-shaped routes are
      all GETs. (4) .f: the page polls, it does not stream, and 0 of 13 operator Playwright
      specs navigate to `/providers`.
      
      TRAPS. (1) A NAME IS NOT AN IMPORT: counting callers by function name reported three for
      `decideRetry` and one for `evaluateSecretRotation`, all in `@oshun/workbench-kit`,
      `@euterpe/elevenlabs`, `@euterpe/api` and `@neith/security`, which export their own
      functions of those names and never touch this module. Callers are resolved by import
      specifier. (2) A DOC COMMENT THAT QUOTES THE OLD CODE STILL CONTAINS IT — an instrument
      test asserting the absence of `.slice(-window.samples)` fails on the prose that explains
      why it was removed; assert on the code shape, not the substring.
      
      1 of 9 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: eight of .a's ten
      topology subjects reach no route, .c has thresholds and decisions nowhere, .d has no
      inspector, .e has no action routes at all, and .f neither streams nor has a failover
      Playwright test. -->
      
  • I7.5 Complete the job/run explorer with stage logs, resource use, cost, output, retry, cancel, and replay. Evidence: injected failure journey.

    • I7.5.a Build deterministic search/filter/sort/pagination for request, plan, job, run/attempt, workflow/model/provider, actor/project, status, asset kind, cost, date, and incident. Evidence: explorer list tests.

    • I7.5.b Build run detail timeline/DAG with queue, reservation, stage, checkpoint, attempt, provider/tool, output, gate, decision, and terminal state identities. Evidence: journey projection tests.

    • I7.5.c Present structured redacted stage events/logs/errors with levels, timestamps, correlations, retries, downloadable evidence, and bounded streaming. Evidence: log view/security tests.

    • I7.5.d Present estimated/reserved/actual CPU/GPU/VRAM/storage/egress/time/ cost by stage and run, including pending/unbilled/variance. Evidence: resource/cost view tests.

    • I7.5.e Present output previews/manifests/hashes/provenance/gates/ quarantine/partial/missing state appropriate to each asset kind. Evidence: output projection tests.

    • I7.5.f Implement cancel/retry/replay with state/role eligibility, impact/cost preview, reason/confirmation, concurrency, and immutable new attempt/run lineage. Evidence: command tests.

    • I7.5.g Verify injected provider refusal/timeout/worker loss/malformed result/late output/recovery plus cross-tenant, accessibility, and long- history performance behavior. Evidence: failure Playwright journey. <!-- MEASURED 2026-08-18 (I7.5.a–g). The parent's evidence is an injected failure journey over stage logs, resource use, cost, output, retry, cancel and replay, so the measurement asks which of those the served job has a field for. Instrument: generate-isis-job-explorer.mts (+7 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-job-explorer.json.

      A BUNDLE BUILT TO LEAVE THE CONSOLE IS THE ONE THAT MUST BE REDACTED. The explorer's
      "Support Bundle" and "Report Issue" buttons serialized the job VERBATIM into a block for
      an operator to copy into a ticket: `input` with the prompt and negativePrompt, the
      presigned `inputUrl`, the `callbackUrl`, `metadata` and the error string. The audit view
      one route away states "PII and sensitive prompt fields are automatically redacted in this
      view" and calls `redactSensitiveData` to do it — so the estate had both the policy and
      the function, and applied them to the surface that stays on screen and not to the artifact
      built to be exported. FIXED: the builder moved to `jobs-utils.ts` and redacts prompts,
      nested secrets, personal data and URL credentials while keeping ids, statuses, progress
      and timestamps. DRIVEN: a job whose prompt contains an email and a card number, whose
      callbackUrl carries `token=sk-live-…`, whose parameters carry an inline `apiKey` and whose
      inputUrl is presigned now yields a bundle containing none of the eight secrets and all
      four identifiers.
      
      A PRESIGNED URL CARRIES ITS CREDENTIAL UNDER A VENDOR PREFIX. `redactSensitiveExportUrl`
      compared raw query keys against a fixed list, so `X-Amz-Signature`, `X-Amz-Credential`,
      `X-Amz-Security-Token` and `X-Goog-Signature` — every credential MinIO and S3 actually
      mint — passed through untouched, in the output export centre's NINE call sites as well as
      here. Keys are now normalized (vendor prefix stripped, separators removed) before an exact
      comparison, so those match and `keyframe`/`monkey` still do not. Red-check: 4 tests fail
      against the old comparison.
      
      A JOB IS A REQUEST AND A STATUS, NOT A RUN. The served `Job` is id, type, status,
      priority, progress, input, output, error, callbackUrl, metadata, workerId and four
      timestamps. Of .b's ten run-detail identities 4 are served; reservation, stage,
      checkpoint, attempt, gate and decision have no field anywhere. `JobListOptions` is
      `status`, `type`, `sortBy`, `sortOrder` — so 3 of .a's eleven filter subjects are
      supported and request, plan, run/attempt, actor/project, cost and incident are not.
      0 of .d's five cost and resource subjects are served and the page does not mention cost.
      Of .f's three commands, cancel and retry have routes and REPLAY has none; retry lineage
      rides in the client-supplied metadata key `retriedFromJobId`, not a column, so nothing
      joins on it. 20 job routes exist and none serves events, logs or stages. 0 of 13 operator
      Playwright specs navigate to `/jobs`.
      
      NOT A DEFECT, CHECKED: the console's retry eligibility (`failed | cancelled`) is exactly
      the server's, and `retryJobWithFallback` falls back to a fresh submit only on 404/405 and
      does carry `retriedFromJobId`.
      
      TRAPS. (1) A page util that imports `@/…` is not importable by a plain `tsx` run — the
      instrument is pointed at the web app's tsconfig (`--tsconfig apps/isis/web/tsconfig.json`)
      so its drive can call the real builder. (2) Blanket-redacting the whole bundle would eat
      the job id through the credit-card pattern; the fix redacts the sensitive parts and keeps
      the identifiers, which is what makes a support bundle useful.
      
      2 of 9 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`: the run identities,
      the cost subjects, the log producer and the replay verb do not exist, and there is no
      failure Playwright journey. -->
      
  • I7.6 Complete the output comparison workspace for each supported asset kind. Evidence: synchronized, accessible comparison tests.

    • I7.6.a Resolve authorized exact candidate revisions and common typed comparison/provenance/gate/decision state with incompatible-kind handling. Evidence: workspace contract tests.

    • I7.6.b Implement synchronized image/portrait/texture pan/zoom/fit/pixel/ alpha/diff and region annotation. Evidence: raster comparison tests.

    • I7.6.c Implement synchronized video/animation/sequence frame/timecode/ playback/rate/loop/audio/caption and shot/pass alignment. Evidence: temporal comparison tests.

    • I7.6.d Implement synchronized audio/voice/music playback/waveform/ loudness/channel/stem/transcript and time-region annotation. Evidence: audio comparison tests.

    • I7.6.e Implement synchronized mesh/rig/splat/scene camera/light/material/ animation/LOD/wireframe/bounds and object-region annotation. Evidence: 3D comparison tests.

    • I7.6.f Present raw gate values/units/thresholds/uncertainty, prompt/ workflow/model/policy diffs, rights, provenance, cost, and target fit. Evidence: metadata comparison tests.

    • I7.6.g Implement select/reject/park/regenerate/comment decisions against exact revisions with stale/concurrent state handling. Evidence: decision tests.

    • I7.6.h Provide text/table alternatives to every canvas, full keyboard/ screen-reader/zoom/reduced-motion parity, responsive inspection, and large-media performance. Evidence: synchronized accessible E2E. <!-- MEASURED 2026-08-18 (I7.6.a–h). The parent asks for a comparison workspace "for each supported asset kind" with SYNCHRONIZED inspection per modality, so the measurement is which synchronization affordance each of .b–.e names exists as a control. Instrument: generate-isis-comparison-workspace.mts (+7 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-comparison-workspace.json.

      AN AVERAGE OF NOTHING IS NOT ZERO. `calculateAverageConsistencyScore` returns 0 for an
      empty metric set — the same number it returns when every metric scored zero — and the
      workspace's key-metric tile rendered that as "Overall Score 0.0%" before either side of
      the comparison had been chosen, which reads as the worst possible result rather than as
      no result. FIXED: `formatConsistencyScorePercent` answers "Not measured" for an empty set
      and the section states how many metrics were computed; the numeric function is unchanged
      because twenty-six comparison profiles fold it into their own report score. Driven: empty
      → "Not measured", one metric scoring 0 → "0.0%", two averaging 0.75 → "75.0%".
      
      A PRESET IS NOT A VIEWER. The workspace is 4969 lines carrying 32 "Apply <asset kind>
      Preset" buttons over 32 `*-comparison-utils` profiles whose consistency reports are real
      computations — jaccard over tags and name tokens, type equality, metadata identity
      comparison — and then it shows the two outputs as a plain `<img>`, `<video>` or
      `<audio>`. There is NO canvas anywhere in it. Of the 26 synchronization affordances
      .b–.e name — pan, zoom, fit, pixel, alpha, diff, frame, timecode, playback, rate, loop,
      caption, waveform, loudness, channel, stem, transcript, camera, light, material,
      animation, LOD, wireframe, bounds and the two annotation kinds — exactly ONE is a
      control: Add Annotation. Of .g's five decisions only comment has one; select, reject,
      park and regenerate do not. Of .f's twelve metadata subjects 3 are presented — prompt,
      workflow and model diffs, as "A vs B" lines — and gate values, units, thresholds,
      uncertainty, policy diffs, rights, provenance, cost and target fit are not. 0 of 13
      operator Playwright specs open the workspace.
      
      TRAPS. (1) AN ASSET-KIND PRESET IS NOT A VIEWER AFFORDANCE: counting the preset buttons
      as controls reported a temporal `frame` stepper (from "Apply Storyboard Frame Set
      Preset") and a 3D `material` toggle (from "Apply Material PBR Pack Preset"); the preset
      controls are excluded from the affordance vocabulary. (2) "Reset Selection" is not a
      select DECISION — decision verbs match whole words only. (3) A LOCAL VARIABLE IS NOT A
      PRESENTATION: testing the metadata markers against the module reported "unit" as
      presented because a byte-size formatter declares `const units = ['KB','MB','GB','TB']`
      two thousand lines above the surface. (4) A COMPARISON OPERATOR LOOKS LIKE A TAG
      BOUNDARY: extracting rendered text with `>([^<>{}]+)<` pulls in `>= 1024 && unitIndex <`
      from that same formatter, so the extractor is anchored on named text-bearing tags.
      (5) "Costume Outputs" is not a cost — whole-word matching again.
      
      3 of 10 properties present. ALL EIGHT CELLS AND THE PARENT LEFT `[ ]`: 25 of 26
      synchronization affordances, 4 of 5 decisions and 9 of 12 metadata subjects have no
      surface, and there is no accessible comparison E2E. -->
      
  • I7.7 Complete the release-gate dashboard with freshness and remediation. Evidence: stale or unavailable evidence cannot look green.

    • I7.7.a Build candidate/target gate-set projection with mandatory/ advisory/N/A/unavailable/error/manual states, raw measurements, uncertainty, overrides, and exact versions. Evidence: projection tests.

    • I7.7.b Present overall readiness as a derived explanation with counts and blockers, never as a standalone color or client-authored boolean. Evidence: readiness tests.

    • I7.7.c Calculate and display subject/evaluator/registry/profile/policy/ rights/consent freshness and exact invalidating changes. Evidence: freshness fixtures.

    • I7.7.d Provide filters/grouping by category/owner/severity/target/status, accessible measurement details, evidence links, regions/timecodes, and comparison. Evidence: dashboard interaction tests.

    • I7.7.e Route each blocker to authorized remediation, rerun, review, or exception request with impact/cost, state eligibility, and immutable outcomes. Evidence: remediation journey.

    • I7.7.f Test stale/unavailable/error/unknown/cancelled/partial data, direct release attempt, concurrent evidence change, accessibility, and large gate matrices. Evidence: fail-closed Playwright suite. <!-- MEASURED 2026-08-18 (I7.7.a–f). The parent's evidence clause is "stale or unavailable evidence cannot look green", which presumes a surface on which evidence could look green. Instrument: generate-isis-release-gate-dashboard.mts (+5 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-release-gate-dashboard.json.

      THERE IS NO RELEASE-GATE SURFACE. The operator console mounts 23 routes and NONE of them
      is a gate, release or readiness view, so .a's projection, .d's filters and grouping, .e's
      remediation routing and .f's fail-closed suite have nothing to attach to. Of .e's four
      remediation destinations — remediate, rerun, review, exception — none has an API route
      either, and 0 of 13 operator Playwright specs open a gate route.
      
      A REQUIREMENT NOBODY ASKED FOR IS NOT A REQUIREMENT THAT PASSED. The readiness model that
      does exist, `buildIsisReleaseReadiness` in `libs/oshun/generation-control-isis`, computed
      `const modelVerified = input.model ? input.model.releaseState === 'eligible' : true` — a
      request that names no model reported the SAME value as a verified, release-eligible one —
      and `provenanceReady: true` whenever a new output was expected. .a asks a gate projection
      to carry an N/A state distinct from a pass, and both of those were passes. FIXED
      additively: the readiness object carries `notApplicableRequirements`, populated with
      `model_not_named`, `provenance_pending_new_output` and `quality_gate_not_required`.
      DRIVEN: no model → `modelVerified=true` with `model_not_named` recorded; a named eligible
      model → `modelVerified=true` with nothing recorded. 20/20 test files in that library stay
      green.
      
      MEASURED, NOT FIXED. (1) .a: 5 of the 6 states are representable (mandatory via
      `missingRequirements`, N/A via the new list, manual via `humanReviewRequired`, plus
      unavailable and error in the type vocabulary) and ADVISORY is not — every requirement is
      a blocker or nothing. Raw measurements, uncertainty and exact versions do not travel with
      the verdict at all: the readiness object is booleans, reasons and requirement lists.
      (2) .c: 0 of the 7 freshness subjects — subject, evaluator, registry, profile, policy,
      rights, consent — has an instant compared anywhere in the readiness path; the gate-side
      measurement of the same absence is `evidence/v1-workbenches/isis-gate-result-freshness.json`
      (I6.8), which recorded that the contract compares instants 14 times and never the one
      that decides publishing. (3) The one gate that could pass, `modelVerified`, requires
      `isVerified && hash`, and I1.2.c recorded that `is_verified` has NO WRITER — so no model
      in the registry reaches `eligible` by any path.
      
      2 of 9 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: there is no dashboard,
      no freshness calculation, no advisory state, no remediation route and no fail-closed
      suite. -->
      
  • I7.8 Complete the provenance inspector and recall workflow. Evidence: tamper, revoke, and blast-radius journey.

    • I7.8.a Build a typed lineage graph/list projection for sources/ references, briefs/specs, workflows/nodes, models/adapters, runs/stages, outputs/edits, credentials, bundles, collections, and releases. Evidence: inspector contract tests.

    • I7.8.b Render forward/reverse paths with stable identities/versions/ hashes, relationship reasons, redactions, cycles, pagination, and a complete noncanvas tree/table alternative. Evidence: lineage UI tests.

    • I7.8.c Verify artifact/manifest/credential signatures/hashes and show valid/invalid/missing/stripped/unknown status with tool/version/time. Evidence: tamper journey.

    • I7.8.d Build blast-radius query/preview for source/consent/license/model/ workflow/output changes with direct/transitive items, targets, owners, actions, and unknowns. Evidence: preview tests.

    • I7.8.e Implement authorized quarantine/recall case creation, scoped confirmation, progress, partial failures, target receipts, notification, and restoration review. Evidence: recall journey.

    • I7.8.f Enforce sensitive provenance field controls, tenant isolation, large-graph performance, keyboard/screen-reader navigation, and immutable case history. Evidence: security/accessibility tests. <!-- MEASURED 2026-08-18 (I7.8.a–f). The parent's evidence is a "tamper, revoke, and blast-radius journey", so the measurement asks for each of the three whether the operator console can reach it. Instrument: generate-isis-provenance-inspector.mts (+6 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-provenance-inspector.json. No fix in this cell: everything that exists is correct, and everything else is absent.

      A LINEAGE OF OUTPUTS IS NOT A LINEAGE OF THE PIPELINE. `LineageNodeType` is
      `source | derived | composite` and every node carries an `outputId`, so of the ten
      families .a asks a lineage projection to carry — sources/references, briefs/specs,
      workflows/nodes, models/adapters, runs/stages, outputs/edits, credentials, bundles,
      collections, releases — exactly ONE is expressible. Five edge types and seven provenance
      routes exist (`GET/POST /outputs/:id/provenance`, `GET /outputs/:id/lineage`,
      `/ancestors`, `/descendants`, `POST/DELETE /outputs/lineage…`).
      
      THE INSPECTOR EXISTS, IN THE OTHER PRODUCT. `apps/oshun/web` has an
      `IsisProvenanceInspector` fed by a real chain verification: `verifyProvenanceLedger`
      reports `ok` and the index the chain broke at, and the Oshun server loader is its one
      product consumer. The Isis operator console has NO inspector, NO verification status, NO
      blast-radius preview and NO recall case — its entire provenance surface is the
      register/update forms and a single "Refresh Lineage" button on the outputs page. 0 of
      .c's five statuses (valid/invalid/missing/stripped/unknown) are shown against a
      signature, hash or integrity check; 0 of .e's seven case steps are named; 0 recall or
      quarantine routes exist anywhere in the Isis services; 0 of 13 operator Playwright specs
      mention provenance, lineage or recall.
      
      CHECKED AND FOUND CORRECT — recorded so a later pass does not re-flag them. (1) The
      lineage walk keeps a `visited` set and a depth cap, so a cyclic graph terminates.
      (2) The lineage route resolves an owner and a tenant scope before reading. (3) The
      fabricated "broken ledger" verdict in the Oshun loader is reachable only through
      `FixtureIsisProvenanceLedgerStore`, which is returned only when
      `OSHUN_ENABLE_TEST_HARNESSES=true` AND `OSHUN_ISIS_PROVENANCE_LEDGER_FIXTURE=clean` — a
      gated test double at a dependency boundary, which the quality bar explicitly permits.
      
      TRAP. A STATUS WORD IS NOT A VERIFICATION STATUS: matching the bare literals credited
      `'unknown'` (an id fallback at OutputsPage:850) and `'missing'` (a bundle-presence table
      cell) as verification statuses; a status counts only where it sits beside a signature,
      hash, integrity or verification word.
      
      3 of 9 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: nine of .a's ten
      lineage families cannot be nodes, .c has no verification surface in this console, .d has
      no blast radius at all, .e has no recall route, and .f has no case history to be
      immutable. -->
      
  • I7.9 Integrate AAA brief, graph, model, training, voice, music, 3D, and post-process surfaces into Yemaya Studio navigation. Evidence: no duplicated Isis operator controls.

    • I7.9.a Map each AAA surface to a Yemaya project/asset/shot/sequence context, authorized roles, Isis route/capability, and return location. Evidence: navigation integration matrix.

    • I7.9.b Add restrained project-level workbench navigation and command palette entries for brief, graph, model, training, voice, music, 3D, and post-process without a second global nav. Evidence: navigation tests.

    • I7.9.c Reuse Isis APIs/contracts and audience-safe projections while keeping provider fleet, global registry lifecycle, incidents, billing, and security administration in operator hosts. Evidence: boundary tests.

    • I7.9.d Preserve selected project/asset/shot/sequence/revision and unsaved state warnings across tool transitions and deep links. Evidence: context journey.

    • I7.9.e Render feature unavailable/degraded/not-configured/unauthorized states from capabilities with safe remediation and no hidden-feature hints. Evidence: capability state tests.

    • I7.9.f Verify direct URL/role/property/function authorization, keyboard/ screen-reader/responsive navigation, and that no operator control/API is copied into Yemaya. Evidence: AAA boundary Playwright suite. <!-- MEASURED 2026-08-19 (I7.9.a-f). The parent's evidence is "no duplicated Isis operator controls", which is two questions wearing one coat: does Yemaya Studio HOST the AAA surfaces, and does it AVOID the operator controls. The second answer is largely good. The first is the finding. Instrument: generate-yemaya-aaa-navigation.mts (+11 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/yemaya-aaa-navigation.json.

      THE BOUNDARY EVICTED THIRTY-ONE CAPABILITIES TO A HOST THAT MOUNTS NONE OF THEM.
      `AAA_ONLY_STUDIO_ROUTES` names 31 route segments the contemplative product must never serve
      because they are AAA controls, and BASE.5.e ratified that list against
      `docs/oshun/studio-vs-yemaya-studio.md`. Yemaya Studio is where they were sent: 0 of the 31
      have a route in `apps/yemaya/studio-{web,desktop}` and 2 appear in their sources at all
      (`music-generation` and `text-to-audio`, generation-type strings inside a store). Of the
      eight surface families .b names - brief, graph, model, training, voice, music, 3D,
      post-process - the navigation vocabulary names 2: music through "Score Editor", and
      post-process through "Object Removal"/"HDR Trim"/"Relight Studio". Brief, graph, model,
      training and voice are named nowhere in the navigation of either host.
      
      .b ASKS FOR NO SECOND GLOBAL NAV AND THERE ARE ALREADY TWO: a `<nav>` in the header (6
      entries) and a `<nav>` in the sidebar (19 entries), neither scoped to a project, plus 0
      command palettes - the only mention of one in the estate is an `@example` in a doc comment
      on `useKeyboardNavigation`, a hook with no caller. Three of those controls land on the 404:
      the sidebar's "Assets" and the header's persistent "Templates" and the welcome page's "From
      Template", all pointing at routes that have never existed. So do 3 of the 7 in-app PWA
      manifest entry points (`/ai`, `/editor/script`, `/protocol`) - and `/editor/script` is a
      registered OS file handler, so opening a script file with the app lands on Not Found.
      
      NO AAA WORKBENCH KNOWS WHICH SHOT IT IS EDITING. 7 routes mount a page that imports an
      `@isis/*` library; 0 of them read a route param or the project store, so .a's mapping to a
      project/asset/shot/sequence context and .d's preservation of it across a transition have
      nothing to preserve. 1 of .d's 5 context dimensions reaches an Isis job at all - a
      `metadata.projectId` the store attaches when a project happens to be open. The
      unsaved-changes modal is translated into 3 locales, 5 strings each, and has 0 consumers.
      
      AN AUTHORIZATION FUNCTION WITH NO CALLER IS NOT AN AUTHORIZATION. `UserEntitlementTier =
      'contemplative' | 'aaa-creator' | 'operator'` exists in exactly one place, the score
      editor's `core.ts`, and `canAccessSceneScoreEditor`/`assertSceneScoreEditorAccess` have 0
      product callers and 2 test callers. Both hosts mount the scene score editor: the web one
      behind `RoleGuard roles="admin"` (Yemaya's role vocabulary is `admin|user|viewer`, with no
      AAA role in it), the desktop one behind nothing - 6 of 6 desktop routes have no guard, and
      that renderer has no auth store at all. The two `score-editor/core.ts` files are
      byte-identical 490-line copies. On the web, 24 routes admit any authenticated user
      including `viewer`, and before this pass the four guards had 0 tests.
      
      A DECLARED SURFACE IS NOT A RENDERED ONE. A walk of relative imports from `main.tsx` reaches
      85 modules; 26 UI modules are outside it, including all five of `src/layouts/` - `AppLayout`,
      `AuthLayout`, `DashboardLayout`, `SettingsLayout` and `ProjectLayout`, the last being exactly
      the project-context shell .d would have needed. That is also why the operator answer is
      unambiguous rather than arguable: `SettingsLayout` declares Billing and API Keys sections,
      but nothing renders it, and the settings page that IS rendered has four sections (General,
      Appearance, Account, About). Of .c's five operator concerns, 3 have no control anywhere and
      2 have exactly one surface between them - the desktop settings page's "AI Provider" heading
      over an "Enter your API key" field, in the host with no authentication.
      
      A CONTROL THAT DISCARDS ITS INPUT STILL LOOKS LIKE A SETTING: across the two settings pages
      the entry reaches, 2 of 15 form controls are bound to a handler. The other 13 - six
      `defaultChecked` checkboxes claiming features are on, two `defaultValue` selects, and the
      API-key password field - read nothing and write nothing.
      
      FIXED IN THIS CELL. All four domain client stores (Isis, Sophia, Hathor, Bellona) named
      `clientConfig` in `partialize`, and every one of those configs declares an `apiKey` plus a
      bearer or access token, so `persist` would write a service credential into localStorage in
      plaintext where it outlives the session. `IsisClient.getConfig()` already refuses to hand
      its own credentials back out, reporting `hasApiKey`/`hasBearerToken`; the stores that supply
      them now follow the same rule through `withoutClientCredentials`, which keeps the endpoints
      and drops `apiKey`/`bearerToken`/`accessToken`. Driven both ways: with the redaction
      removed, the new spec finds the key in `localStorage.getItem('yemaya-isis-store')`. Nothing
      on a product path sets a credential today, so this is what the persisted shape would have
      done to the first caller of `updateConfig` - which is when it would have been too late.
      
      Also added, as rules rather than sentences: `RouteGuard.spec.tsx` (5 assertions) for .f's
      direct-URL and role authorization, including the characterisation that `AuthGuard` admits
      `viewer`, and `client-credential-persistence.spec.ts` (7 assertions).
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The Isis
      store's failure path records the real submission error on the job and fabricates no success.
      (2) `metadata.projectId` is spread into the POST body by `submitJob` and carried by the
      generation API's job route, so the one context dimension that is sent does persist. (3)
      `RouteGuard` evaluates its role check exactly once per mount: `RoleGuard` builds a fresh
      `requireRoles` array every render and `RouteGuard` lists it in an effect dependency array,
      which looks like a re-check loop and is not one, because the array is built by the parent
      and the parent does not re-render when the guard's own state settles (measured: 1 render, 1
      role check, over 100ms). (4) The Isis client store touches 1 of the client's 4 resources
      (`generation`) and 5 methods, and no operator resource. (5) The e2e suite's axe fixture runs
      a WCAG 2.1 AA analysis after every spec.
      
      TRAPS. A REACT CONTEXT PROVIDER IS NOT A PROVIDER FLEET: matching "provider" credited 24
      files, 13 of them `<X.Provider>` wrappers. A WORD IN AN UNRELATED ENUM IS NOT A CAPABILITY
      STATE: `'unavailable'` is a touch-support enum member in the desktop's multi-monitor code
      and a rollback-dialog title in its updater, and `'Unauthorized'` is an HTTP status label -
      adding "the file also says capability" still credited the first, because a DISPLAY
      capability is not a feature capability. A macOS ENTITLEMENT IS NOT A PRODUCT ENTITLEMENT: 8
      files carry the word, 4 code-signing and 4 score-editor, and 0 consult a product
      entitlement. A SUBSTRING IS NOT A WORD: `rig` matches `right` in 10 pages. And
      `<RouteSuspense` starts with `<Route`, so splitting the router on the bare string reported a
      null element for all 33 web routes.
      
      2 of 17 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: .a has no context
      binding and no AAA role, .b has two global navs and no palette, .c reaches Isis by importing
      its libraries rather than calling its API, .d has nothing to preserve, .e has no capability
      source to render a state from, and .f has no boundary suite - 0 of the 24 e2e tests asserts
      that an operator control is absent. -->
      
  • I7.10 Implement handoff from Yemaya project context to an Isis run and back. Evidence: project, shot/asset, actor, and lineage persist.

    • I7.10.a Define handoff intent/envelope for tenant/project, asset/shot/ sequence/revisions, brief/spec, requested operation, actor/audience, return route, expiry, nonce, and signature. Evidence: handoff schema.

    • I7.10.b Validate Yemaya context access and current revisions before issuing the minimal signed Isis link/request. Evidence: issuance tests.

    • I7.10.c Reauthorize Isis actor/object/function, validate link freshness/ replay, resolve canonical inputs, and preview substitutions/missing dependencies before plan creation. Evidence: acceptance tests.

    • I7.10.d Create Isis request/plan/run with immutable Yemaya source links and carry only authorized project/shot/asset metadata. Evidence: run lineage tests.

    • I7.10.e Return via a signed intent to the exact permitted Yemaya context and highlight resulting proposal/output/revision without assuming it was accepted. Evidence: return journey.

    • I7.10.f Handle logged-out, expired, changed/deleted/quarantined context, cancellation, failure, multiple outputs, and cross-tenant attempts. Evidence: handoff Playwright matrix. <!-- MEASURED 2026-08-19 (I7.10.a-f). The parent's evidence is "project, shot/asset, actor, and lineage persist", so the measurement follows one generation the whole way round: what leaves Yemaya, what Isis is asked to check before it acts, what the run records about where it came from, and what comes back. Instrument: generate-yemaya-isis-handoff.mts (+9 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/yemaya-isis-handoff.json.

      THERE IS NO ENVELOPE; THERE IS A JOB SUBMISSION. The only request Yemaya sends Isis is
      `JobSubmitRequest` (10 fields), and 1 of .a's 9 envelope elements is expressible in it -
      the requested operation, as `type`. No tenant, no shot/sequence/revision, no brief, no
      actor, no return route, no expiry, no nonce and no signature. Two callers issue it: the
      studio-web Isis store, 4 `submitJob` calls attaching `{projectId}` and only when a project
      happens to be open, and `yemaya asset generate`, 1 call attaching nothing at all. Neither
      reads a revision before issuing, which is .b's whole subject. On the receiving side 0 of
      .c's 5 acceptance checks appear anywhere on the job route or job service: no
      reauthorization of the actor against the object, no link freshness, no nonce or
      consumed-token set, no canonical input resolution and no substitution preview.
      
      A COLUMN NAMED FOR A LINK IS NOT A LINK. `Asset.isisGenerationId` and
      `Project.isisWorkflowIds` are the only persisted Yemaya-to-Isis references in the estate.
      Both are declared in the Prisma schema; outside that schema and its generated client,
      0 files in `apps/yemaya` or `libs/yemaya` mention either one. Nothing writes them, so .d's
      "immutable Yemaya source links" are two empty columns - and
      `scripts/verify-phase-9-completion.mjs` passes "9.19 schema capability references" by
      regex-matching those column names in the schema file, which is a check that cannot tell a
      filled column from an empty one.
      
      FIXED: THE RETURN LEG DEAD-LETTERED EVERY EVENT AND ATE EVERY FAILURE. `apps/yemaya/api`
      subscribes `isis.asset.generated` and `isis.job.failed` on the bus, and those two handlers
      are the entire path by which a finished run reaches the project that asked for it. Both
      were written against a payload Isis has never published, and the unit test fed each handler
      its own consumer-side interface, so both stayed green.
      
      - `isis.asset.generated`: the handler read `metadata.format` and `metadata.fileSize`. The
        producer (`apps/isis/output-registry`) sends `filename`, `mimeType`, `sizeBytes`,
        `provenance` and `thumbnails`, and marks `metadata` OPTIONAL. Driven with the producer's
        own literal: `Cannot read properties of undefined (reading 'format')`, dead-lettered,
        0 asset references recorded, 0 metrics, 0 notifications, 0 workflow actions.
      - `isis.job.failed`: the handler read `errorCode`/`errorMessage`/`retryCount`/`retryable`
        flat; the producer nests them under `error` and calls the count `attemptCount`. Every one
        read `undefined`, `categorizeError` threw on `undefined.toLowerCase()`, and THAT handler's
        catch acknowledges unconditionally - "informational, don't create a failure loop". Driven:
        `acked=1, published=[]`. Every Isis failure was consumed and discarded in silence.
      
      Both consumer payloads now mirror `@oshun/contracts`, both handlers read the fields the
      producer sends, and `isis-return-leg.spec.ts` (+9) drives them with the producer's own
      literals - including the optional-`metadata` case, the `model_3d` type the old union could
      not express, and `workflow: 'default'`, which Isis sends when a job ran under no workflow
      and which is not a workflow to ask for an action. The instrument compares the producer's
      field set with each consumer's on every verify run.
      
      STILL OPEN, AND THE SAME DEFECT: `libs/bellona/event-handlers` and `libs/hathor/event-handlers`
      subscribe `isis.asset.generated` too (through `apps/bellona/build-api` and
      `apps/hathor/world-api`), and both still declare `generationParams` and `thumbnailUrl`,
      which Isis sends under `provenance.parameters` and `thumbnails` respectively. Neither
      crashes - both feed the missing field to an extractor that returns null - so Hathor never
      links a generated asset to a world element and Bellona never routes one to an engine. That
      is Bellona and Hathor scope, not this cell's, and it is left for their owners.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) Both
      producers' payload literals match their canonical `@oshun/contracts` schemas field for
      field. (2) `metadata.projectId` travels Yemaya store -> job submission -> output row ->
      event payload, so the one context dimension that is sent does survive the round trip.
      (3) `IsisEventPublisher` swallows publish failures deliberately and logs them, rather than
      breaking the generation it is reporting on.
      
      TRAPS. A JOB REPLAY IS NOT A LINK REPLAY: matching `replay` on the job route credited 56
      hits, every one the re-run verb, so .c's replay check reads 0 only once a nonce or
      consumed-token set is what is asked for. A SHORTHAND PROPERTY HAS NO COLON: both producers
      write `projectId,` and `userId,` in shorthand, and a `key:` pattern reported the two
      payloads as diverging from the contract they satisfy. A COMMENT STRIPPER THAT DOES NOT KNOW
      ABOUT STRINGS EATS CODE: the producer literal contains `` `s3://${bucket}/${key}` ``, and a
      naive `//` rule deletes the rest of that line. AN EMPTY SET AGREES WITH EVERYTHING: Bellona
      and Hathor declare no job-failed payload at all, and "no field the producer never sends"
      reported them as agreeing with a contract they never read. A WORD THE DOMAIN ALREADY OWNS
      IS NOT A REFERENCE TO YOUR FEATURE: "handoff" is ordinary film vocabulary - a second-unit
      camera handoff, a shift handoff, an AI-to-human session handoff - and matching it credited
      2 specs. A PRODUCT NAME INSIDE A STRING IS NOT A VISIT TO THAT PRODUCT: the Yemaya smoke
      spec renders `isis-relight-compositor-export` and types `isis://references/...` into a
      field without leaving its own origin. 0 of the 14 Yemaya and Isis e2e specs navigates to
      the other product.
      
      1 of 9 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: .a has no envelope,
      .b validates nothing before issuing, .c performs none of the five acceptance checks, .d's
      two link columns have no writer, .e's return is an unsigned bus event with no return route
      (the leg now works, but a working leg is not a signed intent), and .f has no cross-host
      journey at all. -->
      
  • I7.11 Complete curated Oshun Studio generation cards for illustration, narration, ambient audio, explainer, caption/dub, accessibility, music, and approved 3D briefings. Evidence: tier allowlist tests.

    • I7.11.a Define the curated generation-card contract for approved operation, audience/host, typed high-level inputs, references, cost/time estimate, confirmation, progress, output, and safe errors. Evidence: card schema tests.

    • I7.11.b Implement illustration briefing for subject/purpose/style/ composition/aspect/palette/required/excluded elements and authorized references. Evidence: illustration journey.

    • I7.11.c Implement narration briefing for approved voice/language/tone/ pace/pronunciation/script/timing/format with consent-safe choices. Evidence: narration journey.

    • I7.11.d Implement ambient-audio briefing for environment/events/mood/ perspective/duration/loop/loudness/channel/format. Evidence: ambient audio journey.

    • I7.11.e Implement explainer briefing for audience/objective/script/ scenes/style/duration/narration/captions/aspect/target. Evidence: explainer journey.

    • I7.11.f Implement caption/dub briefing for source/language/transcript/ translation/speaker mapping/timing/style/format/review. Evidence: caption/dub journey.

    • I7.11.g Implement accessibility briefing for alt text, audio description, captions/transcript, reading level, language, timing, and target. Evidence: accessibility asset journey.

    • I7.11.h Implement music briefing for purpose/mood/tempo/form/duration/ instrumentation/hit points/loop/stems/format with rights-safe references. Evidence: music journey.

    • I7.11.i Implement approved 3D briefing for asset/use/subject/style/ dimensions/units/views/materials/poly/texture/rig/format/target without graph/model controls. Evidence: 3D journey.

    • I7.11.j Derive availability/defaults/validation/estimate from canonical capabilities and submit through the same governed plan/dispatch APIs. Evidence: tier allowlist/contract tests. <!-- MEASURED 2026-08-19 (I7.11.a-j). The parent's evidence is "tier allowlist tests", so the measurement asks, for each of the eight briefings .b-.i name, whether a card of its own exists in the ratified catalog and what its typed inputs can express. Instrument: generate-oshun-curated-cards.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-curated-cards.json.

      FIXED: A BORROWED ENTITLEMENT IS NOT THIS CARD'S ENTITLEMENT. `CARD_KINDS` had eight
      members and two of them were not among the eight briefings this cell names: `video` and
      `video-from-image` are cards, and MUSIC and the 3D BRIEFING were not. Both nevertheless
      ship - `/studio/generation/music` and `/studio/generation/nyx-3d`, each with its own BFF
      route - and both resolved a DIFFERENT card's spec to do their entitlement check:
      `music-route.ts` called `getCuratedCardByKind('ambient-audio')` and `nyx-3d-route.ts`
      called `getCuratedCardByKind('illustration')`. The same borrowed spec is what the route
      hands back as the provenance preview, and `CardShell` prints it: a music generation
      displayed `Provenance: wfc-curated-ambient-audio v1.0.0`, and a sky briefing displayed
      `wfc-curated-illustration`. Kill-switching the 3D briefing was impossible without killing
      illustration; a tenant allowlisted for ambient audio was allowlisted for music. Both now
      have their own `CardKind`, their own catalog entry (`curated-music`,
      `curated-sky-3d-briefing`), their own kill switch, tenant allowlist, minimum tier and
      provenance. Entitlement-neutral by construction: the new entries carry the same
      `curated-creator` / `all` / `killSwitchActive: false` defaults the borrowed ones did.
      Driven both ways - with the routes pointed back at the borrowed kinds, the two new
      assertions go red. The two kinds carry no `CuratedCardInputs` variant on purpose: each
      validates its own request shape at its own route, and `POST /v1/isis/curated-cards/generate`
      already rejects any submission whose `inputs.kind` does not match the card's kind.
      
      A SUBMIT SPINNER IS NOT GENERATION PROGRESS, AND AN EMPTY LITERAL IS NOT OUTPUT. 7 of .a's
      9 contract elements are real - approved operation, audience/host, typed inputs, references,
      cost/time estimate, confirmation and safe errors all exist and are server-authoritative.
      The two that are not: `CardShell` flips its button to "Generating…" while `busy` is true,
      and `busy` is true only between the POST and its 202, after which the label reverts and the
      status line reads "Accepted — queued for generation"; nothing follows the enqueued job.
      And `CardShell` renders a variants grid that its only caller feeds
      `variants: [] as const`, always. So a creator submits a card and never sees the thing it
      made.
      
      WHAT THE EIGHT BRIEFINGS CAN EXPRESS: 24 of the 75 fields .b-.i name. illustration 3/9
      (no purpose, composition, palette, required or excluded elements, or authorized
      references), narration 4/8 (no pace, pronunciation, timing or format), ambient audio 3/9
      (no environment, events, perspective, loudness, channel or format), explainer 2/10 (no
      audience, objective, script, scenes, duration, narration, captions or target), caption/dub
      2/9 (no transcript, translation, speaker mapping, timing, style, format or review),
      accessibility 2/7 (no audio description, reading level, language, timing or target), music
      6/10 (no form, instrumentation, hit points or format), 3D 2/13 - the sky-event briefing is
      a lunar-phase visualisation with an event id and a viewer angle, not the general approved
      3D briefing .i describes.
      
      .j IS THREE ENDPOINTS, NOT ONE. `/v1/isis/curated-cards/generate`, `/v1/isis/music/generate`
      and `/v1/isis/nyx-3d/generate`. Availability, defaults and validation ARE derived
      server-side from the catalog, the entitlement gate and the per-card validators; the
      estimate is the one that is not, and honestly so - see below. The curated-cards studio
      renders a form for 6 of the 10 approved kinds; it now skips the four it cannot render
      instead of emitting an empty wrapper for each.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The catalog
      is server-authoritative and frozen, so a client can never assert its own kill switch, tier
      or tenant allowlist. (2) `PENDING_COST_COEFFICIENTS` and `areCostCoefficientsPending` are
      an honest seam, not a stub: until a deployment publishes per-workflow-class pricing the
      pre-flight estimate reads "pending" rather than a fabricated number, and the file says so.
      (3) The generate route rejects a submission whose `inputs.kind` does not match the catalog
      card's kind. (4) Every curated submission passes `auditRawGraphLeak` - model hashes, LoRA
      ids, scheduler names, ComfyUI node keys - before field validation.
      
      2 of 13 properties present. ALL TEN CELLS AND THE PARENT LEFT `[ ]`: .a is missing progress
      and output, .b-.i can express 24 of 75 asked fields between them, and .j dispatches through
      three endpoints with no published estimate. -->
      
  • I7.12 Ensure curated cards expose only approved high-level controls. Evidence: raw-graph/model/provider leak scan passes.

    • I7.12.a Ratify a per-card allowlist of field names/types/values/defaults/ validation/errors and permitted provenance/capability summaries. Evidence: curated control matrix.

    • I7.12.b Generate curated projections from typed server contracts rather than hiding AAA/operator controls with client CSS or feature flags. Evidence: projection tests.

    • I7.12.c Reject unknown/nested/alias/raw JSON fields, workflow/model/node/ sampler/provider IDs, advanced parameter names, and mass assignment at the curated API. Evidence: adversarial request tests.

    • I7.12.d Remove hidden details from HTML/JS bundles, responses, errors, URLs, state stores, telemetry, exports, deep links, and capability hints. Evidence: static/runtime property leak scan.

    • I7.12.e Keep approved high-level values mapped deterministically to governed specs/presets and show audience-safe effective summaries. Evidence: mapping golden tests.

    • I7.12.f Test every card as curated/contemplative/AAA/operator actors, direct routes, crafted payloads, stale links, and changed tier. Evidence: tier authorization Playwright suite. <!-- MEASURED 2026-08-19 (I7.12.a-f). The parent's evidence is a "raw-graph/model/provider leak scan", so the measurement asks two things: what a crafted request can push IN, and what the curated responses carry OUT. Instrument: generate-oshun-curated-control-surface.mts (+7 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-curated-control-surface.json.

      FIXED: THE VALIDATORS CHECKED THE VALUES AND NOTHING CHECKED THE SET. Each per-card
      validator ended `return issues.length === 0 ? ok(kind, i) : fail(issues)` where `i` is the
      caller's whole object, and `curated-generation-executor` forwards "the admitted bare inputs
      unchanged". Driven: an illustration submission carrying `workflowClassId:
      'wfc-aaa-comfy-graph'`, `providerKey: 'replicate'`, `steps: 150`, `cfgScale: 22` and
      `sampler: 'heun'` returned `verdict=pass` with all five forwarded into the job. The
      raw-graph leak audit does not see any of them: it looks for 64-hex model ids, `lora-…`
      patterns, exactly three scheduler names (`ddim`, `dpmpp_2m`, `euler_a` - `heun` is not one)
      and ComfyUI numeric node keys. A crafted value SHAPE was worse than accepted: `subject:
      { $ne: null }` made `i.subject.trim()` throw `i.subject.trim is not a function`, which the
      route turns into a 500 rather than a 422.
      
      `CARD_INPUT_FIELD_ALLOWLIST` now ratifies 39 field names across the 8 form-driven kinds
      (illustration 5, narration 4, ambient-audio 4, explainer 5, caption-dub 2,
      accessibility-pass 2, video 8, video-from-image 9), and `checkCardFieldAllowlist` runs
      before the per-card switch: an unapproved field name is a 422 issue naming the field, and a
      value that is not a string, number, boolean or array of strings/numbers is a 422 rather
      than a throw. Keyed on `CuratedCardInputs['kind']` rather than `CardKind`, because `music`
      and `sky-3d-briefing` have no form on this API. Driven both ways - with the gate disabled
      the two new route assertions go red, and both confirm the executor was never reached. 7 of
      7 of .c's refusal classes are now refused.
      
      STILL OPEN, AND A RATIFICATION RATHER THAN A BUG. .a asks for an allowlist of "permitted
      provenance/capability summaries" and there is none, so ALL SEVEN identifier fields in the
      card spec travel out to a curated audience: `GET /v1/isis/curated-cards/catalog` returns
      each entry's whole `CardSpec` (`binding.workflowClassId`, `binding.personaId`,
      `binding.voiceProfileAllowlist`) and whole `ProvenancePreview`, and `CardShell` PRINTS five
      of them - `Provenance: {workflowClassId} v{workflowClassVersion}`, `Provider:
      {providerKey} / {providerRegion}`, `Cost class: {costClass}`. Nothing real leaks today
      because the catalog ships `providerKey: 'pending'`, but the surface is built to display the
      tenant's resolved provider the moment a deployment fills it in. Which of the seven a
      curated audience may see is exactly the decision .a asks a human to make; .a's "defaults"
      clause has no declaration either.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) 0 of the 11
      generation components hides a control with `display: none`, `visibility: hidden` or a
      `hidden={…}` prop: the six forms are generated from the typed catalog, which is what .b
      asks for. (2) The card-to-workflow-class mapping is a pure function of the card kind
      (`wfc-curated-${kind}`), so .e's determinism holds. (3) `auditRawGraphLeak` runs on every
      curated submission before field validation.
      
      .f HAS NO SUITE AT ALL: 0 of the 561 Oshun web e2e specs visits
      `/studio/generation/{curated-cards,music,nyx-3d}`, so no card is exercised as a
      contemplative, curated, AAA or operator actor, by direct route, or after a tier change.
      
      3 of 6 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: .a is missing defaults
      and the provenance-summary allowlist, .d ships seven identifiers to the client and prints
      five, and .f has no suite. .b, .c and .e hold. -->
      
  • I7.13 Complete the Oshun output gallery with lineage, status, quality, rights, replay eligibility, and editorial promotion. Evidence: live-data journey.

    • I7.13.a Define gallery list/detail projection for exact asset/output revision, preview, source, status, target, generated/manual identity, audience, and current availability. Evidence: gallery contracts.

    • I7.13.b Implement deterministic search/filter/sort/pagination and collection/group views with restrained media-led hierarchy and complete loading/empty/error/stale/offline states. Evidence: gallery list tests.

    • I7.13.c Present authorized brief/run/output/ingredient lineage, decision history, transformations, provenance credential, and unavailable/ redacted links. Evidence: lineage detail tests.

    • I7.13.d Present quality raw measures/verdicts/freshness, rights/consent/ attribution/expiry, safety state, target compatibility, and recall status without false green aggregation. Evidence: status tests.

    • I7.13.e Calculate replay eligibility from current inputs/dependencies/ policy/rights/capacity with mode, drift, estimate, and exact reason. Evidence: replay projection tests.

    • I7.13.f Implement editorial promotion request/preview/submit/status with destination scope, derivative lineage, role separation, and no direct publication. Evidence: promotion journey.

    • I7.13.g Verify live generation-to-gallery-to-promotion, cross-tenant/ property authorization, keyboard/screen-reader/zoom/responsive behavior, and large-gallery performance. Evidence: gallery Playwright E2E. <!-- MEASURED 2026-08-19 (I7.13.a-g). The parent's evidence is a live-data journey over lineage, status, quality, rights, replay eligibility and editorial promotion, so the measurement asks what the gallery's own record projection can SAY about an output, and what the surfaces on top of it can therefore show. Instrument: generate-oshun-output-gallery.mts (+9 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-output-gallery.json. No fix in this cell: what exists is correct, and what is missing is a projection nobody has designed.

      THE RECORD IS A PRODUCTION FACT, NOT A JUDGEMENT ABOUT ONE. `OutputRecord`'s 18 fields say
      who made an output, with what, when, under which licence tag and at what cost. 4 of .a's 8
      projection properties are carried - preview (`thumbnailUrl`), source
      (`generatorKey`/`workflowClassId`), audience (`entitlementTier`) and current availability
      (`takenDownAtUnixSeconds`). There is no revision, no status, no target and no
      generated/manual identity, so the gallery cannot name the exact thing it is showing you or
      say what state it is in. And of .d's ten status families exactly ONE is on the record:
      rights, as a `licenseTag` string. No quality measure, no verdict, no quality freshness, no
      consent state, no attribution, no expiry, no safety state, no target compatibility and no
      recall status. .d's "without false green aggregation" is therefore satisfied the only way
      it can be - nothing is aggregated because nothing is measured.
      
      THE LINEAGE KNOWS TWO KINDS OF THING. `LineageNodeKind` is `output | job`, so .c's brief
      and ingredient lineage have no node to be, and there is no decision history. The six edge
      kinds (`derived-from`, `cited-by-claim`, `used-in-lesson`, `embedded-in-passage`,
      `used-in-scene`, `send-to-editorial`) are the transformations it can express. A withheld
      link is COMPUTED AND NEVER SHOWN: `buildRenderableLineage` returns `droppedForeignOutput`
      and `droppedInvalid` beside the admitted edges, and the client renders neither - so a
      redacted link is silently absent rather than shown as redacted, which is the one thing .c
      asks for by name.
      
      REPLAY PLANS A BRANCH; IT DOES NOT DECIDE ELIGIBILITY. 3 of .e's 9 parts: current inputs
      (`OriginalInputs`), an estimate (`estimatedCostCents`) and an exact reason
      (`BranchValidationError` with `out-of-range`/`unknown-parameter` codes). Dependencies,
      policy, rights, capacity, mode and drift appear nowhere in `branch-replay.ts`, so nothing
      asks whether the run may be repeated - only whether the tweaked parameters are in range.
      
      PROMOTION IS AN OPERATOR BULK ACTION, NOT A WORKFLOW. 5 of .f's 8 parts hold, and the three
      that hold hardest are the ones that matter: role separation (`AuthorizationCheck` gates the
      action by operator tier), no direct publication (the verb emits a `send-to-editorial` edge
      to a `queueId`, never a publish), and derivative lineage (that edge is a lineage kind).
      What is absent is the workflow around them - no promotion request a creator can make, no
      preview of what promotion would do, and no status to follow afterwards; an undo window is
      not a status.
      
      .b IS HALF BUILT AND THE HALF THAT EXISTS IS THE HARD HALF. `applyGalleryFilter` sorts by
      `producedAtUnixSeconds` descending with a stable `outputId` tiebreak and pages by cursor -
      deterministic, which is what .b asks first. The surface renders 2 of the 5 states it names:
      loading and error. There is no empty state at all (the grid never checks for an empty
      result), no stale state, no offline state, and no collection or group view.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1)
      `buildRenderableLineage` now admits an edge that is ANCHORED (an endpoint is visible) and
      does not REACH OUT (every output endpoint is visible), with job nodes exempt - the fix for
      the earlier defect in which the only edge the estate emits, `output → job:<jobId>`, was the
      only edge both readers dropped. (2) The deterministic sort and cursor pagination above.
      (3) The live provenance inspector fetches the SELECTED output's real bundle and renders
      honest loading / not-recorded / not-authorized / unreachable states rather than the canned
      fixture bundle it used to mount.
      
      TRAP. A MENTION IS NOT A VISIT: 4 e2e specs name a gallery route and 1 navigates to one -
      and that one is the ADMIN gallery (`/isis/output-gallery`), not the member
      `/studio/generation-gallery`. The other three use the route as an entitlement-bounce
      target, a boundary assertion and a provenance smoke string. So .g's
      generation-to-gallery-to-promotion journey has never been walked.
      
      1 of 9 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`. -->
      
  • I7.14 Implement honest fixture, provider-gated, degraded, and offline states in every Isis-related host. Evidence: state matrix.

    • I7.14.a Define canonical data-origin and service-state discriminators for live, fixture/demo, provider-gated, not-configured, degraded, partial, stale, offline-cached, offline-unavailable, and unauthorized. Evidence: state schema.

    • I7.14.b Label fixture/demo data persistently in UI/API/export and prevent its mutation/promotion/release from masquerading as production. Evidence: fixture boundary tests.

    • I7.14.c Present provider-gated state with audience-safe dependency/ configuration reason, available alternatives, and no simulated success. Evidence: provider state tests.

    • I7.14.d Present degraded/partial/stale state per capability/data region with timestamp, missing functions, safe actions, and no global healthy implication. Evidence: degraded state tests.

    • I7.14.e Define offline read cache scope/freshness/classification and disable network-required/unsafe mutations with queued-action policy made explicit. Evidence: offline behavior tests.

    • I7.14.f Exercise every state across Isis operator, Yemaya AAA, Oshun Studio, and gallery with reconnect/recovery and accessible announcements. Evidence: cross-host state matrix. <!-- MEASURED 2026-08-19 (I7.14.a-f). The parent's evidence is a state matrix across four hosts, so the measurement builds one: for each of .a's ten discriminators and each host, whether the state is DECLARED (a union member or a discriminant assignment) and whether it is RENDERED. Instrument: generate-isis-host-state-matrix.mts (+6 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-host-state-matrix.json.

      THE STATES EXIST HUNDREDS OF TIMES OVER AND CANONICALLY ZERO TIMES. .a asks for CANONICAL
      discriminators. No exported union in any of the four hosts names five or more of the ten;
      what the estate has instead is one local discriminated union per component. `'unauthorized'`
      alone is declared independently in 359 Oshun Studio files - `{ status: 'unauthorized';
      reason: string }`, written out again in every workspace, each with its own neighbours and
      its own renderer. Nothing can be checked against anything, which is why .f's cross-host
      matrix has nothing to be a matrix OF: 0 state-matrix artifacts exist in `evidence/` or
      `docs/`.
      
      THE MATRIX, HOST BY HOST. 14 of the 40 (host × state) cells declare their state.
      Oshun Studio 6/10 (live 9, fixture/demo 1, not-configured 3, degraded 13, partial 9,
      stale 7, unauthorized 359), the Oshun BFF 5/10, the ISIS OPERATOR CONSOLE 1/10 (degraded,
      in 2 of its 142 files), and the YEMAYA AAA HOST 0/10 - across 123 files it cannot express
      a single one of the ten states, which is the same answer §I7.9 got from the other
      direction. The gallery, measured as its own directory, is also 0/10. And THREE STATES ARE
      ABSENT FROM ALL FOUR HOSTS: `provider-gated`, `offline-cached` and `offline-unavailable`.
      .c's provider-gated state has no declaration anywhere, so a surface cannot render the
      dependency reason .c asks for; what the estate does instead is fail closed - 264 sources
      answer `not_configured` or a 503 - which refuses honestly and tells the surface nothing it
      can show. That is the right refusal and the wrong vocabulary.
      
      A LABEL TRAVELS WITH THE DATA. .b asks for persistent fixture/demo labelling in UI, API and
      export. 4 sources carry a data-origin discriminant (`source: 'fixture' | 'bff'` in the
      Lilith data layer is the real one); 264 gate a fixture store behind an env flag, which
      HIDES fixture data rather than labelling it. A record that arrives from a gated fixture
      store is indistinguishable from a live one once it is in hand, so nothing downstream can
      prevent its promotion or release.
      
      .e's OFFLINE STORY IS A SERVICE WORKER AND NOTHING IT PROMISES. Yemaya AAA and Oshun Studio
      each ship one service worker; between them 10 sources mention a queued action or background
      sync and 2 declare a cache freshness. The Isis operator console has no service worker at
      all. Neither `offline-cached` nor `offline-unavailable` is a state any host can be in, so
      the read-cache scope, freshness and classification .e asks for have nowhere to be recorded.
      
      TRAPS. A WORD IN A MESSAGE IS NOT A STATE: the census counts a union member or a
      discriminant assignment, never the word inside a sentence the UI prints. A `data-` HOOK
      NAMED FIXTURE IS A TEST HOOK, NOT A DATA-ORIGIN LABEL: allowing the spelling anywhere
      inside an attribute name credited `data-fixture-mount`, `data-texture-fixtures` and
      `data-yemaya-remote-actor-home-capture-white-balance-fixture` — 319 "renders" of a state
      nothing declares. A FILENAME IS NOT A LABEL either: matching `'fixture'` anywhere credited
      every `__fixtures__` helper.
      
      1 of 7 properties present - .b's label check passes only because four sources carry a
      data-origin discriminant at all, and none of them is on an Isis path. ALL SIX CELLS AND THE
      PARENT LEFT `[ ]`. -->
      
  • I7.15 Add keyboard and screen-reader parity to media, graph, and 3D surfaces. Evidence: semantic alternative journeys.

    • I7.15.a Define logical focus order, landmarks, headings, labels, instructions, status announcements, shortcuts, and escape behavior for each complex surface. Evidence: accessibility interaction spec.

    • I7.15.b Provide image/video/audio controls, timelines, regions, annotations, tracks, captions/transcripts, and measurements through native/ARIA semantics and keyboard operation. Evidence: media tests.

    • I7.15.c Provide graph node/port/edge/subgraph/errors/selection/reorder/ connect/delete operations through synchronized tree/list/forms rather than pointer-only canvas. Evidence: graph alternative journey.

    • I7.15.d Provide 3D scene hierarchy/objects/components/transforms/ materials/animation/camera/measurements/annotations through synchronized tree/table/forms. Evidence: 3D alternative journey.

    • I7.15.e Synchronize visual and semantic selection/focus/state without duplicate announcements, keyboard traps, lost focus, or hidden updates. Evidence: focus/state tests.

    • I7.15.f Verify screen reader/browser combinations, keyboard-only, 200– 400% zoom/reflow, contrast, reduced motion, high contrast, and long- content virtualization. Evidence: semantic E2E suite. <!-- MEASURED 2026-08-19 (I7.15.a-f). The parent's evidence is "semantic alternative journeys", so the measurement enumerates the complex surfaces the three hosts actually ship and asks each the same question: is there a non-pointer way to do what the canvas does? Instrument: generate-isis-semantic-parity.mts (+5 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-semantic-parity.json.

      A CANVAS WITH NO KEYBOARD HANDLER HAS NO ALTERNATIVE TO SYNCHRONIZE WITH. 16 complex
      surfaces: 9 media, 6 canvas, 3 three-dimensional (two files are both). Of the 7 canvas and
      3D surfaces, 1 declares a key handler or a `tabIndex` (`InteractiveSkyMap`), and 0 expose a
      tree, grid or table alternative — so .c's "synchronized tree/list/forms rather than
      pointer-only canvas" and .d's 3D hierarchy/transform/material tables have nothing to be
      synchronized WITH, and .e's selection synchronisation has no second selection to
      synchronize. Not one of the three 3D surfaces (Yemaya's `MotionCameraControlPage`, Oshun's
      two Hathor virtual-blocking panels) is keyboard operable at all;
      `StudioIsisGaussianSplattingWorkspace` is 3,034 lines with no `onKeyDown`, no `tabIndex`
      and no `aria-label`.
      
      MEDIA IS THE CLASS WHERE THE ANSWER IS BETTER. 8 of the 9 media surfaces use native
      `controls`, 4 ship a caption track and 2 a transcript — `SessionPlayer` and
      `StudyWorkspaceApp` are the two that answer .b's captions/transcript clause properly. The
      Isis operator console's three media surfaces sit inside real tables, which is why they
      score a semantic alternative where the canvases do not.
      
      .f IS ONE DIMENSION OUT OF FIVE IN TWO OF THREE HOSTS. Every spec in the Isis operator
      suite (13) and the Yemaya suite (1) ends in a WCAG axe analysis, and 554 of Oshun Studio's
      561 do. Beyond axe: keyboard-only 0/0/43, zoom-reflow 0/0/263, contrast 0/0/1, reduced
      motion 0/0/540 across isis-operator / yemaya / oshun-studio. So the two hosts that own the
      3D surfaces verify nothing about keyboard, zoom, contrast or motion.
      
      TRAPS. AN `@example` IN A DOC COMMENT IS NOT A SURFACE: two Yemaya modules — a responsive
      layout helper and a fullscreen hook — carry `<video src="video.mp4" />` inside a `@example`
      block, and matching the raw source counted both as media surfaces. A CHECK IN THE FIXTURE
      COVERS EVERY SPEC THAT IMPORTS IT: Yemaya's whole suite ends in an axe analysis because
      `e2e/fixtures.ts` wraps the base `test`, and reading only the spec files reported 0 axe
      coverage for a suite that is 100% covered.
      
      1 of 6 properties present (.a — 13 docs define a focus order or an accessibility
      interaction spec, though none is the per-surface spec .a asks for by name). ALL SIX CELLS
      AND THE PARENT LEFT `[ ]`. -->
      
  • I7.16 Add responsive review surfaces without forcing full graph/3D editing onto mobile. Evidence: mobile can inspect, comment, approve, and act on incidents.

    • I7.16.a Define wide, compact, and single-column review information/ action hierarchy with mobile scope limited to safe inspection and review operations. Evidence: responsive behavior spec.

    • I7.16.b Build mobile asset/media preview and metadata/gate/provenance/ rights inspector with progressive sections and persistent subject identity. Evidence: mobile inspection tests.

    • I7.16.c Provide graph and 3D semantic tree/list/detail/diff snapshots on mobile without loading or exposing the full authoring editor. Evidence: mobile graph/3D tests.

    • I7.16.d Implement anchored comment create/edit/resolve and attachment/ mention behavior with accessible target context. Evidence: mobile comment journey.

    • I7.16.e Implement approve/reject/request-changes/exception/incident actions only where role/state/evidence permit, with impact, reason, confirmation, and stale-state conflict. Evidence: mobile action tests.

    • I7.16.f Verify touch targets, orientation, safe areas, keyboard, screen reader, zoom, reduced motion, slow/offline/reconnect, and desktop/mobile decision parity. Evidence: responsive Playwright matrix. <!-- MEASURED 2026-08-19 (I7.16.a-f). The parent's evidence is "mobile can inspect, comment, approve, and act on incidents", and .a asks for mobile SCOPE to be limited to safe inspection and review — so the measurement asks the scope question first: does anything narrow what a surface will DO when the viewport is narrow? Instrument: generate-isis-responsive-review.mts (+5 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-responsive-review.json.

      A BREAKPOINT THAT ONLY MOVES THINGS IS NOT A SCOPE. The three hosts lay out responsively —
      101 sources carry a Tailwind breakpoint, a media query or a viewport hook — and 0 of them
      use that knowledge to withhold an action. A phone gets the same verb set as a workstation,
      including the destructive ones, so .a's "mobile scope limited to safe inspection and review
      operations" is not a restriction the code can make. 2 docs describe a responsive review
      behaviour, neither of them the per-surface spec .a asks for.
      
      THE OPERATOR CONSOLE HAS NO REVIEW VERB AT ALL. Of .e's five verbs as CONTROLS — a label
      someone can press, not a word in a file — Isis operator has 0 (approve, reject, request
      changes, exception and incident are none of them a control in its 142 sources), Yemaya AAA
      has 3 (approve 7 files, reject 2, request changes 1) and Oshun Studio has all 5. The
      guarantees .e attaches to those verbs do exist somewhere in the estate — confirmation 147
      sources, reason capture 699, impact preview 68, stale-state conflict 117 — but not as a
      contract any one verb is held to.
      
      .c IS ANSWERED BY .I7.15. That cell measured 0 of the 7 canvas and 3D surfaces exposing a
      tree, grid or table alternative at ANY viewport, so there is no semantic view to serve on a
      narrow one; a mobile graph/3D tree cannot be built out of nothing. .d fares better: 27
      sources carry an anchored-comment identity.
      
      .f IS ONE HOST OF THREE. Oshun Studio's 561 specs set a viewport in 269, exercise
      offline/reconnect in 106, reduced motion in 52 and orientation in 13. The Isis operator
      suite (13 specs) verifies NONE of the five dimensions, and the Yemaya suite (1 spec)
      verifies one (orientation). And TOUCH IS THE DIMENSION NO HOST VERIFIES: 0 specs anywhere
      use `hasTouch`, `tap()` or a touchscreen context, so .f's touch-target clause has never
      been exercised on any surface.
      
      TRAPS. A CATCH BLOCK IS NOT A REVIEW VERB: matching the word `exception` across the three
      hosts hit 384 files — `catch (exception)`, `…Exception` classes, exception-handling prose —
      so the verbs are counted as control labels instead. A BREAKPOINT PREVIEW PICKER IS NOT A
      VIEWPORT SCOPE: allowing the bare word `breakpoint` as the guard credited three Hathor
      authoring workspaces whose match was `data-breakpoint-preview-select … disabled`, a
      `<select>` for previewing breakpoints in a design tool.
      
      3 of 6 properties present (.d's anchored comments, .e's verbs existing somewhere, and .e's
      guarantees existing somewhere). ALL SIX CELLS AND THE PARENT LEFT `[ ]`. -->
      
  • I7.17 Complete parity/coverage monitors for generation types, workflows, models, outputs, provider APIs, SDKs, and UI controls. Evidence: every divergence links to the exact missing contract/route/control and cannot be dismissed without a disposition.

    • I7.17.a Define canonical coverage dimensions/identities/versions for asset kind, operation, workflow, model, provider capability, API/SDK, host/audience, control, output/target, and required evidence. Evidence: coverage schema.

    • I7.17.b Ingest source-of-truth registries, admitted workflow/model/ provider runtime capability, API schemas, SDK surfaces, UI manifests, and automated-test evidence with freshness. Evidence: adapters tests.

    • I7.17.c Compute expected-versus-observed rows and classify missing, incompatible, stale, unauthorized exposure, undocumented, untested, or intentionally unavailable. Evidence: parity fixtures.

    • I7.17.d Link each divergence to exact registry entity, contract field, route/function, SDK method, UI control/host, test, owner, and source revision. Evidence: traceability tests.

    • I7.17.e Require disposition with reason/owner/expiry/evidence as fix, planned, accepted gap, not applicable, or policy-forbidden; prohibit dismissal/deletion. Evidence: disposition tests.

    • I7.17.f Alert on new/regressed/expired/stale divergence and gate releases according to severity without conflating hidden features with missing ones. Evidence: monitor/release tests.

    • I7.17.g Present filters/trends/diffs/ownership and export the same typed evidence accessibly at production scale. Evidence: monitor UI tests. <!-- MEASURED 2026-08-19 (I7.17.a-g). The parent's evidence is that "every divergence links to the exact missing contract/route/control and cannot be dismissed without a disposition", so the measurement asks what the estate's parity monitors actually COMPARE, and what happens to a divergence once one is found. Instrument: generate-isis-parity-monitors.mts (+7 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-parity-monitors.json.

      A PAGE NAMED PARITY IS NOT A PARITY COMPARISON. The operator console ships six pages whose
      names end in "Parity" — Generation, Workflow, Outputs, Models, RunPod and ApiParityMonitor.
      FIVE OF THE SIX COMPUTE NO DIVERGENCE AT ALL: they are API exercisers, forms that call a
      method and show you the response (GenerationParityPage makes 9 live client calls and
      computes nothing; ModelsParityPage, RunPodParityPage and WorkflowParityPage make none).
      
      AND THE ONE THAT COMPARES ASKS ONE SIDE ABOUT ITSELF.
      `buildApiParityDomainSnapshot` decides a method is "available" with
      `typeof resourceRecord?.[method] === 'function'` — a property of the SDK object the page
      already imports — and compares it against `GENERATION_METHODS`, `WORKFLOW_METHODS`,
      `OUTPUT_METHODS`: hand-written string arrays in the page component. Both sides of the
      comparison ship in the same bundle. Remove a route from the generation service and neither
      side changes; coverage still reads 100%. Nothing reads an OpenAPI document, a registry or a
      runtime capability, which is .b's entire subject.
      
      AN EXPECTATION OF NOTHING IS NOT FULL COVERAGE: `formatCoverage` returns 100 when
      `totalCount <= 0`, so a dimension nobody has described reads as fully covered.
      
      WHAT THE MONITOR CAN SAY. 5 of .a's 10 coverage dimensions are named anywhere in it
      (workflow, model, API/SDK, control, output/target — and that is a generous floor, counting
      a dimension the monitor merely NAMES). 1 of .c's 7 divergence classes is emitted
      (`missing`); incompatible, stale, unauthorized exposure, undocumented, untested and
      intentionally unavailable have no representation. 2 of .d's 8 trace targets are linked
      (route and SDK method); a divergence names no registry entity, contract field, UI control,
      covering test, owner or source revision. Of .e's five dispositions, `accepted gap` and
      `policy-forbidden` exist as values NOWHERE in the estate, so a divergence cannot be
      dispositioned in the two ways that would otherwise let it survive honestly — and there is
      no prohibition on dismissal because there is nothing to dismiss it INTO. No release gate
      reads a divergence, and the page exports nothing.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The live
      probe beside the comparison is real: it calls `generation.listJobs` and `workflows.list`
      for true and reports healthy / degraded / unavailable with the actual error, never a
      fabricated verdict. (2) `buildApiParityDriftEvents` compares a previous snapshot with a
      current one and names the methods that were lost or regained rather than only a count, at
      three severities.
      
      TRAP. A WORD INSIDE A METHOD NAME IS NOT A DIMENSION: without word boundaries, `operation`
      matched inside the SDK method name `getProviderOperations`.
      
      0 of 9 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`. -->
      
  • I7.18 Complete RunPod fleet and economics operations for endpoint classes, model/node images, warm/cold policy, GPU topology, cache, canary, drain/rollback, SLO, and cost attribution. Evidence: live topology and controlled failover journey.

    • I7.18.a Define endpoint class/pool/worker/GPU/image/model/node/cache/ lifecycle/capacity/cost/SLO contracts and ingest authenticated live topology. Evidence: fleet schema/integration tests.

    • I7.18.b Manage signed/scanned/pinned model-node images and exact endpoint compatibility, rollout state, dependencies, and supply-chain evidence. Evidence: image admission journey.

    • I7.18.c Configure warm/min/max/idle/cold-start/autoscale/concurrency/ region/GPU topology policy with capacity/cost/SLO preview and bounds. Evidence: scaling policy tests.

    • I7.18.d Inspect cache layers/keys/contents/size/hit/miss/eviction/ integrity/warmup by authorized scope without exposing tenant data. Evidence: cache operations tests.

    • I7.18.e Roll out canaries with bounded traffic/capacity, comparison metrics, abort thresholds, observation window, and automatic/manual promotion rules. Evidence: canary journey.

    • I7.18.f Execute drain/failover/rollback with in-flight policy, reservation migration, health/capability validation, immutable decisions, and post-action verification. Evidence: controlled failover journey.

    • I7.18.g Present SLOs, queue/runtime/reliability/cold-start/resource efficiency, complete fixed/variable/egress/cache costs, allocation, and variance by class/pool/workflow/model/tenant. Evidence: economics tests.

    • I7.18.h Enforce operator roles, two-person high-impact controls, confirmation, rate limits, concurrency, emergency stop, audit, and accessible large-fleet UI. Evidence: abuse/operations suite. <!-- MEASURED 2026-08-19 (I7.18.a-h). The parent's evidence is a "live topology and controlled failover journey", so the measurement asks the binding question first: is there an authority to read a topology FROM? Instrument: generate-isis-runpod-fleet.mts (+6 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-runpod-fleet.json.

      THE ANSWER IS AN HONEST NO, WRITTEN DOWN — and this cell records it as CORRECT rather than
      as a gap in the code. `apps/oshun/bff/src/isis/runpod-endpoints-store.ts` is an explicitly
      unbound projection: every field `null` or `[]`, `readonly never[]` row types so a caller
      cannot push a fabricated endpoint past the compiler, and a frozen
      `RUNPOD_ENDPOINTS_DISPOSITION` declaring `mode: 'unbound_empty_projection'`,
      `authorityBinding: 'not_configured'`, `authoritativeState: false` and all 8 faculties —
      endpoint inventory, runtime health, tenant cost, budget envelopes, queue telemetry,
      failover history, secret-rotation authority and mutations — `unavailable`. That is the
      honest seam the quality bar asks for. .a's live topology is therefore not ingested, and
      the estate says so instead of inventing a fleet.
      
      FIXED: A DELIBERATE EMPTY PROJECTION READ AS A MALFORMED RESPONSE. The admin loader's
      `isContextShape` required `typeof record.operatorId === 'string'` with a NON-EMPTY value
      and `typeof record.nowUnixSeconds === 'number'`, and the unbound projection sends `null`
      for both. Driven with the exact body the route serves: `{ok:false,
      reason:'malformed-response'}` — on every request. The page then rendered a red alert
      telling the operator the BFF had misbehaved, when the BFF had answered exactly as
      designed. The loader also DROPPED the disposition it was sent, so nothing downstream could
      explain the empty tables. The guard now accepts the nulls the producer sends, the
      disposition is carried through, and the page renders an honest "no RunPod operations
      authority is bound to this deployment" notice naming the unavailable faculties — instead
      of a failure alert or a silent empty grid. `isis-runpod-endpoints-loader.spec.ts` (+6)
      drives the route's own body and keeps the genuinely-malformed and forbidden cases refusing.
      
      WHAT THE FLEET SURFACES CAN NAME: 7 of the 21 terms .a-.h operate on appear anywhere
      across the five surfaces (store, route, loader, page, panel) — GPU, SLO, cold start, cache,
      failover, rollback and cost. Nothing names an endpoint class, a pool, capacity, a
      model-node image, a signature/scan/pin, a warm or idle policy, autoscale bounds,
      concurrency, cache eviction, a canary, a drain, egress cost, or a two-person control. So
      .b's image admission, .e's canary rollout and .h's two-person high-impact control have no
      vocabulary at all, let alone a journey.
      
      THE FAILOVER JOURNEY IS REAL, AND THERE IS NO FLEET BEHIND IT. One admin e2e navigates to
      `/isis/runpod-endpoints`, opens the controls tab, asserts `record-failover` is DISABLED,
      fills a rationale, asserts it becomes enabled, clicks it and asserts the entry appears in
      the failover list. That is a properly gated, audited control — acting on an endpoint
      inventory the disposition declares unavailable.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The unbound
      projection above. (2) The admin route requires an authenticated caller with an Isis
      workspace scope (`admin:*`, `admin:studio` or `admin:workspace:isis`) before serving even
      the empty projection. (3) The panel treats a missing average runtime as ABSENT and falls
      back to the job's own estimate rather than reporting zero.
      
      TRAP. A MENTION IS NOT A JOURNEY, and a visit is not one either: the census requires a spec
      to navigate to a RunPod route AND exercise a drain, failover or rollback before it counts.
      
      3 of 11 properties present. ALL EIGHT CELLS AND THE PARENT LEFT `[ ]`: the fleet has no
      authority bound, so .a-.g have nothing to operate on, and .h's two-person control does not
      exist. -->
      
  • I7.19 Complete model discovery and supply workflows for registry search, Civitai intake, collections, license/model cards, malware scanning, comparison, quality evaluation, model merging, fine-tuning/LoRA, promotion, quarantine, and revoke. Evidence: no external model or derivative reaches production without a complete evidence chain.

    • I7.19.a Build governed internal/external model discovery with allowlisted sources, authenticated metadata, filters, stable source/version IDs, snapshots, and no automatic admission. Evidence: discovery tests.

    • I7.19.b Intake Civitai/other approved artifacts by immutable download hash, source snapshot, uploader, files/dependencies, license/model card, training provenance, tags/triggers, and quarantine state. Evidence: external intake journey.

    • I7.19.c Build versioned curated collections with membership rationale, audience/use/target, compatibility, lifecycle, owner, and no lifecycle authority over member models. Evidence: collection tests.

    • I7.19.d Run malware/archive/format/serialization/signature/dependency/ vulnerability scans in isolated bounded processing and retain results. Evidence: hostile artifact suite.

    • I7.19.e Compare models/versions on rights, provenance, compatibility, quality/safety calibration, latency/cost/resources, outputs, and uncertainty using pinned fixtures. Evidence: comparison journey.

    • I7.19.f Define model merge inputs/weights/method/config/tool/environment/ license compatibility, generate a new artifact identity, and rerun full admission evidence. Evidence: merge lineage tests.

    • I7.19.g Define fine-tune/LoRA dataset/splits/consent/rights/filtering, base model, method/config/seed/compute, checkpoints/evaluations, and new artifact lineage. Evidence: training journey.

    • I7.19.h Promote exact admitted versions only after complete license/ provenance/security/quality/safety/compatibility approvals; quarantine/ revoke with blast radius and fail-closed routing. Evidence: lifecycle journey. <!-- MEASURED 2026-08-20 (I7.19.a-h). The parent's evidence is that "no external model or derivative reaches production without a complete evidence chain", so the measurement asks the binding question first: is there a path by which an external model reaches production at all, and does anything on that path require the chain? Instrument: generate-isis-model-supply.mts (+13 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/isis-model-supply.json.

      THE SUPPLY BOUNDARIES ARE HONESTLY UNBOUND AND SAY SO — recorded as CORRECT, not as a gap.
      All three (`civitai-intake-store.ts`, `model-merging-store.ts`, `lora-training-store.ts`)
      are explicit empty projections: `readonly never[]` rows so a caller cannot push a fabricated
      candidate past the compiler, and a frozen disposition per store declaring
      `mode: 'unbound_empty_projection'`, `authorityBinding: 'not_configured'`,
      `authoritativeState: false` and every faculty `unavailable` — the LoRA store names 8 of them
      (run inventory, dataset rights, execution, checkpoint telemetry, evaluation authority,
      promotion authority, lifecycle history, mutations). So .b's external intake, .f's merge
      execution and .g's training runs are not performed, and the estate says so instead of
      inventing an intake queue.
      
      FIXED: TWO READERS REJECTED THE PROJECTION THEIR OWN PRODUCER SENDS — the §I7.18 defect
      again, in two more routes. `isis-civitai-intake-loader.ts` and `isis-model-merging-loader.ts`
      both required `typeof record.operatorId === 'string'` with a NON-EMPTY value; both stores
      send `operatorId: null` because no reviewer directory resolves one. Driven with the exact
      bodies the two routes serve: `{ok:false, reason:'malformed-response'}` from both, on every
      request. Each page then rendered a red alert telling the operator the BFF had misbehaved,
      when the BFF had answered exactly as designed, and both loaders DROPPED the disposition, so
      nothing downstream could explain the empty queue or the empty component catalog. Both guards
      now accept the null the producer sends, both carry the disposition through, and both pages
      render an honest "no external model intake authority / no model-merging authority is bound
      to this deployment" notice naming the unavailable faculties. `isis-civitai-intake-loader.spec.ts`
      (+7) and `isis-model-merging-loader.spec.ts` (+7) drive the routes' own bodies and keep the
      genuinely-malformed, unauthenticated, forbidden and http cases refusing. The LoRA loader was
      already correct — it models the projection exactly via `isExactDisposition` — and is the
      pattern the other two now follow.
      
      A CATALOG IS NOT A GATE. `getModelService()` is called from exactly two non-test files in
      the generation API — `routes/models.ts` and the service itself — and ZERO of them are on the
      generation path. Nothing between a submitted job and a provider call asks the model registry
      whether an artifact was admitted, so .h's "fail-closed routing" has no site to fail closed
      at. The registry is a catalog the operator browses, not a gate generation passes through.
      
      A DEPRECATION IS NOT A REVOCATION. .h asks for quarantine and revoke; the canonical admin
      model-governance contract (`libs/contracts/src/common/admin-model-governance.ts`) can express
      NEITHER. Version statuses are production/staging/preview/deprecated/archived; promotion
      events are review_opened/review_approved/review_rejected/promoted/rolled_back/
      emergency_override; review states are not_required/pending/approved/rejected; rollout
      statuses are not_started/in_progress/completed/rolled_back. 0 of 2 withdrawal verbs are
      expressible in any of the four unions. `deprecated`/`archived` are voluntary retirement and
      `rolled_back` is a rollout state; an operator who discovers a poisoned artifact has no state
      to put it in.
      
      A REVIEW DECISION THAT NEVER LEAVES THE BROWSER IS NOT A REVIEW. `IntakeQueuePanel.tsx`
      wires 8 click handlers — approve, reject, conditional, takedown, denylist and three
      selection controls — and makes 0 network calls. Every decision calls `applyReviewDecision`
      and advances React state; `POST /v1/admin/isis/civitai-intake/decisions` exists, is
      authenticated and scope-checked, and answers 503 `intake_authority_not_configured` — and the
      panel never calls it. The panel then prints "OK for runtime" and "· staged" off that local
      state. `ModelMergingPanel.tsx` is the other shape: its one mutation control, "Save as
      managed model", has 0 click handlers at all, and `summarisePreview` — the A/B-against-a-
      fixture-set the panel's own doc comment claims — has 0 product callers, so the fixture-set
      selector selects nothing.
      
      A CONSUMER THAT NOTHING REACHES IS NOT REACH. .c's collections, .e's comparison and quality,
      .f's merging and .g's LoRA training are implemented — 13 modules, 23,393 lines, under
      `libs/isis/ai-providers/src/providers/civitai/`, every one of them published on the
      `@isis/ai-providers` barrel. Product code imports 29 runtime bindings from that package
      across 40 files, and 0 of the 29 come from any of the 13. Counted from import bindings whose
      SPECIFIER is this package, never from bare identifier matches: `interpolate`, `formatBytes`
      and `calculateOverallScore` are declared independently all over this monorepo, and a name
      census reports them as reach.
      
      A COMPILED SHADOW BESIDE ITS SOURCE WINS THE RESOLVER. All 13 of those modules have a
      tracked `.js` and `.d.ts` committed next to the `.ts`. Driven through Vite's own resolver
      against that directory: `./lora-training-provider` and four siblings resolve to the `.js`,
      because Vite's default extension order puts `.js` before `.ts`. Vitest uses that resolver,
      so the five provider tests in the directory grade the compiled snapshot rather than the
      source. 10 of the 13 `.js` were last regenerated BEFORE their `.ts` was last committed, and
      at least one diverges behaviourally: `model-quality-provider` emits trend entries as
      `{date,count}` in the `.js` and `{date,value}` in the `.ts`. Recorded, not removed:
      deleting 26 tracked build outputs out of a published package is a decision about that
      package's build contract, not a repair this cell is entitled to make.
      
      FLAGGED, NOT FIXED: .e's quality instrument manufactures its measurements.
      `ModelQualityProvider.ensureDownloadMetrics`/`ensureGenerationMetrics` build a 31-day series
      from `createHistory(modelId * 17 + 300, 31, 12)` and `createHistory(modelId * 31 + 1200, 31,
      28)` — a sine wave over the subject's own identifier — and the object around it carries 15
      literal rates and shares (`successRate: 97/95/93/91`, `percentage: 55/45/40/60/45/30/15/10`,
      `totalModelsInCategory: 100`) presented as telemetry, with no `not_configured` seam. It has
      0 product consumers, so nothing displays these today; it is recorded here rather than
      rewritten because .e has no bound quality authority to rewrite it against.
      
      WHAT AN INTAKE RECORD CAN CARRY: 3 of .b's 9 subjects. `CivitaiIntakeCandidate` and
      `ManagedModelRecord` carry the immutable download hash, the uploader (as
      `creatorAttribution`) and the license tag. They carry no source snapshot, no files or
      dependencies, no model card, no training provenance, no tags or trigger words, and no
      quarantine state.
      
      WHAT THE ADMISSION PATH SCANS: 2 of .d's 7 kinds. `model.service.ts` runs a malware scan and
      a format check (filename-matches-format plus a content-type allowlist); there is no archive,
      serialization, signature, dependency or vulnerability scan. Bounded yes — `assertAllowedUploadSize`
      — and isolated only in the sense that the one malware-capable engine is ClamAV, a separate
      daemon.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The three
      unbound projections above. (2) The model upload-complete path downloads the object,
      re-computes the digest over the bytes and refuses on a mismatch, rather than recording the
      hash the caller asserted; it also refuses on a size mismatch and on a disallowed content
      type. (3) Malware scanning is fail-closed on an engine that cannot see malware (the §I7.3
      fix): `MALWARE_CAPABLE_SCAN_ENGINES = {clamav}`, so an incapable engine refuses the upload
      instead of recording a clean verdict. (4) Both admin supply routes require an authenticated
      caller with an Isis workspace scope before serving even the empty projection. (5)
      `buildProvenanceBundle` in `@isis/lora-training-surface` computes a real SHA-256 over the
      canonical component set — .f's new artifact identity is honest arithmetic, it simply has
      nowhere to be written.
      
      TRAP. A WORD IN A COMMENT IS NOT A SCAN: `pickle` appears on the admission path exactly once,
      inside a doc comment explaining what the builtin engine MISSES, and an unstripped census
      reported a serialization scan that does not run. The stripper tracks string and template
      literals, because the naive version eats code. TRAP. A REFUSAL JOURNEY IS NOT AN ADMISSION
      JOURNEY: `isis-lora-training-boundary.spec.ts` navigates `/isis/lora-training`, exercises
      `promote`, and asserts 503 with 0 controls rendered — a properly built proof that the
      boundary refuses, and the opposite of the evidence the parent asks for. 6 e2e specs name a
      supply surface, 2 navigate to one, 1 exercises a decision, and 0 carry an artifact through a
      decision that succeeds.
      
      3 of 16 properties present. ALL EIGHT CELLS AND THE PARENT LEFT `[ ]`: no external supply
      authority is bound, nothing on the generation path checks admission, and the governance
      vocabulary cannot express the withdrawal .h requires. -->
      
  • I7.20 Complete workflow and asset marketplaces only for authorized audiences, including submission, review, compatibility, versioning, pricing/entitlement, creator attribution, abuse, takedown, and payout reconciliation. Evidence: marketplace and creator-monetization journeys cannot bypass generation or release gates.

    • I7.20.a Define audience/host/region/tenant marketplace enablement and separate workflow-template, model/modifier, preset, source asset, and generated asset item contracts. Evidence: authorization matrix.

    • I7.20.b Implement creator enrollment/identity/tax/payout/rights authority, item submission, immutable files/hashes/dependencies, metadata, license, pricing, and attribution. Evidence: creator submission journey.

    • I7.20.c Run type-specific security/rights/consent/provenance/quality/ safety/compatibility/generation/release gates and independent human review before listing. Evidence: review gate tests.

    • I7.20.d Version items immutably with compatibility matrices, dependency locks, migration/deprecation/revoke behavior, consumer notices, and historical entitlement. Evidence: version lifecycle tests.

    • I7.20.e Implement offers/prices/currencies/taxes/fees/refunds/promotions, entitlements/seats/terms/regions, purchase, receipt, and server-side use authorization. Evidence: commerce tests.

    • I7.20.f Preserve creator/source/version attribution through download/use/ derivative/generation/release and calculate royalty/payout ledger from verified events. Evidence: attribution journey.

    • I7.20.g Detect/report abuse, malware, fraud, stolen content, impersonation, review manipulation, license breach, and policy violations with case evidence and anti-retaliation controls. Evidence: abuse tests.

    • I7.20.h Execute takedown/quarantine/recall/appeal/restore with blast radius, consumer state, downloads/caches, payouts/holds, notifications, and immutable decisions. Evidence: takedown journey.

    • I7.20.i Reconcile order/refund/fee/tax/royalty/hold/payout/chargeback per item/version/creator/currency and expose discrepancies without double payment. Evidence: payout reconciliation tests.

    • I7.20.j Verify direct API/deep-link/import entitlement bypass, cross- tenant access, gate bypass, creator self-approval, accessibility, and marketplace scale. Evidence: monetization E2E suite. <!-- MEASURED 2026-08-20 (I7.20.a-j). The parent's evidence is that "marketplace and creator-monetization journeys cannot bypass generation or release gates", so the measurement asks the binding question first: is there a marketplace journey at all — can an item be listed, bought, or turned into an entitlement? Instrument: generate-oshun-marketplace-surface.mts (+9 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-marketplace-surface.json.

      THE FIVE MARKETPLACE ROUTES ARE CALCULATORS, NOT A MARKETPLACE. Three validate a listing
      draft the caller submits (asset, ai-model, plugin), one filters a `listings` ARRAY the
      caller submits in the request body, and one computes a collaborator revenue split for
      collaborators the caller submits. All five are request-time pure functions; 0 of 5 persist
      anything, there is no catalog, no order, no receipt and no payout ledger anywhere in the
      estate. So the parent's evidence HOLDS VACUOUSLY: no purchase can bypass the generation
      gate because no purchase exists. Recording the difference matters — a later reader must not
      take a green parent for a marketplace that was built and then locked down.
      
      FIXED: TWO LISTS FOR ONE JUDGEMENT, AND THE ONE THAT RUNS WAS MISSING A PERMISSION THE
      OTHER CALLS DANGEROUS. `@yemaya/marketplace`'s own submission checker named
      `{system:process, filesystem:write, network:websocket}` dangerous;
      `apps/oshun/bff/src/yemaya/plugin-manifest-store.ts` — the validator an operator actually
      reaches at `POST /v1/admin/yemaya/plugin-marketplace/validate` — hand-copied a DIFFERENT
      set, `{filesystem:write, system:process, system:env, crypto:sign}`. The two disagreed on 3
      of 5 members: a manifest requesting `network:websocket` cleared the validator that runs
      with zero warnings, and one requesting `system:env` or `crypto:sign` cleared the library's
      checker. Neither side imported the other; the store also hand-copied the 18-member
      permission union and the 7-member type union out of a contract it only cited in a comment.
      `HIGH_RISK_PLUGIN_PERMISSIONS` and `PLUGIN_PERMISSIONS` are now published from
      `@yemaya/marketplace` and read by BOTH — the high-risk set is the union of the two prior
      lists, because each side had independently judged its members to warrant review and
      dropping either side's would be a downgrade. `plugin-permission-vocabulary.spec.ts` (+6)
      drives the library's checker over every member and holds the union closed;
      the route test (+1) drives `network:websocket` through the BFF and asserts the warning.
      
      FIXED: A FIELD-SHAPE VERDICT IS NOT PUBLISH APPROVAL. The three validators answered
      `{valid: true, errorCount: 0, warningCount: 0, issues: []}` from routes documented as
      "publish-readiness", with nothing at all saying which gates were still owed — no security,
      rights, consent, provenance, quality, safety, compatibility, generation or release gate had
      run, no human had reviewed it, and nothing had been listed, priced, sold or entitled. Each
      store now publishes a frozen `*_VALIDATION_DISPOSITION` naming all ten of .c's reviews and
      gates `unavailable` alongside `listingPublication`, `pricingAuthority`, `entitlementGrant`
      and `mutations`, and each route serves it BOTH on the catalog and beside every verdict, so
      `valid: true` never travels alone. Three route tests (+1 each) assert the pairing.
      
      A CONSUMER THAT NOTHING REACHES IS NOT REACH. `libs/isis/3d-marketplace-ops/src` holds 25
      modules that answer .c-.i almost name for name — `moderation-surface-policy`,
      `policy-enforcement`, `publisher-verification`, `rights-declaration`, `listing-lifecycle`,
      `listing-versioning`, `listing-commercial-terms`, `chain-of-custody`,
      `listing-abuse-reporting`, `duplicate-detection`, `leak-tracing`, `trust-scoring`,
      `takedown-intake`, `listing-hold-policy`, `payout-ledger`, `refund-dispute-workflow`,
      `seller-payout-readiness`, `sales-analytics`. Product code imports exactly ONE runtime
      binding out of all 25: `createCollaboratorRevenueSplitPlan`. Counted from import bindings
      whose SPECIFIER is the package — `getRecipe`, `checkCompatibility` and `interpolate` are
      declared independently across this monorepo and an identifier census reports them as reach.
      
      WHAT THE REACHABLE SURFACES CAN NAME: 2 of .a's 5 item contracts (model/modifier and source
      asset; no workflow template, no preset, no generated asset — and `plugin`, which .a does
      not name, is the third that exists). 0 of 4 enablement dimensions — audience, host, region
      and tenant appear nowhere, so .a's authorization matrix has no axes. 1 of 9 commerce nouns
      (`entitlement`; no offer, tax, fee, refund, promotion, seat, purchase or receipt). 1 of 10
      lifecycle verbs (`payout`, and only as a disposition faculty declaring itself unavailable;
      no deprecate, revoke, abuse report, takedown, quarantine, recall, appeal, restore or
      chargeback).
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The
      generation creator tier is resolved from the AUTHENTICATED token's scopes; the
      `x-oshun-creator-tier` header elevates only outside production, and 0 marketplace nouns are
      read in `generation-admission.ts`. (2) `parseEntitlementsFromRequest` likewise honours the
      `x-oshun-tier` header only outside production, and per-domain overrides — which can
      ELEVATE — are dev-only while suspensions, which only restrict, are honoured everywhere.
      (3) Every marketplace route requires an authenticated caller with a workspace admin scope
      before answering. (4) The creator-monetization planner declares itself an ephemeral
      request-time pure function with plan persistence, rights authority, listing publication,
      payout activation and settlement all unavailable, and collaborator identity `unverified`
      and terms acceptance `self_asserted_unverified`. (5) The unified search says in its own
      summary that it searches the listings the caller supplies — it does not present itself as a
      catalog.
      
      TRAP. A WORD IN A COMMENT IS NOT A CAPABILITY: every one of these surfaces describes in
      prose what a marketplace would do, so an unstripped census reports takedowns and payouts no
      line of code performs. TRAP. A ROUTE HELD IN A CONST IS STILL A ROUTE: the monetization
      suite interpolates `const ROUTE = '/studio/isis/creator-monetization'` into `page.goto`,
      and a literal-only match reported 0 visits to a page it walks in depth. 3 e2e specs name a
      marketplace surface, 1 navigates to one, and 0 exercise a purchase or entitlement grant.
      
      2 of 14 properties present. ALL TEN CELLS AND THE PARENT LEFT `[ ]`: nothing is listed, so
      nothing can be reviewed, versioned, priced, attributed, taken down or reconciled. -->
      
  • I7.21 Complete operator security and financial controls for entitlements, quotas, budgets, cost optimization, chargeback fraud, suspicious activity, account protection, anomaly triage, feature flags, and emergency kills. Evidence: every recommendation is explainable and every mutation is authorized, audited, reversible where possible, and tested against abuse.

    • I7.21.a Build effective-entitlement inspector/change workflow with source, scope, conditions, expiry, conflicts, preview, reason, approval, propagation, and revocation. Evidence: entitlement journey.

    • I7.21.b Build quota/budget inspector and bounded change workflow for hierarchy, window, usage/reservations, forecasts, thresholds, alerts, exceptions, and effective date. Evidence: quota/budget journey.

    • I7.21.c Generate cost-optimization recommendations from attributed versioned measurements with assumptions, confidence, projected savings/ quality/latency/capacity impact, alternatives, and no automatic apply. Evidence: recommendation explainability tests.

    • I7.21.d Detect chargeback/payout/refund/purchase/usage fraud signals, preserve event/feature/rule/model versions, avoid protected-attribute inference, and route to human cases. Evidence: fraud triage tests.

    • I7.21.e Build suspicious-activity/anomaly triage for auth, data access, model/workflow intake, dispatch, spend, provider, marketplace, and release with baseline/threshold/uncertainty. Evidence: anomaly journey.

    • I7.21.f Implement account/session/API-key/device protection, step-up authentication, lock/recovery, session revoke, and high-risk action confirmation without exposing secrets. Evidence: account abuse tests.

    • I7.21.g Govern feature flags by typed scope/audience/environment/ prerequisite/owner/expiry, staged rollout, evaluation, rollback, and audit; prohibit flags from bypassing code floors. Evidence: flag tests.

    • I7.21.h Implement scoped emergency kills for route/workflow/model/ provider/tool/marketplace/release with two-person authority where required, immediate propagation, in-flight policy, and restore gate. Evidence: kill-switch game day.

    • I7.21.i Make every mutation previewable, permission/state checked, reasoned, confirmed, optimistic-concurrency safe, immutable in audit, and reversibly compensated where possible. Evidence: control invariant suite.

    • I7.21.j Test insider, confused-deputy, mass-assignment, replay, race, enumeration, cross-tenant, cost-evasion, fraud, and emergency-abuse paths plus accessible operator recovery. Evidence: adversarial E2E suite. <!-- MEASURED 2026-08-20 (I7.21.a-j). The parent's evidence is that "every mutation is authorized, audited, reversible where possible, and tested against abuse", so the measurement enumerates the mutations first and asks the invariants of each. Instrument: generate-oshun-mutation-invariants.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-mutation-invariants.json.

      A CENSUS IS ONLY AS GOOD AS ITS EXTRACTOR, AND THIS ONE WAS WRONG THREE WAYS BEFORE IT WAS
      RIGHT. The first pass reported 28 mutations, 4 of them audited, and no emergency kill
      switch anywhere in the estate. All three numbers were artefacts of the reader.
      (1) A ROUTE BUILT FROM A TEMPLATE LITERAL IS STILL A ROUTE — `admin-agentic-operations.ts`
      registers its kill switch as `${prefix}/agentic-operations/kill-switches` through a helper
      called with '/admin' and '/v1/admin', and a literal-only extractor cannot see it.
      (2) A HANDLER PASSED BY NAME IS STILL THE HANDLER — those routes pass `killSwitchHandler`
      by identifier, so every invariant measured on the registration call reads absent.
      (3) A GUARD HELD IN A HELPER IS STILL A GUARD, TRANSITIVELY — `requireScenePublish`
      delegates to `hasScenePublishScope`, and a one-level scan calls its handlers unscoped.
      Resolving all three moved the population to 81 and the audited count to 52. Two more traps
      were needed to keep it from over-counting: a `.push()` ON A LOCAL ARRAY IS NOT A MUTATION
      (371 of the 407 admin POST handlers are calculators), and a `reason` FIELD IN AN ERROR
      REPLY IS NOT AN OPERATOR'S RATIONALE (`reason: 'missing_auth_context'` matched every
      handler in the estate).
      
      THE ANSWER, ON 81 DURABLE ADMIN MUTATIONS: 81 authorized, 52 audited, 12 reasoned, 8
      optimistic-concurrency safe, 0 previewable. The authorization half of the parent's claim
      holds outright — every one runs behind an authentication preHandler and checks an admin
      scope in the handler. The rest does not.
      
      THE DECISIONS THAT MOST NEED A RECORD ARE THE ONES WITHOUT ONE. The 29 unaudited mutations
      are not a random tail. They are every consent act (`aja/consent/grant`,
      `aja/content-security/grants` and its revoke, `rail/ori/consents/:userId/:scope/revoke`),
      every moderation act (`moderation/queue`, `.../:itemId/decide`, `appeals/:appealId/resolve`,
      `bans/:banId/revoke`), the appeal decision in `themis`, all six Isis ledger writes and both
      reference-bank writes, the model-governance upload, the output-gallery bulk action and the
      crisis-suppression toggle. Meanwhile the communications, integrations and
      notification-template CRUD is audited end to end. An operator can grant a consent, decide a
      moderation case and revoke a ban, and the estate keeps no record that they did.
      
      FIXED: AN INVITE ID IS NOT A TENANT SCOPE. `POST /v1/admin/tenant-console/members/invite/
      :inviteId/revoke` looked its subject up by id in the store's global invite map and revoked
      whatever it found, with no check that the invitation belonged to the caller's tenant and no
      audit event — while the sibling create in the same file resolves the acting tenant and
      writes `tenant.member_invite`. So a tenant-console admin acting in one tenant could
      withdraw another tenant's pending invitation and leave no trace of having done so. The
      store's `revokeInvite`/`revokeInviteDurably` now require the tenant and answer 404 for an
      invitation belonging to a different one — indistinguishable from an id that does not exist,
      so the boundary does not leak which ids are real. The route resolves the tenant like its
      sibling and records `tenant.member_invite_revoked` with the actor. Driven end to end:
      the route spec (+1) creates invitations in two tenants, gets 404 for the cross-tenant
      revoke, confirms the other tenant's invitation is still pending, revokes its own, and reads
      exactly one revocation event back out of the audit store; the store spec (+1) holds the
      same boundary at the unit level.
      
      A PROMPT COMPRESSOR IS NOT A COST RECOMMENDATION. .c asks for recommendations carrying
      assumptions, confidence, projected savings, latency and capacity impact, and alternatives.
      The one cost surface an operator can reach, `cost-optimization-store.ts`, is a stateless
      token-budget optimiser: it runs prompt text through a real compression technique and
      returns a compression ratio and a quality score. 0 of .c's 6 explainability subjects appear
      on it. It is honest about what it is — `mode: 'stateless_request_engine'`,
      `authoritativeState: false` — it simply is not the thing .c describes.
      
      WHICH CONTROL WORKFLOWS EXIST: 1 of 5. There is a real emergency kill switch (.h), audited
      and flushed before the reply. There is NO mutating route for an entitlement change (.a), a
      quota or budget change (.b), or a feature flag change (.g) — the three routes whose paths
      carry those words (`studio/experimentation-flags/evaluate`,
      `studio/performance-budgets/evaluate`, `yemaya/budget-management/optimize`) are evaluators
      that write nothing.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) All 81
      durable admin mutations are authenticated and admin-scoped; there is no unauthenticated
      write in the admin surface. (2) The agentic-operations kill switch, gold-set promotion and
      champion-challenger rollout each record an audit event and await a persistence flush before
      replying. (3) The tenant-console invite create resolves the acting tenant and records a
      tenant-scoped audit event, which is why its revoke's silence was visible at all.
      
      2 of 10 properties present. ALL TEN CELLS AND THE PARENT LEFT `[ ]`: authorization holds,
      auditing covers 52 of 81, nothing is previewable, and four of the five control workflows
      the cells describe have no route to mutate through. -->
      
  • I8.1 Define review templates by asset kind, risk, audience, and delivery target. Evidence: required roles and quorum matrix.
    • I8.1.a Define immutable review-template identity/version/lifecycle, applicability selectors, required review dimensions, roles, quorum, sequence/parallelism, deadlines, escalation, and code floors. Evidence: review template schema.

    • I8.1.b Map every publishable asset kind/profile to the applicable visual/audio/3D/editorial/technical review dimensions. Evidence: asset coverage matrix.

    • I8.1.c Add risk-class requirements for child/identity/voice/biometric/ sensitive content, protected sources, high-cost, high-reach, and novel workflows. Evidence: risk matrix.

    • I8.1.d Add audience/host/use/territory/channel and delivery-target- specific reviewer roles, evidence, and acceptance criteria. Evidence: audience/target matrix.

    • I8.1.e Validate separation of requester/creator/reviewer/approver and quorum independence, conflicts, substitutions, and abstention. Evidence: role/quorum policy tests.

    • I8.1.f Resolve the exact template/version deterministically at candidate creation and invalidate eligibility when applicability changes. Evidence: resolution/freshness tests. <!-- MEASURED 2026-08-20 (I8.1.a-f). The parent's evidence is a "required roles and quorum matrix", so the measurement asks the binding question first: can a review template in this estate name a required role, or a number of them? Instrument: generate-oshun-review-templates.mts (+9 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-review-templates.json.

      TWO REVIEW-TEMPLATE CONTRACTS, AND NEITHER CAN. `admin-review-stage.ts`'s
      `AdminReviewTemplate` carries six fields — templateId, artifactClass, label, description,
      requiredSections, defaultChecklist. `review-package.ts`'s `ReviewTemplateSchema` carries
      five — templateId, label, description, artifactTypes, requiredStageTypes. 3 of .a's 11
      subjects are expressible across BOTH of them: an immutable identity, an applicability
      selector and a set of required dimensions. There is no version, no lifecycle, no role, no
      quorum, no sequence or parallelism, no deadline, no escalation and no code floor on either.
      A CHECKLIST IS NOT A QUORUM: the template says what to look at, never who must look or how
      many of them.
      
      THE VOCABULARY EXISTS AND IT ATTACHES SOMEWHERE ELSE. `ReviewRoleSchema` enumerates eleven
      review roles — collaboration_lead, program_operator, domain_integrator, security_reviewer,
      compliance_auditor, moderator, editor, support_operator, privacy_officer, release_manager,
      reviewer — and the estate validates that a delegation or escalation actually CHANGES role.
      All three consumers are `libs/shared/review-persistence` modules: the roles attach to stage
      assignments and to recorded decisions, never to a template's requirement. So an operator
      can record which role decided, and no template can say which role had to.
      
      FIXED: A QUORUM IS A COUNT OF PARTIES, NOT OF SIGNATURES.
      `apps/oshun/bff/src/studio/approval-workflow-store.ts` computed
      `approvals.filter(approve).length` and compared it to `requiredApprovals`. `approver` is
      captured on every entry and was never compared — no dedup, no distinctness check. Driven on
      the evaluator itself: a stage with `requiredApprovals: 3` and three entries all from
      `ops-1` returned `{status: 'approved', approveCount: 3, approvalsNeeded: 0}` — byte-identical
      to the verdict for three independent reviewers. Approving entries are now collapsed by
      approver before the count, so one reviewer contributes one vote however many entries they
      submit; a genuine two-hat reviewer still counts once and keeps BOTH roles available to
      satisfy `requiredRole`, and the discarded repeats are reported as `duplicateApprovals` so
      the operator can see them rather than having them vanish. The route spec (+1) drives all
      three cases through the real HTTP surface; every pre-existing expectation is unchanged
      because every existing fixture already used distinct approvers.
      
      AND THE EVALUATOR STILL CANNOT SEPARATE DUTIES. .e asks for separation of requester,
      creator, reviewer and approver. `ApprovalInput` carries `approver`, `role` and `decision`
      — 1 of the 4 duties. There is no requester and no creator anywhere on the input, so the
      evaluator has nothing to compare an approver AGAINST. Distinctness within the quorum is now
      enforced; independence from the work under review is not expressible.
      
      A GOVERNANCE CLASS IS NOT AN ASSET KIND. .b asks that every publishable asset kind be
      mapped to its visual/audio/3D/editorial/technical review dimensions. The applicability axis
      both contracts actually carry is a six-member governance class — release_gate,
      model_promotion, runtime_risk, policy_change, content_review, default — with exactly one
      template each. 1 of .b's 5 review dimensions (editorial) appears anywhere in either
      contract.
      
      1 of .c's 9 risk classes and 1 of .d's 6 audience axes appear in either contract. TRAP. A
      WORD IN A CONTRACT IS NOT A REQUIREMENT: the one risk hit is `voice_profile`, an artifact
      TYPE, not a rule that voice work needs a particular reviewer; the one audience hit is
      `ReviewReleaseChannelSchema`. And bare `target` matched `targetId` and `targets` until the
      spelling was tightened, which would have reported a delivery-target axis that does not
      exist.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1)
      `getAdminReviewTemplate` resolves exactly one template per artifact class with an explicit
      `default` fallback that throws if the default is missing, and `ReviewTemplatePanel` resolves
      it from the review record's OWN class rather than from a caller-supplied one — .f's
      deterministic resolution holds. (2) A reject on any approval rejects the stage outright,
      whatever the approval count. (3) The delegation/escalation validator refuses a transition
      that leaves the role unchanged.
      
      TRAP. A CONSTANT WITH NO IMPORTERS CAN STILL BE READ THROUGH ITS ACCESSOR: grepping for
      `ADMIN_REVIEW_TEMPLATES` finds only the contract and its own spec, which reads as six
      templates nothing consumes. The panel calls `getAdminReviewTemplate`, so the templates are
      rendered — ask of the accessor, not only of the constant.
      
      WHAT .f CANNOT DO: neither contract carries a version or a lifecycle, so a resolved template
      cannot be pinned to the candidate and a later change in applicability has nothing to
      invalidate. Resolution is deterministic and re-runs from scratch every render.
      
      2 of 9 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: the quorum now counts
      parties and resolution is deterministic, but no template can require a role or a quorum,
      which is the matrix the parent asks for. -->
      
  • I8.2 Add anchored review comments for regions, frames, timecodes, graph nodes, mesh parts, and audio ranges. Evidence: anchor survival tests.
    • I8.2.a Define immutable comment/thread/reply identity, subject revision, typed anchor, author, visibility, status, mentions, attachments, resolution, and supersession. Evidence: comment schema tests.

    • I8.2.b Implement image/video region anchors in canonical pixel/ normalized/frame coordinate space with crop/orientation/version context. Evidence: region anchor tests.

    • I8.2.c Implement frame/timecode/shot/range anchors with rate/time-base/ sequence revision and retime/edit mapping. Evidence: temporal anchor tests.

    • I8.2.d Implement graph node/port/edge/parameter/subgraph anchors using stable graph entity/version identities rather than screen coordinates. Evidence: graph anchor tests.

    • I8.2.e Implement mesh/scene object/component/face/vertex/UV/material/ animation anchors with stable semantic IDs and transform/LOD context. Evidence: 3D anchor tests.

    • I8.2.f Implement audio sample/timecode/clip/track/channel/speaker/ transcript-token ranges with sample/time-base context. Evidence: audio anchor tests.

    • I8.2.g Map anchors across compatible revisions with exact/shifted/ ambiguous/orphaned status and preserve the original anchor. Evidence: revision survival tests.

    • I8.2.h Provide synchronized accessible anchor lists/navigation, authorization, notifications, resolve/reopen history, and no comment authority over approval. Evidence: review UI/API tests. <!-- MEASURED 2026-08-20 (I8.2.a-h). The parent's evidence is "anchor survival tests", so the measurement asks the binding question first: is there a typed anchor at all, and does anything map one across a revision? Instrument: generate-oshun-review-anchors.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-review-anchors.json.

      AN UNTYPED KIND-AND-VALUE PAIR IS NOT A TYPED ANCHOR. The estate has exactly one
      review-comment anchor: `AdminEntityCommentAnchorSchema = { kind: string(1..80), value:
      string(1..400) }`, and its own doc comment says the values are "intentionally free-form so
      each workspace can carry its own target shape (line, field, region) without a rigid
      contract". It can HOLD a serialisation of a region, a timecode, a graph node, a mesh face
      or an audio range, and it can VALIDATE none of them — `kind: 'anything-at-all'` is a
      well-formed anchor. So .b's coordinate space, .c's time base, .d's stable graph identity,
      .e's semantic mesh ID and .f's sample rate are not merely absent from the contract; there
      is nowhere in it for them to go. 0 of the 18 context terms those five families need appear
      in either source that references the anchor.
      
      AND THE COMMENT CANNOT SAY WHICH REVISION IT WAS ANCHORED TO. `AdminEntityCommentSchema`
      carries 6 of .a's 11 subjects — comment identity, thread/reply identity, author, status,
      mentions and resolution. It carries no subject revision, no visibility, no attachments, no
      supersession, and no typed anchor. `revision` on the record is the COMMENT's own edit
      counter, not the revision of the thing it points at. Without a subject revision, .g's
      survival mapping is not unimplemented — it is unaskable, because nothing can tell whether
      the anchored thing moved. 0 sources carry any re-anchoring or anchor-status logic, and 0 of
      .g's four statuses (exact/shifted/ambiguous/orphaned) appear beside an anchor.
      
      FIXED: A RESOLVE IS A REVIEW ACT. `POST .../entities/:entityId/comments` recorded
      `admin.entity_comment.created` (or `.entity_annotation.created`) and flushed it; `PATCH`
      and `DELETE` on the same comment recorded NOTHING. So the record held the moment a finding
      was raised and not the moment it was closed, reopened, reworded, moved to a different
      anchor or deleted — and the PATCH handler is the one that can change the anchor. The
      update now records `admin.entity_comment.resolved`, `.reopened` or `.updated` by what the
      request actually did, with `bodyChanged`/`anchorChanged`/`mentionsChanged` in the payload,
      and the delete records `admin.entity_comment.deleted`; both flush before replying, like the
      create. The route spec (+1) drives create → move the anchor → resolve → reopen → delete and
      reads all five events back in order out of the audit store.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The
      approval evaluator never mentions a comment: a stage is decided by approvals alone, so
      .h's "no comment authority over approval" holds. (2) A comment update preserves the
      existing anchor when the request omits it and clears it only on an explicit null. (3) A
      body edit by a non-author is refused with 403.
      
      TRAP. A SUBSTRING IS NOT A SUBJECT. Scanning `libs/` and `apps/oshun/` for the word
      "anchor" finds 97 files and 43 more carrying re-anchoring logic — every one of them an AR,
      XR, cinema or brand-strategy anchor with nothing to do with a review comment, and counting
      them would have credited .b-.g with 12 of 18 context terms and 43 survival sources. The
      honest population is the sources that reference THIS contract by name: two of them. A
      GENERATED CLIENT CONTAINS THE WHOLE API — `libs/openapi/src/generated/oshun-bff.ts`
      mentions the anchor and also every unrelated `trackId`, `channel` and `sampleRate` in the
      estate; it is excluded for the same reason. And bare `normalized` matched a local
      mention-identity variable, crediting a coordinate space that does not exist.
      
      2 of 11 properties present. ALL EIGHT CELLS AND THE PARENT LEFT `[ ]`: the comment
      lifecycle is now fully recorded and comments hold no approval authority, but the anchor is
      untyped, the subject revision is absent, and nothing maps an anchor across a revision. -->
      
  • I8.3 Add safety, rights, identity, quality, brand/editorial, and technical signoff gates. Evidence: separation-of-duties policy tests.
    • I8.3.a Define distinct signoff contracts for safety, rights/consent, identity, quality, brand/editorial, and technical/target dimensions with required evidence and decision codes. Evidence: signoff schemas.

    • I8.3.b Resolve eligible reviewers/approvers, role separation, conflict- of-interest, quorum, delegated/substitute rules, and expiry per dimension. Evidence: duty matrix.

    • I8.3.c Bind each signoff to exact candidate/bundle revision and hashes, target, measurements/verdicts, policy/template versions, comments, and evidence snapshot. Evidence: binding tests.

    • I8.3.d Implement approve/reject/request-changes/abstain with immutable decision, rationale, conditions, and no cross-dimension authority. Evidence: decision tests.

    • I8.3.e Calculate aggregate readiness only when every required dimension has fresh valid quorum and no blocking conflict/condition. Evidence: aggregation tests.

    • I8.3.f Invalidate affected signoffs on subject/evidence/policy/role/ rights change without deleting prior decisions. Evidence: freshness tests. <!-- MEASURED 2026-08-20 (I8.3.a-f). The parent's evidence is "separation-of-duties policy tests", so the measurement asks the binding question first: does anything on the review signoff path refuse an approval from the person who requested or created the work? Instrument: generate-oshun-signoff-gates.mts (+9 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-signoff-gates.json.

      THE ESTATE KNOWS HOW TO REFUSE A SELF-APPROVAL — IN NINETEEN PLACES, NONE OF THEM THIS ONE.
      The exemplar is Oshun's own: `libs/oshun/workbench-kit/src/gate-waivers.ts` flags
      `approver-is-the-requester`, `approver-is-the-accountable-owner` AND
      `approver-role-is-not-the-one-the-gate-names` — .b's duty matrix, implemented, in this
      estate. Neith's privileged-access elevation throws 'requester cannot approve own
      elevation'; the vault escrow, PAM approval, SCIM lifecycle, IDP recovery, Metis peer
      review, Mnemosyne assessment, Psyche's tool-call approvals and the Yemaya study-workspace
      release path all compare a decider against a requester, author or owner. 0 of the 19 are on
      the review signoff path — and the reason is structural: `ApprovalInput` is
      `{approver, role, decision}`, so the evaluator has no second identity to compare against.
      
      RECORDED AS A DECISION, NOT FIXED. Adding an optional `requestedBy` to the workflow input
      and refusing that party's approval would be five lines, but the whole input is
      caller-supplied: a caller who omits the field gets no protection while the response looks
      identical to one that was checked. A half-guarantee that reads like a full one is worse
      than a stated gap, and choosing where the requester identity comes from is a contract
      decision this cell is not entitled to make alone.
      
      THREE OF THE SIX DIMENSIONS HAVE NO STAGE TYPE. `ReviewStageTypeSchema` enumerates twelve —
      review, approval, security, compliance, release_gate, rights, privacy, safety, editorial,
      accessibility, localization, appeal. .a's safety maps to `safety`, rights/consent to
      `rights` and `privacy`, brand/editorial to `editorial`. Identity, quality and
      technical/target map to nothing, so a signoff in those three dimensions cannot be raised
      under a type that names it, let alone carry its own required evidence.
      
      THE DUTY MATRIX AND THE BINDING DO NOT REACH THE EVALUATOR THAT RUNS. 3 of .b's 6 duty
      subjects appear in the review contracts (quorum, delegation, expiry) and 1 reaches the
      evaluator an operator can call — quorum, as `requiredApprovals`. Eligible reviewers, role
      separation and conflict of interest appear in neither. 1 of .c's 7 binding subjects appears
      in the contracts and 0 on the evaluator: it receives no candidate revision, no hash, no
      target, no measurement, no policy or template version and no evidence snapshot. .f's
      invalidation is therefore not unimplemented but unaskable — the evaluator cannot tell that
      any of those changed because it never saw them.
      
      THERE IS NO WAY TO ABSTAIN. .d names approve/reject/request-changes/abstain. Both contracts
      stop at three: `ReviewDecisionOutcomeSchema` is approved|rejected|changes_requested and
      `ADMIN_REVIEW_DECISION_OUTCOMES` is none|approved|rejected|changes_requested. A reviewer
      with a conflict has no code to record that they stood down, so the only ways out of a
      quorum are to approve, to reject, or to leave the stage pending forever.
      
      A QUORUM WITH NO CLOCK CANNOT BE STALE. The aggregate is real — the workflow reports
      `approved` only when every stage is approved — but `ApprovalInput` carries no decision
      time, so .e's "fresh valid quorum" has nothing to measure freshness against and an approval
      counts forever.
      
      THE MODEL THAT COULD HAVE DONE ALL OF THIS HAS NO CONSUMER. `@oshun/review-persistence`
      publishes 18 modules — a decision lifecycle with prior-decision lineage, a delegation
      policy, a release-validity service, a review queue with SLA weighting, a template registry
      and in-memory repositories — and 0 product files import it. Its own contracts carry
      attestations with `expiresAt` and validation that an expired attestation must name when it
      expired. None of that is reachable from any route.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) Each stage
      is evaluated over its own approvals array, so one dimension's quorum cannot be filled from
      another's — .d's "no cross-dimension authority" holds. (2) A reject on any approval rejects
      that stage outright, and any rejected stage rejects the workflow. (3) The aggregate
      requires every stage approved before reporting the workflow approved.
      
      TRAP. A COMPARISON IS NOT ALWAYS A REFUSAL: the 19 include eligibility checks
      (`accountable-signoff.ts` compares the signer against the owner and their delegates) as
      well as outright self-approval refusals. Both are duty checks, which is why the census
      counts "compares a deciding identity against another party's" rather than claiming
      nineteen self-approval guards.
      
      1 of 9 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: dimensions do not
      separate duties, signoffs bind to nothing, and the only thing the reachable evaluator
      enforces is that one dimension cannot borrow another's approvals. -->
      
  • I8.4 Lock approved release candidates to exact revisions and hashes. Evidence: post-approval mutation invalidates approval.
    • I8.4.a Define release-candidate manifest listing exact asset/output/edit/ metadata/rights/provenance/gate/target-profile/dependency revisions and cryptographic hashes. Evidence: candidate schema.

    • I8.4.b Canonicalize and hash the complete candidate manifest and required files before review begins. Evidence: hash golden tests.

    • I8.4.c Bind every review/signoff/exception/confirmation to the manifest digest and immutable candidate identity. Evidence: approval binding tests.

    • I8.4.d Make approved candidate records immutable and route any content/ metadata/dependency change through a new revision/candidate. Evidence: mutation refusal tests.

    • I8.4.e Rehash/revalidate manifest and stored bytes at release planning, immediately before execute, and after target assembly. Evidence: TOCTOU tests.

    • I8.4.f Mark prior approvals stale when any bound version/hash/policy/ evidence changes while retaining historical approval. Evidence: invalidation tests. <!-- MEASURED 2026-08-20 (I8.4.a-f). The parent's evidence is that a "post-approval mutation invalidates approval", so the measurement asks the binding question first: is an approval bound to anything a mutation could change? Instrument: generate-oshun-release-locking.mts (+7 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-release-locking.json.

      THE ESTATE CAN LOCK A PUBLISH AND DOES NOT LOCK A RELEASE.
      `apps/oshun/bff/src/studio/studio-scene-publish-store.ts` is the exemplar and it is exactly
      what .b, .d and .e describe: it hashes the exact submitted manifest into `manifestSha256`,
      derives the publish identity AND the release-edition stream FROM that digest, and on every
      rehydration recomputes `sha256(manifestJson)`, refuses a record whose digest does not match
      its bytes, then rebuilds the identity from the canonical fields and refuses a record whose
      `publishId` or `releaseStream` is not the one those fields imply. That is a locked
      candidate. Recorded as CORRECT so a later pass does not re-flag it.
      
      THE RELEASE-READINESS DECISION, TWO FILES AWAY, BINDS TO A MUTABLE POINTER.
      `AdminReleaseReadinessEntry` declares releaseId, releaseName, targetDate, tier, score,
      summary, gates, goNoGoAt, approverUserId, approverName — no digest, no revision, no
      manifest. `AdminRecordGoNoGoDecisionInputSchema` is {reportId, releaseId, decision,
      approverUserId, approverName, rationale}. So .c has no digest to bind a signoff to, and .f
      cannot mark an approval stale because the approval was never bound to anything that could
      change. 0 of .a's 9 manifest subjects — asset, output, edit, metadata, rights, provenance,
      gate, target profile, dependency — are pinned to a revision or a hash in either release
      contract.
      
      FIXED: AN APPROVER NAMED IN THE BODY IS NOT THE APPROVER WHO CALLED.
      `recordAdminGoNoGoDecision(userId, input)` receives the authenticated principal as `userId`
      and then wrote `release.approverUserId = input.approverUserId` — the value straight off the
      request — using `userId` only to choose which per-operator record to mutate. The two were
      never compared. So an authenticated operator could record a GO decision attributed to any
      principal they cared to name, and the release would read as decided by someone who never
      called.
      
      The named approver is KEPT as given, because transcribing a decision taken in a meeting is
      legitimate — `approverName` is a human name, not a principal. What was missing was the
      disclosure. The record now carries `recordedByUserId` (the authenticated caller, never
      caller-supplied) and `approverIsRecorder`, and the appended summary line reads "GO decision
      by <name>, recorded by <caller>" whenever they differ. This is a seam, not a guarantee: it
      makes an existing ambiguity visible rather than pretending to close it, and the field
      documentation says plainly that nothing downstream should read `approverUserId` as an
      identity the platform verified. The route spec (+1) names a principal the caller is not and
      asserts all three.
      
      NOTHING REHASHES ON THE RELEASE PATH. .e asks for a rehash at release planning, immediately
      before execute, and after target assembly. 5 sources in the estate recompute a digest and
      refuse a mismatch — Bellona's remote-host snapshot verifier ("refusing to restore a
      tampered snapshot"), the scene-publish store, an Isis 3D pipeline, the workbench kit's
      rebuild-identity and Yemaya's resumable upload — and exactly 1 of the 5 is on the release
      path. 0 release sources carry any approval-staleness marker.
      
      CHECKED AND FOUND CORRECT. (1) The scene-publish lock above. (2) A release whose `goNoGoAt`
      is already set refuses a second decision with `release_already_decided` — .d's immutability
      for the decision itself holds, even though a content change is not routed through a new
      candidate. (3) A `go` is refused unless every gate is passing or waived, and a `no_go` is
      refused unless at least one gate is not passing; the decision cannot contradict its own
      gates.
      
      TRAP. AN IDENTIFIER CLASS WITHOUT DIGITS CANNOT MATCH `manifestSha256`. The first version of
      the rehash predicate used `[A-Za-z.]*`, which stops at the `2` of `Sha256`, and it reported
      2 rehashers where there are 5 — missing Bellona's snapshot verifier, the estate's clearest
      example of the thing .e asks for. A census that undercounts the estate's own capability is
      as wrong as one that overcounts it.
      
      3 of 8 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: one publish path locks
      correctly and the release decision refuses to be retaken, but the decision binds to a
      releaseId and nothing else, so there is no manifest to rehash and no binding to invalidate. -->
      
  • I8.5 Assemble deterministic output bundles per target. Evidence: golden bundle round-trip for every publishable asset kind.
    • I8.5.a Define target bundle manifest for entry assets, derivatives, dependencies, metadata, captions/transcripts, provenance, rights/ attribution, checksums, target profile, and tool versions. Evidence: bundle schema.

    • I8.5.b Resolve only exact authorized release-candidate files/ dependencies and fail on missing/mutable/ambiguous/quarantined inputs. Evidence: input resolution tests.

    • I8.5.c Transform/package each publishable asset kind with pinned tools, configs, locale/time/random/order controls, and reproducible filenames. Evidence: per-kind assembly fixtures.

    • I8.5.d Generate normalized metadata, accessibility files, rights/ attribution, C2PA/ledger links, target descriptors, and checksums from canonical records. Evidence: metadata mapping tests.

    • I8.5.e Prevent secrets/private fields/path traversal/symlinks/hidden files/unsafe archives and enforce file/count/size/resource budgets. Evidence: hostile bundle tests.

    • I8.5.f Canonically order and serialize manifests/archive entries/ timestamps/permissions/compression so identical inputs produce identical hashes where format permits. Evidence: determinism tests.

    • I8.5.g Parse and validate every golden bundle through the target consumer/importer and compare semantic round-trip. Evidence: per-target round-trip suite. <!-- MEASURED 2026-08-20 (I8.5.a-g). The parent's evidence is a "golden bundle round-trip for every publishable asset kind", so the measurement asks the binding question first: what does a bundle manifest in this estate declare, and could two runs over identical inputs produce identical bytes? Instrument: generate-oshun-bundle-assembly.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-bundle-assembly.json.

      FIXED: A PATH THAT ONLY DESCRIBES AN ARCHIVE IS STILL A PATH SOMEBODY WILL EXTRACT. All
      four bundle file-entry schemas — `EngineBundleFileEntrySchema` and
      `DCCBundleFileEntrySchema` in the output registry, `JobEngineBundleFileEntrySchema` and
      `JobDCCBundleFileEntrySchema` in the generation API — declared `path: z.string().min(1)`.
      Those manifests arrive on `RegisterOutputRequestSchema`, a CALLER-SUPPLIED registration
      request, so `../../etc/passwd`, `/etc/passwd`, `..\..\windows`, `.ssh/config`, `a//b` and
      `trailing/` were all well-formed bundle entries, and nothing downstream that unpacks a
      bundle by these paths could tell them from ordinary ones. The two `superRefine` validators
      that already read `manifest.files[].path` check for a README entry and a mesh role — never
      for a traversal.
      
      The rule now lives once, in `libs/shared/security/src/bundle-entry-path.ts`, and all four
      schemas read it. A bundle entry path must be a relative POSIX path of ordinary segments: no
      absolute, drive or UNC prefix; no backslash (a `\` is a legal POSIX filename character, so
      an archive writer and reader can disagree about whether `a\b` is one segment or two); no
      `.`, `..`, empty or leading-dot segment; no control characters; no trailing separator; and
      bounded at 1024 characters overall and 255 per segment. Every applicable refusal is
      reported, not just the first, so a caller fixing a manifest sees the whole problem at once.
      ONE RULE IN ONE PLACE WAS THE POINT: putting a copy in each app would have recreated the
      §I7.20 defect — two lists for one judgement — so `@oshun/security` was added to the output
      registry's dependencies rather than duplicating the predicate. `bundle-entry-path.spec.ts`
      (+8) pins every refusal, and a schema-level test in each app (+1 each) drives six hostile
      paths through the real request schema and asserts the ordinary path still parses.
      
      A GENERATION TIMESTAMP IN THE MANIFEST DEFEATS THE DETERMINISM THE CELL ASKS FOR. All 4
      manifests carry `generatedAt: z.string().datetime()`, so an identical bundle assembled
      twice differs in exactly the field that records when it was assembled. TRAP. A
      CALLER-OVERRIDABLE CLOCK IS NOT A PINNED ONE: both assemblers
      (`engine-export-bundles.ts`, `dcc-export-bundles.ts`) accept `options.generatedAt` — which
      is how their own fixtures stay stable — and both fall back to `new Date().toISOString()`,
      so a production bundle is not reproducible even though a test bundle is. Neither sorts its
      `files`, so entry order is whatever the caller passed; 0 of the 23 export-bundle sources
      carry any canonical ordering or stable serialisation.
      
      AN ENTRY CAN SAY WHERE A FILE SITS AND NOT WHAT ITS BYTES HASH TO. 0 of the 4 file-entry
      schemas declare a checksum; the entry carries `path`, an optional `sourceUrl`, a `role`, an
      optional `mimeType` and an optional `sizeBytes`. `checksums` is also one of the 6 of .a's 10
      manifest subjects that is absent — with derivatives, dependencies, captions/transcripts,
      rights/attribution and tool versions. The 4 present are entry assets, metadata, provenance
      (as `sourceWorkflowId`) and the target profile (as `engine`/`dcc`).
      
      3 of .e's 7 hostile-bundle subjects still have no guard anywhere: symlinks, unsafe archives
      (no decompression-ratio or entry-depth bound) and secrets or private fields. Path
      traversal, absolute paths, hidden files and size budgets are guarded by the validator added
      here.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The engine
      bundle schema refuses a manifest whose `engine` disagrees with the bundle's, whose
      `sourceMesh.format` disagrees with the declared format, or whose `entryCount` disagrees with
      `files.length` — real cross-field validation, not shape checking. (2) The engine export
      bundle requires a `README.md` entry and at least one entry with the `mesh` role.
      
      TRAP. A WORD IN A DIFFERENT BUNDLE IS NOT THIS BUNDLE. Searching `apps/` and `libs/` for
      "bundle" plus a round trip reports 21 suites — investigation bundles, data-export bundles,
      offline bundles — and searching for a stable serialiser finds Isis's PROVENANCE bundle. The
      honest population is the 23 sources that name an engine or DCC export bundle, and 0 of them
      round-trips a golden bundle through a target importer, which is the parent's whole evidence.
      
      1 of 7 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`: entry paths are now
      safe, but a bundle carries no checksums, its manifest embeds the moment it was built, its
      entries are unordered, and nothing parses one back through the consumer it was built for. -->
      
  • I8.6 Verify outputs after write in output registry, gallery, marketplace, or downstream store. Evidence: read-back mismatch fails release.
    • I8.6.a Define per-destination write receipt and verification contract for canonical object/version, bytes/hash/length/content type, manifest, metadata, visibility, URL, status, and timestamp. Evidence: receipt schemas.

    • I8.6.b Write idempotently using destination-native conditional/versioned operations and retain immutable request/response correlation. Evidence: duplicate-write tests.

    • I8.6.c Read back from the authoritative destination rather than local request/cache and verify bytes/hash/length/format/manifest/dependencies. Evidence: read-back integrity tests.

    • I8.6.d Verify destination metadata, permissions/audience, publication/ listing state, target URL/download behavior, cache/CDN state, and provenance credential. Evidence: visibility tests.

    • I8.6.e Treat timeout/partial/unknown/mismatch as unverified and block release success; retry or compensate idempotently. Evidence: failure injection tests.

    • I8.6.f Persist signed/hashed verification evidence and recheck after downstream transformations or configured propagation delay. Evidence: delayed verification tests. <!-- MEASURED 2026-08-20 (I8.6.a-f). The parent's evidence is that a "read-back mismatch fails release", so the measurement asks the binding question first: does anything fetch the bytes back from the authoritative destination and compare them? Instrument: generate-oshun-write-verification.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-write-verification.json.

      IT DOES, AND THAT PART IS RIGHT — recorded as CORRECT rather than as a gap. On registration,
      `output.service.ts` builds the storage object URL, fetches it over HTTP, stream-hashes the
      body with a reader rather than buffering the whole file, and throws `HashMismatchError`
      when the computed digest disagrees with the one the caller declared. A wrong digest has
      always blocked registration, and `verifyIntegrity` is routed as a later recheck that fetches
      the same way. That is .c, implemented properly.
      
      FIXED: A VERDICT FROM AN INSTRUMENT THAT COULD NOT LOOK IS NOT A VERDICT. The read-back had
      three exits and only one of them was a measurement. A genuine mismatch threw. But an
      algorithm this runtime cannot compute — `blake3` falls back — logged "skipping server-side
      hash verification" and RETURNED; and unreachable storage logged "could not verify upload
      hash from storage; proceeding with client hash" and returned too. Both left an output
      registered with a digest nobody had checked, byte-identical in the database to one whose
      bytes had been fetched and hashed. .e asks for exactly timeout/partial/unknown to be treated
      as unverified and to block release, and nothing can block on a fact nobody wrote down.
      
      The read-back now returns a typed verdict — `verified`, `unsupported_algorithm` or
      `storage_unreachable`, carrying the algorithm, the expected digest, the computed digest (or
      null) and the reason — and the registration records it on the output under
      `metadata.hashVerification`. A mismatch still throws. Driven end to end in
      `output.service.hash-verification.spec.ts` (+6): matching bytes → `verified` with the
      computed digest; different bytes → `HashMismatchError`; a refused connection → 
      `storage_unreachable` with the reason and a null digest; a 404 → the same with the status in
      the reason; `blake3` → `unsupported_algorithm`; and a property that `verified` is never
      returned without a computed digest.
      
      RECORDING IS NOT BLOCKING, AND THE CELL ASKS FOR BOTH. Registration still proceeds after an
      unreachable destination — that was a deliberate availability choice by the original author,
      stated in its own comment — so .e is only half closed. What changed is that a release gate
      now has a fact to refuse on instead of a warning in a log.
      
      ONE READ-BACK FOR FOUR DESTINATIONS. .a asks for a per-destination write receipt across the
      output registry, gallery, marketplace and downstream stores. Exactly 1 source in Isis and
      Oshun fetches from a destination and refuses a digest mismatch, and it is the output
      registry. The other three destinations verify nothing after writing.
      
      WHAT THE RECEIPT RECORDS: 4 of .a's 10 subjects — the canonical object, the bytes hash, the
      status and the timestamp. It does not record the content length, the content type, the
      manifest, the destination metadata, the visibility or the URL. 0 of .d's 6 destination
      checks exist at all: nothing verifies destination metadata, permissions or audience,
      publication or listing state, target-URL download behaviour, cache or CDN state, or a
      provenance credential. Nothing beyond the bytes is checked.
      
      A RESULT YOU RETURN IS NOT EVIDENCE YOU KEPT. `verifyIntegrity` is routed and real, and its
      `IntegrityVerificationResult` is returned to the caller and never written down — there is no
      INSERT or UPDATE anywhere in its body. So .f's "persist signed/hashed verification evidence"
      holds only for the registration verdict added here, and there is no scheduled recheck after
      a propagation delay. .b's destination-native conditional write is absent too: nothing on the
      path uses an `If-None-Match`, `If-Match` or versioned-write operation, so a duplicate write
      is not refused by the destination.
      
      TRAP. A FIELD SOMEWHERE IN A THREE-THOUSAND-LINE SERVICE IS NOT A FIELD ON THE RECEIPT.
      Matching the receipt subjects against the whole of `output.service.ts` reported 9 of 10 —
      `manifest`, `metadata`, `visibility` and `url` all appear in the file, none of them on a
      verification result. Scoped to the verification contract and the two methods that produce
      one, it is 4. TRAP. A REGEX THAT SEARCHES PAST A FUNCTION FINDS THE NEXT FUNCTION'S CODE:
      looking for an INSERT "within 4000 characters of verifyIntegrity" reported the result as
      persisted; brace-matching the method body reports the truth.
      
      3 of 9 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: the read-back is real
      and a mismatch blocks, an unverifiable write is now recorded as unverified, and everything
      else — per-destination receipts, idempotent conditional writes, destination state, persisted
      evidence and delayed recheck — is absent. -->
      
  • I8.7 Implement collection and campaign curation through versioned bundles. Evidence: no direct catalog mutation.
    • I8.7.a Define collection/campaign bundle identity/version, purpose, audience, owner, schedule, ordered members/exact revisions, metadata, targets, constraints, and lifecycle. Evidence: curation schema.

    • I8.7.b Add/remove/reorder/group/feature members only through a new draft bundle revision with reason and optimistic concurrency. Evidence: curation edit tests.

    • I8.7.c Validate each member’s availability, target compatibility, rights/consent, gates, provenance, editorial approval, and campaign steering constraints. Evidence: membership tests.

    • I8.7.d Preview deterministic catalog/gallery/channel diffs, assets/ metadata/order/schedule/dependencies, and removals before approval. Evidence: publication preview tests.

    • I8.7.e Review/approve/release the exact bundle manifest through governed release APIs; prohibit direct catalog/gallery mutation. Evidence: boundary tests.

    • I8.7.f Preserve released bundle history and propagate member quarantine/ recall/expiry through explicit successor/incident state. Evidence: lifecycle tests. <!-- MEASURED 2026-08-20 (I8.7.a-f). The parent's evidence is "no direct catalog mutation", so the measurement asks the binding question first: can the catalog be mutated without going through a curation bundle? Instrument: generate-oshun-curation-bundles.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-curation-bundles.json.

      IT CAN, AND THAT IS THE ONLY WAY IT IS MUTATED. `POST
      /v1/admin/isis/output-gallery/bulk-actions` takes a list of output ids and one of four
      action kinds — `re-tag`, `send-to-editorial`, `revoke-consent`, `takedown` — and applies it
      to the gallery. The route names no bundle, no collection and no campaign anywhere; there is
      no manifest to approve, no draft revision, and no optimistic-concurrency check. That route
      IS the catalog mutation .e says to prohibit, and it is the only catalog-editing surface an
      operator has.
      
      THE CURATION MODEL EXISTS AND NOTHING REACHES IT. `libs/contracts/src/common/collection.ts`
      is 1030 lines: a 24-field collection with sections, items, collaborators, smart
      configuration, and real cross-field validation that checks domain membership, collaborator
      and section coherence, smart-config consistency, and that `itemCount` equals the number of
      items. 0 product files import a `Collection` binding from `@oshun/contracts`, and 0 admin
      routes name a collection. `CollectionManagementProvider` in `@isis/ai-providers` — the other
      1814-line collection implementation, measured under §I7.20 — has 0 importers too. Two
      curation models, neither reachable, while the catalog is edited by a route that knows about
      neither.
      
      A COLLECTION CANNOT SAY WHICH REVISION IT IS. 6 of .a's 11 bundle subjects are expressible:
      identity, purpose, audience, owner, metadata and lifecycle. There is no version or revision,
      no schedule, no ordered members at exact revisions, no targets and no constraints. .b asks
      that a member edit happen only through a NEW draft bundle revision under optimistic
      concurrency — with no revision field there is nothing to draft and nothing to compare, so
      .b is unaskable rather than unimplemented, and .f's released-bundle history has nothing to
      keep.
      
      NOTHING VALIDATES A MEMBER WHERE MEMBERS ARE ACTUALLY EDITED. .c asks each member to be
      checked for availability, target compatibility, rights and consent, gates, provenance,
      editorial approval and campaign steering. 1 of those 7 appears on the gallery route, and
      only as the `revoke-consent` ACTION — an operator withdrawing consent, not a check that a
      member holds it. The other six appear nowhere on the editing path.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The
      collection contract validates its own shape thoroughly, including the `itemCount`-versus-
      `items.length` agreement that a hand-maintained count would otherwise drift on. (2) The
      gallery bulk action requires an authenticated caller with an Isis workspace scope, and its
      `takedown` kind requires a reason.
      
      TRAP. `etag` IS A SUBSTRING OF `removeTags`. Testing the gallery route for an
      optimistic-concurrency check with `/expectedVersion|ifMatch|etag|revision/i` matched inside
      the identifier `removeTags` and reported a concurrency guard the route does not have; word
      boundaries give the truth. TRAP. A FIELD BAG IS NOT THE WHOLE SCHEMA: `CollectionSchema`
      spreads `CollectionFields` and then adds seven more fields of its own, so reading only the
      bag reported 17 fields and called `owner` absent when it is present — the census reads both.
      TRAP. A NAME IS NOT AN IMPORT: `CollectionSchema` is declared independently in Athena,
      Hestia, Cybele and Veritas, and a bare identifier search reports six consumers this contract
      has not got; counting import bindings whose specifier is `@oshun/contracts` gives zero.
      
      0 of 6 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: the catalog is edited
      directly by a route with no bundle, no revision and no membership validation, and the two
      curation contracts that could have carried one are unreachable from any surface. -->
      
  • I8.8 Implement supersession, archive, unpublish, quarantine, and recall. Evidence: downstream state and lineage remain coherent.
    • I8.8.a Define separate lifecycle command/state/reason/authority/ effective-time/target/notification/verification contracts for supersede, archive, unpublish, quarantine, and recall. Evidence: lifecycle schemas.

    • I8.8.b Supersede by linking an eligible replacement revision/bundle, migration/effective behavior, prior availability, and consumer notice without rewriting the predecessor. Evidence: supersession tests.

    • I8.8.c Archive inactive authoring/catalog records from routine discovery while retaining authorized history, links, and restore rules. Evidence: archive tests.

    • I8.8.d Unpublish distribution/listing visibility by explicit targets while preserving source asset/release evidence and recording residual copies. Evidence: unpublish tests.

    • I8.8.e Quarantine suspected versions from reuse/review/release/delivery with incident scope, restricted evidence, and fail-closed dependency propagation. Evidence: quarantine tests.

    • I8.8.f Recall confirmed affected versions through complete blast radius, downstream disable/removal, notification, receipt, and verification. Evidence: recall tests.

    • I8.8.g Project coherent current/historical state across lineage, search, gallery, marketplace, collections, campaigns, URLs, caches, and exports. Evidence: cross-system lifecycle tests. <!-- MEASURED 2026-08-20 (I8.8.a-g). The parent's evidence is that "downstream state and lineage remain coherent", so the measurement asks the binding question first: which of the five withdrawal verbs exists as a state a record can be in, and does any of them reach anything downstream? Instrument: generate-oshun-lifecycle-withdrawal.mts (+7 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-lifecycle-withdrawal.json.

      AN ARCHIVE THAT ONLY MOVES THE BYTES IS NOT AN ARCHIVE THAT REMOVES THE RECORD FROM
      DISCOVERY. `OutputStatus` is pending | available | archived | expired | deleted, and 1 of
      .a's 5 verbs — archive — has a state at all. That one is annotated "Moved to cold storage",
      which is a storage-tier fact `StorageTier` already carries separately as
      hot | warm | cold | glacier. The three list and search condition sets in the output service
      each exclude exactly one status, `deleted`, so an ARCHIVED output is still returned by
      routine discovery. .c asks archive to take a record out of routine discovery while retaining
      authorized history; the status names a place the bytes went, and nothing anywhere changes
      what a caller browsing the catalog sees.
      
      SUPERSEDE, UNPUBLISH, QUARANTINE AND RECALL HAVE NO STATE. A record cannot be in any of
      them, so .a's per-verb contracts — command, state, reason, authority, effective time,
      target, notification, verification — have nothing to attach to for four of the five verbs.
      .b's supersession is the sharpest case: nothing on the output service carries a
      `supersededBy`, `successorId` or `replacedBy`, so a replacement cannot be linked and a
      predecessor cannot be left intact-but-superseded, which is exactly what .b asks for.
      
      THERE IS A REAL TAKEDOWN WORKFLOW, AND IT IS NOT THE CATALOG'S. `/v1/admin/takedowns`,
      `/v1/admin/takedowns/:id/process` and `/v1/admin/takedowns/:id/counter-notice` are a proper
      notice-and-counter-notice surface, and `/v1/living-scenes/admin/artifacts/:id/takedown`
      withdraws a living-scene artifact. None of them sets a state on an output, and the output
      service does not read them. So .f's recall exists as a process for a different subject.
      
      NOTHING PROPAGATES. No quarantine propagation, no recall cascade, no successor link. 11
      sources across the estate name a blast radius — incident contracts, reliability engineering,
      the Tara workbench — and 0 of them are in the output registry. A CASCADE THAT STOPS AT THE
      RECORD IT WAS CALLED ON IS NOT PROPAGATION, and here there is no cascade to stop.
      
      4 of .g's 9 projection surfaces are even nameable from the output service: lineage, search,
      URLs and exports. Gallery, marketplace, collections, campaigns and caches appear nowhere in
      it, so a withdrawal could not be projected onto them coherently even if one could be issued.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) A delete is
      a SOFT delete — the row is marked `deleted` rather than removed, and every list, search and
      lookup query excludes it, so history survives a deletion. (2) The provenance backfill scan
      deliberately includes archived and expired outputs, which is right: historical records need
      provenance too, and a sweep is not a discovery query.
      
      TRAP. A VERB THAT MEANS SOMETHING ELSE IN ANOTHER DOMAIN IS NOT THIS VERB.
      `/v1/admin/hathor/npc-memory-system/recall` is an NPC retrieving a memory, and counting it
      reported a recall command the catalog has not got. TRAP. A SWEEP IS NOT A DISCOVERY QUERY:
      the expiry sweep also excludes `expired`, and aggregating status filters across the whole
      service reported discovery hiding expired records when the list paths do not. Reading the
      list/search condition arrays specifically gives 3 sites excluding 1 status, out of 7 status
      filters in the file.
      
      1 of 7 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`: one verb has a state
      and it means something else, four have no state at all, nothing links a successor, and
      nothing propagates to any downstream surface. -->
      
  • I8.9 Implement scheduled drops with generation and review capacity reservation. Evidence: overbook and missed-dependency tests.
    • I8.9.a Define scheduled-drop identity/version, timezone/locale, release window, targets, campaign bundle, required generation/gates/reviews, deadlines, dependencies, fallback, owner, and status. Evidence: drop schema.

    • I8.9.b Build a dependency DAG and backward schedule for generation, evaluation, review, correction, packaging, distribution, verification, and notification with buffers. Evidence: schedule fixtures.

    • I8.9.c Forecast and reserve provider/queue/GPU/storage/budget plus named reviewer/approver capacity for the required windows. Evidence: capacity reservation tests.

    • I8.9.d Reject overlapping/overbooked reservations and detect holidays/ working hours/region/target embargo conflicts with exact shortfalls. Evidence: overbook tests.

    • I8.9.e Monitor dependency completion/freshness/slippage/capacity and recalculate risk without silently moving the ratified release window. Evidence: missed-dependency tests.

    • I8.9.f Execute hold/delay/reduce-scope/cancel/fallback only through authorized impact preview and immutable schedule revision/decision. Evidence: contingency tests.

    • I8.9.g Trigger release only when current candidate/approvals/gates/ rights/capacity remain valid and verify all targets. Evidence: scheduled release journey. <!-- MEASURED 2026-08-20 (I8.9.a-g). The parent's evidence clause names two tests — "overbook and missed-dependency" — so the measurement asks those two questions first, of the surfaces that actually exist. Instrument: generate-oshun-drop-scheduling.mts (+10 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-drop-scheduling.json.

      A LONGEST PATH IS COMPUTED FROM DURATIONS, NOT DECLARED. The estate's one reachable
      dependency scheduler is `POST /v1/admin/yemaya/scheduling/schedule` over
      `@yemaya/budget-management`'s TimelineManager, and its `findCriticalPath` returned the tasks
      a caller had flagged `critical` — and, when none were flagged, every task carrying any
      dependency. It read no duration at all: given a(1) -> b(1) -> c(1) beside a lone d(100), it
      answered `["b","c"]`, which omits the head of the chain it names and drops the task 33x
      longer than that chain. Given a 1000-unit `render` and a 1-unit `trivial` flagged critical,
      it answered `["trivial"]`. The studio workspace joins that list with arrows and prints it as
      `critical path: b → c`, so A SET RENDERED WITH ARROWS READ AS A PATH. FIXED: the longest-
      duration chain through the graph, with `findCriticalPathDuration` beside it so the number is
      checkable, and the flag is deliberately not consulted. The route test that pinned the old
      answer — `expect(v.criticalPath).toEqual(['A','B']); // the flagged-critical tasks`, and a
      whole case named "falls back to dependency-carrying tasks when none are flagged critical" —
      WAS GRADING THE ECHO OF ITS OWN INPUT, and is replaced.
      
      FIXED, SAME FILE, TWO MORE. (1) `generateScheduleFromTasks` marked a task visited only AFTER
      recursing into its dependencies, so `x -> y -> x` recursed until the stack ended: a caller
      supplying two mutually-dependent task ids to an admin route got `RangeError: Maximum call
      stack size exceeded`, driven. Now Kahn ordering that raises `ScheduleDependencyCycleError`
      naming the tasks that could not be ordered; the route answers `400 dependency_cycle`.
      (2) `propagateDelay` recursed once per path, so A DIAMOND REACHED ITS JOIN TWICE: for
      a -> {b,c} -> d a 500ms delay returned `["b","d","c","d"]` and left d moved by 1000, driven
      before and after. Now breadth-first over a shifted set — `["b","c","d"]`, d at 500 — and it
      terminates on a cycle. (3) A dependency naming a task the request did not supply ordered
      nothing and was dropped in silence; it is now reported as `unresolvedDependencies` and named
      by the studio workspace, and a duplicate task id is a 400 rather than a quietly smaller
      schedule. THAT IS THE PARENT'S "missed-dependency" TEST, and it is the 1 of 13 properties
      this cell can record as present.
      
      A QUOTA IS NOT A RESERVATION AND A PRIVATE CALENDAR CANNOT COLLIDE. The nearest thing to a
      scheduled drop is `AdminEditorialReleaseStream` (embargo | hotfix | emergency_correction;
      pending | blocked | in_progress | released | cancelled), and its store keeps ONE BUCKET PER
      OPERATOR: all 5 read paths are `buckets.get(userId)`, and the only bucket-wide expression is
      `exportState(): EditorialReleaseStreamsPersistedState`, which no query calls. Two operators
      can hold streams over the same subject in the same window and never meet, so .d's
      overlapping-reservation refusal has nothing to refuse with. The one limit enforced is
      `ADMIN_EDITORIAL_RELEASE_STREAM_QUOTA_PER_USER` — a per-user count, not a capacity.
      
      5 of .a's 14 subjects are expressible over the record's 18 fields: identity, required
      reviews (a nullable `reviewPackageRef`), deadlines, owner, status. Absent: version,
      timezone/locale, release WINDOW (only single instants, `embargoLiftsAt` and
      `targetReleaseAt`), targets, campaign bundle, required generation, required gates,
      dependencies, fallback. 0 of .b's 8 stages — generation, evaluation, review, correction,
      packaging, distribution, verification, notification — is a value the scheduler holds; its
      tasks are opaque caller-supplied ids. No backward pass and no buffer quantity, so .b's
      backward schedule from a ratified window does not exist.
      
      1 of .c's 6 resources has a reservation the Oshun product can reach: budget, through
      `GenerationBudgetGate`. Provider, queue, storage and named reviewer/approver capacity have
      none. GPU reservations DO exist — `reserveGPU`, `reserveModel` — in two Isis libraries no
      Oshun source imports, so A RESERVATION IN A LIBRARY NOTHING IMPORTS HOLDS NOTHING. And 0 of
      12 `reserve*` sites can hold a FUTURE window: the budget gate's window key is
      `day:YYYY-M-D` derived from its own clock, so a drop cannot be promised capacity it will
      need on Friday. 0 of .d's 5 conflict classes is checked anywhere on the drop path.
      
      THE SLIPPAGE MACHINERY EXISTS AND THE CONSOLE CALLS NONE OF IT. TimelineManager declares 24
      methods and the Oshun store calls 6 — addTask, generateScheduleFromTasks, findCriticalPath,
      findCriticalPathDuration, findUnresolvedDependencies, getTaskCount. `assessScheduleRisk`,
      `recordDelay`, `propagateDelay`, `generateRecoveryPlan`, `compressSchedule`,
      `levelResources`, `negotiateDeadline`, `optimizePhaseOverlaps` and the whole milestone
      vocabulary are unreached by any route, so .e's monitoring is not absent from the estate — it
      is absent from the product. And where it does run, A SCHEDULE EDITED IN PLACE HAS NO
      BASELINE TO SLIP AGAINST: `recordDelay` writes `endDate: task.endDate + delayMs` onto the
      task and `propagateDelay` shifts every dependent, with no copy of the ratified dates kept
      beside the working ones. .e asks for risk recalculated WITHOUT silently moving the ratified
      window; the only way this library records a slip is by moving it.
      
      1 of .f's 5 contingency verbs is expressible: cancel. Hold, delay, reduce-scope and fallback
      appear nowhere on the record, `revision` is an in-place counter rather than a history, and
      there is no impact preview. A REMINDER IS NOT A RELEASE: the 4 surfaces carrying a publish
      moment (`content-drop-*`, `admin-reminders-produce`) send notifications and release nothing,
      so .g has no trigger to revalidate against and 0 of its 5 preconditions is asked.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The
      generation budget gate is a real reservation: `reserveSpendUsd` is atomic, `settle`
      reconciles the estimate against the provider's captured cost, and an unreachable store fails
      closed when configured to — its two refusals are `generation_daily_budget_exceeded` and
      `generation_budget_store_unavailable`. (2) The release-stream store enforces its quota,
      requires `embargoLiftsAt` at creation for embargo streams, refuses a blocked status with no
      blocker, and validates every status transition. (3) No project sharing
      `configs/vitest/node-test-coverage.ts` holds a `.spec.ts` file today, so its `.test.ts`-only
      include — a preserved contract for 35 projects — is skipping nothing; this cell added a
      local vitest config for the one project it needed rather than widening the preset.
      
      TRAPS, all four found by disagreeing with a first reading. (1) `= z` AND `.object({` SIT ON
      DIFFERENT LINES, so `indexOf('z.object({')` skipped this schema and anchored on the NEXT
      one: the census read the create-request envelope's 12 fields and reported them as the
      record's 18. (2) A WORD INSIDE A SENTENCE IS NOT A THING THE CODE HAS — a substring search
      reported `review` as a named stage (its one occurrence is the template `${n} subsequent
      milestones need review`) and a buffer in the schedule (its one occurrence is inside the
      justification string 'Extended buffer provides safety margin'). Whole string literals for
      values, identifiers-only for quantities. (3) AN `overlap` IN A TEXTURE EXECUTOR IS NOT A
      RESERVATION OVERLAP: estate-wide the pattern matches 57 files and 0 of them are on a
      schedule. (4) A PARAMETER NAMED `at` IS THE MOMENT THE RESERVATION IS MADE, NOT A WINDOW IT
      HOLDS — matching `at:` reported a forward reservation in a workbench cost fixture.
      
      1 of 13 properties present. ALL SEVEN CELLS AND THE PARENT LEFT `[ ]`: the drop record
      cannot say when its window opens or closes, the scheduler names none of the stages it would
      order, five of six resources cannot be reserved at all and none can be reserved ahead, no
      conflict class is checked, and nothing triggers a release. -->
      
  • I8.10 Implement rights/consent expiry monitoring for released assets. Evidence: warnings and automatic policy actions fire at known times.
    • I8.10.a Index exact rights/consent grant versions, effective/expiry/ review dates, territory/use/target scope, revocation, and dependent released assets/bundles/targets. Evidence: expiry index tests.

    • I8.10.b Calculate next action using timezone-safe instants and versioned warning/action windows, including missing/indefinite/unknown dates. Evidence: fake-clock boundary tests.

    • I8.10.c Run fenced idempotent monitoring with missed-run catch-up, duplicate prevention, clock skew, pagination, and production-scale throughput. Evidence: scheduler tests.

    • I8.10.d Notify configured owners/reviewers/legal/subjects before expiry with safe asset/scope/action/deadline details and escalation. Evidence: warning tests.

    • I8.10.e At action time revalidate grant/dependencies and automatically hold/unpublish/quarantine/recall only as canonical policy authorizes. Evidence: timed action tests.

    • I8.10.f Cancel/reschedule obsolete timers on renewal/replacement/ revocation/supersession and retain immutable event/action history. Evidence: lifecycle tests. <!-- MEASURED 2026-08-20 (I8.10.a-f). The parent's evidence is that "warnings and automatic policy actions fire at known times", so the binding question is asked first: is there any clock-driven, subject-independent reader of a rights or consent expiry at all? Instrument: generate-oshun-rights-expiry.mts (+10 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-rights-expiry.json.

      AN EXPIRY THAT ONLY HAPPENS WHEN THE SUBJECT LOOKS IS NOT A MONITOR. The transition itself is
      real and complete: `applyExpirations` in `apps/oshun/bff/src/consent/state.ts` marks a
      `granted` record `expired` once `expiresAt` has passed, raises `renewalRequired`, stamps
      `renewalRequestedAt`, and appends a typed `policy_expiration` history event with
      `actorType: 'system'`. What it is not is a monitor. All 7 of its call sites are request
      handlers — getSnapshot, getFlowState, grant, withdraw, revoke, renew, getAudit — and every
      one takes a caller-supplied `userId` and reads only `this.users.get(userId)`. A SWEEP THAT
      TAKES A SUBJECT ID IS NOT A SWEEP. A consent that lapsed six months ago is still `granted` in
      the store until that subject next touches a consent endpoint, and nothing anywhere fires at
      the moment `expiresAt` names. 0 timers, 0 cron entries, 0 interval sweeps on this path.
      
      A FENCED SWEEP EXISTS IN A SERVICE THAT HOLDS NO RELEASED ASSETS. `apps/yemaya/svc-study-
      workspace` has exactly what .c describes — `STUDY_EXPIRY_SWEEP_INTERVAL_MS` bounded to a
      60s floor with 0 disabling the timer, plus `POST /api/study/expiry-sweep` for a manual or
      catch-up run. It sweeps study-report retention, not rights or consent, and it is one of 2
      interval-sweep services in the estate, neither under `apps/oshun`. On the Oshun consent path
      0 of .c's 5 properties — fenced, idempotent, missed-run catch-up, duplicate prevention,
      pagination — is present, because there is no run to make idempotent.
      
      A GRANT CANNOT NAME WHAT WAS RELEASED UNDER IT. 7 of .a's 10 index subjects are expressible,
      and the record is better than expected: `record.version`, `record.grantedAt`,
      `record.expiresAt`, `record.target`, `record.revokedAt`, and — on the permission grant
      itself — `permission.allowedRegions` for territory and `permission.useScopes` for use. What
      is missing is the half that makes this cell about RELEASED ASSETS: no review date, no
      dependent released assets, no dependent bundles or targets. Nothing links a grant to the
      outputs published under it, so .e's "revalidate grant AND DEPENDENCIES" has nothing to walk
      and .a's dependency index does not exist.
      
      THE PROJECTOR THE ESTATE DOES HAVE IS CALLED BY NOTHING. `libs/contracts/src/common/
      entitlement-inspection.ts` contains a genuine warning-window calculator — `projectGrantState`
      with a configurable `expiringSoonHours` over `DEFAULT_EXPIRING_SOON_HOURS` — and 0 files
      outside its own module call it. The consent path does not reach it, and there is no warning
      window of any kind on the consent state, so .b's versioned warning/action windows have no
      implementation to version.
      
      0 of .d's warnings fire. 31 messaging producers live under `apps/oshun/bff/src/reminders` and
      `libs/oshun/messaging-channels`, and 0 of them names a rights, consent, licence or grant
      expiry. No owner, reviewer, legal or subject is addressed, and there is no escalation.
      
      0 of .e's 4 automatic actions exists to be taken. `OutputStatus` is
      pending | available | archived | expired | deleted, so hold, unpublish, quarantine and recall
      are not states an output can be in — §I8.8's finding, re-measured here so this cell stands on
      its own evidence — and the consent path invokes none of them. .f's obsolete timers cannot be
      cancelled because none was ever scheduled; 3 of its 4 invalidating events ARE recorded on the
      record (`renewalRequired`, `revocations`, `supersededById`; replacement is not spelled).
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The expiry
      transition is honest end to end: status, renewal flag, renewal timestamp and a typed history
      event with actor, prior status and new status. (2) `computeExpiry` returns null when the
      taxonomy sets `defaultRenewalDays` to null, so an indefinite consent gets no invented expiry
      — the taxonomy's cadences are 30, 180, 365, 730 and null. (3) It is epoch-millisecond
      arithmetic on an ISO instant with no local-calendar walk, so the boundary does not shift
      across a DST transition. That is .b's timezone-safe instants and its missing/indefinite
      dates, and it is 1 of the 2 properties this cell can record present. (4) The consent history
      is append-only through a private `appendHistory` over an 11-member typed vocabulary
      including `expired`, `renewal_requested`, `renewed` and `superseded` — .f's immutable event
      history, and the other present property.
      
      TRAPS, all four found by disagreeing with a first reading. (1) AN EMPTY PARSE AND A MEASURED
      ABSENCE LOOK IDENTICAL: `OutputStatusSchema` and `ConsentHistoryEventTypeSchema` are
      `z.enum([...])`, not `as const` arrays, so reading them the wrong way reported an 11-event
      history as no history and made "0 of 4 policy actions" an unmeasured zero that happened to
      match the truth. Both parses now throw on an empty result. (2) A SCHEMA NAME GUESSED IS A
      POPULATION OF ZERO — `ConsentPermissionSchema` and `ConsentCollectionSchema` do not exist
      (they are `…PermissionGrantSchema` and `…CollectionContextSchema`), so the grant's own
      territory and use scope read as absent from a contract that carries both. (3) WHERE A
      CONSENT WAS COLLECTED IS NOT THE TERRITORY IT COVERS: an identifier sweep over the whole
      contract answered territory scope with `collection.geoCountry`, which records where the
      subject was standing. Fields are now read off named schemas and carry the object they sit
      on. (4) AN EMAIL SUBJECT IS NOT A DATA SUBJECT — a proximity match over the messaging
      surface reported that subjects are warned; the hits are a webPush VAPID `subject` and a
      per-locale reminder `subject` line for something else.
      
      2 of 10 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: the expiry moment is
      known and nothing is watching it, no warning window is computed on this path, no notification
      names an expiry, no policy action exists to take, and the grant cannot say which released
      asset depends on it. -->
      
  • I8.11 Implement release notification and incident routing. Evidence: protected metadata is redacted.
    • I8.11.a Define versioned event-to-recipient/channel/template/escalation rules for planned/succeeded/partial/failed/verified/recalled releases and incidents. Evidence: routing schema.

    • I8.11.b Resolve recipients by tenant/project/target/role/on-call/ subscription/purpose and reauthorize protected detail at delivery time. Evidence: recipient tests.

    • I8.11.c Render localized accessible templates from allowlisted audience- safe fields with explicit status, asset/release ref, target, time, action, owner, and deep link. Evidence: template tests.

    • I8.11.d Redact biometric/child/source/provider/credential/private prompt/ rights evidence and cross-tenant metadata from subject/body/attachments/ URLs. Evidence: notification leak suite.

    • I8.11.e Deliver idempotently over in-app/email/webhook/paging channels with retry/backoff/rate limits/dedup/threading/preferences and receipts. Evidence: delivery integration tests.

    • I8.11.f Escalate unacknowledged/high-severity/failed routes by policy and preserve delivery/acknowledgement/audit without treating notification as remediation. Evidence: escalation tests. <!-- MEASURED 2026-08-26 (I8.11.a-f). The parent's evidence is one clause — "protected metadata is redacted" — so the binding question is asked of the surface that would have to redact: what does the router actually carry? Instrument: generate-oshun-release-notification.mts (+10 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-release-notification.json.

      A ROUTER THAT CARRIES NO CONTENT CANNOT LEAK IT, AND CANNOT DELIVER IT EITHER.
      `POST /v1/admin/studio/notification-routing/route` takes exactly three fields per
      notification — `notificationId`, `priority`, `channel` — plus `{ withinQuietHours }` and
      `{ quietHoursMinPriority, digestLowPriority }`, and answers
      `deliver_now | digest | suppressed`. There is no event kind, no recipient, no template, no
      locale and no body anywhere in it. Nothing protected passes through, so the parent's clause
      is UNFALSIFIABLE on the one surface that routes, and 1 of .a's 4 rule targets — channel —
      is bound, and that one is supplied by the caller rather than resolved by a rule.
      
      0 OF THE 7 EVENTS THE RULES MUST ROUTE IS A VALUE ANYTHING HOLDS. Planned, succeeded,
      partial, failed, verified, recalled and incident appear in no vocabulary on this path. The
      template contract's nearest thing is `lifecycle` — transactional | lifecycle | announcement |
      crisis | billing | support — which classifies a MESSAGE, not a release outcome. So .a's
      event-to-rule table has no events to key on.
      
      THE ON-CALL CATALOG EXISTS AND THE ROUTER CANNOT REACH IT.
      `apps/oshun/bff/src/channel-ops/oncall-schedule-catalog.ts` has exactly 1 importer,
      `routes/admin-oncall.ts`, and the routing store and route name no on-call anything. 0 of
      .b's 7 recipient dimensions — tenant, project, target, role, on-call, subscription, purpose
      — is present on the notification path, and nothing reauthorizes protected detail at delivery
      time, because nothing resolves a recipient at all.
      
      THE TEMPLATE SURFACE IS THE STRONG ONE, AND THIS CELL REPAIRED TWO HOLES IN IT.
      `AdminNotificationTemplate` carries per-locale subject and body across 6 channels, a declared
      variable set, A/B variants with traffic share, a four-state review lifecycle, and a
      deterministic lint pass over 7 rules including alt text, unbalanced HTML, reading grade and
      an unsubscribe requirement. Two of those rules did not do what they said.
      
      FIXED. AN `alt` ANYWHERE IN THE BODY EXCUSED EVERY IMAGE IN IT. `missing_alt_text` tested
      `body contains an <img>` AND `body contains no alt= anywhere`, so a body holding one
      described image and one undescribed one linted clean — driven. Each `<img>` is now asked for
      its own `alt`, and the message counts how many lack one.
      
      FIXED. AN ALLOWLIST THAT CANNOT SEE A REFERENCE CANNOT REFUSE IT. The variable-reference
      regex was `/\{\{(\w+)\}\}/`, which matches neither `{{rights.evidenceUrl}}` nor
      `{{provider-key}}` nor `{{ name }}`. An undeclared reference containing a dot, a hyphen or a
      space was not collected and therefore never raised `forbidden_variable` — and .d's protected
      classes are spelled exactly that way. Driven: both passed lint clean before, both are errors
      now. The lint runs on the real create and update path in
      `admin-notification-templates-store.ts`, and no committed template uses such a reference, so
      the allowlist tightened without breaking one.
      
      0 OF .c's 7 REQUIRED FIELDS IS REQUIRED. Status, asset/release ref, target, time, action,
      owner and deep link are not fields the contract obliges a template to render; `variables` is
      free-form per template. So a release notification can be authored that states none of them.
      
      1 OF .d's 8 PROTECTED CLASSES IS EVEN NAMEABLE on the notification path, and it is
      `provider`, appearing only as `providerMessageId` on the push payload. Biometric, child,
      protected source, credential, private prompt, rights evidence and cross-tenant metadata are
      not words this path knows, and A CLASS THE PATH CANNOT NAME IS ONE IT CANNOT WITHHOLD. There
      is no redactor on any of it, across all 4 carriers .d lists. The estate's designed
      allowlist projector — `projectAllowed` in `libs/oshun/workbench-kit/src/redaction.ts`, built
      for S8.12 with a disclosure ladder and a withhold-by-default register — has 0 product
      callers.
      
      5 OF .e's 8 DELIVERY PROPERTIES are present across the messaging stack: retry, rate limit,
      threading, preferences and receipts. Idempotency, backoff and dedup are absent, so a
      redelivery after a partial failure is not distinguishable from a second notification.
      
      ESCALATION IS ON THE INCIDENT LEDGER AND NOT ON THE ROUTER. `/operator/incidents` has all
      five verbs — acknowledge, assign, escalate, snooze, resolve — and the router has none of
      them, so .f's "escalate unacknowledged and failed ROUTES" has no route state to escalate
      from.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The lint
      pass is honest about what it cannot know: `low_contrast_color` detects inline colours and
      asks an operator to verify WCAG AA rather than returning a contrast verdict it did not
      compute, and `hasUnbalancedHtml` documents that it errs toward false positives. (2)
      `buildPushPayload` refuses a silent push carrying user-visible actions rather than dropping
      them quietly. (3) A NOTIFICATION IS NEVER TREATED AS REMEDIATION here — resolving an
      incident requires a resolution class and a rationale from the caller, and nothing closes one
      because a message went out. That is the one property of the twelve that this cell can
      record present on the product's own behaviour.
      
      TRAPS. (1) A CONCEPT SPELLED INSIDE A COMPOUND IS STILL SPELLED: `\bpreferences\b` never
      matches `NotificationPreferences` and `\bretries\b` never matches `maxRetries`, so a
      word-boundary search reported a messaging stack containing a `preferences.ts` module as
      having no preferences and put .e at 3 of 8 instead of 5. Identifiers are now segmented on
      camelCase. (2) AN HTML `<link>` TAG IS NOT A DEEP LINK — the bare word matched `'link'` in
      the lint pass's self-closing-tag set and reported .c's deep link as a required field. (3) AN
      ABSENT ANTI-PATTERN IS ONLY EVIDENCE IF THE THING IT WOULD CORRUPT EXISTS: "notification is
      not remediation" was true of a file before the incident ledger's five verbs were read, which
      would have been true of a file with no resolution path at all.
      
      3 of 12 properties present. ALL SIX CELLS AND THE PARENT LEFT `[ ]`: the router carries three
      fields and none of them is an event or a recipient, no protected class is nameable let alone
      withheld, no required field is required, and delivery cannot tell a retry from a repeat. -->
      
  • I8.12 Implement recovery of failed multi-target release. Evidence: compensation and retry preserve idempotency.
    • I8.12.a Persist per-target/per-step planned/executing/written/verified/ failed/unknown/compensating/rolled-back state and immutable receipts. Evidence: release state tests.

    • I8.12.b On interruption, reconcile authoritative destination state by idempotency key/version/hash before deciding retry or compensation. Evidence: crash recovery tests.

    • I8.12.c Classify targets/steps as safely retryable, compensatable, manually recoverable, or irreversible with exact reason. Evidence: recovery policy matrix.

    • I8.12.d Retry only missing/failed idempotent writes against the unchanged candidate/plan and never duplicate verified successful targets. Evidence: partial retry tests.

    • I8.12.e Compensate configured successful targets in dependency-safe order and verify removal/reversion while recording residual copies. Evidence: compensation tests.

    • I8.12.f Require fresh authorization/evidence/rights/target eligibility and impact preview for resume/retry/compensate/manual-complete commands. Evidence: stale recovery tests.

    • I8.12.g Derive final succeeded/partial/failed/rolled-back/manual state honestly and retain every attempt/receipt/cost/notification. Evidence: recovery history tests. <!-- MEASURED 2026-08-26 (I8.12.a-g). The parent's evidence is one clause — "compensation and retry preserve idempotency" — and both nouns presuppose a third: a release that reached more than one target and stopped part way. So the binding question is asked of the fan-out first. Instrument: generate-oshun-release-recovery.mts (+19 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-release-recovery.json.

      A RELEASE WITH ONE DESTINATION HAS NO PARTIAL FAILURE TO RECOVER FROM. Across the six files
      the preceding cells established as the release path — I8.4's publish store and its route,
      the go/no-go contract, the two publishers an approved item reaches, and I8.6's one
      destination that reads its write back — there are 0 declared collections of targets, steps
      or attempts. Not a short list: no list. So there is no per-target state to persist, no
      subset to retry and no successful target to compensate, and the parent's clause is
      UNFALSIFIABLE for the same reason I8.11's was — the thing it constrains does not pass
      through. The detector is not blind: asked the same question, it finds `steps` in
      `workbench-kit/publishers.ts`, `targets` in `withdrawal.ts` and `attempts` in the wearable
      store.
      
      NO VOCABULARY IN THE ESTATE CAN SAY WHAT .a's RECORD IS SUPPOSED TO HOLD. Every declared
      `as const` array, `z.enum` and string union in the 141 non-spec sources that spell a
      compensation or rolled-back state was scored against .a's eight. The best is 3 of 8 —
      `libs/aje/account-abstraction/src/eip7702/types.ts#MigrationStatus` — and none of the top
      five sits on a publishing surface. The release path itself reaches 1 of 8, and that one hit
      is I8.6's read-back verdict `verified | unsupported_algorithm | storage_unreachable`, which
      says `verified` about bytes rather than about a target.
      
      THE ESTATE DESIGNED ALL OF I8.12 AND NO PRODUCT CODE CALLS ANY OF IT.
      `libs/oshun/workbench-kit/src/publishers.ts` (S6.9) is .a-.g almost line for line: a plan
      that is a value with a digest, a per-step durable record with idempotency key and
      request/response digests, an execute admission that rechecks the lock, the capability, the
      gates, rights, consent and target health, a read-back over seven dimensions including
      discoverability, a compensation order that runs in reverse and stops at the first
      irreversible step, and a reconciler for an outcome nobody observed. 9 analyzers, 59 named
      fault kinds, 0 callers outside the kit. 6 of 6 receipt fields are in it and 0 are on the
      release path; 3 of .c's 4 classes are in it and 0 are on the release path.
      
      EVEN THE DESIGN HAS NO STATE FOR A COMPENSATION IN FLIGHT. Its `STEP_STATES` are pending,
      sent, succeeded, failed, unknown, compensated, compensation-failed — 2 of .a's 8. `compensated`
      is the terminal one; .a asks for `compensating`, which is the state a crash DURING the undo
      leaves behind, and it is the one state whose absence makes a crashed compensation
      indistinguishable from one that never started.
      
      THE IDEMPOTENCY CLAIM IS WIRED, AND NOT TO ANYTHING THAT PUBLISHES. The kit's idempotency
      module is imported by `assistant/workbench-kit-read.ts` and enforced on the five commands
      `workbench-kit-commands.ts` exposes — capture, promote and archive a spark, transition and
      schedule a concept. 0 of the five publishes or releases.
      
      0 OF .f's FOUR RECOVERY COMMANDS IS A ROUTE ON A RELEASE, over 1597 route literals mounted
      under `/v1/` and the legacy `/admin/` prefix. `resume` exists three times — iris
      continuation tokens, a psyche adapter session, a routines execution — and retry, compensate
      and manual-complete exist zero times. A whole path segment is the rule, so
      `/analytics-reporting/retry-stats` and `/multi-agent-orchestration/retry-delay` are recorded
      as seen and not counted. Of .f's five freshness subjects only rights is nameable on the
      release path, and there is no impact preview because there is no command to preview.
      
      .g CANNOT BE DERIVED BECAUSE ITS OUTCOMES DO NOT EXIST. No release contract anywhere spells
      `partial`; the best final-state vocabulary in the estate reaches 3 of 5 and belongs to
      `libs/psyche/disaster-recovery`. Of .g's four retained subjects the release path carries
      receipt and cost and not attempt or notification — and I8.11 already established that
      nothing notifies about a release outcome at all.
      
      THE CONTROL: THE ESTATE CAN BUILD THIS AND HAS, ON A PATH THAT IS NOT A RELEASE.
      `POST /v1/wearable/health-sessions` → `wearable-health-store.deliverSessionDurably` is a
      routed two-target durable delivery. Its pending per-target attempt reaches durable storage
      BEFORE either platform is written; both operation ids derive from the durable session id so
      the key does not move between attempts; a replay returns the recorded platform record for a
      target already written instead of calling the bridge again; the partial outcome names both
      the failures and the delivered; and the erasure path deletes at both platforms by that same
      operation id, covers partial attempts as well as completed sessions, and throws rather than
      removing the local rows when a platform delete fails — it never records an erasure it did
      not perform. That is .d and most of .e, working, in this codebase, for health data. What it
      still cannot do is the same thing the release path cannot: nothing declares whether either
      platform HONOURS the key it is sent, and nothing sweeps — the resume happens only when the
      caller sends the same session again.
      
      FIXED. A CHECKER MUST NOT FAULT THE DECIDER'S OWN CORRECT ANSWER. `reconcilePublish` returns
      `retry-under-the-same-key` with no probe when the target cannot be asked and the adapter
      declares `returns-the-first-result` — the module's own test calls it "the one case where a
      retry IS the probe" — and feeding that record straight into `publishReconciliationFaults`
      produced `unknown-outcome-retried-before-a-probe`. Driven. The fault now excludes a target
      that promises the first result, because under that declaration repeating the request is how
      you ask; it still fires for `rejects-the-duplicate`, `ignores-the-key` and `unknown`.
      
      FIXED. AN EFFECT THAT STAYS BEHIND A BLOCKED ONE BECAME NOBODY'S TASK. Decision 4 says every
      effect that stays becomes an instruction naming the step, the reference and what a person
      must do. `compensationOrder` kept that promise for the step that blocked and dropped it for
      everything preserved behind it — so a plan with two irreversible steps preserved both and
      instructed about one, and feeding its own output back to `compensationFaults` raised
      `irreversible-effect-with-no-manual-instruction`. Driven. A COMPUTED PLAN MUST PASS ITS OWN
      CHECKER: it does now, and the checker still reports the plan when nobody was told anything.
      
      FIXED. A MAP KEYED BY STEP OVER A LOG OF ATTEMPTS HOLDS WHICHEVER ATTEMPT CAME LAST.
      `executionFaults` reads one row per ATTEMPT — `retry-that-changed-the-idempotency-key`
      exists precisely because it does — and built `new Map(executions.map(…))` to answer the
      dependency question, so it read whichever row the caller put last. The same three-row log —
      upload failed, upload succeeded, publish — reported `step-executed-out-of-dependency-order`
      newest-first and nothing oldest-first. Driven both ways. A dependency is now satisfied when
      ANY attempt of it succeeded, and the finding is raised once per pair rather than once per
      attempt.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1)
      `reconcilePublish` never settles an unknown outcome by guessing: with no probe and a target
      that ignores the key it escalates to a person, and `mark-failed-from-the-probe` is in the
      vocabulary ONLY so the checker can report a system that took it. (2) `compensationOrder`
      stops at the first irreversible step counting backwards and preserves everything before it,
      which is the right answer and not the convenient one. (3) The wearable erasure fails loud
      rather than reporting a removal it did not perform.
      
      TRAPS. (1) A CLASSIFICATION IS A VALUE A RECORD HOLDS, NOT A FIELD NAME: an identifier scan
      blanks literal contents, so `REVERSIBILITY = ['reversible', 'reversible-with-residue',
      'irreversible', 'unknown']` read as absent from the module that declares it, and .c went to
      0 of 4 in the design instead of 3. Classes are now read from declared vocabularies and whole
      literals. (2) `safely retryable` is genuinely not a stored class even so — the design
      DERIVES it from the adapter's declared idempotency at the moment it is needed, which is
      recorded rather than scored as a hole. (3) A WHOLE PATH SEGMENT, NOT A PREFIX: `/retry-stats`
      reports on retries and is not one. (4) AN EMPTY PARSE AND A MEASURED ABSENCE LOOK IDENTICAL —
      the population and both rankings throw when they come back empty, because every file in the
      population spells a state both required lists contain.
      
      0 of 12 properties present. THE PARENT AND ALL SEVEN CELLS LEFT `[ ]`: the release path has
      no second target, so it has no partial failure; nothing anywhere holds .a's states; the
      module that holds the design has no callers; and no route lets anyone resume, retry,
      compensate or manually complete a release. -->
      
  • I8.13 Build campaign/program steering overlays that may tighten workflow, quality, safety, rights, budget, and audience limits but never loosen platform floors. Evidence: property tests and impact preview.
    • I8.13.a Define versioned overlay identity/scope/priority/lifecycle and typed workflow allow/deny, quality threshold, safety/rights requirement, budget/quota, audience/target, schedule, and owner fields. Evidence: overlay schema.

    • I8.13.b Resolve applicable platform/tenant/program/campaign/project policies in canonical precedence with conflict and cycle detection. Evidence: resolution tests.

    • I8.13.c Merge overlays only by intersection, stricter threshold, lower ceiling, additional requirement, or narrower audience/target semantics. Evidence: field-level merge property tests.

    • I8.13.d Reject any direct/indirect/default/omission/inheritance attempt to loosen code-owned safety, rights, consent, privacy, security, spend, or authority floors. Evidence: monotonicity property tests.

    • I8.13.e Preview affected workflows/models/controls/gates/budgets/users/ scheduled work/releases, conflicts, and new blockers before activation. Evidence: impact preview tests.

    • I8.13.f Review/approve/version/schedule/deactivate with role separation, exact effective dates, audit, rollback to prior stricter state, and dependent revalidation. Evidence: overlay lifecycle tests. <!-- MEASURED 2026-08-26 (I8.13.a-f). The parent's evidence is two artifacts — property tests and an impact preview — over one claim: an overlay may TIGHTEN a limit and may never LOOSEN a platform floor. That claim presupposes three things and this cell asks for each: an overlay object, a floor for it to sit above, and a merge whose result is provably no looser than either input. Instrument: generate-oshun-steering-overlays.mts (+16 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-steering-overlays.json.

      A CAMPAIGN THAT ASKS PEOPLE TO SIGN A POLICY IS NOT A CAMPAIGN THAT STEERS ONE. The only
      thing in the estate called a campaign is `AdminPolicyAttestationCampaign` — 20 fields that
      assign a policy to an audience, count required attestations and track a compliance target
      percentage. 0 of the 20 is a threshold, a budget, a quota, an allow/deny or a limit. It is a
      campaign in the sense an email campaign is one.
      
      THE NEAREST REAL STEERING OBJECT IS THE POLICY BUNDLE, and it declares 28 fields covering 7
      of .a's 11 subjects. The four it misses are the four that make an overlay an overlay:
      PRIORITY, BUDGET/QUOTA, SCHEDULE and OWNER. An owner exists on the operator-side editor
      record as `ownerTeam` and not on the contract that travels, so the shipped overlay has
      nobody attached to it.
      
      NO PRIORITY MEANS NO PRECEDENCE, AND NONE OF .b's FIVE SCOPES IS A SCOPE. The scope
      vocabulary is global | domain | surface | persona | workflow_template | model_version |
      region | experiment and the 13 target kinds add provider, entitlement, feature_flag,
      memory_scope and policy_bundle. Platform, tenant, program, campaign and project are 0 of
      21. There is no priority field, so two applicable bundles have no declared order between
      them, and there are 0 resolvers in the estate — nothing anywhere picks among applicable
      bundles, so the precedence .b asks to be canonical is not computed at all.
      
      A BUNDLE MAY TARGET A BUNDLE AND NOTHING LOOKS FOR A CYCLE. `policy_bundle` is a target
      kind, so a → b → a is expressible, and the contract has neither cycle nor conflict
      detection.
      
      0 MERGE SITES. There is no `mergePolicyBundles`, no overlay intersection, nothing that
      combines two applicable policies — so .c's five permitted moves (intersection, stricter
      threshold, lower ceiling, additional requirement, narrower audience) have nothing to
      constrain, and .d's monotonicity property has no function to be a property OF.
      
      A FLOOR IS A VALUE A CHANGE IS COMPARED AGAINST, NOT A SUBJECT THE SCHEMA HAS A FIELD FOR.
      The overlay's validation is 16 `superRefine` passes raising 82 distinct refusal messages,
      and every one was read: not one refuses a value for being below a minimum. There are 0
      `MIN_`/`FLOOR_` constants in the contract, and no mutation reads the prior policy before
      writing the new one, so nothing can tell a tightening from a loosening. 3 of .d's 7 floors —
      privacy, security and spend — are not even expressible on the overlay; 4 are named and 0 are
      enforced.
      
      THE ONE SPEND LIMIT ON THE GENERATION PATH IS A SINGLE UNSCOPED NUMBER THAT IS OFF WHEN
      UNSET. `resolveGenerationBudgetConfig` reads `OSHUN_GENERATION_BUDGET_DAILY_USD`; a value
      `<= 0` disables the gate entirely, and there is no tenant, project, program or campaign
      dimension. So "an overlay may never loosen the spend floor" is vacuous twice over: there is
      nowhere to attach a tighter number, and by default there is no floor.
      
      9 PREVIEW SURFACES ARE MOUNTED AND 0 OF THEM PREVIEWS A POLICY. Over 1597 route literals the
      BFF mounts under `/v1/` and the legacy `/admin/` prefix, the preview surfaces are an
      incident impact view, an Aja avatar preview, an Isis live preview, a OneRoster preview, a
      digest preview and a Nyx ritual preview. None names a policy, an overlay, a bundle, a
      campaign or a program, so .e's ten subjects have no surface to be shown on.
      
      2 OF .f's FIVE LIFECYCLE VERBS ARE COMMANDS. Review and approve exist as store mutations;
      version, schedule and deactivate do not. The approver is stamped straight from the caller
      and compared to nobody, so there is no role separation on the one approval that exists —
      consistent with §I8.3's finding that the estate refuses a self-approval in 19 places and
      none of them is a review signoff. There are no effective dates because there is no schedule
      field, and no dependent revalidation.
      
      FIXED. A PRODUCTION PROMOTION WAS GATED ON A STRING BEING NON-EMPTY.
      `promoteAdminPolicyBundle` refused a production promotion unless
      `governance.rollbackBundleId !== null`. The gate exists so nothing reaches production
      without a state to fall back to, and it was satisfied by text: the seeded consumer-default
      bundle sits in production on the strength of `policy-pack-consumer-default-v2`, and the
      workspace holds three bundles, none of which is it. `rollbackAdminPolicyBundle` had the same
      hole — it stored whatever string it was handed and appended a `rolled_back` event naming it.
      Both now resolve the target against the workspace's bundles and refuse with
      `rollback_target_not_found`.
      
      FIXED. THE SAME DEFECT IN THE MODEL REGISTRY. `promoteAdminModelGovernance` and
      `rollbackAdminModel` accepted any non-empty `rollbackModelId`. Fixed the same way. The
      contrast is the useful part: the model seed DOES retain its prior version
      (`model-oshun-gpt-5-3`, which `model-oshun-gpt-5-4` points at), so its rollback test still
      exercises the success path against a target that exists — the same convention, in the same
      file, that the policy seed does not follow.
      
      FIXED. THE ESTATE'S OWN TEST CALLED A PHANTOM ID KNOWN-GOOD. `rolls back the
      consumer-default production rollout to a known-good bundle` asserted a 200 for
      `policy-pack-consumer-default-v2`. It now asserts the refusal, and a second case exercises
      the success path against a bundle the workspace holds — with the note that the store has no
      lineage between bundles, so the rule it can enforce is "the target is a bundle this
      workspace holds" and not "the target is a prior version of this one". .f's "rollback to
      prior stricter state" needs both a lineage and a strictness order and the model has neither.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The
      allowlists surface is a real workflow allow/deny: workflowTemplateIds, modelVersionIds and
      providerIds allow, blockedProviderIds denies, every list is uniqueness-checked, and a
      provider appearing in both lists is refused rather than silently resolved. (2) The
      governance schema will not let promotion metadata exist without an approval, will not let
      `promotedAt` precede `approvedAt`, and requires a changeSummary for an emergency override.
      (3) The generation budget gate reserves before the work and settles after, and its
      reservationId is the job id so a retried admission cannot consume the same estimate twice.
      
      TRAPS. (1) A SUBSTRING IS NOT THE CONCEPT: `/cycle/` matches `lifecycle`, and a contract
      declaring a lifecycle enum reported cycle detection it does not have. Concepts are matched
      on identifier SEGMENTS. (2) A bare `loosen|monotonic` sweep returned 77 files, almost all of
      them monotonic clocks; the monotonicity question is now asked of the overlay's own refusal
      messages instead. (3) A FIELD BAG IS NOT THE WHOLE SCHEMA — .a's subjects are read off the
      declared `PolicyBundleFields` key set rather than looked for as words in a 1686-line file.
      (4) AN OWNER ON THE EDITOR RECORD IS NOT AN OWNER ON THE OVERLAY.
      
      0 of 13 properties present. THE PARENT AND ALL SIX CELLS LEFT `[ ]`: no campaign or program
      is a scope, no overlay carries a priority or a budget, nothing merges two policies, no floor
      is a value anything is compared against, no surface previews a policy change, and three of
      the five lifecycle verbs do not exist. -->
      
  • I8.14 Build UI-governed asset-kind, use-case, campaign, style, delivery, and gallery taxonomy registries over code-owned floors. Evidence: versioned review, alias, deprecation, localization, and migration journey.
    • I8.14.a Define separate typed registry entry/version/lifecycle contracts for asset kind, use case, campaign, style, delivery target/profile, and gallery taxonomy with stable code. Evidence: registry schemas.

    • I8.14.b Declare code-owned noneditable identities, safety/rights/ compatibility semantics, required fields, forbidden relationships, and fallback behavior as registry floors. Evidence: floor tests.

    • I8.14.c Build operator list/detail/create-draft/edit/diff/validation UI with typed fields, restrained hierarchy, ownership, dependencies, usage, and no arbitrary JSON. Evidence: authoring tests.

    • I8.14.d Implement alias creation/resolution with scope, uniqueness, collision/cycle checks, canonical identity retention, and no silent semantic merge. Evidence: alias tests.

    • I8.14.e Implement localization for labels/descriptions/help/search terms with locale fallback, completeness, translator context, and no localized identity keys. Evidence: localization tests.

    • I8.14.f Review/approve/publish/deprecate/supersede/archive registry versions with role separation, effective dates, audit, and consumer compatibility checks. Evidence: lifecycle journey.

    • I8.14.g Preview and execute resumable migrations of dependent briefs/ assets/workflows/collections/releases with exact mapping, conflicts, unchanged history, and rollback. Evidence: migration journey.

    • I8.14.h Verify unknown/deprecated/missing entries fail or degrade safely in every consumer, with contract/parity/accessibility/performance tests. Evidence: cross-host registry suite. <!-- MEASURED 2026-08-26 (I8.14.a-h). The parent asks for six REGISTRIES over code-owned floors, with a versioned review, an alias, a deprecation, a localization and a migration journey as its evidence. A registry is a thing entries are added to and retired from under a lifecycle, so the binding question is what shape each of the six vocabularies actually has. Instrument: generate-oshun-taxonomy-registries.mts (+14 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-taxonomy-registries.json.

      A CONST ARRAY IS NOT A REGISTRY. 0 of the six is one, and one of them does not exist:
      asset kind is `ASSET_KINDS = ['image', 'video'] as const` in a rendition store; use case is
      a 6-member `THREE_D_QUALITY_USE_CASE_VALUES` array in the 3D quality-gates library;
      campaign is §I8.13's 20-field attestation interface; style is a 6-entry
      `Readonly<Record<string, MediaStyleGuide>>`; delivery target is a 4-member Nyx reminder
      channel array; gallery taxonomy is nothing at all. A new asset kind is a code change and a
      deploy. The shape detector is not blind — it tells a const array, a z.enum, a record map and
      an interface apart, and is asserted doing so.
      
      A NOTIFICATION CHANNEL IS NOT A DELIVERY TARGET. The only thing in the estate spelled
      `DELIVERY_TARGETS` is `['mobile-push', 'email', 'sms', 'calendar-sync']` on the Nyx
      event-action store — where a reminder goes, not what a release is delivered to.
      
      0 OF .a's THREE ENTRY PROPERTIES IS ON ANY OF THE SIX. None carries a stable code, a version
      or a lifecycle, and a sweep of every contract file naming one of the six that declares BOTH
      a version and a lifecycle returns 0 outside the study contracts.
      
      1 OF .b's FIVE FLOORS is declared on the six declarations, and it is "required fields".
      Recorded beside it because it is real and sits outside any declaration: the style map
      resolves an unknown preset through a named, documented `DEFAULT_STYLE_GUIDE_KEY =
      'editorial-illustration'` rather than returning undefined.
      
      0 ROUTES GOVERN ANY OF THE SIX, over 1597 route literals. Two taxonomy routes exist —
      `/v1/admin/hathor/lore-validation-taxonomy` and its `/validate` — and neither is one of the
      six, so .c's six authoring acts, .f's six lifecycle verbs and .g's migration have 0 surfaces
      between them. 0 aliases and 0 translations on any of the six, so .d and .e have nothing to
      resolve or localize.
      
      A REFUSAL SOMEWHERE IS NOT A REFUSAL EVERYWHERE. .h asks that an unknown, deprecated or
      missing entry fail or degrade safely IN EVERY CONSUMER; each of the three cases is covered
      by exactly one of the six. Unknown: the asset-rendition route refuses a kind outside
      `ASSET_KINDS` at the edge. Deprecated: only the campaign interface has a status field at
      all — A VOCABULARY WITH NO LIFECYCLE HAS NO DEPRECATED STATE TO DEGRADE FROM. Missing: the
      style map's documented default.
      
      THE GOVERNED REGISTRY EXISTS AND IS WIRED TO ANOTHER WORKBENCH. Yemaya's study taxonomy is
      every one of .a-.g: `TaxonomyTermFieldsSchema` declares 14 fields including an owner, a
      semver `vocabularyVersion`, `aliases`, BCP-47 `translations`, provenance,
      broader/narrower/related relations, applicability, and a lifecycle discriminated union where
      a deprecation MUST carry a `migrationRule`. `register-taxonomy-term.ts` refuses a relation
      pointing at a term that is not registered and refuses a deprecation whose replacement is
      not, and versions through envelope superseding. 13 files reach it; 0 of them is an Isis or
      Oshun consumer.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The
      control's registration refuses a second address for one term, refuses an address that
      reassigns ownership, refuses a term specialising itself, and requires a tenant. (2) Its
      self-reference check is DEPTH-1 only — a to b to a is not refused at registration — but the
      search expansion records every reference before queueing it and caps depth, so the walk
      terminates rather than hanging; the bounded consequence is recorded rather than reported as
      a defect. (3) The style map's fallback is named, documented and actually on the resolution
      path. (4) The asset-kind route refuses an unknown kind with an explicit message rather than
      passing it through.
      
      TRAPS. (1) A DECLARATION'S FILE IS NOT THE DECLARATION: `apps/oshun/bff/src/admin/state.ts`
      is 30,000 lines holding the whole admin workspace, so asking the FILE whether the campaign
      carries translations or a fallback finds both — belonging to something else. .b, .d and .e
      went from 3/5 and two spurious trues to 1/5 and two falses once every question was asked of
      the declaration's own slice. (2) A REGISTRY ROUTE THAT IS ANOTHER DOMAIN'S IS NOT ONE OF THE
      SIX — the two Hathor lore-validation taxonomy routes are recorded as seen and not counted.
      (3) AN EMPTY PARSE AND A MEASURED ABSENCE LOOK IDENTICAL: the control's fields come back
      empty unless the reader knows `z.object({`, and the throw caught it.
      
      0 of 9 properties present. THE PARENT AND ALL EIGHT CELLS LEFT `[ ]`: none of the six is a
      registry, one of them does not exist, no route governs any of them, and the governed
      registry the estate did build serves a different workbench. -->
      

I9 — Isis observability, evaluation, and launch gates#

  • I9.1 Instrument plan, dispatch, provider, stage, gate, review, release, and recall traces with bounded attributes. Evidence: trace walkthrough.
    • I9.1.a Define canonical trace/span names, parents/links, status, events, sampling, retention, and schema versions for every listed lifecycle. Evidence: telemetry semantic convention.

    • I9.1.b Propagate approved trace/correlation/request/job/run/stage/release context across HTTP, queue, worker, provider, storage, outbox, and downstream target boundaries. Evidence: propagation tests.

    • I9.1.c Allowlist bounded low-cardinality tenant-safe attributes for asset/workflow/model/provider classes, state, failure class, target, and policy versions. Evidence: cardinality review.

    • I9.1.d Exclude prompts/media/source titles/identity/biometrics/secrets/ URLs/raw IDs and apply centralized scrubbing to events/errors. Evidence: trace privacy tests.

    • I9.1.e Instrument success/failure/cancel/retry/cache/partial/rollback/ recall transitions and asynchronous span links without false parentage. Evidence: lifecycle trace tests.

    • I9.1.f Walk a complete brief-to-recall journey and reconstruct every seam from traces plus canonical records. Evidence: trace walkthrough. <!-- MEASURED 2026-08-26 (I9.1.a-f). The parent's evidence is a TRACE WALKTHROUGH: walk one journey and reconstruct every seam. A walkthrough presupposes spans and spans presuppose something that starts them, so the binding question is asked of the product rather than of the manifest — on the eight lifecycles I9.1 names, what starts a span? Instrument: generate-oshun-lifecycle-tracing.mts (+15 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-lifecycle-tracing.json.

      NOTHING IN THE BFF STARTS A SPAN, AND NEITHER DOES `apps/isis`. 18 files in the estate call
      `startSpan`/`startActiveSpan`, across 10 domain roots — lilith, v2, cybele, iris, isis
      libraries, oya, psyche, shared, yemaya — and 0 of them is in `apps/oshun` or `apps/isis`.
      Neither app imports `libs/shared/tracing`, which is a complete tracer with decorators, GenAI
      conventions and middleware. The sweep is not blind: it finds the roots that do trace, and it
      is asserted doing so.
      
      A TRACER THAT DEFAULTS TO RECORDING NOTHING IS NOT INSTRUMENTATION. The one tracing seam on
      the Isis side is `libs/isis/llm-providers/src/client/tracer.ts`, whose default is a
      `NoopTracer` and whose header documents exactly how to plug a real OTel tracer in. 0 product
      call sites do. The seam itself is honest — it records nothing rather than pretending to —
      and that is the whole of the Isis generation path's tracing.
      
      2 OF THE EIGHT LIFECYCLES HAVE A SPAN FAMILY. `libs/oshun/analytics/src/tracing-manifest.ts`
      declares 22 families across 11 subsystems, each with a SpanKind, a description, required
      propagation channels and typed attributes. Of I9.1's eight — plan, dispatch, provider,
      stage, gate, review, release, recall — only review (`review.queue.item_processed`) and
      release (`bff_release.request.handled`) are named, and neither is an Isis span. Isis has
      `isis.generation.run` and `isis.workflow.run`, which are the run and not the lifecycle.
      
      A SPAN DEFINITION FIXES 2 OF .a's SEVEN SUBJECTS. Read off `SpanFamilyDefinition`'s own
      declared fields — name, subsystem, kind, description, propagation, attributes — .a's name
      and parents/links are there and status, events, sampling, retention and schema version are
      not fields a family can carry.
      
      NAMING A CONTEXT FIELD IS NOT PROPAGATING IT. All seven of .b's identifiers — trace,
      correlation, request, job, run, stage, release — appear somewhere across the BFF's files,
      and that says nothing: S10.2's rule is to name the field the context rides in and then show
      that field has both a WRITER and a READER. The HTTP carrier is `traceparent`, and across
      `apps/oshun/bff/src` it has 0 readers and 0 writers. 0 of .b's 7 boundaries — HTTP, queue,
      worker, provider, storage, outbox, downstream target — carries a trace context at all.
      
      A TRACE ID MINTED PER CALL IS A NEW TRACE EVERY TIME. The BFF's one tracing port is the
      workbench-kit command router's `tracing.begin()`, and it mints a fresh `randomBytes(16)`
      trace id and a fresh `randomUUID()` correlation id on every call, accepting nothing inbound.
      Two commands in one operator session are two unrelated traces, and nothing joins them —
      which is exactly the missing edge S10.2 was written about, on the one path in the estate
      that has a tracing seam at all.
      
      THE ATTRIBUTE DESIGN IS 3,200 LINES WITH 0 PRODUCT CONSUMERS.
      `libs/oshun/workbench-kit/src/telemetry-attributes.ts` is a cardinality and unit design —
      signal kinds, cost shapes, bounded-domain requirements, unit dimensions and conversions —
      and no product code reaches it, so .c's allowlist is a design nothing draws from. For .d the
      estate holds exactly one scrubber, `apps/v2/telemetry-privacy-compliance`, and A SCRUBBER IN
      ANOTHER APP IS NOT ON THIS PATH: 0 of .d's 8 excluded classes is withheld anywhere Isis
      writes.
      
      3 OF .e's EIGHT TRANSITIONS ARE A VALUE THE MANIFEST HOLDS — success, failure and partial.
      Cancel, retry, cache, rollback and recall do not appear in the manifest at all, prose
      included, so the transitions a lifecycle trace exists to show are the ones it cannot.
      
      THERE IS NO WALKTHROUGH, AND THE THREE DESIGNS HAVE 0 CONSUMERS BETWEEN THEM. No evidence
      artifact anywhere is a trace walkthrough; a remediation note under `docs/domains/isis`
      describes end-to-end trace ids across api, queue, provider adapters and handler runtimes,
      and A DOCUMENT IS NOT A WALKTHROUGH. The span-family manifest, S4.13's correlation link
      graph (`atIngress`, `linkFaults`) and S10.2's `trace-continuity.ts` — 2,752 lines whose
      opening statement is A TRACE IS A CLAIM ABOUT CAUSATION AND IT IS MADE OF EDGES, NOT SPANS —
      have 0 product callers between the three of them.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The Isis LLM
      client's tracer port is an honest seam rather than a fabrication: it mirrors the OTel
      interface, documents how to plug a real tracer in, and its default records nothing instead
      of pretending to record. (2) The span manifest is a real machine-readable definition, not a
      list of names. (3) S10.2's module states the vein this cell turns on, and states it before
      the estate had anything to apply it to.
      
      TRAPS. (1) NAMING IS NOT PROPAGATING — the context-field sweep over the joined BFF sources
      returned 7 of 7 and would have made .b pass; asked of the carrier instead, it is 0 writers
      and 0 readers. (2) A WORD IN PROSE IS NOT A VALUE THE CODE HOLDS: transitions are read from
      the manifest's declared literals, and the five missing ones turn out to be absent from the
      file entirely. (3) A SCRUBBER IN ANOTHER APP IS NOT ON THIS PATH.
      
      0 of 9 properties present. THE PARENT AND ALL SIX CELLS LEFT `[ ]`: the lifecycle emits no
      spans, six of its eight stages have no span family, the carrier that would join them has
      neither a writer nor a reader, and the one trace id the BFF mints is new on every call. -->
      
  • I9.2 Track queue delay, runtime, success, failure class, quality, moderation, cache, cost, and release conversion by safe dimensions. Evidence: metric schema review.
    • I9.2.a Define each counter/gauge/histogram name, unit, numerator/ denominator, event source, aggregation window, owner, and schema version. Evidence: metric catalog.

    • I9.2.b Instrument queue wait/age/depth, execution/stage duration, throughput, success/partial/failure/cancel/retry, and stable failure taxonomy. Evidence: metric tests.

    • I9.2.c Instrument raw quality/moderation band distributions and unavailable/manual-review rates without exposing sensitive categories to unauthorized dimensions. Evidence: safety metric review.

    • I9.2.d Instrument cache lookup/eligible/hit/miss/corrupt/quarantine/ recompute and estimate/reservation/actual/pending/variance costs. Evidence: cache/cost metric tests.

    • I9.2.e Define release funnel denominators from generated through gated/ reviewed/bundled/written/verified/retained/recall states without hiding failed or excluded items. Evidence: conversion fixtures.

    • I9.2.f Enforce allowlisted low-cardinality tenant-safe dimensions, aggregation/privacy thresholds, and cardinality/volume budgets. Evidence: metric schema/privacy review. <!-- MEASURED 2026-08-26 (I9.2.a-f). The parent's evidence is a METRIC SCHEMA REVIEW, and a schema review presupposes a schema of something that is written. A metric is a counter, a gauge or a histogram — a named series with a unit and an aggregation window that a process increments — so the binding question is asked of the emitters: on the queue, the run, the gate and the release, what increments? Instrument: generate-oshun-lifecycle-metrics.mts (+14 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-lifecycle-metrics.json.

      THE BFF PUBLISHES NINE SERIES AND EVERY ONE OF THEM IS ABOUT AN HTTP REQUEST. `GET
      /metrics` is mounted and serves Prometheus text from `BffMetricsCollector`, fed by an
      `onResponse` hook: requests_total, errors_total, error_rate, latency_ms, latency_mean_ms,
      rps, route_error_rate, domain_success_rate and domain_latency_p95_ms. (It published eight
      when this cell first measured; §I9.3's fix split the mean out from the quantile summary.)
      The record they aggregate has six fields — method, route, statusCode, durationMs, timestamp,
      domain — and not one of the nine names a queue, a run, a stage, a gate, a cache, a cost or a
      release, which is the whole of what I9.2 asks for. 0 files under `apps/oshun` or `apps/isis` construct a Counter,
      a Gauge or a Histogram; the same expression finds 14 elsewhere in the estate.
      
      A DEPENDENCY OF ANOTHER LIBRARY IS NOT A DEPENDENCY OF THIS APP. `prom-client` and
      `@opentelemetry/sdk-metrics` are declared by seven package.json files — shared/metrics,
      shared/tracing, lakshmi/core, cybele/api, yemaya/core, yemaya's study service and nyx's
      pipelines — and by neither app that would emit here.
      
      A DASHBOARD PANEL IS A QUESTION AND SOMETHING HAS TO ANSWER IT.
      `dashboards-queue-health.ts` names seven canonical queues — review, moderation, support,
      generation, appeal, incident communication, audit export — and nine metric kinds: depth,
      oldest-item age, throughput, p50 and p99 latency, dead-letter volume, retry rate, SLA breach
      rate and consumer lag. 4 of .b's eleven series are a kind on that dashboard and 0 of the
      eleven is emitted.
      
      THE ALERTS FIRE ON EVENTS AND THE LIFECYCLE EMITS NONE. All 13 alert definitions name
      `sourceEvents` from the canonical taxonomies rather than a series or a query, which is an
      honest design — but of the five event taxonomies, the release one has exactly 1 consumer and
      it is `feature-flag-expiry-alerts.ts` in the workbench kit. The BFF has one server-side
      event emitter, the studio-workspace telemetry route, and it forwards what a client posted.
      0 emitters are on the generation, gate or release path.
      
      A METRIC DEFINITION FIXES 4 OF .a's SEVEN SUBJECTS. Read off `QueueHealthMetricDefinition`
      and `AlertConditionDefinition` together: name, unit, event source and aggregation window are
      declared; numerator/denominator, owner and schema version are not fields either can carry.
      
      0 BAND, CACHE, COST AND FUNNEL SERIES. .c's quality and moderation distributions, .d's seven
      cache states and five cost states, and .e's eight funnel denominators are none of them a
      counter. The cost arithmetic .d would count DOES exist — the generation budget gate reserves
      an estimate before the work, settles the actual after, and reports the difference rather
      than pretending the estimate was right — it is simply never counted. And .f's cardinality
      design, `telemetry-attributes.ts`, is 3,200 lines with 0 product consumers, the same finding §I9.1
      recorded.
      
      FIXED. A METRIC LABEL FED FROM THE RAW URL IS UNBOUNDED, AND IT CARRIES THE QUERY STRING.
      The `onResponse` hook recorded `request.routeOptions?.url ?? request.url`. `routeOptions.url`
      is the matched route PATTERN and is bounded by the route table; the fallback is the raw path
      AND its query string, so every request that matched no route became its own label value in
      `oshun_bff_route_error_rate`. Driven: a label read back as
      `/v1/unmatched/490?token=secret-490`. The label is now the pattern, and one fixed
      `unmatched` value for everything else — asserted against a real Fastify instance: a matched
      request is labelled `/v1/thing/:id`, two different unmatched paths collapse to one value,
      and neither the id nor the query string reaches the label.
      
      FIXED. A WINDOW THAT PRUNES ONE STRUCTURE AND NOT THE OTHER IS NOT A WINDOW. `errorMessages`
      and `domainErrors` are keyed on that same caller-supplied route string and were never
      pruned, so the rolling window discarded the requests while the maps kept one permanent entry
      per distinct key — memory whose size was decided by whoever was sending requests. Driven:
      2,000 distinct 5xx routes left 2,000 entries behind a window holding 10. Both maps now
      honour the same time window and a hard 500-entry cap, oldest first, and the entries kept are
      the most recent ones.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The alerts
      manifest is honest about what it fires on: every condition names its sourceEvents, its
      threshold, its unit and its evaluation window, so an alert is a rule over the event taxonomy
      rather than over an imagined series. (2) The queue-health dashboard names its seven queues
      and nine metric kinds explicitly rather than leaving the panels to a query somebody writes
      later. (3) The generation budget gate reserves before the work and settles after, and its
      reservationId is the job id so a retried admission cannot consume the same estimate twice.
      
      TRAPS. (1) The first pass of this cell asserted there was no metrics collector at all; there
      is one, it is mounted, and it publishes eight real series — the finding is not that metrics
      are absent but that every series is about a request. (2) A DEPENDENCY OF ANOTHER LIBRARY IS
      NOT A DEPENDENCY OF THIS APP — asking the repository rather than the two apps made ".a has a
      metric library" pass. (3) THE SERIES COUNT COMES FROM THE EXPOSITION: series are read from
      the `# TYPE` lines a scrape returns, not from the word "counter" appearing in the module.
      
      0 of 8 properties present. THE PARENT AND ALL SIX CELLS LEFT `[ ]`: the eight series the BFF
      publishes are all about HTTP requests, no app declares a metric library, no queue or cache
      or cost or funnel state is a counter, and the dashboard that asks nine questions about seven
      queues has nothing answering it. -->
      
  • I9.3 Define SLOs for plan, dispatch acknowledgement, queue age, provider success, artifact durability, gate freshness, and release verification. Evidence: error-budget policy.
    • I9.3.a Define user/system journey, population, eligibility, exclusions, measurement point, SLI formula, window, target, and owner for each SLO. Evidence: SLO specification.

    • I9.3.b Define plan latency, dispatch acknowledgement, queue age, provider success, artifact durability, gate freshness, and release verification SLIs from canonical metrics/events. Evidence: SLI query tests.

    • I9.3.c Segment only by actionable safe service tier/region/provider/ workflow/target classes and prevent small-tenant disclosure. Evidence: SLO dimension review.

    • I9.3.d Define multi-window burn alerts, paging/ticket thresholds, missing-data behavior, maintenance/exclusion rules, and ownership. Evidence: alert simulation.

    • I9.3.e Define error-budget consumption, release/rollout/freeze policy, exception authority, recovery criteria, and historical amendments. Evidence: ratified error-budget policy. <!-- MEASURED 2026-08-26 (I9.3.a-e). The parent's evidence is a RATIFIED ERROR-BUDGET POLICY, and a budget is spent against an SLI: a formula over a population, measured somewhere, in a window, against a target. So the binding question is asked of the seven SLIs .b names — plan latency, dispatch acknowledgement, queue age, provider success, artifact durability, gate freshness, release verification — from what canonical metric or event would each be computed? Instrument: generate-oshun-slo-error-budget.mts (+13 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-slo-error-budget.json.

      AN SLO FOR A SURFACE IS NOT AN SLO FOR A STAGE. `V1/planning/SLO_AND_DR.md` is a real
      ratified document — ten surfaces with an SLI definition, a target and a monthly error
      budget, a tiering principle that says why each tier is what it is, DR objectives per data
      class, a drill program and a named owner for all seven areas. Its rows are the BFF API,
      auth, two Psyche planes, two Living Scenes planes, the Metis gateway, the public viewer, the
      payments plane and the event bus. 2 of .b's seven SLIs have a row and both are incidental —
      `queue` appears in the render plane's BLACK-queue state and `gate` in the cue gate. 0 of the
      seven is computable from the nine series the BFF publishes.
      
      THE ONE SLI THAT IS COMPUTABLE, recorded so those zeros read as a distance rather than a
      blank: the BFF API row's availability SLI is "non-5xx responses ÷ valid requests", and both
      numbers are published as `oshun_bff_requests_total` and `oshun_bff_errors_total`. Its
      latency SLI is stated at p99, which the exposition did not carry until this cell.
      
      THE BURN ARITHMETIC IS CORRECT AND ITS INPUT IS A NUMBER SOMEBODY TYPED. Two surfaces
      implement the multiwindow multi-burn-rate model and both read their inputs from a request
      body. `POST /v1/admin/oncall/burn-decision` applies the thresholds — page at 14.4x over 1h
      only when the 5m window agrees, ticket at 6x over 6h only when the 30m agrees — to four burn
      rates in the body. `POST /v1/admin/studio/observability-dashboards/evaluate` goes further
      and DERIVES the rate, `errorRate / (1 - sloTarget)`, requiring two windows to page — and
      reads the error rate and the target out of the body too. Neither reads a published series.
      The estate can evaluate a burn rate; nothing measures one.
      
      A DOCUMENT THAT MENTIONS A WORD IS NOT A SPECIFICATION THAT FIXES IT. 4 of .a's nine
      subjects are columns the SLO table fixes per SLO — the surface, the SLI formula, the monthly
      window and the target. Population, eligibility, exclusions and the measurement point live
      inside some rows' free-text SLI cell and not others', and ownership is a separate table
      keyed by ROLE, so no SLO row carries an owner.
      
      3 OF .c's FIVE SEGMENTATION CLASSES appear in the specification and there is no
      small-tenant protection: no k-anonymity, no minimum cohort, no suppression rule. 3 of .d's
      six alert subjects are present — the four windows and both thresholds — with no
      missing-data rule, no maintenance exclusion and no owner on the alert itself.
      
      A SPEND BUDGET IS NOT AN ERROR BUDGET. The policy states a feature freeze — budget exhausted
      means only reliability and rollback changes merge, and canary promotions are blocked while
      any RED-tier surface is over budget — and 0 sites enforce it. Seven files mention a budget
      being exhausted; the two a looser sweep first returned are a USD retry budget in
      `ai-platform/efficiency.ts` and a verification budget in the kit. 4 of .e's five policy
      subjects are stated; exception authority is not.
      
      A NAME THAT COLLIDES IS NOT A CONSUMER. `libs/oshun/workbench-kit/src/service-objectives.ts`
      is 4,026 lines that designed I9.3.b's own vocabulary — its `SLI_CLASSES` are availability,
      correctness, latency, deadline, QUEUE-AGE, DURABILITY and FRESHNESS — with observation
      points, blind spots, exclusions with named deciders, ratio faults and outcome grades. It has
      0 importers. A sweep on `blindSpots|ratioFaults|gradeFaults` returns 11 files and every one
      is another domain's identically-named thing, Veritas' Ghana regions with no source coverage
      among them.
      
      AN SLO FOR ANOTHER JOURNEY IS NOT THIS JOURNEY'S SLO.
      `apps/oshun/bff/src/agentic/human-authority-slo-store.ts` is a genuine 658-line SLO surface
      with six numeric targets — takeover acknowledgement p95 2s, checkpoint resume p95 30s,
      latency success rate, edit incorporation rate, drill success rate, drill max age — and six
      required drills. 0 of its targets is one of .b's seven.
      
      FIXED. A MEAN PUBLISHED AS A MEDIAN IS A WRONG NUMBER WITH A CORRECT NAME. The Prometheus
      exposition emitted `avgLatencyMs` under `oshun_bff_latency_ms{quantile="0.5"}`. Driven with
      99 requests at 10 ms and one at 10 s: the median is 10 and the published value was 109.9,
      wrong by eleven times, under a label that says exactly which number it is. The quantile now
      carries the median, the mean is published as `oshun_bff_latency_mean_ms` under a name that
      says what it is, and `quantile="0.99"` was added because the ratified SLO states the BFF
      latency SLI at p99 and `percentile()` was already computing it. §I9.2's record was updated
      from eight published series to nine.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The
      multiwindow burn policy is implemented and not merely described: the page condition requires
      both the 1h and the 5m window, the ticket condition both the 6h and the 30m, and the window
      evidence is returned either way so a decision can be read back. (2) The SLO document says
      why each tier is what it is rather than asserting numbers. (3) Crisis-path availability is
      deliberately NOT given a budget — any unavailability is a Sev-1 regardless of duration —
      which is a decision recorded rather than an omission. (4) The human-authority SLO store is a
      genuine SLO surface for the journey it is about.
      
      TRAPS. (1) A SUBSTRING IS NOT THE CONCEPT, for the second time this phase: `plan` is inside
      "control plane", `queue` inside "BLACK queue state", `gate` inside "Metis gateway" and "cue
      gate", and `target` inside "targets ratchet". Substring matching put .b at 3 of 7 and .a at
      9 of 9. (2) A DOCUMENT THAT MENTIONS A WORD IS NOT A SPECIFICATION THAT FIXES IT — .a is now
      read off the table's columns. (3) A NAME THAT COLLIDES IS NOT A CONSUMER — an import
      specifier, not a symbol name. (4) The first pass of this cell reported that nothing computes
      a burn rate; a 993-line store does, correctly, and the finding is about its input.
      
      0 of 8 properties present. THE PARENT AND ALL FIVE CELLS LEFT `[ ]`: no stage of the
      lifecycle has an SLO, no SLI it owes is computable from anything published, the burn model
      is fed from a request body, the freeze the policy ratifies is enforced by nothing, and the
      4,000-line module that designed exactly these SLI classes has no importers. -->
      
  • I9.4 Build provider and worker fault-injection tests. Evidence: throttle, timeout, malformed output, worker loss, and split-brain cases.
    • I9.4.a Build hermetic controllable provider/worker/queue/callback/storage fault fixtures with deterministic clocks, IDs, and failure points. Evidence: fault harness tests.

    • I9.4.b Inject rate limit/throttle/capacity/refusal with Retry-After and verify taxonomy, backoff, circuit, reservation, and UI state. Evidence: throttle tests.

    • I9.4.c Inject connect/read/overall/deadline timeouts before, during, and after provider acceptance/output write. Evidence: timeout tests.

    • I9.4.d Inject malformed status/schema/content type/oversize/artifact/ manifest/hash/callback data and verify quarantine/no success. Evidence: malformed output tests.

    • I9.4.e Kill/evict/restart workers at each checkpoint and verify fenced lease, explicit resume/new attempt/failure, no duplicate output/cost. Evidence: worker-loss tests.

    • I9.4.f Simulate lease expiry/network partition/two workers/late callback and prove stale actors cannot commit terminal state. Evidence: split- brain tests.

    • I9.4.g Verify telemetry, alert, cancellation, cleanup, resource release, and immutable history for every injected case. Evidence: fault coverage report. <!-- MEASURED 2026-08-26 (I9.4.a-g). The parent names five cases as its evidence — throttle, timeout, malformed output, worker loss, split-brain — so the binding question is asked case by case: is the fault INJECTED into the real path, and what is asserted once it lands? Instrument: generate-oshun-fault-injection.mts (+13 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-fault-injection.json.

      THE PROVIDER HALF IS REAL AND THE WORKER HALF HAS NOTHING TO LOSE. Across 60 generation
      sources and 59 tests, 16 modules take an injectable `fetchImpl`, 6 tests put a 429 on the
      wire through it and 37 assert a rejection rather than a success. That is a genuine hermetic
      seam and the first three cases, on the real modules — not a network test wearing a mock.
      
      `lease` IS A SUBSTRING OF `release`. Matched as a whole word, the generation path has no
      lease, no fence, no heartbeat and no checkpoint; 1 of .e's five subjects is present and it
      is `attempt`. The executors run in process. A worker that cannot be evicted cannot be
      killed at a checkpoint, and two workers that cannot exist cannot both commit, so .e and .f
      have no subject rather than a failing one. The estate DOES have leases — 44 files carry a
      `leaseId` or a `fencingToken` — and none of them is here.
      
      A CODE ASSEMBLED FROM A STATUS IS NOT A TAXONOMY. Five providers throw
      `` `…_http_${response.status}` ``, so the failure CLASS is whatever the provider returned
      rather than a member of a closed set a caller could switch on, and .b asks the throttle case
      to verify a taxonomy. 0 declared failure vocabularies exist on this path. Of .b's five
      downstream properties only backoff is present: no Retry-After is read, no circuit exists,
      and nothing releases a reservation when a throttle lands.
      
      1 OF .c's FOUR TIMEOUT KINDS IS DISTINGUISHABLE, and it is the coarsest. Only an overall
      timeout is expressible; connect, read and deadline are three different failures with three
      different remedies and this path cannot tell them apart, let alone place one before, during
      or after provider acceptance.
      
      5 OF .d's EIGHT MALFORMED KINDS ARE NAMEABLE — schema, content type, artifact, manifest and
      hash — and status, oversize and callback are not. None is quarantined: quarantine is a state
      an output is IN, 241 files in the estate know the word, and the generation path is not one
      of them.
      
      0 OF .g's SIX COVERAGE SUBJECTS IS ASSERTED in any generation test. Not telemetry — §I9.1
      established there are no spans and §I9.2 no lifecycle series — not an alert, not
      cancellation, not cleanup, not resource release, not an immutable history.
      
      THE KIT HOLDS 5,031 LINES OF FIXTURES FOR EXACTLY THESE SEVEN CELLS, WITH 0 IMPORTERS.
      `throttle-fixtures` (652), `timeout-fixtures` (735), `malformed-fixtures` (832),
      `lease-fixtures` (545), `partial-fixtures` (615), `retry-exhaustion-fixtures` (605),
      `cancellation-fixtures` (577) and `duplicate-fixtures` (462) — every one substantial rather
      than nominal, and not one imported by product code.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The provider
      fault tests hold the bright line: an injected 429 or an empty response produces a typed
      error naming the provider and the status, and never a fabricated success. (2) Sixteen
      modules taking an injectable fetch is a real hermetic harness, and .a's provider seam is
      genuinely satisfied even though its queue, callback and storage seams are not. (3) The kit's
      fixtures are written for these cells and are real work, not placeholders.
      
      TRAPS. (1) `lease` IS A SUBSTRING OF `release` — a substring sweep for a lease over a
      release path finds one in every file. The whole of .e and .f turns on the distinction, and
      the control is recorded: `release` appears on the path, `lease` as a word does not. (2) AN
      ERROR CODE IS A LITERAL, NOT AN IDENTIFIER: an identifier scan blanks literal contents, so
      the taxonomy .b asks about was invisible until literals were read from the SOURCES — and
      from the sources only, because a test's `it('cancels the job')` title is prose and counting
      it would let a description stand in for an assertion.
      
      0 of 9 properties present. THE PARENT AND ALL SEVEN CELLS LEFT `[ ]`: three of the five
      named cases are injected and two have no subject, the refusal a throttle produces is
      assembled rather than classified, only the coarsest timeout is expressible, nothing is
      quarantined, and none of the six things .g asks to be verified is asserted anywhere. -->
      
  • I9.5 Build known-value tests for every technical quality gate. Evidence: gate registry coverage test.
    • I9.5.a Enumerate every active technical gate and asset-kind/target/ profile applicability row directly from the gate registry. Evidence: generated coverage manifest.

    • I9.5.b Create licensed versioned positive/negative/boundary/N/A/ unsupported/corrupt fixtures with expected raw measurements and units. Evidence: fixture manifests.

    • I9.5.c Pin evaluator code/config/dependencies and numeric tolerance/ determinism rules per fixture. Evidence: reproducibility configuration.

    • I9.5.d Assert raw measurements, regions/timecodes/object refs, uncertainty/applicability before asserting policy verdicts. Evidence: known-value suites.

    • I9.5.e Assert threshold below/equal/above and target-profile changes recompute correct verdicts without reprocessing where permitted. Evidence: boundary tests.

    • I9.5.f Fail CI when an active registry row lacks a current fixture or a fixture points to retired/unknown gate versions. Evidence: registry coverage gate. <!-- MEASURED 2026-08-26 (I9.5.a-f). The parent's evidence is a GATE REGISTRY COVERAGE TEST: every active row has a current fixture and CI fails when one does not. A known-value test is one where a fixture with a KNOWN raw measurement produces an expected verdict, so the binding question is asked of the registry's rows — how many are exercised across a threshold, and what happens to the ones that are not? Instrument: generate-oshun-gate-known-values.mts (+13 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-gate-known-values.json.

      THE REGISTRY IS REAL AND ITS COVERAGE IS NOT ENUMERATED.
      `libs/oshun/generation-control-isis/src/release-gate-model.ts` declares nine canonical gate
      kinds, an applicability list per output kind, a quality floor and a watermark floor per
      output kind and one safety floor — 9 x 11 is 68 applicability rows, and every threshold
      names the measurement it bounds and varies honestly by kind (a 3d-model's watermark floor is
      0 because watermarking does not apply; an image's is 1). 19 of the 68 rows are exercised by
      a test that names both the row's output kind and its gate. Nothing enumerates the other 49,
      there is no generated coverage manifest, and 0 CI steps fail when a row has no fixture.
      
      FIXED. A NUMBER THAT IS NOT A NUMBER PASSES A RANGE CHECK WRITTEN AS TWO COMPARISONS.
      `NaN < 0` and `NaN > 1` are both false, so `if (x < 0 || x > 1)` admitted it — and every
      threshold below it, `x < floor`, was false as well. Driven: a measurement whose safety,
      watermark and quality scores were all `NaN` validated as `{ ok: true }` and cleared every
      numeric gate, the release mode read `allow`, and the verdicts announced the comparisons they
      had not performed — "safety score NaN ≥ floor 0.9" and "watermark coverage NaN ≥ floor 1".
      Infinity was caught, because `Infinity > 1` is true; only NaN slipped. All three range
      checks and the validator now use one `Number.isFinite` predicate. The same measurement now
      blocks on all three gates with three validation errors. This matters because
      `apps/oshun/bff/src/generation/release-gate.ts` consults that validator to decide whether a
      measurement is canonical at all, so an unmeasured score was releasable.
      
      FIXED. ASSERTING A CONSTANT IS NOT TESTING THE THRESHOLD. The suite asserted
      `CANONICAL_SAFETY_SCORE_FLOOR === 0.9` and never drove the evaluator across it. Five cases
      added: each of the three floors below / exactly at / above, the non-number case, and a
      control that the infinities and out-of-range values it always refused are still refused. All
      three thresholds now have a failing verdict AND a passing verdict asserted against the
      evaluator, which is .e's first clause.
      
      .e HAS TWO CLAUSES AND ONLY THE BOUNDARY ONE IS SATISFIED, SO THE CELL STAYS `[ ]`. Its
      second asks that a target-profile change recompute correct verdicts without reprocessing
      where permitted. The registry has no profile — consistent with §I8.4's finding that 0 of 9
      candidate subjects are pinned to a target profile — so there is nothing for a change to
      recompute against. The kit declares four recompute TRIGGERS (automatic, on-next-save,
      manual, scheduled) and no product registry keys off them.
      
      5 OF .b's SIX FIXTURE CLASSES are in the suite; `unsupported` is not. 1 OF .c's FIVE PINNING
      SUBJECTS is pinned — a numeric tolerance — and the evaluator version, its config, its
      dependencies and a determinism rule are not. 2 OF .d's FIVE RAW SUBJECTS are fields the
      measurement carries: the raw score and the object ref. A region or timecode, an uncertainty
      and an applicability statement are not, so a verdict cannot say WHERE in the asset it was
      measured or how sure it is.
      
      A GATE THAT REPORTS UNAVAILABLE IS NOT A GATE THAT PASSED. The second registry,
      `libs/oshun/tara-workbench/src/gate-registry.ts`, records its own absences: four gates are
      `available` and one is a `seam` waiting on a bound audio-analysis pipeline, and a seam
      reports unavailable rather than passing. That is the shape .a's enumeration wants and it
      exists on the smaller of the two registries.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The floor
      comparisons are consistent and self-documenting: `< floor` fails and the pass note reads
      `≥ floor`, so the boundary is stated in the verdict a reader sees. (2) Every threshold
      varies by output kind rather than being one number applied to everything. (3) The Tara
      registry's seam is the honest kind — no evaluator at all, recorded as such.
      
      TRAPS. (1) A `gateCoverage` IDENTIFIER IN ANOTHER DOMAIN IS NOT THIS REGISTRY'S COVERAGE
      GATE: the bare name matches 21 files, Maya's standards modules among them, and .f would have
      passed on them. A site has to name the registry it covers — and then A RULE THAT NAMES WHAT
      IT WATCHES COUNTS ITSELF: this instrument spells `CANONICAL_OUTPUT_GATE_REQUIREMENTS` in
      order to parse it, so the moment it was committed it matched its own tightened sweep and .f
      passed on the existence of the census that reports .f absent. The instrument now excludes
      itself. (2) A ROW IS A PAIR: `image`
      appears in a dozen tests that say nothing about watermark coverage, so a row counts as
      exercised only when the test names both halves. (3) Prettier wraps a long assertion, so a
      verdict literal can sit on its own line and a `\.toBe\('` pattern misses it — the boundary
      detector reported 2 of 3 until it allowed the wrap.
      
      1 of 8 properties present, and it is the one this cell earned. THE PARENT AND ALL SIX CELLS
      LEFT `[ ]`: 49 of 68 registry rows are unexercised, nothing enumerates them, no CI step
      fails when one has no fixture, the evaluator is not pinned, the measurement cannot say where
      or how surely it measured, and .e's target-profile clause has no profile to change. -->
      
  • I9.6 Build calibrated evaluation suites for model-based safety, similarity, and perceptual gates. Evidence: precision/recall, threshold, subgroup, and uncertainty report.
    • I9.6.a Define licensed/consented representative train-free calibration, threshold-selection, and held-out evaluation corpora with provenance, labels, subgroup/risk/style slices, and versioning. Evidence: corpus manifests.

    • I9.6.b Define expert annotation guidance, multi-rater process, adjudication, inter-rater reliability, ambiguous/N/A labels, and blinded sampling. Evidence: labeling report.

    • I9.6.c Run safety, modality-specific similarity, and perceptual evaluators with pinned versions/configs and retain per-item raw outputs. Evidence: reproducible evaluation runs.

    • I9.6.d Report precision/recall/specificity/FPR/FNR/PR or ROC/calibration and confidence intervals at candidate thresholds. Evidence: metric report.

    • I9.6.e Report subgroup/content/style/device/quality slices with sample counts, uncertainty, disparity, worst cases, and privacy thresholds. Evidence: subgroup report.

    • I9.6.f Select fail/review/pass bands against ratified harm/cost criteria using threshold corpus only, then verify held-out performance. Evidence: threshold decision record.

    • I9.6.g Define drift monitoring/recalibration triggers and route uncertain/ low-quality/out-of-distribution cases to human review. Evidence: uncertainty/drift plan. <!-- MEASURED 2026-08-26 (I9.6.a-g). The parent's evidence is a precision/recall, threshold, subgroup and uncertainty REPORT, and every one of those is a statement about a labelled corpus. So the binding question is asked of the corpus: which labelled set was the safety gate's threshold chosen against? Instrument: generate-oshun-gate-calibration.mts (+13 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-gate-calibration.json.

      A THRESHOLD IS A NUMBER SOMEBODY PICKED UNTIL A CORPUS CHOOSES IT.
      `CANONICAL_SAFETY_SCORE_FLOOR = 0.9` is declared with a comment saying what the number MEANS
      — the minimum safety score an output must earn — and 0 records anywhere say how 0.9 was
      arrived at. 0 of .a's three corpora exists: no calibration set, no threshold-selection set,
      no held-out set. So .d's precision and recall have no population to be computed over, .e's
      subgroup report has no slices to report, and .f's decision record has no decision to record.
      
      2 OF THE THREE GATE FAMILIES HAVE AN EVALUATOR. Safety has sixteen modules that touch it and
      similarity two; PERCEPTUAL has none on the generation path at all, so the family the parent
      names third has nothing to calibrate.
      
      A `-latest` ALIAS IS NOT A PINNED VERSION. The one evaluator whose identity is declared is
      `DEFAULT_MODERATION_MODEL = 'omni-moderation-latest'`, a floating alias — so the model
      answering the gate can change under any calibration that would justify the threshold. Its
      per-item raw outputs ARE retained (`categoryScores` per category), which is the other half
      of .c. Whether to pin a dated snapshot here is a product decision with real consequences —
      a dated snapshot can be retired by the provider — so it is recorded rather than changed.
      
      3 OF .d's EIGHT METRICS reach the path. Specificity, false-positive rate, false-negative
      rate, a PR or ROC curve and a confidence interval do not, and four of them are computed
      elsewhere in the estate for something else. 2 of .b's six annotation subjects and 2 of .e's
      six subgroup subjects are present; there is no multi-rater process, no adjudication, no
      inter-rater statistic, no blinded sampling, no disparity and no privacy threshold.
      
      A CALIBRATION FOR ANOTHER JUDGE IS NOT THIS GATE'S CALIBRATION.
      `libs/oshun/agentic-studio/src/feedback/gold-set-calibration.ts` is exactly the shape .6
      asks for: a held-out gold set re-scored by an injected machine judge, Cohen's kappa over a
      threshold sweep, a recommended cutoff that maximises agreement with the human verdict, and a
      fail-loud seam that throws rather than fabricate a calibration when the judge is absent. It
      calibrates the agentic-studio content judge. It names no release gate.
      
      BOTH STATISTICS ENGINES HAVE 0 IMPORTERS. `trust-safety/evaluation.ts` implements an
      expected-calibration-error over bins, a gold set and live cohorts; `search-discovery`'s
      offline evals implement nDCG, MAP and recall@k. Neither is imported outside its own library.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The image
      safety classifier is honest about what it is: a fail-closed adapter to a real moderation
      service that returns null without a key so the gate BLOCKS, and throws on a moderation
      outage rather than releasing ungoverned — its header says in as many words that a fabricated
      passing score would be the forbidden stub. (2) The safety score is a stated transformation
      of the provider's own per-category numbers — one minus the maximum violation — rather than a
      probability it claims to have calibrated. (3) The trust-safety module computes a real
      expected calibration error rather than reporting a calibration it did not perform.
      
      TRAPS. (1) A DOCUMENT THAT NAMES THE ARTIFACT AS OUTSTANDING IS NOT THE ARTIFACT: a sweep
      for "threshold decision record" returned `ISIS_GENERATION_WORKBENCHES_DESIGN` — the table
      where I9.6.f is listed as work still to do — and .f passed on it. And then A RULE THAT NAMES
      WHAT IT WATCHES COUNTS ITSELF, for the second time this phase: this cell's own evidence
      artifact quotes the checklist text, so writing it moved the candidate count from 1 to 2 and
      the report stopped being idempotent. The sweep now excludes its own output, and idempotence
      is checked by running the generator twice. Naming the constant is not
      deciding it either: `V1/AUDIT_2026-06-24.md` names all three floors in order to observe that
      the docs never surface them. (2) `corpora` IS A SUBSTRING OF `corporate`, so the marketing
      page read as a corpus; the survivors of a tightened sweep are a docs-retrieval fixture and
      the kit's credential- and stripped-content scanners, and none is a labelled evaluation set.
      (3) A PROPERTY OF A CORPUS IS ASKED OF THE CORPUS — over the joined generation path all five
      of .a's properties appear somewhere and .a read 5 of 5. (4) A NAME THAT COLLIDES IS NOT AN
      IMPORTER: `computeMap` matches ten files across asase, iris, nisaba and uzume.
      
      0 of 8 properties present. THE PARENT AND ALL SEVEN CELLS LEFT `[ ]`: there is no corpus, so
      there is no precision, no recall, no subgroup slice and no threshold decision; the one
      evaluator whose identity is declared floats; and the calibration machinery the estate did
      build serves a different judge. -->
      
  • I9.7 Build BFF contract tests for every route, role, seam, and failure. Evidence: route inventory coverage is complete.
    • I9.7.a Generate/reconcile the complete Isis/Yemaya/Oshun/gallery BFF route-method-operation inventory with handlers, schemas, owners, audiences, consumers, and tests. Evidence: route coverage manifest.

    • I9.7.b Assert request/response/content/error schemas, discriminators, versions, units, property absence/redaction, and unknown-field handling. Evidence: schema contract tests.

    • I9.7.c Run anonymous/curated/AAA/operator/reviewer/admin and cross- tenant/project object/property/function authorization matrices. Evidence: role contract tests.

    • I9.7.d Exercise service/provider/storage/queue/search/publisher seams in success, empty, partial, stale, not-configured, degraded, timeout, conflict, refusal, and malformed modes. Evidence: seam tests.

    • I9.7.e Assert pagination/idempotency/concurrency/cursors/deep links/ streaming/reconnect/rate/quota/budget semantics. Evidence: protocol contract tests.

    • I9.7.f Fail CI for unclassified routes, untested role/operation rows, or consumer/provider drift. Evidence: inventory coverage gate. <!-- MEASURED 2026-08-26 (I9.7.a-f). The parent's evidence is one sentence — ROUTE INVENTORY COVERAGE IS COMPLETE — so the binding question is a ratio: of the routes the BFF mounts, how many does a test actually call? Instrument: generate-oshun-bff-contract-coverage.mts (+9 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-bff-contract-coverage.json.

      A ROUTE THAT NOTHING CALLS IS NOT A COVERED ROUTE. 522 route modules declare 1,591 distinct
      paths under `/v1/` and the legacy `/admin/` prefix. Every declared path is reduced to the
      same shape as every URL any spec passes to `app.inject` — a parameter is `:id` on the mount
      and `${id}` in the injection, and a query belongs to neither — and matched. 602 of the 1,591
      paths are exercised and 989 are not; 203 of the 504 modules that declare a path have at
      least one of theirs called and 301 have none. The sweep is not blind: it found 1,491
      distinct injected URLs across 1,137 spec files.
      
      A WORD SOMEWHERE IN A THOUSAND SPECS IS NOT AN ASSERTION ABOUT A ROUTE. All 7 of .c's
      audiences, 9 of .d's 10 seam modes and 9 of .e's 10 protocol subjects appear somewhere
      across those spec files, and a first pass graded .c, .d and .e present on exactly that. Each
      of those cells is a statement about the ROUTE SURFACE — run the matrices, exercise the
      seams, assert the semantics — so each is now graded alongside the coverage ratio, and none
      can pass while 989 paths are never called. 4 of .b's seven schema subjects are named at all;
      a request, a response and an error schema are not asserted by name anywhere.
      
      A DISPOSITION RATCHET IS NOT A COVERAGE GATE. The estate does gate on routes: §I0.18's Isis
      route register, its disposition ratchet and the effective-route matrix are all three wired
      into `verify:inventory:v1-workbenches`, and each refuses a drifted CLASSIFICATION. 0 gate
      refuses an untested route. 5 of .a's six inventory attributes appear across the six route
      registers the estate holds; a handler is not one of them.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The specs
      that exist drive the real app through `app.inject` rather than calling handlers directly, so
      a covered route is covered end to end through its middleware. (2) The three route gates are
      real ratchets over real registers, and the classification work they protect is genuine — it
      is simply a different question from this one.
      
      TRAPS. (1) A RULE THAT NAMES WHAT IT WATCHES COUNTS ITSELF, for the THIRD time this phase:
      wiring this census into the verify script made .f pass on its own name, because the step
      that runs it contains the words a coverage gate would. Two steps match and both of them are
      this one; the census now excludes itself, and the count including it is published so the
      exclusion is legible. §I9.5's instrument matched its own sweep and §I9.6's evidence artifact
      quoted the checklist text — the standing lesson is to run the generator twice and `--check`
      before committing. (2) The first pass of this cell graded .c, .d and .e on word presence
      over a joined thousand-file corpus, which is the same file-scale error §I9.3 and §I9.6 each
      made once.
      
      0 of 7 properties present. THE PARENT AND ALL SIX CELLS LEFT `[ ]`: 989 of 1,591 mounted
      paths are never called by a test, 301 route modules have no exercised path at all, three of
      .b's seven schema subjects are not asserted by name, and the three route gates that do run
      ratchet a classification rather than a coverage. -->
      
  • I9.8 Build real-database and object-store integration tests. Evidence: durable run/output/provenance/recall round-trip.
    • I9.8.a Provision isolated real schema/object store/queue/outbox fixtures with migrations, two tenants, encrypted buckets, and deterministic cleanup. Evidence: integration harness.

    • I9.8.b Persist brief-to-plan-to-job/stages/checkpoints/attempts/variant/ decision with transactional audit/outbox and restart recovery. Evidence: durable run test.

    • I9.8.c Upload/read/hash/cache/quarantine/derive/package output objects and verify metadata, permissions, signed URL, corruption, and retention. Evidence: object round-trip tests.

    • I9.8.d Persist source/model/workflow/consent/rights/run/output/credential/ release lineage and verify forward/reverse/blast-radius queries. Evidence: provenance round-trip.

    • I9.8.e Execute release write/read-back then revoke/quarantine/recall and verify database/outbox/search/object/gallery target reconciliation. Evidence: recall round-trip.

    • I9.8.f Inject commit/outbox/publish/object/queue/process failures and assert idempotent recovery/no orphaned success/cross-tenant leak. Evidence: durability fault tests. <!-- MEASURED 2026-08-26 (I9.8.a-f). The parent's evidence is a DURABLE RUN / OUTPUT / PROVENANCE / RECALL ROUND-TRIP, so the binding question is asked of the four by name and then of the thing that would demonstrate them: does anything run these tests? Instrument: generate-oshun-durable-integration.mts (+8 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-durable-integration.json.

      THE SPECS ARE REAL AND NOTHING AUTOMATIC RUNS THEM. There are 144 integration specs across
      Oshun, 123 of them in the BFF, and 85 read a database URL from the environment —
      `OSHUN_V1_DATABASE_URL` in 87 places and `OSHUN_ADMIN_DATABASE_URL` in 80. They open a pool
      against it, migrate, write and read back, and they SKIP rather than pass when there is none,
      which is the honest behaviour for a local unit run.
      
      A CI STEP THAT PROVISIONS A DATABASE IS NOT A CI STEP THAT RUNS THESE TESTS. Of 106
      workflows, two set a database URL. `oshun-ci.yml` stands up a `postgres:16-alpine` service,
      sets `OSHUN_V1_DATABASE_URL`, applies `prisma migrate deploy` — and then runs the
      persistence EXPLAIN checks. It runs no integration spec. `v1-workbench-kit-verification.yml`
      sets the same variable for a loop of `npx tsx` scans. The only workflow in this repository
      that executes an `*.integration.spec.ts` is `nisaba-ci.yml`, for `libs/nisaba/schemas` — a
      different domain.
      
      AND THE JOB THAT DOES RUN INTEGRATION TESTS RUNS A TARGET THESE PROJECTS DO NOT DEFINE,
      which makes the absence narrower and sharper than "no integration job exists". `ci.yml`
      stands up its own Postgres, exports `DATABASE_URL` and `DATABASE_URL_DIRECT`, and runs
      `nx affected --target=test:integration`. Exactly 10 projects in the monorepo define that
      target — six Aphrodite apps, `libs/iris/testing`, `libs/maya/engine-core`,
      `libs/uzume/protocol-engines` and `testing/` — and 0 of them is under `apps/oshun`,
      `libs/oshun`, `apps/isis` or `libs/isis`. The BFF's own project declares `test` and
      `test:watch` and no integration target at all. So even the job that provisions a database
      and asks for integration tests asks for them by a name these projects do not answer to, and
      the variables it exports are not the two the 85 specs read. The round-trip the parent's
      evidence names is demonstrated by a person running a command, and by nothing else.
      
      3 OF THE FOUR ROUND-TRIPS HAVE A SPEC AND PROVENANCE HAS NONE. Four specs are named for a
      run, three for an output, one for a recall, and 0 for provenance or lineage — which is .d's
      whole subject.
      
      7 OF .a's EIGHT HARNESS SUBJECTS are present; the one missing is TWO TENANTS, so the
      cross-tenant leak .f asks to be refuted has no second tenant to leak to. 9 of .b's ten run
      subjects and 8 of .c's ten object subjects appear.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The specs
      skip rather than pass without a database, which is the difference between an honest local
      guard and a green tick over nothing. (2) Each documents the exact command that runs it,
      migration included, in a header comment, so a person can reproduce a round-trip without
      guessing. (3) The one CI job that provisions a Postgres does apply the migrations first, so
      the EXPLAIN checks it performs run against a real, migrated schema.
      
      TRAPS. (1) A WORD SOMEWHERE ACROSS 144 SPECS IS NOT AN ASSERTION IN THE ONE THAT ROUND-TRIPS.
      All five of .e's recall targets and all six of .f's faults appear somewhere in the joined
      corpus, and both cells passed on exactly that. Asked of the specs whose own names say they
      ARE a recall or a durability test, it is 1 of 5 and 5 of 6; asked of a lineage spec, .d's
      nine subjects are 0 of 9 because there is no lineage spec. This is the same file-scale error
      §I9.3, §I9.6 and §I9.7 each made once, and it is now the standing shape to check first.
      (2) A spec's own FILENAME is what decides which round-trip it is, because a durability spec
      is named for the thing it round-trips.
      
      0 of 7 properties present. THE PARENT AND ALL SIX CELLS LEFT `[ ]`: no CI step runs an Oshun
      integration spec, no spec round-trips provenance, the harness has no second tenant, and the
      recall spec that exists reconciles one of the five targets .e names. -->
      
  • I9.9 Build golden deterministic bundles for each delivery target. Evidence: validated build/import/read-back.
    • I9.9.a Enumerate every supported publishable asset-kind/delivery-target/ profile/version pair and its bundle consumer/validator. Evidence: target coverage manifest.

    • I9.9.b Create canonical minimal/full/edge fixtures with pinned source bytes, dependencies, metadata, accessibility, rights, provenance, and expected manifest. Evidence: golden inputs.

    • I9.9.c Build twice in clean isolated environments with pinned tools/ locale/time/order/config and compare hashes or declared nondeterministic fields. Evidence: determinism report.

    • I9.9.d Validate schema, files, hashes, formats, technical gates, signatures/credentials, target profile, and unsafe-package absence. Evidence: validation reports.

    • I9.9.e Import/consume through the target tool/runtime, read back semantic state, and compare expected identity/units/material/timing/metadata. Evidence: import/read-back tests.

    • I9.9.f Fail coverage when a supported target pair lacks a current golden, pinned validator, declared loss profile, or current proof. Evidence: bundle coverage gate. <!-- MEASURED 2026-08-26 (I9.9.a-f). The parent's evidence is a VALIDATED BUILD / IMPORT / READ-BACK, so the binding question is asked of the pair the cell is indexed by: for each supported asset-kind x delivery-target x profile x version, is there a golden that was built twice and compared? Instrument: generate-oshun-golden-bundles.mts (+11 tests), wired into verify:inventory:v1-workbenches, writing evidence/v1-workbenches/oshun-golden-bundles.json.

      THE BUILD IS REAL. `libs/oshun/workbench-kit/tools/golden-round-trip.mjs` is 473 lines that
      resolve a layout from a profile, write artefact bodies, serialise a manifest, archive with
      GNU tar, hash the bytes, and repeat the whole build under shuffled input order, touched
      mtimes, a Turkish locale, a Kiritimati time zone, a 0077 umask and interleaved writes —
      comparing tar HEADERS as well as digests, asserting `localeCompare` disagrees with byte
      order before it starts so the locale axis is known to be live rather than assumed, and
      extracting each archive to rebuild from the extracted bytes byte-identical. 5 of .c's 6
      perturbation axes are driven and the sixth is DISCLOSED: platform cannot be varied on one
      machine, and it is named as never perturbed rather than dropped from the list.
      
      A BUILD THAT AGREES WITH ITSELF IS NOT A BUILD THAT AGREES WITH A RECORDED ANSWER. Every
      comparison in that runner is `sha256(perturbed)` against `sha256(baseline)`, both computed
      in the same process moments apart by the same code. There are 0 checked-in digests.
      
      AND THE ONE TOOL THAT PERFORMS THE BUILD IS THE ONE THE CI GLOB DOES NOT MATCH. The kit's
      verification workflow partitions `tools/*-scan.mjs` into a files-only job and a
      needs-a-database job, and refuses an empty partition — a careful design. 197 of the 242
      tools in that directory end in `-scan.mjs`; 45 do not, and `golden-round-trip.mjs` is one of
      them. Its three golden siblings — `build-golden-scan`, `golden-manifest-scan`,
      `golden-runner-scan` and `golden-update-gate-scan` — all end in `-scan.mjs` and all run. An
      nx target `golden-round-trip` is declared in the kit's project.json and 0 workflows name it.
      So the determinism run this cell is about is performed when a person types the command.
      
      A DECLARED UNION IS NOT THE ONLY SHAPE AN AXIS TAKES. All four of .a's axes are declared,
      and two of them are FIELDS on a profile rather than an enum: 7 golden fixture classes
      (minimal, full, boundary, localized, provenance-rich, legacy-supported, invalid), 2 worked
      profiles, 3 delivery targets carried as `BundleProfile.targets` (wire, archive, lakehouse,
      each with a version range and required slots) and 2 `profileVersion`s. 84 pairs.
      
      A KIT PROFILE IS NOT A PRODUCT TARGET. 0 modules outside the workbench kit reach
      `BundleProfile` or `WORKED_BUNDLE_PROFILES`, so none of the 84 pairs has the bundle consumer
      or validator .a asks to be enumerated beside it — consistent with §I8.14's finding that the
      only thing in the product spelled `DELIVERY_TARGETS` is a four-member Nyx reminder-channel
      array. 0 importers read a bundle back through a target runtime, so .e's five semantic
      comparisons have nothing to compare.
      
      4 OF .b's SEVEN GOLDEN INPUTS are pinned — source bytes, rights, provenance and the expected
      manifest — and dependencies, metadata and accessibility are not. 3 OF .d's EIGHT VALIDATION
      SUBJECTS are covered: schema, hashes and the target profile. Files, formats, technical
      gates, signatures and the absence of an unsafe package are not.
      
      CHECKED AND FOUND CORRECT - recorded so a later pass does not re-flag them. (1) The runner
      is a real build, not a description of one, and its honesty about the platform axis is the
      difference between five axes driven and six claimed. (2) THE GOLDEN MANIFEST IS THE STRONG
      PART: `GoldenEntry` carries a contentHash and a `LicensedSource`, `content-has-drifted` is
      one of eight declared fault kinds, and `golden-manifest-scan.mjs` supplies a real negative
      control — an entry whose `recomputedHash` is deliberately wrong — so the worked manifest's
      agreeing hashes are a positive control rather than a comparison that cannot fail. That
      distinction was checked rather than assumed. (3) The workflow refuses an empty partition,
      which is the failure mode a glob-driven runner is most exposed to.
      
      TRAPS. (1) A TYPE ANNOTATION CONTAINS BRACKETS: `WORKED_BUNDLE_PROFILES: readonly
      BundleProfile[] = [ … ]` has a `[` before its initialiser, and a reader that scanned to the
      first bracket stopped inside the annotation and reported the declaration absent. (2) A
      DECLARED UNION IS NOT THE ONLY SHAPE AN AXIS TAKES — searching for `DELIVERY_TARGETS` and
      `BUNDLE_VERSIONS` put .a at 1 of 4 when the worked profiles carry both as fields. (3) An
      `importBundle` in another domain is not an import of this bundle: the bare names match an
      offline-bundle service in Lilith mobile and a user-model persistence service in Iris.
      
      0 of 8 properties present. THE PARENT AND ALL SIX CELLS LEFT `[ ]`: the build is real and
      nothing automatic runs it, it compares against itself rather than a recorded digest, the 84
      pairs it could be indexed by have no product consumer, and no bundle is read back through
      the runtime that would receive it. -->
      
  • I9.10 Build full-journey Playwright for operator model intake to release. Evidence: live backing with fixtures explicitly scoped.
    • I9.10.a Seed a quarantined model artifact with explicit external source, license/model card/provenance, security scan, evaluation fixtures, and operator/reviewer roles. Evidence: E2E fixture manifest.

    • I9.10.b Intake the artifact through the real UI/API/storage path and assert quarantine, hash, dependency, scan, and evidence states. Evidence: intake journey.

    • I9.10.c Complete license/provenance review and calibrated evaluations, assert blocking failures, correct them through new evidence/version, and preserve history. Evidence: review journey.

    • I9.10.d Promote the exact version to an approved environment with named approval, then select it in an admitted workflow and dispatch a run. Evidence: promotion/dispatch journey.

    • I9.10.e Compare output, inspect lineage/gates/cost, approve a locked candidate, assemble/write/read-back a target release. Evidence: release journey.

    • I9.10.f Assert direct bypasses, stale evidence, wrong role/tenant, secret/property leak, accessibility, and immutable trace/audit across the journey. Evidence: full Playwright trace. <!-- I9.10 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-intake-release-journey.mts (evidence evidence/v1-workbenches/oshun-intake-release-journey.json, 10 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: A JOURNEY IS A WALK, AND NOTHING HERE WALKS. The six children are six STAGES of one walk, not six subjects — .f's "immutable trace/audit ACROSS THE JOURNEY" cannot be asserted by a spec that only saw one stage, and .c's "correct them through new evidence/version, and preserve history" needs the artifact .a seeded to still be the subject three stages later. So the measurement is how many of the six stages any SINGLE spec reaches. Across all 995 e2e specs in the estate the answer is TWO, and only two specs manage it: operator-editorial-review-approval (review + release) and tenant-integrations-api-keys-webhooks-bff (release + negatives). Neither has a model artifact as its subject, and neither touches intake. Every spec that does touch model intake reaches exactly one stage.

      .a — 3 specs name quarantine and NONE is a model intake path: a studio quality/reliability sweep, a V2 moderation console, and a Neith CI insights dashboard. A SPEC NAMED FOR A SUBJECT IS NOT A SPEC THAT WALKS IT.

      .b — the surface the journey would have to walk EXISTS AND IS MOUNTED: apps/isis/web/src/pages/ModelsPage.tsx, routed at models in App.tsx, naming quarantine, with model-scan-utils.ts carrying seven scan verdicts (clean, failed, infected, skipped, unknown, unrecognized, unsupported). It has 0 e2e specs. The entire Isis web e2e suite is 13 generation control panels — kitbash packs, turnaround sheets, terrain, vehicles. And the one spec anywhere about external model intake, studio-isis-civitai-integration.spec.ts, exists to assert the surface is UNBOUND: mode: 'unbound_empty_projection', producer: 'none', persistence: 'none', authorityBinding: 'not_configured', authoritativeState: false. That is an honest fail-loud seam and the spec is a correct lock on it — it is simply not an intake journey, because there is no intake to journey through.

      .d — A CHILD THAT NAMES TWO ACTS IS NOT EARNED BY ONE. 5 specs promote to an environment; 0 of them dispatch a run. Exactly 1 contains the word "dispatch" and it is dispatchEvent firing pointerdown in a Yemaya drawing smoke test. dispatchEvent IS NOT DISPATCHING A RUN, and the matcher in the instrument excludes it.

      .e — 36 specs read back or assemble something; 0 of them also inspect lineage AND cost, which is half of what the child asks before it gets to approving a locked candidate.

      Four of the cell's named states are asserted by NOTHING, each recorded with a POSITIVE CONTROL on the same corpus so an all-zero column and a broken matcher cannot look identical: locked candidate 0 (control "candidate" 46), stale evidence 0 (control "evidence" 244), SBOM / bill-of-materials 0 (control "dependency" 14), named approval 0 (control "approve" 113).

      THE PARENT'S OTHER HALF HOLDS. "Live backing with fixtures explicitly scoped" is real where specs exist: of 98 admin e2e specs exactly ONE intercepts the network (redesign-incident-signoff.spec.ts). The suite is live-backed. What is missing is the journey, not the backing.

      TWO TRAPS WORTH KEEPING, both of which produced a wrong number here first. \| INSIDE AN ERE IS A LITERAL PIPE: a first pass scanned this corpus with grep -E 'read-back\|readBack' and reported 0 for every multi-alternative term when the true count is 36, and licence and lineage were likewise reported absent when they are in 11 and 40 specs. Every matcher in the instrument is a real alternation and a test asserts none contains \|.

      And A URL IS NOT A COMMENT. The first comment stripper treated the // in 'http://127.0.0.1:3020' as the start of a line comment and deleted the rest of the line, which removed both front-end ports from the one spec in the estate that names two of them. The stripper is now string-aware and a test locks it. Correcting it moved this cell's stage counts (.c 4->6, .e 31->36, .f 22->36) and three controls, and moved nothing that the grading rests on: the deepest walk is still two of six and no child is earned. -->

  • I9.11 Build full-journey Playwright for Yemaya brief to gallery promotion. Evidence: cross-host lineage assertions.
    • I9.11.a Seed authorized Yemaya project/shot/assets/roles and live or explicitly hermetic Isis capabilities with rights/provenance. Evidence: fixture manifest.

    • I9.11.b Author and validate a typed brief in Yemaya, hand off exact context to Isis, preview/confirm the plan, and dispatch. Evidence: outbound handoff test.

    • I9.11.c Observe queue/run/stages, return to exact Yemaya context, inspect generated proposal/output, and accept a project revision. Evidence: return journey.

    • I9.11.d Promote the exact Yemaya output into an Oshun curated derivative and complete editorial review without inheriting AAA authority. Evidence: editorial boundary test.

    • I9.11.e Assert immutable project/shot/brief/plan/run/output/original/ derivative/decision lineage, rights, gates, and audit across hosts. Evidence: cross-host lineage assertions.

    • I9.11.f Exercise expired/deleted context, generation failure, stale output, wrong tenant/role, back navigation, accessibility, and no duplicated operator controls. Evidence: negative E2E matrix. <!-- I9.11 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-brief-gallery-journey.mts (evidence evidence/v1-workbenches/oshun-brief-gallery-journey.json, 10 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: A CROSS-HOST ASSERTION NEEDS TWO HOSTS, AND NO SPEC IN THE ESTATE OPENS A SECOND ONE. The parent's entire evidence clause is "cross-host lineage assertions", and .e spells out what must survive the crossing — project, shot, brief, plan, run, output, original, derivative, decision, "across hosts". That is not a vocabulary question but a question about whether one Playwright context ever drives two application front ends, because a lineage assertion that never leaves one origin is an assertion about one host. Of 995 e2e specs, 346 bind a host by some spelling and 0 name a second. Exactly one spec names two front-end ports — phase-5-6-operator-review-queue pairing Oshun web (3010) with Oshun admin (3020) — which is one product's two consoles, not the Yemaya -> Isis -> Oshun span this cell describes. cross-host itself appears in 0 specs against a live control of 99 for "host".

      .b — AN EVIDENCE ARTIFACT IS NOT A PRODUCT CONTRACT. The typed generation brief is real and carefully specified: 11 facet interfaces (BriefIdentity, BriefIntent, BriefAudience, BriefAssetKind, BriefLocale, BriefReference, RightsClaim, BriefSafety, QualityTarget, BriefBudget, BriefDelivery) with 8 validators, from I1.4. It lives at evidence/v1-workbenches/generation-brief/schema.ts, and 0 files under apps/ or libs/ name it. There is no typed brief in Yemaya to author, so there is no handoff to preview, confirm or dispatch.

      .a/.c/.d — Yemaya's entire e2e footprint in the estate is TWO specs: apps/yemaya/studio-web/e2e/studio-web.smoke.spec.ts (1,727 lines, 24 test titles — welcome page, dark mode, a BIM scene review, a scene-graph lighting workspace, colour references, a Disney timing chart, HDR trims, a PM dashboard, translation memory; 0 of the 24 names a brief) and apps/oshun/web/e2e/studio-yemaya-budget-godot.spec.ts. Neither seeds a project for a journey, and a return to a Yemaya context needs a departure no spec makes.

      .f — A PHRASE DISCUSSED IN A COMMENT IS NOT AN ASSERTION. "no duplicated operator controls" matches 3 specs in the raw corpus and 0 once comments are stripped: all three are prose. Both numbers are recorded, and the grading uses the stripped one.

      A TRAP THIS CELL PAID FOR: A URL IS NOT A COMMENT. The first comment stripper treated the // in 'http://127.0.0.1:3020' as a line comment and deleted the rest of the line, which removed both front-end ports from the one spec in the estate that names two and reported the cross-host port measure as 0 instead of 1. The stripper is now string-aware in this instrument and in I9.10's, and a test in each locks it. -->

  • I9.12 Build full-journey Playwright for curated card generation and editorial promotion. Evidence: tier boundary assertions.
    • I9.12.a Seed curated/contemplative and editorial actors, approved card capabilities/presets, provider-gated/live states, and destination collection. Evidence: E2E fixture manifest.

    • I9.12.b Complete one representative text/image, audio, temporal, and 3D curated briefing with only approved high-level controls. Evidence: curated card journeys.

    • I9.12.c Validate estimate/confirmation/dispatch/progress/failure/output gallery states and audience-safe errors/provenance. Evidence: generation journey.

    • I9.12.d Select an exact output and request editorial promotion; complete separate reviewer decision and verified collection release. Evidence: promotion journey.

    • I9.12.e Assert raw graph/model/provider/sampler/node/LoRA/private prompt data is absent from UI, payloads, URLs, bundles, errors, and telemetry. Evidence: tier property leak assertions.

    • I9.12.f Assert crafted hidden fields/direct routes/deep links/tier switch/ stale output/cross-tenant attempts fail, with accessible responsive journey. Evidence: boundary Playwright matrix. <!-- I9.12 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-curated-card-journey.mts (evidence evidence/v1-workbenches/oshun-curated-card-journey.json, 20 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: THE TIER IS AN ATTRIBUTE, NOT A GATE. The parent's whole evidence clause is "tier boundary assertions", and .e names what the boundary must hold back — raw graph, model, provider, sampler, node, LoRA and private prompt data, absent from UI, payloads, URLs, bundles, errors and telemetry. Unlike I9.11, the subject here is real and good: ten approved cards in a server-authoritative catalog, per-card validators, an entitlement gate, a raw-graph leak auditor, a fail-closed BFF route with twelve route tests (nine of them refusals), and a real executor wired at server.ts:2929. What is missing is the walk and the boundary.

      The customer-facing generation gallery reads the caller's tier exactly ONCE, and writes it into a data-gallery-tier attribute. Zero conditionals branch on it. Its filter rail ships data-filter-model and a data-branch-param="guidance" (the CFG scale) to every caller, plus generator / workflow-class / tenant / entitlement-tier inputs, and the per-output panel prints Workflow class, Model and Generator. The §24.11 boundary that would refuse a tier is keyed on /studio/isis/, and 0 of the 5 curated generation routes are under it. The card shell itself prints Workflow class: <id> and Provider: <key> / <region> on every card.

      AND NOBODY WALKS IT. Of the five curated generation front ends, TWO are named by any Playwright spec: /studio/generation/living-scene (one spec) and /studio/generation-gallery (two). The curated-cards studio hosting six of the ten cards is named by ZERO, as are the music and nyx-3d studios. 0 of the 10 approved card kinds is briefed by a spec, so 0 of .b's four modality classes — text/image, audio, temporal, 3D — is walked.

      .e — the estate's entire stock of tier leak assertions is FOUR expect() calls, lines 100-103 of living-scene-card.spec.ts, against innerHTML, on a card that is not in the catalog: not.toContain of lora, scheduler, workflowgraph, comfy. One of .e's six surfaces is covered and five are empty. The honest denominator: 2,788 absence assertions of every kind live in the corpus. AN OPERATOR PANEL HIDDEN FROM AN UNAUTHORIZED VIEWER IS NOT A CURATED SURFACE PROVEN CLEAN — 38 more absence assertions naming raw machinery run across 13 Isis specs (lora-training, comfyui-nodes, model-registry, provider-factory...), and they ask whether the OPERATOR route refuses a principal, which is the opposite question about a different surface. The populations are counted apart, not merged. A further 28 matches are incidental: \bprompt\b matches data-pwa-install-prompt, so A PWA INSTALL PROMPT IS NOT A GENERATION PROMPT — reported as one number the non-curated population was 66 and looked like coverage.

      .c/.d — AN INSPECTOR IN ITS IDLE STATE IS NOT AN OUTPUT INSPECTED. The two gallery specs select 0 outputs and assert 1 of the 6 provenance states the panel can reach: idle ("Select an output to inspect"). They cannot do better: bindGalleryStore and bindGalleryContext are the gallery's only way to hold records and their sole callers are in gallery.test.tsx — A DESTINATION BOUND ONLY BY ITS OWN TEST, while the loader's docstring says production wires it. And A BUTTON WITH NO HANDLER IS NOT AN ACTION: 4 of the gallery's 6 action buttons — launch-branch, send-to-editorial, send-to-asset-library, takedown — carry no onClick at all. The estate's one promotion verb, send-to-editorial, is an operator-admin bulk action in the admin store that 0 e2e specs call.

      .a/.f — all 10 catalog cards are allowlisted to every tenant with the kill switch off and a single minimum tier, so there is no provider-gated state to seed and no cross-tenant refusal to exercise. 3 of .f's 7 negative classes (crafted hidden fields, direct routes, tier switch) are covered by the BFF route suite; deep links, stale output, cross-tenant and the accessible responsive journey are covered by nothing, and 0 classes by a Playwright matrix. A ROUTE TEST IS NOT A PLAYWRIGHT JOURNEY.

      FIXED: A PENDING PRICE PRINTED AS ¢0.00 IS A FABRICATED ESTIMATE. The catalog ships PENDING_COST_COEFFICIENTS on purpose and exports areCostCoefficientsPending to recognise them; the predicate had ONE caller in the estate and it was its own test. All three studios (curated-cards, music, nyx-3d) passed the sentinel to CardShell, which formatted it as Est. ¢0.00 · 0.0s mean (0.0s p95) — a free, instant generation stated to two decimal places — while the catalog test carried the comment (UI shows "pending"). The shell now renders Est. pending with data-preflight-estimate="pending"; two tests lock it, both confirmed RED against the old shell, and the control proves a published 0.5c/s x 4s still prints ¢2.00.

      TRAPS THIS CELL PAID FOR. A SUBSTRING IS NOT THE IDENTIFIER: matching the catalog cardId with includes credited the illustration card with a briefing, because an admin workflow-template example is called curated-illustration-prompt. Anchors are now identifier-bounded on both sides with hyphen as a word character, so curated-video no longer swallows curated-video-from-image. A WORD IN A SPEC THAT WALKS ANOTHER CARD IS NOT A BRIEFING OF THIS ONE: the loose measure credited video to the Living Scene spec, which says "video" about an audio-description track; the loose count is kept as a labelled control. A SUBJECT NAMED ON ANOTHER LINE IS STILL THE SUBJECT: classifying expect(html).not. toContain('lora') line-at-a-time returned unclassified and reported all six of .e's surfaces as zero, hiding the one thing the estate has — the binding two lines up decides it. And EVALUATE THE CATALOG, DO NOT GREP IT: counting tenantAllowlist: 'all' literals reported ONE entry allowlisted to all tenants, because a single factory writes the field once for all ten; the instrument now imports the catalog and holds the values. lora is a substring of exploration (24 specs unanchored, 5 with a word boundary), so every property token is word-boundary matched. -->

  • I9.13 Run WCAG 2.2 AA automation over critical routes and semantic alternatives. Evidence: reports for themes/viewports/locales.
    • I9.13.a Maintain a critical-route/state/action/role inventory covering operator, AAA, curated, gallery, errors, dialogs, live updates, media, graph, and 3D alternatives. Evidence: accessibility coverage manifest.

    • I9.13.b Run automated rule scans after stable load and key interactions for every supported light/dark/high-contrast theme. Evidence: theme reports.

    • I9.13.c Run keyboard-only focus/order/visibility/trap/shortcut/drag alternative and screen-reader name/role/value/status announcement tests. Evidence: interaction reports.

    • I9.13.d Run narrow/mobile/tablet/desktop and 200%/400% zoom/reflow/ orientation/touch-target checks without horizontal two-dimensional scrolling except essential content. Evidence: viewport reports.

    • I9.13.e Run default/long-text/RTL/CJK/number/date/unit locale cases plus reduced motion and browser text spacing overrides. Evidence: locale reports.

    • I9.13.f Verify noncanvas semantic alternatives provide equivalent inspect/comment/compare/decide operations for media/graph/3D. Evidence: parity journeys.

    • I9.13.g Triage every violation with owner/severity/route/state/artifact, block applicable release issues, and retain current reproducible reports. Evidence: zero-blocker accessibility gate. <!-- I9.13 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-wcag22-run-coverage.mts (evidence evidence/v1-workbenches/oshun-wcag22-run-coverage.json, 16 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 7 children earned; every box stays unchecked.

      THE VEIN: A REGISTER OF THE HARNESS IS NOT A REPORT FROM A RUN. Five of the seven children begin with "Run" and each names its evidence a REPORT — theme, interaction, viewport, locale, parity. The estate has something rare here: a careful accessibility REGISTER from §S11.3 — 8 axe harness sites with engine bindings, tag selections, impact policies and exclusions, a WCAG 2.2 criterion map, a findings register, a waiver register, and a matrix that separates a row a machine can fill from a row that needs a person. None of that is a run. Across the five axes this cell names, the answers are 0, 0, 3-of-4, 0 and 0.

      A SCAN CONFIGURED FOR 2.1 CANNOT REPORT A 2.2 FAILURE. axe-core 4.11.4 ships exactly ONE rule tagged wcag22aa — target-size, SC 2.5.8 — and ships it enabled: false, so an untagged run evaluates 0 of 1 and only a tag selection naming wcag22aa turns it on. The DEFAULT gate, the axeAfterEach fixture 620 Oshun specs inherit, selects ['wcag2a','wcag2aa','wcag21a','wcag21aa']. 2 of the 8 registered harnesses select 2.2 and both are opt-in helpers. Of the three specs named wcag-aa-signoff, two pass the 2.1 list back into the 2.2-capable helper by hand; the third (admin, 3512) inherits the default and is the estate's only sign-off that evaluates target-size — under a title that reads "WCAG 2.1 AA signoff". None of this is dishonest: those specs were built for a 2.1 TODO and say so. It is simply not 2.2.

      .a — A SIGN-OFF OVER TWELVE SHELL SURFACES IS NOT AN INVENTORY OF TEN STATE CLASSES. The three sign-offs enumerate 12 surfaces (Home, Library, Search, Notifications, Settings, six domain entries, and the admin dashboard/inbox/operators/review-queue/audit/compliance/reports) against 719 member-app routes. 9 of the 10 classes .a names appear in none of them; the one that does is operators. Of 646 Oshun e2e specs, 54 run a scan, and AAA, curated, live-update, media, graph and 3D-alternative states are named by 0 of them.

      .b — A THEME NOBODY SETS CANNOT BE SCANNED. The product ships three variants (cream, dusk, cream-hc). cream-hc appears in 0 specs. forcedColors emulation: 0. And 0 specs run a scan inside the test that sets a variant — the file-scope control says 1, and that one spec scans in the test at line 60 and switches to dusk in the test at line 284.

      .c/.d/.e — 49 specs drive the keyboard, 0 name a drag alternative (WCAG 2.2 SC 2.5.7), 0 assert focus-visible. Scans run at 3 of 4 viewport classes — 320x720, 390x844, 1440x960, 1600x960 — with no tablet width at all; 1 spec touches zoom or device scale, 1 names reflow. The 5 locales passed to Playwright's locale option are de/en/en-US/es/fr, all Latin and left-to-right; 2 specs reach RTL another way, 1 names ja-JP, 0 override text spacing, and 0 of the 13 locale-setting specs runs a scan.

      .f/.g — 1 spec names a semantic alternative and 0 of .f's four operations (inspect/comment/compare/decide) is paired with one. The gate fails on 2 of axe's 4 impact levels and reads an unset impact as moderate, the level it passes. 0 axe report artifacts are retained in the repository.

      FIXED: the estate's own axe-policy-scan was RED. It re-derives its reach counts every run and fails on drift — by design — and commit 62564a5ca6 added a 561st web spec (and a 51st scan call site across a 21st file) under a register that still read 560 and "50 call sites across 20 files". It runs on every push in v1-workbench-kit-verification.yml and is not on that workflow's KNOWN_FAILING list, so CI had been failing on it. Three reach strings are current and the scan reports "the register agrees with the engine and the source, 17 negative controls detected". A REGISTER THAT RECORDS A COUNT GOES STALE, which is exactly why that one re-derives.

      TRAPS THIS CELL PAID FOR. A HELPER IS NOT A SPEC: the first withTags census ran over *.spec.ts only, found one call site in the whole estate, and would have reported that nothing in Oshun ever asks for 2.2 — both Oshun helpers do, in e2e/support/accessibility.ts, which is not a spec file. A VERSION TAG AND A CRITERION TAG SHARE A PREFIX: axe encodes success criteria as wcag<major><minor><criterion>, so wcag222 is SC 2.2.2 Pause Stop Hide, a WCAG 2.0 rule — a /^wcag22/ prefix test reported FIVE 2.2 rules with three enabled by default and inverted the cell's central fact. A FILE IS NOT A TEST (the dusk measure above). A PLURAL IS NOT A WORD-BOUNDARY MATCH: \boperator\b rejects operators, so the one covered inventory class read as zero. A LOCALE PREFIX IS A COMMON IDENTIFIER FRAGMENT: a bare \bar- matched data-isis-ar-boundary and reported an Isis analytics panel as an RTL spec. And axe is a substring of relaxed and maxed: across the 995-spec e2e corpus an unanchored search finds 302 and a word-bounded one finds 127. -->

  • I9.14 Run performance and load profiles for comparison, large galleries, queue views, and high-volume jobs. Evidence: budgets pass.
    • I9.14.a Ratify client/server/end-to-end latency, interaction, memory, CPU, network, payload, query, throughput, concurrency, and error budgets per scenario/device/network. Evidence: performance budget spec.

    • I9.14.b Profile image/video/audio/3D comparison load/sync/scrub/annotation with large candidates and semantic alternatives. Evidence: comparison profile.

    • I9.14.c Profile gallery search/filter/pagination/preview/lineage/ virtualization with production-scale assets/collections/metadata. Evidence: gallery profile.

    • I9.14.d Load queue/dashboard/run explorer live updates, reconnect, filtering, pagination, and operator actions at production job/event volume. Evidence: queue view profile.

    • I9.14.e Load plan/dispatch/worker/provider/callback/gates/storage/outbox/ search/release paths with high-volume mixed jobs/failures/cancels. Evidence: backend load report.

    • I9.14.f Measure backpressure/fairness/tail latency/resource/cost/error/ stale behavior and stop tests before unsafe host pressure. Evidence: resource-safe test logs.

    • I9.14.g Compare against ratified budgets, capture profiles/traces/query plans, assign regressions, and rerun to current passing reports. Evidence: performance gate. <!-- I9.14 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-performance-load-profiles.mts (evidence evidence/v1-workbenches/oshun-performance-load-profiles.json, 10 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 7 children earned; every box stays unchecked.

      THE VEIN: A PAGE LOAD IS NOT A LOAD, AND THERE IS NO LOAD GENERATOR HERE. Four of the seven children ask for something to be put UNDER LOAD — large candidate sets, production-scale galleries, production job and event volume, high-volume mixed jobs with failures and cancels. The estate's entire performance apparatus is Lighthouse, and a real one: a committed per-scenario budget file, numeric assertions, and a BLOCKING lighthouse-web job in oshun-ci.yml that fires whenever @oshun/web is affected. Lighthouse loads a page three times and measures the paint. Of eight known load generators — k6, @grafana/k6, autocannon, artillery, loadtest, clinic, wrk, vegeta — 0 are declared in any package manifest and 0 are installed. The gap in .d/.e/.f is one missing tool, not four missing efforts.

      .a — A BUDGET PER SCENARIO IS NOT A BUDGET PER DEVICE AND NETWORK. lighthouse-budget.json is better than a first read suggests: three path patterns each with their own tightened numbers (/domains/nisaba gets FCP 1500 where /* gets 1800). What it has one of is the other two axes — one device preset (desktop) and one throttling profile (rtt 150ms, 1638.4kbps, 4x CPU), across both apps. 6 of .a's 12 dimensions carry a number anywhere: client latency, end-to-end latency, interaction, CPU, network and payload. Server latency, memory, query cost, throughput, concurrency and error budgets have none. The V3 first-load budget does carry two device/network profiles (M1 Mac broadband, Pixel 7 cellular) with recorded p95 measurements — one dimension, and for V3.

      .b/.c/.d — 8 routes are profiled (3 web, 5 admin). 0 are a comparison surface, 0 are a gallery, and 4 are queue-shaped: the member home dashboard and the admin review, moderation and incident queues. AND A PAGE LOAD OF AN EMPTY QUEUE IS NOT A PROFILE AT PRODUCTION VOLUME: neither Lighthouse config and neither Lighthouse workflow contains a seed step, a database service or a fixture load, so what those four measure is the pages with nothing in them.

      .f — 0 Oshun tests read host pressure, though CLAUDE.md makes it an operating rule for humans and agents. There is no load to measure backpressure, fairness or tail latency under.

      .g — the blocking client gate is real and the only Lighthouse workflow that auto-triggers is the admin one (the web workflow was deliberately demoted to workflow_dispatch so the identical budget check does not run twice per PR). What is not retained is any run output: 0 profile, trace or flamegraph artifacts and 0 captured query plans across apps/oshun and libs/oshun. The error-budget half of .g is settled elsewhere — I9.3 measured the SLO and error-budget apparatus and earned 0 of 8 — and this report reads that count out of that cell's committed evidence rather than re-deriving it.

      TRAPS THIS CELL PAID FOR. AN ARRAY OF 501 ITEMS IS NOT 501 CONCURRENT REQUESTS: counting bare Array.from({ length: N }) reported the estate's largest load level as 501, and it is a 501-element request BODY in a reminders route test. A fan-out now counts only when the literal-length array is built inside a Promise.all that issues a request — which makes the declared fan-out 0 — and the dynamic form (Promise.all(rows.map(...)), 23 sites) is reported beside it, because reporting only the literal form would say "zero concurrent requests" about an estate that has some. AND READ THE WHOLE BUDGET FILE: the first forty lines show one /* entry, and the cell was one edit away from recording "a budget for one path pattern" about a file that carries three. -->

  • I9.15 Run adversarial prompt, graph, upload, SSRF, provenance, auth, tenant, spend, and fixture scans. Evidence: zero release blockers.
    • I9.15.a Test prompt/reference/import/retrieved-content injection against system-policy separation, tool grants, data exfiltration, hidden control, and approval claims. Evidence: prompt adversarial suite.

    • I9.15.b Test graph/node/subgraph/parameter/custom package injection, unsafe deserialization/code, dependency confusion, cycles, resource abuse, and sandbox escape. Evidence: graph suite.

    • I9.15.c Test media/archive/model/package/document uploads for malware, parser bombs, traversal/symlink, polyglot/spoofed type, oversize, and metadata/script payloads. Evidence: upload suite.

    • I9.15.d Test URL/provider/webhook/deep-link/import fetch paths for SSRF, DNS rebinding, redirects, credential metadata, internal networks, scheme/ port/host bypass, and response bombs. Evidence: SSRF suite.

    • I9.15.e Test provenance hash/signature/manifest/ingredient/watermark/ ledger stripping, substitution, collision, replay, and false recovery. Evidence: provenance suite.

    • I9.15.f Test authentication/session/API key, route/function/object/ property authorization, mass assignment, enumeration, CSRF, replay, step-up, and confused deputy. Evidence: auth suite.

    • I9.15.g Test two-tenant identifiers/cursors/search/counts/cache/blob/ queue/events/logs/export/recall timing and content isolation. Evidence: tenancy suite.

    • I9.15.h Test quota/budget/reservation/cost/priority/batch/cache/refund/ marketplace/chargeback evasion and concurrency. Evidence: spend suite.

    • I9.15.i Scan fixture/demo/fallback/placeholder/mock state for production reachability, unlabeled data, simulated success, and release eligibility. Evidence: fixture honesty suite.

    • I9.15.j Triage by exploitability/impact with reproducible evidence, owners, fixes, regression tests, and zero unresolved release blockers. Evidence: security gate report. <!-- I9.15 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-adversarial-matrix.mts (evidence evidence/v1-workbenches/oshun-adversarial-matrix.json, 10 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 10 children earned; every box stays unchecked.

      THE VEIN: A VERB IS NOT AN ATTACK, AND traversal IS USUALLY A KEYBOARD. The cell names roughly eighty attack classes across ten suites, and the tempting measurement is to grep the corpus for each word. That method is worthless here and the corpus proves it: 21 test titles say "traversal" and three are a path escaping a directory — the rest are KEYBOARD traversal and tree walks. 79 say "enumerat" and two are account enumeration — the rest are lists the geography enumerations. Every probe in the matrix therefore comes in two forms, a STRICT one requiring the attack's context and a LOOSE one that is the naive word, and both counts are reported per class because the difference IS the finding. Every family is overstated by the naive word, and a test locks that.

      The population is test TITLES, not files: 74,045 of them across 7,679 test files in apps/oshun, libs/oshun, apps/isis, libs/isis and tests/security. A title is where a test says what it tests, and "the word appears somewhere in this 900-line file" has been the wrong denominator four times in this phase.

      AND A SUITE OUTSIDE THE APP TREE IS STILL A SUITE. tests/security/ sits at the repository root and holds a real, good adversarial suite — 6 files, 47 titles, over cross-domain token audiences, per-domain API-key vaults, tenant-scoped queries including the prefix-collision case, a PII scanner and a secret-shaped-string scan. An apps/**-scoped corpus would have missed all of it and reported .f and .g as empty. They are not empty.

      THE RESULT: 52 of 75 attack classes are named by a test title, and the shape of the gap is the point. The estate is strongest exactly where a web product is normally strongest — .f auth 10/10 and .i fixture honesty 4/4 — and thinnest exactly where THIS phase is named: .a prompt adversarial 3/7 (no retrieved-content injection, no system-policy separation, no tool-grant abuse), .b graph 4/8 (no graph/node injection, no dependency confusion, no custom-package injection, no unsafe code execution), .h spend 5/9, .g tenancy 6/10, .e provenance 6/9, .c upload 6/7 (no parser bomb), .d SSRF 6/7 (no response bomb), .j gate 2/4.

      .f IS THE CLOSEST AND STILL NOT EARNED, for a reason worth stating: A TITLE NAMING AN ATTACK IS NOT AN ADVERSARIAL TEST. The matrix measures whether the estate has SOMETHING pointed at each class, not whether that something attacks — and .j, the gate that would certify a suite complete, names two of its four classes. Every class in the artifact carries an example title so a reader can check the claim rather than take it.

      TRAPS THIS CELL PAID FOR, and it paid for many. A PORT IS A SUBSTRING OF A REPORT (also support, export, transport, important): the unanchored scheme/port/host probe reported 141 SSRF bypass tests, and the first example was "reports block mode when cost is unavailable". A CROSS IS A SUBSTRING OF ACROSS: every tenancy probe matched "moves the cursor across rows". search is a substring of research, count of counterclaim and account, cost of costume, log of catalog. A REVIEW CYCLE IS NOT A GRAPH CYCLE, a FALSE-POSITIVE RATE is not a FALSE RECOVERY, a MISSING config is not a STRIPPED manifest, an IDENTITY collision is not a HASH collision, a REFUND FEATURE is not CHARGEBACK EVASION, and "fires the simulator successfully" is not a SIMULATED SUCCESS. Three tightening rounds took the count from 60 to 52.

      AND TWO TRAPS IN THE INSTRUMENT ITSELF. A DOUBLED BACKSLASH IS A LITERAL BACKSLASH: a patch script that emitted \\b into the probe table produced regexes matching a literal backslash-b, which no title contains, and three families silently read zero; a test now compiles every probe and rejects one. A TABLE-DRIVEN TEST STILL HAS A TITLE: 91 files use it.each([...])('title', ...), the first extractor expected the quote immediately after the opening paren, and every one of those titles was dropped — the argument list is now skipped by balancing brackets, which recovered 277 titles. -->

  • I9.16 Re-run generation-tier and Studio-boundary suites. Evidence: no raw machinery is reachable in contemplative Oshun Studio.
    • I9.16.a Enumerate contemplative/curated/AAA/operator host routes, operations, controls, contracts, capabilities, bundles, and expected tier availability from canonical policy. Evidence: boundary manifest.

    • I9.16.b Crawl/render contemplative Oshun Studio routes and bundles for raw workflow/model/provider/node/sampler/LoRA/graph/training/fleet/ operator identifiers and commands. Evidence: static/runtime leak scan.

    • I9.16.c Send direct/guessed/linked/aliased routes and crafted query/body/ nested/unknown fields for every higher-tier operation. Evidence: route/ function authorization tests.

    • I9.16.d Test capability/health/error/not-configured/degraded/telemetry/ export/deep-link/search/autocomplete channels for hidden feature detail. Evidence: side-channel tests.

    • I9.16.e Verify approved curated cards map only high-level inputs through governed APIs and AAA surfaces remain Yemaya-scoped. Evidence: positive tier journeys.

    • I9.16.f Test actor tier/entitlement change mid-session, stale client state/cache, back navigation, service worker/offline cache, and revoked links. Evidence: dynamic boundary tests. <!-- I9.16 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-studio-tier-boundary.mts (evidence evidence/v1-workbenches/oshun-studio-tier-boundary.json, 11 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: THE BOUNDARY IS NAMED FOR ONE DOMAIN AND THE STUDIO HOSTS TWENTY. The §24.11 boundary is genuinely good and better than its own documentation says. STUDIO_ISIS_ALLOWED_ROUTE_SEGMENTS is EMPTY and isAaaOnlyRoute returns true for every segment not in it, so all 87 segments mounted under /studio/isis/ are fail-closed for a contemplative session — AN INVERTED ALLOWLIST IS NOT A DENYLIST WITH GAPS, and this cell nearly recorded the opposite. What the boundary does not reach is the other 416 Studio routes: the member app mounts 504 pages under /studio/ and 88 are under the governed prefix. aja, bellona, hathor, neith, nous, concordia and more are outside it, as is /studio/generation-gallery, which I9.12 measured shipping a Model filter, a Generator filter and a guidance-scale override with no tier condition on them.

      .a — the boundary matrix (BASE.5.e) is real, passes, and enumerates 31 route segments for ONE host (Yemaya AAA) and ONE of .a's seven dimensions. .a asks for four host tiers and for operations, controls, contracts, capabilities and bundles besides routes.

      .b — A DOOR THAT IS LOCKED STILL HAS A SIGN ON IT. The Studio index (app/studio/page.tsx, 1,490 lines) is not under the governed prefix, contains ZERO tier or entitlement references, and renders 87 links to governed /studio/isis/* routes — 9 of whose visible labels name raw machinery: "Open Isis LoRA Training workspace", "Open Isis GPU Worker Architecture workspace", "Open Isis Civitai Integration workspace", "Open Isis ComfyUI Nodes workspace". The destinations 404 for a contemplative session. The names are printed to it. Beyond the index, 37 of the 504 Studio routes name raw machinery within one import hop and 6 of those are outside the boundary. 0 bundle scans exist anywhere in the repository, so the "and bundles" half of .b has no instrument at all.

      .c/.d/.f — 85 specs drive a governed route and 82 assert a boundary outcome, which is real coverage of the prefix that is already fail-closed. At LINE scope, 1 of .d's ten side channels and 1 of .f's five dynamic conditions are paired with the boundary by any spec, and 0 specs name a mid-session tier or entitlement change at all.

      .e — 0 of the 10 approved curated card kinds is walked by a Playwright spec; read out of I9.12's committed evidence rather than re-derived.

      FIXED: A FUNCTION NAMED FOR A CATALOG IT DOES NOT READ. AAA_ONLY_STUDIO_ROUTES lists 31 segments and enforces nothing — the kit's own route-integrity.ts says so in as many words — while isAaaOnlyRoute, which sounds like its reader, consults the empty allowlist instead. Two docstrings on AaaGate.tsx told readers an unknown segment renders its children as a "defensive default"; it does not, it is gated, and that is the safe behaviour. A comment describing the opposite of the enforced behaviour on the estate's #1 no-regress boundary is an invitation to "fix" the code toward it and open the other 56 segments. Both docstrings are corrected and isAaaOnlyRoute now carries the warning at its definition. No behaviour changed; 30 entitlements tests and 1,478 studio component tests stay green.

      TRAPS THIS CELL PAID FOR. EVERY ROUTE REACHES EVERYTHING THROUGH THE SHELL: the first static leak scan walked each route's whole import closure and reported 501 of 503 routes reaching raw machinery — through components/library/librarySavedItems.ts, where the word was checkpoint and it meant a STUDY checkpoint. A CHECKPOINT IS A LEARNING MILESTONE and A SAMPLER IS ALSO A DEPTH SENSOR (an Aja depth-sensing catalog field). The report now scans comment-stripped sources at two depths and prints both. A WORD SOMEWHERE IN A SPEC IS NOT A PAIRING: file-scope pairing reported 70 specs testing the "error" side channel against the boundary, which is only the observation that 85 specs touch the boundary and most files contain the word "error"; line scope reports none, and both are kept. AN export KEYWORD IS NOT AN EXPORT CHANNEL: every TypeScript module begins with one, and the bare probe paired 19 specs on lines that were declarations. -->

  • I9.17 Publish Isis operator and incident runbooks. Evidence: game-day validates provider disable, model revoke, and output recall.
    • I9.17.a Inventory operational scenarios, owners/on-call/escalation, prerequisites, permissions, dashboards/queries, commands, evidence, communications, and recovery criteria. Evidence: runbook index.

    • I9.17.b Publish provider degradation/disable/drain/failover/restore procedure with in-flight/reservation/cost and verification steps. Evidence: provider runbook.

    • I9.17.c Publish model/workflow security/license/safety quarantine/revoke/ replacement/blast-radius procedure. Evidence: model/workflow runbook.

    • I9.17.d Publish output/bundle/gallery/marketplace/release quarantine/ recall/downstream verification/notification procedure. Evidence: recall runbook.

    • I9.17.e Publish queue/worker/storage/cache/database/search/outbox/ credential/provenance/dependency outage and recovery procedures. Evidence: infrastructure runbooks.

    • I9.17.f Publish child/identity/voice/biometric/rights/privacy/security/ spend incidents with restricted evidence and authority/escalation. Evidence: sensitive incident runbooks.

    • I9.17.g Execute timed game days for provider disable, model revoke, and output recall; capture gaps, update runbooks/automation, and verify closure. Evidence: game-day reports. <!-- I9.17 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-isis-runbooks.mts (evidence evidence/v1-workbenches/oshun-isis-runbooks.json, 7 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 7 children earned; every box stays unchecked.

      THE VEIN: A GAME DAY FOR THE INFRASTRUCTURE IS NOT A GAME DAY FOR THE GENERATION LIFECYCLE. The parent names three drills as its evidence — provider disable, model revoke, output recall — and the estate HAS run a real game day. On 2026-08-14 it injected five faults against a disposable estate through scripts/operations/v1-fault-control.mjs and recorded timings, an acceptance matrix, a follow-up register and retests: gd:database-loss, gd:queue-loss, gd:blob-store-loss, gd:regional-loss, gd:relational-corruption. Every one is a piece of infrastructure going away. 0 of the 3 named drills appears in the report, and model revoke is described by 0 runbooks anywhere in the estate.

      A FOLLOW-UP REGISTER FOR ONE SCENARIO IS NOT A REGISTER FOR THE GAME DAY. The register carries 30 follow-ups, all scoped to gd:regional-loss — one of the five exercised — and stamped estate: production while the executions ran on estate: disposable.

      The runbook estate itself is good. docs/runbooks/index.md is a real on-call index with owner, applies-to, incident shape and severity per row, and the directory holds procedures for provider failover, model/workflow rollback, provenance failure, watermark verification failure, privacy incidents and ten infrastructure losses. 19 of the 29 procedures .b–.f name are described by some runbook.

      AN UNINDEXED RUNBOOK IS INVISIBLE TO ON-CALL. The index tells its own readers "you're writing a new runbook — add a row here in the same PR", and 14 of the 49 runbooks beside it have no row — nine of them the Yemaya study family, including yemaya-study-model-replacement.md and yemaya-study-provider-change.md, which are the closest things in the estate to two of the three drills the parent asks for.

      AND A DIRECTORY NAMED RUNBOOKS IS NOT A DIRECTORY OF RUNBOOKS. docs/domains/isis/runbooks/ holds 1,001 files; 7 are operational procedures. The rest are per-asset-kind design documents — workflow graphs, typed request contracts, budget guardrails, endpoint sizing policies — filed under a path that makes any count of "Isis runbooks" meaningless until it is split.

      Per child: .a — the index carries 3 of .a's 9 dimensions (scenarios, owners/on-call, permissions); commands, dashboards, evidence and communications live in the procedures instead, and recovery criteria are stated by 4 runbooks. .b 4/5 (no provider DISABLE, only degradation and failover). .c 3/4 (no model revoke). .d 1/4 — the recall child is the thinnest in the cell: recall appears in one runbook and marketplace in none. .e 7/9 (no worker-stall or cache-loss procedure). .f 4/7 (no identity-incident, spend-incident or restricted-evidence procedure).

      A TRAP THIS CELL PAID FOR: A KEY THAT DOES NOT EXIST AND AN EMPTY LIST READ THE SAME. The follow-up register's array is followUps; guessing entries/register reported 0 follow-ups for a register that holds 30, and a zero from a wrong key looks exactly like a measured absence. -->

  • I9.18 Produce engineering-readiness evidence. Evidence: all required reports are current and linked.
    • I9.18.a Define the required readiness index by requirement, capability, host/audience, asset kind, provider/endpoint, target, standard, owner, evidence type, and freshness rule. Evidence: evidence schema.

    • I9.18.b Collect current contract/schema/migration/unit/integration/E2E/ accessibility/performance/security/evaluation/DR/runbook reports by immutable commit/artifact links. Evidence: automated evidence index.

    • I9.18.c Validate artifact existence/hash, test result/pass state, environment/dependency versions, coverage scope, timestamp/freshness, and no fixture/live ambiguity. Evidence: evidence validator.

    • I9.18.d Map every Isis parent/child task and ratified capability/matrix row bidirectionally to evidence or explicit not-applicable decision. Evidence: zero-orphan traceability report.

    • I9.18.e Summarize open blockers/risks/exceptions/limitations/owners/ deadlines without allowing summary success to hide failed cells. Evidence: readiness dashboard/report.

    • I9.18.f Reproduce the packet from repository and approved external records, sign/version it, and retain prior packets immutably. Evidence: reproducibility test. <!-- I9.18 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-readiness-evidence.mts (evidence evidence/v1-workbenches/oshun-readiness-evidence.json, 10 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: A LEDGER FOR ANOTHER CHECKLIST IS NOT THIS CHECKLIST'S LEDGER. Almost everything this cell asks for has been built, carefully, and points at a different programme. ISIS_V2_EVIDENCE_LEDGER.schema.json is a real evidence schema — nine required fields per entry including commitSha, verificationCommands, artifacts and owner. The ledger beside it holds 1,262 entries with every field populated, citing 5,490 distinct artifact paths of which 5,440 exist. A validator beside THAT checks the ledger against the ticket set and against the filesystem. Its sourceChecklist is ISIS_TODOS_2.md, its ids are DOD2-### and MXD-####, and 0 of the 1,262 match this checklist's I<n>.<m> numbering. This checklist holds 181 parents and 1,391 children, 760 of them ticked, and 8 carry a measurement record a machine can find.

      .b — A REQUIRED FIELD SATISFIED BY A PLACEHOLDER IS NOT AN IMMUTABLE LINK. All 1,262 rows carry the same commitSha and its value is the string PENDING_COMMIT. 0 of the 1 distinct value is a sha.

      .c — A GATE NOBODY RUNS IS NOT A GATE. scripts/isis/validate-v2-compliance-gate.mjs exits 1 today with 50 errors, every one an evidence entry citing a script that no longer exists. It is named by 0 package scripts and 0 workflows, so nothing has read those 50 errors. Of .c's seven validations it performs one — artifact existence. Its commit check is isNonEmptyString(entry.commitSha), which PENDING_COMMIT passes: the same shape as the NaN that cleared a range check written as two comparisons in I9.5.

      .a — the entry schema carries 3 of .a's 10 dimensions (requirement, owner, evidence type). Capability, host/audience, asset kind, provider/endpoint, target, standard and the FRESHNESS RULE — the field that decides whether a report is still current — are absent.

      .e — A LEDGER WHERE EVERY ROW IS DONE HAS NO FAILED CELL TO HIDE. The only status value across all 1,262 rows is DONE, which is precisely the shape .e asks a report not to have.

      .f — the estate proves its artifacts are CURRENT and not that they are SIGNED: 295 of the 309 generators under scripts/v1-workbenches/ carry a --check invocation in verify:inventory:v1-workbenches, which is a real reproducibility proof per artifact, while 21 of the 320 artifacts carry a hash or signature field and 2 prior packets are retained.

      WHAT IS ACTUALLY ANSWERED: .b's vocabulary half. 10 of the 12 report kinds .b names are named by some artifact path in the ledger (unit and evaluation by none), and 8 of 12 by the workbench evidence tree — though that measure is about the INDEX's vocabulary and not about whether the estate has that kind of test, since a unit-test report is not filed under a path containing unit.

      TRAPS THIS CELL PAID FOR. A * IN A GIT PATHSPEC CROSSES A SLASH: evidence/v1-workbenches/*.json returns all 415 JSONs under the tree, not the 320 at the top level. AND git ls-files CANNOT SEE THE FILE THIS RUN IS WRITING: counting the top level with git made the census change the instant it was committed — 319 while its own artifact was untracked, 320 after — so the two self-referential counts are read from the filesystem and the tree control stays a git read. -->

  • I9.19 Obtain named security/privacy, trust-and-safety, rights, and engineering approvals. Evidence: human approval records.
    • I9.19.a Define each approver role/person, authority/scope, required packet sections, independence/conflict rules, decision options, conditions, and expiry. Evidence: approval matrix.

    • I9.19.b Freeze the exact commit/build/config/environment/host/audience/ capability/target readiness packet presented for approval. Evidence: approval subject manifest.

    • I9.19.c Obtain security/privacy decision over threat model, tests, findings/exceptions, data flows, retention, tenancy, biometrics, and incident readiness. Evidence: named decision.

    • I9.19.d Obtain trust-and-safety/rights decision over child/identity/ voice/moderation/originality/consent/licenses/provenance/recall evidence. Evidence: named decisions.

    • I9.19.e Obtain engineering decision over architecture, migrations, reliability/SLO, capacity/performance, operations/DR, tests, and known limitations. Evidence: named decision.

    • I9.19.f Record approve/reject/conditional/request-changes separately, bind conditions/exceptions/expiry, and invalidate on material subject or evidence change. Evidence: approval lifecycle tests. <!-- I9.19 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-release-approvals.mts (evidence evidence/v1-workbenches/oshun-release-approvals.json, 18 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: THE TEMPLATES ASK WHO; THE RECORDS ANSWER WHAT ROLE. The parent's evidence sentence is "human approval records", and the estate has a corpus of them: 9 records under docs/releases/v1/verification/, 71 decision rows, all dated 2026-05-15, each written in exactly one commit and never revisited. docs/launch/go-no-go.md — the template they were filled in from — asks for five columns: Role, SIGNATORY, Decision, TS, Evidence. The record filed under verification/go-no-go/ has three: Role, Decision, Evidence. The two it dropped are the two that make an approval attributable to a person at a moment. docs/launch/safety-review.md asks for signatories: <name> (role); the three board records carry a column headed "Required role", which is the ROSTER table from its §2.1 rather than its signoff block.

      AND WHERE A NAME WAS ACTUALLY ASKED FOR, THE ANSWER IS A MACHINE. Across every role-and-name signature table in the repository — 47 cells in 11 files — 0 name a person: 37 are blank underscore rules, 6 are the literal <name> placeholder in the template itself, 2 read Automated Audit and Automated Validation, and 2 name a team. The zero is falsifiable: anything the classifier cannot place lands in the person bucket and is listed, and that list is empty.

      A ROLE CANNOT BE CHECKED FOR INDEPENDENCE. .a asks for independence and conflict rules; the 9 records mention independence, conflict, recusal and expiry 0 times, and carry 0 identifiers that resolve to an account. You cannot ask whether the Privacy Lead who approved the dogfood drill is the same person as the Privacy Lead on the Privacy Review Board when neither is a person. Every one of the 71 decisions is APPROVE or PASS.

      THE ONE GATE THAT DEMANDS A NAME IS RED, AND THE GA APPROVAL SAID GO ANYWAY. docs/security/pentest/signoff.md is the artifact this cell asks for, built properly: 10 fields per engagement, 3 of them "Name + ISO date", append-only rows, six release-gate rules including a 14-day signoff expiry, and a CI step in .github/workflows/release.yml that refuses the tag. Its engagements table reads _none yet_; the gate exits 1 today. The GA approval recorded on the day that gate was wired says "Security and compliance: pass", "waivers: none", "resulting decision: GO".

      FIXED, IN THAT GATE: A MISSING FOOTER IS NOT A CLEAN ONE. Rule 4 blocks a release on an open AuthN, RBAC or privacy finding, and ran only if (tagsMatch) — so an absent tags-summary: footer and a footer declaring every category clear reached the same verdict, the shape of I9.17's key that does not exist reading the same as an empty list. AND THE THIRD SIGNATURE WAS OPTIONAL IN CODE: rule 5 checked signoff_sec and signoff_eng and never signoff_priv, so the head of privacy was the one of the three the document names whose signature the gate did not want. Both enforced now, the record's rule text aligned with what the gate actually anchors on, and 9 tests that fail four ways against the previous code.

      .b — A SUBJECT NAMED BY A BRANCH IS NOT FROZEN. 1 of 8 subject dimensions appears anywhere in the corpus, and the one that does is release commit: main against a template whose field is <sha> — the shape of the PENDING_COMMIT that satisfied a required commitSha in I9.18. The tag it names, v1-launch-rc, is not one of the repository's 1 tags. 20,390 commits have landed since (measured 2026-08-26; not in the artifact, because a HEAD-distance would go stale on every commit anywhere and --check would fail for reasons having nothing to do with approvals).

      .a — the estate's approval matrix, docs/domains/isis/remediation-v2/ISIS_V2_SIGNOFF_MATRIX.md, carries 5 of .a's 7 dimensions, and it is scoped to ISIS_TODOS_2.md — the same different-checklist trap as I9.18's ledger. The 2 dimensions it lacks are the 2 that make an approval revocable: independence/conflict and expiry. Its "Primary Owner" column names 7 team handles and no person, and its decision-record template is quoted in exactly 1 file: itself.

      .c — A PRIVACY BOARD IS NOT A SECURITY BOARD. There is no security board record; the privacy board's names 2 of .c's 8 subjects (tests, tenancy), with threat model, findings/exceptions, data flows, retention, biometrics and incident readiness absent. The Security Lead appears only as a one-line APPROVE row with a phrase for evidence. .d — the two safety boards name 3 of 9 (child, moderation, provenance). .e — the records the Engineering Lead signs name 4 of 7, missing architecture, capacity and known limitations.

      .f — NO VOCABULARY IN THE ESTATE CAN SAY ALL FOUR THINGS. 219 literal lists across libs/ and apps/ carry both an approval and a rejection spelling; 173 carry 2 of .f's 4 options, 46 carry 3, and 0 carry 4. Only 3 can say conditional and none of those 3 can say changes_requested. THE ESTATE BUILT THE LIFECYCLE .f DESCRIBES, TWICE, AND IMPORTS NEITHER: libs/oshun/workbench-kit/src/signoff.ts has validForHours, declared independence dimensions and material-change detection, and libs/shared/review-persistence/ publishes a decision-lifecycle and a release-validity service — 0 product importers each.

      A SCENARIO LIST IS NOT AN EXERCISE REPORT. §3 of the safety-review template makes 10 contents mandatory before a board may sit; 2 name no path, and 4 of the remaining 8 are absent — the moderation backlog, the persona-tone audit, the DSAR packet, and the red-team exercise report: docs/security/red-team/ holds the scenario set and 0 dated reports.

      WHAT IS ACTUALLY ANSWERED: the estate does check its approval records. libs/oshun/analytics/src/v1-33-launch-readiness-evidence.ts indexes 11 of them and a green vitest suite opens each file and asserts its decision string. All 11 strings are pass strings — so a board that changed its mind would not be recorded as a rejection, it would break the build.

      TRAPS THIS CELL PAID FOR. A NAME THAT COLLIDES IS NOT A CONSUMER, twice: signoff.js is a substring of accountable-signoff.js, which is Yemaya's own module, and signoffStatus is a field on Annapurna's safety task type — a substring probe and a bare-identifier probe each reported 2 importers where there are 0. A SIGNATURE TABLE IS ONE WITH A THIRD COLUMN: a name-and-role header alone admits V1/architecture/glossary.md, whose "Role" is a description of a domain library, and an escalation contact list; requiring a date, signature, decision or status column cut 61 cells to the 47 that are somewhere a person signs. AND A RULE THAT NAMES WHAT IT WATCHES COUNTS ITSELF, the fourth time on this checklist: wiring this census into verify:inventory:v1-workbenches put the string verify-signoff into a package script, and the probe promptly reported that script as running the gate. It runs the gate's tests. -->

  • I9.20 Obtain named product launch approval per audience and host. Evidence: human signoff remains separate for operator, AAA, and curated surfaces.
    • I9.20.a Define distinct launch cells for Isis operator host, Yemaya AAA integration, Oshun curated Studio, and Oshun gallery with audience, regions, capabilities, providers, targets, flags, and rollout plan. Evidence: launch matrix.

    • I9.20.b Assemble cell-specific product evidence for user journeys, utility/usability/accessibility, honest states, documentation/support, analytics, limitations, rollback, and prerequisite governance approvals. Evidence: launch packets.

    • I9.20.c Obtain a named authorized product decision independently for each cell rather than inheriting another audience/host’s signoff. Evidence: separate human approval records.

    • I9.20.d Bind approval to exact release subject, enablement/flag/region/ provider/target scope, conditions, success/stop metrics, and expiry. Evidence: scope validation.

    • I9.20.e Revalidate every cell immediately before rollout and pause only affected cells when approval/evidence/dependency changes. Evidence: prelaunch gate tests.

    • I9.20.f Record phased rollout/hold/rollback/outcome and require new approval for material scope expansion. Evidence: launch history. <!-- I9.20 measured 2026-08-26 by scripts/v1-workbenches/generate-oshun-launch-cells.mts (evidence evidence/v1-workbenches/oshun-launch-cells.json, 10 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: A SUBSTRATE IS NOT AN AUDIENCE. .a asks for four launch cells — Isis operator host, Yemaya AAA integration, Oshun curated Studio, Oshun gallery — each with its own audience, regions, capabilities, providers, targets, flags and rollout, so .c can approve each separately and .e can pause one without the others. The estate's rollout plan partitions the release 7 ways and every one is a SUBSTRATE: customer web, iOS, Android, BFF, domain services, workers, ML inference. That is a partition by what runs the bytes, not by who receives them. All four cells ride on customer web, and 0 of the 4 is a row in the plan, so the smallest thing it can hold back is all four at once.

      AND AN ENTITLEMENT IS NOT A LAUNCH SWITCH. The estate knows exactly who a user is: libs/isis/entitlements/src/generation-tier.ts resolves an entitlement bundle to operator-admin, aaa-creator, curated-creator or contemplative and returns a surface allowlist, and 37 modules spell the same four values. None of that vocabulary reaches the enablement layer. The 4 flag modules that decide what is switched on target a tenant, a domain, a locale, a country, a platform, a rollout percentage and a SUBSCRIPTION tier — a billing band, not a generation tier. Generation tier, region, provider and launch cell are targetable in 0 of them, and the kill switch's two scopes are global and agent_family. No cell has a switch of its own, which is .e's whole request.

      .c — A DOMAIN PM IS NOT A CELL PM. exit-criteria/2026-05-15.md is the closest thing: 11 owner signoffs, 6 of them product managers for a named surface. There is no Isis row — Isis is the subject of this entire phase — no Yemaya row and no gallery row. One row names a cell's host without saying which of its audiences it approves ("Studio PM"), and above all 11 sits a single aggregate GO / NO-GO = GO, which is exactly the inherited signoff .c asks each cell not to take. All 11 decisions are APPROVE.

      .f — A SAMPLE OF ONE STAGE IS NOT A LAUNCH HISTORY. The plan approves 28 stages across the 6 substrates whose schedule is enumerable (domain services says "per-service ECS schedule" and states none). The verification tree holds 1 stage record, titled "Canary Analysis Sample", whose verdict is a recommendation to promote once a manual review completes. 0 promotions, 0 holds, 0 rollbacks, 0 outcomes recorded, and no rule requiring a new approval for material scope expansion.

      .a/.b/.d — 2 of .a's 7 per-cell dimensions (flags, rollout) are named by the launch documents at all; 0 of .d's 9 binding dimensions are, including the exact release subject, conditions, success and stop metrics and expiry. There is 1 launch packet rather than 4, and it names 7 of .b's 8 subjects — honest states is the one it does not.

      WHAT IS ACTUALLY ANSWERED: the cell concept exists, once, built for something else. libs/contracts/src/content-coverage/seed-registry.ts declares a surface classification carrying BOTH introducingTier and productShell, and content-coverage/seed-registry.v1.json fills it in over 28 rows. Its 4 distinct pairs are aaa-creator x yemaya-studio, curated-creator x oshun-studio, operator-admin x oshun-admin and contemplative x oshun-contemplative-products — 2 of .a's 4 cells, with the operator host named oshun-admin rather than Isis and no gallery pair at all. Every tier occurs with exactly one shell, so it is a bijection recorded as data rather than a matrix somebody can add a cell to, and it is a content-coverage registry: it says which shell first introduces a surface at which tier, and nothing about launching.

      TRAPS THIS CELL PAID FOR. A VALUE IS NOT A FILE: asking whether the shell contract "names" the Isis operator host by testing isis and admin against its 665 lines returns true for every cell, because the two words occur in different places. Scoped to the enum's 4 values, 0 name a cell — they are host names with no audience half. Scope the probe to the thing, not to the file that contains it. -->

  • I9.21 Run matrix-generated contract, UI, workflow, runtime, bundle, and release tests for every asset family, conditioning mode, consistency strategy, provider/endpoint class, and delivery target ratified in I0.14-I0.17. Evidence: the Cartesian coverage report has no silent, inherited, or untested supported cell.
    • I9.21.a Generate canonical matrix dimensions/values/versions and explicit supported/unsupported/not-applicable constraints from I0 decisions and source-of-truth registries. Evidence: matrix manifest.

    • I9.21.b Generate only valid constrained cells and require a reasoned disposition for every excluded combination. Evidence: cell enumeration tests.

    • I9.21.c Map each supported cell to typed contract/schema/round-trip and audience/host property-authorization proof. Evidence: contract coverage.

    • I9.21.d Map each exposed cell to UI control/state/accessibility/ responsive and curated/AAA/operator tier proof. Evidence: UI coverage.

    • I9.21.e Map each cell to exact workflow/model/conditioning/continuity/ provider capability negotiation, execution, cancellation, and failure proof. Evidence: workflow/runtime coverage.

    • I9.21.f Map each publishable cell to technical/safety/rights gates, deterministic target bundle, write/read-back, and lifecycle proof. Evidence: bundle/release coverage.

    • I9.21.g Execute with bounded sharding/resource limits, collect per-cell evidence/failures/skips, and prohibit aggregate pass from hiding a cell. Evidence: resource-safe test report.

    • I9.21.h Fail release for supported cells with missing/stale/inherited/ generic-only proof and publish the complete Cartesian coverage report. Evidence: zero-silent-cell gate. <!-- I9.21 measured 2026-08-26 by scripts/v1-workbenches/generate-isis-matrix-cells.mts (evidence evidence/v1-workbenches/isis-matrix-cells.json, 8 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 8 children earned; every box stays unchecked.

      THE VEIN: FIVE ALPHABETS AND NO PRODUCT. All five dimensions this cell names are ratified and separately measured — 32 asset families in isis-asset-family-matrix.json, 31 canonical conditioning controls and 8 consistency dimensions in isis-conditioning-mode-matrix.json, 9 provider kinds in ISIS_PROVIDER_KINDS, 8 delivery targets in the target-profile catalogue. They multiply to 571,392 cells before constraints. Nothing multiplies them: 0 files declare all five, 0 artifacts enumerate a cell, and so 0 excluded combinations carry the reasoned disposition .b asks for. There is no cell for .h's report to call silent.

      AND NO TEST IN THE REPOSITORY NAMES TWO DIMENSIONS. Across 959,226 test titles in 27,332 spec files (the workbench instrument tree excluded): 497 name an asset family, 176 a provider kind, 10 a conditioning control, 3 a consistency strategy, and 0 a delivery target. Exactly 1 title names two dimensions, and it does so because identity-lock is both a canonical control and the lock form of the identity consistency dimension — one dimension counted twice. 0 titles name two DISTINCT dimensions. The estate tests along the axes and never at a point, which is what .c through .f each ask for.

      THE CLOSEST THING TO A MATRIX MANIFEST IS A ROUTE FILE. Two files name three of the five alphabets. The product one is apps/isis/generation-api/src/routes/jobs.ts, 9,821 lines, carrying all 32 families, 11 of 31 controls, all 8 consistency dimensions, 2 of 9 providers and 0 delivery targets — the endpoint that accepts a job, which is where three alphabets meet because a request mentions all three. The other is evidence/v1-workbenches/inventory/tests.json, an index.

      .f — THE DELIVERY-TARGET ALPHABET LIVES IN THE EVIDENCE TREE. Its 8 values are declared across 18 modules under evidence/v1-workbenches/target-profiles/, which 0 product files import. The one dimension along which a bundle is actually delivered is, in the product, a dimension of one — the same finding I8.12 recorded when it looked for a partial failure to recover from.

      .g — A SHARD OF A SUITE IS NOT A SHARD OF A MATRIX. 5 workflows shard a test suite (e2e, iris-ci, oshun-admin-e2e, oshun-web-e2e, tara-ci) and 0 shard an Isis matrix. .h — 0 zero-silent-cell gates and 0 Cartesian coverage reports exist.

      WHAT IS ACTUALLY ANSWERED: the asset-family axis, completely. All 32 of the 32 ratified families are named by at least one test title, and 498 titles name one. Against that: 4 of 31 conditioning controls, 2 of 8 consistency strategies, 7 of 9 provider kinds and 0 of 8 delivery targets have a test title that names them.

      TRAPS THIS CELL PAID FOR. A SUBSTRING IS NOT THE CONCEPT: seven of the eight consistency dimensions are ordinary English words — identity, environment, motion, prop, lens, lighting, wardrobe — and prop sits inside character-prop-interaction-sheet. Matched bare they report 13,002 titles and 69 apparent two-dimension tests, every one of them an asset family whose name contains a common noun; matched in the form that names a strategy (<dimension> lock, <dimension> consistency) they report 3. AND MEASURE TITLES, NOT FILES — with the .each([...]) argument skipped before the title is read, so a table-driven test contributes its template rather than its table. AND A RULE THAT NAMES WHAT IT WATCHES COUNTS ITSELF, the fifth time on this checklist: this census's own known-value tests live in a .test.mts file, so the instant they were staged the corpus grew by one file and eight titles and the artifact went stale for a reason that had nothing to do with the estate — and I9.22's instrument bumped it again the next day. The subject is the ESTATE's test titles, so the whole of scripts/v1-workbenches/ is excluded rather than just this file. -->

  • I9.22 Run full journeys for specialized 3D, cinematic sequence, voice/dubbing, foley/music, marketplace, model intake/training, and parity-monitor workflows. Evidence: each live/exposed capability has current host-specific proof rather than relying on generic generation coverage.
    • I9.22.a Run reconstruction-to-semantic-edit/material/LOD/rig/splat/ engine-DCC-package/read-back journey with non-destructive lineage and target gates. Evidence: specialized 3D E2E.

    • I9.22.b Run storyboard/animatic/shot continuity/render passes/patch rerender/editorial interchange/release journey with stable unaffected shot identities. Evidence: cinematic E2E.

    • I9.22.c Run verified consent through voice enrollment/TTS/clone/dub/ pronunciation/sync/multitrack mix/release/revocation. Evidence: voice/ dubbing E2E.

    • I9.22.d Run SFX/foley/music cue/generation/timeline/stems/loudness/mix/ rights/output bundle journey. Evidence: foley/music E2E.

    • I9.22.e Run creator enrollment/submission/gates/review/list/purchase/ entitlement/use/attribution/payout/takedown journey. Evidence: marketplace E2E.

    • I9.22.f Run external model intake/quarantine/license/security/evaluation/ compare and fine-tune/LoRA/merge lineage through promotion/revoke. Evidence: model supply E2E.

    • I9.22.g Run coverage-source ingest/divergence detection/traceability/ disposition/expiry/regression/release blocking journey. Evidence: parity-monitor E2E.

    • I9.22.h Execute each journey in its real authorized host/audience with live or explicitly hermetic dependencies, negative tier/tenant cases, accessibility, trace/audit, and current per-capability evidence. Evidence: per-journey live trace bound to the current evidence index. <!-- I9.22 measured 2026-08-26 by scripts/v1-workbenches/generate-isis-specialized-journeys.mts (evidence evidence/v1-workbenches/isis-specialized-journeys.json, 8 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 8 children earned; every box stays unchecked.

      THE VEIN: NOT ONE OF THE SEVEN JOURNEYS IS WALKED FURTHEST BY ITS OWN HOST. The parent's evidence sentence asks for host-specific proof "rather than relying on generic generation coverage". Isis owns all seven of these workflows and has 13 e2e specs. Across the estate's 1,013-spec Playwright corpus, the deepest any Isis spec gets into any of the seven journeys is ONE stage. The cinematic, voice/dubbing and foley/music journeys are each led by the same file — apps/yemaya/studio-web/e2e/studio-web.smoke.spec.ts, 1,727 lines, which really does drive a storyboard, a shot scene graph, an F5-TTS provider selection and a SyncNet lip-sync, and is a smoke test for another host.

      A JOURNEY IS A WALK, and 0 of 7 are walked. Deepest single spec per journey, against the stages the checklist spells out: specialized 3D 2 of 8, cinematic 3 of 7, voice/dubbing 4 of 10, foley/music 4 of 9, marketplace 2 of 11, model supply 3 of 9, parity monitor 1 of 7. 16 of the 61 stages are reached by no spec in the corpus at all — semantic mesh edit; animatic and editorial interchange; voice enrollment; the final mix; creator enrollment, submission, listing gates and entitled use; model licence, security scan and revoke; and four of the parity monitor's seven, including divergence detection itself.

      THE CORPUS IS ONE APP. 751 of the 1,013 e2e specs are apps/oshun/web; apps/isis/web has 13, and 11 of those are per-family generation control panels.

      .h — 1 spec of 1,013 opens a second host and 1 asserts a negative tier case; 9 assert a negative tenant case; 3 declare themselves hermetic. The two conditions the estate does meet broadly are accessibility (85 specs run axe) and trace/audit evidence (75).

      TRAPS THIS CELL PAID FOR. A NAME THAT COLLIDES IS NOT A DCC: \bmaya\b matched "Maya Torres", a teacher in a visual-regression fixture, and made a page-screenshot spec the deepest walker of the 3D journey at 3 stages. Requiring the application to be doing something to a 3D asset — or the match to be a file format — drops it to 2. The same discipline is why rig is anchored (it is inside trigger and original), clone is anchored (structuredClone), and mix is matched only as mixdown or mix bus. AND A URL IS NOT A COMMENT — the comment stripper leaves // alone when a colon precedes it, which is what I9.11 paid for. --> Evidence: specialized journey index.


Phase M — Metis courseware and assessment workbench#

Metis must support source-to-course authoring, accessible assessment-item production, psychometric review, adaptive sequencing, integrity governance, and standards-compliant institutional delivery. Learner delivery remains a separate surface from authoring and operator adjudication.

M0 — Product scope, host, and domain decisions#

  • M0.1 Reconcile /metis/*, /domains/metis, /operator/metis, apps/metis/admin, apps/metis/web, and proposed Studio routes. Evidence: one canonical route/host map. <!-- partial: current route inventory (a) done in evidence/v1-workbenches/metis-route-inventory.json. b-f (audience/host classification, canonical target map, dispositions, boundary review, ratification) are semantic/design/human and stay unchecked. --> <!-- SPEC 2026-09-11: .c is a product decision (Section 1 table: "canonical host to ratify in P0"), and M9 was built before it was taken, so authoring now lives in two apps (§2.5.1). An agent drafts docs/adr/ADR-M0.1-metis-canonical-host.md (Proposed) enumerating the three real options — keep the split (apps/metis/web/create + apps/metis/admin), consolidate authoring into apps/metis/admin and role-gate or retire /create/*, or an Oshun Studio facade in V1.2 — and builds .f's drift check against the option the ADR proposes; .c–.f flip only after a named human accepts the ADR. The M0.1 generators need no re-run. -->
    • M0.1.a Inventory every current route, app entry point, navigation link, deep link, API consumer, audience, mutation, and ownership for the named Metis surfaces. Evidence: current route inventory. <!-- evidence: .routes — 39 tracked pages across all 4 named surfaces (metis-admin 15, metis-web 13, oshun /metis 10, /operator/metis 1), each with source/host/mount/Next.js-route-path, apiConsumers (20 distinct /api/metis|admin|auth), navLinks, dynamic flag, and test. /domains/metis has no tracked pages. audience/mutation/ownership are the M0.1.b classification (semantic). generate-metis-route-inventory.mjs + 7 tests, wired into inventory/verify. -->

    • M0.1.b Classify each surface as authoring, institutional setup, educator operations, learner delivery, observer, or operator adjudication. Evidence: audience/host classification. <!-- evidence: metis-audience-classification.json — all 39 routes classified into the 6 canonical audiences via auditable host+path rules (see classify()): oshun-web /metis/* =learner-delivery except the two that author, /operator/metis=operator-adjudication, metis-web=learner-delivery(+authoring on /create), metis-admin spans institutional-setup/operator-adjudication/educator-operations/observer by path. No blank cells. Canonical target-host selection (M0.1.c), owner/migration disposition (M0.1.d), and named ratification (M0.1.f) are measured in evidence/v1-workbenches/metis-canonical-host-map.json. generate-metis-audience-classification.mjs + 8 tests.

                                                                                                                                                                                                                                                                      CORRECTED 2026-08-26 — THE KEY SAID STUDIO AND THE HOST FIELD BESIDE IT SAID
                                                                                                                                                                                                                                                                      DOMAIN SURFACE. The route inventory keyed this surface `oshun-studio-metis`
                                                                                                                                                                                                                                                                      while its own `host` field read `oshun-web (/metis/* domain surface)`, and the
                                                                                                                                                                                                                                                                      classifier read the key: one branch returned `authoring` for all ten routes
                                                                                                                                                                                                                                                                      with the rationale "Oshun Studio /metis authoring workbench". Nothing in the
                                                                                                                                                                                                                                                                      estate mounts a `/studio/metis` route. They are top-level `/metis/*` on the
                                                                                                                                                                                                                                                                      consumer web app, every one rendering a Lilith surface and titled "the
                                                                                                                                                                                                                                                                      schoolroom", "lesson", "lessons", "today's session", "tutor", "assessment" or
                                                                                                                                                                                                                                                                      "BYOM — bring your own materials". Two of the ten author and say so:
                                                                                                                                                                                                                                                                      /metis/ingest calls itself a "Creator-facing surface", /metis/courses/new is
                                                                                                                                                                                                                                                                      "propose a course". The key now says what the host field always said and the
                                                                                                                                                                                                                                                                      classifier reads the page; authoring fell 11 to 3, learner-delivery rose 12 to
                                                                                                                                                                                                                                                                      20. Two traps paid for on the way: A LINK TO AN AUTHORING SURFACE IS NOT AN
                                                                                                                                                                                                                                                                      AUTHORING SURFACE — /metis/lessons matched "propose a course" on the text of a
                                                                                                                                                                                                                                                                      button pointing at /metis/courses/new — and A REDIRECT IS NOT A SURFACE:
                                                                                                                                                                                                                                                                      /metis/upload renders nothing, permanently redirecting to /metis/byom, and now
                                                                                                                                                                                                                                                                      takes its target's audience. -->
      
    • M0.1.c Choose canonical route/host for every retained capability and distinguish embedded Studio facade from standalone app/service owner. Evidence: target route map.

    • M0.1.d Assign existing routes reuse, redirect-with-expiry, compatibility shim, retire, or net-new disposition with owner and migration date. Evidence: route disposition ledger.

    • M0.1.e Validate identity/tenant/context/deep-link/navigation boundaries and prohibit learner/answer-key/integrity/operator leakage. Evidence: route boundary review.

    • M0.1.f Ratify the canonical map and add route integrity/drift checks. Evidence: approved map and automated route test. <!-- M0.1.c-f measured 2026-08-26 by scripts/v1-workbenches/generate-metis-canonical-host-map.mts (evidence evidence/v1-workbenches/metis-canonical-host-map.json, 11 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 4 children earned; every box stays unchecked.

      THE VEIN: THE ENDPOINTS THAT RETURN THE ANSWER KEY ARE THE ONLY TWO IN THEIR FILE THAT DO NOT ASK WHO IS CALLING. .e asks the boundary review to prohibit learner, answer-key, integrity and operator leakage. apps/metis/service/src/metis/api/assessments.py has 14 endpoints; 12 take Depends(get_current_instructor_user) or Depends(get_current_user). The 2 that take db and nothing else are list_assessments and get_assessment, whose AssessmentResponse nests QuestionResponse carrying correct_answer and explanation, and whose docstring is "Get full assessment details including all questions".

      The committed contract says it without being asked: in apps/metis/service/openapi/metis.openapi.json those two operations are the only ones on the assessment paths with no security requirement while every neighbouring mutation declares HTTPBearer. Of the API's 183 operations, 37 declare no security and exactly 2 of the 37 return a schema carrying correct_answer — those two. And apps/metis/web/src/app/assessments/[assessmentId]/page.tsx is 'use client' and fetches that endpoint from the browser, so the key is in the client for the whole attempt; the client only declines to DRAW it until the results view (showCorrectAnswers ? question.correctAnswer : null).

      NOT FIXED HERE, DELIBERATELY. The repair is an authorization change on a Python service whose dependencies are not installed in this environment, on the read path the learner UI depends on. An unverified auth change there is worse than a measured one; the fix wants an environment that can run its tests. 0 of .e's 4 leakage classes is closed in that file.

      .c — A CAPABILITY ON TWO HOSTS WITH NO CANONICAL CHOICE IS NOT A MAP. The 39 routes cover 28 capabilities across 4 surfaces, and 4 capabilities sit on more than one: courses on three (metis-admin, metis-web, oshun-web-metis), plus assessments, login and the host root. 0 artifacts in the estate record which host is canonical for any of them, and none tells the embedded facade apart from the standalone owner.

      .d — A REDIRECT WITHOUT AN EXPIRY IS NOT A DISPOSITION. 0 of 39 routes carry one of .d's 5 dispositions, an owner or a migration date. The estate has exactly 1 redirect — /metis/upload to /metis/byom — and it is a permanentRedirect with no expiry, which is the disposition .d names redirect-with-expiry missing its second half. The Isis programme has a real disposition ledger and gate in 4 files; Metis has none.

      .f — A DRIFT CHECK OVER AN UNRATIFIED MAP KEEPS AN UNRATIFIED MAP CURRENT. Both Metis inventories genuinely carry a --check gate and both run in verify:inventory:v1-workbenches — that half of .f is real. There is no ratified map for them to keep current, and 26 of the 39 routes have no test: metis-web, oshun-web-metis and oshun-operator-metis have no tested route at all, so the 13 that are tested are all metis-admin.

      TRAPS THIS CELL PAID FOR. A DECORATOR IS NOT A SIGNATURE: the guard lives in the function's parameter list, so a probe reading @router.get(...) finds current_user in no endpoint and reports all 5 GETs as unauthenticated where the answer is 2. AND A WORD IS NOT A LEDGER: metis-package-inventory.json contains the string retire, which makes a naive probe report a disposition ledger where there is a package list. -->

  • M0.2 Resolve the domain-registry planned status and launch path before exposing the workbench in the shell. Evidence: approved registry decision. <!-- SPEC 2026-09-11: the registry facts are in M10.19's note. .d is buildable now (shell boundary tests over release-scope.ts and the apps/oshun/web middleware); .a–.c and .e are the decision record and stay open until ratified (§2.5.9). -->
    • M0.2.a Inspect registry status semantics, current Metis deployments/ routes/capabilities, shell visibility logic, feature flags, and consumer assumptions. Evidence: status current-state report.

    • M0.2.b Define objective entry/exit criteria for planned, internal, preview, limited availability, generally available, paused, and retired states. Evidence: lifecycle vocabulary.

    • M0.2.c Choose Metis launch state/path by host/audience/tenant/region and list prerequisites, dependencies, flags, rollback, and owners. Evidence: registry decision record.

    • M0.2.d Ensure shell navigation/capabilities derive from the canonical registry decision and never expose planned routes as usable. Evidence: shell boundary tests.

    • M0.2.e Define transition approval/evidence and automatic regression/ incident behavior before changing the registry. Evidence: ratified launch gate. <!-- M0.2 measured 2026-08-26 by scripts/v1-workbenches/generate-metis-registry-launch-state.mts (evidence evidence/v1-workbenches/metis-registry-launch-state.json, 9 known-value tests, wired into verify:inventory:v1-workbenches). 0 of 5 children earned; every box stays unchecked.

      THE VEIN: THREE REGISTRIES, AND THEY DO NOT AGREE THAT METIS EXISTS YET. domains.json carries a status field and Metis's value is active — so is every one of the other 57 domains, which makes a field with exactly one value across the whole file a column rather than a decision. The real cut sits in a neighbouring key on the same entry: customerReleaseScope: "deferred-to-v1.2". libs/oshun/domain-registry/src/registry.ts carries a different field, availability, whose union is 'active' | 'beta' | 'planned', and Metis is planned there. libs/oshun/navigation/src/release-scope.ts carries a third answer: V1_DEFERRED_DOMAIN_IDS is ['veritas','metis'].

      AND THE RECONCILIATION GATE READS TWO OF THE THREE SOURCES. generate-registry-reconciliation.mjs is real and runs in inventory:v1-workbenches, comparing domains.json against a ledger classification table; it reports 4 contradictions today. It never reads availability, so the one contradiction that decides whether this workbench may be exposed is the one it cannot see.

      FIXED: A ROOM HELD OUT OF THE RELEASE WAS REACHABLE THROUGH /operator. .d asks that shell navigation never expose planned routes as usable, and apps/oshun/web/src/proxy.ts — the Next middleware — enforces exactly that: it resolves a deferred room from the path, 404s the API and redirects the page to /release-scope/<domain>, before auth, so the refusal is identical for everyone. Its prefix list was ['', '/domains', '/d', '/api'], and apps/oshun/web/src/app/operator/metis/page.tsx mounts "Metis · operator" at /operator/metis — inside the proxy's matcher, guarded by nothing in app/operator/layout.tsx, and invisible to the resolver. /operator is now a prefix; 11 of the 11 Oshun-hosted Metis routes are covered, and four tests in libs/oshun/navigation/src/release-scope.spec.ts fail against the previous list. The same list is what caught 31 e2e-mined journeys in August, so this is the second time the gap was in a prefix nobody thought of, not in the rule.

      .b — A STATE A CONSOLE CANNOT REPRESENT IS A STATE A DOMAIN CANNOT BE IN. All 7 of .b's lifecycle states are spelled somewhere: 37 declarations across libs/ and apps/ carry two or more, and general_availability and limited_rollout both appear in libs/contracts/src/metis/index.ts. No single declaration carries more than 3 of the 7, and the union the shell actually consults carries exactly 1 — planned. There are no entry/exit criteria for any of them.

      .c — A COMMENT FIELD IS NOT A DECISION RECORD. The decision as written is $comment_customerReleaseScope on the domains.json entry. Across 15 candidate documents all 9 of .c's dimensions are named somewhere, which says only that the estate writes about Metis; 0 documents name all 9, and the best single one falls short.

      .e — A DRIFT REPORT IS NOT AN APPROVAL. The reconciliation runs in 2 package scripts and CODEOWNERS covers the registry, which is the half of .e that exists. 0 workflows gate a registry change on an approval, the report's own summary says contradictions "require human resolution", and nothing regresses automatically when the status moves.

      TRAPS THIS CELL PAID FOR. A DIMENSION NAMED SOMEWHERE ACROSS FIFTEEN DOCUMENTS IS NOT A DECISION RECORD — asked that way all 9 are present, so the measure is the best single document. AND A TWO-LETTER TOKEN IS NOT A LIFECYCLE STATE: ga is a substring of dozens of identifiers, so generally available is matched as a phrase or an explicit enum member and never as ga. -->

  • M0.3 Define creator, instructional designer, teacher, SME, psychometrician, accessibility reviewer, integrity reviewer, publisher, and tenant-admin roles. Evidence: permission matrix. <!-- SPEC 2026-09-11: the service knows three JWT roles (student, instructor, admin); the nine roles this cell names exist only as M7 assignment records. .a–.e are the matrix, drafted as Proposed; .f is buildable now as a parametrized suite over the served runtime and is the first thing to write, because it will fail on the M0.15.d leak. -->
    • M0.3.a Define each role’s purpose, eligible actor type, tenant/project/ course scope, prerequisites, delegation, conflicts, and lifecycle. Evidence: role definitions.

    • M0.3.b Map create/read/edit/comment/review/approve/publish/export/ configure/adjudicate operations for course/source/lesson/item/rubric/ assessment/credential records. Evidence: operation matrix.

    • M0.3.c Map protected answer/rationale/grade/learner/profile/integrity/ accommodation/PII/psychometric fields to property-level visibility. Evidence: property matrix.

    • M0.3.d Define requester/author/reviewer/approver/publisher separation, conflict-of-interest, quorum, substitute, and break-glass rules. Evidence: separation policy.

    • M0.3.e Define tenant-admin boundaries that permit local configuration and assignment but not platform floors, other tenants, or hidden learner/ integrity data. Evidence: admin boundary tests.

    • M0.3.f Test direct route/object/property/function, role combination, mid-session revocation, and two-tenant authorization. Evidence: permission test matrix. <!-- M0.3 measured 2026-08-27 by scripts/v1-workbenches/generate-metis-role-permission-matrix.mts (evidence evidence/v1-workbenches/metis-role-permission-matrix.json, 9 known-value tests plus the public-registration schema regression, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box stays unchecked.

      THE VEIN: THE PUBLIC DOOR LET A VISITOR WRITE "INSTRUCTOR" ON THEIR OWN BADGE. RegisterRequest accepted student | instructor; the unauthenticated /api/auth/register route passed body.role to AuthService.register_user, which persisted it and signed user.role into the access token. CurrentUser.is_instructor() treats either instructor or admin as privileged, so a new visitor could mint an instructor token and reach the service's instructor-guarded authoring operations. FIXED THIS PASS: public registration accepts exactly student; instructor, admin, and super_admin each fail validation. Administrator/institutional role assignment remains separate. The test follows the whole downstream chain so widening the public enum cannot later look like an innocent contract change.

      .a — SIX IDENTITY ALPHABETS, AND NONE DEFINES THE NINE PEOPLE THE WORKBENCH ASKS FOR. The Python principal has student/instructor/admin; the internal user schema has the same three; admin management adds moderator; the TypeScript gateway has student/instructor/content_creator/admin/ super_admin; agent visibility has learner/instructor/admin/operator; and the learner web profile adds teaching_assistant. Across those six are 9 distinct values. Only creator (as content_creator) and teacher (as instructor) have an identity/authorization spelling. SME and psychometrician appear only as curriculum WORKFLOW seats (discipline_methods, assessment_calibration), not assignable principals; the other five required roles have neither. 0 declarations carry all 9. Across 40 Metis-named document candidates, 0 documents name all nine roles and all seven definition dimensions.

      .b — WRITE IS NOT CREATE, EDIT, COMMENT, REVIEW, APPROVE, PUBLISH, EXPORT, CONFIGURE, AND ADJUDICATE. The gateway's real RBAC table has 5 roles and 48 cells over 14 permissions. Its only action vocabulary is access/admin/manage/read/write: exactly 1 of the 10 requested operations (read) is represented by its own decision. Of the 7 requested record families, exactly course and assessment are represented directly; source, lesson, item, rubric, and credential are not.

      .c — ONE SAFE PROJECTION IS NOT A PROPERTY-BY-ROLE MATRIX. Credit where it is due: QuestionResponse carries correct_answer and explanation, while SourceRenderedAssessmentQuestionResponse, the signed learner publication projection, carries neither. That is one correct projection. There are 0 complete declarations mapping the 9 protected families to roles, and the ordinary assessment reads identified in M0.1 still return the answer-bearing response.

      .d — A ROLE CHECK IS NOT SEPARATION WHEN ONE ADMINISTRATOR MAY REVIEW AND PUBLISH. The curriculum reviewer pool genuinely requires a quorum of at least two and a conflict-free result. The source-publication path limits candidate creation to instructor/admin and review/publication to admin, but compares creator, reviewer, and publisher identities in 0 of the 3 required pairings. One administrator may review and publish, and an admin creator may do all three. There is no substitute or break-glass rule.

      .e — THE GATEWAY TOKEN NAMES AN ORGANIZATION AND THE SERVICE PRINCIPAL DROPS IT. MetisJwtClaims requires organizationId, but that field is never consulted in gateway authorization. The Python CurrentUser has exactly id/email/role/name, the user aggregate has no tenant/organization owner, and tenant-admin is not an authorization role. Agent-runtime records do carry tenant IDs; that does not tenant-scope the human principal or admin authority.

      .f — 5 of 7 requested test dimensions have at least one textual candidate: direct route, object, function, role combination, and two-tenant. Property authorization and mid-session role revocation have none, and no suite generates the required role x operation x object x property x tenant matrix. The new public-registration regression closes one escalation path; it does not make this matrix complete.

      TRAPS THIS CELL PAID FOR. A ROLE WORD IS NOT AN IDENTITY ROLE — SME and psychometrician exist as review-seat labels but cannot authenticate. A WORD IN A DOCUMENT IS NOT A ROLE DEFINITION — the best Metis document names only 3 of 9 roles and 4 of 7 definition dimensions. AND THREE PROTECTED FIELD NAMES BESIDE THREE ROLE WORDS IS NOT A PROPERTY MATRIX: broad co-occurrence finds many long schemas/services, so the artifact records those as rejected lexical candidates and reports 0 complete matrices. -->

  • M0.4 Separate authoring data, learner data, assessment evidence, grades, integrity records, and institutional configuration. Evidence: data ownership and classification map.
    • M0.4.a Inventory entities/fields/stores/events/APIs/search/indexes/ exports/analytics/backups for each data category. Evidence: data inventory. <!-- evidence: metis-data-model.json — 14 SQLAlchemy entities (141 fields, 16 FKs, 14 relationships) with per-field type/nullability/index/foreign-key, grouped by M0.4 data category (learner-data 5, assessment-evidence 3, authoring-data 3, identity 1, institutional-config 1). exports/analytics/ backups + owner/classification are the M0.4.b-e governance half. generate-metis-data-model.mjs + 6 tests. -->

    • M0.4.b Assign canonical owner/store and classification, purpose, lawful/policy basis, access roles, encryption, retention/deletion, and residency per category. Evidence: ownership/classification map.

    • M0.4.c Define minimal versioned projections from learner/grade/integrity/ institutional owners into authoring previews and simulations. Evidence: projection contracts.

    • M0.4.d Prohibit copying raw learner history, grades, diagnoses, accommodations, integrity evidence, or institution secrets into course/ item authoring records. Evidence: schema/property tests.

    • M0.4.e Define cross-owner identifiers, transactions/events, correction/ deletion/legal-hold propagation, and audit without shared authority. Evidence: boundary flow tests.

    • M0.4.f Threat-test search, export, logs, analytics, fixtures, backups, tutor prompts, and previews for cross-category leakage. Evidence: data separation suite. <!-- M0.4 re-audited 2026-08-27 by scripts/v1-workbenches/generate-metis-data-ownership-boundary.mts (evidence evidence/v1-workbenches/metis-data-ownership-boundary.json, 10 known-value/trap tests plus 2 ORM boundary tests, wired into verify:inventory:v1-workbenches). 0 of 6 children earned; every box is unchecked, including M0.4.a, whose earlier evidence did not cover the surfaces named by the cell.

      THE VEIN: RESPONSE OMISSION HAPPENED AFTER PROTECTED DATA HAD ALREADY CROSSED THE BOUNDARY. User used eager selectin loading for authored courses, enrollments, submissions, tutoring sessions, and achievements. Course.enrollments, Lesson.progress_records, and Assessment.submissions did the same. A login/identity lookup therefore fetched the learner's whole adjacent graph; an authoring course read fetched enrollments/progress/submissions even when its DTO omitted them. FIXED THIS PASS: all 8 cross-owner relationships use lazy="raise" and require a caller to opt in with an explicit loader. Same-owner authoring edges remain explicitly eager, and all 443 Metis tests pass.

      TWO MORE CONTAINMENT FIXES. The tutor prompt previously received exact progress percentage, completed/total lesson counts, time spent, latest score, and per-attempt score/pass strings. It now receives versioned progress_band and performance_band projections plus purpose-relevant aggregate/topic signals; regression assertions prove raw metrics are not rendered. The assessment_submitted application log no longer records learner ID, score, pass state, or attempt number.

      .a — A MODEL PARSER IS NOT THE INVENTORY THE CELL NAMES. The checked metis-data-model.json reads only models/*.py. It covers entities, fields, tables/stores, and field indexes: 4 of the required 9 surfaces. It does not inventory events, APIs, exports, analytics, or backups. Its module-wide categories also hide important contradictions: Submission is labeled assessment-evidence while one table co-locates learner answers, grades, and an academic-integrity verdict; grades and integrity therefore have no independent category. XapiStatementRecord is labeled institutional-config even though it stores learner identity/activity and result JSON. The old M0.4.a check is revoked rather than grandfathered.

      .b — 0 documents assign all 9 required owner/classification dimensions across all 6 categories. Metis deployment notes mention fragments such as classification, retention, and residency; they do not assign canonical owner/store, purpose and basis, access roles, encryption, deletion, and residency per category.

      .c — The new tutor projection is genuinely versioned and minimal, but its destination is a learner tutor prompt. There are 0 versioned contracts from learner/grade/integrity/institutional owners into authoring previews or simulations.

      .d — Making 8 ORM edges explicit prevents accidental graph traversal; it is not a record/property prohibition. There are 0 schema/property suites rejecting raw protected payloads placed into authoring records, so this remains open.

      .e — The current estate has 20 identifier-like cross-record fields, but 0 cross-owner event contracts and 0 legal-hold, deletion-propagation, or correction-propagation declarations. Foreign keys and cascading deletes are shared database authority, not a boundary protocol.

      .f — Textual test candidates touch 5 of 8 named channels, but logs, fixtures, and backups have none and there is no unified leakage suite. This pass closes the observed application-log and tutor-prompt leaks; it does not claim that two repairs threat-test the estate.

      TRAPS THIS CELL PAID FOR. AN ENTITY'S MODULE NAME IS NOT ITS DATA OWNER: a learner xAPI record does not become institutional configuration because it lives in institutional_delivery.py. A FIELD ENUMERATION IS NOT A SURFACE INVENTORY: explicitly naming exports/analytics/backups as later "governance" work contradicted M0.4.a itself. AND DTO OMISSION IS NOT SEPARATION when the ORM eagerly fetched the protected graph first. -->

  • M0.5 Decide Metis persistence ownership and database strategy. Evidence: ADR resolves shared oshun_dev versus a dedicated database. <!-- 2026-08-27 audit and decision pass: 4/5 child properties are now evidenced by metis-persistence-inventory.json, metis-persistence-strategy.json, and docs/adr/ADR-M0.5-metis-persistence-ownership.md. The selected target is a bounded hybrid: a dedicated Metis database for Metis-owned aggregates, with externally owned identity and source records crossing the boundary only by canonical ID, event, or rebuildable projection. The ADR is deliberately Proposed with 0/6 approvals, so ratification and this parent cell remain open. -->
    • M0.5.a Inventory current schemas/tables/migrations/connections/ repositories/data volumes/tenancy/backups and production consumers. Evidence: persistence inventory. <!-- evidence: metis-persistence-inventory.json v2 explicitly covers all 9 named surfaces: one 2-migration Alembic chain, 13 tables in implicit public on shared oshun_dev, 4 source engine consumers, 4 Compose consumers, no repository layer, the shared PostgreSQL volume plus Redis/ object/local adjacent stores, 0/13 tenant-keyed tables, global backup/ restore posture, and 5 production source/deployment consumers. Production row and byte volumes are explicitly unknown rather than reported as zero. generate-metis-persistence-inventory.mjs + 9 tests. -->
    • M0.5.b Define aggregate ownership and transaction/consistency/ performance/retention/residency/security requirements for Metis authoring and protected adjacent data. Evidence: requirements matrix. <!-- evidence: ADR-M0.5 requirements matrix has 10 complete aggregate rows across canonical owner/store, classification, purpose, policy basis, access, transaction, consistency, performance, retention/deletion, residency, and security; parsed into metis-persistence-strategy.json. -->
    • M0.5.c Compare shared database, dedicated database, and bounded hybrid options for isolation, operational cost, migrations, joins/events, backup/restore, scaling, and ownership. Evidence: option analysis. <!-- evidence: ADR-M0.5 compares exactly the three requested options across all 7 requested dimensions and selects bounded-hybrid; the parsed option table has one selected row. -->
    • M0.5.d Decide canonical IDs and cross-database references/events, avoiding distributed transactions and shadow copies where possible. Evidence: integration design. <!-- evidence: ADR-M0.5 records 8 integration decisions: canonical IDs, cross-database references, event envelope, transactional outbox/inbox, projections, correction/deletion/hold propagation, shadow-copy prohibition, and the mixed Submission split. -->
    • M0.5.e Ratify an ADR selecting the owner/database strategy, migration/ rollback path, operational owner, and rejected alternatives. Evidence: approved ADR. <!-- open honestly: the Proposed ADR includes a 6-phase migration/ rollback plan, 6 operational ownership rows, and rejected alternatives, but all 6 named human reviewer approvals remain pending. Proposed is not Accepted. --> 2026-09-18: the packet exists (the Proposed ADR, its migration plan and its rejected alternatives); what is left is the six named approvals, which no agent may supply (rule 9). blocked:human
  • M0.6 Map all libs/metis/* packages and canonical contracts to workbench capabilities. Evidence: complete package ledger. <!-- evidence: metis-package-inventory.json v2 records all 28 packages and 6,651 reviewed public capabilities with complete four-audience mappings, owner/API/ledger links, dispositions, and bidirectional drift controls. package-capability-decisions.json is the explicit, hash-ratcheted decision source. -->
    • M0.6.a Enumerate package manifests/exports/source/tests/dependencies/ consumers/build targets/owners and detect nested or untracked packages. Evidence: deterministic package inventory. <!-- evidence: all 28 direct libs/metis/* manifests with public exports, source, 341 tests, declared/source dependencies, consumers, build targets, and owners; 0 nested, untracked, or duplicate manifests. -->
    • M0.6.b Extract public contracts/entities/operations/events/jobs/gates/ adapters from each package and link their exact source locations. Evidence: package capability catalog. <!-- evidence: all 6,651 reviewed public symbols classified across the seven requested kinds and linked to exact declaration/re-export file and line; 0 fallback or unanchored entries. -->
    • M0.6.c Map every package capability to workbench/learner/operator/ institutional feature, canonical owner contract, route/API, and ledger parent task. Evidence: package ledger. <!-- evidence: every capability inherits an explicit workbench, learner, operator, and institutional feature profile plus canonical owner contract, resolvable API surface, and resolvable parent task. -->
    • M0.6.d Assign reuse, refactor, move, merge, deprecate, retire, or net-new gap disposition with owner/evidence for duplicates and orphan packages. Evidence: disposition review. <!-- evidence: 9 reuse, 9 refactor, 5 move, 4 merge, and 1 retire decisions cover all packages, including 12 no-consumer candidates and 8 reviewed semantic duplicate groups; 0 unresolved orphans. -->
    • M0.6.e Add bidirectional drift checks so new exports/packages or ledger capabilities cannot remain unmapped. Evidence: zero-orphan audit. <!-- evidence: 28/28 package decisions and all 6,651 public surfaces are count/hash pinned; missing, extra, changed, unanchored, broken- reference, and unmapped audits are all zero. -->
  • M0.7 Map existing course creation, BYOM, ingest, assessment, tutor, gradebook, and integrity routes. Evidence: reuse/retire/gap dispositions. <!-- evidence: metis-api-inventory.json v2 maps all 254 real service, BFF, and frontend routes through 28 reviewed, hash-pinned flow profiles. The ledger explicitly preserves all observed control gaps; completing M0.7 records their dispositions, not their remediation. -->
    • M0.7.a Inventory frontend/BFF/service routes, methods, schemas, handlers, auth/roles, data owners, consumers, tests, and live/fixture status for each capability family. Evidence: route inventory. <!-- evidence: Python AST extraction finds 183 canonical Metis service operations plus 2 separate Manim-service routes, excluding 2 old docstring false positives; 30 BFF registrations and 39 frontend pages complete the 254-route inventory with exact lines, tests, consumers, schemas/roles, and honest live/fixture status. -->
    • M0.7.b Trace each route through owner service/store/provider and classify authoring, learner delivery, educator operations, institutional, or adjudication authority. Evidence: route flow map. <!-- evidence: 254/254 rows map canonical contract, owner service/store, providers, and one of the six authority classes (including the separately required infrastructure class). -->
    • M0.7.c Compare routes against canonical contracts and requirements for tenancy, property auth, idempotency, audit, protected data, and honest failure states. Evidence: gap analysis. <!-- evidence: all 254 rows compare 8 controls. Gaps remain explicit: authentication 35, tenancy 184, property authorization 84, idempotency 97, audit 92, protected projection 11, canonical contract 28, and honest failure 62. -->
    • M0.7.d Assign per-route reuse, refactor, facade, redirect, retire, replace, or net-new disposition with compatibility/consumer migration. Evidence: route disposition ledger. <!-- evidence: every route has a target/rationale, compatibility plan, and consumer migration: 233 refactor, 19 facade, 2 reuse. Two reviewed BFF helper modules with zero HTTP registrations retire. -->
    • M0.7.e Add route/consumer/ledger traceability checks and negative boundary tests for answer keys, grades, integrity records, and BYOM providers. Evidence: route coverage gate. <!-- evidence: 28/28 source/surface profiles are count/hash pinned; all 20 frontend API calls resolve by method/path exactly once; missing, extra, duplicate, drift, reference, unresolved/ambiguous-consumer, and sensitive-boundary violations are zero. Negative tests cover 6 answer-key, 6 grade, 11 integrity, and 3 BYOM-sensitive routes. -->
  • M0.8 Define content hierarchy: source collection, course, module, lesson, activity, assessment, section, item, stimulus, rubric, objective, standard, and prerequisite. Evidence: hierarchy contract. <!-- evidence: @oshun/contracts/metis content-hierarchy.ts defines all 13 kinds in one strict versioned contract; generated evidence reports all 6 M0.8 controls present with zero drift. -->
    • M0.8.a Define stable identity/version/tenant/owner/lifecycle/common metadata for every hierarchy entity. Evidence: base entity schema. <!-- evidence: the full golden carries all 13 kinds (16 version rows), and every row parses through the shared 8-field base envelope plus 9-field common metadata contract. -->
    • M0.8.b Define allowed containment/cardinality/order/reuse relations from source collections through courses/modules/lessons/activities/ assessments/sections/items. Evidence: containment matrix. <!-- evidence: 10 allowed ordered parent/child rules state min/max and reuse; validators enforce allowed pairs, required cardinality, dense zero-based order, unique edges, reference-only multi-parent reuse, exact revisions, tenant binding, and acyclic containment. -->
    • M0.8.c Define stimulus-to-item and rubric-to-item/assessment reuse with exact revision pinning and no implicit answer exposure. Evidence: reuse invariants. <!-- evidence: exact content-hash/revision pins cover stimulus→item and rubric→item/assessment; item/assessment answer bytes are replaced by metis-assessment-security never-inline refs. Strict schemas and negative goldens reject inline answers and drifted hashes. -->
    • M0.8.d Define objective/standard alignment and prerequisite graph edges, scope, weight/coverage, versions, and cycle/orphan semantics. Evidence: graph contracts. <!-- evidence: alignment edges pin objective, standard, and scope with weight/coverage ≤1; versioned prerequisite entities pin from/to/ scope with required/recommended strength. Orphans, self-edges, cycles, and over-coverage fail. -->
    • M0.8.e Define clone/reference/embed/fork/supersede/archive/delete/ restore behavior and downstream invalidation per relationship. Evidence: lifecycle invariants. <!-- evidence: an 8-row policy fixes target identity/revision, upstream tracking, lifecycle state, and exact downstream invalidation for every named operation; invalidation drift fails validation. -->
    • M0.8.f Publish schemas/goldens for minimal/full/reused/version-diverged/ invalid hierarchies and reconcile existing shapes. Evidence: hierarchy contract suite. <!-- evidence: 4 valid goldens and 6 intended-invalid goldens pass their expected acceptance/rejection checks. A 13-row reconciliation maps every kind to current TS/Python shapes or an explicit net-new gap; all source/symbol anchors resolve. Package tests (10) + generator tests (7) and a root inventory --check gate keep the contract current. -->
  • M0.9 Define the authoring maturity model independently from learner progress and grade states. Evidence: vocabulary collision review. <!-- evidence: @oshun/contracts/metis authoring-maturity.ts plus the generated collision review inventory 27 current vocabularies in all 9 requested categories, with 30 normalized collisions and zero drift. -->
    • M0.9.a Inventory current authoring, review, publication, learner progress, attempt, mastery, completion, grade, and integrity status vocabularies with owners. Evidence: vocabulary inventory. <!-- evidence: every row names its declaring owner, canonical owner, authority/disposition, source, symbol, labels, and rationale; all 27 source/symbol/label anchors resolve. -->
    • M0.9.b Define authoring maturity states/transitions for draft, in review, changes requested, approved, scheduled, published, superseded, archived, quarantined, and retired as applicable. Evidence: state machine. <!-- evidence: the independent 10-state machine has 14 named policies, reaches every state, starts at draft, and terminates at retired. -->
    • M0.9.c Define transition actors, prerequisites, guards, concurrency, version effects, audit, rollback/supersession, and invalidation. Evidence: transition matrix. <!-- evidence: every transition fixes allowed actors, exact prerequisite evidence, tenant/version/idempotency/separation guards, content- revision effect, audit binding, rollback effect, and exact downstream invalidations. All 14 conforming transitions parse; stale version, missing evidence, forged creator, self-approval, and invalidation drift fail. -->
    • M0.9.d Keep learner progress/attempt/mastery/completion and grade/ integrity state in their canonical owner vocabularies and projections. Evidence: separation contracts. <!-- evidence: 6 protected owners keep revisioned projections only: learning progress, assessment attempts, mastery, completion, gradebook, and Themis integrity. Each has a distinct analytics namespace; cross-domain mutation tuples are structurally empty. -->
    • M0.9.e Detect identical/ambiguous labels, enum/code reuse, API/UI mapping, analytics conflation, and migration risks. Evidence: collision review. <!-- evidence: the artifact reviews all 30 case/separator-normalized collisions, publishes 27 complete legacy API/UI mappings, requires owner+vocabulary analytics dimensions, and records 7 concrete migration risks/mitigations. -->
    • M0.9.f Test no authoring transition can directly mutate learner/grade/ integrity state and vice versa. Evidence: boundary tests. <!-- evidence: strict authoring and protected-owner event schemas reject foreign mutation fields in both directions; package tests (9) and deterministic generator tests (8) cover the boundary and zero- drift evidence. -->
  • M0.10 Threat-model prompt injection, poisoned source material, answer-key leakage, test theft, bias, inaccessible items, model cheating, PII, minors, BYOM abuse, and grade tampering. Evidence: mitigation/test map. <!-- evidence: @oshun/contracts/metis threat-model.ts plus generated metis-threat-model.json define 60 owned threat cases in 7 categories, 23 anchored controls, 7 incident playbooks, and 7 launch gates with zero vector/reference/anchor drift. Six gates remain honestly blocked and the academic-integrity gate conditional; the model does not turn partial controls into a false launch green. -->
    • M0.10.a Map assets, actors, trust boundaries, entry points, providers/ models/tools, data stores, exports, integrations, and high-impact actions. Evidence: threat model context. <!-- evidence: 78 owned context items cover all 9 named dimensions: 9 assets, 12 actors, 10 trust boundaries, 11 entry points, 7 providers/models/tools, 8 stores, 6 exports, 6 integrations, and 9 high-impact actions. -->
    • M0.10.b Model injection/poisoning through uploads, URLs, LMS/repos, extracted text, citations, prompts, BYOM responses, tutor context, and generated media/code. Evidence: injection threat cases. <!-- evidence: 11 explicit cases cover upload, URL, LMS, repository, extracted text, citation, prompt, BYOM response, tutor context, generated media, and generated code paths. -->
    • M0.10.c Model answer/rationale/rubric/test-bank exposure and assessment theft through UI/API/search/export/cache/logs/analytics/tutor/accessibility projections. Evidence: content security cases. <!-- evidence: 9 cases each bind all 4 protected material classes to one named exposure surface; protected references exist, while the gate blocks until cross-surface no-secret tests cover all 9 surfaces. -->
    • M0.10.d Model bias/unfair/inaccessible items, psychometric misuse, disability inference, translation drift, cultural harm, and automation overclaim. Evidence: fairness/accessibility cases. <!-- evidence: 8 cases cover bias, unfair and inaccessible items, psychometric misuse, disability inference, translation drift, cultural harm, and automation overclaim with explicit limitations. -->
    • M0.10.e Model learner model-assisted cheating, impersonation, collusion, integrity evidence manipulation, false accusation, and appeal denial. Evidence: integrity cases. <!-- evidence: 6 cases bind learner-safe evidence, probabilistic verdict limitations, independent human review, non-conflicted appeal, and grade correction recovery; the launch gate is conditional. -->
    • M0.10.f Model PII/minor/guardian/educational record/biometric/voice data exposure, overretention, cross-purpose use, and cross-tenant access. Evidence: privacy cases. <!-- evidence: 9 critical cases cover direct PII, minor, guardian, educational, biometric, voice, retention, purpose, and tenant boundaries; M0.11 lifecycle/policy ratification remains a launch blocker rather than being silently assumed. -->
    • M0.10.g Model BYOM credential/model/provider/endpoint/SSRF/spend/ availability/supply-chain abuse and unsafe outputs. Evidence: BYOM cases. <!-- evidence: 8 cases cover credentials, model, provider, SSRF, spend, availability, supply chain, and unsafe output. Existing sandbox, interlock, audit, quarantine, and kill controls are mapped; private- network SSRF, secret storage, and hard spend ceilings block launch. -->
    • M0.10.h Model grade/credential/approval/publication tampering, replay, mass assignment, concurrency, confused deputy, and audit deletion. Evidence: authority cases. <!-- evidence: 9 cases cover every named authority vector and map guarded authoring transitions, tenant/idempotent grade corrections, signed credentials, and Merkle audit anchors; unresolved M0.3 authority and write-path atomicity remain explicit launch blockers. -->
    • M0.10.i Assign prevention/detection/response/recovery control, owner, residual risk, verification, incident playbook, and launch gate to every threat. Evidence: mitigation/test traceability matrix. <!-- evidence: every one of 60 cases has nonempty prevention, detection, response, and recovery assignments; owned controls with exact source and test anchors; residual risk/review trigger; verification IDs; a category-matched 7-step incident playbook; and a derived launch gate. Package tests (14) + generator tests (11) reject dangling, phase-incompatible, missing-vector, duplicate, and false-green data. -->
  • M0.11 Define regional education/privacy obligations and institutional policy hooks. Evidence: legal/privacy review with data-retention rules. <!-- evidence: metis-regional-policy.json is the deterministic technical review packet. M0.11.a/c/d/e are complete, but the parent remains open: all 8 applicability cells and the approval record fail closed pending the real named counsel/privacy review required by M0.11.b/f. -->
    • M0.11.a Inventory intended regions/jurisdictions, learner age groups, institution types, data subjects, delivery modes, vendors, and data transfers. Evidence: applicability inventory. <!-- evidence: 38 owned scope items across all 7 named dimensions, with candidate/pending status, explicit assumptions, unknown-age and unknown-jurisdiction fail-closed entries, and 8 launch cells. -->
    • M0.11.b Record counsel/privacy-reviewed education, child/minor, accessibility, assessment, records, consent, AI, biometric/voice, residency, and deletion obligations per applicability cell. Evidence: obligation matrix. <!-- blocker: every cell carries all 11 review categories and open review questions, but none is mislabeled reviewed; no named counsel/privacy reviewer or reviewed contract version has been supplied. --> 2026-09-18: the matrix carries every category and its open questions; what is left is a named counsel or privacy reviewer reading it (rule 9). blocked:human
    • M0.11.c Define code-owned nonwaivable floors and UI-governed institution policy hooks for retention, access, approvals, proctoring, integrity, providers, exports, locales, and communications. Evidence: policy schema. <!-- evidence: 16 strict nonwaivable rules cover all 9 domains, each with a code owner, UI hook, constraint direction, and platform value. -->
    • M0.11.d Define precedence/merge semantics so institution/region policies can tighten but never loosen applicable platform/legal floors. Evidence: monotonicity tests. <!-- evidence: platform→region→institution merge rejects maximum, minimum, allowed-set, required-set, invariant, duplicate-set, unknown-rule, missing-floor, and reversed-precedence violations; 14 package tests plus 10 generator tests pin the behavior. -->
    • M0.11.e Define retention/deletion/legal hold/correction/export/guardian/ consent workflows and downstream processor propagation. Evidence: lifecycle rules. <!-- evidence: 7 lifecycle state workflows define triggers, prerequisites, conflicts, audit events, deadlines, receipts, and 31 propagation actions across 8 processor classes; destructive deletion is refused under legal hold and rights workflows require authority. -->
    • M0.11.f Obtain named legal/privacy review with version/effective date, assumptions, open issues, and reevaluation triggers. Evidence: approval record. <!-- blocker: the versioned packet records 3 assumptions, 4 open issues, and 5 reevaluation triggers, but intentionally has no fabricated reviewer, review date, effective date, or approval decision. --> 2026-09-18: the versioned packet exists; what is left is the named reviewer, the dates and the decision (rule 9). blocked:human
  • M0.12 Define success measures for authoring quality, accessibility, psychometrics, time-to-publish, interoperability, and learner outcomes. Evidence: metric definitions avoid incentivizing harmful shortcuts. <!-- evidence: metis-success-metrics.json is the deterministic proposed catalog. M0.12.a-f are technically complete with 32 metrics across 6 families, 12 exact source/test anchors, 4 outcome hypotheses, and fail-closed privacy/staleness/guardrail evaluation. The parent remains open because M0.12.g requires real accountable-owner ratification. -->
    • M0.12.a Define authoring quality measures tied to source grounding, objective coverage, correctness, review revisions, gate outcomes, and post-publication corrections. Evidence: quality metric specs. <!-- evidence: 6 authoring-quality metrics separately measure grounding, objective coverage, correctness, review revision, release-gate passage, and post-publication correction, each with countermetrics. -->
    • M0.12.b Define accessibility coverage/defect/learner-use measures by item kind, content type, accommodation, locale, severity, and resolution. Evidence: accessibility metric specs. <!-- evidence: 5 accessibility metrics all require all 6 named slicing dimensions; coverage, open/critical defects, resolution time, and consented learner-use success are distinguished. -->
    • M0.12.c Define psychometric measures with sample/uncertainty/privacy, difficulty/discrimination/reliability/DIF/exposure and no small-group misuse. Evidence: psychometric metric specs. <!-- evidence: 5 psychometric metrics pin difficulty, discrimination, reliability, DIF, and exposure with required uncertainty methods, aggregate N >= 100, subgroup N >= 30, privacy suppression, and a prohibition on individual high-stakes automation. -->
    • M0.12.d Define time-to-publish stage durations/rework/wait/abandonment with quality/safety/accessibility guardrails so speed cannot reward bypass. Evidence: flow metric specs. <!-- evidence: 5 flow metrics cover stage cycle, wait, rework, abandonment, and accepted throughput. Every result is withheld on correctness, release-gate, or critical-accessibility failure and every metric rejects standalone incentive use. -->
    • M0.12.e Define standards/package/LMS/import/export compatibility and semantic round-trip measures per version/profile. Evidence: interoperability metric specs. <!-- evidence: 6 interoperability metrics cover profile conformance, package validation, import/export fidelity, semantic round-trip, and LMS exchange, sliced by pinned standard/profile/vendor version. -->
    • M0.12.f Define learner-outcome hypotheses and measures with baseline, confounders, consent/privacy, subgroup uncertainty, causal limitations, and no individual high-stakes automation. Evidence: outcome metric review. <!-- evidence: 5 aggregate outcome metrics and 4 explicit hypotheses pin baselines, at least 5 confounders, consent, privacy thresholds, uncertainty, no causal claim, and no individual high-stakes use. -->
    • M0.12.g Define owners, sources, formulas/denominators, freshness, privacy thresholds, targets, countermetrics, and review cadence. Evidence: ratified metric catalog. <!-- blocker: all 32 proposed specs contain the required catalog fields, but all targets remain proposed and the versioned approval record intentionally has no fabricated product, learning-science, psychometric, accessibility, privacy, interoperability, or data- governance reviewer/effective date. --> 2026-09-18: all 32 specs carry the catalog fields; what is left is ratification of the proposed targets by named reviewers (rule 9). blocked:human
  • M0.13 Author and ratify V1/METIS_COURSEWARE_WORKBENCH_DESIGN_2026-07-23.md with current-state, host, roles, pedagogy, data, API, standards, UI, trust, migration, rollout, and decision records. Evidence: approved doc and resolved review comments. <!-- evidence: the generated design now exists with 11 explicit current- state rows, 6 governed journeys, 7 unresolved ADRs, and 33 hashed source links. M0.13.a-e are technically complete; parent/f remain open because no named cross-functional reviewers approved an effective revision or resolved a review-comment log. -->
    • M0.13.a Establish document scope, requirements traceability, owners, reviewers, source links, terminology, decision log, and evidence format. Evidence: design-doc outline. <!-- evidence: §1 defines scope, accountable team functions, terminology, evidence/comment formats, a-f traceability, ADR format, and a content-addressed source ledger without inventing named reviewers. -->
    • M0.13.b Document current apps/routes/packages/contracts/stores/flows/ tests/deployments and explicitly separate live, partial, fixture, and proposed behavior. Evidence: current-state review. <!-- evidence: §2 classifies 11 rows across all 4 required truth states, including 39 routes/4 surfaces, 28 packages/6,651 capabilities, 183 API operations, 13 tables, 547 app/library test files, proposed hierarchy/maturity/policy/metrics, and explicit deployment gaps. -->
    • M0.13.c Document canonical hosts/audiences/roles/boundaries, pedagogy/ content model, data ownership/contracts/APIs/jobs/events, and standards profiles. Evidence: architecture sections. <!-- evidence: §3 records all 6 audiences and 9 required roles, denies a false canonical-host claim, defines pedagogy/content/data/API/job/ event boundaries, and separates implemented standard operations from still-pending conformance profiles. -->
    • M0.13.d Document UI workbench journeys/responsive/accessibility/offline/ localization, trust/safety/privacy/integrity/rights, and evaluation/ observability/operations. Evidence: experience/trust sections. <!-- evidence: §4 follows the restrained operational-app thesis and defines 6 end-to-end journeys plus desktop/tablet/mobile, keyboard, screen-reader, reduced-motion, offline, localization, trust, metric, observability, incident, and operations behavior. -->
    • M0.13.e Document migration/compatibility/rollout/flags/deprecation/ rollback/readiness and all unresolved options as explicit ADRs. Evidence: migration/decision sections. <!-- evidence: §5 defines a 7-phase reversible rollout, compatibility and default-off flag rules, deprecation evidence, and 7 owned ADRs for host, roles, persistence, data, standards, experience, and rollout. -->
    • M0.13.f Review against feature inventory/deep dives/code/contracts and resolve or disposition every comment with named domain/security/privacy/ accessibility/engineering/product ratification. Evidence: approved doc. <!-- blocker: §6 pins all source versions and six required review roles, but approvals=0, effective date=none, and no review comments may be fabricated by the generator. --> 2026-09-18: the document pins its sources and names six review roles; what is left is the six approvals (rule 9). blocked:human
  • M0.14 Reconcile V1/features/domain-metis.md, the V1 feature inventory, deep-dive specifications, standalone apps, service OpenAPI, and every libs/metis/* capability at atomic-feature level. Evidence: explicit dispositions include course authoring, curriculum discovery, assessments, tutoring/live voice, mastery/adaptation, knowledge graphs, credentials, prerecorded lectures, agentic media, correctness verification, gradebook, integrity, institutional interoperability, administration, and training-data flywheel. <!-- evidence: metis-feature-reconciliation.json + its hashed rows.jsonl map 10,021 source-native feature rows bidirectionally to 30 canonical capabilities with 0 source or canonical orphans. M0.14.a-f are technically complete; M0.14.g and the parent remain open because the zero-orphan report has no real owner approval or effective date. -->
    • M0.14.a Build deterministic atomic-feature inventories from each named document, app/route, OpenAPI operation/schema, package export, job/event, and test with source anchors. Evidence: source inventories. <!-- metis-code-census.json: 11 artifact classes with source anchors and exact file lists where the upstream evidence does not already own atomic rows: 51 documents; 39 app routes; 183 OpenAPI operations, 174 with response schemas; 558 schemas; 6,651 exports across 28 packages; 14 entities / 141 fields over 13 tables / 2 migrations; 296 atomic job/event entries across 63 source files (286 public capabilities + 10 runtime modules); 547 app/library tests; 906 implementation files (180 py + 717 ts + 9 js); 5 unique apps + 28 libs. Generator + 8 known-value tests are wired into inventory/verify. -->
    • M0.14.b Normalize aliases only through reviewed mappings while retaining source-specific identity and semantics. Evidence: alias map. <!-- evidence: feature-alias-decisions.json defines 30 unique, repository- reviewed canonical mappings grounded in six source documents. The generator retains original source/ref/line/text and permits only a reviewed alias, reviewed package profile/API family, or explicit platform fallback; M0.14.g owner approval is not implied. -->
    • M0.14.c Map course/curriculum/assessment/item/rubric/source and standards authoring features bidirectionally to canonical capabilities. Evidence: core authoring rows. <!-- evidence: 1,323 source rows map bidirectionally to 8 core-authoring capabilities with 1,042 implementation rows and explicit authoring/ delivery authority boundaries; every named capability is nonempty. -->
    • M0.14.d Map tutoring/live voice, learner modeling/mastery/adaptation, knowledge graphs, simulations, correctness verification, and agentic media features with authoring-versus-delivery authority. Evidence: intelligent learning rows. <!-- evidence: 3,418 source rows map bidirectionally to 8 intelligent- learning capabilities with 2,967 implementation rows; every row resolves to distinct authoring and learner/runtime authorities. -->
    • M0.14.e Map credentials, prerecorded lectures, gradebook, integrity, institutional interoperability/administration, analytics, and training- data flywheel features with data/authority boundaries. Evidence: institutional rows. <!-- evidence: 886 source rows map bidirectionally to all 8 named institutional capabilities with 709 implementation rows and explicit grade, integrity, credential, institution, analytics, and governed- training authority boundaries. -->
    • M0.14.f Assign reuse/refactor/move/merge/retire/net-new/defer/reject disposition, canonical owner, parent/child task, dependencies, and proof to every source feature. Evidence: disposition ledger. <!-- evidence: all 10,021 source rows carry an allowed disposition, canonical owner, M0.14 child, downstream parent tasks, dependency list, and exact proof anchor; 111 broad/future/unsupported document assertions are explicitly deferred rather than counted as code. -->
    • M0.14.g Review collisions/orphans/unsupported assertions and add a deterministic bidirectional zero-orphan drift gate. Evidence: approved reconciliation report. <!-- blocker: the deterministic gate resolves 1,466 collisions with a fixed reviewed rule, reports 111 unsupported/broad assertions, and proves 0 source/canonical orphans in both directions. It remains pending real owner approval and an effective date; the generator cannot approve its own report. --> 2026-09-18: the deterministic gate exists and proves zero orphans both ways; what is left is the owner's approval of its report and an effective date (rule 9). blocked:human
  • M0.15 Define the boundary among authoring, institutional setup, educator operations, learner delivery, guardian/observer views, and operator adjudication for every Metis feature. Evidence: no learner progress, grade, answer key, integrity evidence, or credential authority is exposed through an authoring convenience route. <!-- evidence: metis-surface-boundaries.json + its hashed ownership JSONL define all 6 logical surfaces and assign all 10,021 reconciled feature rows plus all 39 current routes to one mutation owner. M0.15.a-c/e are technically complete. The parent/d/f remain open: 3 open sensitive API operations expose answer-key or grade classes, learner submission shapes can carry detailed integrity evidence, and 0/324 required served-runtime adversarial cells cover all 36 ordered surface pairs. -->
    • M0.15.a Define each surface’s actors, purpose, canonical host/routes, owner services/stores, permitted subject projections, and high-impact actions. Evidence: surface boundary matrix. <!-- evidence: 6 strict surface rows define actors, purpose, logical canonical host/route namespace, current routes, owner services and stores, permitted projections, and high-impact actions. Logical host effectiveness remains explicitly pending M0.1 ratification. -->
    • M0.15.b Assign every reconciled atomic feature and route to exactly one mutation owner and any read-only facades/deep links. Evidence: ownership ledger. <!-- evidence: metis-surface-ownership.rows.jsonl contains 10,060 unique rows: 10,021 feature assignments + 39 route assignments, each with exactly one mutation owner/service/store, command surfaces, read-only facades, canonical command route, deep link, and proof; 0 current routes are left as boundary violations. -->
    • M0.15.c Define minimal authoring projections of learner progress/profile, grades, integrity, accommodation, and credentials for preview/planning with purpose/field/aggregation limits. Evidence: projection schemas. <!-- evidence: @oshun/contracts/metis exports 6 strict discriminated projection schemas with preview/planning purpose, exact coarse-band fields, source owner/policy/revision context, N >= 30, small-cell suppression, coarse precision, and bounded freshness. Property tests inject every forbidden field family and fail closed. -->
    • M0.15.d Keep answer keys/rationales/items, detailed integrity evidence, grade corrections, learner PII, and credential issuance/revoke authority within their approved surfaces. Evidence: property/function tests. <!-- blocker (re-measured 2026-09-11): the 8-class policy and property tests deny protected data on every authoring convenience route, but runtime wiring is not complete. GET /api/assessments and GET /api/assessments/{id} (api/assessments.py) now require a bearer token (since 2026-09-06) but only get_current_user, whose default role is student, and AssessmentResponse still nests QuestionResponse.correct_answer and explanation — a learner token reads every answer key. The learner app also serves the whole authoring tree apps/metis/web/src/app/create/* behind login-only middleware (no role gate). The open LTI score route carrying grades and the learner submission shapes carrying integrity evidence were not re-measured. Closing order: project the learner-facing assessment read through M9.15's projection, role-gate or relocate /create/*, then the LTI score route and submission shapes; M10.16.a must be red on all of these before this cell closes. --> <!-- one of the four measured 2026-09-12 (M10.6.e): the LTI score route is closed. POST /lti/assignments/scores is unreachable without a bearer token, refuses a token that bought any other scope, and the grades it writes read back only with the AGS result scope — measured over the served app in tests/test_lti_services.py:: test_the_score_route_is_reachable_only_with_a_bearer_token_that_bought_the_scope. The other three vectors (learner-readable answer keys, /create/*, submission shapes) are untouched and this cell stays open. -->
    • M0.15.e Define signed handoffs with current authorization, tenant/ institution/course/subject/revision context, expiry, replay, return intent, and immutable audit. Evidence: handoff contracts. <!-- evidence: MetisSignedSurfaceHandoffSchema signs every claim with Ed25519, caps TTL at 300s, rechecks current principal/session/authz revision and full context, consumes tenant+subject+session+nonce once, constrains target/return namespaces, and fails closed unless an append-only externally anchored audit record is appended. -->
    • M0.15.f Test direct routes, crafted fields, search/export/logs/ notifications, deep links, role changes, and two-tenant leakage across all surface pairs. Evidence: adversarial boundary suite. <!-- blocker: contract tests cover crafted fields, policy, tamper, current authorization, context, expiry, replay, return namespace, and audit failure. They do not prove direct routes plus all 8 secondary vectors against the served runtime: 36 ordered surface pairs x 9 vectors = 324 required cells, 0 complete. --> <!-- SPEC 2026-09-11: generate the 36 ordered surface pairs × 9 vectors as a parametrized pytest from libs/contracts/src/metis/surface-boundaries.ts (export the pair list as JSON through the existing generator) against the served runtime; never hand-write 324 tests. Each row records pass, fail or not-drivable with reason; the parent needs 324 rows none of which is not-drivable. -->
  • M0.16 Map discipline-specific capabilities for mathematics, physical science, life science, social science, language/humanities, and computing—or the current canonical discipline registry—to common and specialized editors, verifiers, media, simulations, and accessibility needs. Evidence: no discipline inherits unsupported generic behavior. <!-- evidence: metis-discipline-capabilities.json + its hashed 540-row JSONL bind 36 capability requirements to all 15 taxonomy 1.0.0 entries (6 headline core disciplines + 9 supporting subjects). Exactly 144 cells admit proven semantics; all 396 other specialist cells deny generic fallback. Missing institution mappings remain an explicit zero-count, deny-unmapped state rather than an invented map. -->
    • M0.16.a Reconcile named disciplines to the current canonical registry, hierarchy/aliases/versions, institution mappings, and extensibility. Evidence: discipline registry map. <!-- evidence: a stdlib-AST extractor compares every ID, title, placement, alias, taxonomy ID/version, and core/supporting order byte-for-byte with learning_subject_taxonomy_helpers.py. Extension policy requires versioning/non-shadowing/capability coverage; 0 observed institution mappings fail closed as deny-unmapped. -->
    • M0.16.b Define common content/objective/source/item/rubric/review/export/ accessibility capabilities and declare their real cross-discipline support. Evidence: common capability rows. <!-- evidence: 8 shared typed capabilities x 15 exact registry entries = 120 supported common rows, each with implementation + automated test proof and an explicit warning that shared structure does not imply specialist semantics. -->
    • M0.16.c Map mathematics requirements for semantic math, proofs/worked steps, symbolic/numeric verification, graphs, manipulatives, and accessible notation. Evidence: mathematics rows. <!-- evidence: 6 requirements distinguish the bounded algebra/process verifier and parameterized function plots from deferred semantic notation, manipulatives, and accessible math. Unsupported grammar fails loud; no CAS/MathML claim is made. -->
    • M0.16.d Map physical/life science requirements for equations/units, datasets, labs/field evidence, simulations, diagrams, uncertainty, safety/ethics, and accessible alternatives. Evidence: science rows. <!-- evidence: 8 requirements map data figures/diagrams, exactly 3 physics simulation templates, and neuroscience/astronomy safety policies while deferring absent unit, lab/field, measurement- uncertainty, discipline-template, and equivalent-access semantics. -->
    • M0.16.e Map social science/language/humanities requirements for primary/ secondary sources, claims/perspectives, chronology/maps, text/audio/oral performance, interpretation, and cultural/linguistic review. Evidence: humanities rows. <!-- evidence: 6 requirements preserve the narrow anthropology emic/etic policy and four core reviewer pools; general primary/ secondary classification, chronology/maps, oral performance, interpretation, and supporting-subject review remain deferred. -->
    • M0.16.f Map computing requirements for code/editor/runtime/tests/ dependency/sandbox/security, interactive systems, and nonvisual output. Evidence: computing rows. <!-- evidence: 8 requirements admit only structural code validation and bounded JavaScript expected-case execution. The code explicitly disclaims node:vm as an untrusted security sandbox; editor, dependencies, isolated runtime, interactive systems, and nonvisual output remain denied/deferred. -->
    • M0.16.g Map specialized editor, verifier, media/Isis/Yemaya/Euterpe, simulation, item type, export loss, accessibility, and reviewer role for every supported discipline capability. Evidence: discipline matrix. <!-- evidence: every supported cell has all requested facet bindings, real proof anchors, explicit export loss, and reviewer roles; absent Isis/Yemaya/Euterpe or accessible-specialist bindings are recorded as missing/not-required, never implied by a package name. -->
    • M0.16.h Mark unsupported/not-applicable/deferred cells explicitly with reasons and fail generic fallback when specialized semantics are required. Evidence: no-inherited-behavior coverage test. <!-- evidence: all 540 registry/capability pairs have exactly one state: 144 supported, 77 deferred, 1 unsupported, 318 not-applicable. Schema, example, and fast-check coverage prove only exact supported cells admit execution; every other specialist cell denies and sets genericFallback=forbidden. -->

M1 — Contracts, standards, and content model#

  • M1.1 Verify all canonical Metis contracts and remove duplicate workbench shapes. Evidence: contract ownership matrix. <!-- current M1.1 evidence: metis-contract-ownership.json plus the 1,370-row field ledger prove M1.1.b-d. The parent remains open because 92 app-local declarations across 91 logical shapes still require migration and public parse/serialize/round-trip goldens cover only 80/558 schemas. -->
    • M1.1.a Inventory all Metis entities/schemas/types/OpenAPI/event/job/ database/workbench shapes and their definitions, versions, owners, consumers, serializers, and tests. Evidence: contract inventory. <!-- superseded M1.1.a v1 evidence: metis-contract-inventory.json: inventory of the CURRENT Metis contract shapes across representations with DUPLICATE DETECTION (M1.1's goal). From the committed OpenAPI spec (apps/metis/service/openapi/metis.openapi.json, read once) + shared TS contracts (libs/contracts/src/metis) + DB entities (metis-data-model evidence). 558 OpenAPI component schemas collapse to 518 logical shapes (429 Response / 68 Request / 57 base / 4 Input). DUPLICATE-SHAPE FINDING: 40 logical shapes have >1 OpenAPI variant (the Response/Request/base duplication M1.1 targets, e.g. AgentExecutionIntervention/ AgentRubricScore); 0 cross-representation duplicates (the shared TS contracts cover a different canonical slice than the API DTOs, so no OpenAPI base name matches a shared type). 7 shared-contract files (115 Zod schemas + 108 types), 14 SQLAlchemy entities. baseName() strips DTO suffixes for logical grouping. generate-metis-contract-inventory.mjs + 6 tests, wired into inventory/verify + CI. M1.1.b (group by aggregate) + M1.1.c+ (consolidation/duplicate removal) are design and stay unchecked. --> <!-- current M1.1.a v2 evidence: 558 OpenAPI schemas collapse to 518 logical shapes; 40 multi-variant groups are Request/Response pairs, not competing OpenAPI owners. The AST-backed cross-representation scan finds 92 app-local declarations across 91 logical shapes, zero shared-contract collisions, 25 shared files (364 Zod schemas and 351 types), and 14 SQLAlchemy entities. Seven tests cover determinism, normalization, import exclusion, and known inventory values. -->
    • M1.1.b Group shapes by semantic aggregate/operation and distinguish valid command/read/export projections from competing canonical definitions. Evidence: equivalence analysis. <!-- evidence: all 40 multi-variant OpenAPI groups map to non-empty request/response operation usage and classify as valid directional projections; zero competing OpenAPI groups and all 92 app-local declarations are explicitly classified as competing definitions. -->
    • M1.1.c Choose one canonical owner contract per entity/operation and map every duplicate field/enum/version/unit/optional/default difference. Evidence: ownership matrix. <!-- evidence: 131 logical-shape ownership rows name one Metis service OpenAPI/Pydantic owner; 1,370 field rows map presence, requiredness, type, enum, version, unit, optionality, default, and format with zero missing owners or incomplete comparisons. -->
    • M1.1.d Assign reuse/adapter/migrate/deprecate/remove disposition and consumer migration/compatibility window for every duplicate. Evidence: disposition ledger. <!-- evidence: 172 dispositions retain/reuse 80 directional wire projections and freeze 92 app-local declarations against new consumers pending generated-client migration before M1.1.e closure. -->
    • M1.1.e Add public parse/serialize/round-trip/schema goldens and drift/ forbidden-import checks preventing new workbench-local canonical shapes. Evidence: contract ownership gate. <!-- blocker: 80 reviewed public schemas have valid parse/serialize/ round-trip/schema-hash goldens, leaving 478 public schemas uncovered; 92 forbidden workbench-local declarations also remain to migrate. --> <!-- SPEC 2026-09-11: the 478 uncovered public schemas are closed by a generator, not by hand: emit one golden per component schema from openapi/metis.openapi.json (minimal and full instance, parse/serialize/round-trip, schema hash), commit the goldens, gate drift. The 92 workbench-local declarations migrate to libs/contracts/src/metis one logical shape at a time with the forbidden-import check switched on for that shape as it lands. The evidence block reports both counts before and after. -->
  • M1.2 Define immutable course and item revisions with stable identifiers. Evidence: versioning and restore invariants. <!-- evidence: revision-history.ts + metis-revision-invariants.json/JSONL define and exercise an append-only, hash-sealed course/item history; all 6 child invariants and 6 negative/positive probes pass. -->
    • M1.2.a Define stable logical IDs separately from immutable revision IDs, tenant/owner, parent/base revision, sequence/label, state, author/reason, timestamps, and integrity hash. Evidence: revision schemas. <!-- evidence: strict Zod envelopes separate logicalId/revisionId and require owner, lineage, state, authorship, semanticHash, and full integrityHash; tampering and historical ID reuse fail. -->
    • M1.2.b Define which course/module/lesson/activity/assessment/item/ stimulus/rubric/objective/metadata relations are captured by a course or item revision manifest. Evidence: revision boundary map. <!-- evidence: 20 boundary rows cover all 10 relations for both subject kinds with exact revision+content-hash pin, inline immutable metadata snapshot, or explicit exclusion. -->
    • M1.2.c Create revisions by copy-on-write or explicit references with optimistic concurrency and prohibit in-place changes to consumed/ reviewed/published revisions. Evidence: immutability tests. <!-- evidence: commitMetisContentRevision is append-only and checks the expected logical/branch head, dense sequence, unique revision ID, exact refs, locked states, consumer pins, and integrity hashes. -->
    • M1.2.d Define fork/branch/merge/conflict/supersede/archive/restore/ compare semantics without reusing a historical revision ID. Evidence: lifecycle invariants. <!-- evidence: 8 operation rows enforce parent/base/identity/state rules; merges require a common ancestor and explicit conflict resolutions, while comparison reports deterministic manifest-level differences. -->
    • M1.2.e Pin assessments/builds/delivery/attempts/grades/evidence to exact revisions and preserve resolvability after supersession. Evidence: downstream linkage tests. <!-- evidence: all 6 consumer kinds pin logicalId, revisionId, semanticHash, and integrityHash; the historical target remains resolvable after a superseding revision is appended. -->
    • M1.2.f Restore by creating a new revision semantically equal to the old one with restoration provenance, not by mutating history. Evidence: restore invariants. <!-- evidence: restore creates sequence 3 under a new revision ID, preserves the source semantic hash, creates a distinct integrity hash, records exact source/actor/time/reason, and leaves all 3 historical records byte-stable and resolvable. -->
  • M1.3 Define learning objectives with taxonomy, level, measurable verb, criteria, prerequisite, standard alignments, and evidence requirements. Evidence: schema tests. <!-- evidence: learning-objectives.ts + metis-learning-objectives.json; 6 contract test cases and 8 independent invariant probes cover all children and reject scope, taxonomy, threshold, graph, review, and coverage faults. -->
    • M1.3.a Define objective identity/version/scope/locale/title/description/ owner/lifecycle and course hierarchy membership. Evidence: base schema. <!-- evidence: stable objectiveId is separate from revisionId/revision/ contentHash; strict envelopes pin tenant, exact course hierarchy membership, locale, owner, lifecycle, creation, and descriptive fields. -->
    • M1.3.b Define taxonomy identity/version plus objective classification, level, measurable verb, subject/content, context, and allowed extensions. Evidence: taxonomy field tests. <!-- evidence: generic classifications require taxonomyId/version, classification/level codes, measurable verb, subject/content/context, and only namespaced x-owner:key extensions; pedagogy is not hard-coded. -->
    • M1.3.c Define observable performance criteria, conditions, mastery/ quality thresholds, units, rubric refs, and acceptable evidence kinds. Evidence: criteria tests. <!-- evidence: criteria require observable statements, unique conditions, bounded mastery/quality thresholds and units, exact rubric revisions, and at least one of 11 explicit acceptable evidence kinds. -->
    • M1.3.d Define prerequisite/requisite/related objective edges with version, strength, rationale, cycle/orphan/reachability rules. Evidence: prerequisite graph tests. <!-- evidence: three edge-policy rows pin exact objective revisions, edge version/strength/requirement/rationale, reject dependency cycles and orphans, and require publication reachability from declared roots. -->
    • M1.3.e Define standard/framework/version/jurisdiction/grade-band/code alignments with coverage/strength/rationale/evidence and human review. Evidence: alignment tests. <!-- evidence: strict alignment records include framework/version, jurisdiction, grade bands, code, coverage, strength, rationale, exact evidence refs, and complete approved/rejected human review. -->
    • M1.3.f Define content/activity/item/assessment/evidence coverage links and prevent publication of required objectives without ratified coverage. Evidence: coverage invariants. <!-- evidence: publication requires approved exact objective revisions, any mandatory approved standards alignment, and human-approved content/activity/item/assessment/evidence links; removing any one kind produces a blocking invariant. -->
  • M1.4 Define Bloom or ratified cognitive-level vocabulary without hard-coding it as the only pedagogy. Evidence: extensibility test. <!-- evidence: taxonomy-registry.ts + fixtures + metis-taxonomy-registry.json; 6 contract test cases and 8 independent probes cover all children. All fixture approvals are not production ratification; production entries still require their own complete human approval evidence. -->
    • M1.4.a Register cognitive/pedagogical taxonomy identity/version/owner/ lifecycle/source/license/locale and stable level codes. Evidence: vocabulary registry schema. <!-- evidence: strict registry entries pin immutable version hashes, owner/lifecycle/ratification, cited sources, conservative license usage, complete locales, and unique stable codes/order/relations. -->
    • M1.4.b Model Bloom levels/verbs/relations as one admitted vocabulary, including revised/classic profile distinctions where ratified. Evidence: Bloom fixtures. <!-- evidence: classic-1956 and revised-2001 are separate fixture-only versions/profiles in one family, each with six stable codes, localized labels, verbs, and explicit within-profile relations. -->
    • M1.4.c Allow other approved taxonomies and institution mappings through versioned registry entries rather than enum/code changes. Evidence: alternate taxonomy test. <!-- evidence: a bilingual institution-owned fixture admits five novel codes through the generic schema and maps them without adding a Bloom or institution-specific enum to the contract. -->
    • M1.4.d Define cross-taxonomy mappings as exact/broader/narrower/related/ unmapped with version/rationale and no false equivalence. Evidence: mapping tests. <!-- evidence: all five relations are versioned, rationalized, reviewed, and exact mappings additionally require full-meaning scope, confidence 1, evidence, and one non-conflicting target. -->
    • M1.4.e Make objective/item/UI/API/search/export consume taxonomy+version+ code and render locale labels without relying on Bloom-specific fields. Evidence: consumer extensibility tests. <!-- evidence: all six named consumers resolve the same exact generic taxonomy/version/hash/code reference, render requested or default locale labels, and strict requests reject bloomLevel fields. -->
    • M1.4.f Handle unknown/deprecated/unavailable taxonomy versions safely and preserve historical meaning. Evidence: lifecycle tests. <!-- evidence: deprecated/unavailable registry snapshots preserve label, version/hash/code, fallback locale, and replacement metadata; unknown taxonomy/version/code returns a null-label safe status. -->
  • M1.5 Define assessment item kinds including selected response, constructed response, essay, math, code, media, simulation, interaction, portfolio, and oral/performance evidence. Evidence: kind registry. <!-- evidence: assessment-kind-registry.ts + fixtures + metis-assessment-kind-registry.json; 7 contract test cases and 10 independent probes cover 15 kinds and all children. The contract registry is not runtime adoption; legacy service/QTI support remains separately measured and fixture approvals are not production approval. -->
    • M1.5.a Define versioned kind-registry entry with response/evidence model, interaction capabilities, scoring modes, delivery/runtime requirements, accessibility, standards mappings, security risk, and lifecycle. Evidence: kind schema. <!-- evidence: each strict version pins identity/hash/family/owner/review, response/evidence models, capabilities, scoring, runtime, WCAG 2.2 AA requirements, QTI/LTI dispositions, threats/controls, and lifecycle. -->
    • M1.5.b Define selected-response variants for single/multiple choice, ordering, matching, hotspot, and other ratified interactions with option/key invariants. Evidence: selected-response fixtures. <!-- evidence: five mode fixtures enforce unique option/target/region ids, exact single/multiple/ordered/matching/spatial key shapes, declared references only, accessible labels, and mode-specific fields. -->
    • M1.5.c Define constructed-response/essay response formats, length/ language/format constraints, scoring guides, exemplars, and moderation. Evidence: written-response fixtures. <!-- evidence: constructed/essay fixtures bound formats and MIME types, ordered length/word limits, locales, exact rubric/exemplar refs, and one- or two-assessor moderation rules. -->
    • M1.5.d Define math/code response, semantic/executable artifacts, equivalence/tests/sandbox/runtime, partial credit, and accessible alternatives. Evidence: math/code fixtures. <!-- evidence: math pins formats/equivalence/tolerance/weighted tests; code pins language/runtime image/template/tests and a no-network bounded deterministic sandbox; both carry partial credit and construct-equivalent accessible alternatives. -->
    • M1.5.e Define media/simulation/interaction inputs, states/events/ recordings, deterministic seeds, telemetry, scoring hooks, fallback, and sandbox. Evidence: interactive fixtures. <!-- evidence: three fixtures define modalities/state/events, consented recordings, required seeds, telemetry, versioned scoring hooks, exact fallbacks, origin isolation, and network allowlists. -->
    • M1.5.f Define portfolio and oral/performance evidence submissions, artifact/take/observation/assessor/rubric/consent/retention fields. Evidence: performance evidence fixtures. <!-- evidence: portfolio/oral/observed-performance fixtures pin artifacts, take bounds, observation fields, qualified assessors, exact rubrics, revocable consent, minor guardianship, and legal-hold-aware retention. -->
    • M1.5.g Require explicit unsupported/not-applicable behavior across authoring, delivery, accessibility, QTI/LTI/export, analytics, and scoring for every kind. Evidence: registry coverage matrix. <!-- evidence: 15 kinds x 8 named dimensions = 120 exact dispositions; supported rows require evidence while unsupported/not-applicable rows retain explicit rationale and QTI/LTI mapping concordance. -->
  • M1.6 Define stimuli and shared passages separately from items. Evidence: one stimulus can safely version and serve multiple items. <!-- technically complete: a separate immutable stimulus contract covers 12 required kinds, exact revision links, three-item/two-assessment passage reuse, change blast radius, and safe projections. The existing hierarchy shape remains a coarse compatibility boundary; the immutable stimulus contract is not runtime adoption. -->
    • M1.6.a Define stimulus logical/revision identity, kind, content blocks, source/citation/rights, locale, accessibility, metadata, lifecycle, and integrity hash. Evidence: stimulus schema. <!-- evidence: MetisStimulusRevisionSchema separates stable stimulus and immutable revision identity and governs typed blocks, provenance, citations, rights, locale, accessibility review, metadata, lifecycle, predecessor pins, content hash, and integrity hash. -->
    • M1.6.b Define text/passage, image/diagram, audio/video, dataset/table/ chart, code, document, and interactive stimulus-specific fields. Evidence: kind fixtures. <!-- evidence: 12 valid kind fixtures cover text, passage, image, diagram, audio, video, dataset, table, chart, code, document, and sandboxed interactive content with kind-specific accessibility and fallback fields. -->
    • M1.6.c Define item-to-stimulus link with exact revision, role, order, excerpt/locator, presentation context, shared instructions, and dependency invalidation. Evidence: link schema. <!-- evidence: strict MetisItemStimulusLinkSchema pins exact item and stimulus revisions, structured non-inline excerpts, presentation, exact shared-instruction refs, five invalidation domains, and rejects answer/scoring/exposure fields. -->
    • M1.6.d Permit one immutable stimulus revision to serve multiple items/ assessments while keeping each item key/scoring/exposure separate. Evidence: reuse tests. <!-- evidence: one passage revision is linked to three exact item revisions across two assessments; each item owns three unique protected-resource refs outside the links, and missing/shared security bindings fail graph validation. -->
    • M1.6.e Version source/content/accessibility/localization changes through a new stimulus revision and compute affected items/builds/attempts. Evidence: blast-radius tests. <!-- evidence: the r1-to-r2 fixture changes all four protected dimensions under a new-immutable-revision policy; deterministic traversal identifies three items, two assessments, two builds, and two attempts while preserving released/historical pins. -->
    • M1.6.f Prevent stimulus delivery/search/export from exposing linked item answers or unreleased test membership. Evidence: isolation tests. <!-- evidence: strict delivery/search/export projection tests serialize zero protected resource values and zero unreleased assessment ids; learner/search/export membership is empty and authorized delivery can reveal released membership only. -->
  • M1.7 Define answer, rationale, distractor rationale, scoring, feedback, hint, and exposure metadata. Evidence: answer-key fields are protected. <!-- technically complete: strict target contracts separate answer keys, rationales, scoring, learner feedback/hints, exposure, and property-level access. The current service still persists and can serialize inline correct_answer/explanation/hint fields; this protected assessment contract is not runtime enforcement. -->
    • M1.7.a Define protected answer/key forms per item kind with exact item revision, acceptable variants/equivalence/tolerance, ordering, and partial-credit semantics. Evidence: answer schema tests. <!-- evidence: 15 exact item-kind fixtures cover seven protected form families with immutable key identity, value/equivalence/tolerance/ ordering semantics, and priority-ordered partial-credit rules. -->
    • M1.7.b Define correct-answer rationale and per-distractor/response-pattern rationales separately from learner-visible feedback. Evidence: rationale schemas. <!-- evidence: the assessment-secret rationale schema has separate correct, option-distractor, and response-pattern refs and rejects learner feedback; feedback carries a distinct learner-visible classification and cannot include protected rationale refs. -->
    • M1.7.c Define deterministic/rubric/model-assisted/manual scoring method, steps/rules/weights/rounding, evidence, uncertainty, moderation, and version. Evidence: scoring schemas. <!-- evidence: four immutable scoring fixtures pin exact item/key revisions; contiguous weighted steps sum to one, rounding and evidence are explicit, uncertainty is bounded, and model-assisted scoring requires human moderation. -->
    • M1.7.d Define preattempt/postresponse/postsubmission/postdeadline/ instructor feedback and progressive hint ladders with penalty/exposure conditions. Evidence: feedback/hint tests. <!-- evidence: all five feedback stages are explicit; a three-step hint ladder pins availability, exposure, and increasing penalty while prohibiting final-answer and protected-rationale inclusion. -->
    • M1.7.e Define item/key/explanation exposure state, audience, window, attempts, secure-delivery/export/search/log/cache rules, and breach response. Evidence: exposure policy tests. <!-- evidence: item/key/explanation rows govern state, audience, window, and attempts; five surface rules default-deny/redact/no-store, and eight mandatory critical-breach actions seal, purge, rotate, notify, and preserve evidence. -->
    • M1.7.f Enforce property-level permissions for answer/rationale/scoring/ distractor fields across UI/API/tutor/search/export/events/analytics and two tenants. Evidence: protected-field suite. <!-- evidence: 8 protected properties x 7 surfaces produce 56 explicit default-deny rows; 112 same/cross-tenant learner probes deny and audit every value, 16 author UI/API grants are purpose-bound, and 14 safe projections serialize no protected fixture sentinel. -->
  • M1.8 Define rubric criteria, levels, descriptors, weights, scoring method, moderation, and evidence refs. Evidence: deterministic score invariants. <!-- technical evidence: libs/contracts/src/metis/rubric.ts, fixtures, six contract tests, and deterministic metis-rubrics.json define the immutable assessment-rubric authority and five exact score outcomes. The assessment rubric contract is not runtime adoption; hierarchy summaries and agent-evaluation rubrics remain distinct boundaries. -->
    • M1.8.a Define rubric logical/revision identity, scope, owner, locale, analytic/holistic/single-point method, lifecycle, applicable items/ evidence kinds, and integrity hash. Evidence: rubric schema. <!-- evidence: three method fixtures validate exact hierarchy projection, item/assessment/evidence scope, owner, locale, lifecycle, immutable revision identity, content hash, and integrity hash. -->
    • M1.8.b Define stable criteria with objective links, title/description, required flag, weight/points, ordering, dependencies, and evidence requirements. Evidence: criterion tests. <!-- evidence: two stable analytic criteria pin objective revisions, localized text, 0.6/0.4 weights, ten-point maxima, contiguous order, an acyclic dependency, and exact artifact/annotation requirements. -->
    • M1.8.c Define ordered performance levels with codes, descriptors, score/point/range, boundary inclusivity, exemplars, and locale variants. Evidence: level tests. <!-- evidence: range, point, and rating fixtures enforce one score mode, unique contiguous levels, localized descriptors, exemplars, and adjacent boundaries with exactly one inclusive side. -->
    • M1.8.d Define weighted/points/custom rule aggregation, missing/N/A/ unscorable handling, rounding, caps/floors, and total-scale mapping. Evidence: exact score fixtures. <!-- evidence: deterministic results are weighted 72, N/A-renormalized 80, missing/incomplete, points 85, and custom 70; wrong revision, level, maximum, and custom-rule inputs fail closed. -->
    • M1.8.e Define assessor evidence refs, annotations, confidence, criterion decisions, comments, moderation/second marking, reconciliation, and audit. Evidence: scoring evidence tests. <!-- evidence: the exact evidence bundle resolves artifacts and annotations per criterion, captures primary/second confidence and comments, reconciles by consensus, and preserves a five-event audit. -->
    • M1.8.f Version rubrics immutably and pin assessment/scoring/grade records to exact revisions; changes cannot recalculate historical grades without correction workflow. Evidence: versioning invariants. <!-- evidence: r1-to-r2 creates a new immutable revision and hashes; assessment, scoring-policy, and grade consumers pin r1 exactly, while historical r2 use requires correction-not-mutation workflow. -->
  • M1.9 Define accommodations and Personal Needs and Preferences references. Evidence: support requirements remain separate from disability diagnoses. <!-- technical evidence: libs/contracts/src/metis/accommodations.ts, fixtures, six contract tests, and deterministic metis-accommodations.json define functional supports only. The accommodation contract is not runtime adoption or a standards conformance claim; current study preferences remain separate. -->
    • M1.9.a Adopt/version the applicable PNP/accommodation vocabulary and map support codes to delivery capabilities without copying diagnoses. Evidence: standards mapping. <!-- evidence: pins 1EdTech AfA PNP 3.0 Public Candidate Final Draft v2 (2021-01-18), aligns 17 functional support codes to QTI 3 published accessibility supports or a namespaced Metis extension, maps every code to a versioned capability, and explicitly claims no conformance. -->
    • M1.9.b Define authoring-side support requirement for presentation, content, navigation, timing/scheduling, response, setting, language, and assistive technology compatibility. Evidence: requirement schema. <!-- evidence: 17 immutable requirements cover all eight categories with exact support/vocabulary/capability refs, activation and semantics, alternatives, evidence, previewability, and fail-closed behavior. -->
    • M1.9.c Define purpose-bound learner PNP projection with opaque learner/ profile ref, requested/approved support codes, source authority, effective term, and no medical detail. Evidence: projection contract. <!-- evidence: restricted synthetic projection uses opaque learner/profile refs, four requested and three approved functional codes, exact institutional authority/effective context, and rejects unknown, unrequested, non-opaque, or diagnosis-bearing fields. -->
    • M1.9.d Define conflicts, mutually exclusive supports, item invalidation, equivalent alternatives, human review, and cannot-deliver behavior. Evidence: compatibility tests. <!-- evidence: evaluator proves deliver, equivalent-alternative, human-review, and cannot-deliver outcomes; theme conflict, item invalidation, missing capability, and unsupported switch input fail closed with exact reasons. -->
    • M1.9.e Enforce field-level access/retention/audit and prohibit inference or storage of disability diagnosis from support selection/use. Evidence: privacy tests. <!-- evidence: seven protected fields have exact retention/delete rules; 98 author/cross-tenant probes deny values, 22 learner/delivery grants are explicit and audited, and diagnosis/inference/training use are prohibited across seven surfaces. -->
    • M1.9.f Verify authoring preview and learner delivery apply the same supported semantics while protected learner records remain external. Evidence: preview/delivery parity tests. <!-- evidence: preview and delivery deep-equal the same three-row semantics plan; preview contains no learner/profile ref, delivery embeds only an opaque projection ref, and any semantic drift fails validation. -->
  • M1.10 Define locale variants, bidirectional text, MathML, SSML, transcripts, captions, long descriptions, tactile refs, and sign-language assets. Evidence: QTI accessibility mapping. <!-- technically complete 2026-08-28: accessible-content.ts, fixtures, 17 contract tests, the deterministic generator/test, and metis-accessible-content.json cover all eight children and 106 canonical QTI field/asset paths. QTI 3.0.1, MathML 3, SSML/Data-SSML, WebVTT draft status, and BCP 47/UAX #9 are pinned as contract profiles; no standards conformance/certification is claimed. Protected reviewer, presenter, and consent refs are omitted from packages and evidence. This accessible-content contract is not runtime adoption; the current stimulus/lecture/QTI implementations remain a documented migration. -->
    • M1.10.a Define locale/language/script/region/direction variant identity, source revision, translation status, fallback, reviewer, and change lineage. Evidence: locale variant schema. <!-- evidence: two immutable exact-ref variants exercise canonical language/script/region tags, LTR/RTL direction, source/translation linkage, human review, fallback, source revision, and acyclic change lineage; stale hashes and fallback cycles fail validation. -->
    • M1.10.b Define Unicode/bidirectional isolation/base direction/mixed- script/numeral/punctuation behavior for content, answers, code, math, metadata, and input. Evidence: bidi fixtures. <!-- evidence: six exact context rows require markup/plain-text isolates, context-specific code/math/input behavior, neutral punctuation and numeral policies; embeddings, overrides, duplicates, and unbalanced isolates are rejected. -->
    • M1.10.c Define semantic MathML source, presentation/content profile, speech/braille alternatives, equation refs/labels, and fallback image/ text. Evidence: MathML fixtures. <!-- evidence: stable MathML 3 parallel presentation/content markup binds semantic source, equation/label, reviewed speech and braille, image/ text fallback, and exact objectives; active/external/nonallowlisted markup fails the renderer profile. -->
    • M1.10.d Define allowlisted SSML profile, language/voice/pronunciation/ prosody/mark semantics, text equivalent, and safe renderer behavior. Evidence: SSML fixtures. <!-- evidence: eight instruction kinds cover Data-SSML plus profiled language/voice/mark semantics, always bind reviewed text fallback, and prohibit external audio/lexicons/scripts; unknown and unordered instructions fail validation. -->
    • M1.10.e Define transcript and caption tracks with language, speaker, timebase/cues, non-speech audio, review, synchronization, and formats. Evidence: timed-text fixtures. <!-- evidence: versioned caption/transcript tracks bind exact media and locale refs, speakers, millisecond cues, non-speech audio, human review, <=500ms sync evidence, and required WebVTT/structured-HTML primaries; unresolved speakers/formats and overrun cues fail. -->
    • M1.10.f Define alt text, extended/long description, tactile graphic/ braille-ready refs, data/source linkage, and equivalent learning outcome. Evidence: nonvisual alternative tests. <!-- evidence: the nonvisual fixture binds short/long descriptions, tactile graphic/tour, braille-ready rendition, exact source data and objective refs, and affirmative equivalence review; missing tactile/braille form or non-objective linkage fails. -->
    • M1.10.g Define sign-language asset language/community/presenter/consent/ rendition/time alignment and fallback semantics. Evidence: sign-language fixtures. <!-- evidence: an ase fixture requires declared community, opaque presenter, exact consent, one primary rendition, positive source/ rendition alignment, written locale and transcript fallback; missing consent, multiple primaries, and unresolved fallback fail. -->
    • M1.10.h Map every field/asset to QTI accessibility metadata/content, document explicit loss/unsupported behavior, and round-trip. Evidence: QTI accessibility mapping tests. <!-- evidence: all 106 canonical paths map exactly once with exact, normalized, lossy, unsupported, or policy-omitted disposition and matching round-trip behavior; missing rows, false exactness, or reviewer/presenter/consent export fail closed. -->
  • M1.11 Define QTI 3 package/import/export profile. Evidence: every supported item kind maps or fails with an explicit incompatibility. <!-- CLOSED 2026-09-18, all seven children closed and the parent's own sentence measured rather than inherited: of the 15 canonical item kinds, 7 map and hold a passing non-synthetic receipt from each of the two pinned tools, and the other 8 fail with an explicit code the coverage manifest names. The conformance gate over real exported bytes reads passed: true, 0 blockers. WHAT THIS DOES NOT SAY: no product route calls the QTI 3 runtime, so no runtime adoption is claimed here and M10.5's parent stays open on exactly that. Certification by 1EdTech is not claimed either; what is held is two pinned tools' own output over our own artifacts. --> <!-- technical progress 2026-08-28: qti-package.ts, fixtures, 54 contract tests, the deterministic generator/test, and metis-qti-package.json complete M1.11.a-f. The profile pins QTI 3.0.1 and both external validator roles, covers 11 canonical objects, all 15 item kinds, 90 capability cells, 11 content/accessibility dimensions, 18 hostile imports, and a deterministic schema-validated export golden. M1.11.g remains blocked because no repository-installed official QTI validator bundle or independent consumer has produced the required 14 non-synthetic receipts. The parent remains open; no conformance, certification, or current runtime adoption is claimed. -->
    • M1.11.a Pin supported QTI 3 specification/version/conformance modules, namespaces, package/resource/manifest profile, extensions, and validator toolchain. Evidence: profile decision. <!-- evidence: pins QTI 3.0.1/BPIG document v2.0 (2024-10-01), nine modules, QTI/package/MathML/XML/Metis namespaces, three SHA-256-pinned root schemas, six resource types, two allowlisted extensions, and official-schema plus independent-consumer validator roles. -->
    • M1.11.b Map canonical assessment/test/section/item/stimulus/response/ outcome/processing/rubric/objective/metadata identities and versions to QTI constructs. Evidence: field mapping. <!-- evidence: eleven exact-ref rows map every canonical identity once; rubric/objective/metadata normalization is explicit, response and processing are protected, and no protected values enter evidence. -->
    • M1.11.c Map each item kind/interaction/scoring/feedback/hint/adaptive/ template capability with exact supported/lossy/extension/unsupported disposition. Evidence: item-kind matrix. <!-- evidence: all 15 immutable registry kind refs map exactly once over 90 capability cells; four exact, one normalized, two lossy, five unsupported, and three inapplicable kind rows have matching round-trip or explicit incompatibility behavior. -->
    • M1.11.d Map media/files/MathML/accessibility/locale/bidi/captions/ transcripts/PNP/dependencies and package paths securely. Evidence: content/accessibility matrix. <!-- evidence: eleven exact dimension rows require immutable resource hashes, declared offline dependencies, allowlisted MIME types, no learner PNP values, and reject absolute/traversal/encoded/URI/ backslash/control/empty/dot/overlong package paths. -->
    • M1.11.e Import with archive/schema/size/path/resource protections, preview/migration/conflict/reporting, immutable source package, and no answer-key exposure. Evidence: hostile import tests. <!-- evidence: bounded ZIP/XML/resource evaluation rejects traversal, symlink, encryption, bombs, case collisions, undeclared/missing or external resources, schema/entity failures, mutable source, source-hash/MIME mismatch, embedded active/external content, and answer/private preview exposure across 18 hostile probes. -->
    • M1.11.f Export deterministic validated packages with exact revisions, standards metadata, checksums, rights, declared extensions/loss report, and no secrets/private fields. Evidence: golden exports. <!-- evidence: a no-wall-clock golden pins three source revisions, sorted payload-derived file hashes/bytes, canonical ZIP metadata, descriptor hash, QTI metadata, schema receipts, rights, two extensions, explicit losses, and nine prohibited private/secret field classes; tamper, reorder, rights denial, or exposure fails closed. -->
    • M1.11.g Round-trip every supported item-kind/profile through pinned independent validators/consumers and fail explicitly for incompatible cells. Evidence: QTI conformance suite. <!-- CLOSED 2026-09-18 WITH M10.5.d (§2.5.6: one gate, together). 14 of 14 non-synthetic passing receipts are held, one per supported kind per pinned tool, under evidence/v1-workbenches/metis-conformance/, and tests/test_qti3_conformance.py asserts the held set EQUALS the set the contract's matrix requires. The two that were missing were hotspot's: Metis now has a canonical spatial item, and the consumer caught a live defect in its first export (the image written as img, which the XSD admits inside a hotspot interaction and the consumer refuses). All eight incompatible cells still fail with explicit codes; the conformance gate evidence/v1-workbenches/metis-qti/conformance-gate.json reads passed: true with 0 blockers. --> <!-- blocked 2026-08-28: the contract requires a passing non-synthetic receipt from both the pinned official QTI 3.0.1 schema toolchain and independent QTI consumer for each of seven mapped kinds (14 total). Neither tool is installed and no genuine receipt exists; all eight incompatible cells already fail with explicit codes. --> <!-- still blocked 2026-09-11, at 12 of 14: M10.5.d installed and pinned both tools (apps/metis/service/tools/qti3/) and retained 12 passing non-synthetic receipts under evidence/v1-workbenches/metis-conformance/, verified against the contract's own version and integrity pins by tests/test_qti3_conformance.py. The two that remain are hotspot's, and they cannot be taken: Metis has no canonical hotspot item, so there is nothing to hand a validator. Closing this needs a canonical hotspot model, not another validator run. §2.5.6 binds this to M10.5.d. -->
  • M1.12 Define LTI 1.3/LTI Advantage launch, deep-link, roles, and grade service profile. Evidence: standards matrix.
    • M1.12.a Pin LTI 1.3/LTI Advantage service versions/conformance, deployment/issuer/client/key registration, platforms, privacy profile, and supported flows. Evidence: profile decision. <!-- evidence: the versioned target profile pins Core 1.3 document v3, Deep Linking 2.0 document v2, NRPS 2.0, AGS 2.0 document v3, Security Framework 1.1, and certification guide v3; asymmetric registration binds tenant/platform/issuer/client/deployment, OIDC/token/JWKS/login/redirect endpoints, RS256/2048-bit keys, five scopes, six platform targets, no-PII defaults, six flows, and an explicit no-conformance/no-runtime-adoption posture. -->
    • M1.12.b Define OIDC login and resource-link launch validation for issuer, audience/client, deployment, nonce/state, message type/version, target link, roles, context, resource, user, locale, and custom claims. Evidence: launch contract tests. <!-- evidence: unknown-input validation requires one-time state/nonce, issuer/client/audience/authorized-party/deployment/tenant binding, RS256 signature/kid/pinned JWKS, bounded iat/exp/skew, LTI 1.3 type/version, signed target link, subject policy, official role, context/resource, BCP47 locale, and string-only custom claims; replay, substitution, expiry, HS256, and cross-tenant probes fail. -->
    • M1.12.c Map institutional/course/section/role claims to canonical identities and least-privilege authoring/learner/educator operations without trusting display fields. Evidence: role mapping matrix. <!-- evidence: eleven official system/institution/context/marker role URIs map to bounded canonical grants; learner gets launch/own-result, mentor gets launch only, and tenant/institution/course/section/ principal/placement projection uses signed opaque identifiers with no trusted display fields. -->
    • M1.12.d Define deep-link request/response for course/content/assessment selection, exact revision/resource, placement, availability, grade settings, signing, and return validation. Evidence: deep-link tests. <!-- evidence: request/response contracts bind deployment, exact HTTPS return URL, opaque data echo, accepted item/target/multiplicity/ line-item capabilities, tool/platform issuer/audience, RS256/kid, one-time nonce, bounded iat/exp/skew, exact course/content/assessment revision+hash, placement, availability, grade settings, and string custom refs while excluding answer keys and learner-private data. -->
    • M1.12.e Define Names and Role Provisioning scope/pagination/freshness/ privacy plus Assignment and Grade Services line-item/score/result directionality, idempotency, correction, and visibility. Evidence: service contract tests. <!-- evidence: private_key_jwt token exchange binds registered client iss/sub, token-endpoint aud, deployment, exact scopes, RS256/kid, one-time JTI, and 300-second lifetime; NRPS GET/media/scope/context binds role/resource filters, limit/next/differences, origin, freshness, minimal fields, and purpose receipts for PII; AGS pins line-item revision, score-write/result-read direction, visibility, increasing timestamps, correction lineage, and no answer material. -->
    • M1.12.f Define key rotation/JWKS cache, replay/clock skew, launch expiry, logout/session, error/redaction, and tenant/deployment isolation. Evidence: security tests. <!-- evidence: policy requires RS256/2048-bit keys, bounded Cache-Control JWKS caching, one refresh/retry for unknown kid, fail-closed stale/ revoked keys, unique kid with overlap rotation, encrypted private keys, durable atomic state/nonce/JTI consumption, 120-second skew, 600-second launch lifetime, bounded sessions, local logout/token revocation, public redacted errors, and tenant/issuer/client/ deployment/context isolation. -->
    • M1.12.g Run platform/tool conformance and consumer-driven tests for each supported flow/role/failure, declaring unsupported cells. Evidence: LTI standards matrix. <!-- blocked 2026-08-28: every one of six flow cells declares its role and ten failure-class coverage, but the gate requires a passing non-synthetic receipt from both the pinned 1EdTech certification suite and an independent LTI 1.3 platform per flow (12 total). Neither external system has been executed and no genuine receipt exists; the target contract is not runtime-adopted or certified. --> <!-- half the gate is met 2026-09-12 (M10.6): an independent LTI 1.3 platform — Moodle 4.5.13+, installed on this server and driving Metis through its own API — produced non-synthetic receipts for the launch, deep-linking, grade-service and key-rotation flows, retained under evidence/v1-workbenches/metis-conformance/moodle-lti-1p3/. The grade is in Moodle's own gradebook table, not in a report of ours. The 1EdTech certification suite half is untouched: it is hosted and member-only, so §2.5.6 says record the attempt and leave the cell open. It is open. -->
  • M1.13 Define OneRoster, Caliper, Open Badges, SCORM, and xAPI profiles. Evidence: source-of-truth and directionality for each field.
    • M1.13.a Pin version/conformance/transport/auth/profile/extensions/ validators and system-of-record ownership for every listed standard. Evidence: standards decision matrix. <!-- Evidence: libs/contracts/src/metis/interoperability-profiles.ts, libs/contracts/src/metis/interoperability-profiles.fixtures.ts, and evidence/v1-workbenches/metis-interoperability-profiles.json pin OneRoster REST 1.2 + CSV 1.2.1, Caliper 1.2, Open Badges 3.0 document 1.4.5, SCORM 1.2, SCORM 2004 4th Edition, and IEEE 9274.1.1-2023/xAPI 2.0. Every row declares transport, authentication, direction, extensions, validator, owner, no runtime adoption, and no conformance claim. -->
    • M1.13.b Define OneRoster organization/course/class/term/user/enrollment/ academic-session/category/line-item/result mappings, directionality, identifiers, delta, deletion, and conflict rules. Evidence: OneRoster mapping tests. <!-- Evidence: the shared contract, its 37-test focused suite, and the deterministic evidence cover all ten entity families, tenant + connector + sourcedId identity, SIS-owned inbound rosters, Metis-owned outbound gradebook records, REST watermarks, complete CSV snapshots, tombstones, parent references, and fail-closed ambiguity/conflict rules. -->
    • M1.13.c Define Caliper entity/event/action/actor/object/generated/ membership/assessment/grade semantics, profile versions, privacy, event ordering, and delivery. Evidence: Caliper mapping tests. <!-- Evidence: Caliper 1.2 Assessment, Grading, Session, and Tool Use rules bind actor/object/generated/membership semantics, pairwise identity, purpose, eventTime + sequence ordering, TLS Bearer Sensor delivery, envelope idempotency, and bounded retry/failure policy. -->
    • M1.13.d Define Open Badges issuer/profile/achievement/alignment/criteria/ evidence/assertion/recipient/status/endorsement mapping and verification direction. Evidence: Open Badges tests. <!-- Evidence: the Open Badges 3.0 VC Data Model 2.0 target validates JSON-LD safe mode, Profile issuer/creator, exact achievement and evidence revisions, alignment/criteria, pairwise recipient, Bitstring revocation status, endorsement, EdDSA Data Integrity proof/key authority, expiry, and presenter-to-verifier direction. -->
    • M1.13.e Define SCORM package/manifest/SCO/asset/sequencing/runtime data/ completion/success/score/interaction/bookmark/error mappings and sandbox. Evidence: SCORM profile tests. <!-- Evidence: distinct SCORM 1.2 and 2004 4th Edition package/runtime schemas cover root manifests, SCO/assets, 2004 sequencing and navigation, API vs API_1484_11, edition-specific CMI status/score/ interaction/bookmark/error mapping, commit lineage, archive-path safety, isolated package origins, CSP, and capability bridging. -->
    • M1.13.f Define xAPI actor/verb/object/result/context/authority/statement refs/attachments/state/activity/profile mapping, LRS direction, voiding, and privacy. Evidence: xAPI profile tests. <!-- Evidence: the IEEE 9274.1.1-2023/xAPI 2.0 contract covers one pairwise actor IFI, verb/object/result/context/authority, immutable statements and StatementRefs, hash-verified retained attachments, ETag-protected State/Activity Profile/Agent Profile documents, explicit provider/consumer direction, tenant-bound voiding, and a truthful no-full-LRS-conformance boundary. -->
    • M1.13.g Document lossy/unsupported/conflicting fields, canonical precedence, version negotiation, and round-trip/conformance evidence per profile. Evidence: cross-standard matrix. <!-- Technical matrix complete; external gate open. The deterministic artifact classifies 12 canonical fields across all six profiles, pins owner-first precedence and fail-closed version negotiation, and records synthetic semantic round trips with zero undeclared losses. Synthetic results are not conformance receipts. One genuine official-validator receipt plus one genuine independent-system receipt is still required for each profile (12 genuine external receipts total); none exists, so M1.13.g and M1.13 remain open and the target contract is not runtime-adopted or conformant. -->
  • M1.14 Define psychometric metadata for classical item statistics, 1PL, 2PL, 3PL, calibration sample, standard error, DIF, exposure, and status. Evidence: schema invariants and privacy review. <!-- Evidence: libs/contracts/src/metis/psychometric-metadata.ts, its synthetic fixtures and 39-test focused suite, and evidence/v1-workbenches/metis-psychometric-metadata.json define the complete metadata envelope, exercise 18 fail-closed classes, and record 19/19 passing invariant probes. The current runtime audit remains explicit: partial CTT/IRT/DIF/exposure helpers exist, but the target contract is not runtime-adopted. -->
    • M1.14.a Define analysis/calibration identity/version, exact item revision, population/sample/filter/window, method/software/config, owner, and lifecycle. Evidence: psychometric run schema. <!-- Evidence: the run schema pins tenant, run/version, exact item, population/sample, analysis kinds, method, software/version/build, config/hash/seed, assumptions, owner, lifecycle, snapshots, and chronologically valid terminal output. -->
    • M1.14.b Define classical difficulty/facility, discrimination, point- biserial, option/distractor statistics, reliability contribution, missingness, counts, and uncertainty. Evidence: CTT schema tests. <!-- Evidence: CTT schemas reconcile facility/difficulty and 13 count fields, upper/lower and point-biserial discrimination, protected keyed/distractor categories, reliability-if-deleted, missingness, standard errors, and uncertainty intervals. -->
    • M1.14.c Define 1PL/2PL/3PL ability scale, difficulty, discrimination, guessing, information/fit, parameter constraints, transformation/linking, standard errors, and convergence. Evidence: IRT schema tests. <!-- Evidence: exact 1PL/2PL/3PL invariants govern the versioned logit scale, a/b/c estimation/fixing, constraints, item information and conditional SE, fit/held-out evidence, exact anchor linking and transformation uncertainty, and convergence before approval. -->
    • M1.14.d Define calibration sample inclusion/exclusion, size, weights, representativeness, administration/item versions, privacy threshold, and data quality. Evidence: sample invariants. <!-- Evidence: the sample schema pins population/frame/filter/window, rules, reconciled counts, weighting/effective n, representativeness, exact administration/item versions, quality, institution-approved aggregate/subgroup floors, disclosure review, suppression, retention, and pseudonymous linkage. -->
    • M1.14.e Define DIF group descriptors through privacy-preserving approved dimensions, method, reference/focal groups, effect/significance/ uncertainty, sample floors, and review outcome. Evidence: DIF tests. <!-- Evidence: DIF groups use opaque ids and hashed value tokens with approved dimensions/labels/consent; results pin matching, method, effect, multiple-comparison significance, uncertainty and floors, suppress undersized cells, and require a non-causal human review. -->
    • M1.14.f Define exposure count/rate/window/population/administration/ security classification and active/field-test/calibrating/review/ approved/retired/compromised status. Evidence: exposure/status tests. <!-- Evidence: exposure binds exact item/window/population/administration versions, opportunity count/rate/uncertainty, control policy, classification and incident; an append-only, decision-backed state machine governs all seven named statuses. -->
    • M1.14.g Separate raw/estimated statistics from policy/quality decisions, pin versions, protect small groups/learner data, and prohibit unsupported individual inference. Evidence: privacy/invariant review. <!-- Evidence: immutable estimate bundles are separate from versioned, non-automatic human policy decisions; the privacy review prohibits direct identifiers, raw responses/demographics/answer material, small-cell release, estimate mutation, and six classes of unsupported individual inference with no override. -->
  • M1.15 Define academic-integrity evidence, classifier outputs, decisions, appeals, and audit references. Evidence: competing integrity vocabularies are reconciled without lossy mapping. <!-- Evidence: libs/contracts/src/metis/academic-integrity-evidence.ts, its synthetic fixtures and 43-test focused suite, and evidence/v1-workbenches/metis-academic-integrity-evidence.json define the target envelope, reconcile 12 source vocabularies and 64 source values with explicit loss status, and exercise 20 fail-closed classes through 21/21 invariant probes. The runtime audit remains explicit: legacy verdict, detector, incident, appeal, and grade projections are partial, and the target contract is not runtime-adopted. -->
    • M1.15.a Inventory/reconcile plagiarism, proctoring, authorship, similarity, behavior, device/session, assessment-security, incident, decision, and appeal vocabularies with versioned mappings. Evidence: vocabulary map. <!-- Evidence: the registry preserves every namespaced source id/version/value/meaning, covers all ten requested concept families and seven semantic layers, and marks exact versus context-required mappings without permitting automatic adjudication. -->
    • M1.15.b Define evidence identity/version/source/collector, learner/ attempt/item refs, event/time range, integrity hash, classification, access, retention, chain of custody, and redaction. Evidence: evidence schema. <!-- Evidence: versioned evidence pins a pseudonymous learner, exact attempt/assessment/item revisions, event and source collection, protected artifact/hash, policy-owned access/retention/redaction, and chronological hash-linked custody with audit refs. -->
    • M1.15.c Define classifier/rule output with model/code/config/version, input refs, scores/features, uncertainty, applicability, limitations, threshold, and no direct misconduct conclusion. Evidence: output schema. <!-- Evidence: output schemas pin producer/model/code/config, exact evidence inputs, source label/mapping, scores/scales, protected features, calibrated uncertainty, applicability/OOD, limitations, and a review-prioritization-only threshold; four automatic uses are prohibited. -->
    • M1.15.d Define human case/decision with allegation/category, evidence reviewed, standard/policy, actor/authority, rationale, outcome, consequence recommendation, and separation from grade mutation. Evidence: decision schema. <!-- Evidence: a human-filed allegation is never a finding; adverse outcomes require a conflict-checked authorized human and material verified/corroborated evidence. Consequences remain recommendations and grade mutation requires a separate workflow. -->
    • M1.15.e Define appeal filing/evidence/access/independent reviewer/ deadlines/stay/outcome/remedy and link it without overwriting the original decision. Evidence: appeal contract. <!-- Evidence: appeals pin filing/grounds, evidence access, an independent authorized reviewer, policy deadlines/timezone, stay, outcome/remedy, and a distinct superseding decision while the original decision remains immutable. -->
    • M1.15.f Define append-only audit refs for collection/access/redaction/ sharing/decision/appeal/correction/deletion/legal hold and institution boundary. Evidence: audit tests. <!-- Evidence: a contiguous, chronological, hash-linked audit trail defines all nine actions, resolves custody and target refs, gates cross-institution sharing on an agreement/minimum-necessary check, and blocks deletion under legal hold. -->
    • M1.15.g Preserve source vocabulary and semantic loss status through mappings; reject lossy automatic adjudication. Evidence: reconciliation invariants. <!-- Evidence: source vocabulary/meaning is mandatory on every mapping; non-exact mappings disclose loss and require human interpretation. Classifiers cannot open cases, find misconduct, recommend consequences, mutate grades, or resolve appeals automatically. -->
  • M1.16 Define gradebook correction and visibility contracts. Evidence: append-only correction and role-redaction tests. <!-- Evidence: libs/contracts/src/metis/gradebook-corrections.ts, its synthetic fixtures and 49-test focused suite, and evidence/v1-workbenches/metis-gradebook-corrections.json define the target envelope across 10 immutable histories/correction types, eight roles, 17 fields, nine actions, six aggregate dimensions, and LTI AGS 2.0 plus OneRoster 1.2 receipts. Twenty-nine fail-closed classes pass 30/30 invariant probes. The runtime audit remains explicit: five legacy grade surfaces are partial or missing, and the target contract is not runtime-adopted. -->
    • M1.16.a Define grade record identity/version for learner, course/ assessment/item/attempt revisions, score/value/status, scale, source, grader, timestamps, and evidence refs. Evidence: grade schema. <!-- Evidence: every grade record pins a pseudonymous learner, exact course/assessment/item and submission revisions, attempt, bounded state/scale, source, grader, timestamps, protected evidence, and exact-hash contiguous prior-version lineage. -->
    • M1.16.b Define append-only correction command with target grade version, reason/category, old/new values, evidence, requester, approver, effective time, and idempotency. Evidence: correction schema. <!-- Evidence: immutable version-one commands bind the exact expected grade, policy, old/new state, protected reason/evidence, institution-scoped actors, approval/application chronology, and a tenant+grade+version+category idempotency key. -->
    • M1.16.c Define recalculate, regrade, override, void, excuse, late/ incomplete, moderation, appeal, and import correction semantics separately. Evidence: correction-type tests. <!-- Evidence: ten discriminated command schemas preserve separate formula, rubric, exceptional authority, void, eligibility, late, incomplete, panel, appeal-decision, and import-conflict meaning; the complete fixture exercises every category exactly once. -->
    • M1.16.d Enforce eligible states/roles, separation of duties, expected version, policy, scale/range, downstream aggregate impacts, and audit. Evidence: correction guard tests. <!-- Evidence: versioned policy rows govern eligible states and actors, high-risk commands separate requester/approver, scale/range and exact-version guards fail closed, all six downstream dimensions receive one impact, and 79 append-only hash-linked events audit request/approval/application/release/notification/propagation. -->
    • M1.16.e Define field/action visibility for learner, guardian, teacher, assessor, administrator, integrity reviewer, support, and integrations by release window/privacy. Evidence: visibility matrix. <!-- Evidence: the default-deny matrix defines all 17 fields and nine actions for eight roles; every grant requires privacy policy, learner/guardian/integration access is release-gated with its specific authority, and support requires a legitimate active case while sensitive support/integration fields remain denied. -->
    • M1.16.f Project current grade from immutable history, preserve prior values/reasons, notify affected actors safely, and propagate LTI/ OneRoster corrections idempotently. Evidence: end-to-end tests. <!-- Evidence: current state is derived from exact immutable history; prior state/reasons stay protected; 20 idempotent notifications disclose no narrative/evidence/internal actor/direct identifier; and 20 exact-lineage LTI/OneRoster receipts use unique keys and strictly newer timestamps without claiming platform adoption or standards certification. -->
  • M1.17 Define course-build and assessment-evidence bundles. Evidence: canonical contract goldens validate. <!-- Evidence: libs/contracts/src/metis/canonical-bundles.ts, its synthetic fixtures and focused suite, six byte-canonical JSON goldens, and evidence/v1-workbenches/metis-canonical-bundles.json define distinct publishable course and protected assessment-record envelopes across 23 artifact kinds. All six valid bundles and 29 fail-closed classes pass 35/35 invariant probes. The runtime audit remains explicit: six adjacent or legacy surfaces do not adopt the target contract, which is not runtime-adopted. -->
    • M1.17.a Define course-build bundle manifest with exact course hierarchy/ content/objective/standard/item/stimulus/rubric/media/localization/ accessibility revisions, dependencies, target profile, and hashes. Evidence: course bundle schema. <!-- Evidence: every course golden binds exact source revision/content hashes, exact dependency refs, target consumers/features/locales, and canonical payload hashes; the full golden covers assessment, item, stimulus, rubric, media, localization, accessibility, and interactive artifacts in addition to the base hierarchy. -->
    • M1.17.b Define assessment-evidence bundle manifest with assessment/item/ stimulus/rubric revisions, delivery config, response/evidence refs, scoring/moderation, integrity, accommodations projection, grades/ corrections, and protected-field partitions. Evidence: evidence schema. <!-- Evidence: the protected golden requires all 13 assessment-record kinds, exact attempt/root refs, a pseudonymous subject hash, and exactly one learner, answer, integrity, and review partition. -->
    • M1.17.c Define encryption/signature/access/retention/redaction/selective disclosure for learner/answer/integrity evidence separately from publishable course content. Evidence: security tests. <!-- Evidence: course payloads are publishable canonical JSON with no protected record kinds; every protected payload is ciphertext in an A256GCM envelope partition with unique nonce/key-ref hashes, default-deny access, retention/redaction policy refs, and minimum-necessary disclosures. Detached EdDSA JWS values remain explicitly synthetic-unverified, not production crypto proof. -->
    • M1.17.d Resolve exact dependencies, normalize manifest/order/filenames/ timestamps, compute hashes, and declare external/redacted/missing refs. Evidence: determinism tests. <!-- Evidence: the profiled I-JSON/JCS serializer rejects ambiguous values and pins object/array/ref/path/timestamp/archive ordering; manifests, payload bytes, and ordered payload sets are rehashed, dependency graphs are exact and acyclic, paths are portable, and external/redacted/optional-missing descriptors are explicit. -->
    • M1.17.e Define version/migration/compatibility/loss/validator/import/read- back semantics for each target consumer. Evidence: compatibility tests. <!-- Evidence: every target consumer has an exact, migration-required, lossy, or unsupported result with source-preserving migration and explicit loss rules, required internal schema/round-trip checks, and hash-bound import/read-back receipts. External Common Cartridge validation and certification remain explicitly not run. -->
    • M1.17.f Create minimal/full/multilocale/accessibility/interactive/ protected-evidence goldens and validate independent round-trip. Evidence: canonical bundle suite. <!-- Evidence: six deterministic files under libs/contracts/src/metis/goldens/canonical-bundles/ cover the five named course profiles and protected assessment evidence; separate disk readers parse, reserialize, and rehash each file byte-for-byte without using the generator write path. -->
  • M1.18 Add machine-readable standards/version metadata to every exported package. Evidence: compatibility probe. <!-- Partial evidence: libs/contracts/src/metis/package-metadata.ts, fixtures/spec, ten canonical goldens under libs/contracts/src/metis/goldens/package-metadata/, and evidence/v1-workbenches/metis-package-metadata.json complete the target schema, native mappings, capability vocabulary, integrity binding, and compatibility probe. The parent remains open because runtime export/import adoption is not complete. -->
    • M1.18.a Define common package metadata for format/profile/spec version, conformance modules, extensions, producer/version, build/contract version, target, locale, generated time policy, and manifest digest. Evidence: metadata schema. <!-- Evidence: the strict common envelope pins current metadata, producer/build/contract/target/locale fields, deterministic UTC generation policy, exact conformance modules/extensions, and a manifest digest whose scope omits the injected metadata node to avoid circular hashing. -->
    • M1.18.b Map common metadata into QTI/LTI-deep-link/OneRoster/Caliper/ Open Badges/SCORM/xAPI/course-build/lecture package-native locations. Evidence: per-standard mapping. <!-- Evidence: ten registry rows pin the exact container, native path, encoding, binding mode, round-trip result, and caveat. OneRoster CSV is truthfully modeled as a lossy non-standard sidecar rather than an invented standard field; the other external bindings are namespaced metadata/extension or declared package-resource locations, SCORM sidecar loss is disclosed, and both Metis formats are native. -->
    • M1.18.c Declare supported/required/optional/extension/lossy/unsupported features and dependency/runtime versions machine-readably. Evidence: capability manifest tests. <!-- Evidence: all six capability states are schema-pinned, extension declarations and exact loss reasons are enforced, required and optional dependencies carry resolved versions, and fixtures/tests exercise every state. -->
    • M1.18.d Sign/hash metadata with the package and prevent divergence between human documentation, filename, manifest, and payload. Evidence: integrity tests. <!-- Evidence: the logical integrity profile recomputes canonical metadata, manifest, ordered payload-set, and native-binding hashes; their package digest binds the detached signature and derives the filename and exact human-readable projection. Twenty-five hostile mutations prove fail-closed divergence checks. Fixture signatures remain explicitly synthetic-unverified. -->
    • M1.18.e Build a pre-import compatibility probe returning compatible, migration-required, degraded/lossy, unsupported, invalid, or unknown with exact reasons. Evidence: probe tests. <!-- Evidence: deterministic probes cover every named outcome, validate current envelopes before negotiation, identify registered migrations, reject required feature/dependency gaps, disclose optional/extension/loss degradation, and return unknown rather than guessing unrecognized metadata or format versions. -->
    • M1.18.f Verify every exporter emits current metadata and every importer enforces it rather than guessing from content. Evidence: export coverage gate. <!-- Blocked by implementation: all ten audited runtime adapters still lack the common envelope on export and fail-closed compatibility enforcement on import. The evidence runtime audit names each surface; target fixtures are not runtime-adoption proof. --> <!-- SPEC 2026-09-11: the ten runtime adapters are named in the runtime audit inside evidence/v1-workbenches/metis-package-metadata.json. Each exporter emits the common envelope from libs/contracts/src/metis/package-metadata.ts; each importer refuses a package without it (fail-closed, typed reason); the coverage gate fails on any adapter not doing both. -->
  • M1.19 Define credential template, requirement, evidence, issuer, recipient projection, achievement, badge, endorsement, issuance, verification, expiry, revocation, and replacement contracts. Evidence: Open Badges mapping and authority/privacy invariants.
    • M1.19.a Define credential template/revision identity, title/description/ image, type, locale, issuer eligibility, achievement alignment, criteria, evidence policy, validity, and lifecycle. Evidence: template schema.
    • M1.19.b Define nested all/any/count/sequence/time/grade/mastery/course/ assessment/experience requirements with exact revisions and deterministic evaluation. Evidence: requirement tests.
    • M1.19.c Define evidence refs/types/issuer/source/subject/integrity/ verification/access/retention and minimize learner detail exposed to authoring or public verifiers. Evidence: evidence contracts.
    • M1.19.d Define issuer identity/authority/keys/status/endorsements plus privacy-preserving recipient identifier/proof and subject binding. Evidence: issuer/recipient tests.
    • M1.19.e Define achievement/badge/assertion/endorsement mapping to Open Badges with alignments, criteria, evidence, dates, status, signatures, hosted/signed forms, and extensions. Evidence: mapping goldens.
    • M1.19.f Define issuance request/evaluation/approval/signing/delivery/ receipt idempotency and prohibit course authors from self-granting issuer or recipient authority. Evidence: issuance tests.
    • M1.19.g Define verification for signature/hash/issuer/status/recipient/ evidence/expiry/revocation and privacy-safe public responses. Evidence: verifier tests.
    • M1.19.h Define expiry/renewal/revocation/replacement/supersession with reason, effective time, status lists/notifications, lineage, and no history rewrite. Evidence: lifecycle invariants. <!-- Completed by the versioned credential-lifecycle schema/evaluators, 4 canonical Open Badges mapping goldens, 34 hostile probes, and generated evidence artifact. These are synthetic target-contract evidence, not production runtime adoption, cryptographic proof, external validation, or Open Badges certification. The runtime audit remains false and names the legacy migration blockers. -->
  • M1.20 Define prerecorded lecture packages with script, segment, speaker or avatar, voice/likeness consent, slides/diagrams, demonstrations, captions, transcript, audio description, translations, media renditions, sources, checks, and playback manifest. Evidence: versioned package goldens.
    • M1.20.a Define lecture/package/logical/revision identity, course/lesson/ objective links, locale, duration/timebase, owner, lifecycle, target, and integrity manifest. Evidence: lecture schema.
    • M1.20.b Define versioned script and segment/chapter records with text, objectives, sources/citations, timing, speaker, visual/demo cues, and revision dependencies. Evidence: script/segment tests.
    • M1.20.c Define human/synthetic/avatar presenter identity, role, language, pronunciation, performance, voice/likeness/model/source consent scope, disclosure, expiry, and revocation. Evidence: presenter contracts.
    • M1.20.d Define slides/diagrams/charts/code/labs/demonstrations/media ingredients with exact source/data/code/renderer/rights/provenance and segment alignment. Evidence: visual ingredient tests.
    • M1.20.e Define transcript/caption/audio-description/sign-language/ translation variants with language, timing, speaker/non-speech cues, review, and source-revision linkage. Evidence: accessibility contracts.
    • M1.20.f Define video/audio/slide/text rendition profiles with codec/ dimensions/rate/color/loudness/bitrate/device/bandwidth/offline and quality checks. Evidence: rendition tests.
    • M1.20.g Define correctness/source, safety, consent/rights, accessibility, sync, technical, target, and reviewer evidence required per package. Evidence: check registry.
    • M1.20.h Define playback manifest for chapters/tracks/renditions/ captions/descriptions/translations/download/streaming/fallback with exact hashes and no hidden authority. Evidence: playback tests.
    • M1.20.i Produce deterministic minimal/full/multilingual/accessible/ avatar/demo goldens and validate package/playback round-trip. Evidence: lecture package suite. <!-- Completed by the immutable lecture-package target schema/evaluator, 6 canonical package/playback goldens, 30 named failure classes, source decisions, runtime audit, and generated evidence artifact. This is synthetic target-contract evidence, not runtime adoption, external media/C2PA/HLS/consent validation, WCAG conformance, or an operational-readiness claim. -->
  • M1.21 Define tutoring configuration for persona, pedagogy, hint ladder, source set, course/item revision, learner-context projection, live voice/text mode, tools, escalation, safety, integrity mode, and session retention. Evidence: authoring settings cannot grant the tutor forbidden learner or answer-key access.
    • M1.21.a Define immutable tutor configuration/revision identity, owner/ approver, applicable course/lesson/objective/item/locale/audience, lifecycle, and code/institution floors. Evidence: config schema.
    • M1.21.b Define bounded persona/tone/language/disclosure and versioned pedagogy/strategy/scaffolding rules without impersonation or hidden human claims. Evidence: persona/pedagogy tests.
    • M1.21.c Define hint ladder stages, eligibility, attempt/timing triggers, source/rationale boundaries, penalties, exhaustion, and protected-answer nonexposure. Evidence: hint tests.
    • M1.21.d Define authorized source collection/course/item revision set, retrieval/citation policy, update/retraction invalidation, and fail- closed grounding behavior. Evidence: source-set tests.
    • M1.21.e Define minimized learner-context projection fields/purpose/ freshness/retention and prohibit raw profile/grade/integrity/diagnosis/ answer-key access. Evidence: context property tests.
    • M1.21.f Define text/live voice input-output, speaker/voice consent, transcript/caption, interruption, latency, recording, disclosure, and degraded-mode behavior. Evidence: modality tests.
    • M1.21.g Define allowlisted versioned tools, typed grants, resource/ network/data limits, confirmation, outputs, audit, and no configuration- granted privilege. Evidence: tool authorization tests.
    • M1.21.h Define safety, child/privacy, self-harm/escalation, academic- integrity modes per assessment state, refusal, educator handoff, and incident paths. Evidence: policy matrix.
    • M1.21.i Define session duration/context/history/summary/retention/ deletion/export/analytics and exact configuration/course/model versions. Evidence: session lifecycle tests. <!-- Completed by the immutable tutoring-configuration target schema, deny-by-default learner/content/tool authorization evaluator, external typed runtime-grant boundary, governed hint evaluator, 3 session-lifecycle goldens, 32 named failure classes, source decisions, runtime audit, and generated evidence artifact. This is synthetic target-contract evidence, not runtime adoption, external policy/consent/grant validation, crisis integration, deletion execution, WCAG conformance, or an operational-readiness claim. -->
  • M1.22 Define learner-profile and preference projections used by preview, adaptation, accessibility, and tutor simulation without copying protected learner records into authoring storage. Evidence: minimization and purpose-boundary tests.
    • M1.22.a Inventory candidate learner/profile/preference/PNP/mastery/ progress/grade/integrity fields, canonical owners, classifications, and prohibited authoring uses. Evidence: field inventory.
    • M1.22.b Define separate purpose-bound projections for anonymous author preview, authorized adaptation, accessibility delivery, tutor session, and synthetic simulation. Evidence: projection schemas.
    • M1.22.c Use opaque subject/profile/simulation identities and expose only minimum locale, support capability, prior-knowledge band, preferences, and contextual state required for each purpose. Evidence: minimization tests.
    • M1.22.d Represent unknown/withheld/not-applicable/stale values explicitly and prevent inference of diagnoses, protected traits, grade, or integrity status. Evidence: inference tests.
    • M1.22.e Authorize issuer/consumer/purpose/tenant/course/session, bind source/freshness/expiry, prevent caching/authoring persistence, and audit access. Evidence: purpose-boundary tests.
    • M1.22.f Define correction/revocation/deletion/session-end invalidation and prove projections disappear from temporary stores/logs/prompts/exports. Evidence: lifecycle tests.
    • M1.22.g Build synthetic profile fixtures for authoring simulations so real learners are never required for general preview/testing. Evidence: simulation isolation tests. <!-- M1.22 completion evidence: learner-profile-projections.ts and deterministic generated evidence inventory 42 candidate fields, define five separately scoped projection profiles, preserve all unavailable-value states, deny 17 purpose/storage/inference boundary cases, sweep six temporary locations for five lifecycle triggers, and provide three hand-authored synthetic profiles. This is synthetic target-contract evidence, not runtime adoption, operational authorization, production erasure execution, FERPA/COPPA compliance, AccessForAll conformance, or legal advice. -->

M2 — Source ingest, grounding, and curriculum planning#

  • M2.1 Implement source intake from upload, URL, feed, LMS, repository, DOI, and manual record. Evidence: each path records provenance and checksum.
    • M2.1.a Define immutable source/intake/version/provenance/checksum, connector, authority, rights, classification, and lifecycle contracts. Evidence: intake schemas.
    • M2.1.b Implement bounded multipart upload with content sniffing, hash- while-streaming, temporary quarantine, and tenant isolation. Evidence: upload tests.
    • M2.1.c Implement allowlisted URL/feed retrieval with SSRF/redirect/DNS/ size/type/time protections and immutable retrieval receipt. Evidence: network intake tests.
    • M2.1.d Implement LMS/repository connectors with scoped credentials, stable external IDs/versions, pagination/rate limits, and deletion/ update semantics. Evidence: connector tests.
    • M2.1.e Resolve DOI metadata/content through approved registries with authoritative identifier, edition/version, access/rights, and no false full-text assumption. Evidence: DOI fixtures.
    • M2.1.f Implement manual bibliographic/source record with required provenance/evidence and explicit no-bytes/unchecked status. Evidence: manual intake tests.
    • M2.1.g Normalize all paths to one source-version record with original locator, retrieval time, content hash, connector/tool versions, and audit. Evidence: cross-path parity tests. <!-- M2.1 completion evidence: SourceVersionRecord and the Metis source-intake runtime normalize upload, URL, feed, LMS, repository, DOI, and manual-record paths into one immutable, tenant-scoped provenance/checksum/audit contract. Focused API, streaming, SSRF, connector, DOI-fixture, manual-record, parity, OpenAPI, and generated-client checks provide runtime implementation evidence; M2.2 scan admission is separately evidenced and production deployment remains an operational gate. -->
  • M2.2 Quarantine and scan source uploads. Evidence: malicious document, archive, macro, and oversize fixtures.
    • M2.2.a Store new bytes in nonpublic nonexecutable quarantine with opaque names, tenant scope, encryption, limits, and no downstream parser access. Evidence: quarantine tests.
    • M2.2.b Scan malware/signatures, content type/polyglot, macros/scripts/ active content, archive traversal/symlinks/nesting/bombs, and parser exploit indicators. Evidence: malicious fixture suite.
    • M2.2.c Enforce compressed/uncompressed size, file/count/depth, page/ object/media, CPU/time/memory, and password/encryption rules. Evidence: boundary tests.
    • M2.2.d Record scanner/database/version/config/hash/results and classify clean/quarantined/rejected/manual/unavailable without false clean state. Evidence: result schema tests.
    • M2.2.e Admit exact scanned hash only after all required scans; rescan on scanner/policy/freshness change and retain prior evidence. Evidence: admission/race tests. <!-- M2.2 completion evidence: encrypted AES-256-GCM tenant quarantine, bounded built-in structural inspection, a versioned fail-closed ClamAV command adapter, immutable self-hashed scan/admission receipts, exact-hash parser gating, superseding-rescan revocation, and malicious document/archive/macro/oversize fixtures are covered by the Metis runtime and generated evidence inventory. This is runtime implementation evidence, not a security certification; production key custody and live ClamAV database delivery remain deployment gates. -->
  • M2.3 Extract text, structure, tables, equations, media, and metadata with page/section anchors. Evidence: golden corpus.
    • M2.3.a Define extraction document/node/anchor/confidence/error contracts with exact source hash and extractor/config versions. Evidence: extraction schema.
    • M2.3.b Extract ordered text, headings/sections/lists/footnotes/citations/ reading order with page/region/character anchors. Evidence: structure goldens.
    • M2.3.c Extract tables with cells/spans/headers/captions/units/footnotes and page regions; preserve image fallback. Evidence: table goldens.
    • M2.3.d Extract equations as source/math representation plus visible text/ image, labels, variables, and page regions without inventing semantics. Evidence: equation goldens.
    • M2.3.e Extract embedded images/audio/video/files with hashes, captions/ alt text, time/page anchors, technical metadata, and source rights. Evidence: media goldens.
    • M2.3.f Extract bibliographic/document metadata with source-versus-derived provenance, normalization, confidence, and conflicts. Evidence: metadata fixtures.
    • M2.3.g Test multilingual/RTL/scanned/OCR/complex layout/malformed/ password/partial cases against licensed golden corpus. Evidence: extraction quality report. <!-- M2.3 completion evidence: the source-bound 2.3 extraction graph, typed W3C-aligned character/page/region/time selectors, bounded Markdown/HTML/PDF/DOCX/image/archive/media readers, Tesseract TSV seam, explicit table/equation/media/metadata provenance, and the synthetic CC0 golden corpus are exercised by runtime, OpenAPI, generated-client, and evidence checks. This is implementation and deterministic corpus evidence, not an extraction-quality certification; production parser and OCR binaries/language packs remain deployment gates, alongside representative institution-owned corpus calibration. -->
  • M2.4 Preserve original bytes and extraction version. Evidence: re-extraction does not rewrite historical evidence.
    • M2.4.a Store original bytes immutably by content hash with source-version manifest, encryption, access, retention, and legal-hold metadata. Evidence: storage integrity tests.
    • M2.4.b Define extraction-run identity with exact source hash, parser/OCR/ model/config/runtime versions, status, outputs, errors, and timestamps. Evidence: extraction run schema.
    • M2.4.c Persist extracted artifacts/anchors under the run identity and forbid in-place correction or replacement. Evidence: immutability tests.
    • M2.4.d Re-extract as a new run, compare/diff outputs, migrate accepted anchors explicitly, and retain consumers of old evidence. Evidence: re-extraction tests.
    • M2.4.e Verify stored source/extraction hashes on read/reuse and quarantine corruption without rewriting history. Evidence: corruption tests. <!-- M2.4 completion evidence: tenant-partitioned content-addressed originals use authenticated AES-256-GCM envelopes and immutable source-version manifests with retention, legal-hold, access, and disposition policy. Exact source/extractor/runtime/config/model-use identities, outputs, bindings, errors, retries, parents, supersession, comparisons, selection transitions, accepted anchor migrations, and corruption quarantine receipts are append-only and hash-verified on reuse. Deterministic/concurrent retry, historical re-extraction, partial/failed, refresh invalidation, tenant access, and tamper tests cover the local runtime. Production KMS/HSM custody and replicated object-lock storage remain deployment gates, alongside backup/restore exercises, approved retention schedules, and operational legal-hold procedures. -->
  • M2.5 Implement source authority, license, currency, jurisdiction, and audience metadata. Evidence: validation and review UI.
    • M2.5.a Define publisher/author/institution/issuer authority identity, evidence, role, credentials, reputation limits, and verification state. Evidence: authority schema.
    • M2.5.b Define source/license/version/owner, access/use/reproduce/adapt/ teach/train/distribute rights, attribution, territory, term, and evidence. Evidence: license schema.
    • M2.5.c Define published/updated/retrieved/reviewed/effective/expired dates, supersession/retraction, subject volatility, and currency review interval. Evidence: currency tests.
    • M2.5.d Define jurisdiction/geographic scope/applicability/conflicts and intended learner age/grade/language/prerequisite/accessibility audience. Evidence: jurisdiction/audience schemas.
    • M2.5.e Validate completeness/compatibility/freshness and present raw metadata/evidence/conflicts/reviewer decision separately. Evidence: validation/review UI tests. <!-- Implemented in the source-governance 2.5.0 contracts, deterministic validation helpers, append-only source-version-bound governance service, tenant and admin APIs, generated OpenAPI client, and the Metis admin Source Evidence Review workspace. Authority identity, credentials, evidence, explicitly bounded reputation, the seven rights actions, attribution/territory/term, currency lifecycle, jurisdiction conflicts, BCP-47-shaped audience language, learner age/grade/prerequisite/accessibility, raw declarations, evidence, conflicts, live findings, and authenticated decisions remain structurally separate. Python service/API/tamper/idempotency/ tenant tests, admin component tests, responsive Playwright flows, and automatic WCAG analysis cover the local runtime; see evidence/v1-workbenches/metis-source-governance.json. Production registry integrations and legal license clearance remain deployment and governance gates, together with ratified institution-specific jurisdiction/audience policy, qualified reviewer operations, replicated durable storage, and backup/ restore exercises. -->
  • M2.6 Implement claim extraction behind the fenced provider seam. Evidence: proposals require human confirmation.
    • M2.6.a Define claim proposal with exact source/extraction revision, text/normalized proposition, type/scope/qualifiers, locators, confidence, provider/model/prompt/config version, and status. Evidence: schema tests.
    • M2.6.b Send only authorized bounded source passages/instructions through the injected provider, fenced as untrusted data with no tools/authority. Evidence: injection/data-minimization tests.
    • M2.6.c Validate provider schema/locators/quotes/size, reject unsupported source assertions, and persist refusal/malformed/unavailable honestly. Evidence: provider seam tests.
    • M2.6.d Present proposals with source context, uncertainty, duplicates/ conflicts, and accept/edit/reject/defer actions. Evidence: review tests. <!-- Implemented in the source-claims 2.6.1 contracts, exact-extraction- bound append-only service, injected OpenAI-compatible provider protocol, tenant and admin APIs, generated OpenAPI client, and the Metis admin Claim Confirmation workspace. Only current, governance-approved source passages with explicit access/use/ reproduce/adapt permission are selected by exact node character range. Source and instructor inputs are rendered as separately labeled untrusted non-instruction data; the provider receives no tool surface or authority. Quote offsets, substrings, schema, relationships, numeric/lexical support, and a two-MiB output bound fail closed, while refused, malformed, unavailable, and completed runs remain immutable and visible without partial proposals. Explicit administrator accept/edit decisions atomically create hash-linked canonical revisions that preserve the exact proposal, provider invocation, extraction/run hashes, and authority decision; reject/defer, unauthorized, or stale decisions cannot create a revision, and later withdrawal never rewrites revision history. Python provider/service/API/tamper/idempotency/tenant tests, admin component tests, three-browser Playwright edit/stale/mobile flows, and automatic WCAG analysis cover the local runtime; see evidence/v1-workbenches/metis-source-claims.json. Production provider credentials and availability remain deployment gates, together with representative-corpus quality calibration, adversarial red-team exercises, qualified reviewer operations, legal source clearance, replicated durable storage, backup/ restore exercises, and production monitoring. -->
    • M2.6.e Create canonical claim revisions only from explicit authorized human decisions; preserve original proposal/model provenance. Evidence: authority boundary tests.
  • M2.7 Implement citation attachment with page, section, paragraph, timecode, dataset slice, or URI fragment locators. Evidence: locator validation fixtures.
    • M2.7.a Define citation identity/version, claim/content target, exact source/extraction revision, locator kind/value, quote/context, role, author, and verification state. Evidence: citation schema.
    • M2.7.b Implement page/region, heading/section, paragraph/character, media timecode/range, dataset row/column/filter/version, and URI fragment locators as typed variants. Evidence: locator fixtures.
    • M2.7.c Validate locator syntax/bounds/source-kind/revision and resolve to exact content without mutable URL/page assumptions. Evidence: validation tests.
    • M2.7.d Detect source/extraction changes that make locators exact, shifted, ambiguous, orphaned, or invalid and require re-review. Evidence: anchor survival tests.
    • M2.7.e Render/export accessible citation context and preserve rights/ excerpt limits without exposing protected sources. Evidence: citation UI/ export tests. <!-- Implemented in the source-citations 2.7.0 discriminated contracts, offline exact-extraction resolver, append-only citation service, tenant and admin APIs, generated OpenAPI client, and the Claim Confirmation citation provenance ledger. Citation versions bind claim/content targets to exact extraction outputs, source-version records, original-locator hashes, typed locators, resolved anchors, quote/context, roles, authenticated authors, review state, and rights snapshots. Page/normalized-region, section/path, paragraph/half-open character, media/half-open millisecond, dataset row/column/filter/version, and RFC 3986 fragment fixtures resolve only inside immutable extraction output without network dereference or mutable page assumptions. Selected extraction, source, claim-authority, and rights changes reclassify exact, shifted, ambiguous, orphaned, or invalid locators and require an append-only reviewed reattachment. Exports apply the stricter of snapshotted/current rights, cap eligible excerpts at 500 characters, and redact protected context; every caller-facing attachment, reattachment, package, and admin workspace response projection omits stored quote, surrounding context, locator, and rights objects entirely. Strict Python schema/resolver/service/API/tenant/idempotency/tamper tests, OpenAPI drift checks, admin typecheck/lint/component tests, and three-browser Playwright attachment/redaction/mobile journeys with automatic WCAG analysis cover the local runtime; see evidence/v1-workbenches/metis-source-citations.json. Production deployment and legal rights clearance remain gates, together with representative-corpus anchor-survival measurement, qualified reviewer operations, replicated durable storage, backup/restore, monitoring, and re-review service levels. -->
  • M2.8 Implement high-risk severity floors for health, legal, safety, and other ratified domains. Evidence: authoritative-source requirements.
    • M2.8.a Define versioned risk-domain registry, severity levels, claim/use/ audience applicability, and nonwaivable code floors. Evidence: risk schema.
    • M2.8.b Define minimum source authority, recency, jurisdiction, corroboration, citation, reviewer qualification, disclaimer/escalation, and update monitoring per severity. Evidence: requirement matrix.
    • M2.8.c Classify curriculum claims/lessons/items/tutor settings using typed evidence and route ambiguous/high severity to human review. Evidence: classification tests.
    • M2.8.d Block planning/publication/tutor use when authoritative-source or qualified-review requirements are missing/stale/conflicting. Evidence: fail-closed tests.
    • M2.8.e Test institution overlays can tighten but never loosen floors and that source withdrawal propagates. Evidence: monotonicity/lifecycle tests. <!-- Implemented in the source-risk-policy 2.8.0 contracts, deterministic self-hashed domain registry, immutable tenant-scoped assessment and review service, administrator decision workspace, generated OpenAPI client, and planning/generation/publication/tutoring read-back gates. Health, legal, safety, and financial policies each define general, elevated, high, and critical floors across exact target, use, audience, and consequence applicability. Current governed source authority and age, exact jurisdiction, independently governed corroboration, verified target-bound citations, qualified review, disclaimer, escalation, monitoring, withdrawal, and critical tutor prohibition are evaluated independently per requested use. Claims, lessons, assessment items, tutor settings, and curriculum scope bind exact revisions or hashes; ambiguity and high severity require a qualified evidence-hash-bound human decision. New institution overlay versions may only tighten both code floors and their own previous version. Current evidence is re-derived at each downstream boundary, so changed overlays, stale review hashes, source withdrawal, or target/citation drift block use without rewriting history. Python schema/helper/service/API/idempotency/tamper/tenant/lifecycle tests, OpenAPI drift checks, admin typecheck/lint/component tests, and responsive Playwright decision/review journeys with automatic WCAG analysis cover the local runtime; see evidence/v1-workbenches/metis-source-risk-policy.json. The recorded WHO, NIST, GovInfo, OSHA, CFPB, and UNESCO sources inform a conservative engineering baseline rather than jurisdictional, clinical, legal, safety, or financial ratification. Production ratification, qualified reviewer staffing, representative-corpus calibration, durable replicated storage, backup/restore, monitoring operations, and deployment validation remain explicit gates. -->
  • M2.9 Implement source update/retraction blast radius across lessons, items, feedback, tutor prompts, and published courses. Evidence: known-value propagation test.
    • M2.9.a Maintain forward/reverse versioned dependency edges from source/ extraction/claim/citation into all named consumers and release bundles. Evidence: dependency graph tests.
    • M2.9.b Ingest verified update/supersession/retraction/rights change with effective time, reason, authority, scope, and source identity matching. Evidence: change intake tests.
    • M2.9.c Traverse direct/transitive impacts with stable pagination, reason/path/current revision/status, unknown dependencies, and tenant/ property authorization. Evidence: blast-radius fixtures.
    • M2.9.d Mark affected draft/review/published/tutor/delivery eligibility stale or blocked according to risk policy without deleting history. Evidence: propagation tests.
    • M2.9.e Route owners to re-source/revise/review/rebuild/unpublish and verify all downstream states/read-backs. Evidence: known-value end-to-end test. <!-- Implemented in the source-blast-radius 2.9.0 contracts and service, with immutable content-addressed graph snapshots, forward/reverse versioned source-to-consumer edges, explicit unknown dependencies, exact source-version identity checks, verified authority evidence, append-only hash-chained change events, deterministic shortest-path traversal, signed snapshot-bound pagination, tenant/property authorization, and immutable remediation receipts. Effective updates, supersessions, retractions, and rights changes mark the exact current lesson/item/feedback/tutor/publication/delivery revisions stale or blocked; planning, generation, promotion, publication, tutoring, and delivery re-check unresolved impacts without rewriting historical artifacts. The administrator lineage ledger exposes exact paths, owners, authority evidence, snapshot provenance, and re-source/revise/review/rebuild/unpublish routing. A known-value retraction fixture proves the complete source -> extraction -> claim -> citation -> lesson path, every named consumer, stable pagination and tamper rejection, access isolation, all downstream gates, unpublish safety, and unchanged historical bytes. OpenAPI/client drift checks, Python integration tests, admin typecheck/lint/component coverage, and responsive accessible Playwright journeys cover the local runtime; see evidence/v1-workbenches/metis-source-blast-radius.json. W3C PROV-O, DataCite metadata, and Crossref Crossmark inform the provenance and lifecycle model. Production publisher/webhook integrations, durable replicated storage, alert delivery, operator staffing, and deployment validation remain explicit gates. -->
  • M2.10 Implement curriculum brief capture with audience, prerequisites, outcomes, constraints, standards, duration, modality, locale, and assessment strategy. Evidence: required-field gate.
    • M2.10.a Define versioned brief identity/scope/owner/course/program, intended learners/grade/age/context, enrollment scale, and source refs. Evidence: brief schema.
    • M2.10.b Capture prerequisite knowledge/skills/evidence and target learning outcomes/objectives with measurable criteria. Evidence: prerequisite/outcome tests.
    • M2.10.c Capture standards/framework/version/coverage, institution/region/ safety/accessibility/rights/technology/budget/schedule constraints. Evidence: constraint tests.
    • M2.10.d Capture total/calendar/seat/self-study duration, synchronous/ asynchronous/in-person/online/blended modality, locale/language, device/ bandwidth/offline requirements. Evidence: delivery tests.
    • M2.10.e Capture formative/summative/diagnostic/authentic assessment strategy, weighting/frequency/security/accommodation/feedback constraints. Evidence: assessment strategy tests.
    • M2.10.f Validate required/conditional fields, conflicts, units, compatibility, code/institution floors, and reviewer acceptance before planning. Evidence: required-field gate. <!-- Implemented in the curriculum-brief 2.10.0 contracts and service, with exact course/program/owner and source-version identities, append-only self-hashed revisions and reviews, deterministic validation evidence, a code floor, monotonic tightening institution floors, idempotent author/reviewer mutations, tenant authorization, and tamper/chain read-back checks. Draft revisions may retain policy-floor findings, while submission, administrator acceptance, and source-outline planning fail closed on required/conditional fields, conflicts, units, modality compatibility, current policy, exact current revision hashes, exact validation hashes, and source package binding. The instructor/admin dossier exposes four progressively disclosed stages, a live institution-aware readiness ledger, revision register, and exact reviewer hash read-back. Python service/API/OpenAPI tests, generated-client drift checks, web validation tests, production build verification, and responsive axe-backed Playwright journeys across Chromium, Firefox, and WebKit cover the local runtime; see evidence/v1-workbenches/metis-curriculum-brief.json. 1EdTech CASE 1.1, WCAG 2.2, RFC 5646/BCP 47, and CEDS inform the standards, accessibility, locale, and education-domain boundaries. Production identity federation, durable replicated storage, institution policy ratification, reviewer staffing, backup/restore, monitoring, and deployment validation remain explicit operational gates. -->
  • M2.11 Implement objective and concept-map authoring. Evidence: cycle, orphan, unreachable objective, and prerequisite checks.
    • M2.11.a Build typed objective create/edit/version/link UI/API with taxonomy, criteria, standards, evidence, ownership, and validation. Evidence: objective authoring tests.
    • M2.11.b Define concept node identity/type/version/source/locale and prerequisite/part-of/example/contrast/causal/related relationship kinds. Evidence: concept graph schema.
    • M2.11.c Build accessible graph plus synchronized tree/list/table operations for add/link/unlink/reorder/group/annotate/import. Evidence: semantic UI tests.
    • M2.11.d Detect cycles where forbidden, self/duplicate/conflicting edges, orphan objectives/concepts, missing roots, and unreachable outcomes. Evidence: graph validation fixtures.
    • M2.11.e Validate prerequisite ordering against modules/lessons/ assessments and surface remediation without automatic destructive edits. Evidence: sequencing tests. <!-- Implemented in the curriculum objective-map 2.11.0 contracts, service, API, and instructor workbench. Objective and concept identities carry typed ownership, locale, taxonomy, criteria, standards, evidence, source, and embedded immutable revision chains inside append-only self-hashed map snapshots bound to one exact accepted curriculum-brief revision. Deterministic validation fails closed for validated maps and planning readiness while draft revisions may retain explicit cycle, self/duplicate/conflicting edge, orphan, missing-root, unreachable, parent-structure, duplicate-position, dangling-placement, and prerequisite-order findings with human-readable remediation and no automatic graph or sequence rewrite. The warm-paper workbench synchronizes one selection across a visual SVG graph, keyboard tree, semantic list, and native table, and exposes reversible add/edit/link/unlink/ reorder/group/annotate/import operations plus exact validation and revision hashes. Python service/API/OpenAPI tests, generated-client drift checks, web contract tests, production build verification, and responsive axe-backed Playwright journeys across Chromium, Firefox, and WebKit cover the local runtime; see evidence/v1-workbenches/metis-curriculum-map.json. 1EdTech CASE 1.1, W3C SKOS, and WAI-ARIA APG tree and keyboard patterns inform the standards and interaction boundaries; this is runtime evidence, not CASE/SKOS conformance or assistive-technology certification. Production identity federation, replicated durable storage and backup, institution policy ratification, representative graph-scale performance and assistive-technology testing, monitoring, and deployment validation remain explicit operational gates. -->
  • M2.12 Implement standards-alignment suggestions behind the provider seam. Evidence: human acceptance and source/version attribution.
    • M2.12.a Define alignment proposal for exact objective/content/item and standard framework/version/code, strength/coverage, rationale, source locators, confidence, and provider provenance. Evidence: schema tests.
    • M2.12.b Retrieve only authorized current standard text/metadata and send bounded inputs through the fenced provider with injection resistance. Evidence: provider input tests.
    • M2.12.c Validate proposed codes/versions/locators against the canonical standards registry and deduplicate/conflict-check. Evidence: validation tests.
    • M2.12.d Present source text/diff/confidence/limitations and require human accept/edit/reject with reviewer authority. Evidence: review journey.
    • M2.12.e Persist accepted alignment separately from proposal and invalidate/review on objective or standard revision change. Evidence: lineage tests. <!-- Implemented in the curriculum standards-alignment 2.12.0 contracts, canonical registry, fenced provider seam, append-only service, API, generated client, and instructor/admin review workbench. Exact objective/content/item snapshots and canonical records retain resource/revision/content/text hashes, framework/version/code/locator, strength, coverage, rationale, confidence, limitations, and provider/prompt/config provenance. Only current, explicitly authorized records can cross the zero-tool trust boundary; target, registry, and reviewer inputs are bounded, data-only fenced, and rejected rather than truncated, while every proposed identity is validated against the selected canonical set and checked for duplicates and conflicts. Named administrators must inspect exact source and target evidence and accept, edit, or reject with hash readback; accepted alignments are immutable records separate from model proposals and decisions, with fail-closed review-required lineage when objectives disappear or target/standard revision, content, lifecycle, or authorization changes. Python service/API/OpenAPI tests, generated-client drift checks, web contract tests, production build verification, and responsive axe-backed Playwright journeys across Chromium, Firefox, and WebKit cover the local runtime; see evidence/v1-workbenches/metis-standards-alignment.json. 1EdTech CASE 1.1, OWASP LLM01:2025, NIST AI RMF, and WAI-ARIA APG tabs and keyboard practices inform the contract, trust, authority, and interaction boundaries; this is runtime evidence, not external standards conformance, security certification, or deployment validation. Seeded records are repository-authored normalized metadata abstracts versioned by the exact Metis catalog; they do not reproduce full upstream standards or claim an upstream framework edition, which reviewers must verify at the linked authoritative source. Production provider credentials and allowlisting, institution-owned registry ingestion/versioning and rights review, federated identity and reviewer staffing, replicated durable storage and backup, monitoring, representative accessibility testing, and deployment validation remain explicit operational gates. -->
  • M2.13 Implement course-outline divergence. Evidence: N variants preserve generated structure and score provenance.
    • M2.13.a Define outline variant/set contracts with stable module/lesson/ objective structure IDs, source brief/revisions, generation plan, and provider provenance. Evidence: outline schemas.
    • M2.13.b Generate N bounded variants with explicit diversity strategy, seeds/controls, budget/cost confirmation, and unique identities. Evidence: divergence tests.
    • M2.13.c Validate every variant for hierarchy, objective/prerequisite/ standards/coverage/workload constraints before comparison. Evidence: outline validation tests.
    • M2.13.d Record provider/evaluator/human scores with rubric/version/ uncertainty separately; manual variants show absent generated scores. Evidence: score provenance tests.
    • M2.13.e Compare/select/reject/merge/fork variants through decision records while retaining all original structures. Evidence: decision journey. <!-- Implemented in the course-outline divergence 2.13.0 contracts, zero-tool provider seam, deterministic validator, append-only service, API, generated client, and instructor workbench. Every set binds one exact planning-ready curriculum-brief revision, accepted brief review, objective-map revision, validation hashes, source objective/prerequisite/standard identities, generation plan, and confirmed cost ceiling. Requests require two through six candidates with explicit diversity strategies, unique seeds, workload and structure bounds, input/output limits, and pricing provenance; provider results must match the exact plan and count, remain inside the confirmed cost envelope, and retain prompt, model, config, trust-label, token, outcome, and response hashes. Source, plan, and reviewer guidance cross a bounded data-only prompt fence with no tool access and fail closed on oversize, refusal, malformed output, count/identity drift, or storage tampering. Stable variant/module/lesson/objective structure IDs preserve each generated shape. Deterministic validation records hierarchy and sequence limits, lesson and objective prerequisite ordering, exact objective and standards references, required objective/standard/assessment coverage, and workload bounds before comparison, selection, merge, or fork. Provider, evaluator, and named human scores are separate immutable records with distinct lineage, rubric/version, dimensions, rationale, and uncertainty; manual and derived structures inherit no generated scores. Optimistic content, validation, and workspace hashes bind select/reject/merge/fork decisions while all originals and exact source contributions remain readable. The warm-paper blueprint workbench exposes the complete decision journey through a semantic captioned comparison table in its own labelled scroll region, roving keyboard tabs, narrow document reflow, and reduced motion. Python service/API/OpenAPI tests, generated-client drift checks, web contract tests, production build verification, and responsive axe-backed Playwright journeys across Chromium, Firefox, and WebKit cover the local runtime; see evidence/v1-workbenches/metis-course-outline.json. 1EdTech Common Cartridge, NIST AI RMF, OWASP LLM01:2025, and W3C table/reflow guidance inform the identity, measurement, trust, and interaction boundaries; this is runtime evidence, not Common Cartridge or WCAG conformance, security certification, or deployment validation. Production provider credentials and allowlisting, institution policy and rubric governance, federated identity and reviewer staffing, replicated durable storage and backup, monitoring, representative accessibility and scale testing, and deployment validation remain explicit operational gates. -->
  • M2.14 Implement workload and seat-time estimation. Evidence: deterministic calculation with modality-specific assumptions.
    • M2.14.a Define versioned effort model for reading/listening/viewing, instruction, discussion, practice, assessment, project/lab/simulation, feedback, and navigation with units/assumptions. Evidence: model schema.
    • M2.14.b Derive content/activity quantities from exact outline/course revision and apply modality, difficulty, learner-band, language, accessibility, and synchronous/self-paced factors. Evidence: calculation fixtures.
    • M2.14.c Calculate learner effort, instructor/facilitator workload, seat time, independent time, calendar duration, range/uncertainty, and bottlenecks separately. Evidence: estimate tests.
    • M2.14.d Expose per-component inputs/formulas/assumptions/overrides/source and never hide unknown work as zero. Evidence: explainability tests.
    • M2.14.e Compare estimate to brief/institution constraints and require authorized reasoned overrides without changing raw calculations. Evidence: constraint tests. <!-- M2.14 evidence: schema version 2.14.0 and effort-model version 2026.08.1 define 18 modality/unit rates across nine component families, each with an explicit unit, lower/expected/upper rate, scaling rule, source references, and assumptions. Requests bind an immutable accepted-brief revision plus either exact outline identity/content hashes or a captured, self-hashed course revision whose text witnesses preserve derived word counts; media durations and other quantities not present in the source remain unknown, never inferred from URLs or converted to zero. Decimal interval arithmetic applies explicit difficulty, learner-band, language, accessibility, and delivery factors. Accessibility factors describe only documented accommodation, format, and support-service work rather than treating disability as learner slowness. Results separate learner effort, enrollment-aware instructor/facilitator work, seat time, independent time, calendar duration, uncertainty ranges, and bottlenecks; incomplete totals are labelled known subtotals with unknown component counts. The warm-paper engineering ledger exposes raw and adjusted component formulas, quantities, rates, factors, assumptions, sources, constraints, and immutable authorized reasoned overrides without mutating raw arithmetic. Tenant and role gates, idempotency, artifact tamper checks, course-revision capture, OpenAPI/client contracts, focused and full Python service tests, web unit tests, lint, typechecking, production build, and axe-backed responsive Playwright journeys across Chromium, Firefox, and WebKit cover the local runtime; see evidence/v1-workbenches/metis-workload-estimation.json. 34 CFR 600.2, the Rice Course Workload Estimator, NIST TN 1297, and W3C reflow guidance inform the credit-hour vocabulary, transparent assumptions, range reporting, and interaction design. This is inspectable planning evidence, not a credit-hour or clock-hour determination, statistical confidence interval, labor rule, accessibility/WCAG conformance claim, or deployment validation. Institution-approved rates and policy rules, production identity and storage, representative learner validation, governance, monitoring, and deployment testing remain explicit operational gates. -->
  • M2.15 Implement objective-to-content-to-assessment coverage matrices. Evidence: undercoverage and overassessment gates.
    • M2.15.a Build exact revision links from objectives to modules/lessons/ blocks/activities/practice/items/assessments/rubrics and evidence. Evidence: coverage graph tests.
    • M2.15.b Define instructional/introduced/practiced/reinforced/assessed/ mastered coverage roles, depth/cognitive level, weight, and required thresholds. Evidence: coverage vocabulary.
    • M2.15.c Compute accessible objective-by-content-by-assessment matrices with direct/indirect, missing, stale, unsupported, and duplicated links. Evidence: matrix fixtures.
    • M2.15.d Detect required objective undercoverage, content without objective, assessment without instruction, excessive weight/frequency, level mismatch, and overassessment. Evidence: gate tests.
    • M2.15.e Present raw links/measures and remediation separately from policy verdict; recompute on revision/threshold changes. Evidence: UI/API tests. <!-- M2.15 evidence: schema version 2.15.0 and coverage-policy version 2026.08.1 bind immutable curriculum-map and outline revisions plus their hashes to self-hashed objective, module, lesson, block, activity, practice, item, assessment, rubric, evidence, and relationship snapshots. Each link preserves exact endpoint revisions/hashes, direct or indirect relation, one of six versioned instructional/introduced/practiced/reinforced/ assessed/mastered roles, cognitive level, weight, frequency, rationale, and evidence revisions. Deterministic raw measurement retains every link and the complete objective-by-resource Cartesian matrix, including empty, stale, unsupported, and duplicated relationships; policy credit excludes stale/unsupported evidence and deduplicates repeated objective/resource/ role links without hiding the raw records. Institution-sourced effective thresholds drive explicit source-traced blocking and warning findings for required-objective undercoverage, content without objectives, assessment without instruction, excessive weight or frequency, cognitive-level mismatch, overassessment, stale links, unsupported links, and duplicates. Source facts, raw measures, policy verdicts, and non-automatic remediation suggestions are separate immutable layers. Exact source or governed threshold changes create a linear, self-hashed recomputation successor; no-op, stale-workspace, divergent-successor, tenant, role, idempotency, and tamper checks fail closed. The warm-paper audit workbench splits content, practice, and assessment into semantic captioned matrices with scoped headers and labelled internal scroll regions, supplies a linear objective register and raw-link inspector, preserves raw/policy selection, reflows at 390 px, and honors reduced motion. Ruff, Mypy, 644 Python service tests (10 environment-dependent skips), OpenAPI and generated-client drift checks, 113 web unit tests, lint, typecheck, production build, 10 evidence tests, and axe-backed Playwright journeys across Chromium, Firefox, and WebKit cover the local runtime; see evidence/v1-workbenches/metis-curriculum-coverage.json. 1EdTech CASE, CAST UDL assessment guidance, and W3C table/reflow guidance inform identity, planning, and interaction boundaries. This is inspectable runtime planning evidence, not CASE certification or conformance, proof of instructional or assessment validity or learner mastery, accreditation determination, accessibility/WCAG conformance, or deployment validation. Institution vocabulary and threshold governance, expert curriculum review, representative-user evaluation, production identity and durable storage, backup, monitoring, security, scale, and deployment evidence remain explicit operational gates. -->
  • M2.16 Implement curriculum decision records. Evidence: selected, rejected, and superseded outlines remain traceable.
    • M2.16.a Define immutable decision identity, actor/authority, compared outline revisions, selected/rejected/deferred/merged disposition, rationale, evidence, conditions, and timestamp. Evidence: decision schema.

    • M2.16.b Bind selection to exact outline/brief/objective/standards/ workload/coverage versions and retain all alternatives. Evidence: selection tests.

    • M2.16.c Record rejection/defer reasons and requested changes without deleting or modifying the outline. Evidence: negative decision tests.

    • M2.16.d Record merge/fork as a new outline revision with explicit source variant contributions and decisions. Evidence: merge lineage tests.

    • M2.16.e Supersede through a new decision linked to the prior decision and preserve current/historical projection. Evidence: traceability tests. <!-- M2.16 evidence (2026-08-29): Metis now persists tenant-scoped, immutable, self-hashed 2.16.0 curriculum decision records with stable identity, named actor and role, explicit authority scope and basis, all compared outline revision witnesses, selected/rejected/deferred/merged/forked disposition, rationale, exact review evidence, conditions, negative reason codes and requested changes, server timestamp, and request and record digests. Selection fails closed unless the chosen subject has current exact workload and coverage workspaces whose outline, brief, objective-map, content, source, model/policy, calculation, and verdict hashes agree; the accepted brief revision supplies exact versioned standards while every alternative remains embedded in the record. Rejection and deferral write only the decision ledger and retain byte-equivalent outline records. Merge and fork delegate to the M2.13 outline revision engine, then bind its immutable structural decision, created revision, parent structures, and explicit source contributions. A new decision must name the exact current predecessor; current and historical status is derived from one linear successor chain, and missing, stale, cyclic, divergent, cross-tenant, unauthorized, idempotency-reused, or tampered evidence fails closed.

      The warm-paper decision workbench at /create/decisions exposes a captioned semantic comparison register, exact workload/coverage readiness, selected-row evidence inspector, signed authority and evidence form, conditional rejection/defer requests, merge/fork revision controls, and a linear current/historical authority timeline. The comparison is the only labelled two-dimensional scroll region, the rest reflows at 390 px, reduced-motion preferences are honored, and shared axe automation runs on every browser journey. Runtime OpenAPI and the generated TypeScript client include list, detail, and decision-write contracts. Automated service, web, contract, build, evidence, and cross-browser verification is recorded in evidence/v1-workbenches/metis-curriculum-decisions.json.

      W3C PROV-O entity/activity/agent, attribution, revision, and derivation concepts; NIST AI RMF human governance and documentation boundaries; and W3C table and reflow guidance informed the provenance and interaction design. This is inspectable local runtime evidence, not PROV-O or NIST framework conformance, instructional validity, accreditation or governance determination, WCAG conformance, or production deployment validation. Institution authority policy, expert curriculum review, representative-user evaluation, production identity and durable transactional storage, backup, monitoring, scale, security, and deployment evidence remain explicit operational gates. -->

  • M2.17 Build curriculum discovery across existing courses, open resources, standards, concept graphs, prerequisites, credentials, and institution catalogs with rights, authority, version, locale, and reuse diagnostics. Evidence: discovery never silently imports or republishes content.
    • M2.17.a Define discovery index/result/source/version contracts and connectors for each named catalog type with provenance/freshness. Evidence: discovery schemas.

    • M2.17.b Ingest only authorized metadata/excerpts/fingerprints with rights, authority, locale, standards/objectives/concepts/prerequisites/ credential relations. Evidence: connector tests.

    • M2.17.c Implement permission-aware search/filter/facet/rank/compare with exact source, version, match rationale, currency, and no protected answer/learner leakage. Evidence: query tests.

    • M2.17.d Diagnose reuse compatibility for license, audience, locale, standards, prerequisites, level, accessibility, target, source quality, and update risk. Evidence: diagnostic fixtures.

    • M2.17.e Implement preview/bookmark/request-import/link-as-reference as explicit actions; require review/rights/validation before canonical authoring use. Evidence: no-silent-import tests. <!-- M2.17 evidence (2026-08-29): Metis now exposes an immutable, self-hashed 2.17.0 curriculum-discovery boundary over exactly seven connector profiles: existing courses, authorized open resources, CASE-shaped standards frameworks, governed concept graphs, prerequisite maps, credential catalogs, and institution registrar catalogs. Every snapshot retains connector profile and version, authority basis, catalog version, manifest URI and digest, retrieval and index time, tenant owner, exact external source/version URIs, fingerprint, rights evidence, locale, accessibility metadata, quality, visibility, and typed standard, objective, concept, prerequisite, and credential relations. Input contracts require explicit metadata and fingerprint authorization, allow excerpts only with preview authorization, structurally prohibit protected answers and learner data, reject contradictory or uncleared copy/adapt grants, require catalog-appropriate authority, and preserve every historical snapshot while projecting only the latest per connector owner.

      Search filters visibility before query matching, facets, ranking, or comparison so hidden tenant records do not affect counts or evidence. Results bind the exact index record to source version, retrieval currency, rank components and rationale, ten fixed reuse checks (license, audience, locale, standards, prerequisites, level, accessibility, target, source quality, and update risk), and deterministic result and diagnostic hashes. Invisible comparisons, stale item or diagnostic hashes, unauthorized public snapshots, idempotency divergence, lineage divergence, and disk tampering fail closed. Preview, bookmark, import request, and reference link are separate immutable actions. Import approval requires a different administrator, copy/adapt rights, a non-blocked current diagnosis, and rights and validation hashes matching the exact current evidence; even an approved review records canonical_write_performed=false, copies no content, and creates no canonical artifact.

      The frontend-skill-guided warm-paper finding aid at /create/discovery uses a restrained ranked register and synchronized exact-source inspector, exposes visible facets, authority, rights, version, fingerprint, currency, match and typed-relation evidence, reveals a semantic ten-row comparison only after selecting two to four sources, and makes the four action verbs explicit alongside their human/rights/validation gate states. Content reflows without page overflow at 390 px; only the labelled two-dimensional comparison scrolls internally, reduced-motion preferences remove entrance motion, and the shared axe fixture runs after every browser journey. Runtime OpenAPI and the generated TypeScript client include connector, snapshot, search, exact-item, action, action-ledger, and independent-review contracts. Focused Ruff/Mypy/service/API/contract/release-gate tests, web lint/type/unit checks, deterministic evidence tests, OpenAPI/client drift checks, and 12 axe-backed journeys across Chromium, Firefox, and WebKit cover the affected local runtime; see evidence/v1-workbenches/metis-curriculum-discovery.json.

      1EdTech CASE and LTI Resource Search metadata, Schema.org LearningResource, W3C ODRL 2.2, Credential Engine CTDL, and Creative Commons license guidance informed identity, relation, rights, and action boundaries. This is inspectable local runtime evidence, not standards or accessibility conformance, a legal or license determination, proof of instructional validity, accreditation, or production deployment validation. Institution connector credentials and source-specific terms, legal and expert curriculum review, representative-user evaluation, production identity and durable transactional storage, backup, monitoring, scale, security, and deployment evidence remain explicit operational gates. -->

  • M2.18 Build discipline-aware planning templates and verification requirements for proofs/worked solutions, lab or field evidence, data analysis, source interpretation, language performance, and executable computing artifacts. Evidence: specialized requirements remain typed and survive standards export where supported.
    • M2.18.a Define versioned discipline template with applicability, specialized typed blocks, verification/gate/reviewer/accessibility/ standards-export requirements, inheritance, and code floors. Evidence: template schema.

    • M2.18.b Define proof/worked-solution premises/steps/justifications/ notation/equivalence/checker/manual review and semantic math alternatives. Evidence: mathematics templates.

    • M2.18.c Define lab/field question/method/materials/safety/ethics/ observations/data/uncertainty/analysis/reproducibility evidence. Evidence: lab templates.

    • M2.18.d Define dataset/source/version/variables/units/cleaning/method/ code/results/visualization/uncertainty/reproducibility for data analysis. Evidence: data templates.

    • M2.18.e Define primary/secondary source context/authority/perspective/ claim/citation/corroboration/counterevidence for interpretation. Evidence: source interpretation templates.

    • M2.18.f Define language/oral performance prompt/context/language/skills/ recording/consent/rubric/assessor/accessibility evidence. Evidence: performance templates.

    • M2.18.g Define code/runtime/dependencies/input/output/tests/seed/resource/ network/security/accessibility/reproducibility for computing artifacts. Evidence: computing templates.

    • M2.18.h Map each typed field to applicable QTI/course/standards export or explicit extension/loss/incompatibility and round-trip it. Evidence: export survival tests. <!-- M2.18 evidence (2026-08-29): Metis now exposes a strict, immutable 2.18.0 discipline-planning registry with one self-hashed abstract code floor and six self-hashed concrete templates. Applicability pins subject, artifact kind, and depth tier. Every concrete template inherits the same reviewer-count, reviewer-role, typed-completeness, human-review, accessibility-alternative, and export-survival requirements; catalog construction rejects any weakened floor, incomplete field mapping, stale template hash, mismatched artifact kind, or inapplicable subject.

      The discriminated artifact union gives proof/worked solutions typed premises, steps, justifications, notation, equivalence claims, checker evidence, manual review, and presentation/content MathML plus speech, braille, and plain-text alternatives. Lab and field plans type question, method, materials and hazards, safety, ethics, observations, data, uncertainty, analysis, and reproducibility. Data plans type the exact dataset source/version, variables/types/units/roles, cleaning, method, pinned code, results, accessible visualization, uncertainty, and reproducibility. Interpretation plans keep primary and secondary source kinds separate while typing context, citation, authority, perspective, claim, corroboration, and counterevidence. Language plans type prompt, context, locale, skills, recording digest and transcript/captions, consent/retention, rubric, calibrated assessor, moderation, and non-audio alternatives. Computing plans type source, runtime/container, dependencies/integrity, input/output contracts, tests, deterministic seed, resource limits, fail-closed network policy, sandbox and scans, accessible errors/output, and reproducibility evidence.

      Every leaf in every specialized block has exactly one QTI, course, and standards disposition. Metis course JSON carries the typed plan natively. QTI 3 packages retain it through a declared JSON sidecar because QTI natively describes assessment interactions, while CASE 1.1 retains it in a declared extension because CASE natively describes competencies and rubrics. Preserved-extension exports deserialize to the identical plan hash across all three channels; native-only exports preserve course JSON and report every removed QTI/CASE extension leaf as explicit loss. No unsupported artifact evidence is mislabeled as a native standards field.

      The frontend-skill-guided /create/disciplines workbench uses a restrained six-template rail, typed scalar and structured evidence editor, exact template pin ledger, non-weakening gate inspector, extension-policy control, and per-field export-survival table. Structured JSON errors are stable across engines. The page reflows without document overflow at 390 px; the labelled field table alone scrolls horizontally and is keyboard-focusable; reduced-motion preferences remove entrance motion; and the shared axe fixture runs after every browser journey. Runtime OpenAPI and the generated TypeScript client include the catalog, validation, export, six-way discriminator, and field-disposition contracts. Focused Ruff/Mypy/service/API/OpenAPI/release-gate tests, web lint/type/unit checks, deterministic evidence tests, OpenAPI/client drift checks, and 12 axe-backed journeys across Chromium, Firefox, and WebKit cover the affected local runtime; see evidence/v1-workbenches/metis-discipline-templates.json.

      Official 1EdTech QTI 3 and CASE 1.1, W3C MathML 4, and WCAG 2.2 informed the local mapping and accessibility decisions. This is inspectable local runtime evidence, not QTI, CASE, MathML, WCAG, instructional, accreditation, or production conformance. Expert discipline, accessibility, safety, ethics, assessment, legal, representative-user, production identity, transactional persistence, backup, monitoring, scale, security, and deployment review remain explicit operational gates. -->

M3 — Lesson, activity, media, and rubric authoring#

  • M3.1 Build course and module boards with guarded maturity transitions. Evidence: server-enforced move tests.
    • M3.1.a Define board columns/grouping/order/filter view from canonical course/module revision and maturity state. Evidence: board view tests.

    • M3.1.b Implement accessible create/open/reorder/group operations with keyboard/non-drag alternatives and optimistic concurrency. Evidence: board interaction tests.

    • M3.1.c Resolve allowed maturity transitions, roles, prerequisites, blockers, review/approval, reason, and version effect server-side. Evidence: transition matrix tests.

    • M3.1.d Reject crafted moves, stale revisions, missing gates, unauthorized roles, and cross-tenant IDs; preserve immutable transition history. Evidence: server-enforced negative tests. <!-- M3.1 evidence (2026-08-29): Metis now projects course and module boards from the existing SQL Course/Module hierarchy. Alembic 003 adds canonical maturity, revision, group, and order metadata plus an append-only maturity transition ledger; a clean temporary Postgres rehearsal passed empty-to- head upgrade, 003-to-002 downgrade, and 002-to-003 re-upgrade. Alembic's synchronous URL now selects the declared psycopg driver explicitly.

      Every board card carries its exact entity revision and content witness. The board witness is calculated over the complete unfiltered canonical entity set, so grouping, maturity filtering, and search cannot weaken create or reorder concurrency checks. Course and module create, open, exact-set reorder, and group moves use expected board hashes and/or entity revisions. Earlier/Later buttons are the primary reorder control, making pointer dragging unnecessary; filtered views disable reorder rather than submitting an incomplete set.

      A server-owned eleven-edge matrix governs Idea, Planned, Authoring, In review, Approved, Published, and Archived. Clients submit only entity ID, expected revision, target, and a bounded reason. The authenticated server derives role eligibility, canonical content prerequisites, review state, author-approval conflict, parent/publication gates, blockers, review effect, and version effect. Row locks and optimistic witnesses reject stale, crafted, missing-gate, unauthorized, duplicate/incomplete reorder, and cross-tenant requests. Accepted moves append actor, reason, complete resolved evidence, previous-event hash, and event hash; history reads recompute and report chain validity. There is no update/delete history API.

      The frontend-skill-guided /create/boards workbench uses a restrained board-first register, accessible forms, explicit open/reorder/group controls, a gate-and-version-effect inspector, stale-write reload, and an immutable-history view. At 390 px the document remains contained while the labelled board region alone scrolls horizontally; reduced-motion preferences remove entrance animation. Focused Ruff/Mypy/service/API/ OpenAPI/persistence/release-gate checks, generated-client drift checks, web lint/type/unit checks, deterministic evidence tests, and 12 axe-backed Playwright journeys across Chromium, Firefox, and WebKit cover the affected local runtime. See evidence/v1-workbenches/metis-course-module-boards.json.

      This is inspectable local runtime evidence, not proof of production migration, deployment, scale, availability, backup/recovery, monitoring, identity-provider or institutional role governance, representative-user usability, accessibility conformance, instructional validity, legal review, or accreditation. Those remain explicit operational gates. -->

  • M3.2 Build the lesson dossier with objectives, prerequisites, sources, content, activities, assessments, assets, gates, review, and history. Evidence: live-data browser journey.
    • M3.2.a Define one typed lesson dossier projection joining exact lesson/ course revision and all named relationships with freshness/redaction. Evidence: dossier contract.

    • M3.2.b Present objectives/prerequisites/coverage and source/citation/ authority/currency/rights with conflicts and missing requirements. Evidence: curriculum/source panels.

    • M3.2.c Present ordered content/activities/assessments/rubrics/assets/ accessibility variants and dependency state. Evidence: content panels.

    • M3.2.d Present quality/safety/accessibility/correctness gates, comments/ reviews/approvals, maturity, publication readiness, and remediation. Evidence: gate/review panels.

    • M3.2.e Present revision/diff/decision/audit/published-build/use history with stable deep links and no answer-key exposure. Evidence: history tests.

    • M3.2.f Verify live edit-review-preview journey, direct property auth, responsive accessibility, and large-dossier performance. Evidence: browser E2E. <!-- M3.2 evidence (2026-08-29): Metis now joins the canonical SQL Lesson, Module, Course, assessment metadata, and an append-only dossier ledger into one typed 3.2.0 projection. Every projection and mutation carries exact lesson, module, course, dossier revision, tenant, observation-time, and dossier-hash witnesses. Legacy course/module/lesson mutations advance the same hierarchy revisions, so older endpoints cannot silently bypass dossier concurrency. Alembic 004 adds lesson maturity/revision plus hash-chained dossier revision and event tables; an isolated temporary Postgres rehearsal passed empty-to-head upgrade, 004-to-003 downgrade, and 003-to-004 re-upgrade with two columns, two tables, and seventeen indexes verified at the boundary before the temporary database was removed.

      The strict dossier contract binds stable, unique objectives, prerequisites, governed source versions/citations, ordered content, activities, assessment links, rubrics, assets, and accessibility variants. It rejects broken references, non-contiguous order, and protected source excerpts at the persistence boundary. The server derives dependency state, coverage/source/rights/accessibility/assessment conflicts, missing requirements, relationship freshness, explicit redactions, and quality, safety, accessibility, and correctness gates. Assessment projections carry only identity, metadata, question count, objective/rubric links, and dependency state: questions, options, explanations, answer keys, and correct answers are absent. Service and browser fixtures include a sentinel secret answer and prove it never appears in dossier or preview JSON, the page, or browser mutation payloads.

      Authenticated reads and writes authorize the exact course property; row locks plus all expected revisions and the dossier hash reject stale, cross-tenant, crafted, and missing-gate requests. Author self-approval is forbidden, approval is administrator-only and requires a review request for the exact current revision, and edit/review/approval transitions append bounded actor/reason/diff evidence. Preview, published-build, and recorded- use events preserve stable deep links, exact artifact hashes, and validated revision/event hash chains without a history update/delete API.

      The frontend-skill-guided /create/lessons/[lessonId] workbench is a calm, dossier-first evidence register rather than a dashboard mosaic. It exposes exact scope and counts, canonical and structured editing, curriculum/source conflicts, ordered learning relationships, derived gates/remediation, immutable review, learner-safe preview, revision diffs, decisions, audits, builds, uses, redactions, and a labelled freshness register. Course Studio links newly created lessons directly to their stable dossiers. Large registers progressively disclose beyond forty rows; at 390 px the document remains contained while only the labelled freshness table scrolls, and reduced-motion preferences remove transitions.

      Focused Ruff/Mypy/service/API/OpenAPI tests (including sixty-six final service/API contract checks), a fresh migration rehearsal, generated-client drift checks, web lint/type/unit checks, deterministic evidence tests, and 12 axe-backed Playwright journeys across Chromium, Firefox, and WebKit cover the affected local runtime. See evidence/v1-workbenches/metis-lesson-dossiers.json.

      This is inspectable local runtime evidence, not proof of production deployment, production migration, scale, availability, backup/recovery, monitoring, identity-provider or institutional role governance, representative-user usability, accessibility conformance, instructional validity, legal review, or accreditation. Those remain explicit operational gates. -->

  • M3.3 Build a structured lesson block editor for prose, callout, example, equation, code, image, audio, video, interaction, practice, reflection, and assessment link. Evidence: block registry coverage.
    • M3.3.a Define versioned block registry entry, shared identity/order/ locale/source/objective/accessibility fields, per-kind schema, renderer, validator, and export mappings. Evidence: registry schema.

    • M3.3.b Implement typed editors/previews for prose/callout/example and semantic equation/code blocks. Evidence: text/technical block tests.

    • M3.3.c Implement governed image/audio/video attachment blocks with captions/transcripts/descriptions/rights/provenance. Evidence: media block tests.

    • M3.3.d Implement sandboxed interaction, practice, reflection, and exact assessment/item link blocks without copying protected answers. Evidence: activity block tests.

    • M3.3.e Implement insert/duplicate/reference/reorder/nest/convert/delete/ restore with stable IDs, depth rules, dependency invalidation, and non-drag controls. Evidence: editing tests.

    • M3.3.f Fail coverage for any registered block lacking editor, learner renderer, validation, accessibility, localization, import/export, preview, and test disposition. Evidence: block registry gate. <!-- M3.3 completion evidence (2026-08-29):

      • metis.schemas.lesson_blocks originated registry 3.3.0 and now exposes backward-compatible registry 3.5.0, strict discriminated schemas for all twelve requested kinds, stable shared identity/order/parent/locale/source/objective/reference/origin and accessibility fields, per-kind editor/learner-renderer/validator dispositions, localization/import mappings, six export mappings, preview fixtures, and test dispositions. Server and browser coverage gates fail when any registered capability is absent.
      • Prose/callout/example/reflection editors, semantic sanitized MathML, and deliberately inert code previews are typed end to end. Image, audio, and video blocks pin exact asset/revision/rendition/hash plus rights evidence, provenance, captions/transcripts/descriptions, and accessibility alternatives. Unsafe or incomplete blocks render as inert unavailable placeholders rather than leaking ungoverned media.
      • Interactions use the allowlisted isolated-v1 sandbox profile and declared capabilities. Practice and assessment links carry exact activity/assessment/item IDs only; learner renders and browser fixtures prove the protected-answer sentinel is absent.
      • The authenticated block-operation boundary applies named insert/duplicate/reference/reorder/nest/convert/edit/delete/restore commands against exact course/module/lesson/dossier/revision/block-hash witnesses. Server application preserves stable IDs, normalized sibling order, maximum depth three, cycle/reference rules, append-only revision and event history, tombstone restore, and dependency invalidation. Full dossier PUT cannot bypass the operation boundary.
      • The frontend-skill-informed manuscript canvas, structural rails, and focused inspector expose typed controls for all twelve kinds, explicit non-drag reorder/nesting actions, deletion history, exact learner preview, validation findings, and registry proof. The canvas discloses blocks progressively after sixty and remains keyboard-operable, reduced-motion aware, and contained at 390 CSS pixels.
      • Automated proof passes Ruff, Python 3.12 MyPy, 46 focused service/API/ OpenAPI tests, OpenAPI export and generated-client drift checks, Metis web ESLint/typecheck/unit tests, twelve Chromium/Firefox/WebKit M3.3 Playwright journeys (including 180-block performance, 390-pixel containment, reduced motion, axe, and direct-property refusal), and twelve M3.2 dossier regression journeys. Deterministic traceability and current input checksums are published at evidence/v1-workbenches/metis-lesson-blocks.json.

      This is inspectable local contract and browser-runtime evidence, not proof of production deployment, production media/activity integration, production load or availability, backup/recovery, institutional policy configuration, representative-user usability, accessibility conformance, instructional validity, legal review, or accreditation. M3.5-M3.6 retain the governed gallery and hostile sandbox verification gates. -->

  • M3.4 Build semantic MathML and accessible code/example editing. Evidence: screen-reader and copy/export tests.
    • M3.4.a Author validated MathML source with visual preview, semantic tree, labels/numbering, variable/units context, speech/braille text, and image fallback. Evidence: math editor tests.

    • M3.4.b Preserve semantic MathML through save/diff/copy/paste/QTI/export/ learner render without raster-only substitution. Evidence: round-trip tests.

    • M3.4.c Author code with language/version, plain text, line semantics, annotations, input/output/example/test refs, and no formatting-only meaning. Evidence: code editor tests.

    • M3.4.d Provide accessible syntax/error/result descriptions, keyboard operation, copy behavior, wrapping/zoom/high contrast, and screen-reader navigation. Evidence: accessibility tests.

    • M3.4.e Sanitize markup/code, fence execution separately, and verify exports cannot inject scripts/styles or leak hidden answers. Evidence: security/export tests. <!-- M3.4 completion evidence (2026-08-29):

      • Technical schema 1.1.0 and the technical milestone's registry 3.4.0 now participate in backward-compatible registry 3.5.0 while retaining stored 1.0.0 compatibility and requiring complete nonvisual semantics for newly saved equation, code, and worked-example blocks. The strict MathML parser accepts a bounded MathML namespace/tag/attribute profile, rejects declarations, processing instructions, custom entities, foreign/active elements, unsafe attributes, structural text, more than 2,000 nodes, depth beyond 32, and sources beyond 100 KB, then stores canonical namespace-qualified MathML and a typed semantic tree.
      • Equation authoring presents validated source, a React-reconstructed visual preview without HTML injection, keyboard-navigable semantic nodes, label and equation number, variable meanings and unit systems, speech/braille/plain alternatives, and an exact governed fallback image revision/rendition/digest/alt-text witness. Learner projections retain the MathML tree and alternatives and explicitly report raster_only=false.
      • Exact JSON-pointer diffs, append-only named edits, SHA-256 integrity clipboard envelopes, new-ID paste with origin lineage and renewed accessibility review, and learner projection preserve semantic source. The authenticated export boundary binds lesson/module/course/dossier, dossier hash, and block-document hash before serializing learner HTML, plain text, semantic PDF source, QTI, SCORM, or Common Cartridge. Every non-plain equation export embeds semantic MathML, and all eighteen kind-by-target cases are parsed back to verify their declared round trip rather than accepting raster substitution.
      • Code remains plain text with explicit language/version, exactly one role and accessible description per line, typed annotation ranges, input/output/example/test references, and independent syntax, error, and result descriptions. The editor exposes keyboard line navigation, wrapping, 80-200% zoom, and an axe-verified high-contrast mode. Worked examples expose ordered analysis steps, result descriptions, exact input/output/test references, and an ordered screen-reader preview.
      • Rendering never grants execution: code carries an immutable disabled fence with no sandbox, network, filesystem, credentials, or learner- data access, and the UI exposes no run control. XML/HTML serializers escape code and author text, protected answer keys are forbidden in clipboard/schema/API projections, stale or invalid blocks cannot export, and direct-property authorization is enforced before any technical payload is produced. Denied Clipboard API access falls back to explicit selection copy without an unhandled browser rejection.
      • Focused Ruff/format, Python 3.12 strict MyPy, 71 service/API/OpenAPI tests, OpenAPI export and generated-client drift/type checks, Metis web ESLint/typecheck, 141 web unit tests, and 24 axe-backed block-editor journeys across Chromium, Firefox, and WebKit pass. Those comprise 12 M3.4 math/code/example/security journeys plus 12 retained M3.3 block journeys; 12 cross-browser M3.2 dossier journeys also pass. Generated traceability and current input checksums are published at evidence/v1-workbenches/metis-lesson-technical.json.

      This is inspectable local contract and browser-runtime evidence, not proof of production deployment, formal assistive-technology or representative-user validation, accessibility conformance, production LMS interoperability, semantic print fidelity, institutional policy configuration, instructional validity, legal review, or accreditation. Those remain explicit operational gates. -->

  • M3.5 Build media attachment through the governed Isis gallery. Evidence: source, rights, alt text, transcript, captions, and provenance gates.
    • M3.5.a Query/select only curated audience-authorized gallery asset revisions by kind/locale/target/status with current capability state. Evidence: gallery selector tests.

    • M3.5.b Attach exact immutable asset/output revision and approved derivative/rendition role without copying or rewriting Isis lineage. Evidence: attachment schema tests.

    • M3.5.c Require source/rights/consent/provenance/quality/safety/target eligibility and propagate expiry/quarantine/recall. Evidence: gate tests.

    • M3.5.d Require kind-appropriate alt text/long description/transcript/ captions/audio description/sign-language or reviewed N/A reason. Evidence: accessibility gate tests.

    • M3.5.e Preview through the learner renderer and verify media, metadata, accessibility tracks, permissions, and offline/bandwidth variants. Evidence: attachment E2E. <!-- M3.5 completion evidence (2026-08-29):

      • Isis now owns a curated lesson-gallery projection over immutable output records. Its BFF derives tenant, audience, and capabilities from the authenticated request; filters current revisions by media kind, locale, target, lifecycle status, and delivery capability; computes totals only after authorization; and discloses no generation workflow, model, prompt, raw evidence, or lineage internals.
      • Metis queries that projection only after direct lesson authorization. Attachment re-resolves the opaque selection against current Isis state inside the dossier row lock, pins exact asset/output revision, derivative/rendition role, rendition and content digests, and authority fingerprints, and appends the dossier asset and media block atomically. Generic insert, convert, edit, and whole-dossier replacement paths cannot fabricate or rewrite governed media or copy Isis lineage.
      • Selection and every learner render require current source, rights, consent, provenance, quality, safety, target, and rendition gates. Expiry, supersession, quarantine, recall, or lost eligibility causes exact re-resolution to fail and learner media plus its top-level stale delivery locator to be withheld. Service and authority tests cover concurrent state change, forged selections, digest mismatch, tenant or audience mismatch, and authorization-before-disclosure behavior.
      • Isis validates kind-appropriate alternative text, long descriptions, transcripts, captions, audio descriptions, and sign-language references or an explicit reviewed N/A reason with reviewer and review time. Metis revalidates that complete access matrix rather than trusting a partial projection, and stores authority references and fingerprints instead of copied source/rights/provenance evidence.
      • The frontend-skill-informed editorial contact sheet supports filtered keyboard selection, restrained governance evidence, exact attachment, and a read-only immutable inspector. The same learner renderer presents image/audio/video media, caption and accessibility tracks, rights and credit metadata, download permissions, and online/offline/ low-bandwidth variants; recalled media becomes an unavailable notice.
      • Automated proof passes 60 focused Python service/API/OpenAPI/client tests, 38 Isis gallery tests, five focused BFF route/OpenAPI-runtime tests, Ruff, strict targeted MyPy, Isis and Metis-web typechecks, task-scoped ESLint, both OpenAPI generated-client drift checks, 145 web unit tests, and 33 axe-backed block-editor journeys across Chromium, Firefox, and WebKit. Nine of those browser journeys exercise query, keyboard selection, exact attachment, governance evidence, learner variants/tracks, recall withholding, tenant boundaries, and narrow layout. Deterministic traceability and current input checksums are in evidence/v1-workbenches/metis-lesson-gallery.json.

      This is inspectable local contract and browser-runtime evidence, not proof of production Isis curation population, deployment, representative catalog scale, CDN/offline delivery, institutional rights policy, assistive-technology or representative-user validation, accessibility conformance, legal review, or accreditation. Task-scoped BFF lint is clean; the app-wide BFF lint target remains red on 150 errors and 265 warnings in unrelated pre-existing files. Those operational and baseline gates remain explicit. -->

  • M3.6 Build interactive-activity configuration with preview isolation. Evidence: untrusted embeds cannot access parent credentials or data.
    • M3.6.a Define allowlisted activity/runtime version, typed parameters, initial/reset state, objectives, events, data needs, sizing, offline, accessibility alternative, and target. Evidence: config schema.

    • M3.6.b Validate parameter ranges/dependencies/seed/resource/network/ browser permissions/content security and prohibit arbitrary code/URLs. Evidence: validation tests.

    • M3.6.c Render preview in origin/process sandbox with explicit message protocol, capability grants, time/memory/storage/network limits, and teardown. Evidence: sandbox tests.

    • M3.6.d Use synthetic sample identity/context and prevent access to parent cookies/storage/credentials/DOM/network, real learner data, or answer keys. Evidence: escape tests.

    • M3.6.e Capture deterministic events/results/errors for preview validation without recording as real attempts or grades. Evidence: preview tests. <!-- M3.6 completion evidence (2026-08-29):

      • Metis now owns a URL-free activity registry 3.6.0 with two pinned activity/runtime/implementation tuples. Each entry fixes typed integer, number, boolean, and enum definitions; exact initial/reset state; objectives, events, synthetic-only data needs, pinned package dependencies, capability grants, responsive sizing, offline posture, learner-web target, CSP profile, and resource ceilings. Interaction blocks require a complete keyboard-usable non-interactive alternative.
      • Pydantic discriminated values and server-owned validators reject wrong types, missing/extra IDs, duplicates, ranges, step misalignment, dependency violations, invalid seeds, ungranted browser capabilities, widened sizing or resource limits, weakened network/storage/ credential/DOM/answer-key policy, arbitrary code, and runtime URLs. Block validation also blocks protected assessment-answer text copied into activity authoring fields.
      • Read-only preview issuance binds current lesson, module, course, dossier, dossier-hash, and block-document-hash witnesses. It projects a constant sample learner/tenant, objective aliases, and deterministic session/instance/contract hashes, scans the finished payload against canonical IDs and assessment secrets, and creates no attempt, grade, dossier revision, domain event, or database flush. Author-entered title and instruction prose remains in the parent; the iframe receives only server-owned runtime copy, removing that answer-key escape channel.
      • The frontend-skill-informed workbench edits only server-allowlisted fields and opens a static runtime on a distinct origin in an allow-scripts-only, credentialless iframe with an opaque origin. The sandbox route sends deny-by-default CSP and Permissions-Policy, credentialless embedder policy, cross-origin resource policy, no-referrer/no-store headers, and Origin-Agent-Cluster opt-in. Runtime fetch/XHR/WebSocket/EventSource/beacon/open access is disabled in addition to zero-network CSP and zero persistent-storage grants.
      • Host/runtime communication accepts only the explicit versioned initialize/reset/teardown and runtime-event protocol, exact session, source and opaque origin, contiguous sequence, allowed event type, bounded payload and event count. Timeout, invalid messages, runtime errors, and explicit teardown fail closed and remove the browsing context. Result payloads contain null score/grade/attempt fields and a preview-only recording scope; stable canonical payload serialization makes the validation ledger deterministic.
      • Automated proof passes focused Ruff/format, Python 3.12 strict targeted MyPy, 61 service/API/OpenAPI tests, OpenAPI export and generated-client drift checks, Metis web ESLint/typecheck, 149 web unit tests, nine deterministic evidence tests, and 36 axe-backed block-editor journeys across Chromium, Firefox, and WebKit. Three hostile activity journeys verify distinct opaque origin, exact sandbox headers, missing parent cookie/credentials/DOM/storage/network access, no answer-key escape, deterministic events, reset, no grade/attempt event, and teardown. The inspected full-page Chromium capture preserves the restrained editorial hierarchy and makes the isolation proof legible. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-activities.json.

      This is inspectable local contract and browser-runtime evidence, not proof of production origin routing or deployment. Opaque cross-origin sandboxing and Origin-Agent-Cluster establish browser isolation requests, while exact renderer-process placement and hard OS memory enforcement remain browser/deployment controlled and are not certified here. It also is not representative assistive-technology or learner validation, accessibility conformance, institutional policy approval, instructional validity, legal review, or accreditation. The repository's default Python 3.11 MyPy mode currently stops in installed NumPy stubs that use Python 3.12 syntax; the changed service surface is clean in the actual Python 3.12 environment. A production Next build was not run under the small-host hard memory gate; targeted typecheck, unit tests, and three- engine dev-runtime compilation/browser verification pass. Those operational and baseline gates remain explicit. -->

  • M3.7 Build rubric authoring with weight validation and exemplar evidence. Evidence: exact weighted-score fixtures.
    • M3.7.a Build typed rubric/criterion/level/descriptors/points/weights/ ordering/objective/item/evidence authoring with immutable revisions. Evidence: authoring tests.

    • M3.7.b Validate unique criteria/levels, complete descriptors, ordered/ nonoverlapping score ranges, weights/totals, required evidence, locale, and scoring method. Evidence: validation tests.

    • M3.7.c Attach authorized exemplar artifact/revision, criterion/level mapping, annotation, rights/consent, redaction, and reviewer approval. Evidence: exemplar tests.

    • M3.7.d Preview learner/assessor versions with protected exemplar/key fields separated and accessible table/nonvisual semantics. Evidence: role preview tests.

    • M3.7.e Calculate exact scores for boundary/missing/N/A/rounding/ moderation cases and pin delivered items to rubric revision. Evidence: weighted-score fixtures. <!-- M3.7 completion evidence (2026-08-30):

      • Metis now owns strict typed rubric, criterion, level, evidence, objective, assessment-item, scoring-policy, exemplar, review, preview, item-binding, moderation, and score contracts. Server-issued rubric revision IDs, sequence numbers, timestamps, digests, dossier revisions, and domain events make rubric history append-only; exact lesson, module, course, dossier, prior-rubric, and content-hash witnesses reject stale authoring.
      • Authoring validation requires unique contiguous criteria and levels, complete locale-matched descriptors, monotonically increasing points, contiguous ordered nonoverlapping inclusive ranges, unique evidence requirements, valid objective/item/assessment links, the fixed analytic weighted-points policy, and exactly 10,000 basis points globally and for every delivered item. Generic dossier replacement cannot mutate the typed rubric-revision ledger.
      • Exemplars resolve an exact current dossier artifact revision and digest and an exact criterion/level pair. Rights, consent evidence, redaction evidence, annotation, and independent approval fail closed. A new or changed approved exemplar must be submitted by the authenticated admin named as reviewer, and that reviewer cannot be the course author.
      • Distinct learner and assessor response models provide an ordered table caption, explicit row/column semantics, and nonvisual summary. Learner schemas structurally omit assessor notes, score keys, review identity, and assessor-only exemplars; the service constructs each role projection independently and regression tests scan the serialized learner result for protected values.
      • Exact Decimal scoring resolves inclusive boundary levels, verifies all required evidence, scores missing criteria as zero, reweights N/A criteria, applies independently justified moderation, rounds HALF_UP at the configured precision, and emits exact string-valued contributions, totals, a calculation digest, and the immutable delivered assessment/ item/rubric revision binding. Known-value fixtures cover boundary, missing, N/A, rounding, moderation, and stale or mismatched delivery.
      • The frontend-skill-informed rubric workbench uses the existing warm manuscript studio language as a restrained assessor ledger: criterion ordering and exact weights, level ranges, evidence requirements, governed exemplar review, role-switched semantic tables, immutable witness receipts, and non-persisting score proofs remain legible at 390px and under reduced motion. The dossier projects real safe item metadata for selection without answer keys or explanations.
      • Automated proof passes focused Ruff/format, Python 3.12 targeted MyPy, 79 service/dossier/block/OpenAPI tests, OpenAPI export and generated- client drift checks, Metis web ESLint/typecheck, 154 web unit tests, deterministic evidence tests, 15 axe-backed rubric journeys, twelve dossier regression journeys, and 36 block-editor regression journeys across Chromium, Firefox, and WebKit. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-rubrics.json.

      This is inspectable local contract and browser-runtime evidence, not proof of production deployment, production identity or persistence integration, institutional rubric or moderation policy, representative learner/assessor usability, assistive-technology validation, accessibility conformance, instructional validity, legal review, or accreditation. The repository's default Python 3.11 MyPy mode currently stops in installed NumPy stubs that use Python 3.12 syntax; the changed service surface is clean in the actual Python 3.12 environment. A production Next build was not run under the small-host hard memory gate; targeted typecheck, unit tests, and three-engine dev-runtime compilation/ browser verification pass. Those operational and baseline gates remain explicit. -->

  • M3.8 Build authoring-time readability and language-complexity analysis. Evidence: server-computed known-value tests.
    • M3.8.a Define supported language/locale/age/grade/content-kind applicability, metrics/features, units, versions, thresholds, and limitations. Evidence: analysis schema.

    • M3.8.b Compute server-side token/sentence/word/syllable/lexical/syntactic/ cohesion/domain-term features on exact block/lesson revision. Evidence: known-value fixtures.

    • M3.8.c Return raw measures/regions/assumptions/uncertainty/N/A separately from institution policy bands and suggestions. Evidence: result tests.

    • M3.8.d Avoid penalizing proper nouns, technical terms, quotations, code, math, multilingual/learner-level intentional text without applicability context. Evidence: false-positive tests.

    • M3.8.e Present advisory evidence with source/version and author accept/ dismiss/revise decision; never auto-rewrite. Evidence: review UI tests. <!-- M3.8 completion evidence (2026-08-30):

      • Metis now owns a strict v1 readability profile for English, the exact en, en-US, and en-GB locales, ages 11–120, four grade bands, six analyzable prose-bearing block kinds, six automatically N/A block kinds, nineteen versioned raw metrics across token, sentence, word, syllable, lexical, syntactic, cohesion, and domain feature families, four grade-sensitive policy references, 500-block and two-million- prose-character request bounds, and six disclosed limitations. Unsupported languages, English locales, audiences, content kinds, and locale mismatches return explicit N/A evidence rather than fabricated values.
      • The deterministic server engine analyzes exact typed block fields and binds every result to lesson, module, course, dossier, dossier digest, and block-document digest witnesses. A known-value fixture proves all nineteen unique outputs for cats nap. dogs run., including four tokens/words/syllables, two sentences, mean length 2.00, reading ease 120.21, and grade proxy -3.01. The content digest is stable when only observation time changes.
      • Raw lesson/block measures, exact field paths and character offsets, excerpts, regions, exclusions, assumptions, uncertainty, and N/A reasons are structurally separate from the institution policy result. Policy carries its own ID/version, observed/reference comparisons and advisory prompts, reports aggregate_score=not_computed, and never represents a learner diagnosis, quality score, or accessibility score.
      • Declared technical/domain terms, quotations, likely proper nouns and acronyms remain visible as excluded raw regions but leave policy-grade arithmetic. Parameterized fixtures prove explicit N/A behavior for code, MathML equations, locale mismatch, unsupported English locale, unsupported language and audience, and author-declared intentional or multilingual complexity. Every N/A case yields no policy prompt.
      • Analysis is a read-only named operation. Accept, dismiss, and revise are authenticated human decisions: the service recomputes the exact analysis, rejects stale digests and unknown suggestions, appends an immutable review record to the dossier, and confirms content_mutated=false. Generic dossier replacement cannot bypass the review ledger. Suggestion contracts contain evidence and author prompts while fixing proposed replacement to null and automatic rewrite to false.
      • The frontend-skill-informed language proof sheet uses a restrained warm manuscript layout with exact witnesses, applicability controls, protected vocabulary, raw measures and anchored regions, explicit N/A, a separate policy table with formulas/limits, and a dark immutable human-decision ledger. The inspected full-page Chromium capture keeps the evidence hierarchy legible. The UI contains no horizontal overflow at 390px, honors reduced motion, and fixes the contrast and definition- list semantics found by automated accessibility review.
      • Automated proof passes focused Ruff/format, Python 3.12 strict targeted MyPy, 130 service/block/interaction/rubric/technical/API/OpenAPI tests, live OpenAPI and generated-client drift checks, Metis web ESLint, typecheck, 157 unit tests, nine deterministic evidence tests, and 78 axe-backed browser journeys across Chromium, Firefox, and WebKit: 15 readability, twelve dossier, 36 block-editor, and 15 rubric cases. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-readability.json.

      This is inspectable local contract and browser-runtime evidence, not proof of production deployment or persistence integration, representative learner/author usability, assistive-technology validation, formal accessibility conformance, institutional threshold approval, comprehension, instructional validity, legal review, or accreditation. The v1 engine is deliberately English-only and uses disclosed punctuation, orthographic syllable, marker-word clause, and adjacent-word-overlap heuristics rather than a full linguistic parser. The repository's default Python 3.11 MyPy mode currently stops in installed NumPy stubs that use Python 3.12 syntax; the changed service surface is clean in the actual Python 3.12 environment. A production Next build was not run under the small-host hard memory gate; targeted typecheck, unit tests, dev-runtime compilation, and three-engine browser verification pass. Those operational and baseline gates remain explicit. -->

  • M3.9 Build terminology and glossary lint. Evidence: locale-aware, course-scoped results.
    • M3.9.a Define versioned course/locale glossary entries, preferred/ allowed/deprecated/forbidden terms, definitions, variants, context, capitalization, links, and ownership. Evidence: glossary schema.

    • M3.9.b Tokenize/normalize by locale/script while excluding code, math, citations, quoted source, names, and configured contexts. Evidence: locale fixtures.

    • M3.9.c Detect undefined/inconsistent/deprecated/forbidden/ambiguous terms, acronym first use, spelling/casing, and glossary-link drift with exact ranges. Evidence: lint tests.

    • M3.9.d Scope results to exact course/glossary/lesson revisions and distinguish inherited/institution/domain terms and conflicts. Evidence: scope tests.

    • M3.9.e Present replace/add-definition/link/allow-context/dismiss actions with human decision and no destructive bulk edit. Evidence: lint UI tests. <!-- M3.9 implementation evidence (2026-08-30): - Added immutable metis.terminology-lint v1 contracts for exact institution/domain/course glossary snapshots and entry revisions. The schema binds locale, scope identity, inheritance, publication, preferred/allowed/deprecated/forbidden status, definition, spelling/ alias/acronym variants, contextual dispositions, capitalization, definition/authority/usage/translation links, replacement lineage, and accountable owners. The lesson dossier projection is version 3.5.0 and carries an append-only terminology-review ledger. - The deterministic server engine resolves exact-locale entries by course > domain > institution precedence without hiding inherited resolutions or status/definition/capitalization/ambiguous-surface conflicts. Exact lesson/module/course/dossier/block witnesses and every glossary revision/hash are joined into one glossary-set digest; stale, foreign-course, missing-block, and oversized scopes fail with stable errors. - Unicode NFKC plus case-fold tokenization preserves original field offsets and identifies Latin, Han, Hiragana, and Katakana evidence. Parameterized en-US, es-ES, fr-FR, and ja-JP fixtures cover accented, apostrophe, and full-width forms. Code, math, citations, quotations, configured names/contexts, unsupported locales/content, and empty analyzable text become exact exclusions or explicit N/A rather than silent findings. Limits are disclosed and enforced at 500 blocks, 2,000,000 analyzed characters, 100 glossaries, and 100,000 entries. - All nine required categories—undefined, inconsistent, deprecated, forbidden, ambiguous, acronym-first-use, spelling, casing, and glossary-link drift—are covered by server-known fixtures with exact block, JSON field path, start/end offsets, excerpt, entry/glossary/ conflict provenance, severity, and grounded actions. The analysis digest excludes observation time; every finding declares automatic_edit=false and the analysis declares content_mutated=false. - Added authenticated analysis and review commands to the dossier API, live/checked-in OpenAPI, and generated TypeScript client. Review submission recomputes the exact analysis and rejects stale witnesses, unknown findings, mismatched actions, or out-of-scope link targets. Exactly one replace/add-definition/link/allow-context/dismiss proposal is appended with reviewer and full witness evidence; generic dossier updates cannot bypass the ledger, and both content_mutated and glossary_mutated remain false. - Following the required frontend skill, the author surface uses an editorial terminology-ledger thesis: restrained paper-like lesson evidence, a dark provenance rail, one teal action accent, visible layer conflicts, and a separate preserved-decision ledger instead of generic dashboard cards. It exposes locale/scope controls, definition candidates, name exclusions, block treatments, exact ranges, exclusions/N/A, method limits, all five human actions, immutable history, stale-witness reload, direct role restriction, keyboard focus, responsive 390px layout, and reduced-motion behavior. Automated axe review also drove dark-rail contrast and opacity-free motion fixes. - Automated proof passes focused Ruff/format, Python 3.12 strict MyPy for all six affected service files, 147 terminology/readability/ dossier/block/interaction/rubric/technical/API/OpenAPI tests, live OpenAPI and generated-client drift checks, Metis web ESLint, typecheck, 161 unit tests, nine deterministic evidence tests, and 96 affected axe-backed browser journeys across Chromium, Firefox, and WebKit: 18 terminology, 15 readability, twelve dossier, 36 block editor, and 15 rubric cases. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-terminology.json.

      This is inspectable local contract and browser-runtime evidence, not
      proof of production deployment or persistence integration,
      institutional glossary approval, representative author usability,
      assistive-technology validation, formal accessibility conformance,
      semantic or linguistic correctness, legal review, or accreditation.
      The v1 engine supports disclosed en/es/fr/ja locale families but uses
      literal normalization/token patterns, deterministic quotation/citation/
      name controls, and no morphological parser or semantic disambiguation;
      Japanese script runs are not word segmentation. Undefined-term evidence
      is conservatively limited to configured definition candidates and
      unregistered uppercase acronyms. A production Next build was not run
      under the 15 GiB host's hard memory gate; targeted typecheck, unit tests,
      dev-runtime compilation, and three-engine browser verification pass.
      Those operational and baseline gates remain explicit. -->
      
  • M3.10 Build inclusivity and bias assist as review evidence, not an unchallengeable truth score. Evidence: model/version/uncertainty visible.
    • M3.10.a Define versioned review dimensions/rubrics, applicability, evidence units, evaluator/provider/model/config, thresholds/manual bands, limitations, and protected-data policy. Evidence: assist schema.

    • M3.10.b Analyze exact content/item/media revision for representation, stereotypes, exclusionary language/context, cultural/linguistic bias, accessibility, and assessment construct-irrelevant barriers. Evidence: calibrated fixtures.

    • M3.10.c Return attributed findings with text/region/item anchor, category, rationale, score/confidence/uncertainty, counterevidence, and N/A/error. Evidence: result tests.

    • M3.10.d Display evaluator/model/version/corpus limits and avoid one aggregate “bias score” or automatic truth/author blame. Evidence: UI tests.

    • M3.10.e Require authorized human accept/reject/reframe/request-specialist review decisions, preserving proposal and disagreement. Evidence: authority tests. <!-- M3.10 implementation evidence (2026-08-30): - lesson_inclusivity.py defines the immutable 1.0.0 assist contract: six separately reported review dimensions, typed evidence units, applicability, three human-review bands, exact evaluator, provider, model, configuration, corpus, and calibration identities plus hashes, and explicit limitations. Its protected-data policy permits authored content and public course metadata only, forbids learner identity/enrollment/grade/attempt/accommodation/demographic and inferred-trait inputs, keeps provider transmission and excerpt persistence off, prohibits trait inference, and requires an authorized human decision. - lesson_inclusivity_helpers.py provides a bounded server-local deterministic English evaluator over the exact lesson/module/course/ dossier/block-document witnesses and selected block, content-item, governed-media-revision, and assessment dossier-revision subjects. Calibrated literal and structural fixtures cover representation, stereotyping, exclusionary language and sensory assumptions, cultural/linguistic context, missing media accessibility support, and construct-irrelevant assessment barriers. Quoted, historical, and configured-exclusion contexts remain explicit rather than silently changing the evidence. - Results preserve original text offsets or typed item/media/assessment anchors, exact revision/hash references, category, rationale, evaluator rule/runtime attribution, evidence-strength score, confidence, uncertainty, suggested human-review band, and linked counterevidence. Every unit is independently analyzed, not_applicable, or error; every dimension remains separate; the contract computes no aggregate bias score and makes no truth, author-blame, automatic edit, or learner-impact claim. Stable subject-set and analysis digests make the same exact inputs reproducible. - The dossier schema is now 3.6.0. Analyze and review endpoints bind exact witnesses, reject stale or cross-tenant requests, and append immutable inclusivity review/event/revision evidence. Direct generic mutation of the review ledger is rejected. Authorized reviewer roles may accept, reject, reframe, or request specialist review; each record preserves the evaluator proposal, human rationale, disagreement or reframing, specialist request, reviewer identity/role, and confirms that neither the finding nor authored content was mutated. - The Metis Challenge folio workbench follows the mandatory frontend-skill: a warm editorial evidence surface with one muted violet accent and a dark provenance margin, not a dashboard scorecard. It exposes exact scope and witnesses, six independent dimensions, anchored findings, counterevidence, N/A/error states, all model/ provider/config/corpus/calibration versions and limitations, all four human decisions, preserved disagreement, immutable history, stale reload, role refusal, 390px reflow, and reduced-motion behavior. Automated axe review also drove a provenance-rail contrast correction. - Automated proof passes focused Ruff/format, Python 3.12 strict MyPy for six affected service files, 171 inclusivity/readability/ terminology/dossier/block/interaction/rubric/technical/API/OpenAPI tests, live OpenAPI and generated-client drift checks, Metis web ESLint, typecheck, 164 unit tests, nine deterministic evidence tests, and 114 affected axe-backed browser journeys across Chromium, Firefox, and WebKit: 18 inclusivity, 18 terminology, 15 readability, twelve dossier, 36 block-editor, and 15 rubric cases. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-inclusivity.json.

      This is inspectable local contract and browser-runtime evidence, not
      proof of production deployment or persistence integration, cultural or
      linguistic validity, demographic representativeness, lived impact,
      institutional policy approval, legal review, representative specialist/
      author usability, assistive-technology validation, or formal
      accessibility conformance. The v1 evaluator supports disclosed en,
      en-US, and en-GB locales and deliberately uses a small versioned literal/
      structural corpus; it cannot infer intent, speaker position, dialect,
      pragmatics, cultural meaning, the assessed construct, or whether an
      alternative is equivalent. Assessment analysis is bound to the exact
      dossier revision because the legacy assessment link has no separate
      upstream version field; media without a governed asset revision is an
      explicit error. A production Next build was not run under the 15 GiB
      host's hard memory gate; targeted typecheck, unit tests, dev-runtime
      compilation, and three-engine browser verification pass. Those
      operational and baseline gates remain explicit. -->
      
  • M3.11 Build lesson duration and cognitive-load estimates. Evidence: assumptions and confidence shown.
    • M3.11.a Define per-block/activity/media/assessment time and intrinsic/ extraneous/germane or ratified load model with versions/units/ assumptions/applicability. Evidence: estimate schema.

    • M3.11.b Compute active/passive/synchronous/independent duration and load by segment from exact content quantities, modality, difficulty, prerequisites, language, device, and supports. Evidence: fixtures.

    • M3.11.c Aggregate sequence peaks, sustained load, switching, redundancy, pacing, breaks, and total ranges with uncertainty. Evidence: sequence tests.

    • M3.11.d Compare against brief/institution/learner-band constraints and surface assumptions/data gaps/remediation without treating estimates as fact. Evidence: explainability tests.

    • M3.11.e Record author override with reason and preserve calculated values separately; recompute on revision/model change. Evidence: lifecycle tests. <!-- M3.11 implementation evidence (2026-08-30): - lesson_duration_load.py defines the immutable 1.0.0 estimate contract and the metis.lesson-duration-load / 1.0.0 model plus metis.lesson-duration-load.en-v1 / 1.0.0 configuration. Eleven versioned rates cover block, activity, media, and assessment subjects in word, line, equation, image, media-second, activity, and assessment-item units. Every rate carries applicability, bounded seconds per unit, active share, baseline intrinsic/extraneous/germane planning-proxy ranges, assumptions, and source references. The model profile exposes its arithmetic, hashes, supported locales, learner- band uninterrupted thresholds, and limitations; duration/load observation and automatic fact flags are structurally false. - lesson_duration_load_helpers.py walks exact depth-first authored blocks, then dossier-only activities and assessments, and binds every segment to a subject revision/hash plus exact quantity source. Block quantities come from exact words, lines, equations, images, or typed activities; governed media require an exact asset revision and declared duration; assessments require a declared exact version, item count, and bounded seconds per item. Missing or mismatched evidence is an explicit data_gap, not_applicable, or error, never a silent zero. Difficulty, prerequisite readiness, instruction language, device, supports, modality, delivery mode, synchronous assignments, and breaks produce attributed factors. Active plus passive and synchronous plus independent ranges remain arithmetically balanced. Scope is bounded to 10,000 segments and two million characters per text subject, and stable digests exclude observation time. - Sequence aggregation reports lower/expected/upper total and channel ranges, complete/known-subtotal/unavailable status, confidence, missing segment IDs, separate intrinsic/extraneous/germane peaks, sustained high-load and uninterrupted runs, modality switches, adjacent redundancy, pacing changes, and planned breaks with uncertainty and remediation. Brief and institution constraints carry policy code/version/evidence, while the learner-band uninterrupted constraint is bound to the model profile. Interval comparisons return within/below/above/straddles/inconclusive/not-available and explicitly state that they are not observed learner results; missing policy sources do not become passes. - The dossier schema is now 3.7.0. Estimate requests are stateless and bind exact lesson/module/course/dossier/block-document, subject-set, context, model, and configuration witnesses. The named override operation locks and recomputes the current estimate, rejects stale or unknown segments and cross-tenant access, snapshots the raw calculated duration/load vector, and appends a reasoned replacement plus evidence reference, actor, and exact hashes. Calculated and effective segments, summaries, and constraints remain separate. An override applies only to the exact subject/context/model tuple; later exact overrides supersede without erasing history, and content/context/model revisions surface stale-subject/stale-context/stale-model dispositions. Direct generic mutation of the override ledger is rejected, and receipts confirm content_mutated=false and calculated_values_mutated=false. - The Metis Pacing sheet workbench follows the mandatory frontend-skill: a warm editorial rehearsal surface with restrained rust/navy accents, a calculated staff beside an author-adjusted staff, explainable segment sheets, and a dark provenance rail rather than a generic score dashboard. Authors can declare every context axis, supports, synchronous segments, breaks, exact media and assessment quantities, and policy ceilings; inspect all four duration channels, all three separately labeled planning proxies, sequence signals, constraint comparisons, assumptions, uncertainty, gaps, remediation, versions, and hashes; then append a balanced adjustment while the raw estimate stays visible. It includes stale reload, role refusal, 390px reflow, reduced-motion handling, confined sequence overflow, and an explicit “estimate, never observation” boundary. Automated axe review drove darker accent and secondary-text contrast before acceptance. - Automated proof passes focused Ruff format/lint, Python 3.12 strict MyPy for six affected source files, 185 estimator/inclusivity/ readability/terminology/dossier/block/interaction/rubric/technical/ API/OpenAPI tests, live OpenAPI and generated-client drift checks, Metis web ESLint, typecheck, 167 unit tests, nine deterministic evidence tests, and 132 affected axe-backed browser journeys across Chromium, Firefox, and WebKit: 18 duration/load, 18 inclusivity, 18 terminology, 15 readability, twelve dossier, 36 block-editor, and 15 rubric cases. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-duration-load.json.

      This is inspectable local contract and browser-runtime evidence, not
      proof of production deployment or persistence integration, empirical
      timing or cognitive-load validity, observed learner state, outcome or
      ability prediction, institutional policy approval, representative
      author or learner usability, assistive-technology validation, formal
      accessibility conformance, or accreditation. The eleven v1 rates and
      context multipliers are coarse disclosed deterministic planning
      assumptions over en/en-US/en-GB content, not a validated psychometric or
      learning-science instrument. No learner records, demographics,
      telemetry, attempts, grades, accommodations, or measured behavior are
      inspected. Media duration must be supplied against a governed asset
      revision. Assessment version and timing are author-declared because the
      legacy dossier assessment link has no separate upstream version field.
      A production Next build was not run under the 15 GiB host's hard memory
      gate; targeted typecheck, unit tests, dev-runtime compilation, and
      three-engine browser verification pass. Those operational and baseline
      gates remain explicit. -->
      
  • M3.12 Build learner preview using the real delivery renderer and permission-safe sample identity. Evidence: author/learner parity test.
    • M3.12.a Invoke the same versioned renderer/runtime/target contracts used by learner delivery against an exact unpublished course/lesson build. Evidence: renderer parity tests.

    • M3.12.b Issue a synthetic sample identity with no real enrollment/ learner/grade/integrity authority and a bounded preview session. Evidence: identity isolation tests.

    • M3.12.c Project selected synthetic locale/preferences/PNP/prerequisite/ progress states without copying real learner records. Evidence: context tests.

    • M3.12.d Prevent preview interactions from creating real attempts, grades, credentials, analytics, notifications, integrity cases, or LTI/ LMS writes. Evidence: side-effect tests.

    • M3.12.e Compare author preview DOM/media/interaction/accessibility/ scoring-safe behavior to learner delivery for the same build. Evidence: parity journey. <!-- M3.12 implementation evidence (2026-08-30): - Metis now has a separate immutable 1.0.0 lesson-delivery contract. One LessonDeliveryDocument binds the exact lesson/module/course/ dossier/block-document revisions and hashes, publication posture, learner-web target contract, block and activity registry versions, interaction protocol, renderer ID/version/hash, selected context, redacted content, governed learner media, filtered runtime entries, semantic block projections, and its own deterministic digest. Author preview and published delivery are distinct authenticated read routes, but both call the same pure document builder. Published access permits a course author/administrator, an authenticated free-preview learner, or an active enrollment and returns only an opaque subject digest; it records no access or progress mutation. Preview can bind an unpublished build through complete optimistic witnesses and never needs a publish state or fake enrollment. - The author route issues a random, bounded 5–30 minute, non-resumable sample session whose state is limited to browser parent memory. Its generated Sample learner identity has structurally false enrollment, grade, and integrity authority, no bearer token, no canonical learner or profile lookup, and no protected-data copy. Locale, explanation depth, media preference, pace, PNP supports, prerequisite state, progress state/percent/completed block IDs, and a rationale are typed, hashed, uniqueness/consistency checked, and constrained to the exact current block set. The service uses for_update=False, does not load Enrollment or Progress on preview, does not append a dossier event or revision, and does not call add/flush. Exact answer values, assessment explanations, and canonical person IDs are scanned out before the contract can leave the service. - Preview write isolation names and exactly denies attempts, grades, credentials, analytics, notifications, integrity, LTI, and LMS sinks; mutation transport is disabled, credentials are not issued, persistence hooks are absent, scoring is local/nonpersistent, and the event budget is bounded to parent memory. The shared React renderer has a single document prop and no preview/delivery mode branch or network client. Completion, practice, reflection, governed activity, and assessment-detail actions use component-local state only. Governed media fail closed on withheld or unsafe locators. Interaction blocks fail closed without the exact filtered runtime-registry entry and expose the same activity/runtime, learner-web target, and opaque-origin sandbox profile in both surfaces. Assessment answers never enter the document. - Following the mandatory frontend-skill, the new delivery rehearsal is an editorial learner canvas rather than a dashboard: the real lesson paper dominates, an intentionally compact sample-context rail sits to one side, and exact build/session/write/parity evidence sits on the other. The author page owns those rails and session countdown outside the renderer. Course play invokes the same renderer for canonical UUID lessons and retains its prior content/Yemaya fallback when typed delivery is absent. The renderer covers all twelve block kinds, semantic math and code alternatives, governed media/accessibility text, exact activity runtime evidence, keyboard-operable local inputs, withheld states, selected synthetic completion, responsive 390px reflow, and reduced motion. Existing fake-ID course fixtures keep their legacy path, while real UUID lessons take typed delivery. - Parity is executable evidence rather than a UI claim. The contract requires equal preview/learner document hashes and the same renderer contract hash. Service tests compare complete author and published documents; React server-render tests compare exact static markup; and the Playwright journey compares exact innerHTML plus document digest between author rehearsal and course play for the same build. Separate browser journeys verify all twelve kinds and media/accessibility alternatives, exact context payloads, zero API requests after local interactions, 409 witness reload, role refusal before dossier access, and 390px/reduced-motion containment. Every journey uses the shared automatic WCAG 2.1 AA axe fixture. - Automated proof passes focused Ruff format/lint, Python 3.12 strict targeted MyPy for the four affected service sources, 190 lesson block/ activity/technical/dossier/duration/readability/terminology/ inclusivity/rubric/API/OpenAPI tests, live OpenAPI and generated-client drift checks, Metis web ESLint and typecheck, 173 web unit tests (six focused context/renderer cases), nine deterministic evidence tests, and 18 delivery-preview browser journeys across Chromium, Firefox, and WebKit. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-delivery-preview.json.

      This is inspectable local contract, renderer, and automated browser
      evidence, not proof of production deployment or persistence integration,
      every external activity or media runtime, representative author or
      learner usability, assistive-technology behavior, formal accessibility
      conformance, or every future target/build combination. The covered
      locale set is deliberately en/en-US/en-GB; device/theme/network/offline
      matrices remain M3.13 scope. The existing production course-progress
      chrome is outside the shared document and is not asserted as preview
      parity. The installed MyPy configuration defaults to Python 3.11 while
      installed NumPy stubs use Python 3.12 syntax; the changed service surface
      is clean under the actual Python 3.12 environment. A production Next
      build was not run under the 15 GiB host hard-memory gate; targeted
      typecheck, unit tests, development-runtime compilation, visual review,
      and three-engine browser verification pass. -->
      
  • M3.13 Build device, theme, locale, accommodation, offline, and bandwidth preview modes. Evidence: preview matrix.
    • M3.13.a Define supported viewport/input/device capabilities, themes/ contrast/motion, locales/direction, PNP supports, connectivity/cache, and network profiles as versioned preview presets. Evidence: preset schema.
    • M3.13.b Apply presets to the real learner renderer without claiming hardware/assistive-technology emulation beyond measured capabilities. Evidence: preview accuracy tests.
    • M3.13.c Verify content/layout/input/media/interaction/accessibility/ localization behavior and explicit unsupported/degraded state for every valid matrix cell. Evidence: preview matrix.
    • M3.13.d Simulate offline/intermittent/slow/high-latency/limited-data behavior with cache provenance, fallback, retries, and no hidden success. Evidence: network tests.
    • M3.13.e Save/share a bounded preview configuration and findings without learner identity or false conformance certification. Evidence: report tests. <!-- M3.13 implementation evidence (2026-08-30): - The immutable metis.lesson-delivery-preview-presets@1.0.0 registry exposes exactly eight finite cells spanning desktop/tablet/phone CSS viewports; pointer, touch, keyboard, and combined requested input semantics; light/dark/high-contrast/reduced-motion scoped appearance; en-US/en-GB plus an explicit en-US RTL direction-only fallback; registered PNP supports; fast/high-latency/limited-data/intermittent/ offline transport; and write-through, explicit-network, warm-session, and empty-session cache modes. The registry and every bound environment carry deterministic SHA-256 witnesses and are available through an instructor/admin route and generated OpenAPI client. - Registered sessions still mount LessonDeliveryRenderer over the exact immutable M3.12 document. LessonDeliveryRuntimeProvider adds only CSS-container dimensions, renderer-scoped tokens/direction/ requested semantics, and controlled media transport. Contracts fix physical-hardware emulation, assistive-technology emulation, and formal accessibility certification to false; saved measurements separately record actual browser container, pointer/hover, theme, contrast, motion, online, locale, direction, and DPR observations. - Every cell returns all seven content/layout/input/media/interaction/ accessibility/localization outcomes with explicit supported, degraded, unsupported, or not-applicable status. Playwright iterates all eight cells in Chromium, Firefox, and WebKit, checks exact container dimensions/direction/theme/transport, containment, and a WCAG 2.1 A/AA axe analysis inside the real learner document for every cell. Existing exact author/published DOM parity remains covered when no matrix cell is selected. - Governed media starts source-less, then uses credential-free, no-store controlled fetch with latency/throughput pacing, deterministic intermittent drops, exact content-type/byte-length/SHA-256 checks, document-bound session-memory cache provenance, explicit limited-data opt-in, visible retry, and offline warm-hit/cold-miss fallback. Browser request ledgers prove held, failed, and offline states make no hidden request. The tests caught and removed an initial-render native request and made retry counters safe under React development effect replay. - Migration 005 and LessonDeliveryPreviewReportRecord define immutable author/admin-scoped, idempotent, exact-build/environment reports with 30-day expiry. POST/GET routes reauthorize course scope, revalidate all witnesses and governed media observations, reject idempotency drift, verify stored report hashes, and reject expired reports. The response, share receipt/link, and JSON download contain a bounded safe context, configuration, actual measurements, and server-derived findings while structurally excluding learner/author identity and all hardware/AT/ conformance claims. - Verification passed: Ruff and Python 3.12 MyPy on every changed service surface; offline Alembic upgrade SQL through revision 005; 83 focused service/API/OpenAPI/export tests; current OpenAPI regeneration; full Metis web lint/typecheck; 177 web unit tests; and 27 focused Playwright cases across Chromium/Firefox/WebKit with both automatic serious/ critical axe gating and the per-cell WCAG A/AA checks. Deterministic traceability is stored at evidence/v1-workbenches/metis-lesson-delivery-preview-matrix.json. - Scope remains explicit: the checks do not prove production deployment, a live database migration, representative author/learner usability, physical hardware, assistive technology, formal accessibility conformance, or arbitrary custom matrix cells. A production Next build was not run under the 15 GiB host approval gate; targeted typecheck, unit coverage, and three-engine development-runtime compilation pass. -->
  • M3.14 Build autosave, revision history, semantic diff, restore, and concurrent conflict resolution. Evidence: Playwright conflict journey.
    • M3.14.a Implement local draft state and debounced idempotent server autosave with expected revision, dirty/saving/saved/offline/error/ conflict indicators, retry, and unload recovery. Evidence: autosave tests.
    • M3.14.b Persist immutable authoring revisions/checkpoints with author/ reason/time/base, changed entities, and audit; avoid revision explosion from transient UI state. Evidence: history tests.
    • M3.14.c Compute semantic block/field/relationship/objective/source/ accessibility diffs using stable IDs and protected-field redaction. Evidence: diff fixtures.
    • M3.14.d Restore by creating a new revision from an authorized historical version with impact/stale approval preview. Evidence: restore tests.
    • M3.14.e Detect concurrent edits at entity/field/list level, present mine/theirs/base, support safe merge/manual resolution, and forbid last- write-wins data loss. Evidence: conflict tests.
    • M3.14.f Run two-editor/offline/reconnect/reorder/delete/restore/ permission-revocation accessible Playwright journey. Evidence: conflict E2E. <!-- M3.14 completion evidence (2026-08-30):
      • The authoring write boundary now separates private mutable autosave heads from immutable canonical checkpoints. Each editor-owned LessonAuthoringDraftRecord carries an immutable exact base witness and full base document, a monotonically compared draft version, the current document digest, and request-bound idempotency evidence. A 700 ms client debounce writes only through compare-and-swap; retries cannot reuse a key with different content, a stale request cannot overwrite the draft head, and another actor cannot read it.
      • The change-ledger client keeps a lesson+actor-scoped, versioned local recovery envelope with a seven-day expiry and no credentials. It persists synchronously on every accepted edit and beforeunload, reconciles the server head without dropping a newer local intent, retries on reconnect, and exposes current/dirty/saving/saved/offline/ error/conflict/access-revoked states. A revoked 403 stops automatic retries while retaining an explicit JSON export. Protected source excerpts remain excluded by the typed dossier boundary.
      • Migration 006 adds the private draft table and enriches every future immutable dossier revision with a full canonical+dossier document and digest, revision kind, base revision number/hash, changed stable entities, restore provenance, and request idempotency evidence. The existing actor, role, reason, timestamp, chained revision hash, and chained audit event remain server-owned. Named content operations now capture the same full snapshot; autosave tests prove that transient keystrokes create zero immutable revisions.
      • lesson_authoring_helpers.py computes deterministic domain-semantic block, scalar field, relationship, objective, source, and accessibility changes. Collections bind stable domain IDs rather than indexes; relationship additions merge as sets with stable order; and delete-versus-edit, reordered positions, same-field edits, and other overlaps retain base/mine/theirs. Source excerpts and answer/secret field families are structurally redacted from diff and conflict projections.
      • Checkpoint creation performs a server-side three-way merge from the immutable base, private head, and newly authorized canonical state. Non-overlapping field and relationship changes merge safely. Every unresolved entity/field/list overlap returns a typed conflict outcome with no projection or revision, and mine/theirs/manual resolutions must bind the current deterministic conflict IDs before validation and commit. Named review/rubric/governed-media evidence still cannot be bypassed through the generic authoring document.
      • Authorized comparison and restore endpoints read hash-verified full historical snapshots. Restore preview rechecks all optimistic witnesses, computes the exact semantic direction and changed entities, lists approval events that will become stale, and issues a stable preview digest. Apply recomputes that preview and creates a new chained restore revision pointing to the historical source; it never mutates old history. Pre-M3.14 rows without a full canonical snapshot fail explicitly instead of fabricating restorable values.
      • The UI follows the frontend-skill's calm editorial change-ledger thesis: the draft is dominant, the immutable timeline is a narrow rail, and conflicts expand in place into base/mine/theirs evidence. Conflict mount transfers keyboard focus to “No silent winner”; all choices are labelled fieldset controls, protected values announce redaction, and the layout reflows without page overflow at 390 px.
      • Verification passed: Ruff on the full Metis service, Python 3.12 MyPy on every changed service surface, offline Alembic SQL through the sole 006 head, deterministic OpenAPI export plus generated Metis TypeScript drift checks, 209 lesson/OpenAPI/provider Python tests, full Metis web lint and typecheck, all 181 web unit tests, and six focused Playwright cases across Chromium/Firefox/WebKit. The browser suite uses the automatic serious/critical axe gate and proves two isolated editor contexts, conflict focus and no last-write-wins, offline unload and reconnect recovery, block reorder+delete, restore, permission revocation with stopped retries, local export, reduced motion, and 390 px reflow. Deterministic traceability is stored at evidence/v1-workbenches/metis-lesson-authoring-revisions.json.
      • Scope remains explicit: tests do not prove a production deployment, a live migration, representative two-person usability, physical assistive-technology behavior, or formal accessibility conformance. The two browser editors use isolated contexts against a deterministic API ledger. A production Next build was not run under the 15 GiB host approval gate; targeted development-runtime compilation, typecheck, unit, and three-engine browser checks pass. -->
  • M3.15 Build explicit AI assist for rewrite, example, hint, feedback, and localization. Evidence: per-block accept and AI attribution.
    • M3.15.a Define assist request/proposal contracts for each operation with exact block/course/source/objective/item/locale revisions, instructions, provider provenance, cost, and status. Evidence: assist schemas.
    • M3.15.b Send only authorized bounded content/source context through the fenced provider with answer-key/integrity/learner-data exclusions and injection resistance. Evidence: context boundary tests.
    • M3.15.c Validate operation-specific output structure, citations, protected fields, locale, accessibility, safety, and malformed/refusal/ unavailable/budget states. Evidence: provider tests.
    • M3.15.d Present per-block diff/source/uncertainty/attribution and require explicit accept/edit/reject; never replace content automatically. Evidence: review UI tests.
    • M3.15.e Accept as a new attributed revision, preserve proposal and human edits, and recompute affected correctness/accessibility/gates. Evidence: lineage tests. <!-- M3.15 evidence (2026-08-30): - lesson_assist.py and the server-authored preparation receipt bind the exact lesson/module/course/dossier/block/registry/locale witness plus hashed source versions, objective revisions, and relevant item records. Provider/model/config/prompt/pricing identity, tokens, USD minor-unit cost, outcome, and zero-tool status remain immutable. - The OpenAI-compatible boundary uses five prompt-trust zones, rejects truncation, preflights the educator-confirmed budget, and excludes answer keys, integrity fields, learner data, protected source text, credentials, and tools structurally. Trusted/current/rights-cleared references and operation-specific output, citation, locale, accessibility-review, and integrity-marker checks fail closed. - The block inspector adds a restrained proposal margin rather than a chat surface. It shows the field diff, source citations, uncertainty, rationale, provider/model/cost, and quarantined provider states. The canonical editor remains unchanged until separate accept, accept-edit, or reject controls record an educator reason. - Accepted output creates only an assist_accept immutable revision carrying proposal/provider/model/operation/human-edit/effective-block attribution. Proposal and educator-edited JSON remain separate rows; block validation, correctness/accessibility findings, readiness, and release gates are recomputed. Reject creates no canonical revision and stale proposals cannot apply. - Automated proof covers all five operations, refusal/malformed/ unavailable/budget/stale states, injection and protected-data fixtures, idempotency, lineage, 390 px reflow, reduced motion, keyboard focus, and automatic axe analysis. Deterministic traceability is stored at evidence/v1-workbenches/metis-lesson-assist.json. - Scope remains explicit: provider and browser transports are mocked; tests do not prove a live model, production deployment, production migration, representative educator usability, physical assistive- technology behavior, or formal accessibility conformance. Migration 007 was verified as offline PostgreSQL SQL. A Next production build was not run under the 15 GiB host approval gate; targeted development compilation, typecheck, unit, service, contract, and three-engine browser checks are the applicable proof. -->
  • M3.16 Build lesson-template inheritance with code/institution floors. Evidence: tenant templates cannot loosen safety or accessibility.
    • M3.16.a Define versioned template identity/scope/applicability, parent, typed block skeleton/defaults/constraints/required gates, owner, lifecycle, and code-floor refs. Evidence: template schema.
    • M3.16.b Resolve exact parent chain with cycle/depth/version handling and explicit per-field/block merge/replace/append/remove/lock semantics. Evidence: inheritance tests.
    • M3.16.c Apply code, regional, institution, program, and course floors monotonically for safety, rights, sources, accessibility, assessment integrity, privacy, and standards. Evidence: floor property tests.
    • M3.16.d Preview inherited/defaulted/overridden/locked/conflicting values and effective requirements before instantiation/update. Evidence: effective template tests.
    • M3.16.e Version/publish/deprecate templates and explicitly pin/migrate lesson consumers without altering historical revisions. Evidence: lifecycle tests. <!-- M3.16 evidence (2026-08-30): - lesson_template.py closes the 1.0.0 identity, exact version ref, code/regional/institution/program/course scope, applicability, ownership, lifecycle, parent, typed value/default, skeleton, constraint, required-gate, code-floor, preview, and consumer-lineage contracts. Every template carries all seven floor domains exactly once, and generated OpenAPI/TypeScript expose only exact typed bodies. - The resolver walks a digest-verified root-to-leaf exact chain with distinct missing-template, missing-version, mismatch, cycle, maximum- depth, scope-inversion, and code-floor-pin failures. Field and block merge/replace/append/remove/lock operations preserve contribution provenance; attempts to alter locked values remain visible as rejected conflicts rather than silently winning. - The immutable code floor seeds safety, rights, sources, accessibility, assessment integrity, privacy, and standards. Floors may only raise their level and add required evidence, gates, and prohibitions. Property coverage exercises every one of the seven domains against all four weakening dimensions (28 fail-closed cases), plus monotonic tightening and a complete five-scope chain. - The frontend-skill-guided editor adds a quiet full-width policy ledger before block insertion. Educators preview the exact parent chain, inherited/defaulted/overridden/locked/removed/conflicting values, block skeleton, constraints, gates, all effective floors, and provenance before the explicit pin action is enabled. Conflict verdicts take focus, pin controls disable fail-closed, motion is removable, and the existing editorial manuscript remains dominant. - Migration 008 stores immutable definitions, hash-chained publish/ deprecate events, and append-only instantiate/migrate pins. Exact template, preview, dossier revision/hash, actor, reason, idempotency, and previous-pin lineage are captured; a storage check requires historical revisions to remain preserved, and stale/no-op/changed- lineage operations fail closed. - Verification passed: targeted Ruff and strict Python 3.12 MyPy, sole Alembic 008 head plus offline PostgreSQL SQL, deterministic OpenAPI export and generated-client drift checks, 88 resolver/service/API/ OpenAPI/exporter tests, full Metis web lint/typecheck, all 188 web unit tests, eight evidence self-checks, and all 54 lesson-editor Playwright cases across Chromium/Firefox/WebKit. Browser proof covers exact five-layer preview, pin then migration without history rewrite, floor/lock conflicts, lifecycle transitions, focus, reduced motion, mobile reflow, and automatic serious/critical axe analysis. Deterministic traceability is stored at evidence/v1-workbenches/metis-lesson-templates.json. - Scope remains explicit: service database and browser APIs are mocked; tests do not prove a production deployment, a live migration, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the 15 GiB host approval gate; targeted development compilation, typecheck, unit, service, contract, migration, and three-engine browser checks are the applicable proof. -->
  • M3.17 Build diagram, chart, graph, presentation, audio/TTS, video, and Manim-style executable-animation authoring pipelines with source/data binding, preview, versioning, cancellation, cost, and per-artifact accessibility alternatives. Evidence: no generated teaching media can be accepted without its source, code/data, renderer, and critic evidence.
    • M3.17.a Define a common teaching-media project/stage/artifact contract with objective/content/source/data/code/spec/renderer/model/tool/version, cost, progress, output, gates, and lineage. Evidence: media schemas.
    • M3.17.b Implement diagram/illustration pipeline with semantic elements/ labels/relationships, source grounding, alt/long description, and editable vector/spec where supported. Evidence: diagram tests.
    • M3.17.c Implement data chart/graph pipeline with immutable dataset slice, transform code, units/scales/labels, statistical checks, data table/ narrative alternative, and reproducible render. Evidence: chart tests.
    • M3.17.d Implement presentation pipeline with slide/notes/source/media/ reading-order/contrast/layout contracts and accessible export. Evidence: presentation tests.
    • M3.17.e Implement audio/TTS pipeline with script/source, approved voice/ consent, pronunciation, timing, loudness, transcript, and downloadable accessible text. Evidence: audio tests.
    • M3.17.f Implement video pipeline with script/storyboard/scenes/media/ narration/captions/transcript/audio description/renditions and sync. Evidence: video tests.
    • M3.17.g Implement Manim-style animation with versioned source/code/data/ environment/seed/assets, sandboxed render, frame/time validation, captions/descriptions, and nonanimated equivalent. Evidence: animation tests.
    • M3.17.h Provide preview/progress/cancel/retry/partial/error/cost/ regeneration and immutable revision semantics per stage. Evidence: lifecycle tests.
    • M3.17.i Require source/correctness/code-data-renderer/provenance/rights/ safety/accessibility/technical/critic evidence before educator acceptance. Evidence: acceptance gate. <!-- M3.17 evidence (2026-08-30): - teaching_media.py closes a single 1.0.0 project, immutable revision, ordered stage, content-addressed artifact/evidence, exact lesson/ objective/source/content witness, versioned renderer/model/tool, budget/progress/cost, worker-receipt, lineage, decision, and ten-domain acceptance contract. The authenticated API and generated TypeScript client expose only these typed request and response bodies. - Six exact deterministic renderer environments bind the governed control plane to the existing Metis semantic-diagram, verified data- figure, presentation-export, TTS, video-compositor, and isolated Manim implementations. Diagram semantics and editable alternatives; chart dataset/code/scales/statistics/table/narrative; slide reading order/contrast/notes/export; voice consent/pronunciation/loudness/ transcript; video storyboard/renditions/captions/audio description/ sync; and animation code/data/environment/seed/sandbox/frame/static- equivalent requirements all fail closed in their typed specs. - Migration 009 stores full append-only project snapshots, artifacts, and educator decisions with exact idempotency, hash-chain, indexed scope, budget, and composite project-revision foreign keys. Mutations lock the current project chain, recheck the live lesson dossier, and preserve every attempt. Content/renderer lineage, worker receipts, output digests, storage bodies, gates, project state, and decision digests are independently reproducible. - The frontend-skill-guided lesson editor adds one restrained editorial production strip: a project rail, ordered stage timeline, accessible artifact preview, exact source/renderer/worker receipt margin, and ten-domain evidence ledger. It exposes progress, retained cost, partial/error/cancel/retry/regenerate states and immutable revision inspection; acceptance disables and focuses the exact missing gate. - Verification passed: targeted Ruff and strict Python 3.12 MyPy, sole Alembic 009 head plus 1,147-line offline PostgreSQL SQL, deterministic OpenAPI export and generated-client drift checks, the focused and adjacent service/lesson-dossier/Manim/OpenAPI suites, the complete Metis web lint/typecheck and 198-unit-test suite, the Metis multimedia test and lint targets, 12 evidence self-checks, all 23 lesson-editor Chromium regressions, and all five teaching-media flows across Chromium/Firefox/WebKit. Browser proof includes 390 px reflow, reduced motion, exact worker receipts, acceptance/missing-critic focus, partial/failure/cancellation/retry/regeneration, sandboxed animation authoring, and automatic serious/critical axe analysis. Deterministic traceability is stored at evidence/v1-workbenches/metis-teaching-media.json. - Scope remains explicit: service database, browser APIs, and external render-worker transports are mocked; ten Manim environment-dependent cases skip when their optional renderer/toolchain is absent. Tests do not prove a production deployment, live migration, production media provider, representative educator usability, physical assistive- technology behavior, or formal accessibility conformance. A Next production build was not run under the 15 GiB host approval gate. A dependency-expanded multimedia build also remains blocked outside this scope by existing Iris/Metis type errors; its own tests and lint target pass, while the changed web and service surfaces pass targeted compilation and strict type analysis. -->
  • M3.18 Build the author→critic→media loop as an inspectable proposal workflow: lesson-to-scene plan, code/spec generation, sandbox render, visual/layout/correctness critique, revision, and educator acceptance. Evidence: the loop stops on budget, policy, renderer, or verification failure and never self-publishes.
    • M3.18.a Define loop/iteration/agent/tool/plan/spec/render/critique/ revision/proposal/decision contracts with exact versions and budgets. Evidence: workflow schemas.
    • M3.18.b Convert exact lesson/objectives/sources/accessibility/target into an educator-reviewable scene/artifact plan before generation. Evidence: planning tests.
    • M3.18.c Generate typed code/spec only through allowlisted tools/models with source/data binding, dependency locks, and fenced untrusted content. Evidence: generation tests.
    • M3.18.d Render in a network/filesystem/resource/time-limited sandbox, retain logs/outputs/checksums, and stop on renderer/security failure. Evidence: sandbox tests.
    • M3.18.e Run separate visual/layout/accessibility/correctness/source/ technical critics with attributed findings, uncertainty, disagreement, and no approval authority. Evidence: critic tests.
    • M3.18.f Produce bounded revisions linked to findings with maximum iterations/token/tool/compute/cost/time and explicit stop/timeout/ abstention/partial state. Evidence: loop-control tests.
    • M3.18.g Require educator inspect/accept/edit/reject for the exact final proposal and route it through normal gates; prohibit self-publication or self-waiver. Evidence: authority boundary tests. <!-- M3.18 implementation evidence (2026-08-31): - Added immutable v1 loop, project/lesson witness, target, six-axis budget/usage, policy-registry, scene-plan, typed-generation, sandbox-receipt, artifact, finding, six-critic, revision-patch, iteration, stop, unpublished-proposal, and educator-decision contracts. Agents pin provider/model/prompt/config versions and have no approval authority; tools are allowlisted with exact versions, configuration hashes, and dependency locks. Every plan, generation, receipt, finding, critique, revision, proposal, decision, and stored loop revision is independently content-addressed. - The loop is an append-only PostgreSQL and authenticated API control plane. Creation binds the exact current lesson dossier and teaching- media project revision; workers may submit only a pre-generation plan or bounded iteration receipt; educators alone approve/edit/ reject the plan and accept/edit/reject the exact final proposal. Row locks, current-project rechecks, idempotency, authorization, monotonic cumulative usage, immutable history, and read-time digest verification fail closed on stale, cross-scope, replay, or tampered state. - Generation cannot start until the exact objective/source/ accessibility/target-bound scene plan is approved. Typed source must use the admitted author/model/tool and exact dependency lock with untrusted input fenced. Sandbox receipts close network access, limit writes to the job directory, bind CPU/memory/timeout, and retain stdout/stderr, artifacts, worker versions, and checksums. Policy, renderer, verification, deadline, timeout, partial, abstention, and all iteration/token/tool/compute/cost/wall-time limits produce an explicit terminal receipt instead of autonomous retry. - Visual, layout, accessibility, correctness, source, and technical critics remain separately attributed and versioned. Findings retain exact artifact/source evidence, severity, confidence, and uncertainty; disagreement groups remain visible, and later iterations must link every prior non-passing finding. Convergence creates only an unpublished_proposal with self-publish, self-waiver, and normal acceptance-decision flags fixed false. Educator acceptance advances the exact proposal only to the existing normal teaching-media gates. - The frontend-skill-guided lesson editor adds a restrained proposal ledger beside normal media production: exact-witness creation, content-addressed revision rail, six budget meters, plan review, allowlisted policy margin, sandbox/log/source inspection, six critic columns, disagreement and stop receipts, and exact educator actions. It reflows at narrow widths, honors reduced motion, and keeps status, decision, and failure state available without color alone. - Verification passed: Ruff over the complete service source/tests, strict configured-Python MyPy for the new schema/helpers and strict Python 3.12 MyPy for all five changed service modules, 20 focused loop tests, 122 adjacent media/agent/render/OpenAPI tests with ten optional Manim environment skips, deterministic OpenAPI and generated-client drift checks, Metis web lint/typecheck and all 202 unit tests, all 30 lesson-editor Chromium regressions, and all seven affected flows across Chromium/Firefox/WebKit (21/21). Every browser test also ran automatic serious/critical axe analysis. Migration 010 is the sole head and passed an isolated PostgreSQL 001→010 upgrade plus 010→009→010 downgrade/upgrade cycle. Eleven evidence self-checks pass at evidence/v1-workbenches/metis-media-loop.json. - Scope remains explicit: database-service, browser-API, and external render-worker behavior uses deterministic fixtures; the migration exercise was isolated local PostgreSQL, not production. Repository- wide Alembic autogenerate still reports pre-existing metadata drift in older tables, while the new loop table reports no drift. Tests do not prove a production deployment/provider, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; changed web and service surfaces pass their targeted compile, type, lint, unit, contract, migration, accessibility, and cross-browser gates. -->
  • M3.19 Build prerecorded lecture-package authoring for script, timing, presenter/avatar, pronunciation, voice, slides, diagrams, demonstrations, chapters, captions, transcript, audio description, localization, renditions, and learner preview. Evidence: media, consent, correctness, accessibility, and target-playback gates.
    • M3.19.a Build structured script/segment/chapter/source/citation/objective editing with timing and dependency-aware revisions. Evidence: script authoring tests. <!-- evidence: metis-lecture-package.json .structuredScript; immutable source/objective/citation-bound segments and chapters, closed ordered dependency/timing validation, optimistic concurrency, append-only SQL revisions, and educator author/revise/history browser coverage. -->
    • M3.19.b Build presenter/avatar/voice/language/pronunciation/performance selection with identity, consent, rights, disclosure, and expiry state. Evidence: presenter tests. <!-- evidence: .presenterGovernance; exact presenter/avatar/voice/locale/ performance/pronunciation controls and content-addressed identity, consent, rights, disclosure, and expiry; structural gate tests prove expiry overrides nominally passing receipts. -->
    • M3.19.c Build slides/diagrams/charts/code/demonstration cue authoring and exact media/source/data/code links synchronized to segments. Evidence: visual authoring tests. <!-- evidence: .synchronizedVisuals; all five cue kinds and all four exact link kinds carry locator/version/SHA-256 material, remain bounded to segment timing, and are inspected in authoring plus shared playback. -->
    • M3.19.d Build captions/transcript/audio description/sign-language/ translation review with speaker/time alignment and source-revision invalidation. Evidence: accessibility authoring tests. <!-- evidence: .accessibilityLocalization; all five human-reviewed track kinds bind source revision, speaker, source segment, and timing; stale/missing tracks fail closed and es-MX translation is exercised in the browser preview. -->
    • M3.19.e Build video/audio/slide/text rendition configuration and multitrack learner preview across device/bandwidth/offline/locale modes. Evidence: rendition preview tests. <!-- evidence: .targetRenditions; all four rendition kinds bind exact locale/device/bandwidth/offline/locator/hash/track targets, every track locale requires offline low-bandwidth phone audio and screen-reader text, and the shared renderer exposes exact/fallback target state. -->
    • M3.19.f Present media/consent/rights/source/correctness/accessibility/ sync/technical/target gates and prevent package acceptance on stale/ unavailable blockers. Evidence: gate tests. <!-- evidence: .acceptanceGates; nine closed gate domains, reproducible receipts/gate digest, structural consent/accessibility/target overrides, stale/expired/unavailable blocker tests, and disabled assembly until all nine current receipts pass in Playwright. -->
    • M3.19.g Assemble/version/validate deterministic playback package and read it through the real learner renderer. Evidence: package E2E. <!-- evidence: .deterministicPlayback; exact authoring revision and gate bind the playback manifest, storage reads reverify body/index/lineage/ manifest hashes, authenticated learner reads recheck current expiry, and educator preview plus course player use LecturePackageRenderer. --> <!-- completion evidence:
      • Deterministic generator and 11 self-checks: evidence/v1-workbenches/metis-lecture-package.json via pnpm run verify:inventory:v1-metis-lecture-package. It content-addresses migration/model/schema/helper/service/API/OpenAPI/generated-client/web/ browser/TODO inputs and maps M3.19.a-g to executable evidence.
      • Backend: migration 011 creates append-only lecture_package_revisions; exact Pydantic contracts and lifecycle helpers cover structured ordered scripts, five cue kinds, five accessibility/localization tracks, four rendition kinds, presenter governance, nine acceptance gates, immutable revisions, deterministic assembly, and storage verification. Seven authenticated API operations expose catalog/create/read/revise/gate/ assemble/learner-playback flows with exact witness, source, objective, idempotency, owner, and optimistic-concurrency enforcement.
      • Frontend: the restrained editorial lecture workbench is integrated into the lesson teaching-media surface; educators edit presenter/avatar/ voice/performance/pronunciation/consent expiry, narration/timing, sign-language notes and reviewed translation, inspect visual/source links, immutable history and nine gate states, then assemble only when current. One responsive, reduced-motion-compatible renderer serves both educator target preview and the real course learner route across locale, device, bandwidth, offline, chapter, timeline, and track controls.
      • Verification passed: Ruff over complete Metis service source/tests; strict configured-Python MyPy on the new schema/helpers and strict Python 3.12 MyPy on all five changed service modules; 13 focused tests; 118 adjacent media/avatar/prerecorded/OpenAPI tests with ten optional Manim environment skips; deterministic OpenAPI and generated Metis TypeScript drift checks; web lint with zero warnings, typecheck, and all 207 unit tests in 56 files; all 32 lesson-editor Chromium regressions; and all three affected flows across Chromium/Firefox/WebKit (9/9). Browser automation includes automatic serious/critical axe analysis. Migration 011 is the sole head and passed isolated local PostgreSQL 001→011 plus 011→010→011 downgrade/upgrade verification.
      • Scope remains explicit: database-service calls, browser API calls, and prerecorded asset locators use deterministic fixtures; the migration exercise was isolated local PostgreSQL, not production. Tests do not prove production deployment/media delivery, physical assistive- technology behavior, representative educator/learner usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; the affected service and web surfaces pass targeted compile, type, lint, unit, contract, migration, accessibility, regression, and cross-browser gates. -->
  • M3.20 Build interactive simulation/widget authoring with typed parameters, initial state, learning objective, expected observations, hints, assessment hooks, reset, deterministic seed, sandbox, telemetry, and non-interactive equivalent. Evidence: security and accessible-equivalence tests.
    • M3.20.a Define registered simulation/widget/runtime version, objective, typed parameters/units/ranges/dependencies, initial state, deterministic seed, expected states/observations, and lifecycle. Evidence: schema tests. <!-- evidence: metis-simulation-authoring.json .registeredContract; registry 3.20 retains two exact legacy versions and publishes two current runtimes with closed objectives, typed units/ranges/dependency rules, seeds, state, observations, and seven-state lifecycle contracts. -->
    • M3.20.b Build parameter/state/scenario/control/layout/instructions authoring with validation, presets, reset, and exact versioning. Evidence: authoring tests. <!-- evidence: .authoring; exact preset/runtime/scenario/control/layout identities, typed parameter and initial-state fields, deterministic seed, reset state, semantic instructions, telemetry controls, and authored equivalent operations are validated and persisted in immutable lesson- dossier revisions; Playwright exercises the complete edit write. -->
    • M3.20.c Define hints/scaffolds and assessment hooks for events/state/ observations/responses/scoring evidence without embedding answer keys in public config. Evidence: learning interaction tests. <!-- evidence: .learningInteraction; tiered scaffolds and finite hint, response, completion, and observation events bind public assessment-hook IDs/response schemas/scoring-evidence fields while answer material stays server-side and the public security contract explicitly denies keys. -->
    • M3.20.d Execute preview/delivery in a versioned sandbox with resource/ network/storage/permission/message limits, cancellation, and teardown. Evidence: security tests. <!-- evidence: .sandboxSecurity; author preview and learner delivery use one versioned renderer in an opaque cross-origin credentialless iframe with allow-scripts only, deny-by-default CSP/network/credentials/storage/ parent-DOM policy, bounded runtime/memory/events/message sizes, host cancellation, teardown, and DOM removal verified in real browsers. -->
    • M3.20.e Define privacy-safe typed telemetry events, sampling, learner/ attempt context, offline buffering, consent/retention, and no arbitrary payloads. Evidence: telemetry tests. <!-- evidence: .telemetry; nine allowlisted event kinds have closed typed fields, pseudonymous learner and ephemeral attempt contexts, explicit consent and sampling, 14-day retention, and a drop-oldest offline buffer bounded to 64 events/32768 bytes/24 hours; schemas and units reject arbitrary payload containers and unconsented event collection. -->
    • M3.20.f Author a noninteractive semantic equivalent with same objective/ information/operation/assessment path and test outcome equivalence with assistive technology. Evidence: accessible-equivalence tests. <!-- evidence: .accessibleEquivalence; block validation requires exact objective, scenario information, operation-step IDs, response schema, assessment hooks, and observations to match the simulation. The shared learner renderer exposes ordered semantic text, labelled response, keyboard access, outcome witnesses, and screen-reader/text-only modes. --> <!-- completion evidence:
      • Deterministic generator and nine self-checks: evidence/v1-workbenches/metis-simulation-authoring.json via pnpm run verify:inventory:v1-metis-simulation-authoring. It content- addresses schemas/helpers/OpenAPI/generated-client/editor/renderer/unit/ browser/TODO inputs and maps M3.20.a-f to executable evidence.
      • Backend: 100 focused service/API/OpenAPI tests pass across interaction, block, delivery-preview, dossier-service, authenticated API-contract, and generated-contract suites. Changed files pass Ruff and mypy; OpenAPI and TypeScript generated-client drift checks pass. Exact interaction config is stored through the existing append-only lesson-dossier revision path, so this extension requires no parallel table or migration.
      • Frontend: all 56 Vitest files/208 tests, ESLint, TypeScript, and changed- file Prettier checks pass. The restrained lesson inspector edits every registered authoring dimension, and the shared simulation component delivers the same security, telemetry, lifecycle, and semantic- equivalent contract in author preview and learner rehearsal.
      • Browser: all 33 lesson-editor and all nine learner-delivery Chromium tests pass serially, including repository automatic axe checks; the simulation sandbox, exact authoring write, and learner contract also pass a nine-test Chromium/Firefox/WebKit matrix. Tests cover typed scenarios/ state/telemetry/equivalence, opaque sandbox headers and permissions, scaffolds/responses/observations/reset/cancel/teardown, no protected material, local-only equivalent response, and accessible outcome IDs.
      • Scope remains explicit: browser API transport and dossier persistence use deterministic fixtures while the actual cross-origin sandbox executes in real browser engines. This does not prove production deployment, representative learner usability, instructional validity, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, security, accessibility, regression, and cross-browser gates pass. -->
  • M3.21 Build media regeneration and surgical replacement by segment, figure, narration range, caption, translation, or rendition while retaining immutable prior artifacts and dependency invalidation. Evidence: unchanged segments keep stable identities and approvals are recomputed.
    • M3.21.a Define replacement request with exact project/package/artifact revision, typed segment/figure/audio range/caption/translation/rendition target, requested change, reason, and expected version. Evidence: schema. <!-- evidence: metis-media-replacement.json .exactRequest; the closed target union binds one component to exact project, package, package revision/hash, specification hash, artifact ID/revision/hash, requested change, reason, and a bounded start/end witness for narration ranges. -->
    • M3.21.b Resolve minimal affected source/spec/code/data/render stages and downstream dependency/invalidation graph before execution. Evidence: impact preview tests. <!-- evidence: .impactPreview; a read-only content-addressed preview resolves exact dependency edges, retained components, affected stages, gate domains, approval receipts, and prior playback before any revision is written. All six targets assert exact stage/gate sets and dependent counts in service tests, and Playwright inspects the educator preview. -->
    • M3.21.c Regenerate/edit only the authorized target with compatible boundaries/timing/layout/style/voice/locale/format and new artifact identity. Evidence: surgical replacement tests. <!-- evidence: .surgicalReplacement; the service rejects zero, multiple, renamed, stale, cross-component, or boundary-drifting changes. Exactly one target changes, six compatibility locks remain true, and the new artifact receives a content-addressed identity at revision n+1. -->
    • M3.21.d Retain unchanged component identities/hashes/revisions and all prior artifacts; never splice over immutable package history. Evidence: stability tests. <!-- evidence: .immutableRetention; unchanged manifest entries and gate receipts remain byte-identical, replacement/acceptance/gate/assembly actions append hash-chained package revisions, and the complete prior artifact plus playback witnesses remain inspectable in history. -->
    • M3.21.e Recompute affected source/correctness/consent/rights/ accessibility/technical/critic gates and mark dependent approvals/builds stale precisely. Evidence: invalidation tests. <!-- evidence: .preciseRegating; only target-dependent receipt domains become stale, unaffected consent/rights/source/correctness/media gates remain current when their inputs are unchanged, affected receipt IDs are audited, and gate execution is blocked until explicit acceptance. Tests prove unaffected receipts survive byte-for-byte while affected receipts are recomputed against the replacement specification. -->
    • M3.21.f Assemble a new package revision, compare old/new components, require educator/reviewer acceptance, and verify learner playback. Evidence: replacement E2E. <!-- evidence: .acceptanceAndPlayback; educator/admin acceptance is an attributed immutable revision, the workbench compares before/after artifact revisions and preserved history, learner lookup serves the last accepted package throughout review, and accepted/regated media assembles a distinct learner package containing the replacement caption. --> <!-- completion evidence:
      • Deterministic generator and nine self-checks: evidence/v1-workbenches/metis-media-replacement.json via pnpm run verify:inventory:v1-metis-media-replacement. It content-addresses schemas/helpers/services/API/OpenAPI/generated-client/workbench/unit/ browser/TODO inputs and maps M3.21.a-f to executable evidence.
      • Backend: 112 affected teaching-media, media-loop, prerecorded-lecture, lecture-authoring, API-contract, and OpenAPI tests pass. Changed files pass Ruff and four focused mypy source checks; live/checked-in OpenAPI and generated TypeScript client drift checks pass. Package revisions reuse the append-only hash-chained JSON body with no parallel payload table; forward migration 012 expands the ledger's database state/action constraints for replacement_review, replace, and accept_replacement.
      • Frontend: all 56 Vitest files/209 tests, ESLint with zero warnings, TypeScript, and changed-file Prettier checks pass. The restrained workbench exposes exact target/artifact selection, read-only blast radius, immutable before/after comparison, explicit acceptance, precise regating, and preserved revision inspection without exposing protected material.
      • Browser: all 34 lesson-editor and all nine learner-delivery Chromium tests pass serially with repository automatic axe checks. Authoring, assembly, and replacement also pass a nine-test Chromium/Firefox/WebKit matrix. The replacement E2E proves exact witnesses, four caption- dependent stale gates, disabled pre-acceptance regating, retained r3 playback during review, accepted recomputation, new r7 learner playback, and the replacement caption without answer/secret leakage.
      • Scope remains explicit: browser API transport and service persistence use deterministic fixtures/repositories. This does not prove production deployment, representative educator or learner usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, security, accessibility, regression, and cross-browser gates pass. -->

M4 — Assessment and item-bank authoring#

  • M4.1 Build item-bank search by kind, objective, standard, difficulty, discrimination, exposure, locale, status, author, gate, and usage. Evidence: permission-aware query tests.
    • M4.1.a Define searchable exact item-revision projection and allowlisted filters/sorts/facets for every named dimension. Evidence: search schema.
    • M4.1.b Index answer-safe metadata with tenant/course/role/exposure field controls and explicit stale/unknown psychometric state. Evidence: index tests.
    • M4.1.c Implement deterministic cursor pagination, query limits, stable ranking, and permission-safe counts/suggestions. Evidence: query tests.
    • M4.1.d Present result/version/use/gate/rights/locale/accessibility state and authorized preview without key/rationale leakage. Evidence: UI tests.
    • M4.1.e Test direct foreign IDs, hidden filters, facets/counts/timing, protected fields, and two tenants. Evidence: authorization suite. <!-- M4.1 completion evidence (2026-08-31):
      • Reproducible provenance: evidence/v1-workbenches/ metis-item-bank-search.json, generated and drift-checked by pnpm run verify:inventory:v1-metis-item-bank-search, maps M4.1.a-e to executable source, schema, contract, authorization, and browser evidence.
      • Projection and persistence: migration 013 adds an append-only exact (item_id, revision) search projection with content-addressed lineage, a private source-change witness, explicit tenant/course/author/use state, and no answer, key, rationale, option, explanation, or hint columns. Unknown/stale/current discrimination and exposure value semantics are enforced in both Pydantic and PostgreSQL constraints. Legacy assessment question mutations, lifecycle changes, and retirement append projection revisions instead of rewriting history.
      • Query and authorization: authenticated instructor/admin search supports every named M4.1 dimension through closed request models, six stable total-order sorts, thirteen permission-safe facets, bounded suggestions, a 50-row page cap, a 120-character query cap, and a 5,000-document scope cap. HMAC cursors bind offset, normalized query, role, tenant, and the exact role-visible snapshot. Latest revision is resolved before role visibility so a newly hidden revision cannot fall back to an older visible one. Foreign-tenant changes do not perturb local snapshots or counts.
      • Direct lookup and leakage boundary: exact preview returns one generic 404 contract for foreign, admin-only, and missing revisions and traverses the same complete bounded tenant snapshot for each case. Search, facets, counts, suggestions, preview, OpenAPI, generated clients, and the browser response guard exclude protected fields; exact exposure remains administrator-controlled.
      • Workbench: following frontend-skill, the instructor surface uses a restrained warm-paper editorial register rather than a generic card grid. It exposes all named search dimensions, permission policy, safe facets and suggestions, exact version/use/gate/rights/locale/accessibility state, signed-page continuation, learner denial, responsive reflow, and an answer-safe exact-revision preview.
      • Automated verification: 61 focused item-bank/OpenAPI service tests and 18 assessment-service regression tests pass together (79/79); Ruff and focused Python 3.12 mypy pass; checked-in OpenAPI and generated TypeScript drift checks pass; all 57 Metis web unit files (214 tests), full web lint, web typecheck, and changed-file Prettier checks pass. Twelve item-bank Playwright flows pass serially across Chromium, Firefox, and WebKit with the repository automatic serious/critical Axe scan, including all filters, usage facet application, exact safe preview, signed pagination, learner denial before transport, and 390px no-overflow behavior.
      • Live PostgreSQL: a clean upgrade through 013, a real 012 -> 013 legacy backfill containing a private answer/rationale/hint, both semantic constraint rejection probes, and an answer-only update/appended revision-2 search smoke pass. The projected prompt is normalized correctly, protected projection-column count is zero, and only safe latest revision 2 is serialized.
      • Scope remains explicit: browser API transport uses deterministic fixtures. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, security, accessibility, responsive, and cross-browser gates pass. -->
  • M4.2 Build zero-form item capture plus structured completion. Evidence: drafts retain provenance and validation state.
    • M4.2.a Define minimal capture for prompt/idea/source/objective/item-kind hint/author/context with offline/local draft identity. Evidence: capture schema.
    • M4.2.b Create a canonical draft revision with author/source/import/AI provenance and explicit incomplete/unvalidated status. Evidence: creation tests.
    • M4.2.c Infer no protected answer or item semantics automatically; route the draft into kind-specific structured fields and requirements. Evidence: boundary tests.
    • M4.2.d Show completion checklist for stem/response/key/scoring/rationale/ objective/accessibility/source/review/gates. Evidence: completion tests.
    • M4.2.e Preserve original capture and every structured edit/validation result through revision history. Evidence: provenance tests. <!-- M4.2 completion evidence (2026-08-31):
      • Reproducible provenance: evidence/v1-workbenches/ metis-item-draft-capture.json, generated and drift-checked by pnpm run verify:inventory:v1-metis-item-draft-capture, maps M4.2.a-e to executable schema, persistence, API, authorization, offline, browser, and live-database evidence. All eight generator self-checks pass and every input is content-addressed.
      • Minimal capture and offline identity: the closed 4.2.0 request requires only local_draft_id; prompt, idea, source references, objective IDs, explicit canonical item-kind hint, and context remain independently optional. The instructor workbench writes an actor-bound local-item-draft:<UUID> recovery record before transport, preserves exact literal text, retains unsynchronized edits, and replays the same identity and capture after reconnect. Local identity reuse with different content fails with an explicit 409 rather than overwriting the original.
      • Canonical provenance and storage: migration 014 creates one immutable original-capture/provenance root and a separate append-only structured revision ledger. Author identity is derived from authentication; source/import/AI provenance is typed and explicit; capture and provenance have independent SHA-256 witnesses. Revision 1 is explicitly incomplete and unvalidated, every revision has request and content witnesses, and PostgreSQL constraints enforce contiguous lineage, idempotency, event, status, and validation state.
      • Protected boundary and routing: capture and structured OpenAPI schemas, ORM columns, generated clients, service responses, local storage, and UI transport contain no answer, key, option, explanation, hint, or rationale payload property. Initial structure copies only the literal prompt, objectives, and sources. Wording such as “multiple choice” or “correct answer” does not select a kind or populate response/scoring/protected fields; only an explicit hint routes to the canonical selected-response, written-response, executable-response, interactive-response, or performance-evidence family, with inference_used=false.
      • Structured completion and revision history: the workbench renders the exact ten-field stem/response/key/scoring/rationale/objective/ accessibility/source/review/gates checklist. Unknown kinds block response, scoring, key, and rationale; explicit kinds expose answer-safe response and scoring fields while key/rationale stay blocked for their protected kind editor (or key becomes not applicable where the registry says so). Structured edits reset validation to unvalidated; explicit validation appends its exact checklist and findings. Reads reverify original, provenance, revision hashes, and the complete contiguous chain.
      • Database immutability: the live clean migration chain was paused at 013 (both draft tables absent), upgraded to 014, and exposed both expected tables plus two mutation-rejection triggers. Real async PostgreSQL capture, structured edit, and validation commits produced three linked revisions; direct capture UPDATE and revision DELETE probes were rejected by the database trigger. The checked-in migration downgrade/upgrade cycle and head 014 pass.
      • Automated verification: all 16 focused M4.2 backend tests and the complete 80-test affected M4.1/M4.2/OpenAPI service set pass; checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Ruff passes 12 affected Python/migration/test files and strict mypy passes five source files. All 58 Metis web unit files (220 tests), full ESLint with zero warnings, and TypeScript pass. Twelve capture/completion Playwright flows pass serially across Chromium, Firefox, and WebKit with the repository automatic serious/critical Axe scan, covering zero-form creation, structured revision and validation, answer-like prose non-inference, offline exact-identity replay, learner denial before transport, revision history, and 390px no-overflow behavior. The adjacent twelve-flow item-bank search matrix also passes across all three engines after capture navigation was added.
      • Workbench design: following frontend-skill, the new surface extends the restrained warm-paper assessment studio with an editorial capture ledger, immutable witness, structured writing column, sticky completion register, protected-boundary notice, and revision folio rather than a generic card dashboard. Search and course-studio navigation link the capture flow.
      • Scope remains explicit: browser API transport uses deterministic fixtures. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, security, accessibility, offline, responsive, and cross-browser gates pass. -->
  • M4.3 Build item-variant generation behind the fenced provider seam. Evidence: not_configured, malformed, refusal, and budget states.
    • M4.3.a Define variant request/set/proposal with exact source item, permitted transformations, invariants, count/diversity, provider provenance, estimate, and budget. Evidence: schemas.
    • M4.3.b Fence source/stem/objectives/rubric/allowed key context and exclude learner/integrity/unrelated bank data from provider input. Evidence: context tests.
    • M4.3.c Validate generated kind/schema/key/rationale/objective/difficulty/ locale/accessibility/source and similarity constraints. Evidence: proposal validation tests.
    • M4.3.d Preserve not_configured, refusal, malformed, timeout, partial, moderation, and budget outcomes without fallback content. Evidence: seam tests.
    • M4.3.e Require per-variant human accept/edit/reject and create new attributed item revisions; no automatic bank admission. Evidence: review tests. <!-- M4.3 completion evidence (2026-08-31):
      • Reproducible provenance: evidence/v1-workbenches/ metis-item-variant-generation.json, generated and drift-checked by pnpm run verify:inventory:v1-metis-item-variant-generation, maps M4.3.a-e to executable contracts, provider fencing, validation, persistence, authorization, browser, and live-database evidence. All eight generator self-checks pass and every input is content-addressed.
      • Exact request, estimate, and provenance: closed 4.3.0 schemas bind one content-addressed source draft revision to an explicit transformation set, hard invariants, one-to-six count, exact diversity specifications, difficulty/locale targets, server-modeled estimate, educator-confirmed maximum budget, and immutable provider/model/config/prompt/request/ response/token/cost/trust-zone witnesses. Preparation records its exact count and total estimate; count changes refetch preparation, while stale estimate or pricing witnesses fail with an explicit 409 before model invocation. The provider is disabled unless explicitly configured.
      • Provider fence: the only provider-visible product context is the exact source kind/stem/response/scoring/objectives/difficulty/locale/ accessibility/source surface plus optional server-authorized key context, confirmed controls, and educator guidance. Each surface is rendered as a labeled untrusted non-instruction zone; tools and browsing are absent. Learner data, integrity telemetry, unrelated bank content, credentials, and tools are enumerated exclusions. Deterministic HTTP transport tests prove learner-secret and unrelated-bank sentinels never enter the request.
      • Item validation: Pydantic rejects extra or malformed provider properties; the service independently revalidates kind, ordered objectives, protected key/rationale authorization, scoring, source references, difficulty, locale, required accessibility text, transformation/diversity/count, and moderation. Source and pairwise Jaccard similarity are recomputed locally; kind, key, objective, locale, count, source-copy, and similarity drift probes fail closed.
      • Explicit outcomes without fallback: not_configured, unavailable, refusal, malformed envelope/payload, timeout, partial output, moderation block, and preflight/actual budget-exceeded outcomes are typed and stored with zero silent candidates for terminal failures. Partial content is a bounded subset, never padded. The browser executes and labels all eight terminal states while retaining invocation cost, tool state, and receipt.
      • Human review and storage: migration 015 separates immutable generation sets, unaccepted proposals, and one append-only decision per proposal. Accept, human edit, and reject are distinct closed requests with a required human reason. Accept/edit create a content-addressed revision-1 record attributed to source draft/revision, proposal, provider/model, and actor; reject creates none. Every accepted revision is explicitly not_admitted; no item-bank projection write exists in the migration or service. PostgreSQL checks bind action/edit/revision semantics.
      • Live PostgreSQL: a clean chain paused at 014, upgraded to 015, downgraded, and upgraded again. Real async capture/edit, exact preparation, generation, proposal, and accept commits passed; the item-bank projection count remained unchanged. Direct set UPDATE, proposal DELETE, and decision UPDATE probes were rejected by all three immutability triggers. The disposable test database was removed afterward.
      • Automated verification: all 12 focused M4.3 tests including live PostgreSQL pass, and the complete 92-test affected M4.1/M4.2/M4.3/OpenAPI service set passes. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Ruff passes all 16 changed Python/migration/test files and strict mypy passes five new source modules. All 59 Metis web unit files (222 tests), full ESLint with zero warnings, and TypeScript pass. Nine item-variant Playwright flows pass serially across Chromium, Firefox, and WebKit with the repository automatic serious/critical Axe scan, covering exact count/source/budget binding, accept/edit/reject, all provider outcomes, learner denial before transport, and 390px no-overflow. The adjacent twelve-flow item-draft matrix also passes across all three engines after variant navigation was added.
      • Workbench design: following frontend-skill, the educator-only surface uses an editorial assessment atelier: warm-paper planning controls beside a dark exact-source folio, a compact source witness, explicit exclusions and estimate, and a numbered proposal ledger rather than generic cards. Every proposal keeps its own rationale and decision controls, and accepted results visibly remain “not admitted.”
      • Scope remains explicit: provider HTTP and browser API transport use deterministic doubles; no production model credential was invoked. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32- GiB host approval gate; affected compile, type, lint, unit, contract, database, security, accessibility, responsive, and cross-browser gates pass. -->
  • M4.4 Build selected-response editor with keyed answer and distractor rationales. Evidence: answer uniqueness and protected-field tests.
    • M4.4.a Author stem/instructions/options with stable option IDs, rich accessible content, locale variants, ordering/randomization, and source refs. Evidence: editor tests.
    • M4.4.b Configure single/multiple/ordered/matching response rules and exact key/partial-credit scoring with uniqueness/cardinality validation. Evidence: answer tests.
    • M4.4.c Author correct rationale and each distractor’s misconception/ rationale separately from learner feedback. Evidence: rationale tests.
    • M4.4.d Preview author/key/learner modes with randomized order preserving stable key identity and accessibility. Evidence: preview tests.
    • M4.4.e Enforce key/rationale property authorization across API/UI/search/ export/log/tutor and direct crafted payloads. Evidence: protected-field tests. <!-- M4.4 completion evidence (2026-08-31):
      • Canonical contracts and scoring: schema 4.4.0 closes every request, localized rich-content surface, stable option:<UUID> and target:<UUID> identity, source reference, randomization/fixed-position rule, and protected key/rationale object. Conditional validators bind single, multiple, ordered, and one-to-one matching keys to unique known IDs and valid cardinality. Exact scoring plus bounded multiple per-correct-less-incorrect, ordered per-position, and matching per-pair partial credit is computed server-side from stable IDs. Tests execute all four kinds, incorrect cardinality/foreign identity, locale completeness, duplicate identity, and closed crafted-property rejection.
      • Feedback and rationale boundary: learner feedback remains an option-local localized surface. Correct rationale and each distractor misconception/ localized rationale live in separate protected properties; choice items require one review for every and only non-key option. The TypeScript builder and browser assertions prove neither representation absorbs the other. Ordered and matching items reject distractor-review payloads.
      • Role and downstream projections: author, key, and learner previews are distinct discriminated response models. Only key mode can contain the response rule, correct rationale, or distractor reviews; learner mode omits feedback before submission and structurally has no protected fields. Randomization derives only from revision witness, attempt seed, and stable identity, while unique first/last positions stay fixed. A separate closed projection for learner, search, safe export, log, and tutor surfaces enumerates excluded properties and cannot serialize key or rationale. Protected export is a separate explicitly labeled endpoint.
      • Authorization and immutable persistence: all preparation, full-history, append, preview, scoring, and protected-export HTTP paths require the instructor/admin dependency before service construction. Parameterized API tests send learner requests to all seven paths, including malformed crafted answer-key bodies, and prove 403 occurs before service access. Migration 016 stores full authoring revisions behind exact draft and optimistic editor witnesses, content/request/revision hashes, contiguous lineage, role constraints, idempotency, and a database UPDATE/DELETE rejection trigger. It never writes the item-bank search projection.
      • Live PostgreSQL: a clean chain upgraded through 015, paused and upgraded to 016, downgraded to 015, and re-upgraded to head. Real async capture/edit/prepare/append/preview/project/score commits passed. Learner and key modes retained identical stable IDs; five safe projections were key-free; item-bank projection count was unchanged; direct authoring-row UPDATE and DELETE probes were rejected. The disposable database was removed afterward.
      • Automated verification: all 27 focused M4.4 tests including live PostgreSQL pass, and the complete 119-test affected M4.1–M4.4/OpenAPI service set passes. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Ruff passes the 12 focused Python/migration/test files and strict mypy passes the four new source modules. All 60 Metis web unit files (228 tests), full ESLint with zero warnings, and TypeScript pass. Nine selected-response Playwright flows pass serially across Chromium, Firefox, and WebKit with the repository automatic serious/ critical Axe scan, covering complete authoring, feedback/rationale separation, bilingual completeness, stable-key role previews, learner denial before transport, deterministic randomization, and 390px no-overflow. The adjacent twelve-flow capture matrix also passes across all three engines after the workbench link was added.
      • Workbench design: following frontend-skill, the educator-only page uses a restrained editorial composition: warm-paper numbered authoring ledger, compact exact-source/locale bands, visually fenced protected reviews, and one sticky dark role-preview rail rather than generic card grids. Matching pairs can be remapped independently, ordered keys can be moved, fixed positions are explicit, and every preview exposes the short stable identity used across randomization.
      • Generated evidence: evidence/v1-workbenches/metis-selected-response-authoring.json is content-addressed over implementation, tests, generated contracts, UI, TODO, and evidence producer/test inputs; its drift and semantic tests pass. Scope remains explicit: browser transport uses deterministic doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, security, accessibility, responsive, and cross-browser gates pass. -->
  • M4.5 Build constructed-response and essay editor with scoring guide, exemplar, acceptable variants, and moderation requirements. Evidence: contract tests.
    • M4.5.a Author prompt/context/response format, length/language/media/ citation constraints, instructions, sources, and accessibility. Evidence: prompt tests.
    • M4.5.b Define acceptable answer variants/elements/tolerances and unacceptable/missing evidence without reducing essays to hidden exact strings. Evidence: response model tests.
    • M4.5.c Attach rubric/scoring guide revision, criterion mappings, partial/ holistic method, and human/model-assisted scoring bounds. Evidence: scoring tests.
    • M4.5.d Attach authorized annotated exemplars at performance levels with rights/redaction and protected learner identity. Evidence: exemplar tests.
    • M4.5.e Define first/second marking, blind review, discrepancy thresholds, moderation/reconciliation, and audit. Evidence: moderation tests. <!-- completion evidence (2026-08-31):
      • Contract and semantic boundary: schema 4.5.0 admits only constructed response and essay kinds. Localized rich prompt, context, instructions, editor labels, format guidance, and media-alternative guidance are complete across every authored locale. Response formats, character/word bounds, languages, document MIME/byte limits, learner media permission, citation policy/minimum/styles, governed source references, keyboard-only support, source-description verification, and assistive-technology review are closed and cross-validated.
      • No hidden essay key: the protected response model is explicitly based on stable evidence elements, acceptable semantic variants and tolerance guidance, plus both unacceptable and missing-evidence boundaries. The schema has no exact-answer property, carries hidden_exact_strings_prohibited: true, rejects crafted extra answer fields, unknown element references, and non-semantic essay tolerances.
      • Scoring: exact rubric ID/revision/SHA-256 witnesses and criterion-to-element mappings support analytic partial or holistic judgment. Analytic points must sum to the maximum; holistic mappings cannot smuggle point keys. Human final judgment is mandatory. Model help is disabled or advisory, cannot finalize, is capped at 3,000 basis points, and requires explicit confidence and evaluation-profile bounds.
      • Exemplars and identity: at least two distinct performance levels require localized response content, criterion annotations, explicit rights basis, holder, license/permission reference, authorizer and timestamp. Verified direct/indirect redaction and literal absence of protected learner identity fail closed in both schema and browser payload tests.
      • Marking and moderation: first marking is required; configured second marking requires at least two blind markers and a positive discrepancy threshold. The executable moderation planner returns awaiting, complete, or reconciliation-required states, computes basis-point discrepancy, routes to lead/third/consensus reconciliation, and enumerates append-only first/second mark, model recommendation, human override, reconciliation, and final-decision audit events.
      • Protected surfaces and authorization: author, marker, and learner previews are discriminated. Learner plus search, safe export, log, and tutor projections structurally omit answer models, scoring, criteria, exemplars, moderation, model assistance, and marker identity. Protected export is a separate labeled route. All seven preparation/history/append/preview/ projection/export/moderation HTTP paths require instructor/admin authority before service construction; parameterized crafted learner requests prove 403 occurs before protected service access.
      • Persistence and live PostgreSQL: migration 017, chained from 016, stores complete authoring revisions with exact source/editor witnesses, content/request/revision hashes, contiguous lineage, role constraints, idempotency, and an UPDATE/DELETE rejection trigger. A disposable database passed clean 001→016→017, 017→016, and re-upgrade-to-head paths. Real async capture/edit/prepare/append/author-marker-learner preview/five safe projections/moderation commits passed; item-bank search count was unchanged; direct UPDATE and DELETE were rejected. The test-only database was removed afterward.
      • Automated verification: all 26 focused M4.5 tests including live PostgreSQL pass; the consolidated affected M4.1–M4.5/OpenAPI service set passes 145 tests. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Ruff passes 14 focused Python/migration/test files and strict mypy passes the four new source modules. All 61 Metis web unit files (232 tests), full ESLint with zero warnings, and TypeScript pass. Nine written-response Playwright flows pass serially across Chromium, Firefox, and WebKit with the repository automatic serious/critical Axe scan, covering complete authoring, semantic-no-exact-key payloads, scoring/model bounds, rights/redaction, blind moderation, bilingual completeness, author/marker/learner field separation, learner denial before transport, and 390px no-overflow. The adjacent twelve-flow capture matrix also passes across all engines after adding the routed workbench link.
      • Workbench design: following frontend-skill, the educator-only surface extends the restrained assessment editorial system: exact-source strip, dark locale ledger, numbered paper authoring sequence, protected semantic/ model/exemplar insets, and one sticky dark role-preview rail. The layout avoids generic dashboard cards and reuses the established responsive and accessible selected-response visual language for a coherent suite.
      • Generated evidence: evidence/v1-workbenches/metis-written-response-authoring.json is content-addressed over implementation, tests, generated contracts, UI, TODO, and its producer/test inputs; its drift and semantic tests pass. Scope remains explicit: browser API transport uses deterministic doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, security, accessibility, responsive, and cross-browser gates pass. -->
  • M4.6 Build math and code item editors with executable/renderer sandboxes. Evidence: resource, network, and escape tests.
    • M4.6.a Author semantic MathML/symbolic/numeric/unit/graph/proof response models with equivalence/tolerance/assumptions and accessible rendering. Evidence: math item tests.

    • M4.6.b Author code language/runtime/dependencies/starter files/API, stdin/arguments/output, visible/hidden tests, scoring, and limits. Evidence: code item tests.

    • M4.6.c Pin renderer/compiler/interpreter/CAS/test harness/container/ dependency versions and deterministic seeds. Evidence: runtime contracts.

    • M4.6.d Execute author preview in isolated CPU/memory/time/process/file/ network/syscall-limited sandbox with teardown. Evidence: sandbox tests.

    • M4.6.e Protect hidden tests/answers/runtime secrets and test infinite loops, fork/memory bombs, filesystem/network/escape, and malicious output. Evidence: adversarial suite. <!-- M4.6 completion evidence (2026-08-31):

      • Contracts and access boundaries: Metis now owns a strict 4.6.0 discriminated contract for all six requested math response forms and executable code. Math records semantic protected answers with verified SHA-256 witnesses, versioned equivalence rules, assumptions, explicit tolerances, graph/proof constraints, localized accessible alternatives, exact renderer/graph/CAS pins, and an opaque no-script renderer policy. Code records exact language/runtime, optional compiler, preinstalled dependency integrity, starter files, invocation/API/I/O, visible and hidden tests, partial scoring, resource policy, and runtime-secret names. Compiler is explicitly not applicable for the one admitted interpreted Python runtime. Instructor/admin authorization precedes all eight routes; learner/search/export/log/tutor payload types structurally omit canonical answers, hidden tests, and runtime secrets, while a separate protected export retains them for authorized authoring and evaluation.
      • Execution isolation: author preview admits only CPython 3.12.14 in python@sha256:d09d15e60962ca365d1cd544a48773bac9d33f2fb1b00f2aa0deec78ade7dc31, with a byte-verified checked-in launcher and default-deny seccomp policy. Docker is invoked with pulls disabled, no network, read-only root and submission mounts, an isolated bounded tmpfs, unprivileged UID/GID, every capability dropped, no-new-privileges, process/memory/CPU/wall/ output/open-file bounds, deterministic hash seed, no secret forwarding, one disposable container per visible case, forced removal, and verified absence after teardown. Authored image/profile/tool substitution is rejected before process launch.
      • Adversarial and persistence evidence: ten live Docker cases pass for normal exact-image execution, no secret inheritance, network and host filesystem denial, namespace/syscall denial, writable-file and process bounds, infinite loop, memory bomb, output bomb, malicious ANSI/control/ HTML output neutralization, and teardown. Migration 018, chained from 017, passed clean 001→018, 018→017, and re-upgrade paths on disposable PostgreSQL; full protected state persisted, automatic item-bank search writes stayed at zero, and direct UPDATE/DELETE were rejected. Test-only databases were removed after verification.
      • Automated verification: all 44 focused M4.6 service/API/migration/live PostgreSQL/live Docker tests pass; the consolidated affected M4.1–M4.6 and OpenAPI service set passes 186 tests. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 17 Python/ migration/test/runtime files and strict mypy covers five source modules. All 62 Metis web unit files (244 tests), full ESLint with zero warnings, and TypeScript pass. Nine math/code Playwright flows pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis on the parent workbench, covering exact runtime/test/scoring payloads, protected execution requests, proof semantics, opaque MathML CSP/sandbox/ role/name contracts, instructor-only transport, bilingual completeness, saved-revision rehydration, and 390px no-overflow. The iframe intentionally rejects script injection, so Axe cannot enter it; its security and accessibility contract is asserted directly. The adjacent twelve-flow capture matrix also passes after adding the routed workbench link.
      • Workbench design: following frontend-skill, the educator-only surface extends the restrained assessment editorial system with an exact-source strip, dark locale ledger, numbered paper authoring sequence, protected hidden-suite inset, pinned-runtime and sandbox controls, semantic learner/ author preview, and one sticky execution-receipt rail rather than a dense generic dashboard.
      • Generated evidence: evidence/v1-workbenches/metis-math-code-authoring.json is content-addressed over implementation, checked-in runtime policy, tests, generated contracts, UI, TODO, and its producer/test inputs; drift and semantic tests pass. Browser API transport uses deterministic doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, formal accessibility conformance, or production evaluation-engine calibration. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, security, accessibility, responsive, and cross-browser gates pass. -->
  • M4.7 Build media and interactive item editor with accessible alternatives. Evidence: QTI export/import and accommodation preview.
    • M4.7.a Author governed image/audio/video/3D/stimulus media revisions, interactions, response regions/timing/state, scoring hooks, and sources. Evidence: media item tests.

    • M4.7.b Author captions/transcripts/descriptions/sign alternatives/ keyboard paths/noninteractive equivalents and support requirements. Evidence: accessibility tests.

    • M4.7.c Configure sandbox runtime, typed events/state/reset/seed/resource/ network limits, and secure response capture. Evidence: interactive tests.

    • M4.7.d Preview permitted accommodation/locale/device/input modes with the real delivery renderer. Evidence: accommodation preview matrix.

    • M4.7.e Map/import/export through the QTI profile or return exact extension/ loss/incompatibility with round-trip tests. Evidence: QTI suite. <!-- M4.7 completion evidence (2026-08-31):

      • Governed authoring contract: Metis now owns a strict 4.7.0 model for media, simulation, and interaction items. Image, audio, video, 3D, and stimulus bindings carry immutable asset and rendition revisions and hashes plus delivery locator, authority, source, rights, and provenance. Normalized response regions, timing/attempt bounds, typed state and event schemas, exact evaluator pins, expected values, and partial-credit hooks are cross-validated; scoring totals exactly 10,000 basis points and the capture allowlist equals the authored response-event set.
      • Accessibility and safe disclosure: every locale must resolve alternative text, long description, transcript, captions, audio description, and sign-language support as either a governed reference or an explicit reviewer/time-stamped not-applicable decision. Keyboard paths, noninteractive equivalents, support requirements, reduced-motion equivalents, and screen-reader review are required. Learner/search/ export/log/tutor projections structurally omit scoring hooks, expected values, signer references, reviewer identities/times, and internal not-applicable reasons; protected export is a separate instructor/admin route. Blank added locales cannot be saved until every surface and alternative is completed.
      • Runtime and response capture: the sealed contract pins a delivery engine, exact bundle digest, deterministic seed, reset policy, HTTPS network allowlist, opaque origin, and CPU/wall/memory/process/writable/event/ payload limits while denying same-origin authority, popups, top navigation, and persistent storage. Preview capture admits only exact typed events and bounded payloads tied to the current revision, then returns an HMAC-SHA256 receipt without learner identity, freeform telemetry, or key exposure. A configured server secret (with the existing server JWT secret as the deployment fallback) provides stable signing; the service fails closed if no signer is supplied.
      • Delivery and QTI evidence: the educator workbench previews locale, desktop/tablet/mobile, keyboard/pointer/touch/switch, full/reduced motion, standard/high contrast, default/captions/transcript/audio-description/ sign-language modes, and online/low/offline bandwidth through the shared MediaInteractiveDelivery learner component and records one evidence decision per dimension. QTI 3.0 Metis extension export includes a checksum-bound full typed sidecar and immediately reimports it to an exact content hash; QTI 3.0 core returns the exact loss paths; QTI 2.2 returns an explicit incompatibility and no misleading artifact.
      • Persistence and automated verification: migration 019, chained from 018, passed clean 001→019, 019→018, and re-upgrade paths on disposable PostgreSQL. Full content persisted, automatic item-bank projection writes remained at zero, and direct UPDATE/DELETE were rejected. All 31 focused M4.7 service/API/migration/live-PostgreSQL tests pass; the consolidated affected M4.1–M4.7 and OpenAPI set passes 210 tests. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 15 Python/migration/test files and strict mypy covers six source modules. All 63 Metis web unit files (248 tests), full ESLint with zero warnings, and TypeScript pass. Nine media/interactive Playwright flows pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis, covering complete payloads, safe real-renderer accommodation switching, signed capture, QTI exact/loss/incompatibility, learner denial, translation completeness, and 390px no-overflow. The adjacent twelve-flow capture matrix also passes after adding the routed workbench link.
      • Workbench design: following frontend-skill, the educator-only surface extends the restrained assessment editorial system with an exact-source strip, numbered authoring ledger, dense but linear governed-media table, explicit alternative decisions, one sticky dark delivery rail, and a contour-map stimulus stage. The first automated accessibility run exposed inherited low-contrast rail evidence; its component colors were corrected before the three-engine gate passed.
      • Generated evidence: evidence/v1-workbenches/metis-media-interactive-authoring.json is content-addressed over implementation, tests, generated contracts, UI, TODO, and producer/test inputs; its drift and semantic tests pass. Browser API transport uses deterministic doubles, and QTI round-trip is in-process rather than external-LMS certification. This does not prove production deployment, representative educator usability, physical assistive- technology behavior, formal accessibility conformance, or third-party QTI certification. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The three disposable verification databases were removed after use. -->
  • M4.8 Build stimulus-set authoring. Evidence: shared-stimulus edits mark every dependent item stale.
    • M4.8.a Build immutable stimulus create/edit/version/localize/source/ accessibility/rights workflow separately from items. Evidence: authoring tests.

    • M4.8.b Build set identity, ordered stimulus revisions, shared directions/ context, item membership/order, and presentation rules. Evidence: set schema tests.

    • M4.8.c Attach items by exact revision/role/locator and prevent cross- tenant, unauthorized, circular, or answer-exposing links. Evidence: link tests.

    • M4.8.d Preview item/set combinations and detect conflicting directions, locale, accessibility, exposure, layout, and target constraints. Evidence: preview tests.

    • M4.8.e On stimulus/set revision, compute dependent item/assessment/build blast radius and mark review/gates stale without rewriting them. Evidence: invalidation tests. <!-- M4.8 completion evidence (2026-08-31):

      • Immutable stimulus workflow: Metis now owns a strict 4.8.0 contract and append-only ledger for stimulus revisions independent of item revisions. Each stimulus records exact locale surfaces, source and rights provenance, governed asset revisions, accessibility supports, reviewed alternatives, author lineage, content hash, and idempotency key. Creating a revision requires an exact optimistic-concurrency pin; partial translations and locale surfaces without matching accessibility alternatives fail closed.
      • Set identity and composition: a separate immutable set ledger preserves stable set identity, exact ordered stimulus revisions, localized shared directions and context, exact ordered item or nested-set members, and presentation constraints for layout, viewport, devices, inputs, locale, and accessibility. The service rejects duplicate or noncontiguous order, no-change revisions, stale editors, missing revisions, and transitive set cycles while retaining complete revision history.
      • Exact authorized graph: item members pin tenant, family, item ID, revision, SHA-256, learner projection, semantic role, and locator. Exact resolution covers selected-response, written-response, math/code, and media/interactive ledgers. Instructor-only API gates run before service dispatch; cross-tenant, missing, unauthorized, circular, and answer-material-bearing references are rejected.
      • Preview and invalidation: exact-revision preview composes shared set context with learner-safe item pins and reports closed conflict codes for directions, locale, accessibility, answer exposure, layout, and target constraints. Revising a shared stimulus appends staleness rows for its current set, item, assessment, and build dependents; revising a set marks its prior item and downstream dependencies stale. Review/gate identities and old revision hashes are retained, and no dependent record is rewritten.
      • Persistence and automated verification: migration 020, chained from 019, passed clean 001→020, 020→019, and re-upgrade paths on disposable PostgreSQL. Three database triggers reject UPDATE/DELETE across the two revision ledgers and staleness ledger. The live PostgreSQL integration observes four exact stale dependency kinds after a shared-stimulus edit and proves the prior set remains unchanged. All 21 focused M4.8 service/API/migration/live-PostgreSQL tests pass; the consolidated affected M4.1–M4.8 and OpenAPI set passes 232 tests, with ten optional sandbox-runtime cases skipped by their existing environment gate. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 16 Python/migration/test files and strict mypy covers five source modules.
      • Authoring studio: following frontend-skill, the new responsive source room separates stimulus and set work, supports bilingual authoring with accessibility/source/rights metadata, pins exact learner-safe items and downstream revisions, previews all six conflict families, and exposes the append-only review blast radius. It is discoverable from both item bank entry points. All 64 Metis web unit files (253 tests), ESLint with zero warnings, and TypeScript pass. Twelve stimulus/set Playwright flows pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis, including bilingual completeness, exact safe pins, conflict reporting, four-kind staleness, no rewrite, learner denial, and 390px reflow. The adjacent twelve-flow item-capture matrix also passes across all three engines.
      • Evidence boundary: the deterministic metis-stimulus-set-authoring.json inventory hashes every implementation, test, contract, UI, and traceability input and asserts M4.8.a–M4.8.e. Browser API transport uses deterministic doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.9 Build item-level objective, standard, concept, and prerequisite linking. Evidence: invalid and cross-tenant refs fail.
    • M4.9.a Define typed link roles/strength/coverage/cognitive level/rationale/ evidence with exact item and target revisions. Evidence: link schemas.

    • M4.9.b Search/select only authorized course/tenant objective, standard, concept, and prerequisite registry entries. Evidence: selector tests.

    • M4.9.c Validate target existence/lifecycle/version, discipline/kind/ grade/locale compatibility, cycles, and conflicting links. Evidence: validation tests.

    • M4.9.d Present links in item, coverage matrix, concept graph, search, and standards export with consistent meaning. Evidence: projection tests.

    • M4.9.e Reject crafted foreign/stale/deprecated/unknown refs and preserve link revision/audit. Evidence: negative tests. <!-- M4.9 completion evidence (2026-08-31):

      • Exact typed links: Metis now owns a strict 4.9.0 contract for a learner-safe exact item revision, exact registry target snapshots, and typed objective, standard, concept, and prerequisite links. Every link records role, strength, basis-point coverage, cognitive classification, rationale, and evidence while preserving item and target revision IDs, content hashes, registry revision, tenant/course scope, and author lineage. Structural validation rejects duplicate targets, incompatible roles, invalid coverage, and any answer-bearing item projection.
      • Authorized selection: the target registry is resolved from current planning-ready curriculum-map revisions and the canonical standards registry. Instructor-only search filters query, target kind, tenant, course, lifecycle, and authorization before returning selectable entries. Objectives, standards, concepts, and prerequisites are all exposed without accepting client-authored registry identities.
      • Closed validation: exact resolution covers selected-response, written-response, math/code, and media/interactive item ledgers. The service re-resolves both item and target snapshots, then rejects missing, unauthorized, foreign-tenant/course, stale, deprecated, unknown, or tampered references. Discipline, grade, locale, item-family, target conflict, registry-cycle, and composed prerequisite-cycle checks all fail closed with stable error codes.
      • Consistent projections and history: one canonical semantic checksum and exact link payload are projected to the item, coverage matrix, concept graph, search, and standards export surfaces, preventing consumers from assigning different meaning to the same revision. Optimistic revision pins, idempotency drift checks, no-change rejection, content and lineage integrity checks, and immutable revision audit preserve every accepted decision, while rejected attempts leave prior revisions untouched.
      • Persistence and automated verification: migration 021, chained from 020, passed clean 001→021, 021→020, and re-upgrade paths on disposable PostgreSQL. Two database triggers reject UPDATE/DELETE on the revision and audit ledgers. The live integration preserves two exact revisions and two audit events and proves both ledgers immutable. All 32 focused M4.9 service/API/migration/live-PostgreSQL tests pass; the consolidated affected M4.1–M4.9 and OpenAPI set passes 265 tests, with ten optional sandbox-runtime cases skipped by their existing environment gate. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 16 Python/migration/test files and strict mypy covers five source modules.
      • Authoring workbench: following frontend-skill, the responsive instructor-only alignment desk combines an authorized registry browser, exact item binding, compact typed link controls, semantic projection comparison, and immutable audit history. It is discoverable from both item-bank entry points. All 65 Metis web unit files (259 tests), ESLint with zero warnings, and TypeScript pass. Twelve M4.9 Playwright flows pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis, covering all target kinds, exact safe saves, projection equality, revision/audit preservation, stale-target draft retention, learner denial, and 390px reflow. The adjacent twelve-flow item-capture matrix also passes across all three engines.
      • Evidence boundary: the deterministic metis-item-learning-links.json inventory hashes every implementation, test, contract, UI, traceability, and producer input and asserts M4.9.a–M4.9.e. Browser API transport uses deterministic doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than- 32-GiB host approval gate; affected compile, type, lint, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.10 Build difficulty, cognitive level, time, and confidence estimates with source provenance. Evidence: estimates are never labeled calibrated before empirical data.
    • M4.10.a Define separate author, rule/model-predicted, expert-reviewed, pilot, and empirically calibrated estimate sources with versions. Evidence: estimate schema.

    • M4.10.b Estimate difficulty, taxonomy+level, response time/range, and per-estimate confidence/uncertainty from exact item/context revision. Evidence: known-value tests.

    • M4.10.c Record method/model/config/corpus/features/assumptions/applicability and human changes; keep different source estimates visible. Evidence: provenance tests.

    • M4.10.d Reserve “calibrated” for ratified empirical sample/method/ uncertainty evidence and reject predicted-to-calibrated promotion. Evidence: status invariant tests.

    • M4.10.e Recompute/preserve history on item/model/data change and surface disagreement/manual review. Evidence: lifecycle tests. <!-- M4.10 completion evidence (2026-08-31):

      • Separate typed sources: Metis now owns a strict 4.10.0 contract for author, deterministic rule/model prediction, expert review, pilot observation, and empirical calibration sources. Each source carries its own identity, version, lifecycle, status, actor, exact item/context hashes, values, provenance, and semantic checksum. Source-kind/status/ evidence combinations fail closed, so authored, predicted, reviewed, pilot-observed, and calibrated states cannot be interchanged.
      • Bounded exact estimates: a deterministic, explicitly untrained and uncalibrated rule profile inspects non-answer features from one exact selected-response, written-response, math/code, or media/interactive item revision in one authorized planning-ready course-map context. It emits ordered basis-point difficulty bounds, Bloom taxonomy/level and measurable verb, ordered response-time seconds, and per-source confidence plus uncertainty. Known-value tests lock the feature vector and all four outputs; learner-safe item and context revision hashes are carried through every prediction and accepted source.
      • Complete visible provenance: every non-calibrated submission records exact method/model/config IDs, versions, hashes, corpus-use status, feature references, assumptions, applicability, and limitations. Expert review pins an exact prior estimate and must declare the complete set of changed dimensions; undeclared or stale changes fail. Rule, author, expert, pilot, and empirical sources remain separately visible, including stale/superseded history, rather than being blended into an unexplained number.
      • Calibration invariant: only a trusted registry record can create an empirical calibrated source. Its ratified evidence requires at least 30 observations, exact data snapshot ID/version/hash and sample size, exact method ID/version/hash, collection window, uncertainty method, exact item/context hashes, and at least two distinct post-collection ratifiers. Registry provenance must match that data snapshot and method. Callers submit only the opaque evidence ID/version; user-authored values or provenance, stale evidence, mismatched registry identity, and predicted-to-calibrated promotion are rejected. The production-safe default registry is empty until a trusted empirical provider is wired.
      • Recalculation, disagreement, and immutable history: item, context, model-profile, pilot-data, and empirical-data triggers are validated against the actual transition. Item/context/model changes stale the affected old source and append a new rule prediction while preserving independent current sources. Difficulty, cognitive-level, and response- time spreads produce explicit manual-review signals. Optimistic pins, idempotency drift/no-change rejection, content/source/revision hashes, and lineage validation preserve every accepted revision and audit event.
      • Persistence and automated verification: migration 022, chained from 021, passed a fresh 001→022 upgrade and a 022→021→022 round-trip on disposable PostgreSQL. Two database triggers reject UPDATE/DELETE on estimate revision and audit ledgers. The live integration preserves two exact revisions, a ratified calibration, and two audit events, and proves both ledgers immutable. All 33 focused M4.10 service/API/migration/live- PostgreSQL tests pass. The consolidated affected M4.1–M4.10 and OpenAPI set passes 289 tests, with 19 existing optional database/sandbox cases skipped by their environment gates. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 16 Python/migration/test files and strict mypy (Python 3.12 target) covers five source modules.
      • Authoring workbench: following frontend-skill, the responsive, instructor-only estimate observatory uses an editorial three-ledger composition for exact scope, bounded source declaration, and source/ history evidence. It exposes all four submission paths, keeps rule and pilot states visibly uncalibrated, accepts empirical records only by registry reference, displays ratifier/sample facts, provenance, disagreement, and immutable audit, and is linked from item-bank, capture, and learning-link entry points. All 66 Metis web unit files (265 tests), ESLint with zero warnings, and TypeScript pass. Eighteen M4.10 Playwright flows pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis; the adjacent 24 capture and learning-link flows also pass across all three engines.
      • Evidence boundary: the deterministic metis-item-estimates.json inventory hashes every implementation, test, contract, UI, traceability, and producer input and asserts M4.10.a–M4.10.e. Browser API transport uses deterministic doubles. This does not prove production deployment, connection of a production empirical provider, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.11 Build distractor plausibility and misconception mapping. Evidence: known-value rules and human review.
    • M4.11.a Define versioned misconception concepts, evidence/source, objective/learner band, locale, lifecycle, and option mapping contract. Evidence: schema tests.

    • M4.11.b Check option uniqueness, grammatical/semantic parallelism, length/ cue imbalance, implausibility, overlap, absolute wording, and correct- answer leakage using known-value rules. Evidence: lint fixtures.

    • M4.11.c Evaluate distractor-to-misconception relevance separately from plausibility and never infer an individual learner misconception solely from a response. Evidence: boundary tests.

    • M4.11.d Present attributed rule/model suggestions with uncertainty and require human accept/edit/reject mappings. Evidence: review tests.

    • M4.11.e Preserve empirical option statistics separately and trigger re- review when distractors/items/misconception vocabularies change. Evidence: lifecycle tests. <!-- M4.11 completion evidence (2026-09-01):

      • Versioned vocabulary and exact mappings: Metis now owns a strict 4.11.0 contract for misconception concept identity, revision, evidence/source identity and hash, objective scope, learner band, locale, lifecycle, correction guidance, indicators, and exact option- to-concept mappings. Optimistic revision pins, semantic and revision hashes, idempotency drift rejection, unique scope values, and exact supersession lineage fail closed; a first revision cannot claim to be superseded without an actual predecessor.
      • Known-value surface evidence: a content-addressed 1.0.0 rule profile always evaluates all eight required categories: normalized option uniqueness, grammatical parallelism, semantic parallelism, length/cue imbalance, overt implausibility, option overlap, absolute wording, and keyed-answer leakage. Findings preserve exact option IDs, severity, evidence, rule identity/hash, and uncertainty. Fixture tests cover every rule family, while bounded plausibility explicitly states that lexical surface checks cannot establish domain plausibility.
      • Relevance and human-review boundary: distractor plausibility and misconception relevance are separate typed results. Rule and submitted model suggestions carry exact method/model/corpus attribution, confidence, uncertainty, concept revision, and a literal human-review requirement. A reviewer must accept the suggestion unchanged, edit it explicitly, or reject it; every suggestion must receive exactly one decision. Keyed options, stale concepts, incomplete decisions, and silent accept-time edits are rejected. Both contracts and UI hard-code that one response neither establishes nor permits an individual learner misconception inference.
      • Separate empirical lifecycle: option counts, rates, point-biserial values, sample/method provenance, collection window, and uncertainty live in a distinct no-inference statistics contract and ledger. Item, distractor-set, vocabulary, and statistics changes independently mark prior evidence for re-analysis/re-review. Preparation and immutable history expose exact current and reviewed hashes instead of blending observed behavior into a mapping claim.
      • Persistence and API verification: migration 023, chained from 022, passed a clean 001→023 upgrade and a 023→022→023 round-trip on disposable PostgreSQL. Four database triggers reject UPDATE/DELETE on misconception, analysis, human-review, and statistics ledgers. The live integration preserves two concepts plus one analysis, review, and statistics record for its isolated tenant and probes every trigger. Seven instructor-only routes publish vocabulary, preparation, analysis, review, statistics, and history operations; learner requests are denied before the service. All 46 focused M4.11 tests, including live PostgreSQL, pass. The consolidated affected M4.1–M4.11 and OpenAPI set passes 331 tests, with 20 existing optional database/sandbox cases skipped by their environment gates. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 17 Python/migration/test files and strict mypy (Python 3.12 target) covers five source modules.
      • Distractor review desk: following frontend-skill, the responsive, instructor-only workbench uses a restrained three-desk editorial composition for exact scope, vocabulary, surface evidence, and human judgment, with observed behavior and immutable lineage kept below in separate ledgers. It supports concept revision, all eight checks, attributed rule/model suggestions, explicit accept/edit/reject, sealed statistics, stale-state notices, and hard non-diagnosis language, and is linked from item-bank, capture, selected-response, and estimate entry points. A rendered desktop visual pass, all 67 Metis web unit files (273 tests), ESLint with zero warnings, and TypeScript pass. Eighteen M4.11 Playwright flows pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis; the adjacent selected-response and estimate suites add 27 passing cross-browser regressions, including 390-pixel reflow and learner-denial paths.
      • Evidence boundary: the deterministic metis-distractor-mapping.json inventory hashes every implementation, test, migration, API/client contract, UI, traceability, and producer input and asserts M4.11.a–M4.11.e. Browser API transport uses deterministic doubles. This does not prove production deployment, individual learner diagnosis, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than- 32-GiB host approval gate; affected compile, type, lint, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.12 Build single-correct-answer and ambiguity verification. Evidence: adversarial crafted-item corpus.
    • M4.12.a Define applicability by item/response/scoring kind and explicit uniqueness/ambiguity verdicts, evidence, uncertainty, and reviewer state. Evidence: verifier schema.

    • M4.12.b Verify keyed response satisfies prompt and every nonkey fails under stated assumptions, tolerances, sources, and response rules. Evidence: known-value tests.

    • M4.12.c Detect multiple defensible interpretations/answers, missing context, contradictory sources, locale/cultural ambiguity, option overlap, and rubric underdetermination. Evidence: adversarial corpus.

    • M4.12.d Use symbolic/executable/source/model-assisted checks only as attributed evidence with timeout/error/N/A and human review bands. Evidence: checker tests.

    • M4.12.e Block single-correct publication on unresolved ambiguity while preserving raw findings and exception limits. Evidence: gate tests. <!-- M4.12 completion evidence (2026-09-01):

      • Exact applicability and proof contract: Metis now owns a strict 4.12.0 contract that pins one protected selected-response revision, item/response/scoring kind, locale, exact option partition, response rule, sources, assumptions, tolerances, interpretations, rubric, uncertainty, automated verdict, and completed reviewer state. The verifier explicitly returns applicable only for exact single-response scoring; all four selected-response kinds have applicability coverage. A stored structural key is never treated as semantic proof.
      • Known-value and adversarial verification: the deterministic truth matrix requires the key to satisfy prompt, source, and response rule while every nonkey independently fails under the submitted boundary. A seven-item crafted corpus contains one unique control and exercises multiple defensible answers, missing context, contradictory source editions, locale/cultural differences, option overlap, and rubric underdetermination. Every result retains exact evidence attribution and uncertainty; ambiguous or incomplete evidence fails closed.
      • Checker and human-authority boundary: symbolic, executable, source, and model-assisted checks carry exact method/configuration and, where applicable, source/model identity. Pass, fail, timeout, error, and N/A shapes are explicit, including time limits, error codes, rationales, evidence, uncertainty, and human-review bands. Checker output remains advisory. Human review must decide every immutable raw finding and cannot declare uniqueness over an incomplete key/nonkey proof.
      • Publication gate, persistence, and API verification: unresolved semantic ambiguity blocks the authoritative single-correct publish gate while raw findings remain unchanged. The sole bounded exception is one checker-unavailable finding, approved by an administrator, for no more than 30 days and one use. Migration 024, chained from 023, passed a clean 001→024 upgrade and a 024→023→024 round-trip on disposable PostgreSQL. Three database triggers reject UPDATE/DELETE on verification, human-review, and gate ledgers. Each isolated live workflow preserves two revisions in every ledger. Five instructor- only routes publish preparation, verification, human review, gate, and history; learner requests are denied before the service. All 45 focused M4.12 tests, including live PostgreSQL, pass. The consolidated affected M4.1–M4.12 and OpenAPI set passes 375 tests, with 11 optional live-database cases skipped by their environment guards. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 17 Python/migration/test files and strict mypy (Python 3.12 target) covers five source modules.
      • Answer proving desk: following frontend-skill, the responsive, instructor-only workbench uses a restrained three-desk editorial composition for the exact claim, adversarial evidence, and authoritative publication gate. It exposes option-by-option status, versioned sources, assumptions, missing context, locale/culture, rubric evidence, all four checker kinds, immutable raw findings, explicit human decisions, stale witnesses, exception limits, and gate state, and is linked from item-bank, selected-response, distractor, and estimate entry points. A rendered desktop visual pass, all 68 Metis web unit files (281 tests), ESLint with zero warnings, and TypeScript pass. Eighteen M4.12 Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis; the adjacent selected-response and distractor suites add 27 passing cross-browser regressions, including 390-pixel reflow and learner-denial paths.
      • Evidence boundary: the deterministic metis-answer-ambiguity.json inventory hashes every implementation, corpus, test, migration, API/client contract, UI, traceability, and producer input and asserts M4.12.a–M4.12.e. Browser API transport uses deterministic doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.13 Build answer-key leakage scan across learner-visible fields, assets, alt text, hints, exports, logs, and URLs. Evidence: negative suite.
    • M4.13.a Enumerate learner-visible UI/API/renderer/search/cache/event/ analytics/export/package/log/error/URL channels and protected key/ rationale/test fields. Evidence: coverage manifest.

    • M4.13.b Scan stem/options/feedback/hints/metadata/filenames/IDs and accessibility text/tracks for exact, normalized, encoded, indirect, and positional key leakage. Evidence: text leakage tests.

    • M4.13.c Scan media pixels/audio/transcripts/captions/descriptions/ documents/code/runtime output/interactive state and hidden metadata. Evidence: asset leakage tests.

    • M4.13.d Inspect QTI/LTI/SCORM/xAPI/course packages, previews, deep links, URLs, client bundles/state, errors/logs/traces, and support exports. Evidence: channel tests.

    • M4.13.e Test answer variations/multiple locales/randomized order/ encryption/role changes and fail release for any unresolved exposure. Evidence: adversarial negative suite. <!-- M4.13 completion evidence (2026-09-01):

      • Closed coverage and protected-material contract: Metis now owns a strict 4.13.0 contract and executable manifest for all 22 learner-visible UI, API, renderer, search, cache, event, analytics, package, preview, client, error, log, trace, and support-export channels. Seven protected key, localized key text, rationale, misconception, response-rule, and test-field kinds are represented only by nonreversible descriptors in preparation and scan responses. Coverage is evaluated across every required channel, surface, authored locale, and learner/anonymous role; missing or unavailable evidence fails closed.
      • Text, asset, and channel negative suite: exact, Unicode-normalized, base64/hex/URL-encoded, indirect, and positional detection scans content, decoded locators, filenames, identifiers, hidden metadata, and attributed scanner output. The narrow option allowlist admits only the exact key identity or localized key text on its own option/accessibility surface. OCR pixels, ASR audio, transcripts, captions, descriptions, documents, code, runtime output, interactive state, archives, QTI/LTI/SCORM/xAPI, client bundles/state, and operational output are all exercised. Scanner method/version/hash, timeout, outcome, and extracted witness are retained; missing observations, timeouts, errors, and invalid decrypted-observation lineage are blocking evidence failures.
      • Adversarial release authority: the exhaustive safe fixture contains 144 attributed artifacts and eight presentation cases spanning two locales, two learner-visible roles, two randomized orders per locale/role, and four encrypted cases. Client-decryptable ciphertext, role or locale gaps, incomplete surfaces, unresolved exposures, incomplete human decisions, stale exact pins, or tampered lineage block release. Every raw finding is decided exactly once, remains immutable, and has no exception path.
      • Persistence, API, and contract verification: migration 025, chained from 024, passed a clean 001→025 upgrade and a 025→024→025 round-trip on disposable PostgreSQL. Three triggers reject UPDATE/DELETE on scan, review, and release-gate ledgers; the isolated live workflow preserves two revisions in each ledger. Five instructor-only routes publish preparation, scan, review, release gate, and history; learner requests are denied before the service. All 48 focused M4.13 tests, including live PostgreSQL, pass. The affected M4.1–M4.13 and OpenAPI set passes 422 tests, with 22 optional integration cases skipped by their existing environment guards. Checked-in OpenAPI and generated Metis TypeScript client drift checks pass. Focused Ruff covers 17 Python/migration/test files and strict mypy covers five source modules.
      • Leakage review desk: following frontend-skill, the responsive, instructor-only workbench uses a restrained three-desk editorial layout for coverage, attributed capture and raw findings, and the authoritative no-exception gate. It exposes the 22-channel register, exact revision pins, capture parsing, findings defaulted to confirm, explicit decisions, immutable history, and stale state without retaining protected answers, and is linked from the item-bank, selected-response, distractor, and answer-verification entry points. A rendered desktop visual pass, all 69 Metis web unit files (286 tests), ESLint with zero warnings, and TypeScript pass. Eighteen M4.13 Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis; adjacent selected-response, distractor, and answer-verification suites add 45 passing cross-browser regressions, including 390-pixel reflow and learner-denial paths.
      • Evidence boundary: the deterministic metis-answer-leakage.json inventory hashes every implementation, test, migration, API/client contract, UI, traceability, and producer input and asserts M4.13.a–M4.13.e. Scanner observations and browser API transport use deterministic test doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, type, lint, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification databases were removed after use. -->
  • M4.14 Build item accessibility checklist and automated lint. Evidence: QTI Elevated Accessibility-relevant support matrix.
    • M4.14.a Define versioned checklist by item/media/interaction kind, applicable WCAG/QTI requirements, PNP supports, severity, evidence, and manual-only criteria. Evidence: checklist registry.
    • M4.14.b Lint structure/semantics/names/instructions/order/focus/keyboard/ timing/drag/motion/color/contrast/zoom/reflow/error behavior. Evidence: interaction lint tests.
    • M4.14.c Lint images/math/code/tables/charts/audio/video/3D/interactions for required equivalents, captions/transcripts/descriptions, and sync. Evidence: content lint tests.
    • M4.14.d Test locale/bidi, screen reader, keyboard, zoom, theme, viewport, reduced motion, and configured accommodation previews. Evidence: automation matrix.
    • M4.14.e Map each support to QTI accessibility metadata/content with supported/degraded/unsupported status and block unresolved mandatory checks. Evidence: support matrix. <!-- M4.14 completion evidence (2026-09-01):
      • Runtime checklist: item_accessibility.py and item_accessibility_helpers.py publish immutable schema version 4.14.0 and checklist version 2026.09.01 for WCAG 2.2, QTI 3.0 Elevated Accessibility, and AccessForAll PNP 3.0. Twenty-seven requirements carry closed item/media/interaction applicability, WCAG criteria, QTI metadata paths, PNP support codes, blocker/error severity, mandatory state, evidence/remediation text, and explicit manual-only state. The registry admits all 15 item kinds, nine media kinds, and 11 interaction kinds while exact source snapshots resolve the selected-response, written-response, math/code, and media/interactive revision families.
      • Executable lint: automatic observations bind a named/versioned/hashed tool and configuration, timeout, elapsed time, outcome, evidence, and an exact observation digest. Negative suites exercise structure, semantics, names, instructions, reading/focus order, keyboard, timing, drag, motion, color, contrast, zoom, reflow, and errors. Timeout, tool error, invalid not-applicable claims, missing observations, missing or failed manual reviews, digest tampering, and duplicate evidence all fail closed. Image, MathML/spoken math, code, table, chart, transcript, captions, audio description, 3D, and custom-interaction equivalence and synchronization requirements are executable and tested together.
      • Automation/support matrix: every locale requires the exact ten-cell locale/BiDi, named screen-reader, keyboard, 200% zoom, 400% zoom, light theme, dark theme, 320-CSS-pixel viewport, reduced-motion, and configured accommodation matrix. RTL direction mismatch and incomplete PNP support previews are rejected. Every applicable PNP support must map to QTI metadata and/or content with evidence and an explicit supported, degraded, or unsupported disposition; degraded, unsupported, and missing mandatory supports block. Current item, registry, audit, and gate hashes are atomic, stale evidence reopens release, and the schema contains no exception path.
      • Persistence/API/UI: migration 026 chains from 025 and adds two constrained append-only audit/gate ledgers with two PostgreSQL mutation triggers. Clean 001->026 and 026->025->026 migrations pass. A live PostgreSQL scenario preserves two audits (clear then failed), two gates (pass then blocked), raw findings, and rejects both UPDATE and DELETE. Four instructor-only endpoints expose preparation, audit append, gate append, and immutable history through closed OpenAPI/generated-client contracts; learner requests are denied before service construction. The instructor workbench renders a keyboard-accessible 27-check register, locale preview matrix, attributed evidence intake, immutable findings, QTI/PNP support dispositions, exact witnesses, history, and a no-exception release decision. Browser-authored empty evidence claims are rejected before transport.
      • Verification: all 40 focused M4.14 backend tests, including live PostgreSQL, pass. The affected M4.1-M4.14 and OpenAPI set passes 461 tests with 13 existing database-gated skips. Ruff passes 16 focused files, strict mypy passes five source files, deterministic OpenAPI export and Metis client drift checks pass, web lint/typechecking pass, and the complete web unit set passes 70 files / 291 tests. Eighteen M4.14 Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis, RTL, clear, failed, timeout, learner-denial, and 390-pixel paths. Eighty-four adjacent item bank, selected/written/math/media authoring, answer-verification, and answer-leakage cross-browser executions pass after constraining the expanded item-bank workflow navigation at desktop widths.
      • Evidence boundary: deterministic metis-item-accessibility.json hashes every implementation, test, migration, API/client contract, UI, traceability, and producer input and asserts M4.14.a-M4.14.e. Scanner captures, named screen-reader cases, and browser API transport use deterministic test doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility certification. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, lint, type, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.15 Build item compare and decision records. Evidence: stem, response, rationale, accessibility, and gate differences are visible.
    • M4.15.a Resolve two authorized exact item/stimulus/rubric revisions and define typed semantic diff model. Evidence: comparison schema.
    • M4.15.b Diff stem/instructions/sources/objectives/standards and response/ options/interaction/scoring/key separately. Evidence: content diffs.
    • M4.15.c Diff rationales/feedback/hints/exposure, accessibility/locales/ media, estimates/psychometrics, gates/reviews/use impacts. Evidence: protected/status diffs.
    • M4.15.d Render role-redacted accessible comparison and exact revision/ provenance/impact without leaking answer data. Evidence: UI tests.
    • M4.15.e Record select/reject/request-changes/merge/supersede with actor, rationale, evidence, immutable source refs, and new revision if merged. Evidence: decision tests. <!-- M4.15 completion evidence (2026-09-01):
      • Exact typed comparison: item_comparison.py and item_comparison_helpers.py publish closed schema version 4.15.0 and resolve authorized selected-response, written-response, math/code, and media/interactive histories by draft, revision, and revision digest. Exact stimulus locators on the immutable item source are resolved through authorized history, optional client stimulus digests are verified, unlinked or wrong-revision pins fail closed, and embedded written-response scoring guides expose exact rubric revision witnesses. The snapshot binds item/stimulus/rubric provenance and exactly 21 separately typed semantic fields.
      • Semantic and protected proof: stem, instructions, sources, objectives, standards, response, options, interaction, scoring, key, rationales, feedback, hints, exposure, accessibility, locales, media, estimates, psychometrics, gates/reviews, and use impacts each retain independent change state, count, digest, and explanation. Scoring, key, rationale, feedback, and hint values never enter the stored snapshot: only a protected structural summary and nonreversible digest remain. Missing psychometric, exposure, and usage state stays explicitly unknown.
      • Decisions and persistence: migration 027 chains from 026 and adds comparison and decision ledgers with same-family/distinct-revision, decision-shape, merge-witness, supersession, and composite comparison/tenant foreign-key constraints plus two UPDATE/DELETE rejection triggers. Select, reject, request-changes, merge, and supersede records bind actor, rationale, attributed evidence, both immutable source revision hashes, optimistic latest-decision pins, and content-addressed witnesses. Merge dispatches through the real typed append service and records its newly attributed item revision; existing authority can change only through an explicit exact-latest supersede. Tampered comparison and decision payloads fail closed.
      • API and workbench: four instructor-only operations publish comparison creation/catalog, exact workspace retrieval, and decision append through deterministic OpenAPI and generated TypeScript contracts. The linked comparison press renders exact revision folios, a keyboard-scrollable and category-filterable 21-row proof sheet, a conspicuous protected boundary, consequence registers that display unknown state, an immutable decision folio, and a responsive authority desk. Learners are denied before comparison transport, invalid evidence is rejected in the browser, and the protected sentinel never renders. A desktop rendered visual pass confirmed the hierarchy and redaction boundary.
      • Verification: 29 focused backend tests including live PostgreSQL pass; the affected M4.1-M4.15 and OpenAPI set passes 489 tests with 14 expected environment-gated database skips. Ruff passes 13 focused files, strict mypy passes five source files, 48 OpenAPI/API contract tests pass, and canonical OpenAPI export plus generated-client drift checks pass. Full Metis web lint and typechecking pass with zero warnings, and all 71 web unit files / 296 tests pass. Eighteen M4.15 Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis; 102 adjacent item-bank, authoring, answer-gate, and accessibility cross-browser regressions also pass, including clean WebKit confirmation of the one transient retried case. Clean 001->027 and 027->026->027 migrations pass. Live PostgreSQL preserves one comparison, one merge decision, and three total item revisions, verifies all constraints and two immutable triggers, and rejects both UPDATE and DELETE.
      • Evidence boundary: deterministic metis-item-comparison.json hashes every implementation, test, migration, API/client contract, UI, traceability, and producer input and asserts M4.15.a-M4.15.e. Browser API transport uses deterministic test doubles. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected compile, lint, type, unit, contract, database, accessibility, responsive, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.16 Build test/section assembly with blueprint constraints and exposure controls. Evidence: deterministic solver fixtures.
    • M4.16.a Define assessment/section blueprint for objective/standard/ content/cognitive level/item kind/difficulty/time/points/locale/ accessibility/security/exposure constraints. Evidence: schema tests.
    • M4.16.b Build authorized candidate pool from exact eligible item/ stimulus/rubric revisions and current gates/rights/calibration state. Evidence: eligibility tests.
    • M4.16.c Solve required/optional/min/max/balance/enemy/group/stimulus/ ordering/randomization constraints deterministically with seed/version. Evidence: solver goldens.
    • M4.16.d Enforce prior/current exposure, overlap, reuse windows, secure pools/forms, and institution policy without leaking bank membership. Evidence: exposure tests.
    • M4.16.e Return complete solution, infeasible diagnostics, alternatives, blueprint coverage, estimate, and human decision; never silently relax constraints. Evidence: assembly tests. <!-- M4.16 completion evidence (2026-09-01):
      • Closed blueprint and authorized pool: item_assembly.py publishes closed schema version 4.16.0 for all ten planning dimensions plus exact count, time, points, required/optional, min/max, balance, enemy, group, stimulus, fixed-position, order, seed, exposure, reuse, secure-pool, and institution-policy constraints. The service resolves each exact item, stimulus, rubric, estimate, and learning-link revision through authorized history, then requires current accessibility, answer, leakage, rights, calibration, exposure, and operational evidence. Institution and receipt issuer UUIDs are bound to the authorized tenant and authenticated actor before candidate resolution; stale, missing, foreign, future, or expired evidence fails closed.
      • Deterministic solving and diagnostics: solver version item-assembly-solver:1.0.0 canonicalizes and content-addresses the candidate pool, uses SHA-256 seed ranks for selection and display, and evaluates the complete constraint set without a relaxation path. The order-independent golden solution digest is 19797660f46e8693e10c63388dca44c292c84e3cd50d41624b12e22882342be9. Search is bounded at 500,000 nodes. Complete results retain exact selection/order, coverage, points, time, section stimulus use, and whole-form prior overlap; infeasible results retain bounded exact witnesses and explicitly unapplied alternatives.
      • Persistence, authority, and API: migration 028 chains from 027 and adds immutable assembly and decision ledgers with closed checks, composite tenant/run ownership, idempotency, and two UPDATE/DELETE rejection triggers. Accept, reject, and request-changes decisions bind attributed evidence, the exact assembly digest, and the exact predecessor digest. Reads independently reconstruct and verify the decision chain and reject forks, cycles, orphans, cross-run witnesses, and acceptance of an infeasible result. Four instructor-only operations publish catalog, assembly creation/read, and decision append through deterministic OpenAPI and generated TypeScript contracts; learner requests are denied before bank or secure-pool transport.
      • Workbench: the linked assembly desk presents the manifest and preflight boundary, explicit exact-form verdict, eligibility ledger, selected order, ten-dimension coverage, points/time/overlap totals, diagnostics, unapplied alternatives, and immutable human-decision folio. Browser builders reject malformed, incomplete, relaxed, duplicate, wrong-tenant, or wrong-issuer manifests before transport. The proof ledger remains keyboard-scrollable without page overflow at 390 pixels, and a rendered desktop visual inspection confirmed its hierarchy and protected-data boundary.
      • Verification: 52 focused M4.16 backend tests including live PostgreSQL pass. The affected item-authoring, answer-gate, QTI, migration, API, OpenAPI, and generated-contract matrix passes 562 tests with no skips. Ruff passes 13 focused files and strict mypy passes five source files. Clean 001->028 and 028->027->028 migrations pass; live PostgreSQL preserves one assembly and two chained decisions and rejects both UPDATE and DELETE. Full Metis web lint/typechecking pass with zero warnings, and all 72 web unit files / 303 tests pass. Twenty-seven M4.16 Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis. The 201-case adjacent item-authoring matrix passes across all three engines; its one retried Firefox timing assertion was synchronized to the new request and its complete stimulus-set spec then passes 12/12 without retry.
      • Evidence boundary: deterministic metis-item-assembly.json hashes all implementation, migration, contract, test, UI, traceability, and producer inputs and asserts M4.16.a-M4.16.e. Browser API transport uses deterministic test doubles. The pinned institution policy is an authenticated actor attestation, not an external policy-registry signature. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, migration, database, accessibility, responsive, visual, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.17 Build secure preview that cannot record a learner attempt or contaminate calibration. Evidence: event and gradebook assertions.
    • M4.17.a Use the real delivery renderer with exact draft build and synthetic preview identity/session marked nonlearner/nonoperational. Evidence: preview contracts.
    • M4.17.b Isolate preview responses/events/files/timers/cookies/storage from learner attempt, gradebook, integrity, analytics, calibration, LTI, and credential systems. Evidence: side-effect tests.
    • M4.17.c Render learner-visible content only; expose protected answer/ scoring overlays through a separately authorized author mode. Evidence: property tests.
    • M4.17.d Simulate scoring/feedback/navigation/accommodations/offline with disposable state and explicit preview labels. Evidence: behavior tests.
    • M4.17.e Assert no real attempt/grade/event/calibration/exposure count is written across completion, refresh, crash, and malicious requests. Evidence: integration assertions. <!-- M4.17 completion evidence (2026-09-01):
      • Exact nonoperational renderer: closed schema version 4.17.0 requires tenant, family, draft, revision, revision SHA-256, content SHA-256, locale, seed, and content-addressed build witnesses. Each selected, written, math/code, and media/interactive revision is resolved through authorized exact history and projected through its existing learner- only builder. The workbench imports the same AssessmentQuestion component used by live assessment; the contract fixes its live import path and requires identical learner-document, build, and renderer witnesses while keeping preview chrome outside that renderer. Synthetic identity preview_nonlearner has no enrollment, attempt, grade, event, calibration, or exposure authority. Sessions expire in 5-30 minutes, issue no learner or bearer token, are nonresumable, and retain at most 64 local events in browser parent memory.
      • Isolation and protected material: all 18 attempt/submission/grade/ gradebook/event/xAPI/Caliper/analytics/integrity/calibration/exposure/ credential/LTI/notification/cookie/local-storage/IndexedDB/file sinks are explicitly denied. Responses, events, files, and timers remain in React parent memory; file responses retain metadata and hashes only, with no upload bytes. Refresh, crash, reset, and actual reload discard the state, and browser tests find no preview cookie or storage key. The service requires a clean SQLAlchemy unit of work, disables autoflush, and fails if resolution dirties it. Learner documents are adversarially scanned for protected key/value families. Answers, scoring, feedback, rationales, misconceptions, hidden tests, expected values, and moderation details appear only in the distinct instructor- only author-overlay operation carrying literal scope item-preview:protected-author-mode; that response never re-embeds the learner surface.
      • Disposable behavior and API: three instructor-only preview operations publish closed OpenAPI and generated TypeScript contracts for prepare, simulation, and protected overlay; learners are rejected before the service or browser preview transport runs. The simulator covers exact selected-response scoring, written human-review state, exact math and media hooks, disconnected code execution, feedback, item/review/ complete navigation, elapsed time, accommodations, online/offline behavior, event bounds, expiry, state tampering, and malicious family mismatch. Every result returns content-addressed state plus a zero- side-effect receipt. The linked responsive workbench states the nonattempt boundary before transport, and rendered desktop inspection confirmed the learner-frame hierarchy and separated protected overlay. No model or migration is introduced because the feature is intentionally stateless.
      • Verification: 19 focused M4.17 backend tests pass, including a live PostgreSQL gate that authors and edits a real draft, appends its real selected-response revision, resolves that exact history through the production service, then applies completion, refresh, crash, malicious response, and protected-overlay paths. Counts remain exactly unchanged before and after for submissions, xAPI statements, estimate revisions, estimate audit events, bank-search exposure revisions, assembly runs, and assembly decisions. The affected 70-file item-authoring, all-family, exact-history, accessibility, assembly, estimate, exposure, QTI, migration, API, and OpenAPI matrix passes 421 tests with no skips. Ruff passes nine focused files, strict Python 3.12 mypy passes four source files, and the deterministic OpenAPI export and generated client checks pass. Metis web lint/typechecking pass with zero warnings, and all 73 web unit files / 308 tests pass. Twenty-one M4.17 Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis, responsive 390-pixel checks, actual reload, reduced-motion/high-contrast, offline, and role-denial coverage.
      • Evidence boundary: deterministic metis-item-preview.json hashes the implementation, contracts, tests, real renderer, workbench, traceability, and evidence producer and asserts M4.17.a-M4.17.e. Browser API transport uses deterministic test doubles; live service and database isolation are exercised separately. Code execution stays disconnected and media scoring simulates the exact hooks without an operational player. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, database, accessibility, responsive, visual, and cross-browser gates pass. The disposable verification database was removed after use. -->
  • M4.18 Build bulk import with per-item diagnostics and no silent drops. Evidence: imported plus rejected equals input count.
    • M4.18.a Define supported source formats/profiles, import batch/item IDs, limits, options, duplicate policy, and per-item outcome contracts. Evidence: import schema.
    • M4.18.b Parse archives/files/rows with type/size/depth/count/path/macro/ formula/external-ref protections and retain immutable source. Evidence: hostile import tests.
    • M4.18.c Validate/migrate each item/stimulus/rubric/media/source/answer/ accessibility/standard field independently with exact locator diagnostics. Evidence: validation fixtures.
    • M4.18.d Preview create/update/skip/reject/conflict/loss/substitution and require authorization before committing successful items. Evidence: preview tests.
    • M4.18.e Commit each accepted item idempotently with provenance and keep rejected inputs/diagnostics without partial silent fields. Evidence: persistence tests.
    • M4.18.f Assert imported+updated+skipped+rejected equals discovered input items and reconcile nested assets/stimuli. Evidence: count invariant. <!-- M4.18 completion evidence (2026-09-01):
      • Closed intake contracts: schema version 4.18.0 publishes three source formats (Oshun JSON, CSV, and a bounded QTI 3 package subset), four exact profiles, content-addressed source/batch/item/preview/commit witnesses, hard nonrelaxable request limits, four policy families, seven preview outcomes, four commit dispositions, and exact bare-host HTTPS allowlists. A new idempotency key creates a distinct preview of the same bytes so an existing binding is visibly reconciled as skip; exact request replay returns the original immutable workspace.
      • Bounded parsing and no silent drops: source bytes, length, media type, path, and SHA-256 are verified before parsing and retained verbatim. JSON/legacy JSON, RFC-style CSV, and every declared QTI item resource are parsed with file/archive/row/item/node/cell/depth/size/compression limits. Traversal, normalized duplicate paths, symlinks, encrypted or nested archives, active macros/scripts, spreadsheet formulas, XML declarations/entities/stylesheets, unsafe or nonallowlisted external references, MIME drift, digest drift, and decompression bombs fail closed. Blank and overflow CSV rows, unknown package/item/CSV fields, non-XML declared QTI item resources, unreferenced hostile XML, assets, and archive entries receive explicit retained diagnostics or loss outcomes rather than disappearing.
      • Independent diagnosis and authorization: item, stimulus, rubric, media, source, answer, accessibility, and standard categories each receive a field status. Diagnostics carry exact source file, path, archive entry, row, JSON pointer, field, code, severity, observed value, and stable ID. Legacy and QTI migrations surface as explicit substitutions. Duplicate state yields create/update/skip/conflict; unknown fields yield reject/loss; invalid inputs remain reject. Commit requires the exact preview hash, one exact item decision per input, a literal instructor authorization scope, and a separate affirmative substitution bit. Protected answers and rationales remain in the immutable import ledger but are excluded from draft capture context.
      • Persistence and reconciliation: migration 029 adds append-only batch, item, commit, and publisher-binding tables with tenant-scoped idempotency, checks, foreign keys, indexes, content hashes, and UPDATE/ DELETE rejection triggers. Accepted creates and updates use the real item-draft service and append hash-chained publisher bindings; rejected raw inputs and diagnostics stay queryable. Live PostgreSQL proved exact commit replay, create+reject, repeat skip+reject, update, and a revert to an earlier publisher version through one draft with four revisions and three binding revisions. Every preview and commit obeys its discovered-accounting equation, and nested asset/stimulus receipts are explicit and hash-verified.
      • Workbench: following the frontend skill, the linked instructor intake desk uses a restrained paper/ink/rust ledger rather than cards. It establishes scope, profiles, and hard limits before file transport; computes exact browser byte witnesses; renders every item, all eight field states, exact expandable diagnostics, and nested reconciliation; separates per-item decisions, substitution acceptance, and final exact-preview authorization; and ends with the immutable commit equation. A rendered desktop visual inspection confirmed the intended hierarchy and protected-material boundary.
      • Verification: 33 focused M4.18 backend tests including two live PostgreSQL gates pass. The affected import, item-draft, QTI, data- ownership, persistence, migration, API, export, and OpenAPI matrix passes 110 tests with no skips. Ruff passes 13 focused files; strict Python 3.12 mypy passes five source files; deterministic OpenAPI export and generated-client checks pass. Clean 001->029 and 029->028->029 migrations pass; all four persistence tables reject UPDATE and DELETE. Metis web lint/typechecking pass with zero warnings, and all 74 web unit files / 312 tests pass. Eighteen M4.18 Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/ critical Axe analysis, keyboard, role-denial, exact-byte, exact- authorization, and responsive 390-pixel coverage.
      • Evidence boundary: deterministic metis-item-import.json hashes all implementation, migration, contract, test, UI, traceability, and evidence-producer inputs and asserts M4.18.a-M4.18.f. Browser API transport uses deterministic test doubles; the live service and database paths are exercised separately. This does not prove production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, migration, database, accessibility, responsive, visual, and cross-browser gates pass. The disposable verification database was removed after use. -->

M5 — Pedagogical, quality, psychometric, and fairness gates#

  • M5.1 Implement objective coverage and construct-alignment gates. Evidence: crafted aligned and misaligned curricula.
    • M5.1.a Compute direct/indirect introduced/practiced/assessed coverage, depth/weight/cognitive-level distribution, and stale/missing links from exact revisions. Evidence: coverage fixtures.
    • M5.1.b Compare content/activity/item/rubric demands to objective criteria and intended construct; detect underrepresentation and construct- irrelevant variance. Evidence: alignment corpus.
    • M5.1.c Apply versioned required/advisory thresholds by course/assessment/ risk and persist raw coverage separately. Evidence: verdict tests.
    • M5.1.d Surface conflicting/unknown/N/A/manual-review state and remediation; block required misalignment. Evidence: gate UI/API tests. <!-- M5.1 evidence (2026-09-01): implemented a closed 5.1.0 exact- revision source contract over the M2.15 coverage snapshot, a deterministic server-side raw measurement, and versioned metis.objective-construct-alignment policy 2026.09.1. The raw layer computes direct/indirect introduced, practiced, and assessed counts; declared weights; fixed depth and cognitive distributions; exact current/stale/missing/unsupported/duplicate/conflicting link state; criterion and intended-construct representation; and construct-irrelevant learner-demand signals across content, activity, item, and rubric revisions. Four course/assessment × standard/high-stakes profiles each carry all 12 required/advisory threshold metrics. Unknown and conflicting evidence pauses in manual review, N/A remains explicit, confirmed required failures block release, and every finding receives named human remediation without an automatic content mutation. - Persistence/API/UI: Alembic 030 adds separate append-only objective_alignment_raw_measurements and objective_alignment_gate_runs ledgers with a tenant-bound foreign key, exact source/raw/policy/verdict/run hashes, idempotency, release constraints, and UPDATE/DELETE rejection. Four instructor-only operations expose the policy register, run catalog, gate creation, and exact workspace retrieval. The generated OpenAPI client is current. /create/alignment-gates follows the mandatory frontend skill as a raw-first curriculum-audit folio with an evidence spine, role distributions, exact link ledger, criterion/construct/variance comparison, explicit state ledger, remediation, history, and sticky release verdict; it is linked from the Create workflow. - Verification: crafted aligned, required-misaligned, stale, missing, unknown, conflicting, and not-applicable curricula pass the focused suite. The affected coverage/OpenAPI/export/release matrix passes 74 tests with no skips; strict Python 3.12 mypy passes five source modules. A fresh PostgreSQL 001->030 migration, 030->029->030 round trip, and live persistence gate pass; raw reuse/new-verdict reapplication, tenant isolation, and both ledgers' mutation guards are exercised. Metis web lint/typechecking pass with zero warnings; all 75 unit files / 315 tests pass. Twenty-one Playwright executions pass serially across Chromium, Firefox, and WebKit with automatic serious/critical Axe analysis, keyboard operation, learner denial before transport, exact policy/source binding, aligned/misaligned/ unknown/conflicting/N/A rendering, remediation, full-page visual inspection, and 390-pixel responsive coverage. - Evidence boundary: deterministic metis-objective-alignment.json hashes every implementation, migration, contract, corpus, test, UI, traceability, and evidence- producer input and asserts M5.1.a-M5.1.d. Browser API transport uses deterministic test doubles; the live service and database path is exercised separately. A passing declared-evidence gate is not proof of learner attainment or independent assessment validity, and this work does not establish production deployment, representative educator usability, physical assistive-technology behavior, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, migration, database, accessibility, responsive, visual, and cross-browser gates pass. -->
  • M5.2 Implement reading-level and sentence-complexity measurements. Evidence: independently derived known values.
    • M5.2.a Pin supported language/content applicability and formula/feature/ tokenizer/normalization versions. Evidence: measurement schema.

    • M5.2.b Compute reading indices and sentence/word/syntactic distributions server-side with exact units and excluded-content rules. Evidence: independent known-value fixtures.

    • M5.2.c Report raw values/ranges/assumptions/uncertainty/N/A separately from learner-band policy. Evidence: result tests.

    • M5.2.d Verify technical terms, math/code, quotations, multilingual, short, and malformed text behavior. Evidence: edge-case corpus. <!-- M5.2 completion evidence (2026-09-01):

      • The existing lesson readability boundary is upgraded in place to analysis schema 1.1.0 and deterministic English engine 2.0.0 rather than duplicated. Its profile independently pins formula and feature sets at 2.0.0 plus tokenizer, terminal-punctuation sentence segmenter, and offset-preserving casefold normalization at 1.0.0. Twelve typed block rules name the exact learner-facing fields for six prose-bearing kinds and return six code/math/media/reference kinds as N/A; seven versioned block/span/word exclusion rules state separately what remains in raw evidence and what leaves policy arithmetic.
      • The server now returns four exact frequency distributions for eligible words per sentence, heuristic syllables per eligible word, subordinate markers per sentence, and clause proxies per sentence. Every result carries its population, exact unit, sample and excluded-observation counts, minimum, P25, median, P75, P90, maximum, mean, frequency points, formula version, and pinned inclusive-linear-interpolation method. Decimal output uses ROUND_HALF_UP. A hand-derived fixture with sentence lengths [2,5], syllables [1,1,3,3,2,3,1], markers [0,1], and clause proxies [1,2] independently proves the four distributions plus Flesch Reading Ease 34.08 and Flesch-Kincaid Grade 9.38.
      • Raw measures, raw ranges, exact anchored/excluded spans, assumptions, uncertainty, block N/A, and derived-measure N/A are structurally separate from the unchanged learner-band policy. Fully quoted prose, for example, retains three raw words while reporting eligible words as zero and the reading indices as explicit N/A, never numeric zero. The policy produces no suggestion without eligible evidence, computes no aggregate score, and never rewrites content. New reviews persist the exact engine version used for analysis while legacy schema/engine 1.0.0 review records remain loadable.
      • A checked-in seven-case corpus executes technical-term, executable-code, mathematical-notation, quotation, intentional-multilingual, short, and malformed-punctuation behavior. It proves raw-versus-policy term and quotation handling, code/math N/A, intentional multilingual N/A, a single unterminated sentence, and deterministic repeated/missing terminal punctuation. The editorial proof sheet exposes the five method versions, exact content/exclusion registries, a selectable frequency/percentile measurement ruler, and metric-level N/A before the policy overlay. The existing exact-witness review and immutable human decision flow remains intact.
      • Automated proof passes Ruff format/lint for six affected Python files, strict MyPy for three affected service modules, 112 readability/dossier/ API/OpenAPI tests, live OpenAPI artifact and generated-client drift checks, Metis web ESLint and TypeScript checks, 316 Vitest tests across 75 files, thirteen deterministic evidence tests, and fifteen axe-backed Playwright journeys across Chromium, Firefox, and WebKit. Browser proof includes exact version/distribution switching, explicit N/A, policy separation, immutable review, role restriction, unsupported code, 390px containment, and reduced motion. Current traceability and input checksums are published at evidence/v1-workbenches/metis-lesson-readability.json.

      This is inspectable local contract, service, and browser-runtime evidence, not proof of production deployment, representative learner/author usability, observed comprehension, institutional band validity, formal accessibility conformance, or assistive-technology behavior. No database migration was required because readability review records remain in the existing typed dossier ledger. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, generated-client, accessibility, responsive, visual, and cross-browser gates pass. -->

  • M5.3 Implement item clarity and ambiguity rules. Evidence: curated adversarial corpus.
    • M5.3.a Define versioned rules for incomplete context, unclear referent/ instruction/response format, negatives, absolutes, double-barreled prompts, cues, language, and construct-irrelevant complexity. Evidence: rule registry.

    • M5.3.b Run deterministic rules with exact text/field anchors and locale/item-kind applicability. Evidence: known-value tests.

    • M5.3.c Run attributed model/human-assist checks only for semantic ambiguity with uncertainty and counterinterpretations. Evidence: calibrated assist tests.

    • M5.3.d Require human resolution for blocking ambiguity and retain dismissed findings/reasons. Evidence: adversarial corpus gate. <!-- M5.3 completion evidence (2026-09-01):

      • Metis now publishes item-clarity schema 5.3.0 and content-addressed rule profile metis.item-clarity-rules 1.0.0. Eleven individually versioned rules cover incomplete context, unclear referents, instructions and response formats, negative and absolute wording, double-barreled prompts, answer cues, language clarity, construct-irrelevant complexity, and semantic ambiguity. Every rule pins its evaluator, default severity, source families, item kinds, locales, field purposes, and rule digest.
      • The service normalizes exact, hash-bound text fields from native selected-response, written-response, math/code, media/interactive, and performance-evidence item revisions. Deterministic findings carry the exact field path, character offsets, excerpt, and field digest. A twelve-case adversarial corpus includes a clean control, every ten deterministic categories, and non-calibrated-locale behavior; direct fixture tests also verify all four specialized revision normalizers.
      • Semantic ambiguity remains a separate attributed assist path. Model and human assists pin provider, method, configuration, and conditional model identities and hashes; pass/finding/uncertain/timeout/error/N/A outcomes remain visible. A semantic finding requires uncertainty and at least two counterinterpretations at an exact current anchor. Assists remain advisory and cannot decide a finding or grant release.
      • Three append-only PostgreSQL ledgers preserve raw scans, complete human decisions, and release gates. Every finding must be confirmed, resolved, or dismissed; resolve/dismiss decisions require explicit evidence, retained dismissed-finding IDs, reasons, and the original raw finding. Confirmed blocking ambiguity fails closed, while revision and lineage pins make stale scans, reviews, and gates non-releasable. Five instructor-only API routes expose prepare, scan, review, gate, and history flows through the closed OpenAPI contract and generated client.
      • The item-bank clarity desk follows the frontend-skill proofing-desk hierarchy: exact source text and selected anchor dominate, the compact finding ledger exposes applicability and assist attribution, and the resolution inspector requires evidence before gating. Existing item authoring desks link into the workflow. Learners are denied before API transport, and the three-pane layout reflows into a linear 390 px flow.
      • Executable evidence is frozen at evidence/v1-workbenches/metis-item-clarity.json. Verification passes: 76 focused backend/schema/helper/service/API/migration/OpenAPI tests; a separate live PostgreSQL 031 probe spanning native authoring, scan, dismissal, and release plus rejected mutation of all three ledgers; targeted Ruff and strict MyPy; byte-current OpenAPI and TypeScript client generation; affected web lint and TypeScript; the full Metis web unit suite (76 files / 318 tests); and 15 Chromium/Firefox/WebKit Playwright executions covering exact requests and anchors, blocking and dismissal gates, retained findings, attributed semantic evidence, automatic serious/critical Axe analysis, role denial, and responsive reflow. Visual inspection at full desktop resolution confirmed the intended restrained proofing hierarchy.
      • These controls establish the bounded M5.3 technical implementation; they do not establish production deployment, representative educator usability, formal accessibility conformance, or physical assistive-technology behavior. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, migration, live-database, accessibility, responsive, visual, and cross-browser gates pass. -->
  • M5.4 Implement distractor plausibility and non-overlap rules. Evidence: known item fixtures.
    • M5.4.a Check duplicates/normalization/subsumption/intersection and mutually compatible options under stated response rules. Evidence: non-overlap fixtures.

    • M5.4.b Check implausibility, grammatical mismatch, length/style/cue imbalance, absolutes, joke options, and source/rationale inconsistency. Evidence: plausibility fixtures.

    • M5.4.c Apply kind/locale/discipline-specific rules and expose exact option pairs/findings without guessing learner ability. Evidence: applicability tests.

    • M5.4.d Bind verdict to item revision/rules and require review of semantic findings. Evidence: gate lifecycle tests. <!-- M5.4 completion evidence (2026-09-01):

      • Metis now publishes distractor-quality schema 5.4.0 and the content-addressed metis.distractor-quality-rules 1.0.0 profile. Thirteen individually versioned rules cover normalized duplicates, subsumption, enumerated intersection, single-response compatibility, facial implausibility, grammar, length/style/cue balance, absolute wording, joke options, and source/rationale consistency. Exact selected-response revision, key partition, cardinality, locale, discipline, sources, option text/rationale, rule definitions, and ordered option-pair digests remain bound to every scan.
      • Deterministic checks run only inside their recorded response-type, locale, and discipline applicability. Compatibility and source/ rationale consistency accept only attributed model or human-assist evidence with exact method/config/model identity, uncertainty, and at least two counterinterpretations. The result explicitly disclaims any learner-ability or domain-truth inference.
      • Human review must decide every semantic finding. Resolve/dismiss decisions require governed evidence; deterministic blockers cannot be dismissed and instead require a new item revision. The release gate fails closed on changed item/rule/scan/review witnesses, confirmed semantic findings, or deterministic blockers while retaining every raw finding and disposition.
      • Alembic 032 chains from 031 and adds scan, semantic-review, and gate ledgers with database-level append-only triggers. Five instructor-only endpoints expose prepare, scan, review, release-gate, and immutable history operations. The generated OpenAPI client is current. The restrained option-quality workbench exposes exact pair selection, applicability reasons, attribution, evidence dispositions, and gate history beside the existing misconception/statistics workbench.
      • Machine-readable evidence is frozen at evidence/v1-workbenches/metis-distractor-quality.json: 13 known-item cases, ten deterministic and three semantic categories, four discipline profiles, five routes, three storage tables, and three immutable triggers. Its generator/test require four satisfied claims for every M5.4 subitem plus storage/API/generated-client/workbench coverage and current input checksums.
      • Verification passed: 74 focused M5.4/OpenAPI tests; 145 selected- response, distractor-mapping, distractor-quality, migration, service, API, and OpenAPI regression tests; one live PostgreSQL migration/ semantic-dismissal/three-trigger mutation test; targeted Ruff and Python 3.12 mypy checks; OpenAPI export and generated-client checks; the complete Metis web unit suite (77 files / 320 tests); zero-warning web lint and typecheck; and 33 Chromium/Firefox/WebKit Playwright executions across the existing and new distractor desks. The new 15 executions cover exact revision/applicability requests, option-pair focus, non-dismissible deterministic defects, attributed competing interpretations, mandatory resolution/dismissal evidence, retained raw evidence, fail-closed/pass gates, learner denial, 390px reflow, and automatic serious/critical Axe analysis. Desktop visual inspection confirmed the intended manuscript/finding/disposition hierarchy.
      • These controls establish the bounded M5.4 technical implementation; they do not establish production deployment, representative educator usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, migration, live-database, accessibility, responsive, visual, and cross-browser gates pass. -->
  • M5.5 Implement answer uniqueness and rationale-consistency gates. Evidence: mutation tests.
    • M5.5.a Verify key cardinality/uniqueness/equivalence/tolerance against item-kind response/scoring contract. Evidence: key tests.

    • M5.5.b Verify keyed answers satisfy the stem/sources/assumptions and nonkeys do not, with explicit unsupported/manual state. Evidence: correctness tests.

    • M5.5.c Verify correct and distractor rationales refer to stable response IDs and agree with key, scoring, feedback, and hints. Evidence: consistency tests.

    • M5.5.d Mutation-test option/key/order/scoring/rationale/source changes and require invalidation/review. Evidence: mutation suite. <!-- M5.5 completion evidence (2026-09-01):

      • Metis now publishes answer-consistency schema 5.5.0 for the four native selected-response kinds. Every scan binds the exact response IDs, authored order/text, key kind and members, scoring method, equivalence/tolerance policy, stem/source assumptions, source references, rationales, learner feedback, and hint policy. Decimal tolerance bands and equivalence groups are evaluated without lossy-number conversion.
      • Deterministic checks enforce key cardinality, stable references, normalized uniqueness, authored order, scoring alignment, and every response-surface reference. Attributed source-checker, human-assist, or exact-identity model evidence records whether the key satisfies the stem/sources/assumptions and each nonkey fails them. Unsupported/manual judgments remain explicit and require two counterinterpretations; they are never silently converted into correctness claims.
      • Correct and distractor rationales, feedback, and the explicitly non-authored hint surface are joined through stable response IDs and checked against key, scoring, and correctness evidence. Human review must decide semantic uncertainty with governed evidence. Deterministic defects cannot be dismissed, raw findings remain immutable, and the release gate fails closed on stale scans or changed witnesses.
      • Alembic 033 chains from 032 and adds append-only scan, review, and gate ledgers with database-enforced update/delete rejection. Five instructor-only prepare/scan/review/release/history endpoints use idempotency, optimistic pins, immutable lineage, and exact witness verification. The generated OpenAPI contract and TypeScript client are current. The manuscript-first workbench exposes the exact key, scoring, truth matrix, assumptions, stable-ID learning surfaces, mutation witnesses, dispositions, and release history.
      • Mutation coverage rebuilds real selected-response sources after option text, key, order, scoring, rationale, source-reference, and learner-feedback edits and proves the exact invalidation reason. A direct component-hash matrix additionally covers the governed hint surface, while item-revision-only mutation has its own reason. Machine-readable evidence is frozen at evidence/v1-workbenches/metis-answer-consistency.json; its generator and tests require all four subitem claims, 15 finding categories, eight component hashes, nine invalidation reasons, storage/API/client/ workbench coverage, and current input checksums.
      • Verification passed: 85 focused M5.5/OpenAPI tests; 158 selected- response, ambiguity, answer-consistency, API, migration, service, and OpenAPI regressions including three live PostgreSQL integrations; the new 033 integration proves all three ledger triggers reject mutation; targeted Ruff and Python 3.12 mypy; OpenAPI export and generated-client checks; the complete Metis web unit suite (78 files / 323 tests); zero-warning web lint and typecheck; and 36 Chromium/Firefox/WebKit Playwright executions across the existing and new answer desks. The new 18 executions cover exact scan witnesses, explicit manual state, stable-ID blockers, evidence-required semantic dismissal, retained raw findings, exact mutation reasons, learner denial, 390 px reflow, and automatic serious/critical Axe analysis. Desktop visual inspection confirmed the intended manuscript/correctness/mutation-gate hierarchy, and the production-source stub-indicator scan is clean.
      • These controls establish the bounded M5.5 technical implementation; they do not establish production deployment, representative educator usability, independent domain-truth proof, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, migration, live-database, accessibility, responsive, visual, and cross-browser gates pass. -->
  • M5.6 Implement rubric weight, criterion coverage, level ordering, and descriptor-distinctness gates. Evidence: exact fixtures.
    • M5.6.a Validate weights/points/totals/ranges/rounding/caps/floors and missing/N/A scoring for the selected rubric method. Evidence: arithmetic fixtures.

    • M5.6.b Verify required objectives/evidence/performance dimensions are covered without duplicate or orphan criteria. Evidence: coverage tests.

    • M5.6.c Verify level codes/scores/ranges are strictly ordered/nonoverlap/ complete and descriptor direction matches scale. Evidence: ordering fixtures.

    • M5.6.d Detect identical/subsumed/contradictory/nonobservable descriptors with human review for semantic distinctness. Evidence: descriptor corpus.

    • M5.6.e Bind raw checks/verdicts to rubric revision and invalidate dependent item/assessment approval on change. Evidence: lifecycle tests. <!-- M5.6 completion evidence (2026-09-01):

      • Metis now publishes rubric-quality schema 5.6.0 over the existing immutable lesson-rubric revision rather than introducing a second rubric authority. The selected analytic weighted-points contract is checked for exact 10000-basis-point criterion and per-item totals, representative points, complete/nonoverlapping ranges, half-up rounding, zero-valued missing scores, N/A reweighting, and explicit unconfigured caps/floors. Native Decimal fixtures prove full = 100.00, missing = 40.00, and N/A reweight = 100.00.
      • Seventeen deterministic/semantic categories cover required objective, evidence, and performance-dimension coverage; orphan and duplicate criteria; ordered codes, scores, and ranges; normalized-identical, subsumed, contradictory, nonobservable, unsupported/manual, and wrong-direction descriptors; and dependent approval freshness. Semantic findings retain attribution, uncertainty, evidence, and two competing interpretations where meaning is unsupported/manual. Human review must disposition every semantic finding, dismissal requires evidence, and deterministic findings cannot be dismissed.
      • Six content-addressed receipts bind scoring, weights, coverage, level order, descriptors, and dependent approvals to the server-issued rubric revision. Those receipts plus rubric-revision lineage map to seven exact invalidation reasons. A changed rubric invalidates item and assessment approvals; explicitly invalidated dependencies may be evaluated but cannot be published. Three append-only ledgers and database triggers preserve scans, raw findings, human reviews, and release gates. Five instructor-only prepare/scan/review/release/history endpoints enforce idempotency, optimistic pins, immutable lineage, tenant scope, and witness verification. OpenAPI, the generated client, and the proof-first authoring desk expose the same closed lifecycle.
      • Machine-readable evidence is frozen at evidence/v1-workbenches/metis-rubric-quality.json; its generator and tests require all five subitem claims, nine arithmetic fixtures, all 17 finding categories, six governed component hashes, seven mutation reasons, storage/API/client/workbench coverage, and current input checksums. Verification passed: 87 focused M5.6/OpenAPI tests; 158 selected lesson-rubric, dossier, M5.6, API, migration, service, and OpenAPI regressions; the live PostgreSQL 034 integration proving all three ledger triggers reject mutation; targeted Ruff and Python 3.12 mypy; OpenAPI export and generated-client checks; the complete Metis web unit suite (79 files / 326 tests); zero-warning web lint and typecheck; and 33 Chromium/Firefox/WebKit Playwright executions across the existing rubric workbench and new quality desk. The 18 new executions cover exact request witnesses, manual descriptor evidence, non-dismissible structural defects, evidence-required dismissal, retained raw findings, evaluation/publication separation, approval rebinding, learner denial before transport, 390 px reflow, and automatic serious/critical Axe analysis. Desktop and mobile visual inspection confirmed the intended manuscript/proof/release hierarchy, and the production-source stub-indicator scan is clean.
      • These controls establish the bounded M5.6 technical implementation; they do not establish production deployment, representative educator usability, independent descriptor-truth proof, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; affected lint, type, unit, contract, migration, live-database, accessibility, responsive, visual, and cross-browser gates pass. -->
  • M5.7 Implement accessibility completeness gates per item modality. Evidence: missing alternative assets block publication.
    • M5.7.a Resolve applicable checklist/support requirements from item kind, media/interaction, target, locale, and PNP profile. Evidence: applicability matrix.
    • M5.7.b Verify structure/keyboard/focus/timing/drag/color/motion/zoom/ reflow/instruction/error requirements. Evidence: interaction gate tests.
    • M5.7.c Verify modality-specific alt/long descriptions, table/math/code semantics, captions/transcripts/descriptions/sign, and noninteractive equivalents. Evidence: asset gate tests.
    • M5.7.d Validate referenced alternative asset revision/locale/sync/rights/ availability and outcome equivalence. Evidence: alternative tests.
    • M5.7.e Block publication on missing/stale/failed mandatory requirements; preserve N/A/manual review with evidence. Evidence: publication gate. <!-- completion-evidence:
      • evidence/v1-workbenches/metis-accessibility-completeness.json is the deterministic M5.7 manifest. It binds the current item revision and M4.14 registry/audit/gate revisions, resolves all 27 inherited checks plus 13 alternative kinds across finite target/locale/PNP cells, and retains explicit attributed automatic, manual, and N/A evidence.
      • The fail-closed evaluator independently verifies requested/resolved asset revision, source hash, locale, target/profile, synchronization, rights, availability, automated or reviewer-attributed outcome equivalence, authority hash, and receipt hash. Missing alternatives are incomplete; stale or failed witnesses are blocked. Sixteen finding categories and five item/baseline/policy/matrix mutation reasons keep publication closed without erasing the original assessment findings.
      • Alembic 035 adds separate append-only assessment and publication-gate ledgers. Four instructor-only API operations, exact OpenAPI/generated TypeScript contracts, and the accessibility publication desk expose preparation, assessment, gate, history, coverage, blockers, immutable lineage, and the explicit Isis authority boundary.
      • Verification passed 69 focused schema/helper/service/API/migration/ OpenAPI tests plus the live PostgreSQL integration at migration head 035. The live probe released a complete item and proved both ledger triggers reject UPDATE and DELETE. The selected M4.7 media-authoring, M4.14 accessibility, M5.7 completeness, and OpenAPI regression set passed 139 tests; Ruff format/lint, strict Python 3.12 mypy, OpenAPI drift, and generated-client checks are clean.
      • The complete Metis web unit suite passed 80 files / 330 tests with zero-warning lint and typecheck. Eighteen new and 36 combined Chromium/Firefox/WebKit Playwright executions cover the contextual matrix at 390 px, explicit N/A/manual rows, exact asset receipts, missing-asset blocking, baseline mutation, learner denial before transport, and automatic serious/critical Axe analysis. The targeted production-source stub-indicator scan is clean.
      • These controls establish the bounded M5.7 technical implementation. Alternative-asset authority receipts remain attributed inputs rather than live Isis reads; M5.8 owns that integration. This does not claim production deployment, physical assistive-technology validation, representative learner usability, or formal conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.8 Implement media technical and caption/transcript gates through Isis. Evidence: live contract integration.
    • M5.8.a Map Metis media kind/target/profile/accessibility needs to canonical Isis asset/gate/profile identities and versions. Evidence: integration mapping.
    • M5.8.b Request/read exact asset revision technical measurements and captions/transcript/sync gates through audience-authorized contracts. Evidence: live contract tests.
    • M5.8.c Revalidate gate freshness against media/profile/evaluator/policy versions and represent unavailable/not-configured/partial honestly. Evidence: freshness tests.
    • M5.8.d Translate results to Metis lesson/item/build eligibility without copying or overriding Isis raw evidence. Evidence: authority tests.
    • M5.8.e Propagate media replacement/quarantine/recall/gate change and block affected publication. Evidence: lifecycle integration test. <!-- completion-evidence:
      • evidence/v1-workbenches/metis-media-readiness.json is the deterministic M5.8 manifest. It binds the canonical Isis gallery projection, authenticated BFF routes, fail-closed Metis HTTP client, exact-read eligibility service, lesson attachment/delivery path, OpenAPI artifacts, generated clients, workbench, and test corpus.
      • Isis now publishes one target/locale-specific media-readiness witness for each exact asset and rendition revision. Three media kinds and four targets resolve canonical profiles over eight accessibility needs. Technical, caption, transcript, and synchronization gates bind current profile, policy, evaluator, asset/alternative revisions, timestamps, expiry, and evidence; missing, unavailable, not-configured, partial, failed, stale, or mismatched required state is ineligible. Non-applicable gates remain explicit and evidence-free.
      • The existing audience-authorized Isis query/resolve contract remains the sole authority. Tenant, actor, and Metis audience are derived from authentication rather than request fields, and exact resolution is private/no-store and re-evaluates recall, quarantine, expiry, profile, evaluator, policy, measurements, and gate state. Metis forwards the caller bearer and signed exact selection; learner access is denied before the authority call.
      • The instructor-only /api/media-readiness/evaluations endpoint translates the live result for lesson, item, and assessment-build subjects. Authorized callers receive an ephemeral Isis snapshot for inspection, while the separate Metis eligibility artifact contains only identities, versions, states, expiry, and content-addressed fingerprints with raw_isis_evidence_copied=false. Governed lesson attachments likewise retain only readiness profile/policy identities, authority revision, and snapshot hash—never raw measurements or Isis evidence locators.
      • Learner delivery re-resolves every governed attachment. Selection, kind, digest, profile/policy, authority-revision, or readiness-snapshot change withholds media and removes its asset from preview. Executed fixtures cover replacement, recall, quarantine, technical unavailability, caption not-configuration, partial transcript, policy change, evaluator change, and readiness gate change without silently substituting another revision.
      • Verification passed 82 focused Metis gallery/readiness/dossier regressions, 12 Isis output-gallery/BFF/runtime-contract tests, and three focused Metis OpenAPI checks. OpenAPI exports and both generated clients are current; targeted Ruff/format/lint, Metis web and BFF typechecks, and changed-source lint are clean. The complete Metis web unit suite passed 81 files / 333 tests. Thirty Chromium, Firefox, and WebKit executions cover exact witnesses, all three eligibility scopes, unavailable/not-configured/partial states, recall/quarantine/gate change, learner denial, the raw-authority boundary, 390 px reflow, and automatic serious/critical Axe analysis. The targeted production-source stub-indicator scan is clean.
      • The full BFF lint target was also executed with a supervised 6 GiB V8 heap and remains red on its unrelated existing backlog (150 errors / 265 warnings across pre-existing Aja, Hathor, Yemaya, and boundary files); no finding names the curated-gallery files changed for M5.8. The BFF typecheck, focused route tests, runtime-drift test, changed production lint, and OpenAPI formatting checks pass.
      • These controls establish the bounded M5.8 live-contract integration; they do not claim a deployed production Isis probe, production deployment, representative educator usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.9 Implement declared-versus-empirical difficulty checks. Evidence: calibrated fixture.
    • M5.9.a Resolve declared target/band/source and empirical item revision, population/sample, method, statistic, uncertainty, and calibration status. Evidence: input schema.
    • M5.9.b Compare on a ratified common scale or explicit mapping with tolerance/uncertainty; prohibit raw incomparable subtraction. Evidence: comparison fixtures.
    • M5.9.c Return aligned, material mismatch, insufficient, stale, N/A, or conflicting state with direction/magnitude. Evidence: verdict tests.
    • M5.9.d Route mismatch to item/blueprint review and never overwrite empirical or author-declared values. Evidence: calibrated lifecycle test. <!-- Completion evidence (2026-09-01): evidence/v1-workbenches/metis-item-difficulty.json is the deterministic, checksum-bound M5.9 manifest.
      • The instructor-only evaluation contract pins one immutable M4.10 estimate-set revision/hash and one exact code-owned policy ID/version. It resolves the existing current author and ratified empirical-calibration sources without introducing a second source of truth. Declared evidence carries target, bounded band/label, scale, exact source identity, method, interval method, and uncertainty; empirical evidence carries exact item/context and evidence pins, population, sample identity/size, method, statistic, interval, uncertainty method, collection window, and ratification state.
      • The sealed policy names the common probability-proxy difficulty scale, tolerance, minimum empirical sample, and two exact, hashed, dual-ratified mappings for author and psychometric scale v1. Mapping schemas require complete bounded coverage and strict monotonicity. Evaluation proceeds only when both sides already share the common scale or exactly one explicit ratified mapping exists; missing or conflicting mappings withhold the comparison and raw_incomparable_subtraction_performed remains literally false.
      • The interval/tolerance evaluator returns all six required states: aligned, material mismatch, insufficient, stale, not applicable, and conflicting. Comparable mismatches report empirical-harder or empirical-easier direction plus the material gap beyond the accepted band. Stale exact pins remain distinguishable from insufficient and contradictory evidence rather than being collapsed into a score.
      • Material mismatch and mapping conflict route explicitly to item and blueprint review. The response is marked evaluation-only with literal false declared/empirical overwrite flags. The calibrated lifecycle test snapshots every source, revision-row count, and audit-row count before evaluation and proves all remain unchanged afterward, so the derived check adds no persistence table or mutation path.
      • The estimate observatory now presents an evidence-led declared versus empirical proof panel with source bands, empirical sample/method, common scale, tolerance, direction, gap, reason codes, immutable-value guard, and review links. Learners remain denied before estimate transport. Visual trace inspection confirmed the desktop composition; the first browser pass also found and fixed a serious color-contrast issue before the final gate.
      • Verification passed 28 focused difficulty/lifecycle backend tests, all 43 Metis OpenAPI contract tests, and the existing estimate migration plus live PostgreSQL immutability gate. Ruff and formatting passed on eight changed Python/test files; strict Mypy passed four production source files. The OpenAPI export and generated Metis TypeScript client are current. Metis web lint, focused E2E lint, and typecheck passed; the complete web unit suite passed 82 files / 335 tests. Thirty-six Chromium, Firefox, and WebKit executions cover the prior estimate lifecycle plus aligned, both-source mismatch proof, insufficient, stale, not-applicable, and conflicting presentations, exact request pins, review routes, immutability copy, learner denial, 390 px reflow, and automatic serious/critical Axe analysis. The targeted production-source stub-indicator scan is clean.
      • These controls establish the bounded M5.9 comparison and review control plane. They do not claim a connected production psychometric provider, production deployment, representative educator usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.10 Implement IRT parameter bounds and fit diagnostics. Evidence: 1PL/2PL/3PL known-value fixtures.
    • M5.10.a Validate model-specific difficulty/discrimination/guessing/ scaling/linking parameter presence, numeric bounds, standard errors, and covariance. Evidence: parameter fixtures.
    • M5.10.b Validate convergence, identifiability, sample/item sufficiency, information, residual/global/item fit, and local dependence. Evidence: diagnostic fixtures.
    • M5.10.c Distinguish 1PL/2PL/3PL applicability and prohibit unsupported parameter interpretation across models. Evidence: model matrix.
    • M5.10.d Persist raw diagnostics/tool/config/version separately from calibrated/flagged/ineligible verdict. Evidence: result tests. <!-- Completion evidence (2026-09-01): evidence/v1-workbenches/metis-irt-diagnostics.json is the deterministic, checksum-bound M5.10 manifest.
      • The closed instructor/admin contract pins an exact learner-safe item revision, run/input/output/sample/population/ability-scale identities, software build and runtime, configuration version/hash/seed, model family, intended use, completion time, and every raw diagnostic. A versioned code-owned policy defines operational bounds for b, a, c, and D, allowed scaling constants, model-specific sample/item minima, identification and information limits, residual/global/item-fit limits, and local-dependence Q3. The policy records its sources and limitations and carries a reproducible SHA-256 witness.
      • Estimated parameters require a finite value, nonnegative standard error, and ordered interval containing the value; fixed parameters require a reason and prohibit estimated uncertainty. The full estimated-parameter covariance matrix must contain exactly every cell, be symmetric and positive semidefinite, and reconcile each diagonal to the corresponding squared standard error. Linked transformations pin an exact target scale, unique anchor revisions, method/evidence, positive slope, uncertainty, and a reconciled 2x2 covariance; unlinked runs cannot claim transformation inputs.
      • Exact known-value 1PL/2PL/3PL fixtures execute the logistic response and item-information formulas. Every submitted information point is recomputed from the raw b/a/c/D values; conditional standard error must equal one over square-root information. The evaluator produces one typed check for each of thirteen categories: model applicability, bounds, covariance, convergence, identification, sample sufficiency, item sufficiency, information, residuals, global fit, item fit, local dependence, and linking. Core failures are ineligible; structurally sound runs with fit/dependence concerns are flagged; otherwise they are calibrated.
      • The complete model matrix fixes a=1 and c=0 for 1PL, estimates a while fixing c=0 for 2PL, and estimates a/c for 3PL; all families estimate b and use one supported fixed D. Unsupported roles, unavailable required values, inadequate covariance, nonconvergence, nonidentification, insufficient sample/items/information, or rejected linking fail closed and expose no parameter interpretations. Review-only fit flags remain distinct from ineligibility.
      • Alembic 036 adds separate append-only item_irt_diagnostic_runs and item_irt_diagnostic_verdicts tables with tenant-bound foreign keys, idempotency, exact hashes, and database triggers rejecting update/delete. Raw JSON contains the exact tool, configuration, and diagnostics but no policy or verdict. Verdict rows persist the exact policy/checks/status separately and enforce promotion_applied = false; the live PostgreSQL test also proves the M4.10 estimate-revision count is unchanged.
      • The read-only calibration ledger presents model/verdict filtering, parameter uncertainty, raw tool receipt, sample/convergence/ identification/linking evidence, all thirteen checks, independent raw, policy, and verdict hashes, and a prominent no-promotion receipt. Its catalog integrity-checks compact run summaries instead of loading every raw information curve; selecting a run loads and revalidates the full evidence workspace. Learners are denied before diagnostic transport. Visual inspection confirmed the desktop composition; the first automated browser pass found and corrected one serious footer contrast defect.
      • Verification passed 28 focused M5.10 backend tests, including the migration and live PostgreSQL separation/immutability/no-promotion gate. Ruff and strict Mypy pass on the production contract, model, evaluator, persistence, and API files; OpenAPI export and the generated TypeScript client are current. Metis web lint/typecheck and focused data tests pass. Twelve Chromium, Firefox, and WebKit executions cover 1PL/2PL/3PL, calibrated/flagged/ineligible, raw/verdict separation, filters, empty state, learner denial, 390 px reflow, and automatic serious/critical Axe analysis. The complete OpenAPI and web-unit gates, targeted production-source stub scan, and directly affected evidence manifests are also current and passing.
      • These controls establish the bounded M5.10 diagnostic control plane; they do not claim a connected production psychometric provider, production deployment, universal IRT thresholds, representative psychometrician usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.11 Implement sample-size and standard-error eligibility for calibration promotion. Evidence: low-evidence items cannot become calibrated.
    • M5.11.a Define minimum effective sample, response distribution, administration diversity, data quality, missingness, and standard-error/ confidence criteria by model/use. Evidence: eligibility policy.
    • M5.11.b Calculate included/excluded/weighted counts and privacy-safe cohort coverage from exact attempt/item revisions. Evidence: sample tests.
    • M5.11.c Verify parameter/statistic standard errors and convergence against all required thresholds without treating missing as zero. Evidence: boundary tests.
    • M5.11.d Block promotion to calibrated when any required evidence is low/ stale/conflicting and route field-test/manual review. Evidence: state transition tests. <!-- M5.11 completion evidence (2026-09-01):
      • The closed 5.11.0 contracts and content-addressed 2026.09.1 policy define all nine 1PL/2PL/3PL x pilot/operational/research profiles. Each profile pins minimum included/effective samples, distinct learners, response-category count/rate, administrations, institutions, coverage, missingness, invalidity, weight ratio, privacy-cell size, freshness, and required parameter/statistic SE, confidence-level, and interval-width limits. Thresholds are explicit Metis operational policy; automatic promotion is literally false.
      • Exact pseudonymous attempt revisions pin the learner, item revision, administration, institution, occurrence, eligibility/exclusion, outcome, weight, and approved coverage dimensions. The pure sample reducer reconciles base/included/excluded and outcome counts, exclusion reasons, sum of weights and squared weights, Kish effective n, distinct learner/administration/institution counts, missing/invalid/coverage rates, and privacy-safe released/suppressed cell counts. Tests exercise nonuniform weights and prove no cohort token or cell identity is returned.
      • The evaluator emits exactly ten typed checks over exact revisions, effective sample, response distribution, administration diversity, data quality, missingness, uncertainty, M5.10 convergence/fit, freshness, and privacy coverage. Every required estimated parameter and reported statistic must carry a non-missing SE and interval; missing remains insufficient, never zero. Exact inclusive policy boundaries are floating-point tolerant, while excessive SE, inadequate confidence, excessive interval width, nonconvergence, and noncalibrated M5.10 verdicts fail closed.
      • Low/sparse evidence routes to field_test / field-data collection; stale, conflicting, duplicate, quality-failed, or convergence-failed evidence routes to psychometric manual review. Only a completely passing assessment receives a content-addressed, time-bounded eligibility receipt for a separate human ratification step. The M4.10 empirical-calibration boundary now requires that exact matching, untampered, unexpired M5.11 receipt; no M5.11 path creates or mutates an estimate and every record asserts promotion_applied = false.
      • Alembic 037 adds separate tenant-bound append-only promotion-input and assessment ledgers linked to the exact M5.10 verdict, with idempotency, policy/sample/check hashes, status/route/expiry constraints, and update/delete rejection triggers. Live PostgreSQL tests prove eligible versus low-evidence routing, idempotent replay, cross-tenant denial, append-only enforcement, compact catalog integrity, no receipt for low evidence, and no new calibrated estimate revision.
      • The instructor-only promotion-control workbench presents the sample funnel, effective n, response balance, diversity, missing/invalid rates, privacy-safe coverage, exact policy profile, all ten checks, the required route, and either a bounded receipt or an explicit blocked receipt. Learners are denied before transport. Visual inspection confirmed the evidence-led desktop composition; responsive automation proves the full ledger reflows without horizontal overflow at 390 px.
      • Verification passed 19 focused M5.11 evaluator/API tests, 24 trusted calibration regression tests, the static migration guard, both live PostgreSQL gates, and all 43 Metis OpenAPI contract tests. Ruff, strict Mypy at the Python 3.12 runtime, OpenAPI export/client checks, targeted web lint/typecheck, the production stub scan, 84 files / 339 web unit tests, eight evidence-generator tests, and twelve Chromium/Firefox/ WebKit workbench executions with automatic serious/critical Axe analysis all pass.
      • These controls establish the bounded M5.11 calibration-promotion gate; they do not claim a connected production psychometric provider, production deployment, universal psychometric thresholds, representative psychometrician usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.12 Implement DIF analysis across ratified cohorts with minimum sample and privacy rules. Evidence: synthetic fairness corpus.
    • M5.12.a Define ratified privacy-reviewed cohort dimensions/reference/ focal groups, method/model, matching variable, thresholds, correction, minimum samples, and authorized purpose. Evidence: DIF policy.
    • M5.12.b Build deidentified analysis inputs with exact item revisions, valid responses/weights, missingness, group privacy thresholds, and no small-cell outputs. Evidence: data tests.
    • M5.12.c Compute uniform/nonuniform DIF effect/significance/uncertainty/ model fit with pinned tools and multiple-testing handling. Evidence: synthetic corpus.
    • M5.12.d Return insufficient/N/A/error/conflict and subgroup-safe results separately from fairness decision. Evidence: result tests.
    • M5.12.e Require psychometric/fairness human review before restrict/ revise/retire; prohibit individual inference. Evidence: decision tests. <!-- M5.12 completion evidence (2026-09-01):
      • The closed 5.12.0 contracts and content-addressed 2026.09.1 policy ratify locale-family, language-support, region, and explicitly consented-demographic cohort dimensions. Each dimension pins its data classification, consent rule, reference/focal groups, authorized aggregate-item-fairness purpose, minimum source sample, and privacy floor. The analysis family pins weighted logistic regression, rest score matching, uniform and nonuniform materiality thresholds, a 0.05 family alpha, Holm correction, quality/weight/model-fit limits, and explicitly prohibits automatic fairness decisions and individual inference.
      • Deidentified rows bind exact response, item, administration, cohort, occurrence-time, and pseudonymous learner revisions. The reducer reconciles valid/missing/invalid responses, binary outcomes, matching scores that exclude the analyzed item, positive weights, duplicate learner participation, weight dispersion, and cohort receipt seals. Release eligibility uses the maximum of the policy floor and both current ratified privacy-receipt floors. Every small cell suppresses counts, rates, weights, effects, significance, and uncertainty; no learner or cohort token enters a released result or API workspace.
      • The pinned pure-Python estimator fits intercept, centered rest score, focal-group, and focal-by-rest-score terms with deterministic weighted Newton-Raphson estimation and final observed information. It reports uniform/nonuniform coefficients, standard errors, two-sided Wald p-values, 95% intervals, convergence/fit evidence, and one Holm family across every eligible item-term hypothesis. The synthetic corpus recovers uniform, nonuniform, and null patterns and covers exact threshold, multiplicity, privacy, duplicate, and failure boundaries.
      • Eligibility is represented independently from the statistical finding and independently again from the human disposition and operational action. eligible, insufficient, not_applicable, error, and conflict remain distinct; all noneligible states fail closed and expose only subgroup-safe reasons. A signal is evidence for human review, never a causal finding of bias.
      • Alembic 038 adds tenant-bound append-only raw-input, analysis-run, and human-decision ledgers with idempotency, exact policy/evidence hashes, update/delete rejection triggers, and revalidation on retrieval. Live PostgreSQL tests prove eligible and suppressed persistence, cross-tenant denial, immutable inputs/runs/decisions, compact catalog integrity, and that review records never apply an item action.
      • Restrict/revise/retire requires an eligible signal plus two independent named approvals with psychometrician and fairness-review authority, matching dispositions/actions, and administrator submission. Null, insufficient, and not-applicable states have narrower allowed routes; error/conflict states allow no action. Every decision retains action_applied = false, leaving the actual lifecycle mutation to a separate operational control plane.
      • The instructor DIF review ledger keeps cohort eligibility, statistical evidence, human disposition, and operational action visibly separate; only administrators can submit dual review. The responsive 390 px flow has no horizontal overflow, the cohort ledger is keyboard-focusable, learners are denied before evidence transport, and automatic serious/critical Axe analysis passes in Chromium, Firefox, and WebKit.
      • Verification passed 27 focused M5.12 contract/evaluator/API/migration/ live-PostgreSQL tests, 45 psychometric-authority regressions, all 43 Metis OpenAPI contract tests, Ruff, strict Python 3.12 Mypy, current OpenAPI/client generation checks, targeted changed-file web lint, the full Metis web typecheck, 85 files / 342 web unit tests, the production stub scan, nine evidence-generator tests, and 15 three-engine browser executions with automated accessibility and mobile coverage.
      • These controls establish the bounded M5.12 DIF review control plane; they do not claim a connected production fairness dataset, production deployment, universal fairness thresholds, causal bias attribution, representative psychometric/fairness-reviewer usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.13 Implement exposure and compromise-risk gates. Evidence: overexposed items are automatically restricted or retired.
    • M5.13.a Define exposure metrics/windows/populations/channels, secure-item classification, thresholds, known breach signals, and lifecycle policy. Evidence: exposure schema.
    • M5.13.b Count delivered/seen/responded/reused/exported/shared occurrences by exact item revision with preview/test exclusion and privacy-safe aggregation. Evidence: counting tests.
    • M5.13.c Ingest verified compromise/leak/reconstruction reports with evidence, scope, confidence, and incident status. Evidence: report tests.
    • M5.13.d Calculate normal/warning/restricted/quarantined/retired state and enforce pool/assembly/delivery/export changes server-side. Evidence: threshold tests.
    • M5.13.e Propagate replacement/notification/blast radius while retaining historical attempts/grades and allowing authorized appeal. Evidence: incident journey. <!-- M5.13 completion evidence (2026-09-01):
      • The closed 5.13.0 contracts and content-addressed 2026.09.1 policy extend the binding metis.protected-assessment:1.0.0 authority. They define delivered, seen, responded, reused, exported, and shared occurrences; seven delivery/export/share channels; learner, instructor, reviewer, institution, and public populations; standard, secure, and high-stakes classifications; 30/90-day rolling windows; per-class privacy floors; ordered warning/restriction/retirement thresholds; six known compromise signals; and the normal, warning, restricted, quarantined, and retired lifecycle.
      • Every occurrence pins an exact item UUID, revision, revision SHA, channel, population, environment, pseudonymous actor hash, context revision, optional administration revision, and occurrence time. Immutable batches accumulate across the exact revision, freeze its classification, reject duplicate occurrence IDs across batches, and serialize concurrent writers with a PostgreSQL advisory transaction lock. Preview/test and outside-window occurrences do not spend the budget. Aggregate and per-metric actor floors independently suppress counts; private weighted points can still automatically restrict or retire an item without releasing a small cell.
      • Administrator-only incident intake requires a known signal, scope, severity, open/contained/remediating/resolved status, at least 75% confidence, authenticated tenant verifier, ordered discovery/ verification times, and one or more security-restricted evidence witnesses. Report, evidence, request, and stored-field hashes are revalidated; released workspaces contain compact incident witnesses, never raw occurrence tokens or forensic material.
      • Metric and incident precedence deterministically resolves all five states: warning/restriction/retirement thresholds, verified compromise restriction, high-severity quarantine, and critical/public/answer-key retirement. One persisted assessment drives four server decisions. Item-bank search and direct preview exclude blocked exact revisions; assembly denies before expensive candidate resolution; assessment GET and submit check the current exact indexed revision; selected-response, written-response, math/code, and media/interactive protected exports all deny from the same state. Standalone admission fails closed when no current M5.13 assessment exists.
      • Alembic 039 adds tenant-bound append-only occurrence-batch, compromise-incident, lifecycle-assessment, and appeal ledgers with idempotency, composite tenant foreign keys, state/enforcement checks, content hashes, and update/delete rejection triggers. Live PostgreSQL proves automatic retirement, all-surface denial, pool exclusion, tenant isolation, idempotent replay, immutable records, authorized appeal, and unchanged historical submission count.
      • Enforced states propagate replacement-required status, the affected context count or its privacy suppression, security/assessment-owner/ institution-reviewer notification audiences, and affected-learner notification for public incidents. Attempts and grades are always preserved. An administrator may append an uphold/overturn decision; recomputation excludes an overturned incident but cannot erase metric exposure, the incident, the prior assessment, or appeal history.
      • The instructor exposure-control ledger presents lifecycle before the four admissions, then privacy-safe measurements, evidence witnesses, response/blast radius, history, and admin-only appeal. Learners are denied before evidence transport; the 390 px layout has no horizontal overflow; and automatic serious/critical Axe analysis passes in Chromium, Firefox, and WebKit.
      • Verification passed 31 focused M5.13 schema/counting/lifecycle/service/ API/migration tests, 100 directly affected service regressions, the migration plus live-PostgreSQL gate, all 43 Metis OpenAPI contract tests, Ruff, strict Python 3.12 Mypy, current OpenAPI/client generation checks, targeted changed-file web lint, the full Metis web typecheck, 86 files / 346 web unit tests, the production stub scan, nine evidence- generator tests, and 15 three-engine browser executions with automatic accessibility and mobile coverage.
      • These controls establish the bounded M5.13 exposure and compromise control plane. They do not claim connected production occurrence telemetry, an external breach feed, delivered outbound notifications, production deployment, universal security thresholds, representative security-reviewer usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.14 Implement content and answer freshness gates. Evidence: source or standard updates mark dependencies stale.
    • M5.14.a Bind lesson/item/answer/rationale/feedback/hint to exact source/ claim/citation/standard/objective/tool/policy revisions and review dates. Evidence: dependency tests.
    • M5.14.b Detect source update/retraction, standard supersession, factual/ temporal review expiry, dependency change, and conflicting current evidence. Evidence: freshness fixtures.
    • M5.14.c Classify current/due/stale/blocked/unknown/N/A by content risk and use, preserving prior review evidence. Evidence: verdict tests.
    • M5.14.d Propagate stale state to tutor prompts, assessments, builds, and publication/delivery according to policy. Evidence: propagation test. <!-- M5.14 bounded completion evidence (2026-09-01):
      • Added the versioned metis.content-answer-freshness policy, exact subject/dependency/review/signal/verdict contracts, deterministic hash sealing and evaluation, tenant-safe service/API, generated OpenAPI client, and Alembic 040. Applicable reviews cover all six named content kinds and bind any of the seven required dependency kinds to exact revision, SHA-256, authority, bound time, current observation, evidence witness, reviewer, and review/evaluation dates.
      • The evaluator detects source update/retraction, standard or other dependency supersession, exact revision drift, unavailable current authority, conflicting current evidence, missing observation, due windows, and factual/temporal expiry. Four risk bands and declared uses deterministically resolve the ordered current/due/stale/blocked/ unknown/not-applicable vocabulary. Not-applicable is an explicit, reviewed low/moderate authoring-only result and cannot conceal learner-facing content.
      • content_freshness_reviews, content_freshness_signals, and content_freshness_verdicts are append-only, tenant-linked ledgers. Idempotency conflicts, cross-tenant reads, reporter mismatches, and evidence/hash drift fail closed. Dependency signals append new verdicts for every affected exact revision while retaining the sealed review and prior verdict history; PostgreSQL advisory locks serialize review and dependency writers.
      • Runtime admission re-evaluates the current clock and signal ledger, not merely the stored label. It executes before tutor session/start generation, assessment question delivery/submission and publication, lesson build, course/admin publication, and published learner delivery. Legacy integrated content without an applicable M5.14 review remains available; the standalone admission API fails closed when an exact review is absent.
      • Added the instructor/admin Freshness dependency desk with a compact exact-revision queue, six-state rail, five-surface admission matrix, bounded review clock, reviewed-versus-observed dependency ledger, findings, immutable history, and admin-only dependency signal intake. Learners are denied before evidence transport. The restrained ink/linen ledger layout follows the frontend skill, uses amber only for operational change, supports reduced motion, and has no horizontal overflow at 390 px.
      • Deterministic evidence is serialized in evidence/v1-workbenches/metis-content-freshness.json. Verification: 38 focused schema/helper/service/API/migration/propagation tests, 68 affected service regressions, Alembic head plus one live PostgreSQL propagation/tenant/three-ledger immutability test, 47 OpenAPI/export tests, generated-client drift check, targeted Ruff and strict MyPy, targeted web lint and full web typecheck, 87 files / 351 web unit tests, production stub/random scan, eight evidence-generator tests, and 15 Chromium/Firefox/WebKit Playwright executions with automatic serious/critical axe analysis and mobile overflow coverage all pass.
      • These controls establish the bounded M5.14 freshness plane. They do not claim a connected production authority registry, periodic reconciliation scheduler, production deployment, representative reviewer usability, or formal accessibility conformance. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.15 Implement cultural, regional, and language review assignments. Evidence: required locale reviewers block publication.
    • M5.15.a Define required reviewer qualifications/locale/region/community/ discipline/independence by content/item/audience/risk/target. Evidence: assignment policy.
    • M5.15.b Match eligible reviewers without exposing protected attributes, apply workload/conflict rules, deadlines, substitutes, and escalation. Evidence: assignment tests.
    • M5.15.c Provide exact locale variant/source/context/comparison/evidence and structured cultural/regional/language review decisions. Evidence: review tests.
    • M5.15.d Require fresh decision/quorum for each applicable locale/region and block publication on missing/expired/conflicted assignment. Evidence: gate tests. <!-- M5.15 bounded completion evidence (2026-09-02):
      • Added the content-addressed metis.cultural-regional-language-review policy and closed contracts for eight content kinds, four audiences, four risk bands, three review dimensions, exact target revision/hash, locale, region, community context, discipline, qualifications, independence, workload, deadline, reserve, escalation, and fresh decision/quorum rules. High/critical scopes require both exact region and community context; audience and risk rules raise quorum up to four distinct reviewers per exact requirement.
      • Matching accepts only closed capability, availability, workload, and conflict/independence envelopes. Eight named protected attributes are rejected as extra input; the response is reduced to opaque reviewer identity, eligible requirement identities, reason codes, active load, and deterministic rank. Declared financial, institutional, or personal conflicts always disqualify; risk-specific independence separately controls authorship. Capacity limits, one reserve per requirement, assignment windows, conflict declarations, reserve activation, and overdue escalation are enforced server-side.
      • Review packets bind the exact target and locale variant to source and comparison hashes, audience/risk/discipline/region/community context, unique dimension-specific evidence, and a canonical packet hash. Structured findings cannot cite evidence outside the packet or evidence captured after the decision; material/change-required findings require exact evidence; approvals cannot conceal blocking or change-required findings. Decisions and their validity windows are hash sealed.
      • Alembic 041 adds tenant-linked append-only assignment-run, lifecycle- event, decision, and quorum-verdict ledgers with idempotency, composite tenant foreign keys, state/value checks, hash identities, and update/ delete rejection triggers. Live PostgreSQL proves exact matching, idempotent assignment and admission replay, full locale/region/community quorum, tenant isolation, clock expiry, runtime denial/admission, and immutable ledgers.
      • Runtime admission recomputes every applicable exact requirement from current events, decisions, expiry, conflict, and target revision. It executes before course/admin and assessment publication, assessment delivery/submission, lesson build, and published lesson delivery. Explicitly registered pending, missing, negative, conflicted, expired, or superseded scopes deny before protected mutation, artifact writing, or learner delivery. Legacy content with no declared M5.15 scope remains available; the standalone admission API fails closed for a missing or mismatched registered witness.
      • Added the instructor/admin Locale review assignment desk: exact target and policy witness, privacy-reduced matching composer, locale/region/ community quorum rail, assignment/deadline/reserve lifecycle, exact source-to-variant context/comparison/evidence sequence, structured decision form, and immutable history. Learners are denied before review transport; the 390 px layout has no horizontal overflow; entrance motion is transform-only with reduced-motion support; and automatic serious/ critical Axe analysis passes in Chromium, Firefox, and WebKit.
      • Deterministic evidence is serialized in evidence/v1-workbenches/metis-cultural-review.json. Verification: 24 focused M5.15 schema/helper/service/API/migration/propagation tests, including the live PostgreSQL gate; 57 affected service regressions; 50 OpenAPI and release-coverage tests; targeted Ruff and strict Python 3.12 Mypy; byte compilation; current OpenAPI export/generated-client checks; targeted zero-warning web lint and full web typecheck; 88 files / 355 Metis web unit tests; production stub/random scan; eight evidence- generator tests; and 18 Chromium/Firefox/WebKit Playwright executions with automatic accessibility, authorization, transport, keyboard, and mobile-overflow coverage all pass.
      • These controls establish the bounded M5.15 cultural-review assignment plane. They do not claim a connected production qualification registry, an independent protected-attribute audit, production deployment, representative reviewer usability, or formal accessibility conformance. Browser API transport is mocked. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.16 Persist raw measures independently from policy verdicts. Evidence: threshold changes recompute without data loss.
    • M5.16.a Define immutable measurement records with subject, evaluator/ model/code/config/corpus versions, inputs, values/units/regions, uncertainty/applicability, and hash. Evidence: schema tests.
    • M5.16.b Define verdict records referencing exact measurements plus gate/ policy/threshold/profile versions and rationale. Evidence: verdict schema.
    • M5.16.c Recompute new verdict identities on threshold/policy change without modifying/reprocessing valid raw measures. Evidence: recompute tests.
    • M5.16.d Require new measurement when subject/evaluator/input changes and retain all historical projections. Evidence: lifecycle invariants. <!-- Evidence (2026-09-02):
      • Added closed, immutable RawMeasurementRequest / RawMeasurementRecord contracts with exact subject revision hashes; evaluator and evaluator-artifact hashes; explicit model, code, config, and corpus used/not-applicable witnesses; ordered input snapshots; bounded regions; typed values and units; uncertainty, applicability, rationale, evidence references, execution time, input-set hash, and a canonical content-addressed raw-measurement:<sha256> identity. Raw contracts reject undeclared gate, policy, threshold, verdict, and outcome fields.
      • Added independent PolicyVerdictRecord contracts whose facts are only exact measurement ID/hash references. Gate, policy, profile, threshold snapshot, ordered rule decisions, rationale, predecessor, evaluation time, outcome, and policy-verdict:<sha256> identity remain in the policy projection. Unit mismatches, missing metrics, cross-subject projections, unchanged-policy recomputes, and history branching fail closed. The recompute request has no raw payload or replacement measurement-ID field.
      • Alembic revision 042 and SQLAlchemy models persist three separate append-only tenant-linked tables: measurement_ledger_raw_measurements, measurement_ledger_policy_verdicts, and measurement_ledger_verdict_measurements. Composite foreign keys bind verdict history to exact same-tenant raw records, a self-reference retains predecessor history, a one-successor constraint prevents branching, and PostgreSQL triggers reject every update/delete.
      • /api/measurement-ledger now exposes authenticated catalog, exact- subject workspace, raw append, initial projection, and policy-only recompute operations. Exact replay is content-idempotent. Live PostgreSQL tests prove a threshold change flips pass to fail while the raw count, raw identity, raw hash, and exact verdict-to-measurement link remain unchanged; all projections stay readable. Subject, evaluator, and input mutations append distinct raw facts. Unit tests additionally cover model, code, config, corpus, region, value, uncertainty, and applicability identity changes.
      • Added /create/measurement-ledger, a restrained fact-ledger workbench that keeps immutable readings, units, uncertainty, exact regions, inputs, component provenance, and hashes visually separate from the versioned policy lens. It supports raw append, first projection, successor recompute, full non-destructive history, role denial, text- labelled outcomes, reduced motion, and 390px responsive use. Browser assertions inspect outbound bodies to prove raw transport is policy- free and recompute cannot replace/resend a measurement.
      • Canonical generated evidence is evidence/v1-workbenches/metis-measurement-ledger.json. Verification: 29 focused M5.16 schema/helper/API/migration tests including three live PostgreSQL lifecycle gates; 58 OpenAPI/export/release-coverage tests; targeted Ruff and strict Python 3.12 Mypy; byte compilation; current OpenAPI export and generated-client drift checks; targeted zero-warning web lint and full web typecheck; 89 files / 361 Metis web unit tests; production stub/random scan; eight evidence-generator tests; and 24 Chromium/Firefox/WebKit Playwright executions with automatic Axe, authorization, transport, lifecycle, and mobile-overflow coverage all pass.
      • These controls establish the bounded M5.16 measurement-ledger plane. The richer M5.17 state vocabulary and M5.18 exact evaluator, corpus, policy, threshold, and environment registry are now integrated. They do not claim production registry deployment, representative steward usability, or formal accessibility conformance. Browser API transport is mocked. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, live-database, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.17 Expose uncertainty, non-applicability, and insufficient-data states. Evidence: no unknown state renders as pass.
    • M5.17.a Define pass/fail/advisory/manual-review/N/A/insufficient/ unavailable/error/stale/unknown states and permitted transitions. Evidence: state vocabulary.
    • M5.17.b Require reason/evidence/uncertainty/sample/applicability/ remediation fields appropriate to each state. Evidence: schema tests.
    • M5.17.c Aggregate mandatory gates without converting missing/unknown/N/A/ unavailable/insufficient into numeric zero or pass. Evidence: aggregation property tests.
    • M5.17.d Render/export every state with text/semantics, freshness, and no color-only or false green representation. Evidence: UI/API tests. <!-- Evidence (2026-09-02):
      • Added a closed, content-addressed ten-state vocabulary for pass, fail, advisory, manual_review, not_applicable, insufficient_data, unavailable, error, stale, and unknown. Each state defines its mandatory-gate effect, required evidence fields, and remediation obligation. Seven explicit transition causes bind before/after state pairs to raw-measurement identity: policy-only recompute may change only derived state; new measurement, freshness expiry, evaluator failure/recovery, and manual review must carry the corresponding identity witness. Invalid transitions fail closed.
      • Measurement schema version 1.1 now carries typed observation/sample, freshness, applicability, uncertainty, reason/evidence, and remediation witnesses. Absolute, relative, and confidence-interval uncertainty is evaluated as an interval; a threshold crossing routes to manual review. Expired current evidence becomes stale, raw non-observed states remain non-observed, and legacy records without state evidence become unknown rather than inheriting a numeric pass.
      • Mandatory aggregation allows release only when every mandatory decision is explicitly pass; advisory-only rows do not satisfy mandatory requirements. Tests exhaust all 729 three-gate combinations over the nine non-advisory mandatory states, plus no-gate, legacy, malformed- evidence, uncertainty-crossing, and freshness-expiry cases. No missing, N/A, insufficient, unavailable, stale, error, manual-review, or unknown decision converts to numeric zero or pass.
      • Authenticated outcome-vocabulary and workspace-export APIs expose the exact vocabulary, current/history ledger, transition witness, and content hash. The measurement-ledger workbench presents the state word, release effect, reason, evidence, uncertainty, sample, applicability, freshness, and remediation as text; color is a secondary border cue. Its JSON export includes all ten states without false green semantics. The restrained evidence-sheet hierarchy and 390px layout were visually inspected under the mandatory frontend workflow.
      • Canonical generated evidence is evidence/v1-workbenches/metis-outcome-states.json. Verification: 50 focused schema/helper/API/migration tests including three live PostgreSQL gates; 58 OpenAPI/export/release-coverage tests; targeted Ruff and strict Python 3.12 Mypy; byte compilation; current OpenAPI and generated-client drift checks; targeted zero-warning web lint and full web typecheck; 89 files / 361 Metis web unit tests; production stub/random scan; eight evidence-generator tests; and 24 Chromium, Firefox, and WebKit Playwright executions covering all ten states, structured evidence, export, release blocking, authorization, exact transport, 390px reflow, and automatic serious/critical Axe analysis.
      • These controls establish the bounded M5.17 outcome-state plane. The M5.18 exact provenance registry is now integrated. They do not claim production registry deployment, representative steward usability, or formal accessibility conformance. Browser API transport is mocked. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, database, lint, type, unit, accessibility, responsive, visual, and cross-browser gates pass. -->
  • M5.18 Version every evaluator, policy, corpus, and threshold. Evidence: gate result provenance.
    • M5.18.a Register evaluator code/model/config/runtime/dependency version/ hash/source/license/lifecycle and supported inputs. Evidence: registry.
    • M5.18.b Register corpus manifest/labels/splits/hashes/provenance/rights/ coverage/privacy and calibration status. Evidence: corpus registry.
    • M5.18.c Register policy/rubric/threshold/band/aggregation/profile versions with approval/effective dates and code floors. Evidence: policy registry.
    • M5.18.d Bind every measurement/verdict to exact versions/environment and reproduce or explicitly label nonreproducibility. Evidence: provenance tests. <!-- M5.18 completion evidence (2026-09-02):
      • metis.schemas.provenance_registry closes immutable evaluator, corpus, policy-bundle, threshold-band, and execution-environment contracts. Used artifacts require version, SHA-256, source URI/source revision, and license; not-applicable artifacts require an explicit rationale and cannot carry false version claims. Evaluators additionally pin runtime, dependencies, supported inputs, determinism, replay, effectivity, ownership, and lifecycle evidence.
      • Corpus records pin manifest, labels, immutable splits, source lineage, rights windows, coverage and known gaps, privacy classification and retention, calibration evidence/approval, and lifecycle. Policy bundles pin gate, policy, rubric, profile, threshold artifact/rule IDs/bands, aggregation, approval/effective dates, code floors, determinism, and replay procedure. Environments pin OS/architecture, runtimes, container, dependency lock, variables, infrastructure, hardware, locale/timezone, network mode, capture, retention, and availability.
      • Alembic 043 creates four append-only PostgreSQL registry tables, five tenant-scoped ledger foreign keys, supporting indexes and uniqueness constraints, and the complete ten-state outcome constraint. Live PostgreSQL gates prove idempotent content addressing, exact measurement and verdict pins, successor recomputation without remeasurement, registry immutability, tenant isolation, and a 5.18 export containing only the referenced registry subset.
      • New writes require schema 1.2 plus exact registry IDs and content hashes. Service validation rejects inactive/expired evaluators, corpora, environments and policies; unsupported inputs; rights expiry; mismatched components, thresholds, rules, approval, code floors, or source replay. Deterministic executions may claim bit-exact replay; seeded executions require a captured seed; nondeterministic, incomplete, and legacy records remain explicitly nonreproducible with reason codes, limitations, and replay instructions.
      • Authenticated instructor/admin APIs expose the registry catalog and append endpoints for every registry kind. Regenerated OpenAPI and TypeScript clients carry the 5.18 contracts. The measurement workbench resolves writes against active/effective registry versions, disables complete-looking mismatches, renders the exact evaluator/corpus/policy/ environment spine with approval, coverage, rights, privacy, calibration, code-floor and content-hash evidence, and states replay truth in text.
      • Verification: 69 focused schema/helper/API/migration/live-PostgreSQL tests; 53 OpenAPI/export/release-coverage tests; Python lint and strict typecheck; OpenAPI and generated-client drift checks; full Metis web lint/typecheck; 89 files / 361 web unit tests; production stub/random scan; eight M5.18 evidence-generator tests plus refreshed eight-test M5.16 and M5.17 generators; and 24 Chromium/Firefox/WebKit Playwright executions with automatic serious/critical Axe, authorization, exact transport, registry mismatch, export, visual, and 390px reflow coverage.
      • These controls establish the bounded M5.18 provenance-registry plane. Browser API transport is mocked. They do not claim production registry deployment, representative steward usability, or formal accessibility certification. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, migration, database, lint, type, unit, accessibility, responsive, visual, and cross-browser gates pass. -->
  • M5.19 Implement atomic-claim decomposition and claim-to-source-span linking for lesson, feedback, hint, worked-solution, and lecture content. Evidence: labeled corpus measures missed and incorrectly split claims.
    • M5.19.a Define atomic claim identity/revision, subject/predicate/object or normalized proposition, qualifiers/scope/modality/time, content kind/ locator, source-span links, and status. Evidence: claim schema.
    • M5.19.b Segment each named content kind into factual, procedural, normative, interpretive, mathematical, attribution, and nonclaim text with content-specific rules. Evidence: segmentation fixtures.
    • M5.19.c Preserve compound relationships, coreference/context, equations/ conditions, negation, examples, quotations, and pedagogical scaffolding without losing meaning. Evidence: adversarial corpus.
    • M5.19.d Link each claim to exact source/extraction revision and one or more typed spans/locators with support role and human confirmation. Evidence: linking tests.
    • M5.19.e Evaluate missed/spurious/over-split/under-split/boundary/link precision/recall by content/discipline/locale and uncertainty. Evidence: labeled corpus report. <!-- M5.19 completion evidence (2026-09-02):
      • metis.schemas.claim_decomposition closes atomic claim/revision/status, normalized proposition or triple, scope, modality, qualifiers, exact content revision/locator, semantic relationships, typed source spans, link confirmation, immutable decision history, engine witness, corpus metrics, error taxonomy, strata, and workspace contracts. Exact content and source spans must match their length and SHA-256; revision history must be contiguous.
      • The deterministic 5.19 rule engine covers lesson, feedback, hint, worked-solution, and lecture content and the closed factual, procedural, normative, interpretive, mathematical, attribution, and nonclaim taxonomy. Content-specific clause rules preserve compound relations, coreference context, equations, conditions, negation, examples, quotations, and pedagogical scaffolding through explicit features, relationships, and exact context locators.
      • Runs accept only current human-authorized M2.6 canonical source claim revisions. Every candidate link pins the canonical claim/revision, extraction/output hash, ingestion item/source hash, node/anchors, half-open character span, exact quote/hash, source proposition, typed role, similarity, and uncertainty. Append-only administrator decisions confirm, correct, reject, or defer a claim and decide every candidate link; stale source authority blocks confirmation without rewriting prior evidence. Idempotency, tenant visibility, integrity hashes, and crash-safe immutable file writes are enforced.
      • The checked-in 12-case adversarial corpus covers all five content kinds, all seven segment kinds, nine disciplines, and en-US/es-ES. Runtime evaluation reports claim detection P/R/F1 0.85/0.85/0.85, boundary detection 0.60/0.60/0.60, and source linking 0.722222/0.8125/0.764706, each with 95% Wilson intervals. It retains three missed, three spurious, two over-split, two under-split, six boundary, three link false-positive, and one link false-negative error receipts, plus content-kind, discipline, and locale strata rather than presenting perfect scores.
      • Authenticated instructor APIs create/read exact runs and expose the report; administrator APIs expose the cross-package workspace, create runs, and append human decisions. Regenerated OpenAPI and TypeScript contracts carry the full surface. The admin evidence ledger aligns the selected content and source highlights, exposes semantic context and immutable history, creates hash-bound runs, locks stale decisions, and displays corpus error/uncertainty evidence at desktop and 390px widths.
      • Verification: 17 focused decomposition/source-authority/API tests; 53 OpenAPI/export/release-coverage tests; checked OpenAPI and generated- client drift; Python lint/compile; admin lint and full typecheck; 27 files / 40 admin unit tests; eight evidence- generator tests; 82 product-graph tests; and six Chromium/Firefox/ WebKit Playwright executions covering exact synchronized spans, corpus results, immutable human link review, stale locks, five-kind creation transport, 390px reflow, and automatic serious/critical Axe analysis.
      • These controls establish the bounded M5.19 decomposition and candidate- linking plane. Candidate links expressly do not claim source support, entailment, faithfulness, contradiction resolution, or factual correctness; those remain M5.20 gates. Browser API transport is mocked. Production deployment, representative reviewer usability, and formal accessibility certification are not established. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.20 Implement retrieval-grounded support, citation sufficiency, faithfulness, contradiction, and cross-source conflict gates separately. Evidence: entailment, partial support, irrelevant citation, conflict, inaccessible source, and abstention fixtures.
    • M5.20.a Resolve exact claim, cited/retrieved source spans, authority/ rights/currency/accessibility, retrieval version/rank, and evaluation context. Evidence: gate input tests.
    • M5.20.b Measure support/entailment per source span with full/partial/ unsupported/contradicted/unclear/N/A and qualifiers. Evidence: support fixtures.
    • M5.20.c Evaluate citation sufficiency for claim coverage, source authority/diversity/recency/directness and detect irrelevant/overbroad/ inaccessible citations. Evidence: sufficiency fixtures.
    • M5.20.d Compare content wording to supported claim for added certainty, omitted qualifiers, scope/number/time/causal/attribution changes. Evidence: faithfulness fixtures.
    • M5.20.e Detect within-source and cross-source contradiction/conflict/ version/jurisdiction/perspective differences without forcing consensus. Evidence: conflict fixtures.
    • M5.20.f Keep each raw result/verifier/version/uncertainty separate and abstain/manual-review on inaccessible/insufficient/ambiguous evidence. Evidence: state tests.
    • M5.20.g Apply risk-specific policy verdicts without averaging one failed dimension into pass. Evidence: aggregate gate tests. <!-- M5.20 completion evidence (2026-09-02):
      • metis.schemas.grounding_gate closes the exact confirmed M5.19 claim, content revision, retrieved/cited source links and spans, query/index/ retriever versions and hashes, rank/score, risk and use context, source authority IDs, rights, currency, accessibility, directness, scope, jurisdiction, raw-result, verifier, uncertainty, policy, review, corpus, and workspace contracts. Response evidence arrays are required even when empty, and every run retains an integrity hash.
      • The append-only service accepts only the current human-confirmed M5.19 claim revision and every confirmed link, re-resolves approved M2.5 governance, rejects false authority/rights/currency/jurisdiction snapshots, prepares governance-bound eligible contexts, locks stale evidence, enforces tenant visibility and idempotency, and recovers interrupted immutable run/decision writes without rewriting raw policy.
      • Five separately witnessed deterministic verifiers retain per-span support, nine citation criteria, wording faithfulness, pairwise source conflict, and risk policy. Semantic support auto-passes only an exact normalized proposition match; deterministic polarity/number conflicts can block, while non-exact or inaccessible evidence abstains or requires manual review. Full/partial/unsupported/contradicted/unclear/N/A, qualifiers, irrelevant/overbroad/inaccessible links, all seven wording changes, and within/cross-source/version/jurisdiction/perspective differences remain explicit. A single source never implies consensus.
      • General/elevated/high/critical profiles set independent authority, diversity, recency, and directness floors. The policy outcome follows the worst required raw dimension without an aggregate score; only an all-dimension pass permits release, and an administrator disposition cannot override a block, abstention, or manual-review result.
      • The checked-in 24-case adversarial corpus covers ten required fixture families, all six support states, all seven faithfulness changes, all five conflict kinds, and all four risks. Runtime evaluation reports support 23/23, citation 24/24, faithfulness 24/24, conflict 6/6, and policy 24/24 fixture matches with 95% Wilson intervals and expressly labels this bounded conformance rather than open-domain NLI accuracy or proof that an authoritative source is true.
      • Authenticated instructor APIs create/read exact runs and expose corpus results; administrator APIs expose the cross-package docket, create runs, and append dispositions. Regenerated OpenAPI and TypeScript contracts publish the full surface. The admin evidence court aligns claim/source highlights, switches among independent dimensions, exposes verifier/version/uncertainty receipts and non-consensus conflict states, creates governance-bound retrieval receipts, locks stale review, and reflows at 390px.
      • Verification: 13 focused service/contract tests; checked OpenAPI and generated-client drift; Python lint, strict typecheck, and compile; admin lint/typecheck and 28 files / 41 unit tests; eleven evidence- generator tests; product-graph gates; and six Chromium/Firefox/WebKit Playwright executions covering synchronized evidence, all dimensions, immutable non-overriding disposition, staleness, critical-risk transport, 390px overflow, and automatic serious/critical Axe.
      • These controls establish a conservative bounded M5.20 grounding-gate plane, not a production semantic-entailment model. Browser API transport is mocked. Open-domain NLI accuracy, source truth, production deployment, representative reviewer usability, and formal accessibility certification are not established. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.21 Implement executable mathematics verification using ratified symbolic/numeric tooling and assumption/domain tracking. Evidence: proofs, equations, units, edge cases, and unsupported constructs do not reduce to an LLM opinion.
    • M5.21.a Define supported math languages/constructs, parser/CAS/numeric/ unit/proof tool versions, assumptions/domains, tolerances, resource limits, and explicit unsupported behavior. Evidence: verifier contract.
    • M5.21.b Parse semantic math, bind variables/units/definitions/assumptions/ domains, reject ambiguous notation, and retain source spans. Evidence: parsing fixtures.
    • M5.21.c Verify symbolic equivalence/identity/equation/inequality/solution sets under stated domains and detect extraneous/missing cases. Evidence: symbolic goldens.
    • M5.21.d Verify numeric computations/approximations/rounding/error bounds/ edge and randomized property cases with deterministic seeds. Evidence: numeric goldens.
    • M5.21.e Verify dimensional/unit consistency/conversion/significant figures and physical-domain constraints. Evidence: unit fixtures.
    • M5.21.f Verify supported proof steps/rules separately; return localized error/unknown/timeout/unsupported rather than provider opinion. Evidence: proof fixtures.
    • M5.21.g Sandbox tooling and persist inputs/outputs/tool versions/traces separately from policy/human verdict. Evidence: security/provenance tests. <!-- M5.21 completion evidence (2026-09-02):
      • metis.schemas.mathematics_verification closes the 5.21.0 target, exact source-span/hash, explicitly bound variable/domain/definition, tolerance, resource-limit, symbolic/numeric/unit/proof obligation, parse artifact, localized diagnostic, raw trace/counterexample/ missing/extraneous case, pinned tool/engine, sandbox receipt, worst- state policy, append-only run/review, corpus, and workspace contracts. The expression language is explicitly the bounded metis-ascii-math-v1 subset; unsupported behavior fails closed.
      • The restricted Python-AST parser constructs only reviewed arithmetic, bounded-power, named-function, relation, and solution-set nodes. It never calls eval, sympify, or parse_expr, never admits implicit multiplication or arbitrary attributes/calls, binds every name before execution, preserves exact source spans, and returns unknown or unsupported for ambiguity, unbound names, conditional sets, and unsupported multivariate/proof shapes.
      • The exact worker uses SymPy 1.14.0 and mpmath 1.3.0 for equivalence, defined-domain comparison, equations, inequalities, and missing or extraneous solution cases; Decimal arithmetic for explicit, edge, rounding, maximum-error, and deterministic seeded property cases; and Pint 0.25.2 for independently traced conversion, dimensionality, significant-figure, and physical-domain checks. Supported proof steps compare each dependency's exact solution set and retain per-step pass/fail/unknown/timeout/unsupported state and localized diagnostics.
      • Bubblewrap 0.9.0 runs the worker with all namespaces unshared, no network, a read-only root, bounded tmpfs, uid/gid 65534, all capabilities dropped, a cleared environment, and wall/CPU/address- space/process/open-file/output/core limits. Tool versions and artifact hashes are checked before launch; missing or unratified tooling has no unsandboxed fallback. Network/root-write/uid probes, bounded stdout, stderr, exit state, timeout, hashes, and forced process-group teardown remain in the receipt.
      • Policy follows the worst executable task state without averaging. A failed task or high/critical unknown state blocks; lower-risk unknown, timeout, or unsupported results require review. Human dispositions are immutable audit entries and cannot change raw mathematics or policy. Tenant-scoped append-only files enforce exact spans, expression IDs, idempotency, integrity hashes, interrupted-write recovery, and stale- revision review locks.
      • The checked-in 22-case adversarial corpus covers 22 fixture families: ambiguity, unbound and unsupported constructs, symbolic equivalence, domain loss, equation/inequality sets, extraneous cases, numeric exact/ edge/rounding/error/seeded cases, unit conversion/dimensionality/ significant figures/physical bounds, valid and invalid proof steps, hidden assumptions, and unsupported proof dependency shapes. Actual sandbox evaluation reports parsing 3/3, symbolic 6/6, numeric 5/5, units 4/4, proofs 4/4, and isolation 1/1 with 95% Wilson intervals.
      • Authenticated instructor APIs expose evaluation and exact run creation/ retrieval. Administrator APIs expose the docket, create runs, and append dispositions. Regenerated OpenAPI and TypeScript contracts publish the closed surface. The admin proof bench synchronizes exact source highlights with parse and proof steps, preserves separate symbolic/numeric/unit/proof traces and counterexamples, exposes pinned tools and isolation receipts, locks stale reviews, and reflows without horizontal overflow at 390px.
      • Verification: 14 focused engine/service/OpenAPI tests; checked OpenAPI and generated-client drift; Python lint, strict typecheck, and compile; admin lint/typecheck and 29 files / 42 unit tests; twelve deterministic evidence-generator tests; product-graph gates; and six Chromium/ Firefox/WebKit Playwright executions covering synchronized spans, separate raw dimensions, localized proof steps, tool/sandbox receipts, immutable non-overriding review, staleness, exact create transport, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-mathematics-verification.json through inventory:v1-metis-mathematics-verification and its verification script. These controls establish deterministic conformance for the ratified subset, not correctness for all mathematics. Browser API transport is mocked. Production deployment, representative reviewer usability, physical assistive-technology testing, and formal accessibility certification are not established. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.22 Implement sandboxed code correctness using compilation/execution, tests, resource limits, language/runtime versions, and deterministic or explicitly stochastic expectations. Evidence: correct, wrong, unsafe, flaky, timeout, and unsupported fixtures.
    • M5.22.a Define supported language/runtime/compiler/dependency/test harness versions, input/output/files/API, deterministic seed/repetition, scoring, and unsupported policy. Evidence: verifier contract.
    • M5.22.b Build immutable source/dependency/test artifacts and compile/ static-check with exact diagnostics in an isolated environment. Evidence: compile fixtures.
    • M5.22.c Execute visible/hidden/example/property/metamorphic/performance tests with independent cases and exact expected/tolerant/stochastic assertions. Evidence: correctness suite.
    • M5.22.d Enforce CPU/memory/wall/process/file/output/network/syscall/ secret limits and terminate entire sandbox on breach/timeout. Evidence: unsafe/timeout tests.
    • M5.22.e Detect nondeterminism/flakiness through repeated seeded runs and report distribution/unstable state rather than lucky pass. Evidence: flaky fixtures.
    • M5.22.f Persist per-test outputs/resources/tool versions and separate raw execution from score/verdict/manual review. Evidence: result tests. <!-- M5.22 technically complete; operational readiness remains open.
      • The closed 5.22.0 contract binds dependency-free Python 3.12.14, CPython interpreter/compiler artifacts, the reviewed harness, exact Python image digest, Docker 29.5.3, the default-deny seccomp profile, immutable source/support artifacts, entrypoint/API/stdin/arguments/ output mode, dependency policy, deterministic or explicitly stochastic repetitions, weighted scoring, resource ceilings, and fail-closed unsupported behavior. There is no host-execution or runtime-install fallback.
      • Every source artifact is content-hash checked, traversal-safe, materialized read-only, compiled without submission execution inside the same pinned least-privilege image, and returned with exact syntax locations, imported-module evidence, conservative boundary warnings, artifact hashes, compiler/tool/image/seccomp witnesses, bounded output, and confirmed container teardown.
      • Fresh containers execute visible and hidden examples, explicit property cases with generator witnesses, metamorphic pairs, performance ceilings, and adversarial security probes independently. Exact, normalized-text, JSON, numeric-tolerance, deterministic, and declared stochastic assertions remain explicit. Hidden expected output is never mounted into submission execution and is redacted from raw attempt projections.
      • CPU share, memory/swap, wall time, process count, open files, bounded writable tmpfs, aggregate output, network-none, read-only root, unprivileged UID, dropped capabilities, no-new-privileges, protected secret non-forwarding, and default-deny syscalls are enforced. Timeout or resource breach forcibly removes the container and terminates the client process group, then proves the container is absent. Live tests cover wall timeout, memory pressure, output bombs, network, filesystem, process, syscall, secret, and teardown boundaries.
      • Repetitions derive stable per-case seeds and persist every attempt plus output distributions. Divergent deterministic output becomes unstable; declared stochastic output must remain inside its bounded allowed set and meet minimum diversity, so a lucky single pass cannot hide flakiness.
      • Tenant-scoped append-only storage binds compiler/test/score receipts back to the exact request, rejects divergent idempotency reuse and tampering, recovers interrupted idempotency receipt writes, locks stale revision review, and appends non-overriding dispositions. Instructor and admin APIs expose immutable runs, workspace/evaluation evidence, and review. Raw compilation, per-test attempts/resources/distributions, the 10,000-basis-point score, five independent policy dimensions, and manual review remain structurally separate.
      • The admin forensic bench synchronizes a line-addressable immutable source receipt, exact compiler diagnostics, separate visible/hidden/ property/metamorphic/performance/stochastic traces, expandable stdout/ stderr/expected/resource receipts, output distributions, tool/image/ seccomp witnesses, score and policy, staleness, and append-only review. It reflows without horizontal viewport overflow at 390px and honors reduced motion.
      • The executable corpus contains 17 correct, wrong, hard-coded-hidden, property, metamorphic, performance, timeout, output-limit, unsafe, secret, flaky, stochastic, compile, and unsupported-dependency cases. All 17 observed outcomes match labels across nine separately reported dimensions with Wilson 95% intervals. This is bounded conformance for the ratified runtime, not correctness for arbitrary languages, dependencies, APIs, or programs; static import warnings are not a complete vulnerability or malware analysis.
      • Verification: 22 focused live Docker engine, append-only service, and closed OpenAPI tests; OpenAPI and generated-client drift checks; Python lint, strict typecheck, and compile; admin lint/typecheck and 30 files / 43 unit tests; eleven deterministic evidence-generator tests; product- graph gates; and six Chromium/Firefox/WebKit Playwright executions covering source/diagnostic synchronization, independent raw tests, hidden redaction, distributions, resources, tool/isolation receipts, stale locking, immutable non-overriding review, exact create transport, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-code-verification.json through inventory:v1-metis-code-verification and its verification script. Browser API transport is mocked. Production deployment, representative reviewer usability, physical assistive-technology testing, and formal accessibility certification are not established. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.23 Implement step/process verification for worked solutions, not just final-answer matching. Evidence: localized error, invalid shortcut, circular reasoning, and correct-alternative-path fixtures.
    • M5.23.a Define solution step identity/order/dependencies, expression/ claim, transformation/rule/justification, assumptions/domain, evidence, and intermediate result. Evidence: process schema.
    • M5.23.b Verify each step follows from permitted prior steps/premises via ratified symbolic/numeric/rule/source tools and localize first/all errors. Evidence: localized error fixtures.
    • M5.23.c Detect hidden assumption/domain loss, division by zero, extraneous solution, unit error, invalid shortcut, circular reasoning, and unproved lemma. Evidence: adversarial fixtures.
    • M5.23.d Accept correct alternative paths/orderings/equivalent forms when independently verified, without exact-template matching. Evidence: alternative-path fixtures.
    • M5.23.e Distinguish correct/incorrect/partially supported/unknown/ unsupported/timeout per step and aggregate without final-answer override. Evidence: aggregation tests.
    • M5.23.f Present verifier trace and human-review corrections; keep final answer verification as a separate dimension. Evidence: review tests. <!-- M5.23 technically complete; operational readiness remains open.
      • The closed 5.23.0 contract states step identity, contiguous ordinals, declared dependencies, rule ID, justification, assumption and evidence bindings, an optional rule argument, and one intermediate result per step. Intermediate results are a discriminated relation, structured claim, or dimensioned quantity; every expression must equal the exact character span of the worked solution it cites, every excerpt and revision is SHA-256 bound, and variables carry an explicit solve domain. The contract is published closed on the instructor and admin OpenAPI surfaces and in the generated TypeScript client.
      • Fourteen ratified rules run inside a pinned bubblewrap 0.9.0 sandbox (all six namespaces unshared, read-only root, bounded tmpfs, uid 65534, all capabilities dropped, CPU/address-space/process/open-file/output rlimits) against pinned SymPy 1.14.0, Pint 0.25.2, and mpmath 1.3.0. In-sandbox network and root-write probes plus the observed uid are checked before any result is accepted; a missing, unratified, or unprobeable runtime fails closed with no host-execution fallback.
      • A step is checked against its declared earlier steps and premises: a forward or missing dependency, an undeclared assumption, an unstated premise, an unratified rule, or an unverified prerequisite each has its own code and its own state. Transformations are judged by exact solution-set comparison over the declared domain, quantities by Pint conversion and dimensionality, claims by exact revision-bound evidence or one explicit Horn implication. Every step carries its localized source span, and the aggregate reports both the first and all failing steps.
      • Hidden assumptions, domain loss, possible division by zero, extraneous solutions, unit errors, invalid shortcuts, circular reasoning, and unproved lemmas are eight separate verdicts, not one rejection. Nonzero factors must be proven against the declared domain rather than asserted, and a cycle in the dependency graph is localized to the steps that participate in it.
      • Correctness is solution-set equality, never template equality, so an alternative rule (divide-by-a-proven-nonzero instead of add-same-both-sides), an equivalent form (x - 3 = 0 closing a conclusion stated as x = 3), and a different display order for two independent premises are all accepted on their own evidence.
      • The six step states stay structurally distinct: a SymPy conditional set is unknown, an unratified rule is unsupported, a representable claim with unavailable evidence is partially_supported, and a sandbox breach is timeout. Aggregation is the worst raw step state, and the service refuses to persist any aggregate that is not a faithful projection of the raw results. The final answer is verified separately and cannot convert a failed, unknown, or unsupported process into a correct one; process, final answer, domain, and evidence are four independently reported policy dimensions.
      • Tenant-scoped append-only storage binds every run to its request hash, rejects divergent idempotency reuse and tampering, recovers interrupted receipt writes, locks disposition on stale revisions, and rejects a correction naming a step outside the run. Reviews append a disposition and per-step corrections with raw_result_overridden structurally false.
      • The admin forensic bench synchronizes the immutable worked solution with the localized span highlighted in place, per-step rule, dependencies, assumptions, evidence, missing bindings, lost and extraneous solutions, the raw tool trace, the separate final-answer dimension, the aggregate, the four policy dimensions, the ruleset and isolation witness, and append-only review. It reflows without horizontal viewport overflow at 390px and honors reduced motion.
      • The executable corpus contains 19 labeled cases across 19 fixture families and seven dimensions. All 19 observed states and codes match their labels, reported per dimension with Wilson 95% intervals. This is bounded conformance for the ratified ruleset, not a proof system for arbitrary natural-language reasoning; multivariate relations and unratified rules fail closed as unsupported.
      • Verification: 35 focused sandbox, append-only service, and closed OpenAPI tests; OpenAPI and generated-client drift checks; Python lint, strict typecheck; admin lint/typecheck and 31 files / 44 unit tests; eleven deterministic evidence-generator tests; 19 shared admin layout and WCAG signoff executions; and nine Chromium/Firefox/WebKit Playwright executions covering first-error localization, span highlighting, distinct dependent-step states, non-overriding final answers, appended step corrections, stale locking, exact create transport, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-solution-process-verification.json through inventory:v1-metis-solution-process-verification and its verification script. Browser API transport is mocked. Production deployment, representative reviewer usability, physical assistive-technology testing, and formal accessibility certification are not established. A Next production build was not run under the less-than-32-GiB host approval gate; targeted contract, lint, type, unit, accessibility, responsive, and cross-browser gates pass. -->
  • M5.24 Implement calibrated multi-evaluator panels only where deterministic verification is unavailable, with independence analysis, correlated-error controls, confidence calibration, abstention, and human-review bands. Evidence: gold-set and champion/challenger reports.
    • M5.24.a Define applicability gate proving deterministic/rule/source/ executable verification is unavailable or incomplete before panel use. Evidence: routing tests.
    • M5.24.b Register evaluator identity/model/prompt/config/provider/training/ data/tool dependencies, rubric scope, calibration, and lifecycle. Evidence: panel registry.
    • M5.24.c Record shared dependencies/prompts/data/providers as correlation groups and minimum independence/quorum/weight rules. Evidence: independence analysis.
    • M5.24.d Collect blinded per-evaluator label/score/reason/evidence/ confidence/abstention/error against exact subject/rubric revisions. Evidence: collection tests.
    • M5.24.e Calibrate confidence/thresholds/manual bands on versioned gold sets and report overall/subgroup/content errors with uncertainty. Evidence: calibration report.
    • M5.24.f Aggregate with correlated-error controls, preserve disagreement/ minority/timeout/abstention, and never manufacture gate authority. Evidence: consensus fixtures.
    • M5.24.g Run champion/challenger drift comparison before replacement and require human review for configured bands. Evidence: lifecycle report. <!-- M5.24 technically complete; operational readiness remains open.
      • The closed 5.24.0 contract is deterministic over recorded evaluator observations. The engine never calls an evaluator, so every result here is a property of the panel rules, not of any model.
      • A panel is admissible only for rubric criteria that no available deterministic verifier fully covers. Every criterion must carry a probe: an unexamined criterion is probe_evidence_missing, full coverage is deterministic_verification_available, and both refuse the run with a not_applicable verdict. Observations for a deterministically covered criterion are rejected, and the service refuses to persist a consensus that does not cover exactly the residual criteria in order.
      • Each evaluator registers its lifecycle, provider, model and version, prompt revision and hash, config hash, decoding seed, rubric scope, and its model/prompt/provider dependencies (training data, tools, and retrieval indexes are declarable). A criterion outside an evaluator's declared scope is refused rather than silently judged.
      • Correlation groups are derived from the declared dependencies and then audited against the declaration: a shared dependency with no declared group is undeclared-correlation, and a declared group no shared dependency justifies is unjustified-declaration. Evaluators are union-found into independent components, each component votes once, and its weight is capped by the declared ceiling.
      • Every observation must present the blinded subject alias derived from the exact content hash and the exact rubric revision hash, and its label must be on the criterion's declared scale. Scored, abstained, timed-out, and errored observations are structurally distinct shapes. Refused observations are kept as findings rather than dropped.
      • Calibration measures accuracy, a Wilson 95% interval, and a ten-bin expected calibration error on the versioned gold-set revision, reported overall and per subgroup and content kind. An evaluator below the sample, accuracy, or calibration-error floor is uncalibrated and carries no weight; the service refuses to persist a consensus in which an uncalibrated evaluator voted.
      • A tie is not a majority. A criterion without a strict plurality reports no majority at all and keeps every tied label. Disagreement, minority weights, abstentions, timeouts, errors, and exclusions all survive aggregation. The verdict is structurally advisory: authority is advisory_only, replaces_deterministic_verification and gate_authority_manufactured are False, and the strongest thing a panel can demand is human review.
      • Champion/challenger comparison runs on the shared gold-set items and reports overall drift, agreement, and the worst per-subgroup drift. A challenger that improves overall but regresses on a subgroup is rejected; an improvement only ever becomes promotion_requires_human_review. automatic_promotion is False and automatic_replacements is 0 by construction.
      • Tenant-scoped append-only storage binds each run to its request hash, rejects divergent idempotency reuse and tampering, recovers interrupted receipt writes, and locks disposition on stale subject revisions. A deterministic override must name the deterministic evidence it relies on, and raw_panel_result_overridden is structurally false.
      • The admin bench synchronizes the applicability probes, derived correlation groups and findings, per-evaluator calibration with subgroup and content-kind intervals, per-criterion consensus with the preserved minority, rejected-observation findings, the lifecycle comparison, the advisory verdict, and append-only review. It reflows without horizontal viewport overflow at 390px and honors reduced motion.
      • The executable corpus contains 25 labeled cases across 24 fixture families and seven dimensions. All 25 observed outcomes and codes match their labels, reported per dimension with Wilson 95% intervals.
      • Verification: 38 focused engine, append-only service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin lint/typecheck and 32 files / 45 unit tests; thirteen deterministic evidence-generator tests; 22 shared admin layout and WCAG executions; and nine Chromium/Firefox/WebKit Playwright executions covering applicability refusal, correlated voices, per-subgroup calibration, a tie that names no majority, an unanswered residual criterion, advisory-only authority, append-only review with deterministic evidence, stale locking, exact create transport, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-evaluator-panel.json through inventory:v1-metis-evaluator-panel and its verification script. This establishes the panel rules and nothing about any live model's judgement quality. Independence is derived from declared dependencies only — evaluators can still correlate through undeclared training data. Browser API transport is mocked; production deployment, representative reviewer usability, and formal accessibility certification are not established. -->
  • M5.25 Implement verifier-guided best-of-N generation without hiding rejected candidates or selection bias. Evidence: candidate, score, verifier version, selection rationale, and educator decision remain inspectable.
    • M5.25.a Define request/candidate-set/count/diversity/seed/budget, generator versions, verifier plan/rubric/versions, and stop conditions. Evidence: workflow schema.
    • M5.25.b Generate bounded unique candidate identities with complete prompt/source/model/config/cost lineage and preserve malformed/refused/ failed candidates. Evidence: generation tests.
    • M5.25.c Run applicable deterministic and calibrated verifiers on every eligible candidate with raw results, unavailable/error/abstention, and no selective rerun. Evidence: evaluation tests.
    • M5.25.d Rank/filter only by the declared versioned rule, expose scores/ uncertainty/rejections/tie/selection bias, and prohibit hidden candidate removal. Evidence: selection fixtures.
    • M5.25.e Present all authorized candidates/evidence/cost and require an educator select/reject/edit/regenerate decision. Evidence: review tests.
    • M5.25.f Create the accepted artifact as a new revision with candidate/ verifier/human decision lineage and normal gates. Evidence: end-to-end trace. <!-- M5.25 technically complete; operational readiness remains open.
      • The closed 5.25.0 contract is deterministic over recorded candidates. The engine never generates a candidate and never runs a verifier, so every result here is a property of the retention, completeness, ranking, bias-reporting, and decision rules rather than of any generator or verifier.
      • The workflow declares candidate count, minimum unique outputs, the diversity rule, a deterministic seed, a budget, a deadline, an output ceiling, the versioned verifier plan with per-verifier weights totalling 10,000 basis points, and its stop conditions. A plan whose weights do not total, that requires nothing, or a run that overspends its budget is refused at intake.
      • Every candidate carries its generator, provider, model and version, prompt revision and hash, config hash, source revisions, token counts, cost, and latency. malformed, refused, and generation_failed candidates must record why they were unusable and are retained: they are scored (as rejected), counted, and displayed. Duplicate outputs are named, not dropped. The service refuses to persist a run whose scored record does not contain every generated candidate.
      • Every eligible candidate must carry exactly one result for every planned verifier. A missing result and a second attempt for the same candidate/verifier pair are both findings, and either one moves the selection to plan_incomplete rather than quietly selecting from a partially evaluated set. unavailable, error, and abstained outcomes structurally carry no score and contribute nothing.
      • Ranking is the declared plan weights over the recorded scores, with the weighted uncertainty interval carried alongside. A failed required verifier or a total below the declared floor rejects the candidate with a named code, and the candidate stays in the record and out of the shortlist.
      • Selection bias is measured against the score, not the rank. Equal scores are ordered by generation position only to make the ranking total, so a rank-versus-position correlation would report that tiebreak back as a bias. Kendall tau-b relates the score to generation position and to output length, and the shortlist's generator concentration is reported beside them. A finding needs both the declared tau threshold and at least four eligible candidates, because a correlation over two points is arithmetic rather than evidence; the statistic is published either way.
      • The engine shortlists and never accepts. A tie is handed to an educator, automatic_acceptance is structurally false, and hidden_candidate_removals is structurally zero. The bench shows every candidate — its output or its failure record, its per-verifier contributions, its lineage and its cost — before the decision control.
      • A revision exists only after an educator acceptance. It binds the source candidate, the verifier results that examined it, the ranking rule, and the decision into one lineage hash, carries those verifier results forward as its gates, and records an educator who accepted a rule-rejected candidate as an explicit override with the codes it overrode. created_without_educator_decision is structurally false. Rejection and regeneration create nothing.
      • Tenant-scoped append-only storage binds each run to its request hash, rejects divergent idempotency reuse and tampering, refuses a decision naming an ungenerated candidate or quoting a stale run hash, and allows exactly one decision per run.
      • The executable corpus contains 31 labeled cases across six dimensions, all matched, reported with Wilson 95% intervals.
      • Verification: 44 focused engine, append-only service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin lint/typecheck and 33 files / 46 unit tests; twelve deterministic evidence-generator tests; 25 shared admin layout and WCAG executions; and twelve Chromium/Firefox/WebKit Playwright executions covering an unresolved tie, a retained unusable candidate, declared weight arithmetic, revision-only-on-acceptance, append-only decisions, exact create transport, 390px overflow, and automatic serious/critical Axe analysis.
      • Fixed while building this cell: the three verification benches faded in with an opacity keyframe, so an accessibility audit taken mid-animation graded every colour under the bench against a blended background. The entrance is now transform-only. code-verifications still animates opacity and carries the same latent flake.
      • Evidence is reproducible at evidence/v1-workbenches/metis-best-of-n.json through inventory:v1-metis-best-of-n and its verification script. The bias report covers two nuisance variables at a declared threshold and minimum sample; a value below either is not evidence of an unbiased selection. Browser API transport is mocked; production deployment, representative educator usability, and formal accessibility certification are not established. -->

M6 — Adaptive sequencing, tutor linkage, and learning analytics#

  • M6.1 Build pathway/program authoring with entry criteria, objective DAG, cadence, pacing, and completion rules. Evidence: cycle and reachability tests.
    • M6.1.a Define versioned program/pathway, cohort/audience, course/module/ objective membership, owner, locale, lifecycle, and credential links. Evidence: pathway schema.
    • M6.1.b Author typed entry prerequisites, prior evidence/credentials/ assessment/mastery conditions, alternatives, expiry, and review. Evidence: entry-criteria tests.
    • M6.1.c Author objective/course dependency DAG with required/optional/ choice/remedial/parallel edges and cycle/orphan/reachability checks. Evidence: graph tests.
    • M6.1.d Author calendar/cadence/start windows/deadlines/pacing/minimum/ maximum/seat-time/break/load rules with timezone and accommodation. Evidence: schedule tests.
    • M6.1.e Author deterministic completion rules over exact objectives/ evidence/assessments/grades/mastery/approvals without credential issuance authority. Evidence: completion tests. <!-- M6.1 technically complete; operational readiness remains open.
      • The closed 6.1.0 contract states a versioned program/pathway revision with its owner, lifecycle, cohort or role audience, locale, and its course/module/objective members, each bound to an exact member revision hash and its declared seat time. Duplicate node IDs, a member included twice, a pathway that requires nothing, and an audience whose scope does not match its kind are each refused at intake.
      • Entry criteria are six typed condition kinds — prior credential, prior pathway, assessment score, objective mastery, evidence artifact, and manual approval — grouped into all-of and any-of requirements, where any_of is how a pathway declares alternatives. Evidence past its expires_after_days is expired, not unsatisfied: the learner met the threshold and only the clock stopped them, and the decision still reports the observed and required values. A pending approval is review_required, a missing subject is evidence_missing, and a blocked alternative still names the routes that remain open.
      • The dependency graph carries required, optional, choice, remedial, and parallel edges, and only four of the five constrain order: an optional edge is a suggestion, so the same edge that closes a cycle as required leaves the order intact as optional. Kahn's algorithm produces the topological order; Tarjan's strongly connected components localize every cycle, and a cyclic graph structurally has no order at all. Self-edges, edges naming a node outside the pathway, orphans, unreachable nodes, a choice group of one, and a remedial edge feeding a required node are each named with the nodes involved.
      • Cadence resolves the declared IANA zone through zoneinfo and reports when it cannot, so a planned day is a real local day rather than an assumed one. The plan lays the topological order out over calendar days from seat time, applies extended-time, reduced-load, deferred-start and alternate-deadline accommodations, and then checks the enrollment window, the deadline, the weekly load ceiling net of required break days, and the pacing bounds. An alternate-deadline accommodation rescues an overrunning plan without changing the plan itself.
      • Completion is exact criteria over the recorded learner evidence under an all-of or at-least-n rule. Entry failure precedes completion: a learner who met every completion criterion but never qualified to enrol is entry_not_satisfied, not complete.
      • The engine has no credential issuance authority. A pathway's credential links carry issued_by_this_pathway = False, the completion rule carries issues_credentials = False, and the decision carries credentials_issued = 0, credential_issuance_authority = "none", and issued_here = False on every eligibility row. An incomplete pathway makes nobody eligible, and the service refuses to persist any analysis that claims otherwise. Issuance stays with the credential service.
      • Tenant-scoped append-only storage binds each authoring record to its request hash, rejects divergent idempotency reuse and tampering, and marks a superseded revision stale. Publishability names every blocking graph finding, schedule finding, unreachable required node, and non-publishable lifecycle.
      • The executable corpus contains 43 labeled cases across five dimensions, all matched, reported with Wilson 95% intervals.
      • Verification: 58 focused engine, service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin lint/typecheck and 34 files / 47 unit tests; eleven deterministic evidence-generator tests; 19 shared admin layout and WCAG executions; and nine Chromium/Firefox/WebKit Playwright executions covering a blocked cyclic graph with no order, a publishable revision's plan and entry evidence, credential eligibility that issues nothing, exact create transport, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-learning-pathway.json through inventory:v1-metis-learning-pathway and its verification script. Entry and completion read a learner evidence snapshot supplied with the request; nothing establishes that the snapshot faithfully reads a production gradebook. Pacing is planning arithmetic at a fixed sixty-minutes-per-day rate, not a prediction of how long a learner takes. Browser API transport is mocked; production deployment, representative author usability, and formal accessibility certification are not established. -->
  • M6.2 Build adaptive sequencing overlays that can tighten but not loosen platform safeguards. Evidence: property tests.
    • M6.2.a Define overlay scope/version/applicability, eligible content/items, sequencing objectives, mastery/pacing/difficulty rules, constraints, owner, and lifecycle. Evidence: overlay schema.

    • M6.2.b Merge platform/region/institution/program/course/learner-support constraints only by narrower eligibility, stricter limits, or additional requirements. Evidence: merge property tests.

    • M6.2.c Prohibit overlays from weakening safety, privacy, accessibility, integrity, exposure, rights, prerequisite, assessment, or human-authority floors. Evidence: monotonicity tests.

    • M6.2.d Preview affected paths/items/learners using privacy-safe aggregates and require review/versioning before activation. Evidence: impact tests.

    • M6.2.e Revalidate effective overlay at each decision and preserve exact version/rationale in sequence history. Evidence: runtime tests. <!-- M6.2 technically complete; operational readiness remains open.

      • CORRECTION (2026-09-05, while building M6.10's privacy cell): the impact preview leaked a suppressed bucket. Subjects sit in exactly one bucket, so the buckets partition the population once and changed_decision_count is their margin; publishing it beside all-but-one bucket handed the withheld one back by subtraction. Demonstrated on this cell's own corpus fixture, where the published margin of 15 minus the published affected count of 12 yielded exactly item:beta's three affected subjects — and on a constructed 20/20/5 case where 45 - 40 recovered the five-subject bucket entirely. The residual_disclosure_note claimed a reader could only bound such a value between one and nine; they could compute it. The margin is now withheld unless nothing is suppressed, or at least two buckets are and their combined affected count clears the floor, so subtraction yields a group rather than a cell. tests/test_sequencing_overlay_privacy.py pins it (7 tests), the note now describes the rule it applies, the workbench distinguishes "population below the minimum cell size" from "publishing it would recover a withheld bucket", and the corpus case that asserted the leak is relabelled case:impact-total-withheld-when-subtractable.
      • The closed 6.2.0 contract states an overlay revision with its owner, semantic version, revision hash, lifecycle, and an applicability that binds exactly one scope key — six scopes ordered platform, region, institution, program, course, learner support, where only platform names no key. Eighteen constraint dimensions across twelve families carry one of five kinds, and each numeric dimension declares the range it is written in, so a ceiling and a floor never share one anonymous integer domain. A duplicated dimension, an unsorted constraint set, a platform floor outside platform scope, a second platform floor, a review naming another overlay, and an overlay reviewing its own baseline are each refused at intake.
      • The merge is the meet of each kind: minimum for a ceiling, maximum for a floor, intersection for an allow-list, union for a required set, and logical or for a required flag. permits is written from the meaning of each kind and never consults meet, so the monotonicity checks are not grading the merge with the merge. The effective value is therefore no more permissive than any overlay that declared it, whatever an author writes, and it does not depend on the order the overlays arrive in.
      • An attempt to widen is named rather than silently clamped. A declaration at a scope is compared against the meet of every broader scope; when it permits more, the effective value still holds and the attempt is reported with the scope, the dimension, the value declared, and the value that stands. A silent clamp would be indistinguishable from an author who meant what they wrote.
      • Only four of the five kinds can express an intent to widen. M6.2.b allows additional requirements, so a required_set declaration is additive by nature: a course that writes {support.reader} has said nothing about the captions the platform already demands, and the union keeps both. The other four are complete statements — an allow-list that omits a member forbids it — so naming an excluded member is a widening.
      • A set meet can land on a value no single overlay wrote, and such a value has no author. Those constraints carry combination = "combined_declarations", no binding overlay or version, every declarer, and a composed rationale. Naming one author would misattribute the constraint.
      • Each of the nine protected families of M6.2.c must carry a platform floor. An unstated floor is not a floor: a missing family leaves the decision indeterminate_missing_protected_floor and evaluates no candidate, rather than treating the family as unconstrained. Weakening a stated protected floor is its own floor severity, distinct from loosening a non-protected broader scope.
      • The impact preview computes which decisions activation would change by admitting every supplied subject under the baseline and again under the proposal. A bucket is published only when its population, its affected count, and its unaffected count are each zero or at least the minimum cell size, so a suppressed cell is not a zero cell and a genuine zero is still published as zero. The published type has no field that can carry a subject reference, and the store refuses to persist a preview that publishes a small cell. The residual inference a reader who already knows a suppressed population can still make is stated rather than denied.
      • Activation requires an approved review pinned to this exact revision hash and version, by someone other than the owner, over a clean merge with every protected floor stated and an impact preview present. A review of a different version is not a review of this one.
      • Every decision recomputes the effective overlay from the overlay set; a caller's asserted hash is only compared and never adopted. A stale assertion refuses to decide at all, which is recorded as distinct from refusing every candidate, and a missing protected floor outranks a stale assertion. The sequence history keeps every contributing overlay's version and revision hash, and every applied constraint's binding version and rationale.
      • The executable corpus contains 59 labeled cases across five dimensions, all matched, reported with Wilson 95% intervals. The property sweep executes six laws over a deterministic, byte-stable generated space: idempotence (26), commutativity (164), associativity (1,160), meet-is-a-lower-bound (164), decision monotonicity (1,864) and protected-floor preservation (21,912), with zero violations. The decision law counts only the cases where the merged set actually admitted, so the implication is never vacuous, and five negative controls break the merge on purpose to prove each law can fail — including one recording that a left projection is idempotent and associative, so only commutativity catches it.
      • Verification: 63 focused engine, property, service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin typecheck and 35 files / 48 unit tests; twelve deterministic evidence-generator tests; 19 shared admin layout and WCAG executions; and 21 Chromium/Firefox/WebKit Playwright executions covering the merge, the combined value with no author, the nine floors, cell suppression, the activation gate, a weakened floor that still loses, refusal on overlay drift, the published property laws, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-sequencing-overlay.json through inventory:v1-metis-sequencing-overlay and its verification script. The engine constrains a sequencing decision and does not rank pedagogy; selection among admitted candidates is the published deterministic rule. Impact aggregates are computed over the subject population supplied with the request, and nothing establishes that population is a faithful sample of a live cohort. The overlay-gated decision path is this engine's own endpoint; metis.services.adaptive_learning_helpers does not route through it, so no claim is made that production sequencing is overlay-gated today. Browser API transport is mocked; production deployment, representative author usability, and formal accessibility certification are not established. -->
  • M6.3 Build item-pool eligibility rules by objective, mastery band, exposure, accommodation, locale, and integrity mode. Evidence: known-value selector tests.
    • M6.3.a Define typed rule inputs/outputs and exact item revision prerequisites for objective/coverage, difficulty/mastery band, exposure/ compromise, locale, PNP support, delivery, and integrity state. Evidence: eligibility schema.
    • M6.3.b Filter stale/unapproved/unavailable/incompatible/quarantined/ overexposed items and unresolved mandatory gates with exact reasons. Evidence: exclusion tests.
    • M6.3.c Apply current minimized learner-context projection without exposing diagnosis/grade/integrity detail to the item bank. Evidence: context boundary tests.
    • M6.3.d Select deterministically from eligible items by declared strategy/ seed with tie-break and no hidden constraint relaxation. Evidence: selector goldens.
    • M6.3.e Persist candidate/rejection/selection/rule/version rationale and test every dimension/boundary combination. Evidence: known-value suite. <!-- M6.3 technically complete; operational readiness remains open.
      • The closed 6.3.0 contract states an eligibility rule with its owner, version, and hash, the revision prerequisites an item must present, and one typed sub-rule per dimension: objective coverage, the difficulty window each mastery band may be served from, exposure and compromise, locale with declared fallbacks, personal-needs supports, delivery, and integrity state. Every candidate is judged on all seven and keeps every reason, so a report never shows only the first thing that went wrong.
      • A verdict about another revision is not a verdict about this one. Each piece of candidate evidence carries the exact item revision it was computed against, and evidence computed against an older revision is stale however the source graded it — the corpus case is an item whose estimate says passed and is refused anyway.
      • Twenty exclusion reasons are declared and each is produced in isolation by its own candidate. Stale, unapproved, unavailable, quarantined, compromised, and overexposed are distinct outcomes rather than one bucket, and an unresolved mandatory gate is neither passed nor failed: an unknown or stale gate excludes under its own reason, separate from a blocked one.
      • The item bank is handed a projection, and the projection type has no field a diagnosis, a grade, or an integrity incident could occupy. The property is checked as indistinguishability rather than inspection: two learner contexts that differ only in those three fields produce byte-identical projections, while a changed accommodation or mastery figure changes the hash. The store scans the serialized bytes as well, because a sensitive value can travel through a field that is allowed to exist. Supports come from accommodations, the band from the mastery figure against declared thresholds, and the integrity mode from the session — never from the incidents.
      • An unobserved mastery is treated as novice and says it was not observed, so a fail-closed default is never mistaken for a measurement. The band cut points (4,000 / 7,000 / 9,000 basis points) are pinned in the engine witness and exercised on both sides.
      • There is no code path that relaxes a constraint to find something to deliver. relax_on_empty and constraints_relaxed are Literal values, the store refuses any analysis claiming otherwise, and an empty eligible set is reported as no_eligible_item with the tally of which dimension excluded how many — distinct again from empty_pool, where nothing was judged at all. Selection orders eligible candidates by the declared strategy and breaks every tie on candidate ID, and the record says whether the strategy or the tie-break decided it.
      • A learner's granted accommodation is an additional required support, merged with the rule's own — the same principle M6.2 applies to a required set.
      • Every dimension is probed on both sides of its exact boundary: one candidate on the permitted edge, one a single step past it. All seven separate. A dimension that did not would be a rule deciding nothing.
      • The M6.2 overlay engine is genuinely called: tighten_rule_with_effective_overlay maps an effective overlay's exposure ceiling, repeat-spacing floor and difficulty ceiling onto the rule by meet, so an overlay can tighten a pool and a looser one leaves it exactly where it was.
      • The executable corpus contains 95 labeled cases across five dimensions, all matched, reported with Wilson 95% intervals.
      • Verification: 64 focused engine, service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin typecheck, lint, and 36 files / 49 unit tests; eleven deterministic evidence-generator tests; 19 shared admin layout and WCAG executions; and 18 Chromium/Firefox/WebKit Playwright executions covering the projection with no learner detail in it, every exclusion reason, the selection and what decided it, an empty eligible set, the seven-way boundary probe, 390px overflow, and automatic serious/critical Axe analysis.
      • Correction to M6.2 and M5.22 while here: the .bench entrance animation on the sequencing-overlay, item-pool and code-verification pages faded opacity, so an Axe run sampling mid-animation graded every colour under it against a blended background and a 5.18:1 text colour measured 4.49:1. All three now animate transform only. The M6.2 browser matrix passed as shipped; it was passing a coin flip.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-item-pool-eligibility.json through inventory:v1-metis-item-pool-eligibility and its verification script. Eligibility is decided from the candidate evidence an item bank presents with the request, and nothing establishes that a bank reports its own approval, availability, exposure, or gate state faithfully. Selection is a reproducible order, not a claim about which item teaches best. Browser API transport is mocked; production deployment, representative author usability, and formal accessibility certification are not established. -->
  • M6.4 Surface BKT, FSRS-5, graph-tracing, and IRT rationale separately. Evidence: UI does not collapse distinct models into one opaque score.
    • M6.4.a Define separate model-result contracts with model/version/config/ input projection/time/item-objective refs, outputs, uncertainty, and applicability. Evidence: model schemas.
    • M6.4.b Present BKT knowledge-state/prior/learn/guess/slip/update evidence and assumptions. Evidence: BKT view tests.
    • M6.4.c Present FSRS-5 memory stability/difficulty/retrievability/review history/desired retention/scheduling evidence. Evidence: FSRS view tests.
    • M6.4.d Present graph-tracing matched/missing/extra/path/state/step evidence and concept dependencies. Evidence: graph view tests.
    • M6.4.e Present IRT ability/item parameters/information/standard error/ calibration/linking evidence. Evidence: IRT view tests.
    • M6.4.f Show model disagreements/N/A/stale/insufficient data and combined sequencing rule explicitly; prohibit a generic mastery score. Evidence: UI contract tests. <!-- M6.4 technically complete; operational readiness remains open.
      • The closed 6.4.0 contract gives each model family its own result type. Every result carries the same envelope — model id, version, config hash, the exact learner projection it read, the objective and item it answered about, and at least one stated assumption — and then nothing else in common. p_known exists only on the BKT result, retrievability only on FSRS, missing_concept_ids only on the graph, ability only on IRT. A measured check confirms the four families share no output field name, and the store refuses a panel that claims otherwise.
      • Abstention is a measured condition, not a label. insufficient_data requires observations_available < observations_required, and stale requires an evidence age past a stated limit. A model that abstains still publishes everything it computed; it simply carries no verdict and names which of not_applicable, insufficient_data, or stale stopped it.
      • BKT applies the Corbett–Anderson update: condition on the observation with guess and slip, then apply the learn rate, publishing the before, the posterior, and the after for every step. A model with guess + slip at or above one is refused at intake, because such a model reads a correct answer as evidence against knowing. The recurrence is recomputed independently in the test rather than read back.
      • FSRS-5 uses the published power forgetting curve with decay -0.5 and factor 19/81. The constants are checked by their own definition: retrievability at an elapsed time equal to stability is exactly 90%, and the interval for 90% retention is exactly the stability, for every stability tried. Stability and difficulty are supplied by the scheduler; this engine trains neither and says so.
      • Graph tracing computes matched, missing and extra as exact set differences over the expected and observed paths, and marks a dependency unresolved when a concept the learner reached depends on one they missed. A concept is matched or missing, never both.
      • IRT sums three-parameter-logistic Fisher information over the answered items, derives the standard error as its reciprocal square root, and reads the lower 95% bound rather than the ability point estimate — the baseline panel is a case where the ability clears the threshold and the bound does not. Zero information means no interval exists, not a very wide one, and a model with no interval is not ready.
      • The combined rule is a conjunction over the models that could answer, published as a literal. Every pair that answered differently is named with both models and both bases in their own units. With nothing applicable the panel is indeterminate_no_applicable_model rather than not_ready, and a model that abstains does not vote — the corpus case is an IRT abstention that leaves the other three agreeing on ready.
      • The prohibition on a generic mastery score is structural and measured. No field on the decision is a cross-model number, generic_score_fields and composite_scores_computed are Literal[0], and an equal-tally separation probe holds the ready-count fixed while varying which model said what: a surface that had collapsed four models into one score would answer identically, and all three probes show the panels still differ.
      • The executable corpus contains 63 labeled cases across six dimensions, all matched, reported with Wilson 95% intervals.
      • Verification: 36 focused engine, service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin typecheck, lint, and 37 files / 50 unit tests; twelve deterministic evidence-generator tests; 19 shared admin layout and WCAG executions; and 21 Chromium/Firefox/WebKit Playwright executions covering each model's own parameters and assumptions, every disagreement, the explicit statement that nothing combines the four, a stale model that abstains without losing its numbers, an all-abstaining panel, the separation probe, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-learner-model-rationale.json through inventory:v1-metis-learner-model-rationale and its verification script. Each model is computed from the parameters supplied with the request: this engine does not fit BKT parameters, train FSRS stability or difficulty, author a concept graph, or calibrate IRT items. The combined rule is a stated policy, not a claim that the four models are jointly calibrated or that their agreement implies correctness. The estate's metis.services.learner_mastery_helpers still collapses a single probability into one MASTERED-to-NOT_STARTED label with no model attribution — exactly what M6.4.f prohibits — and this surface neither reads it nor changed it. Browser API transport is mocked; production deployment, representative author usability, and formal accessibility certification are not established. -->
  • M6.5 Build author-facing simulation with synthetic learners only. Evidence: no real learner PII enters preview.
    • M6.5.a Define versioned synthetic learner persona/profile distributions for prerequisites/mastery/history/preferences/PNP/locale/pacing/ integrity states with no real source identity. Evidence: fixture schema.
    • M6.5.b Generate deterministic seeded individuals/cohorts and label every record/event/result as synthetic. Evidence: generation tests.
    • M6.5.c Run exact pathway/adaptive/tutor/item rules in isolated simulation with bounded steps/time/budget and no delivery/grade side effects. Evidence: simulation tests.
    • M6.5.d Present paths/decisions/model rationales/outcomes/failures/ distributions and allow scenario comparison. Evidence: UI tests.
    • M6.5.e Scan source queries, stores, logs, exports, and analytics to prove no real learner ID/PII/profile/attempt enters simulation. Evidence: isolation suite. <!-- M6.5 technically complete; operational readiness remains open.
      • The closed 6.5.0 contract states a versioned, hashed cohort specification whose personas are distributions: a range for prerequisite satisfaction, mastery, history length, pacing, and learning gain, plus locale, PNP support, integrity mode, and preference sets. Every persona carries derived_from_real_learner = False as a Literal, and persona weights are refused unless they sum to ten thousand basis points — a distribution that does not sum to one is not a distribution.
      • A generated learner reference is a hash of the seed, the specification, and the ordinal, so there is no input through which a caller could supply a real identity. The reference pattern is ^synthetic:[a-f0-9]{16}$ and the type carries no learner_id, email, name, or external_id field for one to occupy.
      • Cohort composition uses largest-remainder Hare quota, so the split is exact rather than drawn: 40/45/15 over twelve learners is 5/5/2, and the same seed rebuilds a byte-identical cohort while another seed does not.
      • Every emitted record is labelled synthetic — learner, cohort, step, outcome, run, delta, and comparison. The isolation scan caught the two comparison types when they were not, which is the point of scanning rather than asserting.
      • A step calls the same item-pool eligibility engine a real decision calls — one engine call per step, counted and bounded. The run is bounded three ways, and it stops at the first bound it reaches and names which: max_steps_per_learner, max_simulated_days, or max_engine_calls. A simulation that silently truncated would report a shorter path as a finished one. delivery_side_effects and gradebook_writes are Literal[0].
      • The response model is declared rather than discovered: a learner answers correctly with their persona's prerequisite-satisfaction probability, drawn from a generator seeded per learner so the same learner draws the same sequence in every scenario, and a correct answer closes the declared learning-gain fraction of the remaining gap. A step that delivered nothing cannot change mastery, and an incorrect answer changes nothing.
      • Both scenarios walk one identical cohort, so a difference between them is the scenario rather than the sample — the comparison carries the cohort hash and cohort_identical as a Literal. The two ratified scenarios differ only in their exposure ceiling and all twelve learners' runs change.
      • M6.5.e is two halves. The static half walks the transitive import closure of the modules a simulation runs and refuses any database, ORM, learner store, gradebook, progress, analytics, or export module; its negative control is metis.api.admin, which is not itself on the denylist and whose closure is, several imports deep. The dynamic half scans everything a run produced for an unlabelled record, an email, a version 1–5 UUID, a learner identifier, or a reference that is not seed-derived. The report cannot call itself isolated while holding a finding, and the store refuses to persist a simulation that is not.
      • Network capability is reported rather than refused, and named. The closure reaches socket, ssl and http.client through metis.services.source_intake_network, which the shared hashing helper imports for one pure function. Nothing on a simulation path calls it. M6.5.e asks about learner data, so this is surfaced as a capability finding instead of being quietly folded into a pass.
      • The executable corpus contains 32 labeled cases across five dimensions, all matched, reported with Wilson 95% intervals.
      • Verification: 46 focused engine, isolation, service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin typecheck, lint, and 38 files / 51 unit tests; twelve deterministic evidence-generator tests; 19 shared admin layout and WCAG executions; and 21 Chromium/Firefox/WebKit Playwright executions covering the seeded cohort, the bounded runs, a run that names the bound it reached, the identical-cohort comparison, the isolation report with its capability finding, a page-wide check that no string looks like a person, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-learner-simulation.json through inventory:v1-metis-learner-simulation and its verification script. The response model reproduces no real learner and predicts no real outcome. Persona distributions are authored, and nothing establishes that any of them resembles a population an institution actually teaches. The isolation claim is about learner data and is not a capability sandbox. A simulation exercises item-pool eligibility; it does not exercise delivery, rendering, or tutoring dialogue. Browser API transport is mocked; production deployment, representative author usability, and formal accessibility certification are not established. -->
  • M6.6 Enforce and measure the adaptive next-item latency budget. Evidence: runtime gate, not a documentation assertion.
    • M6.6.a Define latency SLI boundaries from response commit/context receipt through eligibility/model/selection to next-item usable response, with target percentiles and device/network tiers. Evidence: budget spec.
    • M6.6.b Instrument each server/client/cache/model/store stage with bounded trace/metric attributes and exact timeout/deadline propagation. Evidence: instrumentation tests.
    • M6.6.c Implement bounded candidate pools/model work, caching/fallback/ degraded paths, cancellation, and honest unavailable state. Evidence: runtime tests.
    • M6.6.d Load-test realistic concurrent learners/pools/models/failures/ cold caches and measure p50/p95/p99/end-to-end error. Evidence: profile.
    • M6.6.e Fail release/runtime health when the ratified budget or missing- measurement rule is breached; never assert compliance from design docs. Evidence: performance gate. <!-- M6.6 technically complete; operational readiness remains open.
      • The closed 6.6.0 contract states the SLI boundary as five ordered stages — context receipt, eligibility, model panel, selection, render handoff — and budgets each of four delivery tiers at p50, p95 and p99 with a minimum sample count. A tier whose higher percentile carries a tighter budget is refused, a stage whose timeout falls below its own budget is refused, and a budget whose stage budgets sum past the tightest end-to-end p50 is refused: stages that cannot fit inside the whole are not a budget.
      • Instrumentation is a propagated deadline, not a stopwatch. Each span records what it entered with and what it left with; the contract refuses a span that ends holding more deadline than it started with, one whose spend does not equal its duration, and a trace where the deadline one stage left is not the deadline the next received.
      • Trace attributes are a closed key vocabulary with bounded values. A URL, an email, a UUID and a 64-character string all fail the value pattern, so a metric label fed from a raw identifier is not representable and cardinality cannot run away.
      • Candidate and model work are bounded before a stage runs, and a truncation is accounted rather than silent. A stage cut off at its timeout spends the timeout, not however long the work took; with a declared fallback the decision is served_degraded and names the stage, and without one it is unavailable with stage_timed_out_without_a_fallback. An exhausted deadline cancels with deadline_exhausted_before_next_item, and an exhausted candidate budget is unavailable having timed nothing at all.
      • The cancellation path is reachable because one ratified tier's p99 (220 ms) is tighter than the sum of every stage timeout (245 ms), which is exactly the condition under which cancellation has anything to do. A budget where it could never fire would have left that branch dead.
      • Percentiles are nearest rank over the recorded microseconds, so a reported percentile is a sample that was actually observed rather than an interpolation between two that were. A percentile over no samples raises rather than returning zero, and a tier with no samples publishes no percentile at all. Durations are integers throughout: a float could carry a NaN, and a NaN passes a range check written as two comparisons.
      • A live measurement times the real eligibility, model-panel and selection engines on time.perf_counter_ns. No absolute wall-clock figure is asserted anywhere — the corpus checks the gate's logic, not the machine's speed — and the browser suite reads a fixed deterministic trace rather than whatever this machine happened to do.
      • The gate reads a measured profile and nothing else. asserted_from_documentation is a Literal[False], a gate with no breach and no gap cannot pass having evaluated nothing, and a tier below its minimum sample count fails before any budget is compared, because a measurement that was not taken is not a passing one. A breach names the tier, the percentile, the budget, the observation, and exactly how far past it went. The store additionally refuses a gate that read a different profile or judged a different budget than the request carried.
      • The executable corpus contains 36 labeled cases across five dimensions, all matched, reported with Wilson 95% intervals.
      • Verification: 43 focused engine, service, and closed OpenAPI tests; ruff and strict mypy; OpenAPI and generated-client drift checks; admin typecheck, lint, and 39 files / 52 unit tests; thirteen deterministic evidence-generator tests; 19 shared admin layout and WCAG executions; and 24 Chromium/Firefox/WebKit Playwright executions covering the ratified budget, the measured percentiles, a passing gate that asserts nothing, a breach that says how far over, a short measurement that fails first, an instrumented trace with its propagated deadline and bounded attributes, 390px overflow, and automatic serious/critical Axe analysis.
      • Evidence is reproducible at evidence/v1-workbenches/ metis-adaptive-latency.json through inventory:v1-metis-adaptive-latency and its verification script. A live measurement times this engine's in-process compute; it is not a measurement of a deployed service's network, database, cold start, or contention. The budget is a set of numbers a working group ratified and nothing establishes that they are the right ones for any learner or device. Delivery-tier names stand for a device and network pairing; this engine is told which tier a request belongs to rather than detecting it. Browser API transport is mocked; production deployment, representative author usability, and formal accessibility certification are not established. -->
  • M6.7 Link tutor personas, prompts, tools, sources, and course versions. Evidence: tutor cannot use unapproved or stale course material.
    • M6.7.a Bind each tutor-config revision to exact persona/pedagogy/hint, system prompt/template, allowlisted tools, source set, and course/item revisions. Evidence: binding schema.

    • M6.7.b Resolve only approved/current/published material and track source/ course update/retraction/rights blast radius. Evidence: resolver tests.

    • M6.7.c Validate prompt/tool/source/course compatibility, safety/integrity/ PNP/locale policies, and grants before activation/session. Evidence: configuration tests.

    • M6.7.d Persist effective versions in each session/response and reject stale/retired/direct substituted IDs. Evidence: stale-material tests.

    • M6.7.e Present authorized version/diff/status and migrate/deactivate explicitly without rewriting historical sessions. Evidence: lifecycle tests. <!-- M6.7 evidence (2026-09-05) Built the 6.7.0 tutor configuration binding contract (apps/metis/service/src/metis/schemas/tutor_configuration_binding.py), the resolution/compatibility/session/migration engine (services/tutor_configuration_binding_engine.py), a 45-case labelled corpus over 30 fixture families and five dimensions (data/m6_7_tutor_binding_corpus.json, evaluated by services/tutor_configuration_binding_evaluation.py), an append-only tenant-scoped store (services/tutor_configuration_binding_service.py), instructor + admin routes (api/tutor_configuration_binding.py, api/admin.py), and the admin workbench at /tutor-bindings.

       The evidence clause is structural, not a promise. A MaterialRef has no
       bare-identifier shape: kind, material_id, revision_id and a 64-hex
       revision_sha256 are all required and additionalProperties is false, so
       no client can express "use the current one". Resolution compares the
       bound revision against the catalogue and yields exactly one success
       (`resolved`) and seven refusals (`stale`, `not_approved`,
       `not_published`, `retracted`, `rights_revoked`, `substituted`,
       `unresolvable`), each of which blocks activation.
      
       A CATALOGUE HOLDING A DIFFERENT REVISION UNDER THE SAME IDENTIFIER IS
       A SUBSTITUTION, NOT AN UPDATE. The unknown-revision check runs before
       any status check, so a material the catalogue has "moved on" from is
       named rather than followed — verified over the full 60-row cross
       product of status x published x rights_cleared x revision in
       test_an_unknown_revision_is_substituted_whatever_else_the_record_says.
       A superseded revision the catalogue still holds is `stale` instead;
       that distinction is only possible because MaterialRecord carries
       known_revision_sha256s, not just the current one.
      
       RESOLUTION IS PERFORMED AGAIN AT EVERY SESSION START. Approving a
       configuration is not a licence that survives its material: the engine
       re-resolves before it starts a session, and refuses rather than
       substituting (`refuse_the_session_rather_than_substitute_a_revision`).
       A refused session records no effective material at all. The
       stale-material sweep runs one probe per blocking outcome and asserts
       all three properties per probe — blocks activation, refuses the
       session, leaves the historical session intact: 7/7 separate.
      
       A MIGRATION MINTS A REVISION AND REWRITES NOTHING. MigrationRecord
       carries historical_sessions_rewritten: Literal[0] and TutorSessionRecord
       carries rewritten_by_migration: Literal[False], so a rewrite is
       unconstructible rather than merely unimplemented; both consts appear in
       the published OpenAPI. `revision_current` on the workspace projection is
       computed over what the tenant holds, never a stored flag.
      
       AN ALLOWLISTED TOOL NOTHING REFERENCES IS ADVISORY, NOT BLOCKING — the
       only advisory code in the vocabulary, and CompatibilityFinding's
       validator makes severity a function of the code, so it cannot be set
       wrong (a mutation that forced blocking=True was refused by the contract
       outright). A prompt reaching for an ungranted tool or citing an unbound
       source blocks; so does a locale, support, or integrity mode the course
       does not offer, a draft lifecycle, a missing grant, a self-grant, and a
       grant pinned to another revision.
      
       Verification: 58 focused Python tests (28 engine, 12 property, 11
       service, 7 API contract); corpus 45/45 matched across all five
       dimensions with Wilson 95% intervals; ruff + mypy --strict clean on
       every touched module; OpenAPI and the generated TypeScript client
       regenerated (5 paths, 31 schemas added, none removed); 42 browser
       checks green across chromium/firefox/webkit with automatic serious +
       critical axe analysis and a 390px overflow audit; 40 admin vitest files
       (53 tests); 14 evidence-generator tests, every executable clause
       satisfied.
      
       Mutation-tested rather than merely asserted: making substitution follow
       the catalogue mismatched 2 corpus cases and broke the sweep; letting a
       blocked configuration start its session mismatched 4; treating a
       superseded revision as resolved mismatched 6.
      
       Two defects this cell surfaced and fixed in shared code: the admin
       scroll containers were not keyboard-reachable (an overflow-x region
       needs focus and a name), and the fixed sidebar used minHeight: 100vh
       with no scroll, so every link past the viewport was unreachable by
       mouse or keyboard — it is now bounded with a scrolling nav. A name
       collision was also avoided rather than accepted: M6.7's activation
       types are BindingActivationBlocker/BindingActivationDecision because
       reusing M6.2's names would have silently renamed both in the published
       API and the generated client.
      
       Limitations, stated rather than implied: resolution reads the catalogue
       supplied with the request, and nothing here establishes that a
       catalogue faithfully reports what any content system currently
       approves, publishes, or has retracted. Tool and source references are
       read from the prompt body with a pinned pattern, so a template that
       reaches a capability another way — a tool the model selects at runtime,
       say — is outside what this check can see. Compatibility compares
       declared locales, supports and integrity modes; it does not inspect the
       material itself. The estate's existing tutoring service separately
       refuses source-grounded answers when a linked source package is not
       approved; this cell neither replaces that check nor claims to have
       changed it, and NO PRODUCTION TUTOR RUNTIME IS BOUND TO THIS ENGINE.
       Browser API transport is mocked; production deployment, representative
       author usability, and formal accessibility certification are not
       established.
      

      -->

  • M6.8 Build tutor-response grounding and pedagogical policy evaluation. Evidence: generated guidance cites current approved sources.
    • M6.8.a Decompose tutor output into factual/procedural/hint/pedagogical/ motivational/safety claims and attach exact course/source citations. Evidence: response schema.

    • M6.8.b Verify support/sufficiency/faithfulness/contradiction/currency for claims and required current approved source use. Evidence: grounding fixtures.

    • M6.8.c Evaluate hint-ladder stage, scaffolding, direct-answer exposure, learner agency, cognitive demand, tone, accessibility, and assessment mode against exact policy. Evidence: pedagogy fixtures.

    • M6.8.d Preserve raw evaluator/rule/model outputs, uncertainty/N/A/error, policy verdict, and human review separately. Evidence: result tests.

    • M6.8.e Block/redact/regenerate/escalate according to policy and verify no ungrounded response is presented as supported. Evidence: end-to-end gate. <!-- M6.8 evidence (2026-09-05) Built the 6.8.0 tutor-response grounding contract (apps/metis/service/src/metis/schemas/tutor_response_grounding.py), the decomposition/grounding/pedagogy/enforcement engine (services/tutor_response_grounding_engine.py), an 83-case labelled corpus over 26 fixture families and five dimensions (data/m6_8_tutor_response_grounding_corpus.json, evaluated by services/tutor_response_grounding_evaluation.py), an append-only tenant-scoped store (services/tutor_response_grounding_service.py), instructor + admin routes (api/tutor_response_grounding.py, api/admin.py), and the admin workbench at /tutor-responses.

       THE DECOMPOSITION LOSES NOTHING, STRUCTURALLY. Segments are contiguous,
       carry their own trailing whitespace, start at zero and end at the
       response length, and TutorResponseDecomposition re-hashes the
       concatenation of their exact text against the response digest. A
       decomposition that dropped a sentence is unconstructible, not merely
       untested — the probe that removes the last claim is refused with "claim
       spans must reach the end of the response". Verified over eight shapes
       including leading whitespace, no terminator, blank-line runs and
       "One.Two.Three.".
      
       "CURRENT APPROVED SOURCE" HAS ONE DEFINITION IN THIS ESTATE, NOT TWO.
       Citations carry M6.7's MaterialRef — the shape with no bare-identifier
       form — and currency is decided by calling M6.7's own resolve_material
       and carrying its outcome verbatim. All eight M6.7 outcomes are exercised
       from M6.8, and CitationCurrency's validator makes current_approved a
       function of the outcome, so this cell cannot hold a second opinion about
       what approved means. A mutation pinning current_approved to True was refused
       by the contract before any corpus case ran.
      
       AN INSTRUMENT THAT COULD NOT LOOK DOES NOT RETURN A VERDICT. Three blind
       seams each produce an EvaluatorError and an abstention, never a pass: no
       catalogue to resolve against (an "unresolvable" outcome there is the
       absence of an instrument, not a finding), an unreadable excerpt, and an
       absent answer key. A fourth was found during the build: the cognitive-
       demand dimension reported "the turn asks the learner to do nothing" for
       a turn that plainly did ask, when the Anderson & Krathwohl verb table
       simply had not recognised the task. That is now three distinct states —
       the policy sets no floor (not applicable), the turn asks nothing (not
       applicable), and the floor could not be checked (abstain) — and the
       detail says which.
      
       NOTHING IS AVERAGED. Each of thirteen dimensions keeps its own state and
       TutorPolicyVerdict's validator recomputes the worst one and the
       dimension that produced it; `averaged` is a Literal[False]. A mutation
       switching the engine to a majority vote was refused by the contract
       outright.
      
       REDACTION IS DELETION, AND A CITATION MARKER IS A CLAIM THE GATE STOOD
       BEHIND. presented_text must equal the concatenation of exactly the spans
       that survived redaction, and presented_as_supported is permitted only
       where the support verdict is `supported`, the claim is grounded and
       every cited revision resolved. Both are validators on the response.
       Ten ungrounded-presentation probes attempt the forbidden constructions
       against the real contract — presenting an unsupported claim as
       supported, presenting a redacted claim, keeping a redacted span in the
       presented text, a block or an escalate that still carries text, a
       redaction that removes nothing, a release that also redacts, a verdict
       softened below its worst dimension, a decomposition that drops a
       sentence, and a stale citation presented as supported. 10/10 refused.
      
       Four defects this cell surfaced and fixed in its own code before
       shipping:
       - THE EVALUATION ID DID NOT BIND EVERYTHING THE VERDICT DEPENDS ON. It
         hashed the turn, the policy and the config revision but not the
         catalogue, so the same turn judged against two catalogues that
         disagree about whether its citation is still current collided on one
         identifier — precisely the pair of runs an operator most needs to tell
         apart. Surfaced by generating the browser fixture from the real
         service, which refused the second write. The id now binds the whole
         semantic request.
       - A COMMENT NAMED A RULE THE CODE DID NOT APPLY. The enforcement table
         said "nothing survives a contradiction", but the flag it read covered
         only conflicts *between two cited excerpts*; a turn asserting
         something its own cited source denies was being trimmed and the
         remainder shipped. Both now end the turn: trimming a falsehood and
         releasing the rest treats a fabrication as a layout problem.
       - `safety_reviewer` WAS UNREACHABLE VOCABULARY. Every abstaining safety
         claim was caught by the definite-failure branch first. A safety claim
         whose grounding is merely uncertain now routes to a safety reviewer
         instead of being blocked outright — blocking would leave a learner
         with nothing at the moment they most need something.
       - `TutorPolicyOutcome.not_applicable` WAS A VALUE NO INPUT COULD
         PRODUCE, because the ladder, scaffolding, tone and accessibility
         dimensions always evaluate. Removed; the branch now raises.
       The producibility audit that caught the last two is a test: every value
       of every M6.8 vocabulary — 6 claim kinds, 6 support verdicts, 5
       sufficiency criteria (each also shown failing), 4 faithfulness changes,
       2 contradiction kinds, 12 pedagogy finding codes, 3 evaluator errors, 5
       dimension states, 3 policy outcomes, 5 enforcement actions, 7 reason
       codes, 3 escalation targets and all 5 ladder stages — is produced by a
       named input.
      
       Mutation-tested rather than merely asserted, eight mutations, all
       caught: three by the contract refusing to represent the wrong state
       (currency ignoring the catalogue, a majority-vote verdict, a redaction
       that keeps the text), five by corpus mismatch (an abstention read as
       support, an unclassified segment defaulting to a learner prompt, an
       ignored polarity conflict, a missing answer key read as no exposure, a
       source-contradicted claim trimmed instead of blocked). A methodology
       note worth keeping: two mutations of identical byte length applied
       within the same second reused a stale .pyc and reported the same result;
       the run is now made with bytecode caching off, and the second mutation
       proved to be caught only after that.
      
       Verification: 82 focused Python tests (34 engine, 33 property, 8
       service, 7 API contract) and the whole 2,714-test Metis suite run
       once (2,655 passed, 53 skipped, 6 failed — the OpenAPI staleness gate,
       since regenerated and passing, plus five pre-existing M5.x
       item-difficulty/item-assembly/evidence-bundle fixture failures whose
       every imported module this branch leaves untouched, e.g.
       test_item_difficulty_helpers asserting sample_size 64 against a fixture
       producing 640); corpus 83/83 matched across decomposition,
       grounding, pedagogy, separation and enforcement with Wilson 95%
       intervals; ruff + ruff format + mypy --strict clean on every touched
       module; OpenAPI and the generated TypeScript client regenerated (7
       paths, 36 schemas added, none removed, contract check clean); 30 browser
       checks green across chromium/firefox/webkit with automatic serious +
       critical axe analysis, a 390px overflow audit and a keyboard-reachability
       sweep over every scrolling region; 41 admin vitest files (54 tests). The
       browser fixture is generated from the real service rather than
       hand-written, so the mocked payloads are the contract's own output.
      
       M6.8.d is structural in the storage layout, not only in the schema: a
       verdict document has no review field at all, and a review is written to
       a separate file in the same tenant directory. The service re-reads the
       verdict bytes after appending and refuses if they changed; the
       projection restates the engine's action and outcome beside the reviews
       and validates that neither moved.
      
       Limitations, stated rather than implied: segment classification is a
       declared lexical vocabulary hashed into rules_sha256, and support is
       normalized content-token containment with polarity and numeral checks —
       bounded conformance on a checked-in corpus, not open-domain natural-
       language inference, and the Wilson intervals describe this corpus only.
       Readability is Flesch-Kincaid over a vowel-group syllable estimate and
       cognitive demand is an Anderson & Krathwohl verb table; both are named
       approximations with the tradeoff written down, and a task phrased
       outside the table abstains rather than passing. Tone is judged only
       against phrases the supplied policy bans. Direct-answer exposure is
       normalized containment of the supplied answer key, so a disclosure
       phrased differently is outside what this check can see. Nothing here
       establishes that a catalogue faithfully reports what any content system
       currently approves, nor that an approved source is true. NO PRODUCTION
       TUTOR RUNTIME IS BOUND TO THIS ENGINE, and the legacy per-session
       scorer in services/tutor_evaluation_helpers.py — which does average its
       dimensions into one 0-100 number — is untouched and not replaced by this
       cell. Browser API transport is mocked; production deployment,
       representative author usability, and formal accessibility certification
       are not established.
      

      -->

  • M6.9 Build assessment-mode integrity interlocks for tutor and visualization tools. Evidence: active assessed/exam contexts block forbidden assistance.
    • M6.9.a Define assessment context assertion with exact assessment/item/ attempt, mode, allowed aids/tools/hints/content, time window, institution policy, actor/session, and signature/freshness. Evidence: context schema.

    • M6.9.b Map practice/homework/quiz/graded/exam/proctored/accommodation/ review modes to allowed tutor/visualization/execution/source/answer operations. Evidence: policy matrix.

    • M6.9.c Revalidate mode on every tool request/stream turn/deep link and fail closed on missing/stale/conflicting context. Evidence: runtime tests.

    • M6.9.d Prevent direct routes, alternate modality, cached session, prompt injection, tool alias, graph/media, and accessibility-channel bypass. Evidence: adversarial tests.

    • M6.9.e Provide safe refusal/permitted scaffolding/escalation and log privacy-minimized interlock events without auto-accusation. Evidence: journey tests. <!-- M6.9 evidence (2026-09-05) Built the 6.9.0 interlock contract (apps/metis/service/src/metis/schemas/assessment_interlock.py), the verification/matrix/decision engine (services/assessment_interlock_engine.py), a 67-case labelled corpus over 13 fixture families and five dimensions (data/m6_9_assessment_interlock_corpus.json, evaluated by services/assessment_interlock_evaluation.py), an append-only tenant-scoped store with mint and check (services/assessment_interlock_service.py), instructor + admin routes (api/assessment_interlock.py, api/admin.py), and the admin workbench at /assessment-interlocks.

       ONE PROPERTY ANSWERS ALL SEVEN NAMED BYPASSES. `disposition_for(mode,
       operation)` takes two arguments, and neither of them describes how the
       ask arrived. Modality, route, checkpoint and every free-text field the
       request carries are recorded for the operator and are not inputs to the
       decision, so a direct route, another modality, a resumed session, an
       injected instruction, a tool alias, a media probe and the accessibility
       channel all reach the same cell of the same matrix. The bypass sweep
       asks for one denied operation seven ways and reports whether the
       disposition or the outcome moved: 7/7 invariant. Three mutations that
       made the disposition depend on the channel, the route, or the
       checkpoint were each caught by that sweep and by nothing else.
      
       THE MATRIX IS TOTAL, AND SAYS SO FROM A WALK RATHER THAN A CLAIM.
       8 modes x 13 operations = 104 cells, every one decided, validated at
       import: a hole raises, and so does a mode whose row merely repeats
       another mode's, because a mode that decides nothing differently is a
       label rather than a mode. The report counts decided cells, lists
       undecided ones, names duplicate rows, and checks that all four declared
       dispositions are actually used — removing one cell and making proctored
       identical to exam were both refused at import. The workbench prints the
       engine's own 104 cells rather than reconstructing the grid from
       whichever cells the corpus samples.
      
       AN UNAUTHENTICATED MODE IS A CLAIM, NOT A FACT. When no assertion is
       presented, the signature does not match, or no signing key is
       configured, the decision reports `mode: null` — the contract makes a
       stated mode and a verified context equivalent, so an interlock cannot
       log a mode it never authenticated. A mutation that reported the
       assertion's own mode anyway was refused by the contract outright.
      
       THE TWO HALVES FAIL IN OPPOSITE DIRECTIONS, DELIBERATELY. Minting an
       assertion without a key RAISES (503) — you cannot sign without a key,
       and issuing an unsigned context would hand verification something it
       then has to trust. Verifying without a key DENIES, records
       `signature_checked: false`, and withholds the mode — an interlock that
       cannot check its context must not permit the operation it guards.
      
       AN INTERLOCK EVENT CANNOT CARRY LEARNER CONTENT OR AN ACCUSATION. There
       is no field on InterlockEvent able to hold a message, an answer, a note
       or an actor id — every field is an identifier, a closed-vocabulary
       value, or a digest — and `learner_accused` is a `Literal[False]`. The
       actor appears only as a pseudonym, and `pseudonym_keyed` says whether
       the signing key was available to key it, because an unkeyed digest is
       confirmable by anyone who can guess the actor id and should say so
       rather than implying otherwise. Refusal wording is drawn from a closed
       set, so a refusal can neither leak the answer it withholds nor be
       phrased as an allegation; a test asserts the wording contains none of
       cheat/misconduct/violation/suspicious.
      
       Two defects and two dead vocabulary values fixed before shipping:
       - `interlock.context.payload-mismatch` WAS UNREACHABLE. The assertion's
         validator recomputes `payload_sha256` from its own body, so a mismatch
         is unconstructible and the failure code could never fire. Removed.
       - `interlock.refusal.operation-not-recognised` REPLACED A FABRICATED
         CLASSIFICATION. An operation the alias table does not know now records
         `canonical_operation: null` and denies, rather than being guessed into
         the strictest class — a fabricated classification in an audit trail is
         worse than an honest gap. A mutation that guessed `answer_check` was
         caught by the corpus.
       - `route` WAS TYPED AS AN IDENTIFIER, so no real route could be sent;
         it is now a bounded path type that refuses a traversal sequence.
       - The engine's `matrix_rows()` helper had NO CALLER once the report
         carried the rows itself; removed rather than left as decoration.
      
       Escalation is a routing decision, not a softer refusal: a denial under
       `proctored` escalates to a human because that is what proctoring means,
       and an ungranted aid under `accommodation` escalates because whether an
       accommodation covers something is a person's call rather than a
       matrix's. Uncertainty never becomes a silent no.
      
       Verification: 113 focused Python tests (75 engine, 22 property, 12
       service, 7 API contract), including a sweep of all 104 cells at all four
       checkpoints (416 decisions, none falling through); corpus 67/67 matched
       across context, matrix, revalidation, bypass and journey with Wilson 95%
       intervals; 11 mutations, all caught (2 by the import-time matrix
       validator, 2 by the contract, 7 by corpus or bypass mismatch); a
       producibility audit asserting every context-failure code, outcome,
       refusal reason, escalation reason, disposition, mode and operation class
       is produced by a named input; ruff + ruff format + mypy --strict clean;
       OpenAPI and the generated TypeScript client regenerated (6 paths, 22
       schemas added, none removed, contract check clean); 36 browser
       checks green across chromium/firefox/webkit with automatic serious +
       critical axe analysis, a 390px overflow audit and a keyboard-
       reachability sweep. The browser fixture is generated from the real
       service rather than hand-written.
      
       Limitations, stated rather than implied: the interlock governs operation
       CLASSES. It does not inspect payload content, so a learner who extracts
       an answer from content that an allowed operation legitimately returned
       is outside what this check can see — the graph/media and accessibility
       bypasses are answered at the operation layer, not the content layer.
       The alias table covers the names this estate uses; a caller inventing a
       new name is denied as unrecognised, which is safe but is not the same as
       understanding it. Nothing here establishes that an institution's policy
       revision is authentic, only that the assertion and the request agree
       about which one they are under. NO PRODUCTION TUTOR OR VISUALIZATION
       RUNTIME IS BOUND TO THIS INTERLOCK, and M6.8's separate assessment-mode
       pedagogy dimension is untouched and not replaced by it. Browser API
       transport is mocked; production deployment, representative operator
       usability, and formal accessibility certification are not established.
      

      -->

  • M6.10 Build author analytics for completion, misconception, item performance, objective mastery, fairness, and content drop-off. Evidence: aggregation privacy and minimum cohort rules.
    • M6.10.a Define each metric’s population/denominator/window/source/ revision/unit/missingness/owner and valid descriptive use. Evidence: metric catalog.

    • M6.10.b Compute completion/progression/drop-off by course/module/lesson/ activity/content step and item performance/option/rubric/time/exposure. Evidence: known-value tests.

    • M6.10.c Compute objective mastery/model-specific summaries and aggregate misconception evidence without individual diagnosis. Evidence: learning analytics tests.

    • M6.10.d Compute approved fairness/DIF/accessibility/outcome slices with minimum cohorts, uncertainty, privacy suppression, and no hidden trait inference. Evidence: privacy tests.

    • M6.10.e Present revision-aware trends/comparisons/filters and explicit unknown/suppressed/stale/causal-limit state. Evidence: analytics UI tests. <!-- M6.10 evidence (2026-09-05) Built the 6.10.0 author-analytics contract (apps/metis/service/src/metis/schemas/author_analytics.py), the catalog/aggregation/disclosure engine (services/author_analytics_engine.py), a 38-case labelled corpus over 13 fixture families and five dimensions (data/m6_10_author_analytics_corpus.json, evaluated by services/author_analytics_evaluation.py), an append-only tenant-scoped store (services/author_analytics_service.py), instructor + admin routes (api/author_analytics.py, api/admin.py), and the admin workbench at /author-analytics.

       A METRIC WITH NO STATED DENOMINATOR IS NOT A METRIC. Every catalog entry
       declares its population, window, source revision, missingness policy,
       owner, valid descriptive use and known limitations; a rate or a
       proportion must also state what it divides by, and a unit that is not a
       quotient is refused if it supplies one. `missingness_policy` has NO
       default, because treating a gap as a zero is a decision with
       consequences and has to be written down. The catalog also refuses causal
       prose: a metric whose stated use contains "improves", "causes", "leads
       to" or thirteen other markers cannot be constructed, because author
       analytics counts what happened and does not establish why.
      
       A VALUE EXISTS EXACTLY WHEN IT IS REPORTED. `MeasuredValue.value` is
       `None` unless `state` is `reported`, so a suppressed cell cannot render
       as zero and an unknown cell cannot either — the two absences this estate
       most often collapses into a misleading nought. Five states are
       distinguished (reported, unknown, suppressed, stale, not_applicable) and
       the workbench prints the state rather than a number for four of them.
      
       A SUPPRESSED CELL IS NOT RECOVERABLE FROM A PUBLISHED MARGIN. This is
       the M6.2 defect this session found and fixed, enforced here one level
       higher: `CohortTable` REFUSES a reported margin beside a suppressed
       cell, so the arrangement is unconstructible rather than documented as a
       residual risk. Raising the cohort floor from 10 to 20 in the workbench
       withholds a cohort that was published a moment earlier, and withholds
       the total with it.
      
       A SLICE IS ONLY EVER OVER AN ATTRIBUTE SOMEBODY RATIFIED. Cohort
       dimensions are M5.12's vocabulary, reached through four named optional
       fields on `ObservationCohorts` — no map, no free key, no derived value —
       so an attribute nobody approved has nowhere to arrive and the engine has
       no inference path at all. A probe supplying `inferred_ability` is
       refused by the shape itself.
      
       Known values, worked by hand and then checked: completion 24/30 = 0.8;
       drop-off 6/30 = 0.2, counted independently and summing to 1.0 rather
       than derived from completion; item facility 13/30 = 0.433333 with a
       Wilson interval matching the published formula to six places; mastery
       10/20 for model:bkt beside 15/20 for model:irt, reported separately
       because two models disagreeing is a finding rather than an average;
       misconception prevalence 3/30, on a shape with no learner field at all
       and `individual_diagnosis: Literal[False]`.
      
       Two defects found during the build:
       - THREE MUTATIONS SURVIVED THE FIRST CORPUS. An empty population
         reported as zero, mastery pooled across models, and a comparison
         across content revisions claiming to be like for like all passed 33/33
         because the corpus had no fixture that could see them. Closing the
         gaps turned the comparison into something that answers the question it
         names: it is emitted only when the observations actually span two
         revisions, and `like_for_like` is a function of whether they do.
       - A SCHEMA NAME COLLISION RENAMED M6.3's PUBLISHED SCHEMA.
         `ObjectiveMastery` already existed for one learner's mastery of one
         objective; defining a cohort aggregate by the same name made FastAPI
         emit `ObjectiveMastery-Input` and `ObjectiveMastery-Output` and delete
         the plain name, breaking the generated client for M6.3. Caught by
         diffing the published schema set before and after, which showed
         +28/-1. Renamed to `ObjectiveMasteryAggregate`; a new test walks all
         2,868 model classes under `metis.schemas` and asserts no two modules
         define the same name. The estate holds that invariant everywhere else,
         so it is now pinned.
      
       Verification: 44 focused Python tests (29 engine, 7 service, 8 API
       contract); corpus 38/38 across catalog, known_value, learning_analytics,
       privacy and state with Wilson 95% intervals; 8 mutations, all caught (3
       by the contract refusing to represent the wrong state, 5 by corpus
       mismatch); 5 disclosure probes, all refused; ruff + ruff format + mypy
       --strict clean; OpenAPI and the generated TypeScript client regenerated
       (4 paths, 27 schemas added, none removed, contract check clean); 33
       browser checks green across chromium/firefox/webkit with automatic
       serious + critical axe analysis, a 390px overflow audit and a
       keyboard-reachability sweep. The browser fixture is generated from the
       real service rather than hand-written.
      
       Limitations, stated rather than implied: the disclosure rule is minimum
       cohort plus margin withholding. It defends against recovering a withheld
       cell by subtracting the published ones from the total; it does NOT
       defend against a reader who already knows a cohort's population and can
       therefore bound its value, and it is not differential privacy. Fairness
       slices are computed over supplied, ratified attributes only — nothing
       here establishes that those attributes were collected with consent, only
       that no other kind can be sliced on. DIF proper remains M5.12's cell;
       this one reports completion by cohort and does not estimate item bias,
       and the outcome slices named in M6.10.d are covered only to the extent
       that completion is one. Rubric-level item analytics named in M6.10.b are
       NOT built: the observation shape carries option selection, timing and
       exposure, and no rubric. NO PRODUCTION LEARNING-RECORD STORE IS BOUND TO
       THIS ENGINE; observations arrive in the request. Browser API transport is
       mocked; production deployment, representative author usability, and
       formal accessibility certification are not established.
      

      -->

  • M6.11 Build feedback-to-revision queues. Evidence: learner data is aggregated/redacted and does not auto-edit content.
    • M6.11.a Define feedback signal/source/aggregation/version/severity/ evidence/privacy, queue item, target content revision, owner, status, and decision contract. Evidence: queue schema.

    • M6.11.b Ingest approved surveys/comments/support/corrections/analytics/ performance/accessibility/integrity-safe signals with PII redaction and cohort thresholds. Evidence: intake tests.

    • M6.11.c Deduplicate/cluster/priority-rank with attributed rule/model versions, uncertainty, counterexamples, and no unreviewed learner quote exposure. Evidence: triage tests.

    • M6.11.d Let authors inspect/accept/reject/defer/request research and create a new draft revision linked to signals; never auto-edit published content. Evidence: authority tests.

    • M6.11.e Track revision/gate/review/publication outcome and close the loop with privacy-safe evaluation. Evidence: lineage tests. <!-- M6.11 evidence (2026-09-05) Built the 6.11.0 feedback-queue contract (apps/metis/service/src/metis/schemas/feedback_queue.py), the intake/triage/decision engine (services/feedback_queue_engine.py), a 30-case labelled corpus over 13 fixture families and five dimensions (data/m6_11_feedback_queue_corpus.json, evaluated by services/feedback_queue_evaluation.py), an append-only tenant-scoped store with a decision flow (services/feedback_queue_service.py), instructor + admin routes (api/feedback_queue.py, api/admin.py), and the admin workbench at /feedback-queues.

       A SIGNAL CAN PRODUCE A PROPOSAL, AND THERE IS NO FIELD THROUGH WHICH IT
       COULD PRODUCE AN EDIT. DraftRevisionProposal names the revision it is
       BASED ON and a new draft it mints; it has no field naming a published
       revision to overwrite, and `published_content_modified` is a
       `Literal[False]`. Only an acceptance mints a proposal — a validator ties
       the two together, so attaching a draft to a rejection is
       unconstructible. Six auto-edit probes attempt every route to an edit
       (flipping the literal, adding an overwrite field, giving a rejection a
       proposal, publishing without a passing gate, and two quote routes);
       6/6 refused.
      
       THE QUEUE HAS NO FIELD ABLE TO HOLD A LEARNER'S WORDS. QueueItem carries
       signal identifiers, counts, severity and uncertainty, and nothing a
       learner wrote. A quote reaches it only as a ReleasedQuote, which cannot
       be constructed without the reviewer who released it, when, and the
       redactions applied — so "no unreviewed learner quote exposure" is a
       missing field rather than a rule somebody has to remember. Raw text
       lives in RawFeedbackSignal at intake and stops there: AcceptedSignal,
       the only thing intake emits, has no verbatim field and says
       `verbatim_retained: Literal[False]`.
      
       THE PRIVACY CLAIM IS SETTLED BY LOOKING. The published queue payload is
       walked with M6.5's identity detectors — one definition of what a learner
       identity looks like, reused rather than copied — and the response
       carries the finding count it actually got. The service scans again on
       the way to disk, so a record carrying an identity is not persistable
       whatever the engine reported.
      
       Three defects found in this cell's own code before shipping:
       - A DETECTOR IS NOT A REDACTOR. Reusing M6.5's `LEARNER_ID_PATTERN` to
         *remove* rather than detect left `[redacted-learner]042`: the pattern
         matches `learner:` plus ONE character, which answers "is there an
         identifier here?" and does not remove one. Removal now uses
         maximal-munch patterns, asserted directly per kind.
       - AND THEN THE FIX WAS OVERCLAIMED. The first comment said re-scanning
         proves the redaction worked. It does not: `[redacted-learner]042`
         matches no detector, because the `learner` inside the placeholder is
         followed by `]`. The re-scan catches residue that STILL LOOKS LIKE an
         identity and cannot see residue the placeholder masked; what protects
         is maximal munch. The comment and the tests now say exactly that, and
         a test asserts the re-scan's blind spot rather than hiding it.
       - A VOCABULARY WORD TRIPPED THE SCAN ON ITSELF. The redaction kind
         `learner_identifier` matches the learner detector, so the published
         payload reported one identity finding — its own enum value. Narrowing
         the scan to dodge that would have hollowed it out, so the value is
         renamed `identity_reference` and the scan stays total.
      
       `feedback.intake.nothing-left-after-redaction` was dead until a fourth
       fix: "[redacted-email]" is not an empty string, so a submission that was
       only an address was being accepted as a signal with no content, padding
       a cluster's reach. Intake now checks what survives once the placeholders
       are removed.
      
       Verification: 53 focused Python tests (35 engine, 11 service, 7 API
       contract); corpus 30/30 across intake, triage, authority, lineage and
       privacy with Wilson 95% intervals; 8 mutations — 7 caught (2 by the
       contract, 5 by corpus mismatch) and the eighth, removing the residue
       re-check, provably unobservable because the removal patterns prevent
       what it guards against, so it is exercised instead by simulating the
       pattern-list drift it exists for; ruff + ruff format + mypy --strict
       clean; OpenAPI and the generated TypeScript client regenerated (5 paths,
       20 schemas added, none removed, no name collisions); 33 browser checks
       green across chromium/firefox/webkit with automatic serious + critical
       axe analysis, a 390px overflow audit and a keyboard-reachability sweep.
       The browser fixture is generated from the real service.
      
       Limitations, stated rather than implied: redaction removes the shapes it
       knows — an email, a UUID, a `learner:`-style reference. A submission
       that names somebody in prose still names them afterwards, which is why
       the queue carries no signal text at all by default and a quote needs a
       human to release it. Clustering is by target content revision and source;
       it is not semantic, so two differently-worded reports of the same problem
       on the same revision cluster together and two different problems on one
       revision also do. The cohort threshold applies to learner-voice sources
       only, because one accessibility failure is one accessibility failure.
       `severity` is a function of the source, not of the content. Nothing here
       submits a draft to a gate or publishes one — the lineage record has the
       fields to track that and this cell only ever writes `not_submitted` /
       `not_published`. NO PRODUCTION FEEDBACK SOURCE IS BOUND TO THIS INTAKE;
       signals arrive in the request. Browser API transport is mocked;
       production deployment, representative author usability, and formal
       accessibility certification are not established.
      

      -->

  • M6.12 Build experiment governance for content or sequencing A/B tests. Evidence: hypothesis, metric, guardrail, allocation, stop, and analysis records.
    • M6.12.a Define experiment identity/version, hypothesis, exact variants/ content/algorithm revisions, eligible population, owner, duration, and institutional/legal/ethics approval. Evidence: schema.

    • M6.12.b Pre-register primary/secondary metrics, guardrails, estimand, power/sample assumptions, analysis method, subgroup/privacy rules, and multiple-testing policy. Evidence: analysis plan.

    • M6.12.c Define deterministic privacy-safe assignment, allocation, exclusion, ramp, consistency, contamination, consent/opt-out, and logging. Evidence: allocation tests.

    • M6.12.d Define start/pause/stop/harm/quality/SLO/integrity thresholds, monitoring, authority, and immutable commands. Evidence: guardrail tests.

    • M6.12.e Analyze per preregistration with attrition/missingness/uncertainty/ limitations, human decision, and no automatic rollout. Evidence: analysis records. <!-- M6.12 evidence (2026-09-05) Built the 6.12.0 experiment-governance contract (apps/metis/service/src/metis/schemas/experiment_governance.py), the allocation/guardrail/analysis engine (services/experiment_governance_engine.py), a 59-case labelled corpus over 15 fixture families and five dimensions (data/m6_12_experiment_governance_corpus.json, evaluated by services/experiment_governance_evaluation.py with Wilson intervals per dimension), an append-only tenant-scoped store (services/experiment_governance_service.py), instructor and admin routes (api/experiment_governance.py, api/admin.py), and the /experiment-governance admin workbench.

       M6.12.a — schema. `ExperimentDefinition` requires exactly one control
       arm and an allocation summing to 10,000 basis points; every
       `VariantDefinition` pins a `content_revision_sha256`, so a variant
       cannot name content without naming which revision of it.
      
       M6.12.b — analysis plan. `PreRegistrationIntake` carries everything a
       plan seals and has NO field for its own revision hash; the service
       derives it from the plan's own content and `PreRegistration` refuses a
       declared seal that does not hash the plan. Before this the hash was a
       caller-supplied number only ever compared to itself, and the earlier
       helper hashed the metrics alone — so alpha, the target sample and the
       multiple-testing policy were all editable inside a "sealed" plan. Eight
       parametrised cases prove each sealed field moves the hash.
      
       M6.12.c — allocation tests. The ramp decision reads digest[0:8] and the
       arm decision reads digest[8:16]: drawing both from the same bits leaves
       a partial ramp entirely inside one arm (measured: at a 25% ramp every
       assigned subject landed in control), and widening a ramp then reshuffles
       arms. `ramp_consistency` widens 25%→50/75/100% over 800 subjects and
       reports 0 subjects moved. `ExperimentCohort` states how many subjects
       carry each disqualifying condition, because without it four of the five
       `AssignmentExclusion` values were unreachable through the API and the
       branches that refuse them never ran; `assign_one`'s precedence is tested
       directly against subjects carrying two conditions at once. An
       `AssignmentRecord` has no subject-identifier field, only a per-tenant
       salted digest, and the service refuses to allocate at all when
       `experiment_governance_assignment_salt` is unset rather than emitting a
       guessable one. `contamination_free` was REMOVED: with the report's
       digest-uniqueness rule in place no constructible report could ever carry
       two arms for one subject, so the flag was true by construction.
      
       M6.12.d — guardrail tests. `CommandLog.a_stop_is_final` refuses a resume
       after a stop. A guardrail reads the direction it registered (a drop in a
       higher-is-better guardrail is adverse, a rise is not) and treats its
       tolerance as a limit that may be reached but not passed; both boundaries
       are tested. `AnalysisResult` refuses a `consider_rollout` recommendation
       alongside any breach.
      
       M6.12.e — analysis records. `ExperimentRunResponse` requires an analysis
       to exist exactly when the command log says stopped, so a preregistered
       test cannot be read mid-flight. `MetricAnalysis` makes a metric
       confirmatory exactly when the plan gave it a role and refuses a
       significant verdict on an exploratory one. `AttritionReport` names the
       metric it counted and reports the worst-observed primary; previously
       `build_run` passed observed == assigned, so the report could only ever
       say zero. `MetricObservation` carries rates and coverage, never counts —
       counts are derived from the arms the allocation actually produced, so a
       metric cannot be reported on more subjects than were enrolled (the
       earlier count-carrying shape let a 1,000-observation request accompany a
       10% ramp that reached 106). `RolloutRecommendation` has no field naming
       a variant to promote and carries `automatic_rollout: Literal[False]`.
      
       Three defects the adversarial pass found in my own code:
      
       1. A LOOKUP TABLE CANNOT SERVE A CORRECTION THAT DIVIDES. The critical
          values were a hand-kept dict, and `_effective_alpha` floors — so
          Bonferroni over three tests at 500 basis points asks for alpha 166
          and the table's `167` entry was unreachable. Sixteen further
          reachable alphas had no entry at all, and a plan with six primaries
          would have returned no verdict on anything. `_critical_value` now
          computes from `statistics.NormalDist.inv_cdf` (Wichura AS241 in the
          standard library) and the five ratified constants CHECK that
          computation at import; a test proves every alpha the corrector can
          produce over twenty tests resolves.
       2. THREE REASONS TO TAKE NO VERDICT SHARED ONE EXPLANATION. A missing
          critical value and a zero pooled variance both returned
          `underpowered` with a detail sentence naming a sample-size shortfall
          that had not happened. They are now separate, and the surviving
          `statistic is None` branch's comment was corrected once testing
          showed the empty-arm half of it is shadowed by the sample check.
       3. A TWO-SIDED TEST THAT READS THE SIGN HIDES EVERY HARM. Dropping the
          `abs()` around the statistic survived the suite, because no fixture
          had a treatment that significantly hurt its primary metric. Added
          the `treatment_hurts` scenario: −11.97 points, `significant`,
          `keep_control`.
      
       Verification: 78 service tests (engine 57, store 10, contract 11) plus
       59/59 corpus cases matched and 8/8 governance probes refused; ruff,
       `ruff format` and `mypy --strict` clean on all five modules.
       Mutation testing applied 27 mutations to the engine and schema across
       three rounds and the suite now kills 26; the one survivor ("read the
       last command instead of folding the sequence") is semantically
       equivalent — every command kind maps to one state — and the docstring
       that claimed otherwise was corrected rather than a test written around
       it. The duplicate-model-name scan added in M6.10 caught a real collision
       (`SyntheticCohort` already belonged to M6.5's learner simulation) before
       it could rename both schemas in the published API; renamed to
       `ExperimentCohort`. OpenAPI regenerated and compared semantically
       against HEAD: +26 schemas, 0 removed, 0 existing schemas changed,
       4 new paths. Browser: 16 Playwright specs green on chromium against a
       fixture generated from the real service (28 runs across 7 scenarios ×
       4 ramps), including a 390px viewport check and keyboard reachability of
       every horizontally scrolling region.
      

      -->

  • M6.13 Build curriculum, item-writing, verification, accessibility, and tutoring agent plans with least-privilege tool grants. Evidence: sources, course revision, policy, budget, stop conditions, and approval points are locked before dispatch.
    • M6.13.a Define versioned agent-plan/step contracts and distinct plan templates/output schemas for curriculum, item, verification, accessibility, and tutoring work. Evidence: plan schemas.

    • M6.13.b Bind exact source/extraction/claim/course/item/rubric/standard/ policy/config revisions and purpose-minimized learner context where permitted. Evidence: input binding tests.

    • M6.13.c Resolve each tool/model/provider grant by version, typed operation, data scope, tenant/course, resource/network limit, answer-key visibility, and expiry. Evidence: grant matrix.

    • M6.13.d Define dependencies/preconditions/outputs/evidence, budget per step/total, deadlines/retries, hard stop/refusal, and no silent fallback. Evidence: execution plan tests.

    • M6.13.e Insert required human review/approval points with role separation and unchanged-plan/evidence binding; agents receive no approval/publish/ grade/integrity/credential authority. Evidence: authority tests.

    • M6.13.f Validate graph/grants/inputs/policy/budget/stops/approvals/ destinations before dispatch and reject direct plan injection. Evidence: zero-dispatch tests. <!-- M6.13 evidence (2026-09-05) Built the 6.13.0 agent-plan contract (apps/metis/service/src/metis/schemas/agent_plan.py), the template registry / grant matrix / dispatch engine (services/agent_plan_engine.py), a 54-case labelled corpus over 13 fixture families and six dimensions (data/m6_13_agent_plan_corpus.json, evaluated by services/agent_plan_evaluation.py with Wilson intervals per dimension), an append-only tenant-scoped store (services/agent_plan_service.py), instructor and admin routes (api/agent_plan.py, api/admin.py), and the /agent-plans admin workbench.

       M6.13.a — plan schemas. `OUTPUT_FOR_PLAN_KIND` is a table, not a field:
       a curriculum plan cannot declare that it emits items. Five ratified
       templates cover the five kinds, and `_validate_templates` refuses at
       import a template whose steps ask for an operation the matrix does not
       permit its kind to hold, whose required inputs no step reads, or whose
       approval role equals its authoring role.
      
       M6.13.b — input binding tests. `InputBinding` requires a
       `revision_sha256`; `PlanStep` refuses a binding the step holds no grant
       to read, and refuses a learner-context grant with no field named —
       standing permission over every field is not least privilege.
       `LearnerContextField` has no learner-identifier field, so a context
       grant cannot become a way to fetch a named person's record.
      
       M6.13.c — grant matrix. 75 cells (5 kinds x 15 operations), total by
       construction: `_validate_matrix` fails at import on a gap, because an
       unlisted pair would read as permitted. Only tutoring and accessibility
       may read learner context; tutoring may NOT read item revisions, because
       a tutor that can fetch items can fetch the question a learner is about
       to be asked. Every grant names a tool version, a typed operation, a
       scope, a tenant, a per-operation resource ceiling, a network posture
       and an expiry, and carries `answer_key_visible: Literal[False]`.
      
       M6.13.d — execution plan tests. `StopCondition.on_trigger` has two
       values, `hard_stop` and `refuse_and_report`; there is no value that
       continues on a smaller model. Every step must state what happens when
       its budget runs out. The steps together may not outspend the plan and
       no step deadline may outrun the plan's. Step order is the topological
       order, so a cycle cannot be expressed.
      
       M6.13.e — authority tests. `ToolOperation` contains no verb for
       approving, publishing, grading, deciding an integrity appeal, waiving a
       gate, issuing a credential, releasing an answer key, or modifying
       published content. Those eight are listed in `WITHHELD_AUTHORITIES`
       and the two sets are asserted disjoint at import; each is probed by
       actually attempting to construct a grant naming it. `ApprovalPoint`
       blocks rather than notifies, and an approval by the authoring role is
       refused at construction.
      
       M6.13.f — zero-dispatch tests. `PlanDispatchRequest` has no field for a
       plan, a step, a grant, an approval or a budget, so injection has
       nowhere to land; the service expands the plan from the registry. All 14
       refusal reasons are reachable and each is covered by a test. A refused
       decision carries `plan=None` and `plan_sha256=None`, enforced by
       `DispatchDecision`, so a turned-down request leaves no authorised-
       looking artefact. Every check runs, so a caller fixing one problem does
       not have to dispatch again to find the next.
      
       Three defects the adversarial pass found in my own code:
      
       1. A TUTORING DISPATCH WITH NO LEARNER CONTEXT CRASHED. `_expand`
          issued the template's learner-context grant unconditionally and
          `PlanStep` then refused the empty field set, raising out of the
          engine instead of returning a decision. Fixed by granting only what
          the bindings will use — which is also the narrower plan.
       2. AN UNUSED PARAMETER IN A DECISION FUNCTION IS A SILENT STUB.
          `check_request` accepted `tenant_id` and immediately `del`d it,
          telling a reader the tenancy had been checked there. Removed; the
          tenancy is checked in `check_plan` against the grants the expansion
          issued, and in the store against the caller.
       3. ONE CEILING WEARING FIFTEEN HATS. Every grant was built with the
          same `_ceiling()` defaults, so a resource limit bounded the loosest
          operation and said nothing about the other fourteen. Replaced with
          `CEILING_FOR_OPERATION`, validated total at import; the narrowest
          read in the vocabulary (learner context, 8 KiB) is now provably the
          narrowest.
      
       A fourth was removed rather than fixed: a per-grant matrix re-check in
       the store was unreachable, because `check_plan` already raises
       `grant_exceeds_template_scope` for exactly that case. A defensive
       branch no test can reach reads as coverage it does not give.
      
       Verification: 95 service tests (engine 74, store 11, contract 11) plus
       54/54 corpus cases matched and 23/23 dispatch probes refused across all
       four refusal mechanisms; ruff, `ruff format` and `mypy --strict` clean
       on all five modules. Mutation testing applied 31 mutations across the
       schema, engine and store and the suite kills all 31. OpenAPI compared
       semantically against the previous contract: +20 schemas, 0 removed, 0
       existing schemas changed, 5 new paths. Browser: 15 Playwright specs
       green on chromium against a fixture generated from the real service
       (9 dispatches, 5 dispatched and 4 refused), including the rendered
       75-cell grant matrix, a 390px viewport check and keyboard reachability
       of every horizontally scrolling region.
      

      -->

  • M6.14 Build Metis-scoped agent run, queue, plan, tool-call, budget, evidence, output, and provenance views. Evidence: live traced journey and tenant isolation.
    • M6.14.a Define audience-safe projections and queries for request/plan/ step/queue/run/attempt/tool/budget/evidence/output/decision/lineage. Evidence: view contracts.

    • M6.14.b Build run list/detail and plan DAG with state, dependencies, approvals/stops, queue/reservations, live cursor events, reconnect, and immutable versions. Evidence: run journey.

    • M6.14.c Present tool/version/purpose/status/timing/budget and redacted typed inputs/outputs/errors without prompts/keys/learner data. Evidence: tool-call view tests.

    • M6.14.d Present allocated/consumed/remaining token/tool/compute/cost/time plus evidence/source/citation/gate and proposed output lineage. Evidence: budget/evidence tests.

    • M6.14.e Enforce tenant/course/actor/audience/object/property/function authorization across URLs/search/streams/exports/actions. Evidence: isolation tests.

    • M6.14.f Verify live traced failure/review/accept journey, semantic accessibility, responsive modes, and long-history performance. Evidence: Playwright E2E. <!-- M6.14 evidence (2026-09-05) Built the 6.14.0 run-view contract (apps/metis/service/src/metis/schemas/agent_run_view.py), the audience matrix / redaction / event-chain / budget engine (services/agent_run_view_engine.py), a 39-case labelled corpus over 13 fixture families and six dimensions (data/m6_14_agent_run_view_corpus.json, evaluated by services/agent_run_view_evaluation.py with Wilson intervals per dimension), an append-only tenant-scoped store (services/agent_run_view_service.py), instructor routes (api/agent_run_view.py), and the /agent-run-views admin workbench.

       M6.14.a — view contracts. AN AUDIENCE IS NOT A RANK. The existing
       runtime's `AgentRoleVisibility` is a four-level ladder used once, for an
       approval gate, so one role always sees a superset of another's. This
       cell's `AUDIENCE_MATRIX` states, for every one of 5 audiences x 12
       objects, exactly which of 5 functions apply; `_validate_matrix` fails at
       import on a gap, and `authorize` subscripts rather than defaulting.
       Five audience pairs are provably incomparable (author and operator each
       hold something the other does not), which is what distinguishes an
       audience from a rank.
      
       M6.14.b — run journey. The event stream is a hash chain: each event's
       digest covers its own contents including its predecessor's, so an
       inserted, dropped or reordered event breaks the chain where it was
       touched. A reconnecting reader gets `contiguous` AND `broke_at` — a
       boolean says something is wrong, a sequence says where to go and fetch.
       Both are validated against the page they describe, so a page cannot
       claim to be whole. Run list rows carry state and progress, never the
       whole view.
      
       M6.14.c — tool-call view tests. `RedactedPayload` has no `value`,
       `text`, `prompt` or `body` field. A prompt or an API key is not blanked
       out; there is nowhere for the bytes to sit. Each field is described by
       name, kind, byte size (the JSON encoding, asserted) and a reason drawn
       from a six-value vocabulary, and a field nobody classified defaults to
       withheld — describing an unclassified field is a choice to show it.
      
       M6.14.d — budget/evidence tests. `BudgetLine.remaining` is checked
       against `allocated - consumed`, and consumption for calls, compute and
       cost is summed from the attempts rather than carried as a counter that
       could drift. A ledger omitting any of the five resources is refused.
       `OutputProposal` carries `published: Literal[False]` and can name only a
       proposal or a review queue, and a view carrying an output whose cited
       evidence it does not also carry is unconstructible.
      
       M6.14.e — isolation tests. 16 probes across all five surfaces a caller
       actually has (url, search, stream, export, action), all refused; the
       report refuses to publish if any surface is unprobed. Tenancy is
       exercised against the real store: a run recorded under one tenant is not
       found under another, a view cannot be stored under a tenant other than
       the caller's, and a role the audience table does not name is refused
       rather than defaulted to the narrowest view — showing an unknown role
       the least still shows it something nobody decided it should see. The
       request shape has no `audience` field, so a caller cannot choose one.
      
       M6.14.f — Playwright E2E. 15 specs green on chromium against a fixture
       generated from the real service, walking all five audiences and
       asserting that an author sees outputs but not tool calls while an
       operator sees the reverse, that no prompt, key or learner reference
       reaches the rendered page, and that a viewer who may not stream is still
       told how long the history is. The harness's own axe pass caught a real
       WCAG AA contrast failure (3.48:1 muted text); fixed to 6.8:1 here and in
       M6.13's page, which shares the stylesheet.
      
       Two defects the adversarial pass found in my own code:
      
       1. A REPORTED GAP MUST BE THE GAP. `EventPage` said only whether it was
          contiguous, and my own observation computed the break point as
          `from_sequence + len(events)` — 8 for a chain that broke at 5. Added
          `broke_at`, derived and validated, so a wrong break point is
          unconstructible.
       2. A MODEL AND A BUILDER WITH NO CALLER IS A SPEC-ONLY MODULE.
          `RunListItemView` and `list_item` were built and never wired, while
          M6.14.b asks for a run list. Wired as `GET /agent-run-views/runs`,
          with `overspent` sourced from the budget only when the audience may
          read one — so a list row cannot leak an overspend to a viewer with no
          budget access.
      
       Verification: 76 service tests (engine 53, store 15, contract 11) plus
       39/39 corpus cases matched and 16/16 isolation probes refused across all
       five surfaces; ruff, `ruff format` and `mypy --strict` clean on all four
       modules. Mutation testing applied 34 mutations across the schema, engine
       and store and the suite kills all 34. OpenAPI compared semantically:
       +23 schemas, 0 removed, 0 existing schemas changed, 3 new paths.
       Long-history performance is measured, not asserted: paging the tail of a
       5,000-event chain took ~650 microseconds, taken on the last page because
       a scan from the start is fast at sequence 1.
      

      -->

  • M6.15 Intake agent output as proposals requiring attributed human acceptance. Evidence: agents cannot expose answer keys, change grades, decide integrity appeals, waive gates, or publish.
    • M6.15.a Define signed/hashed result envelope with agent/run/step/tool/ model/config versions, exact inputs, output kind/schema, evidence/ citations, confidence, cost, and status. Evidence: result schema.

    • M6.15.b Validate signature/hash/schema/size/type/expected step/source/ tenant/course and scan content/files before storage. Evidence: hostile result tests.

    • M6.15.c Detect/redact/reject forbidden learner/answer/integrity/grade/ credential/private-source/tool-secret fields and unsupported claims. Evidence: protected-field tests.

    • M6.15.d Store curriculum/item/content/review/accessibility/tutor outputs only as attributed proposals linked to exact authoring revisions. Evidence: proposal persistence tests.

    • M6.15.e Require an authorized human accept/edit/reject/defer decision; acceptance creates a new revision and reruns relevant gates. Evidence: review tests.

    • M6.15.f Prove no result field/action can mutate grades, decide appeals, waive gates, approve/publish content, or issue credentials. Evidence: mass-assignment/authority suite. <!-- M6.15 evidence (2026-09-05) Built the 6.15.0 result-intake contract (apps/metis/service/src/metis/schemas/agent_result_intake.py), the signature / protected-content / review engine (services/agent_result_intake_engine.py), a 35-case labelled corpus over 13 fixture families and six dimensions (data/m6_15_agent_result_intake_corpus.json, evaluated by services/agent_result_intake_evaluation.py with Wilson intervals per dimension), an append-only tenant-scoped store (services/agent_result_intake_service.py), instructor routes (api/agent_result_intake.py), and the /agent-results admin workbench.

       M6.15.a — result schema. `ResultEnvelopeBody` is its own model and the
       signature is `model_dump()` over the whole of it, so a field added
       later joins the signature automatically; a test asserts the signer
       names no individual field, because naming them is how a field ends up
       outside the signature. Six parametrised cases prove each producer
       version moves the signature. `ResultEnvelope` refuses a declared body
       hash that is not the body's.
      
       M6.15.b — hostile result tests. The expected step and output kind come
       from the caller's plan, never from the envelope: taking them from the
       result would accept a result for a step nobody dispatched. Every check
       runs, so a producer fixing one problem does not resubmit to find the
       next. An unset signing secret refuses every result rather than
       accepting unverified ones. A second submission of one envelope id is a
       replay, not an update.
      
       M6.15.c — protected-field tests. `ResultPayload` is a list of named
       fields with a declared kind, never a free-form object, so a payload
       cannot smuggle a nested structure past a shallow scan; `ALLOWED_FIELDS`
       names what each output kind may carry, and anything unlisted is refused
       rather than stored just in case. Eleven patterns cover eight protected
       kinds, each distinct kind reported once per value, and a finding never
       quotes what it matched — a finding that repeated the credential would
       carry the credential into the record meant to prove it was kept out. A
       protected finding is a rejection, not a note.
      
       M6.15.d — proposal persistence tests. A stored proposal is attributed
       (agent, model, tool, config, all versioned), pinned to the exact
       revisions it cited, and sits in `awaiting_review`. A rejected result
       leaves no proposal, enforced by `IntakeResponse`.
      
       M6.15.e — review tests. A decision names the human and the exact
       proposal hash they saw; a decision about a proposal that changed since
       is refused. Acceptance produces a `DraftRevision` attributed to the
       accepter and schedules every gate the output kind names, all `pending`
       — acceptance schedules a gate rather than deciding one. A reviewer's
       edit is scanned on the same terms as the agent's output and cannot
       introduce an unlisted field.
      
       M6.15.f — mass-assignment/authority suite. `FORBIDDEN_FIELD_SUBSTRINGS`
       names fourteen words; `reachable_field_names()` walks all 103 field
       names in the published model tree, nested models included, and
       `_validate_no_forbidden_field_names` refuses at import if one appears.
       The scan admits NO exemption list, because an exemption is exactly the
       hole it closes — two of my own field names (`published_field_names`,
       `replaces_published_revision`) were renamed rather than excused. A
       second import-time guard proves no allowlisted field name contains a
       forbidden substring, so the two defences cannot drift apart. 14 probes
       cover all six named authorities across all four refusal mechanisms
       (absent field, allowlist, scan, state machine), all refused.
      
       Two defects the adversarial pass found in my own code:
      
       1. A SECRET DETECTOR THAT STOPS AT THE FIRST SEPARATOR MISSES THE
          COMMONEST SHAPE. `\b(?:sk|pk|api|bearer)[-_][A-Za-z0-9]{8,}` does
          not match `sk-live-abcdef0123456789`, because `live` is four
          characters and the hyphen ends the run — and that is the exact
          string M6.14's own fixture uses. Split into three shapes with
          separators allowed inside the token, and checked against four real
          key formats and three innocuous phrases.
       2. `\b` BEFORE A SLASH NEVER MATCHES AFTER A SPACE. The private-source
          pattern `\b/(?:home|Users|root)/` could not fire in ordinary prose,
          because a space and a slash are both non-word characters. Replaced
          with a lookbehind that also excludes a public URL path.
      
       Verification: 91 service tests (engine 62, store 15, contract 14) plus
       35/35 corpus cases matched and 14/14 authority probes refused; ruff,
       `ruff format` and `mypy --strict` clean on all four modules. Mutation
       testing applied 43 mutations across the schema, engine and store and
       the suite kills all 43 — including the guards the layer above shadows,
       tested directly rather than removed, because a future allowlist edit is
       exactly what they are for. OpenAPI compared semantically: +20 schemas,
       0 removed, 0 existing schemas changed, 4 new paths. Browser: 10
       Playwright specs green on chromium against a fixture generated from the
       real service, asserting no protected content reaches the rendered page.
      

      -->

  • M6.16 Implement agent pause, redirect, cancel, budget kill, and replay. Evidence: failure and race tests preserve immutable run/evidence history.
    • M6.16.a Define separate command eligibility/role/scope/reason/ idempotency/expected version/outcome contracts. Evidence: command schemas.

    • M6.16.b Pause at safe tool/step boundary, checkpoint resumable state, stop new calls, and resume only after full revalidation. Evidence: pause tests.

    • M6.16.c Redirect through a new plan revision with changed objective/ inputs/tools/budget and preserved completed history. Evidence: redirect lineage tests.

    • M6.16.d Cancel queued/running/downstream work with provider/tool propagation, late-result fencing, cleanup, and partial evidence retained. Evidence: cancellation race tests.

    • M6.16.e Kill before next spend at hard budget ceiling and reconcile nonrefundable in-flight usage without overrun success. Evidence: budget kill tests.

    • M6.16.f Replay as new run identity after current source/course/policy/ grant/budget/rights validation; never mutate source run. Evidence: replay tests. <!-- M6.16 evidence (2026-09-05) Built the 6.16.0 run-control contract (apps/metis/service/src/metis/schemas/agent_command.py), the eligibility / apply / fence engine (services/agent_command_engine.py), a 44-case labelled corpus over 13 fixture families and six dimensions (data/m6_16_agent_command_corpus.json, evaluated by services/agent_command_evaluation.py with Wilson intervals per dimension), an append-only tenant-scoped store (services/agent_command_service.py), instructor routes (api/agent_command.py), and the /run-control admin workbench.

       M6.16.a — command schemas. Six command types, not one with a `kind`
       field: a pause has no field naming a new plan and a redirect has none
       naming a refund, so a command cannot ask for something its verb does
       not mean. Six routes, one per verb, each accepting only its own body.
       `ELIGIBLE_LIFECYCLES` is total over 6 commands x 7 lifecycles and
       validated at import, which also refuses a command eligible nowhere and
       (except replay) any command eligible against a terminal lifecycle.
       Every command carries `expected_version`, and `CommandRecord` records
       both the version named and the version found — a reader of a stale
       refusal needs to know what the issuer was looking at.
      
       M6.16.b — pause tests. A pause takes a `Checkpoint` whose completed
       plus pending equals the plan; a checkpoint that lost a step would
       resume into a run that skipped it. It stops *before* a pending step, so
       a run with nothing left is refused with `no_safe_boundary` rather than
       interrupting work in flight. `RunControlState` requires a checkpoint
       exactly while paused, so one cannot outlive the resume that cleared it.
       `ResumeCommand.revalidated` is a `Literal[True]` and the engine refuses
       a resume whose revalidation did not pass.
      
       M6.16.c — redirect lineage tests. A redirect changes the plan id and
       leaves the completed step ids alone; the store refuses any transition
       that discards completed work, because a reviewer may already have cited
       it. A redirect applies only to a queued or paused run — mid-flight is
       not when you change what a run is for.
      
       M6.16.d — cancellation race tests. A cancel bumps a fence, and
       `check_fence` refuses a result whose fence is not the current one: a
       tool call that started before a cancel can finish after it, and without
       the fence it looks exactly like a timely result. `CommandRecord` ties
       the fence move to the run ending, so a cancel that did not move it is
       unconstructible. Five race probes issue two commands against one
       version and each resolves to exactly one change; the loser is the
       second writer, not the gentler verb, proved by running the same pair in
       both orders.
      
       M6.16.e — budget kill tests. A kill below the ceiling is refused —
       that is a cancel wearing a budget's name. `SpendLedger` carries the
       unrecoverable share separately and refuses more unrecoverable than was
       ever committed; the kill records what was already committed to a
       provider rather than writing it off, and the run's lifecycle becomes
       `killed`, never `completed`.
      
       M6.16.f — replay tests. `ReplayCommand` refuses its own run id and
       requires all six revalidations as distinct values (`min_length=6`
       counts entries; the check counts distinct ones). An applied replay
       moves only `version` and `updated_at` on the source — asserted by
       diffing every field of the state before and after — and
       `CommandRecord` refuses a replay whose lifecycle changed.
      
       Two defects the adversarial pass found in my own code:
      
       1. A REFUSAL REBUILT THE STATE FROM PARTS AND DROPPED THE CHECKPOINT.
          A refused pause against an already-paused run returned a state with
          no checkpoint, which then failed its own validator. A refusal now
          returns the original state object.
       2. A LOSER'S `version_before` IS WHAT IT SAW, NOT WHAT IT NAMED. My own
          race message read "named version 5" for a command that named 4.
          Added `expected_version` to `CommandRecord`, with a validator
          requiring the two to agree on anything that applied.
      
       `command_kinds_in` was written and never called; removed rather than
       left as a helper with no caller.
      
       Verification: 88 service tests (engine 62, store 18, contract 8) plus
       44/44 corpus cases matched and 5/5 races resolved across five verbs;
       ruff, `ruff format` and `mypy --strict` clean on all four modules.
       Mutation testing applied 44 mutations across the schema, engine and
       store and the suite kills all 44. OpenAPI compared semantically: +19
       schemas, 0 removed, 0 existing schemas changed, 11 new paths. Browser:
       10 Playwright specs green on chromium against a fixture generated from
       the real service, including the rendered 42-cell eligibility matrix.
      

      -->

  • M6.17 Build tutor-configuration and preview workspaces for text, live voice, pronunciation, pacing, hint ladders, Socratic guidance, demonstrations, accessibility, human escalation, and session summaries. Evidence: simulated and real learner contexts are visibly distinct and integrity interlocks remain active in every modality.
    • M6.17.a Build versioned persona/pedagogy/language/tone/disclosure and exact course/source/objective/config applicability editor. Evidence: tutor config tests.

    • M6.17.b Build hint-ladder/Socratic/scaffolding/demonstration/direct- answer bounds, triggers, penalties, and assessment-mode rules. Evidence: pedagogy editor tests.

    • M6.17.c Build allowlisted tool/source/citation/model/provider/budget/ retention/safety/escalation editor with effective grant preview. Evidence: configuration guard tests.

    • M6.17.d Build text preview with synthetic learner states, transcript, grounding/pedagogy/integrity findings, and no real side effects. Evidence: text preview tests.

    • M6.17.e Build live voice preview with approved voice/consent, language/ pronunciation/pacing/interruption/captions/transcript/latency and degraded states. Evidence: voice preview tests.

    • M6.17.f Build accessibility modes for keyboard/screen reader/visual/ hearing/cognitive supports and equivalent source/hint/tool behavior. Evidence: accessibility tests.

    • M6.17.g Build human escalation simulation with reason/context summary/ privacy limits/routing/acknowledgement and no fake handoff. Evidence: escalation tests.

    • M6.17.h Build session summary/retention/delete/export settings and show exact context sources, redactions, model/config versions, and costs. Evidence: session tests.

    • M6.17.i Visually and contractually distinguish synthetic preview from authorized real learner context; rerun integrity interlock tests for text, voice, tool, demo, accessibility, and cached/offline paths. Evidence: modality boundary suite. <!-- M6.17 evidence (2026-09-06) Built the 6.17.0 tutor-preview contract (apps/metis/service/src/metis/schemas/tutor_preview.py), the pedagogy / grant / render / boundary engine (services/tutor_preview_engine.py), a 46-case labelled corpus over 13 fixture families and seven dimensions (data/m6_17_tutor_preview_corpus.json, evaluated by services/tutor_preview_evaluation.py with Wilson intervals per dimension), an append-only tenant-scoped store (services/tutor_preview_service.py), instructor routes (api/tutor_preview.py), and the /tutor-previews admin workbench.

       M6.17.a — tutor config tests. `Applicability` pins the course, source
       and config revisions; `DisclosureLevel` has two values and both
       disclose, because a tutor that says nothing about itself has no
       configuration this module can express.
      
       M6.17.b — pedagogy editor tests. `MOVES_IN_MODE` is total over 4 modes
       x 6 moves and validated at import, which also refuses a matrix letting
       an assessment hint or answer, and a mode permitting nothing at all. A
       `HintLadder` must climb: a rung revealing less than the one before, or
       costing less for more, is unconstructible — that is the rung every
       learner takes first. `PedagogyBounds` cannot widen what its mode allows.
      
       M6.17.c — configuration guard tests. `EffectiveGrant` is checked to be
       exactly the intersection of requested and held, and its `dropped` list
       must name every requested tool the grant withheld — an author who is
       not told goes on believing the tutor has it. `RetentionSettings`
       refuses audio kept longer than the transcript it belongs to.
      
       M6.17.d — text preview tests. A move the mode forbids is NOT rendered:
       it is reported as a finding, which is what an interlock looks like from
       outside. Turns are numbered from zero without gaps, and `TextPreview`
       carries `produced_no_side_effects: Literal[True]`.
      
       M6.17.e — voice preview tests. Consent, captions and a transcript are
       `Literal[True]`. The degradation is CHOSEN from the measurement, and a
       latency breach reporting no degradation is unconstructible — a breach
       nobody degraded for is a breach nobody noticed.
      
       M6.17.f — accessibility tests. Every mode offers the same moves and the
       same sources, compared AS SETS across the six modes rather than each
       mode asserting its own adequacy. A mode whose tool need the grant does
       not meet is reported as a gap, and the response refuses any mode
       offering a tool outside the effective grant.
      
       M6.17.g — escalation tests. There is no state meaning "handed off"
       without an acknowledgement: `acknowledged` requires both a name and a
       time, and telling the learner a person has it while nobody has is
       unconstructible. All three of learner words, identifier and attempt
       contents are withheld from the summary.
      
       M6.17.h — session tests. The summary names every context source WITH
       the revision it was read at, the persona and model versions, the
       redactions and the cost, and must count the turns the preview actually
       rendered.
      
       M6.17.i — modality boundary suite. `SyntheticLearnerState` has no field
       that could hold a real learner and its label must start `synthetic-`;
       every rendered turn is prefixed `[synthetic preview]`, checked in
       `PreviewTurn` rather than left to a renderer; and `ModalityCheck` pairs
       the interlock claim with the move attempted and the mode it was
       attempted in, so a path claiming an active interlock while making a
       permitted move is unconstructible. All six paths (text, voice, tool,
       demonstration, accessibility, cached) are checked, and the practice-mode
       run shows the interlock is NOT simply always on.
      
       Verification: 88 service tests (engine 69, store 9, contract 10) plus
       46/46 corpus cases matched and 15/15 boundary probes refused across all
       four mechanisms (absent field, marker check, matrix cell, intersection);
       ruff, `ruff format` and `mypy --strict` clean on all four modules.
       Mutation testing applied 44 mutations across the schema, engine and
       store and the suite kills all 44 — including three guards a layer above
       shadows, tested directly because a stored-bytes route or an allowlist
       edit makes each reachable. OpenAPI compared semantically: +26 schemas,
       0 removed, 0 existing schemas changed, 3 new paths. Browser: 11
       Playwright specs green on chromium against a fixture generated from the
       real service, including the rendered 24-cell pedagogy matrix and an
       assertion that no learner reference reaches the page.
      

      -->

  • M6.18 Build credential pathway authoring that links objectives, assessments, evidence, mastery/grade rules, human approvals, issuer policy, expiry, and revocation. Evidence: course completion alone cannot mint a credential unless the locked criteria and authority are satisfied.
    • M6.18.a Build versioned credential pathway/template selection, issuer/ authority, audience, achievement/standards, lifecycle, locale, and validity editor. Evidence: pathway schema/UI tests.

    • M6.18.b Author nested objective/course/assessment/evidence/mastery/grade/ experience/time requirements with exact revisions and deterministic evaluation semantics. Evidence: criteria tests.

    • M6.18.c Configure acceptable evidence types/sources/verification/ freshness/privacy and required human review/approval/quorum. Evidence: evidence policy tests.

    • M6.18.d Preview current/synthetic subject eligibility per criterion with missing/stale/conflicting/N/A state and no exposed private detail. Evidence: evaluation tests.

    • M6.18.e Lock criteria/issuer/policy/evidence/subject projection on issuance request and require authorized issuer decision/signing. Evidence: issuance guard tests.

    • M6.18.f Prove course completion or author action alone cannot mint, backdate, replace, revoke, or verify credentials. Evidence: authority bypass suite.

    • M6.18.g Define expiry/renewal/revoke/replacement and propagate status without altering underlying learning evidence. Evidence: lifecycle tests. <!-- M6.18 evidence (2026-09-06) Built the 6.18.0 credential-pathway contract (apps/metis/service/src/metis/schemas/credential_pathway.py), the criteria / eligibility / issuance / lifecycle engine (services/credential_pathway_engine.py), a 44-case labelled corpus over 13 fixture families and seven dimensions (data/m6_18_credential_pathway_corpus.json, evaluated by services/credential_pathway_evaluation.py with Wilson intervals per dimension), an append-only tenant-scoped store (services/credential_pathway_service.py), registrar routes (api/credential_pathway.py), and the /credential-pathways admin workbench.

       M6.18.a — pathway schema/UI tests. `CredentialPathway` is versioned
       (semver `pathway_version`), selects one of four credential kinds, and
       carries the issuer authority, audience, standard ids, locale, groups
       and validity days. `completion_alone_is_sufficient` is a
       `Literal[False]`: there is no pathway on which the course finishing is
       the award, and constructing one raises. `IssuerAuthority` requires a
       non-empty `signing_roles`, so an issuer nobody may sign for does not
       exist. UI: 12 Playwright specs green on chromium, firefox and webkit,
       including one that reaches the workbench by following the sidebar link
       rather than by navigating straight to its URL.
      
       M6.18.b — criteria tests. Six criterion kinds cover objective and
       mastery (`objective_mastery`), course (`course_completion`), assessment
       and grade (`assessment_score`), evidence (`evidence_artefact`),
       experience (`supervised_experience`) and time (`elapsed_time`).
       `subject_revision_sha256` is REQUIRED: a criterion naming a course
       without naming which revision would be satisfied by a course that had
       since changed into something else. A criterion that states no bar is
       unconstructible — a threshold kind without a threshold and an hours
       kind without hours both raise. Groups nest one level with `all_of`,
       `any_of` and `at_least_n`; evaluation is pure, so two evaluations of
       the same inputs give the same verdicts.
      
       M6.18.c — evidence policy tests. `EVIDENCE_FOR_CRITERION` is total over
       6 criterion kinds x 5 evidence types = 30 cells, 7 accepted, validated
       at import: a gap stops the service starting rather than admitting an
       attendance record as proof of mastery. `verification_required` is a
       `Literal[True]`, freshness has a policy default and a per-criterion
       override, `withheld_from_projection` must name all three of free-text
       answers, identifier and demographics, and `review_quorum` may not
       exceed the reviewer roles that could meet it.
      
       M6.18.d — evaluation tests. Six criterion states, each reachable and
       each distinguished by a fixture: `missing` is not `not_met`, because
       reporting evidence nobody could find as a shortfall tells an issuer the
       subject failed when nothing was checked. `SubjectProjection` carries a
       salted digest and no identifier, and `EligibilityPreview` evaluates a
       current OR synthetic subject and records nothing. FIXED: `is_synthetic`
       was a declared field nothing read — it now refuses issuance, because a
       credential minted for a made-up subject would attest to learning nobody
       did. A preview of a synthetic subject that did not say so is
       unconstructible.
      
       M6.18.e — issuance guard tests. `IssuanceLock` hashes the criteria, the
       issuer, the policy, the projection and the eligibility outcome, and
       every field in it is covered by its own hash. FIXED: the lock folded in
       the moment it was taken, so no two parties could ever compute the same
       string and the approve-then-issue flow was impossible; the evaluation
       instant is now excluded from both the lock and the eligibility digest,
       while a change in the substance — evidence going stale — still moves
       the lock. The decision is HMAC-signed over the lock (32-character
       minimum, `SigningNotConfiguredError` and a 503 when unset), and nine
       refusal reasons are each reachable and separately evidenced. FIXED:
       approvals were a field on the request, so one caller could invent a
       quorum; they are now recorded acts by the approving party against one
       exact lock, and the quorum counts distinct people.
      
       M6.18.f — authority bypass suite. 21 probes, all refused, spanning all
       five verbs (mint, backdate, replace, revoke, verify), all three actors
       (course completion, author action, unauthorized role) and all four
       mechanisms (absent field, lock binding, state machine, role check). The
       request shape has no field for a status, an issue date or a credential
       id, so a caller cannot mint by asking.
      
       M6.18.g — lifecycle tests. `STATUS_AFTER` is total over 4 actions x 4
       statuses = 16 cells, 7 allowed; an action with nowhere to go refuses
       rather than proceeding. A `LifecycleEvent` carries no criterion, no
       evidence and no subject, and `alters_learning_evidence` is a
       `Literal[False]`. FIXED: `apply_lifecycle` copied the record, leaving
       its digest covering the old status, so every legitimately revoked
       credential verified as tampered; it now re-seals. A renewal must state
       an expiry later than the one it replaces, or it is not a renewal.
      
       Verification: 66 service tests (engine 38, store 16, contract 12) plus
       44/44 corpus cases matched and 21/21 authority probes refused; ruff,
       `ruff format` and `mypy --strict` clean on all five modules. Mutation
       testing applied 49 mutations across the schema, engine and store and
       the suite kills all 49. The store's own re-checks of contract
       invariants were DELETED rather than left unkillable: two places
       computing a lock is two places that can drift, so the engine exports
       `lock_for` and the store calls it. OpenAPI compared semantically: +27
       schemas, 0 removed, 0 existing schemas changed, 9 new paths. Browser:
       12 Playwright specs green on chromium, firefox and webkit against a
       fixture generated from the real service — the rendered 30-cell evidence
       matrix checked cell by cell against the service, every verb's refusal
       count, the sidebar link followed rather than the URL typed, and an
       assertion that no subject identity reaches the page; axe WCAG 2.1 AA
       runs after every one of the 36 test executions with zero serious or
       critical violations.
      

      -->

  • M6.19 Build governed training-data/evaluation feedback intake from educator decisions, aggregated learner outcomes, verifier disagreements, and media critiques. Evidence: consent, minimization, purpose, retention, contamination, licensing, deletion, lineage, and opt-out controls precede any training or fine-tuning use.
    • M6.19.a Define feedback-source registry and intake record for educator decisions, cohort-level outcomes, verifier disagreements, and media critiques with exact originating versions. Evidence: schemas.
    • M6.19.b Verify collection notice/consent or approved basis, creator/ learner/institution rights/license, purpose scope, tenant policy, and opt-out before admission. Evidence: eligibility tests.
    • M6.19.c Minimize/redact/deidentify protected answer, learner, grade, integrity, biometric/voice, source, tenant, and free-text PII with privacy thresholds. Evidence: minimization tests.
    • M6.19.d Filter quality/label validity, duplication, leakage, benchmark/ evaluation contamination, malicious/poisoned content, and subgroup representation. Evidence: quality/contamination tests.
    • M6.19.e Store immutable dataset candidate/record lineage, license/ purpose/retention/delete/opt-out state and prohibit direct training use. Evidence: dataset ledger tests.
    • M6.19.f Require separate dataset curation, governance approval, versioned manifest/splits, security/privacy/fairness review before evaluation or training. Evidence: promotion gate.
    • M6.19.g Propagate deletion/withdrawal/revocation through datasets, experiments, derived models/evaluations and record irrecoverable constraints. Evidence: lineage/withdrawal tests. <!-- M6.19 evidence (2026-09-05) apps/metis/service/src/metis/{schemas/training_intake.py, services/training_intake_{engine,service,evaluation}.py, api/training_intake.py, data/m6_19_training_intake_corpus.json}; admin /training-intake workbench. 72 service tests (engine 48, store 13, contract 11), 49/49 corpus cases, 26/26 bypass probes refused over 5 verbs x 3 actors x 4 mechanisms, 13 Playwright specs on chromium with axe clean; OpenAPI +35 schemas, +7 paths, 0 changed. Nothing that arrives is trainable: DatasetCandidate.trainable is Literal[False]; only a DatasetRelease (separate curator, quorum of 2, three cleared reviews, disjoint manifest) says True. A cohort outcome withholds >=2 buckets with a residual >= the reportable minimum, so no cell is recoverable by subtraction. Promotion re-resolves the lineage against withdrawals on file. Defects found and fixed: the minimisation matrix claimed hashed for source/tenant fields the payload does not carry; the candidate digest folded in the assembly instant (same class as M6.18's lock); one unreachable store re-check deleted; four unread vocabulary constants deleted. -->

M7 — Integrity, review, and institutional governance#

  • M7.1 Integrate Themis adjudication through its canonical contract only. Evidence: no private Themis imports.
    • M7.1.a Inventory Themis public contracts/versions/operations/events/ auth/failure modes and all current Metis integration/private imports. Evidence: integration inventory. <!-- metis-themis-integration.json: Metis↔Themis integration/boundary inventory. Themis exposes 71 public @themis/* packages; Metis integrates EXACTLY 3 (@themis/arbitration, @themis/justice, @themis/policy) through a SINGLE canonical wiring module (libs/metis/integrations/src/themis-canonical- wiring.ts): 16 typed MetisThemis* contracts (Adjudication/Appeal/AuditRecord/ EnforcementAction...) + 2 enforcement consts (METIS_THEMIS_POLICY_DOMAIN, METIS_THEMIS_ENFORCEMENT_ACTION). BOUNDARY FINDING: 3 public bare-package imports, 0 public subpath imports, 0 PRIVATE/deep imports (@themis/x/src|dist| internal) — the boundary is CLEAN (Metis reaches no Themis internals; all integration flows through the sanctioned wiring). generate-metis-themis- integration.mjs + 6 tests, wired into inventory/verify + CI. M7.1.b+ (Themis version/failure-mode contracts) are design and stay unchecked. -->
    • M7.1.b Map Metis case/evidence/subject/policy/decision/appeal fields to the public contract with protected-field and authority boundaries. Evidence: mapping tests.
    • M7.1.c Implement an injected adapter/client with version negotiation, idempotency, timeouts, redaction, audit, and unavailable/degraded state. Evidence: contract tests.
    • M7.1.d Remove/forbid private package/schema/store imports and duplicate adjudication mutation paths. Evidence: dependency-boundary gate.
    • M7.1.e Verify case submit/status/decision/appeal/revocation journey and cross-tenant/role failures. Evidence: integration E2E. <!-- M7.1.b–e evidence (2026-09-05) libs/metis/integrations/src/themis-contract-mapping.ts (b), themis-adjudication-adapter.ts (c), scripts/v1-workbenches/ themis-boundary-gate.{mjs,test.mjs} (d, wired into verify:inventory:v1-workbenches), themis-contract-mapping.spec.ts + themis-adjudication-journey.spec.ts (b/c/e). 21 specs green plus the 9 pre-existing wiring tests; gate test 5/5 and the gate is clean over 1,296 Metis-boundary files. Protected fields: a party is a salted digest and mapEvidence never reads excerpt text; mapped payloads validate against Themis's own zod inputs. Authority: a 15-cell role x verb matrix, total at module load. Adapter: version negotiation (0.1.x), idempotency by key + payload digest, injected- clock timeouts -> degraded, transport throw -> unavailable (never a fabricated case), an audit entry per call with reachedTransport. The journey walks submit -> status -> decision -> appeal -> revocation against the real in-process DisputeManagementSystem; cross-tenant and role refusals never reach the transport. Findings: Themis has no revocation state, so revocation is closed + a structured note and the mapping names it lossy; Themis exports "./*" so the gate forbids EVERY subpath import, not just /src|dist/; the gate's first run flagged apps/oshun/bff/src/themis/moderation-appeal-store.ts, which is the BFF's own Themis product surface, not Metis — scan narrowed to Metis code + Metis-specific BFF paths and the reason is stated in the gate; Themis enforces no status transitions, so Metis does (filed -> adjudicated allowed, as the existing bridge already does). -->
  • M7.2 Reconcile assessment and tutor integrity-mode vocabularies through an explicit lossless mapping. Evidence: round-trip tests.
    • M7.2.a Inventory every assessment/tutor/delivery/institution integrity mode code, definition, version, owner, behavior, and consumer. Evidence: vocabulary inventory.
    • M7.2.b Define canonical mode identities separately from context-specific projections and map exact/broader/narrower/composite/unmapped semantics. Evidence: mapping table.
    • M7.2.c Represent untranslatable fields/constraints explicitly and reject lossy defaults or “closest mode” execution. Evidence: loss tests.
    • M7.2.d Round-trip each supported mode/version through assessment context, tutor config/session, API/events, and UI labels. Evidence: goldens.
    • M7.2.e Version/deprecate mappings with compatibility windows and preserve historical session meaning. Evidence: lifecycle tests. <!-- M7.2 evidence (2026-09-05) apps/metis/service/src/metis/schemas/integrity_mode_registry.py (a, b, e), services/integrity_mode_{engine,goldens}.py (c, d), api/ integrity_mode.py (3 routes), data/m7_2_integrity_mode_goldens.json, tests/test_integrity_mode_registry.py (17) plus 35 neighbouring tests green; OpenAPI +7 schemas, +3 paths, 0 changed. The inventory found SEVEN mode vocabularies on FOUR axes (stakes, posture, stance, style): practice is a stakes level in four of them and a tutor stance in a fifth; proctored is a stakes level in the interlock and a posture in the item pool. Canonical identities live per axis, separate from every projection; the 37-projection table is total at import and every projection states exact/broader/narrower/composite/unmapped with its loss named (Themis-style review and collaborative are unmapped and say why; interlock proctored is composite over stakes+posture). The reverse direction refuses with candidates rather than choosing a closest mode. Loss tests fixed three production paths that defaulted: normalize_integrity_mode (unknown -> teach), derive_integrity_mode_ from_tutor_style (everything -> teach), _build_policy_binding (unknown assessment_type -> assessed), plus the web label fallback ('Teach'); seven lossy defaults are recorded against their vocabularies. Goldens: 37 round trips (27 lossless, 10 with the loss named) and 85 context expressions (31 named unexpressible — e.g. posture has no assessment- context vocabulary) generated from the registry and drift-checked. Lifecycle: the stance code practice is deprecated to practice_coach under a 7.2.0..7.4.0 window; a session recorded under 7.2.0 resolves to practice_coach while the current mapping is inside the window and is refused with the window named once it moves past. Finding: the registry's own import check caught my first table (style:direct had no exact code) — a canonical nobody can say is a word, not a mode. -->
  • M7.3 Build integrity signal review with model version, evidence, confidence, uncertainty, and protected-data controls. Evidence: reviewer journey.
    • M7.3.a Define signal/case projection with exact learner/attempt/item refs, source/collector/model/rule/config versions, raw outputs, evidence/chain, confidence/uncertainty, and state. Evidence: schema.
    • M7.3.b Build permission-filtered queue/detail with priority, deadlines, duplicate/correlation, evidence freshness, and explicit classifier-not- conclusion language. Evidence: review UI tests.
    • M7.3.c Render timeline/media/session/device/similarity/authorship evidence through field-level redaction, secure viewer, and access audit. Evidence: protected-data tests.
    • M7.3.d Support request-evidence/dismiss/escalate/open-case actions with reason/expected version and no automatic grade/consequence. Evidence: action tests.
    • M7.3.e Verify reviewer journey for positive/false-positive/conflicting/ insufficient/appealed/cross-tenant cases accessibly. Evidence: E2E. <!-- M7.3 evidence (2026-09-05) apps/metis/service/src/metis/{schemas/integrity_signal_review.py, services/integrity_signal_review_{engine,evaluation,service}.py, api/integrity_signal_review.py} (7 routes); admin /integrity-signals workbench + e2e. 17 tests in tests/test_integrity_signal_review.py plus both duplicate-model-name guards green; 9 Playwright specs on chromium with axe clean; OpenAPI +21 schemas, +7 paths, 0 changed. (a) the projection pins learner (salted digest), attempt and item to exact revisions and carries collector/model/rule/config versions, raw classifier outputs, a confidence with interval + basis (a calibrated basis must name its sample), a hash-chained evidence trail, and state; is_conclusion is Literal[False] and the only language is classifier_indicates — a summary containing verdict words is unconstructible. (b) the queue derives rank (severity, then narrower interval, then deadline), deadline by severity, overdue, duplicates (same attempt revision+kind+source+config; earliest is original), correlation, and freshness; it carries the digest and no field for a person; filtered by tenant and, for a learner, to their own. (c) a total 15x3 field-group x audience visibility matrix; learner-facing excerpt text is COMPUTED by running the M6.11 redactor plus engine- owned residue detectors (phone, handle) and withheld on residue — a FieldDecision with residue and any treatment but withheld is unconstructible; every view appends an access-audit entry naming the fields shown and the view digest. (d) four verbs over a total 4x6 table, expected_version refused when stale, the original projection returned on refusal, and no field on the action or its record for a grade, score change or consequence. (e) six journeys walked against the engine; cross-tenant and learner action refused before the transport. Findings: redactionState on excerpts was set to learner_safe by its only writer while copying raw answer text, and the learner-facing excerpt response returns text regardless of it — labelled, never enforced; my own field no_automatic_consequence tripped my consequence-word scan (renamed carries_effect); QueueItem collided with M6.11's (renamed SignalQueueItem — the duplicate-name guard scans the schema dir at test time and caught it mid full-suite run). -->
  • M7.4 Build decision and appeal workflows with separation of duties. Evidence: original reviewer cannot decide a restricted appeal. <!-- Verified 2026-09-05: schemas/appeal_workflow.py 7.4.0 + services/appeal_workflow_{engine,evaluation}.py + api/appeal_workflow.py (GET /appeal-workflows/evaluation); tests/test_appeal_workflow.py 22 tests + 2 duplicate-name guards = 24 passed; mypy --strict clean on 4 files; OpenAPI +1 path, +4 schemas, 0 changed. (a) 6 case and 7 appeal states over transition tables validated at import; 7 roles; a total 7x4 evidence-rights matrix; deadlines by standard; a stay extends the deadline by exactly its duration; every state change writes an audit entry (actor, role, versions) through one _move. (b) eight refusal reasons, each reached by its own input; a decision names all six guards passed. (c) only the learner, or a representative holding the learner's unexpired grant, files; the new-evidence ground carries a chained link; accessibility support opens a stay; the status view has no field for raw outputs. (d) MAY_DECIDE_APPEAL 2x7: restricted refuses the original detector owner, reviewer and decider (the original reviewer cannot decide a restricted appeal even holding the appeal_decider role); the original decider never may; a restricted quorum is 2 DISTINCT independent signatories. (e) six outcomes each recorded from its state with a CorrectionTrigger as a SEPARATE record (performed is Literal[False], required_role integrity_lead); the BFF projection is lossy for 4 of 6 and says so. Findings: BFF decideAppeal has no separation-of-duties guard; my generic transition_case could enter decided/appealed/stayed with no decision, filing or stay record (0 callers — the caller scan caught it; now UNGUARDED_TARGETS = evidence_complete, closed and it refuses to leave a stay); the outcome digest was first computed over a hand-built dict (now AppealOutcomeBody.seal()); DecisionRequest/ DecisionRecord collided with M6.11 (renamed Case*). -->
    • M7.4.a Define decision/appeal state machines, roles, evidence rights, standards/policies, deadlines, stays, outcomes, remedies, and audit. Evidence: workflow schemas.
    • M7.4.b Require eligible independent decision maker, conflict check, evidence completeness, learner explanation, and expected case version. Evidence: decision guards.
    • M7.4.c Let learner/authorized representative file appeal, grounds, new evidence, accessibility support, and receive status safely. Evidence: appeal intake tests.
    • M7.4.d Enforce restricted appeal independence/quorum so original detector/reviewer/decider cannot decide where prohibited. Evidence: separation tests.
    • M7.4.e Record appeal affirm/modify/overturn/remand/withdraw/expire and trigger grade/consequence correction separately. Evidence: outcome tests.
  • M7.5 Build immutable integrity audit and learner-visible explanation projection. Evidence: role-based redaction tests. <!-- Verified 2026-09-05: schemas/integrity_audit.py 7.5.0 + services/integrity_audit_{engine,service,evaluation}.py + api/integrity_audit.py (11 routes under /integrity-audit); tests/test_integrity_audit.py 25 tests + 2 duplicate-name guards = 27 passed; mypy --strict clean on 6 files; OpenAPI +11 paths, +44 schemas, 0 changed. (a) ten event kinds with actor, authority basis (total over roles; a grant or authorization must name its record), subject versions (non-mutating kinds leave the version) and a per-subject hash chain whose event digest covers the PAYLOAD'S DIGEST; a case chain's opening link names a link the signal chain contains; four faults each found at its link (gap, relink, cross-subject, on-disk tamper). (b) a total 8-section x 6-audience policy: raw evidence, classifier features, other learners, staff identity and security method are withheld from learner and representative, masked (count/roles/flag) for reviewers, and shown to forensic access only under a stored authorization naming examiner, subject and expiry; a withheld section is empty AND named; the actor digest is absent when staff identity is withheld. (c) seven ordered sections (allegation, policy, evidence+limitations, decision, consequence, appeal rights + deadlines, support), each with its provenance, plus a locale and accessibility profile. (d) the projector takes the case, its chain, the decision, appeal and outcome — there is no parameter for a signal projection; the catalogue is total over locales x keys, pre-decision keys are scanned for accusatory stems at import, an unsupported locale is refused with the supported list, and the same records give the same view digest under a different clock. (e) tamper -> digest_mismatch at the link; a legal hold refuses the whole purge and a release is its own record; a purge drops payloads only and is itself a link, the chain verifies after; the export is sealed and carries the correction history; cross-tenant and other-learner reads refused; every read appends an access link. Findings: the opening link first named the signal chain's MOVING head; learner digest, signal id and the opening link began in the purgeable payload (moved to the header); RedactedPayload collided with M6.14; PLACEHOLDERS as an identifier would have tripped the stub scan (TEMPLATE_FIELDS). -->
    • M7.5.a Record append-only collection/access/change/signal/case/evidence/ decision/appeal/notification/correction events with actor/authority/ versions/hash chain. Evidence: audit schema tests.
    • M7.5.b Protect raw evidence/classifier features/other learners/reviewer identity/security methods while retaining authorized forensic access. Evidence: field policy.
    • M7.5.c Define learner-visible explanation for allegation/category, applicable policy, evidence summary/limitations, decision/reason, consequence, appeal rights/deadlines, and support. Evidence: projection schema.
    • M7.5.d Generate explanations from canonical decisions, not model output, with locale/accessibility and no accusation before decision. Evidence: projection tests.
    • M7.5.e Verify tamper detection, retention/legal hold, role/tenant redaction, export, and correction history. Evidence: audit tests.
  • M7.6 Build content review templates for SME, pedagogy, accessibility, fairness, legal/rights, and publisher signoff. Evidence: required-role matrix. <!-- Verified 2026-09-05: schemas/content_review_template.py 7.6.0 + services/content_review_template_{engine,evaluation}.py + api/content_review_template.py (5 routes under /content-review-templates); tests/test_content_review_template.py 20 tests + 2 duplicate-name guards = 22 passed; mypy --strict clean on 4 files; OpenAPI +5 paths, +30 schemas, 0 changed. (a) versioned templates applying by content kind, risk, locale and delivery target; each of the six dimensions carries criteria, evidence kind, roles, qualifications, quorum, separation, sequence, expiry and gate floors; two platform templates cover every (kind, risk) exactly once (checked at import); the sequence is validated acyclic and the publisher signs after every other dimension; the template is sealed by a digest. (b) 18 criteria over 6 dimensions, each code prefixed by its dimension, so a criterion cannot be lent: a DimensionDecision citing another dimension's criterion, and a signoff carrying another dimension's decision, are both unconstructible. (c) the required-role matrix is the only way into a dimension — every role x every dimension walked; author, wrong locale, missing qualification, a signer of an excluded dimension and a repeat signature are each refused by their own reason; a quorum counts DISTINCT people (one person signing twice leaves the dimension pending); all seven signoff refusal reasons are reached by their own input. (d) a signoff binds content, revision, evidence, the gate-results digest and the template version; a change invalidates only the dimensions CHANGE_AFFECTS names (total over 8 change kinds; every kind reaches the publisher) and carries the rest forward as CarriedSignoff records naming the original, both revisions and the change kinds — a carried record for an affected dimension is unconstructible; expiry and a superseded template invalidate too. (e) 400 generated overlays applied to a platform template: no overlay the type can express lowered a quorum, dropped a criterion or a qualification, relaxed a locale/separation/sequence, lengthened an expiry or lowered a gate floor; OverlayAddition has no removal field, and six weakened EffectiveTemplates were each refused by its validator. Findings: an overlay CAN express a longer expiry, which the type cannot forbid — apply_overlay refuses it (the property probe treats a refusal as not-a-weakening); the platform's institution reviewer role is deliberately absent from the required-role matrix so it can only ever be an addition. -->
    • M7.6.a Define versioned review template applicability by content/item/ course/risk/locale/target with dimension, checklist, evidence, role, quorum, sequence, expiry, and code floors. Evidence: schema.
    • M7.6.b Define SME correctness/source, pedagogy/alignment, accessibility, fairness/cultural, legal/rights, and publisher/target criteria as separate decisions. Evidence: dimension matrices.
    • M7.6.c Resolve required roles/qualifications/locale/independence and prohibit one signoff from satisfying another dimension. Evidence: role tests.
    • M7.6.d Bind signoffs to exact revisions/evidence/gates/template and invalidate only affected approvals on change. Evidence: freshness tests.
    • M7.6.e Test institution overlays can add reviewers/criteria but not remove platform requirements. Evidence: template property tests.
  • M7.7 Build anchored comments on course, lesson, block, objective, rubric, item, response, stimulus, and media. Evidence: anchor survival tests. <!-- Verified 2026-09-05: schemas/anchored_comment.py 7.7.0 + services/anchored_comment_{engine,service,evaluation}.py + api/anchored_comment.py (11 routes under /anchored-comments) + admin /comments + e2e; tests/test_anchored_comment.py 28 tests + 2 duplicate-name guards = 30 passed; 5 chromium e2e specs passed (axe after each, incl. 390px with no page-level horizontal scroll); admin vitest 1 passed; mypy --strict clean on 5 files; OpenAPI +11 paths, +35 schemas, 0 changed. (a) 9 subject kinds x 7 anchor kinds as a total applicability table; a discriminated anchor union (entity, field, text with quote+context+offset, region, timecode, item option, rubric cell); a thread names the exact subject revision, is a tree whose replies must name a comment in it, and carries visibility, mentions bounded by the audience, typed attachments and a status. (b) six verbs over a total 6x2 table with expected_version, an edit only by its author that keeps the digest of what it replaced, and a notification plan where a mute removes the row rather than silencing it; every read gives a position and an anchor DESCRIPTION for a reader who cannot see the highlight; all 11 refusal reasons reached by their own input. (c) each anchor kind driven through every status its kind can reach (18 resolutions): entity/region/timecode/item option are exact-or-orphaned, field also shifts by rename, rubric cell also goes ambiguous, text reaches all four; the resolution always carries the ORIGINAL and an unresolved anchor keeps pointing at it while asking a person. (d) a scan of all 27 comment types' 148 fields against 11 mutation words matched nothing; excerpting a learner's response, live assessment material or a private source is refused by a policy over subject x anchor, a response is never discussed above the reviewers, and a property the reader cannot read cannot be anchored to. Findings: the offset-first text rule matched a DIFFERENT run of the same words at the old offset (context now outranks the offset); an edit built an unvalidated dict into a validated field; the store hashed a clock it serialised differently on the way back in; an inapplicable anchor 500d through the model validator (now a refusal); learner_response_excerpt was unreachable because the applicability table excluded text on a response — a protection that cannot fire is decoration, so the table now admits it and the validator asserts the pair exists. -->
    • M7.7.a Define comment/thread/reply, exact subject revision, typed entity/ field/text/region/timecode/item-option/rubric anchor, visibility, mentions, attachments, and status. Evidence: schema.
    • M7.7.b Implement accessible create/reply/edit/resolve/reopen/navigate with object/property permissions and notification controls. Evidence: UI/API tests.
    • M7.7.c Map stable entity anchors across revisions and text/media anchors as exact/shifted/ambiguous/orphaned while preserving originals. Evidence: survival tests.
    • M7.7.d Prevent comments from changing content/answer/grade/approval and protect learner/assessment/private-source anchors. Evidence: boundary tests.
  • M7.8 Build reviewer conflict-of-interest declarations. Evidence: incompatible conflicts block signoff. <!-- Verified 2026-09-05: schemas/reviewer_conflict.py 7.8.0 + services/reviewer_conflict_{engine,evaluation}.py + api/reviewer_conflict.py (6 routes under /reviewer-conflicts); tests/test_reviewer_conflict.py 27 tests + 4 duplicate-name guards + the OpenAPI contract = 73 passed; mypy --strict clean on 4 files; OpenAPI +6 paths, +29 schemas, 0 changed, 0 removed. (a) 8 conflict types x 4 scopes x 7 relationship bases, self-declared or raised by a detector whose signal is_conclusion=False and speaks as "detector_indicates"; a NIL RETURN is itself a declaration, so silence and "nothing to declare" are different states; 7 statuses over a total transition table with terminal dismissed/withdrawn. (b) an assignment with no current declaration is refused, not assumed clean: four coverage gaps (missing, stale, does-not-cover, unsure-unresolved) each map to their own refusal; freshness tightens with risk (180/180/90/30 days); an unsure reviewer is REFERRED and cannot answer their own referral, and an unanswered referral keeps blocking. (c) a total 6x8 policy matrix reaching all three dispositions, tightened by risk and then by an institution overlay; 200 generated overlays never loosened a cell at any risk, and an overlay that names a looser cell is unconstructible. (d) all 7 refusal reasons reached by their own scenario; a disqualifying conflict blocks assignment AND signoff AND the quorum (3 candidates, 1 disqualified, quorum of 3 not met); a manageable one blocks until a mitigation is approved by SOMEBODY ELSE; a recusal blocks its own dimension and no other; an evaluation whose verdicts contradict its refusals is unconstructible. (e) a total 9x5 detail-visibility matrix: an author learns only that a declaration exists, a panel sees type and identity but never basis, evidence or dates; a withheld field is ABSENT from the view, not blanked; every look leaves an entry accounting for all 9 fields; a withdrawal does not delete while an appeal window is open or the 365-day minimum stands. Findings: my ViewRequest silently renamed M7.3's schema to metis__api__integrity_signal_review__ViewRequest (caught by diffing the schema set: -1) — renamed ConflictViewRequest; VisibilityProbe collided with M7.5's (4th collision in 4 cells); 10 orphan vocabularies became import-time totality checks and producibility tests, and one dead constant was deleted. -->
    • M7.8.a Define conflict types/scope/subject/reviewer relationship, self-declaration, detected signal, confidentiality, evidence, dates, and status. Evidence: declaration schema.
    • M7.8.b Require declaration/update before assignment/decision and provide safe “unsure/request review” path. Evidence: workflow tests.
    • M7.8.c Evaluate conflict policy by review dimension/risk/institution as disqualifying, manageable with mitigation, or permitted. Evidence: policy matrix.
    • M7.8.d Block assignment/signoff/quorum for incompatible conflicts and record reassignment/recusal/mitigation approval. Evidence: enforcement tests.
    • M7.8.e Protect conflict details from unauthorized parties while auditing access and preserving appeal/retention needs. Evidence: privacy tests.
  • M7.9 Build institution-specific policy overlays with platform floors. Evidence: tenants can tighten but never loosen. <!-- Verified 2026-09-05: schemas/institution_overlay.py 7.9.0 + services/institution_overlay_{engine,evaluation}.py + api/institution_overlay.py (6 routes under /institution-overlays); tests/test_institution_overlay.py 22 tests; mypy --strict clean on 4 files; OpenAPI +6 paths, +26 schemas, 0 changed, 0 removed. (a) 17 typed settings covering all 12 families, each declaring its value kind and WHICH WAY IS TIGHTER (higher, lower, superset, subset, on, fixed); a direction that cannot describe a shape is unconstructible, and the import validator refuses the module if any direction, kind or family has no setting — it caught that nothing used superset until a required-formats setting was added. (b) five scopes stack from the platform floor; a resolved value records the level that set it and every level that contributed, in precedence order; a level that loosens produces a conflict naming both levels and both values, and the tighter value stands. (c) 300 generated stacks — a random value per setting at every scope — never resolved looser than any contributing level or than the platform floor; every setting was also walked directly with a tighter, an equal and a looser value, from a TIGHTENED base so a looser value always exists. (d) the preview counts affected items from the inventory across all six categories and is sealed by a digest; an activation is refused when it is unreviewed, approved over a different preview, in the wrong status, backdated, or would loosen something. (e) statuses over a total transition table with terminal superseded/rolled_back; a history is append-only with one active version; a rollback is a NEW draft carrying the old values and naming both what it supersedes and what it restores; revalidation lists what needs a recheck and names the items nothing touched. Findings: the preview and the activation check resolved a stack that resolve then filtered the candidate OUT of (not active yet), so a preview would have reported no impact — fixed with as_if_active; a level stacked twice (live version + candidate) recorded itself twice as a contributor; approver_is_proposer was an unreachable refusal because the type already forbids it (deleted, and proved at the type); my ResolveRequest silently renamed M7.6's OpenAPI schema — the fifth name collision in five cells, caught by diffing removals. -->
    • M7.9.a Define typed versioned institution overlay for content/source/ review/accessibility/integrity/provider/retention/export/assessment/ grade/credential/tutor settings. Evidence: schema.
    • M7.9.b Resolve platform/region/institution/program/course precedence with exact inheritance/merge/conflict semantics. Evidence: resolution tests.
    • M7.9.c Prove field-level monotonic tightening of safety/privacy/rights/ accessibility/integrity/security/authority floors. Evidence: property tests.
    • M7.9.d Preview affected users/content/builds/attempts/integrations/ schedules and require authorized review before activation. Evidence: impact tests.
    • M7.9.e Version/effective-date/rollback overlays and revalidate dependent state without rewriting history. Evidence: lifecycle tests.
  • M7.10 Build gradebook visibility and correction review. Evidence: correction events are append-only and authorized. <!-- Verified 2026-09-05: schemas/gradebook_correction.py 7.10.0 + services/gradebook_correction_{engine,evaluation}.py + api/gradebook_correction.py (7 routes under /gradebook-corrections) + admin /grade-corrections + e2e; tests/test_gradebook_correction.py 23 tests; 5 chromium e2e specs passed (axe after each, incl. 390px); admin vitest 1 passed; mypy --strict clean on 4 files; OpenAPI +7 paths, +26 schemas, 0 changed, 0 removed. (a) a grade IS its events — there is no mutable value field and the current value is derived; a total 9 field-group x 5 audience x 2 release-state matrix (90 cells): a learner and a guardian see no value before release, an assessor marks blind in BOTH states, and a projection carrying a withheld field is unconstructible. (b) a request binds the exact event it disputes, its proposed value and kind, evidence, the requester and the counted downstream impact; the queue derives every delta and puts a correction already published downstream first. (c) all ten refusal reasons reached by their own scenario: the scorer cannot review, the requester cannot decide, the reviewer who moved it on cannot approve, a value off the scale or unchanged, a stale version, a wrong status, a superseded target, a moved item revision, and no notice to the learner. (d) a total 3x5 decision table; an approval APPENDS an event and every earlier one is byte-identical afterwards, so a view shown before the correction is still reproducible via history_as_of. (e) recomputation is a pure function of its inputs (same digest under reordering) and names the grades it excluded rather than treating them as zero; propagation derives its idempotency key from the payload, not the clock, so the second send reports unchanged, and prior views and appeals are carried on the result. Findings: my "a projection cannot carry a withheld field" test was vacuous for a teacher after release, who is withheld nothing — moved to an audience that is; DOWNSTREAM_SYSTEMS was an orphan vocabulary until each system was walked as an impact and a propagation target. -->
    • M7.10.a Build role/field/time-window filtered grade/current/history/ evidence projection for learner/guardian/teacher/assessor/admin. Evidence: visibility tests.
    • M7.10.b Build correction request/detail queue with target grade revision, proposed value/status, reason/evidence, downstream impact, and requester. Evidence: review schema/UI.
    • M7.10.c Enforce scorer/reviewer/approver separation, scale/range/policy/ attempt/item freshness, expected version, and learner notice. Evidence: guard tests.
    • M7.10.d Record approve/reject/request-changes as immutable decision and approved correction as append-only grade event. Evidence: history tests.
    • M7.10.e Recompute aggregates/LTI/OneRoster/credentials idempotently and preserve prior views/appeals. Evidence: propagation tests.
  • M7.11 Build BYOM quarantine, evaluation, approval, kill switch, and abuse monitoring into operator review. Evidence: fail-closed end-to-end journey. <!-- Verified 2026-09-05: schemas/provider_model_admission.py 7.11.0 + services/provider_model_admission_{engine,evaluation}.py + api/provider_model_admission.py (6 routes under /provider-model-admissions); tests/test_provider_model_admission.py 22 tests + the duplicate-name guard = 23 passed; mypy --strict clean on 4 files; OpenAPI +6 paths, +27 schemas, 0 changed, 0 removed, 0 qualified. NOTE: the estate's existing BYOM (libs/metis/integrations/src/byom, apps/oshun/bff/.../metis-byom- decision-store.ts) is bring-your-own-MATERIAL; this cell is bring-your-own-MODEL and shares no vocabulary with it. (a) intake records provider, endpoint (TLS-only, region-pinned, carrying a model FINGERPRINT), credential BY REFERENCE, ownership, licence, security review, data use, residency, cost envelope and capabilities, and starts quarantined; a scan of all 20 models' 134 field names against 8 secret-bearing words matched nothing; provider_default data use and a residency that disagrees with the endpoint region are unconstructible. (b) 8 probe kinds, each accounted for in every run — a probe that did not run is recorded not_run and HOLDS the candidate whatever its kind, because the blocking/advisory split is about a failure and a measurement not taken is not a pass; fixtures must be synthetic and approved, and a gap in fixture coverage holds quarantine too. (c) five named roles, five DISTINCT people, every approval over the same evaluation run and the same fingerprint; a scope by tenant, course, use, data class and region with an expiry, and identified_learner data is unconstructible in a scope. All 12 refusal reasons reached, each carrying fallback=not_configured. (d) a kill is scoped (candidate/tenant/provider/platform) with a blast radius counted from what is admitted and running, a credential revoke, and an in-flight behaviour a platform kill cannot set to drain; restore gates are total over the five bases and every gate is required by some basis. (e) 9 abuse signals, 4 of which suspend before a person looks; a triage decision cannot dismiss or merely watch one of those, and must name the signals and kinds that drove it. (f) the nine-leg journey — direct endpoint, secret substitution, model substitution, stale approval, cross tenant, outage, abuse, kill, recovery — every leg refused, every one falling back to not_configured. Findings: an advisory probe that did not RUN was treated as no obstacle until the outage leg exposed it; AdmissionRefusal collided with M6.19's (sixth name collision, caught by the removal diff) — renamed ModelAdmissionRefusal, and the spec now carries zero qualified names. -->
    • M7.11.a Intake provider/model/endpoint/config credentials by secret ref, ownership/license/security/data-use/residency/cost/capabilities and quarantined state. Evidence: intake tests.
    • M7.11.b Probe/scan/benchmark with synthetic approved fixtures for contract, safety, grounding, answer leakage, privacy, latency/cost, availability, and model/version drift. Evidence: evaluation suite.
    • M7.11.c Require named security/privacy/safety/education/operator approval and scope admission by tenant/course/use/data/region before enablement. Evidence: approval tests.
    • M7.11.d Implement scoped disable/kill with session/job behavior, credential revoke, fallback/not-configured state, blast radius, and restore gates. Evidence: kill-switch tests.
    • M7.11.e Monitor prompt injection/exfiltration, harmful/answer-leaking output, anomalous spend/rate/data, endpoint drift, and complaints with explainable human triage. Evidence: abuse monitoring tests.
    • M7.11.f Run direct endpoint/secret/model substitution, stale approval, cross-tenant, outage, abuse, kill, and recovery journey. Evidence: fail-closed E2E.
  • M7.12 Build review SLA, escalation, workload, and notification views. Evidence: quiet hours never suppress urgent integrity events. <!-- Verified 2026-09-05: schemas/review_sla.py 7.12.0 + services/review_sla_{engine,evaluation}.py + api/review_sla.py (6 routes under /review-sla) + admin /review-sla + e2e; tests/test_review_sla.py 20 tests; 5 chromium e2e specs passed (axe after each, incl. 390px); admin vitest 1 passed; mypy --strict clean on 4 files; OpenAPI +6 paths, +29 schemas, 0 changed, 0 removed, 0 qualified. (a) priority is DERIVED from review type x risk over a total 6x4 table (a high-risk integrity review is urgent whoever files it); a business calendar in a real tzdata zone with working days, hours and holidays; deadlines are WALKED through working hours, so Friday 15:00 to Tuesday 11:00 across a bank holiday is 4 hours and a deadline always lands inside a working day; pauses carry a reason and are subtracted, so time spent waiting on somebody else is not time the reviewer was late; an escalation path must climb and start at the deadline. (b) workload excludes blocked and finished tasks, derives utilisation and reports none when capacity is zero; counts by role, locale or institution are suppressed below 5; assignment picks the least-loaded qualified locale-matched reviewer with the digest as a reproducible tie-break and records what it considered; all five assignment refusals reached. (c) four audiences; a summary reader gets counts and NO titles, and a view carrying a title it withholds — or withholding one it shows — is unconstructible. (d/e) quiet hours are evaluated in the recipient's own zone with zoneinfo: the same instant is night in London and the working day in New York and Kolkata, and the window holds either side of the UK's autumn clock change (asserted by the differing UTC offsets). A ratified urgent integrity or safety event is sent THROUGH quiet hours; the type makes a suppressed or deferred urgent notification unconstructible, and an override without a ratification equally so. A deferral names when it will be sent, and the idempotency key is derived from the payload, so a retry is one send. Findings: my warned state was unreachable in the first probe set (the clock never sat between 75% and 100% of the target); the New York probe conflated "urgent" with "quiet" and proved neither until split. Known unrelated reds in the suite: 5 pre-existing failures in cells this work never touched — test_evidence_bundle_helpers and test_item_assembly_eligibility (stale SimpleNamespace doubles missing maturity/scalar that exposure_control_service now reads) and test_item_difficulty_helpers (a fixture expecting 64 against a computed 640). -->
    • M7.12.a Define review-type/risk/priority SLA clocks, business calendars, pauses, deadlines, breach thresholds, owners, and escalation paths. Evidence: SLA schema.
    • M7.12.b Compute queue age/due/breach/workload/capacity by qualified role/ locale/institution with privacy-safe counts and fair assignment. Evidence: fake-clock/load tests.
    • M7.12.c Build accessible queue/calendar/workload views with filters, current/stale/blocked states, reassignment, and no sensitive titles in unauthorized summaries. Evidence: UI tests.
    • M7.12.d Route reminders/escalations over configured channels/quiet hours/ substitutes/on-call with idempotency and acknowledgement. Evidence: notification tests.
    • M7.12.e Override quiet hours only for ratified urgent integrity/safety events, record reason, and test timezones/DST/failures. Evidence: escalation tests.
  • M7.13 Build human signoff gates that automation cannot complete. Evidence: system actor refusal. <!-- Verified 2026-09-05: schemas/human_signoff_gate.py 7.13.0 + services/human_signoff_gate_{engine,evaluation,service}.py + api/human_signoff_gate.py (5 routes under /human-signoff-gates); tests/test_human_signoff_gate.py 57 tests; ruff + mypy --strict clean on 5 source files and the test; OpenAPI +5 paths, +17 schemas, 0 removed, 0 qualified. The whole cell rests on one property: a HumanSignoff's actor field is annotated HumanActor, not a union, so a machine signature is not merely unauthorised but UNCONSTRUCTIBLE — asserted at import, walked over all 8 non-human kinds plus an impersonator and a delegate, and checked again in the published contract (the actor is a single $ref, not an anyOf). (a) 6 gate kinds over a total requirements table (eligible roles, quorum, step-up), a gate binds its exact subject revision and one digest over all its evidence, and delegable is a Literal[False]. (b) an actor carries an interactive authentication (method, session, second factor, 15-minute step-up freshness); the API builds it from the gateway's session headers and there is NO actor field in any request body. (c) refusal at three layers, each recorded by name in the report — the type (4 probes), the engine (10) and the store (1, which actually writes to a temp root and reads the directory back empty). (d) 15 adversarial probes, all 15 refusal reasons reached: batch replay, subject moved, evidence swapped, one person twice, expired, ineligible role, undeclared conflict, stale step-up, rewritten policy, copied record, two mass-assignment payloads, a frozen-record mutation, a service account in a signoff, and the store write. (e) a material change invalidates the gate and NAMES the signatures it preserves; all 4 change kinds come from a real detection. Findings: 3 public helpers had 0 callers (deleted); policy_changed was unreachable from the detector, so a gate now records the digest of the requirements row it opened under and a rewritten policy is caught with no notification (this also added the gate_policy_changed refusal, named apart from the material-change kind so a log never has to guess which vocabulary a value came from); my first API took the existing signatures FROM THE CALLER, so a signer could hide their own earlier signature and be counted twice — the store holds them now; and GateOutcome collided with best_of_n's, silently renaming both in the published contract — visible only in the removal diff, and renamed to SignoffGateOutcome. Known unrelated reds: the same 5 pre-existing failures recorded under M7.12. -->
    • M7.13.a Define human-only gate kinds, eligible named roles/qualifications/ independence/quorum, exact subject/evidence, decision fields, expiry, and nondelegability. Evidence: gate schema.
    • M7.13.b Require interactive authenticated human action with step-up where needed, conflict declaration, rationale, and unchanged subject/evidence. Evidence: signoff tests.
    • M7.13.c Reject system/service/agent/model/provider/import/migration actor types and impersonated/delegated tokens at API/domain/store layers. Evidence: system actor refusal tests.
    • M7.13.d Prevent mass assignment, event replay, direct state mutation, copied signoff, stale approval, and flag/policy bypass. Evidence: adversarial tests.
    • M7.13.e Invalidate on material change and preserve original human decision/audit. Evidence: lifecycle tests.
  • M7.14 Build review evidence export suitable for institutional audit. Evidence: stable schema, access control, and tamper hash. <!-- Verified 2026-09-05: schemas/review_evidence_export.py 7.14.0 + services/review_evidence_export_{engine,validator,evaluation}.py + api/review_evidence_export.py (5 routes under /review-evidence-exports); tests/test_review_evidence_export.py 55 tests; ruff + mypy --strict clean on 5 source files and the test; OpenAPI +5 paths, +20 schemas, 0 removed, 0 qualified. The claim the cell rests on is that the checks read THE BYTES: a minimisation policy is verified by searching the produced files for the raw value, a checksum by re-hashing the file, and a tamper by handing the edited bundle to a validator that does not import the builder. (a) a manifest totalling all 9 sections — an absent one says whether it was barred by the purpose or simply not asked for — with a redaction ledger and the audit section that no scope may drop. (b) 5 purposes x 6 field classes = 30 minimisation cells, all 4 treatments used; every one of the 30 is probed against the bytes and the walk is non-vacuous (12 kept values ARE found, 18 removed ones are not); a pseudonym counts as removal, is stable within an export and disjoint under a different key; role x purpose and section x purpose are both total, and a learner reaches only their own subject access. (c) byte-identical rebuilds, sorted keys, RFC 4180 CRLF with every field quoted, one UTC spelling of an instant, and all 6 formula leaders (= + - @ TAB CR) guarded by a PREFIX so the auditor still sees what the field said; 402 rows split into 3 numbered parts sharing a header. (d) HMAC-SHA256 over the canonical manifest digest, per-file SHA-256 over the encoded bytes, verification instructions generated from the constants in use, and provenance naming actor, role, time and commit; with no key the build is REFUSED rather than shipped unsigned. (e) 11 tamper probes, 11 of 11 detected, every validation finding reached — and where the edit is to the manifest the probe RE-SIGNS it, because an attacker who can rewrite a manifest can also recompute a signature. Findings: the first injection probe compared against a hand-quoted string and reported a false negative on the payload containing a quote (the guard was fine, the probe was not — it now reads the cell back with a csv reader); the first minimisation probe treated a pseudonym as a disclosure, so a leaked raw value alongside a pseudonym would have passed; and the engine carried a duplicate file-name guard that ExportFile already enforces and no test could reach, so it was deleted rather than kept as decoration. The builder and the validator share no code, so the two markers they agree on are asserted equal in a test. Known unrelated reds: the same 5 pre-existing failures recorded under M7.12. -->
    • M7.14.a Define versioned export manifest for subjects/revisions, templates/gates, measurements/evidence, comments, assignments/conflicts, decisions/appeals/signoffs, actors/roles, audit, and redactions. Evidence: schema.
    • M7.14.b Authorize export purpose/scope/institution/course/time/subjects/ fields and apply learner/answer/integrity/source/privacy minimization. Evidence: access tests.
    • M7.14.c Produce deterministic JSON/CSV/PDF or ratified formats with stable IDs/times/locales, checksums, schema/version metadata, and no spreadsheet/formula/archive injection. Evidence: export tests.
    • M7.14.d Sign/hash manifest/files and include verification instructions, missing/redacted/external refs, generation actor/time/commit, and audit. Evidence: tamper tests.
    • M7.14.e Round-trip through independent validator and test two-tenant/ role/large export/retention/expiry behavior. Evidence: audit export suite.

M8 — Persistence, APIs, jobs, and tenancy#

  • M8.1 Implement missing workbench models using the ratified database. Evidence: migration review and real-DB apply. <!-- Verified 2026-09-06: a-d closed. The gap map (a) is computed from the code and says 98 ratified contracts against 106 tables, with 36 cells writing only to the file store and 23 with no storage at all. The design (b) proposes 81 tables for those and refuses to design 158 wire formats and computed reports. The migrations (c) carry all 81 as readable five-phase plans and freeze the 14 whose design needed no human decision into alembic revision 044. The apply (d) ran that chain against a real PostgreSQL 16.14 database from clean and from revision 043, and checked the constraints and the query plans by asking the server. The 67 aggregates whose design carries an open question have a plan and no licence: a table nobody agreed the shape of is not created by a deriver, and that is recorded rather than quietly resolved. -->
    • M8.1.a Diff ratified contracts/ownership against existing schema and list exact reusable models/fields/indexes and genuine gaps. Evidence: gap map. <!-- Verified 2026-09-05: schemas/persistence_gap_map.py 8.1.0 + services/persistence_gap_map.py + api/persistence_gap_map.py (2 admin-only routes under /persistence-gap-map); tests/test_persistence_gap_map.py 24 tests; ruff + mypy --strict clean; OpenAPI +2 paths, +3 schemas, 0 removed, 0 qualified. The map is COMPUTED from three sources that cannot be edited into agreeing: the contract modules read with ast, the tables read from the SQLAlchemy metadata, and the medium each cell writes to read from its service's imports. Nothing is declared, so it moves when the code moves. Result: 98 ratified contracts against 106 tables — 39 payload_backed, 36 file_store_only, 23 absent, 0 typed, 0 unbacked_fields; by medium 37 ratified_database, 36 file_store, 25 in_memory_only; 86 tables claimed, 20 unclaimed; and 0 tables holding a contract lack a tenant-leading index. Ownership comes from the contract's own version literal (7.14.0 is M7.14). Finding, and the reason the first draft was wrong: it reported 39 contracts as "partial" with hundreds of "missing" fields, because this estate stores payloads in *_json columns beside its typed ones — a field with no column of its own is very likely inside the payload and the schema cannot see which. Reporting that as missing would have been an accusation with no evidence, so the verdict vocabulary now separates payload_backed (untyped fields, and a payload column that may hold them) from unbacked_fields (untyped fields and no payload column anywhere), and the row says untyped, never missing — which is also exactly the input M8.2 needs. Two of the five verdicts have no instance in the estate today; that is a fact about the estate, not the function, so a test asks the decision function directly and every verdict is producible. -->
    • M8.1.b Design tenant-scoped keys, logical/revision IDs, FKs/checks/ uniqueness/classification/retention for each missing aggregate. Evidence: schema review. <!-- Verified 2026-09-05: schemas/persistence_design.py 8.1.1 + services/persistence_design.py + api/persistence_design.py (2 admin-only routes under /persistence-design); tests/test_persistence_design.py 22 tests; ruff + mypy --strict clean; OpenAPI +2 paths, +7 schemas, 0 removed, 0 qualified. Derived from the 60 contracts the M8.1.a gap map says have no table: 81 proposed tables and 158 shapes given none, every aggregate root of every missing contract accounted for as one or the other. The design is a PROPOSAL and most of these tests are about what it declines to decide. The tenant leads every key and the uniqueness is exactly the key in key order; where several identifier fields could be the key it names one, lists the rest and marks the row for ratification (30 of 81 ambiguous); CHECKs are lifted from the Literal vocabularies the contracts already declare (asserted against agent_command's real COMMAND_KINDS); every FK is tenant-scoped by construction and records whether its target is a table this review proposes or one that already exists (both states reached); a name over Postgres's 63-char limit is shortened with a digest rather than truncated, because Postgres truncates silently and two long names become one table. Retention is proposed per classification and ALWAYS carries requires_ratification=True — a schema deriver does not get to decide how long an institution keeps records — and operational_record and ratified are reserved for a person: no pattern produces them, and a test asserts the deriver never does. 67 of 81 rows still need a human decision and the report says so. Finding: the first run proposed 166 tables, 118 of them for HTTP request/response shapes that happened to carry an assessment_id — the suffix test now runs BEFORE the key test, because a wire format with an identifier in it is still a wire format and giving it a table invents an aggregate that does not exist. Also replaced a forward_references field that was always empty by construction with the per-FK target_state that actually distinguishes something. -->
    • M8.1.c Add expand/backfill/verify/cutover/contract migrations with compatibility, idempotency, batching, rollback, and no shadow store. Evidence: migration tests. <!-- Verified 2026-09-06: schemas/persistence_migration.py 8.1.2 + services/persistence_migration.py + scripts/generate_persistence_expand_migration.py + alembic/versions/044_persistence_expand.py (frozen) + 2 more routes under /persistence-design; tests/test_persistence_migration.py 28 tests; ruff + mypy --strict clean; OpenAPI +2 paths, +3 schemas, 0 removed, 0 qualified. Five phases in one order for all 81 designed aggregates, and the order IS the safety property: expand only adds (every statement IF NOT EXISTS), backfill walks the store in sorted path order from a checkpoint in batches of 200, verify compares the two stores and is the only thing that yields a cutover permit, cutover moves the authority, contract removes the old store. A phase that writes owes a rollback and a read-only one may not have a made-up rollback — asserted both ways at the type. The backfill is tested against a real store on disk: 25 records in batches of 10, and an interruption at record 5 resumed twice, each record moved exactly once. All 4 verify verdicts reachable; a changed row at the SAME row count is caught by the digest. A cutover permit binds the source digest it was granted against, so a store that grew or changed between verifying and cutting over invalidates it. And a source still present after cutover is reported as the shadow store, which is exactly the state this sequence exists to avoid. The frozen alembic revision is GENERATED, not hand-written, and the drift test parses its _STATEMENTS with ast and compares tuples — a substring search over the source would have passed or failed on reformatting. It carries only the 14 aggregates the design settled on its own; the other 67 have a plan to read and no licence to run. Finding: phase_order() had 0 callers and was deleted. -->
    • M8.1.d Apply to an isolated real database from clean and production-like prior versions; verify constraints/data/query plans. Evidence: real-DB report. <!-- Verified 2026-09-06 against PostgreSQL 16.14 on 127.0.0.1:5432: tests/test_persistence_real_database.py 11 tests, ALL PASSED, each run against a database created for the run (metis_m81d_<hex>) and dropped after it. METIS_TEST_DATABASE_URL=postgresql+psycopg:// oshun:***@127.0.0.1:5432/oshun_dev PYTHONPATH=src .venv/bin/python -m pytest tests/test_persistence_real_database.py -q -> "11 passed in 11.62s". Applied alembic upgrade head from CLEAN (all 44 revisions) and separately from the production-like prior version: downgrade 043 removed persistence_cutover and all 14 tables, then upgrade head restored them — the upgrade a live database actually runs, which is not the same test as from clean. Constraints checked by asking the server, not by reading the SQL: the tenant-scoped unique index refuses a duplicate and allows the same logical id under a different tenant; a NULL tenant is refused; and a CHECK derived from a contract's Literal vocabulary refuses a value outside it while allowing one inside (this caught my own fixture first — the original test inserted {"kind":"note"} into agent_plan and Postgres rejected it, which is the constraint working). Idempotency: every frozen CREATE re-ran with no change; the ALTER TABLE ... ADD CONSTRAINT is the one statement Postgres has no IF NOT EXISTS for, and the test asserts it raises rather than pretending otherwise. Query plans measured with EXPLAIN over 100,000 rows across 5 tenants after ANALYZE: the tenant-scoped read uses an Index Scan with no Seq Scan, and the contrasting non-tenant read on an unindexed column IS a Seq Scan — without that contrast the first plan would prove nothing. -->
  • M8.2 Separate operational columns from versioned content payloads. Evidence: critical filters use indexed columns. <!-- Verified 2026-09-06: schemas/operational_projection.py 8.2.0 + services/operational_projection.py + api/operational_projection.py (2 admin-only routes) + scripts/generate_operational_columns_migration.py + alembic/versions/045_operational_columns.py (frozen, 126 statements); tests/test_operational_projection.py 34 tests + tests/test_persistence_real_database.py 17 (all against a real PostgreSQL 16.14 database created for the run); ruff + mypy --strict clean; OpenAPI +2 paths, +8 schemas, 0 removed, 0 qualified. (a) 4,490 fields of every ratified aggregate classified into 13 roles — 914 operational, 3,576 content — and content is the DEFAULT rather than a residue, because a field left in the payload is never lost while promoting one to a column asserts we know how to derive and index it. The matcher never uses a bare substring: state is inside statement, risk is inside asterisk, and a field matches only on its whole name or a trailing _-separated segment. (b) 7 typed columns, each naming the ONE payload field it comes from; two columns from one field would be two derivations and the test refuses it, and a source scan asserts no other module in the estate writes an op_ column. An absent field projects to NULL, never to a default nobody wrote. (c) an operational change carries the content digest through untouched and ImmutableUpdate cannot be constructed if it moved; authored content changes by making a revision. (d) tenant-leading composite and partial indexes, benchmarked with EXPLAIN over 100,000 rows across 5 tenants after ANALYZE: the state filter uses the PARTIAL index by name, the owner filter uses the owner index, no ratified filter's plan mentions content_json or jsonb — and the contrasting payload filter does, which is what the columns replaced. (e) the reconciler recomputes from the payload and REPORTS; repaired is a Literal[False] so no caller can read the result as an instruction, and all 4 divergence kinds are reached. Findings: identity was a role no field could reach, and gate_id was being filed as authored content — a trailing _id rule now classifies a record's own identifier, after the specific patterns so tenant_id and actor_id keep theirs. And I wrote the truncation bug M8.1.b warns about: [:63] was applied to the whole CREATE INDEX IF NOT EXISTS ix_... prefix, so the index created was ..._tenant_sta while the plan declared ..._tenant_state — the real-database benchmark caught it, the name is now computed once and digest-shortened, and the type refuses a plan whose declared name is not in its own statement. -->
    • M8.2.a Classify identity/tenant/owner/state/maturity/timestamps/locks/ queue/freshness/risk/visibility fields as operational versus immutable authored content/revision payload. Evidence: field map.
    • M8.2.b Normalize query-critical operational fields into typed columns with constraints and one authoritative derivation. Evidence: schema tests.
    • M8.2.c Store versioned typed content by immutable revision with schema version/hash and prohibit operational mutation from rewriting payload. Evidence: immutability tests.
    • M8.2.d Add tenant-leading composite/partial indexes for ratified filters/ sorts and verify production-scale query plans/no JSON scans. Evidence: index benchmark.
    • M8.2.e Reconcile operational projection after revision/events and detect divergence instead of dual authority. Evidence: consistency tests.
  • M8.3 Implement course, lesson, item, rubric, source, objective, standard, review, and release stores. Evidence: durable-before-ack tests. <!-- Verified 2026-09-06: schemas/aggregate_store.py + services/aggregate_store.py + alembic/versions/046_aggregate_stores.py (5 tables, hand-written because fixed rather than derived); tests/test_aggregate_store_real_database.py 35 tests, ALL against a real PostgreSQL 16.14 database created for the run and dropped after; ruff + mypy --strict clean. There is no unit-test version of this cell: "the four rows commit together", "a second writer at the same revision loses" and "a link to a revision that does not exist cannot be inserted" are properties of a database, and a fake would only prove the fake behaves as written. (a) ONE store for all nine kinds rather than nine copies of the same concurrency and audit code — the kind is a column, so a cross-aggregate reference is one foreign key instead of nine; the repository owns its transactions and no method leaves one open, so a partial write is not something it can be asked to do. All 9 kinds created and read back, and the CHECK on the table is asserted to name all 9. (b) expected_revision is required, so a caller that has not read cannot write; two threads writing at revision 1 leave exactly one winner, one revision_conflict naming the revision the store holds, and two revisions in the table — not three. Exact revision reads, an unknown revision refused rather than approximated, the revision chain linked by previous_sha256, a caller refused a kind it does not hold and a course it is not scoped to. (c) references are a FOREIGN KEY into store_revisions, so a link to a revision that was never written is refused by the DATABASE — nothing in the application checks, and a reference into another tenant is not visible to the key at all, which is stronger than recognising and rejecting it. A reference pins an exact revision that a later revision of the target does not move. (d) the durability test does not check what the store SAYS: it kills a write with a duplicate idempotency key after the revision, aggregate and audit rows are already inserted, then counts all four tables from a fresh connection — 0, 0, 0, 0 — while the earlier successful write is untouched. The acknowledgement is constructed after the with block, so "acknowledged" and "durable" cannot come apart. (e) restart (a rebuilt repository reads what the old one wrote), duplicate/retry (an idempotency key refuses the second attempt, and a byte-identical payload is not a new revision), 64 concurrent writers across 8 threads all committing, two tenants holding the same logical id without meeting, and 100,000 rows across 5 tenants where the tenant+kind+state read uses ix_store_aggregates_tenant_kind_state with no Seq Scan. 7 of 8 refusal reasons reached; cross_tenant_reference is answered by the foreign key as unknown_revision, recorded rather than papered over. Findings: my first scale test loaded ONE tenant and asserted an index scan — the planner was right to scan, because a tenant predicate that matches every row is not selective, so the fixture now loads five tenants and a contrasting no-tenant read is asserted to BE a Seq Scan. Also removed a referenced audit action nothing produced and the module's *_SCHEMA_VERSION marker, which would have made the storage layer count itself among the things needing storage. -->
    • M8.3.a Define repository interfaces and transaction ownership for each aggregate with typed create-revision/read/list/transition operations. Evidence: store contracts.
    • M8.3.b Implement tenant/project/course-scoped persistence, exact revision reads, constraints, optimistic concurrency, and authorization context. Evidence: repository tests.
    • M8.3.c Implement cross-aggregate refs through validated IDs/versions and outbox events, avoiding application-enforced dangling links. Evidence: integration tests.
    • M8.3.d Commit aggregate/revision/audit/outbox atomically before success acknowledgement and roll back on any failure. Evidence: durable-before- ack tests.
    • M8.3.e Test restart/duplicate/retry/deadlock/serialization/two-tenant and production-size access paths. Evidence: real-DB suite.
  • M8.4 Implement immutable revision and semantic-diff queries. Evidence: history integrity integration test. <!-- Verified 2026-09-06: schemas/revision_history.py 8.4.0 + services/revision_history.py + alembic/versions/047_revision_immutability.py (a reason column, two triggers and two keyset indexes); tests/test_revision_history_diff.py 44 tests + tests/test_revision_history_real_database.py 22 tests against a real PostgreSQL 16.14 database created for the run; ruff + mypy --strict clean. (a) pagination is by CURSOR, not offset: an offset means something different the moment a revision is written, so the test pages, writes a revision in between, and asserts nothing from the first page reappears on the second; paging all the way through 10 revisions returns each exactly once. Ancestry walks link by link and reports the revision where a chain breaks rather than trusting it; consumers are the things that pinned an EXACT revision (revision 2 exists and nothing pinned it); a branch point cannot be written by the store, so the query reports damage rather than branching. Protected fields are withheld BY NAME — a learner's row carries no reason and no author and says which, and the digest is never withheld from anybody. (b) immutability is a TRIGGER, not a rule the application keeps: the test runs UPDATE against all six of payload, hash, parent, author, reason and time, and a DELETE, and the database refuses each with "history is immutable"; the audit table too. An approved workflow may redact (retention_expiry, legal_erasure) and an invented name may not — and because the permission is SET LOCAL, the very next statement on the same pooled connection is refused again, which is asserted. (c/d) the diff is over IDENTITIES: three blocks rotated give three moved and zero changed, an insertion in the middle moves the ones after it and changes none, and a list with no identities is answered unknown rather than guessed at. An answer-key path is redacted — that it changed, never what to, and the old value appears nowhere in the serialised diff. All 7 dimensions reached from real payload pairs, all 7 change kinds, and the diff names both revisions and both digests so it cannot be attached to the wrong pair. (e) the corruption fixture drags the projection back to revision 1 and deletes another one entirely; the rebuild replays the log, repairs both, and the log itself is verified unchanged afterwards — the repair only ever goes from the log to the projection. Findings: the entity dimension was unreachable because A and B or C grouped wrongly, so a whole object added was filed as a plain field; and threading the new reason column through M8.3's read left the SELECT one column short, which the rebuild test caught as an IndexError. -->
    • M8.4.a Implement revision list/detail/ancestry/branch/consumer queries with stable pagination and protected-field projection. Evidence: query tests.
    • M8.4.b Verify stored payload/hash/parent/author/reason/time cannot be changed/deleted outside approved retention/legal workflows. Evidence: immutability tests.
    • M8.4.c Compute typed semantic field/entity/list/relationship/source/ accessibility/key-safe diffs using stable identities. Evidence: diff goldens.
    • M8.4.d Represent added/removed/changed/moved/unchanged/redacted/unknown and bound diff to schema/version pair. Evidence: result schema.
    • M8.4.e Rebuild current/history from revisions/events after corruption fixture and verify integrity across restore/supersede. Evidence: integration test.
  • M8.5 Implement source/claim/citation and blast-radius APIs. Evidence: high-risk grounding known-value route test. <!-- Verified 2026-09-06: schemas/source_graph.py 8.5.0 + services/source_graph.py + api/source_graph.py (4 routes under /source-graph) + alembic/versions/048_source_propagation.py; tests/test_source_graph.py 35 tests + 16 against a real PostgreSQL 16.14 database; ruff + mypy --strict clean; OpenAPI +4 paths, +9 schemas, 0 removed, 0 qualified. (a) six typed locator kinds, and a locator nobody could follow is not accepted: a bad DOI, an http URL and a javascript: URL are all refused, and a page range that ends before it begins points at nothing — which a regex over "pages 40-12" cannot see and a typed range can. Property-level authorization by audience: a learner reads a protected source's title and NOT its locator, rights or protected flag, and the withheld properties are NAMED. (b) rights that were never checked are not rights to use (unknown is deliberately outside CITABLE_RIGHTS), currency runs from the last verification and differs by authority, and a stale source is not citable however good its rights. The headline: THERE IS NO PATH TO A CANONICAL CLAIM THAT DOES NOT CARRY A HUMAN DECISION — advance_claim raises on canonical by name, make_canonical requires the decision as an argument, the type refuses a canonical claim without one, and a provider proposal carries is_decision: Literal[False] at any confidence. (c) breadth-first over M8.3's reference rows with a fixed neighbour ordering, so the reported shortest path is the SAME shortest path on every run (asserted by walking twice); a node reachable two ways is reported once at the shorter distance; a walk that hits the depth limit reports the depth it stopped at, because a truncated answer that does not admit it is a wrong answer; and the walk does not cross a tenant even when both hold the same logical id. (d) a retraction writes propagation rows and outbox events in one transaction and NEVER touches the revision log — asserted by re-reading the lesson's digest, checking the ancestry is intact, and counting the revisions afterwards. (e) the known-value route: source -> objective -> lesson -> item -> release plus a tutor review, retracted as protected material, with the exact five (kind, id, state, owner) tuples and the exact distances written down in the test, and another tenant's identical chain untouched. Finding: the outbox idempotency key omitted the reason, so an ESCALATION from stale to blocked could not write its events and killed the transaction — a stale notice and a later blocking notice are different news an owner must get, so the reason is now part of the key, the insert is ON CONFLICT DO NOTHING, and a true retry acknowledges the delivery it already made rather than inventing a second one. -->
    • M8.5.a Define source/extraction/claim/citation list/detail/create-review/ link/verify contracts with exact revisions and property authorization. Evidence: API schemas.
    • M8.5.b Validate typed locators/source rights/authority/currency and provider proposals; canonical claims require human decision. Evidence: mutation tests.
    • M8.5.c Implement forward/reverse dependency and impact preview queries with stable paths/pagination/reasons/current state. Evidence: graph tests.
    • M8.5.d Implement update/retraction command triggering stale/block state, owner queue, outbox/audit, and no history rewrite. Evidence: propagation tests.
    • M8.5.e Run high-risk source-to-lesson/item/tutor/published-build route fixture with cross-tenant/protected-source failures. Evidence: known- value route E2E.
  • M8.6 Implement course/lesson/item/rubric CRUD with property-level authorization and optimistic concurrency. Evidence: contract suite. <!-- Verified 2026-09-06: schemas/authoring_api.py 8.6.0 + services/authoring_api.py + api/authoring.py (6 routes under /authoring) + alembic/versions/049_aggregate_archive.py; tests/test_authoring_api.py 23 tests against a real PostgreSQL 16.14 database; ruff + mypy --strict clean; OpenAPI +6 paths, +11 schemas, 1 removed (my own M8.5 SourceRecord, renamed), 0 qualified. (a) eight operations over M8.3's store, and A PUBLISHED AGGREGATE IS NOT EDITED IN PLACE — edit refuses it by name and says to make a new revision, which is a different route with the same store call underneath. Archive is a SHELF, not a lifecycle decision: a withdrawn lesson can be archived and restored and is still withdrawn afterwards, because withdrawal is terminal on purpose and a shelf is a different question. (b) role x operation is total and no two roles are alike; a role is refused at the route before the object is looked at (proved by showing the object does not exist afterwards); per-property visibility is applied in ONE place on the way out, and it applies to VALIDATION ERRORS too — a learner's rejection of an empty answer_key says the property is empty and does not quote it, which is the easiest place a withheld field escapes. Groups match whole segments, so answerable is not an answer. (c) never last-write-wins: a second editor at the same base gets conflict with the current and base revisions and the store still holds the FIRST writer's value; the conflict diff is the M8.4 differ, so an answer-key change is redacted and the old value appears nowhere in the serialised outcome — a merge hint is not worth handing a key to anybody who can force a conflict. ETags round-trip and refuse one issued for another version. (d) writes go through M8.3's atomic four-row transaction with idempotency. (e) all SEVEN outcome states are published in one type and all seven are REACHED by the suite; success and redacted are deliberately different answers, because a caller reading success with fields missing would take them for fields that were never there. Findings: three model names collided with other cells (PropagationResult, SourceRecord, ImpactPreview) — renamed to SourceRetractionResult, CitedSource, BlastRadiusPreview — and EditRequest collided with the anchored-comment API's. The OpenAPI removal diff caught only the last one: a collision between an exposed model and one that is NOT exposed reaches the contract as a silent rename of the exposed one, so the package-wide duplicate-name guard is the check that sees it, and this cell now carries a copy. Also fixed an ETag parser that split from the left, cutting lesson:0001 in half. -->
    • M8.6.a Define versioned create/read/list/edit/new-revision/transition/ archive/restore contracts and disallow in-place published mutation. Evidence: schemas.
    • M8.6.b Enforce route/function/object/tenant/course authorization and per-property answer/source/learner/review visibility on inputs/outputs/ errors. Evidence: auth matrix.
    • M8.6.c Accept expected revision/ETag and return typed conflict with safe current/base diff; never last-write-wins. Evidence: concurrency tests.
    • M8.6.d Validate relationships/registries/standards/floors and write revision/audit/outbox atomically with idempotency. Evidence: mutation tests.
    • M8.6.e Publish consumer contracts for success/validation/conflict/stale/ forbidden/not-found/redacted states. Evidence: contract suite.
  • M8.7 Implement provider-backed generation/assist APIs. Evidence: 503, 502, cancellation, malformed output, budget, and success cases. <!-- Verified 2026-09-06: schemas/generation_job.py 8.7.0 + services/generation_job.py + api/generation.py (4 routes under /generation) + alembic/versions/050_generation_jobs.py; tests/test_generation_job.py 45 tests against a real PostgreSQL 16.14 database with a provider double at the seam; ruff + mypy --strict clean; OpenAPI +4 paths, +8 schemas, 0 removed, 0 qualified. (a) five operations, typed request/context/estimate/job/proposal/progress/error, and the provider is a Protocol — the only thing faked in the tests, and it is the dependency rather than the thing under test. (b) GRANTS ARE RESOLVED FROM THE OPERATION AND THE REQUEST IS NOT AN INPUT: the server's context holds answer_key, learner_responses and credentials, and no operation reads any of them; the prompt is asserted not to contain them. Money is reserved BEFORE dispatch inside a transaction that counts open holds, so two jobs cannot both fit into room for one (asserted), and a job that does not fit never reaches the provider (the double records zero calls). (c) all nine failure classes carry a fixed status — not_configured and provider_unavailable and rate_limited are 503, malformed_output and upstream_error are 502 — and BOTH the status and the retryability come from the class, refused at the type when a caller tries to set either case by case. malformed_output is deliberately not retryable: the same prompt produces the same mess. There is NO FALLBACK: an unconfigured provider is a 503 over HTTP and is never quietly replaced. (d) durable jobs read back from the database, idempotent by key, and FENCED — the cancel test bumps the fence while a provider is still working and the late result is dropped, with the contrast that an on-time result is accepted. Cancellation releases the money. (e) a proposal carries applied and approved as Literal[False] and there is no field that could say otherwise; a generation writes nothing to the aggregate store (asserted by counting revisions); and every one of the six failing behaviours raises with a planted secret in its message while NONE of the four secrets reaches the failure body or the job document. Findings: my budget tests assumed a 1-cent limit would refuse a job the estimator prices at 1 cent — it fits, and the honest zero-budget case is limit 0; and the forbidden-grants check sat behind the allowed-grants check, making it unreachable, so it now answers first. Two vocabulary values have no producer in this build — rate_limited (the seam cannot report it yet) and queued (the job runs inline) — and that is recorded in the reachability assertions rather than papered over. -->
    • M8.7.a Define operation-specific typed request/estimate/confirmation/job/ proposal/progress/error contracts and injected provider seam. Evidence: API schemas.
    • M8.7.b Resolve canonical source/course/item/context/policy/tool grants server-side and reserve quota/budget before dispatch. Evidence: guard tests.
    • M8.7.c Map not-configured/unavailable to 503, malformed/upstream failure to stable 502 class, plus refusal/timeout/rate/budget/cancel states without fallback. Evidence: failure contracts.
    • M8.7.d Implement idempotent durable jobs, progress stream, cancellation/ late-result fencing, and attributed proposal intake. Evidence: job tests.
    • M8.7.e Prove success cannot auto-edit/approve/publish/expose answers and errors leak no prompts/credentials/learner data. Evidence: boundary suite.
  • M8.8 Implement quality and psychometric gate jobs. Evidence: stale-policy recomputation tests. <!-- Verified 2026-09-06: schemas/gate_job.py 8.8.0 + services/gate_job.py + alembic/versions/051_gate_jobs.py; tests/test_gate_job.py 47 tests + 2 real-database tests; ruff + mypy --strict clean. The cell turns on ONE distinction and everything follows from it: a MEASUREMENT is identified by what could change what was observed (subject revision, evaluator, evaluator version, corpus, input digest) and a VERDICT by that measurement plus the policy and its thresholds. The two input sets are DISJOINT, asserted at import, which is what makes decide_recompute a derivation rather than a judgement. (d) a threshold or policy version moving is verdict_only and the measurement key is literally unchanged; each of the 5 measurement inputs moving is new_measurement_required (parametrised over all five); when both move the measurement wins, because no policy arithmetic can produce an observation of a different subject. The schema enforces the same rule: gate_measurement is keyed by the observation and not the policy, so a policy edit CANNOT write a second row — checked against the real constraint on PostgreSQL 16.14. (a) one item per gate at one exact revision; a gate measuring a different revision, waiting on a gate the job does not run, or waiting on itself is unconstructible, and a dependency cycle is refused at plan time rather than hanging. (b) leases with a holder and an expiry, reclaimable once lapsed, bounded at 3 attempts; a dependant of a failed gate is SKIPPED rather than failed, because it did not fail — nothing ran it; and a cancellation KEEPS what was already measured, since throwing away real observations makes cancelling expensive enough that people stop doing it when they should. (c) a measurement cannot carry a judgement (a verdict/passed/decision key is refused) and a gate that does not apply observed nothing; applicability, uncertainty and freshness are separate fields, and the type refuses a pass that is not applicable, is undetermined, or rests on a stale measurement. A DETERMINISTIC measurement never goes stale on its own — same inputs, same answer — and its absence from the freshness table is asserted. (e) authority is computed from EVERY gate: one failure withholds the whole job, an unfinished or cancelled job authorises nothing, a stale measurement withholds even where the numbers were fine, and advisory is a note rather than permission. All 5 withholding reasons reached, and "mostly passed" is unconstructible. -->
    • M8.8.a Resolve exact subject revisions and required deterministic/model/ psychometric/accessibility gates plus evaluator/policy/corpus versions. Evidence: job plan tests.
    • M8.8.b Execute bounded dependency-aware gate items with leases, idempotency, progress, retries, cancellation, and partial preservation. Evidence: worker tests.
    • M8.8.c Persist immutable raw measurements and separate verdicts/ applicability/uncertainty/freshness. Evidence: result tests.
    • M8.8.d Recompute verdict-only on policy/threshold change and require new measurement on subject/evaluator/input change. Evidence: stale-policy tests.
    • M8.8.e Prevent partial/stale/unknown jobs from authorizing publication or calibration promotion. Evidence: authority tests.
  • M8.9 Implement QTI/LTI/OneRoster/SCORM/xAPI/Caliper/Open Badges import and export jobs. Evidence: per-standard golden fixtures. <!-- Verified 2026-09-06: schemas/standards_exchange.py 8.9.0 + services/standards_exchange.py + services/open_badges_service.py (the seventh standard, which this build had no implementation for); tests/test_standards_exchange.py 42 tests; ruff + mypy --strict clean. (a) one job shape over 7 standards x 2 directions, and A CAPABILITY IS WHAT THE CODE DOES: the registry declares intent, probe_capabilities imports the implementation module, and supported is refused at the type unless both are true. The probe initially reported open_badges UNSUPPORTED because nothing implemented it — so it was implemented (OB 3.0 assertions as Verifiable Credentials over the estate's own credential models) rather than left as a note, and all 14 cells are now supported. (b) the hostile defences live with the CONTAINER rather than the standard, because the attacks are properties of the container: real fixtures build a zip whose member is ../../etc/cron.d/x, a deflated archive that unpacks past the ceiling from a 300KB file, an XML with a billion-laughs entity, and an XML pulling in /etc/passwd — all refused, and a clean payload of every container passes. A preview accounts for everything it was shown and plans NOTHING when the payload itself is refused. (c) exports declare loss DERIVED FROM THE PROFILE rather than from what the exporter noticed, undeclared_loss catches a manifest that hid one, a manifest cannot declare loss against another profile, and two builds of the same export are byte-identical. (d) NO SILENT DROPS: the receipt refuses to exist unless every offered item has exactly one outcome, so "we lost three of forty questions" is not a bug this surface can have — it is a shape it cannot express. A resume passes what it already did, touches none of it again, and still accounts for all of it. (e) minimal/full/edge goldens for the Open Badges profile all parse and round-trip byte-identically through build and parse, with six incompatible cells stated by name — and a document carrying a proof reports proof_present, never verified, because treating the presence of a proof as proof is how badge systems are fooled. Findings: my zip-bomb fixture wrote STORED rather than DEFLATED, so the "small archive" was 32MB; and the completion check I wanted to exercise sat behind the receipt check, so the test now supplies a receipt to reach it. -->
    • M8.9.a Define per-standard operation/direction/profile/version/credential/ input/output/diagnostic/progress/result contracts and capability probes. Evidence: job schemas.
    • M8.9.b Parse/fetch/upload safely, validate conformance/identity/version/ rights, preview mappings/loss/conflicts, and require authorization. Evidence: hostile import tests.
    • M8.9.c Export exact revisions deterministically with standards metadata, privacy/property filtering, manifests/hashes, and declared loss. Evidence: export tests.
    • M8.9.d Run durable idempotent per-item/resource jobs with partial results, retry/cancel/resume, receipts, and no silent drops. Evidence: lifecycle tests.
    • M8.9.e Validate/import/read-back minimal/full/edge golden for every supported profile and explicit incompatible cells. Evidence: conformance suite.
  • M8.10 Implement publication and LMS synchronization as idempotent sagas. Evidence: partial-failure recovery. <!-- Verified 2026-09-06: schemas/publication_saga.py 8.10.0 + services/publication_saga.py + models/publication_saga.py + alembic/versions/052_publication_sagas.py; tests/test_publication_saga.py 54 + tests/test_publication_saga_real_database.py 34 (real PostgreSQL, a destination double at the boundary); ruff + mypy --strict clean; no new routes, so the OpenAPI diff is empty and the package-wide duplicate model-name guard is clean. (a) the plan names exact build/assessment revisions and target IDs, is stored once per key (a re-plan returns the first plan, because re-planning is how duplicates get made), and refuses two steps sharing an idempotency key, a step waiting on one the order never satisfies, and — via duplicate_risk at store_plan — a creating step whose key is not unique to this plan. (b) THE PREREQUISITES ARE RE-READ BEFORE EVERY WRITE, NOT ONCE PER RUN: a probe counts snapshots taken and asserts one per step, and a double that moves rights after the second step leaves steps 1-2 succeeded, step 3 pending, the run partial/compensate, and the destination holding exactly two writes. A plan that goes stale before anything landed is needs_operator/manual rather than compensate: there is nothing to undo and the plan itself is what moved. (c) THE INTENT TO WRITE IS COMMITTED BEFORE THE CALL GOES OUT. A separate connection, opened from inside the destination's write, sees the ledger row already committed; a timeout raised after the resource landed leaves the step verifying, and the resume reads instead of writing — one write under that key, asserted from the double. A refused connection cannot have delivered anything, so its key is released and the retry writes; anything unrecognised is treated as possibly delivered, because an unverified step costs a read and a duplicated one costs a duplicate. (d) A WRITE IS NOT A SUCCESS UNTIL THE DESTINATION IS READ BACK AND AGREES. StepCheckpoint(state="succeeded") is unconstructible with no verification, with one that came back absent, with one that came back different, and with a "matched" verification whose two digests do not match. A destination that acknowledges and silently drops reads back absent (retry, and the key goes back because we now know it did not land); one holding something else reads back different and is never overwritten automatically — a re-execute writes nothing. (e) compensation walks the plan in reverse under its own key suffix, and a compensation the read-back cannot confirm LEAVES THE STEP SUCCEEDED — because it is, the resource is still standing — and asks for a person instead of reporting a rollback that did not happen. reconcile only reads: it reports a resource the destination lost (retry), a write that landed without being recorded (reconcile, adopt it, because writing again is how a course gets two), one holding something else (manual), and what it could not read at all, and a probe asserts it changed neither the run nor the destination. Two vocabulary defects in my own draft, both caught by producibility probes: writing and revalidating were states no checkpoint could ever be written in (checkpoints are persisted at decisions, not during them), and TERMINAL_STEP_STATES described something the engine does not honour — it retries failed. Replaced by RESUMABLE_STEP_STATES, which the engine actually branches on, with a real-DB probe asserting every remaining state and every saga state but running is produced by some scenario. M8.1's gap map then correctly reported the new contract as storage-less, because 052's tables are raw SQL and were not in Base.metadata; adding models/publication_saga.py fixed that, and the gap map's own tenant-index invariant immediately caught that two child tables had no tenant column at all — both now carry tenant_id and a tenant-leading index, with a probe asserting every row a run writes carries it. -->
    • M8.10.a Define immutable publication/sync plan with exact build/ assessment revisions, target/course/context/resource IDs, prerequisite snapshot, steps, compensation, and idempotency. Evidence: saga schema.
    • M8.10.b Revalidate gates/reviews/rights/standards/target/config/ authorization immediately before each external write. Evidence: TOCTOU tests.
    • M8.10.c Execute create/update/link/content/grade/config/visibility steps with durable checkpoints and destination-native idempotency/versioning. Evidence: integration tests.
    • M8.10.d Read back external state/bytes/metadata/permissions/links and mark success only after verification. Evidence: read-back tests.
    • M8.10.e Recover by reconcile/retry/compensate/manual action without duplicating resources or hiding partial state. Evidence: partial-failure tests.
  • M8.11 Implement gradebook corrections and late-event replay. Evidence: duplicate and out-of-order tests. <!-- Verified 2026-09-06: schemas/grade_ledger.py 8.11.0 + services/grade_ledger.py + models/grade_ledger.py + alembic/versions/053_grade_ledger.py, under M7.10's existing grade contracts; tests/test_grade_ledger.py 35 + tests/test_grade_ledger_real_database.py 23 (real PostgreSQL); ruff + mypy --strict clean; no new routes, OpenAPI unchanged. (a) AN ARRIVAL ORDER IS NOT AN EVENT ORDER. The ledger keeps occurred_at (what the source said) and received_at (when we heard) in separate columns and orders by the first only; the permutation probe feeds four events in all 24 arrival orders and every one produces the same sequence numbers and the same digest. AN EXTERNAL IDENTIFIER IS NOT AN INTERNAL ONE: a platform's event id can be an uppercase UUID or a URL, so the source's id is stored verbatim (it is what dedup keys on) and the estate's Identifier is derived from it — deterministic and scoped by system. (b) appends are serialised per grade with an advisory lock, because AN EXPECTED-VERSION CHECK TWO WRITERS CAN PASS AT ONCE IS A COMMENT: two threads at the same version contend and exactly one wins. Audit and outbox rows are written in the append's own transaction. (c) A TIE IS AN AMBIGUITY, NOT A COIN FLIP — two events nothing can separate are refused and the pair is named, rather than ordered by arrival behind a deterministic-looking tiebreak; a source that can emit two events in one instant must supply a sequence, refused at the type. Three real defects the tests caught in my own engine: a retry was reported as AMBIGUOUS WITH ITSELF; a retry carrying the version its first attempt saw would have been REFUSED FOR BEING LATE (a delivery guarantee turned into an error); and the same source event claiming a SECOND GRADE came back as a duplicate, silently dropping it. The duplicate check now answers before every other check, and grade_mismatch — which the producibility probe showed had no producer at all — is now the refusal for the third. (d) FOUND AND FIXED IN M7.10: propagate folded history.current() .sequence into the payload digest, so a late event renumbering the chain changed the idempotency key of a value that had not changed — a rebuild after a late arrival would have written to LTI/OneRoster for a grade nobody touched. The sequence is this estate's position in a derived history, not something a downstream holds; it is out of the payload, with a probe on each direction (unchanged value → same key, changed value → different key) and M7.10's own 23 tests still green. ReplayReport refuses to be built if the outbox count moved across a rebuild, counted from a separate read, and ten rebuilds produce one digest. (e) the ledger feeds M7.10's visibility matrix rather than going round it: the same ledger shows a learner no value before release and their value after, and stale/unauthorized/unspoken-for appends are refused with every reason reported at once. -->
    • M8.11.a Persist immutable grade/source event/correction IDs, exact learner/attempt/item/scale revisions, ordering keys, timestamps, source, and integrity. Evidence: event schema.
    • M8.11.b Apply authorized corrections as append-only events with expected grade version, audit/outbox, and deterministic current projection. Evidence: correction tests.
    • M8.11.c Deduplicate by source/idempotency and order/reconcile late, reordered, retried, corrected, voided, and replayed events. Evidence: event ordering tests.
    • M8.11.d Rebuild projection/aggregates/LTI/OneRoster/credentials from event history without duplicating external writes. Evidence: replay tests.
    • M8.11.e Preserve visibility/release windows/appeals/notifications and reject stale unauthorized mutation. Evidence: boundary tests.
  • M8.12 Implement integrity/appeal API integration without shadow records. Evidence: authoritative ID linkage. <!-- Verified 2026-09-06: schemas/themis_linkage.py 8.12.0 + services/themis_linkage.py + services/themis_boundary_gate.py + models/themis_linkage.py + alembic 054, and a one-line change to services/persistence_design.py; tests/test_themis_linkage.py 56 + tests/test_themis_linkage_real_database.py 18 (real PostgreSQL); ruff + mypy --strict clean; no new routes, OpenAPI unchanged. THE GATE FOUND A REAL SHADOW ON ITS FIRST RUN. Revision 044 created appeal_workflow_integrity_cases — a complete Metis-side integrity case, derived from M7.4's IntegrityCase contract, for a decision Themis owns. 044 and 045 are regenerated without it and 054 drops it from any database that already ran the old chain, with a real-DB test asserting it is absent after the whole chain runs. The deriver does not drop the module in silence: it emits a DerivedShape with a new authoritative_elsewhere reason for each of its aggregates, because A CONTRACT EXCLUDED WITHOUT A STATED REASON IS INDISTINGUISHABLE FROM ONE NOBODY LOOKED AT — my first attempt did drop it, and M8.1's own "designed or explained" test caught that. The only OpenAPI change in this cell is that one enum value; the schema and path sets are identical. (a) the linkage record is an authoritative kind + id + version, a Metis subject, and a delivery state. Its field tuple is asserted against a hand-written list, no field name anywhere in the contract may contain any of twelve content words (rationale, finding, verdict, sanction...), and a real-DB test reads information_schema.columns against a hand-written column list — because a shadow usually arrives as one more column on a table that already exists, not as a new table. (b) the seam is a Protocol with one method and no way to ask for a rationale: WHAT THIS SERVICE CANNOT ASK FOR, IT CANNOT STORE. Authorization carries actor/tenant/purpose/properties and is never taken from a body; linking is idempotent on the authoritative identity, and the same record pointing at a second subject is refused. (c) A DELETION IS NOT AN ABSENCE. deleted (Themis says it is gone), forbidden (it may exist and you may not see it) and missing (the read failed and said nothing) stay three answers, each with its own sentence, because a learner told "no case found" when the truth is "you are not allowed to see it" has been told something false. An unclassified seam failure raises rather than falling into missing. The projection links to an explanation and never copies one, and a blank status without a reason is unconstructible. (d) events are versioned and idempotent: newer applies, equal is a duplicate, older is a reordered delivery and does not wind the record back, a merge is followed and says where the case went, a withdrawal makes the outcome undeliverable, and a forbidden/missing report changes nothing because it is not news about the record. The outcome vocabulary is partitioned into those that record something and those that decline, checked at import — the first draft forbade a withdrawal from moving the version, which was wrong: "this is gone" arrives as a version like anything else. (e) A SHADOW IS DECIDED BY PROVENANCE, NOT BY ITS NAME. The gate's first draft matched table names against appeal and was wrong both ways: it flagged item_exposure_appeals (an appeal about an assessment item leaking, nothing to do with anyone's conduct) and no column-level rule could have caught the real shadow, whose columns are just tenant_id, case_id, revision, content_json. What separates them is which contract each table serves. The name net is kept only for tables no contract claims, which is the one case provenance cannot answer. All three surfaces — Base.metadata plus every CREATE TABLE in the migration history, the deriver's proposals, and imports reaching past a Themis public package — are arguments to the scanners, and each is run over a tree with a shadow planted in it, because M7.1's own lesson is that AN INVENTORY THAT FAILS ONLY ON DRIFT IS NOT A GATE. -->
    • M8.12.a Store only Themis authoritative case/evidence/decision/appeal IDs/ versions plus minimal Metis subject linkage and delivery state. Evidence: linkage schema.
    • M8.12.b Create/query/update through public Themis contracts with current actor/tenant/purpose/property authorization and idempotency. Evidence: contract tests.
    • M8.12.c Project audience-safe status/explanation into Metis without copying raw evidence/decision authority. Evidence: projection tests.
    • M8.12.d Consume versioned events idempotently and handle missing/ forbidden/deleted/merged/corrected authoritative records. Evidence: event tests.
    • M8.12.e Add schema/import guards proving no shadow integrity decision/ appeal tables or private imports are introduced. Evidence: boundary gate.
  • M8.13 Implement permission-aware search and item-bank indexing. Evidence: hidden answer and protected integrity metadata do not leak. <!-- Verified 2026-09-06: schemas/search_index.py 8.13.0 + services/search_index.py; tests/test_search_index.py 92 tests; ruff + mypy --strict clean; no new routes, OpenAPI unchanged. (a) A FIELD REACHES THE INDEX BECAUSE IT WAS CLASSIFIED, NOT BECAUSE IT WAS PRESENT. FIELD_CLASSES is total over the eight kinds the item names; a payload field nobody classified is REFUSED rather than indexed or dropped, because both of those are somebody guessing. Every kind is checked at import to have something searchable and something forbidden — a kind that forbids nothing has not been reviewed, it has been skipped. (b) index events are ordered by source revision and never by arrival: an older one is ignored, a duplicate changes nothing, a quarantine withholds without forgetting, and a delete leaves a tombstone a late upsert cannot resurrect. Reconciliation finds all three populations, including the source with no document at all — the one a rebuild that only walks the index misses entirely. Exposure is a first-class field, not a display string, so the suppression rule cannot be fooled by what an author typed; a worn-out item is withheld from everyone but an admin. (c) filters, sorts, facets and pagination are allowlists, and the index only narrows: every surviving hit is authorized individually afterwards. A facet a caller cannot then filter on is refused at import as a dead end. (d) the protection is structural rather than a filter. A forbidden value never reaches the document, so it cannot appear in a snippet (cut from the searchable map), a facet, a count, an autocomplete suggestion (built from authorized documents only) or a log. A COUNT IS A LEAK: totals and facet buckets are computed from what survived authorization, and the type refuses a response whose facet counts more than its total. A QUERY LOG IS A PLACE PROTECTED TEXT ENDS UP — somebody pastes an answer key in to see whether it is findable — so the log keeps a digest and a term count and has no field text could go in. The residue scan runs over all eight kinds, searching the WHOLE serialized document rather than its declared fields, and a probe proves the scan can actually find a planted leak. (e) A NON-LEAK IS PROVED BY INDISTINGUISHABILITY, so the adversarial tests compare the ENTIRE response for a document that exists and is forbidden against the response for one that does not exist and require them equal — asserting "the answer key is not in the response" passes for a response that says "0 results (1 hidden)". Same for a guessed course id, a quarantined document, and two tenants holding the same entity id (a grant is scoped by the tenant that issued it, not by the id it names). Timing is not asserted with a clock, which would be flaky: THE TIMING ORACLE IS CLOSED BY CONSTRUCTION, NOT BY PADDING — nothing in the search path reads a protected value, so no branch's duration can depend on one, and a scan over the module's own source proves it never names a forbidden field. -->
    • M8.13.a Define index documents per source/course/lesson/item/stimulus/ rubric/objective/review with classified searchable/display-only/ forbidden fields. Evidence: index schema.
    • M8.13.b Build tenant/course/role/exposure-aware indexing events, version/freshness, deletion/quarantine, and rebuild reconciliation. Evidence: index lifecycle tests.
    • M8.13.c Query with allowlisted filters/sorts/facets/autocomplete/ pagination and server-side object/property authorization after hits. Evidence: search tests.
    • M8.13.d Exclude answer/rationale/hidden tests/integrity evidence/learner data/protected sources from index, snippets, counts, spellcheck, and logs. Evidence: leak tests.
    • M8.13.e Test guessed IDs/terms/timing/counts/stale docs and two-tenant collisions. Evidence: adversarial suite.
  • M8.14 Implement tenant-isolation tests for authoring, learner preview, search, export, gradebook, integrity, and analytics. Evidence: two-tenant suite. <!-- Every sub-item is now checked. This parent was briefly unchecked on the belief that learner preview lived in Studio; it does not, and .c records the correction and the work. Verified 2026-09-06: schemas/tenant_isolation.py 8.14.0 + services/tenant_isolation.py; tests/test_tenant_isolation.py 35 + tests/test_tenant_isolation_real_database.py 12 + tests/test_tenant_isolation_authoring_api.py 9 (the last two against real PostgreSQL); ruff + mypy --strict clean; no new routes or tables, OpenAPI unchanged. THE MEASUREMENT FIRST: classify(Base.metadata) walks every table and says how a row reaches a tenant. 102 by a tenant column, all with a tenant-leading index; 0 by inheritance through a foreign key (checked, not assumed); and 13 BY NOTHING AT ALL — users, courses, modules, lessons, assessments, questions, submissions, enrollments, progress, achievements, tutoring_sessions, tutoring_messages, xapi_statements. That is the original schema, written before this estate had tenants, and it holds the learners and the content. It is recorded as a baseline rather than quietly fixed: adding a tenant column to thirteen core tables is a migration, a backfill and a rewrite of every query that touches them, and doing that as a side effect of writing a test suite would be the wrong way round. The baseline is a RATCHET, tested in both directions — a new untenanted table fails, and so does a baseline entry that has since been tenanted, because otherwise the list rots into a permanent excuse. (a) THE ONLY FIXTURE THAT CAN FAIL IS A COLLIDING ONE. Both tenants use the same title, slug, logical id, source event id, Themis case id and learner reference; the only difference is one distinctive word each. With different ids the suite would pass against a service doing no scoping at all, because there would be nothing to confuse. A test asserts the colliding rows really do exist under both tenants, so the rest cannot be passing on an empty database. (b)/(e) the real-database half seeds three stores at once — the M8.11 grade ledger, the M8.12 Themis linkage and the M8.3 aggregate store — so a leak crossing between them has somewhere to show up. Each tenant's reads return one row, not two, under an id both of them used; the two projections' digests DIFFER, because identical ones would mean the projection ignored something it read; and a third tenant that wrote nothing gets nothing, so "returns only mine" is not "returns everything". Every table is counted per tenant and the counts must be equal, non-zero, and leave nothing attributed to nobody. (c) STAYS UNCHECKED. The learner preview surface is Studio's, and SURFACE_HOMES records that rather than letting a Python suite imply it covered a renderer, an offline cache and a media pipeline it does not host. Naming it is the difference between "covered" and "nobody looked". (d) search is exercised over the colliding corpus: each tenant sees one hit, the facet counts one, autocomplete does not read the other tenant's words, and searching for the other tenant's distinctive term returns a response IDENTICAL to the one for a term that is not there. (f) all ten channels are enumerated and the sweep refuses to be built with one missing — a channel nobody enumerated is a channel nobody checked. Three evidence strings are turned into checks rather than left as sentences: no set_attribute call exists anywhere in the service (trace), no @lru_cached function takes a tenant/learner/actor argument (cache), and tenant_storage_key is deterministic, non-reversible and distinct per tenant (blob). The database channel is reported as keyed_with_recorded_exceptions and NAMES ALL THIRTEEN — claiming keyed_by_tenant of a channel with thirteen exceptions would simply be false. -->
    • M8.14.a Seed two tenants with colliding names/IDs-slugs/terms/times/ sources/items/learners/roles plus distinct keys/policies/integrations. Evidence: isolation fixture.
    • M8.14.b Exercise every authoring CRUD/revision/comment/review/build/ provider/agent path with foreign IDs/cursors/links. Evidence: authoring tests. <!-- tests/test_tenant_isolation_authoring_api.py 9 tests over HTTP: all eight /authoring routes driven by two tenants holding the SAME logical id, so a route ignoring the tenant would still have a candidate row to return. Create, read, list, edit, new revision, transition, archive and restore; the listing counts one and not two, an edit is invisible to the other tenant, a transition moves one lifecycle and not the other (observable through the listing, because a read deliberately does not carry the state), an archive leaves the other tenant's listing alone, and a revision number only the other tenant has does not reach across. A CALLER CANNOT NAME ITS OWN TENANT: _tenant_of derives it from the authenticated user with a uuid5, and a test sends both an X-Tenant-Id header and a tenant_id query parameter and still gets its own row. The comment, review, build, provider and agent paths named in this item are Studio's, in TypeScript, and are not routes this service serves. -->
    • M8.14.c Exercise learner preview/session/cache/offline/renderer/media/ tutor synthetic and real-context boundaries. Evidence: preview tests. <!-- I FIRST RECORDED THIS AS OUT OF REACH AND WAS WRONG. The learner preview is not Studio's — it is apps/metis/web, in this repository: src/app/create/lessons/[lessonId]/preview with a delivery renderer, a runtime provider and a session media cache. Naming something out of reach without looking is the same mistake as claiming to have covered it. The tenancy boundary is the session media cache, and it holds: the key is ${documentSha256}:${contentSha256}, both halves digests of the bytes themselves, so A SHARED CACHE KEY MEANS SHARED BYTES — two tenants can only collide when they hold byte-identical content, and the entry they share is what they both already have. A cache keyed by lesson id would be a cross-tenant leak inside one browser session; this one cannot be. The cache also lives in a useRef for one mounted preview and is released on unmount, so a second tenant's session starts empty — tested by asserting every object URL is revoked and the map cleared, because a cache outliving the page would leak whatever its key. Both properties are asserted in apps/metis/web/src/components/lesson-delivery/LessonDeliveryRuntime.spec.ts (4 vitest tests passing). M8.14.f's cache_key evidence is corrected accordingly: it previously said "every @lru_cache in this service", which was true and scoped to Python, while the estate's real cache risk is a TypeScript Map a Python scan cannot see. -->
    • M8.14.d Exercise search/facets/counts/autocomplete/index lag and all standards/audit/content/evidence exports. Evidence: search/export tests.
    • M8.14.e Exercise grade/event/correction/LTI, Themis case/appeal, and analytics/cohort/privacy projections. Evidence: protected data tests.
    • M8.14.f Inspect API/UI/events/queue/blob/cache/log/trace/notification/ timing channels for identifiers, content, existence, or counts. Evidence: two-tenant suite.
  • M8.15 Generate OpenAPI and consumer contracts for Studio, learner, operator, and institutional clients. Evidence: drift gate. <!-- LEFT UNCHECKED behind .b and .d. What is done: schemas/api_contract.py 8.15.0 + services/api_contract.py; tests/test_api_contract.py 53 tests; ruff + mypy --strict clean; no new routes, OpenAPI surface unchanged. AN UNCLASSIFIED OPERATION IS A FAILURE, NOT A DEFAULT — all 767 are classified, and asking why 35 declared no security requirement is how the unauthenticated draft-lesson leak was found and fixed (see fix(metis): stop serving unpublished lesson bodies to anonymous callers). A CONTRACT THAT OMITS A STATUS THE CODE RETURNS IS ONE A CONSUMER CANNOT HONOUR: 56 modules raise a status none of their operations document — 404 forty-five times, 409 forty, 403 twenty-one, 503 five — computed by JOINING THE SOURCE TO THE DOCUMENT THROUGH THE ROUTER TABLE, because neither artefact answers it alone. Held by a ratchet in both directions, like M8.14's untenanted tables: a new gap fails, and a closed one must come off the list or it becomes folklore. Documenting all 56 is a change to ninety route modules' decorators; the honest thing one cell can do is measure it and hold the number. -->
    • M8.15.a Inventory every route/operation/audience/consumer and canonical request/response/error/event/stream schema and owner. Evidence: API inventory. <!-- metis-api-surface.json: the canonical Metis API surface, consolidating committed metis-api-inventory + metis-route-inventory + a per-type breakdown of the OpenAPI component schemas (apps/metis/service/openapi/metis.openapi.json, read once). 181 operations across 8 route families, served to audience tiers (101 privileged / 44 open / 36 authenticated), consumed by 20 distinct API consumers. Canonical schemas: 535 total classified by schemaType() with error/event/stream priority over request/response — 405 response, 68 request, 50 base, 8 event, 3 error, 1 stream. FINDING: the API is RESPONSE-DTO-HEAVY (405/535) with THIN error/event/stream schema typing (12 total) — a canonical-schema gap the M8.15.b OpenAPI drift gate should address. Single service owner (apps/metis/service). Fast (0.2s, committed-evidence reads). generate-metis-api-surface.mjs + 6 tests, wired into inventory/verify + CI. M8.15.b (OpenAPI generation + drift gate) is implementation and stays unchecked. -->
    • M8.15.b Generate OpenAPI with auth/scopes, discriminators, property redaction/absence, versions/units, pagination/idempotency/concurrency, examples, and stable errors. Evidence: schema validation. <!-- STILL UNCHECKED, but the two substantive halves are done. AUTH: every one of 767 operations is classified, the three course routes declare security: [{HTTPBearer: []}, {}] because a token changes their answer rather than gating the call, and the register carries a stated reason for each of the 32 needing none. STABLE ERRORS: this estate put two unrelated shapes on the wire — {error, message, details} from the handlers and FastAPI's bare {detail} from any route raising HTTPException — so a client parsing error.message got nothing from a 404. An HTTPException handler in main.py now adds error and message beside a verbatim detail. ADDITIVE ON PURPOSE: 47 assertions across 40 test files read detail, and rewriting the wire format is not something a documentation cell gets to do as a side effect. Verified by running all 675 API-contract tests — every one passed — and by a wire test asserting a real 403 carries all three keys. A status outside the code table gets a generic code rather than a fabricated one. Driving the contracts against the provider also found one of my own claims wrong: the authoring surface answers a refusal with a typed AuthoringOutcome, not an error envelope, so studio.forbidden is now recorded as a payload and not structured_with_detail. WHAT IS LEFT, and why this stays unchecked: no discriminators, no per-operation examples, and no unit annotations. Examples on 767 operations is a separate piece of work, not a paragraph. --> <!-- SPEC 2026-09-11: what closes .b, exactly: (1) an OpenAPI discriminator on every tagged Pydantic union, derived from the models (M10.1.c records that untagged unions cannot exist); (2) one example per operation generated from the M8.15.c consumer-contract fixtures, never hand-written; (3) x-unit annotations on the 310 unit-bearing fields M10.1.c counted. Regenerate through scripts/export_openapi.py, semantic-diff the schema and path sets, keep {detail} beside {error, message}. -->
    • M8.15.c Publish distinct Studio/learner/operator/institution consumer contracts for success/empty/partial/stale/degraded/not-configured/ forbidden/failure cases. Evidence: contract fixtures. <!-- Four contracts, each total over the eight cases and unconstructible with one missing. A case this service cannot produce carries neither a status nor a body shape, so a silently absent case cannot read as a case that cannot happen. The four are distinct as whole documents, and a separate test asserts the AUDIENCE CHANGES THE WIRE AND NOT ONLY THE PROSE: forbidden is 403 for studio, operator and institution and 404 for a learner, because a 403 confirms the thing exists. Studio and institution share a status shape, which is honest — they are both privileged — so the distinctness test compares whole expectations rather than statuses alone. -->
    • M8.15.d Verify real provider implementation against every consumer and deployed compatibility window. Evidence: provider verification. <!-- LEFT UNCHECKED, but no longer prose: tests/test_provider_verification.py drives NINE of the thirty-two audience-and-case pairs against the running app and checks each against what the contract promised — studio success/empty/stale(409)/forbidden(403)/not_configured(503), learner success/empty/forbidden(404), operator success. A CONTRACT NOTHING EXERCISES IS A DOCUMENT, and the type refuses a verification whose observed status disagrees with the promise: a mismatch is a broken promise, not a result to file. It also refuses one with neither an observed status nor a reason, so A CASE NOBODY DROVE IS RECORDED AS NOT DRIVEN, NOT AS PASSING — the report is total over all thirty-two and each of the other twenty-three names why (eight of them because the standards exchange has no HTTP surface in this build, three because driving a 500 means breaking the service on purpose). VERIFIED_FLOOR is a floor, not a target: the suite asserts at least nine, so the number can only go up. Still unchecked because "every consumer" is thirty-two, not nine, and there is no deployed compatibility window in this repository to verify against. --> <!-- SPEC 2026-09-11: there is no deployment, so the "deployed compatibility window" is defined as the committed openapi/metis.openapi.json at the previous tagged release commit and at HEAD. Drive the remaining 23 audience/case pairs, recording each undrivable pair with its reason. The eight "standards exchange has no HTTP surface" pairs need a decision — add the surface (a new cell) or record a §2.1 N/A with a named approver; an agent drafts either, a person decides. -->
    • M8.15.e Fail CI on unclassified operations, breaking schema/semantic/ auth changes, or protected property exposure. Evidence: drift gate. <!-- services/api_contract.py: drift_gate diffs the committed document against a candidate and fails on an unclassified operation or any of eight breaking-change kinds. Every kind has a test that takes the real 767-operation document, makes exactly ONE change to a copy, and requires that change to be classified — a gate nobody has seen fail is a gate nobody has tested, and a self-diff passing proves only that it is not noisy. PROTECTED PROPERTY EXPOSURE is the one that matters: a response growing an answer_key breaks no existing consumer, which is precisely why it needs a gate rather than a review, and it is parametrised over all twelve protected names with a control showing an ordinary new field is not a finding. -->

M9 — Workbench UI, calendar, localization, catalog, and publication#

  • M9.1 Build the Metis authoring hub around assigned work, recent courses, gate failures, review queue, and scheduled releases. Evidence: utility-first design review.
    • M9.1.a Define audience-safe live view models, priority/ownership/ freshness, actions, and navigation for every named hub region. Evidence: hub contract. <!-- Verified 2026-09-06: schemas/authoring_hub.py 9.1.0 + services/authoring_hub.py; tests/test_authoring_hub.py 62 tests; ruff + mypy --strict clean; no routes yet, so the OpenAPI surface is unchanged. One shape over all five regions, because five bespoke shapes would be five places to forget the count rule. A REGION'S COUNT IS A COUNT OF WHAT THIS VIEWER MAY SEE — M8.13's rule where it is most visible, since a hub's whole job is numbers and "14 gate failures" from a page that filtered eleven is a lie told in large type. A region carrying more cards than it counts is unconstructible, and an empty one counts nothing. AN ACTION A VIEWER MAY NOT TAKE IS ABSENT, NOT DISABLED: there is no enabled field to set false, grants_for returns only what this viewer may do, and the hub refuses a card offering an action its audience may never hold. Every mutating action carries its confirmation and whether it needs a reason ON THE GRANT, so the dialogue and the audit trail cannot disagree; open carries neither, because treating a look like a change trains people to click through. A card that can be acted on can also be opened — acting on something you cannot open is acting blind. PRIORITY IS DERIVED, NOT TYPED, from ownership, state and age. Work waiting on somebody else does not age into due, or another team's queue climbs to the top of this page; unassigned work is not late but is not somebody else's problem either. Every priority value is checked to be producible from some combination. A REGION THAT CANNOT SAY WHEN IT WAS COMPUTED SAYS SO: freshness is required, unknown is one of its values, and a confident freshness with no timestamp is unconstructible. empty (an answer with nothing in it), offline (no answer) and loading are kept apart, and a ready region with no cards is refused as an empty one — "nothing to do" and "still loading" are different sentences. Ordering is deterministic down to the card id, so the same corpus in a different order renders the same list; a hub that reshuffles between reloads is one nobody can work down. services/authoring_hub_sources.py wires four of the five regions to real tenant-scoped tables (store_aggregates joined to store_revisions, gate_job, the content-review signoffs, publication_saga_run) with tests/test_authoring_hub_sources_real_database.py 15 tests against a real PostgreSQL. A REGION WITH NO SCOPEABLE SOURCE IS OFFLINE, NOT EMPTY: recent_courses would read courses, one of the thirteen M8.14 found untenanted, so it returns offline with that reason in the message rather than claiming there are no recent courses. assigned_work is degraded for a smaller reason — the aggregate store records who revised a thing and nothing about who it is assigned to, so ownership means "last touched by" and the message says so instead of letting the page claim an assignment model the estate does not have. Pyright caught three readers taking a viewer_id they never read; the parameter is gone, because AN UNUSED PARAMETER IN A DECISION FUNCTION IS A SILENT STUB. Two tenants seeded with colliding ids across all four sources see one row each. -->
    • M9.1.b Compose a restrained task-first workspace with assigned blockers primary, recent context/reviews/schedule secondary, and no decorative hero or repeated card grid. Evidence: responsive design review. <!-- apps/metis/admin/src/app/hub/{page.tsx,page.module.css}, served at /hub. A NEW surface: the existing / dashboard is a different page with its own specs (visual-rhythm asserts it is "Dashboard" with four stat cards) and was not replaced. Two unequal columns rather than a grid — 3fr primary against 2fr secondary — because a grid of equal tiles says every region matters the same amount, which is the opposite of this page's claim. The browser asserts it rather than a reviewer: the primary column's measured boundingBox is wider than the secondary's, and the assigned-work region is wider than the gate region. No hero and no images at all. -->
    • M9.1.c Implement deterministic filters/sorts/counts/deep links/live updates and loading/empty/stale/degraded/offline states. Evidence: interaction tests. <!-- All six region states render a different sentence, asserted one by one in the browser; the one region with nothing to explain renders no message. AN OFFLINE REGION IS NOT AN EMPTY ONE — offline reports "age unknown" where empty reports its freshness. FILTERING NARROWS WHAT IS DRAWN AND LEAVES THE COUNT ALONE: with the blocking filter on, the assigned-work region shows zero cards and still reports 3, because a count that moved with the filter would be counting the page. Ordering is deterministic server-side down to the card id. Deep links carry the exact revision the card was built from. -->
    • M9.1.d Gate inline claim/reassign/review/remediate actions by current role/state/version with reason/confirmation/audit. Evidence: action tests. <!-- AN ACTION A VIEWER MAY NOT TAKE IS NOT ON THE PAGE AT ALL. The server sends only the grants this viewer holds and the page renders only those: a card carrying just open has no buttons, and a browser assertion requires zero button[disabled] inside any card — a greyed-out control would still say the thing exists and that somebody else may move it. Every mutating action opens a confirmation carrying the server's own sentence, and one that requires a reason cannot be confirmed until one is typed (the confirm button is disabled, which is an incomplete form rather than a withheld action, and the test scopes the no-disabled rule to cards for exactly that reason). Cancelling records nothing. -->
    • M9.1.e Verify keyboard/screen-reader/zoom/responsive and production-scale task/course/review volumes. Evidence: utility-first Playwright journey. <!-- e2e/authoring-hub.spec.ts, 20 tests, chromium, workers=1, all passing. Keyboard: tab reaches the next filter and Enter operates it. Focus is visibly ringed (computed outlineStyle is not none). Screen reader: the dialogue is a real role="dialog" with aria-modal, the filter group is a named role="group", and the five regions are five level-two headings so the page is navigable by heading. Responsive: at 720px there is no horizontal overflow and the work still comes first — narrow does not mean reordered. Zoom: 200% at 1280 (i.e. 640 CSS pixels) keeps the primary region visible with no overflow. Production scale: a region of 20 drawn cards reporting 4,137 visible renders without overflow and still says "Showing 20 of 4137". The 503 case renders the alert and NOT half a page beside it. One locator bug of my own, found by the browser: getByRole('alert') also matches Next's route announcer, so the assertion is scoped to the page's own error element. -->
  • M9.2 Build inbox capture for sources, curriculum briefs, lesson ideas, item ideas, and feedback. Evidence: keyboard-first journey.
    • M9.2.a Define one inbox item contract with typed capture kind, raw input, source/provenance, context, owner, status, validation, and conversion linkage. Evidence: inbox schema.
    • M9.2.b Implement keyboard-first quick capture for each named kind with paste/upload/link, minimal required fields, shortcuts, autosave, and accessible feedback. Evidence: capture tests.
    • M9.2.c Quarantine/scan/fence external content, protect learner feedback, label AI/imported sources, and prevent raw capture from publishing. Evidence: intake boundary tests.
    • M9.2.d Triage/search/filter/assign/defer/archive/deduplicate and convert into canonical source/brief/lesson/item/feedback records as explicit new revisions. Evidence: conversion tests.
    • M9.2.e Verify offline/reconnect/duplicate/conflict/permission/large-paste keyboard-only journey. Evidence: Playwright E2E.
  • M9.3 Build course and item workspaces using shared shell primitives. Evidence: route/selection/deep-link restoration.
    • M9.3.a Define canonical routes and shared shell layout for navigator/ primary editor/inspector/status/command palette across course and item workspaces. Evidence: IA contract.
    • M9.3.b Reuse shared selection/revision/dirty/save/conflict/capability/ permission state and avoid a second app shell or duplicated owner UI. Evidence: shell integration tests.
    • M9.3.c Compose course hierarchy/dossier/content/coverage/review/build routes and item stem/response/key/rationale/accessibility/gate/use routes. Evidence: workspace tests.
    • M9.3.d Persist/restore tenant/course/item/revision/selected entity/panel/ filter/focus through authorized URLs and back/forward/reload. Evidence: restoration tests.
    • M9.3.e Verify direct link, deleted/stale/forbidden target, unsaved navigation, responsive modes, accessibility, and route ownership. Evidence: Playwright suite.
  • M9.4 Build source registry and grounding workspace. Evidence: unsourced high-risk claim block-to-clear journey.
    • M9.4.a Build permission-aware source/extraction/claim/citation search and detail with authority/rights/currency/jurisdiction/audience/retraction state. Evidence: registry tests.
    • M9.4.b Build extraction/original comparison, typed locator viewer, claim-span linking, source conflict, update/retraction, and version history. Evidence: grounding interaction tests.
    • M9.4.c Build high-risk claim queue showing risk/source requirements, support/sufficiency/faithfulness/conflict gates, uncertainty, and reviewer role. Evidence: gate view tests.
    • M9.4.d Implement attach/replace/corroborate/edit/abstain/request-review actions with exact revisions and blast-radius preview. Evidence: remediation tests.
    • M9.4.e Run unsourced high-risk claim through authoritative citation, human review, cleared gate, rebuild, and stale-on-retraction journey. Evidence: Playwright E2E.
  • M9.5 Build objective, standard, concept, and prerequisite graph workspace with semantic list parity. Evidence: accessible graph journey.
    • M9.5.a Build typed graph projection with stable objective/concept/ standard nodes, prerequisite/coverage/alignment edges, revisions, provenance, validation, and permissions. Evidence: graph contract.
    • M9.5.b Implement visual create/link/unlink/select/filter/layout/compare with no screen-position-as-data and explicit validation. Evidence: graph interaction tests.
    • M9.5.c Provide synchronized tree/list/table/forms for every inspect/add/ link/edit/delete/reorder/navigation operation. Evidence: semantic parity tests.
    • M9.5.d Surface cycles/orphans/unreachable/missing prerequisites/ standards drift/coverage impacts and non-destructive remediation. Evidence: validation tests.
    • M9.5.e Verify keyboard/screen reader/non-drag/zoom/responsive/large graph and direct authorization journey. Evidence: accessible E2E.
  • M9.6 Build psychometric and DIF review workspace. Evidence: raw sample, uncertainty, policy, and cohort privacy are visible.
    • M9.6.a Build permission-filtered item/calibration/run queue/detail with exact sample/method/model/tool/item versions and eligibility status. Evidence: projection tests.
    • M9.6.b Present raw CTT/IRT parameters, standard errors, fit/information, sample/data quality, exposure, and declared-versus-empirical comparison. Evidence: measurement view tests.
    • M9.6.c Present DIF method/reference-focal cohorts/effect/significance/ uncertainty/sample counts with privacy suppression and no individual inference. Evidence: DIF view tests.
    • M9.6.d Show thresholds/policy/verdicts/manual bands/disagreement/ unavailable state separately from raw results. Evidence: state tests.
    • M9.6.e Implement request-rerun/promote/restrict/revise/retire/review decisions with role separation and immutable evidence. Evidence: review journey.
  • M9.7 Build integrity and appeals deep links without duplicating /operator/metis. Evidence: role-directed navigation.
    • M9.7.a Define allowed link intents from course/item/attempt/grade/support contexts to authoritative Themis/operator case/appeal views and safe return paths. Evidence: route map.
    • M9.7.b Issue/validate signed tenant/subject/context/intent/expiry/nonce links and reauthorize current actor/object/function at destination. Evidence: deep-link tests.
    • M9.7.c Show only audience-safe status/explanation/next action in Metis; keep raw evidence/adjudication/appeal decisions in owner surface. Evidence: property boundary tests.
    • M9.7.d Route learner/teacher/integrity reviewer/operator actors to their permitted views and handle logged-out/expired/deleted/forbidden cases. Evidence: role navigation tests.
    • M9.7.e Scan Metis routes/components/APIs for duplicated integrity mutation UI/private contracts. Evidence: ownership gate.
  • M9.8 Build editorial calendar for course launches, assessment windows, standard updates, item refresh, and review capacity. Evidence: time-zone and committed-slot tests.
    • M9.8.a Define versioned calendar event/resource/reservation/dependency contracts for each named event type with owner/state/timezone/window/ recurrence/visibility. Evidence: calendar schema.
    • M9.8.b Build month/week/list/timeline views, filters, accessible non-drag create/move/resize, conflict indicators, and role-filtered details. Evidence: calendar UI tests.
    • M9.8.c Model launch/assessment embargo/deadline, standards effective dates, source/item freshness, and dependency DAG/backward planning. Evidence: scheduling tests.
    • M9.8.d Reserve named reviewer/approver/localizer/publisher and system build/provider/LMS capacity; reject overbooked committed slots. Evidence: capacity tests.
    • M9.8.e Test IANA time zones/DST/all-day/locale/recurrence/quiet hours/ missed dependency/committed slot concurrency. Evidence: calendar suite.
  • M9.9 Build taxonomy registry for subject, discipline, objective, standard, credential, modality, level, and audience. Evidence: impact preview.
    • M9.9.a Define versioned typed entries/relations/aliases/locales/ ownership/lifecycle and code-owned floors per named taxonomy. Evidence: registry schemas.
    • M9.9.b Build permissioned author/edit/diff/review/publish/deprecate/ supersede UI with uniqueness/cycle/collision validation. Evidence: lifecycle tests.
    • M9.9.c Show exact use across courses/items/objectives/standards/builds/ credentials/search/integrations before change. Evidence: impact query.
    • M9.9.d Preview alias/deprecation/migration effects, conflicts/loss and require versioned approval; preserve historical codes. Evidence: impact preview tests.
    • M9.9.e Verify consumer fallback/unknown behavior and cross-registry drift without collapsing distinct taxonomies. Evidence: registry E2E.
  • M9.10 Build localization workflow with translation memory, glossary, MathML, media, accessibility, and parity gates. Evidence: RTL and long locale journey.
    • M9.10.a Define locale-variant job/unit/source/target/status/translator/ reviewer/version contracts and immutable source-revision binding. Evidence: localization schema.
    • M9.10.b Integrate permissioned translation memory and course glossary suggestions with source/version/match/rights and human acceptance. Evidence: TM/glossary tests.
    • M9.10.c Localize structured text, bidi, MathML, SSML, media text/tracks/ voice, descriptions, interaction labels, answers/rubrics, and metadata without breaking stable IDs. Evidence: content tests.
    • M9.10.d Run source/citation/meaning/key/scoring/terminology/layout/ accessibility/media-sync/standards-export parity gates per locale. Evidence: parity suite.
    • M9.10.e Manage source-change invalidation, translation diff/review/ approval/build and prohibit silent fallback in assessed content. Evidence: lifecycle tests.
    • M9.10.f Run RTL/long/CJK/plural/date-number/unit/voice/caption/offline learner-render journey. Evidence: locale E2E.
  • M9.11 Build course/item catalog curation via reviewed bundles. Evidence: no direct production writes.
    • M9.11.a Define versioned catalog bundle with exact course/item/build revisions, audience/locale/metadata/order/collections/schedule/targets, owner, and manifest hash. Evidence: bundle schema.
    • M9.11.b Build draft add/remove/reorder/group/feature metadata editor with eligibility, rights/gates/reviews/credential/integration checks. Evidence: curation tests.
    • M9.11.c Preview deterministic catalog/search/learner-facing diffs, removals, dependencies, visibility, and affected links. Evidence: preview tests.
    • M9.11.d Require separate review/approval and publish through idempotent release API; prohibit catalog database/direct target mutation. Evidence: boundary tests.
    • M9.11.e Preserve released versions and propagate supersession/unpublish/ source withdrawal/recall explicitly. Evidence: lifecycle E2E.
  • M9.12 Build publication candidate, approval, schedule, execute, and verified-in-target views. Evidence: failed LMS sync cannot appear published.
    • M9.12.a Build candidate manifest view with exact content/build/standards/ locale/accessibility/gates/reviews/rights/integration versions and hash. Evidence: candidate tests.
    • M9.12.b Build separate dimension approvals bound to candidate digest, reviewer roles/quorum/conflicts/conditions/freshness. Evidence: approval tests.
    • M9.12.c Build schedule/dependency/capacity/embargo/timezone/target preview and confirmation against the unchanged candidate. Evidence: schedule tests.
    • M9.12.d Build live per-target saga progress/receipts/partial/failure/ retry/compensation with immutable attempts. Evidence: execution tests.
    • M9.12.e Show published only after authoritative LMS/catalog/store read- back verifies bytes/metadata/visibility/links; expose failed/unknown honestly. Evidence: verified-target E2E.
  • M9.13 Build QTI/LMS export diagnostics with per-item incompatibilities. Evidence: no silent downgrade.
    • M9.13.a Define diagnostic model per package/assessment/section/item/ stimulus/rubric/media field with standard/profile/version, severity, support/loss/extension, locator, and remediation. Evidence: schema.
    • M9.13.b Run preflight mappings/validators/target capability probe on exact revisions before export. Evidence: preflight tests.
    • M9.13.c Present per-item unsupported interaction/scoring/accessibility/ media/locale/metadata and aggregate counts without hiding failures. Evidence: diagnostics UI tests.
    • M9.13.d Allow explicit supported substitution/exclusion only through reviewed new build revision; never silently downgrade. Evidence: substitution tests.
    • M9.13.e Link output manifest/loss report/import/read-back evidence to diagnostics and block incompatible targets. Evidence: export E2E.
  • M9.14 Build command palette and Studio IA registration. Evidence: route integrity and permission filtering.
    • M9.14.a Register Metis routes/actions/entities/keywords/shortcuts through canonical Studio IA with stable IDs/audiences/capabilities. Evidence: registration tests.
    • M9.14.b Build permission/capability/context-filtered navigation and typed non-destructive commands with current selection/revision. Evidence: palette tests.
    • M9.14.c Exclude unauthorized/disabled/operator/learner/protected-answer commands from discovery and execution server-side. Evidence: auth tests.
    • M9.14.d Implement keyboard search/grouping/help/focus return and localized accessible labels without shortcut conflicts. Evidence: accessibility tests.
    • M9.14.e Verify every entry resolves to an owned canonical route/action and stale/deprecated registrations fail CI. Evidence: route integrity gate.
  • M9.15 Build responsive reviewer mode for tablet/mobile without exposing answer keys to unauthorized contexts. Evidence: viewport and auth tests.
    • M9.15.a Define mobile review scope/layout for subject identity, preview, sources/gates/diffs/comments, decision, and escalation; exclude complex authoring. Evidence: responsive spec.
    • M9.15.b Build media/content/item/rubric semantic inspection and anchored comment flows with stable revision context. Evidence: mobile tests.
    • M9.15.c Project answer/key/rationale/integrity/learner fields by current role/purpose and remove them from payloads/DOM/cache, not CSS alone. Evidence: property tests.
    • M9.15.d Implement approve/reject/request-changes/abstain only for eligible roles/states with evidence, reason, confirmation, and stale conflicts. Evidence: decision tests.
    • M9.15.e Verify touch/keyboard/screen reader/zoom/orientation/slow-offline/ direct URL/tier-change viewports. Evidence: responsive auth E2E.
  • M9.16 Build all loading, empty, unauthorized, stale, offline, conflict, not-configured, degraded, and failed states. Evidence: UI state matrix.
    • M9.16.a Define canonical state discriminators, safe messages, available/ forbidden actions, timestamps, recovery, announcements, and telemetry for each named state. Evidence: state contract.
    • M9.16.b Implement route/list/detail/editor/preview/media/job/review/ publication/integration variants without fabricated data or false green. Evidence: component matrix.
    • M9.16.c Distinguish unauthorized from authenticated-empty safely, stale from current, offline cache from live, conflict from failure, and degraded from not-configured. Evidence: semantic tests.
    • M9.16.d Preserve user input/focus/context where safe through retry, reconnect, reauth, refresh, conflict resolution, and service recovery. Evidence: recovery tests.
    • M9.16.e Exercise every critical route/state/theme/viewport/locale with accessible announcements and permission assertions. Evidence: UI state Playwright matrix.
  • M9.17 Build lecture/media production, render, critic, accessibility, and rendition workspaces with per-segment status and replacement. Evidence: authoring-to-real-learner-renderer journey.
    • M9.17.a Build lecture/media project list/detail with script/segment/ artifact/rendition versions, source/objective links, owner, cost, gates, review, and package status. Evidence: workspace contract.
    • M9.17.b Build per-segment plan/spec/code/data/media/presenter/consent/ render job timeline with progress/logs/outputs/failures/cancel/retry. Evidence: production tests. Cancel and retry are version-fenced tickets naming the exact request against the service that owns the production; this surface reads productions and does not write them.
    • M9.17.c Build synchronized media preview and visual/layout/correctness/ source/accessibility/technical critic findings with attributed versions and educator decisions. Evidence: critic workflow tests.
    • M9.17.d Build captions/transcript/audio description/sign/translation/ pronunciation and rendition status/diff/review across locales/targets. Evidence: accessibility/rendition tests.
    • M9.17.e Build surgical replacement impact preview, regenerate/edit, unchanged identity proof, invalidated gates/approvals, and new package comparison. Evidence: replacement tests.
    • M9.17.f Run lesson-to-media-to-reviewed-package-to-real learner renderer with failure/cancel/consent/source withdrawal/accessibility cases. Evidence: Playwright E2E — 18 admin walks plus a member-web walk of the same package through LecturePackageRenderer, served from a playback fixture the service generates from the production the workspace judged.
  • M9.18 Build credential template, pathway, candidate evidence, review, issue, verify, expire, revoke, replace, and audit views with issuer-role boundaries. Evidence: Open Badges verification and learner-privacy tests.
    • M9.18.a Build template/pathway authoring with criteria/evidence/ achievement/alignment/issuer/validity/locale/version/diff/lifecycle and floor validation. Evidence: authoring tests.
    • M9.18.b Build candidate queue/detail with privacy-minimized subject, per-criterion exact evidence/current/missing/stale/conflict state, eligibility calculation, and requester. Evidence: candidate tests.
    • M9.18.c Build independent reviewer/issuer approve/reject/request-evidence decisions with authority/conflict/quorum/step-up and locked subject. Evidence: review tests.
    • M9.18.d Build issue/sign/deliver/receipt and public/private verifier views for signature/issuer/recipient/evidence/status with privacy-safe results. Evidence: Open Badges verification tests. The signed document itself is the estate's existing OB 3.0 assertion builder; what this cell adds is the durable status a verifier reads and the audience projection around it.
    • M9.18.e Build expiry/renew/revoke/replace/supersede status/impact/ notification and immutable lineage/audit. Evidence: lifecycle tests.
    • M9.18.f Prove authors/course completion/agents cannot issue or revoke, and verify learner/guardian/public/institution field boundaries. Evidence: authority/privacy E2E.
  • M9.19 Build institutional integration operations for roster sync, deep links, grade services, exports, webhooks, credentials, failures, replay, and reconciliation without duplicating operator incident tooling. Evidence: role-directed, live-status journey.
    • M9.19.a Build institution-scoped connection/status/capability/config/ credential-ref/version/owner views for LTI, OneRoster, LMS, LRS, webhook, and credential integrations. Evidence: connection tests. Capabilities are probed with a shelf life; an unprobed one is unknown, never supported, and the status is derived from the runs rather than from the connection's own summary columns.
    • M9.19.b Build roster sync and deep-link/launch operation views with per-run/resource progress, mappings, changes, conflicts, diagnostics, and receipts. Evidence: roster/link tests.
    • M9.19.c Build grade service/current-correction/outbound status and exports/webhooks/credential event delivery views with privacy/redaction/ idempotency. Evidence: data-operation tests.
    • M9.19.d Build failure/partial/unknown/retry/replay/reconcile controls with authoritative read-back, impact, expected version, role, and audit. Evidence: recovery tests. This surface plans and enqueues; the delivery itself is the integration worker's, and a recovery answers 202 with a queued operation and no receipt rather than claiming a far side it has not reached.
    • M9.19.e Route platform security/outage/incident/drain/kill operations to operator tools through governed links; do not duplicate them here. Evidence: ownership tests.
    • M9.19.f Run institution admin/teacher/operator journeys across live success, version skew, duplicate/late event, outage, recovery, and tenant isolation. Evidence: role-directed E2E.

M10 — Metis verification, operations, and launch#

<!-- SPEC 2026-09-11: M10 is bound by §2.5. Journeys name the host per stage (§2.5.1), contain no page.route (§2.5.2), run against the venv + METIS_TEST_DATABASE_URL stack (§2.5.3), bind evaluators to OpenRouter (§2.5.4), and treat M10.20 as human-only (§2.5.9). The JWT role model is three values — student (the default when the claim is absent), instructor, admin (deps.py CurrentUser) — so any cell that separates reviewer/decider/appeal/grade roles (M10.11, M10.12, M10.22) separates them by actor id through the M7 assignment, conflict-of-interest and signoff records, never by JWT role alone. -->

  • M10.1 Run contract tests for every canonical and workbench schema. Evidence: no unreviewed snapshot updates.
    • M10.1.a Generate the complete canonical/workbench entity/API/event/job/ bundle/standard schema inventory with owner/version/consumer/test. Evidence: coverage manifest — openapi/schema-inventory.json, 3,907 models across 258 modules, generated by importing every schema, route and service module rather than declared by hand.
    • M10.1.b Test public parse/validation/serialization/round-trip for minimal, full, locale/accessibility, protected-field, unknown/invalid fixtures. Evidence: contract suites. The estate-wide suite covers minimal, round-trip and unknown-field for the 2,560 models a minimal instance can be synthesized for, with the declines and refusals counted; full, locale, accessibility and protected-field fixtures are the per-cell contract modules', and the manifest records which models have one.
    • M10.1.c Test cross-package/API/database/standards adapters for semantic field/version/unit/optional/discriminator compatibility. Evidence: adapter tests. Database: every mapped model's table is created by a migration, read from the migration's syntax tree. Units: 310 unit-bearing fields hold quantities. Optional: required-field sets are recorded per model so a diff shows a breaking addition. Discriminator: Pydantic refuses a union with an untagged member at class-definition time, so the guarantee is recorded rather than duplicated. Cross-package is pnpm --filter @oshun/codegen run openapi:types:check.
    • M10.1.d Fail missing/duplicate/unowned/unmapped schemas and require reviewed rationale/diff for every snapshot/golden update. Evidence: contract gate. Hard failures: a published name declared twice, a component FastAPI had to qualify, a component with no model, a model with no migration. Measured ratchets that may not grow: unowned, permissive, untested, unconsumed, unmapped tables. The manifest drift gate is what forces the reviewed diff.
  • M10.2 Run known-value tests for rubric, readability, coverage, item quality, IRT, BKT, FSRS-5, DIF, exposure, and selection logic. Evidence: function-to-test coverage index.
    • M10.2.a Inventory exact functions/versions/units/applicability and create independent positive/negative/boundary/N/A/error fixtures per named algorithm. Evidence: function-to-test index — metis/algorithm_registry.py, eleven algorithms each with its reference, units and applicability. Eight are checked by a scalar table with all five categories; three read a domain object too large for one and name the tests that carry their exact answers, which the index verifies exist.
    • M10.2.b Verify rubric arithmetic, readability/complexity, objective/ construct coverage, item clarity/distractor/key/rationale rules, and exposure thresholds. Evidence: deterministic suites. Flesch and the exposure thresholds are point cases derived from the published coefficients; the rubric's weighted score, its N/A reweighting and its half-up rounding are indexed to the four tests that already prove them.
    • M10.2.c Verify 1PL/2PL/3PL parameters/fit/SE, BKT updates, FSRS-5 scheduling, DIF methods/privacy floors, and calibration eligibility. Evidence: psychometric goldens. BKT, FSRS-5 and the 3PL information function are point cases with the arithmetic written beside each expectation. DIF is an iterative maximum-likelihood fit with no closed form to compare against, so it is registered as property-checked and indexed to the recovery, Holm-family, privacy-floor and revision-conflict tests rather than given a second implementation of its own iteration to agree with.
    • M10.2.d Verify pool/blueprint/adaptive selection constraints, tie-breaks, seeds, eligibility reasons, and no silent relaxation. Evidence: selector goldens — indexed to the six assembly tests that cover the order-independent golden, the blueprint coverage, the seeded tie-break, the unrelaxed enemy and stimulus constraints, the infeasible blueprint and the fail-closed pool eligibility.
    • M10.2.e Fail coverage for any active algorithm/version without current fixture/tolerance/source derivation. Evidence: coverage gate. Every algorithm pins a digest of its own logic, taken from the syntax tree with docstrings stripped, so a reformat changes nothing and an altered branch or constant fails the gate until the values are re-derived. A tolerance without a reason, a case with no derivation, a category with no case and an algorithm with nothing checking it each fail on their own. Both halves were verified by mutation: changing FSRS_FACTOR from 19/81 to 19/80 fails five known values, and wrapping a return in a min that changes nothing for valid inputs fails only the pin.
  • M10.3 Run calibrated evaluation for AI item writing, bias assist, tutor, and integrity classifiers. Evidence: threshold and subgroup report. <!-- SPEC 2026-09-11: §2.5.4 binding and §2.5.5 corpus honesty apply. The four evaluators are the item-variant provider (services/item_variant_provider.py), bias/cultural review (services/cultural_review*), tutor preview evaluation (services/tutor_preview_evaluation.py) and the integrity signal classifiers; every run records the exact METIS_*_MODEL slug, OPENROUTER_PROVIDER_SORT, the provider that served it (from the usage payload) and the prompt version. M10.3.a's licensed/consented representative corpus cannot be produced by an agent: build the synthetic manifest, label it synthetic and un-adjudicated, and leave .a and the parent open. .b–.e are earnable on the synthetic corpus with that limitation stated in the report. --> <!-- STILL UNCHECKED — .a cannot be earned by an agent (the SPEC note above: a licensed/consented representative corpus). .b–.d are to be earned on a synthetic, un-adjudicated corpus with that limitation stated; .e is built and today refuses all four evaluators. -->
    • M10.3.a Define versioned licensed/consented representative held-out corpora, labels/rubrics, annotator/adjudication process, slices, privacy, and contamination controls for each evaluator. Evidence: corpus manifests. <!-- STILL UNCHECKED — the synthetic manifests the SPEC note asks for exist (data/m10_3_*_corpus.json: synthetic, single-labelled by "fixture construction", consent synthetic); the representative one still needs a person. BLOCKED on a corpus nobody has assembled. The manifest contract exists (schemas/evaluation_corpus.py: per-item licence, consent basis, annotators/adjudicator, declared slices, content digests for contamination) and refuses what the estate has. No held-out set exists for any of the four evaluators. The nearest in-repo corpora (m6_8/m6_17 for the tutor, m5_24/m5_25 for item writing) test deterministic engines around those AIs, not the AIs themselves. Each was written and labelled once by the session that built the engine it checks. None of the 26 corpus files records a sampling frame, annotators, adjudication or consent, and a test pins that. Bias assist and integrity classification have no corpus at all. Sourcing, consent and annotators are human work. -->
    • M10.3.b Run exact model/prompt/config/tool versions and retain per-item proposals/scores/labels/reasons/uncertainty/abstention/errors. Evidence: reproducible runs. Evidence: four runs kept whole in evidence/v1-workbenches/metis-evaluation/<evaluator>-run.json. They come from scripts/run_m10_3_evaluations.py, services/evaluation_runs.py and the synthetic manifests data/m10_3_*_corpus.json. RE-MEASURED 2026-09-11: only the item-variant provider calls a model. - cultural_review* is human quorum logic, so the bias assist measured is lesson_inclusivity_helpers. - The tutor preview is canned text. - Integrity is keyword rules. What each run records: - The model run: the dated slug, price sort, reasoning disabled, and the prompt and config digests. Per item it keeps the served model@provider, cost, latency and the whole proposal. - Deterministic runs are versioned by the digest of their deciding code. Reproducibility, all tested: - deterministic runs re-run byte-for-byte; - every item-writing judgement re-derives from its retained proposal; - all four round-trip through the M10.3.e store on a real database. The first real run is kept (item_writing-run-20260911T163341Z-prompt-v1-all-malformed.json): all 24 answers came back malformed. The prompt named an internal class instead of its fields, and the reasoning model spent all 1,500 output tokens thinking. Both are fixed (prompt v2, METIS_ITEM_VARIANT_MODEL_REASONING), and a length finish is now reported as a truncation. Every manifest states it is synthetic.
    • M10.3.c Report task-appropriate correctness/validity/harm/precision/ recall/FPR/FNR/calibration plus subgroup/language/discipline/item-kind slices with confidence intervals. Evidence: reports. Evidence: <evaluator>-report.json and REPORT.md, computed by services/evaluation_reports.py from the committed runs and manifests only. - Item writing: validity 21/24 = 0.88 [0.69, 0.96], with one key changed and one given away. An earlier run on the same route scored 18/23. - Bias assist: recall 12/30, FPR 6/12, ECE 0.31 over its 18 flags. It missed every paraphrase and flagged every refuting quotation. - Tutor preview: 46/46. - Integrity: recall 4/22 = 0.18 [0.07, 0.39]. All six verbatim copies of the product's own tutor lines passed, as did all eight model-written answers. Slices: - item writing: discipline, locale (en/es), item kind, pressure, transformation; - bias: category, phrasing; - integrity: action, length, URL; - tutor: dimension. Every slice has its size and a Wilson interval, and all are under 20 items, so each is marked as unable to carry a conclusion. Calibration is reported only where a confidence is stated. The gate refuses all four.
    • M10.3.d Select fail/review/pass thresholds against ratified criteria, validate held-out performance, and document limitations/manual bands. Evidence: threshold decisions. <!-- STILL UNCHECKED — BLOCKED on ratification, which is a person's act (§2.5.9). The readiness packet is ready with the approver blank: evidence/v1-workbenches/metis-evaluation/thresholds-readiness.json, from services/evaluation_thresholds.py. - It proposes criteria. - It chooses bands on a fixed 60/40 split, stratified by label (by dimension for the tutor), and checks them on the held-out 40%. - It proposes an automatic band only when that band's precision lower bound clears the floor on the selection half. No band earns one today. The item writer and tutor preview state no confidence, and the bias and integrity rules' calls never reach a 0.90 lower bound. So the manual band is 100% for all four, and the packet says why. The gate already refuses criteria ratified after the run began, a decision made under other criteria or before the run ended, and a review-band cost the run does not reproduce. -->
    • M10.3.e Define drift/champion-challenger/recalibration triggers and block launch on unresolved material subgroup or safety failure. Evidence: evaluation gate. Evidence: services/evaluation_gate.py, with 72 tests in tests/test_evaluation_gate.py. The gate is given a corpus and a run and measures them itself. It takes no report, so a caller cannot hand it figures. What it refuses: - a missing or late ratification, or a decision made under other criteria or against another run; - a serving version different from the one measured; - a convenience or synthetic sample, or more than 10% single-annotated labels; - contamination, or skipped, foreign or errored items; - a floor missed on the interval's lower bound; - any safety finding (there is no way to accept one); - any material subgroup gap nobody answered for in this run. Slices too small, or holding too few positives to compare a miss rate, are marked unchecked. So are runs that mostly abstain. A pass carrying any of these cannot be constructed. drift() calls for recalibration on movement past the trigger or age past the limit. It refuses comparisons across corpora or against the wrong champion, and flags a challenger that is better on average but worse for one slice. promotion_refusals blocks on any regression. assert_evaluators_may_launch raises unless all four evaluators pass, and one fully passing launch is tested. The evidence is durable (migration 069, services/evaluation_store.py, 13 real-database tests). Foreign keys carry the id and the digest, so a run of changed labels, or a decision on a run that never happened, cannot be written. M8.4's trigger refuses UPDATE and DELETE, every read re-checks the digest, and evidence_for_decision feeds the gate from stored records. Today it refuses all four: no corpus or run exists (see .a). NOT WIRED INTO SERVING: nothing in the product calls the gate. The tutor answers learners through its OpenRouter model with no verdict behind it, and Metis has no launch switch to put the gate in. Whether the tutor should stop serving until it passes is a product decision.
  • M10.4 Run real-database and object-store integration tests. Evidence: source-to-course durable round-trip. <!-- SPEC 2026-09-11: harness = the §2.5.3 stack. Two tenants seeded with colliding ids following tests/test_tenant_isolation_real_database.py; object store = MinIO from docker/docker-compose.dev.yml (port 9000, METIS_S3_ENDPOINT); queue/outbox = what the service actually runs (Redis/Celery). Where no search index exists for an entity, the cell records "no index" rather than asserting one. Every evidence block reports passed/skipped/failed counts; a run whose real-database tests were skipped is a failed run. --> OPEN 2026-09-11, with all five children closed. The durable round-trip runs through services/source_pipeline.py, and no HTTP route calls it. The product's source ingestion routes, and about 20 other services, still write source_ingestion_storage_dir on local disk (M8.1's gap map's file store). So the path a user's source takes is not the one proven durable. Closing this needs those routes cut over to the pipeline. services/durable_reconciler.py also has no scheduled caller. Cells: claim, citation and release have no index, and the relay reports them not_indexed. RE-VERIFIED 2026-09-12 AND THE BLOCKER HOLDS: no module under src/metis/api/ mentions source_pipeline, and nothing outside durable_reconciler.py mentions the reconciler, so neither has gained a caller. THE CUTOVER IS LARGER THAN "ABOUT 20". 68 modules under src/metis/services/ reference source_ingestion_storage_dir and 65 of those also contain a filesystem write. That count is by grep and is an upper bound — it does not prove every write targets that directory — but it is far enough above twenty to change the size of the job, so it is recorded with its method rather than swapped in as a corrected figure. 2026-09-18: TWO OF THE THREE BLOCKERS ARE CLOSED; THE PARENT STAYS OPEN. What moved: - The source's own leg is cut over. POST /api/source-ingestion now promotes every item whose admitted bytes the service holds onto the durable path BEFORE the manifest is written (services/source_durable_intake.py, services/durable_runtime.py): the plaintext the quarantine service scanned becomes a content-addressed blob in the new s3_bucket_sources bucket, the source record an aggregate revision naming it, and the extracted text a second blob and revision. The item's durable_source says where it landed, and its absence means it did not land - it is never filled in from the intake record. The scan is CARRIED, not repeated: ingest_scanned takes the admission's verdict with its receipt, policy and scanner-set digests, because re-scanning here would let a host with no ClamAV overrule a real clean verdict with unavailable. A missing or unreachable bucket is a 503 that registers nothing, never a fall back to the local directory; the pipeline is built lazily, so an intake the service itself refuses still answers 409 for that reason. Evidence: tests/test_source_ingestion_durable_real_stack.py, 3 tests passing on a real stack (0 skipped), reading back through a second engine and a second S3 client. 2 of the 3 were re-run against the route with the promotion removed and failed, so they are not vacuous. The doubles are a stub clamscan at the ClamAV boundary and FastAPI's own dependency override for the caller; the quarantine service, the ingestion service, the pipeline, PostgreSQL and MinIO are real. - The reconciler has a caller: metis.tasks.durable.reconcile_estate, on beat hourly, over every tenant the store holds a record for and every publication run each has. It repairs nothing, by construction. Evidence: tests/test_durable_reconcile_task_real_stack.py, 3 tests passing on a real stack - it reports the owed outbox events and the unindexed source, says the same thing twice (so it repaired nothing), goes clean after a drain, and names a blob deleted out from under an acknowledged revision. WHAT IS STILL OPEN, and why the parent is not checked: - The journey stops at the source. ground and build_course still have no product caller, so claims, citations and the course are reachable only from a test. The parent's evidence sentence is a source-to-COURSE round trip, and half of it is not on the product path. - The ingestion package manifest is still a JSON file on local disk, so GET /api/source-ingestion still reads the file store. - The 68/65 file-store figures above are unchanged, re-measured 2026-09-18 by the same method.
    • M10.4.a Provision isolated real migrated database/object storage/search/ queue/outbox with two tenants and deterministic fixture/cleanup. Evidence: harness. Evidence: apps/metis/service/tests/real_stack.py, with 12 tests across tests/test_real_stack.py and tests/test_search_store_real_database.py, all passing and none skipped. Run with METIS_TEST_DATABASE_URL, METIS_TEST_S3_ENDPOINT and METIS_TEST_REDIS_URL set; without them the tests skip and name what is missing. Each stack is: - its own PostgreSQL database, migrated to head by the service's alembic. The test checks alembic_version against the script head. The outbox (store_outbox) lives there, and so does search: this estate has no engine, so the M8.13 index is now durable in search_index_document (migration 070, services/search_store.py). M8.1's gap map moved the search_index contract from absent to payload_backed; - two MinIO buckets. The server itself refuses bytes that do not match the digest sent with them; - a Redis key prefix for the queue; - two tenants whose ids derive from a seed. The fixture is deterministic: two stacks provisioned at once share nothing, yet produce identical fixture digests. Provisioning is all-or-nothing: a stack failing at its second bucket is watched undoing the database and the first bucket. Teardown is checked: it reports what it could not remove, shown by making removal fail. The search store keeps revision order over arrival, keeps tombstones, stops serving quarantined documents, survives a new engine, serialises two appliers on the row lock, and scopes by tenant. Its product writer, the outbox relay, belongs to M10.4.b/.c.
    • M10.4.b Ingest/quarantine/store/hash/extract/version a source and persist claims/citations/objectives/course/lesson/item/rubric/reviews/build. Evidence: source-to-course test. Evidence: tests/test_source_to_course_real_stack.py, 3 tests passing on a real stack (0 skipped). RE-MEASURED 2026-09-11: every stage of the existing source pipeline wrote JSON files under /tmp. None of it was durable, so the durable journey was built: - services/blob_store.py: content-addressed per tenant, the server verifies each ChecksumSHA256, a HEAD read-back precedes every receipt, and reads re-hash. - services/source_pipeline.py: two scans (the structural scanner plus a malware scanner), then bytes before the record that names them, a source revision, then extraction to a derivative blob and a second source revision. - Claims come from the deterministic M5.19 decomposer, linked to sentence spans of the stored text. - A citation binds each claim to its span, and outside works (URL/DOI) are cited separately. - Then objective, course, lesson, item, rubric, review, and a release whose manifest is itself a verified blob. Claims and citations are now stored aggregates (migration 071). AuthorableKind keeps the authoring API and generation jobs at the original nine, since a claim becomes canonical only by a human decision. Everything is read back through a second engine: - both revisions; the bytes and derivative by digest; each span's quote by digest; - the manifest equals the release; - the source retraction walk reaches every record built; - one audit and one outbox row per write, with nothing in the other tenant. With the scanner unavailable or reporting infection, the source is stored quarantined and extraction refuses it. The HTTP ingestion routes still write the file store. Moving them onto this pipeline is a cutover.
    • M10.4.c Verify transactions/audit/outbox/restart/idempotency/concurrency, exact revisions, blobs/derivatives, search index, and tenant isolation. Evidence: durability tests. Evidence: tests/test_durability_real_stack.py, 13 tests passing on a real stack (0 skipped), plus the M10.4.a/.b suites; 428 pass across the store-dependent suites. RE-MEASURED: nothing ever drained the outbox (published_at was never set), so services/outbox_relay.py was built. Each event's index change and its delivery mark commit in one savepoint; claims use SKIP LOCKED; the relay reads the exact revision an event names and indexes only its index map. Proven: - A relay killed after applying leaves the event undelivered and the index unchanged, then redelivers. - Two relays at once deliver disjoint sets that together cover every event. - A poison event fails loudly and does not starve the events behind it. - A withdrawal tombstones the document. A learner's search finds the lesson, and the item's answer_key never reaches the index. - A write referencing a missing revision leaves no row in any of four tables. - A re-run step writes nothing, and the same bytes stay one object. - Every old revision reads back exactly; each audit entry names its actor; a changed or missing object is refused. - The derivative equals a fresh extraction of the stored original. - Two tenants with identical ids share nothing across revisions, outbox, audit, index or blob prefixes. FIXED a live M8.3 defect: the loser of a race for one revision was told unknown_aggregate instead of revision_conflict. The cause was FOR UPDATE OF a on a join, which Postgres re-checked against the old revision row after the lock wait. The race test caught it before the fix.
    • M10.4.d Publish/write/read-back then source retract/quarantine/restore and reconcile every owner/store/index. Evidence: lifecycle round-trip. Evidence: tests/test_lifecycle_real_stack.py, 5 tests passing on a real stack (0 skipped), plus 3 restore tests in tests/test_source_graph_real_database.py and 2 in tests/test_search_index.py. One course is published through the M8.10 saga to the exports bucket by services/object_store_publication.py (S3 destination; prerequisites read from the stores) and read back from the bucket. Then: - its source's rights lapse (M8.5 retraction): the export becomes a reconciler finding, compensation hides it without deleting it, and prerequisites report rights as the only move; - the source is quarantined: extraction refuses and the index stops serving it; - the source is restored (source_graph.restore, a restore index action, migration 072): owners' queues empty and search serves the lesson again; - the course is published again. After each step, services/durable_reconciler.py must find nothing across blobs, index, outbox, exports and owner queues. It is shown to name each kind of disagreement it looks for when one is injected. FIXED (latent until something resolves a propagation): owner_queue and impact_preview read resolved rows as open. The relay, reconciler and destination have no product caller yet: the HTTP ingestion cutover is open (M10.4.b), and Celery wiring is .e's queue.
    • M10.4.e Inject commit/blob/outbox/queue/search/process failures and prove no acknowledged orphan/false success. Evidence: fault suite. Evidence: tests/test_faults_real_stack.py, 10 tests passing on a real stack (0 skipped, 0 failed). With the lifecycle, durability, harness, search and saga suites, 229 pass. The faults are real: - PostgreSQL triggers raise at the revision (deferred to COMMIT), outbox and audit rows. No row lands in any of the four tables. The bytes, written first, are reported as an unacknowledged orphan, and a retry adopts them. - An unreachable object store: no record names the bytes. - A trigger on the index. Only that event stays undelivered. - A deferred trigger on the relay's delivery mark: nothing is marked delivered and nothing is indexed. - A dead broker: no event is lost, and the reconciler reports every one still owed. - The service's own Celery worker is SIGKILLed while blocked mid-drain, with the course's event already applied and uncommitted (its row is locked). Nothing is marked, the write rolls back, and a fresh worker delivers all 3 exactly once. - A publisher is SIGKILLed after MinIO stored its PUT; a relay withholds the answer. The step stays pending. Resuming adopts the object with 0 writes, leaving one object. - A down destination never reports success; it recovers with one write. The relay now runs as metis.tasks.durable.drain_outbox, scheduled by beat on queue metis.durable. Before, nothing in the service drained the outbox. FIXED: - One index error rolled back the relay's whole batch on every pass, so delivery stopped for everyone (the test failed before the fix). - Same-instant revisions were delivered in random-UUID order (ignored_older; the .c test failed 4 of 4 alone). - Via e884e9ff9fa: every authoring-API edit nulled the aggregate's course, and a course-scoped caller could revise another course's aggregate.
  • M10.5 Run QTI 3 import/export goldens, including accessibility features. Evidence: external validator or pinned conformance tool results. <!-- SPEC 2026-09-11: same external gate as M1.11.g — close them together or not at all (§2.5.6). .a–.c and .e are earnable from libs/contracts/src/metis/qti-package.ts, services/qti*, tests/test_qti_*; .d needs the two pinned validators' own output over our export: 1edtech-qti3-xsd = the public QTI 3.0 XSD bundle validated with the xmlschema package (pin the download URL and sha256 under apps/metis/service/tools/qti3/), longsight-qti3-cli = installed from its public repository at a pinned commit. PNP, MathML and caption features must survive the validator, not only our own parser. -->
    • M10.5.a Enumerate supported QTI profile/item-kind/scoring/feedback/ stimulus/media/locale/accessibility cells and explicit unsupported ones. Evidence: coverage manifest. Evidence: evidence/v1-workbenches/metis-qti/coverage-manifest.json. It is generated by scripts/export_qti_coverage.py from the registry in services/qti3_profile.py, which export consults. - 63 cells. Every supported cell names the function that implements it (tested to exist), and every refused cell names its code. - The contract column is tested against M1.11's generated evidence. 19 cells fall short of it, and all are listed: - Metis has no canonical hotspot item. The router sends hotspot to the selected-response family, which has no hotspot kind. - Written scoring is a loss, not the contract's extension. - Per-option feedback on ordering and matching is lossy. - No draft carries a hint. - Stimulus, media, captions, transcripts and PNP are unmapped. The runtime is schemas/qti3.py and services/qti3_{content,render,map,parse,profile}.py, with 33 tests in tests/test_qti3_runtime.py. It maps single-choice, multiple-choice, ordering, matching, constructed-response and essay drafts into QTI 3.0.1, one item per locale, and reads them back. It refuses by code what it does not model. Measured outside the repo; the receipts are .d's: - 12 files in four packages pass the official QTI 3.0.1 XSDs. - The Longsight consumer validates the packages, and its scores agree with Metis's scorer on all 42 valid responses. No product route calls the runtime yet; that is .c's work.
    • M10.5.b Build minimal/full/edge/multilingual/RTL/MathML/caption/ transcript/PNP/interactive goldens with exact expected mappings/loss. Evidence: golden corpus. Evidence: apps/metis/service/tests/fixtures/qti3/goldens/, seven packages written by scripts/export_qti_goldens.py from the canonical inputs in tests/qti3_golden_fixtures.py, and compared byte for byte by tests/test_qti3_goldens.py (30 tests). Each golden's descriptor and its exact set of loss codes are committed with it, so a change in what leaves Metis fails a test. - minimal, full (partial credit with a penalty, feedback, cards), edge (escaped characters, a share that does not divide), multilingual (en/fr/ja), RTL (ar, dir=rtl), MathML (islands carried through the lesson canonicaliser). - caption, transcript, PNP and interactive are one package, as they are in practice: a stimulus set whose video carries WebVTT captions, an HTML transcript and a long description in both locales, each declared as a content-package variant with an Access for All resource description, with manifest dependencies and items that reference the stimulus in their own language. - The media is real: a 32x32 one-second H.264/AAC clip, a PNG, WebVTT and HTML, all on the profile's MIME allowlist. Media is content-addressed, so the goldens hold no duplicate bytes. Measured outside the repo (receipts are .d's): all 19 XML files pass the official QTI 3.0.1 XSDs; all seven packages pass the Longsight consumer's strict package validation; and its scores equal Metis's own scorer on all 36 valid responses to the goldens' items. New runtime: services/qti3_stimulus.py (an asset resolver, digest and MIME checks, adaptations per media kind) plus stimulus rendering, parsing and manifest dependencies/variants. The coverage manifest's disagreements with M1.11 fall from 19 to 14: media, captions, transcripts, PNP and dependencies now match it. Hotspot stays unmapped, because Metis has no canonical hotspot item.
    • M10.5.c Import safely, validate canonical semantics/protected keys, then export deterministically with version metadata/manifests. Evidence: round-trip tests. Evidence: apps/metis/service/tests/test_qti3_packages.py (40 tests) over services/qti3_package.py, plus the parse and render changes it exercises. Every golden package is exported, read back, and the items, stimuli, media and canonical refs that come out are compared with the ones that went in; then the same package is broken one way at a time and each break is refused by its own code. - Import safely: the archive is bounded before anything is parsed (entries, entry size, total expansion, compression ratio, nested archive by MAGIC not extension, encrypted entry, symlink, duplicate normalized path) and every path is a plain relative path inside the package. The manifest is reconciled with the archive BOTH ways: a file a resource declares must exist, a file no resource declares is refused, and every dependency, variant, stimulus reference and media reference must name something present. - Canonical semantics: a key that names a choice the item does not have, a single-response item with two correct values, an order key that is not a permutation of the choices, a match key that is not a choice/target pair, a mapping that scores something nobody can respond with, or a mapping whose bounds can exceed the item's maximum are each refused by name rather than imported as an item no learner can answer. - Protected keys: contains_protected_answer_material was the literal True for every item file. It is now computed, because an externally scored item renders with neither a key nor response processing — the descriptor was declaring protected material in documents that held none. A test proves the declaration matches what the document contains, and that the key appears in no other file. - Deterministic export with version metadata: every resource now carries qmd:qtiMetadata (interaction type, feedback type, scoring mode, tool name/version/vendor), so a consumer can tell what wrote a package. Two builds of the same inputs are byte-identical, and the M1.11 archive policy is now read back OFF THE BYTES — lexicographic order, 1980 timestamps, mode 0644, no extras or comment, and a compressed size equal to level-9 raw deflate (9 of 14 entries distinguish level 9 from level 1, so the check can fail). Re-measured after the manifest change: all 19 committed golden XML files validate against the official QTI 3.0.1 XSDs with 0 errors, and all 7 packages still pass the Longsight consumer — 11 items checked, 0 failed, 0 package errors. Those are .d's receipts to retain. CORRECTION (.d, same day): this line first read "exit 0", which was not a measurement. Run through node_modules/.bin/qti3, the consumer's entry guard declines to run and it exits 0 in silence; the numbers above are from re-running its resolved dist/index.js, which prints a JSON verdict. .d's runner refuses to write a receipt for a report with no verdict in it. The coverage manifest grows to 65 cells (qti-metadata, bounded-package-import). STILL NO PRODUCT ROUTE CALLS THE RUNTIME, and .a said this was .c's work. Measured why a cutover today would be a regression, not a migration: services/item_import_helpers.py resolves item-level media out of an imported package (ItemImportRaw.media, alt text required) and the runtime's Qti3Item carries no media — media lives on stimuli. Its archive reader is also not weaker than the runtime's; it checks magic bytes per entry, an extension allowlist, depth, ratio and encryption already. The runtime parser is deliberately narrow — it refuses response processing it cannot regenerate — so pointing the import route at it would refuse foreign QTI the product accepts today. The product path that assembles and reimports a package belongs to M10.10.d, and the parent M10.5 stays open behind it and .d/.e. 2026-09-18: .d and .e are closed, so THE PARENT'S ONE REMAINING BLOCKER IS THIS PARAGRAPH — no module under src/metis/api/ mentions qti3_profile, qti3_package or export_package (re-measured by grep on 2026-09-18; api/qti.py calls the older qti_service, not this runtime). A profile nothing serves is a profile, not an export.
    • M10.5.d Validate through pinned independent QTI schema/conformance tool and at least one supported consumer where required. Evidence: reports. CLOSED 2026-09-18 AT 14 OF 14, WITH M1.11.g. The two receipts that could not be taken were hotspot's, and the reason was not the validators: Metis had no canonical spatial item to hand them. It has one now. - The canonical item: schemas/selected_response.py gains the fifth mode M1.5.b ratified — regions with shapes and coordinates, the image they are measured on held by digest with its pixel size, and a region key. A coordinate count that does not fit its shape, a region that runs off the picture, a key naming a region the item does not have, a label or an alt text missing a locale, and a maximum above the region count are each refused at authoring time. Scoring, preview and the safe projection carry regions the way they carried options; the projection shows every label and the image's text equivalent and never the key. - The three option-shaped scanners (answer consistency, leakage, ambiguity) now REFUSE a spatial item by code. Each is keyed by OptionId, so a source built from a region-based item would have held no responses at all and every rule would have passed over nothing. - The export: qti3_map._hotspot_item writes a qti-hotspot-interaction with the image, each region's shape, coords and hotspot-label, and the region key; qti3_parse reads it back; qti3_profile moves the cell from unsupported to normalized with its handler, and the coverage manifest's contract disagreements lose hotspot. - A LIVE INTEROP DEFECT, FOUND BY THE CONSUMER AND FIXED: the image was first written as img, which the pinned XSD admits inside a hotspot interaction. The Longsight consumer rejects it — interaction.child.unsupported, "qti-hotspot-interaction does not allow img as a direct child" — so the package was schema-valid and unreadable by a real consumer, which is the silent downgrade this gate exists to catch. It is object now, with the text equivalent as its content, and both tools pass. Evidence: evidence/v1-workbenches/metis-conformance/, 14 receipts, all passed, all synthetic: false, regenerated whole by scripts/export_qti_conformance.py over the two pinned tools. The six pre-existing kinds' inputHash values are UNCHANGED, so nothing about the other exports moved. tests/test_qti3_conformance.py now asserts held == needed rather than held < needed. tests/test_selected_response_hotspot.py (27 tests) covers the item itself, and tests/test_qti3_* (137) the package and the gate. The fixture image is real and committed (tests/fixtures/qti3/media/animal-cell.png, 240x160): every coordinate is measured on it, so the picture and the regions cannot drift apart.
    • M10.5.e Fail silent extension/drop/default/downgrade and any supported cell lacking current proof. Evidence: conformance gate. Evidence: services/qti3_gate.py, run by scripts/export_qti_gate.py into evidence/v1-workbenches/metis-qti/conformance-gate.json, with 15 tests in tests/test_qti3_gate.py. The gate is computed over real exported bytes — seven per-kind packages and all seven goldens, 40 documents — never over a description of them. IT PASSES SINCE 2026-09-18: it failed on exactly the two receipts that could not be taken (missing-passing-receipt:assessment-kind:hotspot:*) until M10.5.d gave Metis a canonical spatial item. test_qti3_gate.py now asserts the pass, with 7 supported kinds and 14 receipts, rather than the two blockers, so the verdict is still checked and not merely produced. The five checks, and what each one is shown catching in a test: - silent extension: a namespace that neither the pinned schema closure defines nor the profile declares. The allowlist is the 11 targetNamespaces of the 21 pinned schemas — verified against the files themselves by fetch_xsd.py, not maintained by hand — plus the profile's own. Caught: an invented xmlns:x. - silent drop: a loss code the registry does not declare, or a document that does not come back byte-identical through the parser. Caught: an invented code, a re-indented document, and a document the parser refuses (reported as a blocker, never raised — the gate stays total over a hostile package). - silent default: a score maximum, mapping default or mapping bound left implicit. Caught: normal-maximum and upper-bound removed. - silent downgrade: the interaction in the document, the qmd:interactionType in the manifest and the contract's mapping for that kind must be one thing. Caught: an ordering item rewritten as a choice list, and a manifest that declares a different interaction than its document carries — which is the shape of the live defect M10.5.a found on the product's import path. - unproven cell: a supported kind with no current passing receipt from each pinned validator. Caught: no receipts at all (14 blockers), and a receipt whose inputHash is bytes this repository no longer builds (stale proof proves nothing). NEW, because the gate needed it: the runtime's loss vocabulary is now declared once, in qti3_profile.LOSS_CODES (26 codes, published in the coverage manifest). Before this, a refused cell named its code but a LOST thing did not: the 15 codes the corpus emitted were nowhere in the registry, so a drop could be recorded under any name at all and nothing would notice. A test scans the runtime's own source for every Qti3Loss code and requires the declared set and the emitted set to be equal, so the list cannot drift in either direction. The committed decision is re-computed by a test and must equal what the gate says today, so a package that starts extending, dropping, defaulting or downgrading in silence fails the suite rather than the next reader.
  • M10.6 Run LTI 1.3/LTI Advantage security and flow tests. Evidence: OIDC, JWKS rotation, nonce, state, roles, deep link, and grade service. <!-- SPEC 2026-09-11: same external gate as M1.12.g (§2.5.6). .a–.e are earnable against our own platform simulator only as security/flow unit coverage (tests/test_lti_service.py, tests/test_lti_api_contract.py, libs/contracts/src/metis/lti-advantage.ts); the parent's evidence line additionally needs one launch per flow from an independent LTI 1.3 platform — retain the receipt or record the failed attempt. The open LTI score route named in M0.15.d must be proven closed before the grade-service suite (.e) counts. --> <!-- the independent-platform half is DONE 2026-09-12; the parent stays open because §2.5.6 binds it to M1.12.g, which also wants the 1EdTech certification suite, and that is member-only. Moodle 4.5.13+ was installed headlessly on this server (apps/metis/service/tools/lti/README.md reproduces it), registered Metis as an LTI 1.3 tool through Moodle's own API, and drove every flow the evidence line names. Receipts, each the platform's own behaviour, are in evidence/v1-workbenches/metis-conformance/moodle-lti-1p3/: - OIDC + nonce + state: Moodle posts its login to /api/lti/login/initiate, we answer 303 with a state and nonce we minted, Moodle authorizes and posts a real RS256 id_token, and the launch verifies (receipt-resource-link.json). - JWKS: we fetch Moodle's keys; Moodle fetches ours to verify what we sign. ROTATION was exercised live — Moodle minted a new key pair and published only the new kid, the running tool held the old one cached, and the next launch verified on the refresh (receipt-jwks-rotation.json). - Roles: Moodle's Learner maps to learner (launch, view own result); its Instructor to educator with select-content and post-scores. - Deep link: Moodle's LtiDeepLinkingRequest is accepted, our signed response is posted back, and MOODLE PARSED IT into a real activity configuration (receipt-deep-linking.json). - Grade service: the tool exchanged a private_key_jwt for an access token at Moodle's token endpoint, created a line item, posted a score and read the roster through NRPS. The grade is in Moodle's own grade_grades table — learner1, 7.00000 of 10.00000, read straight out of its database (receipt-grade-service.json). Running it found three things our own tests could not, all fixed: a platform never sends our platform_id, a platform drives a browser rather than a JSON API, and a platform on loopback publishes over plain http. The tool-side AGS/NRPS client did not exist and now does (services/lti_services_client.py, 14 tests). WHAT REMAINS: the 1EdTech certification suite half. It is hosted and member-only, so §2.5.6's "record the attempt and leave the cell open" applies; no part of it was run and none is claimed. -->
    • M10.6.a Test registration/issuer/client/deployment/JWKS discovery/cache/ rotation/revocation and algorithm/key/confusion failures. Evidence: trust tests. Evidence: apps/metis/service/tests/test_lti_trust.py (33 tests) over the new services/lti_trust.py. Every test signs a real RS256 token with a real 2048-bit key, because the questions are about keys. THE CELL COULD NOT BE EARNED BY WRITING TESTS: there was nothing to test. lti_service.py verified a platform's id_token with HS256 against a platform_secret taken at registration, and never fetched the jwks_url it stored. A LIVE DEFECT, not a gap: any holder of that secret — both sides, and anyone either leaked it to — could mint a launch for any learner in any role, and LTI 1.3's whole trust model is that no such secret exists. Fixed here; platform_secret is gone from the API. What the trust layer now refuses, each with the contract's own error code (libs/contracts/src/metis/lti-advantage.ts) and a test: - algorithm-denied: anything but RS256, refused on the header before a key is fetched — including alg: none, and HS256 signed with the platform's own public key, which is the classic confusion. - key-id-missing: no kid. Without one, verification becomes "does any published key match", which keeps honouring a withdrawn key. - registration-inactive, issuer-mismatch, audience-mismatch (exactly one registered client in aud), authorized-party-mismatch (azp rules for multiple audiences), deployment-mismatch, issued-in-future, launch-expired, launch-lifetime-exceeded (600 s), with a minute of clock skew. JWKS discovery, cache, rotation and revocation, all measured by counting fetches at the boundary: - discovery fetches the registration's URL and no other; a key published at another platform's URL does not verify here. - three launches, one fetch: a launch is not a network call. - after the TTL the next launch fetches again. - ROTATION: an unfamiliar kid triggers exactly one refresh and then verifies, and the old key keeps working while it is still published. - REVOCATION: a refresh REPLACES the key set. Merging is how a withdrawn key outlives its revocation, so a withdrawn key stops verifying at the next refresh, and invalidate() drops it at once. - a stream of invented kids costs one refresh, not one per token: otherwise a tool is an amplifier pointed at its platform's JWKS. The tool's own service calls now use private_key_jwt against the tool's published keys (tool_jwks_url), so neither side stores a secret that could mint the other's traffic. HttpsJwksFetcher is the only thing that reaches a network: https only, no redirects, bounded body, and the protocol is the seam the tests replace. Registration gained platform_jwks_url, tool_jwks_url and status (re-registering as suspended stops every launch), and re-registering with a new JWKS URL drops what was cached for the old one. OpenAPI, the generated client and the schema inventory are regenerated. SIDE FINDING, left for M1.1: evidence/v1-workbenches/ metis-contract-ownership.json is stale. Regenerating it today gives 198 multi-variant OpenAPI groups and 58 competing definitions against the committed 40 and 0, and scripts/v1-workbenches/generate-metis-contract-ownership.test.mjs already fails 7 of 11 at HEAD — measured by running it against HEAD's own OpenAPI, so it is not this change. It was left as it is rather than half-regenerated here.
    • M10.6.b Test OIDC login/launch issuer/audience/message/version/target/ nonce/state/time/replay/session and malicious claim cases. Evidence: launch suite. Evidence: apps/metis/service/tests/test_lti_launch.py (27 tests) over the new services/lti_login.py and the rewired POST /api/lti/launch. As in .a, the suite could not be written against what was there: THERE WAS NO LOGIN. /lti/launch took state and nonce as request fields and checked the token's nonce against the one the caller had just supplied — a check that passes for anyone who sends both halves. There was nothing a replay could fail. A launch is now the second half of a conversation this tool starts. POST /api/lti/login is the OIDC third-party login: it validates the issuer, client and target against the registration, mints a 256-bit state and nonce from the system CSPRNG, remembers what it asked for, and returns the authorization URL (scope, response_type, response_mode, prompt, redirect_uri, login_hint, state, nonce). /launch then consumes the state and judges the token against that login. nonce is gone from the launch request: a nonce the caller supplies is a value the caller chose to match. What each case now costs, with the contract's error codes: - state-replayed for a state nobody asked for, for the same launch posted twice, and — deliberately — after a REFUSED launch, because a state that survives a wrong answer is a retry oracle. - state-mismatch for an expired login (10 minutes) and for a state belonging to another platform. - nonce-mismatch for a token carrying any other nonce, and nonce-replayed for a nonce already answered, which is the case a fresh login would otherwise launder. - target-link-mismatch when the token's target is not the one the login named or is not a registered redirect (registration now carries tool_redirect_urls, a set, as the contract's toolRedirectUris does, and exact membership is required). - login-deployment-mismatch when the deployment drifts between the two halves; anonymous-resource-launch-denied for a resource-link launch with no subject, while deep linking rightly needs none. - lti.version-unsupported and lti.message-type-unsupported are Metis codes, said to be Metis codes: the contract's vocabulary has no code for either, and filing them under one that means something else would be worse than naming them. Malicious-claim cases covered: a launch signed by a key the platform does not publish, a target pointing at another origin, an invented state, a stolen nonce, a version downgrade, an unsupported message type, and an anonymous attempt. KNOWN AND RECORDED, not hidden: the state and nonce stores are in process, so a restart forgets which were spent. It fails closed — an unknown state is refused — but a restart mid-attack reopens the replay window for tokens still inside their lifetime. The stores are a seam (LoginStore, SpentNonces) so a durable one can replace them; that is the same class of defect already recorded against credential_service in M9.18.
    • M10.6.c Test role/context/resource/user/privacy mappings for learner/ instructor/admin and unknown/custom/cross-tenant claims. Evidence: role tests. Evidence: apps/metis/service/tests/test_lti_roles.py (20 tests) over the new services/lti_roles.py. Before this, launch_tool copied the roles claim into its response verbatim: eleven role URIs, seven canonical roles and ten operations were pinned in the contract and none of them existed in the service, so every launch had the same authority as every other — none, or whatever a caller chose to infer. The mapping is not written here twice: the first test reads evidence/v1-workbenches/metis-lti-advantage.json's own role matrix and requires all 11 rows to match this module's — scope, canonical role, operation list, context binding, authoritative identity and displayClaimsTrusted: false — so the code cannot drift from the contract without failing. The three rules, each with the case that motivates it: - ONLY RECOGNISED ROLES COUNT. A launch carrying only roles this profile does not map is refused with unknown-role-denied; a custom vendor role beside a known one is carried for display and grants nothing. An unrecognised default would make "unknown" the widest privilege instead of the narrowest. - A CONTEXT ROLE WITHOUT A CONTEXT GRANTS NOTHING. An instructor role sent with no context claim is an instructor of nothing: the canonical role is still reported, and the operation set is empty. Operations are the union of what each recognised role grants, so no single role wins by sounding senior. - IDENTITY IS SIGNED, DISPLAY IS NOT. The principal is issuer#subject; name and email became display_name and display_email, are never matched on, and a launch that carries neither still identifies its user. The roster stores them as display beside the signed subject it actually matches on. The matrix's own headline claims are asserted: a learner can neither author nor grade, a mentor reads neither roster nor grades. Cross-tenant: the same context id arriving from two platforms is two rosters, and an access token minted for one platform reads only that platform's roster — tested with two registered platforms, each with its own keys.
    • M10.6.d Test deep-link request/response/signature/content/resource/ placement/return/stale/deleted/forbidden cases. Evidence: deep-link suite. Evidence: apps/metis/service/tests/test_lti_deep_link.py (14 tests) over the new services/lti_deep_link.py, plus the API's new GET /api/lti/jwks. TWO DEFECTS FIXED, both of which made the response unusable or unsafe: - THE RESPONSE WAS SIGNED WITH METIS'S OWN SESSION SECRET (HS256, and iss: "metis"). No platform holds that secret, so no platform could ever verify a deep-linking response this service produced. It is now signed with the tool's own RSA key, names its kid, and the tests verify it exactly as a platform would — through the public half served at /api/lti/jwks. A deployment with no key configured signs nothing and says so (lti.tool-key-not-configured, 503), instead of signing with the session secret; the pattern and the wording follow the integrity-deep-link key already in config.py. - THE PLATFORM'S OPAQUE data CAME FROM THE CALLER. It is how the platform recognises its own request, and the request body had a data field anyone could fill. The field is gone; the value travels in the launch's deep_linking_settings and is echoed from there. LTI_DEEP_LINKING_SETTINGS_CLAIM was defined in the service and read by nothing, so everything the platform said it would accept was ignored. The settings are now captured at launch and honoured: lti.deep-linking-type-denied for an item type it does not accept, lti.deep-linking-target-denied for a presentation target it does not accept, lti.deep-linking-multiple-denied for a second item where one was asked for (and several where it allows them), lti.deep-linking-return-url-invalid for a return URL that is not https, and lti.deep-linking-settings-missing for a deep-linking launch that carries none. The response carries the platform's own return_url, which is where the browser must post it. Forbidden and stale: selecting content now requires the launch to hold the select-content operation from .c's mapping, so a learner's launch is refused (lti.deep-linking-forbidden); an expired launch token is refused, as is one signed by anything but this service. While fixing the stale case: _decode_metis_token checked expiry against the library's idea of now rather than the clock the call was given, so a caller passing a time got an expiry judged against a different one. It now uses the call's clock, which is also what makes the AGS and NRPS token checks testable in time.
    • M10.6.e Test NRPS and AGS line items/scores/results/pagination/scopes/ idempotency/corrections/late events/visibility. Evidence: service suite. Evidence: apps/metis/service/tests/test_lti_services.py (17 tests). Four of the named behaviours did not exist and were built: - RESULTS. There was no way to read what a line item holds; scores went in and nothing came out. GET /api/lti/assignments/line-items/ {id}/results reads them, gated on the AGS result scope, which is also new (.../lti-ags/scope/result.readonly). - IDEMPOTENCY. A score carries a timestamp — when it was true, not when it arrived — and re-posting the same timestamp returns the same score_id marked duplicate, without applying it again. A retry after a timeout used to create a second score; a retried correction would have read as two corrections. - CORRECTIONS AND LATE EVENTS. A newer timestamp replaces the result and increments its revision, so a reader can see it was corrected. An older one is refused: arrival order is not event order, which is the lesson M8.11's grade ledger already paid for. - PAGINATION. A roster and a result set are both unbounded — a lecture course is thousands of people — and both were returned whole. Both are now paged in subject order, with a next_page token that is ours: a token this service did not issue is refused rather than parsed. Visibility, per scope, each tested both ways: posting a score takes the score scope and no other; reading results takes the result scope and no other (a score-scoped token cannot read back what it wrote); the roster takes the NRPS scope; a token minted for one platform cannot touch another's line item; an expired access token buys nothing. M0.15.d's PRECONDITION FOR THIS CELL, measured over the served app rather than assumed: POST /lti/assignments/scores is unreachable without a bearer token (422), refuses an empty, non-bearer or unparseable one, refuses a token that bought a different scope (403 missing required scope), and accepts only a score-scoped one; the grades it writes are readable only with the result scope. That is the vector M0.15.d left unmeasured; M0.15.d itself stays open on its other three (the learner-readable answer keys, /create/*, and the submission shapes). While writing it: the AGS scope checks ignored the clock they were given, so an expired access token still bought a score write in any caller that passes a time. Fixed with _decode_metis_token's clock from .d.
  • M10.7 Run OneRoster, SCORM, xAPI, Caliper, Open Badges, and gradebook interoperability fixtures. Evidence: compatibility report. CLOSED 2026-09-18, with all six children closed AND the parent's own sentence measured rather than inherited. The compatibility report is computed from the code today: 49 cells across all six profiles — 38 over HTTP, 8 library, 3 unsupported — with every implemented cell naming a handler that exists, every unsupported one naming the symbol whose absence makes it true, 0 silent losses, and every declared loss demonstrated by calling the code and watching the field not arrive. WHAT IS STILL UNSUPPORTED, AND SAYS SO: three OneRoster record types — academic session, term and category. They are not among the seven entities M10.7.b names, they are reported as gaps rather than omitted, and the ratchet fails the day one of their symbols exists. WHAT THIS PARENT DOES NOT CLAIM: no certification by any of the six bodies. What is held is fixtures, a generated matrix nothing can edit into agreeing with itself, and — for QTI, which is M10.5's — two pinned external tools' own output. <!-- SPEC 2026-09-11: service modules exist for all six (schemas/{oneroster,scorm,xapi,caliper}.py, api/{oneroster,caliper}.py, services/open_badges_service.py, services/standards_exchange.py, services/lms_interop_matrix.py) with tests/test_{oneroster,scorm,xapi,caliper}_{service,api_contract}.py; Open Badges has NO test file today, so .d starts by writing one. The standards exchange has no HTTP surface in this build (M8.15.d): .a's matrix says per cell whether the direction is served over HTTP, library-only, or unsupported, and a library-only cell is "compatible" for the library alone. -->
    • M10.7.a Enumerate supported version/profile/operation/direction/entity cells and source-of-truth ownership per standard. Evidence: matrix. Evidence: evidence/v1-workbenches/metis-interop/coverage-matrix.json, generated by scripts/export_interop_matrix.py from the registry in services/interop_coverage.py, with 16 tests in tests/test_interop_coverage.py. 47 cells over the six profiles M1.13 pins when this was written: 21 served over HTTP, 4 library-only, 22 unsupported. The registry is generated, so the counts move as later items build things — .c and .d record the moves, and the committed matrix is the current number. Every cell names its profile, entity, operation, direction and the owner of the record, and a cell cannot describe itself incoherently — an unsupported cell that names a handler, an HTTP cell with no route, or an implemented cell owned by nobody all fail to construct, and the test proves each refusal. Every claim is checked against something outside the registry: each handler against the import system (a method on a class included), each route against the OpenAPI document the service really publishes, and the profile set against the contract's own decision matrix. FOURTEEN DISAGREEMENTS WITH M1.13, computed from its evidence rather than restated, and the last one is a defect: - Five profiles the contract records as runtimeAdopted: false are served over HTTP today (OneRoster, Caliper, both SCORM editions, xAPI). Only Open Badges is genuinely library-only. - Two declared directions nothing implements: OneRoster metis-to-external (M1.13 makes Metis the gradebook provider and nothing publishes a line item or result) and xAPI bidirectional-explicit-owner (no state or profile documents). - Six of OneRoster's ten record types are unmodelled: organization, academic-session, term, category, line-item, result. OneRosterEntityType carries user, course, class and enrollment. - IDENTITY: M1.13 pins tenant+connector+sourcedId for every entity and emailAsJoinKey: false. The runtime matches users by normalized email and classes by course slug, AND NO METIS MODEL PERSISTS A sourcedId at all. A re-sync cannot recognise what it created: change an email in the SIS and the next sync makes a second user. That last row is why .b is not a test-writing task — a round trip needs a stable identity to round-trip through. Also recorded as cells, because an absent row reads as "no such cell" rather than "not built": no OneRoster delta watermark or tombstone, no SCORM sequencing, interactions or objectives, no xAPI voiding, no Open Badges revocation or endorsements, and a badge's proof is detected but deliberately never treated as verified.
    • M10.7.b Run OneRoster org/course/class/user/enrollment/line-item/result snapshot/delta/conflict/delete round-trips. Evidence: OneRoster suite. CLOSED 2026-09-18 ON ALL SEVEN. It was open on three. organization, line-item and result are not modelled at all (M10.7.a's matrix), so there is nothing to round-trip them through; line-item and result are the gradebook direction M1.13 makes Metis the provider for, and nothing publishes one. Closing this needs those built, not tested. AND THE GRADEBOOK HALF IS BLOCKED ON SOMETHING SHARPER THAN "UNBUILT", found 2026-09-13 by starting to build it. line-item and result are gradebook-publish / metis-to-external: Metis SERVES them and a SIS reads them. A published result names a STUDENT — OneRoster wants the student's sourcedId — and the grade ledger records whose grade a grade is ONLY as learner_digest, an irreversible sha256. Nothing in the estate maps a digest back to anybody: every use compares it to one the caller already computed ("a learner may only see signals whose learner_digest is their own"), which is right for privacy and is exactly why it cannot be published. THE JOIN THAT WOULD RESCUE IT DOES NOT EXIST EITHER. A digest can be RECOMPUTED for a student you can already name, so a roster cohort could in principle be matched. But the only grades the ledger holds are the LTI AGS ones M10.7.e's road now carries, and their digest is over the PLATFORM's user id, while roster_external_reference maps a sourcedId to a METIS user. Nothing connects a platform user id to either. So the chain runs sourcedId → Metis user → nothing. THIS IS A TENSION AND NOT AN OVERSIGHT, worth stating as one: M10.7.d found this estate serialising a learner's NAME to a Caliper consumer, and the answer was to identify people by digest. The same decision is what stops Metis being a gradebook provider. Closing the cell means deciding which of the two wins for this direction — a published gradebook needs a nameable student — and that is a design decision with a privacy consequence, not a missing model. THE DECISION WAS TAKEN 2026-09-18 (user): a scoped, auditable join. line-item AND result ARE SERVED NOW. Two of the three unmodelled entities are built; organization is what keeps this cell open. - The join: roster_grade_identity (migration 079) maps a learner digest to a SIS sourcedId, per tenant and per connector, unique in BOTH directions — two sourcedIds sharing a digest would put one learner's grades under two names, and two digests sharing a sourcedId would put two learners' grades under one. - THE SCOPE IS THE DESIGN, and a test holds it. record_identity_async has exactly one caller (oneroster_service), resolve_for_publication exactly one (oneroster_gradebook), and that one exactly one route. There is deliberately NO resolve_one(digest) helper: a general-purpose reverser is what a later caller reaches for, and it would end the digest's usefulness for every other surface at once. Every read takes an actor, a purpose from a fixed vocabulary and a detail, none of them defaulted, and writes its audit row in the SAME transaction as the read — so a publication that happened and an audit that did not cannot both be true. - A SECOND GAP FOUND BY BUILDING IT: the ledger could not say what a grade is ABOUT. grade_id is a digest over platform, line item and user, and a LedgerEvent carries neither the line item nor the learner, so nothing in the database could enumerate one line item's results. grade_subject (migration 079) is that record, written by the ingress when it appends and never rewritten; it is pseudonymous — the same digest project() already takes — so it adds a lookup and never an identity. - AN UNRESOLVED LEARNER IS REPORTED. A grade whose learner the roster has never named comes back in unpublishable_line_items, because a SIS receiving a short gradebook with no way to tell would record those learners as having done no work. - GET /api/oneroster/gradebook publishes both, tenant DERIVED from the caller. A result is derived from the ledger at read time and never cached here: a correction appended a second ago is in the next read, and this surface does not become a second record of a grade. Evidence: tests/test_oneroster_gradebook_postgres_integration.py, 13 tests on a live PostgreSQL — publication, the unresolved learner, the orphan line item, a correction landing in the next read, the audit row per read, a refused purpose, a digest already published as somebody else, an identity change as an update, two tenants, two connectors, write-once subjects, and the two scope ratchets over the source. The matrix moves with the code: oneroster-1.2:metis-to-external is no longer an unserved direction, and the coverage test that asserted it was now asserts that it is served. AND organization IS BUILT TOO (migration 080, roster_organization). Metis has no organisation of its own — a tenant is not a school — so an org is kept as what it is: the source's record under the source's id, per tenant and per connector. That is enough for the thing the entity was needed for. A LIVE DEFECT IT CLOSES: OneRosterUserRecord.org_sourced_ids existed and NOTHING READ IT, so a user could claim membership of an organization nobody had ever declared and the sync took it. The claim is checked now — against this payload's orgs and against the ones this connector sent before — and an unknown one is a WARNING rather than a block, because the org may be in the next page and losing a real person over a record that is merely late would be worse. An org takes the same three round trips every other entity does: snapshot (create/update), delta (date_last_modified against the caller's watermark), conflict (two orgs with one sourcedId) and delete (tobedeleted deactivates and KEEPS the row, so the same sourcedId arriving again is not adopted as a new organization). Evidence: 9 more tests in tests/test_oneroster_identity.py (21 in the file), and 257 across the coverage, compatibility, contract, route-matrix, gap-map, migration and grade-interop gates. WHAT "SNAPSHOT/DELTA/CONFLICT/DELETE" MEANS FOR THE TWO PUBLISHED ENTITIES, said rather than implied: line-item and result are the direction where a SIS READS Metis. There is no delta, conflict or delete round trip to run on them, because this estate is not the party that changes them — it answers with what the ledger currently says, and a correction appended a second ago is in the next read. THREE OF THE TEN CONTRACT RECORD TYPES REMAIN UNMODELLED — academic session, term and category — and each still names the symbol whose absence makes its row true. They are not among the seven this cell names. WHAT WAS BUILT, because .a measured that a round trip had no identity to travel on: a sync could not recognise what it had created. sourcedId was never persisted; users were matched by email and classes by course slug, so an address change in the SIS made a SECOND PERSON, and a renamed class made a second course. - roster_external_reference (migration 073, applied to a live PostgreSQL) stores one row per external identity — tenant + connector + entity + sourcedId — against the Metis record it resolved to, with the source's own date_last_modified and a status. Two connectors numbering independently are two identities. - Identity now resolves through the reference, and email or slug only ADOPTS a record the first time a connector sees it. An email change is an update to that person; the planner compares emails now, which it never had to when the email was the identity. - Two sourcedIds claiming one Metis record is a blocking identity_owned_by_another_source_record conflict, not a merge. - DELTA: records carry date_last_modified, the request takes modified_since, and the response returns the watermark to ask with next time. It is the source's clock, so a re-run here cannot move it, and a record the source dated nothing is never skipped — silence is not a statement that nothing happened. - DELETE: tobedeleted deactivates the record and TOMBSTONES the reference rather than dropping it, so the same sourcedId arriving again is not adopted as somebody new. Evidence: tests/test_oneroster_identity.py (16 tests) for the rules and tests/test_oneroster_postgres_integration.py against a live database for the round trip: sync, change the email, sync again — one user, one reference, the new address — then withdraw and find the tombstone. A LIVE DEFECT FOUND BY RUNNING THAT, unrelated to rostering and worse: AuthService.hash_password RAISED ON EVERY INPUT. passlib 1.7.4 reads bcrypt.__about__ to pick its backend and bcrypt removed it in 4.1 (this venv has 5.0), so the error came back as a misleading "password cannot be longer than 72 bytes" for a 57-byte string. Registration could not work at all, and the suite was green because the only test that reaches this service patches it out. Now called through bcrypt directly, with over-long passwords refused rather than silently truncated to something a shorter password would also open, and hashes written by passlib still verify. tests/test_auth_passwords.py (10 tests) hashes and verifies real passwords with nothing mocked.
    • M10.7.c Run SCORM package/runtime/sequencing/bookmark/score/status/ interaction/error and xAPI statement/state/profile/voiding fixtures. Evidence: SCORM/xAPI suites. WAS OPEN ON SCORM SEQUENCING AND THE xAPI DOCUMENT RESOURCES. Sequencing and navigation are not written or honoured — a package is one launchable activity — and there are no state or profile documents, so an activity cannot keep a learner's place here. Both are cells in M10.7.a's matrix with their reasons; closing this means building them. DONE, and both were defects rather than gaps: - xAPI VOIDING. A voiding statement (verb voided, object a StatementRef) was stored as an ordinary statement and the statement it withdrew KEPT READING AS CURRENT — the one thing voiding exists to prevent. Now the target is marked voided and carries the id of the statement that withdrew it; a query stops answering with it, and a replay still carries it, because a consumer catching up has to learn that it was voided. The verb alone voids nothing (it must be a StatementRef), a void names something this LRS holds, and a voiding statement cannot itself be voided — otherwise a retraction could be retracted and nothing would settle. tests/test_xapi_voiding.py, 11 tests. - SCORM INTERACTIONS. A commit carried runtime_data, a free-form dict, so a player's per-question detail arrived in whatever shape it liked and nothing checked it. An interaction is a typed record now: a response that fits its type (a true-false answer is true or false, a numeric answer is a number), a result from SCORM's vocabulary or a number, no question reported twice in one commit, and NO TYPE THE PACKAGE'S OWN EDITION LACKS — a 1.2 package reporting long-fill-in is reporting something no 1.2 player wrote. tests/test_scorm_interactions.py, 24 tests. Already covered by the existing suites and re-checked here: package and manifest, runtime commit, bookmark (location and suspend_data), score (raw/min/max scaled) and status (lesson and success), and a replayed commit staying one commit. The matrix moves with the code: scorm-*/interaction and xapi-2.0/statement/void are now served cells rather than reasons, so the coverage matrix is 24 HTTP cells and 19 unsupported. THE xAPI DOCUMENT RESOURCES ARE BUILT. Half of what kept this open is closed: state-document/put-get-delete and profile-document/put-get-delete are served cells now rather than witnesses named by their own absence, and an activity can keep a learner's place. Migration 075 (xapi_documents), services/xapi_document_service.py, thirteen operations under /api/xapi/activities/{state,profile} and /api/xapi/agents/profile, and tests/test_xapi_documents.py — 45 tests against a freshly migrated Postgres, service and HTTP both. The matrix is 30 HTTP, 7 library and 12 unsupported. DURABLE FROM THE FIRST LINE, deliberately. This estate keeps losing the same argument with itself — LTI line items, LTI scores until M10.7.e, credential revocations (M9.18) — and a document store a restart empties would tell a learner they had never begun, which is the one thing a state document exists to prevent. THE CONCURRENCY IS THE RESOURCE. A learner with two tabs open is two writers. PUT replaces and POST merges, which are different verbs and not a preference; a merge is defined over two JSON OBJECTS, so a POST onto text or an array is refused rather than quietly made a replace that loses everything the client did not send; and a write over an existing document that names no expected version is refused 409 WITH the ETag to re-read, while one that names a version that has since moved is 412. A client cannot retry correctly unless it is told which of those two happened. A delete obeys the same preconditions, because deleting what somebody has just written loses their work exactly as overwriting it would. THE IDENTITY IS A TYPE AND AN INDEX, NOT A CONVENTION. A state belongs to one learner on one activity, an activity profile to the activity and every learner shares it, an agent profile to the learner across activities — so an activity profile that names an agent WILL NOT CONSTRUCT, because that mistake is how one learner's place gets served as everybody's. A registration makes a different document, so a second attempt does not resume the first. The whole identity is folded into one storage_key with a unique index over it, so two writers racing lose in the database rather than in an if. A CROSS-TENANT HOLE THIS ITEM OPENED AND M8.14's RATCHET CAUGHT. The first draft's table had no tenant_id, and every other part of a document's name — the activity, the agent, the document id — arrives from the caller. Any authenticated caller could therefore have read any tenant's state by naming its activity and its learner. Nothing in this item's own tests would have found it; the tenancy ratchet failed the full suite with newly_untenanted=('xapi_documents',) and that is the only reason it was found. The tenant is part of the identity now, derived from who is asking and never from a parameter, and folded into storage_key so the unique index enforces it rather than a query remembering to. Three tests assert the boundary directly: a read, a scoped delete and a listing all stop at it. A LIMIT FOUND BY WRITING THE TEST, AND KEPT AS A LIMIT. XapiActor requires a non-empty account_name, so an agent identified by a MAILBOX alone — legal xAPI, and what a platform that issues no accounts sends — cannot be constructed here at all, and no document can be keyed by one. The first draft carried an mbox fallback; it was unreachable code. It is gone, the constraint is asserted by a test, and the key still carries its account: namespace so that adding an mbox branch later cannot silently merge an account x@y with a mailbox x@y. The cells are compatible rather than lossy because such an agent is REFUSED 400 rather than stored with its mailbox dropped — a refusal is not a silent mapping. AND THE OBJECTIVES ARE TYPED NOW TOO. cmi.objectives.n was the other free-form casualty beside the interactions: a package reported what a learner had ACHIEVED and it arrived inside runtime_data, a dict nothing checked, so an objective scored twelve out of ten and carrying a status its own edition does not define was stored exactly like a good one. ScormObjective — the symbol M10.7.a named as the witness — is real, and a score that does not hold together against its own range will not construct. AN OBJECTIVE IS NOT AN INTERACTION, and the two are kept apart on purpose: an interaction is one question, an objective is a thing the course says the learner can now do, and several interactions bear on one. That is why an interaction NAMES objectives by id rather than repeating them — if it carried the objective's own score, two questions could disagree about what a learner had achieved and nothing would say which was the record. AND NOTHING AN EDITION COULD NOT HAVE WRITTEN. SCORM 1.2 has ONE objective status word and no scaled score, so a 1.2 package reporting a completion apart from its success, or a scaled score, is refused — the same rule the interactions got. tests/test_scorm_objectives.py, 13 tests. The matrix is 32 HTTP, 7 library and 10 unsupported. SCORM SEQUENCING IS BUILT, 2026-09-18, and this cell closes with it. A package was one launchable activity; it is an activity tree now, and the manifest carries imsss:sequencing for a 2004 package. A DOCUMENTED SUBSET, WITH THE REST REFUSED BY NAME. schemas/scorm_sequencing.py lists 30 constructs honoured and 7 refused, each with the section of the book that defines it, and services/scorm_manifest.py RAISES on a manifest asking for randomisation, constrained choice, rollup considerations, delivery controls, auxiliary resources, an exit-condition rule or retryAll. Dropping one silently would deliver a different course from the one the author sequenced, and the learner would have no way to tell. The rulebook is services/scorm_sequencing.py, with the book's own process identifiers on its functions (OP.1, NB.2.1, SB.2.3, RB.1.5) and the book's own exception codes on its refusals, so it can be read against the specification rather than against its author. Honoured: the four control modes; flow forward and back including descent into a cluster; choice, with hiddenFromChoice, disabled and choiceExit each refused by their own code; skip and stopForwardTraversal; attempt limits; post-conditions (continue/previous/retry/exitParent/exitAll); measure, objective and completion rollup with all/any/none/atLeastCount/ atLeastPercent child sets and objective measure weights; and start, resume, suspend, exit and abandon. TWO RULES THAT ARE EASY TO GET WRONG AND ARE TESTED FOR: - An unknown is not a false. SN 4.5 evaluates a rule combining an unknown condition to unknown, and an unknown rule does not fire; a "not satisfied" retry that read silence as failure would send a learner round again on a judgement nothing had made. - A cluster with no rollup rule keeps what it held. SCORM gives no cluster a default, and inventing one would satisfy a section nobody said was satisfied. DURABLE FROM THE FIRST LINE (migration 078, services/scorm_sequencing_store.py). A sequencing session is a learner's place in a course; in process memory — where the rest of the SCORM service still is, which is M8.1's separately recorded gap — it would be lost on every deploy, sending the learner back to the first activity and handing back the attempts an author limited. Two tabs are two writers: every write names the version it read and the loser is told rather than silently overwriting. Over HTTP: POST /api/scorm/sequencing/{launch,navigate,tracking} and GET /api/scorm/packages/{id}/activity-tree, launch-token authenticated like the runtime commit and declared as such in DECLARED_POSTURES. A player may only report on the activity it is RUNNING — otherwise one SCO could set another's objectives, which is how a course marks itself complete. Evidence: tests/test_scorm_sequencing.py (29, the rulebook), tests/test_scorm_sequencing_manifest.py (16, write and read back, including a whole-tree round trip and every refusal) and tests/test_scorm_sequencing_postgres_integration.py (10, over HTTP against a live PostgreSQL, read back through a second engine). 173 more across the contract, route-matrix, gap-map and migration gates. THE MATRIX MOVES WITH THE CODE: sequencing/sequencing-and-navigation is a served cell for both profiles — compatible for 2004 and version-skew for 1.2, because SCORM 1.2 has no imsss:sequencing and writing 2004 elements into a 1.2 manifest would send a player rules its own edition does not define.
    • M10.7.d Run Caliper entity/event/profile/privacy/order and Open Badges achievement/assertion/evidence/signature/status/revocation fixtures. Evidence: Caliper/badge suites. CLOSED 2026-09-18. The two things that kept it open were built in the body below and the checkbox was never flipped; re-verified today rather than taken on the note's word: - verify_proof and read_credential_status exist in services/open_badges_proof.py and are exposed through open_badges_service — the two symbols M10.7.a's matrix named as the witnesses whose ABSENCE made those rows true. - The registry, read from the code today, holds NO unsupported cell and NO remaining witness for either profile: open-badges is 2 compatible and 6 lossy, caliper 2 compatible and 3 lossy. - 138 tests pass across test_open_badges_service.py, test_open_badges_proof.py, test_caliper_profiles.py, test_caliper_service.py and test_outbound_ssrf.py, and the adversarial grep over both badge modules returns nothing. The honest limits below are unchanged and are limits, not gaps: no third-party-issued badge has been verified here, and the Caliper acknowledgement position is as volatile as the event log it points into. WAS OPEN ON BADGE SIGNATURE AND STATUS/REVOCATION. Neither is a fixture that failed; both are things this estate does not do for somebody else's badge, and they are the two that decide whether an inbound badge means anything. Closing this means verifying a foreign issuer's proof and reading a credential status list. Both are cells in M10.7.a's matrix with their reasons. CALIPER DONE — and all three of entity/privacy/order were defects the contract already forbade: - THE GRADING PROFILE COULD NOT BE RECORDED. metis-interoperability-profiles.json pins GradeEvent/Graded → GradingProfile, and the request schema's literals listed neither, so the one profile the contract names for grades was rejected at the door. The contract row and the code now agree, and the first test reads the rows out of the contract so they cannot drift apart again. - EVENTS CAME BACK IN ARRIVAL ORDER. The contract pins eventTime-ascending-then-sequence; _filter_events returned insertion order, so an event that HAPPENED an hour earlier and arrived five minutes later sorted last — the exact case ordering exists for. Now sorted by when it happened, with arrival only breaking ties, in queries and in the exported envelope alike. - THE LEARNER'S NAME LEFT THE ESTATE. The contract pins directPiiSerialized: false and every payload carried actor.name — including the sensor envelope, which goes to somebody else. What leaves now identifies the actor without naming them; Metis's own record still has the name, and an OBJECT may still be named, because a lesson is not a person. tests/test_caliper_profiles.py, 11 tests, covering entity identity, per-event profile, dedup, privacy and order. OPEN BADGES: the module had NO tests, which for an importer is the worst place to have none — everything it does is decide whether somebody else's document is trustworthy enough to read. tests/test_open_badges_service.py, 27 tests: the round trip, the deterministic write (sorted keys, no whitespace drift, so a badge can be hashed), the required contexts and types, and fourteen documents that are badge-shaped and are not badges (a credential of another type, an achievement with no criteria, a subject that names nothing, a validFrom reading "one tuesday"), plus an alignment with no target, which is dropped rather than half-read. A REAL GAP CLOSED: EVIDENCE. An achievement says what was required; evidence says what was actually SHOWN, and it was dropped on read and had nowhere to go on write, so an imported badge was a claim with nothing behind it. It round-trips now, and an entry with no narrative is a link rather than a record and is not read as one. A NON-GAP CONFIRMED: proof_present is recorded as CLAIMED and never as verified. The suite proves the distinction with a forged document — an impostor issuer and the string obviously-not-a-signature — which parses as present and unverified, because treating a proof's presence as proof is how badge systems are fooled. The matrix moves with the code: caliper-1.2/event/ordered-read and open-badges-3.0/evidence/read-and-write are served cells now, and what remains unsupported says so more precisely — verify-proof because credential_service.verify_credential checks a proof against METIS'S OWN verification method, so a foreign issuer's badge comes back not_found rather than verified. 25 HTTP, 5 library, 19 unsupported. BOTH ARE BUILT NOW, AND THE ITEM STAYS OPEN ON ONE STEP OF TRANSPORT. services/open_badges_proof.py verifies a foreign issuer's eddsa-jcs-2022 proof and reads a BitstringStatusList entry, exposed as open_badges_service.verify_proof and .read_credential_status — the exact two symbols M10.7.a's matrix named as the witnesses whose ABSENCE made those rows true, so the rows flipped by being made real. tests/test_open_badges_proof.py, 29 tests. The matrix is 28 HTTP, 7 library and 14 unsupported (the sentence above was written when it was 25/5/19 and later items moved cells without updating it). WHY THAT SUITE AND THAT KEY FORM. eddsa-jcs-2022 over a did:key is the combination checkable with NO NETWORK — the public key is INSIDE the identifier — so the first badge verification this estate performs adds no fetch, no resolver and no registry to trust. A VERDICT WITH TWO VALUES WOULD BE A LIE. A badge signed with a suite this estate does not implement is UNCHECKED, and a verifier that answers only valid/invalid has to call it a forgery. no_proof, unsupported_cryptosuite, unresolvable_verification_method and uncanonicalizable all mean "nobody here looked"; signature_invalid is the only outcome that accuses the issuer, and NOT_CHECKED names the difference so a caller cannot read one as the other. AND A VERIFIED PROOF NAMES THE KEY, NOT THE ISSUER. It says the bytes were signed by that key. Whether that key belongs to the institution on the badge is a trust decision, and there is no issuer registry to make it against — the same absence M10.13.f, M10.14.e and M10.20.a each record. The verdict says so in its own detail string. A CANONICALIZATION THAT IS NEARLY RIGHT REPORTS A GOOD BADGE AS A FORGERY, so JCS is implemented properly rather than borrowed. MEASURED: this estate's existing canonical_json_bytes sorts by code point and RFC 8785 sorts by UTF-16 code unit — 😀 and ＀ come out in opposite orders — and Python writes 1.0, 1e-07 and 1e-06 where ECMAScript writes 1, 1e-7 and 0.000001. A number this canonicalizer cannot reproduce EXACTLY is refused as uncanonicalizable rather than hashed differently from the way its issuer hashed it. STATUS: a revoked badge no longer reads as valid. A supplied list is decoded (multibase, gunzip, bounded at 16 MiB against a compression bomb) and the bit read most-significant-first, which a test checks at all eight positions in a byte — read the other way round and every badge in the first byte gets somebody else's status. A suspension list is refused as an answer to a revocation question. AND THE LIST IS FETCHED NOW, THROUGH THAT GUARD. fetch_status_list is the THIRD fetch somebody outside aims, and the most attacker-chosen of the three: an LTI fetch needs a registration an administrator made or a launch a platform signed, and this one needs only somebody handing over a badge. It goes through refuse_unless_public and the same pin the LTI paths got in M10.8.e — dial the address that was approved, keep the hostname for Host and the TLS SNI — with a bounded read and a bounded decompression behind it. FETCHING IS OPT-IN, because it makes this service issue a request somebody else chose; without fetch=True the badge's URL is never dialled. And a list that cannot be fetched reads unreadable_list, never unset — the failure that would matter most is a badge whose status could not be read coming back looking like a badge that is fine, and a test asserts it does not. test_outbound_ssrf's "the two fetches somebody else aims" became "named one by one", which is the assertion that was always meant. AND A CALIPER CONSUMER CAN TELL A GAP FROM AN EMPTY PAGE. Export was a query by cursor with no acknowledgement, so a consumer that asked for the next page and received an empty one could not tell whether there was nothing new or whether something had gone missing: both look identical from outside. POST /api/caliper/deliveries records how far a consumer has processed and answers with outstanding — so an empty page beside a non-zero count is a problem it can SEE. TWO REFUSALS, BOTH ABOUT NOT HIDING A GAP. An acknowledgement that goes BACKWARDS is refused: a consumer that has processed through 40 and then claims 20 would make work already done report as work still to do, which reads exactly like a gap that is not there. And one PAST the highest sequence recorded is refused: nobody can have processed an event this estate never sent, and taking the claim would mark real events as delivered. A repeat of the SAME position is not a refusal, because a retry after a timeout is the ordinary case. A consumer that has never acknowledged reads zero rather than absent — nothing processed is a position, not a missing record — and two consumers keep their own, because one being behind is not the other's problem. RECORDED AS A LOSS: the position is held beside the event log and is as volatile as it is. An acknowledgement of events a restart forgets cannot outlive them, so making this durable means making the log durable first, and doing only half would let a consumer's position outlive the events it points into. AND ENDORSEMENTS ARE READ NOW. parse_endorsement — the last symbol M10.7.a named for this profile — reads an EndorsementCredential into who vouched, what they vouched for, and what they said, so who else stands behind an achievement is no longer lost on the way in. THE SUBJECT IS THE WHOLE POINT. An endorsement whose subject names NOTHING is refused rather than parsed, because an endorsement of nothing reads as vouching for whichever credential happened to be beside it — which is exactly how an endorsement gets borrowed. And a document claiming to be an OpenBadgeCredential AND an EndorsementCredential at once is refused, because that would let an issuer award a badge and endorse it in one document and have the endorsement counted as somebody else's word. An endorsement with no comment still endorses: weaker, and still somebody's word, and refusing it would drop a real one for saying little. Its proof is CLAIMED, like a badge's — but unlike when that rule was written there is a verifier now, and a caller that wants an endorsement checked passes it to verify_proof rather than reading proof_present as an answer. The matrix is 32 HTTP, 8 library and 9 unsupported. HONEST LIMIT OF THE EVIDENCE: the suite signs with the same canonicalizer the verifier uses, so it proves the chain agrees with ITSELF. No badge issued by a third-party implementation has been verified here — none exists offline and inventing one proves nothing. What IS independently checked is the serializer, against canonical strings written out by hand. The residual risk is a canonicalization difference on a shape no test covers, and its symptom would be signature_invalid on a genuine badge.
    • M10.7.e Run internal gradebook import/correction/visibility/LTI- OneRoster reconciliation with duplicate/out-of-order events. Evidence: grade interoperability suite. CLOSED 2026-09-18. It was open on the prerequisite below, and M10.7.b closed it: result is modelled and published, so reconciliation across the two roads is a JOURNEY a grade takes rather than a capability of the ledger. A score arrives on the LTI AGS road, a retry of it is one event, a rescore that was true earlier and arrived last never becomes the grade in force, and what the SIS reads is what the ledger DERIVES — asserted from the other side, against project(), rather than restated from the publisher. AND THE DEFECT THIS CELL RECORDED AND DID NOT FIX IS FIXED: THE GRADE WAS DURABLE AND THE LINE ITEM IT HANGS ON WAS NOT. Line items lived in a class-level dictionary and sync_score refuses one it cannot find, so after a restart a platform posting to a line item METIS ITSELF created was answered line item is not registered, 403, forever, with nothing it could do about it. lti_line_item (migration 081) is that table, written where create_line_item writes. It is NOT the binding table, deliberately: grade_ingress_binding says what a line item SCORES and somebody has to assert that; this says the line item EXISTS, which is a thing this tool did. Conflating them would let a registration nobody made be inferred from a binding somebody did. THE STORE IS CHOSEN BY NAME AND NEVER FALLEN BACK TO. LtiService.use_line_item_store(InProcessLineItems()) is how a suite with no database asks for an in-process one, the way use_jwks_cache already worked; a store that silently became in-process when the database was unreachable would put the defect straight back. reset() now forgets WHICH store is in use rather than emptying one — a reset that cleared the table would be one import away from clearing a real tenant's line items. Evidence: tests/test_grade_ingress_journey.py 28 tests (the new one creates a line item, clears exactly what a restart clears, and posts a score to it successfully), tests/test_oneroster_gradebook_postgres_integration.py 14, tests/test_grade_interoperability.py 16, and 115 across every suite that constructs an LtiService. WHAT WAS OPEN, AND WHY IT IS NOT NOW: the OneRoster road could not be built because result was not modelled at all (M10.7.b). It is. WHAT WAS OPEN, AND IS NOW CLOSED FOR LTI AGS. The durable ledger M8.11 built HAD NO WRITER: nothing under src/ imported services/grade_ledger.py except the model registry, and a score posted over LTI AGS landed in LtiService's class-level dicts, which reset() clears and a restart forgets. POST /lti/assignments/scores now carries the accepted score into the ledger, and tests/test_grade_ingress_journey.py proves it the only way worth proving: post a real score with a real AGS token, clear exactly what a restart clears, and read the grade back out of Postgres. 26 tests against a freshly migrated database. THE HARD PART WAS A FIELD NOTHING CARRIES. A LedgerEvent names the attempt and the item_revision_sha256 the value was arrived at against — that is the point of M7.10's chain — and an AGS post carries a number, a user and an opaque line-item id. There is no honest derivation from an identifier to a content digest, so the digest is NOT derived: grade_ingress_binding (migration 074) is where somebody says what a line item scores, and a score against an unbound line item is refused rather than recorded with an invented revision. A ledger holding a confident digest of nothing is worse than an empty one, because the empty one is visibly empty. AND THE TENANT COMES FROM THE SAME PLACE, for the same reason. An AGS post says nothing about whose records it belongs in — the platform registration does not record who registered it — so binding a line item is a tenant CLAIMING it, and a line item two tenants claimed is refused naming both rather than assigned to whichever sorted first. WHAT IS DERIVED, IS DERIVED AND NOT GUESSED. The grade and the attempt are the (platform, line item, user) triple, which is the platform's own notion of one person's work on one assignment; the learner and the actor are digests, not names, which is M10.7.d's Caliper finding applied before it could recur; and the source sequence is the count of posts applied to that result, because lti_ags is a bursting source and the protocol supplies no counter — two posts bearing one instant are an ambiguous_order refusal without it, which a test asserts. THREE REFUSALS THAT ARE NOT GAPS, each with its own test: a line item nobody bound; a score the platform reports as Pending, PendingManual or Failed, because a provisional number recorded as scored is the grade in force until something else arrives; and a line item whose maximum disagrees with the bound scale, because rescaling 40/50 into 80/100 on the way in is exactly the silent mapping M10.7.f blocks. A REFUSED SCORE DOES NOT FAIL THE POST. AGS is the platform's protocol and its rules were met; failing a well-formed post because nobody on this side bound the line item would make a Metis omission look like a platform error. The post stands and LtiScoreResponse.ledger carries the receipt, so the platform is told exactly what became of the number instead of being told nothing. The read path is GET /api/gradebook-ingress/projection, derived from the entries on every read rather than served from the stored row — a projection that has fallen behind its own ledger is the failure a read path can hide — and scoped to the caller's own claim, so a line item somebody else bound answers the same 404 as one nobody bound. RECORDED, NOT FIXED, AND FOUND BY BUILDING THIS: THE GRADE IS DURABLE AND THE LINE ITEM IT HANGS ON IS NOT. _line_items is the same kind of class-level dict the scores were in, and sync_score refuses a line item it cannot find — so after a restart a platform posting to a line item Metis itself created is told line item is not registered and answered 403, forever. The grade road is durable from the post onward and the thing being posted about evaporates. The fix has the shape this item just built: a table, written where create_line_item writes today. The binding table is NOT that table — it says what a line item scores, not that it exists — and conflating them would let a registration nobody made be inferred from a binding somebody did. The absence is held as its inverse now: test_the_grade_ledger_has_a_writer_under_src fails the day the ingress is deleted, because every ledger test would still pass without it and the store would quietly go back to being one no road reaches. tests/test_grade_interoperability.py, 16 tests against a freshly migrated Postgres, driving import, correction, visibility, duplicates and out-of-order arrival across all five sources at once. THE DEFECT IT WAS WRITTEN TO FIND: THE ALPHABET WAS DECIDING WHICH GRADE STOOD. SourceRef.ordering_key ended in the system's NAME, and the ambiguity check compared that same key — so a tie between two systems was invisible to the check that exists to catch ties, and was broken by lti_ags < oneroster. A platform post and a roster record landing on the same second ordered by spelling, and the one that sorted last became the grade in force. The module's own docstring says a tie is an ambiguity and not a coin flip; across sources it was a coin flip with the alphabet as the coin. A SEQUENCE IS A COUNTER INSIDE ONE SOURCE — an import's first event is not before a platform's seventh — so separates_from now asks what the sources actually said: a different instant separates them, the same instant from two systems does not, and a source's own counter separates only its own events. Such a pair is refused as ambiguous_order naming BOTH systems, and as_history refuses to derive over one too, so a ledger that somehow holds one cannot quietly project a winner. A SECOND, SMALLER ONE: a projection over entries that do not yet make a history — a rescore whose opening score has not arrived, which is ordinary while an import catches up — reached the caller as a pydantic validation dump from two layers down. It is a LedgerError naming the grade and how many entries are waiting; the entries stay, and the projection works the moment the score lands. What the fixtures confirmed already worked: a retry is one event, and a retry carrying the version its first attempt carried is still accepted because a delivery guarantee is not a stale-version error; one identifier from two systems is two events; a term replayed BACKWARDS into a second tenant produces a byte-identical history; one system corrects another's event by the id that system used; and M7.10's visibility matrix over a ledger-derived history withholds the value from a learner before release and shows them the correction after it.
    • M10.7.f Publish per-cell compatible/lossy/extension/unsupported/version- skew results and block silent mapping. Evidence: compatibility report. Evidence: evidence/v1-workbenches/metis-interop/compatibility-report.json, generated by scripts/export_compatibility_report.py from the same registry M10.7.a's matrix comes from, with 21 tests in tests/test_interop_compatibility.py and 15 in tests/test_interop_loss.py. All 49 cells classified: 15 compatible, 13 lossy, 4 extension, 1 version-skew, 16 unsupported. fidelity is a REQUIRED field, so a new cell cannot arrive unclassified, and a label its own fields contradict will not construct — a compatible cell that names a loss, a lossy one that names none, an extension that says nothing it adds. BLOCKING SILENT MAPPING, and it was not blocked: EVERY INBOUND STANDARDS PAYLOAD WAS SILENTLY DISCARDING WHATEVER IT DID NOT MODEL. Pydantic's default is to drop an undeclared field without a word, so a SIS sending middleName, grades and agents, a platform sending an xAPI authority, a sensor sending a Caliper session, or a player committing cmi.exit got back a success over a record quietly smaller than the one it sent. Refusing the whole sync over middleName is not interoperability either — so the payload is now walked against the model that parsed it and every unread path comes back in fields_not_carried on the response. services/interop_loss.py, wired into all four ingresses (OneRoster sync, xAPI statements, Caliper events, SCORM commit), bounded in depth, breadth and list length because the payload is somebody else's. The rule is then enforced in the type: a cell that RECEIVES over HTTP and declares a loss nothing announces WILL NOT CONSTRUCT, and the report publishes silent_losses (empty) so a reader can check it without running the suite. The rule binds where a sender exists to be told — an HTTP request with a response. A library call (the badge parser) has no sender to answer, so its losses are published in the report instead, which is the other way of not being silent; the three cells it applies to say so in their own rows. THE SECOND RATCHET, WHICH IS WHY THIS ITEM FOUND ITS OWN CHECKLIST STALE: an implemented cell proved itself by naming a handler that exists, and an UNSUPPORTED cell proved nothing at all. M10.7.b built roster_reference, the delta watermark and tobedeleted tombstones, and three cells went on reporting all three as unbuilt for an entire item — the matrix over-reported gaps and nothing could contradict it. Every unsupported cell now names the symbol whose ABSENCE makes it true (OneRosterLineItemRecord, ScormSequencing, XapiStateDocument, open_badges_service.verify_proof), and the test fails the day one of them exists. The three stale rows are corrected: sourcedId persistence and tombstones are compatible, the delta is an extension (a payload-level watermark, not OneRoster's REST filter), and the identity join's own row now says what it does — sourcedId first, a normalized email fallback ONLY on a first sync, which M1.13's emailAsJoinKey: false does not sanction and without which a first sync copies every user the estate already has. 17 losses named as field paths, each demonstrated by calling the code: SCORM 1.2 refusing long-fill-in (the one version-skew cell), the Caliper envelope leaving actor.name behind, fields_not_carried naming what a badge cannot hold — and A BADGE'S validUntil IS NEVER READ, so an expired badge parses exactly like a live one, which the report now says out loud. The matrix moves with it: 28 HTTP, 5 library, 16 unsupported.
  • M10.8 Run BFF authorization, idempotency, concurrency, pagination, rate, SSRF, and tenant negative suites. Evidence: route matrix complete. <!-- SPEC 2026-09-11: .a's inventory is generated from apps/metis/service/openapi/metis.openapi.json (767 operations) joined to the M8.15 classification register in services/api_contract.py; never a hand-written route list. .b runs anonymous plus the three JWT roles plus a second tenant, and because student is the default role every learner case is also the "token with no role claim" case. .e enumerates every outbound fetch by listing the httpx call sites under src/metis/services (source connectors, LMS/LTI, webhooks, media) before writing SSRF cases. .f is a verify gate: a route/role/failure row with no test fails it. -->
    • M10.8.a Generate full route-method-operation/role/audience/object/ property/function inventory from runtime/OpenAPI and map tests. Evidence: route matrix. Evidence: evidence/v1-workbenches/metis-bff/route-matrix.json, generated by scripts/export_route_matrix.py from services/route_matrix.py, with 17 tests in tests/test_route_matrix.py. 882 operations — the SPEC's 767 was 2026-09-06 and the surface has grown — each joined to the function that serves it, the roles that reach it, its audience, its object, the protected properties its responses can reach, and the test modules that name its path. Nothing is hand-written: the operations and statuses come from the published document, the handler and its dependencies from the AST of metis/api/*.py and metis/main.py, the posture from M8.15's register, and the test mapping from path literals in tests/*.py. 653 operations are named by a test and 229 ARE NAMED BY NONE. The mapping is deliberately literal — a test reaching a route through a helper is not seen — so 229 is an upper bound on what is untested and the right direction for a coverage gate to be wrong in. Worst by object: admin 57 of 198, item-bank 28 of 137, and measurement-ledger 12 of 12 — a whole surface no test names. That list is M10.8.f's. THREE DEFECTS IN THE INVENTORY ITSELF, each found by a check written to fail: 4 operations matched no handler (a trailing slash normalised away, a {asset_path:path} converter published without its converter, and /health and /ready hanging off the app rather than a router); an f-string's trailing piece was being read as a path of its own; and, the one that mattered, NINE ROUTES WERE READ AS ANONYMOUS BECAUSE THEIR AUTHENTICATION WAS ONE DEPENDENCY DEEPER — export_diagnostics routes depend on export_tenant, which depends on get_current_user. A scan that reads a route as open when it is not is wrong in the worst direction, so the dependency graph is walked, metis/deps.py included. Roles now: 537 instructor, 179 any-authenticated, 125 admin, 41 anonymous — and every anonymous one has a posture M8.15 declared open, checked against the register rather than restated. The protected-property column reads 0, and 0 is only worth reading because the report names what was looked for: one published schema carries one (AssessmentContext.answer_key) and only a REQUEST body reaches it, so it is accepted and never returned. The walk that found nothing is proven able to find something — a synthetic schema three deep through a list and a union — and to survive a reference cycle.
    • M10.8.b Test anonymous/learner/guardian/author/reviewer/teacher/admin/ operator and cross-tenant IDs/fields/search/counts/exports. Evidence: authorization suite. CLOSED 2026-09-18. Both halves are done: the cross-tenant read sweep below (2026-09-12) and the persona vocabulary the last paragraph was open on. THE EIGHT ARE EXPRESSIBLE NOW (schemas/personas.py). A capability table says what each persona may do, and CurrentUser asks it rather than asking which of two words a token carried — which is why five personas could not be added before without rewriting every route. A NEW PERSONA WIDENS NOTHING. student still means learner and instructor still means teacher, asserted by a test over all three legacy words, so no token that exists reaches anything it did not reach yesterday. is_instructor() asks for teach_class rather than author_content deliberately: that seam gates 537 operations, and granting all of them to the new author and reviewer would be the one thing adding a vocabulary must not do. A ROLE NOBODY DEFINES RESOLVES TO NOTHING, not to the narrowest persona: a token carrying role: "superuser" must not quietly become a learner and work, because then nobody finds out that the issuer and this service disagree. THE SWEEP RUNS ALL EIGHT: tests/test_route_authorization.py offers every published operation to anonymous, a token with no role claim, a learner, a guardian, an author, a reviewer, an operator and a teacher, each below its bar — 18 tests, every offer refused 401 or 403 and never 422, because a validation error means the request reached the body. WHAT WAS DELIBERATELY NOT DONE, and is now asserted so a later item re-takes the decision rather than drifting past it (user decision 2026-09-18): the 537 operations behind the instructor seam KEEP it. Which of them are authoring, which reviewing and which teaching is a per-route judgement the checklist does not supply, so an author, a reviewer, a guardian and an operator reach what a learner reaches and no more — which the sweep proves rather than assumes. WAS OPEN ON FIVE OF THE EIGHT PERSONAS AND ON THE CROSS-TENANT READ SWEEP. CurrentUser.role holds one of THREE values — student, instructor, admin — so guardian, author, reviewer, teacher and operator are distinctions this item asks the surface to make and it cannot make them: every route that should tell a guardian from a learner, or a reviewer from an author, sees the same caller. And the two-tenant read sweep (create in A, read as B, over ids/fields/search/counts/exports) exists for the eight authoring routes M8.14.b built it for, not for the other 874. Closing this means a role vocabulary that can express the eight and a generated two-tenant sweep. THE CROSS-TENANT SWEEP IS NOW GENERATED AND RUN, 2026-09-12. tests/test_cross_tenant_read_sweep.py takes every GET operation in the published document that needs no path parameter — 197 collection reads — and offers each to a second instructor who has created nothing. 138 of them answered 200, 46 refused with 403, and NOT ONE returned any of the 8 identifiers belonging to the first tenant. The assertion is over BYTES, not counts: a count passes whenever two tenants happen to hold the same number of rows, and a field check misses an id echoed inside a nested payload or an error message. The sweep refuses to pass vacuously — it asserts it had identifiers to look for and that it reached more than twenty permitted reads. AND IT FOUND A LIVE 500. /api/admin/authoring-hub/workspace answered internal_server_error for the second instructor and 200 for the first. AuthoringHub.viewer_id was typed Identifier, whose pattern is ^[a-z][a-z0-9._:-]{2,159}$ and so requires a leading lowercase LETTER — but a viewer id is a USER id, which is a UUID. Ten of the sixteen possible leading hex characters are digits, so the authoring hub was unopenable for most real callers and fine for the rest, which is why it looked like it worked. The unit tests never caught it because they passed viewer_id="actor:one", a slug shape the route never produces. The field is a uuid.UUID now, the tests pass a real user id, and both a digit-leading and a letter-leading account answer 200 against the running service. One schema moved in the published document. WAS STILL OPEN ON: the five personas. CurrentUser.role held three values and the item names eight. Closed 2026-09-18, above. DONE, and it found a cross-tenant hole. tests/test_route_authorization.py drives the REAL application with REAL bearer tokens — nothing overridden, because the dependency you would override is the thing under test — and offers all 882 operations to every caller below their bar: anonymous, a token with NO ROLE CLAIM (which defaults to student, so it is a learner's token), a student and an instructor. 1,628 offers, every one refused 401 or 403 and NEVER 422, because a validation error means the request reached the body, which means it passed the gate. Four bad tokens — signed by somebody else, expired, carrying no subject, and not a token at all — are each 401. THE DEFECT: A CALLER COULD NAME ITS OWN TENANT. oneroster_service opened with tenant_id = body.tenant_id or actor_user_id, so an authenticated administrator could put any tenant's id in the request body and read, write and re-point THAT tenant's roster_reference rows — the mapping that decides which Metis user a SIS sourcedId becomes on every later sync, so re-pointing it lands the next sync's name, email and withdrawal on an account of the attacker's choosing. M8.14 established the property this breaks: a caller cannot name its own tenant. The tenant is now derived from the authenticated user, by a deps.tenant_of shared with the twenty modules that had each written the same uuid5 inline — which is how a twenty-first came to take it out of a request body — and a body that names a different one is REFUSED 403 rather than quietly ignored, because ignoring it would leave a caller believing it had written somewhere it had not. A SECOND FINDING, IN THIS SESSION'S OWN GATES: three drift gates (M10.7.a's coverage matrix, M10.7.f's compatibility report, M10.8.a's route matrix) compared the committed artifact to the generator BYTE FOR BYTE. The commit hook reformats generated JSON after the generator writes it, so all three were green in the working tree and RED at HEAD. They compare documents now, and the fix is verified by running prettier over the artifacts and re-running the gates.
    • M10.8.c Test idempotency key replay/conflict/scope, ETag/expected-version races, simultaneous edits/decisions, and event dedup/order. Evidence: mutation suite. CLOSED 2026-09-18 ON WHAT THE DOCUMENT SAYS — the last open half. ALL 177 KEYED OPERATIONS PUBLISH THE CONFLICT NOW; it was 0, then 95, and the remaining 82 were held as a ceiling with the note that shrinking it meant "reading those mappings rather than running a script over them". They were read, one at a time, and every one of the 82 could answer 409: - twenty-six modules map a domain error whose status is a runtime attribute — five of those classes DEFAULT to 409, so any bare raise is one, and the rest pass 409 explicitly at the raise; - agent_runtime maps any code its table does not list to 409 by default; - generation's cancelled failure class is 409 in FAILURE_STATUS. Each is documented against that evidence and not one was documented for tidiness — documenting a status an operation cannot return is the same error pointing the other way. THE CEILING IS GONE RATHER THAN SET TO ZERO, which is the rule M8.15's own status ratchet was retired under: a baseline over nothing is a gate that cannot fail. The test asserts the thing itself, so an operation that gains a key without its conflict fails immediately. THE EXPECTED-VERSION HALF: 25 of 26 publish 409 or 412; it was 2 of 14. THE ONE THAT DOES NOT IS CORRECT AND IS NAMED RATHER THAN COUNTED — POST /api/gradebook-corrections/decisions answers 200 with a typed CorrectionRefusal in the body, because a stale version there is a DECISION about a correction and not a transport failure: the reviewer is told what the store holds and why their decision was not taken, and a 409 would lose the reason. Evidence: tests/test_mutation_concurrency.py, 11 tests, and 142 across the contract, openapi and route-matrix gates. WAS OPEN ON WHAT THE DOCUMENT SAYS. 177 of the 463 mutating operations take an idempotency key and NOT ONE publishes the status a replay or a reused key answers with; 12 of the 14 taking an expected_version publish neither 409 nor 412. The behaviour is right now — the last test in the suite measures the silence and fails the day any operation starts documenting it — but a client generated from this document has no branch for the conflict, so it reads a 409 as an unknown error. Closing this means the statuses reaching the published contract. THE STATUSES HAVE REACHED IT, AND THE RATCHET THAT HELD THEM IS GONE. M8.15 recorded FIFTY-SIX modules that raise a status none of their operations document — 404 forty-five times, 409 forty, 403 twenty-one — and held the number as a ratchet because documenting them was "a change to ninety route modules' decorators". That change is made. The gap is ZERO modules, and the ratchet's own rule said what to do next: delete it rather than empty it, because a baseline over nothing is a gate that cannot fail. So STATUS_GAP_BASELINE and status_gap_ratchet are deleted and test_no_module_raises_a_status_it_never_documents asserts the thing itself — a module that starts raising something it does not publish fails immediately instead of being written down and lived with. DOCUMENTED PER HANDLER, NOT PER MODULE, which is the part that needed judgement rather than a script. A function's statuses are what it can actually raise: the HTTPExceptions in its own body plus those its module's error mapper can produce for it. A read that catches nothing gets nothing — /evaluation on six of these modules promises no 404, because it cannot answer one. Documenting a status an operation cannot return is the same error pointing the other way. AND THE MODULE-LEVEL GATE WAS COARSER THAN IT LOOKED. admin was not in the baseline at all — it raises 409 and some one of its eighty-nine operations documented it, which satisfied a gate that asks about modules. Fifty-one of its handlers still published nothing. The operation-level measurement in test_mutation_concurrency is the finer instrument and it is the one this item is written against. KEYED OPERATIONS PUBLISHING THE CONFLICT: 0 OF 177 BEFORE, 95 NOW. The remaining 82 are held as a ceiling that may shrink and must never grow, and they are not one kind of thing. Some handlers genuinely cannot conflict. The others raise a domain error whose status arrives as a RUNTIME ATTRIBUTE — ContentFreshnessError defaults to 409 and never writes it as a literal — so the operation answers 409 and no static pass can see that it does. Shrinking the ceiling means reading those mappings one at a time, which is what keeps this open. The refusals-unpublished debt fell with it: 121 operations required a role and documented no failure at all; 102 do now. DONE, AND THE FIRST CLAUSE WAS A DEFECT. AN IDEMPOTENCY KEY IS A RETRY TOKEN AND THIS ONE WAS A UNIQUENESS TOKEN. aggregate_store consulted the key LAST, through a unique constraint on the outbox, so a caller that did not hear the answer and sent the same request again was refused stale — indistinguishable from somebody else having got there first, which is the ONE distinction the key exists to make. The key is now asked before the revision checks: the same intent is answered with the first attempt's acknowledgement (same revision, same ETag, same audit line, same outbox event, replayed: true), and a key pointed at a different intent is refused, because answering it with the first attempt's result would confirm a write that never happened. The store's own outbox row carries the audit id now, so a replay answers with the acknowledgement that was issued rather than a fresh one that merely looks like it. A SECOND, SMALLER ONE: that refusal reached the caller as 422 validation_failed — the authoring layer's fallthrough for every refusal it had no case for. A well-formed request that names an intent the caller has already named differently is a conflict, so the published outcome vocabulary gains duplicate_key (409). M8.6's own rule then did the rest: a state in the contract that nothing produces is a promise nobody keeps, so its reachability test obliged a case that produces it. tests/test_mutation_concurrency.py, 10 tests against a freshly migrated Postgres: a retry written once and answered twice, an edit retried against its own first attempt, a key pointed elsewhere, a key with a changed payload, two tenants using ONE key and getting two writes, a transition replayed under its own key, a stale revision told what the store holds, two threads at one revision producing one winner and one refusal with nothing written by the loser, and the same retry run twice at once still writing once. Event dedup and order were done in M10.7 and M8.11 and are not re-measured here: xAPI statement conflict, Caliper event dedup and ordering, SCORM commit replay, and the grade ledger's (tenant, system, source_event_id) with its advisory lock.
    • M10.8.d Test cursor signing/scope/stability/expiry, filter/sort limits, rate/quota/budget/concurrency bypass, and privacy-safe errors. Evidence: protocol suite. THE RATE LIMITER EXISTS NOW; QUOTA AND CURSOR EXPIRY STILL DO NOT. services/request_throttle.py joins the three pieces that were lying about separately, and main.py applies it as middleware — a middleware and not a dependency, because the finding was about every authenticated route and a dependency would have had to be added to eight hundred and ninety-odd operations to answer it. tests/test_backpressure.py — renamed from test_backpressure_absence.py, because it is not about an absence any more — drives the REAL application against the REAL Redis and watches the fourth request of a three-request window come back 429 with a Retry-After. Its old assertions are the same claims pointing the other way: the limiter HAS a caller, the configured number IS read, a route CAN answer 429. IT IS A FIXED WINDOW AND THE CODE USED TO CALL IT SLIDING. Measured: check_rate_limit sets a counter with a TTL on the first request and increments until the TTL expires, so the window is anchored to that request rather than moving with the clock. The docstring said sliding and the implementation never was. The consequence is the ordinary one and it is now written down: a caller can spend the whole allowance at the end of one window and the whole allowance at the start of the next, so the real short-term ceiling is twice the configured number. rate_limit_burst IS STILL READ BY NOBODY, and a test asserts that on purpose. A fixed-window counter has no burst parameter to read; spending it would mean a token bucket, where a burst allowance is a real thing with a refill rate. Folding it into the window maximum would be inventing a meaning the name does not have. (That test had to stop grepping for the name to find this out — every mention of rate_limit_burst is now in prose EXPLAINING that nothing reads it, so it asks the AST for an attribute access instead.) IT FAILS OPEN WHEN REDIS CANNOT ANSWER, deliberately and with a test. A limiter keeps load off a service; one that refuses everything when its own dependency is down converts a Redis outage into a total outage. Every other way of not running names itself too — disabled, no_limiter, unthrottled_path, nonpositive_maximum — because a throttle that quietly does nothing is the state this cell was opened on, and "it did not run" should be a value a test can read rather than a silence it has to infer. IT IS ON IN PRODUCTION AND OPT-IN ANYWHERE ELSE, and that policy was WRONG THE FIRST TIME. Enabling it everywhere but the testing environment looked obviously right and failed TWENTY-SEVEN suites, because they build the real application with default settings — environment defaults to development — and drive hundreds of routes at one address in one process. test_route_authorization alone offers every operation to every caller. Those suites measure the service, and a throttle firing inside them measures the suite. The finding was about a DEPLOYED service accepting requests without limit, so that is where it runs by default; rate_limit_enabled turns it on anywhere. The policy is a named function rather than an expression inline in the middleware, because it is a decision somebody should be able to read and test on its own — and because the first version of it was wrong. test_protocol_limits's "nothing here limits a rate" flipped with it, and the new sentence is the precise one: the limiter is INSTALLED here and not ARMED here, so sixty requests still pass and the reason has changed from "nothing limits anything" to "this configuration asked for no limit". AND THE CURSOR EXPIRES NOW. The item-bank cursor — the one opaque cursor in the estate, the rest being positions — carries an iat and is refused past thirty minutes. IN BOTH DIRECTIONS: a cursor issued in the FUTURE is refused too, because otherwise a forged issued-at would be the way to make one last for ever and the expiry would check only honest cursors. Version 1 cursors are refused rather than honoured — one with no issued-at cannot be expired, so accepting it would leave the hole open for as long as anybody held one, and for a page token the cost of that is asking for the first page again. NOTHING ABOUT IT IS IN THE PUBLISHED DOCUMENT, deliberately: a lifetime is the server's business and a cursor_ttl field would be a promise a client could rely on. A TEST HAD TO STOP ASSERTING DETERMINISM TO ALLOW IT. An issued-at makes two otherwise identical searches produce different cursor strings, and test_signed_cursor_is_deterministic_… compared the bytes. It passed only while both calls landed inside the same second — a test that fails on a slow morning. It asserts what actually mattered now: the same question asked twice continues at the same place, checked by decoding rather than by comparing strings. STILL OPEN ON QUOTA. A quota is a budget over a longer period than a window — a month's generation spend, a term's exports — and nothing keeps one. The rate limiter is a window and cannot be read as one. WHAT THE ORIGINAL MEASUREMENT SAID. Sixty requests in a row to an authenticated route are all served; the only middleware this application installs is a response timer and CORS; no route answers 429. There is no limiter to bypass because there is no limiter, and the test that says so is written to fail the day one appears. No cursor carries an issued-at or a lifetime either: the signed one binds a SNAPSHOT, so it stops working when the data moves, which is stability and not expiry. DONE, AND ONE PARAMETER WAS BOTH UNFLOORED AND UNCAPPED. GET /api/authoring/{kind}?limit was a bare limit: int = 100 that reached LIMIT :limit unchecked: ?limit=-1 came back a 500 — the database's complaint, rendered as "an unexpected error occurred", for what is plainly a bad request — and ?limit=100000000 was simply served. It is Query(default=100, ge=1, le=500) now, so the bound is ENFORCED and PUBLISHED, and ?revision is floored at 1 because revisions start there. The rest of the limits hold and are now checked against the document rather than believed: every page size a caller can name, in a query parameter or a request body, declares a maximum AND a minimum; every sort/order_by is a closed vocabulary rather than a column name; and the only integer parameter left without a floor is a stream position, recorded by name so a new one has to be looked at. CURSORS. One opaque cursor exists — ItemBankSearchRequest.cursor — and it is signed and bound to the query digest, the snapshot digest, the tenant and the role. Every other paginated surface hands back a POSITION, which is forgeable into nothing a caller could not already ask for, because the query is re-scoped server-side on the next call. The test pins that there is exactly one opaque cursor, so a second one arrives with a signer or fails here. BUDGET is built and CONCURRENCY was proven in M10.8.c: reserve_budget takes a reservation inside a transaction, so a job that does not fit never reaches a provider — but reserved_cents appears in NO published schema, so a caller cannot see the reservation that binds it. PRIVACY-SAFE ERRORS. Another tenant's aggregate answers not_found rather than forbidden — a 403 on a record you cannot see confirms that it is there — and carries none of its payload; a validation failure is a fixed sentence with a field list and no traceback. tests/test_protocol_limits.py, 11 tests.
    • M10.8.e Test URL/feed/LMS/repository/webhook/media fetch SSRF, redirects, DNS rebinding, internal metadata, schemes/ports, size/time bombs. Evidence: SSRF suite. OPEN ON DNS REBINDING FOR EIGHT OF THE NINE FETCHES. Source intake resolves, checks, and then PINS the peer address it validated, so a name that answers differently on the second look is caught. The LTI paths resolve and check and then hand the URL to httpx, which resolves again — the window between the two lookups is exactly what a rebinding attack aims at. Closing this means a transport that dials the address that was validated, which is what source_intake_network already has and nothing else does. THE REBINDING WINDOW IS CLOSED. outbound_targets.pin_to_validated_address returns a request aimed at the LITERAL the policy approved, and both LTI fetches use it, so neither hands a NAME to a client that would resolve it a second time. The hostname rides in Host so the server still routes, and in the TLS SNI so the CERTIFICATE is still verified against the name — drop that and pinning would have bought safety by giving up authentication. httpx.stream cannot carry extensions, so both fetches now go through an httpx.Client; the SNI hostname travels there. AND THE PIN IS ASSERTED BY WATCHING WHERE A REQUEST IS AIMED, which is the part the existing cases could not do. Every test here already passed whether the address was pinned or not: they drive a fetch at a 0.2-second timeout and assert it FAILS to connect, which proves the policy let it through and says nothing about where it went. The nine new cases capture the outgoing request instead — the fetcher, the policy and the pin are all real, and only DNS and the HTTP client are substituted at their boundaries. One of them answers public on the first lookup and private on every one after, and the request still goes to the public literal, because there is no second lookup to move it. A SEAM THAT WAS MISSING ON ONE SIDE. HttpxTransport took a resolver and HttpsJwksFetcher did not, so the JWKS path could only have been tested by the internet agreeing with the test. It takes one now, the same way, and the two fetchers are symmetric. WHAT KEEPS THIS OPEN, AND A CORRECTION TO WHAT THAT MEANS. Every fetch somebody OUTSIDE aims is pinned now — the two LTI ones here and the badge status list M10.7.d added, which is the third and the most attacker-chosen. The seven model providers read their base URL from settings and are not. THE FIRST VERSION OF THIS SENTENCE SAID THEY SHOULD GO THROUGH "THE SAME POLICY AND THE SAME PIN", AND THAT IS WRONG. Reading the call sites to do it showed why: refuse_unless_public would REFUSE an operator who points a model provider at an internal gateway on 10.0.0.5, which is an ordinary and deliberate deployment. The policy that is right for a URL a platform chose is the wrong policy for a URL an operator chose, and applying it uniformly would have broken working estates in the name of closing a checkbox. So what remains is smaller and different: pinning without the globality check buys little on its own — each request re-resolves, so an operator whose DNS is hostile is compromised from the first answer either way — and the useful control for these would be an operator-declared allowlist of model hosts, which is a new mechanism rather than a reuse of this one. That is what this stays open on. DONE, AND TWO OF THE NINE FETCHES POINTED WHEREVER SOMEBODY ELSE SAID. The spec's first instruction was to enumerate every httpx call site under src/metis/services before writing a case, and enumerating found nine: seven read their base URL from settings, and two do not. lti_trust fetches the JWKS an ADMINISTRATOR registered, and lti_services_client calls the AGS and NRPS endpoints that arrive in CLAIMS INSIDE A LAUNCH TOKEN — a platform that signs its own launch chooses those. Neither checked where it was going, so https://169.254.169.254/latest/meta-data/iam/security-credentials/ was a URL this service would have called WITH A BEARER TOKEN ATTACHED, and https://10.0.0.5/lineitems was another. services/outbound_targets.py makes the check source_intake_network already made, written once so the two agree: resolve the name and refuse unless EVERY address it answers with is public. One public answer beside a private one is refused, which is the shape of a rebinding attempt. A TEST WRITTEN TO ASK FOUND A HOLE IN THE CHECK ITSELF: Python reports 224.0.0.1 as is_global, because multicast is not in the special-purpose registry it consults, so multicast, reserved and unspecified addresses are now refused by name. LOOPBACK IS A DECISION EACH CALLER MAKES rather than a global rule. M10.6 allowed it for JWKS deliberately — RFC 8252's exception, and the only reason a platform on this machine is testable — and this item does not undecide it. The AGS client refuses it: a key document is read and discarded, while those calls carry a token and write. SIZE BOMBS. Both LTI fetches capped the body with response.content[:limit], which caps what is KEPT and not what arrives — the whole body is in memory by the time it is sliced. Both read with a budget now and stop mid-stream, proven against a response that never ends. Already held and re-checked rather than rebuilt: no fetch anywhere follows a redirect (asserted across all nine), and source intake's own SSRF suite covers private DNS, redirect revalidation, peer mismatch, oversize, deadlines, credentials, fragments, ports and the host allowlist. tests/test_outbound_ssrf.py, 32 tests, including an inventory that FAILS WHEN A NEW httpx CALL SITE APPEARS until somebody says who aims it.
    • M10.8.f Fail CI for any untested route/role/failure row or protected field leak. Evidence: coverage gate. Evidence: evidence/v1-workbenches/metis-bff/coverage-gate.json, computed by services/coverage_gate.py from M10.8.a's inventory, runnable as pnpm run metis:coverage:gate (exit 1 on any failing row), with 11 tests in tests/test_coverage_gate.py. Three rows, each failing in BOTH DIRECTIONS — a new bad row fails it, and a recorded row that got fixed and was left on the list fails it too, because a list that can only grow is an excuse with a file name (M8.14 settled that for tenancy): - AN UNTESTED ROUTE. 229 of 882 operations are named by no test. That is today's debt, recorded operation by operation in src/metis/data/coverage_gate_baseline.json, and the gate is what stops the 230th. - A REFUSAL NOBODY PUBLISHED. 121 operations require a role and document no 4xx at all, so a client generated from the document has no branch for the 403 it WILL get — M10.8.b proved every one of them refuses. An anonymous route is not asked for one, and a test proves the gate makes that distinction. - A PROTECTED FIELD IN A RESPONSE. No baseline for this one and no way to record an exception: answer_key and the eleven properties beside it may not become reachable, and the gate fails with nowhere to write it down. Zero today. Every check is paired with a doctored surface that fails it: an invented untested route, a baseline entry that gained a test, an operation that stopped documenting its refusal, and a response that reaches answer_key. The baseline lives beside the code rather than in evidence/, because evidence is written by a generator and a baseline is edited by a person.
  • M10.9 Run full-journey Playwright: source ingest to grounded course publish. Evidence: live backing and verified target. <!-- SPEC 2026-09-11: host per stage. Today source intake, grounding and authoring exist under BOTH apps/metis/web/src/app/create/* (brief, outlines, lessons, item-bank; port 3020) and apps/metis/admin (grounding, workspaces, hub, publication; port 3021). Until M0.1.c is ratified the journey is written against the pages that exist and each stage names its app and route; a stage with no page in either app is a recorded gap, never a mocked step. Live backing per §2.5.2/§2.5.3; next build && next start and chromium only per §2.5.7. The "authoritative target" in .d is the LMS the M9.12 publication saga reads back from; a local LTI platform stub counts only if it is the same stub the saga already trusts, and the evidence says so. --> THE ROOT BLOCKER IS CLOSED 2026-09-13. It held for two days and it was real: source intake had no page in either application — apps/metis/web/src/app/create had no source or intake route, and the admin app's source-impact and training-intake are different surfaces (impact analysis over a source already held, and training-data intake). The admin app now serves /source-intake, and the stage map it is measured by reports ELEVEN stages walkable out of eleven, where it reported ten yesterday. WHAT THAT DOES AND DOES NOT UNBLOCK. The first stage can now be walked through a page by a person, which is what .a–.e each waited on. It does not settle the rest of any of them: .a still owes the standard and LMS seeds and the extraction assertion, .b still needs content seeded and the gateway — a third process — stood up, .c still owes three of its four gates, and .d/.e still have no authoritative target to read back from. This parent stays unchecked on those, not on the page.
    • M10.9.a Seed roles/tenant/standard/LMS and upload one valid plus malicious source through real backing; assert quarantine/extraction/provenance. Evidence: intake journey. OPEN ON TWO THINGS, DOWN FROM THREE. The standard and LMS seeds are not done (the roles and the tenant are), and EXTRACTION is not asserted: this stage proves quarantine, scanning and provenance, and the extraction step is the next endpoint along. THE PAGE BLOCKER IS CLOSED 2026-09-13, AND IT WAS THE BIG ONE. It read: the source-ingestion API publishes 41 operations — 52 when re-counted from app.openapi() on 2026-09-13; the API grew, the finding did not change — and NEITHER apps/metis/web NOR apps/metis/admin references one of them, so there is nothing to walk in a browser — recorded, never mocked, as this item's spec requires, and measured by a test that fails the day a page appears. That test has now flipped, on purpose and in both places it was taken: test_a_page_reaches_the_source_intake_api in tests/test_source_intake_journey.py, and test_every_stage_of_the_journey_now_has_a_page in tests/test_journey_stage_map.py. Both assert the opposite of what they asserted yesterday and are worth more that way round: a deleted route, or a page that stops naming the intake API, puts the first stage back out of reach and fails there. - THE PAGE. apps/metis/admin/src/app/source-intake/page.tsx, routed at /source-intake and reached from the sidebar, over a proxy at src/app/api/source-ingestion/[[...path]]/route.ts and a client of its own — admin-client is pinned to /api/admin and this API is not under it. The proxy forwards await request.arrayBuffer() and not .text(), because a multipart body does not survive being decoded to a string and re-encoded. - WALKED IN CHROMIUM, 4 tests in apps/metis/admin/e2e/source-intake-workspace.spec.ts, against the real page and a stubbed intake API — the routing, the upload control, the digests and the findings are the application's own. The stub returns the receipt the service really produces, INCLUDING malware_scanner_unavailable, so a page that quietly dropped the finding an estate with no scanner gets would fail there. Each test also carries the suite's axe pass: zero serious or critical WCAG 2.1 AA violations. - THE TWO DIGESTS ARE SHOWN SIDE BY SIDE, because their difference is the evidence that what is stored is the envelope and not the plaintext, and the page says which of the two things it is looking at rather than leaving a reader to compare hex by eye. - A REFUSAL IS CARRIED, NOT SWALLOWED. The service refuses an upload outright until an intake key is configured; the page renders that detail in an alert and shows no receipt. That is the fail-loud seam .a already proved at the service, now visible to the person holding the file. DONE, AGAINST A RUNNING SERVICE AND A REAL DATABASE. tests/test_source_intake_journey.py, 7 tests, driven over HTTP at METIS_JOURNEY_BASE_URL; receipts in evidence/v1-workbenches/metis-journey/intake-receipts.json. - QUARANTINE. An upload never lands as plaintext: the artifact is an AES-256-GCM envelope under a named key, and the receipt carries the content digest AND the envelope digest, which are different numbers. The service REFUSED to accept an upload at all until a key was configured — a fail-loud seam, met before any of this could run. - PROVENANCE. The source version records intake_kind, the original locator, the retrieval time, the quarantine and evidence URIs, the version number and a tenant DERIVED from the uploader. - THE SCAN DISCRIMINATES. An XML declaring an external entity pointed at file:///etc/passwd is xml_external_entity, CRITICAL; the same pipeline over a Markdown file raises it not at all, which is the check that proves the finding is about the file. - AND IT SAYS WHEN IT CANNOT SEE. With no malware scanner configured every receipt carries malware_scanner_unavailable as a CRITICAL finding rather than a clean bill. An instrument that is not there does not get to agree. - The receipt is bound to the bytes: quarantine id, content digest, envelope digest and size all match the upload it describes, and it names the policy, the policy digest and the scanner-set digest.
    • M10.9.b Review source authority/rights/currency, confirm claims/citations, author objectives/outline/lesson/item/media and clear high-risk grounding. Evidence: authoring journey. OPEN ON THE WALK ITSELF. Every stage page now RENDERS against a live service — which it could not this morning — but the journey database holds no content yet, so the pages render empty. Walking .b means seeding an objective, an outline, a lesson, an item and a media asset through the live APIs first, and the authoring half of that lives in the WEB app, which calls /api/metis/v1/* — the gateway, a third process this journey has not yet stood up. DONE: THE MAP, AND THE FIX THAT MADE ANY OF IT POSSIBLE. evidence/v1-workbenches/metis-journey/stage-map.json, generated by scripts/export_journey_stage_map.py from the two applications' own source, with 4 tests in tests/test_journey_stage_map.py. Eleven stages, and as of 2026-09-13 ELEVEN of them walkable. It said ten until then, and the eleventh — source intake — had NO PAGE IN EITHER APPLICATION, which was M10.9.a's finding arrived at from the other direction: .a searched for references to the API, this asks which stages have a route, and the two agreed. They agree still, in the other direction: the admin app serves /source-intake and this map now reports no unwalkable stage at all. THE JOURNEY CROSSES TWO HOSTS, and the map says which: authoring (objectives, outline, lesson) is the web app against the gateway, and reviewing (grounding, claims, publication) is the admin app against the service. Neither application serves both halves. The five admin stage pages — grounding-gates, rights, claims, media, publications — were driven in chromium against the live service and every one rendered its own heading with no authentication alert. That is only true because of the defect this stage found and fixed: React runs a CHILD's effects before its PARENT's, so every admin page fetched before AdminAuthProvider installed the API client's token accessor, every first request went out unauthenticated, and nothing re-fetched. Against mocks — which is all any existing admin spec has ever run against — it never showed. Against a live service the whole admin application read "Authentication required".
    • M10.9.c Run correctness/accessibility/quality/review gates, resolve a blocker through a new revision, and bind approvals. Evidence: review journey. OPEN ON THREE OF THE FOUR GATES AND ON THE BROWSER. This walks the REVIEW gate — human signoff — end to end over HTTP; the correctness, accessibility and quality gates have their own routes and are not walked here, and none of it is driven through a page, because the review stage's pages need content the journey database does not hold (M10.9.b's blocker). DONE: THE SIGNOFF GATE, AND FOUR ROUTES NOTHING HAD EVER CALLED. M7.13 built the gate and tested the ENGINE; /signoffs, /attempt-checks, /outcomes and /invalidations were four of the 229 operations M10.8.a found that no test names. tests/test_review_gate_journey.py, 11 tests against a running service. Every refusal was produced by the service while the test was being written, and each is a door somebody could otherwise walk through: no interactive session (not_interactive), no tenant header ("nothing to scope it to"), no conflict declaration (conflict_not_declared), an evidence digest that is not the one the attempt was prepared against (evidence_changed), a second factor older than fifteen minutes (step_up_required), and a gate opened under a policy digest that is not the rules in force (gate_policy_changed). A PASSWORD IS REFUSED AT THE DOOR rather than counted and found wanting: a stale passkey reaches the engine and comes back step_up_required, while a password never becomes an actor at all. AND THE TWO PROPERTIES THAT MATTER. A quorum of two counts PEOPLE: the same person signing twice leaves one signatory and an uncleared gate, and two people in two roles clear it. A signoff binds the revision it was given, so a NEW REVISION starts unsigned — otherwise "approved" would mean "approved once, for something else", which is the blocker this item asks to be resolved by revising.
    • M10.9.d Build/schedule/publish/sync and verify authoritative target read-back before published state. Evidence: publication journey. OPEN ON THE AUTHORITATIVE TARGET, WHICH IS NOT AN LMS. This item's spec names the LMS the M9.12 saga reads back from; the saga's Destination protocol has exactly ONE implementation in this estate and it is an OBJECT STORE. There is no LMS destination to read back from, so the read-back property is proven against reported facets rather than against a live platform, and the sync half is unwalked. And nothing here is driven through a page: M10.9.b's blocker. DONE: PUBLISHED IS COMPUTED, NOT DECLARED. tests/test_publication_journey.py, 11 tests over HTTP against a running service, including schedules/{id}/preview and schedules/{id}/confirm — two of the 229 operations M10.8.a found that no test names. - A candidate pins nine dimensions into one digest, and the same nine readings of the same revision produce the same digest. - MATCHING BYTES ALONE IS NOT PUBLISHED. Four facets are required — bytes, metadata, visibility, links — and the saga verifies only the first, so a target with bytes matched and nothing else read is unknown: not published, and not failed either. All four matched is published; a facet that came back different is failed; a facet nobody checked keeps it unknown; and a failed attempt cannot read as published however the facets look. - A facet that was NOT CHECKED may not carry a time it was checked. The model refuses that contradiction, which is how the record stays true rather than merely complete. - A schedule is planned against THIS candidate's digest and says so; confirming answers with the REFUSALS, so an empty list is the confirmation and there is no second place for the same fact to disagree with itself. NOBODY CAN APPROVE ANYTHING TODAY, measured over HTTP. The approval route refuses the assembler BY IDENTITY, and publication_tenant makes every user their own tenant — so the second person does not get a refusal about roles, they get not_found. The two rules together mean no candidate can be approved by anyone at all. The derivation documents this limitation in its own docstring; this stage is the first thing to measure it from outside.
    • M10.9.e Retract the source and assert stale/blast-radius/unpublish/ notification/history plus accessibility/tenant/tier boundaries. Evidence: full E2E trace. NOT ATTEMPTED, and recorded as such rather than part-done. This is the one stage that cannot be walked in isolation: retracting a source and watching the blast radius requires a source, content derived from it, and a publication — .a, .b and .d in ONE database, which is exactly what M10.9.b is blocked on. What it will need, from the routes that exist: POST /api/admin/catalog/source-withdrawals, POST /api/admin/source-impact/{package_id}/changes — which takes an expected_graph_sha256, so the source graph has to be there to be named — then GET /api/source-ingestion/{package_id}/source-changes/{change_event_id}/blast-radius, POST /api/admin/catalog/bundles/{bundle_id}/unpublish and the remediation route. The accessibility, tenant and tier boundaries are assertions over that trace rather than separate calls.
  • M10.10 Run full-journey Playwright: item authoring to accessible QTI export. Evidence: validation report. <!-- SPEC 2026-09-11: item authoring pages are apps/metis/web/src/app/create/item-bank/* (capture, stimulus-sets, previews, accessibility, answer-leakage) and apps/metis/admin/src/app/items/[itemId]/* (stem, response, key, rationale, accessibility, gate, use); export diagnostics is apps/metis/admin/src/app/exports. .d's external validation is M10.5.d's receipt. .e's key-absence assertions read page.content() and the network log, never visibility (M9.15's method), and cover the learner preview served by apps/metis/web. --> OPEN, AND NOT BECAUSE NOTHING WAS WALKED. Five children are recorded: .a captured and completed an item, .b proved a preview writes no learner record, .c broke and mended it through three of four gates, .e found and fixed a learner reading the answer key. What the PARENT claims is the end-to-end sentence — authoring THROUGH TO accessible QTI export — and both halves of the ending are missing: .d's export validation is M10.5.d's receipt and M10.5.d is open, and NONE of this is driven through a page. Every stage here is HTTP against a running service; the item-bank pages exist and no journey opens one. The "Playwright" in this item's own title is the part not done.
    • M10.10.a Capture/complete a representative stimulus-linked item with objective/standard/key/rationales/rubric/source and protected fields. Evidence: item journey. OPEN ON THE STANDARD, THE RUBRIC AND THE STIMULUS. The capture carries an OBJECTIVE and a SOURCE; a standard alignment, a rubric and a linked stimulus set are separate surfaces this stage does not reach, and none of it is driven through a page. DONE, OVER HTTP AGAINST A RUNNING SERVICE. tests/test_item_authoring_journey.py, 11 tests: capture a draft, author a complete selected-response revision against that exact draft revision, then read the item back on all four of its surfaces as an author and as a learner. - A capture keeps the objective, the source reference and WHO wrote it, with the tenant derived from the caller. A kind outside the vocabulary is refused by name. - A revision is bound to the draft revision AND its digest; a stale digest is refused, and content naming a source the draft does not have is refused as selected_response_source_reference_drift. The item's sources are the draft's sources, not a second list. - THE PROJECTION IS A LEARNER SURFACE AND SAYS SO: surface: "learner", protected_properties_visible: false, and an excluded_properties list that NAMES the five it removed — response_rule, answer_key, correct_rationale, distractor_rationales, misconceptions. None of them is in the bytes. The protected export names the authority it answered to. - A learner reaches NONE of the four surfaces, including the projection built for a learner: the item bank is an authoring surface and a learner meets items through an assessment. WHICH IS THE POINT OF READING THIS BESIDE M10.10.e. The mechanism to withhold a key — a named learner projection with an explicit excluded list — already existed one module away from the assessment read that handed the key over.
    • M10.10.b Add modality-specific accessibility/locale/PNP content and verify real learner preview creates no attempt/grade/calibration event. Evidence: preview assertions. OPEN ON ADDING THE CONTENT. The accessibility and locale surfaces the authored revision carries come from this suite's own fixture rather than from a caller adding them modality by modality, and PNP (personal needs and preferences) is not touched at all. What IS proven is the second half of the clause, which is the one with a property in it. A PREVIEW IS A LOOK, NOT AN ATTEMPT. tests/test_item_preview_journey.py, 4 tests against a running service with the journey's database read directly — because the thing under test is A WRITE THAT MUST NOT HAPPEN, and the only witness to that is the table. Rendering a preview and scoring a response against EVERY option, including the right one, leaves submissions, grade_ledger_entry and item_calibration_promotion_inputs exactly as they were. The scoring is real while it happens: one option comes back correct: true and another false, so the preview is scoring rather than declining, and it reports stable_option_identity_used, meaning it scored the option's identity and not its position. AND THE COUNTER MOVES FOR A REAL ATTEMPT — a learner submitting through /api/assessments/{id}/submit takes submissions from n to n+1. Without that half, "nothing was written" could just as well have meant "nothing can be written here".
    • M10.10.c Trigger ambiguity/key/accessibility/QTI incompatibility, fix via revisions, and complete independent review/gates. Evidence: remediation journey. THREE OF THE FOUR WALKED; OPEN ON QTI INCOMPATIBILITY, which is M10.5.d's receipt and M10.5.d is open — the same blocker M10.10.d carries. tests/test_item_remediation_journey.py, 10 tests over HTTP against a running service, plus evidence/v1-workbenches/metis-journey/remediation-journey.json. Each class is walked the same way, and the second half is the half that can fail: make the gate say NO, fix it, make the SAME gate say YES. A gate that only ever says NO proves nothing. - Ambiguity: a distractor assessed as also satisfying the stem. The service derives nonkey_correctness_not_rejected from the key partition itself — nothing in the request names a category or a severity — and the gate refuses to publish. - Answer key: the keyed option's own text planted on a learner-visible ui/stem, found across a complete 22-channel, 144-artifact sweep as an exact match. - Accessibility: one automated check failing on the live item; a11y-check:structure-semantics, severity error, gate blocked. A CLEARANCE IS ABOUT ONE REVISION OF ONE ITEM. After the remediation revision the gate computed over the previous text stopped applying by itself: mutation_reasons: ['responses_changed'], and prepare turned review_required on. Nothing had to tell it. FOUND AND FIXED: THE AUTHOR COULD WAVE AWAY THEIR OWN ANSWER-KEY LEAK. The leakage gate answered pass / release_allowed: true over a scan that said leak_detected, carrying in its own raw_findings an exact exposure marked exception_eligible: false. The review counted only confirm as unresolved and never asked WHO decided, so the one instructor who wrote the item and submitted the artifacts could type "false positive" and ship the key. Every layer declares this is not exceptable — the finding, the review content and the gate content all carry a Literal[False] — and the one path around them was relabelling. review_answer_leakage now refuses a resolve or false_positive from the scan's own submitter; confirm, the safe direction, stays open to everybody. WHO THE SECOND PERSON CAN BE, measured rather than assumed: an instructor IS their own item-bank tenant, so a second instructor gets 404 in their own scope and 403 naming the author's — peer review is not a thing this service can express. An admin naming the tenant it acts into is the only shape independence has here, and the journey drives the release through one.
    • M10.10.d Assemble QTI package, inspect per-item diagnostics/loss report, validate externally, reimport, and compare semantics. Evidence: validation report. NOT ATTEMPTED, AND BLOCKED ON M10.5.d. This item's external validation IS M10.5.d's receipt, and M10.5.d is open: there is no canonical hotspot item to validate, so the conformance cell cannot close. The assemble-and-diagnose half is built — M10.5's qti3_* runtime publishes 26 loss codes behind a conformance gate — and the reimport-and-compare half needs the package this stage would assemble.
    • M10.10.e Assert answer-key absence from learner/UI/package-unprotected/ logs/URLs and role/tenant/accessibility behavior. Evidence: E2E suite. A LEARNER COULD READ THE ANSWER KEY. FIXED. GET /api/assessments/{id} and GET /api/assessments returned correct_answer and explanation to EVERY AUTHENTICATED CALLER, and both docstrings said so as though it settled the matter — "so this read requires an authenticated caller". The learner sitting the assessment is authenticated. A student calling either route was handed the answers to the quiz they were about to take. Proven live: a student's GET came back with correct_answer: "Chlorophyll" and the explanation, then the same request after the fix came back with neither. The fix is a projection, not a filter on one field. A caller who is not an instructor gets the questions with correct_answer, explanation AND the evidence bundle removed — the bundle's claims are extracted from the explanation's own text, so redacting the field and leaving the bundle would MOVE the answer rather than withhold it. Measured M9.15's way: a word planted in the explanation and nowhere else, looked for in the RESPONSE BYTES. The correct option's text is one of the options a learner must see, so its presence proves nothing. TWO WRITES ANSWERED 500 FOR EVERY CALLER, found on the way in. POST /api/courses serialises modules and POST /api/assessments serialises questions; both are lazy relationships, and reading one during serialisation happens outside the async session's greenlet. Every create raised MissingGreenlet and returned "an unexpected error occurred". Both load the relationship before the response is built now. Nothing caught it because no test creates either over HTTP. tests/test_answer_key_absence_journey.py, 6 tests against a running service, including the one that makes the others mean something: an instructor STILL sees the key, so a fix that redacted for everybody would fail here. OPEN ON THE REST OF THE CLAUSE: the UI, the unprotected part of a QTI package, logs and URLs are not covered, and neither are the tenant and accessibility behaviours. M10.10.c reached the machinery that WOULD cover them and found the limit. answer-leakage enumerates 22 learner-visible channels — log, url, deep_link, qti package payloads, export, error, trace among them — and a scan of a complete 144-artifact set across all of them returns coverage.complete: true. But A COVERAGE RESULT ATTESTS TO THE SUBMISSION, NOT TO THE SYSTEM: every artifact is supplied by the caller. LearnerVisibleLeakageArtifact has NO producer in this estate — it is defined in the schema, consumed by the helpers, typed in the generated client, and constructed only by test fixtures. The workbench page at apps/metis/web/src/app/create/item-bank/answer-leakage is a textarea a human pastes a capture into, and it fails closed on an empty one, which is the right posture and not a harvest. So the honest reading is: the SCANNER is proven across 22 channels and the gate that judges it is now sound (M10.10.c), and nothing walks the running system's logs, URLs or packages to feed it. That harvest is what this clause still needs.
  • M10.11 Run full-journey Playwright: psychometric review to item revision. Evidence: calibrated data fixture remains isolated. <!-- SPEC 2026-09-11: review pages are apps/metis/admin/src/app/psychometrics and apps/metis/web/src/app/create/item-bank/estimates/* (dif-review, IRT). The synthetic calibrated sample is generated by the M6 learner-simulation service with the isolation suite it ships (learner_simulation); .e proves isolation at the database (rows tagged synthetic never join analytics, grade or training tables), not by UI absence. --> OPEN ON TWO THINGS, WITH .a, .c AND .d DONE. The word in this item's own title is Playwright and nothing here is driven through a page: the review pages named above exist and no journey opens one, which is also what leaves .e's accessibility clause unwalked. And .b is open on parameters, fit and standard errors, because the review contract REFUSES an estimate without uncertainty and nothing in this estate fits a model to responses — the IRT surface attests an external tool's output and no tool is wired to it. What the stage did establish is that the fairness analysis is sound under both questions: it invents nothing on a null sample carrying real sampling noise (.b), and it finds a planted shift on the one item it was planted on (.c). Those are only worth anything together.
    • M10.11.a Seed isolated synthetic calibrated response sample with exact item/model/method versions, privacy-safe cohorts, and known CTT/IRT/DIF/ exposure outputs. Evidence: fixture manifest. THE M6.5 SIMULATION COULD NOT PRODUCE ONE, and that had to be fixed first. Its response rule drew against the learner's prerequisite satisfaction alone, so the delivered item could not move accuracy — while the pool engine beside it bands eligibility by a mastery DIFFICULTY window and offers a nearest_target_difficulty strategy. A simulation whose selection layer targets difficulty and whose response layer cannot feel it measures nothing it varies. Measured on the baseline corpus: the correlation between an item's observed p-value and its declared difficulty was +0.575, the WRONG SIGN, on two to nine responses an item. The model is now Rasch — ability is the logit of the persona's declared satisfaction, difficulty is the item's declared centilogits, and at the average difficulty of zero the probability is the persona's satisfaction exactly, so that field keeps its meaning. response_method and its schema literal are v2. DONE: THE TRUTH CAME FIRST. src/metis/schemas/calibration_fixture.py, src/metis/services/calibration_fixture.py, scripts/export_calibration_fixture.py (pnpm run metis:calibration:fixture, --check compares documents not bytes), tests/test_calibration_fixture.py (13 tests), and evidence/v1-workbenches/metis-calibration-fixture.json. 400 synthetic learners give 4,100 responses over 16 items, generated by the M6.5 engine through the same item-pool eligibility engine a real delivery calls. Item difficulties and persona abilities are DECLARED and the responses are drawn from them, which is what makes the outputs known rather than asserted: the observed proportions correct recover the declared difficulty ordering at a Spearman correlation of −0.9794. - Exact versions: engine id, engine version, response_method, eligibility engine id, the cohort seed, a digest of the whole request, and each item pinned to its revision sha256. - Privacy-safe cohorts: a cohort here is a PERSONA, an authored distribution, so it names no attribute of a person. All 16 items clear the 30-per-cell floor in both cohorts — the same floor dif_analysis will judge the sample by, because a sample built to feed an analysis may not be assembled under a laxer rule than the one that will judge it. The model refuses a released cell below the floor, and refuses a suppressed cell that still carries the counts it suppressed. - CTT: proportion correct and the item-REST correlation, so an item is never correlated with a total containing itself. - Exposure: per-item count and rate in basis points. - DIF: BOTH GROUPS ARE ONE POPULATION WEARING TWO LABELS, drawn from a single declared ability range, so the ground truth is none_injected. A fixture whose groups genuinely differ only tests that a detector finds a planted difference; this one tests the direction that matters, that a detector does not INVENT one. And it publishes the size of its own noise: with no DIF present at all, sampling alone still produced a largest group gap of 0.2036. A detector whose threshold sits under that reports unfairness against an innocent item on every null sample it is ever shown. - Isolation: the M6.5 import-closure and output-identity scan over this exact run — isolated, 0 unlabelled records, 0 real-identity matches — and the manifest refuses to validate if it is not.
    • M10.11.b Review raw parameters/fit/SE/sample/uncertainty/DIF/privacy/ declared difficulty without exposing learner records. Evidence: review journey. DIF, PRIVACY AND AUTHORITY WALKED; OPEN ON PARAMETERS, FIT AND SE, and the reason is the honest one: the review contract REFUSES what an unfitted sample can offer. Estimate will not accept a difficulty, a discrimination, a guessing parameter or a point-biserial as a bare number — "a point estimate on its own reads as precise and a reviewer compares two spurious digits" — and the standard error it wants means fitting a model to responses. NOTHING IN THIS ESTATE FITS ONE. The IRT surface is an attestation of an external tool's output and no tool is wired to it, so M10.11.a's sample can put a proportion correct, a sample size and an exposure rate in front of a reviewer, and must report difficulty and discrimination as unavailable with the reason. tests/test_psychometric_review_journey.py, 14 tests. THE MULTIPLICITY CORRECTION IS LOAD-BEARING, MEASURED. DIF is the one analysis here that really computes — weighted logistic regression, Newton-Raphson, inverse observed information, two-sided Wald, Holm — so M10.11.a's sample was run through it for real. In a sample with NO differential item functioning in it, 11 of 16 items produced a uniform effect at or beyond the policy's materiality threshold of 0.5, the largest reaching 1.45, and the largest |z| was 2.05 — whose UNADJUSTED p-value clears alpha. Holm over the family of 32 hypotheses takes the smallest adjusted p to 1.0 and all 16 come back no_statistical_signal, direction: none, no human review required. Without the correction this run publishes a fairness finding against an innocent item. THE EXISTING DIF SUITE COULD NOT HAVE SHOWN THAT. Its null fixture gives reference and focal the identical correct-count in every stratum — (4, 7, 10, 13, 16) against (4, 7, 10, 13, 16) — a null with no sampling variation in it at all. Nothing in it ever asks what the correction is for. M10.11.a's sample wobbles as much as chance makes it wobble, which is why it can. Privacy: no result carries a learner reference or a per-response row; every released cohort sample is at or above the floor it declares; and M9.6's complementary suppression holds — one suppressed cell beside a published total gives the cell back, so a second is withheld and recoverable_by_subtraction comes back empty. Authority: promoting is a psychometrician's alone, restricting and retiring need a psychometrician AND a fairness reviewer, and an analyst may only ask for a re-run or record what they saw. A statistical signal is not a cause and never an individual: every result carries bias_cause_established: false and individual_inference_supported: false.
    • M10.11.c Trigger insufficient/misfit/DIF or exposure blocker and record psychometric/fairness decision. Evidence: gate journey. DONE, WITH A MATCHED PAIR. tests/dif_control_fixtures.py and tests/test_dif_gate_journey.py, 10 tests over the real analysis. Both samples come from ONE generator, one seed and one set of declared item difficulties; the only difference is a shift of 120 centilogits planted on one item for the focal group. So a detector that answers differently is answering to the plant, and one that answers the same is not detecting anything. M10.11.b asked the negative question — does it invent a finding where there is none — and this asks the positive one. A gate that never fires and a gate that always fires each pass half of these tests and neither is a gate. - DIF BLOCKER: exactly one item flagged, the planted one. Uniform effect −1.065, Holm-adjusted p 0.0019, material_signal, human_review_required: true, and direction: reference_favored — the right way round, since the focal group met that item harder. The same item in the null answers no_statistical_signal, and the other seven are unmoved in both. - INSUFFICIENT BLOCKER at 40 learners a group against the locale_family floor of 50: status: insufficient, finding: inconclusive, effect and significance both suppressed, and the sample, model_fit and multiplicity checks fail. AN INSUFFICIENT RUN HIDES ITS OWN CELL SIZES — both cohort samples come back suppressed with source_n: null, because reporting "too few to analyse: 40 and 40" would publish the very counts the floor exists to protect. - THE DECISION NEEDS TWO DISCIPLINES AND TWO PEOPLE. Recording one requires a psychometrician AND a fairness reviewer, two distinct reviewer ids, both recommending the action actually recorded. One person signing twice is refused, one discipline signing twice is refused, and a second signature on an action only the first recommended is refused. No decision may set establishes_bias_cause or individual_inference_supported: a statistical signal says an item behaved differently, not why, and never anything about one person.
    • M10.11.d Create a new item revision/remediation, retain old calibration/ use/decision, and set new revision uncalibrated until evidence exists. Evidence: lifecycle assertions. DONE. tests/test_calibration_lifecycle_journey.py, 7 tests. A CALIBRATION BELONGS TO A REVISION, NOT TO AN ITEM. Three things have to hold at once and they pull in different directions: the new revision is not calibrated because nobody measured it, the old calibration is not deleted because it was true of the item it was about, and the old evidence does not quietly carry over. - The revision that has its own evidence is eligible_for_ratification with every check passing — the half that can fail, so the refusals below mean something. - Present the same 640 well-formed attempts against a revision they were not gathered on and exact_revisions reads stale, the status drops to manual_review, and the route is never calibration_ratification. It reads stale rather than fail because nothing is wrong with the evidence — it is about something else. - THE OLD ATTEMPTS STOP COUNTING RATHER THAN MERELY BEING FLAGGED. effective_sample, response_distribution and administration_diversity all go from pass to insufficient. This is the distinction that decides whether the guard holds: a system that flagged the mismatch and still counted 640 attempts would report a well-powered calibration with a warning attached, and a warning beside a green number is read as a green number. - SUPERSEDING IS NOT ERASING. The old receipt still names the revision it was measured on, its sample and its attempt count, and it is sealed over its own content — repointing it at the new revision breaks the seal detectably. It also never claims promotion_applied, and it expires, so an old calibration cannot be leaned on forever. The failure this guards against is one of omission: keeping a calibration keyed to the item rather than the item REVISION means a one-word fix to a stem silently inherits a difficulty estimate measured on different words, and nothing anywhere reports an error.
    • M10.11.e Prove synthetic fixture cannot enter real analytics/grade/ training data and verify role/tenant/accessibility boundaries. Evidence: isolation E2E. ISOLATION, ROLE AND TENANT PROVEN; OPEN ON ACCESSIBILITY, which needs a browser against the psychometrics pages and is the same unwalked UI half M10.10's parent carries. tests/test_calibration_isolation_journey.py, 10 tests. THE STRUCTURAL PROOF IS THE STRONGER ONE. The transitive import closure of everything this stage runs — the M6.5 simulation, the fixture generator and the DIF analysis — contains no database, ORM, learner store, gradebook, analytics or export module. That says the sample CANNOT reach a learner record, which is a better claim than "it did not this time"; counting rows can only ever show the second. The scan is proved able to fail by walking outward from aggregate_store, which is built on SQLAlchemy and is duly caught. NO SANDBOX IS CLAIMED. The closure does pull in socket and TLS transitively through the standard library, and the test asserts that it does, so the claim cannot quietly grow into one about confinement. Dynamically: the sample's isolation report is isolated over a non-empty scan with every record synthetic-labelled, 0 unlabelled and 0 real-identity matches; no learner reference survives into a DIF result; and generating the fixture and running both analyses moves no row in any of the nine tables a person's work lands in — submissions, the four grade-ledger tables, the two gradebook-correction tables, training intake and calibration promotion inputs. FOUND AND FIXED: A STUDENT REACHED AN /api/admin/ ROUTE. psychometric_review took plain get_current_user while all four sibling psychometric surfaces — dif_analysis, irt_diagnostics, calibration_promotion, exposure_control — already required instructor or admin. AUTHORITY_FOR_USER maps "student" to analyst, an authority that may request a re-run and record a review. Measured over HTTP: a student got 200 and authority: analyst from /api/admin/psychometrics/queue. Nothing had leaked, because the tenant is derived from the caller and a student's queue is their own and empty — but the exposure was one scoping parameter away and the route is mounted under /api/admin/. It requires instructor authority now and the same request answers 403, measured against the running service. OPEN, AND THE SAME SHAPE AS M10.9.d: NOBODY CAN BE A FAIRNESS REVIEWER. restrict and retire are the two dual-control actions and they are permitted to a psychometrician and a fairness reviewer. fairness_reviewer is reached only by a user whose platform role is "reviewer", and this service issues student, instructor and admin. So the discipline that exists to check a psychometrician cannot be occupied by anybody, and the second signature on a high-impact fairness action is unobtainable. That is a gap in the role vocabulary rather than in the mapping, and closing it needs a platform role this service does not have.
  • M10.12 Run full-journey Playwright: integrity decision, appeal, and gradebook correction. Evidence: role separation. OPEN. .a and .b are done; .c has no surface to walk, .d has no store, and .e's answer is a defect rather than an assertion. Nothing here is driven through a page, which is this item's own title again. <!-- SPEC 2026-09-11: pages are apps/metis/admin/src/app/integrity-signals, .../governance (decisions and appeals) and .../grade-corrections. Role separation is by actor id through the M7 conflict-of-interest and signoff records (see the M10 heading note); the JWT admin role alone must NOT be enough to decide an appeal one has reviewed, and .e asserts exactly that. -->
    • M10.12.a Seed a synthetic learner attempt, attributed uncertain integrity signal/evidence, separate reviewer/decider/appeal/grade roles, and policy. Evidence: fixture. DONE over HTTP against a running service. tests/test_integrity_decision_journey.py records a signal carrying its model id, its model version, its collector version, a confidence INTERVAL with a stated basis and sample size, and the raw classifier output beside the threshold it was judged against — so a reviewer can see how close to the line it fell rather than only which side. A SIGNAL IS AN INDICATION AND THE TYPE WILL NOT LET IT BE A FINDING. The projection carries is_conclusion: false and language: classifier_indicates as literals, so no payload, however constructed, can present a detection as a conclusion. The evidence is a chain in which each link names the digest of the one before it, so nothing can be inserted into a settled chain without breaking every link after it. The four roles are separate in the engines rather than in one fixture: the reviewer and integrity-lead audiences come from the caller's role, the decider independence is by actor digest in the appeal engine, and the grade roles are the scorer, requester, reviewer and approver of the correction engine.
    • M10.12.b Review signal/evidence/model/version/uncertainty, request evidence, and record initial human decision/explanation. Evidence: decision journey. DONE. tests/test_integrity_decision_journey.py, 11 tests over HTTP. The queue shows the signal, a view names the audience that saw it and leaves an access-audit trace with a digest and a time, asking for evidence moves the signal new to evidence_requested and records who asked, the history reads back, and an action naming a version that has moved is refused stale_expected_version. NO REVIEW VERB CAN CARRY A GRADE OR A CONSEQUENCE, and not because a check forbids it: the four verbs are request_evidence, dismiss, escalate and open_case, and ReviewAction has no field for one. FOUND: THE ESCALATION THIS AUDIENCE MODEL EXISTS FOR CANNOT HAPPEN. escalate is one of the four verbs and there is nobody to escalate to. _authority sets tenant_id to the caller's own user id for every audience, and this router — unlike the item-bank surfaces, which take ?tenant_id= for exactly this reason — offers no parameter to act into another tenant. Measured: an integrity lead gets 404 on both the history and the view of a signal a reviewer recorded. AND THE ACCOUNTABILITY RECORD IS THE CASUALTY. The access audit exists so somebody can see which reviewer opened which signal. A reviewer is correctly refused it, 403. The integrity lead who may read it gets 200 and an empty list, because it is scoped to their own tenant and the reviewer's views sit unreadable in another. The views it exists to record are the exact views nobody can read.
    • M10.12.c File learner appeal, enforce independent restricted reviewer, add evidence, and overturn/modify/remand with immutable history. Evidence: appeal journey. NOT WALKABLE: THE APPEAL WORKFLOW HAS NO SURFACE AND NO STORE. apps/metis/service/src/metis/api/appeal_workflow.py is 27 lines and mounts exactly ONE route, GET /api/appeal-workflows/evaluation, which returns the engine's own labelled-corpus report. There is no route to file an appeal, add evidence, decide, overturn, modify or remand, and no table holds an appeal case — the only appeal table in the schema is item_exposure_appeals, which belongs to M6 exposure control and is a different thing. The engine itself is real and well made: appeal_workflow_engine is a state machine whose decide refuses a decider who owned the detector or reviewed the signal (not_independent), who declared a conflict with this learner, who names a stale version, or who acts before the evidence and the learner's explanation are settled — all BY ACTOR DIGEST rather than by JWT role, which is what this stage's spec asks for. It is imported by its own evaluation corpus and by integrity_audit_engine, and by nothing that a person can reach. So the independence rule this item exists to walk is implemented and unreachable. Closing this needs the routes and the persistence, not a test.
    • M10.12.d Request/approve append-only grade correction and reconcile aggregates/integration/learner visibility. Evidence: correction journey. APPEND-ONLY AND LEARNER VISIBILITY WALKED; OPEN ON RECONCILIATION, because there is nothing durable to reconcile. tests/test_grade_correction_journey.py, 7 tests over HTTP. A grade is a sequence of events and an approval appends one — the disputed event stays in the history beside the value that replaced it, and the sequence numbers stay ordered and unique. A correction cannot be approved without a notice to the learner (learner_not_noticed), and a decision naming a version that has moved is refused. THE TABLES HAVE A READER AND A WRITER AS OF 2026-09-13. They did not before: POST /decisions read the request, the history and the prior decisions out of the request body, computed, and returned, loading nothing and saving nothing. Migration 044 created gradebook_correction_grade_events and gradebook_correction_correction_decisions and NOTHING IN THIS REPOSITORY READ OR WROTE EITHER — no ORM model, no repository, no statement naming them, checked repo-wide — and a test asserted that absence so the day a writer appeared it would say so. That test has now flipped to test_the_store_the_server_consults_is_real, and it asserts the property that was actually missing: not that a row exists, but that a decision recorded by one HTTP call CONSTRAINS THE NEXT ONE. models/gradebook_correction_store.py, services/gradebook_correction_ store.py, and migration 076, which indexes the content_json::jsonb ->> 'request_id' expression the lookup filters on so the separation check does not become a scan of every decision a tenant ever made. Measured against a live service: 20 decisions and 4 grade events durable in the two tables after the journey suite ran twice, and the suite passes on the second run because the rules read state rather than reset with the process. A DEFECT IN THE ABOVE, FOUND AND FIXED THE SAME DAY. The first version of the route resolved its tenant with resolve_item_bank_tenant, which returns the actor's RAW UUID, while the grade ledger these corrections correct is written by api/gradebook_ingress.py under uuid5(NAMESPACE_URL, "metis-tenant:" + user.id) — the derivation deps.tenant_of documents as "derived and never accepted". For the SAME signed-in instructor those are two different UUIDs, so a correction and the grade it corrects were landing in different partitions. 23 API modules use the uuid5 derivation; this route was the only one using the item bank's, which is that family's convention for its own services. Nothing misbehaved — no code joins the two stores yet, and no exploit is claimed — but a reconciliation across two partitions is impossible BY CONSTRUCTION rather than merely unbuilt, which is the thing this very clause asks for. Now tenant_of(current_user.id); instructor authority was already required by the route's dependency, so the resolver's role check was redundant. Re-measured against a live service on a fresh database: the 9 journey tests pass over HTTP, all 9 decisions and 2 grade events land under tenant_of(the instructor) — the ledger's partition — and the same caller approving their own request is still REFUSED ['reviewer_cannot_approve']. RECONCILIATION IS STILL NOT WALKED. Two corrections to one grade now share state, which is what this clause said was impossible, so the blocker as written is gone. But POST /aggregates still recomputes over inputs the caller hands it, so nothing yet reconciles an aggregate ACROSS the corrections now on record. That is what remains here, and it is a smaller thing than what it replaced.
    • M10.12.e Assert original reviewer/system actor cannot decide appeal/ grade, protected evidence redaction, tenant isolation, and accessibility. Evidence: separation E2E. ASSERTED. THE GRADE HALF NOW HOLDS; THE APPEAL HALF CANNOT BE ASKED. A SEPARATION OF DUTIES ENFORCED AGAINST HISTORY THE ACTOR SUPPLIES IS NOT A SEPARATION OF DUTIES. The three grade rules are right and each has its own refusal — the scorer cannot review, the requester cannot decide, the reviewer who moved a request on cannot approve it — and all three fire when the server is told. But only ONE of their four inputs came from the session. actor_digest is derived from the token and no body field can move it, which is the half that was always right. The grade history, the correction request and the prior decisions were all read out of the REQUEST BODY, and prior_decisions defaulted to an empty tuple. MEASURED OVER HTTP, ONE CALLER, ONE REQUEST, BEFORE: they ask for changes and the service records their digest as the reviewer; they then APPROVE THE SAME REQUEST and it SUCCEEDS, because the default value of a field is to mention nothing. Only the identical call that disclosed the first decision was refused. The rule was asking the person it is meant to constrain. FIXED 2026-09-13, AND THE FIX WAS PERSISTENCE RATHER THAN A CHECK. The same walk against the same service now answers reviewer_cannot_approve on the second call, with NOTHING in that call mentioning the first: POST /decisions loads the prior decisions from the store and prior_decisions is gone from the published request shape — refused, not ignored, because a silently dropped field would leave a caller believing they had disclosed something. The store is .d's, and the tenant it is scoped by comes from the session through resolve_item_bank_tenant, never from a payload; scoping it by a body field would have put the hole back one level down. WHAT IS STILL BODY-SUPPLIED, said plainly: the grade history and the correction request. Those are two of the four inputs and they are unchanged, so this clause's finding is half closed, not closed. The appeal half of this clause cannot be asserted at all — see .c, where the independence rule exists and no route reaches it. Redaction and tenant isolation hold on the signal surface: a view names the fields it showed and withheld, a reviewer is refused the access audit (403), and a learner is placed in the learner audience and sees an empty queue. Accessibility stays open with every other UI clause in §M10.
  • M10.13 Run WCAG 2.2 AA automation, keyboard journeys, and screen-reader announcement tests across authoring and preview. Evidence: report matrix. OPEN, with .b done and the rest measured. THE THREE DEFECTS THIS STAGE FOUND WERE ALL INVISIBLE TO THE GATE THAT EXISTED — two because the rule was switched off, and one because no scan can see when a live region is mounted. That is the report matrix's real finding: the coverage was wide (every route visited) and shallow (one state, one rule set, no traversal). A real assistive-technology run belongs to S11.5.f and is not claimable here; this stage's own spec note says so, and .d is DOM assertions by design. <!-- SPEC 2026-09-11: §2.5.8 — 2.2 tags through the kit policy, not axeAfterEach's 2.1 set. The screen-reader cell (.d) is DOM/live-region assertions here; a real assistive-technology run is S11.5.f's matrix and is not claimable on this Mac, so the report matrix names it as not run. -->
    • M10.13.a Maintain critical route/state/role/action inventory for hub, course/item/source/graph/review/calendar/localization/media/credential/ integrations and learner preview. Evidence: coverage matrix. ROUTE COVERAGE IS ESSENTIALLY COMPLETE AND IS NOT THE GAP. Measured across both applications: all 65 admin route groups are visited by at least one spec, and 10 of the 11 web route groups are — only saved is never opened. 71 admin specs and 66 web specs. THE GAP IS STATES AND ACTIONS, NOT ROUTES, and M10.13.b/.d proved it twice over. Both contrast defects lived in an EMPTY STATE, so every spec that visited the route with data in it passed; and the 41 unannounced live regions are a mounting pattern no route visit can see. A route-level matrix would have shown green for all three. So the inventory this item asks for is the wrong shape to catch what was actually wrong here. What the stage needed, and what .b and .d now have, is a matrix over STATES — empty, loading, error, populated — and over the mount-time behaviour of announcements. Recording that rather than producing a route table that would have said everything was fine.
    • M10.13.b Run automated rule scans across themes/viewports/locales/zoom/ reduced motion after key interactions and error states. Evidence: scan reports. THE SIGNOFF WAS GREEN BECAUSE THE RULE WAS OFF. FIXED. wcag-aa-signoff-v1-p2-3513.spec.ts is named "Metis admin WCAG 2.1 AA signoff" and ran with disableRules: ['color-contrast'], deferring to a design-token migration named with no owner and no date. 1.4.3 Contrast (Minimum) is AA and is the rule a real interface is likeliest to break, so the signoff was silent about exactly the thing it was named for. The carve-out is REMOVED and the spec passes with the rule on. TWO REAL DEFECTS BEHIND IT, both in shared components and both in EMPTY STATES, which is why every scan of a populated page passed: DataTable's "no rows" message was #9ca3af on white, 2.53:1, and ActivityFeed's was #9ca3af on #f9fafb, 2.42:1. Both are #64748b now, 4.76:1 and 4.63:1. The empty state is the only message a reader has when there is nothing else on the card, so it is the worst place to be faint. The failures looked intermittent until the cause was found: they appeared whenever a fetch failed and left a table empty. WCAG 2.2 RAN FOR THE FIRST TIME. Both Metis applications take their gate from axeAfterEach, which declares the 2.1 terms, and the one hand-rolled helper in apps/metis/web declares the same four. So nothing in Metis had ever asked for the 2.2 terms. Measured against the axe-core installed here (4.11.4), they add EXACTLY ONE RULE: 68 rules to 69, and the addition is target-size — 2.5.8 Target Size (Minimum). It also ships enabled: false in axe's metadata, so naming the tag is what turns it on; scanning with no tags at all leaves it off. Result: NO target-size violations across eight authoring surfaces at desktop width, nor on the hub at 390px, which is where pointer targets are smallest. That is a fact nobody had established. e2e/support/wcag22.ts, e2e/wcag22-target-size.spec.ts (10 tests) and e2e/contrast-carve-out-census.spec.ts (13 surfaces, a ceiling of zero that can only be ratcheted down). 52 of 52 green over two repeats, chromium. OPEN: themes, locales and reduced motion are not varied here — the scans run one theme, one locale and default motion.
    • M10.13.c Run keyboard-only focus/order/visibility/shortcut/non-drag/ editor/graph/calendar/media/modal/save/conflict journeys. Evidence: keyboard reports. TRAVERSAL WALKED ON THE LEARNER APP; the widget-by-widget list above is not, and the counts say why that matters. MEASURED FIRST: across the 65 specs in apps/metis/web, twelve press a key and NONE PRESS TAB. The keys they press are arrows, Enter, Home and End — widget keys, pressed inside a component that already holds focus. Nothing asked the question a keyboard user asks first, which is whether Tab reaches the controls at all and in what order. The admin app is thin rather than empty: 4 of 69 specs press Tab, and the one named "every control is reachable and operable from the keyboard" presses it once, between two adjacent filters. apps/metis/web/e2e/keyboard-traversal.spec.ts, 5 tests, 15 of 15 green over three repeats on chromium. The assertions are properties rather than one page's layout, so they keep meaning when it changes: every visible control is reachable by Tab alone; the order Tab visits them in is document order, with no positive tabindex lifting a control out of it; focus is visible on whatever holds it; nothing traps the keyboard; and the sign-in form can be completed and submitted without a pointer at all. The login page passes all five. Two things had to be right in the harness before that meant anything: Next's dev overlay is focusable and Playwright's locators pierce its shadow root, so it had to be stepped over on both the walk and the enumeration or the test measures the harness; and Tab WRAPS to the top after the last control, so an order check spanning the wrap reads a correct cycle as a jump backwards. OPEN: the editor, graph, calendar, media, modal, save and conflict journeys the item names, and shortcuts and drag alternatives. This walks one page thoroughly rather than eleven surfaces shallowly.
    • M10.13.d Run screen-reader name/role/value/instruction/status/live update/ validation/selection/decision and semantic alternative journeys. Evidence: announcement reports. FOUND: 41 LIVE REGIONS THAT ARRIVE WITH THEIR MESSAGE AND ARE NEVER ANNOUNCED. aria-live="polite" and role="status" tell assistive technology to watch an element and read what CHANGES inside it. NVDA and JAWS attach that watch when the region enters the accessibility tree, so a region inserted already containing its text has nothing to change — the text was there when the watch began. The shape across apps/metis/web is the silent one: {feedback ? <p role="status">{feedback.message}</p> : null} and the shape that works is the region rendered from the start, empty, with the message written into it. The consequence is not subtle: a learner submits a form, a sighted user sees the confirmation, and a screen-reader user is told nothing. 41 of these across 37 files. role="alert" is excluded and is NOT a defect — an alert is defined to be announced when it appears, so inserting one with its text is correct; there are 13 of those and they are left alone. Getting that distinction wrong would have condemned a correct pattern. src/lib/live-region-insertion.spec.ts is a ceiling rather than a zero, because each site has to be moved by hand with its styling. Three of the most learner-facing are fixed — the standards alignment, objectives and decisions pages — and the ceiling stands at 38 and may only go down. The gate also proves it can fail and refuses to count an alert. WHAT WAS ALREADY GOOD: 24 of the 65 web specs assert on live regions and 62 product files declare one, so the vocabulary is used widely and the defect is in HOW the region is mounted, not in whether anybody thought about announcements. OPEN: name/role/value and the validation, selection and decision journeys are not walked, and a real assistive-technology run is S11.5.f's matrix rather than this one — this stage's own spec note says so, and it is not claimable here.
    • M10.13.e Verify real learner preview parity and PNP/MathML/code/table/ media/interaction alternatives. Evidence: content accessibility report. NOT WALKED AS A JOURNEY, but two precise findings came out of reading the delivery renderer, and both are about the same image. FOUND: AN IMAGE NOBODY DESCRIBED IS PRESENTED AS DECORATION. LessonDeliveryRenderer renders <img alt={media.alt_text ?? ''}>. An empty alt is not "no description" — it is an assertion that the image is decorative and assistive technology should skip it. The model cannot tell the two apart: alt_text is nullable and there is NO decorative flag anywhere in the web app or the service schemas, so a null can only mean nobody wrote one. A learner using a screen reader is silently given nothing, and is not told that anything is missing — while the same component renders a visible "Reason: …" when media is withheld, so the codebase plainly knows how to say a thing is absent. AND THE FALLBACK THAT WOULD FIX IT EXISTS AND IS UNREAD. image_fallback_alt is declared on lesson_technical.py, published in the OpenAPI document and typed in the generated client — and no component in apps/metis/web reads it. Its only other appearance in the repository is one test fixture, set to null. MediaBlock is also handed a fallbackText it does not use for the image either. No test covers the null case, which is why this survived: the one renderer spec sets alt_text to a real sentence. What IS present and good: role="math" carries aria-label from speech_text, code blocks are labelled by language, and audio and video carry titles — so the alternatives vocabulary is real and it is the image path that drops it. OPEN: preview parity between the learner view and the authoring preview, PNP, and table and interaction alternatives are unwalked, and only 3 files in the web app mention PNP at all.
    • M10.13.f Triage owners/severity and block release on current unresolved applicable violations. Evidence: zero-blocker report matrix. RELEASE IS BLOCKED ON THEM NOW, WHICH IT WAS NOT BEFORE. The gate that was supposed to do this ran with color-contrast disabled, so the one rule most likely to be violated could not block anything. That carve-out is removed (.b) and three gates stand in its place, each of which fails rather than warns: - wcag-aa-signoff-v1-p2-3513.spec.ts, 13 surfaces, now with contrast on. - contrast-carve-out-census.spec.ts, the same 13 surfaces against a contrast ceiling of ZERO that may only be ratcheted down. - wcag22-target-size.spec.ts, 8 surfaces plus a 390px viewport under the WCAG 2.2 terms. - live-region-insertion.spec.ts, a ceiling of 38 on regions that arrive with their message, down from the 41 measured. CURRENT UNRESOLVED APPLICABLE VIOLATIONS, all found this stage: - contrast on two shared empty states — FIXED, both components. - 38 live regions that arrive with their message — OPEN, held at a ceiling; each needs moving by hand with its styling. - a null alt_text rendered as alt="", and image_fallback_alt published by the API and read by nothing — OPEN (.e). - keyboard traversal on ten of the eleven surfaces the item names — OPEN (.c). NO OWNER FIELD EXISTS TO TRIAGE INTO, which is the honest answer to the first half of this item. The estate has no accessibility owner registry and skipAxe — the control this would be recorded against — is a bare boolean with no owner, reason or expiry, 247 uses across 241 files estate-wide. Metis is the good citizen there: 1 of its 137 specs uses it, and the one new use in wcag22-target-size.spec.ts replaces the scan with a WIDER one rather than giving anything up. A severity ordering is available and used — everything above gates on axe's serious and critical — but owners and expiries are not expressible, so the triage half of this item stays open.
  • M10.14 Run performance budgets for large courses, item banks, graphs, search, save, preview, and export. Evidence: trend reports. OPEN ON A CHAIN THAT STARTS OUTSIDE THIS REPOSITORY. Production volume is unrecorded, so .b cannot build a representative corpus; .c and .d profile against it; .e needs both a baseline and a ratified budget, and the budget type deliberately refuses to exist without a ratifier. Each link is measured rather than assumed, and .a still found and gated a real defect on the way: a script budget that cannot bind. <!-- SPEC 2026-09-11: §2.5.7. Budgets in .a are ratified numbers written before measuring; fixtures in .b go into an isolated database, never the dev one; .c profiles one scenario at a time with the memory checkpoint between scenarios. -->
    • M10.14.a Ratify scenario/device/network budgets for load/interaction/ render/save/search/export latency, payload/query, CPU/memory/network, throughput, and error rate. Evidence: budget spec. BLOCKED ON AN AUTHORITY, AND THE TYPE IS RIGHT TO BLOCK IT. LatencyBudgetSpec requires ratified_by and ratified_at, so a budget cannot exist in this estate without naming who ratified it. That is the correct design and it means this item cannot be closed by writing numbers: I can propose them and cannot ratify them. Recording rather than inventing a ratifier. WHAT METIS ALREADY HAS, measured: - A good budget CONTRACT. LatencyBudgetSpec carries p50/p95/p99 per tier and per stage, pairs device WITH network because "a p95 for a phone on 3G is not a p95", refuses a higher percentile with a tighter budget, and refuses a stage that times out before its own budget is spent. - Exactly ONE ratified instance: budget:adaptive-next-item v1.0.0, ratified by the Performance Working Group on 2026-08-01 — and it lives in adaptive_latency_evaluation.py, an evaluation corpus, so it is a budget the engine is scored against rather than one the running system is held to. - Client budgets on 5 routes in lighthouse-budget.json: paint and interaction timings, resource sizes and counts. NONE OF THE SEVEN SCENARIOS THIS ITEM NAMES HAS A BUDGET — large courses, item banks, graphs, search, save, preview and export. Nor does any Metis artefact budget query count, memory, CPU, throughput, concurrency or error rate; the estate's own I9.14.a recorded 6 of 12 dimensions budgeted anywhere and Metis is not in that count. FOUND: A BUDGET THAT CANNOT BIND IS NOT A BUDGET. On all five budgeted routes the script resource budget EXACTLY EQUALS the total, so it is satisfied by every page the total already admits and can never be the constraint that fails. Scripts are the dominant payload in a Next application, so the one resource type most worth constraining was the inert one. image and font sit below their totals and do bind — checking that stopped this being overstated. src/lib/performance-budget-coherence.spec.ts, 5 tests: every routed path is budgeted, every route states a total, paint metrics are ordered, and the inert count is held at a ceiling of 5 that may only go down. It is a ceiling rather than a zero because correcting it means choosing numbers, and this item's own rule is that budgets are ratified BEFORE being measured against — a number picked to fit the bundle that exists is a description.
    • M10.14.b Create production-scale course/item/stimulus/media/source/graph/ locale/revision/review fixtures with documented distributions. Evidence: fixture manifest. BLOCKED AT ITS ROOT: NOBODY KNOWS WHAT PRODUCTION SCALE IS. The estate's own persistence inventory records it — evidence/v1-workbenches/metis-persistence-inventory.json carries observedProductionBytes: null, observedProductionRows: null and productionVolumeEvidence: "not present in the repository; capacity baseline required before migration". A fixture built without that is not production-scale, it is a fixture of a size somebody chose, and profiling it would produce numbers that look like measurements and answer no question. The distributions this item asks to be documented are exactly the thing that is missing. The machinery to build one exists and is proven: M10.11.a generates 400 synthetic learners and 4,100 responses into an isolated database with a manifest, which is the shape this item wants. It is the target SIZE that is unknown, not the method.
    • M10.14.c Profile course/item workspaces, graph/list, search/facets, autosave/diff/conflict, learner preview/media, and export/diagnostics independently. Evidence: scenario reports. BLOCKED ON .b. Profiling a fixture of an arbitrary size measures the fixture. The per-scenario discipline this item asks for — one at a time, with a memory checkpoint between — is the right method and is already the rule this repository works under; what it has nothing to run against is a corpus anybody can defend as representative.
    • M10.14.d Capture client/server traces, query plans, bundle/network/ hydration/virtualization costs and memory leaks over sustained use. Evidence: profiles. BLOCKED ON .b for the query plans and the sustained-use leak hunt, both of which are statements about behaviour AT a volume. The bundle and network half is separable and partly budgeted already — the five-route lighthouse budget in .a covers resource sizes and counts — but .a also found that its script budget cannot bind, so the bundle number it reports is not currently constrained by anything.
    • M10.14.e Compare trends to baselines/budgets, assign regressions, rerun, and retain current passing evidence. Evidence: performance gate. BLOCKED ON .a AND .c: a trend needs a baseline to move from and a budget to be judged against, and neither exists for the seven scenarios this stage names. Assigning a regression additionally needs an owner field, and the same gap M10.13.f recorded applies here — the estate has no performance owner registry either.
  • M10.15 Run load tests for imports, gate jobs, adaptive selection, LMS sync, grade events, and analytics. Evidence: backpressure and SLO results. OPEN, AND .f ANSWERED ITS OWN QUESTION WITHOUT A LOAD RUN: there is no backpressure in this service to verify. The rest waits on .a, which waits on a production baseline that does not exist, exactly as M10.14 does. Running load anyway would produce numbers about a corpus nobody chose against limits nobody set. <!-- SPEC 2026-09-11: runs on the Linux dev server (free -m checkpoint, stop threshold written into .a's plan); a load run attempted on this Mac is recorded as a failed attempt, never as evidence (§2.5.7). -->
    • M10.15.a Define realistic arrival rates/concurrency/payload distributions/ tenant mixes/dependencies/failures/SLOs/resource stop thresholds per workload. Evidence: load plans. OPEN. "Realistic" is the word that blocks it, and for the same reason M10.14.b is blocked: the estate records observedProductionRows: null and "capacity baseline required", so an arrival rate written here would be a number somebody chose rather than a rate anybody observed. A target capacity could be DECLARED instead — that is a design decision rather than an observation and would be honest — but it needs the same ratifying authority M10.14.a is blocked on. The resource stop threshold is the one part that is already real and is written down outside this checklist: this repository's own rule is a free -m checkpoint before every heavy invocation and a teardown below roughly 3 GB available on this box.
    • M10.15.b Load source/QTI/standards imports and extraction/media/ quality/psychometric gate queues with bounded parallelism. Evidence: import/gate report. BLOCKED ON .a's plan and on .f: loading a queue that has no backpressure measures how fast it can be overwhelmed.
    • M10.15.c Load adaptive next-item selection/tutor context at learner concurrency with cold/hot caches and degraded models/stores. Evidence: adaptive report. BLOCKED ON .a. This is the one workload with a ratified latency budget — budget:adaptive-next-item, p50/p95/p99 per stage — so it is the first that becomes runnable once an arrival rate exists.
    • M10.15.d Load LMS/roster/grade/webhook sync, duplicate/out-of-order/ corrected events, replay/reconciliation, and provider limits. Evidence: integration report. BLOCKED ON .a AND ON THE ABSENT DESTINATION M10.9.d recorded: the saga's Destination protocol has one implementation in this estate and it is an object store, so there is no LMS to sync against.
    • M10.15.e Load privacy-safe analytics ingestion/aggregation/query with high-volume attempts/events and suppression. Evidence: analytics report. BLOCKED ON .a. The suppression half is separately proven: M10.11.e and M10.12 show cohort cells withheld below the floor and complementary suppression enforced, but neither was measured under volume.
    • M10.15.f Verify backpressure/fairness/retry/circuit/queue age/lag/ resource/cost/tail SLOs and stop before unsafe host pressure. Evidence: resource-safe SLO results. THERE IS NO BACKPRESSURE TO VERIFY. This service has three pieces of a rate limiter and no rate limiter: - config.py declares rate_limit_requests_per_minute = 60 and rate_limit_burst = 10; - deps.py implements check_rate_limit, a correct sliding-window counter over Redis that sets a key with a TTL, refuses past the maximum and increments otherwise; - main.py maps 429 to rate_limited and "Too many requests." And nothing joins them up. check_rate_limit has ZERO callers in src/ or tests/; the two configured numbers are read NOWHERE outside their own declaration; and no route or middleware can produce the 429 the error mapper knows how to render. The only middleware installed is CORS and a request timer. Somebody reading the configuration would reasonably conclude callers are limited to sixty requests a minute. They are not limited at all, and a load run would discover that the service accepts everything until it falls over — which measures the absence of a control rather than the presence of a limit. tests/test_backpressure_absence.py, 6 tests, asserts the three pieces exist AND that nothing connects them, so the day it is wired the suite says so and the tests flip into assertions about the limit. Better than a comment, which is what an unwired control usually gets. THE DAY CAME (M10.8.d, 2026-09-13). The pieces are joined: services/request_throttle.py with a middleware in main.py, and the file is tests/test_backpressure.py now — 27 tests, the last of which drives the real application against the real Redis and watches the fourth request of a three-request window come back 429. Every assertion above flipped as it was written to. Read M10.8.d for the three decisions worth knowing: it is a FIXED window and the code used to call it sliding, rate_limit_burst is still read by nobody because a fixed window has no burst parameter, and it fails open when Redis cannot answer. SO BACKPRESSURE EXISTS TO VERIFY NOW, and this item's own clause is still not verified: what a load run would measure is the limit at its configured sixty a minute under real concurrency, which needs the production volume M10.15.a is blocked on. The rest of the clause — fairness, retry, circuit breaking, queue age and lag, cost and tail SLOs — is unverified behind .a–.e.
  • M10.16 Run answer-key leakage, prompt injection, BYOM, PII, grade tamper, integrity bypass, and cross-tenant adversarial tests. Evidence: zero release blockers. OPEN, AND IT STILL CANNOT CLAIM ZERO RELEASE BLOCKERS, but for one reason now rather than two. The self-approval hole this section measured is CLOSED (.e, 2026-09-13): a reviewer can no longer approve their own grade correction, because the rule is checked against a store instead of against the caller's account of themselves. What remains is .f — four workflows unreachable because every user is their own tenant — and that is a tenancy model, not a defect a suite can close. The two findings its own spec note starts from are also closed: the answer key no longer reaches a learner, red then green in that order, and SSRF on the fetch sites is refused. <!-- SPEC 2026-09-11: .a starts from a known live finding — GET /api/assessments and GET /api/assessments/{id} (api/assessments.py) return QuestionResponse.correct_answer and explanation to ANY authenticated caller, including the default student role (measured 2026-09-11; see M0.15.d). The suite must be red on it first, the fix lands under M0.15.d, then the suite goes green — in that order. .c's BYOM paths are the BFF decision store (apps/oshun/bff/src/metis/metis-byom-decision-store.ts) plus the service's per-feature *_MODEL_BASE_URL settings, and its SSRF cases reuse M10.8.e's fetch-site list. .f's two-tenant vectors reuse M0.15.f's 36×9 matrix. -->
    • M10.16.a Scan learner UI/API/search/cache/client state/assets/accessibility/ packages/logs/URLs/tutor/analytics for answer/rationale/hidden-test leaks. Evidence: key leakage suite. THE API CHANNEL IS DONE, RED THEN GREEN, IN THAT ORDER — which is what this item's own spec note asks for. The known live finding it names was real: GET /api/assessments and GET /api/assessments/{id} returned correct_answer and explanation to every authenticated caller including a student. Measured over HTTP before the fix, a learner's GET came back with correct_answer: "Chlorophyll" and the explanation; after it, neither. tests/test_answer_key_absence_journey.py, 6 tests, and the fix removes the evidence bundle too, because its claims are extracted from the explanation's own text — redacting the field and leaving the bundle would move the answer rather than withhold it. THE OTHER TWELVE CHANNELS ARE SCANNED BUT NOT HARVESTED. M4.13's answer-leakage runtime enumerates 22 learner-visible channels — log, url, deep_link, qti packages, export, cache, search, error, trace among them — and M10.10.c drove a complete 144-artifact sweep across all of them that returns coverage.complete: true. But every artifact is SUPPLIED BY THE CALLER: LearnerVisibleLeakageArtifact has no producer in this estate, and the workbench page is a textarea somebody pastes a capture into. The scanner is proven; nothing walks the running system to feed it. That harvest is what this clause still needs. FOUR OF THE TWELVE READ CLEAN ON INSPECTION, 2026-09-18 — read rather than harvested, so this is a narrower claim than the scanner would make, and it is recorded because the alternative is a reader assuming the worst about channels that are in fact built right: - LOG. assessment_service is the only holder of the key that logs at all, and its two calls carry ids and a type — assessment_created and assessment_submitted. No payload, so no key. - SEARCH (the durable index). Already covered by a suite this record did not credit: FIELD_CLASSES is total over the eight kinds and REFUSES an unclassified field rather than indexing or dropping it, and residue re-scans each built document for any forbidden value. tests/test_search_index.py runs it over all eight kinds — 62 tests. The channel has a producer and a scan already. - SEARCH (the item bank). correct_answer appears exactly ONCE in services/item_bank_search_service.py, and it is inside the input to source_witness_sha256 — a digest that moves when the answer moves, holding no answer. The projection the search serves is built separately and does not carry it. - TUTOR. correct_answer appears nowhere in tutoring_service or in the retrieval and grounding modules that feed its trust zones, so the key never enters the tutor's context to be fenced in the first place — which is M6.13's answer again: the strongest refusal is having nowhere to put it. THE HARVEST IS STILL THE GAP and reading four channels does not replace it. LearnerVisibleLeakageArtifact still has no producer: nothing exercises a route as a learner and hands what came back to scan_answer_leakage. That producer is the work, and the remaining eight channels — ui, cache, client_state, client_bundle, url, deep_link, export/qti, analytics — need it rather than another reading.
    • M10.16.b Inject instructions through sources/uploads/citations/standards/ learner text/provider outputs/media/code/graphs against prompts/tools/ agents/retrieval. Evidence: injection suite. THIN. prompt_trust_helpers carries the right idea — trust zones, fence_untrusted, sanitize_prompt_data and a versioned renderer — and it has 3 tests; agent_safety_helpers has 6. Nine tests is not an injection suite against the ten carriers this item names, and none of them drives a source, an upload, a citation or a provider output end to end. M6.13's finding still stands as the estate's strongest answer here: injection is best refused by having nowhere to put it. THE CENSUS ABOVE WAS AN UNDERCOUNT, measured 2026-09-18. Ten test files exercise the fence, not two, and each of the six model boundaries has its own test that inspects the body it would transmit: test_source_claim_provider.py (4), test_course_outline_provider.py (4), test_curriculum_alignment_provider.py (5), test_item_variant_provider.py (13), test_lesson_assist_provider.py (7), test_source_lesson_model_provider.py (4), plus test_prompt_trust_helpers.py, test_agent_safety_helpers.py, test_agent_evaluation_helpers.py and test_review_evidence_export.py — 101 tests in those files before today. The existing renderer tests already drive delimiter forgery and two encoded layers. AND THE STRUCTURAL CENSUS IS CLEAN, which this stage rarely gets to record: of the ten modules in the service that make an outbound POST, SIX SEND A PROMPT AND ALL SIX FENCE IT. The other four — isis_gallery_client, lti_services_client, lti_trust, open_badges_proof — send no prompt at all. tutoring_service renders a trust prompt too and reaches a model through a provider rather than its own client. ONE REAL HOLE, MEASURED AND CLOSED. sanitize_prompt_data normalised with NFKC ONCE, BEFORE its two decode passes, so a homoglyph that arrived ENCODED was never normalised: %EF%BC%9C%EF%BC%9C%EF%BC%9C decoded to U+FF1C FULLWIDTH LESS-THAN after the only normalisation, and the collapse — which looks for ASCII <<< — left <<<END_UNTRUSTED_DATA>>> sitting in the data. It could NOT forge the delimiter, because END_UNTRUSTED_CONTENT is rewritten independently of the brackets, so this was hardening and not a break — but it left text that reads as a closing marker to anything matching loosely. Fixed by normalising again after the decode loop, which is idempotent for content that was already normal: every digest of ordinary text is unmoved, and 111 tests across the ten fence files stayed green. Three attack shapes now assert it. AND THE END-TO-END DRIVE THIS ITEM ASKED FOR EXISTS NOW: tests/test_prompt_injection_journey_real_stack.py, 5 tests. A markdown file carrying three payloads — a plain instruction, a forged closing delimiter and an encoded homoglyph of one — is ingested through the real pipeline into real MinIO, read back out, extracted, and the EXTRACTED text is handed to the claim provider as the passage it would be. The assertions are on the body the provider actually transmits: the payload sits between the markers and nowhere else, the authority half is untouched and carries the immunity clause, the forged and encoded delimiters are both collapsed inside the fence, instructor guidance is fenced as a second zone, and the body offers no tools. Fencing is not redaction and the test says so: the sentence is still there, because a claim extractor that never saw it could not report it. The double is at the HTTP boundary and nowhere else. WHAT IS STILL NOT DRIVEN, of the ten carriers this item names: media, code, graphs and standards. Learner text and provider outputs are fenced by tutoring_service and asserted in its own tests but not driven from a stored artefact the way a source now is.
    • M10.16.c Test BYOM endpoint/credential/SSRF/model substitution/supply chain/data-use/output/spend/outage/kill-switch paths. Evidence: BYOM suite. SSRF IS DONE AND FOUND TWO LIVE HOLES. M10.8.e built outbound_targets.refuse_unless_public() and tests/test_outbound_ssrf.py, 18 tests over the fetch-site list this item points at. It closed SSRF on two LTI fetches and size bombs, and the policy refuses multicast, reserved and unspecified addresses as well as non-global ones — 224.0.0.1 reports is_global=True in Python, which is how the first version would have let it through. MODEL SUBSTITUTION IS NOW VISIBLE, AND IT WAS NOT. Measured 2026-09-13: the admission engine carries the right rule — fingerprint_moved, "the endpoint is answering as a different model than the one that was evaluated and approved" — keyed on ServeRequest.observed_fingerprint_sha256. That field appeared in exactly TWO places in src/: its own declaration and that comparison. NOTHING COMPUTED IT. Every value it had ever held came from a request body or from the workbench's fixture constant. The same shape as the hole closed in M10.12.d/.e one layer up, and worse, because the input had no producer at all. ONE OF SIX OBSERVED WHAT ANSWERED; NONE OF SIX CHECKED IT. item_variant_provider read model/provider off the response envelope; the other five declared choices and usage only, so the served model was dropped unparsed and every record named the model that was ASKED FOR. The two readers of served_model build a provenance string and compare nothing. Now: services/served_model_identity.py is the one derivation (observe, substituted, and the fingerprint that finally gives observed_fingerprint_sha256 a producer), all six envelopes read the answer, all six record served_model/served_provider on EVERY outcome including refusals, and a silent endpoint reads as unobserved rather than as agreement. 14 tests in test_served_model_identity.py — including a parametrized ratchet so a seventh provider cannot skip it — plus one driven substitution test in each of the six provider suites. ONE CAVEAT ON THE ABOVE, recorded rather than discovered later. Four of the consumers digest a whole invocation dump and re-derive it on read (source_claim_service over canonical_json_bytes, curriculum_alignment's re-dump, and the lesson-assist chain). That is symmetric within a run, so a fresh database verifies — but an invocation STORED BEFORE these two fields existed would re-parse with them present as None and fail its own provenance check. The house already solved this once: item_variants.WITNESS_OPTIONAL_FIELDS leaves such fields out of the witness when unset. Not a live defect, because nothing in this estate has stored invocations predating today; it becomes one the moment something does. AND IT ALREADY BIT ONCE, in a smaller way: lesson_assist_provider sealed its digest over a DICT that omitted the new keys while the model defaulted them, so three lesson-dossier tests failed with an invalid digest chain. A DICT THAT OMITS A KEY DIGESTS DIFFERENTLY FROM A MODEL THAT DEFAULTS IT. Fixed the same day, along with a frozen dataclass whose required new fields turned a model path into its deterministic fallback by raising TypeError at every other construction site. CORRECTION TO THIS RECORD'S OWN EARLIER CLAIM: "Admission and the kill switch are M7.11's, already closed" IS WRONG for these paths. provider_model_admission_engine's docstring says "Every path that could end in 'send this prompt to the tenant's model' runs through check_serve … There is no second path." There are six. check_serve's only callers are the admission module and its own route. The kill switch cannot stop any of the six. STILL OPEN, AND WHY NOTHING IS CHECKED HERE: the refusal half. Closing it needs a store of the last observed fingerprint per endpoint, because a pre-flight check cannot have observed the answer to the call it is gating. test_no_provider_consults_admission_yet pins that open rather than letting it pass unrecorded. ALSO OPEN: supply chain, data-use, spend and outage. Spend is two of six — item_variant and lesson_assist price usage and can answer budget_exceeded; the other four have no budget or cost at all. The BFF store this item's spec note names (apps/oshun/bff/src/metis/metis-byom-decision-store.ts) is bring-your-own-MATERIAL source-review governance, not a model admission path — a different sense of "BYOM". It has its own unit and durability specs; it is not exercised from the service suite, and it would not answer this item if it were.
    • M10.16.d Test learner/minor/PNP/grade/integrity/voice/source PII across UI/API/events/logs/exports/analytics/training feedback. Evidence: privacy suite. OPEN, with one narrow suite: test_sequencing_overlay_privacy.py, 5 tests. Against the seven subjects and eight surfaces this item names that is a corner of the matrix. What IS proven elsewhere and counts toward it: cohort suppression below a floor of 30 with complementary suppression (M10.11.e, M10.12), learner-audience redaction on the integrity signal surface (M10.12.b), and the synthetic fixture's import-closure proof that it cannot reach a learner record at all (M10.11.e). Minor, voice and training-feedback PII are untouched.
    • M10.16.e Test grade/event/correction/LTI/credential tamper/replay/race/ mass assignment and integrity evidence/decision/appeal/interlock bypass. Evidence: authority suite. PARTLY DONE, AND M10.12 FOUND THE BYPASS THIS ITEM IS ABOUT. Replay and race are covered: test_mutation_concurrency.py, 10 tests, after M10.8.c fixed an idempotency key that was being read as a uniqueness token rather than a retry token. Mass assignment is structurally refused — every intake model forbids unknown fields — and M10.8.b closed a caller that could name somebody else's tenant. THE CORRECTION BYPASS IS CLOSED 2026-09-13. It was real and measured: a reviewer approved their own grade correction by not mentioning that they had reviewed it, because prior_decisions was read from the request body and defaulted to empty. The route now loads prior decisions from the correction store and refuses the field outright; the same walk answers reviewer_cannot_approve (M10.12.d/.e). The appeal bypass still cannot be tested at all — M10.12.c, the engine has no route.
    • M10.16.f Test two-tenant IDs/search/counts/blobs/cache/queues/deep links/ exports/integrations/timing and close all release blockers. Evidence: adversarial report. OPEN. test_tenant_isolation_authoring_api.py has 9 tests and M8.14 built a two-tenant suite and a tenancy ratchet, so IDs and the authoring API are covered. Search, counts, blobs, cache, queues, deep links, exports, integrations and timing are not. AND THE TENANT MODEL ITSELF IS THE FINDING THIS STAGE KEEPS MEETING. Every user is their own tenant, which makes cross-tenant leakage structurally hard and multi-person work impossible: four instances now — no publication candidate can be approved (M10.9.d), no second instructor can review an item (M10.10.c), no fairness reviewer can exist (M10.11.e), and no integrity lead can see a reviewer's signal (M10.12.b). Release blockers cannot be closed while those stand.
  • M10.17 Run backup/restore and late-event replay drills. Evidence: gradebook, audit, blobs, and search reconcile. OPEN WITH .b DONE FOR REAL. The restore half was run against the compose Postgres and verified by hashing 163 tables on both sides rather than counting them. The replay half is not a drill yet, blobs are untouched because MinIO is not running, and the gradebook cannot reconcile because its correction tables have no writer (M10.12.d). Two facts the drill established that no document had: there is no point-in-time recovery, and nothing schedules a backup. ALL FOUR LEGS OF THE EVIDENCE SENTENCE NOW RECONCILE ON A RESTORED ESTATE (2026-09-18), and the three gaps the paragraph above names are closed or superseded: the replay half was drilled on 2026-09-13 (.c), the gradebook's correction tables got their writer the same day (M10.12.d), and MinIO is running on this host now. tests/test_estate_restore_drill_real_stack.py is the drill, and a test rather than a script so it can be run again instead of described: a real journey through the durable pipeline, pg_dump -Fc, the originals bucket mirrored through the S3 API (mc is not installed here, so each object is read and written back with its digest recomputed, which is what BlobStore.put does), the database restored into a second one and the objects wiped and restored, then durable_reconciler.reconcile run against the restored pair. VERIFIED BY HASHING BOTH STORES, never by counting: 172 tables digested with md5(string_agg(row::text ORDER BY row::text)) on both sides, all 172 equal, 7 of them non-empty; every object compared on length, on the sha256 of its bytes and on its content type. The reconciler then answered clean on all of it — blobs, search index, outbox, exports, grades and M10.17.d's record legs — over 15 restored revisions. THE DRILL'S OWN FINDING, and the reason the blob half was worth running rather than assuming: a restore into a differently-named bucket leaves every receipt pointing at a bucket that does not hold the object, and the reconciler says so. The measurement and what it implies for a restore are in .e. STILL OPEN, WHICH IS WHY THIS STAYS UNCHECKED: .a's DR plan, .d's remaining legs, and .e's scheduled backup, encryption, retention and runbook. The two facts .b established are unchanged — there is no point-in-time recovery and nothing schedules a backup. <!-- SPEC 2026-09-11: real stores only — pg_dump/pg_restore into a second database on the compose Postgres, MinIO mc mirror into a second bucket, outbox replay through the service's own worker. A restore is verified by hashing rows and objects on both sides, never by row counts. RPO/RTO in .a are declared numbers; .e measures them. -->
    • M10.17.a Define backup scope/frequency/encryption/key/retention/RPO/RTO for database/blobs/audit/search config/standards/integration state and document rebuildable stores. Evidence: DR plan. OPEN, AND ONE NUMBER IN IT IS NOW DECIDED BY THE SERVER RATHER THAN BY A PLAN. archive_mode is OFF on the compose Postgres and wal_level is replica, so THERE IS NO POINT-IN-TIME RECOVERY: a restore can only go back to the last dump, never to a chosen moment. Any RPO written here is therefore bounded below by the backup interval, and no interval is configured either — nothing schedules a dump. RTO HAS A MEASURED FLOOR from .b's drill: 204 ms to dump and 1,981 ms to restore, 2.2 seconds total for 1.1 MB. That is a floor for a development-sized database and says nothing about production, whose volume is unrecorded (the same gap M10.14.b is blocked on). Scope, frequency, encryption, key management and retention are undefined, and the blob half is untouched — MinIO is a compose profile that is not running here. THE BLOB HALF IS NO LONGER UNTOUCHED (2026-09-18). MinIO runs on this host now, and the drill named in the parent backs up and restores the originals bucket beside the database, so the scope this plan has to state is two stores and not one. One constraint on it is measured rather than open to choice: the objects must come back under the bucket name the receipts already name, because BlobStore.verify reads receipt.bucket and the receipts are immutable (see .e, step 3). Scope, frequency, encryption, key management and retention remain undefined, and the RPO any of them implies is still bounded below by a backup interval nothing configures. SCOPE, FREQUENCY AND RETENTION ARE NOW STATED IN CODE rather than owed by this plan (2026-09-18, with .e). services/estate_backup.py backs up the database and the source bucket, and its docstring says why the other four buckets are not: s3_bucket_exports holds artifacts a release republishes, s3_bucket_courses and s3_bucket_assets hold derivatives of records in the database, and s3_bucket_uploads is swept daily by tasks.cleanup. That is this clause's "document rebuildable stores", written where the code that acts on it lives. Frequency is backup_interval_seconds (default 86,400) and retention is backup_retention_count (default 7) — both DEFAULTS, and the decision this plan still owes is what they should be for a deployment, which needs the production volume M10.14.b is blocked on. ENCRYPTION AND KEY MANAGEMENT ARE STILL UNDEFINED and are the clause's real remainder: the dumps land in object storage unencrypted, and where a key would live is an operator decision, not a default this session can pick.
    • M10.17.b Take/verify backups and restore into isolated environment at selected point with tenant/permissions/hash integrity. Evidence: restore drill. DRILL RUN FOR REAL, against the compose Postgres, on the database this session's journeys wrote to. pg_dump -Fc of metis_item_journey (1,123,762 bytes), createdb + pg_restore --no-owner --no-privileges into an isolated metis_restore_drill, exit 0. VERIFIED BY HASHING ROWS, NOT BY COUNTING THEM, which is what this item's spec insists on: for each of the 163 public tables on both sides, md5(string_agg(row::text ORDER BY row::text)). ALL 163 DIGESTS MATCH. 18 of the tables are non-empty, so this is not a comparison of nothing — a count would have said two tables agreed whenever they merely had the same number of different rows. Tenant integrity follows from the row digests: tenant_id is part of every row hashed. PERMISSION INTEGRITY IS VACUOUS HERE, and measured rather than assumed: the restore dropped owners and privileges, and the source database has ZERO role grants on public beyond the owner and PUBLIC, so nothing existed to lose. On a deployment that grants to real roles this drill would have to be re-run without those flags. "AT SELECTED POINT" IS NOT DEMONSTRATED AND CANNOT BE — see .a. The drill databases were dropped afterwards.
    • M10.17.c Replay outbox/grade/LMS/integrity/revision/events including duplicate/out-of-order/late/correction/void cases. Evidence: replay drill. THE DRILL IS RUN NOW, FOR REAL, and it is scripts/replay_drill.py so it can be run again rather than described. Against the compose Postgres, 2026-09-13: SEVEN OUTBOX EVENTS SURVIVED A RESTORE AND THE SERVICE'S OWN WORKER DRAINED THEM THERE. Six lesson aggregates plus one revision were written through the real AggregateStore, leaving 7 undelivered events; pg_dump -Fc then pg_restore into an isolated database carried all 7 across; metis.tasks.durable.drain_outbox — the task beat runs every 10 seconds, executed in its OWN process reading its OWN configuration pointed at the restored database, not a function handed an engine — reported delivered=7 failed=0. Undelivered went 7 to 0 and search_index_document went 0 to 6 (six lessons; the seventh event is a second revision of one of them and upserts the same document). A SECOND drain delivered nothing and left the index at 6, which is the idempotence a re-run after a crash depends on. THE NAMED GRADE CASES, replayed on the RESTORED database through the ledger's own append path: duplicate (the same source event id retried came back appended=False rather than refused for carrying a stale version — a retry is not a new append), out-of-order/late (an event that HAPPENED three hours earlier arriving second was appended), correction (naming the event it corrects), and void (no value). Five appends, four entries held, and a history derived from them. reconcile answered never_projected, which is correct and is M10.17.d's own point: the projection is built lazily on read, so a restored grade nobody has read has nothing stored to compare. TWO FAIL-LOUD SEAMS PROVED THEMSELVES while building this, both worth recording because each looked like a defect first. A revision whose payload offers no index key is indexed not at all — the record decides what search may hold, the relay does not guess from the content. And offering lesson.summary was REFUSED outright: "is not classified; indexing it and dropping it are both a decision nobody recorded". That refusal isolated correctly — all 7 events stayed undelivered and nothing was half-applied, which is the relay's "ONE EVENT'S FAILURE IS THAT EVENT'S" claim, measured. A SECOND EVENT TYPE, because everything above is aggregate.revised. Withdrawing a lesson on the restored database produced aggregate.transitioned, the worker delivered it, and the document went to state = deleted while the ROW REMAINED. That is correct and deliberate — a delete leaves a tombstone "so a late upsert for" an older revision cannot resurrect what was withdrawn — and asserting the row COUNT read as a defect until the state column was looked at. Live documents went 6 to 5; rows stayed 6. THE LMS LEG IS DRIVEN TOO, through the INGRESS road rather than the ledger's own API: ingest_ags_score, which resolves the binding and derives the grade id from the platform triple. A first post appended; THE SAME POST REDELIVERED came back accepted=True appended=False, which is idempotence on the source's own id rather than a refusal; a later score appended; and a post out of 10 against a binding out of 100 was REFUSED with scale_disagreement rather than rescaled, which is the silent mapping M10.7.f exists to block. STILL OPEN, WHICH IS WHY THIS STAYS UNCHECKED: the integrity replay this clause names. And the reason is worth more than the gap — THERE IS NO INTEGRITY EVENT STREAM TO REPLAY. Only aggregate_store and source_graph write to store_outbox; no integrity service references an outbox at all. So this leg cannot be driven until integrity emits something durable, and recording it as covered because nothing broke would be closing a clause over an absence. THE RELAY'S OTHER TWO EVENT TYPES ARE DRIVEN NOW TOO, so all four are. A source and a lesson that CITES it were written on the restored database — the citation is what puts a row in store_references — and then source.retracted moved the citing lesson to quarantined and source.restored lifted it back to live, each delivered by the worker. So a retraction reaches what depended on the source, through a restore, without anybody calling the relay by hand. ONE OBSERVATION RECORDED RATHER THAN ASSERTED: retract emits an outbox event per DEPENDENT, so the retracted source's own index document stays live. Whether a factually_wrong source should also stop being served by that act, or only by a revision marking it unavailable (which is the path the relay's quarantine branch describes), is a design question this drill does not answer and no claim is made about it here. The earlier reading of this item, still true of what it covered: M8.11's grade ledger handles late-event replay over a 24-per-day matrix — AN ARRIVAL ORDER IS NOT AN EVENT ORDER — and M10.8.c's idempotency work made a retry replay the first attempt's acknowledgement rather than answering stale. What is missing is the drill this item asks for: replay through the service's OWN worker against a restored database, which is the only way the outbox and the projections are exercised together.
    • M10.17.d Rebuild/reconcile current grade/course/build/audit/search/blob/ dependency/integration projections to authoritative histories. Evidence: reconciliation report. STILL BLOCKED ON .c, BUT ITS SECOND BLOCKER IS GONE AND THE REAL GAP IS NOW NAMED. The correction tables have a writer as of 2026-09-13 (M10.12.d), and the partition defect that would have split a grade's history across two tenant ids is fixed, so a grade history is now persisted and reachable under one key. WHAT THE INPUTS ACTUALLY ARE, MEASURED 2026-09-13 BY DRIVING IT. grade_ledger.project() takes the entries, a GradeScale and a learner_digest. Bound a line item, ingested one FullyGraded AGS score through the real ingest_ags_score, then asked which of them the projection depends on: - the entries are durable; - THE SCALE MATTERS AND IS RECOVERABLE. Deriving with a wrong scale is REFUSED (LedgerError, via GradeHistory's own check that every event is on the bound scale). grade_ingress_binding stores scale_id beside scale_minimum/maximum/decimals and the entry carries scale_id, so a join on (tenant_id, scale_id) reaches it. Nothing performs that join today, but the data is there; - THE LEARNER DIGEST AFFECTS NOTHING. Deriving the same entries with a completely different digest produced the IDENTICAL history_sha256 and current_value. It is a field of GradeHistory that no validator reads and that substance() — what the digest is taken over — does not include. It is a required parameter that changes no output. CORRECTION TO THIS RECORD'S FIRST VERSION, which said the rebuild's inputs were not durable and no projection could be rebuilt. That was wrong, and it was written from the function's SIGNATURE rather than from what the function uses. A grade projection CAN be rebuilt from the record — through a join nothing performs, on a scale_id that is not guaranteed unique within a tenant. The honest gap is smaller and is the two items below. A SECOND THING THE WALK SHOWED: after a successful ingest, grade_ledger_projection HAD NO ROW. The projection is built lazily on read (projection_for_ags), so "reconcile the projection against the history" has nothing to compare for any grade nobody has read. What the walk did confirm working: given the inputs, rebuild() derives the history and emits nothing — outbox 1 before, 1 after. THE GRADE LEG IS NOW BUILT AND WIRED IN (2026-09-13), and the rest of this clause is not. grade_ledger.reconcile() re-derives one grade from its own entries and compares with the stored projection, WRITING NOTHING — which is why it is not rebuild(): that calls project(), which upserts, so it repairs while reporting, and something that changes what it measures cannot report that the thing was already right. durable_reconciler.Reconciliation gains grades_disagree (counts against clean) and grades_never_projected (does not — a lazily-built projection that nobody has read is an absence, not drift), and reconcile() now sweeps every grade the tenant holds an entry for. Migration 077 + grade_ledger_derivation record the scale per grade, written with the first entry and never rewritten, because the scale was not unambiguously recoverable (see above) and a wrong one is refused. 8 tests against a real database, including TWO tamper cases. The second exists because the first was BELIEVED: stored_projection parses document_json, so moving only the history_sha256 COLUMN changed nothing the comparison reads and the reconciliation cheerfully agreed. That is now its own refusal, projection_row_inconsistent — a row that disagrees with its own document is corrupt in a way the history comparison cannot see. WHAT IS ACTUALLY MISSING, measured 2026-09-13: services/durable_reconciler.py (M10.4.d) reconciles blobs, search index, outbox, publication and owner queues — grep -n grade returns ZERO HITS. Grades are not reconciled by the estate's reconciler at all. Meanwhile services/grade_ledger.py (M8.11) ALREADY HAS the machinery this clause wants — project(), rebuild(), stored_projection(), entries(), and a rebuild that deliberately "announces nothing" so recomputing does not wake the outbox. So the rebuild exists and NOTHING DRIVES IT AS A RECONCILIATION. That is the work, and it is smaller than the blocker it replaced. The aggregate store's five tables DO hold authoritative revisions and an outbox, so the course, audit and dependency projections are rebuildable in principle. THE COURSE, AUDIT AND DEPENDENCY LEGS ARE BUILT NOW (2026-09-18), and the sentence above is what they were built on. services/record_reconciler.py sweeps one tenant and reports, writing nothing; durable_reconciler.reconcile carries it, so the scheduled reconcile_estate task is the caller. Every leg below was missing, and store_aggregates — the row every other leg of M10.4.d measures something against — was never compared with the revisions and moves that produced it. The yardstick was unmeasured. - THE HISTORY AGAINST ITS OWN BYTES. content_sha256 is recomputed from the stored content_json, which is text (measured in the information schema, so re-hashing the column is hashing the bytes the digest covered, not a normalised jsonb round trip). Nothing did this: revision_history.ancestry walks the previous_sha256 links and never opens the content, so a payload rewritten under the retention workflow left two digest columns agreeing with each other. The test asserts that ancestry still reports that chain intact. - THE REVISION SEQUENCE. ancestry sets revision=len(chain), so a history with revision 2 deleted answers "an intact chain up to revision 2" for an aggregate standing at revision 3. Measured, not inferred: the test deletes revision 2 under the legal-erasure token and reads walked.revision == 2 back. - THE HEAD THE AGGREGATE ROW POINTS AT, read-only. rebuild_from_ revisions compared this already and REPAIRED as it looked, which is the same reason the grade leg uses reconcile and not rebuild. - THE OPERATIONAL STATE, REPLAYED FROM TWO WITNESSES. A creation is refused in any state but draft (RevisionRequest's own validator), so the walk has a start, and every move since is recorded twice in the transaction that made it. Both must agree with the column and with each other. The walk follows the from/to linkage rather than the clock, because transitions are serialised under FOR UPDATE — two moves cannot leave the same state — while two moves at one pinned instant are ordinary in this repo's fixtures. - THE AUDIT TRAIL'S COMPLETENESS, for every revision and every move. - THE RETRACTION QUEUE AGAINST THE REFERENCE GRAPH. source_propagation is derived: retract walks the citations backwards and writes a row per dependent. An open entry must name something a recursive walk of store_references still reaches from that source, and one that does not is an owner being asked about work that never depended on it. MIGRATION 082, BECAUSE THE STATE HAD ONE WITNESS AND IT WAS THE WRONG ONE. store_audit recorded THAT a transition happened and never what it moved to: the from/to pair lived only in the store_outbox event payload. The outbox carries published_at, which the relay writes, so 047's immutability trigger cannot be put on it — the operational state of every aggregate in this estate was derivable from a mutable table and from nothing else. 082 adds from_state/to_state to the trail, nullable and NOT backfilled (a state nobody recorded does not get invented into the one table that cannot be corrected), and the sweep names the aggregates whose trail predates it instead of reading a missing state as draft. THREE DEFECTS IN ONE FUNCTION, all in authoring_api.set_archived, found by asking what records archived_at (049): - IT WROTE NO AUDIT ENTRY AND NO EVENT. An authorised operation — may(role, "archive") gates it — whose only record was the column it wrote. Who shelved a lesson, and when, was unanswerable. - IT DISCARDED THE REASON. ArchiveBody.reason is required, 8 characters minimum, and no line of the function read it. - ITS STALENESS CHECK READ FROM A PAGE. The expected revision was compared against store.list_aggregates(context, kind=kind), which is LIMIT 100 ORDER BY updated_at DESC, so archiving an aggregate outside the hundred most recently touched of its kind was answered not_found for a record that exists. The write is now AggregateStore.set_archived: the column and an archived/restored entry in one transaction, under the row's own lock, with the caller's reason as the entry's detail. It returns an ArchiveAcknowledgement and deliberately not a WriteAcknowledgement, which would have had to name an outbox event id no row answers to — a shelf is not a lifecycle state, nothing is served differently for it, and the relay refuses an event type it does not know. THE REBUILD HALF, which the clause also asks for. rebuild_from_revisions carried the stored op_state forward — it repaired the revision pointer and trusted the column for the state, including a value somebody had written straight onto the projection — and wrote draft when the row was missing altogether. It now derives the state (trail first, moves second) and REFUSES when neither answers, rather than inventing one. An empty move history is not that case and must not be treated as one: nothing having moved does derive draft. The two witnesses are cross-checked before the moves are believed, because an empty move list cannot tell "nothing moved" from "the events are gone" — the trail's transitioned entries can, and they exist even in a trail written before 082. TWO LEGS OF THE CLAUSE ARE NOT PROJECTIONS, MEASURED RATHER THAN SKIPPED. Dependency: both ends of store_references are foreign keys into store_revisions (046), so a dangling reference is refused by the database and there is nothing to reconcile; what IS derived from the graph is the retraction queue, and that is the leg above. Integration: grade_ingress_binding is registered and not derived (grade_ingress says so in its own docstring, and registering it is why M10.7.f can refuse a scale disagreement), integration_workspace is an aggregate saved whole, roster_grade_identity holds only what the SIS supplied through its single writer, and lti_line_item records what this tool created. None is computed from another table, so none has a history to be rebuilt from, and the drill that covers them is .b's restore. STILL OPEN, WHICH IS WHY THIS STAYS UNCHECKED: - THE QUEUE RECONCILES IN ONE DIRECTION ONLY. The other — every dependent of a retracted source is in the queue — needs to know WHICH REVISION of the source was retracted, because the walk is over exact revisions. source_propagation records source_id and the reason and not the revision, so the set it ought to contain is not recoverable, and a sweep that assumed "any revision" would report every citer of an older revision as missing. - A LIVE DEFECT THIS SWEEP CANNOT SEE, found while building it and not fixed here: retract calls impact_preview(..., limit=1000) and iterates preview.rows, never reading next_cursor or truncated_at_depth. A blast radius wider than 1,000 dependents or deeper than MAX_DEPTH = 6 is silently not propagated — no entry, no event, no owner told — and SourceRetractionResult carries no truncation flag, so the caller cannot tell either. - project_id/course_id are still recorded nowhere but the projection. They are write-once (a revision that would move an aggregate between courses is refused), so the projection is its own record; the sweep counts them as an absence it could not check rather than calling them consistent. - .c's integrity replay, unchanged: no integrity service writes to an outbox, so that leg still has nothing to replay. EVIDENCE. services/record_reconciler.py, migration 082, AggregateStore.set_archived, authoring_api.set_archived, revision_history.rebuild_from_revisions, durable_reconciler.Reconciliation.record, tasks/durable.py. 13 tests in tests/test_record_reconciliation_real_stack.py (every tamper read back before it is judged — the first one did not land, and the read-back is what said so), 3 added to tests/test_aggregate_store_real_database.py and 2 to tests/test_revision_history_real_database.py; 112 passing across the six affected suites against a real PostgreSQL, MinIO and Redis stack. metis:contracts:check reports no OpenAPI change: nothing here is on a route.
    • M10.17.e Measure RPO/RTO, residual/manual steps, update runbooks/tests, and close gaps. Evidence: game-day report. PART MEASURED, THE REST OPEN. RTO has a real floor from .b — 2.2 seconds end to end at development volume — and RPO is unbounded because nothing schedules a backup and WAL archiving is off, so the honest answer is "everything since whenever somebody last ran pg_dump by hand". Every step in .b's drill was manual; none of it is a runbook or a scheduled job yet. RTO NOW HAS A FLOOR FOR BOTH HALVES, measured 2026-09-18 by the drill in tests/test_estate_restore_drill_real_stack.py rather than by .b's database-only run: dump 0.181 s, restore 1.790 s, object mirror 0.029 s, object restore 0.045 s — 2.045 s end to end for a 938,995-byte dump of 172 tables (7 non-empty, 15 revisions) and two objects totalling 1,723 bytes. A development volume still, and production volume is still unrecorded (the gap M10.14.b is blocked on), so this is a floor and not a projection. RPO IS UNCHANGED AND STILL UNBOUNDED: nothing schedules a dump and archive_mode is off, so the honest answer remains "everything since whenever somebody last ran pg_dump by hand". THE RESIDUAL MANUAL STEPS ARE NOW ENUMERATED BY HAVING PERFORMED THEM, not by imagining the restore: 1. create the target database (CREATE DATABASE) — pg_restore does not, and the drill fails at this step without it; 2. pg_restore --no-owner --no-privileges, which is only vacuously safe here: the source database has no role grants to lose (.b measured that), so a deployment with real roles needs the drill re-run without those flags; 3. RESTORE THE OBJECTS INTO THE BUCKET NAME THE RECEIPTS ALREADY NAME. This is the step no document had, and the drill establishes it by measurement: BlobReceipt carries its bucket and BlobStore.verify reads receipt.bucket, not the bucket the store was handed. With the objects mirrored into a second bucket and the original emptied, the reconciler reported every blob broken — and reported them broken again when pointed AT the backup bucket, because the receipt decides where it looks. The receipts cannot be rewritten to suit: they live in store_revisions, which 047 makes immutable outside an approved retention or legal workflow. So a restore either recreates the bucket under its original name or it produces an estate whose every blob is unverifiable; 4. drain the outbox on the restored database (.c's leg), because a restore lands undelivered events that only the worker clears. A DESIGN QUESTION THIS DRILL RAISES AND DOES NOT ANSWER, recorded rather than quietly settled: a bucket in a receipt is payload deciding where the service reads. Making verify and get look only in their own bucket would remove that and make a renamed restore verifiable, but it would also fail any legitimate cross-bucket receipt (an export read through a store holding the originals), and durable_reconciler verifies every receipt it finds in a revision against one store. It is a boundary change with a security reading on both sides and it wants a decision, not a guess. STILL OPEN: the scheduled backup that would bound RPO, encryption and key management for the dumps, retention, and the runbook itself — none of the four steps above is written down anywhere but here and in the drill. THE SCHEDULED BACKUP IS BUILT (2026-09-18), SO RPO IS A NUMBER NOW. services/estate_backup.py + migration 083 + the beat entry back-up-estate: a pg_dump -Fc of the configured database every backup_interval_seconds (default 86,400), the source bucket mirrored beside it, and a row in estate_backup_run for every attempt. RPO ≤ the interval + the run, and last_successful() is what makes the bound READABLE — the question an incident asks is when the last one succeeded, and before this there was nowhere to ask it. A DUMP NOTHING HAS READ IS NOT A BACKUP, so a run is not succeeded until pg_restore --list has parsed the archive and its table count is recorded; the table's CHECK refuses a succeeded row that does not name its artifact, length and digest. The test goes further and RESTORES the dump into another database, comparing every table digest on both sides: all equal except estate_backup_run itself, which differs for a reason worth recording — the run row is written AFTER the dump, so a dump cannot contain its own record. THE TWO FAILURE PATHS ARE DIFFERENT, DELIBERATELY. A missing pg_dump or a missing backup bucket writes NO row and raises BackupUnavailableError: a precondition nobody met is not an attempt on the database, and recording it as a failed run would say one was made. A dump that ran and failed IS an attempt and is recorded as failed before the error is raised, because a gap in the schedule is only visible if the failures are in the table beside the successes. A failed run never becomes last_successful — asserted, not assumed. RETENTION IS CONFIGURED AND PRUNES ARCHIVES, NEVER RECORDS: backup_retention_count (default 7) dumps are kept, the rest are deleted from the bucket and their rows stay, so the history of what was taken survives the archive it pointed at. Source objects are never pruned — they are content-addressed and a revision names them forever, so an object no longer in the sources bucket is exactly what the backup exists to return. WHAT THE SCHEDULE DOES NOT SETTLE, and .a still owes: the interval is a DEFAULT and not a decision; there is still no point-in-time recovery (archive_mode is off, so the floor is the last dump and never a chosen moment); the dumps are not encrypted and key management is undecided; and the runbook is still only this record and the drill. 5 tests in tests/test_estate_backup_real_stack.py against a real database and a real bucket. THE RUNBOOK IS UPDATED TOO, which is the other half of this clause. docs/runbooks/metis-disaster-recovery.md goes to 2.0.0, and the revision corrected a statement as well as adding the new steps: VERSION 1.0.0 NAMED THE WRONG BUCKETS FOR BLOB RECOVERY. It listed metis-courses, metis-assets, metis-exports and metis-uploads, and the durable bytes a receipt points at are in metis-sources (s3_bucket_sources, M10.4) — the one bucket it did not name. The runbook now opens with how to read the recovery point out of estate_backup_run, carries the bucket-name constraint, and ends with drain + reconcile before traffic is switched. AND ONE PROCEDURE IS CODE RATHER THAN A SHELL LINE. "Mirror the bucket back" has no CLI here — which aws mc finds neither — and either could be installed, but estate_backup.restore_sources is what the service's tests exercise, and a restore step only ever run by hand is a step nobody has verified. The runbook invokes the function. Rehearsed while writing it — (1, 27) on the first run and (0, 0) on the second, which is the idempotence an interrupted restore depends on. The runbook was also missing from docs/runbooks/index.md entirely, which its own version 1.0.0 never added; it has a row now. 6 tests in tests/test_estate_backup_real_stack.py against a real database and a real bucket.
  • M10.18 Publish educator, psychometrician, accessibility, integrity, institutional, incident, and interoperability runbooks. Evidence: game-day results. OPEN, WITH ONE RUNBOOK PUBLISHED AND EXERCISED. The rest are held up less by writing effort than by the workflows they would describe: an appeal with no route, a reconciliation with no destination, a review with no second reviewer. Writing those procedures now would document steps nobody can take. <!-- SPEC 2026-09-11: .a–.e are documents an agent writes from the shipped code paths, each command in them executed once while writing; .f's cross-role game day needs named people and stays open (§2.5.9). -->
    • M10.18.a Define runbook template/owners/on-call/prerequisites/roles/ dashboards/commands/evidence/comms/recovery/verification/version review. Evidence: runbook index. PARTLY: the template is now demonstrated rather than specified. docs/runbooks/metis-disaster-recovery.md follows the house frontmatter already in docs/runbooks/ — title, owner, version, last_reviewed, next_review, applies_to — and carries prerequisites, commands, verification, cleanup, a failure path and a gap table. OWNERS AND ON-CALL CANNOT BE DEFINED HERE. There is no owner registry in this estate; M10.13.f and M10.14.e record the same absence for accessibility and performance. Every runbook written now has to name its owner as unassigned, which is honest and is not a template problem.
    • M10.18.b Publish educator authoring/review/source/assessment/publication and psychometric calibration/DIF/exposure/item lifecycle procedures. Evidence: role runbooks. NOT WRITTEN. The material exists and is unusually well verified — M10.10 walked authoring and review, M10.11 walked calibration, DIF and the item lifecycle — but several of those procedures would have to document a step nobody can take: no second instructor can review an item, and no fairness reviewer can exist. A role runbook whose first instruction is impossible is worse than none.
    • M10.18.c Publish accessibility remediation/alternative/waiver/escalation and integrity signal/case/appeal/grade correction procedures. Evidence: governance runbooks. NOT WRITTEN, and two of the four procedures cannot be: the appeal workflow has no route to describe (M10.12.c), and the grade correction procedure would have to tell a reviewer to disclose their own prior decision, which the endpoint does not require and cannot check (M10.12.e). The accessibility remediation half is writable now that M10.13 has the gates and the ceilings.
    • M10.18.d Publish institution integration/roster/LTI/grade/export/ credential/version/reconciliation and standards interoperability procedures. Evidence: integration runbooks. NOT WRITTEN. The reconciliation procedure is blocked on the same absence M10.9.d found: the saga's only Destination implementation is an object store, so there is no LMS to reconcile against.
    • M10.18.e Publish source withdrawal/answer leak/BYOM/privacy/grade/ integrity/outage/recall/DR incident procedures. Evidence: incident runbooks. THE DR PROCEDURE IS PUBLISHED AND EVERY COMMAND IN IT WAS EXECUTED WHILE WRITING IT, which is what this item's spec asks for. docs/runbooks/metis-disaster-recovery.md: take a backup, restore into an isolated database, and verify by hashing 163 tables on both sides rather than counting rows, with the measured numbers from that run — 204 ms to dump, 1,981 ms to restore, all 163 digests identical, 18 tables non-empty. It also carries what could NOT be executed and why: no point-in-time recovery because archive_mode is off, no schedule, and no blob procedure because MinIO was not running. A runbook that described those anyway would be the failure this checklist keeps finding. It names its own owner as unassigned, because the estate has no owner registry and a runbook cannot invent an on-call rota. OPEN: source withdrawal, answer leak, BYOM, privacy, grade, integrity, outage and recall. The answer-leak procedure is the next one worth writing — M10.10.c and M10.16.a have the verified material for it, including a fix that is already live.
    • M10.18.f Execute cross-role game days, record time/gaps/unsafe ambiguity, update artifacts, and verify closure. Evidence: game-day results. STAYS OPEN BY ITS OWN SPEC: a cross-role game day needs named people. §S10's finding is the one to keep in view here — every game day in that phase reported a result without performing it, and kill -9 on PID 1 exits 0.
  • M10.19 Produce engineering-readiness evidence and reconcile the domain registry. Evidence: no planned/live contradiction. OPEN, AND ITS EVIDENCE LINE IS CURRENTLY FALSE: there IS a planned/live contradiction, measured — three registries answer for Metis and none agree. The readiness evidence this item would produce also already exists in the estate at 1,262 entries and contains zero for this checklist. Both halves are answered by artefacts that point somewhere else. <!-- SPEC 2026-09-11: the "domain registry" is libs/oshun/navigation/src/release-scope.ts (Metis scoped to V1.2) plus the M0.2 state artifact evidence/v1-workbenches/metis-registry-launch-state.json; domains.json is a library-area index, not a status registry. .d compares those two to what the apps/oshun/web middleware actually serves; .e's ratification is human (§2.5.9). -->
    • M10.19.a Define requirement/capability/host/audience/standard/discipline/ integration evidence index and freshness/owner rules. Evidence: schema. AN EVIDENCE INDEX EXISTS AT SCALE AND IS ABOUT A DIFFERENT CHECKLIST. oshun-readiness-evidence.json carries 1,262 ledger entries over 1,573 checklist cells citing 5,490 artifact paths — and ledgerEntriesForThisChecklist: 0. Not one entry is for this checklist, and the artefact mentions Metis zero times. Its own properties: 0 of 6 present, 1 of 7 validations performed, and commitShasThatAreASha: 0 — the commit identifiers it records are not shas. So the schema half of this item is answered by something that already exists; what is missing is an index that points at THIS work. Freshness and owner rules are unwritten, and the owner half runs into the same absent registry as M10.13.f, M10.14.e and M10.18.a. THE INDEX AND BOTH RULES EXIST NOW (2026-09-18), and the freshness rule turned out to be the one with teeth. scripts/v1-workbenches/generate-metis-evidence-index.mts → metis-evidence-index.json beside its own metis-evidence-index.schema.json, wired into both inventory chains. The schema is the deliverable this clause names: one entry per artefact, the seven dimensions required as KEYS, and a dimension the estate does not record carried as null rather than omitted — so an absence is counted instead of missing. FRESHNESS IS COMPUTABLE WITHOUT THE ARTEFACTS DATING THEMSELVES, which matters because M10.19.c measured that only 1 of 20 does. These artefacts CITE THEIR OWN SUBJECTS — the threat model names 44 source paths, the contract inventory 83, and every one of them exists — so an artefact is fresh when its last commit is at or after the newest commit touching any source it cites, and stale names the source that moved. MEASURED OVER THE DOMAIN'S WHOLE EVIDENCE TREE, 175 files citing 4,146 sources: 22 fresh, 115 stale, 38 that cite nothing that could date them. The worst is 22 days — metis-authoring-maturity.json against apps/metis/service/src/metis/schemas/scorm.py. Two of the 115 are stale because of THIS session: metis-credential-lifecycle.json and metis-package-metadata.json sit behind credential_service.py, which M10.22's correction touched today. That is the rule working, not noise. WHAT STALE MEANS, WRITTEN INTO THE ARTEFACT so nobody reads 115 as 115 wrong: a cited source changing for an unrelated reason flags its artefact too, so stale is "a reviewer has to look again" and the number is an upper bound on freshness — the direction a readiness gate should be wrong in. THE OWNER RULE IS STATED AND UNSATISFIED BY EVERY ROW: 175 of 175 carry owner: null, because this estate has no owner registry — the same gap M10.13.f, M10.14.e and M10.18.a record. The schema requires the field so the absence is counted; putting a team name nobody agreed to into 175 rows would have been the alternative. AND THE INDEX POINTS AT THIS WORK, which is what this clause said was missing: each entry lists the §M10 cells that cite it. 23 of 175 are cited by any cell, which is the same finding .c reports from the other side. Capability, audience and discipline are null for all 175 — nothing in the estate records them — and 50 artefacts name an interoperability standard in their own text. 8 tests, including the trap this rule could most easily have fallen into: an artefact that cites nothing answers null, never fresh. Calling an undatable artefact fresh would have rubber-stamped exactly the 38 nobody can check. STAYS UNCHECKED. The schema, both rules and the index are here, and the three dimensions no registry feeds — capability, audience, discipline — are still empty for every row, as is the owner. This clause asks for the index AND its rules; the rules now measure how far the estate is from satisfying them, which is not the same as satisfying them.
    • M10.19.b Collect current contract/migration/test/evaluation/accessibility/ performance/security/DR/runbook/approval artifacts by commit/hash. Evidence: readiness packet. NOT COLLECTED as a packet, though most of the artefacts now exist and are addressable: contracts and migrations in the OpenAPI document and alembic head 073, tests in the service suite, the accessibility gates from M10.13, the DR runbook from M10.18.e, and the 190-odd generated inventories under evidence/v1-workbenches. Collecting them BY COMMIT is the part with no mechanism — the existing ledger records commit identifiers that are not shas. THE PACKET EXISTS NOW (2026-09-18) AND THE MECHANISM IS THE POINT. scripts/v1-workbenches/generate-metis-readiness-packet.mts → evidence/v1-workbenches/metis-readiness-packet.json, wired into both inventory:v1-workbenches and verify:inventory:v1-workbenches, because a generator nobody runs is not a gate. MEASURED, FIRST RUN: the ten classes this clause names, 39 named paths (0 absent, 0 untracked) and 3 collections over 661 files — and 42 of 42 commit addresses are a 40-character sha. That number is the whole answer to "the part with no mechanism": the addresses come from git log -1 --format=%H, which cannot produce PENDING_COMMIT, and the packet counts the shape rather than checking the field is non-empty — which is the check that let the existing ledger's 1,262 placeholder rows through. EVERY ARTEFACT IS ADDRESSED TWICE: the sha256 of its bytes now, and the commit that last touched its path. The file lists come from git ls-files, so an uncommitted artefact is ABSENT from the packet rather than present with an invented history. ONE THING THE CLAUSE CANNOT HAVE, and the packet says so rather than pretending: a packet addressed by commit cannot be drift-checked by its commit addresses, because the commit that will carry an artefact does not exist while the packet naming it is written. --check therefore compares the CONTENT projection — paths, digests, counts, citations — and reports how many addresses have moved since, as information. A test asserts both halves of that: a changed digest moves the projection and a changed address does not. AND IT ALREADY ANSWERS HALF OF .c: each path is checked against this checklist's text, and 14 of 39 are cited by it. The other 25 are artefacts no cell's evidence sentence points at — the orphan direction .c asks for, measured here rather than asserted there. 8 tests in generate-metis-readiness-packet.test.mts, including the exact trap: the address set must not contain PENDING_COMMIT, and the citation count must be neither 0 (the checklist was not read) nor all of them (a count that proves nothing). STAYS UNCHECKED, AND THE DOUBT IS THE REASON RATHER THAN AN ABSENCE OF WORK. The ten classes are the clause's own, but WHICH artefacts belong to each class is a judgement this session made: the estate holds about 140 metis-* evidence files and the packet names 39 of them. The mechanism is complete and the enumeration is an opinion, so the box waits on somebody who owns the domain agreeing that these are the artefacts — which is also what .a's freshness and owner rules would settle.
    • M10.19.c Validate evidence pass/current/scope/environment/version and bidirectional mapping to every supported matrix/ledger row. Evidence: zero-orphan report. BLOCKED ON .b, and the orphan count is already known to be total rather than zero: with ledgerEntriesForThisChecklist: 0, every row of this checklist is an orphan of that ledger and every one of its 1,262 entries is an orphan of this checklist. A bidirectional mapping between two sets that share nothing is empty in both directions. THE REPORT EXISTS NOW (2026-09-18) AND THE ORPHAN COUNT IS NOT TOTAL. .b's packet gave this something to map against, so the mapping is over §M10's own cells rather than over a ledger for another programme: scripts/v1-workbenches/generate-metis-evidence-orphans.mts → evidence/v1-workbenches/metis-evidence-orphans.json, wired into both inventory chains. MEASURED, BOTH DIRECTIONS, 145 cells (48 checked). The figures below are the run that includes this record, because writing a record CHANGES THE MEASUREMENT — the first run said 61 cells named nothing and this cell was one of them; citing the artefacts here made it 60. An observation that moves what it observes is worth stating rather than hiding behind a timestamp. And every later edit to any §M10 record moves them again, which is why metis-evidence-orphans.json is the number and this paragraph is only the reading that prompted it — the counts here are the 2026-09-18 run and the shape of the finding, not a live total: - 60 cells name no artefact at all. Their evidence lines say "Evidence: schema" or "Evidence: reconciliation report" and point at nothing a reader can open. This is the orphan direction that matters, and it is 41% of the stage; - 20 of the packet's 36 artefacts are named by no cell — evidence nobody's evidence sentence points at; - 260 citations. 38 are not paths at all (routes, elided paths, build outputs, and identifiers that merely contain a slash), 251 resolve, 7 are a bare filename several files answer to, and 2 do not resolve; - 170 of the 251 resolve only after guessing a root, and 29 more only by searching the index for a unique basename. The cells cite tests/test_backpressure.py and services/record_reconciler.py, which are relative to apps/metis/service and its src/metis — so a tool reading this checklist cannot follow most of its citations without knowing where each record was written from. That is the finding under the finding: the traceability this clause asks for is defeated by the citation convention before any ledger is involved. THE FIVE DIMENSIONS .c NAMES ARE CHECKED AND 0 OF 20 ARTEFACTS CARRY ALL FIVE: version 19, environment 12, pass 10, scope 10, and currency 1. One artefact of twenty records when it was generated (metis-approval-subject.json, which carries a commit and a branch), so "is this evidence current?" is unanswerable for the other nineteen. The check is key presence and is labelled as a lower bound — it says the field is there to be read, not that its value is meaningful. TWO CORRECTIONS THIS SESSION MADE TO ITS OWN MEASUREMENTS, both worth keeping because each was wrong in a direction that flattered the report: - the first run reported 30 dead citations. A citation pattern that matches a slash matches /api/lti/jwks, xapi-2.0/statement/void and .../governance too. Classified now, and the real figure is 2; - and those 2 are not broken links either. Both are deliberate mentions of a file RENAMED to tests/test_backpressure.py, and each cell names the current path in the same breath. So the stale-link count in §M10 is ZERO, which a generator cannot tell and a reader can — recorded in the artefact's own method so nobody reads the JSON and concludes otherwise. AND A THIRD CORRECTION, ON THIS RECORD ITSELF. The first draft of the paragraph above illustrated the root problem with two invented example paths, in backticks, and the next run counted them: 4 dead citations, two of them written by the sentence explaining the other two. Naming them here in backticks re-created them a second time, so this sentence does not. An illustrative path in a checklist is a citation to anything that reads the checklist; the examples above are real files now. 9 tests in generate-metis-evidence-orphans.test.mts, asserting the classifier by value rather than by count. STAYS UNCHECKED: the report is the deliverable this clause asks for and it reports a non-zero orphan count in both directions. "Zero-orphan" is the state, not the artefact — closing this needs the 61 cells to cite their evidence and the 20 unnamed artefacts to be cited or dropped.
    • M10.19.d Compare implemented/exposed/deployed state to domain registry planned/internal/preview/live flags by host/audience/region. Evidence: reconciliation report. THE COMPARISON IS DONE AND THE ANSWER IS THAT NOTHING AGREES. evidence/v1-workbenches/metis-registry-launch-state.json, which M0.2 generates and this session regenerated: registriesAnsweringForMetis: 3, registriesThatAgree: 0. Three registries answer for Metis and no two of them say the same thing. lifecycleStatesRequired: 7 against lifecycleStatesTheShellCanRepresent: 1 — the shell has one flag where the lifecycle has seven states, so most of the distinctions this item asks to be compared BY cannot be expressed at all. propertiesMeasured: 5, propertiesPresent: 0. What IS clean is the route half: 11 Oshun-hosted Metis routes, all 11 covered, routesNotCovered: [], and the proxy is the middleware, calls the resolver, and refuses before auth. So the reconciliation exists and reports a contradiction. Closing this needs the contradiction resolved, not another comparison.
    • M10.19.e Block contradictory shell/routes/capabilities/registry and ratify exact registry change/rollback. Evidence: approved decision. THE CONTRADICTION IS IDENTIFIED AND CANNOT BE RATIFIED HERE. .d shows three registries disagreeing and a shell that can express one of seven lifecycle states; blocking on that is correct and is what this item asks for. The ratification half is human by this stage's own spec, and the same authority gap recurs — M10.14.a's budget type refuses to exist without a ratifier for the same reason.
  • M10.20 Obtain named education, accessibility, privacy/security, integrity, engineering, and product approvals. Evidence: human signoff. OPEN BY DESIGN — the evidence line is human signoff and no agent may supply it. The packet half is built: .b freezes and hashes the subject, and names the five fields this repository cannot answer rather than defaulting them. What an approver would be signing is now identifiable; who may sign it is not, because the estate has no approver registry. <!-- SPEC 2026-09-11: §2.5.9 — packet only. Every child stays open until a named person acts; .a and .b are the packet. --> 2026-09-18: open by design, as its note says: the packet half is built and the evidence is human sign-off (rule 9). blocked:human
    • M10.20.a Define each named approver’s authority/scope/independence, required packet, decision options/conditions/expiry, and invalidation. Evidence: approval matrix. OPEN ON THE SAME MISSING REGISTRY THIS PHASE KEEPS MEETING. Defining "each NAMED approver's authority" needs a register of approvers, and this estate has none — the same absence recorded at M10.13.f for accessibility, M10.14.e for performance and M10.18.a for runbook owners. What the estate does have, and what an approval matrix should reuse rather than reinvent, is M7.13's human-signoff contract and M7.8's conflict-of-interest engine, both closed, plus the independence rule M10.12.c's appeal engine already states by actor digest. The required-packet half is answerable and .b answers it.
    • M10.20.b Freeze exact commit/build/config/environment/host/audience/ standards/integration/capability subject manifest. Evidence: subject hash. THE COMPUTABLE HALF IS FROZEN AND HASHED. scripts/export_approval_subject.py (pnpm run metis:approval:subject) writes evidence/v1-workbenches/metis-approval-subject.json: the commit, the branch, whether the working tree was clean, the migration head, the OpenAPI digest, 882 operations, 3,685 published schemas, the three applications, and a subject_sha256 over all of it. An approval can now name a digest, and a later reader can tell whether what is in front of them is what was approved. FIVE FIELDS THIS ITEM ASKS FOR ARE NAMED AS NOT COMPUTABLE RATHER THAN DEFAULTED, because a defaulted field in a frozen subject is the one that makes the freeze a lie: there is no build identifier, no captured runtime configuration, no addressable deployed environment, no agreed host/audience (three registries disagree — M10.19.d), and no capability manifest scoped to this checklist. The manifest also records working_tree_clean, because a subject hash taken over a dirty tree names a state nobody can check out.
    • M10.20.c Obtain separate education/pedagogy/psychometrics, accessibility, privacy/security, and integrity approvals with conditions. Evidence: governance decisions. STAYS OPEN BY THIS STAGE'S OWN SPEC: a named person has to act. Recording an approval here would be the exact fabrication this checklist exists to catch. What CAN be said is what each approver would be reading: accessibility has gates and two open ceilings (M10.13), integrity has a measured self-approval hole (M10.12.e), and psychometrics has a fairness analysis proven sound in both directions (M10.11.b and .c).
    • M10.20.d Obtain engineering/operations/interoperability and per-host/ audience product launch decisions. Evidence: launch decisions. STAYS OPEN, and per-host/audience decisions are additionally unanswerable while three registries disagree about what Metis is (M10.19.d).
    • M10.20.e Record approve/reject/conditional/request-changes immutably and revalidate on material evidence/scope/subject change. Evidence: approval lifecycle. OPEN. The revalidation trigger is now mechanical, which is the part worth having early: .b's subject_sha256 changes whenever the commit, the migration head or the published contract does, so "material subject change" is a digest comparison rather than a judgement. The immutable record itself needs the store and the approver registry that .a is blocked on.
  • M10.21 Run full-journey Playwright for lecture/agentic-media creation, correctness critique, accessible renditions, localization, and learner playback. Evidence: real sandbox/render adapters or explicitly hermetic dependencies and exact provenance assertions. NOT WALKED. This is the same unwalked UI half M10.10 and M10.12 carry, with an added dependency: the render stage needs Isis and Euterpe through their canonical adapters, and this stage's spec is explicit that hermetic never means skipping the render. The lecture delivery path itself is real and was read closely under M10.13.e, which found the image alternative it drops. <!-- SPEC 2026-09-11: "explicitly hermetic" means the real service code path with the provider stubbed at its HTTP boundary (the seam the M6/M9 lecture tests already use), a persistent fixture badge in the evidence, and the stub named in .a's manifest; it never means skipping the render stage. Isis and Euterpe dependencies go through their canonical adapters, never private imports (X.5 and X.17 later depend on this). -->
    • M10.21.a Seed exact lesson/objectives/sources/data/presenter consent/ locales/targets and live or explicitly hermetic render/Isis/Euterpe dependencies. Evidence: fixture manifest.
    • M10.21.b Author script/scene plan/spec/code/media, execute sandbox render with progress/cost, and retain all stage/source/tool provenance. Evidence: creation journey.
    • M10.21.c Run visual/layout/source/math-code/correctness/accessibility/ technical critics, reject one iteration, revise, and require educator acceptance. Evidence: critique journey.
    • M10.21.d Author/review captions/transcript/description/translation/ pronunciation/renditions and surgical replacement with stable unaffected segments. Evidence: localization/replacement journey.
    • M10.21.e Assemble package, clear gates, publish, play via real learner renderer across locale/device/bandwidth/accessibility. Evidence: playback journey.
    • M10.21.f Assert budget/policy/render/verification stop, no self-publish, source/consent withdrawal, tenant/role boundaries, and exact lineage. Evidence: E2E trace.
  • M10.22 Run full-journey credential tests from pathway authoring through evidence review, Open Badges issuance/verification, expiry, revocation, replacement, and audit. Evidence: issuer, privacy, and stale-course boundaries. NOT WALKED. The clause above was spliced in half by the record that followed it — "from pathway authoring through" ran straight into a paragraph and the rest of the sentence appeared after it. Rejoined 2026-09-18. AND THE BLOCKER THAT RECORD NAMED IS NOT WHAT HAPPENS. It said the credential service keeps revocations in process dictionaries "so a restart resurrects a revoked badge". The dictionaries are real — nothing in services/credential_service.py survives a restart — but the direction was measured on 2026-09-18 by driving it, and it is the opposite: after the class attributes are cleared, the same credential and the same document verify as not_found, valid=False. verify_credential hands anything the registry does not hold to _verify_unregistered_document, which returns valid=False whatever else passes — and the canonical hash and the HMAC proof DO still pass, because the signing secret is configuration and survives the restart. That is why the claim was plausible and it is still wrong. THE REAL DEFECT IS THE SAFE DIRECTION AND IS STILL A DEFECT: every issued badge becomes unverifiable after a restart, not just the revoked ones, so a learner's legitimate badge stops verifying. A journey that restarts the process would find that, which is what this item should walk. Four tests pin the direction now (tests/test_credential_service.py::CredentialRestartDirectionTest), including the one that matters: a revoked badge must not verify as valid when the issuer cannot account for it. A docstring cannot enforce that, and the one that tried had it backwards. <!-- SPEC 2026-09-11: .d's clock advance requires the credential service's clock to be injectable; if it is not, that injection is the first sub-step and is recorded in the evidence. Signatures are real Ed25519 over Open Badges 3.0 assertions (services/open_badges_service.py); .c's verification runs the verifier against the published JSON, never against the issuing object in memory. -->
    • M10.22.a Seed credential issuer/roles/keys/template/objectives/course/ assessment/evidence policies and synthetic recipient. Evidence: fixture.
    • M10.22.b Author/approve pathway and complete locked criteria with exact learning/assessment/human evidence; test missing/stale/conflicting cells. Evidence: eligibility journey.
    • M10.22.c Review independently, issue/sign/deliver Open Badge, and verify signature/issuer/recipient/evidence/status through public/private views. Evidence: issuance journey.
    • M10.22.d Advance fake clock through warning/expiry/renew and execute revoke/replacement/supersession/status notification. Evidence: lifecycle journey.
    • M10.22.e Assert author/agent/course completion cannot issue, protected learner evidence is minimized, stale course invalidates request, and audit remains immutable. Evidence: boundary E2E.
  • M10.23 Run known-value and calibrated suites for atomic claims, citation OPEN WITH .f DONE. The coverage gate this stage's evidence line asks for exists and passes, and it found the one asymmetry worth knowing: five of the six enumerated verifiers are scored against a versioned corpus and cultural review is not. Building the corpora .a–.e ask for is the remaining work, and .c's calibrated panel sets additionally need the human judgements no agent can supply. sufficiency, faithfulness, contradiction, math, code, worked solutions, evaluator panels, media critics, and discipline-specific verification. Evidence: no generic "content quality" test substitutes for an enumerated verifier. <!-- SPEC 2026-09-11: §2.5.4 and §2.5.5. The enumerated verifiers exist as services (claim_decomposition, grounding_gate, mathematics_verification with pinned sympy/pint/mpmath, code_verification, evaluator_panel, cultural_review); .f's matrix names each one with its fixture file, and a verifier whose only coverage is a generic content-quality test fails the gate. -->
    • M10.23.a Build versioned labeled corpora and coverage rows for atomic claim segmentation/linking and support/sufficiency/faithfulness/ contradiction/conflict. Evidence: grounding suites.
    • M10.23.b Build executable math/code/worked-solution goldens for correct/ wrong/partial/alternative/unsafe/flaky/timeout/unsupported cases. Evidence: deterministic suites.
    • M10.23.c Build calibrated panel/critic gold sets with evaluator dependency/correlation, confidence/abstention/manual bands, champion/ challenger, and subgroup reports. Evidence: panel reports.
    • M10.23.d Build diagram/chart/lecture/media visual/layout/accessibility/ source/correctness critic fixtures and renderer/version goldens. Evidence: media critic suites.
    • M10.23.e Build per-discipline proof/lab/data/source/language/code specialized verification fixtures and standards export survival tests. Evidence: discipline suites.
    • M10.23.f Publish a verifier-to-fixture/report/threshold/owner matrix and fail any enumerated verifier covered only by generic content quality. Evidence: dedicated coverage gate. THE GATE EXISTS AND ALL SIX CLEAR THE BAR THIS ITEM SETS. tests/test_verifier_coverage_matrix.py, 7 tests over the six verifiers this stage enumerates by name. Each has a service of its own, each has at least two dedicated test files, no generic content-quality suite exists to stand in for any of them, and every corpus a verifier names resolves on disk — a corpus path that does not resolve is a verifier scored against nothing, and it would fail only when somebody ran it. THE MATRIX, measured by reading what each verifier's own modules load rather than by matching filenames: - claim_decomposition → m5_19_atomic_claim_corpus.json - grounding_gate → m5_20_grounding_gate_corpus.json - mathematics_verification→ m5_21_executable_math_corpus.json - code_verification → m5_22_sandboxed_code_corpus.json - evaluator_panel → m5_24_evaluator_panel_corpus.json - cultural_review → NONE FIVE ARE CALIBRATED AND ONE IS MERELY TESTED. The five ship an *_evaluation.py that scores the corpus and reports matched against total per dimension. cultural_review has six dedicated test files and a fixtures module, so it passes this item's "not generic" bar outright — but it has no corpus and no evaluation module, so nothing says how much of its own space those six suites reach. Held as a ceiling that may shrink and must never grow. THRESHOLDS AND OWNERS ARE NOT EXPRESSIBLE and the matrix says so rather than inventing them: the estate has no owner registry, the same absence recorded at M10.13.f, M10.14.e, M10.18.a and M10.20.a.

Phase Y — Yemaya AAA production, study, and interactive-case authoring#

Yemaya is an existing AAA production suite, not a greenfield V1 route. This phase adds coherent study-material and interactive-case authoring to apps/yemaya/studio-web and its desktop host while completing and governing the existing project, script/storyboard, character/world, asset, production, engine-integration, collaboration, rendering, agent, marketplace, canon, case, and human-override systems. Consumer study experiences remain distinct from authoring, but the existing AAA suite is in scope for explicit feature reconciliation rather than being hidden behind the word "integration."

Y0 — Boundary, host, inventory, and production decisions#

  • Y0.1 Inventory the Studio web/desktop route trees and classify live, fixture, documentation, duplicate, and missing surfaces. Evidence: route-to-capability matrix. <!-- partial: deterministic route inventory (a) done in evidence/v1-workbenches/yemaya-route-inventory.json. b-f (runtime live/fixture classification, capability mapping, divergence report, dispositions, coverage gate) are runtime/design/human and stay unchecked. -->
    • Y0.1.a Enumerate every web/Electron route, nested layout, navigation/ deep-link entry, IPC/API call, feature flag, consumer, owner, and test. Evidence: deterministic route inventory. <!-- evidence: .routes — 39 routes (33 studio-web React-Router w/ RouteSuspense pageName + nested Layout, 6 studio-desktop Electron renderer), each with path/element/surface/dynamic/test. Per-route file/API resolves for desktop; web pages lazy-load via the ./lib/lazy barrel so the API surface is captured at app level (.appLevelFootprints: 8 web /api/* endpoints, 18 desktop IPC channels). generate-yemaya-route-inventory.mjs + 6 tests, wired into inventory/verify. -->
    • Y0.1.b Exercise each surface against backing services and label live, partial/degraded, fixture/demo, documentation/mock, unreachable, or missing with proof. Evidence: runtime classification.
    • Y0.1.c Map route operations/fields to atomic capabilities, audience/ role, project type, owner service/package, data store, and desktop/web requirement. Evidence: route-capability matrix.
    • Y0.1.d Identify duplicate mutation/read facades, stale aliases, undocumented live routes, orphan APIs, and asserted capabilities with no route. Evidence: divergence report.
    • Y0.1.e Assign reuse/refactor/redirect/retire/net-new/exclude dispositions with owner, compatibility window, and test. Evidence: approved ledger.
    • Y0.1.f Add route/inventory drift checks and honest-state tests. Evidence: route coverage gate.
  • Y0.2 Define the shared renderer boundary between Studio web and Electron desktop. Evidence: no business rule exists only in a renderer component.
    • Y0.2.a Inventory renderer/shared/domain/service/main-process code, imports, state, validation, permissions, API/IPC handlers, and duplicated business rules. Evidence: dependency map. <!-- evidence: yemaya-desktop-partition.json — 62 studio-desktop files partitioned by process (main 36, renderer 22, preload 1, shared 2, score-editor 1); IPC surface = ~297 ipcMain handler channels ↔ ~102 ipcRenderer call channels via 1 contextBridge namespace; renderer→main leakage = 0 (clean boundary — no renderer file imports electron or a main-process module directly). generate-yemaya-desktop-partition.mjs + 5 tests. Duplicated-business-rules detection is the Y0.2.b/e semantic half. -->
    • Y0.2.b Define renderer-agnostic domain contracts/use cases/state machines/validation/authorization and shared view-model seams. Evidence: boundary interfaces.
    • Y0.2.c Define web transport and Electron preload IPC adapters with typed command/query/events, structured errors, cancellation, and versioning. Evidence: adapter contracts.
    • Y0.2.d Keep native filesystem/device/DCC/render capabilities behind main- process service ports; renderer receives no Node/Electron privilege. Evidence: security boundary tests.
    • Y0.2.e Move or disposition renderer-only validation/workflow/policy/ canon/release rules into canonical domain owners. Evidence: rule ledger.
    • Y0.2.f Run identical domain contract/consumer tests through web and desktop adapters and fail forbidden imports. Evidence: parity gate.
  • Y0.3 Define which native capabilities require desktop: local files, DCC bridge, large media, background render, color management, and device capture. Evidence: capability matrix and browser fallback.
    • Y0.3.a Define versioned native-capability registry with host/OS/device/ permission/security/resource/offline requirements and audience. Evidence: capability schema.
    • Y0.3.b Specify local file/directory/project package operations, path/ symlink/permission/sandbox rules, watchers, and web upload/download/ file-system-access fallback. Evidence: file matrix.
    • Y0.3.c Specify DCC/engine bridge discovery/session/command/transfer/ version/auth behavior and governed Bellona handoff; web fallback is inspect/deep-link/export only. Evidence: bridge matrix.
    • Y0.3.d Specify large media streaming/proxy/chunk/cache/memory/disk/ hardware decode behavior and browser limits/degraded states. Evidence: media matrix.
    • Y0.3.e Specify background render/encode/job resume/notifications/power, color profile/display/OCIO, and capture device permission/calibration/ consent behavior. Evidence: render/color/capture matrices.
    • Y0.3.f Define capability negotiation and honest supported/degraded/ unavailable/install-desktop/open-desktop states with no data loss. Evidence: fallback tests.
  • Y0.4 Define the authoring/consumer boundary for study-workspace. Evidence: authoring never writes consumer read models directly.
    • Y0.4.a Inventory authoring/consumer routes, packages, schemas, stores, caches/indexes, mutations, events, and current direct writes. Evidence: boundary inventory.
    • Y0.4.b Define canonical study project/content/revision/build/release mutation owner and separate consumer delivery/progress/session read owners. Evidence: ownership matrix.
    • Y0.4.c Define versioned compiled study bundle/publication event consumed to build delivery read models and no shared mutable content rows. Evidence: handoff contracts.
    • Y0.4.d Define author preview through the real consumer renderer using synthetic identity and nonoperational events. Evidence: preview boundary.
    • Y0.4.e Remove/prohibit direct authoring writes to consumer database/ index/cache/progress and direct consumer edits to authoring revisions. Evidence: dependency/schema guards.
    • Y0.4.f Test publication/supersession/recall/read-back, two-tenant and failure/replay behavior. Evidence: boundary integration suite.
  • Y0.5 Define the study and interactive-case content tracks plus their shared project, source, asset, review, and release spine. Evidence: domain context diagram.
    • Y0.5.a Define track-specific goals, authors/audiences, content entities, authoring journeys, validation, outputs, consumers, and explicit exclusions. Evidence: track charters.
    • Y0.5.b Define shared project/workspace/member/role/revision/maturity/ schedule/budget/audit contracts without collapsing track semantics. Evidence: shared project map.
    • Y0.5.c Define shared governed source/claim/citation/rights and asset/ variant/provenance/continuity/canon relationships. Evidence: source/asset spine.
    • Y0.5.d Define shared comment/review/gate/decision/approval/exception/ localization workflows with track-specific templates. Evidence: review spine.
    • Y0.5.e Define track-specific compiler/bundle/target/publication/catalog/ consumer handoffs atop shared release candidate/execute/verify/recall. Evidence: release spine.
    • Y0.5.f Draw ownership, context, trust, data, event, and cross-domain boundaries plus valid reuse/forbidden coupling. Evidence: ratified context diagram.
  • Y0.6 Map all libs/yemaya/* packages to the two tracks, shared production systems, other Yemaya products, or out-of-scope. Evidence: complete package ledger.
    • Y0.6.a Enumerate every package/export/source/test/dependency/consumer/ build target/owner and detect nested/untracked/dead packages. Evidence: <!-- evidence: .domains.yemaya.packages — 59 libs/yemaya/* packages enumerated from the BASE.3 graph (exports/inbound/outbound/disposition) + BASE.1 tests; dead=retire-candidate disposition. evidence/v1-workbenches/domain-package-inventories.json (254 pkgs; per-package exports/consumers/deps/disposition/tests + capability catalog with exact source locations). generate-domain-package-inventories.mjs + 6 tests. --> package inventory.
    • Y0.6.b Extract public entities/operations/events/jobs/adapters/UI/ algorithms and link exact source locations. Evidence: capability <!-- evidence: .domains.yemaya.packages[].capabilityCatalog — atomic capabilities (entity/operation/event/job/adapter) extracted per package, each linked to its exact sourceFile + sourceRef (891 caps). UI is captured by the component inventories; "algorithms" is not a discrete capability kind. --> catalog.
    • Y0.6.c Map each capability to study track, interactive-case track, shared production, AAA/organization/marketplace/developer/consumer product, or explicit out-of-scope. Evidence: classification ledger.
    • Y0.6.d Assign canonical owner, reuse/refactor/move/merge/retire/exclude/ net-new disposition and identify duplicate/shadow shapes. Evidence: disposition review.
    • Y0.6.e Map every retained package feature bidirectionally to route/API/ contract/ledger/evidence and add zero-orphan drift gate. Evidence: complete package ledger.
  • Y0.7 Reconcile case contracts, compiler, director, engine, verifier, eval, canon, localization, and pipeline ownership. Evidence: dependency direction ADR.
    • Y0.7.a Inventory contracts/packages/services/stores/jobs/CLIs/routes/ events/tests for each named subsystem and current dependency graph. Evidence: subsystem inventory. <!-- evidence: yemaya-case-inventory.json — the 11 interactive-case pipeline subsystems (case-assets/compiler/contracts/director/engine/eval/localization/ pipeline/suspects/verifier/writers-room; 585 public symbols, 111 code files, 37 tests) each with exports/inbound/outbound/disposition, plus the current intra-pipeline dependency graph (21 case-→case- edges, ACYCLIC — case-contracts is the foundation everyone depends on). The dependency-direction ADR is Y0.7's decision. generate-yemaya-case-inventory.mjs + 3 tests. -->
    • Y0.7.b Define canonical aggregates/operations/outputs and mutation/read ownership for case authoring, compilation, direction/runtime, verification evaluation, canon, localization, and pipeline. Evidence: ownership map.
    • Y0.7.c Define permitted dependency direction, public ports/contracts/ events, versioning, idempotency, failure, and no private-store imports. Evidence: target graph.
    • Y0.7.d Reconcile duplicate case/canon/localization/evaluation shapes and migrate consumers with compatibility/retirement plan. Evidence: contract disposition.
    • Y0.7.e Define compiler output as immutable input to engine/director and verifier/eval evidence as nonauthoritative inputs to review. Evidence: authority invariants.
    • Y0.7.f Ratify ADR with cycle/forbidden-import/drift tests and rejected alternatives. Evidence: approved dependency-direction ADR.
  • Y0.8 Reconcile project, asset-library, dailies, collaboration, rendering, orchestration, budget, safety, and human-override systems. Evidence: no planned shadow implementation.
    • Y0.8.a Inventory existing services/packages/routes/stores/jobs/UIs/events/ deployments for each production system with live/partial/fixture status. Evidence: system inventory.
    • Y0.8.b Map study/case requirements to exact reusable project/asset/ dailies/collaboration/render/orchestrator/budget/safety/override operations and contracts. Evidence: reuse matrix.
    • Y0.8.c Identify gaps, duplicate proposed stores/queues/workflows/ dashboards, conflicting state vocabularies, and authority overlap. Evidence: gap/duplication report.
    • Y0.8.d Assign reuse/facade/extend/refactor/migrate/retire/net-new disposition with canonical owner and dependencies. Evidence: disposition ledger.
    • Y0.8.e Define governed cross-system identity/version/tenant/project/ lineage/audit/transaction/event/recovery handoffs. Evidence: seam contracts.
    • Y0.8.f Add architecture/dependency/schema inventory guards proving no shadow canonical implementation. Evidence: ownership gate.
  • Y0.9 Define author, researcher, writer, designer, editor, canon reviewer, solver, localization reviewer, producer, publisher, and administrator roles. Evidence: permission matrix.
    • Y0.9.a Define each role’s purpose/qualifications, track/project/ organization scope, eligible actor, delegation, lifecycle, and conflicts. Evidence: role definitions.
    • Y0.9.b Map project/source/story/case/canon/solution/asset/script/ production/localization/review/release/admin operations per role. Evidence: operation matrix.
    • Y0.9.c Map unpublished story/protected source/spoiler/solution/ participant/biometric/rights/budget/security fields and actions at property level. Evidence: field matrix.
    • Y0.9.d Define author/reviewer/solver/canon/publisher separation, quorum, conflict, substitute, embargo, blind-review, and break-glass rules. Evidence: duty policy.
    • Y0.9.e Define administrator bounds for membership/configuration versus content/canon/solution/approval/publication authority. Evidence: admin boundary.
    • Y0.9.f Test route/object/property/function, combined roles, role change, direct IDs, embargo/spoiler, and two tenants. Evidence: permission suite.
  • Y0.10 Define maturity models for study projects and cases separately. Evidence: state vocabularies do not overload project production status.
    • Y0.10.a Inventory current study/case/content/review/production/render/ release/consumer states, labels, enums, owners, and collisions. Evidence: vocabulary inventory. <!-- yemaya-state-vocabulary.json: derived from the generated Prisma enums (libs/yemaya/database/src/generated/client/enums.ts). 21 typed status/type enums declaring 115 labels (89 distinct), each with owner = the generated Prisma client. COLLISION detection (the Y0.10.a "collisions" dimension): 21 label values appear in >1 enum with unrelated semantics — ACTIVE×3 (ProjectStatus/SubscriptionStatus/UserStatus), DRAFT×3 (AssetStatus/ InvoiceStatus/ProjectStatus), ADMIN×3, FREE×3, OTHER×3 — a real vocabulary hazard surfaced. Lifecycle-stage coverage (enum→stage by name keyword): typed vocabulary EXISTS for content/production/release/consumer; NO typed state vocabulary for study/case/review/render (verified twice across the whole Yemaya surface — an honest architectural gap, reported not hidden). The study-content state MACHINE (states/transitions/prerequisites, Y0.10.b) and later Y0.10.* DEFINE new structure and stay unchecked. generate-yemaya-state-vocabulary.mjs + 6 tests, wired into inventory/verify + CI. -->
    • Y0.10.b Define study content maturity states/transitions/prerequisites/ roles/effects independently from project schedule/asset production. Evidence: study state machine.
    • Y0.10.c Define interactive-case concept/writing/canon/solvability/ production/review/release maturity independently from runtime play state. Evidence: case state machine.
    • Y0.10.d Define shared project production status and relation/projection to each track’s maturity without one mutating the other. Evidence: mapping.
    • Y0.10.e Define transition concurrency, revisions, review invalidation, supersession/archive/quarantine/restore, audit, and server enforcement. Evidence: transition matrix tests.
    • Y0.10.f Detect API/UI/analytics/export vocabulary collisions and prove no overloaded enum/default mapping. Evidence: collision review.
  • Y0.11 Threat-model source injection, malicious project files, spoiler leakage, canon corruption, unfair or harmful case content, unsafe media, agent runaway, desktop IPC, local filesystem, and supply chain. Evidence: mitigations map to tests.
    • Y0.11.a Map actors/assets/trust boundaries/data flows/entry points/ providers/tools/hosts/integrations/high-impact actions. Evidence: threat model context.
    • Y0.11.b Model prompt/source/retrieval/import injection and poisoned/ retracted/false/protected sources across authoring, agents, compilers, critics, and generated media. Evidence: source threats.
    • Y0.11.c Model spoilers/solutions/unpublished story/canon/embargo leakage via UI/API/search/export/logs/caches/collaboration/local files/consumer builds. Evidence: confidentiality threats.
    • Y0.11.d Model canon tamper/conflicts, unfair/unsolvable/manipulative/ harmful case content, unsafe media/identity/voice/minor/rights use, and false automated approval. Evidence: content threats.
    • Y0.11.e Model agent budget/tool/loop/runaway/injection/approval confusion, hidden side effects, and cross-project/tenant data access. Evidence: agent threats.
    • Y0.11.f Model Electron IPC/preload/navigation/deep-link/update/plugin/ local filesystem/path/symlink/credential/device/DCC/renderer compromise. Evidence: desktop threats.
    • Y0.11.g Model package/model/node/renderer/DCC/engine/marketplace/ dependency/artifact/build/distribution supply-chain tamper. Evidence: supply-chain threats.
    • Y0.11.h Assign prevention/detection/response/recovery, owner, residual risk, test, monitoring, incident runbook, and launch gate per threat. Evidence: mitigation traceability map.
  • Y0.12 Define data classification for unpublished story material, protected sources, case solutions, participant data, assets, and local caches. Evidence: privacy/security decision.
    • Y0.12.a Inventory entities/fields/files/blobs/prompts/logs/events/search/ exports/backups/local caches for each named data category. Evidence: data inventory. <!-- yemaya-data-inventory.json: derived from the generated Prisma client (libs/yemaya/database). 38 named entities / 406 scalar fields. Of Y0.12.a's 11 named data forms, 6 are entity-modeled (entities=38 models; fields=406; files/blobs=Asset storage columns; logs=AuditLog/Activity/UsageRecord/ AssetUsage; events=Activity/Notification/Webhook/WebhookDelivery) and 5 are recorded HONESTLY as infrastructure not modeled as Prisma entities (prompts= request/generation-shaped, search=index, exports=export jobs, backups=DR, local-caches=client/edge cache) — not fabricated to look complete. Each entity carries its field count. generate-yemaya-data-inventory.mjs + 5 tests, wired into inventory/verify + CI. Classification labels + per-category owner/ purpose/access/isolation/encryption/residency (Y0.12.b) and later Y0.12.* are design/human governance and stay unchecked. -->
    • Y0.12.b Define classification labels and per-category owner, purpose, access roles, tenant/project isolation, encryption/key, residency, retention/deletion/legal hold, export, and audit. Evidence: policy map.
    • Y0.12.c Define spoiler/embargo/need-to-know compartments and protected source/solution field-level projections with nonenumerable errors. Evidence: confidentiality matrix.
    • Y0.12.d Define participant/performer/voice/likeness/biometric/consent/ minor and review-comment/telemetry privacy handling. Evidence: participant policy.
    • Y0.12.e Define asset derivative/proxy/watermark/provenance and desktop local cache encryption/OS storage/offline/eviction/logout/device-loss behavior. Evidence: asset/cache policy.
    • Y0.12.f Threat-test UI/API/search/logs/notifications/exports/IPC/local files/backups/support/analytics and obtain privacy/security decision. Evidence: classification tests and approval.
  • Y0.13 Define web/desktop parity, version-skew, update, offline, and project-lock policies. Evidence: compatibility ADR.
    • Y0.13.a Define versioned web/desktop/shared-contract/native-capability/ server compatibility matrix and support windows. Evidence: matrix.
    • Y0.13.b Classify required parity versus desktop-only/web-fallback/ unsupported behaviors, UI states, data semantics, and test proof. Evidence: parity policy.
    • Y0.13.c Define startup/runtime version negotiation, min/max/expiring/ unsupported behavior, protocol/schema migration, and no silent skew. Evidence: skew tests.
    • Y0.13.d Define signed update discovery/download/staging/migration/ restart/rollback/forced-security-update and interrupted update behavior. Evidence: update policy.
    • Y0.13.e Define offline cache/project availability, read/write/queued commands, conflict/reconnect/revalidation, secret expiry, and honest state. Evidence: offline policy.
    • Y0.13.f Define project/revision/entity/asset/render locks, lease/ heartbeat/fencing/expiry/takeover/read-only, desktop crash, and split- brain behavior. Evidence: lock policy.
    • Y0.13.g Ratify compatibility ADR and matrix-generated web/desktop/skew/ update/offline/lock tests. Evidence: approved ADR.
  • Y0.14 Define success measures for authoring throughput, solvability, canon quality, source grounding, collaboration, render reliability, and release correctness. Evidence: metric specification.
    • Y0.14.a Define authoring flow measures by track/stage with denominator, wait/rework/block/abandonment and quality/safety countermetrics. Evidence: throughput metrics.
    • Y0.14.b Define solvability/fairness measures for clues/evidence/paths/ ambiguity/hints/solver outcomes with calibrated sample/uncertainty and no “difficulty equals quality” incentive. Evidence: solvability metrics.
    • Y0.14.c Define canon/source measures for conflicts/drift/coverage/ authority/currency/citation/faithfulness/corrections and review burden. Evidence: trust metrics.
    • Y0.14.d Define collaboration measures for review cycle/response/ conflicts/merge/rework/participation with privacy and no worker surveillance incentive. Evidence: collaboration metrics.
    • Y0.14.e Define render/job reliability/SLO/resource/cost/quality and release planned-written-verified/partial/rollback/recall metrics. Evidence: production metrics.
    • Y0.14.f Specify source/formula/unit/window/dimensions/privacy/freshness/ owner/target/guardrail/limitations and ratify against harmful shortcuts. Evidence: approved metric catalog.
  • Y0.15 Author and ratify V1/YEMAYA_AUTHORING_WORKBENCH_DESIGN_2026-07-23.md with current-state, web/desktop boundary, study/case tracks, data, APIs, agents, production, trust, migration, rollout, and decision records. Evidence: approved doc and resolved review comments.
    • Y0.15.a Establish scope/terminology/requirements traceability/source anchors/owners/reviewers/decision log/evidence format. Evidence: outline.
    • Y0.15.b Document current routes/apps/packages/contracts/stores/jobs/ agents/CLIs/integrations/deployments and live/partial/fixture/proposed truth. Evidence: current-state review.
    • Y0.15.c Document web/desktop/renderer/native boundaries, audiences/roles/ product tracks/project profiles, data/contracts/APIs/events/jobs, and cross-domain ownership. Evidence: architecture sections.
    • Y0.15.d Document study/case authoring/canon/sources/assets/production/ agents/compiler/runtime/UI/responsive/accessibility/localization/offline journeys. Evidence: workflow/experience sections.
    • Y0.15.e Document trust/safety/rights/privacy/security/supply chain, verification/evaluation/observability/operations/DR. Evidence: trust sections.
    • Y0.15.f Document migration/compatibility/rollout/flags/deprecation/ rollback/readiness and explicit ADRs/options. Evidence: rollout sections.
    • Y0.15.g Reconcile code/inventory/deep dives, resolve every comment, and obtain named domain/security/privacy/accessibility/engineering/product ratification. Evidence: approved design.
  • Y0.16 Reconcile the domain feature inventory, all web/desktop pages, 44+ API route groups, every libs/yemaya/* package, worker, CLI, walkthrough, and external integration at atomic-feature level. Evidence: the inventory explicitly covers projects, assets, scripts/storyboards, characters/ worlds, production management, collaboration, autonomous pipelines, engines/DCCs, gameplay/NPCs, virtual production/XR, marketplaces/search, organizations/enterprise, community, study, and cases.
    • Y0.16.a Build deterministic atomic inventories from feature docs, every page/route/operation/schema/package/worker/job/CLI/walkthrough/ integration/test with source anchors. Evidence: source inventories. <!-- yemaya-code-census.json: 12 artifact classes, each carrying a SOURCE ANCHOR to its committed evidence. 77 documents; 79 studio pages; 39 app routes (anchor yemaya-route-inventory.json, studio-web 33/desktop 6); 38 entities/406 fields (yemaya-data-inventory.json); 21 state enums/115 labels (yemaya-state-vocabulary.json); 27,337 package exports across 59 packages, 891 capabilities (domain-package-inventories.json); 67 workers; 1 CLI; 4 integration modules (path-scoped, NOT the ~395 content-grep over-count); 11 case subsystems (yemaya-case-inventory.json); 62 desktop files (yemaya- desktop-partition.json); 1,225 tests. Consolidated from fresh git counts + committed Y0.1/Y0.6/Y0.7/Y0.10/Y0.12 evidence. generate-yemaya-code-census.mjs + 7 tests, wired into inventory/verify + CI. Alias normalization via reviewed mappings (Y0.16.b) and dispositions (Y0.16.c+) are human-review governance and stay unchecked. -->
    • Y0.16.b Normalize aliases through reviewed mappings while preserving source identity/version/semantics and live/fixture/asserted state. Evidence: alias map.
    • Y0.16.c Map projects/assets/scripts/storyboards/characters/worlds/ production/collaboration/dailies/rendering and autonomous pipelines. Evidence: core production rows.
    • Y0.16.d Map engines/DCCs/gameplay/NPCs/virtual production/XR/mocap/ builds/runtime/device operations with Bellona/Aja ownership boundaries. Evidence: specialized production rows.
    • Y0.16.e Map marketplace/search/community/organizations/enterprise/admin/ developer APIs/SDK/webhooks/plugins and security/monetization authority. Evidence: platform rows.
    • Y0.16.f Map study/case/source/canon/compiler/solver/localization/ publication/consumer delivery features. Evidence: track rows.
    • Y0.16.g Assign reuse/refactor/move/merge/retire/net-new/defer/reject, canonical owner, task, dependencies, and evidence per atomic feature. Evidence: disposition ledger.
    • Y0.16.h Review collisions/orphans/unsupported assertions and add a bidirectional zero-orphan drift gate. Evidence: approved reconciliation.
  • Y0.17 Define the product boundary among AAA creation, production operations, organization administration, marketplace/community, developer tooling, study/case authoring, and consumer delivery. Evidence: each capability has an authorized audience/host and governed handoff or explicit exclusion.
    • Y0.17.a Define purpose/actors/hosts/routes/owner services/data/actions for AAA creation, production ops, org admin, marketplace/community, developer, study/case, and consumer delivery. Evidence: surface matrix.
    • Y0.17.b Assign every atomic capability to exactly one mutation owner and authorized audiences, with any read-only facade/deep link. Evidence: ownership ledger.
    • Y0.17.c Define governed signed/versioned project/asset/run/build/review/ release/entitlement handoffs between products and external domains. Evidence: handoff contracts.
    • Y0.17.d Keep organization/billing/security, marketplace/payout, operator fleet/incident, and consumer progress/session authority out of authoring convenience routes. Evidence: boundary tests.
    • Y0.17.e Define excluded/deferred/unsupported cells and safe UI/API/ capability behavior without hidden routes or generic fallback. Evidence: exclusion matrix.
    • Y0.17.f Test direct URL/API/property/function, role changes, deep links, search/export/logs/cache, and tenant/product boundary leakage. Evidence: adversarial suite.
  • Y0.18 Define supported project types and production profiles for film, episodic, animation, game, virtual production, XR, study, interactive case, podcast/AV, and other ratified types. Evidence: every profile maps required entities, pipeline stages, roles, gates, assets, exports, and target hosts without inheriting unsupported defaults.
    • Y0.18.a Define versioned project-type/profile registry with owners, audience, lifecycle, inheritance, target platforms, and code floors. Evidence: registry schema.
    • Y0.18.b Define film/episodic/animation profiles for script/scene/shot/ sequence/editorial/VFX/audio/color/delivery stages, roles, assets, gates, schedules, and targets. Evidence: screen profile rows.
    • Y0.18.c Define game profiles for world/level/character/gameplay/NPC/UI/ audio/cinematics/build/test/performance/engine/platform stages and assets. Evidence: game rows.
    • Y0.18.d Define virtual-production/XR profiles for stage/device/tracking/ camera/render/mocap/engine/session/safety/calibration/build/deployment. Evidence: VP/XR rows.
    • Y0.18.e Define study/interactive-case profiles for sources/content/ canon/clues/solver/compiler/localization/accessibility/review/bundle/ consumer target. Evidence: track rows.
    • Y0.18.f Define podcast/AV and other ratified profiles for script/record/ edit/mix/video/captions/distribution/rights/loudness/target requirements. Evidence: AV rows.
    • Y0.18.g Map exact required/optional/unavailable entities/pipeline stages/ roles/gates/assets/exports/host capabilities/dependencies per profile. Evidence: profile matrix.
    • Y0.18.h Fail missing/unsupported cells instead of inheriting defaults and test profile resolution/overlays/migrations. Evidence: no-default gate.

Y1 — Project, content, revision, and source contracts#

  • Y1.1 Define a canonical authoring project with track, workspace, tenant, ownership, locale, status, production links, and retention policy. Evidence: contract invariants.
    • Y1.1.a Define stable project identity/revision, track/project type/ production profile, title/description, tenant/organization/workspace, and lifecycle. Evidence: base schema.
    • Y1.1.b Define owner/member/team/role assignments, visibility/ compartment/embargo, created/imported/forked provenance, and audit. Evidence: ownership tests.
    • Y1.1.c Define source/default/supported locale, fallback, translation policy, time zone, units/color profile, and accessibility targets. Evidence: locale/config tests.
    • Y1.1.d Define authoring maturity separately from project production/ release/consumer state and valid transitions/guards. Evidence: status invariants.
    • Y1.1.e Define links to assets/production schedule/budget/runs/builds/ targets/external projects by exact version/owner contract. Evidence: production link tests.
    • Y1.1.f Define classification/retention/archive/delete/legal hold/export/ local-cache policy and tenant-scoped key/index invariants. Evidence: lifecycle tests.
  • Y1.2 Define study project, collection, module, note, source node, anchor, concept, reading order, activity, and published artifact entities. Evidence: schema fixtures.
    • Y1.2.a Define study-project specialization and versioned collection/ membership/order/source-scope entities. Evidence: project/collection fixtures.
    • Y1.2.b Define module hierarchy/order/objectives/summary/content/activity/ prerequisite/publication state with cycle/depth rules. Evidence: module fixtures.
    • Y1.2.c Define note/block/annotation/tag/author/source/citation/locale/ visibility/revision and protected-source fields. Evidence: note fixtures.
    • Y1.2.d Define source node/version/kind/authority/rights/checksum/ extraction plus typed anchor and note/canon links. Evidence: source fixtures.
    • Y1.2.e Define concept identity/version/definition/aliases/relationships/ evidence and a separate reading-order/prerequisite graph. Evidence: concept/order fixtures.
    • Y1.2.f Define activities with type/instructions/inputs/outputs/resources/ time/accessibility/consumer behavior and no progress authority. Evidence: activity fixtures.
    • Y1.2.g Define published study artifact/build/manifest/locale/target/ constituent revisions/hashes/provenance/release state. Evidence: full schema goldens.
  • Y1.3 Define case project, premise, mystery model, suspect, relationship, motive, means, opportunity, clue, red herring, scene, timeline event, location, reveal, solution, and ending entities. Evidence: complete case fixture.
    • Y1.3.a Define case-project specialization, premise/theme/tone/setting/ audience/format/duration/difficulty/safety/locale and exact source/canon scope. Evidence: case/premise schema.
    • Y1.3.b Define mystery model/question/crime-event/truth state/solution constraints/allowed interpretations and solver-visible separation. Evidence: mystery schema.
    • Y1.3.c Define character/suspect identity, public/private knowledge, background/traits/goals/arc/alibi plus motive/means/opportunity as typed evidence-backed dimensions. Evidence: suspect fixtures.
    • Y1.3.d Define directed typed relationship revisions with source/canon, temporal interval, visibility/spoiler, confidence, and contradictions. Evidence: relationship fixtures.
    • Y1.3.e Define clue/red-herring identity/kind/content/source/truth link/ discovery prerequisites/placement/salience/interpretations/fairness/ payoff and player-visible state. Evidence: clue fixtures.
    • Y1.3.f Define scene/location/timeline event/order/branch/participants/ actions/dialogue/clue transitions/asset/production/continuity links. Evidence: scene/timeline fixtures.
    • Y1.3.g Define reveal sequence, solution reasoning graph/evidence/ alternatives, ending conditions/outcomes/epilogue and embargoed authority. Evidence: solution/ending fixtures.
    • Y1.3.h Publish minimal/full/branching/multisuspect/multilocale/ intentionally ambiguous/invalid complete-case goldens. Evidence: fixture suite.
  • Y1.4 Define canon assertions with namespace, subject, predicate, object, temporal scope, source, confidence, status, and contradiction links. Evidence: canon contract tests.
    • Y1.4.a Define assertion logical/revision identity, tenant/project/ franchise namespace, subject/predicate/object typed refs/literals, author, and lifecycle. Evidence: base schema.
    • Y1.4.b Define temporal instant/interval/sequence/era/unknown scope, effective story-world versus publication chronology, and uncertainty. Evidence: temporal tests.
    • Y1.4.c Define exact source/citation/creator decision/production artifact evidence, authority, confidence basis, rights, and freshness. Evidence: provenance tests.
    • Y1.4.d Define proposed/under-review/canonical/conditional/disputed/ deprecated/retconned/invalid status and role/guard transitions. Evidence: state tests.
    • Y1.4.e Define supports/contradicts/refines/supersedes/scoped-exception/ alternate-continuity links with exact revisions and no auto-resolution. Evidence: relation tests.
    • Y1.4.f Validate namespace/entity/predicate vocabulary/units/cardinality/ temporal logic/cycles/conflicts and preserve unknown/N/A. Evidence: contract suite.
  • Y1.5 Define typed stable anchors for source pages, study notes, case scenes, dialogue, clues, timeline events, assets, frames, and timecodes. Evidence: anchor round-trip tests.
    • Y1.5.a Define common anchor identity, exact subject/revision/kind, locator payload, context/hash/fingerprint, author, and resolution status. Evidence: anchor schema.
    • Y1.5.b Define page/region/section/paragraph/character source and note block/text-range anchors with extraction/revision context. Evidence: text/source fixtures.
    • Y1.5.c Define case scene/entity, dialogue speaker/line/range, clue/ placement, timeline event/interval, and graph entity anchors by stable IDs. Evidence: case fixtures.
    • Y1.5.d Define asset/object/component/region, image/video frame/region, audio/video timecode/sample/range, shot/sequence anchors with units/ timebase. Evidence: media fixtures.
    • Y1.5.e Validate kind/subject/bounds/permissions and round-trip through API/database/export/localization/build without screen-coordinate-only identity. Evidence: round-trip tests.
    • Y1.5.f Map across revisions as exact/shifted/ambiguous/orphaned/invalid with original preserved and human reattachment. Evidence: survival tests.
  • Y1.6 Define immutable revisions and semantic diffs for text, tree, graph, timeline, and structured case data. Evidence: golden diff fixtures.
    • Y1.6.a Define logical versus revision ID, parent/base/branch, schema version, author/reason/time, payload/manifest hash, state, and immutability. Evidence: revision schema.
    • Y1.6.b Define text/block semantic diff for normalized content/format/ source/citation/locale/accessibility without line-only noise. Evidence: text goldens.
    • Y1.6.c Define tree diff for stable node add/remove/edit/move/reorder/ reparent and descendant impact. Evidence: tree goldens.
    • Y1.6.d Define graph diff for node/edge/type/property/add/remove/change/ merge/split and cycle/reachability/canon impact. Evidence: graph goldens.
    • Y1.6.e Define timeline diff for event/duration/interval/order/branch/ participant/location/causality/continuity changes. Evidence: timeline goldens.
    • Y1.6.f Define structured case diff for premise/suspect/motive/means/ opportunity/clue/red-herring/reveal/solution/ending with spoiler redaction. Evidence: case goldens.
    • Y1.6.g Represent added/removed/changed/moved/unchanged/redacted/unknown and pin schema/comparator versions. Evidence: diff contract tests.
  • Y1.7 Define branch, merge, conflict, and restore rules. Evidence: concurrent authoring known-value tests.
    • Y1.7.a Define branch identity/base/head/purpose/owner/scope/visibility/ lock/lifecycle and branchable entity granularity. Evidence: branch schema.
    • Y1.7.b Define three-way merge by stable IDs for scalar/text/list/tree/ graph/timeline/structured data with exact automatic-safe rules. Evidence: merge fixtures.
    • Y1.7.c Define content/order/delete-edit/reparent/graph/canon/solution/ localization/asset conflicts and mine/theirs/base projections. Evidence: conflict fixtures.
    • Y1.7.d Require human resolution for semantic/canon/spoiler/solution/ source/rights conflicts; preserve both inputs and resolution rationale. Evidence: resolution tests.
    • Y1.7.e Create merged/restored state as a new immutable revision with all parents/provenance; never overwrite branch history. Evidence: lineage tests.
    • Y1.7.f Test simultaneous edits/offline/reconnect/locks/rebase/merge/ abandon/restore and downstream gate invalidation. Evidence: concurrency suite.
  • Y1.8 Define decision records for outlines, case solutions, clue plans, canon resolutions, and production choices. Evidence: schema and UI projection.
    • Y1.8.a Define immutable decision identity/type, actor/authority, exact candidate revisions, chosen/rejected/deferred/superseded outcome, rationale/evidence/conditions/time. Evidence: base schema.
    • Y1.8.b Define outline selection/merge/scope/schedule decision fields and source/constraint/coverage comparisons. Evidence: outline fixtures.
    • Y1.8.c Define solution and clue-plan decisions with solver/fairness/canon/ spoiler evidence, alternatives, reveal/order/placement consequences. Evidence: case decision fixtures.
    • Y1.8.d Define canon resolution with conflicting assertions/sources/ continuity scope, accept/reject/retcon/alternate outcome, and approver. Evidence: canon fixtures.
    • Y1.8.e Define production choice with asset/shot/tool/provider/format/ cost/schedule/quality/rights alternatives and impact. Evidence: production fixtures.
    • Y1.8.f Define role-redacted UI/API projection and supersession without deleting prior decisions or exposing spoilers. Evidence: projection tests.
  • Y1.9 Define source authority, rights, locale, checksum, extraction, and citation metadata. Evidence: source-kind validation.
    • Y1.9.a Define source identity/version/kind/title/creator/publisher/ external IDs/locators/edition/status and provenance. Evidence: source schema.
    • Y1.9.b Define authority/role/expertise/official/primary-secondary/ reliability/review evidence, jurisdiction, currency, and conflicts. Evidence: authority tests.
    • Y1.9.c Define ownership/license/access/use/adapt/train/quote/distribute/ attribution/territory/term/embargo/protected-source evidence. Evidence: rights tests.
    • Y1.9.d Define language/script/region/translation/source locale, audience/ sensitivity/spoiler classification, and locale variants. Evidence: locale tests.
    • Y1.9.e Define original bytes/content hash/length/type/storage, retrieval/ connector receipt, scans, and extraction run/tool/config/output/anchors. Evidence: integrity/extraction tests.
    • Y1.9.f Define citation target/source revision/typed locator/quote/context/ role/verification and validate per source kind. Evidence: source-kind validation.
  • Y1.10 Define release bundles for study content and compiled cases. Evidence: schema/version profiles and compatibility policy.
    • Y1.10.a Define common bundle identity/version/track/project/release candidate/manifest/entry points/files/dependencies/hashes/provenance/ rights/locale/target/profile. Evidence: base schema.
    • Y1.10.b Define study bundle hierarchy/content/notes/concepts/reading order/activities/media/sources/citations/accessibility/search/navigation and consumer configuration. Evidence: study goldens.
    • Y1.10.c Define compiled-case world/entities/scenes/dialogue/timeline/ clues/runtime graph/rules/hints/solution partitions/assets/locales/ save-state compatibility. Evidence: case goldens.
    • Y1.10.d Separate player-visible, progressive/reveal, solver/reviewer, protected-source, and operator diagnostics with encryption/access. Evidence: spoiler partition tests.
    • Y1.10.e Define format/schema/runtime/consumer min-max versions, extensions, migration, degradation/loss, and unsupported behavior. Evidence: compatibility policy.
    • Y1.10.f Produce deterministic manifests/packages and validate/import/read- back minimal/full/edge/multilocale fixtures. Evidence: bundle suite.
  • Y1.11 Define consumer read-model derivation. Evidence: projections are reproducible from approved authoring revisions.
    • Y1.11.a Define each consumer projection’s exact approved bundle/ authoring revision inputs, schema/version, audience/locale/entitlement, owner store/index/cache, and rebuild policy. Evidence: projection map.
    • Y1.11.b Define deterministic transforms for study navigation/content/ search/activity and case runtime/entity/clue/dialogue/timeline/save-state models. Evidence: transform goldens.
    • Y1.11.c Exclude drafts/comments/decisions/protected sources/solutions/ reviewer fields/production secrets according to consumer audience. Evidence: field boundary tests.
    • Y1.11.d Build idempotently from release events/bundles with version/ ordering/checkpoint/retry/dead-letter and no authoring direct writes. Evidence: projection job tests.
    • Y1.11.e Rebuild from scratch and compare hashes/semantics to incremental state; reconcile supersession/unpublish/recall. Evidence: reproducibility tests.
  • Y1.12 Define backward compatibility with existing case and study contracts. Evidence: old-version fixture migration and unsupported version refusal.
    • Y1.12.a Inventory existing schema/protocol/bundle/database/client/ consumer versions, fixtures, deployments, producers, and support windows. Evidence: version inventory.
    • Y1.12.b Map old-to-canonical fields/enums/IDs/semantics/defaults/loss/ unsupported states separately for study and case. Evidence: migration mappings.
    • Y1.12.c Implement explicit version detection, parse validation, staged pure migrations, provenance, warnings, and immutable original retention. Evidence: migration tests.
    • Y1.12.d Define read-old/write-new/dual-read/shim/client-server/runtime negotiation and expiration/telemetry. Evidence: compatibility tests.
    • Y1.12.e Migrate minimal/full/edge old fixtures and semantic round-trip them through current consumers. Evidence: golden suite.
    • Y1.12.f Reject unknown/future/expired/incompatible versions with typed safe diagnostics and no partial import/runtime fallback. Evidence: refusal tests.

Y2 — Study source and material authoring#

  • Y2.1 Build study-source intake from upload, URL, reference, import, and existing Sophia/Yemaya records. Evidence: provenance round-trip.
    • Y2.1.a Define source/intake/version/provenance/rights/checksum/connector/ project-link/status contracts common to every path. Evidence: schemas.
    • Y2.1.b Implement bounded streamed upload with content sniff/hash/ temporary quarantine/tenant isolation. Evidence: upload tests.
    • Y2.1.c Implement allowlisted URL/reference retrieval with SSRF/redirect/ DNS/size/type/time controls and immutable receipt. Evidence: network tests.
    • Y2.1.d Implement safe package/reference import with version/profile/ manifest/path/dependency validation and preview. Evidence: import tests.
    • Y2.1.e Resolve existing Sophia/Yemaya records through public owner contracts, exact IDs/revisions/authorization, and no data copy by URL. Evidence: integration tests.
    • Y2.1.f Normalize paths to one immutable source version and round-trip original locator/bytes/hash/rights/connector/version/audit. Evidence: provenance suite.
  • Y2.2 Quarantine and scan documents, archives, media, and project imports. Evidence: malicious corpus.
    • Y2.2.a Store new content in nonpublic nonexecutable quarantine with opaque names, encryption, tenant scope, limits, and no downstream use. Evidence: quarantine tests.
    • Y2.2.b Scan malware/signature/polyglot/macros/scripts/active content and source/project supply-chain metadata. Evidence: malicious fixtures.
    • Y2.2.c Enforce archive traversal/symlink/nesting/bomb/file-count/ compressed-uncompressed size/password rules. Evidence: archive tests.
    • Y2.2.d Enforce document/page/object/parser/OCR and media codec/duration/ dimension/sample/decode/resource limits. Evidence: resource tests.
    • Y2.2.e Validate project manifest/schema/dependencies/plugins/scripts/ paths/versions/signatures and block executable or unknown constructs. Evidence: project import tests.
    • Y2.2.f Persist scanner/version/config/hash/results and admit exact bytes only after required fresh scans; preserve quarantine/manual/unavailable. Evidence: admission tests.
  • Y2.3 Extract source hierarchy, text, media, tables, headings, pages, and stable anchors. Evidence: golden source corpus.
    • Y2.3.a Define extraction run/tree/node/content/anchor/confidence/error contracts bound to source hash and extractor/config versions. Evidence: schemas.
    • Y2.3.b Extract document/book/site/package hierarchy, order, headings/ sections/pages, reading flow, IDs, and cross-references. Evidence: hierarchy goldens.
    • Y2.3.c Extract normalized text/paragraph/list/footnote/quote/citation with source offsets/page/region and language/direction. Evidence: text goldens.
    • Y2.3.d Extract images/audio/video/files with bytes/hash/caption/alt/ technical metadata/time/page anchors and rights. Evidence: media goldens.
    • Y2.3.e Extract tables with cell/header/span/caption/units/footnotes/ region and semantic/text fallback. Evidence: table goldens.
    • Y2.3.f Generate stable anchor identities/fingerprints/context and test multilingual/RTL/OCR/complex/malformed/partial fixtures. Evidence: golden source corpus.
  • Y2.4 Build source-tree editing with cycle prevention, keyboard reorder, bulk move, and impact preview. Evidence: accessible tree journey.
    • Y2.4.a Define editable collection/source/section/node hierarchy operations, stable IDs/order keys, locks, depth/cardinality, and revision semantics. Evidence: tree contract.
    • Y2.4.b Build tree and synchronized list/outline with create/rename/ reorder/reparent/group/remove/restore by pointer and keyboard/non-drag. Evidence: interaction tests.
    • Y2.4.c Validate self/descendant cycles, duplicates, missing parent, protected roots, cross-project/tenant moves, and stale versions server-side. Evidence: negative tests.
    • Y2.4.d Build selected/all/filter-descendant bulk move with count/scope/ destination validation, atomic/partial policy, and confirmation. Evidence: bulk tests.
    • Y2.4.e Preview affected anchors/notes/claims/materials/reading order/ cases/translations/builds/permissions before commit. Evidence: impact tests.
    • Y2.4.f Verify accessible focus/announcement/zoom/large-tree/concurrent journey and immutable history. Evidence: Playwright E2E.
  • Y2.5 Build note and excerpt capture anchored to exact source locations. Evidence: source update/re-extraction preserves or explicitly orphans the anchor.
    • Y2.5.a Define note/excerpt logical/revision identity, author/project/ visibility/locale/tags/concepts/claims/status and typed exact source anchor. Evidence: schema.
    • Y2.5.b Capture selected text/region/page/time/table slice/URI fragment with exact source/extraction revision, quote/context/hash, and bounds. Evidence: capture tests.
    • Y2.5.c Author note blocks/annotations/commentary separately from immutable excerpt and retain source rights/attribution/spoiler classification. Evidence: editing tests.
    • Y2.5.d Validate source/anchor/access and render citation trail/side-by- side context accessibly without protected-source leakage. Evidence: viewer tests.
    • Y2.5.e On source update/re-extraction map anchor exact/shifted/ambiguous/ orphaned, preserve original, and require reattachment/review. Evidence: survival tests.
  • Y2.6 Build concept and claim extraction behind the fenced provider seam. Evidence: human confirmation before persistence.
    • Y2.6.a Define concept/claim proposal with exact source/extraction/span, normalized term/proposition, aliases/type/qualifiers/relations, confidence, provider/model/prompt/config, and status. Evidence: schemas.
    • Y2.6.b Send only authorized bounded source spans fenced as untrusted data with no tools/project secrets/solution/canon authority. Evidence: injection tests.
    • Y2.6.c Validate proposal schema/locators/source support/duplicates/ conflicts/size and preserve not-configured/refusal/malformed/partial. Evidence: seam tests.
    • Y2.6.d Present source context, similar concepts/claims, uncertainty, suggested relations, and accept/edit/reject/defer per proposal. Evidence: review tests.
    • Y2.6.e Create canonical revisions only from authorized human decisions, retaining original proposal/provider/source lineage. Evidence: authority tests.
  • Y2.7 Build citation attachment and source trails. Evidence: unsupported or missing locators fail validation.
    • Y2.7.a Define citation/relationship contract linking exact note/claim/ block/case/canon target to source revision, typed locator, quote/context, support role, and verification. Evidence: schema.
    • Y2.7.b Implement page/section/paragraph/text range/region/timecode/ dataset/table/URI/entity locators with source-kind compatibility. Evidence: locator fixtures.
    • Y2.7.c Validate syntax/bounds/hash/revision/rights/access and reject unsupported/missing/mutable-only locators. Evidence: validation tests.
    • Y2.7.d Build forward target-to-source and reverse source-to-target trails with intermediate claims/extractions/translations/decisions and stable pagination. Evidence: trail tests.
    • Y2.7.e Render/export accessible citations/trails with authorized excerpts/redactions and update/orphan state. Evidence: UI/export tests.
  • Y2.8 Build outline divergence and compare. Evidence: scores and provider provenance are preserved.
    • Y2.8.a Define outline/variant-set/module/node/order/source/concept/ objective/constraint contracts with exact brief/source revision and manual/generated discriminator. Evidence: schema.
    • Y2.8.b Generate bounded N variants with diversity strategy, seed/ controls, provider/model/prompt/config, estimate/budget, and unique IDs. Evidence: generation tests.
    • Y2.8.c Validate hierarchy/reading order/prerequisites/source coverage/ workload/rights/accessibility and retain invalid candidates. Evidence: validation tests.
    • Y2.8.d Record rule/provider/evaluator/human scores separately with version/rubric/uncertainty/N/A and no score invention for manual outlines. Evidence: score tests.
    • Y2.8.e Build side-by-side semantic structure/content/source/coverage/ cost diff and select/reject/merge/fork decisions preserving all candidates. Evidence: compare journey.
  • Y2.9 Build a structured study-material editor for explanation, excerpt, comparison, question, activity, reflection, media, glossary, and case link. Evidence: block registry tests.
    • Y2.9.a Define versioned block registry with shared identity/order/locale/ source/concept/accessibility fields and per-kind schema/renderer/ validator/export mappings. Evidence: registry.
    • Y2.9.b Implement explanation/excerpt/comparison typed editors with claim/ citation/source-pane and contrasted-dimension semantics. Evidence: content block tests.
    • Y2.9.c Implement question/activity/reflection editors with instructions/ expected engagement/resources/time/response/noninteractive alternative and no learner-state mutation. Evidence: activity tests.
    • Y2.9.d Implement governed media attachment and glossary term/definition/ locale/source links plus exact interactive-case deep link/context. Evidence: media/glossary/case tests.
    • Y2.9.e Implement insert/reorder/nest/convert/duplicate/reference/delete/ restore with stable IDs, depth rules, dependency invalidation, and non-drag controls. Evidence: editing tests.
    • Y2.9.f Fail registry coverage for missing editor/consumer renderer/ validation/accessibility/localization/import-export/preview/test. Evidence: block coverage gate.
  • Y2.10 Build reading-order and prerequisite editing. Evidence: cycle, unreachable node, and missing prerequisite gates.
    • Y2.10.a Define separate ordered sequence/branch and prerequisite DAG edge kinds, optional/required/choice/remedial semantics, conditions, and exact node revisions. Evidence: graph schema.
    • Y2.10.b Build accessible visual plus synchronized list/table create/ connect/disconnect/reorder/group/branch operations. Evidence: UI tests.
    • Y2.10.c Detect self/cycles/duplicates/conflicting order, missing roots/ ends/prerequisites, orphans/unreachable/dead branches. Evidence: validation fixtures.
    • Y2.10.d Validate prerequisite availability before node and reading flow/ locale/accessibility/consumer compatibility. Evidence: sequence tests.
    • Y2.10.e Preview impact on notes/materials/activities/cases/builds and recompute stale reviews/gates after revision. Evidence: impact tests.
  • Y2.11 Build source-hierarchy completeness, anchor coverage, claim grounding, reading-order coherence, accessibility, and rights gates. Evidence: known-value fixtures.
    • Y2.11.a Resolve exact project/material/source/tree/anchor/claim/order/ block/asset/locale/release revisions and applicable gate/profile versions. Evidence: input tests.
    • Y2.11.b Measure required source-node hierarchy/metadata/extraction/ authority/currency completeness and note/block anchor coverage/orphan rate. Evidence: hierarchy/anchor fixtures.
    • Y2.11.c Evaluate atomic claim support/sufficiency/faithfulness/conflict/ citation locators and high-risk source floors. Evidence: grounding fixtures.
    • Y2.11.d Evaluate order/prerequisite cycle/reachability/coherence/ workload and content/asset semantic accessibility/localization parity. Evidence: order/accessibility fixtures.
    • Y2.11.e Evaluate source/content/media rights/attribution/embargo/term and downstream target compatibility. Evidence: rights fixtures.
    • Y2.11.f Persist raw results separately from policy verdicts/uncertainty/ N/A/manual review and block mandatory failures. Evidence: gate suite.
  • Y2.12 Build consumer preview through real study-workspace read models. Evidence: preview parity without writing learner state.
    • Y2.12.a Compile exact draft/approved authoring revisions into an isolated preview bundle and derive the same versioned read models as consumer. Evidence: preview pipeline tests.
    • Y2.12.b Render through real consumer route/components/runtime with synthetic identity, selected locale/device/theme/accessibility/offline profile, and prominent preview state. Evidence: renderer parity tests.
    • Y2.12.c Exercise navigation/reading order/concepts/search/media/ activities/case links and compare semantics to published bundle. Evidence: interaction parity.
    • Y2.12.d Isolate preview events/cache/progress/bookmarks/analytics/ notifications from real learner/consumer systems. Evidence: side-effect tests.
    • Y2.12.e Verify permission/spoiler/protected-source/failed build/stale/ offline and accessibility behavior. Evidence: preview E2E.
  • Y2.13 Build localization with citations, glossary, directionality, source-locale trace, media parity, and review. Evidence: locale bundle golden.
    • Y2.13.a Define locale job/unit/variant/source revision/translation/ reviewer/status/fallback contracts and stable IDs. Evidence: schema.
    • Y2.13.b Translate structured material blocks/concepts/glossary/metadata/ instructions with TM/glossary suggestions and human acceptance. Evidence: content tests.
    • Y2.13.c Preserve citation source locale/exact locator, translated quote/ paraphrase relation, source authority, and no false translated source. Evidence: citation tests.
    • Y2.13.d Handle RTL/bidi/Unicode/numerals/dates/units/layout plus localized media/captions/transcripts/descriptions/voice/assets with parity. Evidence: locale/media tests.
    • Y2.13.e Run meaning/claim/key-term/citation/order/accessibility/media/ target parity gates and required locale/cultural review. Evidence: review tests.
    • Y2.13.f Build deterministic locale bundle and consumer round-trip for RTL/long/CJK/missing-fallback cases. Evidence: golden.
  • Y2.14 Build source update/retraction blast radius across notes, claims, materials, cases, translations, and published items. Evidence: propagation test.
    • Y2.14.a Maintain immutable reverse dependency edges from source/ extraction/anchor/citation/claim into every named consumer and bundle/ read model. Evidence: graph tests.
    • Y2.14.b Ingest verified update/supersession/retraction/rights change with authority/scope/effective time/source matching. Evidence: intake tests.
    • Y2.14.c Traverse direct/transitive impacts with reason/path/revision/ current state/owner/unknown and tenant/property authorization. Evidence: blast-radius tests.
    • Y2.14.d Mark affected draft/review/canon/translation/build/publication/ consumer eligibility stale/blocked without rewriting history. Evidence: propagation tests.
    • Y2.14.e Route re-extract/reattach/re-source/rewrite/review/relocalize/ rebuild/unpublish/recall actions and verify downstream read-back. Evidence: end-to-end propagation.

Y3 — Interactive-case concept, structure, and writing#

  • Y3.1 Build zero-form case-premise capture with genre, audience, constraints, safety, setting, duration, difficulty, locale, and format. Evidence: keyboard-first capture.
    • Y3.1.a Define minimal capture identity/project/author/raw idea/source/ provenance/draft status and typed eventual premise requirements. Evidence: schema.
    • Y3.1.b Implement keyboard-first freeform/paste/voice-safe capture with genre/audience/setting/format hints, autosave/offline, and accessible shortcuts. Evidence: capture tests.
    • Y3.1.c Collect structured audience/age, genre/tone, setting/time, duration/pacing/difficulty, locale/language, format/target constraints. Evidence: completion tests.
    • Y3.1.d Collect safety/sensitivity/identity/rights/source/production/ accessibility/budget/code-floor constraints and explicit unknowns. Evidence: constraint tests.
    • Y3.1.e Create immutable premise draft retaining raw capture and validation/provenance; never treat capture as approved framing. Evidence: lineage tests.
  • Y3.2 Build case-framing divergence through case-writers-room and case-director. Evidence: real provider seam and exact returned variants.
    • Y3.2.a Define framing request/plan/variant-set with exact premise/source/ canon/constraints, requested count/diversity, budget, provider/tool versions, and output schema. Evidence: contracts.
    • Y3.2.b Route writers-room ideation and case-director structural planning through their canonical injected operations, not duplicate prompts or private imports. Evidence: seam tests.
    • Y3.2.c Fence untrusted sources/premise, enforce tool grants/budget/stop/ timeout/safety/spoiler boundaries, and preserve not-configured/refusal/ malformed/partial states. Evidence: provider tests.
    • Y3.2.d Create exact unique returned premise/solution/suspects/clue/ structure/tone/scope variants with full step/model/prompt/config/cost lineage. Evidence: variant fixtures.
    • Y3.2.e Validate schema/canon/safety/fairness/production feasibility without automatically accepting or hiding invalid variants. Evidence: validation tests.
  • Y3.3 Build variant compare for premise, solution, suspect set, clue topology, tone, production scope, and gate estimates. Evidence: comparison fixture.
    • Y3.3.a Resolve authorized exact manual/generated variant revisions and define comparison session/layout/selection contracts. Evidence: schema.
    • Y3.3.b Diff premise/question/theme/setting/audience/format/difficulty and tone/style/safety/source/canon assumptions. Evidence: premise diff.
    • Y3.3.c Diff solution/truth/proof chain/ambiguity, suspects/motive/means/ opportunity/alibis/relationships, and alternative hypotheses. Evidence: solution/cast diff.
    • Y3.3.d Diff clue/red-herring topology, dependencies/discovery/reveal/ evidentiary coverage/fairness and scene/timeline branch structure. Evidence: topology diff.
    • Y3.3.e Diff asset/scene/dialogue/locale/voice/render/engine/target/ schedule/resource/cost production scope. Evidence: scope diff.
    • Y3.3.f Present per-gate raw estimate/availability/uncertainty/provider provenance and accessible table/graph alternatives. Evidence: comparison fixture.
  • Y3.4 Build decision records for selected and rejected case framings. Evidence: downstream case links to the decision.
    • Y3.4.a Define immutable decision with exact variant set/revisions, actor/ authority, select/reject/defer/merge outcome, reasons/evidence/ conditions/time. Evidence: schema.
    • Y3.4.b Record one or multiple selected contributions and every rejected variant/reason without deleting or rewriting them. Evidence: decision tests.
    • Y3.4.c Create a new case/premise revision from the selected or explicitly merged framing with source contribution lineage. Evidence: creation tests.
    • Y3.4.d Link all downstream mystery/cast/clue/scene/production decisions to the framing decision and exact premise revision. Evidence: lineage.
    • Y3.4.e Supersede through a new decision after impact review while preserving historical current/old projections. Evidence: lifecycle tests.
  • Y3.5 Build the case dossier with premise, solution, cast, clues, scenes, timeline, locations, canon, assets, gates, review, and history. Evidence: live browser journey.
    • Y3.5.a Define one permission/spoiler-filtered live projection joining exact case revision and every named relationship/freshness state. Evidence: dossier contract.
    • Y3.5.b Present premise/mystery/solution, cast/relationships/knowledge/ motives, clue/red-herring/proof topology, and authorized differences. Evidence: story panels.
    • Y3.5.c Present scene/timeline/location/branch/state/dialogue and canon/ source/contradiction/continuity state. Evidence: structure panels.
    • Y3.5.d Present assets/production/jobs/budget/schedule/localization/ consumer-build dependencies and failure state. Evidence: production panels.
    • Y3.5.e Present solvability/fairness/safety/canon/technical gates, comments/reviews/decisions/approvals/remediation/maturity. Evidence: review panels.
    • Y3.5.f Present semantic revision/branch/merge/decision/audit/playtest/ build/release history with stable links. Evidence: history tests.
    • Y3.5.g Run live edit-review-playtest-remediate journey with role/spoiler/ tenant/accessibility/responsive/large-case proof. Evidence: Playwright.
  • Y3.6 Build suspect editor with identity, role, motive, means, opportunity, alibi, secrets, relationships, knowledge state, and reveal constraints. Evidence: contract validation.
    • Y3.6.a Author stable character/suspect identity, names/aliases/pronouns/ locale, role/archetype, background/traits/goals/arc, source/canon, and asset refs. Evidence: identity tests.
    • Y3.6.b Author typed motive/means/opportunity with temporal/location/ capability/resource/evidence/strength/contradiction and truth status. Evidence: MMO tests.
    • Y3.6.c Author alibi claim/events/witness/source/reliability/timeline/ verification/contradictions and resolved/false/partial status. Evidence: alibi tests.
    • Y3.6.d Author secret/knowledge fact, learned-at-event, knows/does-not- know/believes/lies/uncertain, visibility/spoiler, and evolution. Evidence: knowledge tests.
    • Y3.6.e Author typed directed relationship history and reveal earliest/ latest/condition/audience/order/forbidden disclosure constraints. Evidence: relationship/reveal tests.
    • Y3.6.f Validate stable IDs, required fields, timeline/canon consistency, circular secrets, impossible knowledge, clue/solution links, and protected fields. Evidence: contract suite.
  • Y3.7 Build clue editor with type, source, discovery condition, timing, visibility, interpretation, dependencies, evidentiary weight, and truth status. Evidence: invariant tests.
    • Y3.7.a Author clue identity/kind/content/representation/source/citation/ asset/locale/rights/canon and intended solver role. Evidence: clue schema.
    • Y3.7.b Author discovery actor/location/scene/action/tool/prerequisite/ state/time/window/branch/repeat/miss/fallback conditions as typed logic. Evidence: discovery tests.
    • Y3.7.c Author visibility/salience/presentation/spoiler/knowledge and interpretation list with supporting/refuting facts, ambiguity, and hints. Evidence: interpretation tests.
    • Y3.7.d Author clue-to-clue/scene/suspect/timeline/claim dependencies, evidentiary dimension/weight/confidence, and proof-chain contribution. Evidence: dependency tests.
    • Y3.7.e Author true/false/misleading/ambiguous/fabricated/altered/ incomplete status with truth/canon/source and reveal/payoff resolution. Evidence: truth tests.
    • Y3.7.f Validate reachable discovery, dependency cycles, chronology, source/asset rights, sufficient resolution, and no unsupported freeform conditions. Evidence: invariants.
  • Y3.8 Build red-herring editor with fairness justification and resolution. Evidence: unresolved or indistinguishable red herrings fail review.
    • Y3.8.a Author red-herring clue/suspect/hypothesis linkage, intended inference, source/canon/truth basis, placement/salience, and scope. Evidence: schema.
    • Y3.8.b Author fairness justification showing why it is plausible yet distinguishable through available evidence, logic, timing, or source. Evidence: fairness tests.
    • Y3.8.c Author required contradiction/correction/disproof/reveal/payoff, discovery dependencies, earliest resolution, and player feedback. Evidence: resolution tests.
    • Y3.8.d Compare to true clue/path for unique signals and detect identical/ strictly indistinguishable/unresolvable/misleading-without-payoff cases. Evidence: comparison fixtures.
    • Y3.8.e Block review on missing justification/resolution/unreachable disproof or harmful/unfair manipulation; preserve manual evidence. Evidence: gate tests.
  • Y3.9 Build scene editor with participants, location, time, objectives, dialogue/actions, clues, state effects, branches, assets, and canon refs. Evidence: structured-block tests.
    • Y3.9.a Define scene identity/revision/order/sequence/branch, title/ summary/purpose/pacing/duration/status/owner/locale and source refs. Evidence: scene schema.
    • Y3.9.b Author participants/entry-exit/presence/role/knowledge/wardrobe/ prop/continuity and location/set/subarea/time/weather/light. Evidence: context tests.
    • Y3.9.c Author objectives/beats/conflicts/outcomes and typed ordered dialogue/action/direction/choice/interaction/transition blocks. Evidence: block tests.
    • Y3.9.d Place clues/red herrings/discovery/reveal/knowledge changes and state preconditions/effects with exact entity refs. Evidence: state tests.
    • Y3.9.e Author branch choices/conditions/targets/merge/fallback/dead-end semantics and sequence/timeline/causal links. Evidence: branch tests.
    • Y3.9.f Link assets/shots/storyboards/audio/locales/production tasks and exact canon assertions/exceptions/continuity profiles. Evidence: production/canon tests.
    • Y3.9.g Validate presence/knowledge/timeline/location/state/canon/clue/ branch/asset constraints and structured block registry coverage. Evidence: validation suite.
  • Y3.10 Build timeline editor with absolute/relative time, uncertainty, witness perspective, evidence refs, and contradiction detection. Evidence: keyboard and semantic-list parity.
    • Y3.10.a Define event/interval identity/revision, absolute instant/range/ time zone/calendar/story era or relative-before/after/during/offset/ sequence constraints. Evidence: temporal schema.
    • Y3.10.b Author duration/uncertainty/earliest-latest/confidence and participants/location/actions/state/clue/scene/canon/source links. Evidence: event tests.
    • Y3.10.c Author separate witness/source perspectives for observed/reported/ inferred/remembered/believed times with reliability/contradictions. Evidence: perspective tests.
    • Y3.10.d Build visual lane/scale/order/zoom/filter editor and synchronized keyboard-operable semantic list/table/forms. Evidence: parity tests.
    • Y3.10.e Solve temporal constraints and flag impossible intervals/order/ travel/presence/alibi/knowledge/age/duration, source/canon conflicts with explanation. Evidence: contradiction fixtures.
    • Y3.10.f Preserve ambiguous/alternate chronologies and require human resolution rather than autochanging events. Evidence: review tests.
  • Y3.11 Build relationship and clue graphs with semantic table parity. Evidence: accessible navigation and graph/list consistency.
    • Y3.11.a Define graph view contracts with stable entity/edge IDs, revisions/types/properties/visibility/evidence/validation and role/ spoiler projection. Evidence: schema.
    • Y3.11.b Build relationship graph filters/layout/time/point-of-view/ knowledge/canon/strength/status and typed edit actions. Evidence: relationship UI tests.
    • Y3.11.c Build clue graph for clues/hypotheses/suspects/scenes/discovery/ dependencies/support/refute/reveal/proof paths/fairness. Evidence: clue UI tests.
    • Y3.11.d Provide synchronized table/tree/list/forms for inspect/filter/ select/add/edit/link/unlink/navigate/reorder with non-drag keyboard parity. Evidence: semantic tests.
    • Y3.11.e Keep visual/table selection/filter/edit/validation consistent and surface cycles/orphans/unreachable/conflicts/unknown. Evidence: consistency tests.
    • Y3.11.f Verify screen reader/keyboard/zoom/reflow/reduced motion/large graph and direct authorization. Evidence: accessible E2E.
  • Y3.12 Build solution editor with proof chain, required clue set, alternative-hypothesis disposition, reveal sequence, and confidence. Evidence: schema invariants.
    • Y3.12.a Define protected solution identity/revision, truth outcome/ culprit/cause/mechanism/timeline/motive/means/opportunity and source/ canon basis. Evidence: solution schema.
    • Y3.12.b Author proof graph steps/claims/inferences/rules/assumptions and support/refute clue/evidence refs with ordering/dependencies. Evidence: proof-chain tests.
    • Y3.12.c Define required/sufficient/optional/corroborating/exculpatory/ misleading clue sets and minimum path/coverage/fallback. Evidence: clue-set tests.
    • Y3.12.d Enumerate plausible alternative hypotheses/suspects and disposition as refuted/less supported/equivalent/unresolved with exact evidence. Evidence: alternative tests.
    • Y3.12.e Author reveal sequence/timing/conditions/presentation/explanation/ payoff and ending linkage without premature spoiler exposure. Evidence: reveal tests.
    • Y3.12.f Record author/solver/calibrated confidence/uncertainty separately and validate proof acyclicity/soundness/clue reachability/uniqueness/ canon/timeline. Evidence: invariants.
  • Y3.13 Build branch and state-condition editor. Evidence: unreachable, dead-end, cyclic, and impossible-condition analysis.
    • Y3.13.a Define typed state variables/domains/defaults/ownership/ visibility, condition expression AST, effects, branch/choice/target/ merge/end contracts. Evidence: schema.
    • Y3.13.b Build allowlisted visual/form expression editor for boolean/ comparison/set/count/temporal/knowledge/clue/inventory flags with no arbitrary code. Evidence: condition tests.
    • Y3.13.c Build branch create/order/condition/priority/fallback/target/ effect/merge authoring with semantic list and graph views. Evidence: editor tests.
    • Y3.13.d Analyze reachability, dead ends, unintended cycles/loops, mutually exclusive/exhaustive conditions, impossible state, shadowed/ nondeterministic branches, and orphan content. Evidence: analysis fixtures.
    • Y3.13.e Simulate bounded seeded paths and show counterexample trace/ assumptions without claiming exhaustive proof when limits are reached. Evidence: simulation tests.
  • Y3.14 Build character/scene dialogue assist with canon and spoiler fences. Evidence: explicit accept and AI attribution.
    • Y3.14.a Define assist request/proposal with exact scene/character/ dialogue/locale/canon/knowledge/spoiler/style revisions, operation, provider provenance, and cost. Evidence: schema.
    • Y3.14.b Build only authorized context from character voice/relationship/ current knowledge/scene objective/canon/source, excluding future reveals/ solutions/private sources outside scope. Evidence: context tests.
    • Y3.14.c Fence content, enforce no tool/approval authority, provider/budget/ timeout/safety/identity/rights controls, and honest failure states. Evidence: seam tests.
    • Y3.14.d Validate speaker/scene/knowledge/canon/timeline/spoiler/style/ locale/safety/accessibility and flag conflicts. Evidence: proposal tests.
    • Y3.14.e Present per-line/block diff/context/attribution and require explicit accept/edit/reject; acceptance creates new revision and reruns gates. Evidence: review tests.
  • Y3.15 Build script/table-read preview. Evidence: roles, timing, missing lines, and accessibility output.
    • Y3.15.a Compile selected case/scenes/branches/dialogue revisions into script format with stable scene/beat/line/speaker IDs and spoiler scope. Evidence: script tests.
    • Y3.15.b Validate cast/role assignments, missing/duplicate/unknown speaker, dialogue order, directions, scene coverage, branch variants, and participant access. Evidence: role tests.
    • Y3.15.c Estimate/read timed scene/line/pause/action totals and flag missing lines/overlap/pacing with assumptions. Evidence: timing tests.
    • Y3.15.d Render text table read and optional consented human/approved TTS voices with pronunciation, playback controls, transcript/captions. Evidence: playback tests.
    • Y3.15.e Export accessible script with semantic headings/speaker labels/ directions/alt text, large print/locale/bidi and no unauthorized solutions. Evidence: accessibility tests.
  • Y3.16 Build playtest configuration with spoiler-protected participants, build version, consent, telemetry scope, and retention. Evidence: privacy-safe test.
    • Y3.16.a Define playtest/session/cohort/participant opaque identity, exact case/build/locale/device/accessibility version, purpose, schedule, owner, and lifecycle. Evidence: schema.
    • Y3.16.b Configure participant eligibility/invite/auth/role/embargo/ spoiler compartment, conflict/prior knowledge, guardian/age rules, and compensation. Evidence: participant tests.
    • Y3.16.c Capture consent for participation/recording/voice/video/ telemetry/quotes/recontact/use/term/withdrawal and provide accessible information. Evidence: consent tests.
    • Y3.16.d Allowlist telemetry events/fields/sampling/purpose, exclude raw secrets/solutions/PII where unnecessary, and define offline upload/ encryption/access. Evidence: telemetry tests.
    • Y3.16.e Define session data/comment/replay/media retention/deletion/ legal hold/export and participant withdrawal propagation. Evidence: lifecycle tests.
    • Y3.16.f Validate build availability/integrity, no admin/debug/solution data, participant entitlements, and privacy review before launch. Evidence: privacy-safe gate.
  • Y3.17 Build playtest feedback and session replay linked to immutable case revision. Evidence: no replay leaks solution to unauthorized users.
    • Y3.17.a Define feedback/event/replay artifact with exact case/build/ participant/session/locale/device/config revisions, timestamps, source, consent, visibility, and integrity. Evidence: schema.
    • Y3.17.b Capture structured difficulty/fairness/confusion/enjoyment/ accessibility/bug/clue/scene feedback plus optional consented text/media. Evidence: feedback tests.
    • Y3.17.c Build deterministic replay from allowlisted ordered events/state checkpoints/input/media refs with versioned runtime and gaps/unknown. Evidence: replay tests.
    • Y3.17.d Apply participant/author/solver/reviewer spoiler/solution/ protected-source/property permissions to timelines, events, state, feedback, search, exports, and URLs. Evidence: leak tests.
    • Y3.17.e Anchor comments/findings to session time/event/scene/clue/build and route aggregated/redacted proposals to authoring without auto-edit. Evidence: feedback-to-revision tests.
    • Y3.17.f Test revoked consent/retention expiry/deleted participant/case supersession/corrupt replay/cross-tenant access. Evidence: lifecycle E2E.
  • Y3.18 Build authoring templates for genre and format with code-owned fairness/safety floors. Evidence: template property tests.
    • Y3.18.a Define versioned template identity/scope/genre/format/audience/ project profile, typed premise/cast/clue/scene/solution/production defaults/requirements, parent, and lifecycle. Evidence: schema.
    • Y3.18.b Define genre-specific conventions/tropes/tone/pacing/clue/fairness/ sensitivity and format-specific duration/branch/interaction/asset/target constraints as guidance or typed requirements. Evidence: fixtures.
    • Y3.18.c Resolve inheritance with exact versions/cycle/depth and field/ collection/graph merge/replace/append/remove/lock semantics. Evidence: inheritance tests.
    • Y3.18.d Apply code/region/platform fairness/safety/rights/consent/privacy/ accessibility/spoiler/authority floors monotonically. Evidence: property tests.
    • Y3.18.e Preview inherited/defaulted/overridden/locked/conflicts and impact, then version/publish/deprecate/migrate consumers explicitly. Evidence: lifecycle tests.

Y4 — Canon, solvability, fairness, and safety gates#

  • Y4.1 Implement canon graph validation for missing entities, conflicting assertions, temporal impossibility, and namespace mismatch. Evidence: known-value canon fixtures.
    • Y4.1.a Resolve exact canon namespace/entity/predicate/assertion/source/ status/version graph and validation profile. Evidence: input tests.
    • Y4.1.b Detect missing/deleted/unresolved/unauthorized entity refs, invalid predicate/cardinality/type/unit/literal, dangling relations, and required assertion gaps. Evidence: entity fixtures.
    • Y4.1.c Detect simultaneously applicable contradicting values/relations/ knowledge/identity/location/state with confidence/status/source scope. Evidence: conflict fixtures.
    • Y4.1.d Solve temporal interval/order/age/duration/presence/state constraints and report minimal impossible sets/counterexamples. Evidence: temporal fixtures.
    • Y4.1.e Detect cross-namespace refs, alias/version ambiguity, continuity/ alternate-universe leakage, and unapproved scoped exceptions. Evidence: namespace fixtures.
    • Y4.1.f Persist per-assertion/path raw findings and human resolution state separately from canon gate verdict. Evidence: known-value suite.
  • Y4.2 Implement canon update and retraction blast radius. Evidence: affected scenes, dialogue, clues, solutions, assets, and releases become stale.
    • Y4.2.a Maintain immutable forward/reverse edges from canon assertions/ entities/sources/decisions to every named consumer, build, and release. Evidence: dependency tests.
    • Y4.2.b Ingest authorized new revision/supersession/retcon/deprecation/ retraction/scoped exception with effective continuity/time and evidence. Evidence: update tests.
    • Y4.2.c Traverse direct/transitive impact with path/reason/revision/ current state/owner/unknown and spoiler/tenant authorization. Evidence: query tests.
    • Y4.2.d Mark affected scenes/dialogue/clues/solutions/assets/locales/ reviews/gates/builds/releases stale or blocked without rewriting history. Evidence: propagation tests.
    • Y4.2.e Route reconcile/rewrite/relink/regenerate/review/rebuild/unpublish/ recall and verify downstream read-back. Evidence: blast-radius E2E.
  • Y4.3 Implement clue reachability from every supported play path. Evidence: graph fixtures with reachable and hidden required clues.
    • Y4.3.a Compile exact case revision into typed state/branch/scene/clue/ discovery graph with supported start states/modes/difficulty/accessibility profiles. Evidence: compiler tests.
    • Y4.3.b Enumerate/explore all paths for bounded cases and use sound symbolic/model checking for larger graphs with declared limits. Evidence: traversal tests.
    • Y4.3.c Verify every required clue is discoverable before needed on every supported path, including miss/revisit/fallback/branch merge/reset. Evidence: reachable/hidden fixtures.
    • Y4.3.d Return unreachable clue/path/state counterexample traces, unsupported/timeout/coverage bounds, and no false exhaustive claim. Evidence: result tests.
    • Y4.3.e Separate optional/bonus/alternative-path clues and profile-specific requirements from universal requirements. Evidence: applicability tests.
  • Y4.4 Implement solution existence. Evidence: at least one evidence-backed solution is provable.
    • Y4.4.a Define solution rule system, hypothesis space, truth/canon facts, required evidence/inference rules, allowed assumptions, and exact case revision. Evidence: solver contract.
    • Y4.4.b Compile clue discovery/interpretations/proof steps/alternative paths into solver inputs without consuming protected author assertions as proof automatically. Evidence: input tests.
    • Y4.4.c Find at least one satisfying hypothesis/proof chain using only reachable admissible evidence and valid inferences under each supported profile. Evidence: positive fixtures.
    • Y4.4.d Return zero-solution unsatisfied core/counterexample, timeout/ unsupported/unknown separately and persist trace. Evidence: negative fixtures.
    • Y4.4.e Verify solution matches locked author truth/canon or explicitly declared open outcome without circular/self-revealing clues. Evidence: existence gate.
  • Y4.5 Implement solution uniqueness under the declared rules. Evidence: zero-, one-, and multi-solution fixtures.
    • Y4.5.a Define what counts as materially distinct solution/hypothesis/ culprit/mechanism/timeline versus equivalent representation. Evidence: equivalence rules.
    • Y4.5.b Enumerate or prove bounded count of satisfying solutions from reachable evidence/rules with solver version/limits. Evidence: cardinality tests.
    • Y4.5.c Validate exactly one when uniqueness is required, at least one/ bounded alternatives when ambiguity is intentional, and allowed outcome constraints. Evidence: policy tests.
    • Y4.5.d Return representative multiple solutions and discriminating/ missing evidence plus zero/unknown/timeout states. Evidence: fixtures.
    • Y4.5.e Distinguish true logical uniqueness from only tested path/sample coverage. Evidence: proof-quality tests.
  • Y4.6 Implement alternative-hypothesis testing. Evidence: each plausible suspect is eliminated or retained by explicit evidence.
    • Y4.6.a Generate/resolve the declared plausible suspect/hypothesis set from exact case entities/rules without hidden arbitrary omissions. Evidence: hypothesis tests.
    • Y4.6.b Evaluate motive/means/opportunity/alibi/timeline/clue/canon/ relationship/knowledge evidence for and against each hypothesis. Evidence: evidence matrix.
    • Y4.6.c Build explicit proof/refutation/insufficient chain with source/ clue/inference refs, contradictions, assumptions, confidence, and reachable timing. Evidence: chain tests.
    • Y4.6.d Classify eliminated/retained/equivalent/unresolved/unsupported and require every plausible hypothesis to have a disposition. Evidence: coverage gate.
    • Y4.6.e Detect unjustified elimination, absence-as-proof, circular clues, hidden solution facts, and post-answer evidence. Evidence: adversarial fixtures.
  • Y4.7 Implement fair-clue timing. Evidence: required evidence appears before the answer is demanded.
    • Y4.7.a Define answer-demand/reveal/commit points and required/sufficient/ disambiguating evidence sets per supported path/profile. Evidence: timing schema.
    • Y4.7.b Compute earliest/latest possible discovery and guaranteed availability for each clue under branches, state, miss/revisit, fallback, pacing, and accessibility. Evidence: timing tests.
    • Y4.7.c Verify required evidence is reachable/perceivable/interpretable before commitment with adequate review time and no future/hidden source. Evidence: fairness fixtures.
    • Y4.7.d Report late/missable/branch-locked/too-early-spoiler/ inaccessible evidence and counterexample path. Evidence: result tests.
    • Y4.7.e Apply profile/difficulty policy to raw timing and require human fairness review for contextual salience/pacing. Evidence: gate tests.
  • Y4.8 Implement suspect coverage and opportunity coherence. Evidence: crafted omissions and contradictions fail.
    • Y4.8.a Resolve all in-scope suspects/hypotheses and required motive/ means/opportunity/alibi/relationship/clue/timeline dimensions. Evidence: coverage schema.
    • Y4.8.b Detect missing/placeholder/unsupported dimensions, omitted plausible suspects, overprivileged culprit evidence, and incomplete alternative dispositions. Evidence: omission fixtures.
    • Y4.8.c Verify opportunity intervals against presence/location/travel/ capability/resources/access/knowledge and event timing. Evidence: coherence fixtures.
    • Y4.8.d Verify alibi/witness/source/canon/scene/clue consistency and surface contradictions/unknowns separately. Evidence: contradiction tests.
    • Y4.8.e Preserve intentional unknown/false alibis with evidence and apply fairness policy separately. Evidence: gate tests.
  • Y4.9 Implement timeline consistency and travel-time constraints. Evidence: independently derived fixture values.
    • Y4.9.a Normalize absolute/relative/uncertain event constraints, duration, time zones/calendars/story era, participant/location and source perspective from exact revision. Evidence: input tests.
    • Y4.9.b Solve ordering/overlap/duration/age/state/knowledge/presence/ simultaneous-action constraints and report minimal conflicts. Evidence: timeline fixtures.
    • Y4.9.c Resolve versioned location graph/transport/mode/speed/range/ accessibility/transition overhead and compute minimum/expected travel times independently. Evidence: travel goldens.
    • Y4.9.d Detect impossible/implausible/unknown travel, teleportation, double presence, missed transition, and alibi inconsistencies with units. Evidence: edge fixtures.
    • Y4.9.e Keep author estimates, calculated values, uncertainty, policy thresholds, and intentional canon exceptions separate. Evidence: result tests.
  • Y4.10 Implement scene reachability, dead-end, soft-lock, and state reset checks. Evidence: exhaustive small-case traversal.
    • Y4.10.a Compile typed start states, scenes/branches/conditions/effects/ choices/inventory/knowledge/clues/checkpoints/resets/endings and supported profiles. Evidence: state model tests.
    • Y4.10.b Exhaustively traverse bounded cases and use bounded symbolic/ property exploration for larger cases with declared state/path limits. Evidence: traversal tests.
    • Y4.10.c Detect unreachable scenes/choices/endings, terminal dead ends without intended ending, and orphan branches. Evidence: reachability fixtures.
    • Y4.10.d Detect soft locks where required progress becomes impossible while runtime remains active, with minimal path/state trace. Evidence: soft-lock fixtures.
    • Y4.10.e Verify restart/checkpoint/scene reset/branch merge/replay clears/ preserves exact state and cannot duplicate/lose required clues. Evidence: reset fixtures.
    • Y4.10.f Return exhaustive/partial/timeout/unsupported coverage and never label bounded exploration complete. Evidence: result tests.
  • Y4.11 Implement spoiler-boundary analysis for UI labels, assets, logs, URLs, notifications, localization, and exports. Evidence: leak corpus.
    • Y4.11.a Classify solution/reveal/clue truth/suspect secret/canon/branch/ protected-source fields by spoiler tier, audience, reveal condition, and embargo. Evidence: data map.
    • Y4.11.b Inventory author/reviewer/solver/playtester/consumer/operator UI/API/search/client/cache/assets/accessibility/telemetry/output channels. Evidence: channel manifest.
    • Y4.11.c Scan labels/tooltips/alt text/transcripts/captions/filenames/ thumbnails/metadata/client state/bundles for direct/encoded/indirect leaks. Evidence: UI/asset corpus.
    • Y4.11.d Scan logs/errors/traces/URLs/deep links/notifications/email/ webhooks/localization TM/glossary/export/debug/save files. Evidence: operational leak tests.
    • Y4.11.e Test role/reveal progression/back navigation/offline cache/ locale fallback/search/count/timing/cross-tenant behavior. Evidence: adversarial suite.
    • Y4.11.f Fail release on unresolved spoiler exposure and preserve authorized evidence with redaction. Evidence: leak gate.
  • Y4.12 Implement harmful-content, representation, age-rating, and accessibility gates. Evidence: policy/version and human-review states.
    • Y4.12.a Define separate versioned gate taxonomies/applicability/ evaluator/rule/human evidence for harmful content, representation, regional age rating, and accessibility. Evidence: registries.
    • Y4.12.b Evaluate text/dialogue/scene/choice/consequence/assets/media/ gameplay for violence/abuse/hate/sexual/self-harm/child/identity/ manipulation and contextual uncertainty. Evidence: safety fixtures.
    • Y4.12.c Evaluate stereotype/tokenism/agency/power/cultural/linguistic/ protected-trait portrayal and harmful mechanics as review evidence, not a truth score. Evidence: representation corpus.
    • Y4.12.d Map content descriptors/intensity/context/interactivity/user content/region/platform to rating evidence and required qualified review. Evidence: age-rating matrix.
    • Y4.12.e Evaluate semantic/keyboard/non-drag/color/motion/cognitive/ captions/descriptions/audio/controls/timing alternatives for every content/runtime path. Evidence: accessibility matrix.
    • Y4.12.f Persist raw results/version/uncertainty/N/A separately, route human review/exception, and block mandatory failures. Evidence: gate tests.
  • Y4.13 Implement study grounding and case canon gates as distinct registries. Evidence: no false mapping between evidence and fiction canon.
    • Y4.13.a Define separate gate identities/contracts/vocabularies/owners/ policies/applicability for real-world source grounding and fictional canon consistency. Evidence: registry schemas.
    • Y4.13.b Map study claims/citations/source authority/currency/faithfulness/ conflict only to grounding measurements/verdicts. Evidence: grounding tests.
    • Y4.13.c Map case entity/assertion/timeline/knowledge/continuity/ contradiction only to canon measurements/verdicts. Evidence: canon tests.
    • Y4.13.d Define explicit “fictional claim informed by source” relation retaining both source evidence and canon assertion without equivalence. Evidence: mixed-content fixtures.
    • Y4.13.e Fail cross-registry ID/enum/threshold/result substitution and consumer aggregation that labels canon as factual support or vice versa. Evidence: false-mapping tests.
  • Y4.14 Persist raw solver traces and measurements separately from policy verdicts. Evidence: threshold/policy recomputation.
    • Y4.14.a Define immutable solver run/trace with subject revision, compiled model, solver/engine/config/environment versions, inputs/assumptions, explored states/paths/proofs/counterexamples/limits, and hash. Evidence: trace schema.
    • Y4.14.b Define immutable measurements for reachability/solution count/ timing/coverage/conflicts/resources with units/uncertainty/applicability. Evidence: measurement schema.
    • Y4.14.c Define verdict records referencing exact traces/measurements plus gate/profile/policy/threshold versions. Evidence: verdict schema.
    • Y4.14.d Recompute new verdict identities on policy/threshold/profile change without rerunning or modifying valid raw traces. Evidence: recompute tests.
    • Y4.14.e Require new solver run when case/compiler/solver/input/assumption changes and preserve all history. Evidence: lifecycle tests.
  • Y4.15 Add insufficient-data, timeout, unsupported, and evaluator-failure states. Evidence: none renders as pass.
    • Y4.15.a Define pass/fail/advisory/manual-review/N/A/insufficient/timeout/ unsupported/evaluator-failure/unavailable/stale/unknown states and transitions. Evidence: vocabulary.
    • Y4.15.b Require typed reason/coverage/limit/missing input/dependency/ evaluator/error/correlation/remediation/evidence per nonresult state. Evidence: schema tests.
    • Y4.15.c Map exceptions/timeouts/provider errors/partial solver coverage/ invalid input to exact states without generic catch-as-pass/fail. Evidence: mapping tests.
    • Y4.15.d Aggregate mandatory gates so none of these states can satisfy review/release, and preserve valid partial measurements. Evidence: property tests.
    • Y4.15.e Render/export text/semantic state with no green/color-only/ zero-value assumption. Evidence: UI/API tests.
  • Y4.16 Version solver, engine, canon, evaluator, corpus, and policy. Evidence: every gate is reproducible or explicitly not reproducible.
    • Y4.16.a Register solver/engine/compiler/code/config/runtime/dependency version/hash/source/build/lifecycle/supported constructs. Evidence: tool registry.
    • Y4.16.b Register canon schema/namespace/predicate/entity/mapping/ assertion-source and validation rule versions. Evidence: canon registry.
    • Y4.16.c Register evaluator/model/prompt/config/provider/training/ calibration and corpus manifest/labels/splits/rights/coverage versions. Evidence: evaluator/corpus registry.
    • Y4.16.d Register gate/policy/threshold/profile/rubric/aggregation/ manual-band versions with approvals/effective dates. Evidence: policy registry.
    • Y4.16.e Bind exact versions/environment/hardware/nondeterminism controls to every trace/measurement/verdict and verify goldens. Evidence: reproducibility tests.
    • Y4.16.f Label nonreproducible with exact uncontrolled dependency and prohibit retired/quarantined versions from new runs while retaining history. Evidence: lifecycle tests.

Y5 — Production, assets, rendering, and agent orchestration#

  • Y5.1 Link authoring projects to existing Yemaya projects and organizations. Evidence: no duplicate project identity.
    • Y5.1.a Resolve exact canonical project/organization/workspace/member/ profile IDs/versions through public owner contracts and current actor access. Evidence: resolver tests.
    • Y5.1.b Define authoring-track extension/link record with canonical project identity, selected content roots, config, and no copied owner/ membership/status fields. Evidence: linkage schema.
    • Y5.1.c Create/open/migrate links idempotently and reject cross-tenant, incompatible profile, archived/deleted, or duplicate identity. Evidence: lifecycle tests.
    • Y5.1.d Project organization/project changes and permissions dynamically, preserving authoring revisions without shadow synchronization authority. Evidence: update tests.
    • Y5.1.e Add schema/search/import guards and end-to-end proof that one project appears under one stable identity across surfaces. Evidence: duplication gate.
  • Y5.2 Link characters, locations, scenes, clues, and study materials to asset-library records. Evidence: stable lineage.
    • Y5.2.a Define typed entity-to-asset-revision link roles for portrait/ reference/model/texture/audio/video/scene/proxy/render/evidence and ordering/primary status. Evidence: link schema.
    • Y5.2.b Select only authorized exact asset revisions compatible with entity kind/project/profile/locale/target/status. Evidence: selector tests.
    • Y5.2.c Persist forward/reverse links with source/generation/import/ transformation provenance, rights/consent/canon/continuity refs and author decision. Evidence: lineage tests.
    • Y5.2.d Version link add/remove/reorder/replace without changing asset or entity history and compute downstream scene/build impacts. Evidence: lifecycle tests.
    • Y5.2.e Propagate asset supersession/quarantine/recall/rights expiry and entity/canon revision to eligibility/freshness. Evidence: propagation tests.
  • Y5.3 Build rights/provenance status into asset selection. Evidence: ineligible assets cannot enter a release candidate.
    • Y5.3.a Define asset-selector projection for owner/source/license/use/ adaptation/training/distribution/attribution/territory/term/consent/ provenance/gates/target and exact versions. Evidence: schema.
    • Y5.3.b Derive eligible/review/expired/revoked/incompatible/incomplete/ quarantined/unknown per project use/audience/locale/target rather than a generic asset approval. Evidence: policy tests.
    • Y5.3.c Present status/reasons/evidence/freshness/required attribution/ restrictions without exposing protected source/identity fields. Evidence: UI tests.
    • Y5.3.d Revalidate on attach, candidate lock, build, and release; reject direct foreign/stale asset IDs server-side. Evidence: TOCTOU tests.
    • Y5.3.e Scan every release-candidate assembly path and prove ineligible assets/derivatives cannot enter. Evidence: release boundary suite.
  • Y5.4 Build brief-to-Isis generation handoff for approved assets. Evidence: project/entity context and return lineage.
    • Y5.4.a Define handoff request for exact project/entity/scene/shot/ continuity/canon/source/asset revisions, approved asset kind/operation, audience, return intent, expiry, nonce, and signature. Evidence: schema.
    • Y5.4.b Validate project/actor/use/right/consent/canon/source/target and minimize authorized context before signed handoff. Evidence: issuance tests.
    • Y5.4.c Reauthorize in Isis, resolve canonical inputs, show plan/cost/ capability/substitution preview, and create governed run. Evidence: acceptance tests.
    • Y5.4.d Return generated proposal/output revision with run/model/workflow/ input/rights/provenance/gates and immutable source context linkage. Evidence: return tests.
    • Y5.4.e Require Yemaya accept/reject/link decision; no automatic asset/ canon/scene promotion. Evidence: authority tests.
    • Y5.4.f Test failure/cancel/multiple output/stale context/expired link/ cross-tenant/back navigation. Evidence: handoff E2E.
  • Y5.5 Build dailies review for rendered case scenes and study media. Evidence: timecode/frame annotations survive revisions.
    • Y5.5.a Define dailies session/item/render revision/shot-scene-material/ participant/reviewer/status/timebase/color/audio/rights projection. Evidence: schema.
    • Y5.5.b Build permission-filtered list/detail/playlist with synchronized image/video/audio/3D playback, color proxy, versions, gates, source/ lineage, and live/partial states. Evidence: viewer tests.
    • Y5.5.c Create typed frame/timecode/range/region/object/track anchored comments, drawings/notes, replies, mentions, resolve/reopen, and accessible anchor list. Evidence: annotation tests.
    • Y5.5.d Map anchors across rerender/trim/rate/crop/replacement as exact/ shifted/ambiguous/orphaned and retain original evidence. Evidence: survival tests.
    • Y5.5.e Record approve/reject/changes/hold/compare/promote decisions against exact render revision with stale/concurrency/role checks. Evidence: review tests.
    • Y5.5.f Run case scene and study media render-to-notes-to-rerender journey with accessibility/responsive/tenant proof. Evidence: E2E.
  • Y5.6 Build render and compilation job views with dependency, capacity, cost, progress, cancellation, and failure detail. Evidence: live fault journey.
    • Y5.6.a Define audience-safe projections for request/plan/DAG/job/stage/ attempt/worker-provider/reservation/budget/cost/output/error/events. Evidence: view contracts.
    • Y5.6.b Build job list/detail with stable filters/URLs, dependency graph, queue/capacity/reservations, inputs/versions, and bounded live updates. Evidence: UI tests.
    • Y5.6.c Present estimate/reserved/actual compute/storage/egress/time/cost, pending/unbilled/variance and attribution by stage/artifact. Evidence: cost tests.
    • Y5.6.d Present typed progress/checkpoints/logs/outputs/gates/refusal/ policy/capacity/timeout/malformed/worker-loss/internal failures with redaction. Evidence: failure view tests.
    • Y5.6.e Implement authorized cancel/retry/replay/redirect with state/ impact/cost/reason/confirmation and immutable attempt/run lineage. Evidence: action tests.
    • Y5.6.f Run live injected dependency/provider/worker/storage failure, reconnect/recovery/late output/tenant/accessibility journey. Evidence: Playwright fault E2E.
  • Y5.7 Build production verification before asset promotion. Evidence: checksum, technical, rights, and content gates.
    • Y5.7.a Resolve exact source/output/asset/proxy/rendition/dependency/ target/profile revisions and required verification registry. Evidence: plan tests.
    • Y5.7.b Verify bytes/manifest/dependencies/checksums/signatures/content type/storage integrity and quarantine corruption. Evidence: integrity tests.
    • Y5.7.c Run asset-kind/target technical format/dimensions/color/audio/ video/3D/engine/DCC/performance gates. Evidence: technical fixtures.
    • Y5.7.d Verify source/license/consent/attribution/provenance/watermark/ canon/continuity and content/safety/age/accessibility gates. Evidence: trust fixtures.
    • Y5.7.e Bind raw measurements/verdicts/freshness/overrides/reviews to exact candidate and block promotion on missing/stale/failed mandatory gate. Evidence: promotion tests.
    • Y5.7.f Promote as a new asset status/revision with decision/audit and revalidate on downstream release. Evidence: lifecycle tests.
  • Y5.8 Build project-Obsidian or external-project import/export only through versioned safe adapters. Evidence: path, conflict, and round-trip tests.
    • Y5.8.a Define adapter interface/version/capabilities/source-target profile, manifest/mapping/loss/diagnostic/security contracts and approved adapter registry. Evidence: adapter schema.
    • Y5.8.b Parse imports in quarantine with archive/path/symlink/absolute/ traversal/hidden/script/plugin/size/depth/version protections. Evidence: hostile path tests.
    • Y5.8.c Map project/entities/tree/graph/timeline/notes/assets/links/ metadata/locales/IDs/versions to canonical contracts with source provenance. Evidence: mapping tests.
    • Y5.8.d Preview create/update/merge/skip/conflict/duplicate/loss/ unsupported/external dependency and require authorized resolution. Evidence: conflict tests.
    • Y5.8.e Export deterministic safe package with exact revisions, relative paths, manifest/hashes/standards metadata/rights and no secrets/local paths. Evidence: export tests.
    • Y5.8.f Round-trip every supported adapter/version/edge profile and reject unknown/unsupported versions explicitly. Evidence: adapter suite.
  • Y5.9 Build writers-room/director agent run plans with tool grants, budgets, sources, canon snapshot, and stop conditions. Evidence: plan validation.
    • Y5.9.a Define distinct writers-room/director plan/step/output contracts for ideation/dialogue/structure/continuity/production planning. Evidence: schemas.
    • Y5.9.b Bind exact project/premise/case/scenes/sources/citations/canon/ continuity/assets/profile/policy revisions and spoiler compartments. Evidence: input tests.
    • Y5.9.c Resolve allowlisted versioned tool/model/provider grants by operation/data/project/tenant/network/resource/output and expiry. Evidence: grant tests.
    • Y5.9.d Allocate per-step/total token/tool/compute/cost/time budgets, dependencies/retries/approvals and hard stops for safety/canon/source/ budget/timeout/verification. Evidence: budget/stop tests.
    • Y5.9.e Validate DAG/types/grants/inputs/freshness/policy/budget/ destinations before dispatch; reject injected/arbitrary plans. Evidence: zero-dispatch tests.
    • Y5.9.f Ensure plans grant no maturity/canon/approval/promotion/publish/ budget exception authority. Evidence: authority tests.
  • Y5.10 Build run/queue/budget/provenance views scoped to authoring project. Evidence: cross-project access test.
    • Y5.10.a Define project-filtered request/plan/step/queue/run/attempt/tool/ event/budget/evidence/output/decision/lineage projections. Evidence: contracts.
    • Y5.10.b Build list/detail/DAG/timeline/live reconnect with exact versions, queue/reservation/status/stops/approvals/failures. Evidence: run UI tests.
    • Y5.10.c Present tool purpose/status/timing and redacted inputs/outputs, allocated/consumed/remaining budgets, sources/canon/provenance/proposals. Evidence: detail tests.
    • Y5.10.d Enforce tenant/organization/project/role/spoiler/object/property/ function authorization on query/search/count/URL/stream/export/action. Evidence: cross-project tests.
    • Y5.10.e Verify long-run performance, accessibility/responsive modes and project switch/back/role revoke/cache behavior. Evidence: E2E.
  • Y5.11 Build human-override pause, redirect, accept, reject, and abort. Evidence: overrides are audited and cannot rewrite history.
    • Y5.11.a Define separate command eligibility/role/scope/reason/evidence/ idempotency/expected version/outcome and human-only operations. Evidence: schemas.
    • Y5.11.b Pause at safe boundary/checkpoint and resume only after full source/canon/grant/policy/budget/dependency revalidation. Evidence: pause tests.
    • Y5.11.c Redirect through a new plan revision with changed objective/ steps/tools/inputs/budget and preserved completed outputs/history. Evidence: redirect tests.
    • Y5.11.d Accept/reject exact attributed proposals into review/new authoring revisions without approving canon/gates/maturity/publish. Evidence: decision tests.
    • Y5.11.e Abort/cancel queued/running/downstream calls with late-result fencing, partial evidence/cost, resource cleanup, and explicit outcome. Evidence: abort race tests.
    • Y5.11.f Record append-only commands/events/actor/authority/plan before- after links and prohibit state/history rewriting. Evidence: audit tests.
  • Y5.12 Build agent-output intake as proposals. Evidence: no agent directly advances maturity, approves gates, or publishes.
    • Y5.12.a Define signed/hashed result envelope with agent/run/step/tool/ model/config/input versions, output kind/schema, evidence/sources/canon, confidence/cost/status. Evidence: schema.
    • Y5.12.b Validate signature/hash/schema/size/type/expected output/project/ tenant/source locators and scan files/content. Evidence: hostile tests.
    • Y5.12.c Reject/redact forbidden secrets/protected sources/solutions/ unrelated project data and unsupported canon/approval claims. Evidence: property tests.
    • Y5.12.d Store text/structure/dialogue/asset/production/review results only as attributed proposals with exact authoring target linkage. Evidence: persistence tests.
    • Y5.12.e Require human inspect/accept/edit/reject/defer; acceptance creates a new revision and reruns applicable gates. Evidence: review tests.
    • Y5.12.f Prove no payload/property/event can advance maturity, resolve canon, pass/waive gates, promote assets, approve budget, or publish. Evidence: authority suite.
  • Y5.13 Build cost and capacity estimates for writing, generation, rendering, compilation, localization, and review. Evidence: planned versus actual reconciliation.
    • Y5.13.a Define versioned price/rate/productivity/resource/capacity models and units for labor/tool/provider/compute/storage/egress/vendor/reviewer work. Evidence: model schemas.
    • Y5.13.b Estimate writing/editorial/research/canon/solver/review labor by volume/complexity/locale/risk/role/rates/availability. Evidence: labor fixtures.
    • Y5.13.c Estimate generation/render/compilation/postproduction/storage/ egress by plan/variants/assets/passes/targets/provider/resources/retries. Evidence: technical fixtures.
    • Y5.13.d Estimate localization/translation/media adaptation/locale review/ rendition/build and named reviewer capacity/schedule. Evidence: locale fixtures.
    • Y5.13.e Return low/expected/high, confidence/unknowns/assumptions/ bottlenecks/calendars/currencies/rate versions and threshold confirmation. Evidence: estimate tests.
    • Y5.13.f Ingest actual labor/jobs/providers/storage/reviews, allocate by project/stage/artifact/locale/target, and report pending/unbilled/ variance/cause without double count. Evidence: reconciliation tests.
  • Y5.14 Build web/desktop background-job handoff and reconnect. Evidence: closing a client never corrupts server job state.
    • Y5.14.a Define server-owned durable job identity/state/event cursor/ subscription/cancellation/action contracts independent of client process/window/session. Evidence: job contract.
    • Y5.14.b Submit idempotently with exact project/input/plan/version and return durable job ref before client background tracking. Evidence: submission tests.
    • Y5.14.c Handoff between web/desktop/window/process/device by signed project-scoped link/notification and current reauthorization. Evidence: handoff tests.
    • Y5.14.d Reconnect from last cursor/snapshot with dedup/out-of-order/gap/ terminal handling; refresh expired credentials without resubmission. Evidence: reconnect tests.
    • Y5.14.e Close/crash/update/offline/reopen clients during queued/running/ callback/finalizing and prove job/checkpoints/outputs unaffected. Evidence: lifecycle tests.
    • Y5.14.f Make cancel/retry/replay explicit server commands and never infer cancellation from client disconnect. Evidence: authority tests.

Y5A — Existing AAA production-suite feature completion#

  • Y5A.1 Complete project creation, template, workspace, membership, favorites/recent, duplicate, import/export, archive, restore, retention, and project-type migration. Evidence: every supported project profile has a durable lifecycle journey and no hidden local-only state.
    • Y5A.1.a Build typed project creation for profile/type/template, organization/workspace, title/code/locale/timezone/units/color, owner, classification/retention, targets, and validation. Evidence: creation tests.
    • Y5A.1.b Build versioned project templates with inheritance/defaults/ required entities/pipeline/roles/gates/folders/config, preview, review, code floors, and lifecycle. Evidence: template tests.
    • Y5A.1.c Build project workspace configuration for navigation/layout/ saved views/default tools/capabilities/integrations without local-only canonical state. Evidence: workspace sync tests.
    • Y5A.1.d Build owner/member/team/service-account/guest invite/accept/ revoke/role/scope/expiry/conflict and project leave/transfer workflows. Evidence: membership tests.
    • Y5A.1.e Build per-user favorite/unfavorite/reorder and derived recent/ last-opened state across web/desktop/devices with privacy/retention. Evidence: favorites/recent tests.
    • Y5A.1.f Duplicate/fork exact project revision with selective content/ assets/config/history/member/link options, new IDs, rights/canon/source checks, progress, and provenance. Evidence: duplication tests.
    • Y5A.1.g Import through approved versioned adapters with quarantine, preview/conflicts/mappings/loss/authorization/idempotency and source package retention. Evidence: import tests.
    • Y5A.1.h Export deterministic project package with exact revisions/ manifests/dependencies/hashes/rights/redactions and no local paths/ secrets, then round-trip. Evidence: export tests.
    • Y5A.1.i Archive project from normal discovery/mutation with active job/ lock/build/release/dependency/member impact, and preserve authorized read/history. Evidence: archive tests.
    • Y5A.1.j Restore archive as explicit state transition after policy/ retention/dependency/capability checks without resurrecting revoked rights/secrets/jobs. Evidence: restore tests.
    • Y5A.1.k Enforce retention/deletion/legal hold/export/backup/local-cache schedules and verified downstream cleanup/irrecoverable copies. Evidence: retention tests.
    • Y5A.1.l Migrate project type/profile through versioned impact preview, entity/pipeline/gate/asset/role/target mappings, conflicts, new revision, rollback, and no silent defaults. Evidence: migration tests.
    • Y5A.1.m Run durable create-to-archive/restore/migrate journey for every supported profile across web/desktop/restart/two tenants. Evidence: lifecycle matrix.
  • Y5A.2 Complete the asset library across the ratified asset-type registry, including upload/capture/generation/import, folders/collections, tags, search, exact/perceptual dedupe, versions, relationships, dependencies, rights, provenance, proxy/rendition, approval, promotion, quarantine, archive, and restore. Evidence: per-asset-kind capability and gate matrix.
    • Y5A.2.a Define ratified asset-kind/version registry with schemas/ formats/roles/technical metadata/preview/editor/gates/targets/rights/ provenance and explicit unsupported cells. Evidence: registry.
    • Y5A.2.b Implement upload with streaming hash/sniff/quarantine/scan/ dedupe/metadata/rights/progress/cancel/resume and asset revision creation. Evidence: upload tests.
    • Y5A.2.c Implement governed device/file/desktop capture with permission, calibration/timecode/color/audio, consent, local temp security, ingest, and failure cleanup. Evidence: capture tests.
    • Y5A.2.d Implement Isis generation handoff/return and manual/external import with exact provenance, safe adapters, quarantine, and proposal acceptance. Evidence: generation/import tests.
    • Y5A.2.e Implement versioned folder hierarchy and collection/static-smart/ ordered membership with cycle, permissions, bulk move, and no ownership ambiguity. Evidence: organization tests.
    • Y5A.2.f Implement governed tag/taxonomy/alias/locale authoring and permission-safe search/filter/facet/sort/pagination/autocomplete. Evidence: metadata/search tests.
    • Y5A.2.g Implement exact cryptographic/content duplicate and calibrated modality-specific perceptual near-duplicate detection with candidate evidence/uncertainty and human merge/keep decisions. Evidence: dedupe tests.
    • Y5A.2.h Implement immutable logical/revision history, semantic/media diff, compare/select/supersede/restore and current projection without overwrites. Evidence: version tests.
    • Y5A.2.i Implement typed asset-to-entity/asset/dependency/ingredient/ derivative/rendition/variant/continuity/canon/production relationships with impact queries. Evidence: relationship tests.
    • Y5A.2.j Implement rights/license/consent/attribution/embargo/provenance/ source/generation/model/workflow/watermark/C2PA status and blast radius. Evidence: trust tests.
    • Y5A.2.k Implement original/proxy/thumbnail/streaming/edit/review/delivery/ locale/LOD/mip/stem/rendition manifests, jobs, hashes, compatibility, and source linkage. Evidence: rendition tests.
    • Y5A.2.l Implement asset-kind/target quality/safety/rights/canon/ continuity/accessibility/technical review, decisions, and stale evidence. Evidence: approval tests.
    • Y5A.2.m Implement promote/link/use only for eligible exact revisions and revalidate in build/release; reject direct stale IDs. Evidence: promotion tests.
    • Y5A.2.n Implement quarantine/recall with admission block/dependency/ project/build/release propagation and restricted evidence. Evidence: quarantine tests.
    • Y5A.2.o Implement archive/restore/retention/delete/legal hold while preserving history/lineage and rejecting active dependencies where required. Evidence: lifecycle tests.
    • Y5A.2.p Run per-asset-kind source-to-approved-to-target journey and fail every registry cell without current capability/gate proof. Evidence: matrix gate.
  • Y5A.3 Complete screenplay authoring for structured scenes, slug lines, action, dialogue, characters, revisions, colored pages, locks, notes, breakdowns, reports, compare, table read, and interchange. Evidence: production-script golden and collaborative editing journey.
    • Y5A.3.a Define screenplay/project/logical/revision/page/scene/block/ character/dialogue/production metadata contracts and supported interchange profiles. Evidence: schemas.
    • Y5A.3.b Build structured scene/act/sequence order and slug line INT/EXT/ location/time/continuity/special qualifier editor with stable IDs. Evidence: structure tests.
    • Y5A.3.c Build typed action/description/transition/shot/text/lyrics/ centered/page-break and format rules with keyboard-first editing. Evidence: action block tests.
    • Y5A.3.d Build character cues/dialogue/parentheticals/dual dialogue/ extensions/voiceover/offscreen/language/pronunciation and canonical character links. Evidence: dialogue tests.
    • Y5A.3.e Implement pagination/layout/font/margins/widow-orphan/continued/ page numbering/title pages and accessible semantic alternative. Evidence: render goldens.
    • Y5A.3.f Implement revisions/sets/colored pages/date/version marks/ changed-line indicators and compare/diff/report without changing content identity. Evidence: revision tests.
    • Y5A.3.g Implement scene/page/line locks, production revisions, omitted/ added scenes, numbering/revisions, conflict/override authority and audit. Evidence: lock tests.
    • Y5A.3.h Implement anchored notes/comments/suggestions/assignments/ mentions/resolve and stable anchor migration. Evidence: note tests.
    • Y5A.3.i Build breakdown categories/items/characters/props/wardrobe/ locations/VFX/SFX/stunts/extras/safety with entity/asset/task links. Evidence: breakdown tests.
    • Y5A.3.j Build scene/character/dialogue/location/page/revision/breakdown/ duration/cast/day-night reports with permission-safe exports. Evidence: report tests.
    • Y5A.3.k Build semantic side-by-side compare/merge and consented table-read cast/voice/timing/missing-lines/captions/script output. Evidence: compare/table-read tests.
    • Y5A.3.l Implement Final Draft/FDX and other ratified import/export with source retention, mapping/loss diagnostics, deterministic round-trip, and no answer/spoiler leak. Evidence: interchange tests.
    • Y5A.3.m Run collaborative two-editor/offline/reconnect/lock/revision/ conflict/restore production-script journey and compare golden. Evidence: E2E.
  • Y5A.4 Complete storyboard and animatic authoring for panels, shots, framing, lens/camera, timing, dialogue/audio, continuity, versions, generated assets, sequence playback, exports, and production links. Evidence: script→storyboard→animatic lineage and round-trip.
    • Y5A.4.a Define storyboard/sequence/shot/panel/version/timeline/track/ media/annotation contracts with exact script/scene/beat links. Evidence: schemas.
    • Y5A.4.b Build panel create/import/draw/generate/replace/crop/annotate/ caption/action/dialogue/shot-number/order with asset provenance. Evidence: panel tests.
    • Y5A.4.c Build stable shot identity/scene/sequence/order/type/coverage/ transition/status/owner/production links and conflict-safe reorder. Evidence: shot tests.
    • Y5A.4.d Author composition/framing/aspect/safe area/subject placement/ eyeline/screen direction/axis/depth/scale and visual annotations. Evidence: framing tests.
    • Y5A.4.e Author camera/lens/sensor/focal/aperture/focus/height/distance/ pose/movement/path/shake in explicit units with preview. Evidence: camera tests.
    • Y5A.4.f Author frame/timecode/seconds duration/handles/rate/transition/ retime and dialogue/SFX/music/ambience/guide audio tracks with sync. Evidence: timing/audio tests.
    • Y5A.4.g Apply character/wardrobe/prop/location/light/palette/lens/camera/ motion/canon continuity profiles and drift/conflict annotations. Evidence: continuity tests.
    • Y5A.4.h Version/diff/branch/merge/restore panels/shots/timeline/tracks and preserve stable unaffected identities. Evidence: revision tests.
    • Y5A.4.i Attach generated panels/audio through governed Isis/Euterpe handoffs with run/input/model/workflow/rights/provenance and human acceptance. Evidence: generation tests.
    • Y5A.4.j Build synchronized sequence animatic playback/scrub/loop/ thumbnails/waveforms/captions/comments/comparison with accessible lists. Evidence: playback tests.
    • Y5A.4.k Export image boards/PDF/video/editorial EDL/XML/AAF or ratified profiles with stable IDs/timebases/media/metadata/loss reports and round-trip. Evidence: export tests.
    • Y5A.4.l Link shots to schedule/tasks/assets/renders/dailies/editorial/ engine/DCC and propagate script/panel/timing/continuity changes. Evidence: production linkage tests.
    • Y5A.4.m Run script-scene-to-board-to-generated-panel-to-animatic-to- editorial round-trip with exact lineage. Evidence: E2E.
  • Y5A.5 Complete character, creature, prop, world, location, environment, lore, relationship, visual-development, reference, canon, and continuity-design workspaces. Evidence: source and generated assets remain linked to approved entity revisions and downstream scenes.
    • Y5A.5.a Define versioned entity-kind registry/common identity/locale/ source/canon/asset/continuity/production fields and per-kind schemas. Evidence: registry.
    • Y5A.5.b Build character/creature anatomy/appearance/identity/age/ personality/goals/arc/voice/performance/wardrobe/rig/consent authoring. Evidence: character tests.
    • Y5A.5.c Build prop/item identity/function/dimensions/materials/state/ ownership/interaction/handedness/variants/continuity authoring. Evidence: prop tests.
    • Y5A.5.d Build world/setting geography/history/cultures/politics/economy/ technology/rules/time/canon/regions and hierarchy authoring. Evidence: world tests.
    • Y5A.5.e Build location/environment layout/scale/architecture/biome/ weather/time/light/set-dressing/accessibility/production constraints. Evidence: location tests.
    • Y5A.5.f Build lore assertion/event/faction/object/ritual/language/ chronology/source/authority/contradiction/retcon/continuity authoring. Evidence: lore tests.
    • Y5A.5.g Build directed relationship type/history/strength/state/knowledge/ conflict/secret/temporal/canon and accessible graph/table views. Evidence: relationship tests.
    • Y5A.5.h Build visual-development brief/mood/palette/silhouette/shape/ material/scale/style/reference explorations/variants/decisions. Evidence: visual development tests.
    • Y5A.5.i Build source/reference ingest/boards/roles/crop/annotation/ rights/attribution/consent/authority/avoidance/conditioning links. Evidence: reference tests.
    • Y5A.5.j Build canon assertion/conflict/source/decision/status/namespace/ temporal review and blast-radius within each entity workspace. Evidence: canon tests.
    • Y5A.5.k Build continuity profiles/revisions for identity/appearance/ scale/state/material/wardrobe/props/environment/lore/relationships across scenes/shots/assets. Evidence: continuity tests.
    • Y5A.5.l Attach source/manual/generated asset revisions with role/ provenance/rights/gates/decision, compare/select/promote, and no auto- canon. Evidence: asset linkage tests.
    • Y5A.5.m Link approved exact entity revisions to script/scenes/shots/ gameplay/NPCs/schedule/tasks/builds and propagate changes/stale state. Evidence: downstream linkage tests.
    • Y5A.5.n Run representative per-entity source-to-design-to-approved-asset- to-scene journey with graph/accessibility/tenant proof. Evidence: E2E.
  • Y5A.6 Complete production scheduling for breakdown items, shots/tasks, dependencies, milestones, calendars, call sheets, locations, crew, availability, time zones, schedule optimization/sharing, conflicts, and actuals. Evidence: schedule-to-call-sheet and change-impact journeys.
    • Y5A.6.a Define schedule/project/version/task/shot/breakdown/resource/ dependency/milestone/calendar/booking/call-sheet/actual contracts and stable IDs. Evidence: schemas.
    • Y5A.6.b Convert approved script/entity/scene/shot/breakdown revisions into estimable schedulable items with quantities/durations/skills/resources/ locations/constraints and change links. Evidence: breakdown tests.
    • Y5A.6.c Build task/shot create/assign/status/priority/estimate/duration/ dates/owner/crew/resource/budget/asset/gate/version and bulk operations. Evidence: task tests.
    • Y5A.6.d Build finish-start/start-start/finish-finish/lag/lead/required/ soft dependency DAG, cycle/reachability/critical-path checks, and impact. Evidence: dependency tests.
    • Y5A.6.e Build milestones/deliverables/reviews/release windows/embargo/ approval criteria and slip/escalation behavior. Evidence: milestone tests.
    • Y5A.6.f Build production/institution/crew/location/stage/device/provider calendars, work rules/holidays/quiet hours/recurrence/availability/ holds/bookings in IANA time zones. Evidence: calendar tests.
    • Y5A.6.g Build location/permit/weather/travel/setup/strike/capacity/ accessibility/safety constraints and crew role/rate/union/rest/ qualification/consent/contact projections. Evidence: resource tests.
    • Y5A.6.h Detect double-booking, impossible dependencies, insufficient turnaround/rest/travel/resources/capacity/skills, milestone misses, and timezone/DST conflicts. Evidence: conflict tests.
    • Y5A.6.i Build deterministic optimization suggestions for critical path, resource leveling, grouping/location moves, cost/quality/risk tradeoffs with assumptions and no automatic apply. Evidence: optimization tests.
    • Y5A.6.j Build schedule version/diff/what-if/approval/publish/share/export/ subscriber access with role redaction and stable links. Evidence: schedule lifecycle tests.
    • Y5A.6.k Generate versioned call sheets from committed scene/shot/ location/crew/cast/equipment/safety/weather/travel/contact/access/ schedule refs with acknowledgements. Evidence: call-sheet tests.
    • Y5A.6.l Capture start/end/progress/completion/actual labor/resource/ incident/cost/delay reasons separately from plan and reconcile. Evidence: actuals tests.
    • Y5A.6.m Run script change→breakdown impact→optimized committed schedule→ call sheet→actuals/replan journey across zones. Evidence: E2E.
  • Y5A.7 Complete budgeting for templates, line items, versions, multi-currency/rates, estimates, approvals, commitments, actuals, variance, cost prediction, contingency, and audit. Evidence: exact arithmetic, rate-version, permission, and reconciliation tests.
    • Y5A.7.a Define budget/template/version/category/account/line-item/rate/ estimate/approval/commitment/actual/contingency/forecast/audit contracts with currencies/units/tax/time. Evidence: schemas.
    • Y5A.7.b Build versioned budget templates/mappings/formulas/required categories/assumptions/floors by project profile/region and inheritance. Evidence: template tests.
    • Y5A.7.c Build line items for quantity/unit/rate/currency/tax/markup/ discount/vendor/crew/resource/task/asset/shot/date/cost center/funding with decimal arithmetic. Evidence: line-item tests.
    • Y5A.7.d Build immutable exchange/pay/union/provider/vendor/tax rate versions with source/effective period/rounding and exact consumed rate. Evidence: rate tests.
    • Y5A.7.e Build low/expected/high estimates/scenarios/quotes/assumptions/ confidence, import, compare, freeze/baseline, and budget revision diff. Evidence: estimate tests.
    • Y5A.7.f Build submit/review/request-changes/approve/reject with thresholds, role/quorum/separation, exact version, evidence, conditions, and expiry. Evidence: approval tests.
    • Y5A.7.g Build purchase order/contract/booking/reservation/commitment, encumbrance, status, vendor, approval, cancellation, and remaining available balance. Evidence: commitment tests.
    • Y5A.7.h Ingest verified invoice/timecard/provider/job/storage/expense/ payment actuals with source/idempotency/allocation/status/corrections. Evidence: actual tests.
    • Y5A.7.i Compute planned/baseline/current/committed/actual/forecast/ remaining variance by amount/percentage/currency/category/task/period with typed cause. Evidence: variance goldens.
    • Y5A.7.j Build cost prediction from versioned measurements/model/ assumptions/uncertainty and compare champion/challenger; never overwrite estimates. Evidence: prediction tests.
    • Y5A.7.k Build contingency reserve/allowance/risk trigger/draw/replenish/ approval/release and prevent hidden negative budget. Evidence: contingency tests.
    • Y5A.7.l Enforce project/org/role/property/spend-limit/separation/ concurrency/idempotency and immutable correction/audit/export. Evidence: permission/audit tests.
    • Y5A.7.m Reconcile budget/commitments/actuals/provider bills/currency/ refunds/credits and run exact arithmetic/rate-version journey. Evidence: reconciliation E2E.
  • Y5A.8 Complete dailies, editorial, timing, cleanup, object removal, relighting/lighting, HDR trim, localization, foley, score, portrait, motion-camera, AV narrative, and video-edit workspaces by integrating existing pages and libraries. Evidence: each page has live backing, immutable review media, per-frame/timecode notes, gates, and handoff.
    • Y5A.8.a Inventory every existing page/library/route/API/store/job/ capability/owner/live-fixture state and assign reuse/refactor/merge/ retire/gap disposition. Evidence: workspace ledger.
    • Y5A.8.b Define shared exact project/sequence/shot/asset/source/render/ edit/timeline/track/version/timebase/color/audio/review/job/gate/handoff contracts. Evidence: shared schemas.
    • Y5A.8.c Complete dailies ingest/playlist/proxy/color/audio/playback/ compare/frame-time notes/decisions/rerender/promotion. Evidence: dailies journey.
    • Y5A.8.d Complete editorial sequence/track/clip/source-timecode/record- timecode/trim/transition/retime/multicam/audio/captions/version/conform/ interchange. Evidence: editorial journey.
    • Y5A.8.e Complete animation/action/audio timing/exposure/beat/marker/ duration/rate/sync/retime comparison and accessible timeline/list. Evidence: timing journey.
    • Y5A.8.f Complete cleanup/paint/roto/mask/stabilize/denoise/artifact/ beauty/fix layers with source frames, versioned patches, preview, gates, and revert. Evidence: cleanup journey.
    • Y5A.8.g Complete object-removal target/mask/tracking/occlusion/background/ temporal consistency/generation provenance/review and non-destructive replacement. Evidence: removal journey.
    • Y5A.8.h Complete relighting/lighting source/environment/light/camera/ geometry/material/exposure/shadow/reflection/render-pass/HDR/color workflows. Evidence: lighting journey.
    • Y5A.8.i Complete HDR trim target display/profile/master/shot controls/ metadata/measurement/compare/version/signoff/export. Evidence: HDR journey.
    • Y5A.8.j Complete localization script/dialogue/text/captions/dub/voice/ timing/lip-sync/graphics/locale review/rendition/package parity. Evidence: localization journey.
    • Y5A.8.k Complete foley cue/object/surface/action/performer/take/timing/ sync/edit/mix/rights/asset handoff and review. Evidence: foley journey.
    • Y5A.8.l Complete score spotting/cue/tempo/meter/key/hit points/stems/ mockup/recording/license/timeline/mix/delivery through Euterpe seams. Evidence: score journey.
    • Y5A.8.m Complete portrait identity/consent/reference/selection/retouch/ crop/color/background/renditions/rights/quality/promotion. Evidence: portrait journey.
    • Y5A.8.n Complete motion-camera tracking/solve/path/lens/distortion/ coordinates/scale/confidence/cleanup/preview/DCC-engine export. Evidence: motion-camera journey.
    • Y5A.8.o Complete AV narrative script/story/sequence/picture/dialogue/ narration/music/SFX/captions/renditions/review/package authoring. Evidence: AV narrative journey.
    • Y5A.8.p Complete video-edit source/sequence/clip/track/effects/color/ audio/titles/captions/version/render/conform/export workflows. Evidence: video-edit journey.
    • Y5A.8.q For every workspace, require live owner backing, immutable source/ review media, stable frame/timecode/object anchors, typed gates, job progress/cancel/failure, and governed upstream/downstream handoff. Evidence: per-page invariant matrix.
    • Y5A.8.r Verify cross-workspace edit/rerender/review/handoff lineage, web/desktop/accessibility/role/tenant and no duplicated mutation owner. Evidence: integration E2E.
  • Y5A.9 Complete real-time collaboration for CRDT document sync, presence, cursor/selection, comments/suggestions, edit locks, voice/video, recording/transcript, reconnect, guest access, and approved production management/video-conference integrations. Evidence: degraded and cross-version collaboration journeys.
    • Y5A.9.a Define collaboration document/room/session/participant/client/ protocol/version/operation/vector/awareness/event/snapshot contracts per editor/entity type. Evidence: schemas.
    • Y5A.9.b Implement CRDT operations/schema validation/stable IDs/order/ transactions/undo/redo/snapshots/compaction/persistence/audit and reject unauthorized fields. Evidence: sync tests.
    • Y5A.9.c Implement online/offline/multi-device concurrent create/edit/ reorder/delete/merge with deterministic convergence and semantic conflict escalation where CRDT cannot decide meaning. Evidence: convergence tests.
    • Y5A.9.d Implement permission-filtered ephemeral presence, identity/ status/active tool/viewport, heartbeat/expiry/privacy, and no hidden participant leakage. Evidence: presence tests.
    • Y5A.9.e Implement cursor/text range/block/entity/graph/timeline/asset/ frame/time selection awareness with stable anchor mapping and accessible participant list. Evidence: cursor tests.
    • Y5A.9.f Implement durable comments/threads/replies/mentions/attachments/ suggestions/accept-reject/resolve/reopen/notifications with exact revision anchors. Evidence: review collaboration tests.
    • Y5A.9.g Implement edit/production/asset/scene/shot locks with lease/ fencing/heartbeat/expiry/read-only/takeover/override/audit and no split- brain commit. Evidence: lock tests.
    • Y5A.9.h Implement voice/video room permission/device selection/mute/ screen share/quality/network/encryption/participant/accessibility/ captions and honest degraded state. Evidence: conference tests.
    • Y5A.9.i Implement explicit consented recording/stop/indicator/storage/ access/retention/export/delete plus attributed transcript/speaker/time/ correction/caption and no covert capture. Evidence: recording tests.
    • Y5A.9.j Implement reconnect from vector/cursor/snapshot with auth refresh, replay/dedup/gap/schema migration/conflict and no duplicated operations. Evidence: reconnect tests.
    • Y5A.9.k Implement guest invite/identity/role/project/entity/spoiler/ duration/download/recording limits, revocation, watermark, and audit. Evidence: guest tests.
    • Y5A.9.l Integrate approved production-management/video-conference systems through versioned public adapters for identity/calendar/task/ meeting/recording/links/events with source-of-truth directionality. Evidence: integration tests.
    • Y5A.9.m Run mixed web/desktop/old-new client/offline/partition/reconnect/ revoked guest/degraded media journey with convergence/accessibility/ tenant proof. Evidence: cross-version E2E.
  • Y5A.10 Complete pipeline-template and step authoring, configuration, validation, dry run, versioning, scheduling, budget/resource policy, human checkpoints, autonomous film/game execution, pause/redirect/cancel, retry, rollback, and output review. Evidence: agents never bypass safety, budget, production-verification, or human-override floors.
    • Y5A.10.a Define immutable pipeline/template/version/step/port/edge/ parameter/dependency/tool/model/host/resource/checkpoint/schedule/run/ output contracts and canonical serialization. Evidence: schemas.
    • Y5A.10.b Build versioned template metadata/profile/operation/input-output/ default/extension/inheritance/owner/lifecycle authoring. Evidence: template tests.
    • Y5A.10.c Build typed step catalog and graph authoring for services/tools/ agents/providers/renderers/DCC-engine/human tasks with stable ports/ dependencies and no arbitrary code. Evidence: editor tests.
    • Y5A.10.d Build parameter/config/environment/secret-ref/target/condition/ retry/timeout/idempotency/concurrency/artifact/resource policies with effective-value provenance. Evidence: configuration tests.
    • Y5A.10.e Validate graph types/cycles/reachability/dead outputs/ dependencies/versions/licenses/security/rights/capabilities/resources/ targets/policy/authority. Evidence: validation suite.
    • Y5A.10.f Dry-run resolution/authorization/dependency availability/ estimates/plan/checkpoints/outputs without external side effects. Evidence: zero-side-effect tests.
    • Y5A.10.g Version/diff/review/approve/promote/deprecate/quarantine/rollback exact signed pipelines and preserve historical run bindings. Evidence: lifecycle tests.
    • Y5A.10.h Schedule one-time/recurring/event/dependency-triggered runs with timezone, capacity/reservation, priority/fairness, concurrency, embargo, and missed-run behavior. Evidence: schedule tests.
    • Y5A.10.i Allocate/enforce quota/budget/token/tool/compute/GPU/CPU/VRAM/ storage/egress/time per step/total with estimates, reservations, stops, and reconciliation. Evidence: resource tests.
    • Y5A.10.j Insert human-only review/approval/decision checkpoints with eligible roles/quorum/evidence/unchanged plan/expiry and no system actor completion. Evidence: checkpoint tests.
    • Y5A.10.k Execute bounded autonomous film pipelines for script/board/ asset/shot/render/edit/audio/localize/build outputs and game pipelines for design/asset/level/gameplay/NPC/build/test. Evidence: execution tests.
    • Y5A.10.l Implement pause/resume/redirect/cancel/abort/budget kill with safe boundaries/new plan revisions/provider propagation/late-result fencing/partial evidence. Evidence: command race tests.
    • Y5A.10.m Implement stage-scoped retry/new attempt and rollback/ compensation to prior admitted versions without rewriting runs or deleting outputs. Evidence: recovery tests.
    • Y5A.10.n Route every attributed output through integrity/technical/ safety/rights/canon/continuity/production verification and human accept/reject/revise before promotion. Evidence: output review tests.
    • Y5A.10.o Adversarially prove agents/flags/imports/API payloads cannot bypass code floors, reservations, checkpoints, verification, or human override. Evidence: authority boundary suite.
  • Y5A.11 Complete Blender, Godot, Unreal, Houdini, and DaVinci workflows through the canonical Bellona/public adapter boundaries, including host capability, project/asset handoff, command status, version compatibility, round-trip loss, and recovery. Evidence: no private bridge duplication in Yemaya Studio.
    • Y5A.11.a Define Yemaya-to-Bellona public handoff/session/command/artifact/ progress/error/capability/compatibility/loss/recovery contracts and ownership rules. Evidence: boundary schema.
    • Y5A.11.b Build host capability discovery for installed/connected host/ plugin/version/OS/features/commands/formats/targets and supported/ degraded/unavailable reasons. Evidence: capability tests.
    • Y5A.11.c Implement Blender project/scene/asset/geometry/material/rig/ animation/camera/render/export handoffs and status through Bellona. Evidence: Blender journey.
    • Y5A.11.d Implement Godot project/scene/resource/gameplay/build/import/ export handoffs and status through Bellona. Evidence: Godot journey.
    • Y5A.11.e Implement Unreal project/content/level/asset/sequence/gameplay/ build/import/export handoffs and status through Bellona. Evidence: Unreal journey.
    • Y5A.11.f Implement Houdini project/hip/network/asset/cache/simulation/ cook/export handoffs and status through Bellona. Evidence: Houdini journey.
    • Y5A.11.g Implement DaVinci project/timeline/media/color/audio/caption/ conform/render/interchange handoffs and status through Bellona. Evidence: DaVinci journey.
    • Y5A.11.h Preserve exact Yemaya project/entity/asset/shot/sequence/build context, revisions/hashes/units/axes/color/timebase/rights/provenance on outbound and returned artifacts. Evidence: lineage tests.
    • Y5A.11.i Present typed queued/running/progress/output/partial/cancelled/ failed/unknown command state, logs/receipts/cost and authorized cancel/ retry/replay without raw shell. Evidence: status tests.
    • Y5A.11.j Enforce host/plugin/server/adapter/project/format/feature version compatibility, migration window, and explicit unsupported behavior. Evidence: version matrix.
    • Y5A.11.k Compute/report semantic round-trip losses/substitutions for hierarchy/transforms/units/materials/rig/animation/camera/timeline/ metadata and require review. Evidence: loss tests.
    • Y5A.11.l Recover from disconnect/host crash/plugin restart/partial transfer/command timeout/artifact mismatch with reconciliation and no duplicate mutation. Evidence: recovery tests.
    • Y5A.11.m Add forbidden-import/IPC/API/schema checks proving no private bridge/session/command implementation in Yemaya. Evidence: ownership gate.
  • Y5A.12 Complete gameplay and NPC design for mechanics, rules, states, behaviors, dialogue, navigation, encounters, inventory/combat hooks, simulation, playtest, balance, safety, and engine export. Evidence: design-to-supported-engine bundles and deterministic simulation fixtures.
    • Y5A.12.a Define versioned gameplay project/system/mechanic/rule/state/ event/action/parameter/entity/component/behavior/encounter/test/export contracts and registries. Evidence: schemas.
    • Y5A.12.b Author mechanics with player/NPC actions, inputs, preconditions, effects, resources, feedback, failure/success, timing, tuning, objective, dependencies, and accessibility. Evidence: mechanic tests.
    • Y5A.12.c Author rule sets with typed conditions/effects/priorities/scope/ conflicts/invariants/cooldowns/probability/seed and no arbitrary code. Evidence: rule tests.
    • Y5A.12.d Author finite/hierarchical state machines and event flows with initial/final states, transitions/guards/effects, parallel regions, resets, cycle/dead/unreachable analysis. Evidence: state tests.
    • Y5A.12.e Author NPC identity/role/goals/personality/knowledge/memory/ perception/needs/emotion/relationship/consent/canon and gameplay links. Evidence: NPC tests.
    • Y5A.12.f Author behavior tree/utility/planner/task/condition/blackboard/ interrupt/fallback/timing/resource/safety contracts and accessible graph/table parity. Evidence: behavior tests.
    • Y5A.12.g Author branching dialogue with speaker/knowledge/relationship/ condition/choice/consequence/voice/localization/spoiler/canon and integrity constraints. Evidence: dialogue tests.
    • Y5A.12.h Author navigation spaces/agents/links/areas/costs/obstacles/ paths/avoidance/traversal/off-mesh/crowds and reachability validation. Evidence: navigation tests.
    • Y5A.12.i Author encounters with participants/waves/spawns/objectives/ triggers/space/resources/difficulty/rewards/fail/reset/pacing/safety and dependency state. Evidence: encounter tests.
    • Y5A.12.j Author inventory/item/equipment/crafting/economy and combat stats/actions/damage/status/targeting/AI/animation/VFX/audio hooks as typed public interfaces. Evidence: hook tests.
    • Y5A.12.k Build deterministic seeded headless simulation with exact design/runtime versions, scenarios/bots, progress/cancel/traces/metrics/ counterexamples and bounded resources. Evidence: simulation fixtures.
    • Y5A.12.l Build playtest config/build/participant/consent/telemetry/ feedback/replay/privacy linkage and compare synthetic versus human evidence. Evidence: playtest tests.
    • Y5A.12.m Build balance analysis for win/loss/time/resources/difficulty/ economy/encounter/NPC outcomes with uncertainty/cohort privacy and explainable human tuning decisions. Evidence: balance tests.
    • Y5A.12.n Run harmful mechanic/dark pattern/age/violence/chat/NPC manipulation/bias/accessibility/content safety and human review gates. Evidence: safety matrix.
    • Y5A.12.o Map canonical design to supported Godot/Unreal/other engine profiles through Bellona with stable IDs/versioned adapters/loss report/ import/read-back. Evidence: engine bundle suite.
  • Y5A.13 Complete virtual-production and XR project workflows for remote actor capture, stage/session planning, camera tracking, timecode/genlock, LED/ICVFX, spatial review, device/target profiles, recording, dailies, and Bellona handoff. Evidence: capability-gated host/device matrix.
    • Y5A.13.a Define VP/XR project/stage/session/device/performer/camera/ tracking/timebase/display/render/recording/artifact/profile/handoff contracts and ownership. Evidence: schemas.
    • Y5A.13.b Build remote actor invite/auth/consent/identity/voice-likeness/ device/network/calibration/rehearsal/capture/monitor/stop/revocation and Aja motion handoff. Evidence: remote capture tests.
    • Y5A.13.c Build stage/session plan for location/stage volume/date/time/ participants/crew/roles/shots/assets/sets/equipment/devices/safety/ permits/schedule/call sheet. Evidence: planning tests.
    • Y5A.13.d Build camera/lens/sensor/distortion/focus/pose/tracking system/ markers/origin/coordinate/units/calibration/quality/drift/redundancy/ recording. Evidence: camera tracking tests.
    • Y5A.13.e Build timecode/genlock/PTP/clock/source/rate/drop-frame/ synchronization/calibration/offset/drift/failover and per-device status. Evidence: sync tests.
    • Y5A.13.f Build LED/ICVFX wall/processor/panel/canvas/nDisplay or profile/ camera frustum/color pipeline/latency/genlock/content/lighting/heat/ safety configuration. Evidence: LED tests.
    • Y5A.13.g Build XR scene/space/anchor/scale/origin/tracking/input/ locomotion/interaction/passthrough/performance/comfort/accessibility/ safety authoring. Evidence: XR tests.
    • Y5A.13.h Build spatial review for synchronized multiuser scene/assets/ cameras/takes/annotations/measurements/versions with desktop/2D semantic alternative. Evidence: spatial review tests.
    • Y5A.13.i Define host/device/OS/runtime/plugin/firmware/model/capability/ input-output/latency/accuracy/profile/target support matrix and honest unavailable/degraded states. Evidence: capability tests.
    • Y5A.13.j Build multistream video/audio/motion/tracking/timecode/metadata/ camera/device recording with immutable take IDs/checksums/proxies/ dropped-frame status/consent. Evidence: recording tests.
    • Y5A.13.k Build take/dailies ingest/sync/playback/compare/notes/select/ reject/retake/promotion and exact stage/shot/performer/device lineage. Evidence: dailies tests.
    • Y5A.13.l Handoff stage/project/assets/cameras/tracks/takes/builds through Bellona public contracts and returned artifacts/status/loss/recovery; no duplicate bridge authority. Evidence: handoff tests.
    • Y5A.13.m Run capability-gated session from plan/calibrate/capture/review/ handoff/recovery across supported host/device cells. Evidence: matrix E2E.
  • Y5A.14 Complete unified search across projects, assets, scripts, shots, characters, worlds, locations, tasks, people, comments, jobs, plugins, models, and marketplace items with permission-safe facets/snippets. Evidence: protected, unreleased, and cross-tenant data cannot leak through counts, suggestions, or caches.
    • Y5A.14.a Define versioned searchable-kind registry and per-kind index document with canonical ID/revision/tenant/project/owner/status/locale/ classification/searchable/display-only/forbidden fields. Evidence: schemas.
    • Y5A.14.b Implement project/asset/script/shot/character/world/location/ task documents and entity-specific title/text/tags/relations/status/ rights/canon/production metadata. Evidence: content index tests.
    • Y5A.14.c Implement people/comment/job/plugin/model/marketplace documents with least privilege, no secrets/private prompt/spoiler/biometric/ billing/protected evidence. Evidence: platform index tests.
    • Y5A.14.d Build transactional/outbox-driven create/update/version/ permission/membership/publish/quarantine/archive/delete events with checkpoints/idempotency/order/rebuild reconciliation. Evidence: index lifecycle tests.
    • Y5A.14.e Implement text/exact/fuzzy/tag/entity/relation/reference search with allowlisted query grammar, stable ranking, locale analysis, typo/synonym rules, and version/source. Evidence: query tests.
    • Y5A.14.f Implement permission-safe kind/project/status/date/owner/tag/ asset type/production/marketplace/plugin/model facets and stable signed cursor pagination. Evidence: facet tests.
    • Y5A.14.g Generate snippets/highlights only from authorized fields after property checks and suppress spoiler/protected-source/solution/unreleased context. Evidence: snippet tests.
    • Y5A.14.h Generate autocomplete/typeahead/recent/popular/recommendations with authorization/privacy thresholds and no hidden-title/existence inference. Evidence: suggestion tests.
    • Y5A.14.i Apply current route/object/property/function authorization after every hit and before counts/facets/cache, not solely index ACL snapshots. Evidence: authorization tests.
    • Y5A.14.j Scope query/result/count/facet/suggestion caches by actor/ tenant/project/roles/permissions/locale/version and purge on access/ lifecycle changes. Evidence: cache tests.
    • Y5A.14.k Test colliding tenant/project/title/tag/ID/timestamps, guessed protected terms, membership revoke, stale index, timing/count/empty/ forbidden differences. Evidence: adversarial suite.
    • Y5A.14.l Load-test production-scale mixed kinds/facets/permissions/index lag/rebuild while meeting search SLO and no leakage. Evidence: performance report.
  • Y5A.15 Complete asset, plugin, workflow/template, and AI-model marketplace flows for submission, packaging, compatibility, security/license review, pricing/entitlement, install/update/remove, ratings/reporting, creator attribution/payout, deprecation, takedown, and project dependency impact. Evidence: marketplace content cannot bypass runtime or release policy.
    • Y5A.15.a Define marketplace enablement/audience/region/tenant/host and separate versioned item contracts for asset, plugin, pipeline/workflow/ template, and model/modifier with lifecycle. Evidence: schemas.
    • Y5A.15.b Build creator enrollment/identity/authority/tax/payout/terms/ rights/contact/security and organization/team publishing roles. Evidence: creator tests.
    • Y5A.15.c Build item/version submission with immutable source/artifact/ hash/dependencies/metadata/docs/examples/license/model card/provenance/ pricing/targets/creator attribution. Evidence: submission tests.
    • Y5A.15.d Build deterministic safe package/manifest/signature/checksum/ relative paths/permissions/entry points/dependency lock and independent unpack/install validation. Evidence: packaging tests.
    • Y5A.15.e Build per-host/version/profile/OS/runtime/engine/DCC/model/ workflow/asset/project-type input-output compatibility and declared loss/ unsupported cells. Evidence: matrix.
    • Y5A.15.f Quarantine and run malware/archive/path/serialization/code/ dependency/vulnerability/signature/sandbox/network/resource/security review by item type. Evidence: security tests.
    • Y5A.15.g Review source/ownership/license/rights/consent/training data/ attribution/territory/term/derivative/commercial/safety/quality/ accessibility and human approvals. Evidence: trust tests.
    • Y5A.15.h Build offers/prices/currencies/taxes/fees/promotions/refunds/ trials/seats/usage/region/term and server-side entitlement grant/check/ revoke. Evidence: commerce tests.
    • Y5A.15.i Build download/install with dependency/permission/impact/ confirmation, isolated staging, verification, lockfile, project linkage, and rollback. Evidence: install tests.
    • Y5A.15.j Build update/migration/change log/compatibility/project impact/ staged canary/backup/rollback and remove/disable constraints for active dependencies. Evidence: update/remove tests.
    • Y5A.15.k Build verified-purchase/use ratings/reviews, moderation, version relevance, anti-manipulation, replies, reporting/appeal and no secret/project leakage. Evidence: rating tests.
    • Y5A.15.l Preserve creator/item/version/source attribution through installation/use/derivative/generation/build/release and calculate royalty/fee/hold/refund/chargeback/payout ledger. Evidence: attribution/ payout tests.
    • Y5A.15.m Build deprecate/supersede/end-of-support/revoke/quarantine/ takedown/appeal/restore with consumer notices/download/entitlement/ project/build/runtime/cache/payout effects. Evidence: lifecycle tests.
    • Y5A.15.n Query exact project dependency/blast radius for item/version/ transitive dependencies and preview breakage/migration/rebuild/recall before mutation. Evidence: impact tests.
    • Y5A.15.o Revalidate marketplace item/license/entitlement/security/ compatibility at runtime/build/release and prohibit listing/review/ purchase from satisfying production gates. Evidence: boundary tests.
    • Y5A.15.p Run submission-to-review-to-purchase/install/use/build/release/ update/takedown/payout journey with cross-tenant/fraud/security failures. Evidence: E2E.
  • Y5A.16 Complete organization, team, user, RBAC, invitations, API keys, enterprise plan/entitlement, audit/compliance, analytics, webhook, and integration administration. Evidence: property-level authorization, separation of duties, billing/entitlement consistency, and audit export.
    • Y5A.16.a Define organization/workspace/team/user/membership/role/ permission/entitlement/key/webhook/integration/admin contracts and canonical ownership. Evidence: schemas.
    • Y5A.16.b Build organization create/profile/domains/regions/locale/ security/data/retention/billing owner/transfer/suspend/archive/delete with step-up and audit. Evidence: org tests.
    • Y5A.16.c Build hierarchical/static/dynamic project teams, membership/ manager/guest/service account/role inheritance and cycle/scope controls. Evidence: team tests.
    • Y5A.16.d Build user status/profile/contact/identity provider/MFA/device/ sessions/recovery/lock/suspend/delete/export with privacy/redaction. Evidence: user tests.
    • Y5A.16.e Build typed RBAC roles/permissions/resource/property/function/ conditions/deny/precedence/effective inspector/change preview/version/ assignment/revocation and code floors. Evidence: RBAC tests.
    • Y5A.16.f Build invite create/role/scope/expiry/domain/seat/send/resend/ accept/decline/revoke/conflict/duplicate/transfer and anti-enumeration. Evidence: invitation tests.
    • Y5A.16.g Build API key/service credential purpose/scope/project/ operations/network/expiry/create-once-secret/rotate/revoke/last-use/ anomaly and no UI/log secret. Evidence: key tests.
    • Y5A.16.h Build enterprise plan/seat/storage/compute/provider/feature/ support/region/marketplace limits and server-side effective entitlement from billing/contract/manual grant. Evidence: entitlement tests.
    • Y5A.16.i Reconcile subscription/invoice/payment/credit/refund/usage/ purchase/payout/plan changes to entitlements idempotently without access ambiguity. Evidence: billing consistency tests.
    • Y5A.16.j Build immutable audit for auth/admin/content/asset/job/spend/ marketplace/build/release/integration actions with protected fields, search/export/retention/tamper evidence. Evidence: audit tests.
    • Y5A.16.k Build compliance policy/control/evidence/request/attestation/ exception/retention/legal hold/export and framework mappings without self-certification. Evidence: compliance tests.
    • Y5A.16.l Build privacy-safe organization/project/usage/cost/performance/ collaboration/marketplace analytics with metric definitions/cohort thresholds/role filters and no worker surveillance. Evidence: analytics tests.
    • Y5A.16.m Build webhook endpoint/event filters/signing secret ref/ verification/retry/backoff/rate/dedup/order/dead letter/replay/log/ rotate/disable and payload redaction. Evidence: webhook tests.
    • Y5A.16.n Build approved integration install/config/OAuth/scopes/ mapping/sync/direction/status/log/revoke/data deletion through public adapters. Evidence: integration tests.
    • Y5A.16.o Enforce maker-checker/separation for ownership/security/RBAC/ billing/spend/key/compliance/export/kill actions and property-level auth throughout. Evidence: authority suite.
    • Y5A.16.p Export deterministic redacted signed audit/compliance/admin evidence and run organization lifecycle/tenant/billing/entitlement E2E. Evidence: audit export journey.
  • Y5A.17 Disposition community/showcase features explicitly: creator profile, publish/share, follow, comment, remix/fork, moderation, report, attribution, rights, privacy, and takedown if retained; otherwise define their separate host or retirement. Evidence: no orphan community capability or accidental publication path.
    • Y5A.17.a Inventory all community/showcase routes/components/APIs/stores/ events/jobs/packages/flags/navigation/consumers/tests and live/fixture state at atomic operation level. Evidence: inventory.
    • Y5A.17.b Decide retain-in-Yemaya, move-to-separate-host/service, retire, or explicitly defer per capability with audience/owner/data/authority/ migration/expiry. Evidence: decision record.
    • Y5A.17.c If retained, build creator profile identity/display/links/ portfolio/visibility/verification/block/mute/privacy and no private project leakage. Evidence: profile tests.
    • Y5A.17.d If retained, build explicit publish candidate/review/gates/ visibility/audience/license/embargo/share-link/revoke/read-back rather than making project/asset visibility publish. Evidence: publish tests.
    • Y5A.17.e If retained, build follow/unfollow/request/block/mute/feed/ notification/privacy/rate/anti-enumeration with no authority inheritance. Evidence: social graph tests.
    • Y5A.17.f If retained, build comments/replies/mentions/reactions/edit/ delete/report/moderation/notifications with audience/asset-version/ accessibility and abuse controls. Evidence: interaction tests.
    • Y5A.17.g If retained, build remix/fork eligibility/license/attribution/ source-version/copy scope/new project identity/rights/gates and source revocation effects. Evidence: remix tests.
    • Y5A.17.h If retained, build content/user moderation classifier/rule/ human queue/uncertainty/decision/appeal/sanction/audit without auto- publication or false accusation. Evidence: moderation tests.
    • Y5A.17.i If retained, build report type/subject/evidence/reporter safety/ triage/status/anti-abuse/privacy/escalation and no reporter authority. Evidence: report tests.
    • Y5A.17.j Preserve creator/source/project/asset/model/tool/version/ derivative attribution and license/consent/rights/territory/term/ disclosure through showcase/remix/download. Evidence: attribution tests.
    • Y5A.17.k Build privacy controls for profile/content/follow/interactions/ discoverability/search/counts/download/contact/minors/blocked users/ retention/deletion. Evidence: privacy tests.
    • Y5A.17.l Build takedown/quarantine/unpublish/appeal/restore with blast radius/caches/links/downloads/remixes/notifications/receipts/history. Evidence: takedown tests.
    • Y5A.17.m If moved, implement signed least-privilege publication/handoff/ return contracts and remove duplicate UI/data authority. Evidence: separate-host tests.
    • Y5A.17.n If retired, disable creation/discovery/mutation, migrate/export/ retain/delete data per policy, remove routes/flags/dependencies, and verify no accidental publication path. Evidence: retirement tests.
    • Y5A.17.o Add bidirectional inventory/route/dependency gate proving no orphan community capability or unreviewed live surface. Evidence: disposition gate.
  • Y5A.18 Complete developer surfaces for REST/WebSocket APIs, webhooks, CLI, TypeScript/Python/C++ SDKs, documentation playground, capability discovery, authentication, examples, rate limits, versioning, and deprecation. Evidence: generated contracts and end-to-end external-client conformance for every supported operation.
    • Y5A.18.a Inventory/version every supported external operation/resource/ event/stream/command, audience/scope, owner, route/protocol, stability, consumer, and evidence. Evidence: developer surface registry.
    • Y5A.18.b Generate/validate REST OpenAPI for auth/scopes, requests/ responses/errors, pagination/idempotency/concurrency, uploads/downloads, versions/units, examples, and redactions. Evidence: REST contracts.
    • Y5A.18.c Define WebSocket/SSE connection/auth/subscribe/filter/snapshot/ cursor/event/order/dedup/reconnect/backpressure/heartbeat/error/version/ close semantics. Evidence: realtime contracts.
    • Y5A.18.d Implement webhook event catalog/payload/version/signature/ rotation/replay/retry/order/dedup/rate/redaction/test/reconciliation semantics. Evidence: webhook contracts.
    • Y5A.18.e Build CLI config/profile/auth/project context, typed commands, input/output formats, interactive/noninteractive, dry run/confirm, progress/cancel, exit codes, version update, and no secret logs. Evidence: CLI tests.
    • Y5A.18.f Generate TypeScript SDK models/clients/pagination/streams/ uploads/errors/retries/idempotency/examples and package/version/browser- Node compatibility. Evidence: TS conformance.
    • Y5A.18.g Generate Python SDK sync/async models/clients/iterators/streams/ uploads/errors/retries/idempotency/examples and supported runtime versions. Evidence: Python conformance.
    • Y5A.18.h Generate C++ SDK typed models/client/auth/async/callback/ streaming/binary transfer/errors/cancellation/build systems/platform ABI support. Evidence: C++ conformance.
    • Y5A.18.i Build documentation from canonical contracts with operation/ concepts/auth/permissions/lifecycle/failure/limits/version/migration/ security guides and runnable examples. Evidence: doc drift tests.
    • Y5A.18.j Build isolated playground using synthetic/sandbox tenant and bounded scoped credentials, editable requests, response/event inspector, rate/cost warning, reset, and no production secret/data access. Evidence: playground tests.
    • Y5A.18.k Build audience/authenticated capability discovery for operations/ asset kinds/hosts/providers/targets/limits/versions/degraded states without hidden feature/provider leakage. Evidence: capability tests.
    • Y5A.18.l Implement OAuth/API key/service identity/session/device flows, scopes/project/tenant/resource/property authorization, rotation/revoke/ expiry, and step-up for dangerous actions. Evidence: auth suite.
    • Y5A.18.m Enforce request/data/compute/cost/concurrency/stream/webhook limits with headers/retry/reset, tenant fairness, and no bypass through batch/SDK/alternate protocol. Evidence: rate-limit tests.
    • Y5A.18.n Define URL/schema/event/SDK semantic versioning, compatibility windows/negotiation, additive/breaking policy, changelog, migration, and consumer testing. Evidence: version matrix.
    • Y5A.18.o Define deprecation announce/headers/docs/telemetry/owner/ replacement/sunset/block/exception and remove only after consumer gate. Evidence: deprecation tests.
    • Y5A.18.p Run generated contract tests and real external TS/Python/C++/ CLI/webhook/realtime journeys for every supported operation; fail untested cells. Evidence: conformance matrix.
  • Y5A.19 Complete production safety for prompts/media, model/provider use, consent/likeness, rights, content classification, age rating, moderation, project policy, incident quarantine, recall, and audit across every AAA lane. Evidence: a feature-specific policy matrix; generic upload moderation is not accepted as proof.
    • Y5A.19.a Enumerate every AAA feature/lane/input/output/action/asset kind/ host/audience/target/provider and map applicable safety controls, owners, stages, and proof. Evidence: policy matrix.
    • Y5A.19.b Validate prompt/brief/reference/retrieved/imported content for injection, disallowed intent, identity/spoiler/protected-source leakage, unsafe tools, and policy before planning. Evidence: prompt tests.
    • Y5A.19.c Moderate generated/uploaded/captured/edited image/video/audio/ voice/3D/text/code/project/package metadata at ingest, preview, promotion, build, release with uncertainty. Evidence: media tests.
    • Y5A.19.d Admit models/providers/endpoints/workflows/nodes/adapters only with identity/version/hash/license/training provenance/security/ evaluation/data-use/residency/target approval and revalidate. Evidence: model/provider tests.
    • Y5A.19.e Enforce subject/performer/actor/voice/face/likeness/biometric/ capture consent with scope/purpose/use/territory/term/disclosure/ revocation at every derivative/use/release. Evidence: consent tests.
    • Y5A.19.f Enforce asset/source/music/script/reference/model/license/ ownership/use/adaptation/training/commercial/distribution/attribution/ territory/term/embargo rights. Evidence: rights tests.
    • Y5A.19.g Classify content by violence/sexual/hate/self-harm/child/ substances/language/fear/gambling/user interaction/context/intensity and preserve raw evidence. Evidence: classification tests.
    • Y5A.19.h Map platform/region/audience/project type/content/interactivity to age-rating descriptors/evidence/reviewer/label/target requirements. Evidence: rating matrix.
    • Y5A.19.i Build input/output/project/user/marketplace/community moderation rules/classifiers/human queue/decisions/appeals/uncertainty/ anti-bias/privacy and no automatic accusation. Evidence: moderation tests.
    • Y5A.19.j Build platform/region/organization/project/profile/campaign typed policy overlays that only tighten code floors, with effective preview/version/approval/freshness. Evidence: policy property tests.
    • Y5A.19.k Apply safety checks at feature-specific typed context points for writing, boards, design, gameplay/NPC, VP/XR/capture, DCC/engine, rendering/post, collaboration, agents, marketplace/community, developer/export/release. Evidence: lane coverage tests.
    • Y5A.19.l Create incident case/scope/severity/evidence/authority and quarantine affected model/provider/workflow/source/asset/project/build/ marketplace/release from new use. Evidence: quarantine tests.
    • Y5A.19.m Compute blast radius and recall/unpublish/disable downstream projects/assets/derivatives/builds/gallery/marketplace/community/ external targets/caches/links with receipts. Evidence: recall tests.
    • Y5A.19.n Persist immutable decision/override/exception/scan/model/policy/ action/notification/verification audit with protected-field controls and nonwaivable floors. Evidence: audit tests.
    • Y5A.19.o Build feature-specific calibrated/adversarial fixtures and fail every matrix cell relying only on generic upload moderation. Evidence: safety coverage gate.
  • Y5A.20 Complete animation learning and motion authoring for references, timing/exposure charts, poses/keyframes, arcs/spacing, cleanup, retiming, camera, performance notes, motion assets, Aja capture/retarget handoff, preview/onion-skin or semantic equivalent, critique, versioning, and export. Evidence: reference→motion→approved shot lineage and accessible timing/curve alternatives.
    • Y5A.20.a Define animation project/sequence/shot/character/rig/control/ clip/layer/pose/key/curve/annotation/reference/motion asset/version/ target contracts and units/timebases. Evidence: schemas.
    • Y5A.20.b Build reference ingest/record/link/trim/sync/loop/annotation/ rights/consent/source/performer/camera/timecode and side-by-side study. Evidence: reference tests.
    • Y5A.20.c Build exposure/timing sheet with frame/timecode/beat/phoneme/ action/pose/dialogue/audio/camera tracks, holds/inbetweens/accents/ breakdowns, versions, and list/table parity. Evidence: timing tests.
    • Y5A.20.d Build pose/keyframe authoring for rig/control/body/face/hands/ props/contacts/constraints/space/IK-FK/keys/interpolation/tangents with stable identities. Evidence: keyframe tests.
    • Y5A.20.e Build motion path/arcs/spacing/velocity/acceleration/easing/ silhouette/line-of-action/weight/balance/contact trajectory visualization and measurable diagnostics. Evidence: arc/spacing tests.
    • Y5A.20.f Build cleanup for curve noise/key reduction/tangent/foot slide/ penetrations/jitter/pops/constraints/contacts/rig errors as reversible layers with before-after. Evidence: cleanup tests.
    • Y5A.20.g Build retiming/ripple/stretch/compress/hold/ramp/offset/loop/ frame-rate conversion/audio sync with key/event/annotation/shot impact preview. Evidence: retime tests.
    • Y5A.20.h Build camera shot/lens/framing/pose/path/focus/shake/cut/ continuity/constraints and motion-camera/Bellona engine-DCC handoffs. Evidence: camera tests.
    • Y5A.20.i Build performance notes with character/shot/frame-range/body/ facial/voice/emotion/intent/reference anchors, assignments/resolution, and actor/consent privacy. Evidence: note tests.
    • Y5A.20.j Build motion asset clip/take/rig/skeleton/source/performer/ consent/units/rate/root motion/contacts/tags/versions/gates/library/ promotion. Evidence: motion asset tests.
    • Y5A.20.k Handoff capture/calibration/solve/clean/retarget jobs to Aja public contracts with exact project/performer/rig/shot context, progress, returned lineage, and no duplicate motion pipeline. Evidence: Aja tests.
    • Y5A.20.l Build retarget source-target skeleton/rig/map/rest pose/scale/ root/IK/contacts/rules/version/preview/loss/validation and acceptance. Evidence: retarget tests.
    • Y5A.20.m Build synchronized playback/scrub/loop/speed/onion skin/ghost/ trajectory/curve/dope sheet/reference/audio/camera overlays plus semantic timing/key/curve/path tables/forms. Evidence: preview/accessibility tests.
    • Y5A.20.n Build rule/model/human critique for timing/arcs/spacing/contact/ continuity/anatomy/weight/performance/camera/technical quality with exact regions/uncertainty/decisions. Evidence: critique tests.
    • Y5A.20.o Implement immutable versions/layers/diff/branch/merge/restore/ compare/approval and downstream shot/render/build invalidation. Evidence: version tests.
    • Y5A.20.p Export ratified animation/motion formats through Bellona with skeleton/rig/keys/curves/events/camera/units/timebase/materials/ metadata/rights/provenance/loss/read-back. Evidence: export tests.
    • Y5A.20.q Run reference→timing→keys→Aja capture/retarget→cleanup/ critique→approved shot→engine/DCC export journey with accessible alternatives. Evidence: E2E.

Y6 — Persistence, APIs, desktop security, and collaboration#

  • Y6.1 Decide authoritative stores for projects, study content, cases, canon, assets, reviews, playtests, and releases. Evidence: ownership ADR.
    • Y6.1.a Inventory existing databases/schemas/tables/blobs/indexes/caches/ repositories/services/events/backups/production data per aggregate. Evidence: storage inventory. <!-- yemaya-storage-inventory.json: storage inventory of Yemaya. 38 relational aggregates over 4 PrismaClients, 1 schema (prisma/schema.prisma). INDEX-HEAVY: 120 DB indexes (@@index/@@unique) = 3.2 indexes/aggregate, fitting a creative studio's query load. Storage layers present: tables, schemas, indexes (120), blobs (20 Asset object-storage files via storageKey/storageBucket), caches (13 Redis files), events (4 entities: Activity/Notification/Webhook/WebhookDelivery). ARCHITECTURE FINDING: 0 repository + 0 manager classes — Yemaya is SERVICE-DIRECT, consistent with Isis (I6.1.a) and Veritas (V6.1.a): all three domains use service-direct Prisma access with no repository layer (a cross-domain pattern). Backups are ops/infra-level (not modeled in application code, recorded honestly). Precise scoped patterns (storageKey/Bucket, new Redis, class Repository/Manager); DB indexes counted from the schema. generate-yemaya-storage-inventory.mjs + 6 tests, wired into inventory/verify + CI. Y6.1.b (one mutation owner per aggregate) + later Y6.1. are design and stay unchecked. -->
    • Y6.1.b Assign exactly one mutation owner/store and identity/version/ tenant/transaction/retention/security requirements for each named aggregate. Evidence: ownership matrix.
    • Y6.1.c Distinguish canonical records from blobs, derived read/search/ consumer projections, caches, audit/outbox, analytics, and desktop local working copies. Evidence: data-flow map.
    • Y6.1.d Resolve cross-store links/events/consistency/failure/recovery/ deletion/legal hold/backup and prohibit distributed shadow writes. Evidence: consistency design.
    • Y6.1.e Ratify owner/store/migration/operational owner/rejected alternative ADR and add schema/import ownership checks. Evidence: approved ADR.
  • Y6.2 Add missing explicit-column schemas and safe migrations. Evidence: production-size rehearsal.
    • Y6.2.a Diff ratified contracts against owner schemas and classify reusable fields versus genuine missing tables/columns/constraints/ indexes. Evidence: gap report.
    • Y6.2.b Separate tenant/identity/state/version/owner/time/lock/freshness/ classification/query-critical columns from immutable typed content payloads. Evidence: schema design.
    • Y6.2.c Add expand migrations and backward-compatible read/write/ conversion code without destructive rename/drop/default ambiguity. Evidence: expand tests.
    • Y6.2.d Backfill resumably in bounded tenant-safe batches with checkpoints, errors/quarantine, throttling, and observable progress. Evidence: backfill tests.
    • Y6.2.e Verify counts/FK/unique/check/null/hash/semantic/read/query-plan/ divergence invariants before cutover. Evidence: verification report.
    • Y6.2.f Cut over reversibly, monitor, then contract after rollback window/ consumer inventory; rehearse at production scale/locks/resources. Evidence: migration rehearsal.
  • Y6.3 Implement durable stores with optimistic concurrency and tombstones. Evidence: real-database tests.
    • Y6.3.a Define store interfaces/transactions for exact aggregate/revision/ relation/read/list/transition/restore operations and authorization context. Evidence: repository contracts.
    • Y6.3.b Persist immutable revisions, operational projection, audit, and outbox atomically before acknowledge with tenant-scoped constraints. Evidence: transaction tests.
    • Y6.3.c Enforce expected revision/ETag/lease/fencing and return safe conflicts; prohibit last-write-wins or stale worker commits. Evidence: concurrency tests.
    • Y6.3.d Represent delete/archive/unlink through typed tombstone with actor/reason/time/prior revision/retention/restore scope rather than dangling hard delete. Evidence: tombstone tests.
    • Y6.3.e Make reads/relations/search/events respect tombstone visibility and restore as new revision without resurrecting invalid dependencies. Evidence: lifecycle tests.
    • Y6.3.f Test real database restart/deadlock/duplicate/retry/branch/merge/ retention/two-tenant/production-size paths. Evidence: integration suite.
  • Y6.4 Implement authoring APIs through versioned public contracts. Evidence: web/desktop consumer-driven contracts.
    • Y6.4.a Inventory every project/study/case/canon/source/asset/review/ playtest/build/release operation and canonical command/query/event/ stream contracts. Evidence: API inventory.
    • Y6.4.b Implement versioned list/detail/create-revision/transition/ relationship/comment/review/job/import-export operations with stable errors. Evidence: API tests.
    • Y6.4.c Enforce route/function/object/property/tenant/project/spoiler/ protected-source authorization, idempotency, optimistic concurrency, pagination, rate/budget. Evidence: security tests.
    • Y6.4.d Generate OpenAPI/realtime/IPC-facing shared schemas with versions/ units/discriminators/absence-redaction and no renderer-local shapes. Evidence: contract generation.
    • Y6.4.e Verify web and desktop consumers for success/empty/stale/conflict/ offline/degraded/not-configured/forbidden/failure/version skew. Evidence: consumer-driven suite.
  • Y6.5 Implement compiler/verifier/eval/canon jobs as durable work. Evidence: retry, cancellation, lease-loss, and stale-input tests.
    • Y6.5.a Define job/plan/stage/DAG/attempt/lease/checkpoint/input/output/ progress/error/cost/event contracts and separate compiler/verifier/eval/ canon operation schemas. Evidence: job schemas.
    • Y6.5.b Create idempotently with exact immutable project/case/study/ source/canon/policy/tool versions and transaction-bound audit/outbox. Evidence: submission tests.
    • Y6.5.c Claim with fenced lease/heartbeat/expiry/CAS and checkpoint only validated stage artifacts/hashes/consumed versions. Evidence: worker tests.
    • Y6.5.d Retry only explicit idempotent eligible stage/new attempt with limits/backoff; resume only supported checkpoints. Evidence: retry tests.
    • Y6.5.e Cancel queued/running/provider work with late-result fencing, resource cleanup, partial evidence and explicit outcome. Evidence: cancellation tests.
    • Y6.5.f Detect lease loss/split brain/worker restart/out-of-order callbacks and prevent stale commits/duplicate assets/cost. Evidence: race tests.
    • Y6.5.g Revalidate input/canon/source/policy/tool/dependency freshness before stage/commit and reject stale results. Evidence: stale-input tests.
  • Y6.6 Implement permission-aware search across study and case content. Evidence: spoiler/protected-source leak tests.
    • Y6.6.a Define per-kind index documents/fields/classification for study sources/notes/materials/concepts and case entities/scenes/dialogue/clues/ canon/assets with exact revisions. Evidence: schemas.
    • Y6.6.b Build outbox-driven version/tombstone/permission/reveal/ retraction/publication indexing and reconciliation. Evidence: lifecycle tests.
    • Y6.6.c Implement allowlisted query/filter/facet/sort/snippet/suggestion/ cursor operations with locale analysis and current object/property authorization. Evidence: query tests.
    • Y6.6.d Exclude/redact protected source/solution/secret/reveal/future clue/ unpublished/canon dispute/playtest participant fields from indexes, counts, snippets, suggestions, caches, logs. Evidence: leak tests.
    • Y6.6.e Test actor reveal progression/role change/project switch/offline cache/stale index/colliding tenants/timing/count inference. Evidence: adversarial suite.
  • Y6.7 Harden desktop IPC with an allowlisted, versioned, validated command surface. Evidence: arbitrary channel/message tests fail.
    • Y6.7.a Inventory every renderer-preload-main/worker/native IPC channel/ direction/message/handler/privilege/caller and eliminate unregistered channels. Evidence: IPC registry. <!-- yemaya-ipc-registry.json: IPC registry of the Yemaya studio-desktop (apps/yemaya/studio-desktop/src). 280 registered channels across 28 main modules (277 ipcMain.handle = invoke/bidirectional, 3 ipcMain.on = send/one-way). DIRECTIONAL orphan detection (the "eliminate unregistered channels" ask): split renderer→main REQUESTS (75 ipcRenderer.invoke/send) from renderer LISTENERS (23 ipcRenderer.on) from main→renderer EMISSIONS (24 webContents/sender.send). ORPHANS: 9 request channels invoked but NOT registered as handlers (bellona:connect/export/sync-assets, dialog:message/ open-directory/open-file, ...) = the broken/unregistered channels to eliminate; 19 renderer listeners with no detected main emission (dead-listener candidates); 214 handlers with no LITERAL invoke (mostly internal/module-to- module). CRITICAL FIX: initially conflated events (bellona:engine-* via ipcRenderer.on) with broken invokes → 32 false orphans; separating by direction (invoke/send vs on vs main-send) gives the accurate 9. dynamicInvoke caveat recorded. generate-yemaya-ipc-registry.mjs + 6 tests, wired into inventory/verify + CI. Y6.7.b (typed preload API) + Y6.7.c (validation) are implementation/design and stay unchecked. -->
    • Y6.7.b Expose context-isolated preload API with stable versioned typed command/query/event capabilities and no raw ipcRenderer/Node/Electron objects. Evidence: preload tests.
    • Y6.7.c Validate channel/version/schema/size/depth/count/unknown fields, origin/window/webContents/session, actor/project/capability/permission at main process. Evidence: validation tests.
    • Y6.7.d Define structured errors/timeouts/cancellation/idempotency/ concurrency/progress/event unsubscribe and protect secrets/paths. Evidence: protocol tests.
    • Y6.7.e Sandbox navigation/window creation/webviews/devtools/clipboard/ shell/external URL/file/device/native modules according to policy. Evidence: Electron security tests.
    • Y6.7.f Fuzz arbitrary/unknown/nested/alias/mass-assignment/oversize/ prototype/serialized object/replay/cross-window messages. Evidence: negative suite.
  • Y6.8 Harden local file access with user grants, canonical paths, safe archives, symlink controls, and scoped bookmarks. Evidence: malicious filesystem corpus.
    • Y6.8.a Define typed open/save/import/export/project/cache/temp operation, allowed file/dir/profile, user gesture, purpose, read/write/create/ overwrite/delete scope, and lifecycle. Evidence: file contract.
    • Y6.8.b Obtain OS picker/grant/bookmark/token with opaque reference, project/actor/device binding, expiry/revocation and never trust renderer raw path. Evidence: grant tests.
    • Y6.8.c Canonicalize existing/nonexisting path parent, normalize Unicode/ case/separators, enforce allowlisted roots/extensions, and reject traversal/absolute/device/network/ADS/special paths. Evidence: path corpus.
    • Y6.8.d Resolve/check every path component and reject symlink/junction/ hardlink/alias/mount/race escape using safe open semantics. Evidence: link/TOCTOU tests.
    • Y6.8.e Parse archives with entry/count/depth/size/ratio/path/symlink/ duplicate/case collision/encryption/hidden/executable protections into temp quarantine. Evidence: archive tests.
    • Y6.8.f Implement atomic save/rename/fsync/backup/overwrite confirmation/ partial cleanup and hash/read-back. Evidence: write tests.
    • Y6.8.g Store scoped bookmarks/recents/local project grants securely, reauthorize on use, revoke/logout/device-loss cleanup, and redact logs. Evidence: bookmark tests.
  • Y6.9 Harden protocol/deep-link handlers against command injection and cross-tenant replay. Evidence: negative suite.
    • Y6.9.a Define allowlisted custom/HTTPS link schemes, hosts/routes/intents, canonical grammar, minimal signed claims, max sizes, and version. Evidence: link registry.
    • Y6.9.b Parse without shell/interpolation, reject credentials/fragments/ unknown params/duplicate keys/control chars/Unicode confusables/encoded traversal/nested URLs. Evidence: parser tests.
    • Y6.9.c Validate signature/key/issuer/audience/intent/tenant/project/ subject/version/expiry/not-before/nonce/replay/host-instance before navigation/action. Evidence: validation tests.
    • Y6.9.d Reauthenticate/reauthorize current actor/object/property/function and require confirmation/step-up for allowed mutations; links never carry commands/scripts/paths/secrets. Evidence: authority tests.
    • Y6.9.e Handle single-instance startup/queued links/login continuation/ window focus/concurrent links/expired/deleted objects safely. Evidence: lifecycle tests.
    • Y6.9.f Test shell metacharacters/URL injection/open redirect/replay/ tenant substitution/role revoke/old links/offline cache. Evidence: negative suite.
  • Y6.10 Harden updater, signing, and version rollback. Evidence: signature, channel, expiry, downgrade, and interrupted-update tests.
    • Y6.10.a Define signed update manifest/artifact/delta/full package with product/platform/arch/channel/version/min OS/server/schema/published/ expiry/hash/size/URL/release notes/rollback rules. Evidence: schema.
    • Y6.10.b Verify HTTPS pinned/allowlisted source, manifest/artifact hash/ signature/certificate/status/timestamp/key rotation/revocation before staging. Evidence: signing tests.
    • Y6.10.c Enforce stable/beta/internal channel authorization and prevent cross-channel injection, expired/future/stale manifests, mirror/ redirect/DNS abuse. Evidence: channel tests.
    • Y6.10.d Enforce monotonic upgrade and explicit authorized compatible rollback; block version/schema/security downgrade and replay. Evidence: downgrade tests.
    • Y6.10.e Download resumably with size/disk/resource limits, verify, stage atomically, quiesce jobs/files/locks, migrate, restart, and health check. Evidence: update tests.
    • Y6.10.f Recover from network/disk/power/process/signature/install/ migration/health failure to previous intact version/data with audit. Evidence: interrupted-update tests.
    • Y6.10.g Test old/new app-server-protocol-project compatibility and forced security update/offline behavior. Evidence: compatibility suite.
  • Y6.11 Implement presence and collaboration across web and desktop. Evidence: reconnect, stale presence, and conflict journey.
    • Y6.11.a Use the shared collaboration protocol/CRDT/session/auth/version contracts through web transport and desktop adapter without forked semantics. Evidence: parity tests.
    • Y6.11.b Implement permission/spoiler-filtered presence/heartbeat/expiry/ participant/tool/viewport and stable cursor/selection awareness across content/graph/timeline/media. Evidence: presence tests.
    • Y6.11.c Implement durable CRDT operations/snapshots/compaction/offline/ schema migration with convergence and semantic conflict escalation. Evidence: sync tests.
    • Y6.11.d Reconnect from vector/cursor/snapshot with auth refresh/dedup/ gap/order/role revoke/project lock and no duplicate edits. Evidence: reconnect tests.
    • Y6.11.e Mark/remove stale presence promptly without deleting edits or treating disconnect as lock release/approval/cancel. Evidence: stale tests.
    • Y6.11.f Run mixed web/desktop/version/partition/offline/conflict/ cross-project/accessibility journey. Evidence: E2E.
  • Y6.12 Implement anchored comments/suggestions on all stable anchor kinds. Evidence: anchor survival and orphan workflow.
    • Y6.12.a Define comment/thread/reply/suggestion/decision identity, exact subject revision/typed anchor/author/visibility/status/mentions/ attachments/time. Evidence: schema.
    • Y6.12.b Support source/page/text, study/tree/concept, case/entity/scene/ dialogue/clue/timeline/graph, asset/object/region/frame/timecode/shot/ curve anchors. Evidence: kind matrix.
    • Y6.12.c Implement create/reply/edit/suggest/accept/reject/resolve/reopen/ assign/mention/navigate with object/property/spoiler permissions and immutable history. Evidence: interaction tests.
    • Y6.12.d Map anchors on revision as exact/shifted/ambiguous/orphaned/ invalid and preserve original context/hash plus reattach/dismiss workflow. Evidence: survival tests.
    • Y6.12.e Ensure suggestion acceptance creates a new authored revision and comments cannot approve/canonize/publish/change production state. Evidence: authority tests.
    • Y6.12.f Provide synchronized accessible comment/anchor list and test all kinds through web/desktop/offline/merge/locale changes. Evidence: E2E.
  • Y6.13 Implement review assignments, quorum, conflicts, escalation, and notifications. Evidence: role-policy tests.
    • Y6.13.a Define review template/dimension/assignment/reviewer qualification/ locale/role/quorum/sequence/deadline/decision/freshness contracts. Evidence: schemas.
    • Y6.13.b Assign eligible reviewers by project/track/content/gate/risk/ locale/skills/workload/availability with conflict/delegation/substitute rules. Evidence: assignment tests.
    • Y6.13.c Calculate quorum from independent eligible fresh decisions per dimension; handle reject/request-changes/abstain/timeout/reassignment. Evidence: quorum tests.
    • Y6.13.d Capture conflict-of-interest declaration/detected conflicts/ recusal/mitigation/reviewer privacy and block incompatible signoff. Evidence: conflict tests.
    • Y6.13.e Compute SLA/business-calendar/quiet-hours/breach and escalate to owner/substitute/on-call without changing decision authority. Evidence: escalation tests.
    • Y6.13.f Send localized role-filtered in-app/email/webhook notifications with idempotency/preferences/urgent override/ack/redaction. Evidence: notification tests.
    • Y6.13.g Test exact subject/evidence change invalidation, system actor refusal, stale role, two-tenant, and time zones. Evidence: role-policy suite.
  • Y6.14 Implement two-tenant/project isolation across APIs, search, blobs, jobs, desktop cache, playtests, and export. Evidence: integration suite.
    • Y6.14.a Seed two tenants/projects with colliding names/IDs-like slugs/ tags/files/times/entities/assets/canon/solutions/actors/roles and separate keys/policies. Evidence: fixture.
    • Y6.14.b Exercise every API/IPC/deep-link/object/property/function/ membership/role/guest path using foreign IDs/cursors/idempotency keys. Evidence: API tests.
    • Y6.14.c Exercise search/facets/counts/snippets/suggestions/caches/index lag and spoiler/protected-source/solution terms. Evidence: search tests.
    • Y6.14.d Exercise blob/original/proxy/rendition/package/signed URL/hash/ dedupe/cache/temp/local desktop files/bookmarks/offline data. Evidence: storage/cache tests.
    • Y6.14.e Exercise queues/jobs/events/progress/cancel/retry/callback/ provider/tool/worker/log/trace/cost and project switching. Evidence: job tests.
    • Y6.14.f Exercise playtest participant/config/build/session/telemetry/ replay/feedback/consent/retention and reveal state. Evidence: playtest tests.
    • Y6.14.g Exercise project/study/case/asset/audit/standards/marketplace/ evidence/export/package/download/notifications. Evidence: export tests.
    • Y6.14.h Inspect UI/error/timing/count/existence/client state/service worker/clipboard/recent/app updater channels for leakage. Evidence: integration suite.

Y7 — Review, compilation, localization, publication, and catalog#

  • Y7.1 Define review templates for study content and cases. Evidence: source grounding, canon, solvability, safety, accessibility, rights, localization, and producer roles are explicit.
    • Y7.1.a Define versioned template identity/track/profile/risk/locale/target applicability, dimensions/checklists/evidence/roles/quorum/order/ deadlines/expiry/floors. Evidence: schema.
    • Y7.1.b Define study source authority/currency/grounding/citation/ structure/order/accessibility/rights/localization/consumer/producer review criteria. Evidence: study matrix.
    • Y7.1.c Define case canon/continuity/solvability/uniqueness/fairness/ safety/age/accessibility/rights/localization/assets/engine/producer criteria. Evidence: case matrix.
    • Y7.1.d Resolve qualified SME/researcher/canon/solver/safety/accessibility/ legal/localization/producer/publisher roles with separation/conflicts/ quorum. Evidence: role tests.
    • Y7.1.e Bind decisions to exact candidate/evidence/template versions and invalidate affected signoffs on changes. Evidence: freshness tests.
    • Y7.1.f Prove project/season/tenant overlays can add/tighten but never remove code-required dimensions/roles. Evidence: property tests.
  • Y7.2 Build the unified review inbox with track-specific saved views. Evidence: assignment, SLA, gate, and role filters.
    • Y7.2.a Define permission/spoiler-filtered queue item projection for exact candidate/project/track/revision/template/gates/assignment/deadline/ priority/state. Evidence: contract.
    • Y7.2.b Build deterministic filters/facets/sort/pagination for track/ profile/content kind/role/assignee/SLA/status/gate/risk/locale/release. Evidence: query tests.
    • Y7.2.c Build personal/team/track study/case/canon/solver/localization/ production/release saved views with versioned query/columns/order/share/ default. Evidence: saved-view tests.
    • Y7.2.d Build claim/reassign/decline/recuse/request info/open subject/ review action with expected version, workload/conflict/permission/audit. Evidence: action tests.
    • Y7.2.e Present loading/empty/stale/blocked/overdue/degraded/offline/ unauthorized without title/spoiler/count leakage. Evidence: state tests.
    • Y7.2.f Verify keyboard/screen-reader/responsive/live-update/large queue/ two-tenant journey. Evidence: Playwright.
  • Y7.3 Lock review candidates to exact content, canon, source, asset, evaluator, and policy versions. Evidence: dependency change invalidates signoff.
    • Y7.3.a Define candidate manifest containing exact project/content/tree/ graph/timeline/case/solution/locale/build schema revisions and hashes. Evidence: manifest schema.
    • Y7.3.b Include exact canon assertions/namespaces/decisions, sources/ extractions/claims/citations/rights, assets/renditions/provenance/ continuity revisions. Evidence: dependency manifest tests.
    • Y7.3.c Include exact solver/compiler/engine/evaluator/model/corpus/ config/gate/policy/threshold/profile versions and results. Evidence: evidence binding tests.
    • Y7.3.d Canonicalize/hash/lock manifest and bind assignment/comments/ decisions/quorum/approvals/exception to digest. Evidence: lock tests.
    • Y7.3.e Rehash/revalidate on decision/compile/publish and mark approvals stale on any bound dependency/status/rights/policy change. Evidence: invalidation tests.
    • Y7.3.f Route changes to new candidate/revision and retain historical locked candidates/signoffs. Evidence: immutability tests.
  • Y7.4 Compile study read models from approved authoring revisions. Evidence: deterministic golden bundle.
    • Y7.4.a Resolve exact approved study candidate/hierarchy/materials/ sources/concepts/order/activities/assets/locales/accessibility/target and reject missing/stale dependencies. Evidence: input tests.
    • Y7.4.b Normalize/validate content blocks, stable IDs/links/anchors/ citations/rights, reading order/prerequisites, locale fallback, and consumer capabilities. Evidence: compile tests.
    • Y7.4.c Derive navigation/content/search/concept/glossary/activity/media/ accessibility/provenance models and exclude author/review/protected fields. Evidence: projection goldens.
    • Y7.4.d Serialize deterministic manifest/files/order/numbers/Unicode/ timestamps/config with compiler/version/input hashes. Evidence: determinism tests.
    • Y7.4.e Validate bundle schema/links/hashes/rights/gates and rebuild/read through consumer workspace. Evidence: golden/read-back tests.
  • Y7.5 Compile case bundles through the real case compiler and pipeline. Evidence: deterministic golden case.
    • Y7.5.a Resolve exact approved locked case/premise/entities/canon/source/ scenes/dialogue/timeline/clues/solution/branch/state/assets/locales/ profile/target. Evidence: input tests.
    • Y7.5.b Invoke canonical real case compiler through public contract with pinned version/config, durable job/checkpoints/progress/cancel/errors and no workbench substitute. Evidence: integration tests.
    • Y7.5.c Compile stable runtime world/entity/scene/dialogue/timeline/clue/ state/branch/rule/hint/save/ending models and spoiler partitions. Evidence: compiler goldens.
    • Y7.5.d Run canonical production pipeline for assets/renditions/ localization/accessibility/runtime/package/provenance with exact dependencies. Evidence: pipeline tests.
    • Y7.5.e Serialize deterministic bundle/manifest/hashes/order/IDs/config/ tool versions and validate through real engine loader. Evidence: golden case/read-back.
  • Y7.6 Validate case bundles with engine, solver, canon, localization, asset, and schema gates. Evidence: broken-bundle corpus.
    • Y7.6.a Validate bundle/manifest/schema/version/files/hashes/refs/IDs/ types/cardinality/size/resource/safe package and runtime compatibility. Evidence: schema fixtures.
    • Y7.6.b Load/run with real hermetic engine for init/navigation/state/ branches/clues/dialogue/save/load/reset/ending/error/performance and no protected partition leak. Evidence: engine fixtures.
    • Y7.6.c Run solver reachability/existence/uniqueness/alternative/fair timing/dead-end/soft-lock/timeline traces on compiled bundle, not source alone. Evidence: solver fixtures.
    • Y7.6.d Validate canon entity/assertion/namespace/temporal/knowledge/ continuity and exact compiled-to-authoring mapping. Evidence: canon fixtures.
    • Y7.6.e Validate every locale’s structure/IDs/branches/clues/solution/ timing/text/assets/accessibility/spoiler partitions and fallback. Evidence: locale fixtures.
    • Y7.6.f Validate every asset/rendition/dependency/hash/format/target/ technical/right/consent/provenance/gate and no quarantined inputs. Evidence: asset fixtures.
    • Y7.6.g Preserve raw results separately, require all mandatory gates, and maintain broken-bundle corpus for each failure family. Evidence: coverage gate.
  • Y7.7 Build localization packages with spoiler segmentation and role restrictions. Evidence: translators see only authorized context.
    • Y7.7.a Define locale package/job/unit/segment/source-target/revision/ role/spoiler tier/context/asset/status/manifest contracts. Evidence: schemas.
    • Y7.7.b Segment study/case text/metadata/dialogue/clues/hints/reveals/ solution/UI/assets by stable ID, required context, translation memory/ glossary, and spoiler boundary. Evidence: segmentation tests.
    • Y7.7.c Generate minimal role/assignment/locale/segment-scoped package with only authorized context/source/canon/character/timing/asset references, pseudonyms where needed. Evidence: minimization tests.
    • Y7.7.d Encrypt/sign/hash/package, issue expiring access, disable download/ copy as policy permits, watermark, audit, and revoke. Evidence: security tests.
    • Y7.7.e Import translated units/assets with stable IDs, source revision/ TM/glossary/provenance, validation/conflicts, reviewer decision, and no hidden solution fields. Evidence: import tests.
    • Y7.7.f Test translator/reviewer/solver/producer roles, escalation for additional context, cross-locale/package/tenant and expired/revoked access. Evidence: boundary suite.
  • Y7.8 Build locale parity gates for structure, clues, solution, accessibility, assets, and timing. Evidence: known mismatch fixtures.
    • Y7.8.a Resolve source/target locale bundles and exact structure/entity/ scene/branch/clue/solution/asset/timebase/accessibility revisions and profile. Evidence: inputs.
    • Y7.8.b Compare stable hierarchy/order/IDs/refs/conditions/effects/ branches/reachability and detect added/removed/orphaned structure. Evidence: structure fixtures.
    • Y7.8.c Compare clue/red-herring/interpretation/dependency/weight/ discovery/reveal/spoiler and solution/proof/alternative/ending semantics. Evidence: case fixtures.
    • Y7.8.d Validate translated text/dialogue/terminology/pronunciation/canon/ source/citation/meaning and intentional adaptations with review. Evidence: language tests.
    • Y7.8.e Validate alt/long descriptions/captions/transcripts/sign/ noninteractive alternatives/keyboard/reading order and locale metadata. Evidence: accessibility fixtures.
    • Y7.8.f Validate asset role/version/locale/renditions/graphics/audio/voice/ rights/consent and duration/timecode/dialogue/audio/branch timing. Evidence: asset/timing fixtures.
    • Y7.8.g Persist raw mismatches/uncertainty/N/A/adaptations separately and block unresolved mandatory parity. Evidence: parity gate.
  • Y7.9 Publish through plan, execute, read-back verify, and compensate. Evidence: partial target failure recovery.
    • Y7.9.a Define immutable per-target release plan with exact locked bundle/ candidate/revisions/hashes, prerequisites, destination IDs/config, steps/compensation/estimate/idempotency. Evidence: schema.
    • Y7.9.b Revalidate approvals/gates/canon/source/rights/assets/locales/ target/capability/embargo/authorization immediately before confirmation and each write. Evidence: TOCTOU tests.
    • Y7.9.c Execute durable target steps idempotently with checkpoints/ receipts/partial state/retries/timeouts/cancel and no hidden success. Evidence: execution tests.
    • Y7.9.d Read back authoritative target bytes/manifest/hash/metadata/ visibility/links/runtime/consumer models and mark success only after verification. Evidence: read-back tests.
    • Y7.9.e Reconcile interruption/unknown destination state before retry; compensate eligible successful writes in dependency-safe order and verify. Evidence: recovery tests.
    • Y7.9.f Preserve per-target succeeded/failed/unknown/compensated/residual state, all attempts/cost/audit/notifications. Evidence: partial failure E2E.
  • Y7.10 Build release calendar for study drops, case episodes, playtests, localization, and production capacity. Evidence: dependency and time-zone tests.
    • Y7.10.a Define versioned event/resource/reservation/dependency contracts for study drop/case episode/playtest/locale/review/build/release/ embargo/capacity with time zone. Evidence: schema.
    • Y7.10.b Build month/week/list/timeline views with filters, role/spoiler projection and keyboard/non-drag create/move/resize. Evidence: UI tests.
    • Y7.10.c Build dependency DAG/backward plan for source/canon/content/ asset/solver/review/localize/compile/validate/publish/verify tasks and buffers. Evidence: dependency tests.
    • Y7.10.d Reserve authors/solvers/reviewers/localizers/producers/stages/ providers/render/compile/targets/budget and reject committed overbooking. Evidence: capacity tests.
    • Y7.10.e Monitor freshness/slip/conflict/risk and authorize reschedule/ scope/fallback/hold/cancel through versioned impact decision. Evidence: contingency tests.
    • Y7.10.f Test IANA zones/DST/recurrence/quiet hours/embargo/dependency/ reserved slot/actuals and scheduled release read-back. Evidence: suite.
  • Y7.11 Build study collection and case-season curation through bundles. Evidence: no direct catalog writes.
    • Y7.11.a Define collection/season/bundle version, purpose/audience/locale/ owner/schedule/ordered exact study/case episode/build members/metadata/ targets/lifecycle. Evidence: schema.
    • Y7.11.b Build draft add/remove/reorder/group/feature/episode numbering/ arc/dependency/locale/metadata editor and eligibility checks. Evidence: curation tests.
    • Y7.11.c Validate member availability/target/rights/canon/source/gates/ reviews/spoiler/order/prerequisite/season continuity and steering policy. Evidence: membership tests.
    • Y7.11.d Preview deterministic catalog/consumer/search/navigation/order/ schedule/metadata/add/remove diffs and impacted links. Evidence: preview.
    • Y7.11.e Review/approve/release exact bundle through governed publication; prohibit direct catalog/search/consumer write. Evidence: boundary tests.
    • Y7.11.f Preserve historical releases and propagate member supersession/ recall/canon/source changes via new bundle/incident. Evidence: lifecycle.
  • Y7.12 Build supersession, correction, unpublish, archive, recall, and canon hotfix paths. Evidence: customer read models and audit reconcile.
    • Y7.12.a Define distinct command/state/reason/authority/scope/effective- time/targets/notification/verification contracts for every lifecycle path. Evidence: schemas.
    • Y7.12.b Supersede with eligible replacement exact revision/bundle, effective behavior/migration/redirect/prior availability/consumer notice. Evidence: supersession tests.
    • Y7.12.c Correct factual/typographic/localization/metadata/content via new revisions, source/canon evidence, review/build and correction notice, never historical mutation. Evidence: correction tests.
    • Y7.12.d Unpublish target visibility/links/downloads/search/catalog/ consumer models/caches with receipts while preserving authoring/release evidence. Evidence: unpublish tests.
    • Y7.12.e Archive inactive authoring/catalog records with retention/ dependencies/authorized history/restore and no accidental delivery. Evidence: archive tests.
    • Y7.12.f Recall affected source/canon/asset/bundle/release through blast radius, quarantine/disable/removal/notification/verification/residual copies. Evidence: recall tests.
    • Y7.12.g Apply canon hotfix as scoped authorized assertion/content/build revision with solver/continuity/localization/asset impact and expedited but nonwaived gates. Evidence: hotfix tests.
    • Y7.12.h Rebuild/reconcile customer read/search/catalog/cache/links/state from authoritative events/bundles and compare audit. Evidence: lifecycle E2E.
  • Y7.13 Build deep links from released items to authorized source, provenance, and authoring history. Evidence: role-based disclosure tests.
    • Y7.13.a Define allowlisted link intents/routes for released study/case/ asset/bundle to source/citation/provenance/lineage/revision/decision/ authoring views with minimal claims. Evidence: schema.
    • Y7.13.b Issue signed opaque tenant/project/subject/version/audience/ intent/expiry/nonce links only after source authorization. Evidence: issuance tests.
    • Y7.13.c Revalidate signature/time/replay/host/route and current actor/ object/property/function/reveal/embargo/source rights at open. Evidence: validation tests.
    • Y7.13.d Project learner/customer/participant/author/reviewer/producer/ operator fields separately, redacting protected source/solution/spoiler/ private authoring/identity details. Evidence: disclosure tests.
    • Y7.13.e Handle logged-out/expired/revoked/deleted/superseded/recalled/ cross-tenant links, back/return context, referrer/log scrubbing. Evidence: lifecycle tests.
  • Y7.14 Build post-release feedback intake without silently modifying canon or source truth. Evidence: triage-to-revision journey.
    • Y7.14.a Define feedback/report identity, released item/build/revision/ locale/target, actor/anonymous/consent, category/severity, text/media/ anchor, visibility, status, and provenance. Evidence: schema.
    • Y7.14.b Ingest consumer/support/playtest/reviewer/partner/analytics/ correction/safety/rights/accessibility feedback with rate/abuse/PII/ spoiler controls. Evidence: intake tests.
    • Y7.14.c Triage/deduplicate/cluster/link sources/canon/issues with attributed rules/models, uncertainty, counterevidence and human priority/owner. Evidence: triage tests.
    • Y7.14.d Route to source update, canon proposal, content/locale/asset/ build revision, bug, incident, support, reject/defer decision; no direct truth/canon edit. Evidence: authority tests.
    • Y7.14.e Require normal source/canon/review/gate/build/release lifecycle and link outcome/decision/notification back. Evidence: E2E.
  • Y7.15 Build program/season/collection steering overlays for cadence, difficulty, canon, source, safety, production, and release constraints. Evidence: tighten-only property tests and impact preview.
    • Y7.15.a Define versioned overlay scope/priority/effective dates/owner and typed cadence/difficulty/canon/source/safety/production/budget/audience/ locale/release fields. Evidence: schema.
    • Y7.15.b Resolve platform/region/organization/program/season/collection/ project/profile precedence with conflict/cycle detection. Evidence: resolution tests.
    • Y7.15.c Merge by narrower eligibility/allowed canon/source, stricter gates/thresholds/reviewer/limits, lower budget/capacity, or additional requirements only. Evidence: monotonicity tests.
    • Y7.15.d Reject omission/default/inheritance/alias attempts to loosen safety/rights/privacy/accessibility/canon/source/authority/code floors. Evidence: property tests.
    • Y7.15.e Preview affected content/cases/clues/difficulty/schedules/assets/ pipelines/budgets/locales/builds/releases/consumers/conflicts. Evidence: impact tests.
    • Y7.15.f Review/approve/version/schedule/deactivate/rollback and revalidate dependent state without history rewrite. Evidence: lifecycle tests.
  • Y7.16 Build UI-governed study, case, genre, clue, audience, difficulty, production, and catalog taxonomies over code-owned floors. Evidence: versioned review, alias, deprecation, localization, and migration journey.
    • Y7.16.a Define separate typed registry entry/version/relationship/ lifecycle contracts and stable codes for every named taxonomy. Evidence: schemas.
    • Y7.16.b Declare code-owned noneditable identities/semantics/required fields/forbidden relations/safety-fairness-rights-accessibility/profile floors. Evidence: floor tests.
    • Y7.16.c Build restrained operator list/detail/create-draft/edit/diff/ validation/dependency/usage UI with typed fields and no arbitrary JSON. Evidence: authoring tests.
    • Y7.16.d Build alias/synonym/parent/related/mapping with scope/uniqueness/ collision/cycle/semantic equivalence and no silent merge. Evidence: relation tests.
    • Y7.16.e Build locale label/description/help/search terms/fallback/ completeness/translator context without localized identity codes. Evidence: localization tests.
    • Y7.16.f Review/approve/publish/deprecate/supersede/archive exact versions with effective dates, owner, audit, compatibility checks. Evidence: lifecycle tests.
    • Y7.16.g Preview/execute resumable dependency migrations across projects/ content/cases/assets/templates/overlays/builds/catalogs with conflicts/ rollback/history. Evidence: migration tests.
    • Y7.16.h Verify unknown/deprecated/missing entries fail/degrade safely in every consumer/export and run accessibility/performance/drift tests. Evidence: registry journey.

Y8 — Yemaya UX, accessibility, offline, and performance#

  • Y8.1 Add study and case entries to the existing Yemaya project navigation. Evidence: no second global shell.
    • Y8.1.a Register study/case routes, labels/icons/keywords/capabilities/ project profiles/roles in canonical project navigation/command palette. Evidence: IA registry tests.
    • Y8.1.b Show entries only for compatible project type/track/entitlement/ role/capability and safe unavailable/degraded states. Evidence: visibility tests.
    • Y8.1.c Preserve current project/revision/entity/return context and dirty state warnings across entry/exit/deep links. Evidence: navigation tests.
    • Y8.1.d Reuse existing global/project shell, breadcrumbs, switcher, identity/notifications/search/help and remove duplicate nav headers. Evidence: shell conformance.
    • Y8.1.e Test direct authorized/unauthorized routes, project switch/profile migration, web/desktop, keyboard/screen reader. Evidence: route E2E.
  • Y8.2 Build project dashboards around actionable work, not a card mosaic. Evidence: frontend conformance review.
    • Y8.2.a Define primary user triage questions and live view models for assignments/blockers/gate failures/jobs/reviews/schedule/budget/incidents/ releases per profile. Evidence: dashboard contract.
    • Y8.2.b Compose a restrained hierarchy with actionable queue/production state first, recent project context and secondary metrics second, no hero or repeated container cards. Evidence: design review.
    • Y8.2.c Implement stable filters/sorts/grouping/URLs/freshness/live updates and clear owner/deadline/severity/status/units. Evidence: UI tests.
    • Y8.2.d Implement eligible inline claim/assign/open/remediate/retry/hold/ review actions with permission/state/version/reason/confirmation. Evidence: action tests.
    • Y8.2.e Render loading/empty/stale/degraded/offline/fixture/error/partial/ unauthorized honestly and verify responsive/accessibility/scale. Evidence: frontend conformance E2E.
  • Y8.3 Build workspaces with primary editor, navigation, inspector, and visible gate/job status. Evidence: responsive browser snapshots.
    • Y8.3.a Define shared workspace shell contract for project/context nav, primary canvas/editor, contextual inspector, status rail, command palette, focus/selection/dirty/capability state. Evidence: shell spec.
    • Y8.3.b Compose study source/material/order and case dossier/entity/graph/ timeline/scene/solution editors using the shared shell. Evidence: workspace tests.
    • Y8.3.c Compose AAA script/board/design/schedule/budget/asset/dailies/ pipeline/marketplace/admin workspaces without nested global shells. Evidence: production tests.
    • Y8.3.d Keep applicable gates/freshness/blockers/remediation and jobs/ progress/cost/cancel/failure visible without obscuring the primary work. Evidence: status tests.
    • Y8.3.e Define wide/compact/single-column layouts with inspector/status drawer/tab, persistent subject identity, safe focus and no unusable squashing. Evidence: snapshots.
  • Y8.4 Preserve route, project, selected entity, tab, graph/timeline viewport, and inspector state in deep links. Evidence: refresh/share tests.
    • Y8.4.a Define canonical allowlisted URL state schema/version for route/ project/track/entity/revision/tab/selection/filter/graph-camera/timeline- range/inspector with size/privacy limits. Evidence: link schema.
    • Y8.4.b Encode stable opaque IDs and bounded viewport/range values while excluding secrets/prompts/protected source/solution/local paths/unsaved content. Evidence: serialization tests.
    • Y8.4.c Parse/validate/canonicalize and reauthorize every referenced object/property/function; handle stale/deleted/moved/revoked state. Evidence: restoration tests.
    • Y8.4.d Restore after refresh/back/forward/share/new window/web-desktop handoff and reconcile local unsaved state/conflicts safely. Evidence: navigation tests.
    • Y8.4.e Test cross-tenant/project substitution, oversized/unknown/bad viewport, role/reveal change, offline cache, link expiry. Evidence: share tests.
  • Y8.5 Provide semantic alternatives for graphs, timelines, rendered previews, and playback. Evidence: keyboard/screen-reader journeys.
    • Y8.5.a Define operation/content parity requirements and synchronized selection/focus/status between each visual surface and semantic alternative. Evidence: parity matrix.
    • Y8.5.b Provide graph tree/table/list/forms for nodes/edges/types/ properties/evidence/errors/path/relationships and all edit/navigation operations. Evidence: graph tests.
    • Y8.5.c Provide timeline ordered/grouped event table/forms with absolute/ relative/uncertain time, duration, participants/location/evidence/ conflicts and editing. Evidence: timeline tests.
    • Y8.5.d Provide rendered image/video/3D/audio source/scene/object/frame/ time/track/measurement/annotation list/table and controls. Evidence: preview tests.
    • Y8.5.e Provide playback native controls, transcript/captions/audio descriptions/chapter/shot/track/marker lists and synchronized keyboard navigation. Evidence: playback tests.
    • Y8.5.f Verify equivalent inspect/comment/edit/compare/decide operations, screen-reader announcements, zoom/reflow, large data performance. Evidence: semantic E2E.
  • Y8.6 Provide non-drag alternatives for hierarchy, graph, clue, scene, and timeline editing. Evidence: WCAG 2.2 dragging coverage.
    • Y8.6.a Inventory every drag/drop/pan/reorder/connect/move/resize/trim/ range operation across named editors and map equivalent outcomes. Evidence: drag inventory.
    • Y8.6.b Add move before/after/into/out/first/last/position commands/forms for hierarchy/scene/clue order with destination preview. Evidence: reorder tests.
    • Y8.6.c Add select source/target/relation/ports/condition forms and keyboard commands for graph/clue connections and node movement. Evidence: graph tests.
    • Y8.6.d Add date/time/duration/order/lane/range numeric/text forms and keyboard step commands for timeline events/clips/markers. Evidence: timeline tests.
    • Y8.6.e Keep validation/impact/permissions/concurrency/undo/audit identical between pointer and non-drag paths. Evidence: parity tests.
    • Y8.6.f Run WCAG keyboard/screen-reader/touch/switch-friendly journeys for every operation with focus restoration/announcement. Evidence: coverage.
  • Y8.7 Implement autosave/offline drafts with visible sync, conflict, and recovery. Evidence: disconnect/reconnect and two-client conflict tests.
    • Y8.7.a Define draft/local operation/server revision/sync checkpoint/ device/session/schema/encryption/retention contracts and eligible content types. Evidence: schema.
    • Y8.7.b Implement debounced idempotent autosave with expected revision, dirty/saving/saved/error/conflict/offline indicators and unload/crash recovery. Evidence: autosave tests.
    • Y8.7.c Store offline drafts/CRDT operations/assets securely with size/ quota/expiry/device/logout/project-access controls and no unencrypted spoilers/secrets. Evidence: cache tests.
    • Y8.7.d Replay/rebase/deduplicate on reconnect after current auth/project/ lock/schema/source/canon validation; never silently last-write-wins. Evidence: reconnect tests.
    • Y8.7.e Present mine/theirs/base semantic conflict and safe merge/manual resolution, preserve copies/export, and create new revisions. Evidence: conflict tests.
    • Y8.7.f Test two clients/web-desktop/offline edits/reorder/delete/role revoke/project archive/quota/crash/reconnect. Evidence: E2E.
  • Y8.8 Define web and desktop performance budgets for large projects, graphs, timelines, assets, video, and search. Evidence: profiling reports.
    • Y8.8.a Define shared client/server/end-to-end load/interaction/save/ search/preview/playback latency, CPU/memory/GPU/disk/network/payload/ query/error budgets per device/network/profile. Evidence: budget spec.
    • Y8.8.b Define production-scale project/entity/revision/collaborator/ source/asset/job/comment/locale/search distributions and representative fixtures. Evidence: fixture manifest.
    • Y8.8.c Define graph/timeline nodes/edges/events/paths/annotations/ updates/layout/validation budgets and measurement points. Evidence: editor budgets.
    • Y8.8.d Define asset/gallery/proxy/large image/video/audio/3D load/decode/ scrub/render/cache/disk/memory/GPU/network budgets by web/desktop. Evidence: media budgets.
    • Y8.8.e Define search indexing lag/query/facet/suggestion/pagination/ cache/permission overhead and large-project save/sync/reconnect/export budgets. Evidence: data budgets.
    • Y8.8.f Profile supported scenarios, capture client/server traces/query plans/resource graphs, trend baselines, and fail regressions. Evidence: reports.
  • Y8.9 Virtualize large lists and graphs without breaking accessibility or selection. Evidence: high-volume tests.
    • Y8.9.a Inventory virtualized candidates and define stable item/node/edge identity/order/size/focus/selection/scroll/announcement/search semantics. Evidence: virtualization spec.
    • Y8.9.b Implement lists/tables/trees/grids with overscan/dynamic size/ grouped rows/sticky headers/infinite cursor loading and no duplicate/ missing entries. Evidence: list tests.
    • Y8.9.c Implement graph semantic view and visual culling/LOD/tile/layout/ cluster/incremental updates without dropping reachable selected entities or validation. Evidence: graph tests.
    • Y8.9.d Preserve focus/active descendant/selected/expanded/anchor/keyboard index and screen-reader position/count when items unmount/remount. Evidence: accessibility tests.
    • Y8.9.e Preserve URL selection/scroll/viewport across filter/sort/update/ pagination/refresh/back/reconnect and handle deleted targets. Evidence: state tests.
    • Y8.9.f Load-test high-volume pointer/keyboard/screen-reader/search/live updates and memory over sustained use. Evidence: performance tests.
  • Y8.10 Add reduced-motion behavior to route, panel, graph, and timeline transitions. Evidence: automated assertions.
    • Y8.10.a Inventory route/page/panel/dialog/menu/toast/loading/progress/ graph layout/pan-zoom/highlight/timeline scroll/playhead/drag/preview animations. Evidence: motion inventory.
    • Y8.10.b Define semantic state change/feedback purpose and reduced-motion replacement as instant, short fade, static highlight, manual control, or removal. Evidence: behavior spec.
    • Y8.10.c Honor OS/browser/app preference before first paint across web/ desktop and update live without layout/focus loss. Evidence: preference tests.
    • Y8.10.d Disable parallax/spring/large translation/auto pan/continuous layout/animated zoom and provide nonmotion graph/timeline state changes. Evidence: component tests.
    • Y8.10.e Preserve essential progress/playback/time/capture indicators with accessible text and user-controlled motion. Evidence: parity tests.
    • Y8.10.f Assert computed styles/animation frames/scroll behavior across routes/panels/graphs/timelines/themes/platforms. Evidence: automation.
  • Y8.11 Test themes, density, zoom, RTL, long locale, and high-contrast modes. Evidence: visual/a11y matrix.
    • Y8.11.a Define supported light/dark/system/high-contrast themes, compact/ comfortable/touch density, 100–400% zoom/reflow, RTL/bidi/long/CJK locales, and OS contrast modes. Evidence: matrix.
    • Y8.11.b Test color tokens/contrast/focus/selection/status/charts/graphs/ media annotations/canvas-semantic alternatives in every theme. Evidence: theme tests.
    • Y8.11.c Test density/touch targets/spacing/editor controls/tables/ timeline/graph labels and user preference persistence. Evidence: density tests.
    • Y8.11.d Test browser/app/text zoom/reflow without clipping/overlap/ hidden controls/two-dimensional scroll except essential surfaces. Evidence: zoom tests.
    • Y8.11.e Test RTL navigation/order/icons/bidi isolation/mixed script/ timeline/graph/tree/forms and long/expanding labels/dialogue/captions. Evidence: locale tests.
    • Y8.11.f Run visual regression plus automated/manual accessibility and keyboard/screen-reader tests across web/desktop/viewports. Evidence: report matrix.
  • Y8.12 Surface fixture, offline, stale, degraded, unavailable, and failed state consistently across web and desktop. Evidence: state matrix.
    • Y8.12.a Define canonical data-origin/service-state discriminators, safe messages/timestamps/reasons/actions/recovery/announcements and telemetry. Evidence: state contracts.
    • Y8.12.b Label fixture/demo/synthetic content persistently and prevent production mutation/promotion/build/release. Evidence: fixture tests.
    • Y8.12.c Show offline cached versus unavailable, sync/queued edits/ prohibited actions/freshness/reconnect and no simulated network success. Evidence: offline tests.
    • Y8.12.d Show stale subject/dependency/evidence/capability with exact invalidating version and block unsafe actions. Evidence: stale tests.
    • Y8.12.e Show per-capability degraded/partial/unavailable/not-configured/ failed/unknown with safe alternatives and no global healthy implication. Evidence: service tests.
    • Y8.12.f Exercise each state/recovery across every critical web/desktop route with identical semantics and accessible announcements. Evidence: state matrix.
  • Y8.13 Reconcile navigation and command-palette reachability for every retained AAA, production, marketplace, organization, study, and case capability. Evidence: route/role/project-type matrix has no reachable unauthorized item and no authorized orphan capability.
    • Y8.13.a Generate retained atomic capability/owner/host/route/project-type/ role/entitlement/capability/command registry from Y0.16/Y5A and source routes. Evidence: reachability manifest.
    • Y8.13.b Register each authorized page/action/deep link/command exactly once in existing shell/project nav/palette with contextual prerequisites. Evidence: IA tests.
    • Y8.13.c Compute expected-versus-observed visibility/reachability for every role/project type/web-desktop/capability state and classify unauthorized exposure/orphan/duplicate/stale/misroute. Evidence: matrix.
    • Y8.13.d Test navigation hides unauthorized items without hint leaks and server rejects direct route/command/field/function attempts. Evidence: auth tests.
    • Y8.13.e Test authorized items resolve to owned live/degraded route with correct project context/back path and no second shell/duplicate mutation. Evidence: positive tests.
    • Y8.13.f Fail CI on unreviewed retained capability, authorized orphan, reachable unauthorized item, unknown registry row, or expired disposition. Evidence: route/role/profile gate.
  • Y8.14 Provide responsive review/incident projections for scripts, storyboards, assets, dailies, budgets, schedules, pipeline runs, marketplaces, study, and cases while reserving precision editing for capable hosts. Evidence: mobile/tablet critical-action journeys.
    • Y8.14.a Define mobile/tablet review versus precision-edit capability matrix, subject identity/context, eligible actions, step-up, and desktop/ host handoff per named surface. Evidence: responsive spec.
    • Y8.14.b Build script/storyboard semantic read/diff/comments/review, asset/dailies preview/annotations/gates/decisions, and study/case content/ canon/solver review projections. Evidence: review tests.
    • Y8.14.c Build budget/schedule current-versus-baseline/variance/conflicts/ approvals and pipeline run/plan/progress/cost/failure/pause/cancel/ incident projections. Evidence: operations tests.
    • Y8.14.d Build marketplace item/security/license/compatibility/report/ takedown and incident scope/evidence/actions/status views with protected data controls. Evidence: marketplace/incident tests.
    • Y8.14.e Implement comment/approve/reject/request changes/hold/escalate/ pause/cancel/kill/ack only for current role/state with impact/reason/ confirmation/concurrency/audit. Evidence: action tests.
    • Y8.14.f Provide semantic graph/timeline/3D/curve snapshots/tables instead of full precision editors and signed open-on-desktop/host return links. Evidence: handoff tests.
    • Y8.14.g Test touch/keyboard/screen reader/zoom/orientation/safe area/slow- offline/reconnect/direct authorization/role revoke across representative critical journeys. Evidence: E2E.

Y9 — Yemaya verification, operations, and launch#

  • Y9.1 Run contract and migration compatibility tests for case, study, project, canon, and release schemas. Evidence: old-version corpus.
    • Y9.1.a Assemble supported historical case/study/project/canon/release records, events, blobs, indexes, manifests, and client payloads with provenance and expected outcomes. Evidence: version corpus.
    • Y9.1.b Test forward migrations one version at a time and from every supported direct upgrade point without silent defaulting or data loss. Evidence: migration matrix.
    • Y9.1.c Test current writers with supported old readers and old writers with current readers at explicit compatibility boundaries. Evidence: mixed-version tests.
    • Y9.1.d Verify identifiers, ordering, uncertainty, evidence, permissions, reveal state, checksums, and release signatures survive round trips. Evidence: semantic diffs.
    • Y9.1.e Test interrupted/retried migrations, rollback or restore, malformed/unknown fields, and unsupported versions fail safely. Evidence: failure tests.
    • Y9.1.f Record compatibility window, irreversible steps, operator prerequisites, and corpus ownership in the release gate. Evidence: compatibility report.
  • Y9.2 Run known-value source, hierarchy, reading-order, canon, timeline, solvability, uniqueness, fairness, reachability, and spoiler tests. Evidence: gate registry coverage.
    • Y9.2.a Curate positive, boundary, counterexample, adversarial, and intentionally indeterminate fixtures with expert-reviewed expected measurements and verdicts. Evidence: known-value corpus.
    • Y9.2.b Verify source integrity, hierarchy/reference resolution, and deterministic reading order across format, locale, revision, and partial-ingest cases. Evidence: source gate tests.
    • Y9.2.c Verify canon contradiction, timeline consistency/uncertainty, and stale-dependency invalidation against expected finding sets. Evidence: canon/timeline tests.
    • Y9.2.d Verify solver completeness, uniqueness/ambiguity, reachability, hint progression, dead ends, and assumption budgets. Evidence: solver tests.
    • Y9.2.e Verify fairness/accessibility and spoiler/reveal boundaries by role, stage, locale, export, search, cache, and notification. Evidence: boundary tests.
    • Y9.2.f Map every retained gate to fixtures, measurement assertions, verdict assertions, remediation, owner, and current execution result. Evidence: gate registry coverage report.
  • Y9.3 Run calibrated safety/AI evaluations with threshold, subgroup, and uncertainty reports. Evidence: review approval.
    • Y9.3.a Freeze representative allowed/refuse/escalate/abstain fixtures, subgroup labels, protected-content classes, and expert ground truth. Evidence: evaluation manifest.
    • Y9.3.b Evaluate extraction/generation/solver/agent quality separately from safety, grounding, leakage, and authorization behavior. Evidence: metric suite.
    • Y9.3.c Report threshold curves, confidence intervals, calibration, abstention, false-positive/negative severity, and subgroup deltas. Evidence: evaluation report.
    • Y9.3.d Test prompt/source injection, spoiler elicitation, sensitive inference, unsafe-media requests, tool misuse, and multi-turn attacks. Evidence: adversarial suite.
    • Y9.3.e Reproduce regressions against prior approved baselines and route misses to mitigation, model/prompt rollback, or human-only disposition. Evidence: regression decisions.
    • Y9.3.f Obtain named domain/safety/editorial approval for thresholds, residual risk, exceptions, and expiry. Evidence: signed review.
  • Y9.4 Run real-database, blob, search, queue, compiler, and read-model integration suites. Evidence: durable round-trip.
    • Y9.4.a Provision supported production-equivalent database/blob/search/ queue/compiler/read-model versions with isolated tenant fixtures. Evidence: environment manifest.
    • Y9.4.b Exercise transactional writes, outbox/events, blob completion, indexing, queued compilation, and consumer projection end to end. Evidence: trace-linked tests.
    • Y9.4.c Verify retries/idempotency/order/deduplication, concurrent edits, partial outages, poison messages, dead letters, and replay. Evidence: fault tests.
    • Y9.4.d Verify tenant/role/field/reveal isolation in primary records, blobs, indexes, queue payloads, compiler artifacts, and read models. Evidence: isolation tests.
    • Y9.4.e Compare authoritative and projected states after create/update/ delete/archive/restore/release and repair induced divergence. Evidence: reconciliation report.
    • Y9.4.f Preserve durable IDs/checksums/provenance/audit across backup, restart, retry, and cold consumer rebuild. Evidence: round-trip proof.
  • Y9.5 Run web full journey: source to published study material. Evidence: live backing and verified consumer read model.
    • Y9.5.a Create an authorized study project and ingest representative files/URLs/media with real blob storage, parsing, provenance, and quarantine outcomes. Evidence: journey trace.
    • Y9.5.b Review source structure/reading order/metadata/rights, correct issues, and approve a pinned source revision. Evidence: review audit.
    • Y9.5.c Author grounded material, ordering, questions/hints/accessibility/ locale variants with citations and protected answer boundaries. Evidence: authoring trace.
    • Y9.5.d Run applicable source/canon/fairness/accessibility/spoiler gates, remediate a failure, and prove verdict freshness. Evidence: gate trace.
    • Y9.5.e Preview as eligible learner/reviewer roles and prove protected fields do not leak through route, API, search, cache, or export. Evidence: role tests.
    • Y9.5.f Publish through real job/release services, then consume the signed immutable read model and reconcile it to the approved revision. Evidence: published-consumer proof.
  • Y9.6 Run web full journey: premise to solved, reviewed, localized, published case. Evidence: real compiler/verifier.
    • Y9.6.a Create a case project from premise and build dossier/entities/ clues/relationships/timeline/scenes with source-backed uncertainty. Evidence: authoring trace.
    • Y9.6.b Define reveal order, red herrings, hints, solution/assumptions, alternate hypotheses, and role-specific spoiler boundaries. Evidence: case revision.
    • Y9.6.c Run real graph/timeline/canon/solvability/uniqueness/fairness/ reachability gates, repair failures, and rerun against pinned inputs. Evidence: verifier trace.
    • Y9.6.d Complete independent editorial/domain/safety review with comments, changes, approvals, audit, and stale-approval invalidation. Evidence: review trace.
    • Y9.6.e Localize content, clues, layouts, captions, reading order, and solution reveal while preserving IDs, evidence, and solver parity. Evidence: locale comparison.
    • Y9.6.f Compile/publish through real services and solve the immutable consumer artifact as intended and plausible alternate paths. Evidence: release/consumer proof.
  • Y9.7 Run desktop full journey with local import, offline edit, sync conflict, render job, and publish handoff. Evidence: signed build.
    • Y9.7.a Install/launch the signed supported desktop build in a clean standard-user profile and authenticate/open an entitled real project. Evidence: build/install trace.
    • Y9.7.b Import representative local files through approved pickers, validate/quarantine/deduplicate/upload them, and preserve local-source provenance without exposing paths. Evidence: import trace.
    • Y9.7.c Disconnect, edit eligible content/assets, close/relaunch, and recover encrypted offline drafts with honest availability state. Evidence: offline trace.
    • Y9.7.d Create a concurrent web edit, reconnect, inspect mine/theirs/base, resolve the semantic conflict, and verify revision/audit continuity. Evidence: conflict trace.
    • Y9.7.e Submit/monitor/cancel/retry and complete a real render/compile job with progress/cost/logs/artifacts and resource limits. Evidence: job trace.
    • Y9.7.f Hand off the approved revision to web publish, consume the released artifact, and verify signed-build/deep-link/updater boundaries. Evidence: desktop journey report.
  • Y9.8 Run agent fault, budget, cancellation, and human-override journeys. Evidence: no unauthorized direct mutation.
    • Y9.8.a Exercise approved agent plans across representative read, recommend, draft, job-submit, and proposed-mutation capabilities. Evidence: journey manifest.
    • Y9.8.b Inject model/tool/provider timeout, malformed output, rate limit, stale subject, dependency outage, and partial-result failures. Evidence: fault traces.
    • Y9.8.c Exhaust token/cost/time/tool/step/concurrency budgets and verify deterministic stop, retained partial artifacts, and no hidden overrun. Evidence: budget tests.
    • Y9.8.d Cancel before execution, during tool work, during durable job, and after commit boundary; verify compensation and truthful terminal state. Evidence: cancellation tests.
    • Y9.8.e Pause/override/edit/reject/escalate from the human control surface and verify revalidation, approval, audit, and ownership transitions. Evidence: override tests.
    • Y9.8.f Attempt prompt/tool/direct API mutation outside the plan/role/ capability/project and prove server rejection with no side effect. Evidence: authorization report.
  • Y9.9 Run WCAG 2.2 AA, keyboard, announcement, and semantic-equivalence suites on web and desktop renderer. Evidence: matrix report.
    • Y9.9.a Define route/workspace/dialog/editor/action coverage across wide/compact/single-column, web/desktop, themes, zoom, RTL, and locales. Evidence: accessibility matrix.
    • Y9.9.b Run automated WCAG rules, contrast, name/role/value, landmark/ heading, form/error, target-size, reflow, and motion assertions. Evidence: automation report.
    • Y9.9.c Complete keyboard-only journeys with logical order, visible focus, no traps, skip/recovery paths, and non-drag alternatives. Evidence: keyboard traces.
    • Y9.9.d Verify live announcements for save/sync/job/gate/filter/error/ conflict/review state without duplication, interruption, or silent change. Evidence: announcement tests.
    • Y9.9.e Compare graph/timeline/render/playback visual and semantic paths for content, selection, editing, decisions, validation, and permissions. Evidence: equivalence tests.
    • Y9.9.f Run named assistive-technology/manual audits on supported web and desktop combinations and close or explicitly block every finding. Evidence: signed matrix report.
  • Y9.10 Run large-project performance, memory, startup, search, graph, timeline, render-monitor, and sync tests. Evidence: budgets pass.
    • Y9.10.a Generate versioned production-scale fixtures matching documented project/entity/asset/revision/collaboration/job distributions. Evidence: fixture report.
    • Y9.10.b Measure cold/warm web and desktop startup, project open, route/ workspace transition, idle, and sustained-session memory. Evidence: startup/memory profiles.
    • Y9.10.c Measure indexed search/facets/pagination/update visibility and large graph/timeline load/layout/query/edit/validation. Evidence: interaction profiles.
    • Y9.10.d Measure proxy/media load, scrub/playback, render-monitor live updates/logs/artifacts, and bounded client/server resource use. Evidence: media/job profiles.
    • Y9.10.e Measure large save/autosave/offline cache/reconnect/rebase/ conflict/export with concurrency and constrained networks. Evidence: sync profiles.
    • Y9.10.f Compare every metric to platform/device/network budgets, retain traces, fail regressions, and assign any exception with expiry. Evidence: signed budget report.
  • Y9.11 Run desktop IPC, filesystem, deep-link, updater, signing, sandbox, and dependency security tests. Evidence: zero blockers.
    • Y9.11.a Enumerate every renderer-to-main IPC channel with input/output schema, caller capability, authorization, timeout, size, and audit requirements. Evidence: IPC security manifest.
    • Y9.11.b Fuzz unknown/malformed/oversized/replayed/concurrent IPC and prove privileged operations cannot be invoked from untrusted content. Evidence: IPC adversarial tests.
    • Y9.11.c Test picker-scoped filesystem access, traversal/symlink/race/ archive/device-file attacks, permissions, quotas, cleanup, and path redaction. Evidence: filesystem report.
    • Y9.11.d Test deep-link allowlists, canonical parsing, reauthorization, expired/replayed links, malicious schemes/arguments, and safe focus. Evidence: link security tests.
    • Y9.11.e Test signed update metadata/download/install/rollback/recovery against downgrade, tamper, interruption, channel/version skew, and revoked keys. Evidence: updater report.
    • Y9.11.f Verify code signing/notarization or platform equivalent, sandbox/CSP/permission boundaries, SBOM/licenses/vulnerabilities, and release artifact provenance with zero blockers. Evidence: security gate.
  • Y9.12 Run source injection, spoiler leak, canon tamper, unsafe media, archive, SSRF, auth, and tenant adversarial tests. Evidence: zero blockers.
    • Y9.12.a Inject active instructions/markup/scripts/formulas/links and Unicode/bidi/confusable payloads through every supported source format and derived field. Evidence: source-injection tests.
    • Y9.12.b Probe spoiler/answer/solution leakage through API fields, errors, search, suggestions, notifications, logs, cache, exports, model context, and timing. Evidence: leak report.
    • Y9.12.c Tamper with canon/evidence/revision/gate/result/read-model/release IDs, hashes, ordering, signatures, and stale approvals. Evidence: integrity tests.
    • Y9.12.d Upload polyglot/malformed/oversized/decompression-bomb/active/ prohibited media and verify quarantine, transform isolation, and safe preview. Evidence: media tests.
    • Y9.12.e Exercise archive traversal/symlink/nesting/ratio and URL import SSRF/DNS rebinding/redirect/private-network/credential attacks. Evidence: import security tests.
    • Y9.12.f Test cross-role/project/tenant object, field, function, blob, search, queue, export, deep-link, and cache access with zero side effects. Evidence: authorization gate.
  • Y9.13 Run backup/restore and web/desktop version-skew drills. Evidence: RPO/RTO and compatibility results.
    • Y9.13.a Define protected data/services, backup cadence/retention/ encryption/immutability/region, recovery order, RPO/RTO, and owners. Evidence: recovery contract.
    • Y9.13.b Restore database/blob/search/queue/config/secrets references and release artifacts into an isolated environment from selected full/ incremental/point-in-time backups. Evidence: restore trace.
    • Y9.13.c Reconcile records, blobs, indexes, projections, jobs, audits, signatures, tenant boundaries, and unresolved edits after restore. Evidence: integrity report.
    • Y9.13.d Exercise supported older/newer web API, desktop client, schema, job worker, compiler, and read-model combinations. Evidence: version-skew matrix.
    • Y9.13.e Verify incompatible combinations block safely with upgrade/ retry/export/recovery guidance and no corrupt or silently dropped edits. Evidence: boundary tests.
    • Y9.13.f Measure achieved data loss and recovery duration, record gaps/ exceptions, and obtain operations/security approval. Evidence: signed drill report.
  • Y9.14 Publish author, canon, solver, localization, producer, desktop, incident, and recovery runbooks. Evidence: game-day.
    • Y9.14.a Publish author/source-ingest and canon/timeline remediation runbooks with prerequisites, diagnostics, safe edits, revalidation, and escalation. Evidence: runbook review.
    • Y9.14.b Publish solver/uniqueness/fairness/spoiler investigation runbooks with known-value comparisons and human disposition. Evidence: runbook review.
    • Y9.14.c Publish localization/release and producer schedule/budget/ pipeline/job/engine handoff runbooks with rollback boundaries. Evidence: runbook review.
    • Y9.14.d Publish desktop import/offline/conflict/filesystem/deep-link/ updater/signing support runbooks without requesting secrets or unsafe filesystem access. Evidence: support review.
    • Y9.14.e Publish security/safety/privacy/availability incident triage, containment, evidence, communication, recovery, and postmortem runbooks. Evidence: incident review.
    • Y9.14.f Run timed cross-role game-days from detection through recovery, capture unclear/missing/unsafe steps, revise, and reapprove. Evidence: game-day report.
  • Y9.15 Produce engineering-readiness evidence per track and host. Evidence: all required reports current.
    • Y9.15.a Define required architecture/contracts/migrations/security/ privacy/safety/accessibility/quality/operations evidence by study, case, production, marketplace, organization, and developer track. Evidence: readiness matrix.
    • Y9.15.b Define additional web, desktop, mobile/tablet review, engine/DCC, virtual-production/XR, and external-host evidence with owners/freshness. Evidence: host matrix.
    • Y9.15.c Link each retained capability and release cell to current test/ evaluation/report/runbook/approval artifacts and immutable versions. Evidence: evidence index.
    • Y9.15.d Verify evidence was produced against the candidate commit, configuration, models/providers, dependencies, schemas, and build. Evidence: provenance check.
    • Y9.15.e Classify missing/stale/failed/waived evidence, block required cells, and require named owner/reason/scope/expiry for exceptions. Evidence: readiness gate.
    • Y9.15.f Publish a per-track/host go/no-go summary with residual risks and explicit unsupported combinations. Evidence: readiness packet.
  • Y9.16 Obtain named editorial, domain, safety, accessibility, privacy/security, engineering, and product approvals. Evidence: human signoff.
    • Y9.16.a Define each approver’s scope, required evidence, decision criteria, independence/conflict rules, delegates, and expiry. Evidence: approval policy.
    • Y9.16.b Route immutable candidate/evidence packets to named editorial and relevant study/case/production/domain reviewers. Evidence: review records.
    • Y9.16.c Route calibrated safety, accessibility, privacy, and security reports with open findings/exceptions to accountable reviewers. Evidence: specialist decisions.
    • Y9.16.d Route engineering/operations and product readiness, support, compatibility, rollout, and residual-risk evidence. Evidence: launch decisions.
    • Y9.16.e Record approve/reject/request-changes/conditional decisions with identity, role, scope, candidate, reason, conditions, and timestamp. Evidence: signed audit.
    • Y9.16.f Invalidate affected approvals after candidate/evidence/config/ model/dependency/schema changes and block launch until renewed. Evidence: stale-approval tests.
  • Y9.17 Run full-journey coverage for project/asset lifecycle, script→storyboard→animatic, character/world→scene, schedule/budget/call sheet, animation/motion, dailies/editorial, collaboration, autonomous film/game pipeline, engine round-trip, gameplay/NPC, virtual production/XR, marketplace, organization administration, community disposition, and developer integration. Evidence: one generic project test cannot stand in for these distinct workflows.
    • Y9.17.a Run project/asset create-import-version-relate-search-review/ archive/restore/release across authorized roles and web/desktop. Evidence: lifecycle journey.
    • Y9.17.b Run script→storyboard→animatic and character/world→scene continuity, provenance, review, change-impact, and handoff journeys. Evidence: creative journeys.
    • Y9.17.c Run schedule/budget/call-sheet and animation/motion with baseline/ variance/conflict/approval and Aja/host round-trip. Evidence: production journeys.
    • Y9.17.d Run dailies/editorial and multi-user collaboration through ingest/proxy/sync/notes/decisions/CRDT conflict/offline recovery. Evidence: review/collaboration journeys.
    • Y9.17.e Run autonomous film and game pipelines through approved plan, budget, jobs, human override, artifacts, gates, and publish handoff. Evidence: agent pipeline journeys.
    • Y9.17.f Run engine round-trip, gameplay/NPC behavior, and virtual production/XR capture/calibration/session/incident journeys. Evidence: runtime/VP journeys.
    • Y9.17.g Run marketplace submit-scan-review-install-license-update/ report/takedown and organization membership/policy/audit/retention/ billing administration. Evidence: marketplace/admin journeys.
    • Y9.17.h Run community submission/moderation/appeal/disposition and developer API/webhook/service-account/key rotation/rate-limit/version journeys. Evidence: community/developer journeys.
    • Y9.17.i Prove each journey uses real owned services, isolates tenants/ roles, emits trace/audit evidence, and cleans or retains fixtures by policy. Evidence: journey coverage ledger.
  • Y9.18 Run pairwise compatibility and loss-profile gates for every supported project type, asset kind, web/desktop version, engine/DCC host, external integration, marketplace package, and release target. Evidence: unsupported combinations are explicit and supported cells have current proof.
    • Y9.18.a Build canonical dimensions/values/constraints from supported project profiles, assets/formats, clients, hosts/versions, integrations, packages, locales, and release targets. Evidence: compatibility model.
    • Y9.18.b Generate deterministic pairwise cells plus required high-risk higher-order combinations and label invalid/unsupported cells with rationale. Evidence: coverage set.
    • Y9.18.c Define loss profile per transform/round-trip for identity, hierarchy, metadata, timing, geometry/materials/animation/audio, interactivity, accessibility, provenance, and rights. Evidence: loss contracts.
    • Y9.18.d Execute import/export/round-trip/install/build/release cells on exact versioned hosts and compare semantic measurements to tolerances. Evidence: compatibility results.
    • Y9.18.e Exercise version skew, optional capability absence, degradation, warnings, refusal, repair, and rollback without silent loss. Evidence: boundary tests.
    • Y9.18.f Publish current supported/conditional/unsupported matrix with evidence age, known loss, prerequisites, owner, and expiry; fail stale or unproven supported cells. Evidence: release gate.
  • Y9.19 EVE-SEAM-EXIT: require "required Eve reads and writes admitted with source and behavior evidence, or an exact task-0.8 explicit non-goal proves zero operations; a named defer preserves ownership but never passes completion" before this phase exits. Evidence: source-derived totality record and passing verifier.
    • Y9.19.a Re-run the route, bridge, state-signal, and complete domain-code envelope census; review every changed candidate instead of carrying a stale phase disposition forward. Evidence: current ratchets.
    • Y9.19.b Bind the exact ratified task-0.8 workflow and requirement set; classify it as a required seam or an exact source-backed non-goal. Evidence: charter-bound disposition set.
    • Y9.19.c For a required seam, admit both read and write operations with byte-pinned source and behavior evidence. A named defer retains its owner and unblock condition but fails completion. Evidence: operation proofs.
    • Y9.19.d Permit a zero-operation boundary only for an exact ratified task-0.8 explicit non-goal; absence or unavailability is not a non-goal. Evidence: source-bound non-goal proof.
    • Y9.19.e Run the totality verifier, every retained gate-case CLI, and the injected durable-candidate control; retain red and green receipts. Evidence: release gate logs.

Phase V — Veritas newsroom workbench#

Veritas is the V1 grounded-story domain. The workbench must preserve source protection, make claims and counterclaims inspectable, enforce the implemented editorial and attestation policies, publish through the real CMS, and carry corrections or retractions to every dependent surface.

<!-- SPEC 2026-09-11: PHASE V ESTATE FACTS (measured 2026-09-11 at 48f4bc8ddb; §2.5 binds). (1) NO NEWSROOM AUTHORING UI EXISTS. apps/oshun/web/src/app/veritas/* (story, claim, source, evidence, counterclaims, retraction, topic, provenance, mobile) is the Lilith CONSUMER room (VeritasRoom, reader-facing); /domains/veritas/* (articles, research, headlines, knowledge-graph) are Lilith consumer tools; both are V1.2-scoped by libs/oshun/navigation/src/release-scope.ts and the BFF middleware/release-scope.ts refuses Veritas scope in V1.0 (the existing veritas-full-journey.spec.ts asserts that refusal). apps/veritas/web is a three-page public reader ([locale], reading-list, offline, reads NEXT_PUBLIC_API_URL, no e2e). The committed route inventory labels the consumer room "newsroom authoring" — that label is wrong and V0.1.a/V0.1.c are re-opened below. (2) THREE DISJOINT PERSISTENCE STORIES, NONE CANONICAL: libs/veritas/database (Prisma, 67 models, 2 migrations, VERITAS_DATABASE_URL — not in env-master, no compose service, so a database must be created on the shared Postgres); apps/veritas/api (Hono :3002, raw SQL through a pg pool + Redis + Elasticsearch, HMAC JWT carrying userId/email and NO role, 19 tests that mock the pool; its SQL reads user_subscriptions, a table the Prisma migration does not create); apps/veritas/cms (Fastify :3004, 28 routes, src/storage/memory-storage.ts is its ONLY storage, no auth on routes, 2 tests, nothing consumes it). "Publish through the real CMS" in this phase's header therefore has no durable CMS behind it today; V6.1 decides, V6.9 builds on the decision, V9.8 proves it. (3) V1-SIDE CODE THAT IS REAL: libs/oshun/domain-veritas (36 specs, built on @oshun/contracts/veritas: SourceKind 15-vocabulary, bands, ClaimClass, RetractionState, EvidenceStance, attestor), libs/oshun/veritas-cascade-worker (live-infra integration tests) wired by apps/oshun/bff/src/veritas/retraction-cascade-runtime.ts, and apps/oshun/bff/src/routes/veritas-briefing.ts. Legacy Zod models live in libs/veritas/models/src/schemas (nested) against Prisma (flattened) — V0.3.a already measured the mismatch. (4) STANDARDS: NewsML-G2 and IPTC Media Topics have ZERO code in the estate; C2PA is libs/veritas/content-auth with an optional c2paValidator seam (1 test) and c2pa-node@0.5.26 installed under pnpm; Ghana NMC and election rules exist only in apps/veritas/api/src/domain/compliance/{nmc,election}.ts; USSD is libs/veritas/ussd. (5) MODELS: libs/veritas/llm defaults to a Claude model table over OPENROUTER_API_KEY; libs/veritas/{claims,fact-checking,bias-detection} import it. Bind VERITAS_LLM_PRIMARY_MODEL and VERITAS_LLM_FALLBACK_MODEL to deepseek/deepseek-v4-flash-0731 with OPENROUTER_PROVIDER_SORT=price for every test (§2.5.4); bias-detection has no tests at all. (6) 13 of 64 libs/veritas/* packages have zero tests: agents-product, archive, b2b-sdk, b2b-sdk-python, bias-detection, billing, content-classification, core, expansion, ingestion-core, knowledge-graph, models, signup. (7) Playwright: apps/oshun/web/e2e/*veritas* are consumer-room specs (only home-veritas-briefing intercepts routes); none exercises authoring. Every V-phase journey cell starts from nothing and must obey §2.5.2. -->

V0 — Newsroom scope, route, roles, and ethics#

  • V0.1 Reconcile Oshun Studio, apps/veritas/web, apps/veritas/cms, and existing BFF routes. Evidence: canonical authoring/reader/operator route map. <!-- partial: route inventory (a) done in evidence/v1-workbenches/veritas-route-inventory.json. b-e (journey diagrams, collision ledger, canonical owner map, boundary tests) are runtime/design/human and stay unchecked. --> <!-- SPEC 2026-09-11: the canonical host does not exist (Phase V header note). .d is a product decision with three real options — a new newsroom under Oshun Studio (the Section 1 table's stated host) backed by the Prisma store, a newsroom UI inside apps/veritas/*, or promoting the headless CMS — and an agent drafts docs/adr/ADR-V0.1-veritas-newsroom-host.md (Proposed) rather than choosing. Until a person accepts it, V2–V8 UI cells may build components and contracts but no route tree; V8.2 registers routes only after .d. -->
    • V0.1.a Inventory Studio, Veritas web/CMS, BFF, public reader, deep-link, redirect, navigation, and command registrations with owners/audiences. Evidence: route inventory. <!-- evidence: .pageRoutes + .cmsServiceRoutes + .bffRoutes — all 4 named surfaces: 11 Oshun-Studio /veritas/* authoring pages, 3 veritas-web public reader pages ([locale]-prefixed), 28 veritas-cms Fastify service routes (headless service, no pages), 2 BFF modules (16 registrations). owners/ audiences classification is V0.1.b/d (semantic). generate-veritas-route- inventory.mjs + 7 tests, wired into inventory/verify. First Phase V task. --> <!-- SPEC 2026-09-11: RE-OPENED. The inventory's host label "oshun-web (/veritas/* newsroom authoring)" is wrong — those pages render VeritasRoom for readers — and it omits /domains/veritas/{articles,research,headlines,knowledge-graph} and apps/oshun/web/src/app/api/veritas/*. The cell's own clause asks for owners/audiences, which the comment defers. Regenerate with audience = consumer/reader/service/bff per row and the missing rows added, then flip. -->
    • V0.1.b Trace each newsroom authoring/review/publish/operations journey through current UI, BFF, service, CMS, and public read-back. Evidence: current-state diagrams.
    • V0.1.c Identify duplicate shells/routes/mutations, split ownership, legacy aliases, dead paths, and protected/public boundary violations. Evidence: collision ledger. <!-- evidence: veritas-collision-ledger.json — cross-surface scan of the V0.1.a inventory: 0 duplicate page routes (studio /veritas/* authoring vs veritas-web [locale] reader don't overlap), 0 protected/public boundary violations (reader is locale-scoped, never reaches a protected /veritas/* path — boundary INTACT), 0 CMS↔BFF operation-path overlaps, 0 dead paths. split-ownership + legacy-alias disposition is V0.1.d (semantic). generate-veritas-collision-ledger.mjs + 5 tests. --> <!-- SPEC 2026-09-11: RE-OPENED. The "0 boundary violations" conclusion was computed over the mislabeled V0.1.a inventory (consumer pages counted as protected authoring). Re-run after V0.1.a is corrected; the CMS routes with no auth (apps/veritas/cms/src/server.ts) are a boundary finding this ledger must list, not a zero. -->
    • V0.1.d Assign one canonical host/route/API owner and migration/redirect/ removal disposition to every retained capability. Evidence: route map.
    • V0.1.e Verify authorized entry/back/deep-link behavior and prevent reader/operator navigation from exposing protected newsroom surfaces. Evidence: route tests.
  • V0.2 Map all libs/veritas/* packages to newsroom capabilities or explicit out-of-scope products. Evidence: complete package ledger. <!-- SPEC 2026-09-11: .c starts from the Phase V header note's list of 13 zero-test packages and the unconsumed CMS; a package with no tests and no consumer needs an explicit keep/remove disposition, not silence. libs/oshun/domain-veritas and libs/oshun/veritas-cascade-worker are V1-side consumers to record as retained. -->
    • V0.2.a Enumerate every package, public export, consumer, runtime target, persistence/integration dependency, test, and documented purpose. Evidence: package inventory. <!-- evidence: .domains.veritas.packages — 64 libs/veritas/* packages with exports/consumer/dependency/test/disposition. owning host/service + accountable team are V0.2.b (semantic). evidence/v1-workbenches/domain-package-inventories.json (254 pkgs; per-package exports/consumers/deps/disposition/tests + capability catalog with exact source locations). -->
    • V0.2.b Map retained exports to atomic newsroom capabilities, canonical contracts, owning host/service, and accountable team. Evidence: capability ledger.
    • V0.2.c Classify duplicate/legacy/dead/demo/reader/commercial/platform code with keep/merge/migrate/deprecate/remove/out-of-scope disposition. Evidence: disposition ledger.
    • V0.2.d Identify protected-source, authorization, standards, provenance, and editorial-policy bypasses at package boundaries. Evidence: boundary review.
    • V0.2.e Verify every package and export has one reviewed disposition and no retained consumer depends on an orphan or forbidden path. Evidence: completeness gate.
  • V0.3 Reconcile canonical V1 contracts with legacy Veritas API/client models. Evidence: translation boundaries and no new third vocabulary. <!-- SPEC 2026-09-11: "canonical" here is libs/contracts/src/veritas (the V1 contracts libs/oshun/domain-veritas already consumes); "legacy" is both libs/veritas/models/src/schemas (Zod, nested) and the Prisma schema (flattened) plus the raw SQL shapes in apps/veritas/api. .b decides per mismatch; .d must not add a fourth vocabulary in the CMS. -->
    • V0.3.a Inventory canonical and legacy story/source/claim/evidence/review/ release field names, types, enums, defaults, nullability, and versions. Evidence: contract diff. <!-- veritas-contract-diff.json: Veritas carries TWO typed representations per entity — the generated Prisma DB schema (libs/veritas/database) and the Zod domain schema (libs/veritas/models/src/schemas). This diffs their field sets for story(Article)/source(Source)/claim(Claim)/evidence (ClaimEvidence)/review(ClaimVerification): 60 shared fields, 16 Prisma-only (DB-internal ids/timestamps/soft-delete/counters), 34 Zod-only. Surfaces a REAL structural mismatch — Zod NESTS (ownership, bias) while Prisma FLATTENS (ownershipName/ownershipType, biasConfidence/biasMethodology/…) — which is exactly the semantic mismatch V0.3.b must reconcile. Nullability captured per Zod field (.nullable()/.optional()). The Zod parser walks brace depth to take only top-level keys (nested object keys excluded). release has no dedicated schema pair (publication is derived, not a first-class entity — recorded honestly). generate-veritas-contract-diff.mjs + 6 tests, wired into inventory/verify + CI. V0.3.b (DECIDE canonical source per mismatch), V0.3.c (transform contracts), V0.3.d+ (migration) are design and stay open. -->
    • V0.3.b Decide canonical source of truth and explicit adapter/migration boundary for each semantic mismatch. Evidence: mapping table.
    • V0.3.c Define lossless conversions or documented refusal/known-loss behavior for legacy-only and canonical-only values. Evidence: transform contracts.
    • V0.3.d Update serializers/clients/routes/storage mappings without publishing a hybrid third vocabulary. Evidence: schema/API tests.
    • V0.3.e Round-trip historical/current fixtures across server, clients, jobs, CMS, and public projections. Evidence: compatibility goldens.
  • V0.4 Reconcile story record states, canonical editorial workflow states, and legacy snake-case states. Evidence: lossless mapping or approved migration. <!-- SPEC 2026-09-11: the nine vocabularies V0.4.a found include the CMS ARTICLE_STATUSES in apps/veritas/cms/src/config.ts and the contracts' RetractionState; .b's canonical machine lives in libs/contracts/src/veritas, and .d's migration rehearsal runs against the Prisma database created per the Phase V header note. -->
    • V0.4.a Inventory every persisted/API/client/job/CMS/public state value, transition, terminal condition, and side effect. Evidence: state inventory. <!-- evidence: veritas-state-inventory.json — 9 editorial/workflow state Zod enums across 9 DISTINCT value vocabularies (kebab-case + lowercase), each with binding name/values/source/case-convention. sharedValues surfaces the reconciliation collisions (draft/pending/completed each recur in 4 enums; published in 3). The state VALUES are fully inventoried; the transition graph + terminal/side-effect mapping is V0.4.b (scattered state-machine logic, not one registry — noted honestly). generate-veritas-state-inventory.mjs + 5 tests. -->
    • V0.4.b Define canonical state machine, orthogonal embargo/review/release/ correction dimensions, actor/guard rules, and invalid transitions. Evidence: transition specification.
    • V0.4.c Map legacy snake-case and story-record states with explicit ambiguity, split/merge, terminal, and unknown handling. Evidence: mapping matrix.
    • V0.4.d Migrate durable records/events/jobs/read models idempotently and preserve history, audit, schedules, and in-flight work. Evidence: migration rehearsal.
    • V0.4.e Run mixed-version and transition/property tests; reject unmapped states rather than defaulting to a publishable state. Evidence: state compatibility suite.
  • V0.5 Define reporter, researcher, producer, assignment editor, copy editor, fact-checker, standards editor, legal reviewer, source-protection officer, publisher, corrections editor, and administrator roles. Evidence: permission matrix. <!-- SPEC 2026-09-11: no Veritas runtime carries a role today — the API JWT has userId/email only, the CMS has no auth. .a–.d are the matrix (Proposed); .e is buildable only after V6.4/V6.5 give the roles a server to enforce them, so it stays open until then and says so. -->
    • V0.5.a Define duties and least-privilege object/field/function rights for each named role across pitch through archive/correction. Evidence: role definitions.
    • V0.5.b Define desk/project/story/source-compartment/locale/territory scope, temporary delegation, break-glass, and separation-of-duties. Evidence: scope policy.
    • V0.5.c Define create/read/edit/review/attest/waive/publish/correct/ retract/export/admin rights per state and data class. Evidence: permission matrix.
    • V0.5.d Define incompatible role combinations, recusal/conflict effects, quorum, and independent specialist approvals. Evidence: SoD tests.
    • V0.5.e Verify UI projection and server enforcement for positive, negative, stale-session, revoked, cross-desk, and cross-compartment cases. Evidence: authorization suite.
  • V0.6 Define protected-source compartment, access-grant, emergency-access, disclosure, retention, and audit policy. Evidence: security/legal review.
    • V0.6.a Define compartment identity/key boundary and which source identifiers, contacts, communications, documents, locations, and linkage metadata are protected. Evidence: data-boundary model.
    • V0.6.b Define nominative/time-bound/purpose-bound access requests, approvals, step-up, session restrictions, revocation, and periodic review. Evidence: grant policy.
    • V0.6.c Define emergency-access eligibility, dual control where required, minimal disclosure, alerts, expiry, evidence preservation, and mandatory post-event review. Evidence: break-glass policy.
    • V0.6.d Define redacted projections for ordinary stories/reviews/search/ notifications/logs/analytics/exports/support and safe reidentification boundary. Evidence: disclosure rules.
    • V0.6.e Define retention/deletion/legal hold/key destruction/backup/ recovery and subject-safety exception behavior. Evidence: lifecycle policy.
    • V0.6.f Define tamper-evident access/attempt/export/disclosure audit with protected details visible only to authorized reviewers. Evidence: security/legal approval.
  • V0.7 Define editorial AI disclosure and acceptable-use policy. Evidence: authoring, publication, correction, and reader-visible requirements.
    • V0.7.a Classify allowed, approval-required, prohibited, and human-only AI uses across research, sourcing, drafting, verification, media, localization, distribution, correction, and moderation. Evidence: use matrix.
    • V0.7.b Define input restrictions for protected sources, personal/ licensed/embargoed/legal data, external providers, retention, and model training. Evidence: data-use policy.
    • V0.7.c Define per-change provenance, model/provider/version, prompt/tool/ source use, human acceptance, and audit requirements. Evidence: provenance policy.
    • V0.7.d Define internal reviewer and reader-visible disclosure thresholds by materiality, content type, target, locale, and correction status. Evidence: disclosure rules.
    • V0.7.e Define prohibited fabricated sources/quotes/citations/attribution, automated attestation/approval/publication/correction, and enforcement. Evidence: acceptable-use controls.
    • V0.7.f Define incident, correction, notification, model rollback, and retrospective duties for AI-caused errors. Evidence: governance review.
  • V0.8 Define conflicts, recusal, attestation independence, false-balance, corrections, takedown, and retraction policies. Evidence: ethics decision records.
    • V0.8.a Define personal/financial/political/source/advertiser/partner/ prior-work conflicts, declarations, scope, updates, and visibility. Evidence: conflict policy.
    • V0.8.b Define recusal effects on assignment, editing, review, attestation, publishing, correction, reassignment, and audit. Evidence: recusal policy.
    • V0.8.c Define attestor credential/domain/independence/conflict/ disagreement/expiry floors and prohibited manufactured independence. Evidence: attestation policy.
    • V0.8.d Define evidence-based counterclaim surfacing and false-balance suppression with uncertainty and human accountability. Evidence: balance decision record.
    • V0.8.e Distinguish clarification/update/correction/takedown/retraction, thresholds, authority, reader notice, distribution, appeal, and restoration. Evidence: remedy policy.
    • V0.8.f Define emergency/legal requests without silent deletion or retraction suppression and require immutable decisions. Evidence: ethics approvals.
  • V0.9 Threat-model source coercion, identity leakage, poisoned evidence, prompt injection, fabricated citations, defamation, legal exposure, embargo leak, account compromise, mass publishing, and retraction suppression. Evidence: mitigation/test map.
    • V0.9.a Map actors, assets, trust boundaries, data flows, privileged actions, public/distribution surfaces, and protected compartments. Evidence: threat-model diagrams.
    • V0.9.b Model coercion, impersonation, correlation, metadata/log/search/ notification leakage, compromised reviewer, and emergency-access abuse. Evidence: source-protection threats.
    • V0.9.c Model poisoned/malicious evidence, prompt/tool injection, fabricated/washed citations, provenance tamper, and correlated agent failure. Evidence: integrity threats.
    • V0.9.d Model defamation/privacy/contempt/rights exposure, unsafe media, false balance, election/emergency misuse, and jurisdictional conflict. Evidence: harm/legal threats.
    • V0.9.e Model embargo/schedule leak, credential/session compromise, mass publish/notification, correction/retraction suppression, and audit tamper. Evidence: operations threats.
    • V0.9.f Assign preventive/detective/recovery controls, owners, residual risk, abuse tests, monitoring, and review cadence to every scenario. Evidence: mitigation/test map.
  • V0.10 Define data classification for pitches, drafts, embargoes, source identities, communications, evidence, legal notes, unpublished media, analytics, and public content. Evidence: retention and residency rules.
    • V0.10.a Inventory each named data class plus derived/indexed/cached/log/ backup/export/model-context forms and accountable owner. Evidence: data catalog. <!-- veritas-data-catalog.json: derived from the generated Prisma client (libs/veritas/database/src/generated/client), the authoritative data model. 67 named data classes: 57 primary entities (Article, Source, Claim, Election, DpaConsent, ...) + 10 satellite forms tagged from model naming (model-context/embedding 1 = ArticleEmbedding; derived render 2 / variant 2 / revision 1 / translation 1 / normalized 1; log outbox 1 / delivery 1), classified against a 45-enum canonical vocabulary (SourceKind, ContentPriority, DpaLawfulBasis, ElectionType, ...). Of V0.10.a's 7 named form types, derived/model-context/log are entity-modeled; indexed/cached/backup/export are recorded HONESTLY as infrastructure forms (DB indexes, cache layer, backup/DR, export jobs — NOT Prisma entities), not fabricated. Owner = the veritas domain package (named human owners are V0.10.b). The sensitivity/ compartment/lawful-basis CLASSIFICATION (V0.10.b) and encryption/residency/ retention DEFINITIONS (V0.10.c+) are design/judgment and stay unchecked. generate-veritas-data-catalog.mjs + 7 tests, wired into inventory/verify + CI. -->
    • V0.10.b Assign sensitivity, compartment, tenant/desk/story scope, lawful basis/purpose, and permitted actors/actions. Evidence: classification matrix.
    • V0.10.c Define encryption/key, region/residency/transfer, provider, masking/minimization, and device/offline restrictions. Evidence: handling rules.
    • V0.10.d Define active/archive/legal-hold/backup retention, deletion/ anonymization, correction preservation, and conflict resolution. Evidence: lifecycle schedule.
    • V0.10.e Define safe public/redacted/aggregate projections and prohibit protected data in analytics/search/logs/notifications/support/model training. Evidence: projection policy.
    • V0.10.f Verify rules against representative end-to-end flows and obtain privacy/security/legal/source-protection approval. Evidence: signed review.
  • V0.11 Define newsroom SLOs and success metrics without optimizing for volume at the expense of accuracy. Evidence: quality guardrails.
    • V0.11.a Define availability/latency/durability/freshness/recovery SLOs separately for protected authoring, publish, correction/retraction, and public read-back. Evidence: service objectives.
    • V0.11.b Define accuracy, citation/support, gate escape, correction propagation, source freshness, accessibility, and safety quality measures. Evidence: quality metric catalog.
    • V0.11.c Define measurement windows, denominators, exclusions, data lineage, privacy floors, uncertainty, and alert thresholds. Evidence: metric contracts.
    • V0.11.d Define balanced scorecard rules so speed/output/engagement/ revenue never override accuracy, protected-source, ethics, or correction obligations. Evidence: guardrail policy.
    • V0.11.e Define error budgets, stop-the-line conditions, escalation, remediation, and release consequences. Evidence: SLO policy.
    • V0.11.f Validate dashboards with known-value fixtures and prevent individual/source identification or automatic editorial scoring. Evidence: observability review.
  • V0.12 Ratify the NewsML-G2, IPTC Media Topics, C2PA, and CMS publication profiles. Evidence: standards decision record. <!-- SPEC 2026-09-11: nothing to ratify exists for NewsML-G2 or IPTC (zero code). .a pins the standard versions and the validators (IPTC NewsML-G2 2.x XSD bundle + the IPTC Media Topics vocabulary release, both public downloads to be vendored with sha256; C2PA via c2pa-node@0.5.26 already installed); .f's approval is human (§2.5.9). -->
    • V0.12.a Pin standard/profile/schema/validator versions, namespaces, required/optional extensions, compatibility window, and update owner. Evidence: standards manifest.
    • V0.12.b Map internal stories/packages/planning/concepts/topics/rights/ embargo/update/correction fields to NewsML-G2/IPTC semantics. Evidence: mapping matrix.
    • V0.12.c Define governed IPTC topic selection, aliases, localization, deprecation, and internal taxonomy relationships. Evidence: vocabulary profile.
    • V0.12.d Define C2PA claims/ingredients/actions/signers/trust validation for media and article packages, including stripped/unknown provenance. Evidence: provenance profile.
    • V0.12.e Define CMS plan/publish/read-back/update/correct/retract behavior, target extensions, idempotency, and partial-failure compensation. Evidence: CMS profile.
    • V0.12.f Validate representative and edge-case fixtures, record deviations/interoperability findings, and obtain editorial/engineering approval. Evidence: decision record.
  • V0.13 Author and ratify V1/VERITAS_NEWSROOM_WORKBENCH_DESIGN_2026-07-23.md with current-state, host, roles, protected-source boundary, data, APIs, editorial policy, UI, standards, migration, rollout, and decision records. Evidence: approved doc and resolved review comments.
    • V0.13.a Document current applications/packages/routes/services/stores/ integrations, duplicate ownership, constraints, and known gaps. Evidence: current-state chapter.
    • V0.13.b Document target hosts/ownership/roles/protected-source and public boundaries, data contracts/stores/APIs/events/jobs, and threat controls. Evidence: architecture chapters.
    • V0.13.c Document editorial/ethics/AI/regulatory/standards/gate policies, raw measurement versus verdict, approvals, and remedies. Evidence: governance chapters.
    • V0.13.d Document task-led UI/IA, responsive/accessibility/offline/ performance states, semantic alternatives, and authorized reachability. Evidence: experience chapter.
    • V0.13.e Document migrations/compatibility, rollout/rollback, operations/ recovery/observability, verification, evidence, and launch criteria. Evidence: delivery chapters.
    • V0.13.f Record alternatives/decisions/owners/dates, resolve every substantive review comment, and obtain named ratification. Evidence: approved design.
  • V0.14 Reconcile every Veritas application, libs/veritas/* package, database entity, BFF/CMS/API route, reader walkthrough, agent, worker, scheduled job, and documented feature at atomic-feature level. Evidence: explicit dispositions cover ingestion, research/tips, editorial/claims, knowledge graph, multimedia/live, newsletter/social, multilingual, election/regulatory, community, notifications/emergency, reader/search/ recommendation, B2B/USSD, subscription/revenue, analytics, archive, and operations.
    • V0.14.a Build a source inventory across code, schema, routes, navigation, walkthroughs, workers/jobs/agents, documentation, flags, and deployed integrations. Evidence: reconciliation inputs. <!-- evidence: veritas-code-census.json — comprehensive launch census: 64 libs packages + 13 app subsystems (9 worker/agent/media: agents/ai-workers/ analytics/audio/ingestion/nlp/notifications/social/video), 1270 code files + 156 tests, 58 routes across studio/reader/CMS/BFF, 9 state enums (boundary clean), 79 docs, and 28 deploy targets (web + mobile android/ios + workers + prisma). Consolidated from git counts + committed V0.1/V0.4 evidence. Atomic-feature split + dispositions are V0.14.b+ (semantic). generate-veritas-code-census.mjs + 4 tests. -->
    • V0.14.b Split each source item into atomic user/operator/system capabilities with host, audience, owner, dependencies, data class, and current implementation status. Evidence: atomic ledger.
    • V0.14.c Cover ingestion/research/tips, editorial/claims/graph, multimedia/live, newsletter/social/messaging, multilingual/election/ regulatory, and community explicitly. Evidence: editorial coverage.
    • V0.14.d Cover reader/search/recommendation/archive, notifications/ emergency, B2B/USSD/voice, subscriptions/revenue/analytics, and operations explicitly. Evidence: audience/operations coverage.
    • V0.14.e Assign keep/consolidate/migrate/replace/deprecate/remove/ externalize decisions with rationale, target owner/host, dependencies, and delivery/expiry. Evidence: disposition ledger.
    • V0.14.f Reconcile ledger to routes/packages/entities/contracts/tests and fail any orphan, duplicate owner, unknown feature, or unreviewed row. Evidence: completeness gate.
  • V0.15 Define the boundary among protected newsroom authoring, public reader/mobile, source/tip intake, live/community moderation, commercial operations, B2B developer, USSD/voice, and platform-operator surfaces. Evidence: every retained capability has one host and audience plus a redacted handoff where appropriate. <!-- SPEC 2026-09-11: the surfaces that exist today are: consumer room (apps/oshun/web/src/app/veritas/*, /domains/veritas/*), public reader (apps/veritas/web), commercial/B2B/USSD routes inside apps/veritas/api (payments, subscriptions, ads, b2b-api, tips), and the unauthenticated CMS. Protected newsroom authoring is absent. .b assigns every one of these; .f's suite runs against apps/veritas/api and the CMS as served. -->
    • V0.15.a Define audience/trust/data/action boundaries for each named surface and prohibit shared UI/service assumptions that cross them. Evidence: surface boundary matrix.
    • V0.15.b Assign every retained capability and privileged mutation to one canonical host/owner, with authorized role/project/channel context. Evidence: host ledger.
    • V0.15.c Define typed redacted handoffs for tips, public feedback, moderation, live operations, commercial conflicts, B2B consumption, and operator incidents. Evidence: handoff contracts.
    • V0.15.d Define identity/session/token/deep-link/storage/logging/ notification separation and step-up transitions between surfaces. Evidence: security model.
    • V0.15.e Define public correction/retraction and emergency propagation without granting public/commercial/channel systems editorial authority. Evidence: control-flow diagrams.
    • V0.15.f Test authorized journeys and attempted cross-surface field/ function/object access; eliminate duplicate mutations. Evidence: boundary suite.
  • V0.16 Ratify Ghana NMC/editorial, election-period, data-protection, accessibility, child/trauma, advertising/sponsorship, payment, emergency broadcast, and platform-channel obligations separately. Evidence: jurisdiction/version/owner matrix and feature-specific enforcement plan. <!-- SPEC 2026-09-11: the only code-owned floors today are apps/veritas/api/src/domain/compliance/{nmc,election}.ts and the nmc_complaints table. .a needs counsel; an agent may fill the register with the obligation text and leave reviewer/effective date blank (§2.5.9). -->
    • V0.16.a Identify current authoritative obligations, jurisdiction/scope/ effective date/version/review cadence, counsel or specialist owner, and conflicts for each named regime. Evidence: obligation register.
    • V0.16.b Translate NMC/editorial/election obligations into content/ source/claim/review/schedule/publish/complaint/remedy controls. Evidence: editorial enforcement map.
    • V0.16.c Translate data-protection/accessibility/child/trauma obligations into collection, minimization, consent, display, distribution, retention, request, and review controls. Evidence: rights/safety map.
    • V0.16.d Translate advertising/sponsorship/payment obligations into separation, disclosure, consent, entitlement, refund, audit, and newsroom-independence controls. Evidence: commercial map.
    • V0.16.e Translate emergency-broadcast and each platform/channel rule into authority, template, targeting, delivery, correction/recall, moderation, rate, and archival controls. Evidence: channel map.
    • V0.16.f Link each obligation to feature/code-owned floor, policy version, tests/evidence/approver, operator runbook, and update trigger; record unresolved legal questions. Evidence: ratified matrix.

V1 — Story, source, claim, evidence, and editorial contracts#

  • V1.1 Define story, pitch, assignment, draft, block, headline, dek, byline, topic, desk, edition, package, media, and publish-target entities. Evidence: schema fixtures.
    • V1.1.a Define stable identity, tenant/newsroom/desk/project scope, lifecycle, version, timestamps, actor, and audit fields for each entity. Evidence: base schemas.
    • V1.1.b Define pitch→assignment→story→draft/block and headline/dek/byline ownership/cardinality/order/reference invariants. Evidence: editorial relationship fixtures.
    • V1.1.c Define topic/desk/edition/package/media/publish-target membership, locale/territory/rights, scheduling, and release relationships. Evidence: distribution fixtures.
    • V1.1.d Define state-transition, deletion/archive, validation, unknown enum, and cross-tenant reference constraints. Evidence: schema invariant tests.
    • V1.1.e Publish storage/API/event/read-model schemas with versioning and representative valid/invalid fixtures. Evidence: contract package.
  • V1.2 Define immutable draft revisions and semantic block/range anchors. Evidence: paragraph, quote, claim, and comment anchors survive edits or become explicitly orphaned.
    • V1.2.a Define immutable revision identity, parent(s), author, cause, schema/editor version, created time, and content hash. Evidence: revision contract.
    • V1.2.b Define stable block identity/type/order and semantic text-range anchors with context/fingerprint/affinity. Evidence: anchor schema.
    • V1.2.c Define deterministic anchor mapping across insert/delete/move/ split/merge/replace and concurrent-operation histories. Evidence: transformation tests.
    • V1.2.d Preserve paragraph/quote/claim/citation/comment/review anchors or mark ambiguous/orphaned with candidates and reason; never silently retarget. Evidence: edit corpus.
    • V1.2.e Define restore/branch/merge/compaction/migration and archival behavior without mutating historical revisions. Evidence: history tests.
  • V1.3 Use the canonical 15-kind source vocabulary. Evidence: no lossy fallback to the smaller legacy type enum.
    • V1.3.a Pin the canonical 15 identifiers, labels, definitions, permitted locators/metadata, and unknown/extension policy. Evidence: vocabulary contract.
    • V1.3.b Inventory every legacy/API/client/storage/index/import/export source enum and its consumers. Evidence: usage ledger.
    • V1.3.c Map each legacy value losslessly or require explicit review/ refusal for ambiguous values; never collapse to an approximate kind. Evidence: migration matrix.
    • V1.3.d Update validation/forms/search/facets/scoring/gates/NewsML/CMS adapters to preserve the canonical kind. Evidence: integration tests.
    • V1.3.e Round-trip all 15 kinds and unknown historical inputs across storage/API/UI/jobs/export. Evidence: kind fixture suite.
  • V1.4 Define source identity separately from protected identity details. Evidence: ordinary source records cannot expose compartment secrets.
    • V1.4.a Define ordinary source pseudonymous/public identity, descriptive metadata, type, publisher/organization, quality, and safe display key. Evidence: ordinary source schema.
    • V1.4.b Define separately encrypted/compartmented legal/contact/ communications/location/safety/linkage identity details and keys. Evidence: protected schema.
    • V1.4.c Define opaque join/reference/tokenization and authorized reveal/ reidentification workflows without inferable sequential identifiers. Evidence: boundary contract.
    • V1.4.d Define redacted projections for stories, claims, reviews, search, notifications, logs, analytics, exports, backups, and public products. Evidence: projection fixtures.
    • V1.4.e Test object/field/function/correlation access across roles, tenants, compartments, revocation, and emergency grants. Evidence: negative suite.
  • V1.5 Define source quality vector, domain weight profile, band, adjustments, reasons, evaluator version, and evidence. Evidence: canonical algorithm fixture.
    • V1.5.a Define each raw factor’s meaning, domain/range/unit, missing/ unknown handling, evidence locator, and measurement provenance. Evidence: factor schema.
    • V1.5.b Define versioned domain weight profiles, applicability, normalization, combination/rounding, band thresholds, and confidence. Evidence: algorithm specification.
    • V1.5.c Define manual adjustment direction/magnitude/floor/cap/reason/ evidence/authority/expiry and keep it separate from raw measurement. Evidence: adjustment contract.
    • V1.5.d Define result identity, source/revision/domain, evaluator/ weights/policy versions, inputs, timestamp, staleness, and supersession. Evidence: result schema.
    • V1.5.e Publish hand-calculated positive/boundary/missing/adversarial fixtures and property invariants. Evidence: canonical goldens.
  • V1.6 Define claim class, text, anchor, severity/risk, confidence band, retraction state, supporting sources, counter-sources, and status. Evidence: canonical contract invariants.
    • V1.6.a Define claim identity/class/text/normalized proposition, story/ draft/revision/anchor scope, author/origin, and lifecycle. Evidence: claim schema.
    • V1.6.b Define independent severity/harm/legal risk, evidential confidence, verification/editorial status, and contested/retraction state vocabularies. Evidence: state model.
    • V1.6.c Define typed supporting/contradicting/contextual/irrelevant source links with citation locator, assessment, version, and reviewer. Evidence: evidence-link schema.
    • V1.6.d Define status-transition, minimum evidence, stale-source, counter-source, merge/split, anchor orphan, correction, and retraction invariants. Evidence: contract tests.
    • V1.6.e Define protected/public/read-model projections and prohibit confidence/status conflation or unsupported “verified” defaults. Evidence: projection fixtures.
  • V1.7 Define citation locator types for page, section, paragraph, quote, dataset slice, court record, transcript, audio/video timecode, image region, and URI fragment. Evidence: validation fixtures.
    • V1.7.a Define a discriminated locator contract with common source/ revision/content-hash identity and kind-specific fields. Evidence: locator schema.
    • V1.7.b Define page/section/paragraph/quote locators with numbering, labels, text/context fingerprints, editions, and range rules. Evidence: document fixtures.
    • V1.7.c Define dataset/court/transcript locators with table/query/row/ column/cell, jurisdiction/docket/filing, speaker/turn/line semantics. Evidence: structured fixtures.
    • V1.7.d Define media/URI locators with time/range/track/frame/region/ coordinate/version/fragment and bounded normalization. Evidence: media/ web fixtures.
    • V1.7.e Define resolution/unreachable/changed/ambiguous/stale handling and prohibit silent relocation to new source content. Evidence: validation corpus.
  • V1.8 Define evidence packs with story/claim/source membership, reviewer, state, version, and integrity hash. Evidence: referential-integrity tests.
    • V1.8.a Define immutable pack identity/version/parent, story/draft/ revision scope, purpose, creator, created time, and integrity algorithm. Evidence: pack schema.
    • V1.8.b Define ordered typed membership for claims/sources/citations/ excerpts/media/measurements/attestations with pinned versions/hashes. Evidence: membership schema.
    • V1.8.c Define draft/in-review/approved/rejected/superseded/invalid state, reviewer/quorum/decision, and mutation-invalidates-approval rules. Evidence: pack state machine.
    • V1.8.d Define canonical serialization/hash/signature verification and behavior for missing, changed, unauthorized, or retracted members. Evidence: integrity tests.
    • V1.8.e Test cross-story/tenant/compartment references, duplicate/ circular membership, stale versions, and tampering. Evidence: referential-integrity suite.
  • V1.9 Define counterclaim and unsupported-claim entities without manufacturing equivalence. Evidence: canonical four-way surfacing result.
    • V1.9.a Define counterclaim identity/proposition/class/origin/anchor/ evidence/confidence/status and relation to one or more subject claims. Evidence: counterclaim schema.
    • V1.9.b Define unsupported-claim reason, missing evidence, attempted verification, severity, owner, remediation, and allowed disposition. Evidence: unsupported schema.
    • V1.9.c Define versioned measurements for relevance, evidential quality, independence, consensus, harm, uncertainty, and false-balance risk. Evidence: assessment contract.
    • V1.9.d Define four distinct surfacing results—surface, suppress false balance, require more evidence, editorial review—with reason/evidence/ policy version. Evidence: decision schema.
    • V1.9.e Test that popularity/number of sources alone cannot create equivalence and that contested/unknown/evaluator failure never becomes settled. Evidence: invariant fixtures.
  • V1.10 Define attestor identity, scope, credential, domain, conflict, validity, disagreement, and decision. Evidence: attestation invariants.
    • V1.10.a Define attestor person/organization/service identity and verified contact/credential references separate from public display identity. Evidence: identity schema.
    • V1.10.b Define attestation scope over exact claim/source/evidence-pack/ draft revision, domain/jurisdiction, question, and permitted conclusion. Evidence: scope contract.
    • V1.10.c Define credential/issuer/verification/domain fit, conflict/ independence declaration, validity window, revocation, and expiry. Evidence: eligibility schema.
    • V1.10.d Define agree/disagree/qualified/abstain/unable decisions with rationale/evidence/uncertainty, signature, and immutable timestamp. Evidence: decision contract.
    • V1.10.e Define disagreement aggregation, quorum, specialist floors, supersession, stale-input invalidation, and public/redacted projection. Evidence: attestation invariants.
  • V1.11 Define correction notes, prior public version, reader-visible change, distribution action, and affected claims. Evidence: chronological and text-match invariants.
    • V1.11.a Define correction identity/type/severity/status/reason/discovery/ owner, exact story/release/locale/target, and immutable chronology. Evidence: correction schema.
    • V1.11.b Pin prior and replacement public revisions/blocks/ranges/text/ media with semantic diff and hashes. Evidence: version-link contract.
    • V1.11.c Define affected claims/citations/sources/packages and whether evidential confidence, status, or downstream content changes. Evidence: impact schema.
    • V1.11.d Define reader-visible note wording/placement/timestamp/history and target-specific update/replace/notify actions. Evidence: notice/ action schema.
    • V1.11.e Enforce prior-text match, chronological append-only history, immutable published evidence, idempotency, and correction-of-correction. Evidence: invariant tests.
  • V1.12 Define retraction cascade, cause, source/claim/story graph, actions, target states, owner, progress, and verification. Evidence: schema.
    • V1.12.a Define cascade identity/cause/scope/severity/authority/trigger/ source revision and required emergency/standard process. Evidence: cascade schema.
    • V1.12.b Represent affected source→citation→claim→story→package→target dependency graph with versioned discovery and completeness state. Evidence: impact graph.
    • V1.12.c Define per-node/edge retain-correct-withdraw-retract-unpublish/ notify/reindex/invalidate actions and desired target state. Evidence: action plan schema.
    • V1.12.d Define owner/approval/dependency/order/attempt/idempotency/ progress/failure/retry/compensation state per action. Evidence: execution schema.
    • V1.12.e Define downstream read-back verification, unresolved/unsupported target handling, closure criteria, audit, and reopen behavior. Evidence: verification contract.
  • V1.13 Define embargo, schedule, exclusivity, rights, territory, and syndication metadata. Evidence: policy tests.
    • V1.13.a Define embargo policy independently from editorial/release state with start/end/event trigger/time zone/authority/beneficiaries. Evidence: embargo schema.
    • V1.13.b Define schedule windows, target/locale/edition dependencies, readiness conditions, retries, missed-window, and reschedule semantics. Evidence: schedule schema.
    • V1.13.c Define exclusivity type/party/window/channel/exception and relationship to embargo, rights, and syndication. Evidence: exclusivity contract.
    • V1.13.d Define rights holder/license/use/format/channel/territory/ language/duration/attribution/fee/expiry and proof. Evidence: rights schema.
    • V1.13.e Define syndication partner/package/transform/attribution/ correction-retraction obligations and validate overlapping/conflicting/ expired policies. Evidence: policy tests.
  • V1.14 Define NewsML-G2 News Item, Package Item, Concept/Knowledge Item, Planning Item, and update/correction mappings. Evidence: profile matrix. <!-- SPEC 2026-09-11: greenfield — no NewsML code exists. Put the mapping in libs/contracts/src/veritas/newsml-g2.ts with goldens under libs/contracts/src/veritas/goldens/newsml/, and validate .f against the vendored IPTC XSDs (V0.12.a), never against our own parser only. -->
    • V1.14.a Map story releases and media renditions to News Item identity, version/status, content metadata, rights, subjects, and content sets. Evidence: News Item mapping.
    • V1.14.b Map editions/collections/related coverage and ordered/ role-qualified membership to Package Items. Evidence: Package mapping.
    • V1.14.c Map people/organizations/places/events/topics/claims and their aliases/relationships/validity/provenance to Concept/Knowledge Items. Evidence: concept mapping.
    • V1.14.d Map assignments/coverage plans/schedules/embargoes/deliverables/ contacts to Planning Items without leaking protected sources. Evidence: planning mapping.
    • V1.14.e Define initial/update/correction/replacement/retraction version chains, signals, prior references, and target read-back. Evidence: lifecycle matrix.
    • V1.14.f Validate required/optional/extension/cardinality/code-list rules against the pinned profile and goldens. Evidence: profile tests.
  • V1.15 Define C2PA media and article-package provenance relationships. Evidence: ingredient and editing lineage fixture. <!-- SPEC 2026-09-11: build on libs/veritas/content-auth (its c2paValidator seam is the boundary) and c2pa-node@0.5.26; a manifest state must be one of valid/tampered/stripped/unknown/untrusted/expired and never "verified" when the validator was not configured. -->
    • V1.15.a Define claim generators/signers/trust stores/identities and supported C2PA version/algorithms/assertions. Evidence: trust profile.
    • V1.15.b Map original media captures/imports as ingredients with hashes, capture/source metadata, rights, and protected-data minimization. Evidence: ingredient schema.
    • V1.15.c Record edit/transform/AI/action lineage, tools/versions/ parameters, ingredient relationships, and derived renditions. Evidence: action lineage.
    • V1.15.d Relate article/package manifests to exact media/article revisions, NewsML/CMS release, corrections, and replacements. Evidence: package relationship.
    • V1.15.e Define valid/tampered/stripped/unknown/untrusted/expired states and display/gate behavior without treating absence as falsity. Evidence: validation fixtures.
  • V1.16 Define version compatibility with existing Veritas records. Evidence: migration goldens.
    • V1.16.a Inventory deployed record/event/blob/index/API/client/CMS/ distribution versions and historical edge cases. Evidence: version corpus.
    • V1.16.b Define support window and per-version read/write/migrate/export/ rollback behavior for every contract in V1. Evidence: compatibility matrix.
    • V1.16.c Write deterministic adapters/migrations for renamed/split/ merged/encrypted/compartmented/versioned fields with explicit loss. Evidence: migration code/tests.
    • V1.16.d Rebuild indexes/read models and reconcile references/hashes/ anchors/state/history after migration. Evidence: rehearsal report.
    • V1.16.e Round-trip goldens through old/current readers and writers, interrupted retries, unknown fields, and downgrade refusal. Evidence: migration suite.
    • V1.16.f Document irreversible steps, backup/restore, rollout/rollback, telemetry, and operator approval. Evidence: migration runbook.

V2 — Pitch, assignment, research, and source registry#

  • V2.1 Build zero-form pitch capture with desk, topic, thesis, public interest, urgency, location, sources, risks, format, and target window. Evidence: keyboard-first journey.
    • V2.1.a Define a fast capture command/inbox entry with minimal required text, reporter identity, timestamp, origin, and draft privacy. Evidence: capture contract.
    • V2.1.b Add progressive structured extraction/editing for desk/topic/ thesis/public interest/urgency/location/format/window without blocking initial capture. Evidence: field tests.
    • V2.1.c Attach existing/new/protected source references and safety/legal/ conflict/embargo risks through authorized redacted projections. Evidence: source/risk tests.
    • V2.1.d Support save/duplicate/merge/submit/park/discard with autosave, stable identity, validation, audit, and explicit state. Evidence: lifecycle tests.
    • V2.1.e Complete capture/triage/edit/submit using keyboard and screen reader on responsive web/desktop without a long mandatory form. Evidence: keyboard-first E2E.
  • V2.2 Build story-angle and headline divergence behind the fenced provider seam. Evidence: exact outputs, attribution, and no auto-selection.
    • V2.2.a Define request contract with pinned pitch/story revision, allowed source excerpts, constraints, locale/desk/style, count, and budget. Evidence: provider request schema.
    • V2.2.b Generate materially distinct angle and headline candidates separately, preserving factual constraints and protected-data policy. Evidence: known-input tests.
    • V2.2.c Return exact candidate text, rationale, source-use/citations, model/provider/prompt/version, cost, uncertainty, and safety results. Evidence: attributed response.
    • V2.2.d Present unranked or transparently scored candidates for human compare/edit/accept/reject with no hidden winner or automatic mutation. Evidence: selection tests.
    • V2.2.e Evaluate duplication, factual drift, sensationalism, source invention, protected leakage, provider failure, and deterministic replay. Evidence: divergence eval.
  • V2.3 Build pitch comparison with novelty, evidence readiness, risk, audience, production scope, and conflict views. Evidence: score provenance.
    • V2.3.a Define comparable raw measures, units, missing/unknown handling, weights/policies, and provenance for each named dimension. Evidence: comparison contract.
    • V2.3.b Select two or more pitch revisions and show normalized side-by- side thesis/source/evidence/risk/audience/scope/conflict differences. Evidence: comparison UI tests.
    • V2.3.c Show measurement evidence, evaluator/weight/policy versions, uncertainty, freshness, and manual adjustments rather than opaque scores. Evidence: provenance tests.
    • V2.3.d Preserve protected-source redaction and prevent count/timing/ similarity leakage to unauthorized comparators. Evidence: auth tests.
    • V2.3.e Record commission/merge/park/kill decision with chosen revision, rationale, conflicts/recusal, dissent, and no automatic selection. Evidence: decision audit.
  • V2.4 Build assignment planning with owner, contributors, deadlines, deliverables, budget, embargo, dependencies, and safety plan. Evidence: state and notification tests.
    • V2.4.a Define assignment identity/story/pitch revision, accountable owner, role-qualified contributors, desk, state, priority, and scope. Evidence: assignment schema.
    • V2.4.b Define milestone/deadline/time-zone/deliverable/format/locale/ target acceptance criteria and dependency graph. Evidence: plan schema.
    • V2.4.c Define budget/currency/approval/spend constraints, embargo/ exclusivity/rights, and fieldwork/source/data/media safety plan. Evidence: constraint schema.
    • V2.4.d Implement propose/accept/decline/reassign/update/pause/complete/ cancel transitions with reason, optimistic concurrency, and audit. Evidence: state tests.
    • V2.4.e Send permission-filtered deadline/dependency/risk/change/ escalation notifications without embargo or protected-source leakage. Evidence: notification suite.
  • V2.5 Build protected-source onboarding in a compartmented workflow. Evidence: step-up authorization and redacted audit.
    • V2.5.a Begin onboarding from an authorized story/tip with purpose, pseudonymous ordinary source record, compartment request, and safety assessment. Evidence: initiation tests.
    • V2.5.b Capture identity/contact/communications/relationship/consent/ risk details only inside the encrypted compartment with minimization. Evidence: storage inspection.
    • V2.5.c Verify source-protection role, step-up, grant/quorum/expiry, device/session restrictions, and emergency contact rules. Evidence: authorization tests.
    • V2.5.d Create opaque linkage/redacted projections for assignment, claims, evidence, reviews, notifications, logs, search, and exports. Evidence: projection tests.
    • V2.5.e Exercise onboard/update/revoke/reidentify/emergency-access/ retention/legal-hold workflows with tamper-evident redacted audit. Evidence: protected-source E2E.
  • V2.6 Build public-source intake from URL, document, feed, wire, dataset, court record, transcript, image, audio, and video. Evidence: kind-specific validation.
    • V2.6.a Define intake request/status/result contracts and kind-specific metadata/locator/rights/language/size/type requirements. Evidence: source intake schemas.
    • V2.6.b Implement URL/feed/wire fetch through allowlisted network policy with redirects, robots/license/rate, authenticity, and capture metadata. Evidence: remote intake tests.
    • V2.6.c Implement document/dataset/court/transcript parsing with format/ structure/encoding/schema/table/docket/speaker validation. Evidence: structured intake tests.
    • V2.6.d Implement image/audio/video ingest with media probing, duration/ dimensions/tracks, metadata privacy, transcripts/proxies, and safe preview. Evidence: media intake tests.
    • V2.6.e Classify canonical source kind, deduplicate without destroying provenance, preserve failures/quarantine, and require human metadata/ rights review. Evidence: intake journey.
  • V2.7 Quarantine and scan remote/downloaded content. Evidence: malicious file, SSRF, redirect, archive, and decompression corpus. <!-- SPEC 2026-09-11: the fetch sites to fence are apps/veritas/ingestion, apps/veritas/api URL imports (sources, tips) and any fetch/undici call in libs/veritas/ingestion-core; the quarantine store is MinIO from docker/docker-compose.dev.yml. SSRF cases here are the same list V9.7.f reuses. -->
    • V2.7.a Place all untrusted bytes in isolated non-executable quarantine before parsing/preview/indexing/model use. Evidence: storage boundary tests.
    • V2.7.b Enforce URL scheme/host/IP/port/redirect/DNS/private-network/ credential/size/time policy at every fetch hop. Evidence: SSRF corpus.
    • V2.7.c Detect type/signature mismatch, malware/active content/polyglots, malformed media, and unsafe parser conditions. Evidence: scan corpus.
    • V2.7.d Enforce archive path/symlink/nesting/member/count/ratio/ decompressed-size limits and scan every extracted member. Evidence: archive bomb tests.
    • V2.7.e Record scan engine/signature/policy versions and release/reject/ manual-review decisions; rescan on updates. Evidence: quarantine audit.
    • V2.7.f Test timeout/crash/partial scan/unknown type/false-positive override with no unscanned content escaping. Evidence: failure suite.
  • V2.8 Preserve original bytes, fetch metadata, checksum, extraction, capture time, and rights. Evidence: immutable source evidence.
    • V2.8.a Store original bytes in immutable content-addressed blobs after quarantine with tenant/compartment encryption and retention policy. Evidence: blob tests.
    • V2.8.b Record origin URL/provider/request/redirect/headers/status/ filename/MIME/size/fetcher/version and trusted clock capture time. Evidence: fetch manifest.
    • V2.8.c Compute/version strong checksums for original and every derived extraction/proxy/transcript/snapshot with parent lineage. Evidence: integrity tests.
    • V2.8.d Record extractor/parser/model/version/config/output/errors and exact source ranges so derived text remains traceable. Evidence: extraction provenance.
    • V2.8.e Capture rights/license/terms/territory/expiry/attribution/proof separately from factual provenance and block unsupported use. Evidence: rights tests.
    • V2.8.f Prove later refetch/update cannot overwrite originals and that export/read-back verifies bytes, metadata, and chain. Evidence: immutable evidence suite.
  • V2.9 Build source-quality scoring from the canonical composite and domain weights. Evidence: independently verified known values.
    • V2.9.a Implement each raw factor as a versioned pure measurement with evidence, unit, missing/unknown state, and no policy verdict. Evidence: factor tests.
    • V2.9.b Resolve the applicable pinned domain-weight profile and validate completeness/ranges/normalization before computation. Evidence: profile tests.
    • V2.9.c Implement the canonical combine/adjust/round/band/confidence algorithm exactly and preserve all intermediate values. Evidence: algorithm tests.
    • V2.9.d Persist result/source revision/evaluator/weight/policy/corpus versions and mark stale when inputs or policy change. Evidence: provenance tests.
    • V2.9.e Compare independently hand-calculated positive/boundary/missing/ adversarial fixtures and cross-language implementations. Evidence: known-value report.
  • V2.10 Build source review with factor evidence, manual adjustments, conflict, attestation, and expiry. Evidence: override authorization.
    • V2.10.a Show source identity/redacted projection, revision, raw factors, locators, measurements, algorithm, band, confidence, freshness, and prior reviews. Evidence: review UI.
    • V2.10.b Require factor-by-factor inspect/confirm/challenge/recompute and preserve reviewer disagreement instead of overwriting measurements. Evidence: review tests.
    • V2.10.c Implement bounded manual adjustment with authorized role, direction/value/reason/evidence, conflict/recusal, and second approval where required. Evidence: override tests.
    • V2.10.d Attach applicable attestation with exact scope, credential, independence, decision, expiry, and redacted protected-source handling. Evidence: attestation tests.
    • V2.10.e Set review/adjustment/attestation expiry and stale triggers for source updates, retractions, policy/weight changes, or new contrary evidence. Evidence: lifecycle tests.
    • V2.10.f Record immutable approve/reject/request-evidence/contest/ override decisions and prohibit unauthorized direct score edits. Evidence: authorization/audit suite.
  • V2.11 Build source search, dedupe, relationship, publisher history, and retraction history. Evidence: protected fields never enter the index.
    • V2.11.a Define permission-aware indexed safe fields, tokenization/ language, facets, freshness, deletion, and compartment exclusion. Evidence: search schema.
    • V2.11.b Implement exact/fuzzy/semantic retrieval with stable pagination, result reasons, current source revision, and no protected count/snippet/ suggestion/timing leaks. Evidence: search tests.
    • V2.11.c Implement deterministic/explainable duplicate candidates using public identifiers, publisher, content hashes, locators, and reviewed merge/link/not-duplicate decisions. Evidence: dedupe tests.
    • V2.11.d Model source/publisher/ownership/derivation/citation/translation/ syndication relationships with temporal validity and provenance. Evidence: relationship tests.
    • V2.11.e Show publisher correction/retraction/reliability history with evidence, scope, version, and no automatic guilt-by-association verdict. Evidence: history tests.
    • V2.11.f Inspect index/documents/logs/cache/analytics under two tenants and compartments to prove protected fields never enter search. Evidence: leakage suite.
  • V2.12 Build source communications references without storing plaintext in ordinary newsroom tables. Evidence: vault/secure-system links and access audit.
    • V2.12.a Define opaque communication-reference contract for external secure-system/provider/conversation/item identity, type, time, and safe label. Evidence: reference schema.
    • V2.12.b Store addresses, message bodies, attachments, keys, and sensitive metadata only in approved compartment/vault systems. Evidence: storage inspection.
    • V2.12.c Implement authorized create/link/open/verify/unlink with step-up, short-lived scoped tokens, current access grants, and safe return. Evidence: access journey.
    • V2.12.d Provide redacted ordinary projections that reveal only existence and workflow state when authorized, never participant/content metadata. Evidence: projection tests.
    • V2.12.e Audit attempts and successful access/export without copying plaintext into ordinary audit/log/notification/search/analytics data. Evidence: audit tests.
  • V2.13 Build research notebook and evidence-board views. Evidence: every excerpt remains anchored to a source.
    • V2.13.a Define notebook note/question/hypothesis/task/excerpt/entity/ relationship/timeline item and board grouping/order/link contracts. Evidence: workspace schema.
    • V2.13.b Create excerpts only from pinned source revisions with exact locator/text/hash/context and protected/redacted identity semantics. Evidence: anchoring tests.
    • V2.13.c Support authoring/linking/grouping/filtering/reordering/status/ assignment without converting notes or hypotheses into verified claims. Evidence: editor tests.
    • V2.13.d Detect source change/retraction/unreachable locator and show stale/orphan/candidate remap with explicit reviewer action. Evidence: update tests.
    • V2.13.e Support collaborative revision/comments/offline conflicts/ export with story/compartment permissions and legal holds. Evidence: collaboration tests.
    • V2.13.f Verify every displayed/exported excerpt resolves to exact source evidence or is visibly unresolved. Evidence: integrity gate.
  • V2.14 Build story cluster and related coverage suggestions. Evidence: human acceptance and model/version attribution.
    • V2.14.a Define candidate inputs/features/exclusions and cluster/ relationship types without exposing protected or embargoed material. Evidence: suggestion contract.
    • V2.14.b Generate cluster membership/related-story suggestions with similarity/evidence/rationale, confidence/uncertainty, and freshness. Evidence: model tests.
    • V2.14.c Attribute model/provider/version/config/index/source revisions, cost, and generation time; retain raw result separately from decisions. Evidence: provenance record.
    • V2.14.d Let editors accept/reject/merge/split/relabel suggestions individually with rationale and no automatic public curation/ranking. Evidence: decision tests.
    • V2.14.e Evaluate duplication, topic drift, feedback loops, subgroup/ language performance, correction/retraction propagation, and stale suggestions. Evidence: calibrated report.
  • V2.15 Build source update/retraction preview. Evidence: read-only blast radius across claims, stories, packages, newsletters, social, audio, video, and archives.
    • V2.15.a Ingest or draft a source update/retraction event with exact source/revision, publisher evidence, cause, scope, time, confidence, and verification state. Evidence: event contract.
    • V2.15.b Traverse versioned source→citation→claim→story→package→target dependencies without performing mutations. Evidence: impact query.
    • V2.15.c Include draft/published stories, packages, newsletters, social, audio/video timecodes, archives/search, B2B/channel consumers, and unknown/unregistered edges. Evidence: coverage tests.
    • V2.15.d Show exact affected locator/claim/release/target, current state, proposed remedy, owner, urgency, confidence, and permission-filtered redaction. Evidence: preview UI.
    • V2.15.e Pin the preview graph/version and require a separately authorized correction/retraction plan before any downstream action. Evidence: read-only enforcement.
    • V2.15.f Verify completeness against registered consumers and display unresolved/stale/unauthorized targets rather than reporting a safe blast radius. Evidence: reconciliation tests.
  • V2.16 Build decision records for pitch commission, kill, park, merge, and reassignment. Evidence: rationale and source state preserved.
    • V2.16.a Define immutable decision identity/type/actor/role/time and exact pitch/story/assignment/source/evidence revisions considered. Evidence: decision schema.
    • V2.16.b Require decision-specific rationale, public-interest/novelty/ readiness/risk/scope/conflict considerations, uncertainty, and dissent. Evidence: policy validation.
    • V2.16.c For commission, create a linked assignment without mutating the original pitch/evidence snapshot. Evidence: commission tests.
    • V2.16.d For kill/park, record retention/reopen/review conditions and preserve sources, safety restrictions, and legal holds. Evidence: disposition tests.
    • V2.16.e For merge/reassign, preserve lineage, duplicate identity, source/compartment grants, deadlines, ownership acceptance, and audit. Evidence: transfer tests.
    • V2.16.f Enforce roles/recusal/concurrency and prohibit model/agent impersonation or silent state change. Evidence: authorization tests.
  • V2.17 Build ingestion operations for feed/source schedules, crawl and social monitors, normalization, language detection, dedupe, clustering, retries, quarantine, source health, rate limits, robots/licensing policy, and operator replay. Evidence: live item-to-story-cluster lineage and failure journey.
    • V2.17.a Define feed/source/crawl/social monitor configuration, owner, credential reference, schedule, cursor, locale/territory, policy, and enabled state. Evidence: monitor contracts.
    • V2.17.b Implement due scheduling, leases, concurrency, backoff/jitter, rate/provider quotas, cancellation, missed runs, and no duplicate active run. Evidence: scheduler tests.
    • V2.17.c Enforce robots/licensing/terms/allowlist/network/redirect/ credential policy before fetch and record policy version/decision. Evidence: fetch policy tests.
    • V2.17.d Preserve raw response/item bytes and fetch provenance, then quarantine/scan/unpack/parse with isolated failures. Evidence: intake trace.
    • V2.17.e Normalize identifiers/times/authors/publishers/topics/content/ media/rights while retaining original values and transform lineage. Evidence: normalization tests.
    • V2.17.f Detect language/script with uncertainty and human correction; deduplicate exact/near/updated/syndicated items without losing sources. Evidence: language/dedupe tests.
    • V2.17.g Generate reviewed story-cluster candidates with versioned model/ features/evidence and preserve correction/retraction relationships. Evidence: clustering tests.
    • V2.17.h Expose source health/run/item/quarantine/error/dead-letter/ retry/replay state with permissioned logs and trace correlation. Evidence: operations UI.
    • V2.17.i Make replay idempotent against pinned raw inputs/config/code and prevent re-notification or auto-publication. Evidence: replay tests.
    • V2.17.j Run live item→raw evidence→normalized item→cluster lineage plus outage/rate/policy/malware/parser/poison-item recovery. Evidence: end-to-end journey.
  • V2.18 Build anonymous tip intake and investigative triage with secure submission, attachment scanning, identity/minimization choices, receipt code, threat/safety assessment, protected assignment, retention, and source conversion. Evidence: no ordinary newsroom log, notification, analytics event, or search index reveals the tipster.
    • V2.18.a Define an isolated public intake surface with current threat guidance, no required identity, locale/accessibility, rate/abuse controls, and safe emergency alternatives. Evidence: intake design.
    • V2.18.b Collect tip text and optional contact/identity/minimization/ consent/retention choices into the protected compartment only. Evidence: storage tests.
    • V2.18.c Upload attachments directly to quarantine with metadata minimization, malware/archive/media scanning, and no ordinary preview. Evidence: attachment tests.
    • V2.18.d Issue a high-entropy nonenumerable receipt code and protected status/reply channel with no existence, timing, or identity oracle. Evidence: receipt tests.
    • V2.18.e Triage credibility/urgency/coercion/retaliation/legal/physical/ digital safety with trained roles, step-up, and protected notes. Evidence: safety triage.
    • V2.18.f Create a redacted protected assignment and least-privilege grants without disclosing tip content/source beyond need to know. Evidence: assignment tests.
    • V2.18.g Convert to a source only through explicit authorized review, preserving tip lineage/choices/consent and not inventing identity. Evidence: conversion tests.
    • V2.18.h Apply retention/deletion/legal hold/emergency access and redacted tamper-evident audit. Evidence: lifecycle tests.
    • V2.18.i Inspect ordinary DB/logs/errors/traces/notifications/search/ analytics/cache/backups/support exports for tipster leakage. Evidence: non-leakage gate.
  • V2.19 Build investigative research plans for document sets, datasets, entities, relationships, timelines, public records, interviews, requests, hypotheses, verification status, and collaboration compartments. Evidence: every assertion remains anchored and export respects source protection/legal holds.
    • V2.19.a Define plan identity/scope/question/owner/team/state/milestones/ risk/budget/legal-hold and exact pitch/story revision. Evidence: plan schema.
    • V2.19.b Define document/dataset collections with immutable inputs, schemas/queries/transforms/snapshots, review status, and rights. Evidence: material contracts.
    • V2.19.c Define entity/alias/relationship/timeline items with temporal validity, uncertainty, provenance, merge/split, and protected identity. Evidence: research graph schema.
    • V2.19.d Define public-record/interview/information-request tasks with jurisdiction/contact/deadline/response/appeal/safety and secure communication references. Evidence: task contracts.
    • V2.19.e Define hypothesis/assertion/counter-hypothesis, required evidence, supporting/contradicting locators, confidence, verifier, and status without promoting notes to facts. Evidence: hypothesis schema.
    • V2.19.f Build compartmented workstreams/membership/grants/comments/ notifications/search with cross-compartment redacted handoffs. Evidence: collaboration tests.
    • V2.19.g Build plan board/timeline/notebook/evidence views with stable filters, dependencies, stale-source warnings, and complete anchor resolution. Evidence: workspace tests.
    • V2.19.h Export an authorized purpose-bound snapshot with redaction/ watermark/expiry/audit while enforcing source protection and legal holds. Evidence: export tests.
    • V2.19.i Verify every assertion resolves to pinned evidence or is explicitly unsupported/contested/unknown. Evidence: integrity gate.
  • V2.20 Build knowledge-graph entity resolution and relationship review for people, organizations, places, events, topics, claims, aliases, temporal validity, merge/split, confidence, provenance, and retractions. Evidence: graph suggestions require review and protected entities remain compartmented.
    • V2.20.a Define canonical entity/relation/alias/identifier/attribute schemas with type, temporal validity, uncertainty, provenance, tenant/ compartment scope, and version. Evidence: graph contracts.
    • V2.20.b Extract resolution/link candidates from pinned source/story/ claim revisions behind the provider seam with model/features/evidence/ confidence attribution. Evidence: candidate tests.
    • V2.20.c Present side-by-side candidate/current records and conflicting identifiers/attributes/relationships/times for human accept/reject/ defer. Evidence: resolution review.
    • V2.20.d Implement merge with surviving identity, field-level provenance, alias/edge/reference redirect, protected-grant union constraints, and reversible audit. Evidence: merge tests.
    • V2.20.e Implement split with explicit record/alias/edge/evidence/ history allocation, downstream impact preview, and no orphaned claims. Evidence: split tests.
    • V2.20.f Review/create/update/delete relationships with source locators, direction/type, validity interval, confidence, contradiction, and authorization. Evidence: relation tests.
    • V2.20.g Propagate source/claim/entity retraction or correction as versioned invalidation/review signals without deleting historical graph evidence. Evidence: retraction tests.
    • V2.20.h Exclude protected entities/identifiers/degree/count/snippets from unauthorized graph/search/API/log/cache projections. Evidence: compartment tests.
    • V2.20.i Provide a semantic table/tree/path alternative and verify keyboard/screen-reader operation for all review actions. Evidence: accessible graph E2E.

V3 — Drafting, media, claims, and authoring assist#

  • V3.1 Build the story board with guarded editorial-state transitions. Evidence: server state-machine tests.
    • V3.1.a Define board columns/views from canonical editorial state plus independent desk/owner/priority/deadline/embargo/gate/review dimensions. Evidence: board view model.
    • V3.1.b Query permission-filtered live story summaries with stable sort/ filter/group/pagination and no protected source or embargo leakage. Evidence: board query tests.
    • V3.1.c Expose only server-authorized transitions for current state/role/ revision/gates/approvals, with reason/impact/confirmation. Evidence: transition UI tests.
    • V3.1.d Enforce transitions atomically server-side with optimistic concurrency, idempotency, audit, notifications, and no drag-only path. Evidence: state-machine suite.
    • V3.1.e Handle stale cards, conflicts, partial/degraded data, revoked access, and real-time updates without implying a transition succeeded. Evidence: resilience tests.
  • V3.2 Build the story dossier with assignment, draft, sources, claims, evidence, media, gates, reviews, schedule, distribution, and history. Evidence: live-data browser journey.
    • V3.2.a Define dossier route/context/read model pinned to story and selected revision with permission-filtered section summaries/freshness. Evidence: dossier contract.
    • V3.2.b Compose primary draft editor with assignment/source/claim/ evidence/media inspectors rather than duplicating owned records. Evidence: workspace tests.
    • V3.2.c Show gates/reviews/schedule/distribution current state, blockers, owners, versions, timestamps, remediation, and eligible actions. Evidence: status tests.
    • V3.2.d Provide chronological state/revision/decision/approval/publish/ correction history with redacted protected events. Evidence: history tests.
    • V3.2.e Exercise authorized/unauthorized live database journey from assignment through current public read-back, including stale/degraded/ empty states. Evidence: browser E2E.
  • V3.3 Build a structured article editor for paragraph, heading, list, quote, fact box, data, table, image, gallery, audio, video, embed, correction, and disclosure blocks. Evidence: block registry.
    • V3.3.a Define versioned block registry with stable ID, kind-specific content, children/order, validation, serialization, migration, and semantic output contract. Evidence: registry schema.
    • V3.3.b Implement text structure blocks—paragraph/heading/list/quote/fact box—with canonical commands, paste normalization, and anchor stability. Evidence: text block tests.
    • V3.3.c Implement data/table blocks with source/dataset/query/transform/ units/uncertainty and accessible table semantics. Evidence: data block tests.
    • V3.3.d Implement image/gallery/audio/video/embed blocks via approved assets/allowlists with captions/transcripts/rights/provenance and safe fallbacks. Evidence: media block tests.
    • V3.3.e Implement correction/disclosure blocks with policy-required immutable links and target renderer behavior. Evidence: policy block tests.
    • V3.3.f Test add/edit/move/split/merge/copy-paste/delete/undo/redo/ collaboration across all kinds and reject unknown/unsafe content. Evidence: block registry suite.
  • V3.4 Build attribution-aware quote editing with speaker, source, locator, verification, context, edit notation, and consent. Evidence: quote validation.
    • V3.4.a Define quote text/language, exact source revision/locator, speaker/public label versus protected identity, and captured context. Evidence: quote contract.
    • V3.4.b Support exact/paraphrase/translated/edited quote types with omission/interpolation/clarification notation and immutable original. Evidence: edit tests.
    • V3.4.c Record verifier/method/time/result, recording/transcript/source relationship, uncertainty, and stale/unreachable state. Evidence: verification tests.
    • V3.4.d Record consent/on-record/background/off-record terms, scope, expiry/withdrawal, rights, safety restrictions, and authorized display. Evidence: consent tests.
    • V3.4.e Gate publication on kind-specific attribution/context/locator/ verification/consent rules and prevent identity leakage. Evidence: quote validation suite.
  • V3.5 Build media attachment with rights, caption, credit, alt text, transcript, content warning, location/privacy, and C2PA status. Evidence: publish gate.
    • V3.5.a Attach only approved immutable media asset/revision/rendition with source/story provenance, hash, scan state, and intended block/target. Evidence: attachment contract.
    • V3.5.b Capture rights/license/holder/territory/channel/locale/expiry/ attribution and block unsupported use or expired rights. Evidence: rights tests.
    • V3.5.c Capture factual caption/credit, language, alt text/long description, transcript/captions/audio description, and review status. Evidence: accessibility tests.
    • V3.5.d Capture graphic/trauma/child/privacy/location/doxxing warnings, redactions/crops, consent, and distribution-specific restrictions. Evidence: harm/privacy tests.
    • V3.5.e Validate C2PA valid/tampered/stripped/unknown lineage separately from editorial truth and show exact provenance status. Evidence: C2PA tests.
    • V3.5.f Compute target-specific publish eligibility from all current fields/reviews without silently dropping media or metadata. Evidence: publish gate suite.
  • V3.6 Build selection-anchored claim creation. Evidence: claim remains attached through unrelated edits.
    • V3.6.a Capture current draft revision, selected block/range/text/context, direction/affinity, author, and intended claim class. Evidence: creation contract.
    • V3.6.b Create normalized claim text separately from quoted selection while preserving exact origin anchor and requiring human confirmation. Evidence: creation tests.
    • V3.6.c Map the anchor through unrelated before/after edits, formatting, block moves, and collaboration operations. Evidence: survival tests.
    • V3.6.d Mark ambiguous/deleted/split/merged replacement as orphaned or candidate-remap requiring review, never silently retarget. Evidence: adversarial edit tests.
    • V3.6.e Preserve stable claim identity and history across revisions, undo/redo, restore, copy, and branch/merge. Evidence: revision tests.
  • V3.7 Build claim extraction behind the fenced provider seam. Evidence: proposals require human confirmation. <!-- SPEC 2026-09-11: "fenced provider seam" is the M2.6 shape: bounded authorized passages in, typed proposals out, human confirmation before anything canonical, provider/model/prompt/config version on every proposal. Bind the model per the Phase V header note (5); libs/veritas/claims is the starting point and imports @veritas/llm. -->
    • V3.7.a Define request with pinned draft blocks/revision, eligible text, locale/domain/classes, excluded protected fields, budget, and policy. Evidence: request schema.
    • V3.7.b Return candidate claim text/class/anchor/severity/confidence/ rationale and source-use without writing canonical claim records. Evidence: response tests.
    • V3.7.c Record provider/model/prompt/config/version, input/output hashes, cost/latency/safety, and run provenance. Evidence: attribution.
    • V3.7.d Present candidates for individual accept/edit/reject/duplicate handling and revalidate anchors against the current revision. Evidence: confirmation tests.
    • V3.7.e Evaluate missed high-risk claims, invented/merged claims, bad anchors, protected leakage, injection, and provider failure. Evidence: extraction eval.
  • V3.8 Build citation and counter-source attachment with exact locators. Evidence: duplicate and invalid link tests.
    • V3.8.a Search/select an authorized source revision and create/validate a kind-specific exact locator with excerpt/context/hash. Evidence: locator journey.
    • V3.8.b Classify relationship as supporting/contradicting/contextual/ irrelevant/counter-source with assessment, confidence, and reviewer. Evidence: link contract.
    • V3.8.c Show source quality/retraction/rights/freshness/protection state without leaking compartmented identity. Evidence: projection tests.
    • V3.8.d Detect exact/overlapping/semantically duplicate links and require deliberate merge/retain decisions. Evidence: duplicate tests.
    • V3.8.e Reject cross-tenant/story-policy violations, invalid/unreachable/ stale locators, wrong revisions, unsupported kinds, and unauthorized fields. Evidence: invalid link suite.
    • V3.8.f Preserve link revision/history and invalidate affected gates/ approvals when citation/source/claim changes. Evidence: lifecycle tests.
  • V3.9 Build source/claim/evidence sidebar with authority, confidence, status, contradiction, attestation, and retraction context. Evidence: screen-reader accessible.
    • V3.9.a Define sidebar view model pinned to draft/selection/claim/source revisions with permissions, freshness, loading, and partial state. Evidence: inspector contract.
    • V3.9.b Show sources/citations/evidence locators, raw quality factors, band/confidence/authority, and protected redaction. Evidence: source inspector tests.
    • V3.9.c Show claim severity/status/support/counter-sources/ contradictions/unsupported state and exact anchors. Evidence: claim inspector tests.
    • V3.9.d Show evidence-pack integrity, attestation scope/credential/ conflict/disagreement/expiry, and retraction/correction impact. Evidence: evidence inspector tests.
    • V3.9.e Support inspect/link/unlink/request review/remediate operations with server authorization and focus returning to the editor. Evidence: action tests.
    • V3.9.f Provide headings/lists/tables/status text/announcements and keyboard synchronization with the editor at zoom/reflow. Evidence: screen-reader E2E.
  • V3.10 Build headline/dek variants and body-accuracy comparison. Evidence: exact gate outputs.
    • V3.10.a Create immutable human or attributed AI headline/dek candidate revisions linked to exact body/claim/evidence revision. Evidence: variant schema.
    • V3.10.b Compare candidates side by side for factual propositions, entities/numbers/time/location, certainty, omissions, tone, length, target, and accessibility. Evidence: comparison UI.
    • V3.10.c Run versioned accuracy/sensationalism/clickbait/SEO/format gates and display raw findings/evidence separately from verdicts. Evidence: gate output tests.
    • V3.10.d Allow edit/select/reject only by authorized humans with rationale and no model/score auto-selection. Evidence: decision tests.
    • V3.10.e Recompute/invalidate comparison and approval when body/claims/ sources/candidate/policy changes. Evidence: freshness tests.
  • V3.11 Build authoring assist for structure, clarity, headline, summary, SEO, localization, and format adaptation. Evidence: per-change accept and AI attribution.
    • V3.11.a Define separate assist operations with pinned input revision/ selection, constraints, eligible sources, target/locale/style, provider policy, and budgets. Evidence: operation contracts.
    • V3.11.b Generate structure/clarity suggestions as minimal anchored diffs that preserve claims/quotes/citations/meaning. Evidence: editing tests.
    • V3.11.c Generate headline/summary/SEO suggestions with factual proposition comparison and no keyword-driven certainty or sensational drift. Evidence: accuracy tests.
    • V3.11.d Generate localization/format adaptations linked to exact source text, terminology, claims/evidence, disclosure, rights, and target constraints. Evidence: adaptation tests.
    • V3.11.e Present each change for accept/edit/reject with preview, rationale/source use, conflicts, and stale-revision rebase. Evidence: interaction tests.
    • V3.11.f Record model/provider/prompt/config/version/input/output/cost and accepted human actor per change; never bulk-apply silently. Evidence: attribution audit.
  • V3.12 Prohibit AI from silently inventing sources, quotes, facts, citations, or named attribution. Evidence: adversarial generation eval.
    • V3.12.a Enforce typed outputs that distinguish verbatim source content, grounded inference, proposal, placeholder, uncertainty, and refusal. Evidence: response schema tests.
    • V3.12.b Allow source/quote/citation/attribution references only to authorized provided stable IDs and validate locators/text exactness. Evidence: grounding validator.
    • V3.12.c Reject or visibly flag invented/altered/unresolvable references, unsupported facts, merged speakers, and implied named attribution. Evidence: enforcement tests.
    • V3.12.d Require human confirmation before canonical claim/citation/quote/ source/attribution creation and retain generator provenance. Evidence: mutation boundary tests.
    • V3.12.e Evaluate prompt/source injection, plausible fake citations, source laundering, paraphrase drift, multilingual names, and multi-turn persistence. Evidence: adversarial eval.
    • V3.12.f Define stop-the-line thresholds, regression ownership, provider/ model rollback, and approved residual exceptions. Evidence: safety gate.
  • V3.13 Build word/character, reading time, media, claim density, citation, and freshness meters from server logic. Evidence: known-value tests.
    • V3.13.a Define canonical eligible content, Unicode/grapheme/locale rules, exclusions, units, rounding, and versions for word/character/reading time. Evidence: metric contracts.
    • V3.13.b Define media counts/durations/types and claim/citation density numerators, denominators, duplicate/overlap, and zero-content behavior. Evidence: density contracts.
    • V3.13.c Define source/citation/claim/evidence/policy freshness inputs, clocks, thresholds, unknown/unavailable state, and invalidation. Evidence: freshness contract.
    • V3.13.d Compute authoritatively on exact server revision and return raw values, units, timestamp, calculator/policy version, and evidence. Evidence: API tests.
    • V3.13.e Show advisory meters without converting them into automatic quality/editorial scores or overriding gates. Evidence: UI tests.
    • V3.13.f Verify hand-calculated Unicode/locale/block/media/anchor/stale/ missing fixtures. Evidence: known-value suite.
  • V3.14 Build preview for reader web, mobile, newsletter, wire/package, audio, video, and social targets. Evidence: renderer parity.
    • V3.14.a Define target/version/locale/device/audience preview request pinned to exact draft/media/policy/template/config revisions. Evidence: preview contract.
    • V3.14.b Render reader web/mobile and newsletter using production-shared renderers with navigation, disclosures, corrections, accessibility, and responsive behavior. Evidence: reader/email previews.
    • V3.14.c Render NewsML/wire/package structured payload and validation diagnostics without hiding unsupported/dropped fields. Evidence: package preview.
    • V3.14.d Render audio/video timeline/transcript/captions/chapters/media/ provenance and platform aspect/safe-area variants. Evidence: AV preview.
    • V3.14.e Render social/channel copy/media/link/cards/disclosures/ corrections/limits per current adapter capability. Evidence: target previews.
    • V3.14.f Compare preview semantic output to published/read-back output, label fixtures/staleness, and block unsupported target claims. Evidence: renderer parity suite.
  • V3.15 Build autosave, undo/redo, revision history, semantic diff, restore, presence, and conflict resolution. Evidence: concurrent edit Playwright.
    • V3.15.a Define local operation/draft/server revision/checkpoint/presence/ conflict schemas and eligible protected/offline storage policy. Evidence: collaboration contract.
    • V3.15.b Implement debounced idempotent autosave with expected revision, dirty/saving/saved/error/offline/conflict state and crash recovery. Evidence: autosave tests.
    • V3.15.c Implement operation-aware undo/redo scoped to the actor/session without reverting collaborators or immutable published history. Evidence: history tests.
    • V3.15.d Provide immutable revision timeline and semantic block/text/ claim/citation/media/disclosure diff with author/cause/provenance. Evidence: diff tests.
    • V3.15.e Restore as a new revision with preview, permissions, invalidated gates/approvals/schedules, and no history rewrite. Evidence: restore tests.
    • V3.15.f Show privacy-filtered presence/selection and resolve concurrent block/text/anchor/source/claim edits using mine/theirs/base with copies. Evidence: conflict tests.
    • V3.15.g Run two-client web/desktop/offline/delete/reorder/role-revoke/ reconnect journeys. Evidence: concurrent Playwright.
  • V3.16 Build disclosure editor for AI assistance, methodology, corrections, conflicts, sponsorship, and limitations. Evidence: target-specific required-field gates.
    • V3.16.a Define typed disclosure records with scope, exact story/block/ media/release/locale/target revision, author, status, and history. Evidence: disclosure schema.
    • V3.16.b Build AI-use and methodology disclosures from attributed provenance/methods while allowing accurate human-authored explanation. Evidence: disclosure tests.
    • V3.16.c Build correction/history and conflicts/recusal disclosures linked to canonical decisions without exposing protected/legal details. Evidence: editorial disclosure tests.
    • V3.16.d Build sponsorship/advertising/partnership and limitations/ uncertainty disclosures separated from editorial content and controls. Evidence: independence tests.
    • V3.16.e Define required types/fields/placement/wording/reviewer by desk/ risk/audience/locale/target and recompute on changes. Evidence: policy matrix.
    • V3.16.f Preview/read-back exact reader-visible disclosures and block release if required content is missing, stale, or silently dropped. Evidence: target gate tests.
  • V3.17 Build journalism, research, fact-checking, standards, copy-edit, format, and distribution agent plans with least-privilege tool grants. Evidence: story revision, allowed sources, policy, budget, stop conditions, and approval points are locked before dispatch.
    • V3.17.a Define distinct plan templates and allowed outcomes for journalism/research/fact-check/standards/copy-edit/format/distribution; prohibit unsupported role blending. Evidence: template registry.
    • V3.17.b Pin tenant/story/draft/source/evidence/policy/locale/target revisions, objective, constraints, expected artifacts, and acceptance criteria. Evidence: immutable plan inputs.
    • V3.17.c Resolve allowed sources/fields and least-privilege read/search/ analyze/draft/propose/job tools with object/function scopes and expiry. Evidence: grant manifest.
    • V3.17.d Lock token/cost/time/step/tool/concurrency/provider budgets, retry/fallback, data-use rules, and protected-source exclusions. Evidence: budget/policy manifest.
    • V3.17.e Define checkpoint/human approval/attestation-required/stop-the- line/cancellation/escalation conditions before dispatch. Evidence: control contract.
    • V3.17.f Preview exact plan/grants/risks to an authorized human, record approval, sign/hash it, and reject stale or mutated dispatch. Evidence: dispatch tests.
  • V3.18 Build Veritas-scoped agent run, queue, plan, tool-call, budget, source-use, output, and provenance views. Evidence: protected-source redaction and live trace.
    • V3.18.a Define run/step/tool-call/queue/event/budget/source-use/output/ artifact/provenance state contracts and retention. Evidence: run schema.
    • V3.18.b Build queue/run list with owner/story/plan/status/progress/ budget/priority/wait reason/failure/created/updated and stable filters. Evidence: operations view.
    • V3.18.c Build live run timeline showing immutable plan, step transitions, tool name/scoped purpose/input-output summaries, retries, decisions, and human interventions. Evidence: trace view.
    • V3.18.d Show token/cost/time/tool/step limits versus actual/forecast and stop/cancel controls with truthful delayed accounting. Evidence: budget view tests.
    • V3.18.e Show exact allowed/used source/evidence IDs, locators, revision, result attribution, and derived artifact lineage. Evidence: provenance view.
    • V3.18.f Apply compartment/field redaction to UI/API/events/logs/export and prevent counts/timing/tool summaries from revealing protected sources. Evidence: non-leakage tests.
    • V3.18.g Test live update reconnect/order/deduplication/stale/partial/ unavailable and terminal-state reconciliation. Evidence: live trace E2E.
  • V3.19 Intake agent output only as attributed suggestions, proposed claims, or draft revisions. Evidence: agents cannot create source identity, attest facts, approve legal/standards gates, publish, correct, or retract.
    • V3.19.a Define an allowlisted output envelope for suggestion, proposed claim/citation, anchored edit set, or complete draft revision with run/ plan/model/tool/source provenance. Evidence: output schema.
    • V3.19.b Validate output type/schema/size/anchors/revisions/source IDs/ locators/grounding/rights/safety and quarantine malformed content. Evidence: intake validation.
    • V3.19.c Store output in a proposal namespace with explicit unreviewed state, no canonical source/attestation/approval/release semantics, and safe rendering. Evidence: persistence tests.
    • V3.19.d Present semantic diff/evidence/source use/uncertainty for per-item authorized human accept/edit/reject; rebase or invalidate stale outputs. Evidence: review tests.
    • V3.19.e Apply accepted content as a human-attributed new draft/claim revision while retaining AI contribution attribution and re-running gates. Evidence: acceptance tests.
    • V3.19.f Attempt source identity creation, attestation, legal/standards approval, publish/correct/retract, privilege escalation, and direct API mutation; prove rejection/no side effect. Evidence: capability tests.
  • V3.20 Implement agent pause, redirect, cancel, budget kill, and replay. Evidence: partial results retain source/run provenance and cannot appear approved.
    • V3.20.a Define legal control transitions by run/step/job state, actor, authority, idempotency, and terminal/commit boundary. Evidence: control state machine.
    • V3.20.b Pause at safe checkpoints, freeze new tools, retain leases/ partial state, and resume only after current plan/grant validation. Evidence: pause tests.
    • V3.20.c Redirect via a reviewed plan amendment/new plan with changed objective/constraints/grants/budgets, preserving the original. Evidence: redirect tests.
    • V3.20.d Cancel queued/running/tool/job phases with provider cancellation, compensation, timeout, and honest requested-versus-completed state. Evidence: cancellation tests.
    • V3.20.e Enforce automatic/manual budget kill at every metering boundary with bounded overshoot and no new unmetered actions. Evidence: kill tests.
    • V3.20.f Replay from pinned plan/inputs/tool versions or explicitly mark substitutions/nonreproducibility; deduplicate committed effects. Evidence: replay tests.
    • V3.20.g Preserve all partial suggestions/artifacts/source use/run lineage as incomplete/unapproved and require fresh human review. Evidence: status tests.
  • V3.21 Evaluate multi-agent handoffs for fabricated citation, source laundering, instruction loss, privilege expansion, and correlated error. Evidence: adversarial corpus and stop-the-line thresholds.
    • V3.21.a Define typed handoff envelope with parent/child run, pinned constraints/policy/grants/budget, source/evidence IDs, claims, artifacts, uncertainty, and unresolved issues. Evidence: handoff schema.
    • V3.21.b Verify receivers cannot reinterpret unavailable/protected/ unverified sources as ordinary evidence or drop source-use provenance. Evidence: laundering tests.
    • V3.21.c Test altered/invented citations/quotes/claims across extraction, research, fact-check, editing, localization, and distribution hops. Evidence: fabrication corpus.
    • V3.21.d Test loss/contradiction of negative instructions, disclosure/ embargo/rights/locale constraints, approval requirements, and stop conditions. Evidence: instruction-retention tests.
    • V3.21.e Test delegated tool/role/field/object/tenant/compartment privilege expansion and confused-deputy attacks. Evidence: capability tests.
    • V3.21.f Measure correlated error/amplification, independence, calibration, detection/abstention, and subgroup/language performance over chained runs. Evidence: evaluation report.
    • V3.21.g Set per-severity stop thresholds, block dispatch/intake/release on breach, assign mitigation/model rollback, and require human approval. Evidence: stop-the-line gate.
  • V3.22 Build audio/podcast production for script/rundown, host/voice, recording, clips, music/SFX, edit, mix, loudness, transcript, chapters, captions, translations, rights, provenance, review, and feed/package outputs. Evidence: story claims and corrections stay linked to timecoded audio.
    • V3.22.a Define audio project/episode/script/rundown/segment/track/clip/ marker/version schemas linked to exact story/claim/evidence revisions. Evidence: audio contracts.
    • V3.22.b Build script/rundown editing with host/guest/voice assignments, pronunciation, fact-check holds, timing, disclosures, and source timecode/citation links. Evidence: preproduction tests.
    • V3.22.c Ingest or capture recording with device/session/files/checksums/ consent, original immutable audio, proxies, and privacy-safe metadata. Evidence: recording lineage.
    • V3.22.d Build clip/select/trim/sequence/fade/music/SFX editing using nondestructive operations, licensed assets, and semantic/non-drag controls. Evidence: edit tests.
    • V3.22.e Build mix/render with track/voice/music/SFX balances, channel/ codec/rate, target loudness/true peak, deterministic job provenance, and quality measurements. Evidence: mix tests.
    • V3.22.f Generate/edit/align transcripts, speakers, chapters, captions, descriptions, and translations with human review and locale parity. Evidence: accessibility/localization tests.
    • V3.22.g Enforce voice/music/recording/source/territory/channel/expiry rights, AI/synthetic voice disclosure, consent, and C2PA/provenance. Evidence: rights/provenance gate.
    • V3.22.h Build editorial/fact-check/legal/accessibility/audio review on exact version with timecoded comments and stale invalidation. Evidence: review journey.
    • V3.22.i Generate/validate feed/enclosure/NewsML/CMS/article/package outputs with renditions, metadata, chapters, transcripts, and read-back. Evidence: distribution tests.
    • V3.22.j Preserve claim/citation/correction/retraction links to timecoded ranges through edits/renders/replacements and cascade every target. Evidence: correction lineage E2E.
  • V3.23 Build video and short-form production for script, shot list, source media, edit/timeline, graphics, captions, voice, aspect renditions, thumbnails, platform-safe areas, rights, provenance, review, and correction replacement. Evidence: YouTube/Shorts/TikTok and article embeds derive from locked story evidence.
    • V3.23.a Define video project/version/sequence/track/clip/shot/graphic/ caption/marker/rendition schemas linked to locked story evidence. Evidence: video contracts.
    • V3.23.b Build script/shot-list/storyboard planning with claim/source/ locator, capture/acquisition, contributor, safety, rights, and target requirements. Evidence: production plan.
    • V3.23.c Ingest source media as immutable scanned assets with checksums/ capture metadata/C2PA/proxies/transcripts and privacy review. Evidence: media lineage.
    • V3.23.d Build nondestructive edit/timeline for clips/audio/voice/music/ graphics/lower thirds with semantic/non-drag controls and source links. Evidence: timeline tests.
    • V3.23.e Build factual graphics, names/numbers/maps/charts, captions/ translations/audio descriptions, and human verification against evidence. Evidence: overlay/accessibility tests.
    • V3.23.f Generate landscape/vertical/square renditions and thumbnails with crop/reframe/safe-area/text-length/codec/bitrate constraints and no meaning loss. Evidence: rendition tests.
    • V3.23.g Enforce media/music/voice/likeness/location/platform/territory/ expiry rights, safety warnings, disclosure, and provenance. Evidence: publish gates.
    • V3.23.h Run editorial/fact-check/legal/standards/accessibility review on exact timeline/renditions with timecoded comments. Evidence: review journey.
    • V3.23.i Publish/read back YouTube/Shorts/TikTok and article embeds through target adapters with partial failure, acknowledgement, and no silent auto-publish. Evidence: channel tests.
    • V3.23.j Replace/correct/retract exact ranges/renditions/thumbnails/ captions across targets while retaining prior lineage and notices. Evidence: correction E2E.
  • V3.24 Build live-program rundown and stream production for segments, guests, sources, lower thirds, fact-check hold, delay, captions, translations, moderator cues, Q&A, emergency stop, recording, clipping, archive, and post-live correction. Evidence: live urgency never becomes a generic editorial-gate bypass.
    • V3.24.a Define live program/rundown/segment/cue/guest/source/claim/ destination/session/version/state contracts and accountable operator roles. Evidence: live schemas.
    • V3.24.b Build ordered/timed rundown with script, guests/consent, sources/evidence, assets, lower thirds, transitions, moderator cues, contingencies, and approvals. Evidence: rundown tests.
    • V3.24.c Validate names/titles/numbers/claims/media/rights/disclosures against locked story evidence and route unresolved items to fact-check hold. Evidence: preflight gate.
    • V3.24.d Build ingest/encoder/stream/destination health, preview/program, delay buffer, captions/translations, recording, and failover controls. Evidence: stream operations tests.
    • V3.24.e Build moderated Q&A with question intake/screening/source status/ identity safety, answer evidence, block/remove/escalate, and audit. Evidence: Q&A tests.
    • V3.24.f Enforce role/step-up/two-step confirmation for go-live, fact- check release, delay dump, destination changes, emergency stop, and restart. Evidence: control authorization.
    • V3.24.g Record synchronized program/clean feeds, captions/translations, operator/decision/incident timeline, and immutable provenance. Evidence: session record.
    • V3.24.h Create reviewed clips/archive/article embeds with timecode/ claim/source linkage, rights, accessibility, and target approvals. Evidence: post-live workflow.
    • V3.24.i Author and propagate post-live correction/retraction to recordings/clips/captions/transcripts/archive/embeds/channels with notices and read-back. Evidence: correction E2E.
    • V3.24.j Prove urgency cannot bypass protected-source, legal/high-risk, evidence, disclosure, accessibility, or human publish floors; test emergency authority separately. Evidence: gate suite.
  • V3.25 Build data/visual journalism authoring for datasets, queries, transformations, charts/maps/tables, methodology, uncertainty, source snapshots, accessible alternatives, responsive embeds, and update detection. Evidence: published visualization is reproducible from locked data/code and has semantic parity.
    • V3.25.a Define data project/dataset snapshot/schema/query/transform/ visualization/view/release contracts linked to story/claims/sources. Evidence: data-viz schemas.
    • V3.25.b Ingest immutable licensed datasets with checksum, acquisition/ refresh metadata, schema/profile, privacy/classification, and quarantine. Evidence: dataset lineage.
    • V3.25.c Build parameterized query/notebook/transform pipeline with pinned engine/dependencies/code/config/random seeds, review, and reproducible outputs. Evidence: compute tests.
    • V3.25.d Build chart/map/table specifications with units/denominators/ baselines/scales/geography/time, annotation, provenance, and no misleading defaults. Evidence: visual contract tests.
    • V3.25.e Author methodology/limitations/uncertainty/missing-data/ statistical notes tied to exact data/code/output revisions. Evidence: methodology review.
    • V3.25.f Provide equivalent accessible table/text/key findings, data download where permitted, keyboard navigation, contrast/pattern, and noncolor encodings. Evidence: semantic parity tests.
    • V3.25.g Generate responsive embeds/static/social/print renditions with stable canonical URL, safe sizing, fallback, performance, and target metadata. Evidence: renderer tests.
    • V3.25.h Detect upstream data/schema/code/dependency changes, compare outputs, mark stale, and require human review before update. Evidence: update detection.
    • V3.25.i Publish/read back locked data/code/spec/output/provenance and verify independent reproduction within declared tolerances. Evidence: reproducibility gate.
    • V3.25.j Propagate corrections/retractions through claims, annotations, tables, embeds, images, downloads, caches, and archives. Evidence: correction journey.

V4 — Verification, balance, attestation, and editorial gates#

  • V4.1 Implement canonical source-quality scoring with published factor weights and adjustments. Evidence: known-value suite per claim domain.
    • V4.1.a Implement versioned factor measurements from pinned source evidence with explicit missing/unknown/unavailable states. Evidence: factor unit tests.
    • V4.1.b Resolve applicable claim-domain weight profile and validate factor/weight ranges, completeness, normalization, and effective date. Evidence: profile tests.
    • V4.1.c Compute canonical composite/intermediate values/rounding/band/ confidence deterministically and separately from manual adjustments. Evidence: algorithm tests.
    • V4.1.d Apply only authorized bounded reasoned adjustments with evidence, actor, expiry, conflict, and immutable raw score. Evidence: adjustment tests.
    • V4.1.e Persist all input/evaluator/weight/policy versions and recompute or mark stale on any change. Evidence: provenance tests.
    • V4.1.f Compare independently calculated boundary/missing/adversarial fixtures for every ratified claim domain. Evidence: known-value suite.
  • V4.2 Implement claim extraction quality evaluation against a labeled corpus. Evidence: precision/recall and missed-high-risk analysis. <!-- SPEC 2026-09-11: §2.5.5 — no labeled representative corpus exists; .a builds a synthetic one with ground truth fixed by the fixture and says so, .c runs on the bound cheap model, .f's named review stays human. -->
    • V4.2.a Build versioned representative corpus with exact draft blocks/ ranges, claim classes, severities, locales/desks, negatives, and expert labels. Evidence: corpus manifest.
    • V4.2.b Define matching rules for span overlap, normalized proposition, duplicate/merged/split claims, class, severity, and anchor correctness. Evidence: scoring spec.
    • V4.2.c Run pinned model/provider/prompt/config against held-out corpus with deterministic request/output provenance. Evidence: evaluation run.
    • V4.2.d Report precision/recall/F-score, calibration, duplicate/bad-anchor rate, latency/cost, and subgroup/locale/desk slices. Evidence: metrics.
    • V4.2.e Review every missed high-risk and dangerous false-positive claim, classify causes/severity, and assign mitigation. Evidence: error analysis.
    • V4.2.f Enforce thresholds/regression comparisons/stop-the-line and named human approval before deployment. Evidence: evaluation gate.
  • V4.3 Implement claim-citation support and locator verification. Evidence: supporting, irrelevant, contradicted, and unreachable source fixtures.
    • V4.3.a Resolve exact authorized source revision/locator/content/hash and classify unreachable/changed/ambiguous/protected state. Evidence: locator tests.
    • V4.3.b Measure proposition-level entailment/support/contradiction/ relevance/context and coverage without treating source quality as support. Evidence: measurement tests.
    • V4.3.c Combine multiple citations with independence/duplication/ conflicts and explicit insufficient/contested/unknown state. Evidence: multi-source tests.
    • V4.3.d Persist claim/source/locator/evaluator/model/prompt/corpus/policy versions, raw findings, evidence excerpts, uncertainty, and freshness. Evidence: result provenance.
    • V4.3.e Apply versioned verdict thresholds separately and invalidate on claim/source/locator/policy change or retraction. Evidence: verdict tests.
    • V4.3.f Verify supporting/partial/irrelevant/contradicted/unreachable/ injection/multilingual fixtures with expert expected results. Evidence: known-value suite.
  • V4.4 Implement confidence-band rules, including source-count and counter-source invariants. Evidence: canonical contract fixtures.
    • V4.4.a Define versioned inputs for source quality/independence/count, citation support, counter-sources, contradiction, attestation, and uncertainty. Evidence: confidence contract.
    • V4.4.b Implement band thresholds and minimum independent-source/ specialist/evidence requirements by claim class/severity/domain. Evidence: rule tests.
    • V4.4.c Enforce counter-source and contested/unsupported/stale/unavailable rules that cap, block, or require review; never count duplicates as independence. Evidence: invariant tests.
    • V4.4.d Persist raw inputs/calculation separately from policy verdict and recompute when any dependency/policy changes. Evidence: recomputation tests.
    • V4.4.e Test exact boundaries, missing data, many low-quality dependent sources, one high-quality source, contradictions, and evaluator failure. Evidence: canonical fixtures.
  • V4.5 Implement statistical-claim checks for denominator, population, period, method, uncertainty, and source dataset. Evidence: crafted corpus.
    • V4.5.a Parse or author-review statistic value/unit/direction/comparison, numerator/denominator, population/sample, geography, period, and claim anchor. Evidence: statistic schema/tests.
    • V4.5.b Resolve exact dataset snapshot/table/query/row/column/cell and verify value/denominator/population/period match. Evidence: data locator tests.
    • V4.5.c Check sampling/design/method/weighting/adjustment/base rate, uncertainty/interval/significance, missingness, and limitations. Evidence: method checks.
    • V4.5.d Check chart/text/headline transformations, relative versus absolute change, percentage points, causal language, rounding, and cherry-picked windows. Evidence: presentation checks.
    • V4.5.e Return findings/evidence/uncertainty separately from domain- versioned risk verdict and specialist requirements. Evidence: result contract.
    • V4.5.f Validate correct/error/ambiguous/adversarial and subgroup- sensitive crafted fixtures with expert review. Evidence: corpus report.
  • V4.6 Implement quote checks for source, speaker, context, edit, locator, and verification. Evidence: altered-context fixtures.
    • V4.6.a Resolve exact source revision/locator and compare quote/paraphrase/ translation text to the immutable excerpt/recording/transcript. Evidence: text-match tests.
    • V4.6.b Verify speaker identity/public label/protected projection and distinguish reporter attribution from source assertion. Evidence: speaker tests.
    • V4.6.c Check surrounding context, question/answer, temporal setting, selective omission, sentence joins, and meaning-changing truncation. Evidence: context tests.
    • V4.6.d Validate edit/ellipsis/interpolation/translation notation, verification status, consent/terms, and rights. Evidence: policy checks.
    • V4.6.e Persist raw match/context findings and versions separately from editorial verdict; mark stale on source/draft change. Evidence: provenance tests.
    • V4.6.f Run exact/edited/paraphrased/translated/misattributed/altered- context/unreachable fixtures. Evidence: quote corpus.
  • V4.7 Implement legal/scientific/high-risk severity and specialist-review floors. Evidence: no general source policy discharges specialist gates.
    • V4.7.a Define versioned legal/scientific/public-health/election/safety/ financial/other high-risk classes and severity indicators. Evidence: risk taxonomy.
    • V4.7.b Classify applicable risk from exact claims/sources/media/context with evidence/uncertainty and authorized human correction. Evidence: classification tests.
    • V4.7.c Map class/severity/jurisdiction/target to named credential/domain/ independence/quorum/recusal/expiry specialist floors. Evidence: review matrix.
    • V4.7.d Enforce specialist review as an independent blocking gate that source score, general fact-check, AI result, waiver, or urgency cannot satisfy. Evidence: bypass tests.
    • V4.7.e Invalidate review on relevant claim/evidence/media/policy/ jurisdiction/target change and retain immutable decisions. Evidence: freshness tests.
  • V4.8 Implement contradiction detection against current and archived claims. Evidence: known contradiction fixtures and uncertainty.
    • V4.8.a Build permission-aware candidate retrieval over current/archived claims by entities, normalized proposition, time, location, units, and topic. Evidence: retrieval tests.
    • V4.8.b Compare exact claim revisions for contradiction/qualification/ update/different scope/uncertain relation with evidence and confidence. Evidence: relation evaluator.
    • V4.8.c Account for temporal change, definitions, population, units, negation, modality, attribution, and correction/retraction state. Evidence: semantic fixtures.
    • V4.8.d Attribute model/rules/corpus/index/source versions and keep candidate measurements separate from reviewed relationships. Evidence: provenance tests.
    • V4.8.e Require human confirm/reject/defer and prevent protected/embargoed archived claim leakage through results or counts. Evidence: review/auth tests.
    • V4.8.f Evaluate known contradiction/noncontradiction/ambiguous/changed- world/multilingual fixtures and calibration. Evidence: corpus report.
  • V4.9 Implement counterclaim surfacing through the canonical policy. Evidence: surface, suppress-false-balance, require-more-evidence, and editorial-review outcomes.
    • V4.9.a Gather exact claim/counterclaim/evidence/quality/independence/ consensus/harm/context measurements with policy version. Evidence: input contract.
    • V4.9.b Compute all four canonical outcomes without reducing them to a binary “both sides” or popularity rule. Evidence: decision tests.
    • V4.9.c Return reason codes, contributing measurements, uncertainty, required evidence/reviewer, and safe presentation constraints. Evidence: result contract.
    • V4.9.d Require authorized human disposition for editorial-review and bounded override with rationale/evidence/conflict/expiry. Evidence: override tests.
    • V4.9.e Recompute on claim/counter-source/source-quality/consensus/policy change and show contested/stale/evaluator-failure explicitly. Evidence: lifecycle tests.
    • V4.9.f Validate each outcome across science/legal/election/identity/ opinion and adversarial false-balance fixtures. Evidence: canonical suite.
  • V4.10 Implement bias/loaded-language assist with uncertainty and human review. Evidence: calibrated corpus and subgroup report.
    • V4.10.a Define limited assist scope, taxonomy, context-sensitive labels, locale/desk/genre applicability, and prohibited ideology/author scoring. Evidence: assist policy.
    • V4.10.b Return anchored language findings with rationale, alternatives, uncertainty, model/prompt/corpus versions, and no silent edit. Evidence: response tests.
    • V4.10.c Distinguish loaded wording from accurate quotation, legal/ scientific terminology, reclaimed terms, direct attribution, and necessary specificity. Evidence: context fixtures.
    • V4.10.d Present accept/edit/reject/false-positive feedback to authorized humans with per-change provenance and stale-anchor handling. Evidence: review tests.
    • V4.10.e Measure calibration/error severity and subgroup/dialect/locale/ identity/desk differences with expert review. Evidence: evaluation report.
    • V4.10.f Block use as author/reporter/source/subject reputation or publish score and define regression/rollback thresholds. Evidence: misuse tests.
  • V4.11 Implement headline/dek accuracy versus body and evidence. Evidence: omission, exaggeration, certainty, and clickbait fixtures.
    • V4.11.a Extract or compare headline/dek propositions, entities, numbers, time/location, causal/attribution/modality/certainty against exact body and verified claims. Evidence: comparison tests.
    • V4.11.b Detect unsupported additions, material omissions, scope drift, exaggeration, certainty inflation, contradiction, and misleading question framing. Evidence: finding tests.
    • V4.11.c Evaluate sensational/clickbait cues separately from factual accuracy and respect genre/locale/desk context. Evidence: style tests.
    • V4.11.d Persist anchored findings/evidence/model/rule/policy versions, uncertainty, and raw measurements separately from verdict. Evidence: provenance.
    • V4.11.e Require edit/review according to risk, invalidate on body/claim/ evidence/headline changes, and prohibit automatic selection. Evidence: gate tests.
    • V4.11.f Run omission/exaggeration/certainty/clickbait/accurate-strong/ multilingual/adversarial fixtures. Evidence: known-value suite.
  • V4.12 Implement plagiarism/similarity and attribution checks. Evidence: quoted, licensed, common-phrase, and copied fixtures.
    • V4.12.a Define permissioned comparison corpus, retention/license/ protected-source exclusions, normalization, and versioned index. Evidence: corpus policy.
    • V4.12.b Detect exact/near/paraphrase/structural overlap with anchored matching passages, source candidates, score/uncertainty, and provenance. Evidence: similarity tests.
    • V4.12.c Classify quotation with valid citation, licensed/reused copy, wire/syndication, press/common phrases, self-reuse, and unattributed copy separately. Evidence: classification tests.
    • V4.12.d Check exact attribution/quotation/rights/transform requirements for source kind and target. Evidence: attribution gate.
    • V4.12.e Require human review of findings, false positives, contested ownership, and protected comparisons; never auto-accuse. Evidence: review tests.
    • V4.12.f Evaluate quoted/licensed/common/copied/paraphrased/multilingual/ adversarial fixtures and corpus leakage. Evidence: evaluation report.
  • V4.13 Implement rights, media, accessibility, disclosure, and localization gates. Evidence: gate registry coverage.
    • V4.13.a Implement rights/license/consent/territory/channel/locale/expiry/ attribution checks per source/media/target. Evidence: rights gate tests.
    • V4.13.b Implement media scan/integrity/provenance/caption/credit/privacy/ harm/rendition/target eligibility checks. Evidence: media gate tests.
    • V4.13.c Implement structure/semantics/alt text/transcript/captions/audio description/contrast/reflow/keyboard checks with manual-review floors. Evidence: accessibility gate tests.
    • V4.13.d Implement required AI/methodology/correction/conflict/ sponsorship/limitation type/content/placement/reviewer checks. Evidence: disclosure tests.
    • V4.13.e Implement locale completeness/terminology/name/number/date/ claim-evidence parity/layout/media-text/cultural-review checks. Evidence: localization tests.
    • V4.13.f Register applicability/input/output/remediation/owner/version/ failure state for every gate and prove each retained target is covered. Evidence: gate registry report.
  • V4.14 Implement attestation validity, credential, conflict, disagreement, expiry, and independence rules. Evidence: canonical tests.
    • V4.14.a Resolve exact attestation scope and reject mismatched or stale claim/source/evidence-pack/draft revisions. Evidence: scope tests.
    • V4.14.b Verify attestor identity/credential/issuer/domain/jurisdiction/ validity/revocation with pinned evidence and verification method. Evidence: credential tests.
    • V4.14.c Evaluate declared/detected conflicts, prior involvement, organizational relationships, compensation, recusal, and required independence. Evidence: conflict tests.
    • V4.14.d Aggregate agree/disagree/qualified/abstain/unable responses under quorum/specialist/independence rules without hiding disagreement. Evidence: decision tests.
    • V4.14.e Enforce expiry and invalidation on scope/evidence/credential/ policy/conflict changes; prohibit automation impersonation. Evidence: lifecycle tests.
    • V4.14.f Verify canonical valid/invalid/conflicted/disagreeing/expired/ revoked/insufficient fixtures. Evidence: invariant suite.
  • V4.15 Implement evidence-pack completeness and integrity validation. Evidence: missing/member/hash mismatch fixtures.
    • V4.15.a Resolve exact pack version and required members by claim class/ severity/desk/target/policy. Evidence: completeness rules.
    • V4.15.b Validate every member identity/version/tenant/story/compartment/ permission/reference and required citation locator. Evidence: referential tests.
    • V4.15.c Recompute canonical serialization/member/root hashes and signatures, detecting reorder/replace/duplicate/missing/tampered data. Evidence: integrity tests.
    • V4.15.d Check source/claim/attestation/review freshness, retraction/ correction state, and applicable quality/specialist floors. Evidence: semantic validation.
    • V4.15.e Return separate missing/unauthorized/stale/unreachable/hash- mismatch/unsupported-version results and block unverifiable packs. Evidence: failure-state tests.
    • V4.15.f Run missing/member/hash/cross-tenant/protected/expired/reordered fixture corpus. Evidence: canonical suite.
  • V4.16 Persist raw measurements and evidence separately from policy verdicts. Evidence: policy recomputation.
    • V4.16.a Define immutable measurement record with subject/input revisions, metric values/units, evidence locators, uncertainty, evaluator/model/ corpus version, and time. Evidence: measurement schema.
    • V4.16.b Define separate verdict record with measurement references, applicable policy/rules/thresholds, outcome/reason/remediation, and time. Evidence: verdict schema.
    • V4.16.c Prevent verdict fields from overwriting/denormalizing raw measurements and prevent missing/error measurements from defaulting to pass. Evidence: storage invariants.
    • V4.16.d Recompute new verdicts from retained measurements after policy/ threshold changes without rerunning evaluators when inputs remain valid. Evidence: recomputation tests.
    • V4.16.e Rerun measurements only on input/evaluator invalidation and retain prior measurement/verdict lineage for audit. Evidence: lifecycle tests.
    • V4.16.f Compare old/new policy decisions and explain deltas to reviewers without rewriting historical approvals/releases. Evidence: policy change report.
  • V4.17 Version every evaluator, model, domain-weight table, threshold, and corpus. Evidence: result provenance.
    • V4.17.a Define immutable identities/digests for rules/code/config, provider/model/prompt/tool, weight tables, thresholds/policies, and corpora/splits/labels. Evidence: version contracts.
    • V4.17.b Resolve and pin every dependency before evaluation; prohibit mutable aliases as sole provenance. Evidence: dispatch tests.
    • V4.17.c Persist complete dependency graph plus runtime/build/environment/ locale/randomness/input/output hashes on every result. Evidence: provenance tests.
    • V4.17.d Define approval/effective/retired/blocked states, compatibility, rollback, retention, and artifact availability for versions. Evidence: lifecycle policy.
    • V4.17.e Detect missing/unknown/revoked/unavailable versions and make results unverifiable/stale rather than current. Evidence: failure tests.
    • V4.17.f Reproduce representative results from provenance and report any nondeterminism/substitution explicitly. Evidence: reproducibility gate.
  • V4.18 Make unavailable, insufficient-evidence, contested, stale, and evaluator-failure states explicit. Evidence: none renders as pass.
    • V4.18.a Define canonical distinct measurement/execution/verdict states, reason codes, timestamps, owners, safe messages, and retry/remediation. Evidence: state taxonomy.
    • V4.18.b Map dependency outage/not-configured/unauthorized/timeout/rate/ malformed/unknown version to unavailable or evaluator-failure. Evidence: execution tests.
    • V4.18.c Map missing/weak/unreachable/conflicting evidence and disagreement/ambiguity to insufficient or contested with uncertainty. Evidence: evidence-state tests.
    • V4.18.d Mark stale with exact invalidating subject/dependency/policy/ evaluator version and prohibit silent use after change. Evidence: freshness tests.
    • V4.18.e Define gate/review/publish behavior per state and require human authority for any bounded exception with scope/expiry. Evidence: policy tests.
    • V4.18.f Verify API/UI/export/notifications/analytics never serialize or display any nonpass state as pass/healthy/current. Evidence: projection suite.
  • V4.19 Implement Ghana NMC complaint/content checks and remedy workflow with current rule/category, evidence, response deadline, transition, decision, correction/apology/right-of-reply, escalation, and audit. Evidence: legal/regulatory decision remains human and versioned. <!-- SPEC 2026-09-11: extend apps/veritas/api/src/domain/compliance/nmc.ts and the nmc_complaints table rather than starting a second complaint model; .a's jurisdiction profile is pinned with a version and an open-questions list, .f is human (§2.5.9). -->
    • V4.19.a Pin current NMC/jurisdiction profile, rule/category identifiers, effective dates, authoritative references, counsel owner, and review cadence. Evidence: regulatory profile.
    • V4.19.b Define content-check applicability/inputs/findings/evidence/ uncertainty/remediation for ratified editorial categories without encoding legal conclusion as model output. Evidence: check registry.
    • V4.19.c Intake complaint with complainant/privacy/consent, story/release/ claim/target, allegation/category, evidence, received channel/time, and authenticity. Evidence: complaint schema.
    • V4.19.d Calculate statutory/policy response deadline/calendar/time-zone, assign accountable owner/specialists, acknowledge safely, and escalate approaching breach. Evidence: SLA tests.
    • V4.19.e Implement triage/investigate/response-draft/review/decide/remedy/ appeal/close transitions with roles, evidence, concurrency, and immutable audit. Evidence: workflow tests.
    • V4.19.f Require named human legal/editorial decision with exact rule/ evidence/rationale/uncertainty/conditions and prohibit automated disposition. Evidence: authority tests.
    • V4.19.g Execute approved correction/apology/right-of-reply/takedown/ retraction/escalation through canonical release/cascade workflows and verify downstream notice. Evidence: remedy journey.
    • V4.19.h Preserve protected/legal notes and complainant/source identity through redacted notifications/exports; retain/reopen/appeal per policy. Evidence: privacy/audit tests.
  • V4.20 Implement election-coverage gates for phase, equal opportunity, silence periods, results/source verification, polling methodology, advertising/sponsorship, incitement, misinformation, candidate treatment, and current jurisdictional rules. Evidence: election and phase fixtures plus specialist signoff.
    • V4.20.a Pin election/jurisdiction/contest/phase/calendar/time-zone/rule profile, authoritative sources, effective dates, and specialist owner. Evidence: election profile.
    • V4.20.b Determine phase/silence/emergency/result-certification state from trusted inputs and handle uncertain/changed/court-ordered calendars. Evidence: phase tests.
    • V4.20.c Check candidate/party treatment, access/equal-opportunity/ response records, coverage context, and documented editorial basis without imposing false numerical equivalence. Evidence: treatment gate.
    • V4.20.d Check polling/survey source, sponsor, field dates, population/ sample/method/weighting/question/order/margin/uncertainty/disclosure. Evidence: polling gate.
    • V4.20.e Check result source/status/geography/contest/precinct/denominator/ timestamp and prohibit unverified projections or certification claims. Evidence: result gate.
    • V4.20.f Check advertising/sponsorship identity/disclosure/separation and silence-period/channel/territory scheduling constraints. Evidence: campaign-content gate.
    • V4.20.g Check incitement/violence/doxxing/manipulated media/ misinformation risks with uncertainty, escalation, and harm-minimizing distribution treatment. Evidence: safety gate.
    • V4.20.h Persist raw findings separately from human legal/editorial verdict, require specialist signoff, invalidate on phase/rule/content/ source changes, and test edge fixtures. Evidence: signed phase suite.
  • V4.21 Implement data-protection consent, lawful basis, purpose, subject request, redaction, retention, and disclosure checks across sources, tips, media, community, analytics, subscriptions, and B2B products. Evidence: end-to-end DPA request and legal-hold conflict tests.
    • V4.21.a Build processing inventory per named domain with data subjects/ fields/sensitivity/source/controller-processor/purpose/basis/recipient/ region/retention. Evidence: processing register.
    • V4.21.b Validate consent where used for specificity, informed choice, granularity, evidence, withdrawal, age/guardian, and no bundled dark patterns; validate alternate lawful bases separately. Evidence: basis tests.
    • V4.21.c Enforce purpose limitation/minimization/access/field projection/ model-use/analytics cohort/export/disclosure rules at collection and reuse boundaries. Evidence: policy tests.
    • V4.21.d Intake authenticated access/correct/delete/restrict/object/ portability requests with scope, jurisdiction, deadlines, identity minimization, and protected third-party review. Evidence: request schema.
    • V4.21.e Discover subject data across ordinary/protected stores, blobs, search, logs, analytics, backups, channels, payment, B2B, and processors without creating a new leakage index. Evidence: discovery tests.
    • V4.21.f Apply authorized redaction/correction/deletion/restriction/export and propagate to processors/consumers while preserving public-interest journalism and correction evidence as legally justified. Evidence: remedy tests.
    • V4.21.g Resolve retention/legal-hold/editorial-record/source-safety/ third-party-right conflicts through named human legal decision, partial action, rationale, and appeal. Evidence: conflict tests.
    • V4.21.h Verify response/read-back/deadline/audit and no protected-source/ other-subject disclosure in an end-to-end DPA journey. Evidence: DPA E2E.
  • V4.22 Implement trauma, graphic-content, child, vulnerable-person, location, doxxing, suicide/self-harm, and public-safety review with warnings, minimization, specialist escalation, and distribution-specific treatment. Evidence: harm-reduction policy and redacted-preview suite.
    • V4.22.a Define versioned harm categories/severity/context/uncertainty/ jurisdiction/age/vulnerability/public-interest factors and trained reviewer roles. Evidence: harm taxonomy.
    • V4.22.b Identify exact text/quote/media/audio/video/data/location/ metadata risks with evidence and protected reviewer annotations. Evidence: finding contract.
    • V4.22.c Apply identity/location/contact/image/voice/metadata minimization, pseudonymization/redaction, delayed publication, and safe source linking. Evidence: minimization tests.
    • V4.22.d Define consent/guardian/public-interest/safety exception and independent specialist/legal/source-protection review floors. Evidence: authority matrix.
    • V4.22.e Define content warnings/help context, thumbnail/autoplay/ notification/search/social/channel/locale treatment, access/age gates, and no gratuitous repetition. Evidence: distribution matrix.
    • V4.22.f Define suicide/self-harm language/method/location/detail/ imagery/link treatment with current specialist policy and escalation. Evidence: self-harm gate.
    • V4.22.g Define imminent public-safety escalation and emergency authority without exposing victims/sources or creating a generic publish bypass. Evidence: emergency tests.
    • V4.22.h Preview redacted versus authorized views for every target and test copy/download/cache/search/log/AI context/correction propagation. Evidence: redacted-preview suite.
    • V4.22.i Persist human decisions/rationale/conditions/expiry and invalidate on content/media/target/policy changes. Evidence: harm-reduction approval.
  • V5.1 Build review templates by desk, claim risk, source sensitivity, media, audience, locale, and distribution target. Evidence: role/quorum matrix.
    • V5.1.a Define versioned template identity/applicability/priority and required checklist/gate/evidence/reviewer/decision fields. Evidence: template schema.
    • V5.1.b Model desk/claim class-risk/source sensitivity/media/audience/ locale/territory/target dimensions and deterministic template composition. Evidence: applicability tests.
    • V5.1.c Define reviewer roles/credentials/domains/independence/recusal/ quorum/order/parallelism/deadlines and delegates. Evidence: role/quorum matrix.
    • V5.1.d Define tighten-only local customization versus code/policy-owned floors, governance, migration, approval, effective date, and expiry. Evidence: governance tests.
    • V5.1.e Preview exact resolved review plan on a pinned candidate and fail missing/ambiguous/conflicting/unversioned templates. Evidence: resolution tests.
  • V5.2 Build anchored comments and suggestions on blocks, ranges, claims, citations, sources, headlines, media, and disclosures. Evidence: anchor survival.
    • V5.2.a Define comment/thread/suggestion identity, subject kind/version/ anchor, author, visibility, state, mentions, and audit. Evidence: collaboration schema.
    • V5.2.b Support each named subject with kind-specific semantic anchor and protected-source/legal-note visibility rules. Evidence: anchor registry.
    • V5.2.c Implement reply/mention/resolve/reopen/edit-with-history/delete- tombstone and accept/reject/edit suggestion transitions. Evidence: workflow tests.
    • V5.2.d Transform anchors through revisions/concurrent edits or mark orphaned/ambiguous with candidates and no silent retarget. Evidence: anchor survival corpus.
    • V5.2.e Send redacted permission-filtered notifications/deep links and restore focus/subject context. Evidence: notification tests.
    • V5.2.f Test cross-role/tenant/compartment visibility, export, audit, archive/legal hold, and responsive keyboard/screen-reader interaction. Evidence: collaboration E2E.
  • V5.3 Build fact-check assignments and signoff. Evidence: claim-level coverage and separation of duties.
    • V5.3.a Resolve in-scope claims by class/severity/status and create assignments pinned to exact claim/draft/source/evidence revisions. Evidence: assignment contract.
    • V5.3.b Assign qualified independent fact-checker, due date/priority/ checklist, conflict/recusal, backup, and escalation. Evidence: role tests.
    • V5.3.c Review claim text/anchor/citations/locators/source quality/ counter-sources/contradictions/uncertainty and request evidence or edits. Evidence: review journey.
    • V5.3.d Record claim-level verified/unsupported/contested/needs-change/ unable decision with evidence/rationale/conditions/signature. Evidence: decision schema.
    • V5.3.e Compute complete coverage and block signoff for missing/stale/ orphaned/high-risk claims or self-review. Evidence: coverage tests.
    • V5.3.f Invalidate signoff on relevant revision/source/policy changes and preserve prior immutable decisions. Evidence: freshness/SoD tests.
  • V5.4 Build standards review for ethics, balance, harm, conflicts, and disclosure. Evidence: checklist and decision record.
    • V5.4.a Resolve versioned checklist from desk/risk/audience/target and pin exact draft/claims/media/disclosures/policies. Evidence: review plan.
    • V5.4.b Review ethics/public interest/method/source treatment, counterclaims/false balance, conflicts/recusals, and AI/commercial independence. Evidence: ethics checklist.
    • V5.4.c Review harm/minimization/warnings/children/vulnerable persons/ location/privacy and distribution-specific treatment. Evidence: harm checklist.
    • V5.4.d Review required methodology/AI/correction/conflict/sponsorship/ limitation disclosures for truth, placement, and target parity. Evidence: disclosure checklist.
    • V5.4.e Record approve/reject/request-changes/conditional/abstain with anchored findings, evidence, rationale, conditions, and dissent. Evidence: decision record.
    • V5.4.f Enforce qualification/recusal/quorum, stale invalidation, and no automated standards approval. Evidence: authorization tests.
  • V5.5 Build legal review with compartmented notes and retention. Evidence: field-level authorization and redacted notifications.
    • V5.5.a Define legal matter/review/note/attachment/privilege/jurisdiction/ hold/decision schema pinned to exact candidate/evidence. Evidence: legal contract.
    • V5.5.b Store notes/attachments/communications in a separately authorized encrypted compartment with restricted search/export/provider access. Evidence: storage tests.
    • V5.5.c Build request/accept/investigate/request-information/advise/ decide/escalate/close workflow with qualified counsel and conflicts. Evidence: workflow tests.
    • V5.5.d Expose only safe legal-review status/blocker/required action to ordinary users; redact content in UI/API/events/logs/notifications/ audit exports. Evidence: projection tests.
    • V5.5.e Apply retention/legal hold/access review/revocation/export/ destruction with immutable decision audit. Evidence: lifecycle tests.
    • V5.5.f Invalidate decision on scoped content/evidence/jurisdiction/policy changes and test cross-role/tenant/compartment access. Evidence: security tests.
  • V5.6 Build source-protection review without exposing identity to unauthorized reviewers. Evidence: role projections.
    • V5.6.a Resolve source-sensitivity/protection checklist and exact ordinary/protected records under review. Evidence: review contract.
    • V5.6.b Present pseudonymized source role/relationship/risk/consent/ evidence needs to ordinary reviewers with no inferable protected data. Evidence: redacted projection.
    • V5.6.c Present identity/contact/communications/location/safety/grants only to source-protection officers after step-up and purpose check. Evidence: protected projection.
    • V5.6.d Record minimize/redact/restrict/delay/approve/request-changes/ emergency/escalate decisions with separate protected and redacted audit. Evidence: decision tests.
    • V5.6.e Apply approved display/distribution/search/notification/export/ model-use controls and invalidate on source/content/target changes. Evidence: enforcement tests.
    • V5.6.f Test unauthorized UI/API/network/log/cache/count/timing/ notification/audit access and role revocation mid-review. Evidence: non-leakage suite.
  • V5.7 Build attestor invitation, verification, response, disagreement, and expiry workflows. Evidence: end-to-end attestation journey.
    • V5.7.a Create invitation for exact scope/question/domain/credential/ independence requirements, expiry, locale, and safe evidence package. Evidence: invitation contract.
    • V5.7.b Deliver a signed single-use/nonenumerable invitation with identity proofing, consent, authentication, and protected-source redaction. Evidence: invitation security tests.
    • V5.7.c Verify identity/credential/issuer/domain/validity/revocation and collect conflict/independence declarations. Evidence: eligibility tests.
    • V5.7.d Capture agree/disagree/qualified/abstain/unable response with rationale/evidence/uncertainty/signature and no broader approval power. Evidence: response tests.
    • V5.7.e Surface disagreement without suppression, apply quorum/specialist rules, and route clarification/replacement/escalation. Evidence: disagreement tests.
    • V5.7.f Expire/revoke/invalidate on time/credential/scope/input/conflict change and notify affected review. Evidence: lifecycle tests.
    • V5.7.g Run invite→verify→respond→review→expire/reinvite across success, decline, bad credential, conflict, and stale scope. Evidence: E2E.
  • V5.8 Build conflict-of-interest declaration and recusal. Evidence: incompatible conflict blocks approval.
    • V5.8.a Define versioned conflict categories, subject/scope/severity, private versus reviewer-visible details, evidence, update/expiry, and false-declaration policy. Evidence: declaration schema.
    • V5.8.b Prompt declaration at assignment/review/attestation/publish/ correction and when relevant relationships change. Evidence: trigger tests.
    • V5.8.c Let actors declare/no-conflict/update with protected details and route uncertain/material conflicts to authorized standards review. Evidence: declaration journey.
    • V5.8.d Compute recusal/incompatibility effects on access, assignment, contribution, quorum, decision, reassignment, and notification. Evidence: recusal rules.
    • V5.8.e Block approval/signoff/attestation where conflict or missing declaration violates independence and prevent self-clearance. Evidence: separation tests.
    • V5.8.f Preserve immutable decisions and redacted audit; invalidate affected approvals when declarations change. Evidence: lifecycle tests.
  • V5.9 Build copy-edit and accessibility review. Evidence: signed exact revision and locale.
    • V5.9.a Resolve copy/style/locale and accessibility checklist/templates for exact draft/media/target revision. Evidence: review plan.
    • V5.9.b Review grammar/style/names/numbers/dates/links/headlines/captions/ consistency without silently changing claims/quotes/meaning. Evidence: copy-edit tests.
    • V5.9.c Review headings/lists/tables/link purpose/alt text/transcripts/ captions/descriptions/reading order/language and semantic alternatives. Evidence: accessibility tests.
    • V5.9.d Submit anchored suggestions, accept changes as new revision, and re-run claim/evidence/accuracy gates for material edits. Evidence: revision tests.
    • V5.9.e Record separate qualified copy and accessibility decisions with exact revision/locale/target, findings, evidence, rationale, and conditions. Evidence: signed reviews.
    • V5.9.f Invalidate on relevant content/media/locale/template/policy change and prohibit inherited signoff across locales. Evidence: freshness tests.
  • V5.10 Build publisher signoff with fresh evidence recomputation. Evidence: stale dependency race test.
    • V5.10.a Create immutable release-review candidate pinning draft/claims/ sources/evidence/media/disclosures/locales/targets/policies/gates/reviews. Evidence: candidate manifest.
    • V5.10.b Recompute applicable raw measurements and verdicts against pinned inputs, resolving unavailable/contested/stale/failure explicitly. Evidence: fresh gate run.
    • V5.10.c Verify required role/quorum/independence/legal/standards/fact- check/accessibility/rights approvals are current and candidate-scoped. Evidence: approval check.
    • V5.10.d Present target eligibility, embargo/schedule, corrections/ conflicts, residual risks, exceptions, and read-back plan to publisher. Evidence: signoff UI.
    • V5.10.e Record named approve/reject/request-changes with reason, conditions, exact manifest hash, step-up, and immutable signature. Evidence: signoff record.
    • V5.10.f Atomically recheck manifest/dependency versions at signoff and publish dispatch; fail injected last-moment changes. Evidence: stale- dependency race test.
  • V5.11 Build emergency publication and correction policy with higher authority and post-event review. Evidence: no generic gate bypass.
    • V5.11.a Define narrowly enumerated emergency types, activation/deactivation authority, jurisdiction/target scope, duration, and required evidence. Evidence: emergency policy.
    • V5.11.b Define nonwaivable protected-source/legal/safety/authenticity/ human authority floors and specifically deferrable checks. Evidence: gate matrix.
    • V5.11.c Require step-up and higher/dual authority, exact impact preview, reason, conditions, expiry, and immutable incident linkage. Evidence: authorization tests.
    • V5.11.d Publish/correct through the same canonical adapters with explicit emergency label, target state, monitoring, read-back, and rollback/ recall. Evidence: emergency journey.
    • V5.11.e Queue deferred reviews/evidence/disclosures/corrections with owners/deadlines and block ordinary closure until complete. Evidence: follow-up tests.
    • V5.11.f Run mandatory independent post-event review, revoke emergency mode, remediate controls, and preserve full audit. Evidence: game-day.
  • V5.12 Build review SLA, escalation, after-hours, and notification routing. Evidence: embargo/source data is redacted.
    • V5.12.a Define SLA by review type/risk/desk/priority/phase/target with business calendars/time zones, pause conditions, and measurement events. Evidence: SLA contract.
    • V5.12.b Calculate due/breach state deterministically from immutable events and current assignments; handle reassign/reopen/change priority. Evidence: timing tests.
    • V5.12.c Define primary/backup/on-call/escalation roles, schedules, acknowledgment, handoff, higher authority, and stop-the-line actions. Evidence: routing policy.
    • V5.12.d Send channel-appropriate reminders/escalations with dedupe, quiet-hours/emergency rules, delivery/acknowledgment, and retry. Evidence: notification tests.
    • V5.12.e Redact story/embargo/source/legal/complainant detail based on recipient/channel/device while retaining safe deep link after auth. Evidence: redaction tests.
    • V5.12.f Report SLA with privacy floors and no reviewer productivity or editorial-quality proxy; test after-hours/time-zone/outage cases. Evidence: operations report.
  • V5.13 Build immutable approval and waiver records. Evidence: automation cannot impersonate a human approver.
    • V5.13.a Define approval/waiver identity/type/scope/subject manifest/ decision/conditions/reason/evidence/actor-role/time/signature/version. Evidence: record schema.
    • V5.13.b Require current authenticated human identity, step-up where needed, qualified role, independence/recusal/quorum, and explicit intent. Evidence: authority tests.
    • V5.13.c Restrict waiver to policy-allowed gate/risk/target scope with compensating controls, owner, expiry, review, and nonwaivable floors. Evidence: waiver policy tests.
    • V5.13.d Store append-only/tamper-evident records with supersession/ revocation but no mutation/deletion of prior decisions. Evidence: integrity tests.
    • V5.13.e Invalidate effective approval/waiver on manifest/input/policy/ role/conflict/condition/expiry change and expose stale state. Evidence: lifecycle tests.
    • V5.13.f Attempt service account/model/agent/header/audit-field impersonation, replay, forged signature, and direct DB/API creation. Evidence: non-impersonation suite.
  • V5.14 Build newsroom audit export with protected-source redaction. Evidence: authorization and tamper hash.
    • V5.14.a Define export purpose/scope/time/story/actor/event/data-class/ format/version request and authorized requester/approver policy. Evidence: export contract.
    • V5.14.b Select immutable ordered audit events and related decision/ revision/job/release references with stable cursor and completeness manifest. Evidence: selection tests.
    • V5.14.c Apply field/event/object redaction or pseudonymization for protected source, legal, tipster, child/vulnerable, security, and third- party data. Evidence: redaction matrix.
    • V5.14.d Generate readable and machine-verifiable output with schema, reason codes, gaps, time zone, chain/manifest hash, signature, and encryption. Evidence: export fixture.
    • V5.14.e Deliver through expiring scoped secure download with step-up, watermark if appropriate, rate/size limits, and export audit. Evidence: delivery tests.
    • V5.14.f Verify authorization, two-tenant/compartment nonleakage, completeness, tamper detection, replay consistency, and legal-hold behavior. Evidence: audit export suite.

V6 — Persistence, APIs, search, and distribution integration#

  • V6.1 Decide authoritative stores for stories, drafts, sources, claims, evidence, attestations, reviews, schedules, releases, and cascades. Evidence: ownership ADR. <!-- SPEC 2026-09-11: the decision is among the three persistence stories in the Phase V header note (2). The in-memory CMS cannot be authoritative for anything; the raw-SQL API and the Prisma schema disagree on at least user_subscriptions. Draft docs/adr/ADR-V6.1-veritas-persistence-ownership.md (Proposed, like ADR-M0.5), build .b–.e's tests against the option it proposes, and leave the parent open for the named approver. -->
    • V6.1.a Inventory current tables/documents/blobs/indexes/caches/queues/CMS records and writers/readers for each named aggregate. Evidence: storage inventory. <!-- veritas-storage-inventory.json: storage-layer inventory of Veritas. Of the 7 named layers, 6 are PRESENT — tables (67 relational Prisma aggregates over 4 PrismaClients), blobs (MediaAsset object storage / storageBucket), indexes (ArticleEmbedding vector index), caches (Redis), queues (transactional outbox PushNotificationOutbox/PushNotificationDelivery + BullMQ), CMS — and 1 is ABSENT: documents (no MongoDB/document store; Veritas is relational-only). FINDING: Veritas is a RICH MULTI-STORE server platform, contrasting Euterpe's desktop-file model (E7.1.a) — same storage-layer method, opposite architecture. 0 repository classes (service-direct, like Isis). Used PRECISE per-layer patterns (new Redis, Outbox, ArticleEmbedding, MediaAsset) scoped to the Veritas tree. generate-veritas-storage-inventory.mjs + 6 tests, wired into inventory/verify + CI. V6.1.b (system/record boundary) + later V6.1.* are design and stay unchecked. -->
    • V6.1.b Define authoritative system/record boundary, aggregate/ transaction ownership, immutable versus mutable data, and durability class per entity. Evidence: ownership matrix.
    • V6.1.c Define derived search/read-model/cache/CMS/distribution projections, freshness/rebuild/reconciliation, and forbidden reverse writes. Evidence: projection contracts.
    • V6.1.d Define protected-compartment, encryption/key/region, retention/ hold/backup, and access-audit storage constraints. Evidence: data placement review.
    • V6.1.e Resolve split/duplicate ownership, failure boundaries, outbox/ saga needs, migration, recovery, and accountable service/team. Evidence: approved ownership ADR.
  • V6.2 Add missing schemas with explicit queryable columns and safe migrations. Evidence: production-size rehearsal.
    • V6.2.a Derive missing tables/columns/types/constraints/indexes/ partitions/foreign keys from canonical contracts and required queries. Evidence: schema diff.
    • V6.2.b Keep identity/state/tenant/desk/story/compartment/version/time/ relationship fields explicit and queryable; avoid opaque core JSON. Evidence: schema review.
    • V6.2.c Design expand/backfill/dual-read-write/switch/contract migrations with chunking, checkpoints, idempotency, throttling, and rollback/ restore boundaries. Evidence: migration plan.
    • V6.2.d Preserve immutable revisions/hashes/audit/protected encryption and validate references/state mappings during backfill. Evidence: migration tests.
    • V6.2.e Rehearse production-volume/cardinality/skew/concurrency on real engine, record locks/runtime/storage/query plans, and meet budgets. Evidence: rehearsal report.
    • V6.2.f Deploy with observability/reconciliation/stop conditions and remove old schema only after all consumers prove compatibility. Evidence: rollout gate.
  • V6.3 Implement immutable revision and evidence stores with durable-before-ack behavior. Evidence: real-database tests.
    • V6.3.a Implement append-only revision/evidence/member/blob-reference/ hash/signature schemas with unique identity and parent/version constraints. Evidence: persistence tests.
    • V6.3.b Write record, required references, blob completion, and outbox atomically or through a recoverable protocol before acknowledging. Evidence: durability tests.
    • V6.3.c Enforce optimistic concurrency/idempotency/deduplication and immutable historical content under retry/race/replay. Evidence: concurrency tests.
    • V6.3.d Verify canonical hashes/signatures on read and detect missing/ changed/mismatched blobs or members without returning approved state. Evidence: integrity tests.
    • V6.3.e Inject database/blob/network/process failure at each boundary, restart/reconcile, and prove acknowledged writes survive. Evidence: fault suite.
    • V6.3.f Run backup/restore/read-back against real database/blob services with tenant/compartment isolation. Evidence: durable round-trip.
  • V6.4 Implement pitch/assignment/story/draft APIs with optimistic concurrency and property-level authorization. Evidence: contract suite.
    • V6.4.a Define versioned CRUD/query/transition/command contracts, stable IDs, errors, pagination/filter/sort, idempotency, and deprecation. Evidence: OpenAPI schemas.
    • V6.4.b Implement request identity/tenant/desk/story/role/state/ capability authorization for every object and function. Evidence: route auth tests.
    • V6.4.c Apply field-level read/write/patch/filter/sort/export projection, including embargo/protected/legal/commercial constraints. Evidence: property auth tests.
    • V6.4.d Require expected version/ETag for mutation and return structured conflict/current revision without silent overwrite. Evidence: concurrency tests.
    • V6.4.e Implement idempotent create/commands/transitions with transactional audit/outbox and safe retry. Evidence: idempotency tests.
    • V6.4.f Test positive/negative/cross-tenant/stale-role/direct-object/ mass-assignment/unknown-field/oversized requests. Evidence: contract suite.
  • V6.5 Implement compartmented source APIs and access-grant audit. Evidence: protected identity negative tests.
    • V6.5.a Separate ordinary source endpoints/contracts from protected identity/contact/communications/safety/grant endpoints and credentials. Evidence: API boundary.
    • V6.5.b Implement pseudonymous safe projection and opaque linkage without protected field/count/timing/existence leaks. Evidence: projection tests.
    • V6.5.c Require current nominative purpose/time/story-scoped grant, step-up/device/session policy, and field/function authorization on every protected call. Evidence: access tests.
    • V6.5.d Implement request/approve/deny/revoke/expire/emergency grant workflow with separation, reason, alerts, and post-review. Evidence: grant tests.
    • V6.5.e Emit tamper-evident redacted attempt/access/export audit without copying secrets into ordinary systems. Evidence: audit tests.
    • V6.5.f Test enumeration, BOLA, mass assignment, filter/sort/error/cache/ log leakage, token replay, revoked/mid-session grants, and two tenants. Evidence: protected identity negative suite.
  • V6.6 Implement claim/citation/evidence/attestation APIs. Evidence: canonical invariant route tests.
    • V6.6.a Define versioned resource/command/query/batch contracts and explicit claim/citation/evidence-pack/attestation state transitions. Evidence: API schemas.
    • V6.6.b Validate exact story/draft/revision/anchor/source/locator/member/ scope references and canonical class/state/value vocabularies. Evidence: invariant tests.
    • V6.6.c Enforce tenant/story/desk/role/field/function/compartment/ independence authorization and protected redacted projections. Evidence: authorization tests.
    • V6.6.d Implement optimistic concurrency/idempotency/transactional audit/ outbox and immutable revision or decision records. Evidence: mutation tests.
    • V6.6.e Invalidate dependent gates/reviews/releases on relevant mutation and publish versioned events with no protected payload. Evidence: dependency tests.
    • V6.6.f Run valid/invalid/stale/duplicate/orphan/hash/scope/cross-tenant/ mass-assignment route fixtures. Evidence: canonical contract suite.
  • V6.7 Implement verification/gate jobs with current policy versions. Evidence: stale input and rerun tests.
    • V6.7.a Define job request/run/attempt/result contracts pinning subject/ inputs/evaluator/model/corpus/weight/policy versions and requested gates. Evidence: job schema.
    • V6.7.b Resolve applicability/current approved versions at dispatch and reject mutable aliases, missing configuration, or unauthorized inputs. Evidence: dispatch tests.
    • V6.7.c Implement durable queue/lease/heartbeat/timeout/retry/backoff/ cancellation/idempotency/resource/budget behavior per gate. Evidence: worker tests.
    • V6.7.d Persist raw measurements/evidence before separate verdicts, terminal state, audit, and dependency invalidation events. Evidence: result tests.
    • V6.7.e Detect subject/input/policy version changes before execution and commit; mark stale/superseded and require explicit rerun. Evidence: stale input tests.
    • V6.7.f Inject provider/evaluator/queue/database/blob failure, cancel, replay, duplicate, and out-of-order events. Evidence: fault/rerun suite.
  • V6.8 Implement protected, permission-aware newsroom search. Evidence: titles, counts, facets, and snippets do not leak.
    • V6.8.a Define per-document/field classification, tenant/desk/story/ compartment/role ACL, reveal/embargo/public state, locale, freshness, and deletion schema. Evidence: search security model.
    • V6.8.b Index only allowlisted safe projections from authoritative events with content/version hash, current ACL, and protected fields excluded. Evidence: indexing tests.
    • V6.8.c Apply authorization before candidate retrieval/aggregation/ ranking/highlighting and stable cursor pagination. Evidence: query tests.
    • V6.8.d Prevent query suggestion/autocomplete/spelling/semantic vector/ title/snippet/count/facet/timing/cache/log leaks. Evidence: leakage suite.
    • V6.8.e Process access revoke/embargo/retraction/correction/delete/ compartment change promptly and reconcile index to source of truth. Evidence: update tests.
    • V6.8.f Run two-tenant/multi-role/existence-oracle/BOLA/injection/ high-volume/stale-index adversarial tests. Evidence: protected search report.
  • V6.9 Implement CMS plan/publish/read-back/compensate adapter. Evidence: idempotent partial-failure recovery. <!-- SPEC 2026-09-11: the adapter targets whichever store V6.1 makes authoritative; if that is the Fastify CMS, its memory-storage.ts must first be replaced by a durable store and its routes given auth, and those are new cells under V6.2/V6.4, not side effects of this one. Read-back (.d) is an HTTP read from a separately started process, never the same in-memory object. -->
    • V6.9.a Define target/versioned adapter capabilities, field/asset/locale/ rights/correction mappings, credentials, rate limits, and unsupported behavior. Evidence: adapter profile.
    • V6.9.b Build a read-only publication plan from immutable candidate with target operations/dependencies/idempotency keys/expected current state/ loss/impact. Evidence: plan tests.
    • V6.9.c Execute create/update/assets/link/schedule/publish actions in dependency order with retries, acknowledgments, and per-action durable status. Evidence: publish tests.
    • V6.9.d Read back public/CMS canonical content, metadata, assets, visibility, timestamps, IDs, version, and correction state; semantic- compare to plan. Evidence: verification tests.
    • V6.9.e Compensate only reversible actions, otherwise retain partial release with operator remediation; never report whole success. Evidence: compensation tests.
    • V6.9.f Test timeout after commit, duplicate request, reordered callback, rate/auth loss, asset failure, schedule race, read-back mismatch, and retry. Evidence: partial-failure suite.
  • V6.10 Implement NewsML-G2 package generation and validation. Evidence: golden News Item, Package, Planning, update, and correction fixtures. <!-- SPEC 2026-09-11: greenfield (V1.14 note); .e validates against the vendored IPTC XSDs and .f's "supported consumers" means at least one independent NewsML-G2 reader run over our output with its output retained (§2.5.6). -->
    • V6.10.a Resolve exact pinned release/story/source/claim/media/rights/ topic/planning/correction data and profile/code-list versions. Evidence: generator inputs.
    • V6.10.b Generate deterministic News Items with identity/version/status/ metadata/content/rights/subjects/associations and supported extensions. Evidence: News Item goldens.
    • V6.10.c Generate deterministic Package and Planning Items with ordered/ role-qualified membership, schedules/embargoes, and protected redaction. Evidence: package/planning goldens.
    • V6.10.d Generate initial/update/correction/replacement/retraction chains with prior references, signals, notices, and target actions. Evidence: lifecycle goldens.
    • V6.10.e Validate schema/profile/code list/cardinality/URI/date/time/ rights/language/ID rules using pinned validator. Evidence: validation tests.
    • V6.10.f Round-trip through supported consumers and semantic-compare without silently dropping unsupported data. Evidence: interoperability report.
  • V6.11 Implement newsletter, social, audio, video, wire, and archive distribution adapters as target-specific jobs. Evidence: target matrix.
    • V6.11.a Define shared job/attempt/artifact/ack/read-back/correction envelope plus per-target capabilities, credentials, limits, and version. Evidence: adapter contracts.
    • V6.11.b Implement newsletter provider payload/delivery/archive/ unsubscribe/correction mapping from an approved edition. Evidence: newsletter tests.
    • V6.11.c Implement each social/channel adapter with text/media/link/ disclosure/schedule/edit-delete/correction/moderation capability and explicit unsupported cells. Evidence: social tests.
    • V6.11.d Implement audio/video platform feed/media/metadata/captions/ rights/provenance/replace/correct behavior. Evidence: AV adapter tests.
    • V6.11.e Implement wire NewsML/package acknowledgement/update and archive immutable deposit/public status/correction/legal-hold behavior. Evidence: wire/archive tests.
    • V6.11.f Run durable idempotent target jobs with per-target status, retry/backoff/rate/cancel/read-back and no aggregate false success. Evidence: target matrix.
  • V6.12 Implement C2PA signing/validation for supported media/package outputs. Evidence: tamper and trust-chain tests. <!-- SPEC 2026-09-11: sign with c2pa-node@0.5.26 through libs/veritas/content-auth; .c's "HSM or equivalent" is a test signing certificate generated in the harness and named as such; .f's tamper cases mutate bytes, not our own status field. -->
    • V6.12.a Define supported C2PA versions/algorithms/assertions/ingredients/ actions, claim generators, signing identities, trust stores, and key lifecycle. Evidence: C2PA profile.
    • V6.12.b Build deterministic manifest from exact immutable media/article package/rendition/edit/AI/rights provenance. Evidence: manifest tests.
    • V6.12.c Sign in isolated authorized service/HSM or equivalent with idempotency, timestamp, certificate chain, audit, and no private-key exposure. Evidence: signing tests.
    • V6.12.d Embed or associate manifest per format/target, preserve hashes through supported transforms, and record stripped provenance. Evidence: packaging tests.
    • V6.12.e Validate cryptographic integrity, ingredient/action lineage, signer trust/revocation/expiry, and expected content binding. Evidence: validation tests.
    • V6.12.f Test valid/tampered/replaced/stripped/unknown/untrusted/expired/ revoked/malformed chains and present precise nontruth status. Evidence: trust-chain suite.
  • V6.13 Implement correction and retraction cascade orchestration through outbox/saga. Evidence: duplicate and out-of-order delivery tests. <!-- SPEC 2026-09-11: REUSE, do not rebuild: executeVeritasSourceRetractionCascade in libs/oshun/domain-veritas, libs/oshun/veritas-cascade-worker and apps/oshun/bff/src/veritas/retraction-cascade-runtime.ts already plan → persist (Postgres) → publish (Redis) → re-ground Sophia/Metis with live-infra tests. This cell extends that saga to corrections and to the targets V7.15 registers. -->
    • V6.13.a Create immutable cascade/plan/actions from a pinned dependency graph, approved remedy, target capabilities, and desired states. Evidence: saga contract.
    • V6.13.b Commit authoritative correction/retraction state and outbox event atomically with unique cascade/action/idempotency identities. Evidence: transaction tests.
    • V6.13.c Execute dependency-aware per-target update/replace/remove/ notify/reindex/invalidate actions with leases/retry/backoff/timeouts. Evidence: worker tests.
    • V6.13.d Handle duplicate/out-of-order/late callbacks/events using expected versions and monotonic state without undoing newer remedies. Evidence: delivery tests.
    • V6.13.e Read back every target, reconcile mismatch/unknown/unsupported/ partial state, retry or escalate, and never close unresolved targets. Evidence: verification tests.
    • V6.13.f Compensate only safe reversible actions, preserve audit/history, and replay after outage without duplicate notices. Evidence: fault suite.
  • V6.14 Implement two-tenant and protected-compartment isolation across API, search, blobs, jobs, exports, and audit. Evidence: integration suite.
    • V6.14.a Create two tenants with overlapping IDs/names/content plus ordinary/protected stories, sources, grants, blobs, jobs, and events. Evidence: isolation fixture.
    • V6.14.b Test API object/property/function/filter/sort/batch/cache/ concurrency access under direct and stale identifiers. Evidence: API isolation.
    • V6.14.c Test search indexing/query/suggestion/snippet/count/facet/vector/ timing and deletion/revocation. Evidence: search isolation.
    • V6.14.d Test blob keys/URLs/ranges/proxies/metadata/uploads/downloads/ transforms/backups and signed-link substitution. Evidence: blob isolation.
    • V6.14.e Test queue payloads/workers/callbacks/artifacts/logs/metrics/ retries/dead letters across tenant/compartment context. Evidence: job isolation.
    • V6.14.f Test audit/export/report/notification/webhook/CMS/distribution projections and redactions. Evidence: output isolation.
    • V6.14.g Attempt service account, admin, emergency grant, role revoke, cache poisoning, guessed ID, and cross-tenant relationship attacks. Evidence: adversarial integration suite.
  • V6.15 Generate OpenAPI and consumer contracts for Studio, CMS, reader, archive, and distribution consumers. Evidence: drift gate. <!-- SPEC 2026-09-11: the API app has apps/veritas/api/src/interfaces/http/openapi.ts; generate from the served router, not by hand, and treat the CMS routes and BFF /v1/veritas/* as separate audiences. Follow M8.15's register shape (every operation classified, unclassified = failure). -->
    • V6.15.a Generate canonical versioned OpenAPI from authoritative route/ schema/error/auth/pagination/idempotency contracts with stable operation IDs. Evidence: OpenAPI artifact.
    • V6.15.b Produce audience-specific Studio/CMS/reader/archive/distribution contract projections without publishing protected internal fields. Evidence: consumer schemas.
    • V6.15.c Add representative requests/responses/errors/events and generated clients/types where owned, pinned to exact spec digest. Evidence: contract fixtures.
    • V6.15.d Run schema lint/breaking-change/unknown route/unimplemented operation and server-versus-spec conformance tests. Evidence: provider gate.
    • V6.15.e Run consumer compile/fixture/semantic contract tests for each supported version and target behavior, including corrections. Evidence: consumer gate.
    • V6.15.f Fail CI on unreviewed drift, protected-field exposure, orphan operation, stale generated client, or unsupported breaking change. Evidence: drift gate.

V7 — Calendar, publication, correction, retraction, and curation#

  • V7.1 Build an editorial calendar for assignments, embargoes, coverage plans, review, publication, newsletters, and distribution windows. Evidence: accessible time-zone journey.
    • V7.1.a Define calendar item/reference contracts for each named entity with start/end/deadline/window/time zone/recurrence/status/owner and exact source revision. Evidence: calendar schema.
    • V7.1.b Build permission-filtered day/week/month/agenda views with stable filters by desk/owner/type/state/target/locale and protected redaction. Evidence: view tests.
    • V7.1.c Create/edit/reschedule/assign from owned workflows with validation, dependency/embargo/conflict/impact preview, concurrency, and audit. Evidence: action tests.
    • V7.1.d Normalize storage and display correctly across newsroom/user/ target time zones, DST transitions, ambiguous/nonexistent times, and locale calendars. Evidence: temporal tests.
    • V7.1.e Provide table/agenda and form alternatives to visual calendar/ drag operations with keyboard/screen-reader parity. Evidence: accessibility tests.
    • V7.1.f Exercise assignment→review→embargo→publish→distribution journey across time zones and responsive web/desktop. Evidence: E2E.
  • V7.2 Model embargo as an independent policy, not a publish status. Evidence: NewsML-G2 and CMS mapping tests.
    • V7.2.a Implement embargo policy entity/versions independently of story, editorial, release, schedule, and target execution states. Evidence: state model.
    • V7.2.b Evaluate access/display/export/model-use/notification/search/ preview/publish constraints from current time/event/party/territory/ channel and authorized exceptions. Evidence: policy tests.
    • V7.2.c Handle set/update/lift/expire/breach/dispute with authority, reason, source agreement, affected objects, audit, and notifications. Evidence: lifecycle tests.
    • V7.2.d Recompute release eligibility/schedules on policy changes without changing editorial approval or falsely marking published. Evidence: independence tests.
    • V7.2.e Map embargo fields/actions to NewsML-G2 and each CMS/distribution target, preserving unsupported or lossy behavior explicitly. Evidence: adapter matrix.
  • V7.3 Build desk capacity and specialist-review reservation. Evidence: committed slots cannot be double-booked.
    • V7.3.a Define staff/role/skill/credential/desk/time-zone/availability/ capacity/calendar and privacy-safe display contracts. Evidence: capacity schema.
    • V7.3.b Forecast assignment/review workload by type/risk/effort/deadline/ phase without turning output volume into performance scoring. Evidence: workload model.
    • V7.3.c Create tentative/held/committed/released/expired reservation with story/review/owner/window/capacity units, priority, and dependencies. Evidence: reservation state machine.
    • V7.3.d Commit atomically under concurrent booking, enforce conflicts/ leave/on-call/quorum/credential constraints, and prevent double booking. Evidence: concurrency tests.
    • V7.3.e Support reschedule/reassign/cancel/overbook-exception with impact, authority, notifications, SLA recalculation, and audit. Evidence: operations tests.
  • V7.4 Build release candidates locked to exact draft, source, claim, evidence, media, attestation, policy, and locale versions. Evidence: mutation invalidation.
    • V7.4.a Resolve complete transitive candidate dependency manifest across exact draft/blocks/claims/sources/citations/evidence/media/attestations/ reviews/disclosures/locales/targets/policies. Evidence: manifest schema.
    • V7.4.b Freeze canonical serialized references/hashes/signatures and target/template/config/adapter versions without copying protected plaintext. Evidence: candidate artifact.
    • V7.4.c Verify every dependency exists/is authorized/current/approved/ integrity-valid and record missing/stale/contested/unsupported state. Evidence: validation tests.
    • V7.4.d Allow explicit candidate create/supersede/abandon only; prohibit mutation of locked manifests or historical approvals. Evidence: immutability tests.
    • V7.4.e Subscribe/reconcile relevant mutations and invalidate candidate, previews, signoffs, schedules, and jobs before dispatch. Evidence: invalidation tests.
    • V7.4.f Recheck manifest transactionally at signoff and execution to close dependency races. Evidence: race tests.
  • V7.5 Build target preview and eligibility diagnostics. Evidence: unsupported or missing fields never silently drop.
    • V7.5.a Resolve exact release candidate/target/account/locale/audience/ template/adapter/policy versions and current credentials/capabilities. Evidence: preview inputs.
    • V7.5.b Render production-equivalent semantic/visual/payload preview with target limits/transforms/fallbacks/disclosures/correction behavior. Evidence: preview tests.
    • V7.5.c Validate required/optional fields, types/lengths/media/renditions/ rights/accessibility/embargo/schedule/approval/channel policies. Evidence: diagnostics.
    • V7.5.d Report unsupported/dropped/truncated/transformed/lossy fields individually with severity, effect, remediation, and no false eligible state. Evidence: loss tests.
    • V7.5.e Recompute on candidate/target capability/template/policy/account changes and label stale/fixture/degraded/unavailable state. Evidence: freshness tests.
    • V7.5.f Compare preview semantic output to actual read-back for supported cells. Evidence: parity gate.
  • V7.6 Build schedule, publish, read-back verification, and multi-target result views. Evidence: one failed target cannot make the release wholly successful.
    • V7.6.a Create per-target scheduled action with candidate/eligibility/ account/locale/time zone/window/idempotency/dependency and authorized publisher. Evidence: schedule contract.
    • V7.6.b Recheck embargo/candidate/gates/approvals/rights/credentials/ target capability immediately before durable dispatch. Evidence: dispatch guard tests.
    • V7.6.c Execute target-specific publish jobs with queue/lease/retry/rate/ cancel/callback and persist target ID/URL/version/acknowledgment. Evidence: job tests.
    • V7.6.d Read back public/target content, metadata, assets, visibility, correction/disclosure/provenance and semantic-compare to candidate. Evidence: read-back tests.
    • V7.6.e Present queued/running/scheduled/published/verified/degraded/ failed/unknown/compensated per target plus aggregate partial status. Evidence: result view.
    • V7.6.f Retry/remediate/recall selected targets with impact/concurrency/ audit, never converting one success/failure into whole-release success. Evidence: partial failure E2E.
  • V7.7 Build visible correction authoring with prior/new text, rationale, affected claims, and reader notice. Evidence: canonical correction invariants.
    • V7.7.a Start correction from exact public release/target or verified feedback with type/severity/discovery/source/owner and preserve prior version. Evidence: correction intake.
    • V7.7.b Select exact prior blocks/ranges/media/captions/metadata and author replacement/removal/addition as a new revision with semantic diff. Evidence: change editor.
    • V7.7.c Link affected claims/citations/sources/packages/targets and recompute evidence/gates/reviews plus blast radius. Evidence: impact tests.
    • V7.7.d Require factual rationale, discovery/correction time, responsible editor, applicable approvals, and target/locale-specific reader notice. Evidence: notice contract.
    • V7.7.e Preview prior/new content and notice at every target, including archive/history and accessible semantic output. Evidence: preview tests.
    • V7.7.f Enforce prior text/version/hash match, immutable chronology, concurrency, correction-of-correction, and no silent overwrite. Evidence: canonical invariant suite.
  • V7.8 Build source/claim/story retraction intake and impact preview. Evidence: complete dependency graph.
    • V7.8.a Intake source/claim/story retraction candidate with issuer/ evidence/cause/scope/severity/verification/time and exact subject revision. Evidence: intake schema.
    • V7.8.b Authenticate/verify external retraction or require authorized internal human decision; classify disputed/partial/unknown state. Evidence: verification tests.
    • V7.8.c Traverse full versioned dependency graph through citations/ claims/stories/packages/locales/media/newsletters/social/channels/audio/ video/archive/search/B2B. Evidence: impact query.
    • V7.8.d Show each affected node/edge/current release/target, proposed remedy, urgency/owner/state, protected redaction, and unknown consumers. Evidence: preview UI.
    • V7.8.e Freeze a read-only impact snapshot and reconcile against consumer registry; never mutate during preview. Evidence: completeness tests.
    • V7.8.f Require separately approved cascade plan and fresh graph check before execution. Evidence: authorization/race tests.
  • V7.9 Build retraction-cascade execution with per-target status and retry. Evidence: CMS, reader, newsletter, social, archive, search, and package fixtures.
    • V7.9.a Convert approved impact snapshot into dependency-ordered, target-specific retain/correct/withdraw/retract/remove/notify/reindex/ invalidate actions. Evidence: execution plan.
    • V7.9.b Dispatch durable idempotent saga actions with expected versions, prerequisites, attempts, rate limits, retries, and operator ownership. Evidence: orchestration tests.
    • V7.9.c Implement CMS/reader and NewsML/package state/notice/version updates with preserved history and canonical URLs. Evidence: core target tests.
    • V7.9.d Implement newsletter/social/channel/audio/video/archive/search/ B2B target capabilities with explicit unsupported/partial behavior. Evidence: channel target tests.
    • V7.9.e Show target/action queued/running/succeeded/verified/failed/ unknown/unsupported/compensated and allow authorized safe retry. Evidence: status UI.
    • V7.9.f Handle duplicate/out-of-order/late callbacks, concurrent newer corrections, auth loss, outage, and replay without reversing remedies. Evidence: fault tests.
  • V7.10 Build correction/retraction verification from downstream read-back. Evidence: stale dependent surface blocks closure.
    • V7.10.a Define expected semantic state per target/action, verification adapter/version, tolerated delay, retry policy, and closure criteria. Evidence: verification contract.
    • V7.10.b Read back public content/metadata/status/notices/history/assets/ search/cache/feed/API/webhook state using independent paths. Evidence: read-back tests.
    • V7.10.c Compare identity/version/text/media/claim/correction/retraction/ visibility/time/links semantically rather than only acknowledgment. Evidence: comparator tests.
    • V7.10.d Mark mismatch/stale/unreachable/unauthorized/unsupported/unknown per dependent surface with evidence, owner, retry, and escalation. Evidence: failure states.
    • V7.10.e Prevent cascade/correction closure while any required registered surface is unresolved or registry reconciliation is incomplete. Evidence: closure gate.
    • V7.10.f Reverify after caches/indexes/feeds propagate and retain signed final reconciliation plus later drift monitoring. Evidence: verification report.
  • V7.11 Build section, topic, package, timeline, related, and featured curation via versioned bundles. Evidence: no direct production writes.
    • V7.11.a Define typed versioned bundle identity/scope/locale/territory/ window/status/owner and ordered/grouped item references by curation kind. Evidence: bundle schemas.
    • V7.11.b Build draft editor for select/search/add/remove/order/group/ annotate/schedule with non-drag controls and protected/embargo rules. Evidence: curation UI.
    • V7.11.c Validate item release/rights/correction/retraction/freshness/ target/locale compatibility and required contextual notices. Evidence: eligibility tests.
    • V7.11.d Provide semantic diff/review/approval/preview for exact bundle revision and invalidate on item or policy changes. Evidence: review tests.
    • V7.11.e Publish through canonical target adapter/job/read-back with immutable prior versions, idempotency, and rollback/new revision. Evidence: bundle release tests.
    • V7.11.f Prohibit client/direct CMS/production mutations and audit every attempted bypass. Evidence: no-direct-write tests.
  • V7.12 Build archive and legal-hold behavior. Evidence: public removal does not destroy required evidence.
    • V7.12.a Define archive tiers/state/retention/access/index/public visibility and legal-hold scope/reason/authority/effective/expiry. Evidence: lifecycle contract.
    • V7.12.b Archive immutable stories/revisions/sources/evidence/media/ reviews/releases/corrections/audit with reference and integrity preservation. Evidence: archive tests.
    • V7.12.c Separate public unpublish/withdraw/retract/search removal from internal retention and show appropriate public notice. Evidence: removal tests.
    • V7.12.d Apply hold to records/blobs/indexes/backups/exports/processors, suspend conflicting deletion, and record conflicts. Evidence: hold tests.
    • V7.12.e Enforce archive/hold role/field/compartment access, protected search redaction, audited export, and secure restore. Evidence: authorization tests.
    • V7.12.f Test retention expiry/deletion request/correction/retraction/ source protection/backup restore under and after hold. Evidence: lifecycle suite.
  • V7.13 Build source and attestation expiry monitoring. Evidence: affected published claims enter review.
    • V7.13.a Define expiry/staleness triggers from source update/retraction/ rights/review interval and attestor credential/conflict/validity/scope. Evidence: monitoring rules.
    • V7.13.b Schedule idempotent scans and event-driven checks against current versions with leases/cursors/rate limits and clock/time-zone handling. Evidence: monitor tests.
    • V7.13.c Traverse affected claim/story/release/target graph and calculate risk/urgency/deadline without exposing protected source details. Evidence: impact tests.
    • V7.13.d Mark dependent evidence/gates/approvals/releases as stale or review-required according to policy; never silently keep current pass. Evidence: invalidation tests.
    • V7.13.e Create deduplicated review assignments/notifications/escalations with safe context, owner, due date, and remediation. Evidence: workflow tests.
    • V7.13.f Track resolve/replace/renew/correct/retract/waive outcomes and reverify downstream state. Evidence: monitoring E2E.
  • V7.14 Build post-publication audience feedback and complaint intake. Evidence: triage cannot directly rewrite published truth.
    • V7.14.a Provide accessible authenticated or permitted anonymous intake with channel/type/story/claim/target/locale/contact/consent/privacy/ evidence fields. Evidence: intake contract.
    • V7.14.b Scan/quarantine attachments and rate-limit/spam/abuse protect without exposing reporter/source identities or creating a public existence oracle. Evidence: security tests.
    • V7.14.c Classify feedback, factual correction request, NMC/legal complaint, right-of-reply, safety/privacy, accessibility, and tip as distinct queues. Evidence: routing tests.
    • V7.14.d Assign owner/SLA/priority/acknowledgment/status/response/ escalation with protected/redacted notes and audit. Evidence: triage workflow.
    • V7.14.e Link evidence to exact public version/claim without modifying canonical story, claim, correction, or public content. Evidence: immutability tests.
    • V7.14.f Convert only via authorized editorial/complaint/correction/tip workflow and send privacy-safe resolution. Evidence: handoff E2E.
  • V7.15 Build cross-domain impact hooks for explainers, saved claims, Telegram cards, and Living Scenes. Evidence: correction cascade reaches registered consumers. <!-- SPEC 2026-09-11: the consumer registry must include the consumers that already exist — the Sophia re-grounder and Metis revalidator in libs/oshun/veritas-cascade-worker and the personal-retraction-acknowledgement store in the BFF — before adding new ones; X.21 later reads this registry. -->
    • V7.15.a Define versioned consumer registration with domain/owner/ endpoint/event types/subject mappings/capabilities/auth/retry/read-back/ SLA. Evidence: registry schema.
    • V7.15.b Publish correction/retraction/source/claim/story lifecycle events with opaque stable IDs, versions, reason/notice, and no protected data. Evidence: event contract.
    • V7.15.c Implement explainer, saved-claim, Telegram-card, and Living Scene adapters/projections with explicit correction/retraction behavior. Evidence: consumer tests.
    • V7.15.d Deliver through durable outbox with idempotency/order/retry/ dead-letter/replay and per-consumer status. Evidence: delivery tests.
    • V7.15.e Require acknowledgment or independent read-back and keep unknown/ stale/unregistered/unsupported consumers visible. Evidence: verification tests.
    • V7.15.f Reconcile capability registry to observed consumers and fail missing owner/hook/proof or protected-field exposure. Evidence: cascade coverage gate.
  • V7.16 Build emergency recall with step-up authorization and later review. Evidence: reversible, audited incident journey.
    • V7.16.a Define eligible emergency recall reasons/severity/scope/targets/ authority/nonwaivable constraints and distinction from ordinary retraction. Evidence: recall policy.
    • V7.16.b Build read-only blast-radius/target-capability/impact preview, alternatives, expected public state, and rollback/restoration plan. Evidence: recall plan.
    • V7.16.c Require current publisher/corrections/emergency role, step-up and higher/dual approval as policy requires, reason, incident, and expiry. Evidence: authorization tests.
    • V7.16.d Dispatch idempotent per-target hide/withdraw/replace/notify/ invalidate actions with progress, failures, and independent read-back. Evidence: execution tests.
    • V7.16.e Restore or replace only through new approved state after risk is addressed; retain recall notice/history and prevent stale reappearance. Evidence: reversibility tests.
    • V7.16.f Conduct mandatory post-event review of trigger/authority/actions/ effects/gaps, close unresolved targets, and preserve audit. Evidence: incident journey.
  • V7.17 Build desk/edition/program steering overlays for sourcing, specialist review, risk, style, audience, distribution, and cadence. Evidence: tighten-only property tests and impact preview.
    • V7.17.a Define versioned overlay identity/scope/priority/effective window/ owner/reviewer and each allowed sourcing/review/risk/style/audience/ distribution/cadence property. Evidence: overlay schema.
    • V7.17.b Define code/regulatory/global policy floors and partial order for stricter versus weaker values per property. Evidence: lattice spec.
    • V7.17.c Resolve applicable overlays deterministically across desk/ edition/program/locale/territory/phase/target with conflicts explicit. Evidence: resolution tests.
    • V7.17.d Preview affected drafts/gates/reviews/schedules/targets/users and newly blocked/required actions before activation. Evidence: impact preview.
    • V7.17.e Govern draft/review/approve/activate/supersede/rollback/migrate/ expire with separation, audit, and no direct policy-floor editing. Evidence: governance tests.
    • V7.17.f Property-test that every overlay composition is equal or tighter than floors and cannot relax protected-source/safety/legal/human authority. Evidence: tighten-only suite.
  • V7.18 Build UI-governed desk, topic, format, source, claim, risk, audience, territory, and catalog taxonomies mapped to IPTC vocabularies and code-owned floors. Evidence: versioned review, alias, deprecation, localization, and migration journey.
    • V7.18.a Define stable concept/code/set/version/label/definition/parent/ relation/alias/status/locale/IPTC mapping and code-owned floor schema. Evidence: taxonomy contracts.
    • V7.18.b Build Veritas taxonomy draft editing for add/edit/reparent/ relate/alias/merge/split/localize with cycle/duplicate/code/meaning validation. Evidence: editor tests.
    • V7.18.c Map to pinned IPTC Media Topics/code lists with exact/broader/ narrower/related/unmapped relation, provenance, and review. Evidence: mapping tests.
    • V7.18.d Preview impact across stories/sources/claims/search/analytics/ policies/routes/feeds/NewsML/targets before version activation. Evidence: impact analysis.
    • V7.18.e Govern review/approve/activate/supersede/deprecate with aliases/ replacements, localization completeness, immutable versions, and no floor weakening. Evidence: governance tests.
    • V7.18.f Migrate references/read models/indexes atomically or recoverably, preserve historical meaning, and reject orphan/unknown values. Evidence: migration journey.

V7A — Channel, audience, community, and business operations#

<!-- SPEC 2026-09-11: V7A cells build on apps/veritas/api routes that already exist (payments, subscriptions, ads, b2b-api, paywall, tips, ussd via libs/veritas/ussd, notifications, social, live-stream libs) — inventory them under V0.14 first and extend, never duplicate. Every provider-facing cell (WhatsApp, payment rails, social platforms, telecom USSD) needs live credentials that env-master does not carry; such a cell records the failed attempt and stays open (§2.5.6). Commercial routes are already co-located with editorial routes in one Hono app: V7A.10.a's separation is a real split, and V0.15 must name it. -->

  • V7A.1 Build newsletter edition authoring for sections, story selection, summaries, subject/preheader variants, sponsorship disclosure, personalization limits, preview/test send, audience/consent, schedule, delivery, bounce/complaint, archive, correction, and unsubscribe. Evidence: edition-to-story revision lineage and downstream correction.
    • V7A.1.a Define newsletter/publication/edition/section/item/variant/ audience/template/schedule/send/version/state contracts. Evidence: newsletter schemas.
    • V7A.1.b Build ordered section/story selection from exact eligible releases with semantic/non-drag controls, duplicate/conflict checks, and correction/retraction visibility. Evidence: edition editor.
    • V7A.1.c Author summaries/subject/preheader variants with factual comparison, locale/length/style rules, provenance, and human selection. Evidence: content tests.
    • V7A.1.d Insert sponsorship/advertising blocks with advertiser/campaign/ rights/targeting, mandatory disclosure/separation, and editorial independence. Evidence: sponsorship gate.
    • V7A.1.e Apply bounded versioned personalization only to approved selection/order/content fields with privacy floors and no fact/ correction/disclosure hiding. Evidence: personalization tests.
    • V7A.1.f Render production-equivalent web/email/client previews and authorized test sends with links/assets/UTM/accessibility/dark-mode/ plain-text validation. Evidence: preview tests.
    • V7A.1.g Resolve consented audience/segment/suppression/unsubscribe/ locale/frequency/quiet policy at send time without exporting protected newsroom data. Evidence: audience tests.
    • V7A.1.h Schedule and execute idempotent provider jobs with per-recipient privacy-safe delivery/bounce/complaint/suppression aggregates, retries, and read-back/archive. Evidence: send tests.
    • V7A.1.i Propagate story correction/retraction to edition archive, web copy, eligible follow-up notice, links/search, and provider state while preserving exact edition→story revision lineage. Evidence: correction E2E.
  • V7A.2 Build social-channel planning for platform-specific copy/media, account, schedule, cadence, rights, disclosure, approval, preview, publish, acknowledgement, moderation handoff, retry, edit/delete where supported, and correction. Evidence: no optimizer can change verified meaning or silently publish.
    • V7A.2.a Define per-platform/account capability/version/credential/ content-limit/media/rendition/link/schedule/edit-delete/moderation/ correction contract. Evidence: channel matrix.
    • V7A.2.b Create post plan pinned to story release/claims/media/locale/ target with account, objective, schedule/cadence, owner, and current correction state. Evidence: plan schema.
    • V7A.2.c Author platform copy/media/crops/captions/alt text/link/cards/ hashtags with factual-proposition and rights checks. Evidence: authoring tests.
    • V7A.2.d Apply sponsorship/AI/correction/graphic-content disclosures and safe areas/length/format constraints without silent truncation/drop. Evidence: eligibility tests.
    • V7A.2.e Preview exact target payload/rendering and require named human approval of immutable version; optimizers remain proposal-only. Evidence: approval tests.
    • V7A.2.f Schedule/publish through durable idempotent jobs with account/ rate/auth checks, acknowledgment/target ID/read-back, retry, and partial status. Evidence: publish tests.
    • V7A.2.g Hand off replies/comments/incidents to moderation with exact post/story context and no editorial-source leakage. Evidence: moderation handoff.
    • V7A.2.h Execute supported edit/delete/replace/correction or follow-up notice with prior history and read-back; keep unsupported targets unresolved. Evidence: correction E2E.
    • V7A.2.i Attempt optimizer/model/agent direct mutation/selection/ scheduling/publish and meaning-changing rewrite; prove rejection. Evidence: independence tests.
  • V7A.3 Build WhatsApp/channel ecosystem operations for subscription consent, segmentation, post preview, template/policy status, schedule, delivery, reply/escalation, unsubscribe, correction, and rate/quality limits. Evidence: current channel APIs and privacy rules.
    • V7A.3.a Pin current WhatsApp/other channel provider/API/template/policy/ pricing/quality/rate/region capabilities and responsible owner. Evidence: channel profiles.
    • V7A.3.b Implement explicit opt-in/source/purpose/channel/topic/locale/ time evidence, preference updates, opt-out/suppression, and consent expiry. Evidence: consent tests.
    • V7A.3.c Define privacy-preserving segments with minimum cohorts, allowlisted attributes, no protected-source/editorial-sensitive data, and reproducible membership snapshot. Evidence: segmentation tests.
    • V7A.3.d Author/select approved template and story/card/media/correction payload with locale, parameters, disclosure, rights, length, and policy status. Evidence: content tests.
    • V7A.3.e Preview exact recipient-visible message/conversation context, estimated audience/cost, quiet/frequency/rate/quality impact, and require approval. Evidence: preview tests.
    • V7A.3.f Schedule/send idempotently with consent/template/account/rate/ quality recheck, per-message status, provider acknowledgments, retries, and aggregate privacy. Evidence: delivery tests.
    • V7A.3.g Route replies by intent to support/moderation/complaint/tip/ correction workflows with safe identity/context and SLA; honor unsubscribe immediately. Evidence: reply tests.
    • V7A.3.h Propagate corrections/recalls with platform-supported update or follow-up, archive status/read-back, and unresolved recipient/target visibility. Evidence: correction E2E.
  • V7A.4 Build push, email, in-app, topic/author/source follow, breaking-news, and emergency-broadcast operations with audience, severity, locale, quiet-hours override authority, dedupe, preview, delivery, acknowledgement, correction, expiry, and recall. Evidence: emergency rights and false-alarm review are separate from ordinary campaigns.
    • V7A.4.a Define notification/campaign/event/delivery/ack/recall contracts and per-channel capabilities for push/email/in-app. Evidence: notification schemas.
    • V7A.4.b Implement consented topic/author/public-source follows and audience/locale/device/entitlement/preferences/suppression resolution with protected-source exclusion. Evidence: audience tests.
    • V7A.4.c Define ordinary/breaking/emergency severity, eligibility, evidence, templates, authority, target scope, expiry, and quiet-hours policy separately. Evidence: authority matrix.
    • V7A.4.d Build factual title/body/link/media/disclosure/correction content pinned to exact public release and preview each channel/device/ locale including accessibility. Evidence: preview tests.
    • V7A.4.e Deduplicate by event/story/claim/target/audience/window and apply frequency/quiet limits unless a specifically authorized emergency override exists. Evidence: dedupe/policy tests.
    • V7A.4.f Dispatch durable per-channel jobs with current consent/authority/ release checks, provider acknowledgment, delivery/open/ack privacy floors, retry, and partial state. Evidence: delivery tests.
    • V7A.4.g Correct/expire/recall notifications through target capability, in-app persistent notice, links/read-back, and no false success for delivered immutable messages. Evidence: remedy tests.
    • V7A.4.h Run emergency activation→broadcast→acknowledgment→recall/ correction→deactivation plus mandatory false-alarm/post-event review, separately from campaigns. Evidence: emergency E2E.
  • V7A.5 Build community/comment moderation for thread state, identity/ pseudonymity, spam/abuse, misinformation links, source protection, appeals, rate limits, moderator notes, legal hold, transparency, and article correction synchronization. Evidence: community content is never presented as newsroom-attested fact.
    • V7A.5.a Define community/thread/comment/reaction/report/decision/appeal/ moderator-note/version/state schemas separately from editorial claims. Evidence: community contracts.
    • V7A.5.b Define identity/pseudonymity/verification/consent/age/privacy/ block/mute protections and prevent correlation with protected sources/ tips. Evidence: identity tests.
    • V7A.5.c Implement submit/edit/delete/reply/report with rate/spam/bot/ abuse/link/media scanning, quarantine, duplicate, and safe rendering. Evidence: intake tests.
    • V7A.5.d Run human-governed moderation queues for harassment/hate/threat/ doxxing/misinformation/spam/legal/safety with evidence, uncertainty, consistency, and emergency escalation. Evidence: moderation workflow.
    • V7A.5.e Record allow/limit/label/remove/lock/suspend/escalate decisions, reason/policy version/actor/expiry, protected moderator notes, and transparent user notice. Evidence: decision tests.
    • V7A.5.f Implement appeal intake/independent review/reversal/confirmation/ notice and preserve immutable prior decisions. Evidence: appeals E2E.
    • V7A.5.g Apply retention/deletion/legal hold/audit/export and moderator access controls without leaking reporter/source/legal data. Evidence: lifecycle tests.
    • V7A.5.h Synchronize story correction/retraction context/links/thread state while never labeling community content as verified newsroom fact. Evidence: correction/parity tests.
  • V7A.6 Build reader feedback, complaints, right-of-reply, tip conversion, fact-check requests, and correction requests as distinct queues with routing, evidence, SLA, privacy, status, response, and audit. Evidence: external submissions cannot directly mutate editorial records.
    • V7A.6.a Define distinct typed intake contracts, required/optional identity/consent/story/claim/evidence fields, and safe anonymous options. Evidence: queue schemas.
    • V7A.6.b Scan/quarantine attachments, validate target/version/link, apply rate/abuse protection, and prevent source/reporter existence leakage. Evidence: intake security tests.
    • V7A.6.c Route by type/rule/jurisdiction/desk/risk/locale to named owner, independent queue, SLA/calendar, escalation, and protected compartment. Evidence: routing tests.
    • V7A.6.d Track received/acknowledged/triaged/investigating/waiting/ decided/remedied/appealed/closed with privacy-safe external status. Evidence: state machine.
    • V7A.6.e Preserve submitted evidence/communications/decisions/responses/ deadlines/audit and apply minimization/retention/legal hold. Evidence: record tests.
    • V7A.6.f Convert only through explicit authorized handoff to anonymous tip, fact-check assignment, NMC/legal/right-of-reply, or correction workflow with lineage. Evidence: handoff tests.
    • V7A.6.g Respond/notify with reviewed redacted outcome and appeal path; prevent external submitters or intake workers from editing/publishing. Evidence: mutation boundary tests.
  • V7A.7 Build multilingual production across the ratified Ghanaian and other locales for translation, transliteration, terminology, names, audio/voice, captions, media text, fact parity, cultural review, publication, and correction cascade. Evidence: each locale locks exact source-story and evidence versions.
    • V7A.7.a Ratify supported locale/script/region/fallback/direction/ orthography/number/date/name/pronunciation/voice/channel capability matrix with owners. Evidence: locale profile.
    • V7A.7.b Create locale variant pinned to exact source story/claims/ citations/evidence/media/disclosures/terminology/policy versions. Evidence: locale manifest.
    • V7A.7.c Build translation/transliteration editor with segment/semantic anchors, protected terms, names/entities, citations, uncertainty, and per-change human/AI provenance. Evidence: editor tests.
    • V7A.7.d Localize images/graphics/tables/maps/alt text/transcripts/ captions/audio description/voice/music disclosures and generate target renditions. Evidence: media localization tests.
    • V7A.7.e Run claim/number/date/entity/quote/citation/uncertainty/ correction/disclosure semantic parity plus terminology consistency. Evidence: fact parity gate.
    • V7A.7.f Run independent qualified language/cultural/editorial/ accessibility review on exact locale revision with sensitive-context handling. Evidence: review signoff.
    • V7A.7.g Preview/publish/read back each locale/target independently with canonical/hreflang or equivalent links, rights, fallback, and no silent source-locale substitution. Evidence: release tests.
    • V7A.7.h Propagate source-story correction/retraction to every locale, invalidate approval, translate/review notices, and keep unresolved locales visible. Evidence: cascade E2E.
  • V7A.8 Build B2B feed, search, fact-check, entity, media, monitoring, and alert product operations for API products, clients, keys, scopes, quotas, webhooks, schemas, sandbox, usage, billing, deprecation, revocation, and correction/retraction events. Evidence: developer SDK conformance and no protected/unlicensed field exposure.
    • V7A.8.a Define each API product/resource/action/version/audience/data/ rights/freshness/SLO/pricing capability and explicit protected/ unlicensed exclusions. Evidence: product catalog.
    • V7A.8.b Build client/application/service-account onboarding with owner/ purpose/contract/region/entitlements/environments/approval/review/ suspension state. Evidence: client lifecycle.
    • V7A.8.c Issue hashed/rotatable scoped keys or tokens with least privilege, tenant/product/environment/IP or mTLS constraints, expiry, last-used, and emergency revoke. Evidence: credential tests.
    • V7A.8.d Implement versioned feed/search/fact-check/entity/media/ monitoring/alert schemas with stable IDs, pagination, field projection, attribution, rights, and current correction state. Evidence: API tests.
    • V7A.8.e Enforce per-client/product/operation quota/rate/concurrency/ payload/retention/export policy with consistent headers/errors and safe analytics. Evidence: limit tests.
    • V7A.8.f Implement signed webhooks with subscription filters, secret rotation, replay protection, idempotent delivery/retry/dead letter, and correction/retraction events. Evidence: webhook tests.
    • V7A.8.g Provide isolated synthetic sandbox, versioned OpenAPI/SDK/ examples/conformance, changelog/deprecation/migration, and no production protected data. Evidence: developer experience tests.
    • V7A.8.h Meter usage/billing/refunds/disputes separately from editorial ranking/access decisions and reconcile without placing payment data in newsroom content stores. Evidence: commerce tests.
    • V7A.8.i Run two-client scope/tenant/BOLA/filter/snippet/media URL/export/ log/cache/schema/version/revocation tests and prove no protected/ unlicensed exposure. Evidence: security/SDK conformance gate.
  • V7A.9 Build USSD and voice-callback content operations for menu/routing, concise story/fact-check variants, language, consent, telecom status, session limits, accessibility, fallback, charge disclosure, emergency, correction, and analytics minimization. Evidence: feature-phone content retains canonical story/claim identity and current correction state.
    • V7A.9.a Define USSD menu/session/step/content/action and voice callback/ script/recording/call state contracts with provider/network/version. Evidence: channel schemas.
    • V7A.9.b Author concise locale-specific story/fact-check/correction/ emergency variants pinned to canonical release/claim/evidence and human-reviewed for meaning parity. Evidence: content tests.
    • V7A.9.c Build bounded numbered menu/routing/back/repeat/more/language/ help/exit flows with session timeout/continuation and no inaccessible free-form dependency. Evidence: USSD journey.
    • V7A.9.d Build consented voice callback request/schedule/attempt/retry/ answer/opt-out with reviewed voice/pronunciation/speed and keypad alternatives. Evidence: voice journey.
    • V7A.9.e Integrate current telecom template/short code/caller ID/routing/ delivery/quality/rate/status/callback capabilities with explicit unavailable state. Evidence: provider tests.
    • V7A.9.f Disclose carrier/service charges, consent/purpose/privacy, emergency authority, accessibility/fallback, and support before commitment. Evidence: policy tests.
    • V7A.9.g Propagate correction/retraction/emergency recall by canonical ID to menu/voice content and prevent stale cached variants. Evidence: update tests.
    • V7A.9.h Minimize phone/session/location/content analytics, aggregate with privacy floors, retain briefly, and exclude protected source/editorial scoring. Evidence: privacy tests.
    • V7A.9.i Test session limits/drop/retry/duplicate/locale fallback/ provider outage/role/consent/current-correction state on representative devices. Evidence: feature-phone E2E.
  • V7A.10 Build subscription, payment, tipping, advertising, sponsorship, partnership, and entitlement administration in a separate commercial surface with newsroom-visible conflict/disclosure projections only. Evidence: commercial actors cannot suppress, prioritize, edit, correct, or retract editorial content, and payment data never enters drafts.
    • V7A.10.a Define separate commercial host/service/store/roles/routes/ credentials/audit and explicit prohibited newsroom capabilities. Evidence: boundary architecture.
    • V7A.10.b Build product/plan/subscription/entitlement/trial/cancel/pause/ renew/refund/support state with current terms/consent and provider reconciliation. Evidence: subscription tests.
    • V7A.10.c Build payment/tip intent/authorization/capture/failure/refund/ dispute/payout/receipt/tax handling using tokenized provider data and idempotency. Evidence: payment tests.
    • V7A.10.d Build advertiser/campaign/inventory/creative/targeting/ schedule/delivery/billing controls with privacy floors and no sensitive editorial/source targeting. Evidence: advertising tests.
    • V7A.10.e Build sponsorship/partnership contracts, deliverables, rights, conflicts, separation, disclosures, expiry, and audited exceptions. Evidence: partnership tests.
    • V7A.10.f Project only advertiser/sponsor/partner/conflict/disclosure/ entitlement eligibility into newsroom workflows through typed redacted read models. Evidence: projection tests.
    • V7A.10.g Enforce newsroom independence server-side: commercial actors/ metrics cannot assign, rank, suppress, edit, approve, publish, correct, retract, or alter gates. Evidence: negative authorization suite.
    • V7A.10.h Keep card/bank/mobile-money/payment/contact/billing data out of drafts/claims/search/logs/analytics/model context and inspect storage/ network. Evidence: data-boundary tests.
    • V7A.10.i Reconcile entitlement/revenue/disclosure state across systems, handle provider outage/webhook replay/version skew, and preserve editorial access to required corrections. Evidence: commercial E2E.
  • V7A.11 Build content, audience, search, recommendation, social, channel, subscription, revenue, correction, and public-interest analytics with metric definitions, cohort/privacy floors, bot filtering, attribution, latency/freshness, experiment governance, and export. Evidence: analytics cannot expose protected sources or become an automatic editorial score.
    • V7A.11.a Define versioned metric/event dimensions, numerator/denominator/ units/windows/attribution/source/owner/purpose for each named domain. Evidence: metric catalog.
    • V7A.11.b Instrument allowlisted pseudonymous events with schema/consent/ purpose/data-class/retention and prohibit protected source/tip/legal/ draft/payment content. Evidence: event contracts.
    • V7A.11.c Filter bots/fraud/internal/testing/retries and deduplicate/ sessionize/attribute across channels with documented uncertainty. Evidence: pipeline tests.
    • V7A.11.d Enforce minimum cohort/noise/suppression/access/field/row/ purpose/region/export controls and deletion/opt-out propagation. Evidence: privacy tests.
    • V7A.11.e Show ingestion/processing/index/report latency, freshness/ completeness/unknown state, corrections to metrics, and data-quality incidents. Evidence: observability tests.
    • V7A.11.f Govern experiments with hypothesis/metrics/guardrails/audience/ consent/randomization/power/duration/stop/review and no high-risk editorial manipulation. Evidence: experiment policy.
    • V7A.11.g Provide dashboards and purpose-bound audited exports with semantic labels/units/uncertainty/accessibility and safe small-cell behavior. Evidence: analytics UX/export tests.
    • V7A.11.h Block use of engagement/revenue/audience/source metrics as automatic story/reporter/source truth, quality, publish, correction, or suppression score. Evidence: editorial independence tests.
    • V7A.11.i Verify known-value attribution/freshness/bot/privacy/deletion/ cross-tenant scenarios and reconcile to source systems. Evidence: analytics validation.
  • V7A.12 Build SEO, search/recommendation, topic/entity pages, related coverage, archive, reading-list/follow, and content-discovery controls through versioned editorial policies. Evidence: ranking/metadata changes cannot fabricate facts, hide active corrections, or break canonical URLs.
    • V7A.12.a Define canonical public content/search/discovery document with stable URL/ID/version/locale/topic/entity/rights/release/correction/ retraction/access fields. Evidence: discovery contract.
    • V7A.12.b Build versioned SEO metadata/structured data/sitemap/canonical/ hreflang/robots/indexing policies derived from exact releases. Evidence: SEO tests.
    • V7A.12.c Build lexical/semantic search and topic/entity/archive pages with explainable facets/sorts/current state/accessibility and no stale correction hiding. Evidence: discovery tests.
    • V7A.12.d Build related/recommendation candidate/filter/rank/diversity/ freshness/public-interest rules with model/features/policy attribution and editorial controls. Evidence: recommendation tests.
    • V7A.12.e Build consented reading-list/follow/save with privacy, entitlement, deletion/export, current correction/retraction notices, and no protected source following. Evidence: reader tests.
    • V7A.12.f Govern policy draft/review/approve/experiment/activate/rollback/ expire with impact preview and code-owned fact/correction/rights floors. Evidence: governance tests.
    • V7A.12.g Reindex/invalidate caches on publish/update/correction/ retraction/rights/URL changes and preserve redirects/history/canonical identity. Evidence: lifecycle tests.
    • V7A.12.h Adversarially test fabricated metadata/snippets, correction demotion/hiding, broken canonical loops, filter bubbles, protected data, and ranking direct-write attempts. Evidence: independence gate.
  • V7A.13 Build live-stream/channel operations for stream health, encoder/ platform status, concurrent destinations, delay, moderation, Q&A, recording/clips, rights, incident, failover, and post-event archive. Evidence: operator controls are step-up authorized and editorial claims remain traceable.
    • V7A.13.a Define live channel/session/destination/encoder/feed/status/ incident/action/recording contracts and exact program/story/rundown revision. Evidence: operations schema.
    • V7A.13.b Integrate encoder/input/output/provider health, bitrate/frame/ audio/caption/delay/drop/keyframe metrics, heartbeat, and calibrated alerts. Evidence: health tests.
    • V7A.13.c Configure concurrent destinations with credentials/capability/ rights/locale/schedule/acknowledgment and independent partial state. Evidence: destination tests.
    • V7A.13.d Implement preview/program/delay/fact-check hold/caption/ translation/moderation/Q&A controls linked to rundown claims/evidence. Evidence: live control tests.
    • V7A.13.e Require step-up/current role/state/impact/confirmation for go-live, route change, delay dump, disconnect, emergency stop, failover, and restart. Evidence: authorization tests.
    • V7A.13.f Open incident automatically/manually with synchronized telemetry/actions/decisions/comms/evidence and safe escalation/ postmortem. Evidence: incident tests.
    • V7A.13.g Record immutable program/clean/caption/transcript feeds and create reviewed rights-cleared clips with timecoded claim/source lineage. Evidence: recording tests.
    • V7A.13.h Build post-event archive/metadata/accessibility/rights/ correction-retraction replacement and per-destination read-back. Evidence: archive E2E.
  • V7A.14 Build governed training/evaluation data intake from public and licensed corpora, newsroom decisions, corrections, fact-check outcomes, language review, channel feedback, and model disagreements. Evidence: source protection, consent/license, purpose, minimization, retention, deletion, contamination, split integrity, provenance, opt-out, and editorial-independence controls precede any training or fine-tuning.
    • V7A.14.a Define dataset/use-case/version/owner/model-purpose/allowed- prohibited content/fields/subjects/sources/licence/consent/retention contract. Evidence: dataset policy.
    • V7A.14.b Intake only approved public/licensed/newsroom-decision/ correction/fact-check/language/channel-feedback sources through quarantine and immutable provenance. Evidence: source ledger.
    • V7A.14.c Enforce protected source/tip/legal/embargoed/private draft/ payment/child-sensitive/opted-out exclusions before extraction, annotation, export, or provider access. Evidence: exclusion tests.
    • V7A.14.d Validate licence/consent/purpose/territory/provider/model-use/ derivative/attribution/deletion rights per item and dataset version. Evidence: rights gate.
    • V7A.14.e Minimize/redact/pseudonymize/deduplicate, detect secrets/PII/ prohibited content, and retain transformations with hashes. Evidence: preprocessing tests.
    • V7A.14.f Detect cross-source contamination, benchmark memorization, near-duplicate train/eval/test leakage, temporal/entity/topic leakage, label disagreement, and poisoning. Evidence: integrity tests.
    • V7A.14.g Define annotation/adjudication/quality/subgroup/locale/split/ version/reproducibility and prevent production feedback from silently entering training. Evidence: curation tests.
    • V7A.14.h Implement data subject/source/partner opt-out, deletion/ restriction/licence expiry, derived dataset/model impact trace, and legal-hold conflict review. Evidence: deletion tests.
    • V7A.14.i Require privacy/security/legal/source-protection/editorial/ model-governance approval with exact manifest, risks, provider terms, and expiry before use. Evidence: approval record.
    • V7A.14.j Prohibit model outputs/feedback/performance from automatically changing editorial policy, source scores, corrections, assignments, or publish decisions. Evidence: independence tests.
    • V7A.14.k Reproduce dataset build from immutable inputs/code/config and publish lineage, exclusions, counts, quality, split hashes, access, and current disposition. Evidence: reproducibility gate.

V8 — Newsroom UI, accessibility, and resilience#

  • V8.1 Build the newsroom hub around assignments, embargoes, gate failures, fact-check queue, corrections, and publishing state. Evidence: utility-first design review.
    • V8.1.a Define primary triage questions and live view models for each named queue/state with owner/desk/priority/deadline/severity/freshness/ blocker/eligible action. Evidence: hub contract.
    • V8.1.b Compose a restrained task-led hierarchy with urgent actionable work first, production context second, no decorative hero or repeated card mosaic. Evidence: design review.
    • V8.1.c Implement stable filters/sorts/grouping/search/URLs and live updates while redacting protected/embargo/legal details. Evidence: hub query tests.
    • V8.1.d Implement authorized claim/open/assign/escalate/remediate/retry/ hold/review actions with state/version/reason/impact/confirmation. Evidence: action tests.
    • V8.1.e Render Veritas route loading/empty/stale/offline/degraded/partial/ unauthorized/fixture/error honestly with accessible recovery. Evidence: state tests.
    • V8.1.f Verify wide/compact/single-column, keyboard/screen reader, zoom/ reflow, scale, live updates, and no sensitive count/snippet leaks. Evidence: utility-first E2E.
  • V8.2 Build inbox, pitch board, story workspace, source registry, claim review, calendar, publishing, corrections, and archive routes. Evidence: route integrity. <!-- SPEC 2026-09-11: no route tree exists (Phase V header note). Register routes only after V0.1.d's host ADR is accepted; before that, components and read models may land behind the S7 shell primitives with no page. Do not put newsroom routes in the consumer room or the public reader. -->
    • V8.2.a Register each named route with canonical path/label/keywords/ host/owner/audience/roles/capabilities/data class and project/story context. Evidence: route registry.
    • V8.2.b Define route-specific loaders/actions/read models/deep-link state, breadcrumbs/back paths, and safe unavailable/degraded states. Evidence: route contracts.
    • V8.2.c Reuse existing Studio/global/newsroom shell, identity/search/ notifications/help/command palette, and remove duplicate navigation or mutations. Evidence: shell review.
    • V8.2.d Enforce direct server authorization and protected/public/ commercial/operator boundary for every route/field/function. Evidence: route auth tests.
    • V8.2.e Test navigation/direct URL/refresh/back/forward/story switch/role revoke/not-found/archived/moved context across web/desktop. Evidence: route integrity E2E.
  • V8.3 Use primary editor plus evidence inspector rather than nested card dashboards. Evidence: frontend conformance review.
    • V8.3.a Define shared story workspace shell with primary structured editor, contextual evidence inspector, story navigation, status rail, selection/focus/dirty/capability state. Evidence: workspace spec.
    • V8.3.b Compose source/claim/citation/evidence/attestation/media inspectors synchronized to selection without nested global shells or duplicate canonical data. Evidence: inspector tests.
    • V8.3.c Keep gate/review/embargo/save/job/publish/correction status and blockers/remediation visible without obscuring the article. Evidence: status tests.
    • V8.3.d Use typography, spacing, dividers, disclosure, and progressive panes for hierarchy; avoid decorative card grids, gradients, oversized titles, and controls in arbitrary containers. Evidence: UI review.
    • V8.3.e Define wide/compact/single-column layouts with inspector/status drawer/tab, stable story identity, safe focus, and no squashed precision editor. Evidence: responsive snapshots.
    • V8.3.f Test long story/claims/sources/live collaboration/protected redaction/zoom/keyboard/screen reader. Evidence: frontend conformance E2E.
  • V8.4 Preserve story, revision, block, selection, tab, filters, and inspector in safe deep links. Evidence: protected links expire and do not reveal source identities.
    • V8.4.a Define versioned allowlisted URL state schema for route/story/ revision/block/range/tab/filter/sort/inspector/compare/return with size and sensitivity limits. Evidence: link schema.
    • V8.4.b Encode stable opaque IDs and bounded range/filter values while excluding source identity/contact, legal notes, embargo secrets, excerpts, unsaved text, local paths, and tokens. Evidence: serialization tests.
    • V8.4.c Parse/validate/canonicalize and reauthorize every referenced object/field/action; handle stale/deleted/moved/retracted/revoked subjects. Evidence: restoration tests.
    • V8.4.d Use expiring signed capability/deep links only where ordinary route authorization is insufficient, with purpose/audience/single-use/ revocation and no secret in URL. Evidence: protected-link tests.
    • V8.4.e Restore refresh/back/forward/share/new window/web-desktop context and reconcile unsaved/conflicting state safely. Evidence: navigation tests.
    • V8.4.f Test cross-tenant/story/compartment substitution, range/ID/ encoding/oversize/replay/expiry/log/referrer/history leakage. Evidence: adversarial link suite.
  • V8.5 Provide keyboard-first editing, claim attachment, source search, review, and publish journeys. Evidence: pointer-free Playwright.
    • V8.5.a Inventory every focusable/editor/selection/menu/dialog/drag/drop/ resize/reorder/compare/review/publish operation and expected keyboard outcome. Evidence: interaction inventory.
    • V8.5.b Implement structured-editor navigation/selection/block commands/ formatting/undo/redo/anchors without trapping standard text or assistive technology keys. Evidence: editor keyboard tests.
    • V8.5.c Implement keyboard claim creation and source search/select/ locator/citation/counter-source attachment with context/focus return. Evidence: grounding journey.
    • V8.5.d Implement review comment/suggestion/evidence inspection/checklist/ decision and non-drag reorder/relationship alternatives. Evidence: review journey.
    • V8.5.e Implement preview/eligibility/signoff/schedule/publish/read-back/ correction controls with state/impact/confirmation and focus recovery. Evidence: publication journey.
    • V8.5.f Run pointer-free Playwright across wide/compact/mobile review, web/desktop, dialogs/errors/conflicts/live updates, and role changes. Evidence: keyboard E2E.
  • V8.6 Provide semantic alternatives for evidence graphs, timelines, analytics, media, and compare views. Evidence: screen-reader tests.
    • V8.6.a Define content/operation/selection/focus/status parity contract and synchronized state between each visual surface and its semantic alternative. Evidence: parity matrix.
    • V8.6.b Provide evidence graph table/tree/path/forms for entities/claims/ sources/edges/types/confidence/provenance/retractions and all authorized edits. Evidence: graph tests.
    • V8.6.c Provide timeline ordered/grouped table/forms with exact/ approximate/uncertain time, events/participants/location/evidence/ conflicts, and edit/navigation. Evidence: timeline tests.
    • V8.6.d Provide analytics tables/text summaries with metric definitions, numerator/denominator/units/uncertainty/freshness/cohort privacy and accessible chart descriptions. Evidence: analytics tests.
    • V8.6.e Provide media native controls/transcripts/captions/descriptions/ chapters/timecode/annotation lists and synchronized playback position. Evidence: media tests.
    • V8.6.f Provide compare semantic diff by block/range/claim/citation/ media/disclosure with before/after/context and decision operations. Evidence: compare tests.
    • V8.6.g Verify keyboard/screen-reader/zoom/search/filter/edit/review parity on large data without protected redaction divergence. Evidence: semantic E2E.
  • V8.7 Provide status announcements for save, conflict, source update, gate, review, embargo, publish, correction, and cascade progress. Evidence: announcement assertions.
    • V8.7.a Define event-to-message taxonomy with priority/politeness, dedupe/coalescing, context, units/progress, terminal state, and privacy redaction. Evidence: announcement spec.
    • V8.7.b Announce dirty/saving/saved/offline/error/conflict and semantic conflict resolution without noisy autosave repetition. Evidence: save tests.
    • V8.7.c Announce source change/retraction, gate/review status/blocker/ staleness, embargo time/state, and required remediation. Evidence: editorial status tests.
    • V8.7.d Announce per-target queued/running/published/verified/partial/ failed and correction/cascade progress with counts and unresolved targets. Evidence: release tests.
    • V8.7.e Preserve current focus/typing; pair announcements with persistent visible status/history and avoid exposing protected data. Evidence: accessibility/privacy tests.
    • V8.7.f Assert exact announcement order/count/content for success, retries, live updates, reconnect, duplicates, errors, and cancellation. Evidence: announcement suite.
  • V8.8 Implement autosave/offline/conflict recovery without caching protected-source details in unsafe storage. Evidence: storage inspection.
    • V8.8.a Define eligible offline content and local operation/draft/ checkpoint/schema/device/session/encryption/quota/expiry/access state; explicitly exclude protected identities/legal secrets. Evidence: cache policy.
    • V8.8.b Implement debounced idempotent autosave with expected revision and visible dirty/saving/saved/error/offline/conflict state. Evidence: autosave tests.
    • V8.8.c Store eligible drafts/operations/assets encrypted and partitioned by tenant/user/story with logout/revocation/archive/retention cleanup. Evidence: storage tests.
    • V8.8.d Render protected source references only as safe opaque/redacted placeholders offline and require online step-up to reveal. Evidence: protected-offline tests.
    • V8.8.e Reauthenticate/re-authorize/rebase/deduplicate on reconnect and present mine/theirs/base semantic conflict with copy/export safeguards. Evidence: recovery tests.
    • V8.8.f Inspect browser/desktop storage/cache/index/telemetry/crash/ backups, then test crash/relaunch/two-client/role revoke/quota/schema update. Evidence: storage inspection E2E.
  • V8.9 Implement protected-field copy/download/print controls. Evidence: policy and audit tests.
    • V8.9.a Inventory protected fields/surfaces/actions and resolve current role/grant/purpose/device/session/story/compartment/policy constraints. Evidence: control matrix.
    • V8.9.b Prevent selection/copy/cut/clipboard where policy requires, sanitize rich/plain representations, and avoid leaking through hidden/ semantic DOM. Evidence: clipboard tests.
    • V8.9.c Gate downloads/exports/attachments/originals/proxies with step-up, scope/format/redaction/watermark/expiry/rate and signed authorization. Evidence: download tests.
    • V8.9.d Provide print-safe redacted layout or block printing, including browser/system preview, PDF, background, headers/URLs, and screenshots limitation disclosure. Evidence: print tests.
    • V8.9.e Handle drag/drop/share/open-external/devtools/accessibility/ cache/log/crash edge paths without pretending client controls prevent all exfiltration. Evidence: threat tests.
    • V8.9.f Audit attempted/allowed copy/download/print with redacted subject/ purpose/result and test revoked/stale/cross-role grants. Evidence: policy and audit suite.
  • V8.10 Test 200%/400% zoom, long stories, long locale, RTL, high contrast, theme, density, and reduced motion. Evidence: visual/a11y matrix.
    • V8.10.a Define browser/app/text zoom, viewport, story/source/claim volume, locale/script expansion, direction, theme, contrast, density, and motion test matrix. Evidence: coverage matrix.
    • V8.10.b Test 200% and 400% reflow across hub/editor/inspector/dialog/ calendar/graph/timeline/preview with no clipped/hidden/unreachable controls. Evidence: zoom tests.
    • V8.10.c Test very long stories/blocks/headlines/source names/URLs/ comments/errors/tables/media transcripts with virtualization and stable editing. Evidence: content stress tests.
    • V8.10.d Test long/CJK/complex-script/RTL/mixed-bidi names/quotes/numbers/ locators/timelines/forms/navigation and isolation against spoofing. Evidence: locale tests.
    • V8.10.e Test light/dark/system/high-contrast and compact/comfortable/ touch density for contrast/focus/status/selection/target size and persistent preferences. Evidence: theme/density tests.
    • V8.10.f Honor reduced motion before first paint; remove parallax/spring/ large translation/auto pan while preserving textual progress/playback state. Evidence: motion tests.
    • V8.10.g Run visual regression plus automated/manual accessibility on supported web/desktop/viewports and close all blocking findings. Evidence: visual/a11y report.
  • V8.11 Define performance budgets for editor, evidence sidebar, large source registry, claim graph, search, compare, and cascade. Evidence: profiling.
    • V8.11.a Define supported device/browser/desktop/network/data-scale/ concurrency scenarios and shared load/interaction/save/error/resource metrics. Evidence: performance profile.
    • V8.11.b Define editor open/type/input latency/autosave/diff/history/ collaboration budgets for long structured stories. Evidence: editor budgets.
    • V8.11.c Define evidence sidebar and source registry query/facet/page/ selection/update/render/memory budgets with protected authorization overhead. Evidence: evidence budgets.
    • V8.11.d Define claim graph retrieval/layout/semantic view/edit/ validation and search indexing/query/suggestion/update visibility budgets. Evidence: graph/search budgets.
    • V8.11.e Define compare computation/render/navigation and correction/ retraction cascade plan/progress/read-back/reconciliation budgets. Evidence: workflow budgets.
    • V8.11.f Profile production-scale fixtures with client/server traces/ query plans/CPU/memory/network, retain baselines, and fail regressions. Evidence: profiling report.
  • V8.12 Build honest fixture, stale, offline, degraded, not-configured, unauthorized, partial, and failed states. Evidence: state matrix.
    • V8.12.a Define canonical data-origin/service-state discriminators, timestamps/reasons/safe messages/eligible actions/recovery/telemetry/ announcements. Evidence: state contracts.
    • V8.12.b Label fixture/demo/synthetic content persistently and prevent publishing, source-quality use, external delivery, or production promotion. Evidence: fixture tests.
    • V8.12.c Show stale subject/source/evidence/gate/review/candidate with exact invalidating version and block unsafe continuation. Evidence: stale tests.
    • V8.12.d Show offline cached versus unavailable, queued/prohibited edits, freshness/reconnect and protected fields not cached. Evidence: offline tests.
    • V8.12.e Show per-capability degraded/not-configured/unauthorized/partial/ failed/unknown with completed and unresolved suboperations; no simulated success. Evidence: service tests.
    • V8.12.f Exercise every state/recovery on critical routes and web/ desktop/responsive projections with identical semantics and privacy. Evidence: state matrix E2E.
  • V8.13 Build responsive reporter/reviewer mode while keeping protected controls safe. Evidence: mobile/tablet auth and workflow tests.
    • V8.13.a Define mobile/tablet reporter/reviewer versus precision-edit capability matrix, visible data, actions, step-up, device trust, and desktop handoff. Evidence: responsive spec.
    • V8.13.b Build assignment/pitch/story/source-redacted/claim/evidence/gate/ review/calendar/publish/correction status projections with stable identity. Evidence: responsive views.
    • V8.13.c Build capture/note/photo/audio/upload/comment/request-change/ approve/reject/escalate/hold/cancel actions only where role/state/device policy permits. Evidence: action tests.
    • V8.13.d Present graph/timeline/compare/analytics/media semantic views and signed open-on-desktop handoff rather than unusable precision editors. Evidence: semantic/handoff tests.
    • V8.13.e Require step-up/current grant for protected source/legal/ publish/recall/export controls and prevent unsafe offline cache/copy/ screenshot claims. Evidence: protected control tests.
    • V8.13.f Test touch/keyboard/screen reader/zoom/orientation/safe area/ slow-offline/reconnect/direct authorization/role revoke. Evidence: mobile/tablet E2E.
  • V8.14 Register permission-filtered Studio navigation and command palette. Evidence: route and role parity. <!-- SPEC 2026-09-11: registration goes through the same IA registry Metis used (M9.14 shape: authorize decides, visible_to is written in terms of it, mutating palette commands are unconstructible), and must respect release-scope.ts — a Veritas entry that appears in a V1.0 shell is a violation the route-integrity gate fails on. -->
    • V8.14.a Build canonical route/action registry with owner/host/path/ labels/keywords/project-story context/roles/capabilities/data class and availability. Evidence: navigation manifest.
    • V8.14.b Register authorized Veritas hub and project/story entries in the existing Studio shell without a second global nav or duplicate action. Evidence: shell tests.
    • V8.14.c Filter navigation/palette results before display by current tenant/desk/story/role/compartment/entitlement/capability and avoid count/ label existence leaks. Evidence: visibility tests.
    • V8.14.d Preserve context/return path/dirty warnings and run server authorization on direct route/command execution. Evidence: action tests.
    • V8.14.e Compare expected route/role visibility and command execution across web/desktop/project/story/protected states. Evidence: parity matrix.
    • V8.14.f Fail CI on authorized orphan, reachable unauthorized item, duplicate registry/mutation, stale alias, or unowned route. Evidence: reachability gate.
  • V8.15 Add permission-filtered routes or governed deep links for ingestion, tips/investigations, entity graph, multimedia/live, newsletters/social/ messaging, localization, compliance, community/complaints, B2B/USSD, commercial projections, and analytics. Evidence: every retained feature is reachable by its authorized audience and no protected surface is reachable through reader navigation.
    • V8.15.a Reconcile retained atomic capability ledger to canonical host/ route/role/audience/project/entitlement/capability and required contextual prerequisites. Evidence: reachability manifest.
    • V8.15.b Register newsroom routes for ingestion, tips/investigations, entity graph, multimedia/live, localization, and compliance with compartment-aware authorization. Evidence: editorial route tests.
    • V8.15.c Register channel/community routes for newsletters/social/ messaging/community/complaints and typed redacted newsroom handoffs. Evidence: channel route tests.
    • V8.15.d Register separately hosted B2B/USSD/commercial/analytics routes or governed deep links with current audience/session/scope and safe return. Evidence: external surface tests.
    • V8.15.e Hide protected newsroom capabilities from reader/public/ commercial navigation and reject direct URLs/commands/fields/functions server-side. Evidence: boundary tests.
    • V8.15.f Exercise positive authorized journey for every retained capability and every role/surface plus unauthorized cross-surface attempts. Evidence: route coverage.
    • V8.15.g Fail reconciliation on unreachable authorized feature, exposed protected surface, unknown/duplicate route, or handoff without redaction/ return contract. Evidence: completeness gate.

V9 — Veritas verification, operations, and launch#

<!-- SPEC 2026-09-11: V9 is bound by §2.5 exactly as M10 was: hosts named per stage (there is none until V0.1.d), no page.route in journeys, real Postgres created for VERITAS_DATABASE_URL plus MinIO/Redis/Elasticsearch from docker/docker-compose.dev.yml (the search profile), model binding per the Phase V header note (5), human cells packet-only. The 19 apps/veritas/api tests mock the pg pool: they are contract tests, never integration evidence. next build && next start, chromium only, memory checkpoint before every sweep. -->

  • V9.1 Run canonical contract, editorial state, source-quality, counterclaim, attestation, correction, and retraction suites. Evidence: current-branch reports.
    • V9.1.a Validate current/historical story/source/claim/evidence/ attestation/correction/retraction schemas and version compatibility. Evidence: contract report.
    • V9.1.b Property-test canonical editorial transitions, orthogonal embargo/release/remedy states, invalid transitions, concurrency, and audit. Evidence: state report.
    • V9.1.c Run hand-calculated source-quality factors/weights/adjustments/ bands and missing/unknown/domain cases. Evidence: scoring report.
    • V9.1.d Run all four counterclaim results across independence/quality/ consensus/harm/false-balance and contested states. Evidence: balance report.
    • V9.1.e Run attestation scope/credential/conflict/disagreement/quorum/ expiry/automation-impersonation invariants. Evidence: attestation report.
    • V9.1.f Run chronological prior-text correction and complete ordered retraction graph/action/read-back/closure invariants. Evidence: remedy report.
    • V9.1.g Pin reports to current branch commit/contracts/policies/corpora/ engines and fail any skipped/unreviewed canonical invariant. Evidence: suite manifest.
  • V9.2 Run known-value tests for claim, citation, statistical, quote, headline, source, balance, evidence, and cascade gates. Evidence: gate registry coverage.
    • V9.2.a Freeze expert-reviewed positive/boundary/negative/adversarial/ unavailable/indeterminate fixtures with exact expected raw measurements and verdicts. Evidence: known-value corpus.
    • V9.2.b Verify claim extraction/class/severity/anchor and citation locator/support/contradiction/reachability gates. Evidence: grounding results.
    • V9.2.c Verify statistical denominator/population/period/method/ uncertainty/dataset and quote speaker/context/edit/consent checks. Evidence: fact-check results.
    • V9.2.d Verify headline omission/exaggeration/certainty/clickbait, source composite, counterclaim/false-balance, and evidence-pack integrity. Evidence: editorial results.
    • V9.2.e Verify correction/retraction dependency/action/idempotency/ ordering/read-back/closure gates across target fixtures. Evidence: cascade results.
    • V9.2.f Map every retained gate to applicability, fixtures, measurement/ verdict assertions, remediation, owner, and current result; fail gaps. Evidence: gate registry coverage.
  • V9.3 Run calibrated evaluations for extraction, contradiction, bias, similarity, and assist models. Evidence: precision/recall, subgroup, threshold, and uncertainty report. <!-- SPEC 2026-09-11: §2.5.4/§2.5.5. Tasks map to libs/veritas/claims (extraction), libs/oshun/domain-veritas/src/contradiction (contradiction), libs/veritas/bias-detection (bias — untested today), plagiarism/similarity (no implementation found; V4.12 builds it) and libs/veritas/llm assist. .a is synthetic and says so; .f's review is human. -->
    • V9.3.a Freeze representative held-out corpora/splits/labels for each named model task, risk class, desk, locale, subgroup, and adversarial condition. Evidence: evaluation manifest.
    • V9.3.b Pin model/provider/prompt/tool/config/index/code versions and run with complete input/output/cost/latency provenance. Evidence: run records.
    • V9.3.c Report task-appropriate precision/recall/F-score/calibration/ ranking/span/meaning-preservation metrics, confidence intervals, and uncertainty/abstention. Evidence: metric report.
    • V9.3.d Report subgroup/locale/dialect/desk/source-kind/risk differences and severity-weighted false-positive/negative error analysis. Evidence: slice report.
    • V9.3.e Test prompt/source injection, fabricated/washed citations, protected leakage, correlated handoff errors, and misuse boundaries. Evidence: adversarial report.
    • V9.3.f Compare prior approved baseline, enforce per-severity thresholds/ stop-the-line, assign mitigation/rollback, and obtain named review. Evidence: signed evaluation.
  • V9.4 Run real-database, secure-store, blob, search, queue, CMS, and distribution integration suites. Evidence: durable round-trip. <!-- SPEC 2026-09-11: .a's environment: createdb veritas_test on the compose Postgres, prisma migrate deploy from libs/veritas/database, MinIO bucket per tenant, Elasticsearch from the compose search profile, Redis; the "secure store" for protected sources does not exist yet (V6.5 builds it) — until then .c records the gap rather than testing a stand-in. Report passed/skipped/failed per suite. -->
    • V9.4.a Provision production-equivalent supported service versions with two tenants, protected compartments, representative records/blobs/ indexes/jobs/targets, and isolated credentials. Evidence: environment manifest.
    • V9.4.b Exercise canonical write→outbox→blob→index→gate/review→queue→CMS/ distribution→read-model flow with trace correlation. Evidence: integration trace.
    • V9.4.c Verify secure-store grants/revocation/redacted projections and no protected data in ordinary DB/index/queue/log/target systems. Evidence: compartment tests.
    • V9.4.d Inject transaction/blob/index/queue/CMS/target timeout/outage/ duplicate/order/auth/version failures and prove idempotent recovery. Evidence: fault tests.
    • V9.4.e Reconcile authoritative records/blobs/indexes/jobs/CMS/public/ distribution state after create/update/correction/retraction/archive/ restore. Evidence: reconciliation report.
    • V9.4.f Restart/replay/backup/restore and verify durable IDs/hashes/ provenance/audit plus tenant isolation. Evidence: durable round-trip.
  • V9.5 Run NewsML-G2 golden validation for initial, update, correction, package, planning, rights, and embargo scenarios. Evidence: pinned validator. <!-- SPEC 2026-09-11: same gate as V1.14.f/V6.10.e (§2.5.6): vendored IPTC XSDs + one independent reader; goldens under libs/contracts/src/veritas/goldens/newsml/. -->
    • V9.5.a Pin NewsML-G2/profile/schema/code-list/validator versions and canonical deterministic generator inputs. Evidence: validator manifest.
    • V9.5.b Validate initial/update/replacement/correction/retraction News Items for identity/version/status/prior links/content metadata. Evidence: News Item goldens.
    • V9.5.c Validate ordered/related Package Items and Planning Items for assignments/windows/embargoes without protected-source leakage. Evidence: package/planning goldens.
    • V9.5.d Validate rights/territory/language/syndication/subjects/concepts/ media associations and supported extensions. Evidence: metadata goldens.
    • V9.5.e Test unknown/invalid codes, missing required fields, time zones, unsupported target loss, malformed/cross-version inputs. Evidence: negative corpus.
    • V9.5.f Round-trip through supported consumers, semantic-compare, and retain current-branch validation artifacts. Evidence: interoperability report.
  • V9.6 Run C2PA valid, tampered, stripped, unknown-trust, and ingredient lineage tests. Evidence: validation reports. <!-- SPEC 2026-09-11: run c2pa-node@0.5.26 as the validator with a harness-generated trust store; .a pins the SDK version and certificate fingerprints; "stripped" is proven by removing the JUMBF box from real bytes. -->
    • V9.6.a Pin C2PA/profile/SDK/algorithm/trust-store/certificate/revocation versions and representative media/package manifests. Evidence: test manifest.
    • V9.6.b Generate/sign/validate supported original and derived media with exact ingredient/edit/AI/action/package lineage. Evidence: valid report.
    • V9.6.c Tamper bytes/manifest/assertions/ingredient links/signatures/ timestamps/content binding and assert precise integrity failure. Evidence: tamper report.
    • V9.6.d Strip manifests or use unsupported/malformed data and verify provenance-absent/unknown behavior, not false or verified. Evidence: absence tests.
    • V9.6.e Test unknown/untrusted/expired/revoked signer and trust-store/ network/status uncertainty. Evidence: trust tests.
    • V9.6.f Verify renderer/gate/API/export/CMS display parity and correction/ replacement lineage across valid and failure states. Evidence: validation reports.
  • V9.7 Run BFF auth, object/property authorization, idempotency, concurrency, pagination, rate, SSRF, and tenant tests. Evidence: route inventory complete. <!-- SPEC 2026-09-11: three route surfaces, all in scope: BFF /v1/veritas/* (routes/veritas-briefing.ts, veritas/retraction-cascade-runtime.ts, routes/domain-stubs.ts, gated by middleware/release-scope.ts), apps/veritas/api (Hono, JWT without roles), and the CMS (no auth). .a generates the inventory from the served routers; .b will be red on the CMS by construction and that red is the finding. -->
    • V9.7.a Generate complete route/method/operation/auth/scope/object/field/ function/data-class/idempotency/pagination/rate/external-fetch inventory. Evidence: route manifest.
    • V9.7.b Test unauthenticated/expired/revoked/stale/session-fixation/CSRF/ service-account/emergency/step-up cases. Evidence: authentication tests.
    • V9.7.c Test BOLA/property/mass-assignment/filter/sort/count/snippet/batch/ export/direct-function access across roles/tenants/compartments. Evidence: authorization tests.
    • V9.7.d Test duplicate/replayed/conflicting idempotency keys, optimistic version races, timeout-after-commit, and concurrent transitions. Evidence: mutation tests.
    • V9.7.e Test stable cursor/page bounds/sort ties/filter auth/large inputs, per-principal/resource rates, quotas, and safe errors. Evidence: protocol tests.
    • V9.7.f Test URL import/webhook/callback SSRF/DNS rebinding/redirect/ private network/credential/scheme/size/time attacks. Evidence: network tests.
    • V9.7.g Reconcile all implemented routes to inventory/OpenAPI and fail missing, unknown, untested, or protected-field-exposing operations. Evidence: completeness gate.
  • V9.8 Run full-journey Playwright: pitch to grounded CMS publication. Evidence: live backing and read-back. <!-- SPEC 2026-09-11: there is no host, no pitch UI and no durable CMS today; this cell is reachable only after V0.1.d, V6.1, V6.9 and V8.2. Until then it stays open with that dependency named, and no consumer-room spec may be relabelled as this journey. -->
    • V9.8.a Create/submit/compare/commission a pitch and build an assignment with owner/deadlines/risks/embargo against live services. Evidence: pitch trace.
    • V9.8.b Ingest/quarantine/review real source artifacts and author draft/ claims/citations/quotes/media/disclosures with immutable provenance. Evidence: grounding trace.
    • V9.8.c Run fact-check/gates/standards/legal/accessibility review, fix a failure, and prove exact-revision approval freshness. Evidence: review trace.
    • V9.8.d Create locked release candidate, preview/diagnose targets, sign off, schedule, and dispatch through real CMS adapter/job. Evidence: release trace.
    • V9.8.e Read back public content/metadata/media/disclosures/provenance/ revision and semantic-compare to candidate. Evidence: read-back proof.
    • V9.8.f Run keyboard/screen-reader responsive journey and negative role/ stale dependency/partial CMS failure paths. Evidence: Playwright report.
  • V9.9 Run full-journey Playwright: protected source to redacted review. Evidence: no identity leak in UI, logs, network, notifications, or audit projection. <!-- SPEC 2026-09-11: depends on V6.5 (compartmented source APIs) and the V0.5 role model; .f's no-leak inspection includes the Elasticsearch index and Redis, which the current API writes to without any compartment. -->
    • V9.9.a Onboard a protected source after authorized step-up/grant and store identity/contact/communications only in compartment. Evidence: onboarding trace.
    • V9.9.b Link pseudonymous source/citations/evidence to a story and assign ordinary fact-check/standards/legal/source-protection reviews. Evidence: editorial trace.
    • V9.9.c Compare authorized officer and ordinary reviewer projections in UI/API/network, including counts/timing/errors/search/sidebar/comments/ notifications. Evidence: role comparison.
    • V9.9.d Revoke/expire grant mid-session and verify reveal/cache/copy/ download/offline/deep-link access closes safely. Evidence: revocation tests.
    • V9.9.e Exercise redacted audit/export and emergency access with higher authority/post-review; preserve immutable protected audit. Evidence: audit trace.
    • V9.9.f Inspect client/server logs/traces/analytics/crash reports/queues/ ordinary DB/index/backups for identity linkage. Evidence: no-leak report.
  • V9.10 Run full-journey Playwright: source retraction to verified cascade. Evidence: every registered target updated. <!-- SPEC 2026-09-11: the cascade already runs end to end for source retraction (Phase V header note 3) — start from libs/oshun/veritas-cascade-worker's integration tests and drive them through the BFF route with real Redis/Postgres; .a's registered consumers are V7.15's registry; .e's deliberately failed target is a real registered consumer taken offline, not a mocked callback. -->
    • V9.10.a Publish a story/source/claim fixture to registered CMS/reader/ search/archive/package/newsletter/social/channel/B2B consumers. Evidence: seeded target manifest.
    • V9.10.b Intake and verify source retraction, preview complete dependency graph, approve per-node/target remedy plan, and freeze it. Evidence: impact trace.
    • V9.10.c Execute durable cascade with duplicate/out-of-order/temporary- failure injection and observe truthful per-target progress/retry. Evidence: saga trace.
    • V9.10.d Read back each registered consumer semantically for state/notice/ content/index/cache/event/version and reconcile. Evidence: target proof.
    • V9.10.e Keep an intentionally failed target unresolved, verify closure blocked, repair/retry, then close only after successful read-back. Evidence: closure test.
    • V9.10.f Verify audit/provenance/prior versions/protected redaction and keyboard/screen-reader operations. Evidence: cascade Playwright report.
  • V9.11 Run full-journey Playwright: correction and NewsML/CMS update. Evidence: reader-visible prior/new version. <!-- SPEC 2026-09-11: depends on V6.9, V6.10 and V7.7; read-back in .d is an HTTP read of the public reader (apps/veritas/web) and of the CMS, from separately started processes. -->
    • V9.11.a Seed a grounded published release with NewsML/CMS/public read- back, claims/citations/media, and downstream target identity. Evidence: baseline fixture.
    • V9.11.b Author correction against exact prior public block/text/version, affected claims, rationale, notice, locale/target, and approvals. Evidence: correction trace.
    • V9.11.c Generate/validate NewsML update/correction chain and execute idempotent CMS update through real adapter. Evidence: standards trace.
    • V9.11.d Read back prior/new public version, semantic diff, visible correction note/timestamps/history/canonical URL and NewsML references. Evidence: reader proof.
    • V9.11.e Test stale prior-text race, correction-of-correction, partial target failure/retry, and no historical overwrite. Evidence: edge tests.
    • V9.11.f Verify accessibility, locale and cache/search/archive propagation plus exact candidate/decision/audit lineage. Evidence: Playwright report.
  • V9.12 Run WCAG 2.2 AA, keyboard, announcement, and semantic-equivalence suites. Evidence: themes/viewports/locales. <!-- SPEC 2026-09-11: §2.5.8 — 2.2 tags through the kit policy; .f's manual assistive-technology review is not claimable on this Mac and the matrix names it as not run. -->
    • V9.12.a Define critical route/workspace/dialog/action matrix across web/ desktop/responsive, themes/density, 200–400% zoom, RTL/long locales, and assistive technologies. Evidence: coverage plan.
    • V9.12.b Run automated WCAG rules, name/role/value, landmark/heading/ form/error, contrast, focus, target size, reflow, and motion checks. Evidence: automation report.
    • V9.12.c Run pointer-free editing/source/citation/review/publish/ correction/protected-control journeys with logical focus and non-drag alternatives. Evidence: keyboard report.
    • V9.12.d Assert save/conflict/gate/review/embargo/publish/cascade/error announcements for content/order/deduplication/privacy. Evidence: announcement report.
    • V9.12.e Compare graph/timeline/analytics/media/diff visual and semantic paths for content/operations/selection/status/permissions. Evidence: equivalence report.
    • V9.12.f Run manual supported screen-reader/zoom/high-contrast/RTL review and close or block all findings. Evidence: signed matrix.
  • V9.13 Run long-story, source-volume, search, claim-graph, autosave, concurrent edit, publish, and cascade performance/load tests. Evidence: SLO/budget reports. <!-- SPEC 2026-09-11: §2.5.7 — load runs on the Linux dev server; budgets are written before measuring; Elasticsearch on the Mac counts against the memory checkpoint and is torn down between scenarios. -->
    • V9.13.a Generate versioned production-scale fixtures for long blocks/ claims/citations/comments/history, source registry, graph, releases, targets, corrections, and collaborators. Evidence: load manifest.
    • V9.13.b Measure web/desktop startup/story open/editor input/scroll/ inspector/diff/history/autosave and sustained memory. Evidence: editor profile.
    • V9.13.c Measure source registry/search index/query/facets/pagination/ update visibility and claim graph load/layout/semantic operations. Evidence: discovery profile.
    • V9.13.d Measure concurrent save/rebase/conflict/offline reconnect under multiple clients, network constraints, DB contention, and long sessions. Evidence: collaboration profile.
    • V9.13.e Measure candidate creation/preview/publish/read-back and large correction/retraction plan/dispatch/progress/verification/replay. Evidence: release profile.
    • V9.13.f Capture client/server/database/search/queue/target traces and CPU/memory/network/storage/query plans, compare to budgets, fail regressions. Evidence: SLO report.
  • V9.14 Run source leak, prompt injection, fabricated citation, defamation guard, embargo leak, mass publish, auth, tenant, and audit tamper adversarial tests. Evidence: zero blockers. <!-- SPEC 2026-09-11: known findings to start red on: CMS routes accept unauthenticated writes; the API JWT carries no role so every "role" case is an "any token" case; the consumer room and /domains/veritas/* must never render protected fields (they have none today — assert that stays true). -->
    • V9.14.a Probe protected source/tip/legal/embargo identity/content through UI/API/search/counts/snippets/errors/logs/cache/notifications/exports/ analytics/model context. Evidence: leakage report.
    • V9.14.b Inject instructions/active content/confusables through sources, documents, datasets, media metadata, comments, tips, provider outputs, and multi-agent handoffs. Evidence: injection report.
    • V9.14.c Attempt plausible fabricated/altered/washed citations, quotes, sources, claims, attribution, and evidence-pack hashes. Evidence: grounding report.
    • V9.14.d Exercise defamation/privacy/contempt/rights/high-risk/specialist/ false-balance guard bypass through draft, import, AI, emergency, target adapters. Evidence: harm/legal report.
    • V9.14.e Attempt embargo/schedule/deep-link/cache leak and bulk/mass/ replayed publish, notification, correction suppression, and target credential abuse. Evidence: operations report.
    • V9.14.f Attempt auth/session/role/service-account/BOLA/property/function/ tenant/compartment and audit event/hash/signature/order tamper. Evidence: authorization/integrity report.
    • V9.14.g Require zero severity-blocking findings; assign remediation, regression tests, owner, and approved expiring exception for lower risk. Evidence: adversarial gate.
  • V9.15 Run backup/restore and correction/retraction replay drills. Evidence: CMS, search, archive, audit, and distributed targets reconcile. <!-- SPEC 2026-09-11: pg_dump/pg_restore of the Veritas database into a second database on the same instance, MinIO mc mirror, Elasticsearch snapshot to a filesystem repository, and replay through libs/oshun/veritas-cascade-worker; verify by hashes, not counts; .f's approval is human. -->
    • V9.15.a Define protected/ordinary data/services/backups/encryption/ retention/region/recovery order/RPO/RTO and target reconstruction limits. Evidence: recovery contract.
    • V9.15.b Restore database/secure store/blob/search/queue/config/key references/audit/release artifacts into an isolated environment. Evidence: restore trace.
    • V9.15.c Reconcile records/blobs/indexes/projections/grants/hashes/audit/ CMS/archive/public/target states and no cross-tenant exposure. Evidence: integrity report.
    • V9.15.d Replay correction/retraction outbox/saga from checkpoints with duplicate/out-of-order/partial prior actions and current expected versions. Evidence: replay tests.
    • V9.15.e Independently read back CMS/search/archive/NewsML/channel/B2B targets and repair or retain every unresolved mismatch. Evidence: target reconciliation.
    • V9.15.f Measure achieved loss/recovery time, record gaps/unsupported external restoration, update runbooks, and obtain operations/security approval. Evidence: signed drill report.
  • V9.16 Publish reporter, fact-check, source-protection, legal, corrections, publishing, incident, and recovery runbooks. Evidence: game-day.
    • V9.16.a Publish reporter/research/source-ingest/authoring/offline/ conflict runbooks with safe diagnostics, remediation, and escalation. Evidence: editorial runbooks.
    • V9.16.b Publish fact-check/gate/evidence/attestation/source-expiry/ disputed-result runbooks with known-value comparisons. Evidence: verification runbooks.
    • V9.16.c Publish source-protection/tip/grant/revocation/emergency-access/ suspected leak runbooks without copying protected details. Evidence: protection runbooks.
    • V9.16.d Publish legal/NMC/election/DPA/harm/complaint/hold runbooks with named counsel/specialist escalation and human decision boundaries. Evidence: compliance runbooks.
    • V9.16.e Publish correction/retraction/publish/target failure/embargo/ emergency recall runbooks with read-back and rollback/repair. Evidence: distribution runbooks.
    • V9.16.f Publish security/availability/integrity/AI/channel incident and backup/restore/replay/reconciliation runbooks. Evidence: recovery runbooks.
    • V9.16.g Run timed cross-role game-days, revise unclear/unsafe/missing steps, verify contacts/access, and obtain named approval. Evidence: game-day report.
  • V9.17 Instrument editorial funnel, gate failures, source freshness, corrections, cascades, distribution, and safe cost metrics. Evidence: privacy review and dashboards.
    • V9.17.a Define versioned privacy-safe events/metrics for pitch/ assignment/draft/review/release/remedy stages with denominators, timestamps, owners, and no individual productivity score. Evidence: funnel contracts.
    • V9.17.b Instrument gate execution/outcome/reason/severity/remediation/ retry/staleness without logging protected evidence or model prompts. Evidence: gate telemetry.
    • V9.17.c Instrument source review/freshness/expiry/retraction and affected claim workload using opaque IDs/cohort floors. Evidence: source telemetry.
    • V9.17.d Instrument correction/retraction detection-to-closure and per- target cascade/read-back/unresolved state. Evidence: remedy dashboards.
    • V9.17.e Instrument per-target publish/distribution/ack/read-back/failure/ rate/latency plus model/job/provider cost with currency/unit/budget. Evidence: operations dashboards.
    • V9.17.f Add SLO/quality/incident alerts with thresholds/uncertainty/ owner/runbook and prohibit volume/engagement/revenue from overriding quality. Evidence: guardrail tests.
    • V9.17.g Validate known-value telemetry, freshness/completeness, access/ redaction/retention/deletion/two-tenant privacy and obtain review. Evidence: privacy-approved observability.
  • V9.18 Produce engineering-readiness evidence. Evidence: all requirements map to current proof.
    • V9.18.a Define required architecture/contracts/migrations/security/ privacy/safety/accessibility/quality/standards/integration/operations evidence with owners/freshness. Evidence: readiness matrix.
    • V9.18.b Map every retained atomic capability and checklist parent/child to implementation, tests/evaluations/reports/runbooks/approvals. Evidence: evidence index.
    • V9.18.c Verify evidence provenance matches candidate commit/build/ contracts/schema/policies/models/providers/dependencies/config/targets. Evidence: provenance check.
    • V9.18.d Classify missing/stale/failed/waived proof, block required cells, and record exception owner/reason/scope/compensation/expiry. Evidence: readiness gate.
    • V9.18.e Reconcile route/capability/gate/consumer/obligation/standard/ channel inventories with no orphan or unknown rows. Evidence: completeness report.
    • V9.18.f Publish go/no-go summary, unsupported/degraded combinations, residual risks, support/rollback, and named decision. Evidence: readiness packet.
  • V9.19 Obtain named editorial standards, legal, source-protection, accessibility, privacy/security, engineering, and product approvals. Evidence: human signoff. <!-- SPEC 2026-09-11: §2.5.9 — packet only; every child stays open until a named person acts. -->
    • V9.19.a Define each approver’s exact scope/criteria/required evidence/ independence/delegate/expiry and nondelegable decisions. Evidence: approval policy.
    • V9.19.b Route immutable candidate/evidence to editorial standards and legal/jurisdictional specialists with open findings/exceptions. Evidence: editorial/legal decisions.
    • V9.19.c Route protected-source/tip/nonleakage/emergency-access evidence to source-protection authority through secure compartment. Evidence: protection decision.
    • V9.19.d Route WCAG/assistive-technology and privacy/security/DPA/threat/ adversarial evidence to named specialists. Evidence: assurance decisions.
    • V9.19.e Route architecture/quality/operations/recovery/support and product scope/UX/rollout/residual-risk evidence. Evidence: engineering/product decisions.
    • V9.19.f Record approve/reject/request-changes/conditional decisions with identity/role/scope/candidate/reason/conditions/time/signature. Evidence: human signoff.
    • V9.19.g Invalidate affected signoffs on commit/evidence/policy/model/ schema/dependency/config/target change and require renewal. Evidence: stale approval tests.
  • V9.20 Obtain named launch approval for newsroom and public correction journey. Evidence: human governance record. <!-- SPEC 2026-09-11: §2.5.9 — packet only; .c's live smoke is runnable by an agent once V9.8 exists, .d–.f are human. -->
    • V9.20.a Assemble final immutable launch packet with newsroom authoring/ protected-source/publish and public correction/retraction journey proof. Evidence: launch packet.
    • V9.20.b Verify candidate commit/build/config/migrations/contracts/ standards/targets/runbooks/monitoring/on-call/rollback/support and open risks. Evidence: launch checklist.
    • V9.20.c Run final live-backed smoke from authorized newsroom entry to verified public publish and correction/read-back without fixture data. Evidence: launch smoke.
    • V9.20.d Hold named editorial/governance/engineering/operations/product review with independent correction authority and no automation vote. Evidence: decision meeting record.
    • V9.20.e Record go/no-go/conditional decision, exact scope/regions/ channels/rollout/conditions/owners/times/expiry/rollback triggers. Evidence: governance record.
    • V9.20.f Verify rollout monitoring and ability to stop, recall, correct, restore, and communicate; invalidate approval on candidate change. Evidence: launch control test.
  • V9.21 Run distinct full journeys for feed/social ingest, anonymous tip to protected investigation, entity merge/split, data visualization, podcast/video/short-form, live program, newsletter, social/WhatsApp, emergency alert, community moderation, multilingual publication, B2B client, USSD/voice, and commercial conflict projection. Evidence: no generic story-publish test substitutes for channel-specific proof. <!-- SPEC 2026-09-11: each journey names its owned live service (Phase V header note) and the external provider it needs; a journey whose provider has no credential in env-master (WhatsApp, telecom USSD, payment rails, social platforms) is recorded as attempted-and-blocked, never simulated to green (§2.5.6). .p is the ledger that says which of the fifteen ran for real. -->
    • V9.21.a Run feed/crawl/social-monitor ingest through policy/fetch/ quarantine/normalize/language/dedupe/cluster/replay with live lineage. Evidence: ingestion journey.
    • V9.21.b Run anonymous submission/receipt/safety triage/protected assignment/research/source conversion and prove nonleakage. Evidence: tip journey.
    • V9.21.c Run entity candidate review/merge/redirect then split/ reallocation/downstream invalidation with protected graph isolation. Evidence: entity journey.
    • V9.21.d Run dataset snapshot→query/transform→chart/map/table→methodology/ semantic alternative→publish/reproduce/update. Evidence: data-viz journey.
    • V9.21.e Run podcast/audio script→record/edit/mix/transcript/review/feed and timecoded correction propagation. Evidence: podcast journey.
    • V9.21.f Run video/short-form plan→ingest/edit/graphics/captions/rendition/ review→platform/embed and replacement. Evidence: video journey.
    • V9.21.g Run live rundown/preflight/hold/delay/captions/moderation/Q&A/ failover/stop→record/clip/archive/correct. Evidence: live journey.
    • V9.21.h Run newsletter author/preview/test/audience/consent/schedule/send/ bounce/archive/unsubscribe/correct. Evidence: newsletter journey.
    • V9.21.i Run social and WhatsApp plans through approval/publish/delivery/ moderation or reply/unsubscribe/rate/correction with current APIs. Evidence: social/channel journeys.
    • V9.21.j Run emergency alert authority→preview→deduplicated multi-channel delivery→acknowledgment→recall/correction→post-event review. Evidence: emergency journey.
    • V9.21.k Run comment/report/moderate/appeal/legal hold/transparency and story correction context synchronization. Evidence: community journey.
    • V9.21.l Run multilingual translation/transliteration/media/captions/ factual parity/cultural review/publish/read-back/cascade. Evidence: locale journey.
    • V9.21.m Run B2B client/key/scope/quota/feed/search/webhook/correction/ revocation/SDK conformance with no protected fields. Evidence: B2B journey.
    • V9.21.n Run USSD menu and voice callback through consent/locale/session/ provider/fallback/charge/emergency/correction state. Evidence: feature-phone journey.
    • V9.21.o Run commercial subscription/payment/ad/sponsorship conflict and newsroom redacted disclosure projection while denying editorial mutations. Evidence: commercial journey.
    • V9.21.p Verify every journey uses owned live services, target read-back, trace/audit, roles/tenants, failure recovery, and current correction identity; retain distinct evidence. Evidence: journey coverage ledger.
  • V9.22 Run NMC complaint, election-phase, DPA subject-request, vulnerable person/child, graphic-content, right-of-reply, advertising disclosure, payment, and emergency-broadcast adversarial/compliance suites. Evidence: current jurisdictional profile and named specialist review. <!-- SPEC 2026-09-11: the code-owned floors are apps/veritas/api/src/domain/compliance/{nmc,election}.ts and routes/v1/{compliance,election-coverage,data-protection}.ts; fixtures exercise those, .a's jurisdiction profile carries a version and open questions, .k is human. -->
    • V9.22.a Pin current Ghana and applicable channel/payment jurisdictional profiles/versions/effective dates/authoritative sources and specialists. Evidence: compliance manifest.
    • V9.22.b Run NMC complaint intake/SLA/investigation/human decision/ correction-apology-right-of-reply/escalation/appeal. Evidence: NMC suite.
    • V9.22.c Run pre-election/campaign/silence/polling/results/post-election phase fixtures for equal treatment/source verification/ads/incitement/ misinformation. Evidence: election suite.
    • V9.22.d Run DPA access/correct/delete/restrict/object/portable discovery and remedy across ordinary/protected/channel/commercial/processors plus legal-hold conflict. Evidence: DPA suite.
    • V9.22.e Run child/vulnerable/trauma/location/doxxing/suicide/self-harm consent/minimization/specialist/distribution/correction fixtures. Evidence: vulnerable-person suite.
    • V9.22.f Run graphic-content media/thumbnail/autoplay/warning/age/access/ notification/search/social/locale handling and redacted preview. Evidence: graphic-content suite.
    • V9.22.g Run right-of-reply authenticity/scope/deadline/editorial review/ publication/notice/appeal without direct external mutation. Evidence: reply suite.
    • V9.22.h Run advertising/sponsorship identity/separation/disclosure/ targeting/phase/target and attempted commercial interference. Evidence: advertising suite.
    • V9.22.i Run payment/subscription/tip consent/charge/tokenization/ idempotency/refund/dispute/privacy and editorial-data separation. Evidence: payment suite.
    • V9.22.j Run emergency-broadcast activation/authority/quiet override/ template/target/delivery/recall/false-alarm review and attempted ordinary campaign bypass. Evidence: emergency suite.
    • V9.22.k Require named legal/editorial/privacy/safety/accessibility/ commercial/channel specialist review and zero unmitigated blockers. Evidence: signed compliance report.
  • V9.23 Run correction/retraction propagation across reader web/mobile, search/recommendation, newsletter, social, WhatsApp, push/email, live archive/clips, community context, B2B APIs/webhooks, USSD/voice, analytics, and commercial placements. Evidence: every registered consumer proves read-back or remains visibly unresolved. <!-- SPEC 2026-09-11: the registered-consumer set is V7.15's registry; today it holds the Sophia re-grounder, the Metis revalidator and the BFF acknowledgement store. A consumer with no live target (WhatsApp, USSD, B2B webhooks without a client) is "visibly unresolved" by the cell's own clause, and .l blocks closure on it — that is the expected end state until those channels exist. -->
    • V9.23.a Seed one versioned story/claim/source/media/package into every registered supported consumer and record target IDs/versions/URLs. Evidence: propagation baseline.
    • V9.23.b Author/approve correction and source/claim retraction scenarios with exact dependency graph, desired target states, notices, and immutable plan. Evidence: remedy inputs.
    • V9.23.c Verify reader web/mobile, canonical/locale/history/cache and search/recommendation/topic/entity/related surfaces update. Evidence: reader/discovery read-back.
    • V9.23.d Verify newsletter archive/follow-up and social/WhatsApp/push/ email update/delete/follow-up/recall capability-specific state. Evidence: messaging read-back.
    • V9.23.e Verify live recording/archive/clips/transcript/captions and podcast/video/article embeds carry corrected ranges/notices/replacements. Evidence: media read-back.
    • V9.23.f Verify community thread context/labels and complaint/feedback references update without rewriting community speech. Evidence: community read-back.
    • V9.23.g Verify B2B feeds/search/APIs/webhooks and developer cache/version expose current state and signed correction/retraction events. Evidence: B2B read-back.
    • V9.23.h Verify USSD/voice menu/script/cache and callbacks present current canonical story/claim/remedy state. Evidence: feature-phone read-back.
    • V9.23.i Verify analytics dimensions/metrics annotate or recompute corrected/retracted content and do not preserve false editorial scores. Evidence: analytics reconciliation.
    • V9.23.j Verify advertising/sponsorship/partnership/commercial placements stop or update links/content/disclosures without vetoing remedy. Evidence: commercial reconciliation.
    • V9.23.k Inject duplicate/out-of-order/outage/auth/version/unsupported/ cache-lag failures, retry safely, and retain precise per-target state. Evidence: fault report.
    • V9.23.l Reconcile registered expected versus observed consumers and block closure until every required target proves semantic read-back or remains visibly owned/escalated/unresolved. Evidence: cascade gate.
  • V9.24 Run training/evaluation-data provenance, license/consent, protected-source exclusion, contamination, deletion/opt-out, split leakage, poisoning, feedback manipulation, and reproducibility tests. Evidence: newsroom or audience data cannot enter a model pipeline merely because it exists in Veritas.
    • V9.24.a Build approved/denied/boundary fixtures across public/licensed/ newsroom/audience/channel/protected/legal/embargoed/payment/child data and purposes. Evidence: policy corpus.
    • V9.24.b Verify item→source revision→licence/consent/purpose→transform/ annotation→dataset/split→model/evaluation lineage with hashes/versions. Evidence: provenance tests.
    • V9.24.c Test protected-source/tip/legal/private-draft/payment/opted-out/ disallowed licence/expired right exclusion at query/export/job/provider boundaries. Evidence: exclusion suite.
    • V9.24.d Detect near/exact duplicates, benchmark or production contamination, entity/topic/time/train-eval-test leakage, label leakage, and memorization canaries. Evidence: contamination report.
    • V9.24.e Execute deletion/withdrawal/opt-out/licence expiry/restriction through datasets/caches/providers/derived artifacts/model impact while handling legal hold. Evidence: deletion journey.
    • V9.24.f Inject poisoned labels/content/metadata and manipulated feedback/ brigading/bot/reward signals; verify isolation, detection, review, and no automatic ingestion. Evidence: poisoning tests.
    • V9.24.g Rebuild exact dataset/splits from immutable inputs/code/config/ seeds, compare manifests/counts/hashes, and disclose nondeterminism. Evidence: reproducibility report.
    • V9.24.h Attempt unapproved direct DB/search/log/analytics/export and model-training job access simply because data exists; prove denied/no side effect. Evidence: purpose authorization tests.
    • V9.24.i Require current privacy/security/legal/source-protection/ editorial/model-governance approval and fail stale/unlicensed/ contaminated/unreproducible candidates. Evidence: data release gate.
  • V9.25 EVE-SEAM-EXIT: require "required Eve reads and writes admitted with source and behavior evidence, or an exact task-0.8 explicit non-goal proves zero operations; a named defer preserves ownership but never passes completion" before this phase exits. Evidence: source-derived totality record and passing verifier. <!-- SPEC 2026-09-11: "task-0.8" is EVE_SOTA_GAP_CLOSURE_TODOS_2026-09-01.md 0.8 (the ratified charter workflow inventory); the totality verifier is tools/eve-everywhere/generate-contextual-invocation-totality.mjs, and the Metis admission gates (tools/eve-everywhere/generate-metis-mutation-admission.mjs, verify-metis-view-admission.test.mjs) are the precedent for what an admitted read/write looks like. Phase M carried no such cell, so this is the first domain to exit through it — do not carry a Metis disposition forward. -->
    • V9.25.a Re-run the route, bridge, state-signal, and complete domain-code envelope census; review every changed candidate instead of carrying a stale phase disposition forward. Evidence: current ratchets.
    • V9.25.b Bind the exact ratified task-0.8 workflow and requirement set; classify it as a required seam or an exact source-backed non-goal. Evidence: charter-bound disposition set.
    • V9.25.c For a required seam, admit both read and write operations with byte-pinned source and behavior evidence. A named defer retains its owner and unblock condition but fails completion. Evidence: operation proofs.
    • V9.25.d Permit a zero-operation boundary only for an exact ratified task-0.8 explicit non-goal; absence or unavailability is not a non-goal. Evidence: source-bound non-goal proof.
    • V9.25.e Run the totality verifier, every retained gate-case CLI, and the injected durable-candidate control; retain red and green receipts. Evidence: release gate logs.

Phase E — Euterpe music production workbench#

Euterpe already has a substantial DAW in apps/euterpe/studio-web; this phase must extend and integrate it rather than create a second music editor. The complete workflow spans concept, score, lyrics, arrangement, recording, performance, mix, master, rights, provenance, release, distribution, and catalog operations.

E0 — Product boundary, existing DAW, and professional workflow#

  • E0.1 Ratify Euterpe's product host and decide explicitly whether it is outside or inside the Oshun V1 shell. Evidence: product/architecture ADR.
    • E0.1.a Inventory existing Studio web/desktop/native entry points, identity/project contexts, navigation shells, deployment boundaries, and audience. Evidence: host inventory. <!-- evidence: euterpe-host-inventory.json — studio-web is a DUAL host: a Next.js app-router web frontend (3 routes, 1 nav shell) wrapped by a Tauri desktop shell (productName "Euterpe Studio", id app.oshun.euterpe-studio, 4 native #[command] handlers load/save_project/export_stems/reveal_in_file_manager, 1 capability file), single Nx project (targets lint/test). audience + inside-vs-standalone are the E0.1.b-e semantic half. generate-euterpe-host- inventory.mjs + 5 tests. -->
    • E0.1.b Evaluate inside-shell versus standalone host for audio lifecycle, permissions, latency, offline/native/plugin needs, ownership, and operational isolation. Evidence: option analysis.
    • E0.1.c Decide canonical host(s), route/deep-link/handoff boundaries, data ownership, shell reuse, and prohibited duplicate editor behavior. Evidence: target architecture.
    • E0.1.d Define migration/redirect/session/context/telemetry/support/ rollout/rollback consequences for the chosen boundary. Evidence: delivery plan.
    • E0.1.e Obtain product/audio/architecture/security/accessibility ratification with alternatives and residual risks. Evidence: approved ADR.
  • E0.2 Inventory every Euterpe Studio route, DAW module, component, test, API, Tauri/native bridge, and service dependency. Evidence: route-capability matrix. <!-- partial: route inventory (a) + module inventory (b) done in evidence/v1-workbenches/euterpe-inventory.json. c-e (service dependency inventory, test-coverage ledger, route-capability reconciliation) are design/human and stay unchecked. -->
    • E0.2.a Enumerate routes/navigation/commands/pages/dialogs and user/ operator journeys across web and installed desktop. Evidence: route inventory. <!-- evidence: .routes — the 3 Euterpe Studio App-Router pages (/, /daw, /realtime) with source + test. The DAW is a small route surface over a large module surface (E0.2.b). generate-euterpe-inventory.mjs + 6 tests. -->
    • E0.2.b Enumerate DAW engines/modules/worklets/workers/components/state/ commands/assets/device/plugin/native bridge surfaces. Evidence: module inventory. <!-- evidence: .moduleInventory — 148 engine/logic modules + 58 components across 6 subsystems (daw=135 modules + 150 tests, plus components/identity- billing/realtime-mrt2/shell-runtime/telemetry-launch); nativeSignals detect worklet/worker/Tauri-bridge (0 in-app AudioWorklet processors or Web Workers — the audio engine lives outside the app shell; 1 Tauri-bridge file). -->
    • E0.2.c Enumerate APIs/realtime protocols/stores/blobs/search/queues/ providers/distributors and service owners. Evidence: dependency inventory. <!-- design: service-owner mapping is semantic. -->
    • E0.2.d Map unit/integration/E2E/audio-golden/device/compatibility/ security tests to each retained capability. Evidence: test coverage ledger.
    • E0.2.e Reconcile routes to capabilities/owners/hosts/data classes and flag duplicate/orphan/dead/fixture/unsafe paths. Evidence: route-capability matrix.
  • E0.3 Map all libs/euterpe/* packages to workbench capabilities, services, consumer products, or out-of-scope. Evidence: complete ledger.
    • E0.3.a Inventory every Euterpe package/export/audio consumer/runtime/ test/documented purpose and dependency direction. Evidence: package <!-- evidence: .domains.euterpe.packages — 49 libs/euterpe/* packages with exports/consumer/dependency/test/disposition/direction. euterpe libs-sourced caps=0 (DAW logic lives in apps/euterpe — see E0.2). evidence/v1-workbenches/domain-package-inventories.json (254 pkgs; per-package exports/consumers/deps/disposition/tests + capability catalog with exact source locations). --> inventory.
    • E0.3.b Map retained exports to atomic production capabilities, canonical contracts, owning host/service, and accountable team. Evidence: capability mapping.
    • E0.3.c Classify consumer/conservatory/discovery/legacy/demo/duplicate/ dead code with keep/merge/migrate/deprecate/remove/out-of-scope disposition. Evidence: disposition ledger.
    • E0.3.d Verify every package/export/consumer has one reviewed disposition and no forbidden dependency crosses product or native security boundaries. Evidence: completeness gate.
  • E0.4 Classify existing DAW behavior as live, local-only, fixture-backed, service-backed, provider-gated, or incomplete. Evidence: browser/network walkthrough.
    • E0.4.a Build capability/scenario inventory for arrange/session/score/ mixer/devices/media/rights/release across browser and desktop. Evidence: walkthrough plan.
    • E0.4.b Observe UI/network/storage/audio worklets/workers/native IPC/ provider calls and trace authoritative state and side effects. Evidence: trace captures.
    • E0.4.c Classify each behavior by exact backing/runtime/feature flag/ configuration/dependency and distinguish partial from live. Evidence: implementation ledger.
    • E0.4.d Exercise loading/empty/offline/not-configured/degraded/failure/ permission/device-loss paths and identify simulated success. Evidence: state walkthrough.
    • E0.4.e Reconcile findings to code/tests/docs and assign remediation/ owner/evidence freshness. Evidence: reviewed classification.
  • E0.5 Define composer, lyricist, arranger, performer, recording engineer, editor, mix engineer, mastering engineer, producer, rights manager, label/publisher, distributor, reviewer, and administrator roles. Evidence: permission matrix.
    • E0.5.a Define duties and least-privilege object/field/function rights for every named role across project through catalog/remedy. Evidence: role definitions.
    • E0.5.b Define organization/project/song/session/release/catalog/ territory scopes, delegation, guest, service account, and temporary grants. Evidence: scope model.
    • E0.5.c Define edit/record/render/device/plugin/review/approve/rights/ deliver/takedown/admin rights by maturity state and host capability. Evidence: permission matrix.
    • E0.5.d Define separation of composition/mastering/rights/label/ distribution approvals, conflicts/recusal, quorum, and nondelegable human actions. Evidence: SoD policy.
    • E0.5.e Verify client projection and server/native enforcement across positive/negative/cross-project/revoked/offline/direct command cases. Evidence: authorization suite.
  • E0.6 Define project, song, composition, arrangement, session, recording, mix, master, release, and catalog maturity models separately. Evidence: state diagrams.
    • E0.6.a Define independent state vocabularies, actors, entry/exit criteria, terminal/archival conditions, and allowed transitions for each aggregate. Evidence: state definitions.
    • E0.6.b Define exact dependency relationships without collapsing creative version, technical readiness, rights approval, delivery, and catalog state into one status. Evidence: dependency model.
    • E0.6.c Define transition commands/guards/side effects/audit/invalidation for draft/review/approved/locked/released/replaced/taken-down flows. Evidence: transition tables.
    • E0.6.d Map existing DAW/API/storage/distributor states and ambiguous/ unknown/in-flight historical values. Evidence: compatibility matrix.
    • E0.6.e Property-test valid/invalid/concurrent/stale transitions and cross-model invariants. Evidence: state diagrams and tests.
  • E0.7 Define supported browser, desktop/Tauri, audio device, MIDI device, sample rate, channel, plugin, and offline configurations. Evidence: compatibility matrix.
    • E0.7.a Define supported OS/browser/desktop/app/runtime/CPU/architecture/ memory/storage/GPU combinations and lifecycle. Evidence: host matrix.
    • E0.7.b Define audio API/driver/device/input-output/channel/buffer/sample- rate/bit-depth/clock/hot-plug configurations and limits. Evidence: audio matrix.
    • E0.7.c Define MIDI 1/2, MPE, control-surface, sync/clock, port/transport, permission, and reconnect combinations. Evidence: MIDI matrix.
    • E0.7.d Define native/browser plugin formats/architectures/signature/ sandbox/licence/UI/state/sidechain/latency support and exclusions. Evidence: plugin matrix.
    • E0.7.e Define online/offline/local/remote/provider/render-worker feature availability, storage quotas, graceful degradation, and export escape. Evidence: offline matrix.
    • E0.7.f Publish supported/conditional/unsupported cells with known loss, evidence, owner, version, expiry, and no silent fallback. Evidence: compatibility profile.
  • E0.8 Define collaboration, latency, clock, sync, and ownership boundaries between browser audio, realtime services, and stored projects. Evidence: architecture diagram.
    • E0.8.a Define authoritative ownership for media bytes, project graph, edit operations, transport, device state, render artifacts, presence, and reviews. Evidence: ownership table.
    • E0.8.b Define browser/desktop audio render/control threads and real-time safety boundaries from UI/network/storage/provider work. Evidence: runtime diagram.
    • E0.8.c Define musical time/sample/timecode/tempo clock authority, synchronization/drift/latency measurement/compensation, and failover. Evidence: timing architecture.
    • E0.8.d Define collaboration operations/checkpoints/ordering/CRDT or locking/conflicts/offline queue/reconnect and transport-leader semantics. Evidence: realtime protocol.
    • E0.8.e Define durable-before-ack save/upload/render/job boundaries and recovery/reconciliation after browser/service/native failure. Evidence: durability model.
    • E0.8.f Validate architecture against solo, concurrent, remote recording, stage, offline, and render-worker scenarios. Evidence: architecture review.
  • E0.9 Threat-model malicious audio/plugins/presets/projects, voice impersonation, sample theft, prompt injection, rights fraud, unreleased leaks, device permission abuse, desktop bridge, and distribution fraud. Evidence: mitigation/test map.
    • E0.9.a Map actors/assets/trust boundaries/data flows/privileged commands across web/audio worklets/native host/plugins/providers/workers/ distributors. Evidence: threat diagrams.
    • E0.9.b Model malformed/polyglot/bomb media/project/preset, plugin code/ license/update/sandbox escape, unsafe DSP/feedback, and parser attacks. Evidence: content/code threats.
    • E0.9.c Model voice/performer impersonation, stolen samples/stems/ masters, fabricated rights/splits/consent/provenance, and similarity evasion. Evidence: identity/rights threats.
    • E0.9.d Model prompt/source/preset injection, model/tool misuse, provider leakage/training, generated malware/audio abuse, and agent privilege expansion. Evidence: AI threats.
    • E0.9.e Model unreleased/contract/voiceprint leakage through search/log/ cache/collaboration/review links/support/distribution and cross-tenant access. Evidence: confidentiality threats.
    • E0.9.f Model device/microphone/MIDI/OSC permission abuse, desktop IPC/ filesystem/deep-link/update attacks, mass delivery/takedown suppression, and account compromise. Evidence: platform threats.
    • E0.9.g Assign preventive/detective/recovery controls, residual risk, abuse tests, monitoring, owners, and review cadence. Evidence: mitigation/test map.
  • E0.10 Define rights and privacy classifications for demos, voice prints, stems, sessions, performer data, contracts, splits, and unreleased masters. Evidence: retention/access rules.
    • E0.10.a Inventory named data plus derivatives/proxies/fingerprints/ embeddings/logs/caches/backups/exports/provider/model inputs. Evidence: data catalog.
    • E0.10.b Assign sensitivity, rights kind, data subject/owner, purpose/ lawful basis, project/territory scope, and permitted roles/actions. Evidence: classification matrix.
    • E0.10.c Define encryption/key/region/provider/device/offline/clipboard/ download/share/watermark and collaboration restrictions. Evidence: handling rules.
    • E0.10.d Define active/archive/contract/legal-hold/backup retention, deletion/revocation/takedown, derivative impact, and conflict handling. Evidence: lifecycle schedule.
    • E0.10.e Define redacted projections for search/logs/analytics/ notifications/support/review/distributors and prohibit hidden training. Evidence: projection rules.
    • E0.10.f Validate representative workflows and obtain rights/privacy/ security/production approval. Evidence: signed policy.
  • E0.11 Define professional loudness, format, metadata, provenance, and distribution profiles by target. Evidence: ratified delivery matrix.
    • E0.11.a Inventory supported streaming/download/broadcast/podcast/video/ physical/spatial/distributor targets and current authoritative specs. Evidence: target inventory.
    • E0.11.b Define loudness/true-peak/dynamics/channel/layout/downmix/ duration/gap/fade constraints and pinned measurement standard per target. Evidence: audio profiles.
    • E0.11.c Define container/codec/sample-rate/bit-depth/dither/naming/ artwork/lyrics/captions/file-size checks and accepted alternatives. Evidence: format profiles.
    • E0.11.d Define title/version/contributor/identifier/genre/language/ advisory/territory/rights/pricing/date metadata and DDEX mapping. Evidence: metadata profiles.
    • E0.11.e Define C2PA/external provenance/signing/ingredient/AI disclosure requirements and unsupported/stripped states. Evidence: provenance profiles.
    • E0.11.f Pin versions/owners/effective dates, validate goldens with distributors, and publish supported/conditional/unsupported delivery cells. Evidence: ratified matrix.
  • E0.12 Define success measures for creative latency, glitch rate, save durability, collaboration, rights completeness, master compliance, delivery acceptance, and release correctness. Evidence: metric spec.
    • E0.12.a Define operation-to-audible-response/input-monitor/transport/edit latency distributions by supported host/device/buffer. Evidence: creative latency metrics.
    • E0.12.b Define dropout/underrun/overrun/click/xrun/glitch metrics with sampling, false-positive controls, severity, and privacy. Evidence: engine quality metrics.
    • E0.12.c Define acknowledged-save survival, autosave/recovery loss, collaboration convergence/conflict/reconnect, and media upload durability. Evidence: durability metrics.
    • E0.12.d Define rights/consent/split/provenance completeness and master target/QC conformance without reducing creativity to one score. Evidence: compliance metrics.
    • E0.12.e Define distributor acceptance plus semantic read-back/catalog identity/territory/correction/takedown correctness. Evidence: delivery metrics.
    • E0.12.f Define units/denominators/windows/slices/uncertainty/budgets/ alerts/owners and balanced quality guardrails. Evidence: metric spec.
  • E0.13 Author and ratify V1/EUTERPE_MUSIC_WORKBENCH_DESIGN_2026-07-23.md with current-state, product-host decision, DAW/audio architecture, data, APIs, rights, standards, UI, migration, rollout, and decision records. Evidence: approved doc and resolved review comments.
    • E0.13.a Document current routes/modules/engine/native/packages/services/ formats/workflows/tests, backing state, constraints, and gaps. Evidence: current-state chapter.
    • E0.13.b Document chosen host/ownership/roles/capability profiles and DAW/ audio/realtime/device/plugin/offline/render architecture. Evidence: target architecture chapters.
    • E0.13.c Document contracts/stores/APIs/events/jobs/providers/ distributors, compatibility/migrations, and recovery. Evidence: platform chapters.
    • E0.13.d Document rights/consent/splits/provenance/AI/safety/privacy/ standards/target profiles and approval boundaries. Evidence: governance chapters.
    • E0.13.e Document extension of existing DAW, task-led operations UI, responsive/accessibility/semantic/offline/performance states. Evidence: experience chapter.
    • E0.13.f Document rollout/rollback/observability/runbooks/verification/ evidence/launch criteria and all decision alternatives/owners/dates. Evidence: delivery chapters.
    • E0.13.g Resolve every substantive review comment and obtain named product/music/audio/rights/security/accessibility/engineering ratification. Evidence: approved design.
  • E0.14 Reconcile every libs/euterpe/* package, Studio route/component, Web Audio/native/Tauri operation, provider, agent, realtime protocol, workflow, test, documented feature, and distribution target at atomic-feature level. Evidence: explicit dispositions cover foundations/ theory, composition/score/lyrics, instruments/effects/samples, recording/ DAW, AI audio/voice, mix/master, spatial, stage/realtime, podcast/video, rights/provenance, marketing/discovery, distribution/catalog, learning, and operations.
    • E0.14.a Inventory code/routes/components/commands/audio/native/provider/ agent/protocol/jobs/tests/docs/flags/deployed integrations and targets. Evidence: reconciliation inputs. <!-- evidence: euterpe-code-census.json — comprehensive launch census: 49 libs packages + 2 app projects, 918 code files + 497 tests, 3 routes over 148 DAW modules + 58 components, dual Next.js+Tauri host (4 native #[command]s, 5 rust files), 52 docs, and DAW-specific deploy targets (build/build:native/ build:worklet/lint/test/typecheck). Consolidated from git counts + committed E0.1/E0.2 evidence. Atomic-feature split + dispositions are E0.14.b+ (semantic). generate-euterpe-code-census.mjs + 4 tests. -->
    • E0.14.b Split sources into atomic user/operator/system capabilities with host/audience/owner/state/dependencies/data class/profile. Evidence: atomic ledger.
    • E0.14.c Cover foundations/theory, composition/score/lyrics, instruments/effects/samples, recording/DAW, AI audio/voice, mix/master, and spatial explicitly. Evidence: production coverage.
    • E0.14.d Cover stage/realtime, podcast/video, rights/provenance, marketing/discovery, distribution/catalog, learning, developer/provider, and operations explicitly. Evidence: extended coverage.
    • E0.14.e Assign each Euterpe item keep/consolidate/migrate/replace/ deprecate/remove/externalize with target owner/host/profile/ dependencies/delivery/expiry. Evidence: disposition ledger.
    • E0.14.f Reconcile to routes/packages/contracts/stores/tests and fail orphan/duplicate/unknown/unreviewed capabilities. Evidence: completeness gate.
  • E0.15 Define the boundary among professional production, live/stage, podcast/video, label/marketing/catalog operations, artist collaboration, developer/provider administration, conservatory/ear-training, discovery, sacred/philosophy experiences, and other consumer products. Evidence: every capability has an authorized host/audience, handoff, or explicit exclusion from the production workbench.
    • E0.15.a Define audience/trust/data/action boundaries and product goals for each named surface. Evidence: surface matrix.
    • E0.15.b Assign every retained capability/mutation to one canonical host/ owner/profile/role, distinguishing production core from projections and separate products. Evidence: host ledger.
    • E0.15.c Define typed redacted handoffs for live/stage, podcast/video, marketing/catalog, collaboration/review, providers/developers, learning, and consumer discovery. Evidence: handoff contracts.
    • E0.15.d Define identity/session/token/deep-link/storage/media/provenance/ rights/logging separation and step-up transitions. Evidence: security boundary.
    • E0.15.e Define shared asset/release/correction/takedown identity without granting consumer/marketing/learning surfaces production authority. Evidence: control-flow diagrams.
    • E0.15.f Test authorized cross-surface journeys and attempted object/ field/function/media/device access; eliminate duplicate mutations. Evidence: boundary suite.
  • E0.16 Define capability profiles for browser, installed desktop/Tauri, native audio/plugin host, remote collaboration, live stage, offline, and render worker. Evidence: unsupported audio/MIDI/plugin/spatial/voice behavior is explicit instead of silently degraded.
    • E0.16.a Define profile identity/version/detection/prerequisites and authoritative capability flags/limits for every named host/mode. Evidence: profile schema.
    • E0.16.b Map audio input/output/latency/channel/sample rate/render/ filesystem/media/device permissions by profile. Evidence: audio capability matrix.
    • E0.16.c Map MIDI/MPE/sync/control/plugin/native IPC/multi-window/stage command availability and safety boundaries. Evidence: device matrix.
    • E0.16.d Map collaboration/provider/AI/offline storage/render-worker/ distribution features, data movement, rights/consent, and fallbacks. Evidence: service matrix.
    • E0.16.e Specify unavailable/degraded/read-only/handoff/export behavior, user messaging, telemetry, and no silent quality/state loss. Evidence: fallback contracts.
    • E0.16.f Detect at runtime, revalidate on device/network/config changes, and test every supported/conditional/unsupported cell. Evidence: capability profile suite.

E1 — Music, session, rights, and release contracts#

  • E1.1 Define project, composition, work, recording, version, arrangement, section, tempo map, meter map, key map, marker, and cue entities. Evidence: schema fixtures.
    • E1.1.a Define stable identity/tenant/project/owner/lifecycle/version/ timestamps/audit and distinguish abstract work, composition, recording, and project. Evidence: core schemas.
    • E1.1.b Define arrangement/version parentage and ordered/nested section identity, range, repetition, alternative, and reference semantics. Evidence: structure fixtures.
    • E1.1.c Define piecewise tempo/meter/key maps with musical/sample/ absolute-time anchors, ramps, pickup, polymeter/polyrhythm, and microtonal context. Evidence: map fixtures.
    • E1.1.d Define marker/cue types, position/range/label/action/target/ performer/control links and arrangement/version scope. Evidence: cue fixtures.
    • E1.1.e Define cross-reference, archive/delete, transition, unknown enum, tenant, and immutability constraints plus API/event serialization. Evidence: invariant tests.
  • E1.2 Define track, bus, return, clip, take, comp, lane, region, automation, device, preset, routing, send, and render entities. Evidence: DAW contract compatibility.
    • E1.2.a Define stable graph identity/order/type/role/channel/layout/state for track/bus/return/group/folder and device/routing ownership. Evidence: mixer graph schema.
    • E1.2.b Define immutable media source and clip/take/lane/region placement, offset/range/loop/stretch/fade/gain/channel/version semantics. Evidence: edit schemas.
    • E1.2.c Define comp segment/source/take/range/crossfade/order/revision and no-destructive-source invariants. Evidence: comp schema.
    • E1.2.d Define automation lane/parameter/path/timebase/point/curve/mode/ scope/version and interaction with clips/tempo. Evidence: automation schema.
    • E1.2.e Define typed device/preset/state/version/latency/tail/bypass/ missing/sandbox and routing/send/sidechain graph edges. Evidence: device contracts.
    • E1.2.f Define render request/range/source revision/profile/dependencies/ outputs/manifest/measurements/status/provenance. Evidence: DAW compatibility fixtures.
  • E1.3 Define score parts, notation events, articulation, dynamics, lyrics, chord, lead sheet, MIDI, and interchange metadata. Evidence: round-trip fixtures.
    • E1.3.a Define score/part/staff/voice/measure/time-position/event identity and pitched/unpitched/rest/grace/tuplet/tie/slur/beaming semantics. Evidence: notation schema.
    • E1.3.b Define pitch spelling/frequency/tuning/transposition/clef/key/ range plus articulation/ornament/dynamic/expression relationships. Evidence: musical event fixtures.
    • E1.3.c Define chord symbol/voicing/bass/function/scale/lead-sheet/ repeat/form/roadmap and explicit analytical assumptions. Evidence: harmony schema.
    • E1.3.d Link lyric syllables/melisma/elision/translation/pronunciation to notation events and timing without collapsing lyric revisions. Evidence: underlay fixtures.
    • E1.3.e Map MIDI/MIDI 2 note/controller/program/expression/articulation/ tempo/meter events to notation with explicit quantization/loss. Evidence: MIDI fixtures.
    • E1.3.f Define MusicXML/notation/MIDI interchange provenance/version/ unknown extensions/loss diagnostics and round-trip tolerances. Evidence: round-trip goldens.
  • E1.4 Define lyric sections, lines, syllables, stress, rhyme, language, translation, explicit-content, and timing anchors. Evidence: schema.
    • E1.4.a Define lyric document/revision/section/line/token/syllable stable identity, order, repetition, performer/voice, and text/script/language. Evidence: lyric schema.
    • E1.4.b Define normalized/pronounced syllable, stress/weight/meter/ elision/melisma and confidence/source fields. Evidence: prosody schema.
    • E1.4.c Define rhyme sound/family/scheme/position/exact-near/internal/ multilingual relation with declared pronunciation assumptions. Evidence: rhyme fixtures.
    • E1.4.d Define translation/transliteration/adaptation relationship, segment alignment, terminology/names, semantic parity, and provenance. Evidence: locale schema.
    • E1.4.e Define explicit/advisory/sensitive-topic finding versus human decision, locale/target scope, evidence, and disclosure. Evidence: content policy schema.
    • E1.4.f Define musical/sample/time anchors to score/MIDI/audio/clip ranges and orphan/ambiguous mapping through edits. Evidence: timing anchor tests.
  • E1.5 Define recording device, performer, microphone/input, room, take, latency, consent, and session provenance. Evidence: capture contract.
    • E1.5.a Define capture session/device/driver/input/channel/sample-rate/ bit-depth/buffer/clock/calibration identity and capability snapshot. Evidence: device contract.
    • E1.5.b Define performer/public-versus-protected identity, role/ instrument/voice, microphone/DI/preamp/input chain, room/location, and setup. Evidence: session schema.
    • E1.5.c Define recording/take/lane/file/channels/range/count-in/punch/ loop/dropout/overdub identity with immutable source checksums. Evidence: take schema.
    • E1.5.d Define measured/compensated round-trip/input/plugin/network latency, confidence/calibration version, clock drift, and alignment. Evidence: latency schema.
    • E1.5.e Define consent purpose/scope/territory/term/media/model/voice/ distribution/revocation/evidence separately from recording metadata. Evidence: consent contract.
    • E1.5.f Define complete app/device/audio-chain/operator/performer/time/ source/import/edit provenance with redacted projections. Evidence: capture fixtures.
  • E1.6 Define mix revision, plugin/device graph, automation, loudness, reference, notes, and approval metadata. Evidence: immutable revision tests.
    • E1.6.a Define immutable mix revision/parent/source session and exact track/media/arrangement/routing/device/preset/automation dependency manifest. Evidence: mix schema.
    • E1.6.b Define ordered typed device/routing graph, sidechains/sends/ latency compensation/tails/missing state and complete state snapshots. Evidence: graph fixtures.
    • E1.6.c Define automation/time maps and recalled manual controls with touch/write/latch provenance and deterministic playback intent. Evidence: automation snapshot.
    • E1.6.d Define loudness/peak/dynamics/spectrum/phase/headroom measurements with units/channels/range/analyzer versions separate from verdicts. Evidence: measurement schema.
    • E1.6.e Define reference tracks/segments/level matching/rights/ assumptions and anchored notes/comments/decisions. Evidence: reference contract.
    • E1.6.f Define review/approval/waiver scope/roles/exact revision/ dependencies/conditions/expiry/invalidation and immutable audit. Evidence: revision invariants.
  • E1.7 Define master spec, source mix, processing, measurements, format, delivery target, QC, and approval. Evidence: target profile invariants.
    • E1.7.a Define master job/revision identity and immutable source mix/ stems/reference/profile/target/dependency manifest. Evidence: master schema.
    • E1.7.b Define processing graph/order/state/plugins/analog steps/ calibration/latency/dither and deterministic or externally documented recall. Evidence: processing contract.
    • E1.7.c Define rendered master/container/codec/layout/sample-rate/bit- depth/channel/duration/gap/fade/checksum and alternate renditions. Evidence: format schema.
    • E1.7.d Define raw loudness/peak/dynamics/spectrum/phase/noise/artifact/ null measurements with analyzer/calibration versions. Evidence: QC measurement schema.
    • E1.7.e Define target-specific verdict/findings/remediation/unsupported/ waiver separately from measurements. Evidence: profile contract.
    • E1.7.f Define mastering review/approval/signature/exact output/ conditions/expiry and invalidation on source/profile/rights change. Evidence: target profile invariants.
  • E1.8 Define musical work, sound recording, sample, interpolation, performer, voice, model, and generated-stem rights records separately. Evidence: rights-kind validation.
    • E1.8.a Define common rights record identity/subject kind/version/ claimant/owner/licensor/source/jurisdiction/territory/term/evidence/state. Evidence: base rights schema.
    • E1.8.b Define composition/work and sound-recording/master rights with distinct parties/shares/administration/licensing uses. Evidence: work/master fixtures.
    • E1.8.c Define sample versus interpolation source/use/range/ recognizability/clearance/attribution/restrictions and derived works. Evidence: sample fixtures.
    • E1.8.d Define performer/neighboring/voice/likeness/session/union or contractual rights, consent, compensation, and revocation. Evidence: performer fixtures.
    • E1.8.e Define model/provider/training/inference/output/derivative/ disclosure rights and generated-stem source/condition/run provenance. Evidence: AI rights fixtures.
    • E1.8.f Enforce kind-specific required fields/transitions/conflicts and prohibit conversion/collapse between records. Evidence: validation suite.
  • E1.9 Define ownership splits with parties, roles, percentages, territories, rights types, dates, disputes, and signatures. Evidence: total/incompatibility invariants.
    • E1.9.a Define split sheet/version/subject work-or-recording/right type/ territory/effective window/currency and status. Evidence: split schema.
    • E1.9.b Define party identity/role/capacity/publisher/label/administrator/ representative/payment reference separately from public credit. Evidence: party schema.
    • E1.9.c Define percentage/fraction/points/recoupment/control/license/ collection shares with exact decimal/rational arithmetic. Evidence: share contract.
    • E1.9.d Define proposal/counter/accepted/signed/disputed/superseded/ expired state, comments/evidence, authority, and immutable revision lineage. Evidence: negotiation schema.
    • E1.9.e Define cryptographic/e-sign identity/intent/time/document hash, quorum, minors/representatives, revocation, and signature validity. Evidence: signature contract.
    • E1.9.f Enforce total exactly per right/territory/window, incompatible overlaps, duplicate parties, missing authority, and release snapshot invariants. Evidence: property tests.
  • E1.10 Define identifiers and metadata: ISRC, ISWC, UPC/EAN, catalog number, contributor IDs, titles, versions, genres, languages, and advisories. Evidence: format/checksum validation.
    • E1.10.a Define identifier type/issuer/value/status/subject/version/ territory/date/provenance and primary/alternate/replacement relations. Evidence: identifier schema.
    • E1.10.b Implement ISRC/ISWC/UPC/EAN syntax/check-digit/normalization/ uniqueness/scope and explicit unassigned/pending/invalid state. Evidence: identifier tests.
    • E1.10.c Define internal catalog/contributor/party/artist IDs and external mappings without using mutable display names as identity. Evidence: mapping tests.
    • E1.10.d Define title/subtitle/version/display/phonetic/sort/localized fields, contributors/roles/order/credits, and language/script. Evidence: metadata schema.
    • E1.10.e Define governed genre/mood/advisory/explicit/territory/target vocabularies with source/version/alias/deprecation. Evidence: vocabulary tests.
    • E1.10.f Validate format/checksum/collision/duplicate/reassignment/ locale/round-trip and distributor mapping. Evidence: metadata suite.
  • E1.11 Define provenance ingredients and transformations for generated, recorded, edited, mixed, and mastered audio. Evidence: C2PA/external ledger mapping.
    • E1.11.a Define immutable ingredient identity/kind/hash/rights/owner/ capture/import/generation source/version and protected/public metadata. Evidence: ingredient schema.
    • E1.11.b Define recording actions with performer/device/room/session/ calibration/consent and original take relationships. Evidence: capture lineage.
    • E1.11.c Define edit/comp/stretch/pitch/repair/separation/generation/ plugin/analog/mix/master/render actions with inputs/parameters/tools/ versions/outputs. Evidence: transformation schema.
    • E1.11.d Represent DAG relationships through clips/stems/mixes/masters/ artwork/packages while preserving ranges and many-to-many lineage. Evidence: provenance graph.
    • E1.11.e Map supported assertions/actions/ingredients/signers to C2PA and external ledgers with privacy minimization and explicit loss. Evidence: mapping matrix.
    • E1.11.f Define valid/tampered/missing/stripped/unknown/untrusted state, signature/hash verification, correction, and read-back. Evidence: provenance fixtures.
  • E1.12 Define release, product, track sequence, territory, date, pricing, rights, assets, delivery, takedown, and correction entities. Evidence: DDEX profile mapping.
    • E1.12.a Define release/product/version/type/title/label/catalog/UPC/ metadata/locale/status and immutable candidate relationship. Evidence: release schema.
    • E1.12.b Define ordered disc/side/track/hidden/bonus/component sequence with recording/version/ISRC/display/territory availability. Evidence: sequencing schema.
    • E1.12.c Define territory/window/street/preorder/embargo/time-zone/ pricing/tier/currency/entitlement and rights eligibility. Evidence: market schema.
    • E1.12.d Define master/artwork/lyrics/credits/metadata/provenance/ advisory/rendition assets with checksums and target requirements. Evidence: asset manifest.
    • E1.12.e Define per-distributor delivery/attempt/ack/content-status/ read-back/error/retry identifiers and partial aggregate state. Evidence: delivery schema.
    • E1.12.f Define takedown/redelivery/metadata correction/audio replacement/ territory change causes/actions/notices/history/verification. Evidence: remedy schema.
    • E1.12.g Map every entity/lifecycle to pinned DDEX ERN profile with supported extensions and explicit loss. Evidence: DDEX matrix.
  • E1.13 Define immutable revisions and semantic diffs for score, lyrics, arrangement, session, mix, master, rights, and release. Evidence: golden diff fixtures.
    • E1.13.a Define Euterpe's immutable audio revision/parent(s)/subject/ schema/tool/actor/cause/time/hash/approval lineage and branch/merge semantics. Evidence: revision contract.
    • E1.13.b Define score/lyrics/arrangement diffs by stable musical/text/ section identity, position, value, move, and semantic context. Evidence: creative diff fixtures.
    • E1.13.c Define session diffs for tracks/clips/takes/comp/routing/devices/ automation/tempo and preserve source media lineage. Evidence: session diff fixtures.
    • E1.13.d Define mix/master diffs for dependency graphs/plugin state/ automation/settings/renders/measurements/profile verdicts. Evidence: audio revision fixtures.
    • E1.13.e Define rights/splits/consent/provenance/release/territory/target diffs with protected redaction and material-change classification. Evidence: business diff fixtures.
    • E1.13.f Define ambiguous/unknown plugin/opaque blob/schema-version loss and test add/delete/move/split/merge/restore/concurrent goldens. Evidence: semantic diff suite.
  • E1.14 Define backward compatibility with current DAW project formats. Evidence: old-project migration corpus.
    • E1.14.a Inventory every deployed project/session/media/preset/plugin/ desktop/browser schema version and producer/consumer behavior. Evidence: version inventory.
    • E1.14.b Build representative historical corpus including large, missing- plugin/media, unusual routing/tempo/automation, corrupt, and partial projects. Evidence: migration corpus.
    • E1.14.c Define read/write/support window and deterministic migrations for renamed/split/merged/timebase/device/state fields with explicit loss. Evidence: compatibility matrix.
    • E1.14.d Migrate one version at a time and supported direct paths, preserving musical timing/media/hashes/routing/automation/undo history or documented boundaries. Evidence: migration tests.
    • E1.14.e Test old/new browser/desktop/realtime/render combinations, interrupted migration/retry/rollback/export, unknown fields, and downgrade refusal. Evidence: skew suite.
    • E1.14.f Publish backup/restore/operator diagnostics and require golden semantic/audio comparison before marking migrated. Evidence: rehearsal report.
  • E1.15 Define pitch spelling, frequency, octave, tuning/temperament, microtonal scale, key/tonality, interval, chord/voicing, rhythmic cycle, groove, form, dynamic, articulation, ornament, and cultural-source metadata. Evidence: notation/playback/analysis round-trips do not force all music into 12-tone equal temperament or Western common practice.
    • E1.15.a Define pitch as distinct written spelling, sounding frequency/ ratio/cents, octave/register, reference pitch, transposition, and tuning context. Evidence: pitch schema.
    • E1.15.b Define arbitrary scale/temperament/tuning tables, period/ nonoctave repetition, degrees/accidentals/glyphs/aliases, and Scala or ratified interchange. Evidence: tuning fixtures.
    • E1.15.c Define key/tonality/mode/raga or other system, interval/chord/ voicing/function only when applicable, with cultural-source and analysis assumptions. Evidence: tonal schema.
    • E1.15.d Define rhythmic cycle/groove/pulse/subdivision/swing/polyrhythm/ additive/free-time and meter representation without common-time floors. Evidence: rhythm fixtures.
    • E1.15.e Define form/dynamic/articulation/ornament/expression terminology, notation/playback mappings, locale, source tradition, and performer interpretation. Evidence: expression schema.
    • E1.15.f Round-trip representative culturally sourced/non-12-TET/free- rhythm notation→MIDI/control→playback→analysis with declared loss and expert review. Evidence: inclusive music goldens.
  • E1.16 Define instrument, sampler, synthesizer, effect, plugin, device, patch/preset, sample zone, modulation, control surface, MIDI/MPE mapping, latency, license, and state-snapshot contracts. Evidence: deterministic recall or explicit non-portability diagnostics.
    • E1.16.a Define device/instrument/effect/plugin identity/kind/format/ vendor/version/build/architecture/capabilities/ports/parameters and host support. Evidence: device schema.
    • E1.16.b Define synthesizer/sampler engine topology, voices/polyphony/ tuning/sample zones/articulations/layers/round robins/streaming and source rights. Evidence: instrument schema.
    • E1.16.c Define patch/preset/state blob plus normalized parameters/ resources/migrations/checksum/factory-user scope and portability. Evidence: state contract.
    • E1.16.d Define modulation source→target/polarity/depth/range/smoothing/ rate/sync/scope and typed audio/CV/MIDI/MPE/control mappings. Evidence: mapping schema.
    • E1.16.e Define control-surface port/device/profile/learn/feedback/ conflict and latency/tail/compensation/calibration state. Evidence: control contract.
    • E1.16.f Define license/signature/entitlement/install/update/quarantine/ missing/crash/safe-mode state and protected credentials. Evidence: lifecycle schema.
    • E1.16.g Validate deterministic recall hashes/audio goldens when possible and emit exact missing/version/resource/licence/opaque-state loss diagnostics otherwise. Evidence: recall suite.
  • E1.17 Define stem-separation, transcription, inpainting, continuation, style transfer, conditional generation, voice conversion, speech, singing, choir, and spatial-upmix jobs with source, mask/condition, identity/rights, model, parameters, output stems, confidence, and provenance. Evidence: job-kind fixtures and consent applicability.
    • E1.17.a Define shared immutable job request/run/attempt/status/budget/ provider-model/version/hardware/input-output/provenance contract. Evidence: base job schema.
    • E1.17.b Define separation/transcription source ranges/channels/targets/ notation/lyrics/speakers and output stems/events/alignments/confidence. Evidence: analysis job fixtures.
    • E1.17.c Define inpainting/continuation/variation/style-transfer exact mask/range/context/condition/reference/strength/preservation and generated outputs. Evidence: generative edit fixtures.
    • E1.17.d Define conditional/full generation prompt/score/MIDI/audio/ structure/style/instrument/seed/negative constraints and editable intermediates. Evidence: generation fixtures.
    • E1.17.e Define voice conversion/speech/singing/choir identity/performer/ voice model/locale/pronunciation/timing/pitch/harmony/direction and outputs. Evidence: voice job fixtures.
    • E1.17.f Define spatial-upmix source layout/object inference/target/ renderer/room/trajectory/downmix constraints and outputs. Evidence: upmix job fixtures.
    • E1.17.g Resolve source/reference/sample/model/voice/performer rights and consent scope/revocation/disclosure before dispatch; define prohibited/ nonapplicable states. Evidence: applicability tests.
    • E1.17.h Define measurements/confidence/uncertainty/quality/safety/raw findings separately from policy verdict/human selection. Evidence: result schema.
  • E1.18 Define object/channel/bed-based spatial sessions, speaker layouts, binaural render settings, ambisonic order/normalization, position/ trajectory/room metadata, head tracking, target device, downmix, and interchange. Evidence: Atmos/binaural/HOA/VR-AR profile goldens or explicit unsupported states.
    • E1.18.a Define spatial session/object/bed/channel/track identity, coordinate/frame/unit, position/orientation/spread/divergence/gain/ priority and automation. Evidence: spatial schema.
    • E1.18.b Define named/custom speaker layout/labels/angles/elevation/ distance/LFE/bed mapping, calibration, room, and target constraints. Evidence: layout schema.
    • E1.18.c Define ambisonic order/dimensionality/channel ordering/ normalization/coordinate orientation and HOA encode/decode metadata. Evidence: ambisonic schema.
    • E1.18.d Define binaural renderer/HRTF/head size/room/head tracking/ sensor/time alignment and device-specific settings. Evidence: binaural contract.
    • E1.18.e Define object/bed→speaker/binaural/ambisonic downmix/render matrices, metadata loss, loudness/peak, and fallback. Evidence: render profiles.
    • E1.18.f Define ADM-BWF/Atmos or ratified/VR-AR interchange mappings, profile/version/IDs/timing/objects and explicit unsupported fields. Evidence: interchange matrix.
    • E1.18.g Validate goldens per supported layout/renderer/device/target and serialize honest unsupported/degraded states. Evidence: spatial conformance suite.

E2 — Core DAW, editing, and device workflow#

  • E2.1 Verify sample-accurate timeline, tempo, meter, marker, loop, and transport behavior. Evidence: deterministic timing tests.
    • E2.1.a Define musical-tick/sample/time conversion, rounding, reference sample rate, tempo ramps, meter changes, pickups, and negative/preroll positions. Evidence: timing spec.
    • E2.1.b Verify event/clip/automation/marker/cue placement and rendering across tempo/meter edits without cumulative drift. Evidence: alignment tests.
    • E2.1.c Verify play/stop/pause/seek/scrub/locate/record/pre-roll/count-in state and exact audible/record sample boundaries. Evidence: transport tests.
    • E2.1.d Verify loop/cycle range, wrap, punch, metronome, latency compensation, and automation at boundaries. Evidence: loop tests.
    • E2.1.e Run long-duration, high/low/variable tempo, odd/additive meter, sample-rate conversion, offline/realtime render, and repeated-start deterministic goldens. Evidence: timing suite.
  • E2.2 Verify audio engine startup, suspend/resume, device change, sample rate, buffer size, dropout, and recovery. Evidence: device fault suite.
    • E2.2.a Define engine lifecycle/state/error contract and capability/ permission/device selection before graph activation. Evidence: lifecycle spec.
    • E2.2.b Verify cold/warm startup, user-gesture resume, background/ foreground, suspend/resume, interruption, and safe graph teardown. Evidence: lifecycle tests.
    • E2.2.c Verify input/output/default device add/remove/switch, channel-map change, permission revoke, and fallback without wrong-device capture. Evidence: hot-plug tests.
    • E2.2.d Verify sample-rate/buffer/block/clock change with resampling, latency recalibration, transport/recording handling, and explicit unsupported states. Evidence: configuration tests.
    • E2.2.e Inject CPU starvation/xruns/worklet crash/device loss/driver error/clock discontinuity and recover without corrupting project/media. Evidence: device fault suite.
  • E2.3 Build/complete track, bus, return, group, folder, send, sidechain, mute, solo, arm, monitor, and routing operations. Evidence: signal-flow tests.
    • E2.3.a Implement create/delete/duplicate/rename/reorder/nest/unnest for audio/MIDI/instrument/aux/bus/return/group/folder with stable IDs and non-drag commands. Evidence: structure tests.
    • E2.3.b Implement typed input/output/channel/layout routing with graph validation, cycle/feedback safety, disconnected/missing endpoints, and visible signal path. Evidence: routing tests.
    • E2.3.c Implement pre/post-fader sends, return routing, sidechain source/ destination/tap/channel mapping, gain/pan, enable, and automation. Evidence: send/sidechain tests.
    • E2.3.d Implement mute/solo/solo-safe/exclusive/group propagation with deterministic audible graph and no destructive state changes. Evidence: mute/solo goldens.
    • E2.3.e Implement record arm/input monitoring/auto/off/source selection with permissions, feedback prevention, latency, and multi-track capture. Evidence: monitor tests.
    • E2.3.f Cover every command with undo/redo/concurrency/persistence/ accessible grid and offline/realtime/render signal-flow goldens. Evidence: operation coverage.
  • E2.4 Build/complete clip create, import, move, trim, split, slip, duplicate, loop, stretch, reverse, fade, crossfade, consolidate, and bounce. Evidence: non-destructive edit tests.
    • E2.4.a Implement clip create/import/record-reference/duplicate with immutable source media, channel mapping, project rate, and provenance. Evidence: creation tests.
    • E2.4.b Implement move/copy across tracks/timebases/snap modes and trim start/end with bounded source offset and no media rewrite. Evidence: placement tests.
    • E2.4.c Implement split/join/slip/slide/duplicate/loop/repeat and preserve take/comp/anchor/automation relationships or explicit impact. Evidence: structural edit tests.
    • E2.4.d Implement nondestructive time stretch/pitch/formant/reverse with algorithm/version/quality/transient/tempo context and render provenance. Evidence: processing tests.
    • E2.4.e Implement fade/crossfade shapes/lengths/gains/equal-power/ overlapping source limits and sample-continuity tests. Evidence: fade goldens.
    • E2.4.f Implement consolidate/bounce to immutable derived media with exact range/tails/routing/effects/automation/profile and source manifest. Evidence: render lineage.
    • E2.4.g Verify semantic diff/undo/redo/restore, pointer/non-drag parity, sample boundary/null tests, and source bytes unchanged. Evidence: non-destructive suite.
  • E2.5 Build/complete MIDI record, draw, quantize, groove, velocity, articulation, expression, sustain, MPE, and humanize. Evidence: event round-trip.
    • E2.5.a Implement timestamped MIDI 1/MIDI 2 UMP/MPE recording with port/ channel/group/note identity, latency compensation, overdub/replace/ loop, and raw capture. Evidence: record tests.
    • E2.5.b Implement note/rest/controller/program/pitch/sustain/aftertouch/ per-note expression draw/edit/delete/duplicate with musical/sample timing. Evidence: editor tests.
    • E2.5.c Implement quantize for start/end/duration/strength/swing/iterative rules and reversible source-versus-playback timing. Evidence: quantize tests.
    • E2.5.d Implement groove extraction/application, velocity curves, articulation maps, sustain interpretation, and instrument-specific expression. Evidence: performance tests.
    • E2.5.e Implement seeded bounded humanize for timing/velocity/duration/ expression with preview, repeatability, and no note collisions/stuck notes. Evidence: humanize tests.
    • E2.5.f Round-trip record→edit→save→export/import→playback across MIDI/ MPE devices with undo and explicit conversion loss. Evidence: event suite.
  • E2.6 Build/complete score notation, parts, layout, transposition, articulation, lyric underlay, playback, and print/export. Evidence: notation goldens.
    • E2.6.a Implement score/part/staff/voice/measure/event authoring for pitched/unpitched/grace/tuplet/tie/slur/repeat/free-time structures. Evidence: notation editor tests.
    • E2.6.b Implement clef/key/tuning/accidental/spelling/transposition/ instrument range and concert/written pitch without 12-TET-only assumptions. Evidence: pitch tests.
    • E2.6.c Implement articulation/ornament/dynamics/expression/technique and playback mapping with instrument/library capability and fallback. Evidence: expression goldens.
    • E2.6.d Implement lyric underlay for syllable/melisma/elision/verse/ translation/pronunciation and synchronized lyric timing. Evidence: lyric tests.
    • E2.6.e Implement page/system/staff/spacing/break/visibility/cue/layout rules, extracted parts, accessibility semantics, and print preview. Evidence: layout tests.
    • E2.6.f Export/import MusicXML/PDF/print/MIDI or ratified formats with deterministic notation/playback comparison and loss diagnostics. Evidence: notation goldens.
  • E2.7 Build/complete arrangement sections, scenes, variants, and section-level move/copy. Evidence: marker/clip/automation coherence.
    • E2.7.a Implement stable section/scene/variant identity, nested/ordered structure, musical range/repetitions/color/label/cue and aliases. Evidence: structure tests.
    • E2.7.b Create/resize/split/merge/duplicate/reorder/move/copy sections with explicit ripple versus reference behavior and non-drag commands. Evidence: edit tests.
    • E2.7.c Move/copy affected clips/takes/comp/MIDI/score/lyrics/markers/ tempo-meter-key/automation consistently or preview exclusions. Evidence: coherence tests.
    • E2.7.d Create/switch/compare arrangement variants using shared media and immutable lineage without cross-variant mutation. Evidence: variant tests.
    • E2.7.e Verify undo/redo/concurrency/export/render/session-scene aliases and exact marker/clip/automation timing. Evidence: arrangement suite.
  • E2.8 Build/complete comping with take lanes, audition, swipe/segment selection, crossfades, and source trace. Evidence: comp provenance.
    • E2.8.a Create ordered take lanes from loop/punch/overdub/import with immutable files, performer/session/device/timing/dropout provenance. Evidence: lane tests.
    • E2.8.b Implement lane/take/range audition with level/pan/routing/ latency context, solo-safe restoration, and keyboard controls. Evidence: audition tests.
    • E2.8.c Implement swipe and form/table-based segment selection with exact source/timeline ranges, overlaps, gaps, priority, and destination. Evidence: selection tests.
    • E2.8.d Generate/edit adjacent crossfades with bounded handles and click-free sample continuity; preserve original takes. Evidence: crossfade goldens.
    • E2.8.e Persist comp revision as ordered source-segment manifest and map anchors through take edits/moves or mark stale. Evidence: provenance tests.
    • E2.8.f Verify flatten/bounce/restore/undo/redo/concurrent edits and null/sample comparisons against source segments. Evidence: comp suite.
  • E2.9 Build/complete automation for track, bus, plugin, tempo, and spatial parameters. Evidence: read/write/touch/latch and curve tests.
    • E2.9.a Discover/register automatable parameters with stable device/path/ unit/range/default/step/bipolar/logarithmic/discrete semantics. Evidence: parameter registry.
    • E2.9.b Implement lane create/show/hide/arm and point add/move/delete/ range/scale/copy with musical/sample time and non-drag forms/commands. Evidence: lane editing tests.
    • E2.9.c Implement step/linear/bezier/hold curves, tangents, smoothing, value clamping, sample/block interpolation, and deterministic playback. Evidence: curve goldens.
    • E2.9.d Implement read/off/write/touch/latch/trim modes with gesture capture, return/ramp, punch, loop, undo, and visible recording state. Evidence: mode tests.
    • E2.9.e Handle tempo and spatial/multidimensional parameters, clip versus track automation, section moves, device reorder/remove/missing/version migration. Evidence: lifecycle tests.
    • E2.9.f Verify realtime/offline render parity, latency/feedback safety, persistence/round-trip, and accessible semantic list. Evidence: automation suite.
  • E2.10 Build/complete plugin/device graph with latency, bypass, preset, missing-plugin, sandbox, crash, and version behavior. Evidence: fault suite.
    • E2.10.a Implement typed device node/port/parameter/state graph with ordered serial/parallel/rack/split/merge/sidechain routing and cycle validation. Evidence: graph tests.
    • E2.10.b Measure/report plugin/device latency and tail, compute graph compensation including live/record paths, and update safely on change. Evidence: latency tests.
    • E2.10.c Implement hard/soft/bypass/suspend behavior with gain/click/tail/ latency semantics and automation. Evidence: bypass goldens.
    • E2.10.d Load/save/compare/migrate factory/user/project presets and full state snapshots with hashes/resources/version/license. Evidence: preset tests.
    • E2.10.e Preserve missing plugin placeholder/state/routing/automation and support locate/substitute/reinstall/version migrate with impact preview. Evidence: missing-device tests.
    • E2.10.f Run plugins in ratified sandbox with CPU/memory/time/file/ network/device limits, crash detection/quarantine/restart/safe mode. Evidence: sandbox tests.
    • E2.10.g Inject scan/load/process/UI/state/crash/hang/malicious/version/ licence failures and prove project/audio recovery. Evidence: fault suite.
  • E2.11 Build/complete asset and sample browser with preview, tempo/key, rights, favorites, tags, search, dedupe, and project use. Evidence: permission-aware search.
    • E2.11.a Define asset/sample source/revision/blob/proxy/waveform/duration/ format/channel/sample-rate/tempo/key/tuning/tags/rights/provenance projection. Evidence: browser contract.
    • E2.11.b Ingest/scan/analyze/deduplicate immutable local/project/library/ provider assets with exact versus derived metadata and confidence. Evidence: ingest tests.
    • E2.11.c Implement permission-aware text/facet/semantic search, stable pagination, favorites/collections/tags, and no unreleased/protected count/snippet leaks. Evidence: search tests.
    • E2.11.d Implement tempo/key-aware low-latency preview with original/ processed comparison, start/loop/level/output controls, and device recovery. Evidence: preview tests.
    • E2.11.e Add to project/clip/sampler with immutable source link, rights/ territory/use prompt, transform/provenance, and duplicate handling. Evidence: use journey.
    • E2.11.f Handle removed/revoked/expired/offline/missing/provider-failed assets with honest state and replacement/locate workflow. Evidence: lifecycle tests.
  • E2.12 Build/complete freeze, flatten, bounce-in-place, stem export, and restore. Evidence: source/revision lineage.
    • E2.12.a Define operation/range/source revision/signal tap/profile/tails/ dependencies/destination/naming/rights/provenance and impact preview. Evidence: render plan.
    • E2.12.b Freeze track/device chain to immutable render while preserving editable graph/state/automation/media and latency/routing behavior. Evidence: freeze tests.
    • E2.12.c Flatten only after explicit warning/authorization into a new revision with restorable source snapshot and declared lost editability. Evidence: flatten tests.
    • E2.12.d Bounce-in-place clips/tracks/regions with exact timing/tails/ routing/effects/automation/channel/sample rate and source links. Evidence: bounce goldens.
    • E2.12.e Export deterministic stems by track/bus/group/role/range with alignment, silence/tail policy, manifest, metadata, rights, and hashes. Evidence: stem tests.
    • E2.12.f Restore/unfreeze to exact source revision/state, reconcile later edits, and verify null/semantic comparison plus undo/redo. Evidence: lineage suite.
  • E2.13 Build/complete undo/redo across all structured editing commands. Evidence: command inventory coverage.
    • E2.13.a Generate canonical command inventory with input/preconditions/ effects/inverse/merge boundary/side effects and undo eligibility. Evidence: command ledger.
    • E2.13.b Implement atomic and transaction/grouped undo for timeline/ mixer/device/score/lyrics/rights/release structured mutations. Evidence: command tests.
    • E2.13.c Preserve immutable media/renders/uploads/jobs/external delivery by undoing references or issuing compensating/new revisions rather than deleting history. Evidence: side-effect tests.
    • E2.13.d Define per-user/session/project undo scope under collaboration, remote operations, branch/merge, offline replay, and conflict. Evidence: concurrency tests.
    • E2.13.e Clear/fork/restore stacks correctly after undo→new edit, reload/crash/schema migration, failed command, and partial transaction. Evidence: lifecycle tests.
    • E2.13.f Test command→undo→redo semantic/audio equality, focus/selection restoration, keyboard announcements, and inventory coverage. Evidence: coverage gate.
  • E2.14 Build/complete autosave, crash recovery, project snapshots, version history, compare, and restore-as-new-version. Evidence: crash journey.
    • E2.14.a Define local operation/checkpoint/server revision/media-upload/ snapshot/schema/device/session/encryption/retention state. Evidence: recovery contract.
    • E2.14.b Implement debounced idempotent autosave with expected revision, durable-before-ack, dirty/saving/saved/error/offline/conflict indicators. Evidence: autosave tests.
    • E2.14.c Persist local journals/media safely before risky recording/edit/ plugin/render actions and recover after tab/app/process/device crash. Evidence: crash tests.
    • E2.14.d Create named/automatic immutable project snapshots with complete dependency/media/plugin/right manifests and retention/cleanup. Evidence: snapshot tests.
    • E2.14.e Show version history and semantic/audio-aware compare across arrangement/session/mix/rights/release with missing dependency state. Evidence: history tests.
    • E2.14.f Restore as a new revision, revalidate plugins/media/rights/gates/ approvals, retain current and historical versions, and test crash journey. Evidence: recovery E2E.
  • E2.15 Build/complete keyboard command system with remapping and collision audit. Evidence: pointer-free core edit journey.
    • E2.15.a Define semantic command registry with stable IDs/context/scope/ availability/default bindings/platform variants and descriptions. Evidence: command schema.
    • E2.15.b Implement layered global/workspace/editor/device/plugin/modal resolution with text-input/assistive technology/browser/OS reserved-key handling. Evidence: resolver tests.
    • E2.15.c Build searchable keyboard preferences for add/remove/remap/reset/ import/export profiles with immediate conflict and unreachable-command diagnostics. Evidence: settings tests.
    • E2.15.d Support chords/sequences/modifiers/repeat/hold where safe and announce unavailable/context-changing/destructive commands. Evidence: interaction tests.
    • E2.15.e Audit exact/shadowed/prefix/context/platform/plugin collisions and fail required commands with no usable binding. Evidence: collision report.
    • E2.15.f Run pointer-free create/edit/record/mix/save/review/export/ recovery journey with focus and screen-reader compatibility. Evidence: keyboard E2E.
  • E2.16 Build/complete non-drag alternatives for clip, automation, routing, section, and marker operations. Evidence: WCAG 2.2 dragging tests.
    • E2.16.a Inventory every drag/drop/move/trim/resize/reorder/connect/curve/ range operation and define equivalent semantic result. Evidence: drag inventory.
    • E2.16.b Add numeric/text forms and keyboard step commands for clip/ region/marker positions, ranges, offsets, fades, loops, and lanes. Evidence: timeline tests.
    • E2.16.c Add point/value/time/curve forms, ordered tables, nudge/scale/ delete commands for automation. Evidence: automation tests.
    • E2.16.d Add source/destination/port/tap/channel/gain/sidechain forms and before/after/into/out commands for routing/device graph. Evidence: routing tests.
    • E2.16.e Add before/after/position/range/move-copy form commands for sections/scenes/markers with affected-item preview. Evidence: structure tests.
    • E2.16.f Keep validation/snap/impact/permissions/concurrency/undo/audit identical and run keyboard/screen-reader/touch journeys. Evidence: WCAG 2.2 suite.
  • E2.17 Build/complete semantic accessible grid/list views for timeline, mixer, device graph, and automation. Evidence: screen-reader journey.
    • E2.17.a Define synchronized content/operation/selection/focus/playhead/ status parity contract between visual DAW surfaces and semantic views. Evidence: parity matrix.
    • E2.17.b Build timeline treegrid/table by track/lane/clip/region/take/ marker with musical/sample positions, ranges, state, and edit commands. Evidence: timeline semantic tests.
    • E2.17.c Build mixer grid/table by track/channel for level/pan/mute/solo/ arm/monitor/sends/meters/routing with units and bounded controls. Evidence: mixer semantic tests.
    • E2.17.d Build device/routing graph tree/table/matrix for nodes/ports/ edges/parameters/latency/errors plus connect/reorder/configure commands. Evidence: graph semantic tests.
    • E2.17.e Build automation table by parameter/point/time/value/curve/mode with add/edit/nudge/delete/range operations. Evidence: automation semantic tests.
    • E2.17.f Preserve focus/selection/playback announcements under virtualization/live updates and run end-to-end screen-reader edit/mix/ route/automate journey. Evidence: semantic E2E.
  • E2.18 Define and meet interactive latency, audio glitch, large-project, memory, and render performance budgets. Evidence: profiler reports.
    • E2.18.a Define supported host/device/buffer/sample-rate/project/network/ plugin scenarios and measurement harness/calibration. Evidence: performance profile.
    • E2.18.b Set input-to-monitor, control-to-audible, transport, edit, UI input/paint, waveform/scroll/zoom interaction latency budgets. Evidence: latency budgets.
    • E2.18.c Set xrun/dropout/click/missed-deadline/CPU audio-thread and realtime allocation/lock/GC budgets with safe instrumentation. Evidence: glitch budgets.
    • E2.18.d Set large track/clip/device/automation/score/media project load/ save/search/render and sustained CPU/memory/GPU/disk/network budgets. Evidence: scale budgets.
    • E2.18.e Set realtime/offline/batch render speed/resource/output- determinism and worker concurrency/cost budgets. Evidence: render budgets.
    • E2.18.f Profile production-scale fixtures, capture thread/trace/query/ memory/audio diagnostics, trend baselines, and fail regressions. Evidence: profiler reports.
  • E2.19 Build/complete audio recording and waveform editing for device/ channel selection, calibration, pre-roll, punch, loop recording, take lanes, comp, sample-accurate cut, spectral repair, time/pitch processing, normalization, denoise, phase alignment, and non-destructive source restore. Evidence: calibrated capture/edit fixtures.
    • E2.19.a Build device/input/channel/clock/sample-rate/buffer permission setup, signal level/phase/noise/latency calibration, and safe monitoring. Evidence: setup tests.
    • E2.19.b Implement count-in/pre-roll/manual-auto punch/loop/multitrack/ overdub capture with sample-accurate files, dropout markers, and crash- safe durable takes. Evidence: recording tests.
    • E2.19.c Generate waveform/peak/RMS/spectral caches from immutable media with channel/time/zoom selection and accessible numeric/semantic views. Evidence: waveform tests.
    • E2.19.d Implement sample-accurate cut/copy/delete/silence/gain/fade/ normalize/channel operations as nondestructive revisions. Evidence: waveform edit tests.
    • E2.19.e Implement spectral selection/repair/attenuation/heal with exact time-frequency masks, algorithm/version, preview, and preserved untouched samples when promised. Evidence: spectral goldens.
    • E2.19.f Implement time/pitch/formant/denoise/declick/dehum restoration with settings/quality/latency/tails, A/B, and provenance. Evidence: processing tests.
    • E2.19.g Implement multi-mic/take phase/time alignment with polarity/ delay/correlation/confidence and manual override, not forced aesthetic choice. Evidence: alignment tests.
    • E2.19.h Restore exact source or prior edit revision and verify calibrated capture/edit/null/undo/device-failure fixtures. Evidence: end-to-end corpus.
  • E2.20 Build/complete virtual instruments and samplers for oscillator/ synthesis engines, modulation, envelopes, filters, sample maps, articulation, polyphony, tuning, preset/version management, automation, freeze, and missing-device recovery. Evidence: state recall and render goldens.
    • E2.20.a Implement ratified synthesis engines/oscillators/noise/wavetable/ granular or other sources with oversampling/alias/CPU/polyphony/voice semantics. Evidence: engine tests.
    • E2.20.b Implement envelopes/LFOs/modulation matrices/filters/amplifier/ pan/unison/voice allocation/glide and per-voice/global state. Evidence: synthesis tests.
    • E2.20.c Implement sampler zones/key-velocity ranges/round robin/layers/ loops/crossfades/root pitch/tuning/articulation/streaming and immutable sample rights. Evidence: sample-map tests.
    • E2.20.d Implement polyphony/voice stealing/sustain/MIDI/MPE/per-note expression/microtonal tuning and deterministic event handling. Evidence: performance goldens.
    • E2.20.e Implement versioned preset/state/resources/licence/automation/ modulation mappings and compare/morph where supported. Evidence: recall tests.
    • E2.20.f Implement freeze/render with exact MIDI/audio/device state/ tails/latency/provenance and reversible source. Evidence: freeze tests.
    • E2.20.g Preserve missing-device state/resources/automation/routing and support locate/substitute/migrate/safe mode with loss diagnostics. Evidence: recovery tests.
    • E2.20.h Verify save/reload/version/host/sample-rate/render audio goldens and CPU/memory budgets. Evidence: instrument suite.
  • E2.21 Build/complete effect and signal-chain authoring for EQ, dynamics, saturation, delay, reverb/convolution, modulation, pitch, restoration, metering, parallel/mid-side/multiband routing, oversampling, latency compensation, A/B, preset, and bypass. Evidence: signal-flow and deterministic-state tests.
    • E2.21.a Implement ratified EQ/filter/dynamics/gate/limiter/de-esser/ saturation/distortion families with units/ranges/channels/detectors/ lookahead. Evidence: processor tests.
    • E2.21.b Implement delay/reverb/convolution/modulation/pitch/restoration families with sync/free timing, IR/resource rights, tails, quality, and latency. Evidence: time/effect tests.
    • E2.21.c Implement ordered serial/parallel/send-return/mid-side/multiband/ sidechain signal chains with split/crossover/mix/phase/feedback safety. Evidence: routing goldens.
    • E2.21.d Implement input/output/GR/loudness/spectrum/phase meters with calibration/units/hold/reset and no audio-thread blocking. Evidence: meter tests.
    • E2.21.e Implement oversampling/quality modes and complete plugin/device latency/tail compensation across live, record, bypass, freeze, and offline render. Evidence: compensation tests.
    • E2.21.f Implement level-matched A/B/compare, versioned presets/full state, automation/modulation, safe bypass, and undo. Evidence: interaction tests.
    • E2.21.g Run impulse/step/sine/noise/reference/null/signal-flow goldens across sample rates/channels/hosts and deterministic state reload. Evidence: effect suite.
  • E2.22 Build/complete MIDI, MPE, clock, sync, control-surface, pedal, hardware instrument, and external effect setup with discovery, mapping, learn, calibration, latency, reconnect, conflict, safe feedback, and project portability. Evidence: supported-device matrix and virtual-device automation.
    • E2.22.a Discover/identify/authorize hot-plug MIDI 1/MIDI 2 UMP/MPE/ control/sync/hardware ports with stable aliases and privacy-safe labels. Evidence: discovery tests.
    • E2.22.b Configure input/output/channel/group/MPE zones/filter/transform/ routing and project/global/default profiles. Evidence: routing tests.
    • E2.22.c Build learn/mapping for notes/CC/NRPN/RPN/program/pitch/pressure/ per-note expression/transport/control-surface feedback with conflicts. Evidence: mapping tests.
    • E2.22.d Configure MIDI clock/MTC/MMC/Link or ratified sync authority, tempo/timecode start/stop/locate/drift/jitter and safe feedback loops. Evidence: sync tests.
    • E2.22.e Configure pedals/hardware instruments/external effects with audio/MIDI routing, patch recall, latency/calibration, reamp/send-return, and recording provenance. Evidence: hardware tests.
    • E2.22.f Handle disconnect/reconnect/renamed/duplicate ports, permission revoke, stuck notes, runaway feedback, clock loss, and safe reset. Evidence: fault tests.
    • E2.22.g Serialize portable logical mappings plus device-specific bindings/substitution/loss diagnostics and test virtual devices across supported hosts. Evidence: device matrix.
  • E2.23 Build/complete tuning, notation, scale/mode, world-rhythm, microtonal, transposition, chord/voicing, orchestration/range, form, and expression editors without Western-only validation assumptions. Evidence: representative culturally sourced and non-12-TET goldens plus named musical-domain review.
    • E2.23.a Build tuning/temperament/reference frequency/scale degree/ accidental/glyph/keyboard-layout editor and audition across score/MIDI/ instruments. Evidence: tuning tests.
    • E2.23.b Build notation/spelling/clef/part/transposition views that keep written and sounding pitch plus arbitrary tunings distinct. Evidence: notation tests.
    • E2.23.c Build mode/raga/maqam or ratified culturally sourced system metadata, ascending/descending/context/ornament/phrase rules as optional declared assistance. Evidence: mode fixtures.
    • E2.23.d Build rhythmic cycle/additive/polyrhythm/groove/swing/free-time editors and metronome/grid/playback without Western meter validation floors. Evidence: rhythm goldens.
    • E2.23.e Build chord/voicing/range/orchestration/form/expression assistance with instrument/cultural assumptions, configurable rules, and creative exception. Evidence: analysis tests.
    • E2.23.f Import/export/playback/analyze representative non-12-TET/world- rhythm/culturally sourced fixtures and report exact loss/unsupported state. Evidence: round-trip suite.
    • E2.23.g Obtain named musical-domain/cultural review and prohibit universal aesthetic score or automatic correction. Evidence: review approval.
  • E2.24 Build/complete a session/clip-launch workspace with clips by scene, per-clip and scene launch, quantization, legato, follow/next actions, probability, capture-to-arrangement, arrangement-to-session aliases, recording, performance lock, and accessible grid operation. Evidence: sample-accurate launch/capture and pointer-free journeys.
    • E2.24.a Define session grid track/scene/slot/clip/alias identity and shared source relationship with arrangement. Evidence: session model.
    • E2.24.b Implement clip/scene launch/stop/retrigger/queue with global/ scene/clip quantization, legato phase, exclusive/multi launch, and exact clock authority. Evidence: launch tests.
    • E2.24.c Implement deterministic seeded follow/next/random/probability/ repeat actions with bounds, stop conditions, and visible upcoming state. Evidence: action tests.
    • E2.24.d Record MIDI/audio/automation and capture live launches/ parameters into arrangement with sample-accurate event/clip/tempo lineage. Evidence: capture tests.
    • E2.24.e Create arrangement-to-session aliases/copies with explicit update/independence rules and no accidental source mutation. Evidence: round-trip tests.
    • E2.24.f Implement performance lock, safe allowed controls, undo boundary, failure recovery, and transport/device/scene emergency stop. Evidence: performance tests.
    • E2.24.g Provide accessible treegrid/table and non-drag clip/scene move/ launch/configure commands with focus and status announcements. Evidence: pointer-free E2E.
  • E2.25 Build/complete unified modulation for LFO, envelope, MSEG, envelope/audio follower, step/random, macro/XY, MIDI/MPE, and automation sources routed to any eligible parameter with polarity, depth, smoothing, rate/sync, per-voice/global scope, visualization, and cycle/range safety. Evidence: source→target matrix coverage and deterministic renders.
    • E2.25.a Define typed modulation source/instance/state for every named generator/control, ownership, per-voice/global scope, phase/reset, and deterministic seed. Evidence: source contracts.
    • E2.25.b Implement LFO/envelope/MSEG/step/random generators with shapes/ points/curves/levels/timing/loop/one-shot/retrigger/rate/free-sync. Evidence: generator tests.
    • E2.25.c Implement envelope/audio followers with input tap/channel/ detector/attack/release/range and macro/XY aggregation. Evidence: follower/macro tests.
    • E2.25.d Implement MIDI/MPE/control/automation sources with channel/note/ expression mapping, resolution, pickup, smoothing, and feedback safety. Evidence: control tests.
    • E2.25.e Discover eligible typed targets and connect with polarity/depth/ offset/range/curve/smoothing/combine order and exact units. Evidence: routing tests.
    • E2.25.f Detect cycles/runaway/audio-rate aliasing/range overflow/ unavailable targets and enforce CPU/polyphony limits. Evidence: safety tests.
    • E2.25.g Visualize current/base/modulated value and provide semantic source-target matrix, keyboard forms, automation handoff, undo, preset/ state recall. Evidence: accessible UI tests.
    • E2.25.h Run complete source→target/profile/sample-rate/realtime-offline matrix with seeded deterministic render goldens. Evidence: coverage report.
  • E2.26 Build/complete native VST3, CLAP, AU and other ratified plugin hosting plus browser WAM where supported, including scan, signature/ license, architecture, sandbox, latency/tail, sidechain, MIDI/MPE, automation/modulation, preset/state, UI, crash/quarantine, missing plugin, version migration, and safe-mode recovery. Evidence: per-format/plugin/ host compatibility and malicious-plugin tests.
    • E2.26.a Ratify plugin formats/versions/OS/architectures/browser host profiles, SDK/licence requirements, and explicit unsupported cells. Evidence: hosting matrix.
    • E2.26.b Implement isolated scan/discovery/metadata/capability/port/ parameter/category/vendor/version/hash/signature/licence inventory with timeout/crash quarantine. Evidence: scan tests.
    • E2.26.c Host realtime/offline audio, MIDI/MPE/note expression, sidechain/ multibus/channel layouts, latency/tail/dynamic IO, and sample-rate/block changes. Evidence: processing tests.
    • E2.26.d Host parameter automation/modulation/gesture/state/preset/ resources/program lists with stable IDs, migration, and deterministic recall where possible. Evidence: state tests.
    • E2.26.e Host plugin UI/editor/window/scaling/focus/keyboard/accessibility safely and provide generic parameter UI/semantic fallback. Evidence: UI tests.
    • E2.26.f Sandbox native/browser code with file/network/device/process/ memory/CPU/time permissions, content security, feedback protection, and audit. Evidence: sandbox tests.
    • E2.26.g Preserve missing/quarantined/licence-expired/wrong-architecture/ newer-version placeholder, state/routing/automation and substitution/ reinstall/migrate diagnostics. Evidence: recovery tests.
    • E2.26.h Inject malicious/hanging/crashing/corrupt-state/huge-allocation/ filesystem/network/UI/scan/plugin-update attacks and recover project/ audio safely. Evidence: adversarial tests.
    • E2.26.i Run per-format/plugin/host/sample-rate/layout/online-offline/ save-reload/render compatibility cells with current evidence. Evidence: compatibility gate.
  • E2.27 Build/complete ARA2 or ratified deep-edit integration and per-clip warp markers for time/pitch, tempo mapping, formants, transient/grid alignment, comp/take context, plugin state, undo, and round-trip. Evidence: supported editor matrix and non-destructive audio goldens.
    • E2.27.a Ratify ARA2/deep-edit protocols/editors/versions/host profiles/ features and define document/region/track/media ownership. Evidence: editor matrix.
    • E2.27.b Exchange immutable audio source, musical context/tempo/meter/ tuning, clip/take/comp ranges, selection, and change notifications with version/authorization. Evidence: integration tests.
    • E2.27.c Persist deep-editor document/plugin state, analysis caches, resources, undo transactions, version migration, missing editor, and safe mode. Evidence: state tests.
    • E2.27.d Implement per-clip warp/transient anchors with source/destination time, lock/type/strength/confidence, monotonic/range/grid validation. Evidence: warp schema/tests.
    • E2.27.e Implement time/pitch/formant/transient/grid/tempo mapping edits and comp/take propagation as nondestructive operations with A/B. Evidence: edit goldens.
    • E2.27.f Coordinate host/plugin undo/redo/save/reload/offline render and concurrent clip/tempo/take changes without double application. Evidence: lifecycle tests.
    • E2.27.g Round-trip original→edit→project save→reopen→render/export and verify untouched regions/declared transforms/audio tolerances. Evidence: non-destructive goldens.
    • E2.27.h Exercise unsupported/missing/crash/corrupt/version-skew/ unavailable-analysis states with preserved source and explicit loss. Evidence: fault matrix.
  • E2.28 Build/complete open and professional interchange for DAWproject, Standard MIDI/MIDI 2.0 UMP, MusicXML/notation, BWF/iXML, AAF, OMF, FCPXML, MXF, ADM-BWF, stems, and every ratified format. Evidence: automation, fades, warp, devices/plugin state, note expression, sends, tempo/meter, markers, clips/scenes, metadata, and loss diagnostics are tested per format.
    • E2.28.a Ratify format/profile/version/import-export scope, libraries/ validators, licensing, and supported host/target cells. Evidence: interchange registry.
    • E2.28.b Implement DAWproject import/export for project graph/tracks/ channels/devices/clips/scenes/tempo/meter/automation/media/metadata with portable plugin state where defined. Evidence: DAWproject tests.
    • E2.28.c Implement Standard MIDI/MIDI 2 UMP and MusicXML/notation for events/expression/articulation/tuning/tempo/meter/key/lyrics/parts with explicit quantization/loss. Evidence: musical interchange tests.
    • E2.28.d Implement BWF/iXML/stems/ADM-BWF/MXF for audio/timecode/channel/ spatial/scene/take/track/rights/provenance metadata. Evidence: audio interchange tests.
    • E2.28.e Implement AAF/OMF/FCPXML or ratified editorial formats for tracks/clips/ranges/handles/fades/gain/automation/markers/media/ timecode with relink. Evidence: editorial interchange tests.
    • E2.28.f Define normalized intermediate mapping for devices/plugin state/ sends/warp/note expression/scenes/spatial/metadata and exact unsupported/ approximated/dropped fields. Evidence: mapping contracts.
    • E2.28.g Generate preflight/loss profile, require human acceptance where material, preserve source, provenance, rights, IDs, and reversible import namespace. Evidence: diagnostics tests.
    • E2.28.h Run format→project→format and cross-host semantic/audio round- trips with missing media/plugin/version/corrupt/large fixtures and validators. Evidence: per-format gate.
  • E2.29 Build/complete synchronization and control for Ableton Link or ratified peer sync, MTC, LTC/SMPTE, MIDI clock, word clock/jam sync, MMC, OSC, MCU/HUI/EUCON or ratified control surfaces, MIDI learn, and headless REST/control APIs. Evidence: clock-drift, feedback-loop, reconnect, authority, and unsafe-remote-command tests.
    • E2.29.a Ratify supported protocols/versions/transports/devices/hosts/ frame rates and define capability/security/latency profiles. Evidence: sync/control matrix.
    • E2.29.b Implement peer tempo/phase/quantum sync with explicit leader/ peer authority, join/leave, start-stop behavior, drift/jitter smoothing, and musical-time mapping. Evidence: peer-sync tests.
    • E2.29.c Implement MTC/LTC-SMPTE/MIDI clock/word-clock or jam-sync source/destination/frame-rate/drop-frame/offset/locate/chase/freewheel/ holdover. Evidence: clock tests.
    • E2.29.d Implement MMC/OSC/control-surface transport/mixer/device/ automation commands, display/meter feedback, touch/ownership, and profiles. Evidence: control tests.
    • E2.29.e Implement MIDI learn and mapping conflict/pickup/feedback/ persistence/portable aliases; prevent echo/runaway/stuck controls. Evidence: learn tests.
    • E2.29.f Implement authenticated scoped rate-limited headless REST/ realtime control with idempotency, expected state, allowlists, step-up for dangerous actions, and audit. Evidence: remote API tests.
    • E2.29.g Exercise drift/jitter/packet loss/disconnect/reconnect/leader conflict/frame-rate change/feedback loop/replay/unsafe command and emergency stop. Evidence: adversarial matrix.
  • E2.30 Build/complete offline-first PWA/OPFS and desktop-local operation for projects, media, undo, renders, collaboration queue, provider-gated features, quota/eviction, encryption, reconnect, merge, and recovery. Evidence: offline create/edit/reopen/export plus conflict and storage-loss journeys.
    • E2.30.a Define local project/media/operation/undo/render/job/cache/ collaboration state, schema/version, ownership, encryption, and supported offline capabilities per host. Evidence: offline contract.
    • E2.30.b Implement PWA service worker/OPFS and desktop-local durable storage with atomic writes, content hashes, journals, locks, crash-safe recording/import, and migrations. Evidence: local storage tests.
    • E2.30.c Implement offline create/open/edit/record/import/undo/render/ export for eligible features and honest disabled/provider-required/ licence-check states. Evidence: offline workflow tests.
    • E2.30.d Manage media/proxy/cache/render quota, reservation, pressure, user cleanup/export, browser eviction detection, and never evict sole acknowledged originals silently. Evidence: quota tests.
    • E2.30.e Encrypt local sensitive/unreleased/voice/rights data, partition by tenant/user/project, enforce device/session/logout/revocation/ retention, and protect paths. Evidence: security tests.
    • E2.30.f Queue collaboration/uploads/jobs with stable IDs/dependencies/ retries/cancel and reauthenticate/reauthorize/deduplicate/rebase on reconnect. Evidence: reconnect tests.
    • E2.30.g Resolve project/media/graph/rights conflicts with mine/theirs/ base semantic merge, retained copies, and no last-write-wins data loss. Evidence: conflict tests.
    • E2.30.h Run offline create→record/edit→close/reopen→render/export plus reconnect/two-client/schema/quota/eviction/storage-corruption/recovery journeys. Evidence: offline E2E.
  • E2.31 Build/complete customizable, multi-window, touch/pen, high-DPI, control-surface, and blind-producer workflows across arrange, session, piano roll, score, mixer, modulation, plugin, and review surfaces. Evidence: saved workspace, keyboard/treegrid, screen-reader, touch-target, zoom, and reconnect matrices.
    • E2.31.a Define versioned workspace layout with panels/windows/tabs/ docks/sizes/focus/context/monitor/display/DPI, named presets, migration, reset, and safe bounds. Evidence: workspace schema.
    • E2.31.b Implement multi-window detach/attach/focus/selection/transport/ project synchronization, restore after display loss, and no duplicate audio engine or mutation owner. Evidence: window tests.
    • E2.31.c Implement touch/pen gestures/targets/hover alternatives/ precision controls/palm rejection and non-gesture forms across all named surfaces. Evidence: touch/pen tests.
    • E2.31.d Implement high-DPI/per-monitor scaling/zoom/reflow/canvas text/ waveform/plug-in UI behavior without blurry or unreachable controls. Evidence: DPI tests.
    • E2.31.e Integrate control-surface focus/bank/track/device/parameter/ transport feedback with visible ownership and keyboard parity. Evidence: hardware workflow tests.
    • E2.31.f Provide blind-producer semantic treegrid/table/forms, command search, structured status/meters, non-drag edits, and concise announcements across all named workspaces. Evidence: screen-reader E2E.
    • E2.31.g Test saved layout/profile/project reopen, reconnect/device/ display change, themes/density, 200–400% zoom, RTL/long locale, and supported input modes. Evidence: workspace matrix.
  • E2.32 Build/complete batch render/export/convert/analyze with range, version, target/profile, naming, metadata, rights, dependencies, capacity, cost, preview, queue, partial result, retry, cancel, and output verification. Evidence: mixed-eligibility batch journey and deterministic manifest.
    • E2.32.a Define batch/item immutable request with project/source version, range, operation kind, target/profile, priority, owner, and expected outputs. Evidence: batch schema.
    • E2.32.b Resolve naming/path/template/collision, metadata/locale/ identifiers/artwork, rights/consent/territory, plugin/media/model/ renderer dependencies per item. Evidence: preflight tests.
    • E2.32.c Estimate capacity/CPU-GPU/memory/storage/network/time/currency/ provider cost and require authorization/budget before dispatch. Evidence: estimate tests.
    • E2.32.d Preview exact eligible/blocked/degraded/lossy item plan and allow scoped selection; never drop ineligible items silently. Evidence: preview tests.
    • E2.32.e Execute durable dependency-aware queue with leases/concurrency/ resource limits/progress/logs/artifacts/retry/cancel/idempotency. Evidence: job tests.
    • E2.32.f Record per-item success/partial/failed/cancelled/unknown and preserve successful outputs while retrying only safe failed work. Evidence: partial result tests.
    • E2.32.g Verify output bytes/hash/format/audio measurements/metadata/ rights/provenance/naming and semantic/source lineage. Evidence: output tests.
    • E2.32.h Emit deterministic batch/item manifest with inputs/versions/ dependencies/settings/cost/results and run mixed-eligibility E2E. Evidence: batch journey.
  • E2.33 Build/complete modular synthesis/control where ratified, including audio, CV/gate, trigger, MIDI/MPE, modulation and event ports; typed modules; patch cables; polyphony; feedback safety; oversampling; latency; preset/version; automation; macro; performance view; and accessible list/matrix parity. Evidence: graph validation, signal goldens, CPU budgets, and non-pointer operation.
    • E2.33.a Ratify modular host/profile/module SDK/trust/licence/runtime/ realtime safety and supported port/data-rate/polyphony semantics. Evidence: modular profile.
    • E2.33.b Define typed module/node/version/state/parameter and audio/CV/ gate/trigger/MIDI-MPE/modulation/event port compatibility. Evidence: graph contract.
    • E2.33.c Implement add/remove/duplicate/group/reorder modules and connect/ disconnect cables with type/range/channel/polyphony/latency validation. Evidence: graph editing tests.
    • E2.33.d Implement polyphonic voice/event allocation, oversampling/ resampling/block boundaries/latency compensation and deterministic scheduling. Evidence: signal tests.
    • E2.33.e Detect algebraic/feedback cycles, DC/runaway/NaN/denormal/high- gain/CPU overload and provide bounded feedback/delay/limiter/emergency safeguards. Evidence: safety tests.
    • E2.33.f Implement versioned patch/preset/resources/automation/macros and a performance view exposing only approved controls with recall/migration. Evidence: state tests.
    • E2.33.g Provide synchronized node/port/cable list/table/matrix and source- target forms for complete keyboard/screen-reader/non-pointer parity. Evidence: accessible graph tests.
    • E2.33.h Run typed graph validation, impulse/oscillator/control/event goldens, save/reload/offline render, CPU/polyphony budgets, crash/ malicious module recovery. Evidence: modular suite.

E3 — Concept, score, lyrics, performance, and AI assist#

  • E3.1 Build musical-concept capture from text, hum, voice memo, MIDI, chord, reference, and project context. Evidence: each input keeps provenance and rights prompts.
    • E3.1.a Define concept/idea/revision/input-kind/source/owner/project/ status/visibility and immutable provenance contract. Evidence: concept schema.
    • E3.1.b Capture text/chord/score snippets with locale/tuning/tempo/key/ theory assumptions and stable semantic anchors. Evidence: structured input tests.
    • E3.1.c Capture hum/voice memo/audio through permission/calibration/ immutable media/checksum/device/session/performer consent workflow. Evidence: audio capture tests.
    • E3.1.d Capture MIDI/MPE with device/timebase/timing/expression/tuning and raw versus interpreted events. Evidence: MIDI capture tests.
    • E3.1.e Attach reference/project context with exact asset/revision/range/ role, rights/consent/use prompts, protected visibility, and no silent provider upload. Evidence: reference tests.
    • E3.1.f Support save/tag/link/compare/park/discard/convert to proposal with autosave/accessibility/offline handling. Evidence: capture journey.
  • E3.2 Quarantine and analyze uploaded audio/MIDI/project files. Evidence: malicious/oversize/invalid corpus.
    • E3.2.a Upload untrusted files into isolated non-executable quarantine with size/quota/chunk/checksum/content-type and tenant/project scope. Evidence: upload tests.
    • E3.2.b Detect signature/type mismatch, malformed/polyglot/active/archive bomb/malware/project traversal and unsafe plugin/resource references. Evidence: malicious corpus.
    • E3.2.c Probe audio/MIDI/project structure, duration/channels/rate/events/ graph/dependencies/version without loading untrusted plugins or code. Evidence: safe parser tests.
    • E3.2.d Analyze waveform/tempo/key/tuning/transients/loudness/MIDI/ metadata/rights hints with algorithm versions/confidence and no authoritative overwrite. Evidence: analysis tests.
    • E3.2.e Release/reject/manual-review with scan/analysis/policy versions, reasons, audit, rescan, and safe proxy; preserve original bytes. Evidence: quarantine workflow.
    • E3.2.f Test timeout/crash/oversize/encrypted/unsupported/corrupt/ decompression/plugin-reference paths and prove nothing escapes early. Evidence: invalid corpus report.
  • E3.3 Build arrangement and composition divergence behind governed lyria/provider seams. Evidence: not-configured, failure, budget, and exact-output tests.
    • E3.3.a Define provider-agnostic request with pinned concept/project/ score-MIDI-audio context, desired parts/structure/style constraints, rights, seed/count, and budgets. Evidence: request schema.
    • E3.3.b Resolve configured eligible provider/model/version/data-use/ capability/rights region and fail honestly when unavailable. Evidence: negotiation tests.
    • E3.3.c Generate materially distinct arrangement/composition candidates with editable score/MIDI/audio/stem/section outputs, rationale, and source/reference use. Evidence: divergence tests.
    • E3.3.d Persist exact output bytes/events/settings/prompt/negative prompt/ seed/model/provider/cost/latency/provenance and safety results. Evidence: output attribution.
    • E3.3.e Present unranked or transparently measured candidates for human audition/compare/edit/accept/reject with no automatic project mutation. Evidence: decision tests.
    • E3.3.f Test not-configured/timeout/rate/malformed/partial/cancel/budget/ retry/injection/rights failure and deterministic replay limits. Evidence: provider fault suite.
  • E3.4 Build lyric divergence and rewrite assist with locale, theme, meter, rhyme, explicit-content, and banned-topic constraints. Evidence: explicit accept and attribution.
    • E3.4.a Define operation request with pinned lyric range/revision, locale/script/theme/perspective/meaning, syllable/stress/meter/rhyme/ timing, prohibited topics, and budget. Evidence: request schema.
    • E3.4.b Generate distinct whole-section or minimal anchored rewrite candidates with syllable/stress/rhyme/prosody/semantic comparison and uncertainty. Evidence: candidate tests.
    • E3.4.c Validate explicit/advisory/banned-topic/safety/identity/defamation/ similarity/rights constraints per locale without universal culture assumptions. Evidence: policy tests.
    • E3.4.d Attribute provider/model/prompt/version/reference use/cost and display exact diff for per-line/per-change accept/edit/reject. Evidence: review tests.
    • E3.4.e Rebase or invalidate on lyric/score/timing change; accepted changes create human-attributed new lyric revision with AI disclosure. Evidence: lifecycle tests.
    • E3.4.f Evaluate meter/rhyme/meaning preservation, stereotypes/ multilingual performance, protected references, injection, and forbidden invention. Evidence: assist evaluation.
  • E3.5 Build score/orchestration assist with instrument ranges, transposition, voicing, playability, and style constraints. Evidence: proposal-only mutation.
    • E3.5.a Define request pinned to score/arrangement/tuning/tempo/ instrument library/profile, desired scope, performer capability, style/ cultural assumptions, rights, and budget. Evidence: request contract.
    • E3.5.b Generate anchored proposal diffs for assignment/doubling/ transposition/range/voicing/articulation/dynamics/register/orchestration. Evidence: proposal tests.
    • E3.5.c Measure instrument range/transposition/technical playability/ balance/collision and expose declared model/rule assumptions separately from creative judgment. Evidence: analysis tests.
    • E3.5.d Preserve arbitrary tuning/non-Western notation and never “fix” intentional techniques/style exceptions automatically. Evidence: inclusive fixtures.
    • E3.5.e Present score/MIDI/audio audition plus semantic diff for per-part/ passage accept/edit/reject with stale revision handling. Evidence: review tests.
    • E3.5.f Store provider/model/prompt/reference/source/seed/cost provenance and enforce proposal-only server boundary. Evidence: mutation tests.
  • E3.6 Build variant compare for audio, MIDI, score, lyrics, arrangement, gates, cost, and provenance. Evidence: synchronized accessible compare.
    • E3.6.a Select two or more immutable variants and resolve compatible musical/sample time maps, levels, sources, target profile, and permissions. Evidence: compare contract.
    • E3.6.b Provide synchronized level-matched audio A/B/solo/loop/scrub/ blind playback with loudness adjustment disclosed. Evidence: audio compare tests.
    • E3.6.c Provide MIDI/score/lyrics/arrangement semantic diff by events/ parts/sections/lines/anchors, with visual and table/text parity. Evidence: structured compare tests.
    • E3.6.d Compare raw gate measurements/verdicts/freshness, render/provider/ storage cost with units/currency, and full rights/provenance lineage. Evidence: evidence compare.
    • E3.6.e Synchronize selection/playhead/range/focus/comments across panes and support keyboard/screen-reader/zoom/nonvisual decision journey. Evidence: accessible compare.
    • E3.6.f Record human select/reject/park/no-decision separately without mutating variants or using aesthetic score as automatic winner. Evidence: decision tests.
  • E3.7 Build decision records for selected, rejected, parked, and superseded ideas/takes/mixes/masters. Evidence: stable links.
    • E3.7.a Define immutable decision identity/type/subject kind/revision/ actor-role/time/project/context and considered alternatives. Evidence: decision schema.
    • E3.7.b Require rationale, comparison evidence, creative/technical/rights/ cost considerations, uncertainty, conflict/recusal, and dissent where applicable. Evidence: validation tests.
    • E3.7.c Define selected destination/version creation, rejected retention, parked revisit trigger/expiry, and superseded replacement lineage. Evidence: disposition tests.
    • E3.7.d Preserve stable links after rename/move/archive/restore/migration and redacted permissions for media/rights. Evidence: link tests.
    • E3.7.e Prevent model/agent/score from impersonating human selection and maintain immutable history under reversal/new decision. Evidence: authority tests.
  • E3.8 Build lyric editor with syllable/stress/rhyme/prosody, timing, translation, and pronunciation views. Evidence: known-value meter tests.
    • E3.8.a Build structured section/line/token/syllable editor with stable IDs, Unicode/script/language, verses/repeats, undo, and revision anchors. Evidence: editor tests.
    • E3.8.b Show/edit syllabification, stress/weight, meter/feet or applicable rhythmic model, melisma/elision, and declared language assumptions. Evidence: meter tests.
    • E3.8.c Show/edit rhyme sound/family/scheme/internal/near relationships using locale/pronunciation sources with uncertainty. Evidence: rhyme tests.
    • E3.8.d Align lines/syllables to score/MIDI/audio musical/sample time with tap/record/import/numeric controls and orphan/conflict handling. Evidence: timing tests.
    • E3.8.e Edit translation/transliteration/adaptation in aligned semantic views with terminology/names/meaning/rhyme/meter parity. Evidence: locale tests.
    • E3.8.f Edit IPA/phoneme/pronunciation/alternate performer guidance and audition where authorized. Evidence: pronunciation tests.
    • E3.8.g Validate expert-reviewed known values across ratified languages/ scripts/meters without universalizing one prosodic system. Evidence: corpus report.
  • E3.9 Build chord/key/theory inspector with explicit assumptions. Evidence: score and MIDI fixtures.
    • E3.9.a Resolve selected score/MIDI/audio-derived events, tuning/scale/ notation/tempo context, analysis scope, and applicable theory profile. Evidence: inspector contract.
    • E3.9.b Show observed pitch classes/frequencies/spellings, candidate key/ mode/tonality with confidence and temporal changes. Evidence: key tests.
    • E3.9.c Show chord/voicing/bass/inversion/function/interval/nonchord tones and alternative interpretations with exact event evidence. Evidence: chord tests.
    • E3.9.d Declare tuning/reference pitch/cultural/theoretical assumptions, algorithm/corpus/version, missing/ambiguous/nonapplicable state. Evidence: provenance tests.
    • E3.9.e Allow human choose/edit/annotate/decline an interpretation; never rewrite score/MIDI or rank aesthetic quality automatically. Evidence: interaction tests.
    • E3.9.f Verify Western/non-Western/non-12-TET/polychord/ambiguous/free- tonal score and MIDI fixtures plus semantic accessibility. Evidence: known-value suite.
  • E3.10 Build performance capture with device calibration, latency compensation, count-in, monitoring, punch, takes, and consent. Evidence: capture tests.
    • E3.10.a Select/authorize audio/MIDI/MPE devices/channels and calibrate levels/noise/phase/clock/round-trip latency with recorded evidence. Evidence: setup tests.
    • E3.10.b Create performer/session/part/arrangement/tuning/tempo/cue/ monitoring plan and verify applicable consent before arm. Evidence: capture plan.
    • E3.10.c Implement count-in/pre-roll/click/cue/monitor mix/direct-versus- processed path with feedback protection and disclosed latency. Evidence: monitoring tests.
    • E3.10.d Implement manual/automatic punch/loop/overdub/multitake capture with sample-accurate alignment, dropout markers, and crash-safe local originals. Evidence: recording tests.
    • E3.10.e Apply measured latency compensation without altering original timing, expose confidence/manual correction, and handle clock drift. Evidence: alignment tests.
    • E3.10.f Record immutable device/performer/consent/session/file/take/ operator provenance and revoke/restrict downstream use. Evidence: provenance tests.
  • E3.11 Build transcription for audio-to-MIDI, score, chords, and lyrics behind honest confidence/proposal states. Evidence: calibrated corpus.
    • E3.11.a Define request with pinned audio range/channels/stems, task, tuning/tempo/language/instrument/voice assumptions, provider/model, and budget. Evidence: request contract.
    • E3.11.b Produce MIDI/note events with pitch/spelling/onset/duration/ velocity/expression/instrument/voice/confidence and aligned locators. Evidence: MIDI output tests.
    • E3.11.c Produce score/parts/rhythm/meter/key/articulation and chord/ harmony candidates with alternative interpretations and explicit quantization. Evidence: notation output tests.
    • E3.11.d Produce lyric/speaker/word/syllable/phoneme/time alignments with language/uncertainty and no invented text hidden as transcript. Evidence: lyric output tests.
    • E3.11.e Present proposals for per-event/region/part accept/edit/reject, retain source audio and full model/run provenance. Evidence: review tests.
    • E3.11.f Evaluate onset/offset/pitch/chord/notation/word/alignment error, calibration, polyphony/noise/locale/instrument slices, and abstention. Evidence: calibrated corpus report.
  • E3.12 Build generated voice/performance only with valid consent and authorized host controls. Evidence: impersonation and revocation tests.
    • E3.12.a Define voice/performer identity/model/enrollment/reference/ consent purpose/scope/territory/term/host/output/disclosure contract. Evidence: authorization schema.
    • E3.12.b Verify current identity/consent/rights/model/provider/host/ project/role and step-up before preview, generation, export, or share. Evidence: preflight tests.
    • E3.12.c Watermark/fingerprint or apply ratified provenance/disclosure, retain exact run/reference/settings, and restrict protected voiceprints. Evidence: output controls.
    • E3.12.d Enforce host allowlists, rate/budget/target/content policy, dangerous impersonation/public-figure/fraud constraints, and human approval. Evidence: policy tests.
    • E3.12.e Propagate revocation/expiry/scope change to queued jobs, stored outputs, projects, mixes, masters, releases, downloads, and providers. Evidence: revocation journey.
    • E3.12.f Test spoofed consent/identity, cross-project/model reuse, voice extraction, prompt injection, unauthorized host/API, and revoked cache. Evidence: impersonation suite.
  • E3.13 Build generated stem and edit lineage down to model, prompt, reference, seed, and transformation. Evidence: provenance inspector.
    • E3.13.a Record immutable input source/stem/clip/range/mask/condition/ reference/rights hashes and their exact revisions. Evidence: input lineage.
    • E3.13.b Record provider/model/weights/prompt/negative prompt/seed/ parameters/tool/code/hardware/version/time/cost and safety policy. Evidence: generation lineage.
    • E3.13.c Record output stem/file/channel/range/hash/confidence/ measurements and transformation parent/child DAG. Evidence: output lineage.
    • E3.13.d Carry lineage through comp/edit/stretch/process/mix/master/ bounce/export/release while preserving range/ingredient relationships. Evidence: propagation tests.
    • E3.13.e Build permission-filtered inspector with graph plus semantic tree/table, source-use/rights/consent/disclosure/tamper/missing state. Evidence: provenance UI.
    • E3.13.f Verify hashes/signatures/replay where possible and label opaque/nondeterministic/stripped/untrusted lineage explicitly. Evidence: integrity tests.
  • E3.14 Build assist cost, progress, cancel, retry, and human-acceptance views. Evidence: provider fault journey.
    • E3.14.a Define run/job step/progress/unit/estimate/cost currency/ provider/budget/attempt/terminal state contract. Evidence: operations schema.
    • E3.14.b Show queued/running/waiting/uploading/generating/downloading/ analyzing progress, elapsed/remaining/unknown, current step, and resource use. Evidence: progress tests.
    • E3.14.c Show estimated versus authorized versus actual token/compute/ storage/egress/currency cost with delayed accounting and budget alerts. Evidence: cost tests.
    • E3.14.d Implement requested/acknowledged/completed cancellation across upload/provider/worker/download/commit boundaries with partial artifact handling. Evidence: cancel tests.
    • E3.14.e Implement retry/resume/replay with pinned inputs/versions or explicit substitution, idempotency, safe dedupe, and attempt history. Evidence: retry tests.
    • E3.14.f Present outputs as unaccepted auditionable proposals with exact compare/provenance/rights and per-output accept/edit/reject. Evidence: acceptance tests.
    • E3.14.g Inject not-configured/rate/timeout/malformed/partial/budget/ cancel/network/rights-revoke/provider-commit failures. Evidence: provider fault journey.
  • E3.15 Build composition, lyric, arrangement, performance, mix, mastering, metadata, and rights agent plans with least-privilege tool grants. Evidence: project revision, assets, policy, budget, stop conditions, and approval points are locked before dispatch.
    • E3.15.a Define distinct plan templates/allowed outputs for every named agent domain and prohibit blending creative, rights, QC, approval, and distribution authority. Evidence: template registry.
    • E3.15.b Pin project/song/session/arrangement/mix/master/release/assets/ rights/policy/profile revisions, objective, constraints, and acceptance criteria. Evidence: plan inputs.
    • E3.15.c Grant least-privilege read/analyze/audition/draft/propose/render/ job tools with object/field/range/host/provider/target scope and expiry. Evidence: capability manifest.
    • E3.15.d Lock token/compute/currency/time/step/tool/storage/egress budgets, provider/data-use/local-cloud, retry/fallback, and rights/consent scope. Evidence: budget policy.
    • E3.15.e Define human checkpoints, audition/approval/rights-QC-required/ stop-the-line/cancel/escalation conditions before dispatch. Evidence: control contract.
    • E3.15.f Preview exact plan/grants/data movement/risks to authorized human, sign/hash approval, and reject stale or mutated dispatch. Evidence: dispatch tests.
  • E3.16 Build Euterpe-scoped agent run, queue, plan, tool-call, budget, output, rights-use, and provenance views. Evidence: live trace and project isolation.
    • E3.16.a Define run/step/tool/queue/event/budget/input-rights/output/ artifact/provenance state and retention contracts. Evidence: run schema.
    • E3.16.b Build queue/run list with project/owner/template/status/progress/ priority/wait/failure/budget/cost and stable filters. Evidence: operations tests.
    • E3.16.c Build live timeline for immutable plan, model/tool calls, source/ reference use, decisions/retries/errors, artifacts, and human controls. Evidence: trace tests.
    • E3.16.d Show token/compute/time/currency/storage/egress limits versus actual/forecast and truthful delayed costs. Evidence: budget tests.
    • E3.16.e Show exact input/output media/revisions/ranges, rights/consent/ model use, hashes, transformations, and accepted status. Evidence: provenance view.
    • E3.16.f Enforce tenant/project/role/field/media/rights redaction in UI/ API/events/logs/export and no cross-project counts/timing. Evidence: isolation tests.
    • E3.16.g Test live reconnect/order/dedupe/stale/partial/offline/cancel/ terminal reconciliation. Evidence: live trace E2E.
  • E3.17 Intake agent output as attributed clips, takes, settings, suggestions, or release-metadata proposals. Evidence: agents cannot assert clearance/consent, sign splits, approve a master, or distribute.
    • E3.17.a Define allowlisted output envelopes for clip/take/stem/score- MIDI/lyric/arrangement/settings/mix-master suggestion/metadata proposal with run provenance. Evidence: output schemas.
    • E3.17.b Validate format/hash/duration/range/timebase/channels/events/ parameters/anchors/schema/size/rights/safety and quarantine media. Evidence: intake validation.
    • E3.17.c Store in a proposal namespace with explicit unaccepted state, immutable source/run links, no release/QC/rights/consent/approval semantics. Evidence: persistence tests.
    • E3.17.d Present level-matched audition/semantic diff/settings/input-use/ confidence/cost/rights for per-part accept/edit/reject. Evidence: review tests.
    • E3.17.e Apply accepted output as human-attributed new revision/reference, retaining AI provenance and rerunning technical/rights/approval gates. Evidence: acceptance tests.
    • E3.17.f Attempt clearance/consent assertion, split signing, master/QC approval, release/distribution/takedown, direct project mutation, or privilege expansion. Evidence: capability rejection tests.
  • E3.18 Implement agent pause, redirect, audition, reject, cancel, budget kill, and replay. Evidence: partial audio and metadata remain explicitly unaccepted and linked to immutable runs.
    • E3.18.a Define legal control transitions by run/step/provider/job/ artifact state, actor, idempotency, and commit/terminal boundary. Evidence: control state machine.
    • E3.18.b Pause at safe checkpoints, stop new tool calls, retain leases/ partial artifacts, and resume only after plan/grant/rights validation. Evidence: pause tests.
    • E3.18.c Redirect through reviewed plan amendment/new plan with changed objective/constraints/grants/budgets while preserving original. Evidence: redirect tests.
    • E3.18.d Audition outputs through sandboxed level-safe routing with no project mutation; reject with retained reason and immutable artifact. Evidence: audition/reject tests.
    • E3.18.e Cancel queued/running/provider/render/download phases with requested-versus-completed state, compensation, and timeout. Evidence: cancel tests.
    • E3.18.f Enforce automatic/manual budget kill with bounded overshoot and replay from pinned inputs/versions or disclosed substitutions, deduping side effects. Evidence: budget/replay tests.
    • E3.18.g Verify partial media/metadata/settings stay unaccepted/ unapproved/unreleased with complete run/source/rights provenance. Evidence: status tests.
  • E3.19 Build stem separation with source/channel selection, stem taxonomy, bleed/artifact/confidence review, manual correction, recombination, rights inheritance, and per-stem lineage. Evidence: stem-sum and known-mixture evaluation.
    • E3.19.a Define separation request pinned to source revision/range/ channels, target stem taxonomy/count, model/profile, rights, quality, budget, and output format. Evidence: request schema.
    • E3.19.b Support channel/region/mask/source selection with preview, minimum signal validation, immutable original, and protected rights/ consent checks. Evidence: selection tests.
    • E3.19.c Run durable provider/local job and produce aligned immutable stems plus residual, exact model/settings/seed/hardware/input-output hashes, and cost. Evidence: job tests.
    • E3.19.d Measure per-stem target/interference/bleed/artifact/silence/ phase/timing/confidence and mark insufficient/unsupported/failure separately. Evidence: analysis tests.
    • E3.19.e Provide synchronized level-matched solo/mute/residual/sum A/B and semantic measurement table for human taxonomy/quality review. Evidence: review UI.
    • E3.19.f Allow manual stem relabel/group/split/merge/mask correction/ spectral edit and rerun as new version without rewriting outputs. Evidence: correction tests.
    • E3.19.g Recombine with gain/pan/phase/alignment handling and compare stem sum to source under declared decomposition/tolerance. Evidence: recombination tests.
    • E3.19.h Inherit/review source/sample/performer/voice/model rights and consent per stem; prevent separation from creating new clearance. Evidence: rights tests.
    • E3.19.i Carry per-stem lineage through edit/mix/master/export/release and run synthetic/real known-mixture plus stem-sum evaluation. Evidence: evaluation report.
  • E3.20 Build audio inpainting, continuation, variation, conditional generation, and style transfer with exact masks/ranges, musical context, reference rights, continuity controls, audition, replace/comp, and provenance. Evidence: untouched samples remain bit-identical where the operation promises surgical editing.
    • E3.20.a Define distinct operation contracts for inpaint/continue/variant/ conditional generation/style transfer, each with allowed scope/output and no semantic aliasing. Evidence: operation registry.
    • E3.20.b Select exact source/mask/range/channel/stem/context handles and musical tempo/meter/key/tuning/form/neighboring audio conditions. Evidence: selection tests.
    • E3.20.c Define prompt/score/MIDI/reference/style/instrument/seed/ strength/negative/preservation/transition/duration constraints and resolve rights/consent. Evidence: condition schema.
    • E3.20.d Generate multiple immutable candidates with exact provider/model/ parameters/input-output hashes/cost/confidence/provenance and safe failure state. Evidence: generation tests.
    • E3.20.e Measure boundary continuity, tempo/pitch/timbre/phase/noise/ loudness, source leakage/copying, artifacts, semantic/style fit, and uncertainty. Evidence: quality tests.
    • E3.20.f Audition synchronized source/candidate/context with level match, waveform/spectral/semantic comparison, and per-candidate decision. Evidence: audition tests.
    • E3.20.g Apply by nondestructive replace/new take/comp layer/alternate arrangement, preserving source and supporting undo/restore. Evidence: edit integration tests.
    • E3.20.h For surgical operations, assemble output from exact source plus generated mask and byte/sample-compare every promised untouched region. Evidence: preservation goldens.
    • E3.20.i Carry reference/model/generated-region rights/disclosure/ provenance through project/mix/master/release and test revocation. Evidence: lineage tests.
  • E3.21 Build voice conversion, speech, singing, choir, harmony, dubbing, pronunciation, timing, vocal processing, and performance-direction workflows with subject/performer consent, identity scope, locale, revocation, disclosure, and misuse monitoring. Evidence: each mode has a distinct consent/policy/evaluation profile.
    • E3.21.a Define separate mode profiles/contracts for conversion, speech, solo singing, choir/harmony, dubbing, pronunciation/timing, processing, and direction with explicit nonapplicability. Evidence: mode registry.
    • E3.21.b Verify subject/performer/voice/model identity, enrollment/ reference provenance, purpose/territory/term/host/target consent, rights, and current revocation before every operation. Evidence: consent gate.
    • E3.21.c Define text/lyrics/phonemes/locale/accent/pronunciation/timing/ pitch/melody/harmony/voice assignment/expression/direction inputs with protected data handling. Evidence: input contracts.
    • E3.21.d Generate immutable takes/stems with exact provider/model/voice/ prompt/seed/settings/input-output hashes/watermark/cost/provenance. Evidence: generation tests.
    • E3.21.e Evaluate intelligibility/pronunciation/alignment/pitch/rhythm/ harmony/identity similarity-drift/artifact/naturalness separately with calibrated uncertainty. Evidence: mode evaluations.
    • E3.21.f Audition/compare/comp/edit/process with human acceptance and prevent identity similarity or quality score from asserting consent/ authenticity. Evidence: review tests.
    • E3.21.g Attach mode/AI/voice disclosure and rights metadata to clips/ mixes/masters/releases; restrict downloads/sharing/model reuse. Evidence: enforcement tests.
    • E3.21.h On revocation/expiry/scope change, cancel jobs, quarantine/ restrict outputs, preview blast radius, replace/takedown, and notify providers/targets. Evidence: revocation E2E.
    • E3.21.i Monitor impersonation/fraud/unauthorized cloning/voiceprint extraction/provider misuse with privacy-safe alerts and human review. Evidence: abuse tests.
  • E3.22 Build melody, harmony, rhythm/groove, arrangement, orchestration, accompaniment, improvisation, and full text-to-music workflows with editable intermediate score/MIDI/audio representations. Evidence: creators can inspect and accept/reject parts rather than receiving an opaque final waveform.
    • E3.22.a Define distinct generation/assist workflows, allowed inputs/ outputs, scope, creative assumptions, rights, provider capability, budget, and human checkpoints. Evidence: workflow registry.
    • E3.22.b Build melody and harmony/voicing generation from text/chords/ motifs/score/MIDI/audio context with tuning/range/form constraints and alternative interpretations. Evidence: melodic/harmonic tests.
    • E3.22.c Build rhythm/groove generation from tempo/cycle/pulse/style/ performance references with editable timing/velocity/expression and culturally declared assumptions. Evidence: rhythm tests.
    • E3.22.d Build arrangement/orchestration/accompaniment generation with sections/roles/instruments/ranges/articulations/texture/density and performer/device capabilities. Evidence: production tests.
    • E3.22.e Build improvisation and full text-to-music as planned sequences of editable form/chords/melody/rhythm/parts/score/MIDI/stems/audio, not a single opaque artifact. Evidence: plan/intermediate tests.
    • E3.22.f Provide synchronized score/MIDI/audio audition and semantic diff for per-section/part/event/stem accept/edit/reject/replace/comp. Evidence: decision UI.
    • E3.22.g Record every intermediate’s model/prompt/reference/seed/settings/ parent/rights/cost/provenance and carry through accepted project revision. Evidence: lineage tests.
    • E3.22.h Evaluate structure/timing/playability/preservation/diversity/ originality/leakage/locale-cultural slices without universal aesthetic score. Evidence: evaluation report.
  • E3.23 Build AI mix and master assistance as versioned, auditionable settings or renders with target/reference assumptions, explanations, confidence, cost, undo, and human approval. Evidence: assist cannot overwrite source, approve QC, or distribute.
    • E3.23.a Define mix versus master assist requests pinned to source revision, selected tracks/stems/range, target/reference/profile, allowed processors, rights/privacy, and budget. Evidence: request schemas.
    • E3.23.b Produce versioned settings/automation/routing/device proposals or immutable renders with exact changed parameters and no source mutation. Evidence: proposal tests.
    • E3.23.c Explain target/reference assumptions, measured issues, recommended actions, alternatives, confidence/uncertainty, and model/provider versions. Evidence: explanation tests.
    • E3.23.d Provide level-matched blindable A/B, loudness-normalized comparison, raw measurement deltas, semantic settings diff, and source restore. Evidence: audition tests.
    • E3.23.e Show estimated/actual cost/progress/cancel/retry and preserve partial outputs as unaccepted. Evidence: operations tests.
    • E3.23.f Accept/edit/reject per setting/range/render into a new human- attributed mix/master candidate with undo and full AI provenance. Evidence: acceptance tests.
    • E3.23.g Re-run independent master/target/rights/QC gates and require qualified human approval; block assist from QC approval/distribution. Evidence: authority tests.
  • E3.24 Build local/on-device and cloud AI capability negotiation with model/weight version, hardware/memory, privacy, latency, quality, energy, offline availability, fallback, and result equivalence policy. Evidence: unsupported hardware fails honestly, local data does not leave the device silently, and provider fallback never changes consent or rights scope.
    • E3.24.a Define operation capability requirements and local/browser/ desktop/native/worker/cloud provider profiles with model/weights/ runtime/version/licence. Evidence: capability contracts.
    • E3.24.b Detect CPU/GPU/NPU/architecture/memory/storage/energy/thermal/ network/offline/runtime support and benchmark or validate minimums. Evidence: hardware tests.
    • E3.24.c Resolve data classification/rights/consent/provider-training/ region/retention/encryption/model-use constraints before considering each execution location. Evidence: privacy policy tests.
    • E3.24.d Estimate latency/quality/resource/energy/cost/output formats and present eligible choices/limits to user or policy. Evidence: negotiation UI.
    • E3.24.e Dispatch locally without network when selected and verify firewall/telemetry/cache/model-download boundaries; fail honestly on unsupported hardware. Evidence: local execution tests.
    • E3.24.f Dispatch cloud only after explicit allowed scope with minimized encrypted data, provider/model pinning, audit, cancel, and deletion. Evidence: cloud execution tests.
    • E3.24.g Allow fallback only within preauthorized privacy/rights/consent/ quality/cost constraints and require reapproval for expanded data movement. Evidence: fallback tests.
    • E3.24.h Define semantic/audio result equivalence/tolerances and disclose nondeterminism/model substitutions rather than silently presenting identical capability. Evidence: parity report.

E4 — Theory, audio quality, mix, and master gates#

  • E4.1 Implement tempo/meter-map consistency and event alignment gates. Evidence: known MIDI/score fixtures.
    • E4.1.a Validate tempo/meter segment order/ranges/overlaps/gaps/ramps/ pickup/time signatures and musical-sample conversion continuity. Evidence: map tests.
    • E4.1.b Check score/MIDI/clip/marker/cue/automation/event anchors against applicable map revision and detect drift/orphan/ambiguous mapping. Evidence: alignment tests.
    • E4.1.c Check bar/beat/subdivision boundaries, tuplets/polyrhythm/ additive/free-time applicability and avoid common-time assumptions. Evidence: rhythmic fixtures.
    • E4.1.d Persist raw timing deltas/units/event IDs/map versions/evidence separately from tolerance/profile verdict. Evidence: result contract.
    • E4.1.e Run known MIDI/score/audio/import/export/tempo-edit/long-duration fixtures and stale-map invalidation. Evidence: known-value suite.
  • E4.2 Implement key/scale, chord spelling, range, voice-leading, parallel, and dissonance diagnostics as configurable theory assistance. Evidence: known score fixtures.
    • E4.2.a Define versioned theory profile with tuning/cultural/style/genre/ period/instrument assumptions, enabled diagnostics, thresholds, and nonapplicability. Evidence: profile schema.
    • E4.2.b Measure key/scale membership/spelling/alternative interpretation, chord/voicing/inversion/function and exact note evidence. Evidence: harmony tests.
    • E4.2.c Measure written/sounding range/transposition/playability/ instrumentation against versioned performer/instrument capabilities. Evidence: range tests.
    • E4.2.d Measure voice-leading/spacing/crossing/parallels/dissonance/ resolution only where declared rules apply and show creative alternatives. Evidence: counterpoint tests.
    • E4.2.e Keep raw findings separate from advisory verdict; allow human annotate/waive/disable and never auto-edit or assign aesthetic quality. Evidence: assistance tests.
    • E4.2.f Validate Western/non-Western/non-12-TET/intentional exception/ ambiguous fixtures with named musical review. Evidence: score corpus.
  • E4.3 Distinguish intentional creative exceptions from technical errors. Evidence: authorized waiver retains original measurement.
    • E4.3.a Classify diagnostics as objective technical integrity, target conformance, configurable theory/style assistance, or subjective preference. Evidence: diagnostic registry.
    • E4.3.b Define exception/waiver scope over exact subject/measurement/ profile/version/range, reason/evidence/actor authority/conditions/expiry. Evidence: waiver schema.
    • E4.3.c Preserve immutable raw measurement and original verdict while storing a separate human exception decision. Evidence: storage tests.
    • E4.3.d Restrict nonwaivable safety/rights/consent/integrity/target constraints and require specialist approval for bounded exceptions. Evidence: policy tests.
    • E4.3.e Invalidate waiver on subject/measurement/profile/target/rights/ approver/expiry change and preserve history. Evidence: lifecycle tests.
    • E4.3.f Show intentional/waived/unresolved state distinctly from pass and include it in review/release evidence. Evidence: UI/audit tests.
  • E4.4 Implement lyric syllable, stress, meter, rhyme, pronunciation, and line-timing measurements. Evidence: language-specific fixture set.
    • E4.4.a Pin lyric/score/audio/timing/pronunciation revision and applicable locale/prosody/rhyme/meter profile with declared assumptions. Evidence: measurement inputs.
    • E4.4.b Measure token/syllable boundaries, elision/melisma, stress/weight/ meter alignment and confidence/source. Evidence: prosody tests.
    • E4.4.c Measure rhyme families/scheme/position/exact-near/internal with pronunciation variants and uncertainty. Evidence: rhyme tests.
    • E4.4.d Measure phoneme/pronunciation/name/locale variants and line/ syllable onset-duration alignment to score/MIDI/audio. Evidence: pronunciation/timing tests.
    • E4.4.e Persist raw values/evidence/algorithm/corpus/profile versions separately from creative assist verdict and mark nonapplicable. Evidence: result contract.
    • E4.4.f Validate expert-reviewed fixtures per ratified language/script/ meter, including dialect/translation/free meter. Evidence: locale corpus.
  • E4.5 Implement recording checks for clipping, DC offset, silence, noise, dropouts, channel, sample rate, bit depth, and phase. Evidence: calibrated WAV corpus.
    • E4.5.a Verify file/container/codec/sample-rate/bit-depth/channel/layout/ duration/checksum and decode completeness against session/profile. Evidence: format tests.
    • E4.5.b Measure sample/inter-sample clipping where applicable, peak locations, DC offset, silence/digital zero, low level, and discontinuity. Evidence: signal tests.
    • E4.5.c Measure noise floor/SNR/hum/buzz/broadband/impulse/dropout/glitch with selected calibration and signal sufficiency. Evidence: noise tests.
    • E4.5.d Measure polarity/phase/correlation/time offset across channels/ microphones and detect channel swaps/duplicates/missing channels. Evidence: multichannel tests.
    • E4.5.e Persist raw measurements/units/ranges/analyzer/calibration/source version separately from capture-profile verdict/remediation. Evidence: provenance tests.
    • E4.5.f Run calibrated WAV corpus across rates/depths/layouts/noise/ deliberate clipping/dropout/phase and analyzer failures. Evidence: conformance report.
  • E4.6 Implement mix checks for clipping, headroom, mono compatibility, phase, spectral balance, dynamics, masking, and true peak. Evidence: reference fixtures.
    • E4.6.a Analyze exact mix revision/range/layout/render path and verify sufficient signal/calibration/analyzer/profile before measurement. Evidence: input tests.
    • E4.6.b Measure sample/true peak/clipping/headroom and integrated/short/ momentary dynamics/loudness distributions without universal targets. Evidence: level tests.
    • E4.6.c Measure mono/downmix compatibility, polarity/correlation/phase/ spatial energy and identify affected channels/times/frequencies. Evidence: compatibility tests.
    • E4.6.d Measure spectrum/tonal balance/masking/dynamic range/transients/ crest with exact references/bands/assumptions and uncertainty. Evidence: balance tests.
    • E4.6.e Persist raw values separately from target/reference/profile verdict and treat aesthetic comparison as advisory. Evidence: result tests.
    • E4.6.f Run calibrated reference/edge/intentional-exception/silence/ multichannel fixtures with known expected values. Evidence: mix corpus.
  • E4.7 Implement integrated, short-term, momentary loudness and loudness range per the pinned BS.1770/R128 implementation. Evidence: official or calibrated test signals.
    • E4.7.a Pin BS.1770/R128/EBU Tech implementation/version/channel weights/ K-weighting/gating/window/LRA/true-peak dependencies. Evidence: meter profile.
    • E4.7.b Implement channel-layout-aware K-weighting and absolute/relative gating for integrated loudness with exact block overlap/rounding. Evidence: integrated tests.
    • E4.7.c Implement short-term and momentary window/hop calculations and time-aligned result series. Evidence: window tests.
    • E4.7.d Implement loudness range percentile/gating/sufficient-duration rules and explicit insufficient-signal state. Evidence: LRA tests.
    • E4.7.e Persist LUFS/LU/true-peak raw outputs/ranges/units/layout/ algorithm/calibration/source hashes separately from target verdict. Evidence: result schema.
    • E4.7.f Validate official or independently calibrated mono/stereo/ multichannel/silence/short/program test signals within tolerances. Evidence: conformance report.
  • E4.8 Implement target-specific loudness and peak profiles rather than one universal LUFS number. Evidence: broadcast, streaming, podcast, and platform fixtures.
    • E4.8.a Define versioned profile per broadcast/streaming/podcast/video/ physical/spatial platform with jurisdiction/target/layout/content type. Evidence: profile registry.
    • E4.8.b Encode integrated/short/momentary/LRA/true-sample peak targets/ ranges, measurement standard, normalization behavior, and exceptions. Evidence: profile contracts.
    • E4.8.c Resolve applicable profile from release target/territory/format/ layout/version and show unknown/conflicting/unsupported explicitly. Evidence: resolution tests.
    • E4.8.d Compute verdict from retained raw measurements with exact delta/ reason/remediation and no reinterpretation of creative mix quality. Evidence: verdict tests.
    • E4.8.e Recompute verdict after profile change without rerendering valid source measurements and invalidate approval/delivery as needed. Evidence: recomputation tests.
    • E4.8.f Validate broadcast/streaming/podcast/platform boundary fixtures and current distributor expectations. Evidence: target suite.
  • E4.9 Implement master format, channel layout, codec, sample rate, bit depth, dither, gap, fade, metadata, and artwork checks. Evidence: per-target goldens.
    • E4.9.a Verify container/codec/profile/extension/MIME/decode/duration/ checksum/file size and target support. Evidence: file checks.
    • E4.9.b Verify sample rate/bit depth/endian/channel count/layout/labels/ ordering/downmix metadata and expected master spec. Evidence: audio format tests.
    • E4.9.c Detect/respect dither/noise shaping, unnecessary repeat dither, truncation, clipping, sample-rate conversion version, and silence. Evidence: signal processing tests.
    • E4.9.d Verify program start/end, pregap/postgap/intertrack gap/continuous album transitions/fade shape/tails and sequence alignment. Evidence: sequencing tests.
    • E4.9.e Verify identifiers/titles/versions/contributors/languages/ advisories/rights/lyrics/credits/provenance embedded and package metadata. Evidence: metadata tests.
    • E4.9.f Verify artwork dimensions/color/profile/format/file size/text/ rights/advisory/target safe areas and associations. Evidence: artwork tests.
    • E4.9.g Run per-target goldens and distinguish unsupported/missing/ invalid/stale/analyzer failure from pass. Evidence: master gate suite.
  • E4.10 Implement spatial/immersive profile checks where supported. Evidence: channel/object layout fixtures and explicit unsupported states.
    • E4.10.a Resolve exact session/layout/renderer/device/target profile and supported object/bed/channel/ambisonic/binaural capabilities. Evidence: applicability tests.
    • E4.10.b Validate channel labels/order/angles/elevation/LFE/bed mapping and object count/IDs/coordinates/trajectories/spread/priority/automation. Evidence: layout tests.
    • E4.10.c Validate ambisonic order/dimensionality/channel ordering/ normalization/coordinate orientation and conversion. Evidence: HOA tests.
    • E4.10.d Validate binaural renderer/HRTF/head-tracking sensor/timing/ fallback/device metadata and object-to-binaural behavior. Evidence: binaural tests.
    • E4.10.e Validate downmix/render loudness/peak/phase/localization/ fold-down/metadata/interchange and record exact loss. Evidence: render tests.
    • E4.10.f Emit explicit unsupported/degraded/unavailable/unknown renderer or layout state and run goldens per supported cell. Evidence: spatial gate.
  • E4.11 Implement reference-track comparison without treating aesthetic similarity as automatic quality. Evidence: raw measures and policy.
    • E4.11.a Define authorized reference/revision/range/rights/target/ alignment/level-match and comparison purpose. Evidence: reference contract.
    • E4.11.b Align tempo/time/sections and level-match transparently while preserving original playback and reporting applied transforms. Evidence: alignment tests.
    • E4.11.c Measure loudness/peak/dynamics/spectrum/tonal balance/stereo- spatial/phase/tempo/structure differences with units/ranges/uncertainty. Evidence: measurement tests.
    • E4.11.d Show synchronized A/B and semantic comparison with exact reference/source/profiles/analyzer versions. Evidence: compare tests.
    • E4.11.e Apply optional advisory target bands separately; prohibit a single similarity/quality verdict or automatic mix/master changes. Evidence: policy tests.
    • E4.11.f Validate identical/level-shifted/genre-different/intentionally different/short/unauthorized reference fixtures. Evidence: known-value suite.
  • E4.12 Implement render null/round-trip and stem-sum diagnostics where applicable. Evidence: known signal fixtures.
    • E4.12.a Define applicability/alignment/gain/polarity/delay/sample-rate/ layout/tail/tolerance and expected nondeterminism for each comparison. Evidence: diagnostic profile.
    • E4.12.b Render source and comparison through pinned realtime/offline/ export/import/plugin/format paths with exact manifests. Evidence: reproducible inputs.
    • E4.12.c Align signals and compute null residual peak/RMS/spectrum/ correlation/time-varying error by channel/range. Evidence: null tests.
    • E4.12.d Sum stems with declared routing/gain/pan/phase/processing/tails and compare to mix bus or source. Evidence: stem-sum tests.
    • E4.12.e Distinguish exact, tolerance-pass, expected nondeterministic, mismatched, insufficient, and analyzer failure with evidence. Evidence: result states.
    • E4.12.f Validate impulse/sine/noise/known plugin/round-trip/stem-mix/ alignment-error fixtures across sample rates/layouts. Evidence: signal corpus.
  • E4.13 Persist raw measurements separately from target verdicts. Evidence: profile change recomputes without rerendering.
    • E4.13.a Define immutable measurement record with source/render/range/ channel/layout/input hashes, values/units/series, analyzer/calibration/ plugin versions, and uncertainty. Evidence: measurement schema.
    • E4.13.b Define separate verdict with measurement references, applicable profile/threshold/rules, outcome/reason/remediation/exception and time. Evidence: verdict schema.
    • E4.13.c Prevent verdict fields from mutating/denormalizing raw values and prevent missing/unsupported/error measurements from defaulting to pass. Evidence: storage invariants.
    • E4.13.d Recompute new verdicts from retained valid measurements after profile/threshold change without rendering/analyzing again. Evidence: recomputation tests.
    • E4.13.e Rerender/reanalyze only when source/render/analyzer/calibration dependencies invalidate and preserve prior lineage. Evidence: lifecycle tests.
    • E4.13.f Compare old/new profile results and explain changes without rewriting historical approval/release evidence. Evidence: profile change report.
  • E4.14 Version algorithms, calibration signals, profiles, thresholds, and plugins. Evidence: every result is reproducible or labeled otherwise.
    • E4.14.a Define immutable IDs/digests for analyzer/DSP code/config/ runtime, calibration files/hardware, target/theory profiles/thresholds, and plugin/state/resources. Evidence: version contracts.
    • E4.14.b Resolve/pin every dependency before measurement/render and prohibit mutable aliases as sole provenance. Evidence: dispatch tests.
    • E4.14.c Persist dependency DAG plus OS/CPU/architecture/audio library/ sample rate/layout/random seed/input-output hashes. Evidence: provenance tests.
    • E4.14.d Define approved/effective/retired/revoked/blocked states, compatibility, migration, rollback, retention, and artifact availability. Evidence: lifecycle policy.
    • E4.14.e Detect missing/unknown/changed/unavailable versions and label result unverifiable/stale/nondeterministic instead of reproducible. Evidence: failure tests.
    • E4.14.f Reproduce representative measurements/renders from provenance, compare within declared tolerance, and disclose substitutions. Evidence: reproducibility gate.
  • E4.15 Add unavailable, unsupported, insufficient-signal, and analyzer failure states. Evidence: none renders as pass.
    • E4.15.a Define distinct applicability/execution/measurement/verdict states, reason codes, safe messages, owners, retry/remediation, and telemetry. Evidence: state taxonomy.
    • E4.15.b Map not-configured/dependency/provider/device/plugin/profile/ permission/authorization outage to unavailable. Evidence: availability tests.
    • E4.15.c Map host/format/layout/sample-rate/operation/profile capability absence to unsupported with supported alternatives/loss. Evidence: support tests.
    • E4.15.d Detect silence/too-short/too-few-events/wrong range/missing channels/low SNR or other insufficient signal with required minimum. Evidence: sufficiency tests.
    • E4.15.e Map timeout/crash/NaN/malformed output/version/calibration/ resource errors to analyzer failure with retained diagnostics. Evidence: fault tests.
    • E4.15.f Assert API/UI/export/review/release/analytics never display any nonpass state as pass/current/healthy. Evidence: projection suite.
  • E4.16 Build accessible quality views with values, units, thresholds, confidence, history, and remediation. Evidence: non-visual parity.
    • E4.16.a Define view model with measurement name/value/unit/range/ channel/time/evidence/analyzer/profile/threshold/verdict/confidence/ freshness. Evidence: quality contract.
    • E4.16.b Build prioritized findings/history/remediation presentation using text/table/structure, not color/graph alone, and avoid one aesthetic quality score. Evidence: UI review.
    • E4.16.c Provide semantic tables/descriptions for waveform/spectrum/ loudness/phase/spatial/time-series charts synchronized to visual selection/playback. Evidence: parity tests.
    • E4.16.d Provide keyboard inspect/filter/sort/range/audition/compare/ waive/remediate/rerun operations with focus and announcements. Evidence: keyboard tests.
    • E4.16.e Show unavailable/unsupported/insufficient/failure/stale/ intentional exception distinctly with eligible action and no pass styling. Evidence: state tests.
    • E4.16.f Test screen reader/zoom/reflow/themes/high contrast/long locale/ RTL/large result sets and compare against visual content/operations. Evidence: nonvisual E2E.
  • E4.17 Implement pitch/tuning, rhythm/groove, harmony/counterpoint, melody, orchestration, form, genre/style, and performance diagnostics as separable, culturally configurable assistants with declared assumptions. Evidence: no universal aesthetic "quality" score and expert-reviewed known-value corpora.
    • E4.17.a Define independent diagnostic modules with versioned cultural/ theoretical/style/instrument/performance profiles, applicability, assumptions, measures, and nonapplicability. Evidence: module registry.
    • E4.17.b Implement pitch/tuning/intonation/spelling/range diagnostics for arbitrary temperaments/reference systems with exact event/frequency evidence. Evidence: pitch corpus.
    • E4.17.c Implement rhythm/groove/pulse/cycle/subdivision/swing/timing/ repetition diagnostics without universal grid or common meter. Evidence: rhythm corpus.
    • E4.17.d Implement harmony/counterpoint/chord/voicing/voice-leading/ dissonance diagnostics only under explicit applicable systems. Evidence: harmony corpus.
    • E4.17.e Implement melody/contour/range/phrase/orchestration/ instrumentation/form diagnostics with performer and cultural source context. Evidence: composition corpus.
    • E4.17.f Implement genre/style/performance timing/dynamics/articulation/ technique comparison as descriptive assistance with uncertainty, not correctness. Evidence: style/performance corpus.
    • E4.17.g Persist raw findings/assumptions/evidence separately from human interpretation/exception and offer suggestions only. Evidence: policy tests.
    • E4.17.h Validate each module on expert-reviewed culturally diverse/ ambiguous/intentional-exception fixtures and prohibit aggregate aesthetic score. Evidence: signed evaluation.
  • E4.18 Implement stem-separation/inpainting/continuation/style-transfer fidelity and leakage checks. Evidence: known source/target masks, bleed, preservation, artifact, and uncertainty fixtures.
    • E4.18.a Define job-kind-specific applicability, source/target/mask/ reference/condition inputs, expected preserved/generated regions, and calibrated fixture truth. Evidence: evaluation contract.
    • E4.18.b Measure separation target capture/interference/bleed/residual/ artifact/silence/phase/timing and stem-sum error per stem. Evidence: separation tests.
    • E4.18.c Measure inpainting mask fidelity/boundary continuity and byte/ sample preservation of untouched regions where guaranteed. Evidence: inpaint tests.
    • E4.18.d Measure continuation/variation musical boundary continuity, structure/timing/pitch/timbre/energy drift and repetition/collapse. Evidence: continuation tests.
    • E4.18.e Measure style-transfer condition adherence separately from content/identity/source copying, reference leakage, and rights risk. Evidence: transfer tests.
    • E4.18.f Detect prompt/reference/training memorization or source stem leakage using ratified similarity/fingerprint methods and uncertainty. Evidence: leakage tests.
    • E4.18.g Persist raw metrics/model/evaluator/corpus versions separately from human acceptance/rights verdict and handle insufficient signal. Evidence: result tests.
    • E4.18.h Run synthetic/recorded/edge/adversarial known masks/mixtures/ references across models and report thresholds/calibration. Evidence: fidelity report.
  • E4.19 Implement voice/singing/choir intelligibility, pronunciation, timing, pitch, identity similarity/drift, artifact, consent, and disclosure checks. Evidence: protected calibrated corpus and explicit non-applicability.
    • E4.19.a Define separate evaluation profiles by speech/singing/choir/ conversion/dubbing/locale/voice type with protected corpus governance and applicability. Evidence: profile registry.
    • E4.19.b Measure word/phoneme intelligibility, pronunciation/name/locale accuracy, alignment/timing/duration and confidence. Evidence: language tests.
    • E4.19.c Measure pitch/intonation/rhythm/expressive contour/harmony/ voice-blend or separation under task-appropriate musical context. Evidence: music tests.
    • E4.19.d Measure identity similarity and unintended identity drift/ cross-voice leakage separately from authenticity/consent, with privacy controls. Evidence: identity tests.
    • E4.19.e Measure artifacts/noise/breath/sibilance/phase/discontinuity/ robotic or collapse indicators and calibrated uncertainty. Evidence: artifact tests.
    • E4.19.f Verify current consent/rights/model-use/territory/target/host/ disclosure/watermark/provenance; mark nonapplicable or block separately from audio metrics. Evidence: policy gate.
    • E4.19.g Protect voiceprints/corpus/identity outputs from unauthorized storage/search/log/export/provider access and enforce deletion/ revocation. Evidence: privacy tests.
    • E4.19.h Report task/subgroup/locale/voice slices, calibration/ thresholds/uncertainty with named voice/performance/rights review. Evidence: signed evaluation.
  • E4.20 Implement Dolby Atmos or ratified object-audio, binaural, ambisonic, immersive-device, head-tracking, downmix, loudness, peak, metadata, and renderer compatibility gates. Evidence: each supported layout/renderer/target pair has its own conformance proof.
    • E4.20.a Pin supported object-audio/binaural/ambisonic/device/interchange/ renderer/target profiles, SDK/validator versions, layouts, and explicit unsupported cells. Evidence: conformance matrix.
    • E4.20.b Validate bed/object/channel IDs/layout/count/routing, position/ trajectory/spread/priority/automation, coordinate frames, and ranges. Evidence: object-audio tests.
    • E4.20.c Validate ambisonic order/dimensionality/channel order/ normalization/coordinate orientation plus encode/decode rotations. Evidence: HOA tests.
    • E4.20.d Validate binaural HRTF/renderer/head-tracking sensor rate/ latency/frame alignment/dropout/fallback/device metadata. Evidence: binaural/device tests.
    • E4.20.e Render supported speakers/headphones/devices and measure loudness/true peak/phase/downmix/fold-down/localization/metadata loss per cell. Evidence: render tests.
    • E4.20.f Validate ADM-BWF/Atmos or ratified package structure/timing/ IDs/metadata/checksums/target constraints with pinned validators. Evidence: interchange tests.
    • E4.20.g Run known layout/object/trajectory/head-motion/downmix/silence/ boundary/malformed fixtures and compare render goldens/tolerances. Evidence: profile goldens.
    • E4.20.h Publish supported/conditional/unsupported layout×renderer×device× target cells with current proof, known loss, owner, and expiry; fail stale cells. Evidence: release gate.
  • E5.1 Build rights registry for works, recordings, samples, stems, performances, voices, models, artwork, and references. Evidence: kind-specific forms.
    • E5.1.a Implement common rights subject/version/party/jurisdiction/ territory/use/term/evidence/state/search projection with stable IDs. Evidence: registry contract.
    • E5.1.b Build separate work/composition, sound recording/master, sample/ interpolation, and generated stem forms with kind-specific validation. Evidence: music rights forms.
    • E5.1.c Build performer/neighboring/voice/likeness/session consent forms with protected identity and model-use scope. Evidence: identity rights forms.
    • E5.1.d Build model/provider/training/inference/output/reference and artwork/image/design rights forms. Evidence: AI/artwork forms.
    • E5.1.e Support create/review/approve/dispute/expire/revoke/supersede/ hold with immutable history, attachments, and property authorization. Evidence: lifecycle tests.
    • E5.1.f Provide permission-aware search/relations/blast radius without exposing contracts, voiceprints, or unreleased media. Evidence: registry E2E.
  • E5.2 Build sample intake with source, license, territory, term, use, attribution, restrictions, document, and checksum. Evidence: validation.
    • E5.2.a Intake original sample bytes into quarantine and record source/ seller/library/URL/purchase/account/date/project/import actor. Evidence: intake provenance.
    • E5.2.b Preserve immutable bytes/checksum/fingerprint/media metadata/ preview proxy and scan/archive/project safety state. Evidence: file tests.
    • E5.2.c Capture license/rightsholder/licensee/use/synchronization/ derivative/model/distribution/monetization/attribution restrictions. Evidence: licence form.
    • E5.2.d Capture territory/term/effective/expiry/platform/release/project/ transfer/seat limits and current entitlement evidence. Evidence: scope tests.
    • E5.2.e Store contract/receipt/document encrypted with checksum/version/ issuer and redacted ordinary projection. Evidence: document tests.
    • E5.2.f Validate missing/conflicting/expired/nontransferable/no-AI/ no-commercial/unknown terms and block unsupported project use. Evidence: validation suite.
  • E5.3 Build sample and interpolation clearance workflow. Evidence: unresolved use blocks release.
    • E5.3.a Detect/register exact sample/interpolation use by source/range/ clip/stem/mix/master/release with recognizability and transform provenance. Evidence: use registry.
    • E5.3.b Create clearance case with parties/rights types/territories/ channels/term/attribution/fees/deadline/risk and source evidence. Evidence: case schema.
    • E5.3.c Implement research/contact/request/negotiate/counter/approve/ reject/withdraw/expire/dispute transitions with secure communications. Evidence: workflow tests.
    • E5.3.d Store executed licence/consent/fees/conditions/signatures and map exact allowed uses to project/release. Evidence: clearance record.
    • E5.3.e Re-evaluate after edit/replacement/territory/target/date/ ownership/licence change and invalidate stale clearance. Evidence: lifecycle tests.
    • E5.3.f Block mix/master/release/delivery as policy requires when any detected required use is unresolved, disputed, expired, or unsupported. Evidence: release gate tests.
  • E5.4 Build performer and voice consent with purpose, scope, territory, term, model use, revocation, and evidence. Evidence: consent matrix.
    • E5.4.a Define performer/subject identity/public projection/ representative/minor-capacity and protected contact/voiceprint boundaries. Evidence: identity contract.
    • E5.4.b Capture informed purpose/operations/content/voice model/ training-inference/output use/hosts/providers/targets/audience and prohibited uses. Evidence: consent form.
    • E5.4.c Capture project/work/session/range/territory/channel/locale/term/ compensation/credit/privacy/transfer/withdrawal conditions. Evidence: scope contract.
    • E5.4.d Record identity verification, signed intent/time/document hash/ witness or representative authority, evidence, and version. Evidence: signature tests.
    • E5.4.e Implement active/limited/withdrawn/revoked/expired/disputed/ superseded lifecycle and downstream blast radius/remedy. Evidence: lifecycle tests.
    • E5.4.f Resolve applicability per recording/generation/conversion/edit/ model/release and test consent matrix including provider fallback. Evidence: matrix suite.
  • E5.5 Build ownership split proposals, negotiation state, signatures, disputes, and locked release snapshot. Evidence: exact-total and authority tests.
    • E5.5.a Create versioned split proposal per work/recording/right/ territory/window with exact rational shares, parties/roles, rationale, and source evidence. Evidence: proposal schema.
    • E5.5.b Implement invite/view/counter/comment/accept/reject/abstain/ withdraw transitions with party-specific authorization and expiry. Evidence: negotiation tests.
    • E5.5.c Verify party/representative identity/capacity and collect signed exact document hash/intent/time with no agent impersonation. Evidence: signature tests.
    • E5.5.d Open/track dispute over party/share/right/territory/window/ authority with evidence, mediator/legal decision, and no silent overwrite. Evidence: dispute tests.
    • E5.5.e Enforce exact totals/overlap/incompatibility/duplicate/minor/ missing authority and preserve immutable version history. Evidence: property tests.
    • E5.5.f Lock approved release snapshot to exact split version/signatures/ territories and invalidate candidate on material change. Evidence: release snapshot tests.
  • E5.6 Build publishing/master ownership distinction. Evidence: contracts and UI cannot collapse the two.
    • E5.6.a Define separate composition/publishing and recording/master rights subjects/parties/shares/administration/licensing uses. Evidence: domain contracts.
    • E5.6.b Map song/work/composition versions to recordings/masters via explicit many-to-many relationships without ownership inheritance. Evidence: relationship tests.
    • E5.6.c Present separate forms/tables/status/search/approvals for each right with distinct labels/explanations and permissions. Evidence: UI tests.
    • E5.6.d Compute release/territory/target eligibility from both required rights independently and show exactly which is missing/disputed. Evidence: gate tests.
    • E5.6.e Test publishing-only/master-only/same-party/different-party/ cover/remix/sample/license/admin fixtures and API/export mappings. Evidence: noncollapse suite.
  • E5.7 Build melodic, lyrical, recording, and sample similarity screening with source corpus/version and uncertainty. Evidence: calibrated evaluation.
    • E5.7.a Define separate screening tasks/features/candidate corpora/ licences/permissions/retention and protected/unreleased exclusions. Evidence: screening policy.
    • E5.7.b Screen melody/harmony/rhythm using symbolic/audio-derived segments with transposition/tempo/variation controls and aligned evidence. Evidence: melodic tests.
    • E5.7.c Screen lyrics using locale/script/phonetic/semantic/exact-near spans while distinguishing common phrases/quotes/licensed reuse. Evidence: lyric tests.
    • E5.7.d Screen recordings/audio fingerprints and samples/interpolations with exact ranges, transforms, confidence, and corpus/source identity. Evidence: audio tests.
    • E5.7.e Persist model/index/corpus/version/features/threshold/raw matches/ uncertainty separately from rights verdict. Evidence: provenance tests.
    • E5.7.f Evaluate precision/recall/calibration/subgroup/genre/locale/ transform slices and protected corpus leakage. Evidence: calibrated report.
  • E5.8 Treat similarity as review evidence under ratified policy, not an automatic infringement verdict. Evidence: UI copy and gate policy.
    • E5.8.a Define versioned policy mapping similarity kind/score/confidence/ corpus/source/use/context to no-action/review/more-evidence/specialist outcomes. Evidence: policy contract.
    • E5.8.b Display aligned match excerpts/ranges/transforms/corpus rights/ uncertainty/model version and plain-language limitations. Evidence: UI tests.
    • E5.8.c Distinguish common elements, public-domain/licensed/authorized reuse, independent creation, reference, sample, interpolation, and unknown context. Evidence: review fixtures.
    • E5.8.d Require named human rights/legal review for material findings and record rationale/evidence/decision/conditions/dissent. Evidence: decision tests.
    • E5.8.e Prohibit automatic infringement accusation/takedown/payment/ project mutation or source/artist reputation score. Evidence: misuse tests.
    • E5.8.f Recompute/invalidate on subject/corpus/model/policy/rights changes and preserve prior measurement/decision separately. Evidence: lifecycle tests.
  • E5.9 Build AI-generated and AI-assisted disclosure metadata. Evidence: source run IDs and accepted edits.
    • E5.9.a Define disclosure subject/scope/mode/generated-versus-assisted/ operation/materiality/model/provider/version/time/locale/target schema. Evidence: metadata contract.
    • E5.9.b Link exact agent/generation/edit run IDs, input/reference/source ingredients, output ranges/stems, and accepted human edits/actor. Evidence: lineage tests.
    • E5.9.c Define internal reviewer, rights/provenance, distributor/DDEX, C2PA, credit, and listener-visible disclosure requirements by profile. Evidence: mapping matrix.
    • E5.9.d Generate/edit reviewed wording and machine metadata without exposing prompts, protected identity, contracts, or proprietary model data. Evidence: projection tests.
    • E5.9.e Invalidate disclosures on content/run/model/rights/target change and block release if current required metadata is missing. Evidence: gate tests.
    • E5.9.f Round-trip disclosure through project/master/package/DDEX/ distributor/catalog/read-back and correction. Evidence: parity tests.
  • E5.10 Build C2PA/external provenance for supported audio/artwork/package formats. Evidence: sign, validate, tamper, strip, recover.
    • E5.10.a Pin supported C2PA/external ledger profiles/formats/algorithms/ assertions/signers/trust stores/key lifecycle and target capabilities. Evidence: provenance profile.
    • E5.10.b Build manifest/ledger entry from exact audio/artwork/package ingredients, recording/generation/edit/mix/master actions, rights, and release revisions. Evidence: manifest tests.
    • E5.10.c Sign in isolated authorized service/HSM or equivalent with idempotency/timestamp/certificate chain/audit and no key exposure. Evidence: signing tests.
    • E5.10.d Embed/associate per format/target, preserve through supported transforms, and link stripped/external provenance explicitly. Evidence: packaging tests.
    • E5.10.e Validate hashes/signatures/ingredient/action lineage/trust/ revocation/expiry/content binding and display precise status. Evidence: validation tests.
    • E5.10.f Test valid/tampered/replaced/stripped/unknown/untrusted/expired/ malformed/recovered cases plus catalog/distributor read-back. Evidence: provenance suite.
  • E5.11 Build rights/consent/model/sample blast radius. Evidence: revoke preview reaches projects, clips, mixes, masters, releases, and catalog.
    • E5.11.a Ingest or draft exact rights/consent/model/sample change/ revocation/expiry/dispute event with evidence/scope/time/verification. Evidence: event contract.
    • E5.11.b Traverse versioned rights subject→sample/reference/run→media/ clip/take/stem→session/mix/master→release/delivery/catalog dependency graph. Evidence: impact query.
    • E5.11.c Include local/offline projects, frozen/bounced renders, collaborations/review links, provider copies, packages/distributors/ territories and unknown consumers. Evidence: coverage tests.
    • E5.11.d Show exact affected range/use/right/consent/model/territory/ target/current state/proposed remedy/owner/urgency with protected redaction. Evidence: preview UI.
    • E5.11.e Freeze read-only impact snapshot, reconcile to consumer/ capability registry, and require separate authorized remedy plan. Evidence: no-mutation tests.
    • E5.11.f Test complete project→clip→mix→master→release→catalog paths, stale/unauthorized/missing/offline nodes, and graph changes before execution. Evidence: blast-radius suite.
  • E5.12 Build quarantine, replacement, re-clearance, takedown, and correction workflows. Evidence: released-item journey.
    • E5.12.a Quarantine affected inputs/outputs/projects/jobs/exports/ delivery candidates by exact version/range while preserving evidence and authorized investigation access. Evidence: quarantine tests.
    • E5.12.b Create remedy plan with replace/edit/remove/re-record/regenerate/ relicense/reconsent/recredit/territory-limit/takedown actions and dependencies. Evidence: plan schema.
    • E5.12.c Implement replacement through new media/clip/mix/master/release revisions with provenance, semantic/audio diff, revalidation, and approvals. Evidence: replacement tests.
    • E5.12.d Re-clear or renew exact rights/consent/splits/model/sample uses with current parties/terms/signatures and no retroactive silent scope. Evidence: re-clearance tests.
    • E5.12.e Execute distributor/catalog/project/share takedown/restriction saga with per-target status, retry, read-back, history, and notices. Evidence: takedown tests.
    • E5.12.f Execute metadata/audio/provenance/credit correction/redelivery with prior/current versions and verified downstream state. Evidence: correction tests.
    • E5.12.g Run a released-item revocation→quarantine→replacement or takedown→re-clearance→redelivery journey with unresolved target closure blocking. Evidence: E2E.
  • E5.13 Protect contracts, identity, voiceprints, and unreleased media in logs, analytics, notifications, exports, and support tools. Evidence: redaction suite.
    • E5.13.a Classify exact protected fields/artifacts and allowed purpose/ roles/projections for each named observability/output/support surface. Evidence: redaction matrix.
    • E5.13.b Apply allowlisted structured logging/tracing/metrics with opaque IDs, payload scrubbing, sampling/retention/access, and no audio/text excerpts. Evidence: telemetry tests.
    • E5.13.c Apply cohort/privacy floors and prohibit identity/voiceprint/ contract/unreleased metadata dimensions or model-training reuse in analytics. Evidence: analytics tests.
    • E5.13.d Build permission-filtered notification templates/deep links/ previews with device/channel/lock-screen redaction and no existence leaks. Evidence: notification tests.
    • E5.13.e Build purpose-bound audited exports/support bundles with field/ file redaction, encryption, watermark, expiry, and safe diagnostics. Evidence: export/support tests.
    • E5.13.f Inspect UI/network/logs/traces/crash reports/cache/search/queue/ backups/notifications/exports/support across roles/tenants. Evidence: redaction suite.
  • E5.14 Build authorized, scoped, time-bounded waivers. Evidence: expired waiver invalidates release.
    • E5.14.a Define waiverable versus nonwaivable rights/consent/provenance/ technical/target rules and required authority by severity. Evidence: waiver policy.
    • E5.14.b Create waiver for exact subject/use/measurement/gate/profile/ project/release/territory/target/version with reason/evidence. Evidence: waiver schema.
    • E5.14.c Require qualified current human approver, step-up/independence/ quorum as applicable, compensating controls, conditions, and explicit expiry. Evidence: authority tests.
    • E5.14.d Store immutable signed record separately from underlying finding/ measurement/rights state and prevent agent/service impersonation. Evidence: integrity tests.
    • E5.14.e Invalidate on expiry/revocation/subject/profile/target/rights/ consent/approver/condition change and notify release owners. Evidence: lifecycle tests.
    • E5.14.f Recompute release eligibility and block delivery/catalog continuation when waiver is missing/stale/expired; preserve history. Evidence: release gate tests.

E6 — Collaboration, review, realtime, and approvals#

  • E6.1 Build project membership and role assignment. Evidence: least privilege and cross-project isolation.
    • E6.1.a Define membership identity/project/person/service/role/scope/ status/inviter/approver/effective/expiry and inheritance. Evidence: membership schema.
    • E6.1.b Implement invite/request/accept/decline/assign/change/suspend/ revoke/expire with identity verification and current project authority. Evidence: lifecycle tests.
    • E6.1.c Enforce least-privilege object/field/function/media/device/ rights/review/release permissions and role incompatibilities. Evidence: authorization tests.
    • E6.1.d Apply role changes to active sessions/offline caches/review links/ collaboration streams/jobs/search/downloads promptly. Evidence: revocation tests.
    • E6.1.e Test overlapping IDs/names and direct object/function/filter/ count/cache/websocket/media access across two projects/tenants. Evidence: isolation suite.
  • E6.2 Build presence, selection, transport-leader, and edit-lock semantics. Evidence: disconnect and stale-lock tests.
    • E6.2.a Define ephemeral presence/session/device/user/window/project/ workspace/last-seen/expiry and privacy-safe display. Evidence: presence contract.
    • E6.2.b Broadcast permission-filtered selection/cursor/playhead/range/ track/clip/score/lyric context with rate limits and no media/rights leakage. Evidence: selection tests.
    • E6.2.c Define transport leadership request/grant/transfer/revoke/expiry/ conflict and follower behavior for play/seek/record/loop/tempo. Evidence: leader state machine.
    • E6.2.d Define edit locks/reservations by data kind/range/object with optimistic operations where safe, owner/reason/lease/heartbeat/override. Evidence: lock contract.
    • E6.2.e Handle disconnect/sleep/crash/network partition/reconnect/role revoke with stale expiry, recovery, and no permanent lock or dual leader. Evidence: fault tests.
    • E6.2.f Announce ownership/lock changes accessibly and test two-client/ multi-window/realtime ordering/deduplication. Evidence: collaboration E2E.
  • E6.3 Build collaboration for score, lyrics, arrangement, clips, mix, rights, and release metadata. Evidence: conflict policy per data type.
    • E6.3.a Define operation schemas/identity/causality/checkpoints/ authorization for each named structured data type. Evidence: protocol contracts.
    • E6.3.b Choose CRDT/OT/optimistic concurrency/locks per data type based on semantic merge safety and document decisions. Evidence: conflict matrix.
    • E6.3.c Implement concurrent score/lyric/arrangement/clip edits with stable anchors/order/ranges/moves and deterministic convergence. Evidence: creative tests.
    • E6.3.d Implement mix graph/automation/device/routing edits with cycle/ parameter/version/transport safety and explicit semantic conflicts. Evidence: mix tests.
    • E6.3.e Implement rights/split/consent/release metadata changes through stricter expected-version/review/signature rules, never field-level blind merge. Evidence: governance tests.
    • E6.3.f Reauthorize/rebase/deduplicate offline operations, present mine/ theirs/base conflicts, preserve copies, audit, and invalidate approvals. Evidence: two-client suite.
  • E6.4 Build anchored comments on bars/beats, score events, lyric ranges, clips, time ranges, automation, devices, mixes, masters, and rights. Evidence: anchor survival.
    • E6.4.a Define comment/thread identity/subject kind/version/semantic anchor/author/visibility/state/mentions and immutable history. Evidence: comment schema.
    • E6.4.b Implement musical-time/score-event/lyric-range/clip-source/ sample-time/automation-point/device-parameter anchors. Evidence: anchor registry.
    • E6.4.c Implement mix/master revision and protected rights/contract/ split field anchors with role-filtered projections. Evidence: governed anchor tests.
    • E6.4.d Map anchors through insert/delete/move/stretch/tempo/comp/revision changes or mark ambiguous/orphaned with candidates, never silently retarget. Evidence: survival corpus.
    • E6.4.e Implement reply/mention/resolve/reopen/suggestion/notification/ deep-link/playback-focus behavior with redaction. Evidence: workflow tests.
    • E6.4.f Test web/desktop/offline/concurrent edits/export/archive/legal- hold and accessible keyboard/screen-reader journeys. Evidence: collaboration E2E.
  • E6.5 Build review versions and synchronized playback. Evidence: comment always refers to immutable audio/version.
    • E6.5.a Create immutable review manifest pinning project/arrangement/ session/mix/master/media/score/lyrics/metadata/rights/profile versions. Evidence: review version schema.
    • E6.5.b Generate or associate checksum-addressed review audio/video/ waveform/transcript/score proxies with exact range/layout/loudness profile. Evidence: proxy tests.
    • E6.5.c Synchronize play/pause/seek/loop/rate/playhead/time display and optional leader/follower across authorized reviewers with drift bounds. Evidence: playback tests.
    • E6.5.d Create comments/annotations only against immutable review version and exact time/musical/semantic anchor. Evidence: anchoring tests.
    • E6.5.e Show stale/superseded/missing/expired review versions distinctly and preserve previous playback/comments without retargeting. Evidence: lifecycle tests.
    • E6.5.f Test network jitter/reconnect/range requests/device change/ unauthorized download and screen-reader media controls. Evidence: review E2E.
  • E6.6 Build take, arrangement, mix, and master compare with level-matched playback where applicable. Evidence: comparison tests.
    • E6.6.a Select immutable compatible variants and resolve time/tempo/ arrangement/range/layout/source/profile/permission alignment. Evidence: compare contract.
    • E6.6.b Implement synchronized A/B/solo/blind/loop/scrub with transparent loudness/peak level matching and clip-safe gain. Evidence: playback tests.
    • E6.6.c Show take/comp/source, arrangement section/clip, mix graph/ automation, master settings/render/measurement semantic diffs. Evidence: structured diff tests.
    • E6.6.d Show provenance/rights/cost/gates/reviews/freshness and do not infer winner from loudness or similarity. Evidence: evidence compare.
    • E6.6.e Support comments and human select/reject/park/no-decision with stable links and no variant mutation. Evidence: decision tests.
    • E6.6.f Verify sample synchronization/level match/mono-spatial layouts, keyboard/screen-reader parity, and large media performance. Evidence: comparison suite.
  • E6.7 Build review templates for creative, performance, engineering, mix, master, rights, accessibility/metadata, label, and distribution. Evidence: role/quorum matrix.
    • E6.7.a Define versioned template identity/applicability/priority and required checklist/measurement/gate/evidence/reviewer/decision fields. Evidence: template schema.
    • E6.7.b Define independent creative/performance/engineering/mix/master/ rights/accessibility-metadata/label/distribution scopes. Evidence: template registry.
    • E6.7.c Resolve templates by project/release profile/risk/territory/ target/host/version and compose code/policy-owned floors. Evidence: resolution tests.
    • E6.7.d Define reviewer role/credential/domain/independence/recusal/ quorum/order/deadline/delegate and nondelegable human decisions. Evidence: role/quorum matrix.
    • E6.7.e Govern draft/review/approve/activate/supersede/deprecate/migrate/ expire with immutable versions and impact preview. Evidence: governance tests.
    • E6.7.f Preview exact review plan and fail missing/conflicting/ unversioned requirements. Evidence: review-plan gate.
  • E6.8 Build approval locks and invalidation on dependent change. Evidence: source mix or rights change invalidates master/release signoff.
    • E6.8.a Build transitive dependency manifest/hash for each approval over exact project/media/score/lyrics/session/mix/master/rights/profile/ release versions. Evidence: lock contract.
    • E6.8.b Acquire immutable approval lock at decision with actor/role/ scope/conditions/signature/time and prohibit manifest mutation. Evidence: approval tests.
    • E6.8.c Subscribe/reconcile dependency changes and map materiality to stale/invalidate/review-required per approval type. Evidence: invalidation rules.
    • E6.8.d Atomically recheck dependencies before render/delivery/release/ takedown execution to close race windows. Evidence: race tests.
    • E6.8.e Surface exact invalidating dependency/version/change/affected approval/remediation and notify responsible reviewers safely. Evidence: status tests.
    • E6.8.f Verify source mix/media/plugin/profile/rights/consent/splits/ metadata/target change scenarios and immutable prior approvals. Evidence: dependency suite.
  • E6.9 Build collaborator invites, guest review links, expiry, watermark, download control, and revocation. Evidence: link security tests.
    • E6.9.a Define invite/link identity/purpose/project/review version/ recipient-role/allowed fields-actions/expiry/max uses/device/session and revocation. Evidence: access schema.
    • E6.9.b Issue high-entropy signed single/multi-use links with no secret in analytics/referrer/logs, optional identity proof/step-up, and safe landing. Evidence: link tests.
    • E6.9.c Apply permission-filtered media/score/lyrics/comment review view, dynamic visible/audible watermark, redacted rights/unreleased metadata, and no project mutation. Evidence: guest view.
    • E6.9.d Enforce stream/range/download/original/proxy limits, expiration/ offline caching/print/copy controls, and truthful screenshot limits. Evidence: download tests.
    • E6.9.e Revoke immediately across active sessions/cache/CDN/signed media URLs and audit attempts/success without sensitive payload. Evidence: revocation tests.
    • E6.9.f Test guessing/replay/forwarding/cross-project/substitution/ expired/revoked/referrer/log/watermark bypass and role change. Evidence: link security suite.
  • E6.10 Build notifications with unreleased-content redaction and quiet hours. Evidence: routing fixtures.
    • E6.10.a Define event/template/channel/recipient/urgency/dedupe/expiry/ action schema and per-data-class redaction rules. Evidence: notification contracts.
    • E6.10.b Resolve current project membership/role/rights access/device/ channel/preferences/locale/time zone before rendering/delivery. Evidence: recipient tests.
    • E6.10.c Redact titles/artists/file names/audio/waveforms/lyrics/ contracts/voice/model/release dates and use opaque authenticated deep links as required. Evidence: redaction tests.
    • E6.10.d Apply quiet hours/frequency/dedupe/digest with narrowly defined recording/live/safety/release emergency override authority. Evidence: routing tests.
    • E6.10.e Deliver idempotently with provider acknowledgment/retry/ suppression/bounce and mark partial/failed honestly. Evidence: delivery tests.
    • E6.10.f Test lock-screen/email/push/in-app/webhook/log/analytics projections, role revoke, project archive, and two projects. Evidence: routing fixtures.
  • E6.11 Build realtime session health and graceful degradation. Evidence: collaboration outage keeps local work safe.
    • E6.11.a Define realtime connection/auth/protocol/peer/leader/clock/ operation queue/ack/lag/presence/quality state and telemetry. Evidence: health contract.
    • E6.11.b Measure RTT/jitter/loss/reconnect/operation backlog/ack age/ clock drift/audio dropout and calibrate healthy/degraded/offline thresholds. Evidence: health tests.
    • E6.11.c Show persistent accessible connection/sync/leader/unsynced edit/ recording safety state with actionable diagnosis and no false online. Evidence: UI tests.
    • E6.11.d Degrade presence/live cursors/synchronized playback/audio monitor first while preserving local edit/record/autosave/export where safe. Evidence: degradation tests.
    • E6.11.e Queue local operations/media durably and reauthenticate/ reauthorize/rebase/dedupe/conflict-resolve on reconnect. Evidence: recovery tests.
    • E6.11.f Inject service outage/partition/high jitter/protocol skew/leader loss/role revoke/storage pressure and prove local work remains safe. Evidence: outage suite.
  • E6.12 Build immutable review and approval audit. Evidence: automation cannot impersonate a named approver.
    • E6.12.a Define append-only review/decision/approval/waiver event with actor identity/role/scope/subject manifest/action/reason/evidence/time/ signature. Evidence: audit schema.
    • E6.12.b Capture comments/checklists/measurements/changes/conditions/ conflicts/recusals/quorum and exact versions through transactional outbox. Evidence: audit tests.
    • E6.12.c Chain/hash/sign events or manifests, preserve ordering/ supersession/revocation, and prohibit update/delete of history. Evidence: tamper tests.
    • E6.12.d Build protected/redacted permissioned audit query/export with stable cursor/completeness/version and no contract/voice/media leak. Evidence: projection tests.
    • E6.12.e Attempt model/agent/service account/header/DB replay/identity/ role/signature impersonation and prove rejection/no side effect. Evidence: non-impersonation tests.
    • E6.12.f Verify backup/restore/replay/version migration/completeness and independent tamper detection. Evidence: integrity report.
  • E6.13 Build remote rehearsal/recording and live collaboration with clock/tempo authority, cueing, low-latency monitor versus quality-recording paths, dropout concealment, local capture, upload, alignment, consent, and reconnect. Evidence: network-impairment tests preserve the authoritative recordings.
    • E6.13.a Define remote session/participant/role/device/route/clock/ transport/tempo/cue/recording/consent/security and capability profile. Evidence: session contract.
    • E6.13.b Enroll/calibrate participant audio/MIDI devices, input levels/ channels/clock/round-trip latency, permissions, and performer consent. Evidence: setup journey.
    • E6.13.c Elect/transfer clock-tempo-transport authority and distribute count-in/click/cues/markers/talkback with drift/jitter bounds. Evidence: sync tests.
    • E6.13.d Separate low-latency monitor stream from local quality recording, label codec/dropout concealment, and never treat monitor capture as authoritative. Evidence: path tests.
    • E6.13.e Record crash-safe local originals with checksums/device/session/ performer provenance, upload resumably, deduplicate, and retain consent scope. Evidence: capture tests.
    • E6.13.f Align takes using clocks/cues/correlation/manual correction with confidence, preserving raw timing and dropout/missing spans. Evidence: alignment tests.
    • E6.13.g Handle disconnect/reconnect/leader loss/network partition/ device change/role revoke with local continuity and no duplicate or lost takes. Evidence: recovery tests.
    • E6.13.h Run latency/jitter/loss/reorder/bandwidth/outage/clock drift/ process crash tests and verify authoritative local recordings. Evidence: impairment report.
  • E6.14 Build stage/performance mode for set list, scene/patch changes, stems/click/cues, MIDI/OSC/lighting control where ratified, redundancy, rehearsal, operator lockout, emergency stop, recording, and post-show ingest. Evidence: safe device-control allowlists and offline/failover drills.
    • E6.14.a Define locked show/set/song/scene/cue/stem/click/patch/control/ device/operator/revision manifest and performance capability profile. Evidence: show contract.
    • E6.14.b Build set-list/scene transition/rehearsal editor with preflight for media/plugins/presets/devices/routes/tempo/timecode/rights and offline package. Evidence: preflight tests.
    • E6.14.c Implement sample-accurate stem/click/cue playback and safe quantized/manual scene/patch changes with redundant local media/ transport. Evidence: playback tests.
    • E6.14.d Implement allowlisted typed MIDI/OSC/lighting or ratified controls with target/range/rate/state/ack/feedback-loop constraints. Evidence: device control tests.
    • E6.14.e Lock operator UI/actions/permissions during show, require step-up for dangerous changes, and provide immediate audio/control emergency stop/manual fallback. Evidence: safety tests.
    • E6.14.f Implement primary/backup device/computer/audio/output/clock/ power/network failover with health, rehearsal, switchover, and rollback. Evidence: redundancy drills.
    • E6.14.g Record program/stems/MIDI/control/operator/incident timeline locally with consent/provenance and ingest/reconcile post-show. Evidence: recording tests.
    • E6.14.h Run offline/device/plugin/media/clock/network/controller/crash/ failover/emergency-stop drills and prove no unauthorized commands. Evidence: stage game-day.

E7 — Persistence, APIs, rendering, and distribution#

  • E7.1 Decide authoritative project, media, rights, collaboration, and release stores. Evidence: ownership ADR.
    • E7.1.a Inventory current local/OPFS/desktop/database/blob/index/cache/ realtime/provider/distributor records and all writers/readers. Evidence: storage inventory. <!-- euterpe-storage-inventory.json: storage inventory of the Euterpe Tauri DAW. Of the 9 named storage layers, 5 are PRESENT — local (localStorage, 4 writers / 4 readers), desktop (4 Tauri commands: save_project/load_project/ export_stems/reveal_in_file_manager), blob (Tauri export_stems + Blob), realtime (yjs/CRDT collaboration state), distributor (delivery targets) — and 4 are ABSENT: OPFS, database (IndexedDB/SQLite), index, cache. FINDING: Euterpe is DESKTOP-FILE-CENTRIC (persistence via Tauri filesystem commands) + localStorage for client state, with NO browser-persistent storage. 14 project-store modules. Used PRECISE API-signature patterns (navigator.storage. getDirectory, indexedDB.open, caches.open, new Y.Doc) — loose patterns (\bidb\b/getDirectory/"indexing") gave false positives that wrongly marked OPFS/database/index present. generate-euterpe-storage-inventory.mjs + 6 tests, wired into inventory/verify + CI. E7.1.b (aggregate/transaction ownership) + later E7.1.* are design and stay unchecked. -->
    • E7.1.b Define authoritative aggregate/transaction ownership, immutable/ mutable data, durability/latency class, and offline replica per domain. Evidence: ownership matrix.
    • E7.1.c Define media/blob versus project graph, rights/protected records, collaboration operations/checkpoints, and release/delivery boundaries. Evidence: data architecture.
    • E7.1.d Define derived waveform/search/read model/cache/distributor projections, freshness/rebuild/reconciliation, and forbidden reverse writes. Evidence: projection contracts.
    • E7.1.e Resolve duplicate ownership/migrations/outbox/saga/backups/ recovery/region/encryption and accountable services. Evidence: approved ADR.
  • E7.2 Add missing schemas and migrations without duplicating the existing DAW project model. Evidence: old-project migration rehearsal.
    • E7.2.a Derive missing durable entities/columns/indexes/constraints from canonical contracts and required queries while reusing existing project aggregate. Evidence: schema diff.
    • E7.2.b Define adapter/reference boundaries from normalized service data to embedded DAW project state; prohibit a parallel session graph. Evidence: model review.
    • E7.2.c Design expand/backfill/dual-read-write/switch/contract migrations with chunking/checkpoints/idempotency/throttling/rollback. Evidence: migration plan.
    • E7.2.d Preserve timing/media hashes/device state/automation/rights/ provenance/history and validate references/state mappings. Evidence: migration tests.
    • E7.2.e Rehearse old/large/missing-plugin/offline/concurrent projects on production-equivalent stores and compare semantic/audio goldens. Evidence: rehearsal report.
    • E7.2.f Roll out with metrics/reconciliation/stop conditions and remove legacy paths only after browser/desktop/realtime/render compatibility. Evidence: migration gate.
  • E7.3 Store large media as immutable checksum-addressed blobs with resumable upload and derived proxies. Evidence: byte rehash and recovery.
    • E7.3.a Define blob identity/hash algorithm/tenant/project/data class/ size/type/original filename/source/rights/retention/encryption/region. Evidence: blob schema.
    • E7.3.b Implement multipart/chunk resumable upload with content range/ per-chunk and final hash/size/type/quota/idempotency/session expiry. Evidence: upload tests.
    • E7.3.c Quarantine/scan/probe originals before available use and commit metadata/reference only after durable complete blob. Evidence: completion tests.
    • E7.3.d Generate waveform/proxy/transcode/analysis derivatives with immutable parent/hash/tool/profile/version lineage and independent failure. Evidence: proxy tests.
    • E7.3.e Serve authorized range/stream/download via short-lived scoped URLs or service path with no cross-project/cache/path leakage. Evidence: access tests.
    • E7.3.f Inject interrupted/duplicate/out-of-order/corrupt upload, process crash, blob/database mismatch, restore, and rehash all bytes. Evidence: recovery suite.
  • E7.4 Implement project/session/revision APIs with optimistic concurrency. Evidence: contract suite.
    • E7.4.a Define versioned create/read/query/patch/command/transition/ snapshot/diff contracts, errors, idempotency, pagination, and limits. Evidence: API schemas.
    • E7.4.b Enforce tenant/project/role/object/field/function/host capability/ maturity authorization and redacted projections. Evidence: auth tests.
    • E7.4.c Require expected version/ETag or operation causality for mutation, return structured current/conflict, and prevent silent overwrite. Evidence: concurrency tests.
    • E7.4.d Persist immutable revisions/operations/audit/outbox atomically with durable-before-ack and stable media references. Evidence: durability tests.
    • E7.4.e Support batch/command transactions and validate timing/routing/ state invariants server-side rather than trusting client. Evidence: invariant tests.
    • E7.4.f Test retries/timeout-after-commit/stale/offline replay/mass assignment/cross-project/unknown/oversized/version skew. Evidence: contract suite.
  • E7.5 Implement media ingest, analysis, proxy, render, bounce, mix, and master jobs. Evidence: durable job fault tests.
    • E7.5.a Define distinct versioned job request/run/attempt/step/result/ artifact contracts pinned to project/media/range/profile/dependencies/ rights. Evidence: job schemas.
    • E7.5.b Implement durable queue/lease/heartbeat/timeout/retry/backoff/ cancellation/idempotency/priority/concurrency/resource/budget controls. Evidence: worker tests.
    • E7.5.c Implement quarantine ingest/probe/analyze/proxy jobs with immutable source/derivative provenance and safe parser isolation. Evidence: ingest pipeline.
    • E7.5.d Implement render/bounce/mix/master jobs with exact graph/plugin/ state/sample-rate/layout/range/tails/profile and deterministic manifest. Evidence: render pipeline.
    • E7.5.e Persist progress/logs/measurements/artifacts/cost/terminal state before acknowledgment and verify hashes/format/audio output. Evidence: result tests.
    • E7.5.f Inject database/blob/queue/worker/plugin/provider/device/resource/ cancellation/timeout-after-commit faults and recover/replay safely. Evidence: fault suite.
  • E7.6 Implement rights, consent, splits, and provenance APIs with property-level authorization. Evidence: protected-field tests.
    • E7.6.a Define versioned resource/command/query/review/sign/revoke/ blast-radius contracts and canonical kind/state/value validation. Evidence: API schemas.
    • E7.6.b Separate public credits/safe rights projection from protected identity/contact/contracts/voiceprints/signatures/payment fields. Evidence: projection contracts.
    • E7.6.c Enforce tenant/project/role/purpose/object/field/function/ territory/state/compartment authorization on read/write/filter/sort/ export. Evidence: property auth tests.
    • E7.6.d Implement expected-version/idempotency/immutable revisions/ signatures/audit/outbox and prevent mass assignment or direct approval state. Evidence: mutation tests.
    • E7.6.e Emit redacted versioned invalidation/revocation events to project/ job/review/release/catalog consumers. Evidence: integration tests.
    • E7.6.f Test BOLA/existence/count/snippet/filter/error/cache/log/batch/ revoked grant/two-project access. Evidence: protected-field suite.
  • E7.7 Implement permission-aware project, asset, sample, and catalog search. Evidence: unreleased metadata does not leak.
    • E7.7.a Define document/field classification, tenant/project/catalog/ role/right/release/territory/compartment ACL, freshness, locale, and deletion schema. Evidence: search model.
    • E7.7.b Index only allowlisted projections from authoritative events with content/version hashes and protected contract/voice/unreleased fields excluded. Evidence: indexing tests.
    • E7.7.c Apply authorization before retrieval/aggregation/ranking/ highlighting/vector search and stable cursor pagination. Evidence: query tests.
    • E7.7.d Prevent autocomplete/suggestion/snippet/count/facet/timing/cache/ analytics/log leaks of unreleased or cross-project items. Evidence: leakage tests.
    • E7.7.e Process role/release/territory/rights/revocation/takedown/delete promptly and reconcile index to source of truth. Evidence: update tests.
    • E7.7.f Run two-tenant/project/multi-role/existence-oracle/BOLA/ injection/stale-index/large-catalog adversarial suite. Evidence: search report.
  • E7.8 Implement realtime collaboration protocol versioning and recovery. Evidence: version-skew tests.
    • E7.8.a Define handshake/protocol/schema/capability/host/client versions, auth/session/project/actor IDs, resume cursor, and support window. Evidence: protocol contract.
    • E7.8.b Version operation/event/error/presence/selection/leader/lock/ checkpoint/ack envelopes with unknown-field/type handling. Evidence: message schemas.
    • E7.8.c Negotiate compatible subset or fail/read-only/export explicitly; never silently reinterpret musical timing or rights operations. Evidence: negotiation tests.
    • E7.8.d Persist/resume operations with ordering/causality/deduplication/ replay/checkpoints/compaction and current authorization. Evidence: recovery tests.
    • E7.8.e Rebase/migrate queued old-client/offline operations or surface semantic conflict/copied project without data loss. Evidence: migration tests.
    • E7.8.f Test old/new web/desktop/service combinations, rolling deploy, reconnect after history compaction, unknown messages, and role revoke. Evidence: version-skew suite.
  • E7.9 Implement deterministic release bundle assembly. Evidence: audio, artwork, metadata, rights, provenance, and delivery hash goldens.
    • E7.9.a Create immutable release candidate manifest pinning sequence/ recordings/masters/renditions/artwork/lyrics/credits/metadata/rights/ splits/consent/provenance/profiles/targets. Evidence: candidate schema.
    • E7.9.b Resolve target/territory/locale-specific files/metadata/ identifiers/advisories/rights/disclosures and explicit loss/ unsupported state. Evidence: resolution tests.
    • E7.9.c Render/copy/package only checksum-verified immutable artifacts with canonical names/paths/order/permissions and no hidden mutable source. Evidence: assembly tests.
    • E7.9.d Generate canonical manifest and per-file/package hashes/ signatures/C2PA or ledger references plus tool/config/build versions. Evidence: hash goldens.
    • E7.9.e Verify target format/loudness/metadata/artwork/rights/DDEX/ provenance and compare package to locked candidate before approval. Evidence: bundle gate.
    • E7.9.f Rebuild from same inputs across workers and compare bytes/hashes or explicitly documented nondeterministic fields. Evidence: determinism tests.
  • E7.10 Implement DDEX ERN generation/validation for the ratified profile. Evidence: new release, update, takedown, and territory fixtures.
    • E7.10.a Pin DDEX ERN/profile/code-list/namespace/schema/validator/ partner-extension versions and internal mapping owner. Evidence: profile manifest.
    • E7.10.b Map parties/contributors/works/recordings/releases/resources/ identifiers/titles/genres/languages/rights/territories/deals/technical details. Evidence: mapping tests.
    • E7.10.c Generate deterministic new-release message with stable message/ party/resource/release IDs and target-specific resource files. Evidence: new release golden.
    • E7.10.d Generate update/redelivery/metadata correction/audio replacement/ territory addition-removal/takedown lifecycle messages with prior relationships. Evidence: lifecycle goldens.
    • E7.10.e Validate schema/profile/cardinality/codes/dates/territories/ identifiers/deals/resources/checksums/rights and safe extensions. Evidence: validator tests.
    • E7.10.f Round-trip through sandbox/contract distributor, reconcile acknowledgments/content acceptance, and expose unsupported loss. Evidence: conformance report.
  • E7.11 Implement distributor adapters with plan, deliver, acknowledge, reconcile, retry, and compensate. Evidence: partial failure tests.
    • E7.11.a Define per-distributor/version/account/territory/target capabilities, credentials, DDEX/API/SFTP transport, limits, schedules, and unsupported operations. Evidence: adapter profiles.
    • E7.11.b Build read-only delivery plan from locked bundle with messages/ files/metadata/actions/dependencies/idempotency/expected current state/ costs/loss. Evidence: plan tests.
    • E7.11.c Deliver in dependency order with resumable transfer, retries/ backoff/rate, checksums, signed requests, acknowledgments, and durable per-action status. Evidence: delivery tests.
    • E7.11.d Parse transport/message/resource/content acknowledgments and errors without treating receipt as release acceptance. Evidence: acknowledgment tests.
    • E7.11.e Query/read back catalog/territory/metadata/resource/availability state and semantic-compare to plan. Evidence: reconciliation tests.
    • E7.11.f Compensate only safe reversible actions, retain partial release and operator remediation otherwise, with idempotent retry/replay. Evidence: partial failure suite.
  • E7.12 Implement read-back/acknowledgement verification per distributor. Evidence: accepted transport without accepted content is not success.
    • E7.12.a Define distributor-specific acknowledgment layers/codes/ correlation IDs/status transitions/timing and independent read-back capability. Evidence: verification contracts.
    • E7.12.b Correlate transport/file/message/resource/release/deal/territory acknowledgments to exact delivery action and immutable bundle. Evidence: correlation tests.
    • E7.12.c Read back identifiers/sequence/titles/contributors/audio/ artwork/metadata/rights/territory/date/advisory/provenance/catalog state. Evidence: content tests.
    • E7.12.d Semantic-compare with target-specific transforms/tolerances and classify accepted/processing/partial/rejected/mismatch/unknown. Evidence: comparator tests.
    • E7.12.e Keep release/target unresolved until required content/territory state is accepted and verified; expose owner/retry/escalation. Evidence: closure tests.
    • E7.12.f Test transport-accepted/content-rejected, delayed processing, contradictory callbacks, stale read-back, rate/auth outage, and manual reconciliation. Evidence: verification suite.
  • E7.13 Implement takedown, redelivery, metadata correction, audio replacement, and territory change sagas. Evidence: idempotent recovery.
    • E7.13.a Define distinct remedy requests/authority/reason/scope/release/ target/territory/effective time/desired state and immutable plans. Evidence: saga schemas.
    • E7.13.b Generate target-specific dependency actions/messages/files/ notices/idempotency keys/preconditions/expected versions and impact. Evidence: plan tests.
    • E7.13.c Execute takedown and territory add/remove/restrict with per- target status/history/read-back and no deletion of internal evidence. Evidence: availability tests.
    • E7.13.d Execute redelivery/metadata correction/audio replacement with new bundle/version/provenance/QC/rights and prior relationship. Evidence: update tests.
    • E7.13.e Handle duplicate/out-of-order/delayed acknowledgments, timeout after commit, partial targets, retries/backoff/cancel/compensation. Evidence: fault tests.
    • E7.13.f Replay from durable checkpoints and reconcile distributor/ catalog/search/read models without duplicate delivery or stale resurrection. Evidence: recovery tests.
  • E7.14 Implement two-tenant/project isolation across APIs, blobs, search, collaboration, review links, rendering, and distribution. Evidence: integration suite.
    • E7.14.a Create two tenants/projects with overlapping IDs/names/media/ rights/roles/plugins/jobs/releases/distributor accounts. Evidence: isolation fixture.
    • E7.14.b Test API object/property/function/filter/sort/batch/cache/offline replay and service-account authorization. Evidence: API isolation.
    • E7.14.c Test blob upload/range/proxy/waveform/download/signed URL/key/ path/metadata/backup substitution. Evidence: media isolation.
    • E7.14.d Test search suggestions/snippets/counts/facets/vectors/timing and collaboration websocket/operations/presence/locks/review links. Evidence: discovery/realtime isolation.
    • E7.14.e Test render/provider/plugin/job queue/payload/artifact/log/ metrics/dead-letter/callback project context. Evidence: worker isolation.
    • E7.14.f Test distributor accounts/credentials/bundles/messages/ acknowledgments/read-back/takedown and catalog projections. Evidence: delivery isolation.
    • E7.14.g Attempt guessed IDs/stale grants/admin/emergency/cache poisoning/ deep links/cross-project relationships. Evidence: integration suite.
  • E7.15 Generate API and realtime consumer contracts. Evidence: Studio, service, desktop, and distributor drift gates.
    • E7.15.a Generate canonical OpenAPI plus realtime message/state/capability schemas from authoritative route/domain contracts with stable IDs. Evidence: contract artifacts.
    • E7.15.b Produce Studio/service/desktop/render-worker/provider/ distributor audience projections without protected internal fields. Evidence: consumer schemas.
    • E7.15.c Add requests/responses/errors/events/commands/fixtures and generated clients/types pinned to exact spec/version digest. Evidence: fixtures.
    • E7.15.d Run schema lint/breaking-change/server conformance/unknown route- message/unimplemented operation tests. Evidence: provider gate.
    • E7.15.e Run web/desktop/service/worker/distributor compile and semantic contract tests across supported version skew. Evidence: consumer gates.
    • E7.15.f Fail CI on drift, protected-field exposure, orphan operation, stale generated code, unsafe incompatibility, or unversioned protocol. Evidence: drift gates.

E8 — Release calendar, catalog, analytics, and operations#

  • E8.1 Build release calendar for recording, mix, master, rights, artwork, marketing, delivery, embargo, and street date. Evidence: dependency and time-zone tests.
    • E8.1.a Define typed calendar item/reference for each named milestone/ window with owner/status/start/end/deadline/time zone/recurrence and exact subject revision. Evidence: calendar schema.
    • E8.1.b Build permissioned day/week/month/agenda views with stable project/artist/label/owner/type/status/territory/target filters and unreleased redaction. Evidence: view tests.
    • E8.1.c Create/edit/reschedule/assign through owned workflows with dependency/rights/embargo/capacity/target/impact preview, concurrency, and audit. Evidence: action tests.
    • E8.1.d Compute critical path/readiness/conflicts/slack/missed window/ escalation without converting the calendar into release authority. Evidence: dependency tests.
    • E8.1.e Handle UTC/local/label/distributor/territory time zones, DST, ambiguous/nonexistent times, street-date rules, and embargo events. Evidence: temporal tests.
    • E8.1.f Provide agenda/table/forms/non-drag alternatives and run record→ mix→master→rights→delivery→street-date journey. Evidence: accessible E2E.
  • E8.2 Build studio, engineer, performer, review, render, and delivery capacity reservations. Evidence: committed slots cannot overbook.
    • E8.2.a Define resource/skill/credential/device/room/worker/distributor capacity, calendar/time zone/availability/privacy, and units. Evidence: capacity schema.
    • E8.2.b Define tentative/held/committed/released/expired reservation with project/session/review/job/delivery/window/units/priority/dependencies. Evidence: reservation contract.
    • E8.2.c Check studio/engineer/performer/equipment/consent/travel/review/ render resource and distributor window constraints. Evidence: eligibility tests.
    • E8.2.d Commit atomically under concurrent booking, enforce exclusive/ pooled capacity and prevent double booking or worker oversubscription. Evidence: concurrency tests.
    • E8.2.e Reschedule/reassign/cancel/overbook-exception with impact, authority, notifications, cost, and audit. Evidence: operations tests.
    • E8.2.f Test cross-time-zone/recurring/partial capacity/outage/role revoke/deadline conflict and release of stale holds. Evidence: capacity suite.
  • E8.3 Build album/EP/single/playlist sequencing and versioned curation. Evidence: deterministic catalog bundle.
    • E8.3.a Define product/sequence/version/type/disc-side/position/ transition/gap/hidden-bonus/locale/territory/window/state schema. Evidence: curation contract.
    • E8.3.b Select exact recording/master/version/ISRC/assets/rights/ metadata and add/remove/reorder/group using drag and equivalent commands. Evidence: editor tests.
    • E8.3.c Validate duplicate/identifier/version/territory/rights/advisory/ loudness-gap-continuity/format/target and playlist-specific rules. Evidence: eligibility tests.
    • E8.3.d Preview synchronized sequence/transitions/crossfades/gaps/ metadata/artwork/credits and accessible ordered list. Evidence: preview tests.
    • E8.3.e Review/approve immutable version with semantic/audio diff and invalidate on track/master/rights/metadata/profile change. Evidence: review tests.
    • E8.3.f Assemble repeated builds from same inputs and compare order/ metadata/files/hashes or declared nondeterminism. Evidence: deterministic bundle.
  • E8.4 Build catalog search, relation, version, territory, rights, status, takedown, and replacement views. Evidence: permission-aware queries.
    • E8.4.a Define permission-filtered catalog read model with work/recording/ release/version/party/identifier/territory/rights/status/remedy/ replacement/freshness fields. Evidence: catalog contract.
    • E8.4.b Implement text/identifier/facet/semantic search with stable pagination/sorts/suggestions and unreleased/protected/territory authorization before retrieval. Evidence: search tests.
    • E8.4.c Show work↔recording↔version↔release↔product/artist/label/ contributor/sample/provenance relationships with temporal validity. Evidence: relation views.
    • E8.4.d Show per-territory/right-type/party/window eligibility, disputes/ expiry/waivers and redacted evidence without collapsing work/master. Evidence: rights views.
    • E8.4.e Show per-distributor/target status/ack/read-back/delivery history/ errors plus takedown/correction/replacement lineage. Evidence: status views.
    • E8.4.f Test unauthorized titles/counts/snippets/facets/relations, stale index, correction/takedown propagation, large catalog, and semantic accessibility. Evidence: query suite.
  • E8.5 Build project/release dashboards around action, gate, schedule, delivery, and incident status. Evidence: no decorative card mosaic.
    • E8.5.a Define primary triage questions and live queues for assigned work/ blockers/gate failures/upcoming dependencies/delivery exceptions/ rights incidents with owners/times/severity. Evidence: dashboard contract.
    • E8.5.b Compose restrained task-led hierarchy with urgent actionable work first, project/release context and secondary metrics second, no hero or repeated card mosaic. Evidence: design review.
    • E8.5.c Implement stable filters/sorts/grouping/URLs/freshness/live updates and permissioned unreleased/rights/cost projections. Evidence: query tests.
    • E8.5.d Implement authorized open/claim/assign/remediate/retry/hold/ cancel/escalate/review actions with state/version/impact/reason/ confirmation. Evidence: action tests.
    • E8.5.e Render loading/empty/stale/offline/degraded/partial/unauthorized/ fixture/error honestly with accessible recovery. Evidence: state tests.
    • E8.5.f Verify wide/compact/single-column, keyboard/screen reader, scale/ live updates, and no sensitive count/title leaks. Evidence: frontend conformance E2E.
  • E8.6 Build metadata/localization workflow for title, contributors, lyrics, advisories, artwork, credits, and territory. Evidence: locale and distributor parity.
    • E8.6.a Define locale variant pinned to release/work/recording/master/ lyrics/credits/artwork/rights/target/profile revisions. Evidence: locale manifest.
    • E8.6.b Edit title/subtitle/version/phonetic/sort/contributor names/roles/ order/credits and identifier display with language/script rules. Evidence: metadata editor tests.
    • E8.6.c Translate/transliterate/adapt lyrics with segment/timing/ pronunciation/prosody/meaning/explicit advisory parity and human provenance. Evidence: lyric locale tests.
    • E8.6.d Localize artwork/text/advisory/credits/descriptions with rights/ territory/target safe areas and accessibility. Evidence: asset tests.
    • E8.6.e Review terminology/names/claims-like factual metadata/rights/ advisories/locale completeness and exact target field transformations. Evidence: locale review.
    • E8.6.f Preview/generate/read back DDEX/distributor/catalog locale cells, report unsupported/drop/normalization, and block parity gaps. Evidence: distributor parity suite.
  • E8.7 Instrument audio engine, save, collaboration, render, gate, review, delivery, rights, and takedown traces/metrics. Evidence: bounded schema.
    • E8.7.a Define versioned allowlisted event/metric/span schemas, units/ dimensions/sampling/retention/access/owner and prohibited audio/lyrics/ rights/identity fields. Evidence: telemetry catalog.
    • E8.7.b Instrument audio engine lifecycle/device/buffer/sample rate/xrun/ glitch/latency/CPU safely outside realtime critical work. Evidence: engine telemetry.
    • E8.7.c Instrument save/autosave/media upload/durability/offline/recovery and collaboration connection/lag/queue/conflict/convergence. Evidence: editing telemetry.
    • E8.7.d Instrument render/bounce/mix/master jobs and gates/reviews/ approvals with versions/progress/resources/cost/outcome/reasons. Evidence: pipeline telemetry.
    • E8.7.e Instrument rights/consent/splits/expiry/revocation workload and delivery/ack/read-back/takedown/correction per-target state using opaque IDs. Evidence: release telemetry.
    • E8.7.f Propagate trace context through web/desktop/native/service/queue/ worker/provider/distributor without protected payload. Evidence: trace tests.
    • E8.7.g Validate known-value metrics/cardinality/overhead/redaction/ deletion/two-tenant access and obtain privacy/audio review. Evidence: bounded schema gate.
  • E8.8 Define SLOs for save durability, audio glitches, collaboration, render, master QC, delivery, and catalog correction. Evidence: error-budget policy.
    • E8.8.a Define acknowledged-save survival/recovery-loss and media upload/ project revision durability objectives. Evidence: durability SLO.
    • E8.8.b Define audio xrun/glitch/dropout/input-monitor/transport objectives by supported host/device/buffer/project scenario. Evidence: audio SLO.
    • E8.8.c Define collaboration availability/operation latency/lag/ convergence/reconnect/conflict and local-safety objectives. Evidence: realtime SLO.
    • E8.8.d Define render queue/start/completion/failure/resource and master QC evaluator availability/freshness/reproducibility objectives. Evidence: production SLO.
    • E8.8.e Define delivery/ack/content read-back and catalog correction/ takedown propagation/closure objectives per target. Evidence: distribution SLO.
    • E8.8.f Define windows/denominators/exclusions/slices/data sources/ uncertainty/alerts and quality guardrails. Evidence: measurement policy.
    • E8.8.g Define error budgets, stop-the-line, rollout/fallback/disable/ remediation/owner and prohibit throughput from overriding audio/ rights/release correctness. Evidence: error-budget policy.
  • E8.9 Build cost attribution for storage, egress, providers, rendering, mastering, collaboration, and distribution. Evidence: reconciliation.
    • E8.9.a Define cost event/source/invoice/usage unit/currency/rate/version/ time/tenant/project/job/provider/target and allocation policy. Evidence: cost contract.
    • E8.9.b Meter original/proxy/render/archive storage and upload/download/ stream/distributor/provider egress with dedupe/shared allocation. Evidence: infrastructure costs.
    • E8.9.c Attribute provider/model/analyzer/mastering/render-worker CPU- GPU/time/licence/token/currency cost to immutable jobs/attempts/outputs. Evidence: job costs.
    • E8.9.d Attribute collaboration/realtime/relay/recording and distributor/ delivery/message/territory/transaction costs. Evidence: service costs.
    • E8.9.e Show estimated/authorized/accrued/final/late/adjusted/refunded costs with units/currency/FX/date and role-filtered project/release rollups. Evidence: cost views.
    • E8.9.f Reconcile usage to provider invoices/accounts/credits with tolerance, unknown/unallocated/duplicate/late charges and auditable corrections. Evidence: reconciliation report.
  • E8.10 Build backup/restore for projects, media, rights, audit, and release records. Evidence: consistent restore drill.
    • E8.10.a Define protected data/services, cadence/retention/encryption/ immutability/region/key access/recovery order/RPO/RTO and owners. Evidence: backup policy.
    • E8.10.b Back up database/project operations/snapshots, immutable media/ proxies/manifests, rights/contracts/signatures, audit, jobs/releases/ delivery state and configuration references. Evidence: backup tests.
    • E8.10.c Restore selected point/full/incremental sets into isolated environment with tenant/project/key/identity mapping and no production side effects. Evidence: restore trace.
    • E8.10.d Reconcile project graph/media hashes/plugin state/rights/ signatures/audit/search/read models/jobs/catalog/distributor state. Evidence: integrity report.
    • E8.10.e Test corrupt/missing/expired key/partial backup/version skew/ legal hold/large media and fallback or explicit unrecoverable state. Evidence: failure tests.
    • E8.10.f Measure RPO/RTO, validate playable/renderable/releasable project and no duplicate external delivery, then obtain operations/security approval. Evidence: restore drill.
  • E8.11 Publish audio device, plugin, render, rights, delivery, takedown, incident, and recovery runbooks. Evidence: game-day.
    • E8.11.a Publish audio device/permission/rate/buffer/glitch/latency/ dropout/clock and recording recovery runbooks. Evidence: audio runbooks.
    • E8.11.b Publish plugin scan/licence/missing/crash/quarantine/version/ sandbox/safe-mode and project recovery runbooks. Evidence: plugin runbooks.
    • E8.11.c Publish render/job/cost/capacity/output mismatch/determinism/ worker/provider failure runbooks. Evidence: render runbooks.
    • E8.11.d Publish rights/consent/split/sample/provenance/voice revocation/ leak/dispute/waiver runbooks with named legal escalation. Evidence: rights runbooks.
    • E8.11.e Publish distributor delivery/ack/read-back/correction/ redelivery/takedown/territory and partial failure runbooks. Evidence: delivery runbooks.
    • E8.11.f Publish security/privacy/availability/integrity/stage/ collaboration incident and backup/restore/replay runbooks. Evidence: incident/recovery runbooks.
    • E8.11.g Run timed cross-role game-days, verify access/contacts, revise unsafe/missing steps, and reapprove. Evidence: game-day report.
  • E8.12 Build support-safe diagnostics bundle with protected-data redaction. Evidence: privacy test.
    • E8.12.a Define purpose/requester/authorization/project/host/time range/ data classes/bundle schema/retention and user preview/consent policy. Evidence: diagnostics contract.
    • E8.12.b Collect app/build/OS/browser/native/audio/MIDI/device/plugin/ capability/config state with paths/serials/accounts/licenses minimized. Evidence: environment diagnostics.
    • E8.12.c Collect bounded structured logs/traces/crash/audio-engine/job/ network/storage/quota/realtime/distributor status without audio/lyrics/ contracts/voiceprints/tokens. Evidence: telemetry bundle.
    • E8.12.d Include project/media/plugin/rights/release only as opaque IDs, counts/hashes/status and allowlisted problem-specific summaries. Evidence: redaction tests.
    • E8.12.e Scan generated bundle for secrets/PII/paths/media/text/contract leakage, encrypt/sign/hash, deliver via expiring audited channel, and delete on expiry. Evidence: delivery tests.
    • E8.12.f Test representative roles/tenants/projects/voice/rights/ unreleased/stage/offline failures and manual privacy review. Evidence: privacy report.
  • E8.13 Build album/release/program steering overlays for creative, technical, rights, budget, territory, delivery, and cadence constraints. Evidence: tighten-only property tests and impact preview.
    • E8.13.a Define versioned overlay identity/scope/priority/effective window/ owner/reviewer and each allowed creative/technical/rights/budget/ territory/delivery/cadence property. Evidence: overlay schema.
    • E8.13.b Define code/regulatory/contract/global policy floors and a stricter-versus-weaker partial order per property. Evidence: lattice specification.
    • E8.13.c Resolve applicable album/release/program/artist/label/territory/ target overlays deterministically with conflicts explicit. Evidence: resolution tests.
    • E8.13.d Preview affected projects/tracks/renders/gates/reviews/rights/ budgets/schedules/deliveries and newly blocked/required actions. Evidence: impact preview.
    • E8.13.e Govern draft/review/approve/activate/supersede/rollback/migrate/ expire with separation, immutable versions, and no direct floor edits. Evidence: governance tests.
    • E8.13.f Property-test every composition is equal or tighter than floors and cannot relax consent/rights/human approval/safety/target constraints. Evidence: tighten-only suite.
  • E8.14 Build UI-governed genre, mood, instrument, role, version, rights, format, territory, and catalog taxonomies over code-owned floors. Evidence: versioned review, alias, deprecation, localization, and distributor-mapping journey.
    • E8.14.a Define stable concept/code/set/version/label/definition/parent/ relation/alias/status/locale/external mapping and code-floor schema. Evidence: taxonomy contracts.
    • E8.14.b Build authorized draft editor for add/edit/reparent/relate/alias/ merge/split/localize with cycle/duplicate/code/meaning validation. Evidence: editor tests.
    • E8.14.c Map rights/format/territory to code-owned contractual/technical floors and map catalog/genre/mood/instrument/role/version to distributor vocabularies with relation/provenance. Evidence: mapping tests.
    • E8.14.d Preview impact across projects/assets/search/analytics/rights/ policies/DDEX/distributors/catalog before activation. Evidence: impact analysis.
    • E8.14.e Govern review/approve/activate/supersede/deprecate with aliases/ replacements/localization/immutable versions and no floor weakening. Evidence: governance tests.
    • E8.14.f Migrate references/indexes/read models/packages/distributor mappings recoverably, preserve historical meaning, and reject orphan/ unknown values. Evidence: migration journey.
  • E8.15 Build podcast and audiovisual project profiles for episodes, rundown/script, guests, multitrack recording, edit, music/SFX, video, chapters, transcript/captions, artwork, feed/video renditions, rights, review, schedule, and correction. Evidence: project-to-distribution package journey.
    • E8.15.a Define podcast/audiovisual show/season/episode/version/rundown/ script/segment/guest/source/claim/media/release profile contracts. Evidence: profile schemas.
    • E8.15.b Build rundown/script authoring with host/guest roles, consent, source/fact-check links, timing, cues, disclosures, and episode continuity. Evidence: preproduction tests.
    • E8.15.c Capture/import multitrack audio/video with immutable media, devices/session/guest consent, proxies/sync/timecode, and provenance. Evidence: ingest tests.
    • E8.15.d Edit dialogue/music/SFX/video/graphics nondestructively with rights, loudness, aspect/safe-area, source/timecoded anchors, and accessible controls. Evidence: edit tests.
    • E8.15.e Create transcript/speakers/chapters/captions/translations/audio descriptions/artwork/metadata with locale and semantic parity. Evidence: accessibility/localization tests.
    • E8.15.f Run editorial/fact-check/audio-video/rights/accessibility/label/ target review on exact episode version and invalidate on change. Evidence: review tests.
    • E8.15.g Schedule/generate/validate feed/audio/video/article/social renditions/packages and publish/read back through target adapters. Evidence: distribution tests.
    • E8.15.h Propagate correction/retraction/replacement through episode/ transcript/captions/clips/feed/video/embed/archive with timecoded notices. Evidence: correction E2E.
    • E8.15.i Run complete project→record/edit/review→package→distribution journey including partial target failure. Evidence: profile journey.
  • E8.16 Build artist/label marketing operations for campaign brief, assets, approved copy, audience/territory, schedule, links, rights, budget, channel delivery, analytics, and takedown without granting marketing actors mix/master or rights authority. Evidence: separation-of-duties tests.
    • E8.16.a Define separate marketing campaign/brief/asset/copy/audience/ schedule/link/channel/budget/delivery/version/state and role boundary. Evidence: campaign schema.
    • E8.16.b Create brief pinned to approved release/artist/label/territories/ dates/objectives/disclosures/constraints with no access to protected session/contracts. Evidence: brief tests.
    • E8.16.c Select rights-cleared approved artwork/audio/video/snippets/ lyrics/credits and author target copy as proposals with exact release identity. Evidence: content tests.
    • E8.16.d Resolve consented audience/territory/locale/channel/quiet/ frequency and build stable campaign links/UTM or equivalent with canonical correction state. Evidence: targeting/link tests.
    • E8.16.e Budget/estimate/approve/schedule/preview per-channel jobs and execute through target adapters with acknowledgment/retry/partial state. Evidence: delivery tests.
    • E8.16.f Instrument privacy-safe delivery/engagement/conversion/cost and prohibit metrics from changing creative/rights/release approval. Evidence: analytics tests.
    • E8.16.g Propagate release delay/correction/takedown/territory/right change to queued/live campaigns, links/assets, and channels with read-back. Evidence: remedy tests.
    • E8.16.h Attempt marketing-role mix/master/media source/rights/split/ consent/release/takedown mutations and prove rejection. Evidence: separation-of-duties suite.
  • E8.17 Build provider, model, sample-pack, plugin/preset, instrument, and workflow administration for compatibility, license, security scan, evaluation, entitlement, install/update/deprecate/quarantine, dependency impact, and cost. Evidence: untrusted code/content cannot enter an active project silently.
    • E8.17.a Define governed catalog item/version/kind/vendor/source/hash/ compatibility/licence/security/evaluation/entitlement/cost/status/ dependency contract. Evidence: admin schema.
    • E8.17.b Onboard provider/model with API/data-use/region/retention/ training/operation capabilities, credentials, quotas/cost, and evaluation approval. Evidence: provider tests.
    • E8.17.c Onboard sample packs/presets/instruments with immutable content, source/rights/licence/territory/use, metadata, scan, and preview. Evidence: content tests.
    • E8.17.d Onboard plugins/code/workflows with signature/build/architecture/ format/dependencies/permissions/sandbox/security scan and malicious tests. Evidence: code tests.
    • E8.17.e Run versioned functional/audio/quality/performance/compatibility/ safety/equivalence evaluations per supported host/profile. Evidence: evaluation matrix.
    • E8.17.f Govern draft/review/approve/entitle/install/activate/update/ deprecate/quarantine/revoke/rollback with separation and immutable audit. Evidence: lifecycle tests.
    • E8.17.g Preview project/preset/render/provider/workflow dependency impact and migration/substitution/loss before update/quarantine/removal. Evidence: impact tests.
    • E8.17.h Enforce approved version/entitlement/licence at project open/ scan/load/job/render and block silent untrusted code/content activation. Evidence: enforcement tests.
    • E8.17.i Reconcile usage/licence/seats/provider/plugin/storage/egress costs and expose current supported/blocked alternatives. Evidence: cost report.
  • E8.18 Expose conservatory/ear-training, discovery, accompaniment, virtuoso/performance, guitar, sacred, and philosophy capabilities only through a ratified production use case or a governed deep link to their separate product surfaces. Evidence: each library has a concrete disposition; none is implied covered by the DAW.
    • E8.18.a Inventory each named library/package/route/component/API/data/ audience/current consumer and claimed production capability. Evidence: source ledger.
    • E8.18.b Split into atomic capabilities and assess genuine professional production use, consumer/learning/discovery purpose, ownership, security, and dependency fit. Evidence: disposition analysis.
    • E8.18.c Ratify retained in-DAW use cases with exact workflow/context/ role/inputs/outputs/rights/quality/accessibility and owned mutation. Evidence: production use cases.
    • E8.18.d Assign separate product host/owner/audience for nonproduction capabilities and define compatible identity/entitlement/support. Evidence: host decisions.
    • E8.18.e Implement governed deep link/handoff with opaque context, reauthorization/redaction/expiry/return path and no shared unsafe state. Evidence: handoff tests.
    • E8.18.f Mark externalized/deprecated/removed/unsupported explicitly in navigation/docs/capability profile instead of implying DAW coverage. Evidence: product UX tests.
    • E8.18.g Reconcile every library/capability to one reviewed disposition and fail orphan/duplicate/unreachable/unauthorized rows. Evidence: completeness gate.
  • E8.19 Build a DJ/performance profile where ratified for decks, library/ crates, analysis/beat grids, cue/loop, sync, stems, sampler, mixer, effects, transitions, recording, broadcast/stream, set history, hardware, offline media, rights/reporting, and emergency recovery. Evidence: supported controller/format/stream matrix and set-recording provenance; otherwise disposition the DJ workstation as a separate product.
    • E8.19.a Decide integrated profile versus separate DJ product from audience/host/audio/device/offline/library/rights/live-safety/ownership analysis and ratify boundary. Evidence: product ADR.
    • E8.19.b If retained, define deck/library/crate/track/analysis/grid/cue/ loop/set/history/controller/broadcast/recording state and capability contracts. Evidence: DJ schemas.
    • E8.19.c Ingest/search/prepare offline-capable media with waveform/tempo/ key/grid/gain/stems/cues/loops, rights/territory/source, and manual corrections. Evidence: library tests.
    • E8.19.d Implement multi-deck playback/sync/pitch/key lock/cue/loop/stem/ sampler/mixer/effects/transitions with sample-accurate clock, latency, and safe gain. Evidence: performance tests.
    • E8.19.e Integrate ratified controllers/MIDI/HID/audio interfaces with mappings/feedback/reconnect/redundancy and accessible semantic/ keyboard alternatives. Evidence: controller tests.
    • E8.19.f Implement record/broadcast/stream target setup, levels/metadata/ rights/reporting/ack/failover and immutable set/track/action provenance. Evidence: capture/stream tests.
    • E8.19.g Implement performance lock, local offline media, device/computer/ audio/clock/network failover, emergency stop, and post-set recovery. Evidence: stage drills.
    • E8.19.h Validate supported controller×format×host×stream cells, audio/ timing/performance budgets, rights reports, and set-recording lineage. Evidence: profile matrix.
    • E8.19.i If externalized, record separate host/handoff/disposition and remove implied DJ support from DAW profiles/routes/tests. Evidence: exclusion gate.

E9 — Euterpe verification and launch gates#

  • E9.1 Run old/current project migration and revision integrity suites. Evidence: corpus report.
    • E9.1.a Assemble supported historical/current browser/desktop project/ session/media/plugin/preset/rights/release records with provenance and expected semantics. Evidence: corpus manifest.
    • E9.1.b Run stepwise and supported direct migrations with interruption/ retry/rollback/export/downgrade refusal. Evidence: migration matrix.
    • E9.1.c Verify timing/tempo/meter/score/MIDI/clips/comp/routing/devices/ automation/media hashes and state recall. Evidence: DAW integrity.
    • E9.1.d Verify rights/consent/splits/provenance/review/approval/release/ catalog identity and protected boundaries. Evidence: governance integrity.
    • E9.1.e Test missing plugins/media/licences/unknown fields/corrupt/large/ offline/version-skew cases with explicit loss. Evidence: edge suite.
    • E9.1.f Compare semantic diffs/audio render goldens/revision hashes and publish supported compatibility results. Evidence: corpus report.
  • E9.2 Run deterministic DAW command, timing, routing, automation, comp, render, and undo/redo suites. Evidence: command coverage index.
    • E9.2.a Generate complete structured command inventory and map each command to preconditions/effect/inverse/persistence/accessibility/tests. Evidence: coverage index.
    • E9.2.b Run sample-accurate tempo/meter/transport/loop/punch/clip/MIDI/ score/session-launch timing goldens. Evidence: timing results.
    • E9.2.c Run track/bus/send/sidechain/device/plugin/modular/control routing and latency/tail/bypass/failure signal-flow goldens. Evidence: routing results.
    • E9.2.d Run automation/modulation/curve/mode/target/state/realtime- offline deterministic renders. Evidence: control results.
    • E9.2.e Run take/comp/warp/ARA/edit/fade/bounce/freeze/stem/null/source- preservation goldens. Evidence: edit/render results.
    • E9.2.f Run command→undo→redo→save/reopen semantic/audio equality, collaboration/offline boundaries, and failed transaction recovery. Evidence: undo results.
    • E9.2.g Fail any retained command without current deterministic or explicitly nondeterministic coverage and provenance. Evidence: command gate.
  • E9.3 Run known-value score, theory, lyric, recording, mix, loudness, peak, format, rights, split, and delivery tests. Evidence: gate registry coverage.
    • E9.3.a Freeze expert-reviewed positive/boundary/negative/exception/ unsupported/failure fixtures with expected raw measurements/verdicts. Evidence: known-value corpus.
    • E9.3.b Verify score/tuning/theory/rhythm/groove/lyric prosody/rhyme/ pronunciation/time alignment across ratified cultural profiles. Evidence: musical gate results.
    • E9.3.c Verify recording signal/format/noise/dropout/phase and mix peak/ headroom/mono/spectrum/dynamics/masking measurements. Evidence: audio gate results.
    • E9.3.d Verify BS.1770 loudness/LRA/true peak, target profiles, master/ spatial format, null/stem-sum, and failure states. Evidence: QC results.
    • E9.3.e Verify rights kinds/licence/consent/similarity/splits/exact totals/ waivers/provenance/disclosure/release eligibility. Evidence: rights results.
    • E9.3.f Verify DDEX/distributor target eligibility/format/metadata/ territory/ack/read-back/correction/takedown gates. Evidence: delivery results.
    • E9.3.g Map every retained gate to fixtures/measurement and verdict assertions/remediation/owner/current execution; fail gaps. Evidence: gate registry coverage.
  • E9.4 Run calibrated transcription, similarity, AI assist, and generated voice evaluations. Evidence: threshold, subgroup, uncertainty report.
    • E9.4.a Pin held-out corpora/splits/labels across task, instrument/voice/ genre/style/tuning/locale/subgroup/risk/adversarial conditions. Evidence: evaluation manifest.
    • E9.4.b Pin model/provider/prompt/config/code/index versions and run with input/output/right/consent/cost/latency provenance. Evidence: run records.
    • E9.4.c Evaluate audio-to-MIDI/score/chord/lyric transcription accuracy, alignment, calibration, missed important events, and abstention. Evidence: transcription report.
    • E9.4.d Evaluate melodic/lyrical/recording/sample similarity precision/ recall/calibration/transforms/common/licensed cases and corpus leakage. Evidence: similarity report.
    • E9.4.e Evaluate assist meaning/source preservation/diversity/playability/ artifacts/leakage/rights and per-change acceptance behavior. Evidence: assist report.
    • E9.4.f Evaluate voice intelligibility/pronunciation/timing/pitch/ identity drift/artifacts/consent/disclosure/misuse by mode. Evidence: voice report.
    • E9.4.g Report confidence intervals/subgroup deltas/severity errors/ thresholds/regressions and obtain named music/voice/rights/safety approval. Evidence: signed evaluation.
  • E9.5 Run official/calibrated BS.1770/R128 meter test signals. Evidence: conformance report.
    • E9.5.a Pin standard/algorithm/profile/implementation/calibration/ tolerance/channel-layout versions and authoritative test signal source. Evidence: test manifest.
    • E9.5.b Verify signal bytes/sample rate/bit depth/channels/duration/hash and independent expected integrated/short/momentary/LRA/true-peak values. Evidence: fixture validation.
    • E9.5.c Run mono/stereo/multichannel K-weighting/gating/integrated tests and compare within published tolerances. Evidence: integrated results.
    • E9.5.d Run momentary/short-term time series/window/hop and LRA/sufficient duration tests. Evidence: temporal results.
    • E9.5.e Run true-peak/oversampling/silence/short/low-level/layout/invalid/ analyzer-failure fixtures. Evidence: edge results.
    • E9.5.f Compare supported host/worker implementations and publish current- branch conformance artifacts with all versions. Evidence: report.
  • E9.6 Run DDEX ERN golden/conformance validation. Evidence: pinned profile reports.
    • E9.6.a Pin ERN/profile/schema/code-list/namespace/validator/partner- extension versions and deterministic generator build. Evidence: conformance manifest.
    • E9.6.b Validate new single/album/EP/multidisc/multiterritory release resources/releases/deals/parties/identifiers/technical details. Evidence: new-release goldens.
    • E9.6.c Validate metadata update/redelivery/audio replacement/territory add-remove/takedown lifecycle and prior relationships. Evidence: lifecycle goldens.
    • E9.6.d Test missing/invalid/duplicate IDs, share/rights/territory/date/ code/cardinality/resource/checksum/unsupported extension cases. Evidence: negative corpus.
    • E9.6.e Round-trip through sandbox/contract distributor and distinguish transport/message/resource/content/catalog acceptance. Evidence: interoperability tests.
    • E9.6.f Publish current-branch validation/loss/read-back results per supported distributor profile and fail stale/unproven cells. Evidence: profile reports.
  • E9.7 Run real-database, blob, search, queue, collaboration, render, and distributor integration tests. Evidence: durable round-trip.
    • E9.7.a Provision production-equivalent supported service versions with two tenants/projects, large media, plugins, rights, collaborators, workers, and sandbox targets. Evidence: environment manifest.
    • E9.7.b Exercise project edit/save→blob/proxy/analyze→search→collaboration →render/QC→bundle→distributor→catalog/read-back with trace linkage. Evidence: integration trace.
    • E9.7.c Verify rights/consent/splits/provenance/property redaction and no cross-project data in indexes/queues/logs/artifacts/targets. Evidence: isolation tests.
    • E9.7.d Inject DB/blob/search/queue/realtime/worker/plugin/provider/ distributor timeout/outage/duplicate/order/auth/version faults. Evidence: fault tests.
    • E9.7.e Reconcile authoritative records/media hashes/indexes/operations/ jobs/bundles/acknowledgments/catalog after create/update/takedown/ restore. Evidence: reconciliation report.
    • E9.7.f Restart/replay/backup/restore and prove acknowledged saves/media/ decisions/deliveries survive without duplicate external effects. Evidence: durable round-trip.
  • E9.8 Run full-journey Playwright: concept to arrangement and saved revision. Evidence: live audio engine where browser automation permits.
    • E9.8.a Create a real project and capture text/hum-or-fixture audio/MIDI/ chord/reference concepts with provenance/rights prompts. Evidence: concept trace.
    • E9.8.b Generate or author multiple arrangement/score/MIDI/lyric candidates, compare synchronously, and record human decision. Evidence: creative trace.
    • E9.8.c Edit sections/tracks/clips/MIDI/score/lyrics/devices/automation through pointer and keyboard paths, including undo/redo. Evidence: editing trace.
    • E9.8.d Autosave to live store, close/reopen, inspect history/semantic diff, restore as new revision, and verify media/plugin state. Evidence: persistence trace.
    • E9.8.e Exercise offline/degraded/provider-not-configured/role/stale/ conflict state and no simulated success. Evidence: resilience tests.
    • E9.8.f Use live browser audio engine/virtual devices where automation supports it; otherwise pin deterministic audio fixtures and record exact limitation. Evidence: Playwright report.
  • E9.9 Run full-journey Playwright: recording/import to mix/master QC. Evidence: deterministic fixture media.
    • E9.9.a Configure virtual/fixture device and calibrate, record or import scanned deterministic multichannel media with take/session provenance. Evidence: capture trace.
    • E9.9.b Perform nondestructive waveform/clip/take/comp/fade/stretch/ repair/alignment edits and prove source restore. Evidence: edit trace.
    • E9.9.c Build track/bus/send/sidechain/device/plugin/automation mix, render it, and verify deterministic signal-flow/null outputs. Evidence: mix trace.
    • E9.9.d Create master candidate with target/reference/profile, process/ render, and run loudness/peak/format/spatial/QC measurements. Evidence: master trace.
    • E9.9.e Review accessible values/units/thresholds/history/remediation, fix a failed gate, rerun, and obtain exact-revision human approval. Evidence: QC trace.
    • E9.9.f Test missing/crashing plugin, device loss, analyzer failure, cancel/retry, stale source mix, and no nonpass displayed as pass. Evidence: Playwright report.
  • E9.10 Run full-journey Playwright: rights/splits to delivery and catalog verification. Evidence: sandbox distributor or contract adapter.
    • E9.10.a Register work/master/sample/performer/voice/model/artwork rights and consents with protected projections. Evidence: rights trace.
    • E9.10.b Negotiate/sign exact ownership splits, resolve sample clearance, and verify publishing versus master totals independently. Evidence: split/clearance trace.
    • E9.10.c Build release sequence/metadata/locales/artwork/territories/ target profiles and lock exact rights/master/provenance candidate. Evidence: release trace.
    • E9.10.d Assemble deterministic bundle, validate DDEX/format/QC/ disclosure/rights, approve, and deliver to sandbox/contract distributor. Evidence: delivery trace.
    • E9.10.e Correlate acknowledgments and read back catalog audio/metadata/ identifiers/territories/status; reject transport-only success. Evidence: catalog proof.
    • E9.10.f Test partial target failure/retry, role/field authorization, protected data nonleakage, and accessible status. Evidence: Playwright report.
  • E9.11 Run full-journey Playwright: rights revocation to takedown. Evidence: affected catalog targets reconcile.
    • E9.11.a Seed a delivered release using rights/consent/sample/model across project/clips/mix/master/bundle/distributor/catalog. Evidence: baseline.
    • E9.11.b Revoke/expire/dispute an exact right or consent and preview the complete read-only blast radius with protected redaction. Evidence: impact trace.
    • E9.11.c Approve quarantine/replacement/re-clearance/takedown plan and invalidate active reviews/release/jobs/links. Evidence: remedy plan.
    • E9.11.d Execute per-target takedown/restriction/notice saga with duplicate/out-of-order/failure injection and truthful progress. Evidence: saga trace.
    • E9.11.e Read back distributor/catalog/search/share states, keep one failed target unresolved, repair/retry, and close only after reconciliation. Evidence: target proof.
    • E9.11.f Verify immutable prior artifacts/audit/provenance, no protected leakage, and optional replacement/redelivery lineage. Evidence: Playwright report.
  • E9.12 Run WCAG 2.2 AA, keyboard, announcement, non-drag, and semantic DAW parity suites. Evidence: themes/viewports/locales.
    • E9.12.a Define critical arrange/session/piano roll/score/mixer/device/ modulation/plugin/review/rights/release route/action matrix across hosts/viewports/themes/locales. Evidence: coverage plan.
    • E9.12.b Run automated WCAG/name-role-value/landmark/form/error/contrast/ focus/target/reflow/motion assertions around code-native DAW surfaces. Evidence: automation report.
    • E9.12.c Run pointer-free create/edit/record/route/automate/comp/mix/ review/render/release with remapped commands and no traps. Evidence: keyboard report.
    • E9.12.d Run non-drag position/range/reorder/connect/curve/scene/device operations with identical validation/undo/audit. Evidence: dragging report.
    • E9.12.e Compare visual and semantic timeline/mixer/device/modulation/ automation/graph/quality/compare views for content and operations. Evidence: parity report.
    • E9.12.f Assert transport/record/save/conflict/device/job/gate/review/ delivery/takedown announcements without audio/content leakage. Evidence: announcement report.
    • E9.12.g Run supported screen-reader/zoom/high-contrast/RTL/long locale/ touch/desktop manual review and close blockers. Evidence: signed matrix.
  • E9.13 Run audio glitch, input latency, large project, memory, waveform, mixer, render, autosave, collaboration, and upload performance tests. Evidence: budgets pass.
    • E9.13.a Build production-scale fixtures/matrices for hosts/devices/ buffers/rates/tracks/clips/media/plugins/automation/score/collaborators/ rights/jobs. Evidence: load manifest.
    • E9.13.b Measure input-to-monitor/control-to-audible/transport timing, audio-thread CPU/xruns/glitches/clicks under increasing load. Evidence: realtime profile.
    • E9.13.c Measure project open/save/autosave/recovery, waveform/spectrum/ scroll/zoom, mixer/device UI input/paint and sustained memory. Evidence: editor profile.
    • E9.13.d Measure realtime/offline/batch render speed/resources/ determinism, upload/proxy/resume/egress and worker concurrency/cost. Evidence: pipeline profile.
    • E9.13.e Measure collaboration operation latency/lag/convergence/ reconnect/conflict with multiple clients and impaired networks. Evidence: realtime profile.
    • E9.13.f Capture audio/UI threads/CPU/GPU/memory/disk/network/query/queue/ plugin traces, compare to budgets, trend, and fail regressions. Evidence: performance report.
  • E9.14 Run malicious media/project/plugin/preset, prompt injection, voice impersonation, rights fraud, leak, auth, tenant, and distribution adversarial tests. Evidence: zero blockers.
    • E9.14.a Test malformed/polyglot/bomb/active audio-project-preset/archive/ metadata/media and unsafe parsers/paths/resources. Evidence: content security report.
    • E9.14.b Test malicious/hanging/crashing plugin/module/UI/state/update/ licence/filesystem/network/sandbox escape and safe-mode recovery. Evidence: plugin security report.
    • E9.14.c Inject prompts/instructions through audio metadata/lyrics/ presets/projects/references/provider outputs and agent handoffs. Evidence: injection report.
    • E9.14.d Attempt voice/performer impersonation, cloned consent/signatures, voiceprint extraction, revoked output/model reuse, and fraud. Evidence: identity report.
    • E9.14.e Attempt fabricated/licence/rights/splits/sample/provenance/ similarity/waiver/disclosure manipulation and takedown suppression. Evidence: rights report.
    • E9.14.f Probe unreleased media/contracts/identity/voice through UI/API/ search/log/cache/notifications/review links/exports/support/provider/ distributor. Evidence: leakage report.
    • E9.14.g Test auth/session/role/BOLA/property/function/tenant/project and mass/replayed distribution/ack/callback/credential attacks. Evidence: authorization/distribution report.
    • E9.14.h Require zero blockers, assign regression/owner/mitigation and expiring approved lower-risk exceptions. Evidence: adversarial gate.
  • E9.15 Run crash recovery, offline sync, backup/restore, and distribution replay drills. Evidence: RPO/RTO.
    • E9.15.a Crash tab/app/audio worklet/native host/plugin/worker during recording/edit/save/upload/render and recover acknowledged project/ local originals/operations. Evidence: crash drills.
    • E9.15.b Run offline create/edit/record/render/export, two-client changes, reconnect/rebase/conflict/role revoke/quota/eviction/storage loss. Evidence: offline drills.
    • E9.15.c Restore project/media/plugin state/rights/contracts/audit/ release/jobs from selected backups into isolated environment. Evidence: restore trace.
    • E9.15.d Reconcile hashes/revisions/routing/timing/rights/signatures/ indexes/catalog/distributor state and prove playable/renderable project. Evidence: integrity report.
    • E9.15.e Replay delivery/takedown/correction saga from checkpoints under duplicate/out-of-order/partial prior actions and read back targets. Evidence: replay tests.
    • E9.15.f Measure achieved RPO/RTO, record unrecoverable external limits, revise runbooks, and obtain operations/security approval. Evidence: signed drill report.
  • E9.16 Produce engineering-readiness evidence. Evidence: requirements map to current proof.
    • E9.16.a Define required architecture/contracts/migrations/audio/ compatibility/security/privacy/rights/accessibility/quality/standards/ operations evidence by host/profile. Evidence: readiness matrix.
    • E9.16.b Map every retained atomic capability/checklist parent/child to implementation/tests/goldens/evaluations/reports/runbooks/approvals. Evidence: evidence index.
    • E9.16.c Verify evidence provenance matches candidate commit/build/ schemas/audio engine/plugins/models/providers/profiles/distributors/ devices. Evidence: provenance check.
    • E9.16.d Reconcile route/package/capability/command/gate/device/plugin/ format/target/consumer inventories with no orphan/unknown rows. Evidence: completeness report.
    • E9.16.e Classify missing/stale/failed/waived evidence, block required cells, and record exception scope/owner/compensation/expiry. Evidence: readiness gate.
    • E9.16.f Publish per-host/profile go/no-go, supported/conditional/ unsupported cells, residual risks, support/rollback. Evidence: readiness packet.
  • E9.17 Obtain named music-production, mastering, rights/legal, accessibility, privacy/security, engineering, and product approvals. Evidence: human signoff.
    • E9.17.a Define each approver’s scope/criteria/required evidence/ qualification/independence/delegate/expiry and nondelegable decisions. Evidence: approval policy.
    • E9.17.b Route immutable creative/production/audio-engine/interchange/ specialized workflow evidence to named music-production reviewers. Evidence: production decision.
    • E9.17.c Route loudness/master/spatial/target conformance and delivery evidence to qualified mastering/release reviewers. Evidence: mastering decision.
    • E9.17.d Route rights/consent/splits/sample/similarity/voice/model/ provenance/distributor evidence to rights/legal specialists. Evidence: legal decision.
    • E9.17.e Route WCAG/blind-producer and privacy/security/threat/ adversarial/source-redaction evidence to named specialists. Evidence: assurance decisions.
    • E9.17.f Route architecture/quality/operations/recovery/support and product scope/host/rollout/residual risks. Evidence: engineering/product decisions.
    • E9.17.g Record signed approve/reject/request-change/conditional decisions and invalidate on candidate/evidence/profile/model/plugin/target change. Evidence: human signoff.
  • E9.18 Obtain launch approval for each supported host and distribution profile. Evidence: human governance record.
    • E9.18.a Build exact host×capability×device/plugin/offline/provider and distribution×territory/format/layout/profile candidate matrix. Evidence: launch scope.
    • E9.18.b Assemble per-cell current compatibility/performance/audio/ security/accessibility/rights/QC/DDEX/read-back/recovery proof. Evidence: launch packets.
    • E9.18.c Verify build/signing/config/feature flags/migrations/models/ plugins/profiles/credentials/runbooks/monitoring/on-call/rollback. Evidence: operational checklist.
    • E9.18.d Run final live-backed smoke per supported host and representative distributor profile, including correction/takedown read-back. Evidence: launch smokes.
    • E9.18.e Record named go/no-go/conditional decision per cell with scope/ limits/known loss/rollout/conditions/owner/expiry/rollback trigger. Evidence: governance record.
    • E9.18.f Fail launch on unproven/stale/unknown supported cell and invalidate approval on candidate/profile/target change. Evidence: launch gate.
  • E9.19 Run distinct full journeys for instrument/effect authoring, hardware/MIDI/MPE, culturally diverse tuning/notation, stem separation, inpainting/style transfer, voice conversion/singing/choir, AI composition, local/cloud AI negotiation, modular patching, immersive/spatial mix, remote recording, stage/DJ mode, podcast/video, and marketing handoff. Evidence: no generic DAW smoke test substitutes for specialized proof.
    • E9.19.a Run instrument/sampler/synth and effect/signal-chain authoring through state/preset/automation/modulation/freeze/render/missing-device recovery. Evidence: device journey.
    • E9.19.b Run hardware audio/MIDI/MIDI 2/MPE/pedal/control/external effect discovery→map/calibrate/record/recall→disconnect/reconnect. Evidence: hardware journey.
    • E9.19.c Run culturally diverse/non-12-TET tuning/notation/world-rhythm/ theory/playback/interchange with expert assumptions and loss proof. Evidence: inclusive music journey.
    • E9.19.d Run stem separation through source rights/job/stem review/manual correction/recombination/stem-sum/lineage. Evidence: separation journey.
    • E9.19.e Run inpainting/continuation/style transfer through exact mask/ condition/reference rights/candidates/audition/comp and untouched-region preservation. Evidence: generative edit journey.
    • E9.19.f Run voice conversion, speech/singing/choir/harmony through consent/identity/locale/job/evaluation/disclosure/revocation. Evidence: voice journey.
    • E9.19.g Run melody/harmony/rhythm/arrangement/orchestration/text-to-music plan through editable score/MIDI/stem/audio intermediates and part-level decisions. Evidence: composition journey.
    • E9.19.h Run local/on-device and cloud execution negotiation through hardware/privacy/rights/cost/quality choice, offline behavior, fallback, and result-equivalence proof. Evidence: AI location journey.
    • E9.19.i Run modular patch create/connect/modulate/automate/performance/ save/render with feedback/CPU/crash safety and semantic list parity. Evidence: modular journey.
    • E9.19.j Run object/bed/ambisonic spatial mix through layout/automation/ renderer/head tracking/downmix/QC/interchange/target delivery. Evidence: immersive journey.
    • E9.19.k Run remote rehearsal/record through device consent/clock/cues/ monitor/local originals/upload/alignment/reconnect under network impairment. Evidence: remote session journey.
    • E9.19.l Run stage mode and, if retained, DJ mode through preflight/ offline package/control allowlist/performance lock/record/failover/ emergency stop/post-ingest. Evidence: live performance journeys.
    • E9.19.m Run podcast/audiovisual project through rundown/guest consent/ multitrack edit/captions/rights/review/feed-video package/correction. Evidence: podcast/video journey.
    • E9.19.n Run marketing handoff through approved release/assets/copy/ audience/budget/channels/analytics/takedown while denying production/ rights mutations. Evidence: marketing journey.
    • E9.19.o Verify each journey uses owned live services/host profile, deterministic or calibrated media, roles/rights, failure recovery, trace/audit, and current proof. Evidence: specialized coverage ledger.
  • E9.20 Run pairwise conformance across supported browser/native host, audio/MIDI device, plugin/instrument, sample rate/bit depth, channel/ spatial layout, renderer, provider/model, project profile, interchange format, and delivery target. Evidence: unsupported cells are explicit and supported cells have current evidence.
    • E9.20.a Build canonical dimension/value/constraint model from ratified hosts/OS/architectures/devices/plugins/formats/rates/layouts/renderers/ models/profiles/targets. Evidence: compatibility model.
    • E9.20.b Generate deterministic pairwise cells plus required higher-order audio-risk combinations and label invalid/unsupported cells with rationale. Evidence: coverage set.
    • E9.20.c Define per-cell semantic/audio/timing/state/rights/provenance/ accessibility/performance/loss tolerances and required fixture/hardware. Evidence: conformance contract.
    • E9.20.d Execute supported cells on exact versioned hosts/devices/ virtual hardware/plugins/models/renderers/formats/targets and retain manifests/goldens. Evidence: conformance results.
    • E9.20.e Exercise version skew/capability absence/degradation/fallback/ warnings/refusal/recovery without silent quality/state/rights loss. Evidence: boundary tests.
    • E9.20.f Publish current supported/conditional/unsupported matrix with evidence age/known loss/prerequisites/owner/expiry and fail stale or unproven supported cells. Evidence: release gate.
  • E9.21 Run full journeys for clip-launch performance/capture, source→target modulation, native/browser plugin scan/load/crash/recovery, ARA/warp/comp, deep interchange round-trips, external sync/control surfaces, offline-first storage/sync, customized accessible workspaces, and batch operations. Evidence: each structural DAW capability has current browser/native proof and fault coverage.
    • E9.21.a Run clip/scene launch/quantize/legato/follow/probability/ performance lock/record/capture-to-arrangement with sample-accurate and pointer-free proof. Evidence: session journey.
    • E9.21.b Run every modulation source family through typed target routing/ depth/polarity/range/smoothing/cycle safety/visual-semantic control and deterministic render. Evidence: modulation journey.
    • E9.21.c Run native VST3/CLAP/AU or ratified and browser WAM scan/ signature/licence/load/state/UI/process/crash/quarantine/recover per supported host. Evidence: plugin journey.
    • E9.21.d Run ARA/deep edit and warp/transient/tempo/formant/comp context through save/undo/reopen/render/missing-editor and source-preservation. Evidence: deep-edit journey.
    • E9.21.e Run DAWproject/MIDI2/MusicXML/BWF-iXML/AAF-OMF-FCPXML/ADM-BWF or ratified import/export cross-host with complete loss profiles. Evidence: interchange journeys.
    • E9.21.f Run Link/MTC/LTC/MIDI clock/MMC/OSC/control surface/headless API authority/drift/feedback/reconnect/unsafe command. Evidence: sync/control journey.
    • E9.21.g Run offline-first create/record/edit/undo/render/export/reopen/ quota/eviction/reconnect/merge/storage-loss across PWA and desktop. Evidence: offline journey.
    • E9.21.h Run saved multi-window/touch/high-DPI/control-surface/blind- producer workspace across arrange/session/score/mixer/modulation/plugin/ review, including reconnect. Evidence: workspace journey.
    • E9.21.i Run mixed-eligibility batch render/export/convert/analyze through preflight/cost/capacity/queue/partial/retry/cancel/output verification/ deterministic manifest. Evidence: batch journey.
    • E9.21.j Reconcile each capability to current browser/native proof, device/profile/format cells, fault injection, accessibility, and owned artifacts; fail generic substitution. Evidence: structural DAW gate.
  • E9.22 EVE-SEAM-EXIT: require "required Eve reads and writes admitted with source and behavior evidence, or an exact task-0.8 explicit non-goal proves zero operations; a named defer preserves ownership but never passes completion" before this phase exits. Evidence: source-derived totality record and passing verifier.
    • E9.22.a Re-run the route, bridge, state-signal, and complete domain-code envelope census; review every changed candidate instead of carrying a stale phase disposition forward. Evidence: current ratchets.
    • E9.22.b Bind the exact ratified task-0.8 workflow and requirement set; classify it as a required seam or an exact source-backed non-goal. Evidence: charter-bound disposition set.
    • E9.22.c For a required seam, admit both read and write operations with byte-pinned source and behavior evidence. A named defer retains its owner and unblock condition but fails completion. Evidence: operation proofs.
    • E9.22.d Permit a zero-operation boundary only for an exact ratified task-0.8 explicit non-goal; absence or unavailability is not a non-goal. Evidence: source-bound non-goal proof.
    • E9.22.e Run the totality verifier, every retained gate-case CLI, and the injected durable-candidate control; retain red and green receipts. Evidence: release gate logs.

Phase A — Aja motion-capture operations console#

Aja is a processing and embodied-instruction subsystem. Its console must manage capture intake, consent, pipeline execution, motion QA, retargeting, annotation, asset curation, and export while preserving the canonical Aja adapter used by Metis and other V1 consumers.

A0 — Host, boundary, pipeline, and risk decisions#

  • A0.1 Decide whether the canonical console lives in a new Aja control room or an Oshun Studio operator route. Evidence: host/ownership ADR.
    • A0.1.a Inventory current Studio routes/Aja services/operator audiences, identity/project context, deployment, security, and preview ownership. Evidence: host inventory.
    • A0.1.b Compare dedicated control room versus Studio route for queue/ worker/live capture/biometric isolation/3D performance/support needs. Evidence: option analysis.
    • A0.1.c Decide canonical host, owning team, routes/deep links/shell/data/ mutation boundaries and prohibited duplicate console behavior. Evidence: target architecture.
    • A0.1.d Define migration/redirect/session/handoff/rollout/rollback/ monitoring/support consequences. Evidence: delivery plan.
    • A0.1.e Obtain product/mocap/architecture/privacy/security/accessibility ratification. Evidence: approved ADR.
  • A0.2 Inventory all static Studio routes, Aja services, CLI commands, libraries, contracts, and existing preview/annotation components. Evidence: route-capability matrix. <!-- partial: route (a) + service (b) + code/CLI (c) inventories done in evidence/v1-workbenches/aja-inventory.json. d-e (test/flag/deployment coverage, route-capability reconciliation) are design/human and stay unchecked. -->
    • A0.2.a Enumerate routes/navigation/commands/pages/deep links and current operator/consumer journeys. Evidence: route inventory. <!-- evidence: .studioRoutes — 98 /studio/aja/* Studio pages (98 distinct segments, flat structure), path/segment/dynamic/test. Aja has no dedicated web app; the console is a Studio-route surface. generate-aja-inventory.mjs. -->
    • A0.2.b Enumerate services/endpoints/workers/queues/models/stores/blobs/ caches/search/providers and accountable owners. Evidence: service inventory. <!-- evidence: .services — the 3 motion services classified: svc-motion-ai (http-service, 9 /api/v1/* + /health routes), svc-motion-pipeline (http-service, 25 routes), svc-reference-video (library/no-http). owners are A0.2.b's accountable-owner assignment (semantic). -->
    • A0.2.c Enumerate CLI commands/libraries/contracts/adapters/formats/ preview/annotation/render components and consumers. Evidence: code inventory. <!-- evidence: .cli — 7 CLI commands (compare, formats, metrics, ping, skeletons, stats, version) across the command modules. -->
    • A0.2.d Map tests/docs/flags/deployments to atomic capabilities and classify live/fixture/local/service/incomplete. Evidence: coverage ledger. <!-- runtime/design: live/fixture classification needs the services run. -->
    • A0.2.e Reconcile route→capability→owner/host/data class/profile and flag duplicate/orphan/dead/unsafe paths. Evidence: route-capability matrix.
  • A0.3 Map all libs/aja/* packages to pipeline stages, console capabilities, consumer adapters, or out-of-scope. Evidence: complete ledger.
    • A0.3.a Inventory every package/export/consumer/runtime/test/documented purpose and dependency direction. Evidence: package inventory. <!-- evidence: .domains.aja.packages — 41 libs/aja/* packages with exports/consumer/dependency/test/disposition/direction. evidence/v1-workbenches/domain-package-inventories.json (254 pkgs; per-package exports/consumers/deps/disposition/tests + capability catalog with exact source locations). -->
    • A0.3.b Map retained exports to typed pipeline stages, console functions, canonical contracts/adapters, hosts, and owners. Evidence: capability mapping.
    • A0.3.c Classify duplicate/legacy/demo/consumer-specific/dead code with keep/merge/migrate/deprecate/remove/out-of-scope disposition. Evidence: disposition ledger.
    • A0.3.d Verify each package/export has one reviewed disposition and no private or circular boundary violation. Evidence: completeness gate.
  • A0.4 Verify the canonical @oshun/embodiment-aja contract boundary. Evidence: console work does not introduce private consumer imports.
    • A0.4.a Inventory package public exports/schema versions/consumers and all direct/transitive private import paths. Evidence: boundary inventory. <!-- evidence: aja-embodiment-boundary.json — @oshun/embodiment-aja declares 3 public export subpaths (. / adapter / canonical-adapter, 76 public symbols); 3 consumer files, each importing only declared public subpaths. Private/deep-import violations = 0 (boundaryClean=true) — the A0.4 invariant ("console work does not introduce private consumer imports") holds today. generate-aja-embodiment-boundary.mjs + 6 tests. -->
    • A0.4.b Define canonical capability/demonstration/overlay/handoff/health/ error contracts and ownership/version policy. Evidence: public contract.
    • A0.4.c Refactor or disposition private consumer imports using public adapters without leaking console/internal stage/storage models. Evidence: dependency changes.
    • A0.4.d Add architecture/import/lint/build/contract tests that fail private subpath, reverse dependency, duplicate contract, or drift. Evidence: boundary gate.
    • A0.4.e Run Metis and other consumer conformance with healthy/degraded/ unsupported states. Evidence: consumer report.
  • A0.5 Define operator, capture technician, animator, rigger, QA reviewer, consent officer, privacy reviewer, annotator, producer, publisher, and administrator roles. Evidence: permission matrix.
    • A0.5.a Define least-privilege duties/object/field/function rights for every named role across capture→package→recall. Evidence: role definitions.
    • A0.5.b Define tenant/project/session/subject/consent compartment/profile/ worker/export scope, delegation, guest, and temporary grants. Evidence: scope policy.
    • A0.5.c Define ingest/live-device/configure/run/pause/kill/annotate/ cleanup/map/review/approve/export/recall/admin rights by state. Evidence: permission matrix.
    • A0.5.d Define separation/qualification/quorum/recusal for consent/ privacy/QA/vertical specialist/publish and nondelegable human actions. Evidence: SoD policy.
    • A0.5.e Verify UI, API, queue, blob, preview, export, and direct command enforcement for positive/negative/revoked/cross-project cases. Evidence: authorization suite.
  • A0.6 Ratify the canonical pipeline DAG from ingest through package. Evidence: stages, optional branches, retry boundaries, and artifacts are data.
    • A0.6.a Inventory current stage implementations/inputs/outputs/models/ configs/resources/caches/failures and profile variations. Evidence: stage ledger. <!-- aja-stage-ledger.json: derived from the TYPED motion-pipeline-sdk contracts (libs/aja/motion-pipeline-sdk/src/types.ts), not imposed taxonomy. 11 canonical PipelineStages (ingestion → validation → preprocessing → pose-estimation → skeleton-fitting → domain-analysis → quality-assessment → retargeting → format-conversion → postprocessing → delivery), each mapped to named-reference modules (validation 4 / pose-estimation 1 / retargeting 6 have dedicated libs; the other 8 run inline via the SDK client — reported honestly as orchestrated, NOT unimplemented). inputs = 7 InputConstraints fields, outputs = 6 OutputSpec fields, configs = StageConfig (params/timeout/deps/parallel), failures = StageRetryConfig (retry/backoff) + failed/skipped statuses + failFast, resources = 5 JobPriorities + per-stage/global timeouts + parallel flag, caches = 6 pipeline-cache modules (incl. pose-estimation-cache, retargeting-cache), profile variations = 5 domains × 4 quality presets = 20 combinations + 8 QualityFeatures toggles. models are honestly noted as stage-specific StageConfig.params, not a top-level typed field. generate-aja-stage-ledger.mjs + 7 tests, wired into inventory/verify + CI. The canonical DAG (A0.6.b), branch/skip data (A0.6.c), and later A0.6.* DEFINE new structure and stay unchecked. -->
    • A0.6.b Define required ingest/preprocess/estimate/reconstruct/cleanup/ retarget/validate/package stages and typed artifact edges. Evidence: canonical DAG.
    • A0.6.c Define modality/profile optional branches, conditions/fan-out-in, required floors, skip rules, and unsupported behavior as data. Evidence: branch matrix.
    • A0.6.d Define transaction/retry/checkpoint/cache/cancel/rerun boundaries, non-destructive lineage, and downstream invalidation. Evidence: execution contract.
    • A0.6.e Version/review/publish/deprecate templates and validate cycles/ ports/resources/capabilities/consent/export floors. Evidence: ratified DAG.
  • A0.7 Define supported input, skeleton, rig, animation, motion, video, calibration, and export formats. Evidence: compatibility matrix.
    • A0.7.a Inventory format/profile/version/container/codec/schema/vendor/ validator/library/licence and actual import/export consumers. Evidence: format registry. <!-- evidence: aja-format-registry.json — @aja/motion-formats 7 formats (fbx/gltf/usd/bvh/alembic/mocap-import/clip-database) over 10 extensions, each with import-capable + export-capable symbols, validators, version constants (e.g. FBX_VERSION_MAP), binary signatures (FBX_BINARY_MAGIC); 5 round-trip formats; 58 downstream consumers across 30 projects. licence/vendor are the semantic half. generate-aja-format-registry.mjs + 5 tests. -->
    • A0.7.b Define video/depth/image/calibration/camera/timecode inputs with resolution/rate/color/orientation/clock/metadata limits. Evidence: input matrix.
    • A0.7.c Define skeleton/rig/mesh/blend-shape/coordinate/unit/bind-pose/ naming/topology/skin/mapping support. Evidence: rig matrix.
    • A0.7.d Define motion/animation/curve/event/contact/root/face/hand/ compression/interchange/export target support. Evidence: motion matrix.
    • A0.7.e Define conversion/round-trip/loss diagnostics and supported/ conditional/unsupported format pairs with fixtures/version/owner/expiry. Evidence: compatibility matrix.
  • A0.8 Define subject consent, biometric/privacy, data residency, retention, moderation, watermark, and revocation policies. Evidence: legal/privacy review.
    • A0.8.a Define biometric/personal/sensitive data classes across raw video/ depth/audio/face/body/hand/gaze/voice/mesh/motion/embeddings/metadata. Evidence: data classification.
    • A0.8.b Define informed consent purpose/intended and excluded use/ processing/modalities/models/recipients/territories/term/withdrawal/ evidence. Evidence: consent policy.
    • A0.8.c Define collection/minimization/redaction/pseudonymization/ encryption/key/region/provider/device/access/export/analytics controls. Evidence: privacy policy.
    • A0.8.d Define retention/deletion/legal hold/backups/derived asset/model/ consumer impact and subject request/revocation response. Evidence: lifecycle policy.
    • A0.8.e Define content/safety moderation, watermark/provenance/disclosure, prohibited outputs/claims, and human review by vertical profile. Evidence: safety policy.
    • A0.8.f Obtain current legal/privacy/consent/biometric/security review and link policies to enforcement/tests/runbooks. Evidence: signed review.
  • A0.9 Threat-model non-consensual capture, biometric leakage, malicious video/assets, model attacks, unsafe anatomical output, worker compromise, cost abuse, export tamper, and adapter spoofing. Evidence: mitigation/test map.
    • A0.9.a Map actors/assets/trust/data/control boundaries across devices/ mobile/desktop/upload/API/queue/workers/models/storage/preview/export/ consumers. Evidence: threat diagrams.
    • A0.9.b Model hidden/mismatched/revoked subject consent, covert capture, identity swaps, reidentification, secondary use, and prohibited profile claims. Evidence: consent threats.
    • A0.9.c Model biometric/raw media leakage through search/log/cache/ preview/annotation/analytics/support/providers/exports/adapters and cross-tenant access. Evidence: privacy threats.
    • A0.9.d Model malicious video/depth/rig/archive/project/parser payloads, model poisoning/evasion/prompt attacks, unsafe generated motion, and export tamper. Evidence: content/model threats.
    • A0.9.e Model worker/GPU/lease/cache/checkpoint/result spoofing, arbitrary code/model/version, resource/cost abuse, queue starvation, and artifact substitution. Evidence: compute threats.
    • A0.9.f Model adapter/deep-link/webhook/SDK impersonation, replay, capability/health spoofing, and consumer misuse. Evidence: integration threats.
    • A0.9.g Assign controls/tests/monitoring/owners/residual risks/review cadence to every threat. Evidence: mitigation map.
  • A0.10 Define real-time versus batch latency/quality profiles. Evidence: explicit SLO and gate differences.
    • A0.10.a Define live preview/feedback/record/handoff and batch processing/QA/package scenarios, users, devices, inputs, and outputs. Evidence: use profiles.
    • A0.10.b Set end-to-end/frame/preview/feedback/queue/stage/export latency, jitter/dropout/resource/cost budgets per profile. Evidence: SLOs.
    • A0.10.c Define allowed live model/resolution/modality/calibration/ smoothing/quality compromises and clearly label provisional output. Evidence: live capability matrix.
    • A0.10.d Define batch model/resolution/iterations/ground-truth/gate/ review/reproducibility floors and no live-preview promotion. Evidence: batch matrix.
    • A0.10.e Define profile-specific measurements/thresholds/failure/ fallback/reprocess/approval/export eligibility. Evidence: gate differences.
    • A0.10.f Test same inputs across profiles, compare timing/quality/cost/ lineage, and show unsupported/degraded honestly. Evidence: profile report.
  • A0.11 Define ground-truth, calibration, and evaluation-corpus governance. Evidence: provenance and access controls.
    • A0.11.a Define corpus/dataset/version/purpose/modalities/subjects/ motions/rigs/capture setups/labels/splits/owners and intended profiles. Evidence: corpus schema.
    • A0.11.b Verify consent/licence/lawful basis/territory/retention/model- evaluation use/withdrawal and protected biometric access. Evidence: rights register.
    • A0.11.c Define ground-truth acquisition/calibration systems/protocols/ synchronization/coordinate mapping/uncertainty/quality and expert qualifications. Evidence: truth protocol.
    • A0.11.d Define annotation instructions/tools/inter-rater/adjudication/ versioning/provenance and correction history. Evidence: labeling policy.
    • A0.11.e Define subject-safe train/dev/test/eval splits, leakage/ duplicate/contamination/poisoning controls and subgroup minimums. Evidence: split policy.
    • A0.11.f Enforce encryption/compartment/access/export/audit/deletion/ revocation/provider controls and approve releases. Evidence: governance review.
  • A0.12 Define operational success measures for throughput, queue age, stage failure, quality, manual correction, cache, cost, consent, and export correctness. Evidence: metric spec.
    • A0.12.a Define item/frame/duration/profile throughput and queue wait/age/ starvation/SLA by priority without hiding rejected work. Evidence: queue metrics.
    • A0.12.b Define per-stage start/runtime/retry/failure/cancel/checkpoint/ recovery/capability mismatch and artifact completeness. Evidence: stage metrics.
    • A0.12.c Define calibrated raw quality/verdict/uncertainty/manual correction/adjudication/drift measures by profile and subject-safe cohorts. Evidence: quality metrics.
    • A0.12.d Define cache hit/valid/invalid/corrupt/evict/resource savings and compute/storage/egress/model/currency cost attribution. Evidence: efficiency metrics.
    • A0.12.e Define consent verified/mismatch/expiry/revocation/enforcement and package/export/adapter semantic read-back correctness. Evidence: governance metrics.
    • A0.12.f Define units/denominators/windows/slices/privacy/cardinality/ freshness/alerts/owners/budgets and quality guardrails. Evidence: metric spec.
  • A0.13 Author and ratify V1/AJA_MOCAP_CONSOLE_DESIGN_2026-07-23.md with current-state, host, pipeline DAG, contracts, data, APIs, queue, UI, consent/privacy, export, migration, rollout, and decision records. Evidence: approved doc and resolved review comments.
    • A0.13.a Document current routes/packages/services/CLI/contracts/ adapters/stores/workers/formats/pipelines/tests/backing state and gaps. Evidence: current-state chapter.
    • A0.13.b Document host/ownership/roles/boundaries/capability and vertical profiles, canonical DAG/stages/artifacts/resources/caches. Evidence: architecture chapters.
    • A0.13.c Document capture/motion/rig/annotation/consent/provenance/package contracts and stores/APIs/events/queue/workers/adapters. Evidence: platform chapters.
    • A0.13.d Document consent/privacy/biometric/safety/moderation/threat/ evaluation/export policies and human authority. Evidence: governance chapters.
    • A0.13.e Document task-led console/preview/semantic alternatives/ responsive/accessibility/live/offline/performance/error states. Evidence: experience chapter.
    • A0.13.f Document migration/compatibility/rollout/rollback/operations/ observability/recovery/verification/evidence/launch criteria. Evidence: delivery chapters.
    • A0.13.g Resolve every substantive review comment and obtain named mocap/animation/privacy/safety/security/accessibility/engineering approval. Evidence: ratified design.
  • A0.14 Reconcile every Aja service/CLI command, static Studio route, libs/aja/* package, canonical/study adapter, contract, format, pipeline, worker, test, and documented feature at atomic-feature level. Evidence: explicit dispositions cover video-to-motion, desktop/mobile live capture, body/hand/face tracking, 2D/3D/mesh recovery, cleanup, retargeting, quality, formats, vertical pipelines, reference video, avatars/rigs, batch/queue, SDK/webhooks, privacy/consent, and delivery.
    • A0.14.a Inventory code/routes/CLI/services/packages/contracts/adapters/ formats/stages/workers/tests/docs/flags/deployments. Evidence: reconciliation inputs. <!-- evidence: aja-code-census.json — comprehensive launch census: 41 libs packages + 7 app projects, 445 code files + 43 tests, 98 Studio routes, 3 classified services, 7 CLI commands, 7 interchange formats, the @oshun/embodiment-aja contract boundary (0 violations), 52 docs, and 7 deployment targets (build/dev/lint/serve/start/test/typecheck). Consolidated from fresh git counts + committed A0.2/A0.4/A0.7 evidence. Atomic-feature split + dispositions are A0.14.b+ (semantic). generate-aja-code-census.mjs + 4 tests. -->
    • A0.14.b Split each source item into atomic user/operator/system capabilities with host/audience/owner/profile/state/dependencies/data class. Evidence: atomic ledger.
    • A0.14.c Cover live/mobile/body-hand-face/multi-person/video-to-motion/ 2D-3D-mesh/cleanup/retarget/quality/reference/avatars explicitly. Evidence: motion coverage.
    • A0.14.d Cover formats/verticals/batch-queue/SDK-CLI/webhooks/privacy- consent/package-delivery/consumer adapters explicitly. Evidence: operations coverage.
    • A0.14.e Assign keep/consolidate/migrate/replace/deprecate/remove/ externalize with target owner/host/profile/dependencies/delivery/expiry. Evidence: disposition ledger.
    • A0.14.f Reconcile ledger to routes/packages/contracts/stores/tests and fail orphan/duplicate/unknown/unreviewed capabilities. Evidence: completeness gate.
  • A0.15 Ratify separate capability and safety profiles for yoga, fitness, dance, martial arts, sports, medical/rehabilitation, film/VFX, gaming, and other current domain pipelines. Evidence: each profile has authorized audiences, intended-use limitations, capture requirements, stages, models, quality thresholds, reviewers, outputs, and prohibited claims.
    • A0.15.a Define profile identity/version/owner/audience/intended and excluded use/disclaimer/jurisdiction for each named vertical. Evidence: profile registry.
    • A0.15.b Define required capture devices/views/calibration/reference/ subject/consent/modalities/motion vocabulary and data classification. Evidence: capture matrix.
    • A0.15.c Define canonical required/optional stages/models/configuration/ resources and unsupported behavior per vertical. Evidence: DAG profiles.
    • A0.15.d Define raw measurements/quality floors/uncertainty/human reviewer qualifications/quorum/adjudication/export conditions. Evidence: quality profiles.
    • A0.15.e Define output projection/fields/format/watermark/provenance/ consumer handoff and prohibited medical/coaching/performance claims. Evidence: output policy.
    • A0.15.f Validate independent known-value/privacy/safety fixtures and obtain named domain review for every profile; no inherited proof. Evidence: ratified matrix.
  • A0.16 Define the boundary among capture-device UI, processing operator console, animator/rigger cleanup, clinician/coach/educator consumer applications, developer SDK/CLI, and platform administration. Evidence: no console preview is misrepresented as medical, coaching, or performance advice.
    • A0.16.a Define audience/trust/data/action/responsibility boundaries for each named surface and authoritative owner of every mutation. Evidence: surface matrix.
    • A0.16.b Define capture-device responsibilities for consent/status/ guidance/record/buffer/upload only, with no unqualified interpretation. Evidence: device boundary.
    • A0.16.c Define console versus animator/rigger cleanup responsibilities, exact handoff artifacts/revisions/locks, and quality/approval boundary. Evidence: production boundary.
    • A0.16.d Define consumer projections with intended-use labels, qualified review, redacted biometric data, and no console preview as advice. Evidence: consumer contracts.
    • A0.16.e Define SDK/CLI and platform admin credentials/scopes/audit/ capabilities without content/consent authority expansion. Evidence: developer/operator boundary.
    • A0.16.f Test signed/deep-link/API/adapter handoffs, reauthorization/ redaction/expiry/return, and cross-surface mutation attempts. Evidence: boundary suite.
  • A1.1 Define capture project, session, take, camera, calibration, subject, consent ref, input asset, and capture metadata. Evidence: schema.
    • A1.1.a Define stable tenant/project/session/take identity, lifecycle/ revision/times/operator/profile/purpose/location and authorization scope. Evidence: capture schemas.
    • A1.1.b Define camera/sensor/device/lens/stream/channel/frame-rate/ resolution/orientation/clock/timecode/capability metadata. Evidence: device schema.
    • A1.1.c Define calibration/version/coordinate/units/intrinsic/extrinsic/ distortion/volume/scale/error/reference relationships. Evidence: calibration schema.
    • A1.1.d Define pseudonymous subject/role/take assignment/multi-person ID separately from protected identity/biometrics and exact consent reference. Evidence: subject boundary.
    • A1.1.e Define immutable input assets/checksums/kinds/channels/time ranges/ capture/import provenance/privacy class/retention/rights. Evidence: asset schema.
    • A1.1.f Enforce cross-project/reference/state/version/missing/unknown/ archive invariants with representative fixtures. Evidence: schema tests.
  • A1.2 Define pipeline template as a versioned DAG with stage IDs, dependencies, input/output kinds, configuration schema, capability, resources, timeout, retry, and cache policy. Evidence: DAG validation.
    • A1.2.a Define immutable template/version/owner/profile/status/parent/ effective window and stage/edge order-independent serialization. Evidence: template schema.
    • A1.2.b Define stable stage ID/type/adapter version/typed ports/ dependencies/conditions/optional-required floor and fan-out/in semantics. Evidence: DAG contract.
    • A1.2.c Define versioned configuration schema/defaults/validation/secrets/ migrations and output-affecting field declaration. Evidence: config contract.
    • A1.2.d Define host/model/GPU/CPU-memory/storage/network/licence/consent/ modality capability requirements and reservations. Evidence: capability schema.
    • A1.2.e Define timeout/heartbeat/retry/backoff/checkpoint/cancel/skip/ failure/cache key/read-write-expiry/quarantine policy. Evidence: execution policy.
    • A1.2.f Validate acyclic graph/ports/kinds/reachability/required stages/ resource feasibility/export floors. Evidence: DAG validation suite.
  • A1.3 Define job, stage run, attempt, lease, checkpoint, measurement, log ref, resource use, cost, and output refs. Evidence: durable-job schema.
    • A1.3.a Define immutable job request/template/profile/input/config/ consent/policy/priority/budget/owner/status/version/idempotency. Evidence: job schema.
    • A1.3.b Define per-stage run/dependency/status/current attempt/cache/ worker/capability/input-output and downstream invalidation. Evidence: stage schema.
    • A1.3.c Define attempt/lease/worker/heartbeat/start-end/exit/failure/ retry/cancel/preemption and fencing token. Evidence: lease schema.
    • A1.3.d Define checkpoint/version/progress/cursor/shard/state/artifact/ checksum/resume compatibility and expiry. Evidence: checkpoint schema.
    • A1.3.e Define measurement/log ref/resource usage/cost units/provider/ time/confidence/protected redaction and raw-versus-verdict relationship. Evidence: telemetry schema.
    • A1.3.f Define output artifact/ref/kind/version/hash/lineage/privacy/ quality/consent/release state and durable-before-terminal invariants. Evidence: job fixtures.
  • A1.4 Define skeleton, joint, hierarchy, coordinate system, units, bind pose, rig, mapping, and retarget profile. Evidence: contract fixtures.
    • A1.4.a Define skeleton/version/joint stable IDs/names/aliases/parent/ order/root/end effectors/semantics and modality. Evidence: skeleton schema.
    • A1.4.b Define coordinate handedness/up-forward axes/origin/rotation order/quaternion convention/linear-angular units/scale and transforms. Evidence: coordinate contract.
    • A1.4.c Define bind/rest/reference pose/local-global transforms/bone lengths/orientations/joint limits and confidence/provenance. Evidence: pose fixtures.
    • A1.4.d Define rig controls/deform skeleton/mesh/skin/blend shapes/ constraints/spaces/IK chains/version/host/format. Evidence: rig schema.
    • A1.4.e Define source-target mapping/strategy/offset/scale/axis/retarget chains/root/contact/twist/finger/face settings and version. Evidence: mapping schema.
    • A1.4.f Validate cycles/duplicates/missing parents/axes/units/bind/ correspondence and round-trip fixtures. Evidence: contract suite.
  • A1.5 Define motion clip, frame range, sample rate, root motion, contacts, events, curves, blend shapes, and source lineage. Evidence: round-trip.
    • A1.5.a Define motion clip/version/skeleton/rig/timebase/start/end/frame count/sample rate/variable timestamps/interpolation and coordinate context. Evidence: clip schema.
    • A1.5.b Define per-frame joint/body/hand/face transforms/confidence/ occlusion/tracking state and sparse/dense storage. Evidence: motion samples.
    • A1.5.c Define root/global trajectory/in-place extraction/ground plane/ scale and local versus world motion. Evidence: root schema.
    • A1.5.d Define contacts/events/markers/segments/labels and scalar/vector/ quaternion/blend-shape/custom curves with exact anchors. Evidence: event/curve schema.
    • A1.5.e Define immutable source takes/model/stages/cleanup/generated ranges/mappings/retarget/render/export lineage. Evidence: provenance chain.
    • A1.5.f Round-trip supported formats/sample rates/coordinates/contacts/ curves/face/hand data with loss diagnostics. Evidence: motion goldens.
  • A1.6 Define annotation layer, segment, frame/time anchor, label, reviewer, status, and provenance. Evidence: schema.
    • A1.6.a Define annotation set/layer/version/purpose/schema/taxonomy/ subject motion revision/visibility/owner and immutable parentage. Evidence: annotation contract.
    • A1.6.b Define frame/time/musical/segment/joint/body/hand/face/object/ region anchors with context and mapping/orphan semantics. Evidence: anchor schema.
    • A1.6.c Define label/value/type/units/confidence/uncertainty/source/ intended profile and free-text restrictions. Evidence: label schema.
    • A1.6.d Define draft/submitted/reviewed/accepted/rejected/adjudicated/ superseded state, reviewer qualification, disagreement, and decisions. Evidence: state machine.
    • A1.6.e Record creator/reviewer/tool/model/version/evidence/time/edit/ import/export provenance and protected subject projection. Evidence: provenance tests.
    • A1.6.f Validate overlaps/required labels/taxonomy version/anchor survival/ cross-project/access and representative fixtures. Evidence: schema tests.
  • A1.7 Define technical measurement separately from validation verdict. Evidence: policy can recompute without reprocessing.
    • A1.7.a Define immutable measurement with subject/input/artifact/frame- joint-range, values/units/evidence/confidence/evaluator/model/calibration versions/time. Evidence: measurement schema.
    • A1.7.b Define separate verdict with measurement references/profile/ rules/thresholds/outcome/reason/remediation/exception/time. Evidence: verdict schema.
    • A1.7.c Prevent verdict mutation/denormalization of raw values and prevent missing/error/unsupported data from defaulting to pass. Evidence: invariants.
    • A1.7.d Recompute new verdicts from retained valid measurements after policy/profile threshold changes without rerunning stages. Evidence: recomputation tests.
    • A1.7.e Rerun measurement only on subject/artifact/evaluator/calibration invalidation and retain full historical lineage. Evidence: lifecycle tests.
  • A1.8 Define consent with subject, collector, purpose, allowed uses, territories, recipients, biometric processing, retention, expiry, revocation, and evidence. Evidence: invariants.
    • A1.8.a Define protected subject/representative/minor capacity and collector identity/authority/relationship/collection context. Evidence: identity schema.
    • A1.8.b Define purpose/intended-excluded use/profile/project/session/ modalities/body-hand-face-gaze-mesh/model processing and claims. Evidence: purpose schema.
    • A1.8.c Define allowed outputs/uses/recipients/providers/consumers/ territories/regions/transfers/exports/publication/model training. Evidence: scope schema.
    • A1.8.d Define retention/deletion/withdrawal/revocation/effective-expiry/ legal hold/derived artifact/model impact and contact path. Evidence: lifecycle schema.
    • A1.8.e Define informed notice/version/locale/accessibility, affirmative signature/intent/time/document hash/witness/guardian and evidence. Evidence: consent proof.
    • A1.8.f Enforce subject/session/purpose/modality/recipient/territory/time match, no scope inheritance, and conflict/unknown invariants. Evidence: consent tests.
  • A1.9 Define privacy transformations and redacted derivatives. Evidence: derivative lineage points to protected source without exposing it.
    • A1.9.a Define transform kind/version/config/purpose/authorized input class/output class/irreversibility/risk and approval. Evidence: privacy transform registry.
    • A1.9.b Define face/body/background/object/license-plate/screen/audio/ metadata redaction, crop/blur/mask/avatar/skeleton-only derivatives and quality tradeoffs. Evidence: transform fixtures.
    • A1.9.c Generate immutable derivative with input/output hashes, exact ranges/regions/tool/model/version/operator and protected opaque link. Evidence: lineage tests.
    • A1.9.d Define authorized role/profile/consumer/search/preview/export/ retention for derivative and prohibit reverse lookup or source leakage. Evidence: projection tests.
    • A1.9.e Evaluate residual reidentification/utility/quality/temporal consistency and require human privacy review at thresholds. Evidence: privacy evaluation.
    • A1.9.f Propagate source revocation/deletion/hold/correction through derivatives/providers/caches/consumers with audit. Evidence: lifecycle tests.
  • A1.10 Define watermark and provenance records for every derived motion asset. Evidence: transformation chain fixture.
    • A1.10.a Define provenance ingredient/action/tool/model/stage/config/ subject consent/rights/output identity/hash/version/time DAG. Evidence: provenance schema.
    • A1.10.b Define watermark kind/payload/key/algorithm/strength/placement/ target/detection/confidence/removal policy without biometric secrets. Evidence: watermark schema.
    • A1.10.c Record ingest/estimate/reconstruct/mesh/cleanup/generate/retarget/ annotate/validate/convert/package transformations and edited ranges. Evidence: action lineage.
    • A1.10.d Bind watermark/provenance to exact motion/skeleton/rig/mesh/ package formats and preserve/reapply or diagnose loss on conversion. Evidence: format tests.
    • A1.10.e Validate hash/signature/watermark/tamper/stripped/unknown/ untrusted state and permissioned semantic inspector. Evidence: integrity tests.
    • A1.10.f Run complete protected source→redacted input→motion→cleanup→ retarget→package transformation chain fixture. Evidence: provenance golden.
  • A1.11 Define export package, target format, target rig, mappings, measurements, gates, consent lock, watermark, and manifest. Evidence: schema.
    • A1.11.a Define immutable package/candidate/version/owner/project/profile/ target/territory/consumer/status and exact source motion revision. Evidence: package schema.
    • A1.11.b Define target format/profile/version/rig/skeleton/coordinate/ unit/rate/mapping/retarget/config and compatibility/loss. Evidence: target contract.
    • A1.11.c Pin output motion/mesh/curves/events/annotations/previews/source assets plus hashes and transformation provenance. Evidence: artifact manifest.
    • A1.11.d Pin raw measurements/verdicts/profile/evaluator/calibration/ reviewer approvals and unresolved/waived findings. Evidence: quality manifest.
    • A1.11.e Pin consent/subject/use/modality/recipient/territory/expiry/ privacy derivative/watermark/provenance lock and recheck at dispatch. Evidence: consent lock.
    • A1.11.f Define canonical serialized manifest/hash/signature/read-back/ supersession/recall and invalidation on dependency change. Evidence: package tests.
  • A1.12 Define immutable revisions and diffs for config, mapping, annotation, cleanup, validation, and package. Evidence: golden diffs.
    • A1.12.a Define common immutable revision/parent(s)/subject/schema/tool/ actor/cause/time/hash/approval lineage and branch/merge semantics. Evidence: revision contract.
    • A1.12.b Define config/template/stage parameter diff with defaults/ secrets redaction/output-affecting/material classification. Evidence: config goldens.
    • A1.12.c Define skeleton/rig mapping diff by stable source-target joints/ offsets/strategies/chains and compatibility impact. Evidence: mapping goldens.
    • A1.12.d Define annotation diff by layer/anchor/label/status/reviewer and cleanup motion diff by exact frames/joints/curves/generated ranges. Evidence: edit goldens.
    • A1.12.e Define validation diff by raw measurements/verdict/profile/ waiver and package diff by artifacts/rig/format/consent/watermark/ manifest. Evidence: release goldens.
    • A1.12.f Test add/delete/move/split/merge/restore/concurrent/schema/ opaque loss and protected projections. Evidence: semantic diff suite.
  • A1.13 Define consumer adapter health/capability projections from actual service state. Evidence: healthy/degraded/unhealthy fixtures.
    • A1.13.a Define versioned adapter capability IDs/operations/modalities/ profiles/formats/limits/prerequisites and unsupported reasons. Evidence: capability schema.
    • A1.13.b Define actual dependency health from API/queue/workers/models/ GPU/storage/cache/consent/gates/export versions and freshness. Evidence: health inputs.
    • A1.13.c Compute per-capability available/degraded/unavailable/ unsupported/not-configured/unknown, reason, safe alternatives, and expected impact. Evidence: projection logic.
    • A1.13.d Redact project/subject/biometric/worker/security/cost details and prevent global healthy status from masking failed capability. Evidence: privacy tests.
    • A1.13.e Test healthy/partial/degraded/stale/outage/version-skew/ permission/consent/config fixtures in Metis and other consumers. Evidence: health contract suite.
  • A1.14 Define contract compatibility and format-version negotiation. Evidence: old fixture migration and unsupported refusal.
    • A1.14.a Inventory deployed capture/motion/rig/annotation/consent/job/ package/adapter/format versions and historical edge fixtures. Evidence: version corpus.
    • A1.14.b Define support window and reader/writer/migrate/export/rollback behavior per contract and format. Evidence: compatibility matrix.
    • A1.14.c Define handshake/Accept or equivalent version/capability/profile/ format negotiation with deterministic compatible selection. Evidence: negotiation tests.
    • A1.14.d Implement migrations/adapters for renamed/split/merged/ coordinate/unit/skeleton/consent fields with explicit loss. Evidence: migration tests.
    • A1.14.e Refuse unsupported/unsafe/ambiguous combinations with required upgrade/convert/export alternatives; never default consent/quality. Evidence: refusal tests.
    • A1.14.f Round-trip old/current fixtures across console/service/worker/ SDK/adapter/format and document irreversible steps. Evidence: compatibility report.
  • A1.15 Define body, joint, hand/finger, face/landmark, gaze, expression, blend-shape, confidence, occlusion, contact, multi-person identity, and tracking-loss contracts separately. Evidence: modality-specific privacy, consent, quality, and absence/unsupported semantics.
    • A1.15.a Define body/root/joint hierarchy/pose/global-local transforms/ visibility/confidence/occlusion/tracking state and source modality. Evidence: body schema.
    • A1.15.b Define hand/finger side/joints/curls/spread/contact/gesture/ confidence/occlusion and per-hand supported/absent state. Evidence: hand schema.
    • A1.15.c Define face/landmark/topology/expression/blend shape/ coefficients/identity-independent versus identity-bearing fields. Evidence: face schema.
    • A1.15.d Define gaze origin/direction/target/confidence/calibration and avoid inferring attention/emotion/health beyond intended use. Evidence: gaze schema.
    • A1.15.e Define contacts with body parts/surface/point/normal/start-end/ confidence and multi-person stable track identity/associations. Evidence: contact/identity schemas.
    • A1.15.f Define tracking loss/partial/occluded/not-observed/ unsupported/not-requested/consent-blocked distinctly from zero/neutral. Evidence: absence semantics.
    • A1.15.g Map consent/privacy/retention/redaction/quality/evaluator/profile independently per modality and test round-trips. Evidence: modality matrix.
  • A1.16 Define live-capture device, stream, sensor, clock, frame, network, calibration, preview, recording, buffer, dropout, reconnect, edge/cloud processing, and session-handoff contracts. Evidence: desktop/mobile and multi-device compatibility fixtures.
    • A1.16.a Define device/app/OS/model/sensor/camera/depth/IMU/audio/ capability/permission/thermal/battery/storage identity and state. Evidence: device schema.
    • A1.16.b Define stream/modality/codec/resolution/rate/orientation/ timestamps/clock/timecode/synchronization/calibration/frame metadata. Evidence: stream schema.
    • A1.16.c Define live frame/payload/order/drop/duplicate/late/confidence/ tracking state and preview redaction/latency/quality warning. Evidence: frame contract.
    • A1.16.d Define immutable local recording/chunks/checksums/buffer/quota/ encryption/upload cursor/dropout markers/recovery. Evidence: recording schema.
    • A1.16.e Define edge versus cloud stage selection, model/version/ hardware/privacy/rights/latency/fallback and no silent data transfer. Evidence: processing profile.
    • A1.16.f Define disconnect/reconnect/resume/deduplication/clock reset/ calibration invalidation and signed session/operator handoff. Evidence: handoff contract.
    • A1.16.g Validate desktop/mobile/multi-device/device-loss/network/ interruption/version/clock fixtures. Evidence: compatibility suite.
  • A1.17 Define vertical-pipeline profile, intended use, excluded use, motion vocabulary, reference protocol, required annotations, evaluator, quality floor, human qualification, output projection, and disclaimer. Evidence: profile schema prevents a film/gaming result from masquerading as a clinical or coaching result.
    • A1.17.a Define stable profile/version/owner/jurisdiction/audience/ intended use/excluded use/prohibited claims/disclaimer and review cadence. Evidence: profile schema.
    • A1.17.b Define motion vocabulary/taxonomy/actions/phases/body regions/ expected variability/cultural or domain sources and aliases. Evidence: vocabulary contract.
    • A1.17.c Define capture/reference protocol/devices/views/calibration/ subjects/consent/modalities/environment/instructions. Evidence: protocol schema.
    • A1.17.d Define required DAG stages/models/configs/resources/annotations/ raw measurements/evaluator versions/quality floors/uncertainty. Evidence: processing contract.
    • A1.17.e Define human reviewer qualifications/independence/adjudication/ approval/waiver and consumer output fields/redaction/format/watermark. Evidence: governance contract.
    • A1.17.f Bind every output/package/adapter projection to profile ID/ intended-use warning/disclaimer and reject cross-profile substitution. Evidence: boundary tests.
    • A1.17.g Validate independent known-value/privacy/consent/safety/ handoff fixtures for every ratified vertical. Evidence: profile suite.
  • A2.1 Build project/session/take capture and bulk ingest. Evidence: keyboard and resumable-upload journey.
    • A2.1.a Create/select authorized project/profile/purpose and capture session with devices/subjects/consent/calibration/time/location/owner. Evidence: setup tests.
    • A2.1.b Create takes with slate/action/performers/cameras/range/markers/ notes and immutable local/remote input placeholders. Evidence: take tests.
    • A2.1.c Build multi-file/folder/archive manifest preview with detected kind/size/hash/associations/duplicates/consent/expected mappings. Evidence: bulk preview.
    • A2.1.d Upload resumably with progress/pause/resume/retry/cancel/quota/ checksum and partial per-item state. Evidence: upload tests.
    • A2.1.e Finalize only after quarantine/validation/metadata assignment and durable blob/reference, preserving rejected/orphan cleanup. Evidence: ingest tests.
    • A2.1.f Run pointer-free/screen-reader single and mixed bulk ingest, interruption/reload/duplicate/permission/recovery. Evidence: E2E.
  • A2.2 Quarantine uploaded video, depth, calibration, skeleton, rig, and project archives. Evidence: malicious corpus.
    • A2.2.a Store every untrusted item in isolated non-executable quarantine before preview/parse/index/model/worker use. Evidence: boundary tests.
    • A2.2.b Verify signature/type/codec/schema/size/count/nesting and scan malware/polyglot/active content/unsafe metadata. Evidence: file corpus.
    • A2.2.c Enforce archive path/symlink/device-file/nesting/member/ratio/ decompressed-size limits and scan each member. Evidence: archive corpus.
    • A2.2.d Parse video/depth/calibration/skeleton/rig/project using sandboxed bounded parsers without loading scripts/plugins/external resources. Evidence: parser tests.
    • A2.2.e Record scanner/parser/signature/policy versions and release/ reject/manual-review decision, rescan on update. Evidence: audit.
    • A2.2.f Test crash/timeout/oversize/encrypted/corrupt/unknown/false- positive override and prove no unscanned escape. Evidence: malicious suite.
  • A2.3 Validate media container, codec, resolution, frame rate, timestamps, duration, orientation, color, audio, and corruption. Evidence: fixture matrix.
    • A2.3.a Probe actual container/streams/codecs/profiles/levels/bit-depth/ dimensions/pixel formats and compare extension/MIME/declared metadata. Evidence: format tests.
    • A2.3.b Validate resolution/aspect/crop/rotation/orientation/interlace/ anamorphic/pixel geometry and profile requirements. Evidence: image tests.
    • A2.3.c Validate nominal/average/variable frame rate, time base, PTS/DTS/ monotonicity/gaps/duplicates/duration/start offset and audio sync. Evidence: timing tests.
    • A2.3.d Validate color primaries/transfer/matrix/range/HDR metadata and depth/disparity/calibration channel relationships. Evidence: color/depth tests.
    • A2.3.e Validate audio channels/rate/codec/duration only where captured and never infer consent from presence. Evidence: audio tests.
    • A2.3.f Decode representative/full sampled frames, detect corruption/ truncation/errors, and emit pass/fail/unsupported/partial with evidence. Evidence: fixture matrix.
  • A2.4 Validate multi-view synchronization and camera identity. Evidence: aligned, drifted, dropped-frame, and swapped-camera fixtures.
    • A2.4.a Resolve expected camera/sensor IDs, serial/position/stream/ calibration assignment and protect identifying metadata. Evidence: identity tests.
    • A2.4.b Compare trusted timecode/hardware clock/timestamps/audio/visual sync events with offsets/jitter/drift/confidence and clock resets. Evidence: sync measurements.
    • A2.4.c Detect missing/duplicate/dropped/late/out-of-order frames and variable rate per camera with aligned frame availability. Evidence: frame tests.
    • A2.4.d Detect swapped/mislabeled/duplicate cameras using calibration/ visual/fingerprint evidence and require human correction. Evidence: swap tests.
    • A2.4.e Persist raw offset/drift/identity evidence separately from profile tolerance verdict and remediation/resync plan. Evidence: result tests.
    • A2.4.f Run aligned/drifted/dropped/swapped/clock-reset/partial-view fixtures. Evidence: synchronization suite.
  • A2.5 Validate calibration completeness, reprojection error, lens, scale, coordinate frame, and capture volume. Evidence: known values.
    • A2.5.a Verify required cameras/sensors/intrinsics/extrinsics/distortion/ image size/reference observations/board or volume metadata and versions. Evidence: completeness tests.
    • A2.5.b Recompute per-camera/global reprojection residuals/distributions/ outliers/coverage/confidence from pinned observations. Evidence: reprojection tests.
    • A2.5.c Validate lens model/parameters/ranges/focus-zoom state/ rolling-shutter where applicable and image-stream match. Evidence: lens tests.
    • A2.5.d Validate scale/reference distance/units/handedness/up-forward/ origin/world-camera transforms and round-trip closure. Evidence: coordinate tests.
    • A2.5.e Validate capture volume bounds/overlap/triangulation geometry/ blind zones/ground plane and expected subject coverage. Evidence: volume tests.
    • A2.5.f Persist raw values separately from profile verdict and run known calibrated/error/missing/stale fixtures. Evidence: known-value suite.
  • A2.6 Capture subject consent before biometric processing. Evidence: missing or mismatched consent blocks stage dispatch.
    • A2.6.a Identify each subject/representative and exact project/session/ take/profile/modalities/purpose/collector before processing. Evidence: consent request.
    • A2.6.b Present current accessible localized notice with biometric/ model/recipient/territory/retention/revocation/excluded-use details. Evidence: notice tests.
    • A2.6.c Record affirmative verified identity/authority/intent/signature/ time/document hash/evidence and allow decline/partial modality. Evidence: consent capture.
    • A2.6.d Bind consent reference/version to subject tracks and every raw input with protected ordinary projection. Evidence: binding tests.
    • A2.6.e Enforce current exact match at queue dispatch and worker start/ commit; block missing/mismatched/expired/revoked state. Evidence: dispatch race tests.
  • A2.7 Verify consent scope against project purpose and intended exports. Evidence: purpose/territory/recipient mismatch tests.
    • A2.7.a Resolve exact project/profile intended use/modalities/stages/ models/providers and proposed output/package/export/consumer plan. Evidence: scope inputs.
    • A2.7.b Compare subject consent purpose/uses/excluded uses/modalities/ biometric processing/model training/derivatives. Evidence: purpose tests.
    • A2.7.c Compare territory/region/recipient/provider/consumer/host/ publication/transfer and data residency. Evidence: recipient tests.
    • A2.7.d Compare term/expiry/retention/revocation/status and allowed watermark/redaction/output format/intended claims. Evidence: temporal/ output tests.
    • A2.7.e Return eligible/blocked/partial/manual-review with exact mismatch, safe alternatives, and no automatic scope expansion. Evidence: result contract.
    • A2.7.f Recheck on project/pipeline/export/recipient/territory/consent change and test mismatch matrix. Evidence: lifecycle tests.
  • A2.8 Run input content moderation and privacy classification. Evidence: versioned results and manual-review state.
    • A2.8.a Define profile/jurisdiction/modality-specific categories, prohibited/restricted/review thresholds, human qualifications, and privacy classes. Evidence: policy registry.
    • A2.8.b Analyze pinned redacted or protected video/depth/audio/metadata/ rig/project inputs for unsafe content, minors/vulnerability/nudity/ violence/identifiers/background/sensitive setting. Evidence: analyzer tests.
    • A2.8.c Persist raw findings/ranges/regions/confidence/model/policy/ corpus versions separately from moderation/privacy verdict. Evidence: result schema.
    • A2.8.d Route uncertain/high-risk/disagreement to authorized protected manual review with minimization and no ordinary thumbnail/log leakage. Evidence: review tests.
    • A2.8.e Apply accept/restrict/redact/quarantine/reject/escalate decision and allowed downstream stage/role/export projections. Evidence: enforcement tests.
    • A2.8.f Recompute/invalidate on input/model/policy/profile change and evaluate calibrated subgroup/privacy errors. Evidence: versioned report.
  • A2.9 Build face/background/identifier redaction options before wider access. Evidence: role projections use the derivative.
    • A2.9.a Define redaction request pinned to source/ranges/subjects/profile/ purpose and face/background/identifier types with authorized reviewer. Evidence: request schema.
    • A2.9.b Detect/track faces, bodies/people, background regions, screens/ plates/documents/logos/location/device identifiers with confidence. Evidence: detection tests.
    • A2.9.c Allow protected human adjust/add/remove tracks/masks/regions and choose blur/pixelate/mask/crop/replace/avatar/metadata strip settings. Evidence: editor tests.
    • A2.9.d Render immutable temporally consistent derivative with exact source link, tool/model/settings/ranges/hashes and residual-risk measurements. Evidence: render tests.
    • A2.9.e Map ordinary roles/search/preview/annotation/model/export to derivative only and block raw fallback when derivative fails. Evidence: projection tests.
    • A2.9.f Test occlusion/reentry/motion blur/profile/edge frames/audio/ metadata/cache/download and role revoke. Evidence: redaction suite.
  • A2.10 Build capture quality preview with exposure, blur, occlusion, framing, dropped frames, and coverage. Evidence: known clips.
    • A2.10.a Generate permission-filtered synchronized low-risk previews/ thumbnails/overlays from exact take/calibration with source time. Evidence: preview contract.
    • A2.10.b Measure exposure/clipping/dynamic range/color/blur/shutter/ focus/noise per camera/time with confidence. Evidence: image tests.
    • A2.10.c Measure subject/body-part visibility/occlusion/truncation/ framing/scale/angle/capture-volume and modality coverage. Evidence: coverage tests.
    • A2.10.d Show dropped/duplicate/late frames, sync/drift/calibration/ camera identity issues and affected ranges/views. Evidence: temporal tests.
    • A2.10.e Present raw values/units/profile thresholds/unknowns/ recommendations separately from intake decision, with semantic tables. Evidence: UI tests.
    • A2.10.f Validate known good/bad/borderline/multiview/multi-person/ protected clips and no unredacted preview leak. Evidence: known clip suite.
  • A2.11 Build take grouping, slate, performer, action, marker, and notes. Evidence: metadata schema and search.
    • A2.11.a Define take group/set/version/order/relationship for repeated actions, multi-camera parts, calibration, sessions, and variants. Evidence: grouping schema.
    • A2.11.b Edit slate/scene-shot-take/action/motion class/profile/date/ location/operator/device labels with governed vocabularies. Evidence: metadata tests.
    • A2.11.c Assign pseudonymous performer/subject tracks and consent status without exposing protected identity. Evidence: performer tests.
    • A2.11.d Add time/frame markers/ranges/cues/events/notes with anchors, visibility, authorship, revision, and protected fields. Evidence: annotation tests.
    • A2.11.e Support bulk edit/reorder/merge/split/duplicate handling with impact/concurrency/undo/audit and immutable input lineage. Evidence: operation tests.
    • A2.11.f Index safe metadata for permission-aware search/facets and test protected title/count/snippet leakage. Evidence: search tests.
  • A2.12 Build duplicate and near-duplicate detection. Evidence: exact hash and perceptual similarity remain distinct.
    • A2.12.a Compute/verify cryptographic content hashes and classify exact byte duplicates independently of metadata/path/name. Evidence: exact dedupe tests.
    • A2.12.b Compute versioned video/depth/motion/audio perceptual fingerprints/features on authorized derivatives with range/alignment. Evidence: near-dedupe tests.
    • A2.12.c Generate candidates with exact versus perceptual reason, aligned ranges, transformations, score/confidence, corpus/index version. Evidence: candidate contract.
    • A2.12.d Present protected-aware compare and let human keep/link/replace/ merge metadata/mark-not-duplicate; never delete automatically. Evidence: review tests.
    • A2.12.e Preserve distinct captures/consent/provenance/calibration even when bytes/content match and enforce cross-project isolation. Evidence: integrity tests.
    • A2.12.f Evaluate precision/recall/transforms/partial overlaps/multi-view/ identity leakage and stale index. Evidence: dedupe report.
  • A2.13 Build resumable ingest with checksum and orphan cleanup. Evidence: interruption/retry test.
    • A2.13.a Define upload session/file/chunk/offset/size/hash/tenant/project/ quota/expiry/idempotency and intended asset metadata. Evidence: upload contract.
    • A2.13.b Upload chunks with range/order/duplicate/retry validation, per-chunk hashes, pause/resume/cancel, and bounded concurrency. Evidence: transfer tests.
    • A2.13.c Finalize only after complete size/hash/type/quarantine scan and durable blob/reference transaction; handle timeout after commit. Evidence: completion tests.
    • A2.13.d Resume after network/browser/app/worker restart with current authorization and no duplicate bytes/assets. Evidence: recovery tests.
    • A2.13.e Track abandoned/expired/cancelled/unreferenced chunks/blobs and clean after safe retention, active lease, legal hold, and reconciliation. Evidence: cleanup tests.
    • A2.13.f Inject interruption/out-of-order/corruption/quota/role revoke/ process crash/DB-blob mismatch and verify checksums/orphan cleanup. Evidence: ingest fault suite.
  • A2.14 Build intake decision records for accept, reject, recapture, quarantine, or redact. Evidence: rationale and affected files.
    • A2.14.a Define immutable decision identity/type/actor-role/time/profile/ project/session/take/affected asset revisions and input findings. Evidence: decision schema.
    • A2.14.b Require rationale/evidence/quality/consent/privacy/moderation/ technical considerations, uncertainty, and conditions. Evidence: validation tests.
    • A2.14.c For accept, create eligible intake revision without asserting downstream quality; for reject, retain evidence/retention disposition. Evidence: disposition tests.
    • A2.14.d For recapture, record exact deficiencies/protocol/settings/ deadline/assignment and link replacement take. Evidence: recapture tests.
    • A2.14.e For quarantine/redact, restrict access/stages, create protected review or derivative workflow, and preserve source. Evidence: enforcement tests.
    • A2.14.f Enforce role/consent/concurrency and immutable reversals/new decisions with audit/notifications. Evidence: authority tests.
  • A2.15 Build desktop live capture with camera/sensor discovery, permission, preview, device settings, calibration, clock/sync, subject/ consent match, recording, pause, reconnect, local buffer, upload, and operator handoff. Evidence: device and network fault journeys.
    • A2.15.a Discover/identify/capability-probe/hot-plug cameras/depth/IMU/ audio/sync devices with stable aliases and privacy-safe metadata. Evidence: discovery tests.
    • A2.15.b Request/revoke OS permissions and configure device/source/ resolution/rate/exposure/focus/orientation/depth/audio/clock with unsupported state. Evidence: device tests.
    • A2.15.c Show low-latency redacted synchronized preview, framing/ exposure/blur/occlusion/coverage/drop/sync/thermal/storage/network quality warnings. Evidence: preview tests.
    • A2.15.d Capture/validate calibration and exact subject-track/consent/ profile/purpose/modalities match before arm/processing. Evidence: preflight tests.
    • A2.15.e Implement record/pause/resume/stop/slate/marker/countdown/ emergency stop with monotonic clocks and immutable crash-safe local chunks. Evidence: recording tests.
    • A2.15.f Handle device/clock/network/process loss with visible state, local buffer/quota, dropout markers, reconnect/recalibration, and no false continuity. Evidence: fault tests.
    • A2.15.g Upload resumably with checksums/deduplication/encryption/ consent scope and reconcile authoritative session/take/assets. Evidence: upload tests.
    • A2.15.h Hand off signed exact session/take/device/calibration/subject/ consent/quality/upload context to authorized operator and run complete fault journeys. Evidence: desktop live E2E.
  • A2.16 Build mobile live capture with supported camera/depth/IMU inputs, orientation, thermal/battery/storage/network state, consent, guidance, quality feedback, encrypted local buffering, background/interruption recovery, and session handoff. Evidence: applicable mobile automation and lost-device/privacy tests.
    • A2.16.a Define supported device/OS/app/camera/depth/IMU/audio/ orientation/profile combinations and capability negotiation. Evidence: mobile matrix.
    • A2.16.b Request least-privilege permissions and configure capture rate/ resolution/orientation/lens/depth/IMU sync while handling rotation/ sensor changes. Evidence: setup tests.
    • A2.16.c Present accessible localized consent/profile/purpose/modalities/ privacy/retention notice and bind each subject before capture. Evidence: consent tests.
    • A2.16.d Provide framing/distance/lighting/occlusion/coverage/motion/ calibration guidance and real-time warnings without medical/coaching claims. Evidence: guidance tests.
    • A2.16.e Monitor thermal/battery/storage/quota/network/background/ interruption and adapt or stop with explicit quality impact. Evidence: resource tests.
    • A2.16.f Record encrypted crash-safe local chunks/metadata/checksums with pause/resume/background-call-lock-screen recovery and dropout markers. Evidence: recording tests.
    • A2.16.g Upload resumably/reconcile/deduplicate, delete by policy after durable receipt, and hand off signed session/consent/quality context. Evidence: handoff tests.
    • A2.16.h Run mobile automation for orientation/background/interruption/ low battery/thermal/storage/network/lost-stolen device/remote revoke/ privacy. Evidence: mobile E2E.
  • A2.17 Build multi-person take identity and tracking assignment with consent per subject, occlusion/crossing review, split/merge correction, and protected biometrics. Evidence: identity swaps cannot silently enter downstream motion.
    • A2.17.a Define pseudonymous subject/track/take identity, protected biometric/reference features, consent link, and noninferable ordinary labels. Evidence: identity schema.
    • A2.17.b Initialize/associate subjects across cameras/modalities/time using authorized evidence/confidence and require consent before biometric matching. Evidence: assignment tests.
    • A2.17.c Detect occlusion/entry-exit/crossing/identity-switch/fragment/ merge ambiguity with affected frames/views/confidence. Evidence: tracking tests.
    • A2.17.d Provide synchronized protected review with candidate tracks/ source views and human confirm/reassign/split/merge/unknown controls. Evidence: review tests.
    • A2.17.e Apply corrections as versioned mapping operations, propagate to body/hand/face/mesh/annotations/consent, and invalidate downstream stages/gates. Evidence: correction tests.
    • A2.17.f Enforce per-subject modality/purpose/recipient consent throughout multi-person outputs; block unconsented track rather than empty it silently. Evidence: consent tests.
    • A2.17.g Exclude biometrics from ordinary previews/search/logs/cache/ exports and test two tenants/roles/revocation. Evidence: privacy tests.
    • A2.17.h Run crossing/occlusion/clothing-similar/exit-reentry/swapped camera/partial consent fixtures and prove no silent identity swap. Evidence: identity safety suite.

A3 — Pipeline templates, execution, and stage control#

  • A3.1 Build pipeline-template registry and lifecycle. Evidence: draft, review, publish, deprecate, and archive journey.
    • A3.1.a Build permissioned registry/search/filter/view for template ID/ version/profile/owner/status/capabilities/formats/models/freshness. Evidence: registry tests.
    • A3.1.b Create/clone/import/edit immutable draft versions with DAG/config/ resource/cache/consent/gate/export metadata validation. Evidence: draft tests.
    • A3.1.c Route review comments/changes/domain-privacy-safety-engineering approvals on exact version with semantic diff. Evidence: review tests.
    • A3.1.d Publish/activate with effective window/compatibility/migration/ rollback, prohibit mutation, and dispatch only approved versions. Evidence: publish tests.
    • A3.1.e Deprecate/supersede/archive with dependent project/job/profile/ SDK impact, migration alternatives, and immutable history. Evidence: lifecycle journey.
  • A3.2 Build DAG editor with typed ports, cycle prevention, required-stage floors, keyboard editing, and semantic list parity. Evidence: a11y and validation tests.
    • A3.2.a Render versioned stage catalog and DAG nodes/ports/edges/ conditions/resources/status with stable selection and schema-driven configuration. Evidence: editor tests.
    • A3.2.b Add/remove/duplicate/configure/connect/disconnect/reorder/group stages with typed kind/cardinality/version/capability validation. Evidence: graph operation tests.
    • A3.2.c Detect cycles/unreachable outputs/missing required input/ incompatible optional branch/fan-in/resource/export path. Evidence: DAG validation.
    • A3.2.d Enforce profile/code-owned ingest/consent/privacy/validation/ package floors that template edits cannot bypass. Evidence: floor tests.
    • A3.2.e Provide source/target/port forms, ordered stage list/dependency table, before-after moves, complete keyboard/screen-reader/non-drag parity. Evidence: semantic tests.
    • A3.2.f Support undo/redo/diff/concurrent draft conflict/large graph performance and fail invalid publish. Evidence: editor suite.
  • A3.3 Build per-stage configuration forms from versioned schemas. Evidence: unknown properties and incompatible versions fail.
    • A3.3.a Resolve exact stage type/adapter/config schema/UI schema/default/ capability/profile version and migration. Evidence: schema resolver.
    • A3.3.b Render accessible typed controls for scalar/enum/range/unit/file/ artifact/model/rig/mapping/secret/conditional/repeated fields. Evidence: form tests.
    • A3.3.c Validate required/type/range/unit/cross-field/port/resource/ consent/profile/format constraints client and server. Evidence: validation tests.
    • A3.3.d Store explicit values separately from defaults/secrets, canonical serialize/hash, show output-affecting change and downstream cache impact. Evidence: persistence tests.
    • A3.3.e Migrate supported older configs with semantic diff; reject unknown properties/incompatible schema/adapter versions without defaulting. Evidence: compatibility tests.
  • A3.4 Implement ingest and preprocessing stage adapters. Evidence: exact input/output contracts.
    • A3.4.a Define typed accepted quarantined input assets/calibration/ metadata/consent/privacy and normalized output kinds/version. Evidence: adapter contract.
    • A3.4.b Implement decode/extract/orient/color/depth/timebase/frame/audio/ metadata normalization without altering immutable originals. Evidence: processing tests.
    • A3.4.c Implement resize/crop/redaction/sync/resample/frame selection/ segmentation-ready derivatives with exact transformation lineage. Evidence: derivative tests.
    • A3.4.d Emit per-view/range quality/availability/dropped/corrupt/ unsupported findings and never fabricate missing frames. Evidence: output tests.
    • A3.4.e Pin tool/library/config/hardware/input-output hashes, resource/ cost, logs, and deterministic cache/replay eligibility. Evidence: provenance tests.
    • A3.4.f Inject malformed/partial/missing calibration/consent revoke/ parser/worker/storage faults. Evidence: adapter fault suite.
  • A3.5 Implement 2D pose-estimation stage adapter. Evidence: provider/model provenance and confidence output.
    • A3.5.a Define accepted image/video views/ranges/subjects/profile/ modality/model/crop/calibration/consent and 2D output schema. Evidence: adapter contract.
    • A3.5.b Negotiate eligible local/provider model/version/body-hand-face/ multi-person capabilities, resources, privacy, and budget. Evidence: capability tests.
    • A3.5.c Execute inference with deterministic preprocessing/batching/ coordinate transform and emit landmarks/joints/tracks/confidence/ visibility/occlusion/tracking loss. Evidence: output tests.
    • A3.5.d Map outputs back to exact source camera/frame/pixel-normalized coordinates and subject identities with uncertainty. Evidence: alignment tests.
    • A3.5.e Persist provider/model/weights/config/hardware/corpus/input-output hashes/resource/cost and attributed proposal state. Evidence: provenance tests.
    • A3.5.f Test no-person/multi-person/occlusion/crop/rotation/low quality/ missing modality/provider failure/consent block. Evidence: adapter suite.
  • A3.6 Implement pose-lifting/3D reconstruction stage adapter. Evidence: coordinate and scale contract.
    • A3.6.a Define accepted synchronized 2D/multiview/depth/calibration/ skeleton/subject inputs and 3D joint/trajectory/confidence outputs. Evidence: adapter contract.
    • A3.6.b Negotiate monocular/multiview/depth/model/profile capabilities, coordinate/scale recovery assumptions, resources, and consent. Evidence: capability tests.
    • A3.6.c Execute triangulation/lifting/optimization with exact camera/ calibration/time associations and uncertainty propagation. Evidence: reconstruction tests.
    • A3.6.d Transform to canonical coordinate/units/ground/root frame and retain ambiguity/scale confidence rather than inventing absolute scale. Evidence: coordinate tests.
    • A3.6.e Persist model/config/calibration/input-output hashes/measurements/ resources/cost and proposal lineage. Evidence: provenance tests.
    • A3.6.f Test known 3D/calibration error/single-view/occlusion/drift/ swapped camera/low-confidence/failure fixtures. Evidence: adapter suite.
  • A3.7 Implement human-mesh-recovery stage adapter. Evidence: topology and subject privacy controls.
    • A3.7.a Define accepted images/video/2D-3D pose/calibration/subject/ profile/consent/model inputs and mesh/body-parameter/texture outputs. Evidence: adapter contract.
    • A3.7.b Negotiate approved topology/model/version/modality/hardware/ privacy/output use and block face/body detail outside consent. Evidence: capability tests.
    • A3.7.c Execute recovery with canonical coordinate/scale/frame alignment, stable topology/vertex-joint correspondence, confidence/occlusion. Evidence: mesh tests.
    • A3.7.d Generate redacted/identity-minimized mesh derivatives or suppress texture/shape fields by role/profile; keep protected source separated. Evidence: privacy tests.
    • A3.7.e Persist model/weights/config/input-output hashes/topology version/ resources/cost/provenance and generated disclosure. Evidence: lineage tests.
    • A3.7.f Test topology mismatch/partial body/loose clothing/multi-person/ privacy restriction/consent revoke/failure. Evidence: adapter suite.
  • A3.8 Implement skeleton mapping and retargeting stage adapters. Evidence: canonical, neural, semantic, and proportional strategy fixtures.
    • A3.8.a Define source/target skeleton-rig/motion/bind/coordinate/unit/ mapping/profile/strategy inputs and retargeted output contract. Evidence: adapter schemas.
    • A3.8.b Implement explicit canonical joint-name mapping with hierarchy/ parent/end-effector/root/twist/finger/face validation. Evidence: canonical fixtures.
    • A3.8.c Implement semantic candidate mapping with confidence/evidence and human-confirmed ambiguous/unmapped joints. Evidence: semantic fixtures.
    • A3.8.d Implement proportional retarget with bone lengths/scale/offset/ orientation/pose/contact/root constraints and disclosed assumptions. Evidence: proportional fixtures.
    • A3.8.e Implement neural mapping/retarget only through approved model/ training domain/consent/capability with attributed proposal output. Evidence: neural fixtures.
    • A3.8.f Preserve source and version mapping/output lineage; run quality measures, missing-joint/coordinate/version/failure tests. Evidence: strategy suite.
  • A3.9 Implement IK/cleanup, contact, jitter, interpolation, and root-motion stage adapters. Evidence: non-destructive revision links.
    • A3.9.a Define distinct stage variants with exact source motion/ranges/ joints/contacts/constraints/config/profile and preserved original. Evidence: stage contracts.
    • A3.9.b Implement IK/end-effector/contact constraints with chain/pole/ limits/weights/iterations, residuals, and failure/overconstraint. Evidence: IK tests.
    • A3.9.c Implement jitter/noise filtering with cutoff/window/adaptive/ confidence handling and no unintended timing/pose drift. Evidence: smoothing tests.
    • A3.9.d Implement gap interpolation for missing frames/joints using bounded duration/method/context/confidence and disclose generated ranges. Evidence: interpolation tests.
    • A3.9.e Implement root extraction/in-place/world trajectory/ground/ orientation/contact preservation and coordinate-unit consistency. Evidence: root tests.
    • A3.9.f Emit immutable new revision with per-frame/joint operation lineage; verify unchanged ranges stable, undo/compare, and quality measurements. Evidence: cleanup suite.
  • A3.10 Implement blend-shape/facial retargeting only under valid consent. Evidence: scope test.
    • A3.10.a Resolve source face/landmark/expression/blend-shape tracks, subject identity, exact consent modalities/uses/target/territory/expiry, and privacy profile. Evidence: preflight tests.
    • A3.10.b Define source-target face topology/landmark/blend-shape/neutral/ range/sign/scale mapping with confidence and version. Evidence: mapping contract.
    • A3.10.c Retarget curves/poses with calibration/smoothing/limits/ temporal alignment and preserve tracking loss/occlusion. Evidence: output tests.
    • A3.10.d Emit redacted identity-minimized motion where required, exact model/tool/config/source-output lineage, watermark, and disclosure. Evidence: privacy/provenance tests.
    • A3.10.e Recheck consent at dispatch/commit/package, cancel/quarantine/ invalidate on revoke, and test purpose/recipient/territory mismatch. Evidence: scope suite.
  • A3.11 Implement validation and packaging stages as explicit DAG nodes. Evidence: export cannot bypass them.
    • A3.11.a Define typed validation node consuming exact motion/rig/mapping/ profile/ground truth and producing raw measurements plus versioned verdicts. Evidence: validation contract.
    • A3.11.b Define review/waiver/adjudication prerequisites and block packaging on nonpass/stale/unavailable/low-confidence according to profile. Evidence: gate tests.
    • A3.11.c Define packaging node consuming validated approved revision, target format/rig/mapping/consent/watermark/provenance and outputs. Evidence: package contract.
    • A3.11.d Enforce DAG reachability/required-stage floors and server/worker authorization so export APIs cannot skip nodes or forge results. Evidence: bypass tests.
    • A3.11.e Recheck all dependencies/consent/gate versions atomically before package commit and invalidate downstream on change. Evidence: race tests.
    • A3.11.f Test direct API/CLI/worker/cache/imported artifact/export target bypass and preserve immutable audit. Evidence: enforcement suite.
  • A3.12 Implement honest unavailable for missing GPU, model, reference, rig, host, or calibration. Evidence: no placeholder output.
    • A3.12.a Define required capability/dependency resolution and distinct missing/not-configured/unsupported/unauthorized/offline/degraded/ unavailable reason codes. Evidence: state contract.
    • A3.12.b Check worker GPU/CPU/memory/architecture/model/licence/version/ host/profile capability before enqueue and lease. Evidence: compute tests.
    • A3.12.c Check required source/reference/calibration/skeleton/rig/mapping/ consent/rights/artifacts and exact versions. Evidence: dependency tests.
    • A3.12.d Return no synthetic/empty/identity/placeholder artifact; retain safe diagnostics, alternatives, remediation, and retry eligibility. Evidence: output tests.
    • A3.12.e Propagate unavailable to stage/job/batch/console/adapter/package without displaying success/healthy. Evidence: projection tests.
    • A3.12.f Exercise every missing combination plus resource loss after enqueue/lease and verify no placeholder output. Evidence: state suite.
  • A3.13 Implement enqueue, priority, dependency, pause, resume, cancel, retry, skip-if-optional, and rerun-from-stage. Evidence: state-machine tests.
    • A3.13.a Define canonical job/stage/attempt control state machine, actor/ authorization/idempotency/expected version and legal transitions. Evidence: state contract.
    • A3.13.b Enqueue exact template/input/config/consent/profile/budget with dependency readiness and bounded priority/fairness. Evidence: enqueue tests.
    • A3.13.c Pause/resume at safe stage/checkpoint boundaries with lease/ partial artifact/local state preservation and dependency revalidation. Evidence: pause tests.
    • A3.13.d Cancel queued/running/worker/provider/commit phases with requested-versus-completed state, fencing, cleanup, and partial outputs. Evidence: cancel tests.
    • A3.13.e Retry failed attempts under policy and allow skip only for declared optional stage with explicit downstream impact. Evidence: retry/skip tests.
    • A3.13.f Rerun from selected stage with new immutable config/input/ versions, cache/downstream invalidation, cost preview, and no history rewrite. Evidence: state-machine suite.
  • A3.14 Implement stage cache keys over every output-affecting input, version, model, and policy. Evidence: cache invalidation tests.
    • A3.14.a Enumerate all output-affecting input artifact hashes/ranges/ configs/defaults/calibration/skeleton/rig/mapping/reference/profile/ consent-privacy policy fields. Evidence: key manifest.
    • A3.14.b Include stage/adapter/code/container/model/weights/library/ hardware-sensitive/runtime/policy/evaluator/format versions and seed. Evidence: version key tests.
    • A3.14.c Canonical serialize/normalize order/units/defaults without secrets and hash with versioned algorithm/namespace/tenant/privacy boundary. Evidence: key tests.
    • A3.14.d Validate cached artifact manifest/hash/completeness/lineage/ consent/access/profile/expiry before hit; quarantine mismatch. Evidence: read tests.
    • A3.14.e Invalidate on any dependency/revocation/policy/version change and propagate to downstream stage/results/gates/packages. Evidence: invalidation tests.
    • A3.14.f Mutate each key dimension individually plus ordering/no-op fields and run collision/cross-tenant/corrupt cache tests. Evidence: cache suite.
  • A3.15 Implement checkpoint/recovery for long stages. Evidence: worker restart test.
    • A3.15.a Define checkpointable stage boundaries/state/shard/range/ progress/model/runtime/random state/artifacts/version and safe frequency. Evidence: checkpoint contract.
    • A3.15.b Write checkpoints atomically with hashes/fencing/job-stage- attempt identity and durable artifact references before acknowledging. Evidence: persistence tests.
    • A3.15.c On worker restart/loss, validate current lease/input/config/ model/policy/consent and resume compatible checkpoint or restart safely. Evidence: recovery tests.
    • A3.15.d Deduplicate/reconcile partial outputs/logs/measurements/resource usage/cost and prevent stale worker commit after reassignment. Evidence: fencing tests.
    • A3.15.e Expire/clean checkpoints only after terminal retention/active lease/legal hold/replay rules; quarantine corrupt state. Evidence: cleanup tests.
    • A3.15.f Kill workers at every boundary and verify final output/order/ lineage/equivalence or disclosed nondeterminism. Evidence: restart suite.
  • A3.16 Implement deterministic replay eligibility. Evidence: missing dependency is explicit.
    • A3.16.a Define determinism profile per stage covering code/model/runtime/ hardware/seed/parallelism/libraries/external provider/input/config. Evidence: replay contract.
    • A3.16.b Verify all immutable input/output-affecting dependencies and artifacts remain available, authorized, consent-valid, and hash-verified. Evidence: eligibility tests.
    • A3.16.c Classify exact deterministic/tolerance deterministic/ nondeterministic/nonreplayable with reason and expected comparison method. Evidence: classification.
    • A3.16.d Replay in isolated namespace using pinned dependencies, prevent cache masking when auditing, and record substitutions. Evidence: replay tests.
    • A3.16.e Compare outputs/measurements/events semantically and numerically within declared tolerances; preserve both versions. Evidence: comparison report.
    • A3.16.f Expose missing model/weight/runtime/calibration/input/seed/ licence/consent/provider as explicit blocker, never approximate silently. Evidence: failure tests.
  • A3.17 Implement resource/cost estimates and reservations per stage. Evidence: over-budget prevents dispatch.
    • A3.17.a Define resource estimator inputs and outputs for CPU/GPU type/ memory/storage/network/time/energy/provider units/currency/uncertainty. Evidence: estimate schema.
    • A3.17.b Estimate from stage/model/input frames-duration-resolution/ modalities/batch/worker history with model/version and confidence. Evidence: estimator tests.
    • A3.17.c Aggregate critical-path and parallel DAG estimates including cache likelihood, retries, fan-out/in, artifacts/egress, and target. Evidence: plan estimate.
    • A3.17.d Authorize budget/quotas/priority and reserve capacity/currency/ storage against owner/project/tenant with expiry and idempotency. Evidence: reservation tests.
    • A3.17.e Recheck before dispatch/lease/scale-up; block over-budget or insufficient capacity and require scoped reapproval. Evidence: dispatch tests.
    • A3.17.f Reconcile estimate/reservation/actual/release/refund/overrun and calibrate estimators without leaking subject data. Evidence: cost report.
  • A3.18 Implement batch submission and fan-out/fan-in. Evidence: partial failure and cancellation semantics.
    • A3.18.a Define batch/item identity/template/profile/input/config/ priority/budget/consent/idempotency/dependency and aggregate state. Evidence: batch schema.
    • A3.18.b Preflight each item independently for formats/consent/ capabilities/resources/cache/gates and show eligible/blocked/degraded/ cost. Evidence: preview tests.
    • A3.18.c Fan out jobs/stages/shards with bounded concurrency/fairness/ capacity/priority and preserve item/subject isolation. Evidence: dispatch tests.
    • A3.18.d Fan in only required successful compatible outputs with deterministic order/deduplication/aggregation and explicit missing/ partial inputs. Evidence: aggregation tests.
    • A3.18.e Cancel whole batch/selected items/queued or running fan-out with truthful partial completion and retained successful artifacts. Evidence: cancellation tests.
    • A3.18.f Retry safe failed items/stages, never rerun successes by default, and report aggregate partial/failed/complete without false success. Evidence: partial failure suite.
  • A3.19 Implement explicit full-body, hand/finger, face/landmark, gaze/expression, and multi-person tracking stages with capability negotiation. Evidence: missing modality/model/consent yields an honest unsupported or blocked state rather than empty tracks.
    • A3.19.a Define separate typed stage contracts/capability IDs for body, hand/finger, face/landmark, gaze/expression, and multi-person identity/ association. Evidence: modality registry.
    • A3.19.b Resolve requested input sensors/views/calibration/subject tracks/ profile/model/hardware/privacy/consent per modality before dispatch. Evidence: negotiation tests.
    • A3.19.c Execute body and hand stages with skeleton/joint/side/contact/ confidence/occlusion/tracking-loss outputs and source-frame alignment. Evidence: body/hand tests.
    • A3.19.d Execute face/expression/blend-shape and gaze stages only under exact biometric scope, with identity-minimized output where required. Evidence: face/gaze tests.
    • A3.19.e Execute multi-person detection/tracking/association across views with protected IDs, crossing/occlusion ambiguity, and correction proposals. Evidence: multi-person tests.
    • A3.19.f Persist independent model/config/input-output/provenance/quality/ cost per modality and never infer one modality from empty another. Evidence: lineage tests.
    • A3.19.g Emit unsupported/not-configured/consent-blocked/no-observation/ tracking-lost/low-confidence distinctly, not empty successful tracks. Evidence: absence tests.
    • A3.19.h Run modality combinations, partial consent, missing sensor/model/ GPU, multi-person crossings, revocation, and privacy projections. Evidence: capability suite.
  • A3.20 Implement video-analysis stages for action/segment/event detection, camera motion, occlusion, contacts, tracking confidence, and review proposals. Evidence: model output remains attributed and correctable.
    • A3.20.a Define separate stage/output contracts for temporal segments/ actions/events/camera motion/occlusion/contacts/tracking-quality proposals. Evidence: analysis registry.
    • A3.20.b Pin source take/views/ranges/profile/taxonomy/reference/model/ calibration/subjects/consent and exclude prohibited inferences. Evidence: request tests.
    • A3.20.c Detect action/phase/segment/event candidates with start/end/ label/confidence/alternatives and exact source locators. Evidence: temporal tests.
    • A3.20.d Estimate camera motion/shots and subject/joint visibility/ occlusion/contact/tracking confidence with raw measurements. Evidence: geometry tests.
    • A3.20.e Persist model/weights/config/taxonomy/corpus/input-output hashes/ uncertainty/cost and proposal-only status. Evidence: provenance tests.
    • A3.20.f Provide synchronized preview/table review and human accept/edit/ split/merge/relabel/reject/unknown with versioned annotations. Evidence: review tests.
    • A3.20.g Evaluate known actions/segments/events/camera/occlusion/contact/ tracking fixtures, subgroup/profile slices, and false claims. Evidence: calibrated report.
  • A3.21 Implement motion enhancement, generative completion, animation blending, retiming, gap filling, segmentation, and IK as non-destructive stage variants with exact edited ranges and source lineage. Evidence: unchanged frames remain stable and generated ranges are disclosed.
    • A3.21.a Define distinct operation stage contracts with exact source clip/ranges/joints/curves/masks/context/constraints/model/seed/profile and allowed output. Evidence: operation registry.
    • A3.21.b Implement filtering/enhancement and segmentation as versioned proposals with raw measurements, boundaries, and no destructive source edit. Evidence: enhancement tests.
    • A3.21.c Implement gap filling/generative completion with bounded missing frames/joints, context/contact/environment constraints, alternatives, and uncertainty. Evidence: completion tests.
    • A3.21.d Implement blend/transition between motions with alignment/ contact/root/phase/duration masks and seam measurements. Evidence: blending tests.
    • A3.21.e Implement retiming/time warp with source-destination map, contact/event preservation, derivative curves, and monotonicity. Evidence: retiming tests.
    • A3.21.f Implement IK constraint correction with chain/joint limits/ end-effectors/contacts/iterations/residuals and overconstraint failure. Evidence: IK tests.
    • A3.21.g Produce immutable candidates with per-frame/joint generated/ changed/unchanged labels, model/tool/config/seed/source lineage and disclosure. Evidence: provenance tests.
    • A3.21.h Audition/compare/accept-edit-reject exact ranges and verify promised unchanged frames/curves stable within zero or declared tolerance. Evidence: preservation suite.
  • A3.22 Implement profile-specific DAG templates for yoga, fitness, dance, martial arts, sports, medical/rehabilitation, film/VFX, and gaming. Evidence: each template satisfies A0.15 requirements and has independent known-value fixtures; shared stages do not imply profile proof.
    • A3.22.a Define the common template contract for version, capture protocol, intended/prohibited use, stages/models, annotations, evaluators/floors, qualified reviewer, outputs, warnings, and handoff. Evidence: template schema.
    • A3.22.b Ratify yoga and fitness templates with distinct pose/exercise vocabulary, repetitions/holds, reference measures, safety boundaries, and reviewer qualifications. Evidence: independent fixtures.
    • A3.22.c Ratify dance and martial-arts templates with distinct choreography/technique, rhythm/phase, partner/contact, weapon/safety, and performance-review semantics. Evidence: independent fixtures.
    • A3.22.d Ratify sport-specific templates without collapsing sports into one profile; declare event/technique measures, equipment/environment, prohibited claims, and qualified review. Evidence: sport registry and independent fixtures.
    • A3.22.e Ratify medical and rehabilitation templates with approved protocol, intended use, contraindication/escalation, measurement validity, privacy, and qualified clinical review. Evidence: independent protected fixtures.
    • A3.22.f Ratify film/VFX and gaming templates with distinct shot/runtime constraints, retarget/loop/root-motion needs, deliverables, and creative versus technical approval. Evidence: independent fixtures.
    • A3.22.g Prove every profile selects only compatible stages/models and cannot inherit another profile's thresholds, claims, review, or release evidence. Evidence: cross-profile negative suite.
  • A3.23 Implement format conversion among every ratified video, mocap, skeleton, animation, and engine/DCC package with loss preview, coordinate/unit mapping, embedded metadata, and round-trip diagnostics. Evidence: per-format-pair compatibility matrix.
    • A3.23.a Register each source/target format and profile version with parse/write capabilities, extensions/MIME, limits, and unsupported constructs. Evidence: conversion registry.
    • A3.23.b Parse and normalize source time base, coordinate system, units, skeleton/rig, animation channels, markers, metadata, and embedded media without discarding unknown data silently. Evidence: parse goldens.
    • A3.23.c Define explicit video, mocap, skeleton, animation, and engine/DCC source-target mappings, including required intermediate representations. Evidence: pairwise mapping matrix.
    • A3.23.d Preview dropped, approximated, baked, resampled, renamed, or synthesized data and require acceptance when the target is lossy. Evidence: loss-decision records.
    • A3.23.e Convert with deterministic settings and embed source, mapping, tool/version, checksum, rights/consent, and transformation provenance. Evidence: package inspection.
    • A3.23.f Read back every supported target and compare timing, hierarchy, transforms, curves, contacts, metadata, and declared loss against the source. Evidence: per-pair round-trip suite.
  • A3.24 Implement video-to-CG replacement for actor segmentation, tracking, pose/face/hand transfer, target avatar/rig, occlusion, lighting/shadow, rendering, compositing, temporal quality, review, and delivery through governed Isis/Yemaya/Bellona seams. Evidence: subject consent, source/ derivative provenance, generated-region disclosure, and per-shot replacement tests.
    • A3.24.a Register the source clip, shot ranges, subjects, consent scope, intended use, target project, and protected-source handling before analysis. Evidence: admissibility tests.
    • A3.24.b Segment and track each authorized actor with stable IDs, confidence, occlusion states, masks, and reviewer corrections. Evidence: tracking fixtures.
    • A3.24.c Select a rights-cleared target avatar/rig and verify body, face, hand, clothing, material, and output compatibility. Evidence: compatibility gate.
    • A3.24.d Transfer body, face, and hand motion as separately attributable stages with source ranges, mappings, confidence, corrections, and unchanged-region guarantees. Evidence: transfer tests.
    • A3.24.e Reconstruct occlusion, camera, lighting, contact shadows, and scene interactions with explicit observed versus generated regions. Evidence: per-shot scene report.
    • A3.24.f Render through the governed Bellona/Yemaya boundary and composite through the governed Isis/Yemaya boundary without granting Aja foreign mutation authority. Evidence: seam contracts.
    • A3.24.g Measure temporal identity, mask, edge, motion, lighting, shadow, occlusion, and source-background continuity across every shot. Evidence: known-value and perceptual gates.
    • A3.24.h Review, revise, approve, package, deliver, and read back exact shots with subject/derivative provenance, generated-region disclosure, rights/consent state, and recall hooks. Evidence: replacement journey.
  • A3.25 Implement text/reference-to-motion and physics-aware generative animation as proposal stages with motion intent, duration, skeleton, contacts, environment constraints, seed/model, alternatives, validation, and human acceptance. Evidence: generated frames are attributable and cannot bypass consent, quality, intended-use, or export gates.
    • A3.25.a Define separate text, image/video reference, pose/keyframe, trajectory, and mixed-conditioning request contracts. Evidence: typed input schemas.
    • A3.25.b Resolve motion intent, duration, motion class, skeleton/rig, coordinate/time basis, environment, contacts, obstacles, forces, and prohibited uses before generation. Evidence: constraint plan.
    • A3.25.c Enforce reference rights/consent and detect requests that would reproduce a protected performer or unsupported medical/performance claim. Evidence: admissibility suite.
    • A3.25.d Generate multiple immutable candidates with model/version, configuration, seed, inputs, constraints, nondeterminism, and per-frame/joint generation provenance. Evidence: candidate manifests.
    • A3.25.e Validate kinematics, contacts, collisions, balance, dynamics, skeleton compatibility, temporal continuity, and profile-specific quality without rewriting a failed candidate. Evidence: gate suite.
    • A3.25.f Let an authorized human compare, annotate, accept, edit, or reject candidates while preserving the original proposal and decision rationale. Evidence: review journey.
    • A3.25.g Recompute consent, intended-use, quality, compatibility, and export gates on the accepted revision and disclose generated ranges in every derivative/package. Evidence: fail-closed release test.

A4 — Motion quality, validation, and calibrated evaluation#

  • A4.1 Define versioned gate profiles by motion type, consumer, and export target. Evidence: every publishable format has a profile.
    • A4.1.a Define the gate-profile schema for motion class, intended use, consumer, target, applicable measurements, evaluator versions, thresholds, human-review bands, and failure policy. Evidence: schema tests.
    • A4.1.b Register a separately owned profile for every publishable motion-type/consumer/export combination. Evidence: coverage matrix.
    • A4.1.c Version, approve, activate, retire, and migrate profiles without rewriting historical verdicts. Evidence: lifecycle tests.
    • A4.1.d Reject validation or release when no current compatible profile exists. Evidence: missing-profile negative test.
  • A4.2 Implement skeleton hierarchy, joint naming, bind pose, unit, axis, and frame-rate validation. Evidence: broken fixtures.
    • A4.2.a Validate hierarchy topology, unique joints, parentage, roots, and required bones against the selected skeleton profile. Evidence: hierarchy fixtures.
    • A4.2.b Validate canonical/aliased joint names and surface unknown, duplicate, missing, or ambiguous mappings. Evidence: naming fixtures.
    • A4.2.c Validate bind/rest pose, transforms, orientation, scale, units, axes/handedness, and coordinate metadata. Evidence: transform fixtures.
    • A4.2.d Validate source and declared frame rate/time base without silently resampling. Evidence: timing fixtures.
  • A4.3 Implement missing, duplicate, discontinuous, and non-monotonic frame checks. Evidence: crafted clips.
    • A4.3.a Parse expected and observed frame/sample indices and timestamps. Evidence: frame-ledger fixture.
    • A4.3.b Detect missing and duplicate frames/samples with exact ranges and counts. Evidence: gap/duplicate fixtures.
    • A4.3.c Detect non-monotonic timestamps, discontinuities, resets, and inconsistent time steps. Evidence: temporal fixtures.
    • A4.3.d Separate raw findings from profile verdict and proposed repair. Evidence: policy-recompute test.
  • A4.4 Implement foot/hand contact and sliding measurements. Evidence: independently derived fixture values.
    • A4.4.a Detect or ingest foot/hand contact events with side, effector, frame range, confidence, and ground/object reference. Evidence: contact fixtures.
    • A4.4.b Measure contact-point displacement, velocity, rotation, and penetration in the declared coordinate/time basis. Evidence: independently calculated values.
    • A4.4.c Classify planted, transitioning, airborne, supported, and unknown intervals without treating uncertainty as pass. Evidence: state tests.
    • A4.4.d Apply motion/profile-specific sliding thresholds and retain raw event measurements for review. Evidence: verdict tests.
  • A4.5 Implement jitter/noise and temporal smoothness measurements. Evidence: synthetic signal fixtures.
    • A4.5.a Define joint/channel/range-specific position, rotation, velocity, acceleration, jerk, and frequency-domain signals. Evidence: measurement spec.
    • A4.5.b Compute signals at the declared sample rate with explicit unwrap, boundary, missing-data, and filter behavior. Evidence: synthetic goldens.
    • A4.5.c Detect local spikes, sustained noise, discontinuities, and over-smoothed loss of intended motion separately. Evidence: crafted fixtures.
    • A4.5.d Report exact offending joints/ranges and raw values before profile thresholds determine verdict. Evidence: diagnostic tests.
  • A4.6 Implement bone-length and joint-limit consistency. Evidence: anatomical fixtures.
    • A4.6.a Resolve skeleton-specific reference bone lengths, tolerances, scale rules, and joint-limit definitions. Evidence: profile fixtures.
    • A4.6.b Measure per-frame bone-length drift and distinguish scale or mapping errors from subject-independent anatomical assumptions. Evidence: known-value tests.
    • A4.6.c Detect angular, translational, twist, and hyperextension limit violations with exact joints/ranges. Evidence: limit fixtures.
    • A4.6.d Mark unsupported anatomy/profile combinations unavailable rather than inferring clinical validity. Evidence: negative tests.
  • A4.7 Implement balance, center-of-mass, ground penetration, velocity, acceleration, and physics plausibility diagnostics. Evidence: known motion corpus.
    • A4.7.a Resolve mass model, support contacts/polygon, ground plane, gravity, environment, time basis, and intended-use assumptions. Evidence: diagnostic manifest.
    • A4.7.b Compute center of mass, support margin, penetration, velocity, acceleration, momentum, and discontinuity signals with uncertainty. Evidence: analytical fixtures.
    • A4.7.c Detect incompatible contact, collision, ballistic, force, and balance behavior by motion phase. Evidence: known-motion corpus.
    • A4.7.d Separate animation plausibility warnings from medical, biomechanical, or safety claims requiring qualified validation. Evidence: claim-boundary tests.
  • A4.8 Implement pose, trajectory, contact, and timing fidelity against ground truth when available. Evidence: calibrated dataset.
    • A4.8.a Admit ground truth only with source, rights/consent, coordinate, skeleton, timing, annotation, and calibration provenance. Evidence: dataset gate.
    • A4.8.b Align prediction and reference across coordinate system, scale, skeleton mapping, temporal offset, and valid comparison ranges. Evidence: alignment goldens.
    • A4.8.c Compute pose, joint, trajectory, contact-event, duration, phase, and timing metrics with missing/occluded labels handled explicitly. Evidence: independently derived values.
    • A4.8.d Report per-sequence/per-joint/per-phase results, uncertainty, and calibration limitations before verdict. Evidence: calibrated report.
  • A4.9 Implement retarget fidelity for end effectors, root, silhouette, contacts, and joint rotations. Evidence: source/target fixture pairs.
    • A4.9.a Lock source motion, target rig, mapping, rest poses, scale, retarget settings, and comparison coordinate/time basis. Evidence: retarget manifest.
    • A4.9.b Measure end-effector, root trajectory/orientation, joint rotation, contact, and silhouette deviations separately. Evidence: pairwise goldens.
    • A4.9.c Detect target-limit violations, foot/hand drift, self- intersection, and proportion-induced unsupported comparisons. Evidence: retarget fixtures.
    • A4.9.d Preserve raw source-target measurements and profile-specific verdicts for exact frame review. Evidence: diagnostic report.
  • A4.10 Implement perceptual quality evaluation with model/version, uncertainty, and human-review band. Evidence: calibrated report.
    • A4.10.a Define the evaluation corpus, labels, sampling, consent, motion and demographic cohorts, exclusions, and human adjudication protocol. Evidence: approved dataset card.
    • A4.10.b Register evaluator model/version, preprocessing, configuration, input/output schema, and deterministic or stochastic behavior. Evidence: evaluator manifest.
    • A4.10.c Calibrate scores and uncertainty against held-out human ratings by supported cohort and intended use. Evidence: calibration report.
    • A4.10.d Set pass, fail, and mandatory-human-review bands without using low-confidence or out-of-distribution scores as pass. Evidence: threshold tests.
    • A4.10.e Monitor drift, subgroup error, calibration expiry, and model replacement with historical replay. Evidence: evaluator release gate.
  • A4.11 Implement motion-loop seam and blend compatibility checks. Evidence: loop fixtures.
    • A4.11.a Resolve loop endpoints or blend windows, phase/contact state, root-motion policy, and channels included in comparison. Evidence: seam manifest.
    • A4.11.b Measure pose, rotation, velocity, acceleration, root, contact, curve, and event discontinuity across the seam. Evidence: analytical loop fixtures.
    • A4.11.c Evaluate pairwise blend compatibility over permitted alignment, duration, contact, and phase choices. Evidence: blend matrix.
    • A4.11.d Report exact incompatible channels/ranges and avoid silently applying a destructive repair. Evidence: failure journey.
  • A4.12 Implement blend-shape/facial quality gates under consent. Evidence: protected evaluation.
    • A4.12.a Admit facial source, labels, and evaluation only under current subject consent, purpose, retention, and reviewer authorization. Evidence: protected-data gate.
    • A4.12.b Validate blend-shape names, ranges, exclusivity/coactivation, neutral pose, temporal completeness, and rig compatibility. Evidence: facial fixtures.
    • A4.12.c Measure facial landmark/expression fidelity, jitter, popping, lip/eye timing, symmetry where applicable, and uncertainty. Evidence: protected known values.
    • A4.12.d Revoke, restrict, or recompute facial evidence when consent, model, rig, or policy changes. Evidence: revocation test.
  • A4.13 Implement export-format conformance and round-trip checks. Evidence: per-format goldens.
    • A4.13.a Select the exact target format/profile/toolchain version and required metadata, hierarchy, timing, naming, and packaging rules. Evidence: conformance registry.
    • A4.13.b Validate the emitted bytes/package against official and pinned custom validators. Evidence: conformance goldens.
    • A4.13.c Re-import/read back through an independent target reader and compare motion, rig, curves, timing, contacts, metadata, and files. Evidence: round-trip suite.
    • A4.13.d Surface unsupported/lossy features and fail release when loss exceeds the accepted target profile. Evidence: loss tests.
  • A4.14 Persist raw metrics separately from verdicts. Evidence: policy recompute.
    • A4.14.a Store immutable measurement identity, inputs, units, ranges, algorithm/version, configuration, and raw values. Evidence: metric schema.
    • A4.14.b Store verdict identity, profile/version, thresholds, policy inputs, status, reason, and reviewer action separately. Evidence: verdict schema.
    • A4.14.c Recompute a new verdict from unchanged raw metrics under a new policy without rewriting historical results. Evidence: recompute test.
  • A4.15 Version algorithms, models, ground truth, calibration, thresholds, and profile. Evidence: result provenance.
    • A4.15.a Assign immutable versions and checksums to evaluator code, model, ground truth, calibration, threshold set, and gate profile. Evidence: registry.
    • A4.15.b Persist the complete version tuple and runtime environment on each measurement and verdict. Evidence: provenance query.
    • A4.15.c Detect stale, revoked, incompatible, or expired dependencies and identify affected assets/runs/releases. Evidence: impact test.
  • A4.16 Add unavailable, insufficient-ground-truth, low-confidence, and evaluator-failure states. Evidence: none renders as pass.
    • A4.16.a Define typed non-pass states with reason codes, diagnostic detail, retryability, owner, and permitted next action. Evidence: state contract.
    • A4.16.b Emit unavailable when a required evaluator, model, profile, dependency, or environment cannot run. Evidence: outage test.
    • A4.16.c Emit insufficient-ground-truth and low-confidence when required comparison or calibration support is absent. Evidence: boundary tests.
    • A4.16.d Emit evaluator-failure separately from domain-quality failure and prevent all non-pass states from satisfying release. Evidence: fail-closed UI/API tests.
  • A4.17 Build manual QA sampling and adjudication. Evidence: reviewer decisions link to exact frames and measurements.
    • A4.17.a Define risk-weighted, random, drift-triggered, disagreement, and mandatory-review sampling rules. Evidence: sampling policy.
    • A4.17.b Create immutable review assignments with qualified reviewer, conflict rules, due state, exact asset/revision, and blinded context where required. Evidence: assignment tests.
    • A4.17.c Capture decisions, reason codes, annotations, exact frames/ joints/measurements, confidence, and requested remediation. Evidence: adjudication record.
    • A4.17.d Escalate disagreement and update calibration/threshold evidence without overwriting original machine or human decisions. Evidence: adjudication journey.
  • A4.18 Build drift monitoring by capture setup, subject-safe cohort, motion class, model, rig, and version. Evidence: privacy-approved report.
    • A4.18.a Define privacy-safe dimensions, minimum cohort size, baselines, windows, statistical tests, and alert thresholds. Evidence: approved monitoring spec.
    • A4.18.b Aggregate quality, failure, uncertainty, and reviewer outcomes by capture setup, safe cohort, motion class, model, rig, and version. Evidence: bounded metrics.
    • A4.18.c Suppress or combine sparse cohorts and exclude subject identity or raw biometric/media data. Evidence: privacy tests.
    • A4.18.d Investigate alerts through versioned slices and link confirmed drift to model/profile rollback, recalibration, or review expansion. Evidence: drift drill.

A5 — Preview, cleanup, retargeting, annotation, and library#

  • A5.1 Build synchronized source video, 2D pose, 3D skeleton, mesh, and retarget preview. Evidence: deterministic time alignment.
    • A5.1.a Define the shared time basis, offsets, rates, dropped/missing samples, coordinate transforms, and stream availability contract. Evidence: synchronization schema.
    • A5.1.b Load authorized source video and derived 2D pose without exposing protected media to unauthorized viewers. Evidence: access tests.
    • A5.1.c Render 3D skeleton, mesh, and retarget streams at the same selected source instant with explicit interpolation behavior. Evidence: alignment goldens.
    • A5.1.d Keep transport, seek, range, loop, and selected subject/joint synchronized across every available view. Evidence: Playwright journey.
    • A5.1.e Surface missing, stale, offset, low-confidence, and unavailable streams without fabricating visual alignment. Evidence: state matrix.
  • A5.2 Build before/after and multi-run comparison with synchronized camera, transport, speed, loop, and overlays. Evidence: accessible comparison.
    • A5.2.a Select immutable source/run/artifact revisions and validate skeleton, timing, and view compatibility. Evidence: selection tests.
    • A5.2.b Synchronize camera, transport, speed, range, loop, subject, skeleton, and overlay settings across compared revisions. Evidence: interaction tests.
    • A5.2.c Support side-by-side, overlay, difference, and A/B switching without losing revision identity. Evidence: comparison journey.
    • A5.2.d Provide a semantic difference table, keyboard controls, and announced current revision/value for non-visual review. Evidence: accessibility suite.
  • A5.3 Build joint/contact/error heatmaps with semantic table parity. Evidence: non-visual access to exact values.
    • A5.3.a Define color-independent scales, units, thresholds, missing/ unavailable states, and selected measure/profile/version. Evidence: legend tests.
    • A5.3.b Map per-frame/joint/contact/error values to the visual skeleton and timeline without aggregating away critical failures. Evidence: mapping goldens.
    • A5.3.c Provide a sortable/filterable semantic table with exact joint, range, raw value, unit, threshold, status, and source measure. Evidence: parity tests.
    • A5.3.d Keep visual and table focus/selection synchronized by keyboard and assistive technology. Evidence: Playwright accessibility journey.
  • A5.4 Build frame/time/joint-range annotation. Evidence: anchors survive unrelated revisions.
    • A5.4.a Define annotation anchors for artifact revision, source time/ frame range, subject, joint/channel, coordinate basis, and selected measurements. Evidence: anchor schema.
    • A5.4.b Create, edit, reply, resolve, reopen, and delete annotations with authorization, audit, and concurrency control. Evidence: API/UI tests.
    • A5.4.c Rebase anchors only through explicit mapping when a related revision changes timing or skeleton; mark ambiguous/orphaned anchors. Evidence: revision fixtures.
    • A5.4.d Verify unrelated revisions do not move or mutate existing annotation anchors. Evidence: stability test.
  • A5.5 Build cleanup operations as versioned commands, never destructive source edits. Evidence: undo/redo and lineage.
    • A5.5.a Register typed cleanup commands with exact input revision, target ranges/joints/channels, parameters, preconditions, and impact preview. Evidence: command registry.
    • A5.5.b Execute each command into a new immutable revision with operator, tool/version, settings, changed ranges, and source lineage. Evidence: mutation tests.
    • A5.5.c Implement undo/redo as revision navigation or compensating commands rather than source mutation. Evidence: history journey.
    • A5.5.d Compare and validate the resulting revision before it can replace a selected candidate. Evidence: acceptance gate.
  • A5.6 Build skeleton and bone mapping editor with source/target validation. Evidence: keyboard and list parity.
    • A5.6.a Load immutable source/target skeleton revisions, naming aliases, required joints, rest poses, units, and mapping constraints. Evidence: editor contract.
    • A5.6.b Create, remove, remap, mirror, and auto-propose mappings without accepting ambiguous or incompatible suggestions silently. Evidence: mapping tests.
    • A5.6.c Validate completeness, uniqueness, hierarchy, orientation, scale, unmapped influence, and target-limit compatibility. Evidence: validation fixtures.
    • A5.6.d Provide keyboard/list operations and semantic source-to-target status equivalent to the visual mapping view. Evidence: accessibility journey.
    • A5.6.e Save a versioned mapping draft and require explicit review/ promotion before production use. Evidence: lifecycle test.
  • A5.7 Build retarget profile compare and decision record. Evidence: selected mapping remains linked.
    • A5.7.a Select compatible source motion, target rig, mapping, retarget profiles, evaluator profile, and exact revisions. Evidence: setup tests.
    • A5.7.b Run and display comparable outputs, raw fidelity metrics, warnings, limitations, and cost for each profile. Evidence: comparison fixture.
    • A5.7.c Capture the selected profile/mapping/revisions, reviewer, rationale, rejected alternatives, and intended consumer. Evidence: decision record.
    • A5.7.d Detect stale decisions when any linked mapping, rig, profile, or motion revision changes. Evidence: invalidation test.
  • A5.8 Build avatar/rig library search, compatibility, preview, rights, and version selection. Evidence: ineligible rig reasons.
    • A5.8.a Index only authorized avatar/rig metadata, versions, skeleton, features, formats, rights, license, consent constraints, and lifecycle state. Evidence: search policy tests.
    • A5.8.b Filter and rank by current project, motion, pipeline, target, and rights/compatibility requirements. Evidence: eligibility fixtures.
    • A5.8.c Preview authorized geometry, skeleton, materials, blend shapes, motions, limitations, and provenance at an exact version. Evidence: preview tests.
    • A5.8.d Explain each ineligible, quarantined, revoked, incompatible, or unavailable rig without leaking protected metadata. Evidence: reason matrix.
    • A5.8.e Lock the selected rig revision and recheck it at execution and export. Evidence: race test.
  • A5.9 Build motion asset library with tags, motion class, skeleton, quality, consent, watermark, lineage, and release status. Evidence: permission-aware search.
    • A5.9.a Define the motion-asset projection from authoritative artifact, skeleton, quality, consent, rights, watermark, lineage, and release records. Evidence: projection contract.
    • A5.9.b Index approved searchable fields while excluding protected subject/source identity and unauthorized media-derived metadata. Evidence: index tests.
    • A5.9.c Filter and facet by governed tags, motion class, skeleton, compatibility, quality freshness, consent/use, watermark, lineage, and release state. Evidence: search fixtures.
    • A5.9.d Show exact version, eligibility reasons, dependencies, preview availability, and current recall state before selection. Evidence: library journey.
    • A5.9.e Reconcile index/library state after supersession, consent change, quarantine, unpublish, or recall. Evidence: propagation test.
  • A5.10 Build exact and perceptual dedupe with merge/keep decisions. Evidence: no automatic destructive merge.
    • A5.10.a Compute exact identity from normalized bytes/checksums while preserving distinct ownership and provenance records. Evidence: exact duplicate fixtures.
    • A5.10.b Compute versioned perceptual candidates from authorized motion features with score, uncertainty, and compatible skeleton/time basis. Evidence: near-duplicate fixtures.
    • A5.10.c Present source, rights/consent, quality, versions, dependencies, and exact differences for reviewer comparison. Evidence: review journey.
    • A5.10.d Record keep-both, relate, consolidate-reference, or reject decisions without deleting source artifacts or lineage. Evidence: non-destructive tests.
  • A5.11 Build collections, sets, variants, related motions, and retirement through curation bundles. Evidence: no direct production writes.
    • A5.11.a Define versioned curation-bundle operations for collection/set membership, variant groups, relationships, metadata, and retirement. Evidence: bundle schema.
    • A5.11.b Validate referenced asset versions, permissions, lifecycle, rights/consent, compatibility, and dependency impact. Evidence: preflight tests.
    • A5.11.c Review, approve, apply, reject, or rollback a bundle with idempotency, concurrency, and audit. Evidence: workflow tests.
    • A5.11.d Prohibit direct production mutation outside the approved bundle applier and verify read-back. Evidence: authorization test.
  • A5.12 Build reference video library with protected-source controls. Evidence: subject identity and raw media do not leak.
    • A5.12.a Register reference videos with subject/owner, consent purpose, retention, geography, access compartment, checksums, and lineage. Evidence: protected-source schema.
    • A5.12.b Produce privacy-minimized searchable projections that omit subject identity, raw frames, biometrics, and sensitive inferred metadata. Evidence: index inspection.
    • A5.12.c Enforce object/property authorization, watermarking, bounded preview, download restrictions, and audited access. Evidence: access tests.
    • A5.12.d Revoke, expire, quarantine, retain under hold, or delete source and derived previews according to policy. Evidence: lifecycle drill.
  • A5.13 Build consent expiry/revocation impact view per asset. Evidence: released and derived assets appear.
    • A5.13.a Traverse from consent grant/subject/source to every capture, artifact, derivative, library entry, package, release, consumer, and retained copy. Evidence: dependency graph test.
    • A5.13.b Distinguish current, expiring, expired, revoked, disputed, unknown, and legally held states with effective times. Evidence: state matrix.
    • A5.13.c Show affected revisions, release status, downstream reach, remediation owner, deadline, and unavailable dependencies without leaking protected data. Evidence: impact-view tests.
    • A5.13.d Launch authorized quarantine/recall workflows only after an exact impact preview and confirm closure by consumer read-back. Evidence: revocation journey.
  • A5.14 Build accessible non-3D alternatives for preview and QA. Evidence: screen-reader/keyboard review journey.
    • A5.14.a Define semantic representations for time, skeleton hierarchy, pose/channel values, contacts, annotations, errors, and comparison differences. Evidence: parity specification.
    • A5.14.b Implement keyboard-operable timeline/range/joint selection and synchronized structured tables. Evidence: keyboard suite.
    • A5.14.c Announce transport, selection, validation, loading, stale, and failure state changes without excessive verbosity. Evidence: screen- reader tests.
    • A5.14.d Prove every QA decision possible in 3D has a non-visual path to the same raw evidence and action. Evidence: parity journey.
  • A5.15 Build UI-governed motion, action, skeleton, rig, capture, quality, consent, format, and catalog taxonomies over code-owned floors. Evidence: versioned review, alias, deprecation, and migration journey.
    • A5.15.a Define separate schemas, code-owned invariants, ownership, and allowed relationships for each taxonomy. Evidence: taxonomy registry.
    • A5.15.b Create draft terms with canonical IDs, labels/locales, definitions, aliases, parent/mapping relations, and usage constraints. Evidence: authoring tests.
    • A5.15.c Review, approve, publish, deprecate, supersede, and retire versioned terms without reinterpreting historical records. Evidence: lifecycle journey.
    • A5.15.d Reject collisions, cycles, invalid mappings, floor violations, and destructive deletion of referenced terms. Evidence: validation suite.
    • A5.15.e Migrate references through previewed, authorized, resumable jobs with per-record results and rollback/compensation. Evidence: migration journey.
  • A5.16 Build taxonomy change impact over pipeline templates, search, assets, gates, exports, adapters, and collections. Evidence: read-only preview precedes an authorized migration job.
    • A5.16.a Build a complete reference index from taxonomy terms/versions to templates, assets, search documents, gates, exports, adapters, and collections. Evidence: zero-orphan fixture.
    • A5.16.b Preview affected objects, incompatible mappings, search changes, gate/profile changes, release risk, and migration cost read-only. Evidence: impact report.
    • A5.16.c Require authorized approval of the exact taxonomy and impact revisions before queuing migration. Evidence: stale-preview test.
    • A5.16.d Execute, checkpoint, resume, partially fail, compensate, and read back migrations without hidden direct writes. Evidence: job tests.
  • A5.17 Build avatar customization and governed marketplace intake for body/ proportions, skeleton, blend shapes, materials, clothing, rig/format compatibility, rights, license, version, security scan, preview, project use, dependency impact, update, quarantine, and takedown. Evidence: marketplace assets cannot enter capture/retarget/export without current compatibility and rights evidence.
    • A5.17.a Define separately governed customization parameters, bounds, dependencies, compatibility rules, and immutable output revisions for body, skeleton, face, materials, and clothing. Evidence: customization schema.
    • A5.17.b Preview customization across representative poses/motions and validate deformation, collision, rig, blend-shape, material, clothing, and target-format compatibility. Evidence: preview fixtures.
    • A5.17.c Ingest marketplace submissions with publisher identity, source, files/checksums, rights/license, versions, dependencies, supported rigs/formats, and declared limitations. Evidence: intake contract.
    • A5.17.d Scan archives, code/plugins, geometry, textures, materials, metadata, dependencies, and licenses in quarantine. Evidence: security and policy suite.
    • A5.17.e Review and promote an exact marketplace version with current compatibility, rights, security, quality, and intended-use evidence. Evidence: promotion journey.
    • A5.17.f Lock eligible versions on project use and recompute evidence at capture, retarget, package, and export boundaries. Evidence: stale- eligibility tests.
    • A5.17.g Preview dependency impact and propagate update, quarantine, revocation, takedown, and recall to every project/package/consumer. Evidence: marketplace lifecycle drill.

A6 — Persistence, APIs, queue, workers, and security#

  • A6.1 Decide authoritative stores for capture, jobs, runs, artifacts, rigs, mappings, annotations, validation, consent, and packages. Evidence: ownership ADR.
    • A6.1.a Inventory every current table, blob, index, cache, event, file, and in-memory record for each named entity. Evidence: storage map. <!-- aja-storage-inventory.json: storage map of Aja. Of the 7 storage kinds, 4 PRESENT (cache, inMemory, file, blob) and 3 ABSENT (table, index, event). KEY FINDING: Aja has 0 relational tables — it is a motion-capture COMPUTE domain with NO relational persistence, unlike Isis(24)/Veritas(67)/Yemaya(38) which are relational service-direct. Aja's storage is the 6-module pipeline-cache subsystem (cache-invalidation/cache-utils/lifting-cache/ pipeline-cache-manager/pose-estimation-cache/retargeting-cache) backed by in-memory Map/LRU stores, plus motion file I/O (fbx/bvh/npz). No search index, no EventEmitter (synchronous pipeline). This COMPLETES the cross-domain storage architecture comparison across all V1 domains: relational service-direct (Isis/Veritas/Yemaya) / desktop-file Tauri (Euterpe) / compute-cache (Aja). Precise scoped patterns (the broad "event" keyword over-counted 110→0 real EventEmitters). generate-aja-storage-inventory.mjs + 5 tests, wired into inventory/verify + CI. A6.1.b (one authoritative store/owner per entity) is design and stays unchecked. -->
    • A6.1.b Assign one authoritative store and owner per entity/property, plus projections, retention, consistency, and mutation boundaries. Evidence: ownership matrix.
    • A6.1.c Resolve duplicate or conflicting authorities and define migration and compatibility windows. Evidence: approved ADR.
    • A6.1.d Trace create, update, read, delete/retain, audit, recovery, and downstream propagation for each authority. Evidence: lifecycle review.
  • A6.2 Add missing explicit-column schemas and safe migrations. Evidence: production-size rehearsal.
    • A6.2.a Specify normalized tables, explicit columns/types, keys, constraints, indexes, tenancy, audit fields, and state invariants. Evidence: schema review.
    • A6.2.b Write additive expand migrations and backfills that are bounded, resumable, observable, and safe under concurrent traffic. Evidence: migration tests.
    • A6.2.c Dual-read/write or translate during the compatibility window and verify old/new representation parity. Evidence: compatibility suite.
    • A6.2.d Rehearse production-size backfill, rollback/roll-forward, lock duration, failure recovery, and contract cleanup. Evidence: rehearsal report.
  • A6.3 Implement immutable blob refs, checksums, encryption, retention, legal hold, and derived privacy variants. Evidence: byte rehash.
    • A6.3.a Define immutable content-addressed refs with tenant/project, media type, size, checksum algorithm/value, encryption key version, source lineage, and lifecycle class. Evidence: blob contract.
    • A6.3.b Upload through staging, validate size/type/checksum/malware, and atomically promote without permitting ref overwrite. Evidence: ingest tests.
    • A6.3.c Encrypt in transit/at rest, authorize object and variant access, rotate keys, and prevent cross-tenant ref inference. Evidence: security suite.
    • A6.3.d Apply retention, expiry, deletion, legal hold, consent freeze, and tombstone behavior to originals and derivatives. Evidence: lifecycle tests.
    • A6.3.e Generate privacy-minimized preview/analysis variants with their own checksums and source/policy lineage. Evidence: variant tests.
    • A6.3.f Stream/read back and independently rehash stored bytes before release use. Evidence: corruption fixture.
  • A6.4 Implement capture/job/stage/artifact/rig/annotation/consent APIs with object/property authorization. Evidence: contract suite.
    • A6.4.a Define resource, list/filter, create, mutation, transition, revision, error, pagination, and idempotency contracts per API. Evidence: OpenAPI review.
    • A6.4.b Enforce identity, tenant/project scope, role/capability, object, property, consent compartment, and current revision on every operation. Evidence: authorization matrix.
    • A6.4.c Implement optimistic concurrency, idempotency, transition invariants, immutable lineage, and transactional outbox/audit behavior. Evidence: race tests.
    • A6.4.d Filter lists, errors, expansions, fields, counts, and related resources so unauthorized existence cannot leak. Evidence: negative contract suite.
  • A6.5 Implement durable queue, distributed worker leases, heartbeats, checkpoint, retry, dead-letter, and cancel. Evidence: fault injection.
    • A6.5.a Persist queue item identity, stage/run revision, requirements, priority, availability, attempt, deadline, idempotency, and state. Evidence: queue schema.
    • A6.5.b Claim work with capability-aware atomic leases, fencing tokens, expiry, and duplicate-delivery tolerance. Evidence: contention tests.
    • A6.5.c Heartbeat and checkpoint resumable progress without allowing a stale worker to commit after lease loss. Evidence: fencing test.
    • A6.5.d Classify retryable/permanent/cancelled failures, apply bounded backoff/jitter/budgets, and retain attempt lineage. Evidence: retry tests.
    • A6.5.e Dead-letter exhausted work with diagnostics and authorized replay from an exact revision/configuration. Evidence: replay journey.
    • A6.5.f Propagate cancel cooperatively, terminate within policy, clean partial outputs, and record late results without publishing them. Evidence: cancellation fault injection.
  • A6.6 Implement worker capability, version, model, GPU, load, and health registration. Evidence: incompatible work never dispatches.
    • A6.6.a Define signed worker identity and heartbeat schema for runtime, adapter, tool/model versions, CPU/GPU/memory/storage, capabilities, locality, load, and health. Evidence: registration contract.
    • A6.6.b Authenticate and authorize registration, rotation, drain, quarantine, and removal without trusting self-asserted privilege. Evidence: spoof tests.
    • A6.6.c Match exact stage requirements, compatibility ranges, resources, data locality, consent constraints, and health before lease grant. Evidence: scheduler fixtures.
    • A6.6.d Expire stale workers and reconcile leased work without double commit. Evidence: heartbeat-loss test.
  • A6.7 Implement queue fairness, priority, capacity, and backpressure. Evidence: load tests.
    • A6.7.a Define tenant/project/class quotas, weights, priorities, aging, reservations, concurrency, and starvation limits. Evidence: scheduling policy.
    • A6.7.b Reserve worker/GPU/memory/storage/license capacity atomically before dispatch and release it on every terminal path. Evidence: reservation tests.
    • A6.7.c Enforce ingest, queue-depth, cost, deadline, and downstream pressure limits with explicit reject/defer states. Evidence: backpressure fixtures.
    • A6.7.d Prove urgent work progresses without starving ordinary tenants and noisy tenants cannot monopolize capacity. Evidence: load report.
  • A6.8 Implement pipeline-cache read/write/invalidate/quarantine. Evidence: corrupt cache test.
    • A6.8.a Define complete cache keys from input bytes/revisions, stage/ adapter/model/tool/config versions, profile, environment, and relevant policy/consent state. Evidence: key tests.
    • A6.8.b Write immutable entries only after successful validation with checksums, output manifest, creation lineage, TTL, and eligibility. Evidence: write tests.
    • A6.8.c Read only compatible, current, authorized, checksum-valid entries and record cache-hit provenance on the run. Evidence: read tests.
    • A6.8.d Invalidate by dependency/version/policy/consent/recall impact without deleting evidence needed for historical reconstruction. Evidence: invalidation graph test.
    • A6.8.e Quarantine corruption or suspicious entries, prevent further hits, and recompute safely. Evidence: poison/corruption test.
  • A6.9 Implement validation and gate APIs with stale-policy semantics. Evidence: old pass cannot release.
    • A6.9.a Expose immutable raw measurements, evaluator provenance, verdict profile/version, freshness, non-pass state, and exact input revision. Evidence: response contracts.
    • A6.9.b Authorize measurement access, validation requests, manual review, waiver, and profile administration separately. Evidence: permission tests.
    • A6.9.c Detect stale inputs, algorithms/models, calibration, thresholds, profile, consent, rights, or target and return typed reasons. Evidence: staleness matrix.
    • A6.9.d Require recomputation/review under the current release profile; never translate an old pass into current pass. Evidence: release race test.
  • A6.10 Implement package/export APIs with idempotency and current consent lock. Evidence: revocation race test.
    • A6.10.a Define package/export request, preview, eligibility, target, mapping, watermark, dependency lock, idempotency, and status contracts. Evidence: OpenAPI fixtures.
    • A6.10.b Preview exact included/skipped assets, versions, gates, consent/ rights, losses, cost, and target limitations without mutation. Evidence: preflight tests.
    • A6.10.c Acquire fenced locks on exact revisions and current consent/ rights/gate state immediately before producing output. Evidence: concurrency tests.
    • A6.10.d Produce one immutable package/result per idempotency key and reject mismatched retries. Evidence: duplicate-request test.
    • A6.10.e Verify target read-back, sign manifest/provenance, publish atomically, and release locks on every outcome. Evidence: export journey.
    • A6.10.f Abort or recall when consent changes during assembly/publication and prove no eligible-looking partial output remains. Evidence: revocation race test.
  • A6.11 Implement canonical embodied-instruction adapter projections. Evidence: capability, demonstration, overlay, and handoff contract tests.
    • A6.11.a Implement the canonical health/capability projection with supported motion/profile/input/output/version/latency/limit detail. Evidence: capability contract.
    • A6.11.b Implement demonstration lookup/request projection with current consent, rights, quality, compatibility, and disclosure eligibility. Evidence: demonstration tests.
    • A6.11.c Implement coaching-overlay projection with bounded semantic measures/annotations and no protected raw media or unsupported claim. Evidence: overlay contract.
    • A6.11.d Implement asynchronous handoff/status/result/cancel projection with stable correlation, idempotency, degraded health, and provenance. Evidence: handoff journey.
    • A6.11.e Map internal evolution behind versioned adapter contracts and fail closed on unsupported negotiation. Evidence: compatibility tests.
  • A6.12 Implement permission-aware search and analytics. Evidence: biometric/protected metadata exclusion.
    • A6.12.a Define approved searchable/aggregatable fields and prohibited raw media, biometric, subject identity, consent detail, and protected metadata per role/purpose. Evidence: field policy.
    • A6.12.b Build tenant/project/consent-compartment filtered projections from authoritative events with deletion/recall reconciliation. Evidence: index contract.
    • A6.12.c Enforce authorization before query, filtering, facets, counts, suggestions, snippets, exports, and analytics dimensions. Evidence: inference tests.
    • A6.12.d Apply minimum cohorts, suppression/noise where required, bounded cardinality, retention, and audit to analytics. Evidence: privacy suite.
  • A6.13 Implement two-tenant/project/consent-compartment isolation across API, blob, search, jobs, cache, preview, and export. Evidence: integration suite.
    • A6.13.a Create fixtures with colliding IDs, names, hashes, tags, models, and resources across two tenants/projects/consent compartments. Evidence: isolation corpus.
    • A6.13.b Prove CRUD, list, filter, expansion, error, count, and audit isolation in every API. Evidence: API suite.
    • A6.13.c Prove blob/ref, signed URL, preview, search/index, cache key/hit, queue/worker, log/trace, and webhook isolation. Evidence: service suite.
    • A6.13.d Prove package/export, recall, backup/restore, and administrative operations cannot cross boundaries without explicit governed authority. Evidence: adversarial integration test.
  • A6.14 Generate OpenAPI and worker/adapter consumer contracts. Evidence: drift gate.
    • A6.14.a Generate OpenAPI from the authoritative API definitions with stable operation IDs, schemas, errors, auth, pagination, idempotency, and examples. Evidence: spec validation.
    • A6.14.b Generate worker protocol and canonical adapter contracts with versions, capabilities, events, commands, and compatibility rules. Evidence: protocol artifacts.
    • A6.14.c Generate client/server fixtures or SDK types and compile representative producers/consumers. Evidence: contract build.
    • A6.14.d Diff checked-in/generated artifacts and block breaking or undocumented drift. Evidence: CI gate.
  • A6.15 Implement webhook subscription, signed delivery, filtering, retries, replay protection, dead-letter, disable, and audit for job, stage, output, consent, quality, package, and recall events. Evidence: external subscriber conformance and secret-rotation tests.
    • A6.15.a Create authorized subscriptions with tenant/project scope, event/filter selection, HTTPS endpoint policy, secret ownership, rate limits, and verification handshake. Evidence: subscription tests.
    • A6.15.b Emit versioned minimal event envelopes from the transactional outbox with stable event/resource IDs, revisions, time, trace, and no protected payload. Evidence: event contract.
    • A6.15.c Sign timestamped deliveries and verify SSRF protections, DNS/IP policy, TLS, body checksum, secret version, and bounded response. Evidence: security suite.
    • A6.15.d Apply ordered/unordered semantics explicitly, retry with bounded backoff, prevent replay, and dead-letter exhausted delivery. Evidence: fault tests.
    • A6.15.e Support secret overlap/rotation, pause/disable, replay from an authorized range, and immutable delivery audit. Evidence: lifecycle journey.
    • A6.15.f Prove an external subscriber handles job, stage, output, consent, quality, package, and recall events compatibly. Evidence: conformance suite.
  • A6.16 Complete TypeScript/Python SDK and CLI parity for capability/format discovery, ingest/live-session handoff, job operations, batch, preview refs, annotations, rigs, validation, export, webhooks, and audit-safe inspection. Evidence: generated public-contract matrix and end-to-end external-client tests.
    • A6.16.a Generate or implement shared authentication, endpoint, pagination, retries, timeouts, idempotency, errors, version negotiation, and observability consistently in both SDKs and CLI. Evidence: core parity tests.
    • A6.16.b Implement capability/format/profile discovery and ingest/live- session handoff in TypeScript, Python, and CLI. Evidence: discovery/ intake client tests.
    • A6.16.c Implement job/stage status, batch, retry/cancel, wait/stream, preview refs, and result retrieval in all clients. Evidence: job parity tests.
    • A6.16.d Implement annotation, rig/mapping, validation/review, export/ package, webhook, and audit-safe inspection operations in all clients. Evidence: operation matrix.
    • A6.16.e Keep protected content out of default logs/output, expose machine-readable CLI modes, and handle partial/degraded states. Evidence: privacy and UX tests.
    • A6.16.f Run end-to-end external TypeScript, Python, and CLI journeys against the same server contract and compare observable behavior. Evidence: parity report.
  • A6.17 Implement video chunking, multi-GPU/stage pipeline parallelism, model compilation/optimization, distributed result aggregation, conflict/quality selection, and final assembly with deterministic boundaries or explicit nondeterminism. Evidence: chunk seam, ordering, duplicate, worker loss, heterogeneous model/version, and aggregation tests.
    • A6.17.a Plan chunks from immutable media/time base with overlap, keyframe/contact/shot boundaries, context, order, checksums, and deterministic plan identity. Evidence: chunk-plan tests.
    • A6.17.b Register compiled/optimized model artifacts by source model, compiler/runtime/hardware/profile versions, checksum, validation, and compatibility. Evidence: model registry.
    • A6.17.c Reserve compatible multi-GPU or stage-pipeline resources and dispatch chunks/stages with fenced leases, bounded concurrency, and backpressure. Evidence: scheduler tests.
    • A6.17.d Persist idempotent chunk outputs with input/worker/model/tool/ seed provenance, overlap context, raw quality, and attempt lineage. Evidence: duplicate/worker-loss tests.
    • A6.17.e Aggregate in canonical order, reconcile overlap/conflicts with versioned selection policy, and mark heterogeneous/incomparable results. Evidence: aggregation goldens.
    • A6.17.f Assemble final timing, tracks, metadata, and provenance; validate seams/completeness/quality and disclose any nondeterminism. Evidence: final-assembly tests.
    • A6.17.g Recover from chunk/stage/worker/GPU loss without gaps, duplicates, stale commits, or hidden partial success. Evidence: distributed fault suite.

A7 — Console, batch operations, export, and scheduling#

  • A7.1 Build operator hub around intake, queue, stage failures, QA, consent, worker health, cost, and export readiness. Evidence: utility-first design review.
    • A7.1.a Define operator questions, critical states, next actions, role visibility, and drill-down routes before layout. Evidence: task map.
    • A7.1.b Project live intake, queue, failed-stage, QA, consent, worker, cost, and export-readiness summaries from authoritative APIs. Evidence: data contract tests.
    • A7.1.c Compose a restrained task-led workspace with one clear queue/ incident focus and no decorative hero or undifferentiated card mosaic. Evidence: frontend review.
    • A7.1.d Preserve filters/time window/tenant/project context through drill-down and back navigation. Evidence: navigation journey.
    • A7.1.e Render loading, empty, stale, partial, unavailable, and failed summaries honestly. Evidence: state matrix.
  • A7.2 Build intake, job, pipeline, QA, annotation, library, rigs, consent, workers, batch, schedule, and export routes. Evidence: route integrity.
    • A7.2.a Define route IDs, hierarchy, parameters, loaders, permissions, breadcrumbs, titles, and canonical deep links. Evidence: route registry.
    • A7.2.b Implement intake/job/pipeline/QA/annotation routes with exact object and revision context. Evidence: route tests.
    • A7.2.c Implement library/rig/consent routes with protected-data and eligibility boundaries. Evidence: authorization tests.
    • A7.2.d Implement worker/batch/schedule/export routes with operation and step-up controls. Evidence: workflow tests.
    • A7.2.e Verify direct load, refresh, forward/back, not-found, forbidden, signed-link, and permission-change behavior for every route. Evidence: Playwright route suite.
  • A7.3 Build live per-job DAG with exact stage state, attempts, metrics, cache, resource, cost, logs, and output. Evidence: live fault journey.
    • A7.3.a Load immutable job/run definition and subscribe to ordered, resumable state updates with snapshot reconciliation. Evidence: live-data tests.
    • A7.3.b Render stage/dependency state, attempt, worker/resource, cache, progress, cost, timing, and output without inventing missing links. Evidence: DAG fixture.
    • A7.3.c Inspect a selected stage's exact configuration, inputs, logs, measurements, attempts, outputs, failure, and lineage. Evidence: inspector journey.
    • A7.3.d Provide an equivalent ordered/dependency semantic table and keyboard selection instead of requiring graph interaction. Evidence: accessibility parity.
    • A7.3.e Reconcile disconnect, duplicate/out-of-order events, retry, cancellation, worker loss, and late output. Evidence: live fault journey.
  • A7.4 Build stage configuration and rerun-from-stage with impact preview. Evidence: downstream invalidation.
    • A7.4.a Load the versioned stage schema, defaults, current configuration, allowed overrides, and permissions. Evidence: form-schema tests.
    • A7.4.b Validate model/tool/profile/adapter/resource compatibility and show changed versus inherited values before save. Evidence: validation tests.
    • A7.4.c Preview rerun start point, reusable upstream outputs, invalidated downstream stages/cache/reviews/packages, cost, capacity, and consent impact. Evidence: impact fixture.
    • A7.4.d Require explicit authorization of the exact job/run/config/ impact revisions and enqueue idempotently. Evidence: stale-preview and duplicate tests.
    • A7.4.e Track the new run separately and preserve comparison/lineage to every prior attempt. Evidence: rerun journey.
  • A7.5 Build queue and worker console with drain, retry, quarantine, pause, backfill, and kill controls. Evidence: step-up authorization.
    • A7.5.a Project queue depth/age/class/tenant/priority/backpressure and worker identity/capability/version/resource/load/health/lease state. Evidence: console contract.
    • A7.5.b Inspect exact queued/running work, requirements, attempts, reservations, checkpoints, and dispatch eligibility. Evidence: inspection journey.
    • A7.5.c Preview scope and impact for drain, retry, quarantine, pause, resume, reprioritize, backfill, and kill operations. Evidence: impact tests.
    • A7.5.d Enforce operation-specific step-up, reason, concurrency, idempotency, audit, and protected-tenant boundaries. Evidence: authorization suite.
    • A7.5.e Display partial execution, late results, compensation, and read-back rather than optimistic success. Evidence: fault journey.
  • A7.6 Build batch preview with eligible/skipped items, cost, capacity, gates, consent, and schedule. Evidence: mixed-result journey.
    • A7.6.a Resolve a stable item set/query snapshot, requested operation, configuration, target, and caller authority. Evidence: batch request contract.
    • A7.6.b Evaluate per-item compatibility, gate freshness, consent/rights, lifecycle, dependency, quota, and schedule eligibility. Evidence: eligibility fixtures.
    • A7.6.c Show included/skipped/blocked items with exact reasons plus cost, resources, capacity window, duration, and downstream impact. Evidence: preview tests.
    • A7.6.d Authorize the exact preview revision and execute idempotently with per-item progress, retry/cancel, partial result, and audit. Evidence: mixed-result journey.
  • A7.7 Build committed GPU/worker/capture/reviewer windows in the calendar. Evidence: capacity conflict tests.
    • A7.7.a Model resource pools, skills/capabilities, locations, operating windows, maintenance, quotas, and time zones. Evidence: capacity schema.
    • A7.7.b Create tentative holds and committed reservations for GPU, worker, capture, and reviewer resources with exact demand and owner. Evidence: reservation tests.
    • A7.7.c Detect overlap, capability, quota, travel/setup, dependency, and downstream schedule conflicts transactionally. Evidence: conflict fixtures.
    • A7.7.d Reschedule, cancel, expire, or compensate reservations and propagate impact to queued/batch/live work. Evidence: lifecycle journey.
    • A7.7.e Provide list/agenda and keyboard alternatives to drag-based calendar manipulation. Evidence: accessibility tests.
  • A7.8 Build export target selection, mapping, validation, package, watermark, provenance, and read-back verification. Evidence: golden target journeys.
    • A7.8.a Select an exact artifact revision, consumer/target/profile/ version, mapping, package options, watermark, and delivery destination. Evidence: target-selection tests.
    • A7.8.b Explain compatibility, loss, current quality/consent/rights/ release gates, dependencies, costs, and target availability. Evidence: preflight journey.
    • A7.8.c Preview package files/manifest, mappings, generated regions, watermark placement, provenance, and excluded items. Evidence: package preview.
    • A7.8.d Authorize and monitor immutable export assembly under revision/ consent locks with idempotent retry and explicit partial/failure state. Evidence: operation tests.
    • A7.8.e Independently read back the target, compare expected contents, signatures, motion/metadata, and record verified delivery. Evidence: golden target journey.
  • A7.9 Build supersession, revoke, unpublish, quarantine, and recall. Evidence: downstream library/consumer state.
    • A7.9.a Define distinct eligibility, visibility, execution, distribution, and remediation effects for each lifecycle action. Evidence: transition specification.
    • A7.9.b Compute the exact asset/revision, derivatives, packages, libraries, projects, consumers, cached copies, and active jobs affected. Evidence: dependency preview.
    • A7.9.c Require action-specific authority, step-up, reason, effective time, replacement where applicable, and exact impact revision. Evidence: authorization tests.
    • A7.9.d Execute the propagation saga with checkpoints, per-consumer results, retry/compensation, and no hidden partial completion. Evidence: recall drill.
    • A7.9.e Verify downstream state by read-back and keep unresolved recipients visible until closure. Evidence: reconciliation report.
  • A7.10 Build signed deep links from Metis/Studio consumers to authorized Aja job/asset context. Evidence: expiry, tenant, permission, and replay.
    • A7.10.a Define allowed destination routes/context fields and omit protected source, subject, biometric, secret, or authority-bearing data. Evidence: deep-link contract.
    • A7.10.b Mint bounded audience/tenant/object/revision/purpose/expiry/ nonce links with key version and return location. Evidence: signing tests.
    • A7.10.c Validate signature, audience, current authorization, tenant, expiry, nonce/replay, object existence/revision, and lifecycle on open. Evidence: negative suite.
    • A7.10.d Land on an honest forbidden/expired/missing/stale state without leaking object existence and preserve a safe return path. Evidence: Playwright journey.
  • A7.11 Build responsive tablet/mobile incident, review, consent, and approve mode. Evidence: no full 3D edit requirement.
    • A7.11.a Prioritize incident triage, assignment, evidence review, annotation, consent decision, approval/rejection, and emergency control for compact viewports. Evidence: mobile task map.
    • A7.11.b Provide semantic thumbnails, measurements, tables, source-safe media, and deep links so no full 3D editor is required. Evidence: compact review journey.
    • A7.11.c Implement touch/keyboard targets, reflow, zoom, orientation, reduced motion, safe-area, offline/reconnect, and interruption resilience. Evidence: responsive accessibility suite.
    • A7.11.d Enforce step-up, exact revision, reason, confirmation, and read-back for consent, approve/reject, terminate, or recall actions. Evidence: mobile authorization tests.
  • A7.12 Build semantic alternatives and non-drag controls for DAG, timeline, mapping, preview, and graphs. Evidence: WCAG 2.2 journeys.
    • A7.12.a Inventory every visual datum, relationship, selection, and operation in DAG, timeline, mapping, preview, and graph surfaces. Evidence: parity inventory.
    • A7.12.b Provide structured tables/lists/trees with labels, ordering, dependencies, exact values, units, statuses, and diagnostics. Evidence: semantic snapshots.
    • A7.12.c Provide keyboard/button/menu inputs for pan/zoom/select, connect/map, range edit, reorder, resize, and other drag interactions. Evidence: non-drag tests.
    • A7.12.d Synchronize focus, selection, filters, and actions between visual and semantic views. Evidence: parity journey.
    • A7.12.e Test screen-reader announcement, focus recovery, zoom/reflow, color independence, reduced motion, and error recovery. Evidence: WCAG suite.
  • A7.13 Build honest fixture, missing-GPU/model/rig/host, offline, stale, partial, degraded, and failed states. Evidence: state matrix.
    • A7.13.a Define each state, source signal, reason code, precedence, permitted action, retryability, owner, and release effect. Evidence: state contract.
    • A7.13.b Render fixture/demo data with persistent labeling and prohibit production-affecting actions. Evidence: fixture tests.
    • A7.13.c Render each missing dependency and offline/degraded condition with last-known time, affected scope, and safe next step. Evidence: outage matrix.
    • A7.13.d Distinguish stale and partial data from empty/success and expose mixed per-item/per-stage outcomes. Evidence: partial-state tests.
    • A7.13.e Preserve diagnostics and retry/recovery state for failures without showing optimistic completion. Evidence: fault journey.
  • A7.14 Register permission-filtered Studio/control-room navigation and command palette. Evidence: route/role parity.
    • A7.14.a Register canonical route and command IDs, labels, keywords, hierarchy, required capabilities, target resolver, and feature availability. Evidence: navigation registry.
    • A7.14.b Filter server- and client-side navigation, search, recent, shortcuts, and command results by current authorization and context. Evidence: role tests.
    • A7.14.c Reauthorize and validate the resolved object/revision at route load or command execution rather than trusting visibility. Evidence: stale-permission test.
    • A7.14.d Prove every authorized route/action is reachable and every unauthorized route/action is absent and non-executable. Evidence: route/role parity matrix.
  • A7.15 Build searchable job/result operations with filters, saved views, queue statistics, bulk retry/cancel/reprioritize, webhook status, and export diagnostics. Evidence: bulk impact preview, partial-result, authorization, and undo/compensation behavior.
    • A7.15.a Define permission-aware job/result fields, full-text terms, filters/facets, sort, pagination, and bounded queue statistics. Evidence: query contract.
    • A7.15.b Build shareable saved views with owner/scope, versioned query, columns, defaults, permissions, and safe deletion. Evidence: saved-view tests.
    • A7.15.c Expose webhook delivery and export diagnostic projections linked to exact job/result revisions without leaking secrets or protected payloads. Evidence: diagnostic tests.
    • A7.15.d Resolve bulk selection as a stable snapshot and preview per-item retry/cancel/reprioritize eligibility and downstream impact. Evidence: bulk preview.
    • A7.15.e Execute authorized operations idempotently with per-item progress/result, partial failures, retry, and compensation/undo where valid. Evidence: mixed-result journey.
  • A7.16 Build live-capture supervision for device/session health, subject/ consent, calibration, preview, buffer/upload, tracking modalities, quality warnings, handoff, and emergency stop. Evidence: desktop and mobile session fault journeys.
    • A7.16.a Load live device/session identity, operator, location, subject- safe consent state, selected protocol/modalities, and immutable configuration. Evidence: supervision contract.
    • A7.16.b Monitor device/stream/clock/network/buffer/upload/worker health and calibration validity with bounded live updates. Evidence: telemetry tests.
    • A7.16.c Show privacy-governed preview, tracking modality/confidence, dropped data, quality warnings, and unavailable states. Evidence: live preview fixtures.
    • A7.16.d Gate record/start/resume/handoff on current subject consent, calibration, storage/capacity, protocol, and operator authority. Evidence: precondition tests.
    • A7.16.e Implement step-up emergency stop that halts capture/upload/ processing, preserves admissible evidence, and reports late data. Evidence: stop drill.
    • A7.16.f Recover from device/network/app interruption and reconcile buffers, uploads, stages, and handoff without duplicate or hidden gaps. Evidence: desktop/mobile fault journeys.
  • A7.17 Build profile-specific review projections for all ratified vertical pipelines, including intended-use warning, protocol, reference, measures, annotations, qualified reviewer, and consumer handoff. Evidence: role and profile boundaries prevent unsupported medical/performance claims.
    • A7.17.a Define a common projection contract while keeping profile- specific vocabulary, measures, thresholds, warnings, and decisions in their owned versions. Evidence: projection schema.
    • A7.17.b Render the exact profile/protocol/intended and prohibited use, input/reference provenance, measurements, uncertainty, annotations, and validation freshness. Evidence: profile fixtures.
    • A7.17.c Enforce profile-specific qualified-reviewer roles, conflicts, required evidence, decision authority, and escalation. Evidence: role matrix.
    • A7.17.d Produce only the approved consumer handoff fields, disclosures, limitations, and claim boundaries for the exact profile. Evidence: handoff contract tests.
    • A7.17.e Prove yoga, fitness, dance, martial arts, sports, medical/ rehabilitation, film/VFX, and gaming projections cannot inherit one another's claims, review, or approval. Evidence: cross-profile negative suite.

A8 — Aja observability, verification, and launch#

  • A8.1 Instrument capture, upload, queue, stage, cache, validation, review, export, adapter, consent, and recall traces/metrics. Evidence: bounded schema.
    • A8.1.a Define bounded event/span/metric names, attributes, units, cardinality limits, sampling, retention, and ownership per subsystem. Evidence: telemetry registry.
    • A8.1.b Propagate tenant-safe trace/run/job/stage/artifact correlation across capture, services, queues, workers, adapters, and consumers. Evidence: traced fixture.
    • A8.1.c Instrument success, latency, load, retries, backpressure, partial results, typed failures, validation/review, consent blocks, and recall closure. Evidence: signal tests.
    • A8.1.d Redact raw media, biometric/subject data, secrets, signed URLs, free text, and unbounded identifiers from telemetry. Evidence: privacy scan.
  • A8.2 Define SLOs for ingest, queue age, stage runtime, worker health, validation, preview, export, adapter health, and consent enforcement. Evidence: error-budget policy.
    • A8.2.a Define user-visible service indicators, populations, exclusions, windows, data sources, and measurement quality for each named surface. Evidence: SLI catalog.
    • A8.2.b Ratify latency, availability, freshness, correctness, durability, and enforcement objectives by profile/tier where genuinely distinct. Evidence: approved targets.
    • A8.2.c Create dashboards and multi-window burn alerts with actionable ownership and protected-data-safe drill-down. Evidence: alert tests.
    • A8.2.d Define error-budget consumption, release freeze, exception, recovery, and retrospective policy. Evidence: policy exercise.
  • A8.3 Run contract and migration compatibility suites. Evidence: input, rig, motion, consent, package, and adapter fixtures.
    • A8.3.a Test supported prior/current API, event, worker, adapter, file, and package contract versions against consumer fixtures. Evidence: compatibility matrix.
    • A8.3.b Test input, rig, motion, consent, validation, lineage, and package schema round-trips including unknown/optional fields. Evidence: contract goldens.
    • A8.3.c Test migration upgrade, backfill, mixed-version operation, rollback/roll-forward, and cleanup at representative scale. Evidence: migration report.
    • A8.3.d Block undocumented breaking drift and expired compatibility windows. Evidence: release gate.
  • A8.4 Run known-value calibration, sync, skeleton, frame, contact, slide, jitter, anatomy, ground-truth, retarget, loop, and format tests. Evidence: gate registry coverage.
    • A8.4.a Register independent analytical or externally verified fixtures with source, generator, expected values/tolerances, rights, and checksum. Evidence: fixture registry.
    • A8.4.b Cover calibration/sync/skeleton/frame and contact/slide/jitter measures with positive, boundary, and broken cases. Evidence: known- value report.
    • A8.4.c Cover anatomy/physics, ground-truth fidelity, retarget, loop/ blend, facial where consented, and format round-trip separately. Evidence: gate matrix.
    • A8.4.d Prove every production gate/profile/version maps to current known-value evidence and unavailable gates cannot pass. Evidence: registry audit.
  • A8.5 Run calibrated perceptual/model gate evaluations. Evidence: threshold, subgroup, uncertainty, and drift report.
    • A8.5.a Freeze evaluator/model/config, consented evaluation corpus, labels/adjudication, splits, cohorts, and baselines. Evidence: eval manifest.
    • A8.5.b Measure task performance, calibration, uncertainty, abstention, subgroup error, out-of-distribution behavior, and human disagreement. Evidence: evaluation report.
    • A8.5.c Select pass/fail/review thresholds from declared costs and prove them on held-out data without tuning leakage. Evidence: threshold rationale.
    • A8.5.d Compare candidate to current model, replay affected historical cases, monitor drift, and document limitations/expiry. Evidence: model release gate.
  • A8.6 Run real-database, blob, queue, worker, cache, search, preview, and export integration tests. Evidence: durable round-trip.
    • A8.6.a Stand up production-like database, object store, queue, workers, cache, search, and target adapters with isolated fixtures. Evidence: integration environment.
    • A8.6.b Exercise capture/ingest through persisted job/stages/artifacts, cache/index, preview, validation, package/export, and target read-back. Evidence: durable round-trip.
    • A8.6.c Exercise retries, duplicate delivery, concurrency, stale cache/ index, restart, partial result, and cleanup. Evidence: fault cases.
    • A8.6.d Verify tenant/consent isolation, checksums, lineage, audit, and reconciliation at every storage boundary. Evidence: invariant report.
  • A8.7 Run full-journey Playwright: ingest/consent through validated export. Evidence: deterministic fixture clip and read-back.
    • A8.7.a Create a deterministic rights-cleared clip, capture metadata, subject consent, rig/profile, known gate values, and target fixture. Evidence: fixture manifest.
    • A8.7.b Ingest, validate admissibility, configure/run stages, and inspect synchronized preview/lineage through the browser. Evidence: Playwright trace.
    • A8.7.c Review raw measurements, annotations, consent, and current gates; approve the exact eligible revision. Evidence: decision assertions.
    • A8.7.d Export/package and independently read back target bytes/metadata/ motion/provenance, asserting UI and API final state. Evidence: journey report.
  • A8.8 Run full-journey Playwright: failed stage, rerun, compare, annotate, approve, package. Evidence: exact state.
    • A8.8.a Inject a deterministic stage failure and assert job DAG, attempt, diagnostics, downstream block, and non-success announcement. Evidence: failure trace.
    • A8.8.b Change a permitted configuration, preview invalidation/cost, and rerun from the failed stage. Evidence: rerun assertions.
    • A8.8.c Compare exact old/new revisions, create durable frame/joint annotations, and resolve review. Evidence: comparison trace.
    • A8.8.d Approve/package only the new eligible revision and assert the failed/stale output cannot release. Evidence: exact-state report.
  • A8.9 Run full-journey adapter tests for Metis demonstration, coaching overlay, and handoff. Evidence: health/degraded behavior.
    • A8.9.a Negotiate adapter version/capabilities and test healthy, degraded, unavailable, incompatible, and stale health projections. Evidence: health matrix.
    • A8.9.b Request and consume an eligible demonstration with consent, rights, quality, compatibility, provenance, and revocation assertions. Evidence: demonstration journey.
    • A8.9.c Request a coaching overlay and assert bounded semantic measures, uncertainty, claim limitations, and protected-data exclusion. Evidence: overlay journey.
    • A8.9.d Submit/status/cancel/complete an asynchronous handoff with idempotency, correlation, retry, late-result, and failure behavior. Evidence: handoff journey.
  • A8.10 Run WCAG 2.2 AA, keyboard, announcement, non-drag, and semantic preview parity suites. Evidence: report matrix.
    • A8.10.a Run automated WCAG checks across authenticated routes, roles, themes, locales, and wide/compact/single-column viewports. Evidence: accessibility report.
    • A8.10.b Complete keyboard-only intake, job/DAG, compare, mapping, annotation, review, batch, schedule, and export journeys. Evidence: keyboard traces.
    • A8.10.c Verify names/roles/values, focus order/restoration, live announcements, errors, loading, stale/partial/failure, and step-up dialogs. Evidence: assistive-technology matrix.
    • A8.10.d Verify non-drag controls, 200–400% zoom/reflow, touch targets, color independence, reduced motion, and semantic 3D/graph parity. Evidence: manual/automated results.
  • A8.11 Run upload, queue, batch, worker, preview, large motion, search, and export performance/load tests. Evidence: backpressure and budgets.
    • A8.11.a Define representative and worst-supported datasets, concurrency, resource limits, warm/cold state, budgets, and success/failure criteria. Evidence: load plan.
    • A8.11.b Measure upload/ingest, queue fairness/age, dispatch, stage/ worker, cache, and batch behavior under sustained and burst load. Evidence: service report.
    • A8.11.c Measure first/useful preview, timeline/skeleton/large-motion interaction, search, and export/package/read-back performance. Evidence: user-performance report.
    • A8.11.d Verify quotas, backpressure, degradation, cancellation, recovery, data correctness, cost, and no tenant starvation. Evidence: overload report.
  • A8.12 Run malicious media/archive/rig, consent bypass, biometric leak, model attack, worker spoof, cache poison, auth, tenant, and export tamper tests. Evidence: zero blockers.
    • A8.12.a Test malformed/oversized/polyglot media, archives, rigs, paths, metadata, parsers, decompression, and dependency resolution in quarantine. Evidence: content security suite.
    • A8.12.b Test consent/rights expiry, revocation races, purpose confusion, protected preview/search/log leaks, and biometric inference/exfiltration. Evidence: privacy suite.
    • A8.12.c Test adversarial model inputs/outputs, prompt/reference abuse, unsafe deserialization, model substitution, and generated-content disclosure bypass. Evidence: model security report.
    • A8.12.d Test worker identity/capability spoof, lease theft, stale commit, cache poison, artifact tamper, and supply-chain compromise. Evidence: compute security suite.
    • A8.12.e Test authentication, object/property authorization, tenant/ project isolation, step-up, export/package/signature, and recall bypass. Evidence: adversarial matrix with zero release blockers.
  • A8.13 Run worker loss, GPU loss, queue replay, cache loss, backup/restore, consent revocation, and recall drills. Evidence: RPO/RTO.
    • A8.13.a Inject worker process/host/network and GPU reset/exhaustion loss during queued, running, checkpointing, and committing states. Evidence: compute recovery report.
    • A8.13.b Replay, duplicate, reorder, pause, and partially lose queue/ event delivery; verify fencing, idempotency, and reconciliation. Evidence: messaging drill.
    • A8.13.c Lose/corrupt cache entries and prove safe miss, quarantine, recomputation, and no invalid release. Evidence: cache drill.
    • A8.13.d Restore database/blobs/config/keys/indexes from backup and reconcile checksums, queues, jobs, search, audit, and targets. Evidence: restore report.
    • A8.13.e Revoke consent and recall during processing/export/distribution; verify bounded stop and downstream closure. Evidence: recall drill.
    • A8.13.f Measure and compare achieved RPO/RTO/data loss/duplicate work to ratified objectives. Evidence: recovery scorecard.
  • A8.14 Publish capture, consent, pipeline, worker, QA, export, incident, privacy, and recovery runbooks. Evidence: game-day.
    • A8.14.a Assign owner, audience, prerequisites, triggers, authority, commands/UI paths, diagnostics, safety checks, escalation, and rollback to each runbook. Evidence: runbook template audit.
    • A8.14.b Write capture/consent and pipeline/worker operational runbooks with protected-data and emergency-stop handling. Evidence: reviewed docs.
    • A8.14.c Write QA/export and privacy/security incident runbooks with gate, quarantine, revocation, and recall procedures. Evidence: reviewed docs.
    • A8.14.d Write backup/restore/disaster recovery and dependency-outage runbooks with RPO/RTO validation. Evidence: reviewed docs.
    • A8.14.e Execute representative game-days, record deviations, remediate, and revalidate runbook commands/links. Evidence: exercise report.
  • A8.15 Produce engineering-readiness evidence. Evidence: all requirements map to proof.
    • A8.15.a Build a bidirectional requirement-to-design/code/test/operation/ owner/evidence index with immutable evidence refs. Evidence: zero- orphan audit.
    • A8.15.b Record contract/migration, quality/calibration, security/privacy, accessibility, performance/SLO, recovery, and runbook results with exact versions. Evidence: readiness packet.
    • A8.15.c Surface failed, stale, waived, unavailable, or missing evidence explicitly and block readiness on unresolved mandatory proof. Evidence: readiness gate.
    • A8.15.d Reproduce the packet from repository and approved external evidence and verify links/checksums. Evidence: clean-room generation.
  • A8.16 Obtain named motion/animation, privacy/consent, safety, accessibility, security, engineering, and product approvals. Evidence: human signoff.
    • A8.16.a Define each approver's scope, competence, independence/conflict rules, required evidence, decision options, and expiry. Evidence: approval policy.
    • A8.16.b Present the exact release candidate, readiness packet, limitations, risks, waivers, and unresolved items. Evidence: review bundle.
    • A8.16.c Capture named approve/reject/conditional decisions with role, time, revision, rationale, conditions, and signature/audit. Evidence: signoff records.
    • A8.16.d Invalidate or renew approvals when candidate, evidence, policy, model/profile, consent posture, or material risk changes. Evidence: staleness test.
  • A8.17 Run distinct full journeys for desktop live capture, mobile live capture, multi-person/body/hand/face tracking, video analysis, generative cleanup/blending, format conversion, SDK/CLI, and webhooks. Evidence: no generic batch-ingest journey substitutes for live or modality-specific proof.
    • A8.17.a Run separate desktop and mobile live-capture journeys through device enrollment, calibration, consent, interruption, handoff, and emergency stop. Evidence: live traces.
    • A8.17.b Run separate multi-person, body, hand, and consented face tracking journeys with identity, occlusion, confidence, modality loss, and protected-data assertions. Evidence: modality matrix.
    • A8.17.c Run video-analysis and generative cleanup/blending journeys with exact edited/generated ranges, alternatives, validation, and human acceptance. Evidence: provenance traces.
    • A8.17.d Run every supported format-conversion pair through loss preview, conversion, validation, package, and read-back. Evidence: conversion matrix.
    • A8.17.e Run external TypeScript SDK, Python SDK, CLI, and webhook journeys including auth, version negotiation, retry, partial failure, secret rotation, and recall. Evidence: client matrix.
  • A8.18 Run independent contract, known-value, calibrated-quality, accessibility, privacy/consent, performance, and consumer-handoff gates for yoga, fitness, dance, martial arts, sports, medical/rehabilitation, film/VFX, and gaming profiles. Evidence: no vertical inherits another vertical's release evidence.
    • A8.18.a Register one release-evidence bundle per exact vertical profile/ version with owners, intended/prohibited use, fixtures, consumers, and expiry. Evidence: profile registry.
    • A8.18.b Run profile-specific contract and independent known-value/ calibrated-quality gates against its own protocol, measures, thresholds, and reviewer population. Evidence: quality matrix.
    • A8.18.c Run profile-specific accessibility, privacy/consent, safety/ claim-boundary, performance, and consumer-handoff tests. Evidence: assurance matrix.
    • A8.18.d Cover yoga, fitness, dance, martial arts, each ratified sport, medical/rehabilitation, film/VFX, and gaming with declared unavailable cells rather than inherited proof. Evidence: coverage audit.
    • A8.18.e Block profile release when any required bundle is missing, stale, failed, or belongs to another profile/version. Evidence: independence test.
  • A8.19 Run full journeys for CG replacement, text/reference-to-motion, avatar customization/marketplace, chunked/parallel inference, optimized model rollout, distributed aggregation, and final package assembly. Evidence: specialized lineage, consent, compatibility, failure, and quality assertions.
    • A8.19.a Run video-to-CG per-shot segmentation/tracking/transfer/render/ composite/review/delivery with consent, generated-region, temporal quality, seam, and recall assertions. Evidence: CG journey.
    • A8.19.b Run text and protected-reference-to-motion through constraints, candidate generation, physics/quality validation, human acceptance, export, and disclosure. Evidence: generation journeys.
    • A8.19.c Run avatar customization and marketplace submit/scan/review/ promote/use/update/quarantine/takedown with compatibility and rights propagation. Evidence: marketplace journey.
    • A8.19.d Run chunked multi-GPU/stage-parallel inference through overlap, duplicate, ordering, worker/GPU loss, retry, and deterministic or disclosed-nondeterministic aggregation. Evidence: distributed trace.
    • A8.19.e Roll out a compiled/optimized model through compatibility, known-value/calibrated comparison, canary, rollback, and historical result provenance. Evidence: model rollout journey.
    • A8.19.f Assemble, validate, sign, publish, and independently read back the final package while preserving all specialized lineage, consent, rights, quality, compatibility, and failure evidence. Evidence: package journey.
  • A8.20 EVE-SEAM-EXIT: require "required Eve reads and writes admitted with source and behavior evidence, or an exact task-0.8 explicit non-goal proves zero operations; a named defer preserves ownership but never passes completion" before this phase exits. Evidence: source-derived totality record and passing verifier.
    • A8.20.a Re-run the route, bridge, state-signal, and complete domain-code envelope census; review every changed candidate instead of carrying a stale phase disposition forward. Evidence: current ratchets.
    • A8.20.b Bind the exact ratified task-0.8 workflow and requirement set; classify it as a required seam or an exact source-backed non-goal. Evidence: charter-bound disposition set.
    • A8.20.c For a required seam, admit both read and write operations with byte-pinned source and behavior evidence. A named defer retains its owner and unblock condition but fails completion. Evidence: operation proofs.
    • A8.20.d Permit a zero-operation boundary only for an exact ratified task-0.8 explicit non-goal; absence or unavailability is not a non-goal. Evidence: source-bound non-goal proof.
    • A8.20.e Run the totality verifier, every retained gate-case CLI, and the injected durable-candidate control; retain red and green receipts. Evidence: release gate logs.

Phase B — Bellona DCC bridge and interchange console#

Bellona is a Studio-only substrate with an existing apps/bellona/control-room. The work is to unify that control room with the bridge hosts, build/render services, OpenUSD interchange, consistency checks, artifact store, release evidence, and governed Studio entry point without creating a duplicate console.

B0 — Host, bridge matrix, ownership, and threat model#

  • B0.1 Ratify apps/bellona/control-room as the canonical operator console or record the replacement and migration. Evidence: host ADR.
    • B0.1.a Inventory current Bellona control-room ownership, users, routes, mutations, deployments, data sources, and operational dependencies. Evidence: current-state map.
    • B0.1.b Compare retain/refactor and replacement options against authority, migration, accessibility, operations, cost, and duplication criteria. Evidence: option record.
    • B0.1.c Ratify one canonical console and its governed Studio projection, with named owner and prohibited duplicate surfaces. Evidence: approved ADR.
    • B0.1.d If replaced, define route/data/permission/operation/test rollout, coexistence, rollback, and retirement milestones. Evidence: migration plan.
  • B0.2 Inventory control-room routes/components/tests, static Studio routes, bridge apps, gateways, hosts, build/render services, CLI, and libraries. Evidence: route-capability matrix. <!-- partial: UI (a) + runtime (b) + code (c) inventories done in evidence/v1-workbenches/bellona-inventory.json. d (capability/disposition mapping) is design/human and stays unchecked. -->
    • B0.2.a Enumerate control-room and Studio routes, components, loaders, actions, tests, flags, permissions, and data dependencies. Evidence: UI inventory. <!-- evidence: .ui — the control-room React-Router routes (4) + module/ component counts + 82 /studio/bellona/* Studio pages. generate-bellona- inventory.mjs + 5 tests. -->
    • B0.2.b Enumerate bridge/remote/gateway/host/build/render apps and their protocols, commands, state, deployment, and owner. Evidence: runtime inventory. <!-- evidence: .runtimeApps — all 10 runtime apps classified by kind (dcc-bridge×4, remote×3, build×2, render×1) AND transport protocol (websocket/grpc/http/ipc): the DCC bridges (blender/godot/unity/unreal) speak WebSocket, not HTTP routes. deployment/owner are the semantic half. -->
    • B0.2.c Enumerate CLI commands and Bellona libraries with public contracts, consumers, persistence, jobs/events, and test coverage. Evidence: code inventory. <!-- evidence: .code — 6 CLI commands + 41 libs/bellona/* packages. Per-lib contracts/consumers already in the BASE.3 library graph; this is the Bellona code-surface count. -->
    • B0.2.d Map every discovered item to one capability/disposition and flag duplicates, orphans, stubs, fixtures, and dead paths. Evidence: route-capability matrix.
  • B0.3 Map all libs/bellona/* packages to console capabilities, bridge runtime, SDK, consumer integration, or out-of-scope. Evidence: ledger.
    • B0.3.a Enumerate each package's entry points, exports, schemas, runtime side effects, dependencies, consumers, owners, and tests. Evidence: <!-- evidence: .domains.bellona.packages — 41 libs/bellona/* packages with entry points/exports/dependencies/consumers/tests/disposition. runtime side effects + owners are B0.3.b classification (semantic). evidence/v1-workbenches/domain-package-inventories.json (254 pkgs; per-package exports/consumers/deps/disposition/tests + capability catalog with exact source locations). --> package census.
    • B0.3.b Classify every public symbol by canonical console, bridge runtime, SDK, consumer seam, shared substrate, migration-only, or out-of-scope ownership. Evidence: symbol ledger.
    • B0.3.c Resolve overlapping implementations and undocumented cross- domain authority. Evidence: disposition decisions.
    • B0.3.d Verify build/dependency graph and consumer imports agree with approved boundaries. Evidence: boundary tests.
  • B0.4 Define the governed Oshun Studio entry/facade and deep-link boundary. Evidence: Bellona never becomes a customer-domain route.
    • B0.4.a Define the Studio audience, permitted summary fields/actions, canonical Bellona destinations, and prohibited bridge/admin detail. Evidence: facade contract.
    • B0.4.b Define signed link context, audience, tenant/project, object/ revision, expiry, nonce, and return behavior. Evidence: link schema.
    • B0.4.c Reauthorize at Bellona route load and prevent the Studio facade from proxying canonical mutation authority. Evidence: boundary tests.
    • B0.4.d Verify navigation, route ownership, analytics, support docs, and command palette never expose Bellona as a customer-domain route. Evidence: route audit.
  • B0.5 Define operator, DCC artist, technical artist, pipeline engineer, build engineer, reviewer, release manager, security operator, and administrator roles. Evidence: permission matrix.
    • B0.5.a Define job duties, least-privilege capabilities, tenant/project/ host scope, separation-of-duties, and prohibited actions for each role. Evidence: role definitions.
    • B0.5.b Map view, command, transfer, configure, approve, waive, release, revoke, quarantine, and emergency actions to object/property authority. Evidence: permission matrix.
    • B0.5.c Define step-up, dual control, break-glass, conflict, expiry, and delegation rules for privileged operations. Evidence: control policy.
    • B0.5.d Test every role across console, API, bridge, CLI/SDK, Studio facade, and in-host UI. Evidence: cross-surface authorization suite.
  • B0.6 Define supported DCC/engine hosts and versions for Blender, Maya, Houdini, 3ds Max, Unreal, Unity, Godot, and ratified extensions. Evidence: compatibility matrix.
    • B0.6.a Define host/plugin/OS/architecture/server/protocol version tuple, support tier, lifecycle, owner, and test environment. Evidence: host registry.
    • B0.6.b Ratify Blender, Maya, Houdini, 3ds Max, Unreal, Unity, and Godot tuples independently with install/update and vendor-support constraints. Evidence: approved rows.
    • B0.6.c Register extensions only with capability, trust, packaging, compatibility, security, and release proof separate from the base host. Evidence: extension records.
    • B0.6.d Define compatibility windows, deprecation/expiry, unavailable cells, and upgrade/downgrade behavior. Evidence: lifecycle matrix.
  • B0.7 Define source and target formats, OpenUSD profiles, schemas, resolver rules, units, axes, color, material, animation, rig, and package expectations. Evidence: interchange profile.
    • B0.7.a Register each source/target format/profile/version, reader/ writer, host support, limits, and unsupported constructs. Evidence: format registry.
    • B0.7.b Ratify OpenUSD core/extensions, schema/API versions, layer/ composition policy, resolver contexts, package layout, and validation. Evidence: USD profile.
    • B0.7.c Define geometry/topology, transforms/units/axes, color, material/ texture, animation/time, rig/skinning, camera/light, metadata, and dependency mappings. Evidence: semantic mapping spec.
    • B0.7.d Define deterministic handling and disclosure for unknown, lossy, approximated, baked, remapped, or synthesized features. Evidence: loss policy.
  • B0.8 Define local, remote, cloud, and on-box UE host trust boundaries. Evidence: network/auth architecture.
    • B0.8.a Diagram actors, processes, devices, networks, stores, secrets, data/command flows, and tenancy for each deployment mode. Evidence: trust diagrams.
    • B0.8.b Define host/device identity, enrollment, attestation, mutual authentication, authorization grants, key storage/rotation, and revoke. Evidence: identity architecture.
    • B0.8.c Define inbound/outbound network policy, discovery, relays, firewall/proxy/NAT, SSRF controls, and offline behavior. Evidence: network architecture.
    • B0.8.d Define additional isolation and command/file safeguards for remote/cloud and on-box Unreal execution. Evidence: deployment threat review.
  • B0.9 Define artifact retention, cache, legal hold, licensing, provenance, release evidence, and deletion policy. Evidence: data lifecycle.
    • B0.9.a Classify source, intermediate, cache, log, preview, build, package, release, and evidence artifacts by sensitivity and authority. Evidence: data catalog.
    • B0.9.b Define retention, expiry, cache eviction, archive, deletion, tombstone, backup, and restoration per class. Evidence: lifecycle schedule.
    • B0.9.c Define legal hold and licensing/redistribution constraints, conflicts, and authorized release/deletion behavior. Evidence: policy matrix.
    • B0.9.d Define immutable source/transformation/dependency/tool/host/ operator provenance and evidence retention beyond artifact deletion. Evidence: provenance contract.
    • B0.9.e Prove expiry, deletion, hold, quarantine, recall, and read-back reconciliation across stores/caches/targets. Evidence: lifecycle tests.
  • B0.10 Threat-model remote command injection, bridge impersonation, compromised plugins/hosts, malicious assets/USD, path traversal, SSRF, credential theft, artifact tamper, supply chain, and build exfiltration. Evidence: mitigation/test map.
    • B0.10.a Enumerate assets, actors, entry points, trust transitions, attacker capabilities, abuse cases, and consequences per deployment/ host profile. Evidence: threat inventory.
    • B0.10.b Model command injection/confusion, bridge/host/plugin identity spoofing, stale grants, and compromised execution. Evidence: command threat scenarios.
    • B0.10.c Model malicious files/USD/archives/dependencies, parser abuse, path/resolver traversal, SSRF, and artifact tamper. Evidence: content threat scenarios.
    • B0.10.d Model credential/secret theft, supply-chain substitution, worker/build compromise, cross-tenant access, and build exfiltration. Evidence: platform threat scenarios.
    • B0.10.e Map each threat to preventive/detective/recovery control, owner, residual risk, adversarial test, and release gate. Evidence: approved mitigation matrix.
  • B0.11 Define emergency controls for host revoke, session terminate, command stop, artifact quarantine, template revoke, and release recall. Evidence: authority/blast-radius matrix.
    • B0.11.a Define trigger, authority, step-up/dual control, scope selector, confirmation, and immutable audit for each emergency action. Evidence: control contracts.
    • B0.11.b Compute active sessions/commands/transfers/jobs/builds, artifacts, templates, releases, targets, and consumers in blast radius. Evidence: impact preview.
    • B0.11.c Specify immediate stop/fencing behavior, late-result handling, quarantine, downstream propagation, and degraded operation. Evidence: response design.
    • B0.11.d Define recovery/re-enrollment/unquarantine/re-release authority and proof requirements. Evidence: recovery policy.
    • B0.11.e Exercise each control and verify target/consumer read-back to closure. Evidence: emergency drill.
  • B0.12 Define success measures for connection, transfer, conversion, fidelity, queue, build, cache, host availability, cost, and release verification. Evidence: metric spec.
    • B0.12.a Define each measure's user outcome, numerator/denominator, population, unit, source, window, exclusions, and owner. Evidence: metric catalog.
    • B0.12.b Define latency/reliability/correctness/fidelity/capacity/cost targets by supported host/profile/tier where warranted. Evidence: target matrix.
    • B0.12.c Define bounded privacy-safe dimensions and prohibit secret, asset-content, path, free-text, and high-cardinality labels. Evidence: telemetry policy.
    • B0.12.d Link measures to dashboards, alerts, error budgets, release gates, and corrective decisions. Evidence: operational trace.
  • B0.13 Author and ratify V1/BELLONA_DCC_CONSOLE_DESIGN_2026-07-23.md with current-state, host, bridge trust, OpenUSD spine, contracts, data, APIs, fleet, UI, release, migration, rollout, and decision records. Evidence: approved doc and resolved review comments.
    • B0.13.a Draft current-state findings, scope/boundaries, personas/ permissions, host/version matrix, and trust/threat architecture. Evidence: reviewed design sections.
    • B0.13.b Draft OpenUSD/interchange, canonical data, protocol/API/events, persistence/artifact, queue/fleet, and consumer contracts. Evidence: reviewed technical sections.
    • B0.13.c Draft task-led console/in-host/Studio experiences, accessibility, honest states, operations, verification, and release. Evidence: reviewed experience sections.
    • B0.13.d Draft migration/compatibility/rollout/rollback/retirement, risks, open decisions, ownership, and milestones. Evidence: reviewed delivery sections.
    • B0.13.e Obtain named cross-functional review, resolve every comment or record a decision, and ratify an immutable version. Evidence: approval log.
  • B0.14 Reconcile every Bellona control-room/Studio route, bridge/remote/ build/render app, CLI/SDK command, libs/bellona/* package, protocol, host plugin, workflow, test, and documented feature at atomic-feature level. Evidence: explicit dispositions cover bridges/interchange/builds, render, gameplay, virtual production, XR, mocap, MetaHuman, Resolve/ Houdini, audio/video, text-to-3D/creative flows, agent/MCP control, cross-domain ingestion, events, and release evidence.
    • B0.14.a Build a normalized inventory of every named code, route, contract, command, feature claim, workflow, test, and release artifact with stable IDs and owners. Evidence: atomic inventory. <!-- bellona-code-census.json: 13 artifact classes, each with a stable ID (bellona.pkg/app/code/test/route.*/runtime-app/protocol/cli/contract/ workflow/doc/deploy) + package-derived owner. 41 libs packages + 23 app projects, 952 code files + 630 tests, 4 control-room + 82 studio routes, 10 runtime apps (dcc-bridge 4/remote 3/build 2/render 1) over grpc/http/ ipc/websocket, 6 CLI commands (bellona/export/list/query/start/submit), 8 contract/schema files, 86 workflow/plugin/bridge files, 51 docs, and domain-specific deploy targets (unreal/blender/remote-mac/cloud-agent smokes). Consolidated from fresh git counts + committed B0.2 inventory. generate-bellona-code-census.mjs + 6 tests, wired into inventory/verify + CI. NAMED human owners, atomic-feature decomposition (B0.14.b), dispositions (B0.14.c), zero-orphan reconciliation (B0.14.d), and signed owner review (B0.14.e) are semantic/human-governance and stay unchecked; parent B0.14 stays [ ]. -->
    • B0.14.b Decompose compound claims into independently classifiable host, interchange, build, render, runtime, VP, XR, mocap, character, editorial/procedural, media, generative, agent, integration, event, and evidence capabilities. Evidence: feature ledger.
    • B0.14.c Assign each item retain/refactor/migrate/replace/merge/remove/ external-owner/defer decisions with rationale, destination, dependency, proof, and target milestone. Evidence: disposition audit.
    • B0.14.d Map inventory to ledger and ledger back to authoritative or explicitly net-new scope; resolve duplicates/orphans. Evidence: bidirectional zero-orphan check.
    • B0.14.e Obtain domain-owner review of every unresolved, conflicting, or high-risk disposition. Evidence: signed reconciliation.
  • B0.15 Define the boundary among control-room operations, in-host artist plugin UI, governed Studio facade, remote/cloud administration, build/release automation, virtual-production stage control, gameplay/XR validation, and developer SDK/CLI. Evidence: every capability has one authoritative mutation surface and consistent projections elsewhere.
    • B0.15.a Catalog read and mutation capabilities for each surface and identify duplicated or missing authority. Evidence: surface matrix.
    • B0.15.b Assign one authoritative mutation surface/service per capability with role, context, audit, and consistency contract. Evidence: authority registry.
    • B0.15.c Define read-only projections, signed deep links, request/ approval seams, and event reconciliation for non-authoritative surfaces. Evidence: projection contracts.
    • B0.15.d Test conflicting concurrent actions, stale projections, permission changes, offline hosts, and partial propagation across surfaces. Evidence: boundary suite.
  • B0.16 Ratify specialized compatibility profiles for asset interchange, engine project/runtime, render, virtual production, XR device, mocap, MetaHuman, Resolve editorial, Houdini procedural, and cross-domain compilation. Evidence: each profile names hosts/versions, commands, assets, measurements, gates, safety controls, and release targets.
    • B0.16.a Define the common profile schema and lifecycle for exact hosts/ versions, commands, assets, mappings, measurements, validators/gates, safety controls, targets, owners, evidence, and limitations. Evidence: schema tests.
    • B0.16.b Ratify separate asset-interchange, engine-runtime, and render profiles with no inherited target proof. Evidence: profile records.
    • B0.16.c Ratify separate virtual-production, XR-device, mocap, and MetaHuman profiles with device/stage/performer safety boundaries. Evidence: profile records.
    • B0.16.d Ratify separate Resolve editorial, Houdini procedural, and cross-domain compilation profiles with source/target authority and deterministic evidence rules. Evidence: profile records.
    • B0.16.e Reject execution or release under missing, stale, incompatible, or cross-profile evidence. Evidence: independence suite.

B1 — Session, command, artifact, interchange, and release contracts#

  • B1.1 Define host registration with DCC/engine, version, plugin, platform, capabilities, identity, attestation, health, load, and maintenance. Evidence: schema.
    • B1.1.a Define immutable host/device identity, owner/tenant, enrollment, certificate/key version, attestation, and revocation fields. Evidence: identity schema.
    • B1.1.b Define host/DCC/engine/plugin/OS/architecture/protocol versions and typed capability/limit declarations. Evidence: compatibility schema.
    • B1.1.c Define heartbeat time, health components, resource/load/license state, network locality, and freshness semantics. Evidence: liveness schema.
    • B1.1.d Define maintenance/drain/quarantine/disabled lifecycle, concurrency rules, and audit invariants. Evidence: state tests.
  • B1.2 Define bridge session with host, actor, project, workspace, purpose, grants, expiry, heartbeat, status, and termination. Evidence: invariants.
    • B1.2.a Define session identity, enrolled host revision, authenticated actor, tenant/project/workspace, client, purpose, and trace. Evidence: session schema.
    • B1.2.b Define least-privilege typed grants with resource/path scope, command capabilities, limits, issue/expiry/revoke times, and approver. Evidence: grant schema.
    • B1.2.c Define connect/active/degraded/draining/expired/terminating/ terminated states, heartbeat, reconnect, and concurrent-session rules. Evidence: state machine.
    • B1.2.d Define user, host, administrator, timeout, revoke, and emergency termination semantics with fenced late commands. Evidence: invariants.
  • B1.3 Define command envelope with version, capability, target, arguments, idempotency, deadline, expected revision, authorization, and correlation. Evidence: contract tests.
    • B1.3.a Define command/protocol version, registry capability/action, session/host/project target, and typed argument schema reference. Evidence: envelope schema.
    • B1.3.b Define command/idempotency/correlation/causation IDs, requested time, deadline, priority, and retry budget. Evidence: identity tests.
    • B1.3.c Define expected object/scene/file revisions, preconditions, resource/path scopes, dry-run/impact-preview reference, and actor grant. Evidence: concurrency contract.
    • B1.3.d Sign/authenticate the immutable envelope and reject unknown, malformed, expired, replayed, stale, or unauthorized commands. Evidence: negative contract suite.
  • B1.4 Define command result with status, output refs, diagnostics, logs, measurements, partial state, and retry semantics. Evidence: schema.
    • B1.4.a Define accepted/running/succeeded/failed/cancelled/expired/ rejected/partial/unknown statuses and legal transitions. Evidence: result state machine.
    • B1.4.b Define attempt/host/tool versions, start/end, progress, checkpoints, output artifact refs/checksums, changed-object revisions, and target read-back. Evidence: result schema.
    • B1.4.c Define typed diagnostics, bounded redacted logs, raw measurements, warnings, limitations, and unsupported-feature detail. Evidence: diagnostic fixtures.
    • B1.4.d Define retryability, safe retry point, compensation/undo, irreversible side effects, partial outputs, and late-result semantics. Evidence: fault contract tests.
  • B1.5 Define interchange job and stage DAG from source ingest through canonical USD and targets. Evidence: DAG validation.
    • B1.5.a Define immutable job identity, source artifact/revision/profile, requested targets, configuration, priority, deadline, and actor. Evidence: job schema.
    • B1.5.b Define typed ingest/inspect/normalize/USD-author/resolve/validate/ convert/package/read-back stages with inputs, outputs, requirements, and failure policy. Evidence: stage registry.
    • B1.5.c Validate dependency acyclicity, input/output compatibility, required stages, target/profile coverage, and terminal gates. Evidence: DAG validator.
    • B1.5.d Define attempts, checkpoints, cache keys, rerun/invalidation, cancellation, partial target results, and immutable lineage. Evidence: execution invariants.
  • B1.6 Define artifact, representation, version, layer, dependency, resolver context, checksum, size, rights, provenance, and retention. Evidence: schema.
    • B1.6.a Define logical artifact identity, immutable version, asset type, lifecycle, owner/tenant/project, rights/license, sensitivity, and retention class. Evidence: artifact schema.
    • B1.6.b Define each representation's format/profile, blob ref, checksum, size, media type, tool/host, coordinate/color/time basis, and validation. Evidence: representation schema.
    • B1.6.c Define USD layer identity/role/order/muting, composition arcs, resolver context, variants, payloads, and package membership. Evidence: layer fixtures.
    • B1.6.d Define typed dependency edges with exact version/checksum, optionality, resolver rule, license/rights propagation, and cycle policy. Evidence: dependency tests.
    • B1.6.e Define source/command/transformation/operator/tool provenance, supersession/quarantine/recall, and retention/legal-hold invariants. Evidence: lifecycle fixtures.
  • B1.7 Define geometry, topology, transform, unit, axis, material, texture, rig, animation, camera, light, metadata, and scene measurements. Evidence: typed measure registry.
    • B1.7.a Define common measurement identity, subject/path, component, algorithm/version/config, unit, coordinate/time basis, raw value/range, uncertainty, and source revision. Evidence: measure schema.
    • B1.7.b Register geometry/topology and transform/unit/axis measures with applicability and independent known values. Evidence: registry rows.
    • B1.7.c Register material/texture/color and rig/skinning/animation measures with applicability and independent known values. Evidence: registry rows.
    • B1.7.d Register camera/light/timecode and metadata/scene/dependency measures with applicability and independent known values. Evidence: registry rows.
    • B1.7.e Define missing, unavailable, unsupported, invalid-input, and evaluator-failure states distinct from numeric results. Evidence: state tests.
  • B1.8 Define consistency result separately from raw measurements. Evidence: policy recomputation.
    • B1.8.a Define immutable raw measurement sets independent of thresholds and release decision. Evidence: measurement persistence schema.
    • B1.8.b Define verdict identity, profile/version, exact measurements, thresholds/rules, pass/fail/review/non-pass status, and reasons. Evidence: consistency schema.
    • B1.8.c Define human review/waiver identity, authority, scope, rationale, expiry, and prohibited waivers. Evidence: review schema.
    • B1.8.d Recompute a new verdict under new policy without mutating prior measurements or verdicts. Evidence: policy-recompute fixture.
  • B1.9 Define build template, parameter schema, target, toolchain, dependency lock, environment, resources, gates, and lifecycle. Evidence: schema.
    • B1.9.a Define template identity/version/owner/purpose, supported profiles/targets, typed parameters/defaults/constraints, and secrets references. Evidence: template schema.
    • B1.9.b Define exact toolchain/container/host SDK/runtime/compiler/ plugin versions, dependency lock/checksums, and environment policy. Evidence: reproducibility schema.
    • B1.9.c Define CPU/GPU/memory/storage/license/device/network/resource requirements, capacity class, timeout, checkpoint, and cost model. Evidence: execution schema.
    • B1.9.d Define required input/intermediate/output gates, approvals, target validation/read-back, package/signing, and evidence outputs. Evidence: gate contract.
    • B1.9.e Define draft/review/approved/deprecated/revoked lifecycle, compatibility windows, impact preview, and immutable historical use. Evidence: lifecycle tests.
  • B1.10 Define release candidate and evidence with source revisions, commands, artifacts, checksums, validators, gates, approvals, toolchain, target, and read-back. Evidence: integrity invariants.
    • B1.10.a Lock candidate identity to exact source/dependency/template/ configuration/command/artifact revisions and checksums. Evidence: candidate schema.
    • B1.10.b Record toolchain/host/worker/environment, validator/profile/ calibration/threshold versions, raw measures, verdicts, and limitations. Evidence: technical evidence schema.
    • B1.10.c Record named approvals/waivers, roles, scopes, conditions, expiry, signatures, and staleness dependencies. Evidence: governance schema.
    • B1.10.d Record target/package/deployment identity, manifest/signatures, delivery result, and independent read-back measurements/checksums. Evidence: target evidence schema.
    • B1.10.e Verify no release evidence can be reused after any locked revision, gate, approval, rights, or target state changes. Evidence: integrity tests.
  • B1.11 Define immutable revisions/diffs for layers, dependencies, settings, templates, mappings, and evidence. Evidence: golden fixtures.
    • B1.11.a Define canonical serialization, stable identity, parent/base revision, author/time/reason, checksum, and optimistic-concurrency invariants. Evidence: revision schema.
    • B1.11.b Define semantic layer and dependency diffs including add/ remove/change/reorder/mute/resolver/version/license effects. Evidence: diff goldens.
    • B1.11.c Define settings/template/mapping diffs with defaults, normalization, secret redaction, compatibility, and downstream impact. Evidence: diff goldens.
    • B1.11.d Define evidence diffs for measurements, policies, verdicts, approvals, limitations, and target read-back without rewriting history. Evidence: diff goldens.
  • B1.12 Define protocol and contract version negotiation. Evidence: version skew and unsupported-client tests.
    • B1.12.a Advertise supported protocol/contract/capability versions and constraints for server, gateway, host/plugin, SDK, and consumer. Evidence: negotiation schema.
    • B1.12.b Select the highest mutually approved version deterministically and bind it to the session/command/job. Evidence: selection tests.
    • B1.12.c Define backward/forward compatibility, optional/unknown fields, capability downgrade, deprecation, and expiry semantics. Evidence: compatibility policy.
    • B1.12.d Fail closed with actionable unsupported/incompatible states rather than guessing or silently coercing. Evidence: version-skew suite.
  • B1.13 Define session/command audit projections with sensitive argument redaction. Evidence: traceability without secret leakage.
    • B1.13.a Define immutable audit events for enrollment, grant, session, command, result, artifact, approval, release, revoke, and emergency action. Evidence: audit schema.
    • B1.13.b Record actor/host/tenant/project, action, object/revision, decision, time, correlation, and evidence refs without raw protected payloads. Evidence: trace fixture.
    • B1.13.c Apply typed allowlist/redaction/hashing/tokenization to command arguments, paths, URLs, secrets, content, and DCC logs. Evidence: redaction tests.
    • B1.13.d Provide role-filtered investigations that reconstruct sequence and causation while preserving compartment boundaries. Evidence: audit journey.
  • B1.14 Define SDK and remote-extension consumer compatibility. Evidence: public-contract fixtures.
    • B1.14.a Define supported SDK language/version/runtime and remote- extension platform/version matrices. Evidence: consumer registry.
    • B1.14.b Define public discovery, enrollment/session, typed command, artifact/job/result, event, error, auth, retry, and cancellation contracts. Evidence: public API spec.
    • B1.14.c Define compatibility windows, negotiation, feature detection, generated clients/types, and deprecation behavior. Evidence: version fixtures.
    • B1.14.d Compile and run representative external consumers against current and supported prior server versions. Evidence: consumer suite.
  • B1.15 Define render job, scene/camera/layer, frame/range, renderer, settings, color/OCIO, output/pass, resource, license, checkpoint, artifact, measurement, and review contracts. Evidence: still/sequence and multi-pass fixtures.
    • B1.15.a Define immutable render source scene/USD, camera, layer/variant, frame/range/handles, resolution/aspect, renderer/version, and settings. Evidence: render request schema.
    • B1.15.b Define OCIO/config/display/view/look, input/output color spaces, tone/exposure, and color-provenance lock. Evidence: color contract.
    • B1.15.c Define beauty/AOV/pass/output naming, format/profile, channels, compression, destination, checksum, and package relationships. Evidence: output schema.
    • B1.15.d Define compute/storage/license requirements, chunking, checkpoint/resume, retries, partial frames/passes, cancellation, and cost. Evidence: execution contract.
    • B1.15.e Define per-frame/sequence measurements, visual review, approvals, artifact lineage, and read-back for still, sequence, and multi-pass outputs. Evidence: render fixtures.
  • B1.16 Define virtual-production stage/session, camera tracking, calibration, lens, LED wall, frustum, genlock, timecode, ICVFX layer, playback, recording, safety, and incident contracts. Evidence: device/clock/fail-safe invariants.
    • B1.16.a Define stage/session identity, location, schedule, personnel/ roles, project/shot, device inventory, topology, and operating mode. Evidence: stage schema.
    • B1.16.b Define camera tracker/calibration/lens/distortion/focus/zoom, coordinate transforms, uncertainty, validity window, and drift. Evidence: camera fixtures.
    • B1.16.c Define LED wall/processors/panels, color/calibration, frustum, ICVFX layers, content versions, playback, and target state. Evidence: display contract.
    • B1.16.d Define genlock/timecode/reference clock, sync quality, discontinuity, recording/media, take identity, and read-back. Evidence: clock/recording fixtures.
    • B1.16.e Define device-safe command limits, interlocks, emergency stop, degraded/fallback states, incident capture, and restart authority. Evidence: fail-safe invariants.
  • B1.17 Define XR target/device, runtime, input, interaction, comfort, performance, spatial asset, entitlement, package, deployment, and validation contracts; define mocap stream/skeleton/retarget/recording and MetaHuman identity/rig/Live Link contracts separately. Evidence: platform-specific fixtures and consent/privacy mapping.
    • B1.17.a Define XR target/device/OS/runtime versions, capabilities, entitlement, input/interaction mappings, spatial permissions/assets, and deployment/package identity. Evidence: XR platform schema.
    • B1.17.b Define XR frame timing/latency, rendering, memory/thermal, tracking, locomotion/comfort, accessibility, safety, and validation measures. Evidence: XR quality contract.
    • B1.17.c Define mocap device/stream protocol, clock, subjects/consent, skeleton/channels, calibration, confidence, retarget mapping, recording, dropout, and lineage. Evidence: mocap schema.
    • B1.17.d Define MetaHuman identity/rights, asset/rig/DNA versions, animation/face channels, Live Link source/timecode, Unreal compatibility, and protected-data handling. Evidence: MetaHuman schema.
    • B1.17.e Keep XR, mocap, and MetaHuman execution, permissions, quality, release, and evidence independently versioned. Evidence: boundary fixtures.
  • B1.18 Define gameplay input/action, save schema, inventory, combat, behavior/AI, navigation, state, test scenario, telemetry, migration, and build contracts. Evidence: versioned runtime/save compatibility and no unsafe arbitrary-script command.
    • B1.18.a Define versioned input devices/actions/bindings/contexts, accessibility remapping, network authority, and replayable input fixtures. Evidence: input contract.
    • B1.18.b Define save/profile/checkpoint/world-state schema, version, integrity, privacy, migration, rollback, and forward/backward behavior. Evidence: save fixtures.
    • B1.18.c Define inventory/economy, combat/damage/abilities, behavior/AI, navigation, quest/state-machine, and deterministic scenario contracts. Evidence: gameplay schemas.
    • B1.18.d Define typed test setup/action/assertion/cleanup commands from an allowlist; prohibit arbitrary scripts or console execution. Evidence: command security tests.
    • B1.18.e Define bounded gameplay/performance telemetry, expected outcomes, build/package versions, target deployment, and result evidence. Evidence: scenario fixtures.

B2 — Host enrollment, liveness, sessions, and remote security#

  • B2.1 Build host enrollment with operator approval and cryptographic identity. Evidence: unapproved host cannot connect.
    • B2.1.a Issue a one-time bounded enrollment challenge tied to tenant, environment, expected host, requester, expiry, and nonce. Evidence: challenge tests.
    • B2.1.b Collect host/DCC/plugin/platform identity, key proof, attestation, capabilities, owner, and requested scopes without trusting claims. Evidence: enrollment request.
    • B2.1.c Verify policy, compatibility, attestation/signatures, duplicate identity, and operator authority before approval. Evidence: review tests.
    • B2.1.d Issue environment-bound cryptographic identity and persist approval/rejection/audit; deny every unapproved or replayed request. Evidence: connection negative suite.
  • B2.2 Build short-lived mutual authentication and key rotation. Evidence: expired, revoked, replayed, and wrong-environment credentials fail.
    • B2.2.a Establish mutual TLS or equivalent proof from enrolled host and environment-specific server identities. Evidence: handshake tests.
    • B2.2.b Mint short-lived session credentials with audience, tenant, host, capability scope, issue/expiry, nonce, and key version. Evidence: token inspection.
    • B2.2.c Rotate host/server keys with bounded overlap, proof of possession, atomic activation, rollback, and audit. Evidence: rotation journey.
    • B2.2.d Revoke credentials/keys immediately and reject expired, replayed, wrong-host/audience/environment, or downgraded authentication. Evidence: adversarial suite.
  • B2.3 Build capability/version negotiation. Evidence: unsupported command is rejected before session creation.
    • B2.3.a Exchange server/host/plugin protocol and typed capability/version/ limit declarations after authentication. Evidence: negotiation trace.
    • B2.3.b Intersect declarations with the approved host compatibility profile and current policy. Evidence: selection tests.
    • B2.3.c Bind the negotiated versions/capabilities to the session grant and expose explicit downgraded/unavailable reasons. Evidence: session contract.
    • B2.3.d Reject session creation when required versions/capabilities are absent and reject later unnegotiated commands. Evidence: negative tests.
  • B2.4 Build liveness, latency, load, DCC state, project state, and maintenance probes. Evidence: health rollup fixtures.
    • B2.4.a Collect signed heartbeat sequence/time, round-trip latency, clock skew, process/plugin connectivity, and protocol health. Evidence: liveness fixtures.
    • B2.4.b Collect bounded CPU/GPU/memory/storage/license/queue load and capacity without secret/path/content leakage. Evidence: load fixtures.
    • B2.4.c Collect typed DCC busy/modal/render/dirty-scene state and active project/workspace/revision compatibility. Evidence: host-state fixtures.
    • B2.4.d Collect maintenance/drain/quarantine/update state and freshness. Evidence: maintenance fixtures.
    • B2.4.e Compute healthy/degraded/unavailable/incompatible/stale rollups with component reasons and no false healthy state. Evidence: health matrix.
  • B2.5 Build session open with scoped project, commands, paths, expiry, resources, and actor. Evidence: least-privilege tests.
    • B2.5.a Authenticate the actor and enrolled host, resolve tenant/project/ workspace/purpose, and load current compatibility/health. Evidence: precondition tests.
    • B2.5.b Intersect requested commands, virtual paths, objects, targets, resources, duration, and limits with role/policy/host capabilities. Evidence: scope tests.
    • B2.5.c Require step-up or approval for sensitive scopes and record exact grant rationale/approver. Evidence: grant journey.
    • B2.5.d Open an immutable expiring session with negotiated versions, fencing epoch, actor, scopes, resource reservations, and audit. Evidence: session fixture.
    • B2.5.e Prove ungranted property, path, command, project, resource, and time use fails server- and host-side. Evidence: least-privilege suite.
  • B2.6 Build heartbeat, reconnect, suspend, resume, drain, expiry, and close. Evidence: clock-controlled session tests.
    • B2.6.a Heartbeat with monotonic sequence, observed session epoch, host/ project state, in-flight command refs, and bounded health data. Evidence: heartbeat tests.
    • B2.6.b Mark degraded/disconnected after controlled thresholds and fence commands after lease/session expiry. Evidence: clock tests.
    • B2.6.c Reconnect through reauthentication, state reconciliation, and same-or-new epoch without duplicate command execution. Evidence: reconnect tests.
    • B2.6.d Suspend/resume or drain with explicit admission behavior, in-flight completion/cancellation policy, and deadline. Evidence: lifecycle tests.
    • B2.6.e Close or expire idempotently, release reservations/grants, retain audit, and classify late host results. Evidence: terminal-state tests.
  • B2.7 Build remote command allowlist with typed schemas and property authorization. Evidence: arbitrary command/argument tests fail.
    • B2.7.a Register each command with stable ID/version, supported hosts, typed arguments/results, defaults/bounds, side effects, prerequisites, permissions, and compensation. Evidence: command registry.
    • B2.7.b Validate envelope/schema/unknown fields, negotiated capability, host/project/object/path scope, expected revision, and deadline before dispatch. Evidence: validation suite.
    • B2.7.c Authorize sensitive argument properties and resulting object/ file mutations independently of command visibility. Evidence: property tests.
    • B2.7.d Translate only to a fixed host API/plugin operation; prohibit arbitrary shell, script, eval, console, path, or network execution. Evidence: injection tests.
    • B2.7.e Audit normalized redacted arguments, authorization decision, host operation/version, result, and changed revisions. Evidence: audit fixture.
  • B2.8 Build safe path mapping and virtual roots. Evidence: traversal, symlink, case collision, UNC/device path, and Unicode tests.
    • B2.8.a Configure named virtual roots by tenant/project/host/purpose with local canonical root, read/write policy, formats, quotas, and lifetime. Evidence: root registry.
    • B2.8.b Parse and normalize separators, dot segments, percent/Unicode encoding, normalization form, case semantics, drive/UNC/device prefixes, and reserved names. Evidence: path corpus.
    • B2.8.c Resolve components without following unauthorized symlinks/ junctions/mounts and verify the final canonical target remains within its root. Evidence: traversal tests.
    • B2.8.d Open/create with race-safe primitives, file type/size/permission checks, collision policy, and atomic write/rename. Evidence: TOCTOU tests.
    • B2.8.e Return only virtual paths outside the host and redact local user/ secret path components from logs/errors. Evidence: disclosure tests.
  • B2.9 Build network destination allowlists for host-side fetches. Evidence: SSRF/DNS-rebinding/redirect tests.
    • B2.9.a Define destination profiles by scheme, exact domain, port, path, method, purpose, size/type, TLS policy, credentials, and tenant. Evidence: allowlist schema.
    • B2.9.b Resolve DNS and reject loopback, link-local, private, metadata, multicast, reserved, unexpected IP families, and disallowed address changes. Evidence: SSRF fixtures.
    • B2.9.c Validate every redirect and connection against the profile, pinning/rechecking resolution to resist DNS rebinding. Evidence: redirect/rebinding tests.
    • B2.9.d Bound time, bytes, decompression, responses, retries, and credentials; quarantine fetched content before host use. Evidence: fetch safety suite.
  • B2.10 Build command deadlines, cancellation, idempotency, and late-result handling. Evidence: race tests.
    • B2.10.a Enforce admission and execution deadlines using synchronized or duration-based clocks and typed expiry states. Evidence: clock tests.
    • B2.10.b Deduplicate command IDs/idempotency keys and reject the same key with a different normalized request. Evidence: duplicate tests.
    • B2.10.c Propagate cooperative cancellation, host abort, checkpoint/ cleanup, and compensation according to command contract. Evidence: cancellation tests.
    • B2.10.d Fence result commit by session/command epoch and classify late output without publishing it as success. Evidence: race suite.
    • B2.10.e Reconcile timeout/cancel/result uncertainty through host state and changed-object read-back. Evidence: unknown-outcome journey.
  • B2.11 Build bounded output/log streaming with backpressure and redaction. Evidence: oversize and secret-leak tests.
    • B2.11.a Define typed progress/log/diagnostic/output chunk envelopes with sequence, command/attempt, time, level, media type, size, and terminal semantics. Evidence: stream schema.
    • B2.11.b Enforce per-event/rate/total/buffer/retention/cardinality bounds and spill authorized artifacts rather than unbounded inline data. Evidence: limit tests.
    • B2.11.c Apply structured allowlisting and secret/token/path/content/PII redaction at the host before transport and again on ingest. Evidence: redaction suite.
    • B2.11.d Implement acknowledgement/backpressure, resumable cursor, duplicate/out-of-order handling, disconnect, and dropped-event markers. Evidence: streaming fault tests.
  • B2.12 Build host quarantine and emergency session termination. Evidence: in-flight work enters explicit safe state.
    • B2.12.a Authorize and record host quarantine with reason, evidence, scope, effective time, step-up, and optional fleet profile impact. Evidence: control tests.
    • B2.12.b Immediately block new authentication/sessions/commands, invalidate credentials/grants, and fence host results. Evidence: quarantine tests.
    • B2.12.c Terminate active sessions/commands according to safe-stop contracts and mark jobs/artifacts/releases explicit partial/unknown/ quarantined states. Evidence: in-flight fault test.
    • B2.12.d Trace downstream artifacts/builds/releases made by the host and launch impact review/recall where required. Evidence: blast-radius test.
    • B2.12.e Require remediation, re-attestation, approval, key rotation, and verification before unquarantine. Evidence: recovery journey.
  • B2.13 Build plugin/extension version, signature, and update policy. Evidence: unsigned/downgrade tests.
    • B2.13.a Register approved package identity/version, publisher/signing keys, host/platform ranges, files/checksums, permissions, dependencies, and lifecycle. Evidence: package registry.
    • B2.13.b Verify signatures, provenance/attestation, checksums, archive/ path safety, malware, dependency lock, and policy before install. Evidence: supply-chain suite.
    • B2.13.c Stage, install, activate, health-check, rollback, and report updates atomically without corrupting the host plugin. Evidence: update journey.
    • B2.13.d Reject unsigned, tampered, revoked, wrong-host/environment, unapproved downgrade, dependency-confused, or expired packages. Evidence: negative tests.
    • B2.13.e Reconcile fleet versions and quarantine vulnerable/incompatible hosts with controlled compatibility windows. Evidence: fleet policy.
  • B2.14 Build audited break-glass access. Evidence: step-up, time limit, reason, notification, and review.
    • B2.14.a Define eligible incidents, permitted emergency capabilities, prohibited actions, maximum duration, approver/reviewer, and automatic expiry. Evidence: break-glass policy.
    • B2.14.b Require strong step-up, incident/ticket, justification, exact host/project/resource scope, and dual approval where required. Evidence: activation tests.
    • B2.14.c Mint isolated short-lived grants, visibly label their use, notify security/owners, and record every view/mutation. Evidence: access journey.
    • B2.14.d Auto-expire/revoke, terminate dependent sessions, rotate exposed secrets if needed, and require independent post-use review. Evidence: closure tests.
  • B2.15 Build allowlisted screen/frame/viewport capture and host-state observation for preview, agent planning, diagnostics, and support with explicit user/session grant, visible capture state, window/region scope, frame/rate/size limits, protected-region redaction, retention, and revoke. Evidence: no hidden desktop surveillance or capture outside the granted host/project context.
    • B2.15.a Define distinct preview, planning, diagnostic, and support capture purposes with allowed host windows/viewports, fields, viewers, processing, retention, and prohibited use. Evidence: purpose policy.
    • B2.15.b Require an explicit current user/session grant for exact application/window/viewport/region, duration, frame/rate/size, and destination. Evidence: consent tests.
    • B2.15.c Display persistent host-side capture indicator, scope, consumer, and stop control that the remote operator cannot hide. Evidence: host UI test.
    • B2.15.d Capture only allowlisted pixels/state; mask protected windows/ regions, notifications, credentials, other projects, and desktop surroundings before egress. Evidence: redaction fixtures.
    • B2.15.e Enforce bandwidth/frame/rate/resolution/time bounds, secure transport/storage, viewer authorization, watermark, and audit. Evidence: limit/access suite.
    • B2.15.f Revoke or expire immediately, stop collection/streaming, reconcile buffered frames, apply retention/deletion, and prove no capture outside scope. Evidence: surveillance negative test.

B3 — Ingest, OpenUSD interchange, conversion, and dependency handling#

  • B3.1 Build asset intake from upload, artifact store, DCC pull, project reference, and previous release. Evidence: provenance per path.
    • B3.1.a Define a common intake manifest for actor, tenant/project, purpose, source kind, source identity/revision, expected format/profile, rights/license, and idempotency. Evidence: manifest schema.
    • B3.1.b Implement bounded resumable upload and existing artifact-store reference with checksum/authorization validation. Evidence: intake tests.
    • B3.1.c Implement DCC pull through an enrolled scoped session and project reference through an approved resolver context. Evidence: host/ resolver tests.
    • B3.1.d Implement previous-release selection by immutable manifest/ checksum without mutating or treating it as source authority. Evidence: release-input test.
    • B3.1.e Normalize all paths into one quarantined source artifact while retaining path-specific actor/command/source provenance. Evidence: lineage matrix.
  • B3.2 Quarantine and scan files, archives, USD packages, scripts, plugins, and metadata. Evidence: malicious corpus.
    • B3.2.a Store intake bytes in an isolated non-executable quarantine with immutable checksum, type claim, size limits, and no resolver/network access. Evidence: quarantine tests.
    • B3.2.b Safely enumerate archives/packages with entry/count/depth/ratio/ total-size bounds and reject traversal, links, devices, collisions, and nested bombs. Evidence: archive corpus.
    • B3.2.c Detect actual media/file types and scan content, scripts, plugins/binaries, macros, metadata, and known malware/suspicious constructs. Evidence: scanner matrix.
    • B3.2.d Parse USD in a sandbox with resolver disabled or constrained and inspect layers, assets, schemas, paths, code-bearing fields, and external references. Evidence: malicious USD corpus.
    • B3.2.e Emit signed per-file findings and promote only policy-approved bytes; preserve rejected/quarantined state and reviewer workflow. Evidence: promotion tests.
  • B3.3 Detect format, version, units, axes, color, dependencies, and encryption. Evidence: fixture matrix.
    • B3.3.a Determine format/profile/version from bytes and structure, compare extensions/MIME/claims, and report ambiguity or unsupported variants. Evidence: detection fixtures.
    • B3.3.b Detect scene/asset units, scale, up/forward axes, handedness, coordinate basis, time codes, and frame rates with source locations. Evidence: transform metadata fixtures.
    • B3.3.c Detect embedded/declared color spaces, OCIO context, texture interpretations, and missing/ambiguous color metadata. Evidence: color fixtures.
    • B3.3.d Enumerate embedded/external dependencies and resolver/path requirements without fetching them. Evidence: dependency fixtures.
    • B3.3.e Detect container/file encryption, compression, password/key requirements, and unsupported opaque regions; never guess secrets. Evidence: encrypted fixtures.
  • B3.4 Build canonical OpenUSD stage creation with pinned profile and schema set. Evidence: deterministic golden.
    • B3.4.a Select exact OpenUSD runtime/profile/schema/plugin versions, resolver policy, meters-per-unit, axes, time codes, and layer strategy. Evidence: stage plan.
    • B3.4.b Create root/session/sublayers with deterministic identifiers, ordering, default prim, metadata, namespaces, and canonical formatting. Evidence: stage golden.
    • B3.4.c Author only allowed schemas/APIs and record source-to-prim/ property mapping plus unsupported data. Evidence: schema validation.
    • B3.4.d Serialize deterministically where promised and attribute unavoidable nondeterministic fields explicitly. Evidence: repeated- build checksum test.
    • B3.4.e Reopen with an independent/pinned reader and verify composition, metadata, checksums, and provenance. Evidence: read-back test.
  • B3.5 Build layer, variant set, reference, payload, inherit, specialize, and composition inspection. Evidence: composition fixtures.
    • B3.5.a Enumerate layer stack, identifiers, roles, permissions, offsets, muting, sublayers, edit target, and resolver context. Evidence: layer fixture.
    • B3.5.b Enumerate variant sets/selections/opinions and compare declared, authored, and composed selections. Evidence: variant fixtures.
    • B3.5.c Enumerate references/payloads with target prim, layer offsets, load state, resolution, and dependency identity. Evidence: arc fixtures.
    • B3.5.d Enumerate inherits/specializes and composition strength/order, cycles, conflicts, blocked values, and provenance. Evidence: composition fixtures.
    • B3.5.e Provide composed-value-to-authoring-layer/source tracing and a semantic non-graph representation. Evidence: inspection journey.
  • B3.6 Build resolver context and dependency capture without environment- implicit paths. Evidence: portable package test.
    • B3.6.a Define versioned resolver-context inputs, virtual roots, search/ mapping rules, credentials references, case semantics, and network policy. Evidence: resolver schema.
    • B3.6.b Resolve every dependency under the explicit context and record requested path, canonical identity/version/checksum, selected rule, and resolution diagnostics. Evidence: resolver trace.
    • B3.6.c Reject undeclared environment variables, current directories, user-home paths, ambient search paths, or network fallback. Evidence: hermeticity tests.
    • B3.6.d Capture or package permitted dependencies with rewritten virtual/ relative refs and license/rights/provenance. Evidence: dependency manifest.
    • B3.6.e Resolve and open the package on a clean environment/OS fixture with no original path context. Evidence: portable package test.
  • B3.7 Build missing, cyclic, duplicate, ambiguous, and external dependency detection. Evidence: broken-package corpus.
    • B3.7.a Build a typed dependency graph from authored and discovered refs, including layer/prim/property source and optional/load behavior. Evidence: graph fixture.
    • B3.7.b Detect missing/unreadable/version-mismatched/checksum-mismatched dependencies with exact referrers. Evidence: missing fixtures.
    • B3.7.c Detect cycles and composition/resolver recursion with cycle path and affected objects. Evidence: cyclic fixtures.
    • B3.7.d Detect duplicate logical identities/content, case/Unicode path collisions, and ambiguous multi-candidate resolution. Evidence: ambiguity fixtures.
    • B3.7.e Classify external network/local absolute/out-of-root dependencies and enforce capture, exception, or rejection policy. Evidence: external dependency tests.
  • B3.8 Build geometry/topology conversion and preservation. Evidence: known mesh round-trips.
    • B3.8.a Inventory source primitives, points, faces/indices, orientation, normals/tangents, subdivision, creases, holes, UV/color/primvars, and instances. Evidence: source measure report.
    • B3.8.b Map each supported geometry/topology construct to canonical USD with explicit interpolation, indexing, precision, and naming. Evidence: mapping registry.
    • B3.8.c Triangulate/tessellate/bake/approximate only when target profile requires it and record source-to-output correspondence/loss. Evidence: conversion fixtures.
    • B3.8.d Measure counts, bounds, topology equivalence, orientation, normals, UVs/primvars, subdivision, instances, and geometric deviation. Evidence: independently known values.
    • B3.8.e Round-trip representative and adversarial meshes through every supported target adapter. Evidence: mesh compatibility matrix.
  • B3.9 Build transform, unit, axis, pivot, hierarchy, and instancing conversion. Evidence: independently derived fixtures.
    • B3.9.a Parse local/world transforms, operation order, pivots, inherit/ reset behavior, hierarchy, units, axes/handedness, and instance prototypes. Evidence: source fixtures.
    • B3.9.b Compute explicit basis/unit conversions and operation mappings with precision, decomposition, negative scale, shear, and rotation-order policy. Evidence: analytical goldens.
    • B3.9.c Preserve hierarchy and instancing where supported or disclose flattening/baking/prototype splitting and dependency impact. Evidence: loss fixtures.
    • B3.9.d Compare local/world transforms, pivots, bounds, hierarchy, and instance identity after target read-back. Evidence: round-trip matrix.
  • B3.10 Build material, texture, shader, color-space, UDIM, and fallback mapping, including MaterialX where supported. Evidence: per-target compatibility fixtures.
    • B3.10.a Inventory material graphs, shader nodes/versions/connections/ parameters, assignments/subsets, textures/samplers/UDIMs, and color metadata. Evidence: material manifest.
    • B3.10.b Map supported native constructs to pinned USD Shade/MaterialX profiles with explicit units, color transforms, and texture resolver paths. Evidence: mapping fixtures.
    • B3.10.c Detect unsupported nodes/features and preview substitute, approximate, bake, fallback, or fail choices with visible semantic/ visual loss. Evidence: loss decisions.
    • B3.10.d Package textures/UDIM tiles with checksums, dedupe, rights/ license, path portability, and missing-tile diagnostics. Evidence: texture tests.
    • B3.10.e Read back in each target and compare graph semantics, assignments, parameters, texture/color interpretation, and reference renders where calibrated. Evidence: compatibility matrix.
  • B3.11 Build skeleton, skinning, blend-shape, animation, constraint, and camera conversion. Evidence: round-trip fixtures.
    • B3.11.a Convert skeleton hierarchy/names/rest/bind transforms and skinning influences/weights with unit/axis/precision and limit policy. Evidence: rig fixtures.
    • B3.11.b Convert blend-shape targets/in-betweens/normals/weights and animation channels, interpolation, layers, clips, events, and time base. Evidence: deformation/animation fixtures.
    • B3.11.c Map constraints/controllers/drivers to supported target constructs or preview explicit bake/approximation/unsupported behavior. Evidence: constraint fixtures.
    • B3.11.d Convert camera projection, aperture/sensor, focal length, lens/ clipping/focus, transforms, and animation. Evidence: camera goldens.
    • B3.11.e Read back and measure hierarchy, rest/bind, deformation, weights, curves/timing, constraints/bakes, and camera framing per target. Evidence: round-trip matrix.
  • B3.12 Build metadata, naming, custom property, and namespace policy. Evidence: loss/rename diagnostics.
    • B3.12.a Define canonical identifier, display-name, path, namespace, casing, Unicode, reserved word, length, collision, and stability rules. Evidence: naming spec.
    • B3.12.b Classify standard, custom, sensitive, internal, target-specific, code-bearing, and unsupported metadata/properties with type and retention rules. Evidence: metadata registry.
    • B3.12.c Map names/properties deterministically with source/output lookup, collision resolution, type/range conversion, and reverse mapping. Evidence: mapping tests.
    • B3.12.d Strip/redact/quarantine prohibited metadata and report every rename, drop, coercion, or namespace change. Evidence: diagnostics.
    • B3.12.e Verify target read-back and mapping reversibility within the declared profile. Evidence: metadata round-trip.
  • B3.13 Build target adapters for ratified DCCs/engines through public bridge interfaces. Evidence: no target-specific logic in generic core.
    • B3.13.a Define the public adapter interface for capability discovery, plan, import/export, inspect, validate, cancel, result, and read-back. Evidence: adapter contract.
    • B3.13.b Implement per-host adapters against typed bridge commands and exact host/plugin/version profiles. Evidence: adapter fixtures.
    • B3.13.c Keep host-specific formats, defaults, paths, command mappings, limitations, and workarounds inside the adapter package. Evidence: dependency boundary test.
    • B3.13.d Run a shared conformance suite plus host-specific known-value, unsupported, authorization, cancellation, and failure cases. Evidence: adapter matrix.
    • B3.13.e Reject incompatible/unavailable adapter selection rather than falling through generic assumptions. Evidence: selection tests.
  • B3.14 Run usdchecker plus pinned custom validators. Evidence: exact report and validator versions.
    • B3.14.a Pin usdchecker/OpenUSD and every custom validator code/config/ dependency version with checksum and supported profile. Evidence: validator registry.
    • B3.14.b Run validators in a sandbox against exact stage/package bytes and explicit resolver context. Evidence: execution manifest.
    • B3.14.c Parse raw output into typed findings while retaining the unmodified bounded report artifact. Evidence: parser goldens.
    • B3.14.d Distinguish tool unavailable/crash/timeout/invalid input from a valid validation failure. Evidence: state tests.
    • B3.14.e Persist validator versions/config/checksums/findings and link them to policy verdict/release evidence. Evidence: provenance query.
  • B3.15 Build incremental conversion and cache eligibility. Evidence: cache key covers source, dependencies, adapter, toolchain, policy, and target.
    • B3.15.a Compute semantic change sets from source layer/property/ dependency revisions and identify affected conversion stages/targets. Evidence: invalidation fixtures.
    • B3.15.b Define cache keys from exact source/dependency bytes, resolver context, adapter/toolchain/profile/config/policy/target/environment versions. Evidence: key coverage tests.
    • B3.15.c Reuse only checksum-valid, compatible, authorized, non- quarantined entries whose stage supports incremental execution. Evidence: eligibility tests.
    • B3.15.d Recompute affected outputs, preserve unaffected immutable refs, and rebuild aggregate manifests/validation. Evidence: incremental conversion tests.
    • B3.15.e Quarantine corrupt or conflicting entries and prove a full rebuild matches declared deterministic output. Evidence: cache fault test.
  • B3.16 Build conversion plan preview with expected losses, substitutions, unsupported features, cost, host, and duration. Evidence: user must accept material losses.
    • B3.16.a Resolve exact source/dependencies, target/profile/adapter/ toolchain, compatible host/resources, policy, and current measurements. Evidence: plan inputs.
    • B3.16.b Produce ordered stages with per-stage inputs/outputs/cache, host/resource/license, estimated duration/cost, and failure/cancel semantics. Evidence: plan schema.
    • B3.16.c Enumerate preserved, renamed, converted, substituted, approximated, baked, dropped, and unsupported features with affected objects/properties. Evidence: semantic loss report.
    • B3.16.d Preview validation/review/release consequences and safer target/ configuration alternatives. Evidence: decision view.
    • B3.16.e Require explicit authorized acceptance of the exact material losses and stale the acceptance when inputs/plan/policy change. Evidence: acceptance tests.
  • B3.17 Build durable interchange jobs with checkpoint, retry, cancel, and resume. Evidence: host/worker loss test.
    • B3.17.a Persist job/DAG/stage/attempt state, exact revisions/plan, requirements, reservations, idempotency, deadlines, and lineage. Evidence: job persistence tests.
    • B3.17.b Lease stages to compatible hosts/workers with fencing, heartbeat, bounded progress, and duplicate-delivery tolerance. Evidence: lease tests.
    • B3.17.c Checkpoint resumable inputs/outputs/tool state atomically and validate before reuse. Evidence: checkpoint fixtures.
    • B3.17.d Retry only classified safe stages under budget/backoff; retain attempts and dead-letter permanent/exhausted failures. Evidence: retry suite.
    • B3.17.e Cancel cooperatively and reconcile partial outputs/host side effects/late results explicitly. Evidence: cancel tests.
    • B3.17.f Resume after service/worker/host loss from the last valid fence without duplicate mutations or hidden gaps. Evidence: recovery journey.
  • B3.18 Build deterministic or explicitly non-deterministic replay classification. Evidence: environment dependencies recorded.
    • B3.18.a Classify each stage deterministic, conditionally deterministic, or nondeterministic with sources of time/randomness/threading/host/ renderer/external/environment variation. Evidence: stage registry.
    • B3.18.b Record all reproducibility inputs: bytes/checksums, versions, configs, seeds, locale/time zone, hardware/runtime, environment, and external dependency snapshots. Evidence: replay manifest.
    • B3.18.c Re-execute deterministic stages and compare canonical bytes/ semantic measurements to zero or declared tolerance. Evidence: replay tests.
    • B3.18.d For nondeterministic stages, disclose variation, retain candidates/selection, and compare bounded quality/integrity invariants. Evidence: nondeterministic replay report.

B4 — Cross-DCC fidelity, validation, preview, and diff#

  • B4.1 Define versioned fidelity profiles by source, target, asset type, and release purpose. Evidence: complete profile registry.
    • B4.1.a Define profile schema for source/target formats, host/tool versions, asset type, release purpose, applicable measures, tolerances, validators, review/waiver rules, and limitations. Evidence: schema.
    • B4.1.b Register every supported source-target-asset-purpose combination with independently owned thresholds and evidence. Evidence: coverage matrix.
    • B4.1.c Version, approve, activate, deprecate, retire, and migrate profiles without rewriting historical results. Evidence: lifecycle tests.
    • B4.1.d Reject validation/release when no exact current compatible profile exists. Evidence: missing-profile tests.
  • B4.2 Implement geometry position, normal, topology, UV, and bounds measurements. Evidence: known fixtures.
    • B4.2.a Align source/target geometry by stable mapping, instance, coordinate/unit basis, precision, and comparable primitive scope. Evidence: alignment fixtures.
    • B4.2.b Measure point/surface position distance, normal/tangent angular deviation, and object/world bounds error. Evidence: analytical goldens.
    • B4.2.c Compare topology counts/connectivity/orientation/subdivision/ holes and distinguish declared tessellation or remeshing. Evidence: topology fixtures.
    • B4.2.d Compare UV/primvar sets, interpolation/indexing, seams, orientation, range, and per-element deviation. Evidence: UV fixtures.
    • B4.2.e Report exact objects/elements and raw values before profile verdict. Evidence: diagnostic tests.
  • B4.3 Implement transform, unit, axis, pivot, hierarchy, and instance measurements. Evidence: known fixtures.
    • B4.3.a Resolve source/target mapping and normalize declared unit/axis/ handedness/time basis without hiding conversion errors. Evidence: normalization tests.
    • B4.3.b Measure local/world translation, rotation, scale, shear, operation order, and pivot deviations. Evidence: analytical fixtures.
    • B4.3.c Compare hierarchy parentage/order, reset/inherit behavior, visibility, purpose, and instance/prototype identity/count. Evidence: scene fixtures.
    • B4.3.d Detect unmapped, duplicated, flattened, expanded, or ambiguous nodes/instances and report raw impact. Evidence: structure tests.
  • B4.4 Implement material parameter, texture, color-space, shader, and render-difference measurements. Evidence: target fixture corpus.
    • B4.4.a Compare material assignments, shader/node types/versions, connections, parameter types/values/units, and unsupported mappings. Evidence: graph fixtures.
    • B4.4.b Compare texture identity/checksum, channels, resolution, sampler/wrap/filter, UDIM tiles, paths, and missing/substituted assets. Evidence: texture fixtures.
    • B4.4.c Compare declared/decoded color spaces, OCIO transforms, linear/ display interpretation, alpha, and precision. Evidence: color goldens.
    • B4.4.d Render under pinned scene/camera/light/renderer/OCIO/settings and compute calibrated pixel/perceptual differences with masks. Evidence: target corpus.
    • B4.4.e Separate structural/material loss from renderer nondeterminism and threshold verdict. Evidence: diagnostic report.
  • B4.5 Implement skeleton hierarchy, skin weight, bind pose, blend shape, and deformation measurements. Evidence: rig pairs.
    • B4.5.a Map source/target joints and compare hierarchy, names/aliases, rest/bind transforms, inverse binds, units/axes, and required bones. Evidence: skeleton fixtures.
    • B4.5.b Compare per-vertex joint influences/weights, normalization, truncation/remapping, and weighted deformation impact. Evidence: skinning goldens.
    • B4.5.c Compare blend-shape targets/in-betweens/deltas/normals, assignments, weights, ranges, and combination behavior. Evidence: blend-shape fixtures.
    • B4.5.d Evaluate deformed geometry under a pinned representative pose/ expression corpus and report per-vertex/surface error. Evidence: rig pair report.
  • B4.6 Implement animation curve, key, timing, interpolation, root motion, and event measurements. Evidence: animation pairs.
    • B4.6.a Align clips/tracks/channels by mapping, time codes, rates, ranges, offsets, loops, layers, and coordinate basis. Evidence: alignment fixtures.
    • B4.6.b Compare key times/values/tangents/weights/interpolation, resampling, infinity/extrapolation, and baked versus authored behavior. Evidence: curve goldens.
    • B4.6.c Measure sampled position/rotation/scale and derivative error over time with discontinuity and precision handling. Evidence: sample fixtures.
    • B4.6.d Compare root trajectory/orientation, contacts, markers/events, clip metadata, duration, and loop seams. Evidence: motion pairs.
    • B4.6.e Report exact channels/ranges and declared bake/loss separately from verdict. Evidence: animation diagnostics.
  • B4.7 Implement camera, light, scene, variant, metadata, and dependency preservation measurements. Evidence: scene fixtures.
    • B4.7.a Compare camera projection, filmback/aperture, focal/lens/ clipping/focus, transform, framing, and animation. Evidence: camera fixtures.
    • B4.7.b Compare light type, transform, intensity/exposure, color/ temperature, shape, shadow, linking, and unsupported attributes. Evidence: light fixtures.
    • B4.7.c Compare scene hierarchy, purpose/visibility, collections, variants/selections, layer composition, and default state. Evidence: scene fixtures.
    • B4.7.d Compare governed metadata/custom properties/namespaces with rename/coercion/drop diagnostics. Evidence: metadata fixtures.
    • B4.7.e Compare dependency identities/versions/checksums/resolution, optionality, packaging, and license/rights propagation. Evidence: dependency fixtures.
  • B4.8 Implement target build/import/runtime smoke validation. Evidence: asset opens and loads in supported target version.
    • B4.8.a Provision the exact supported target host/runtime/toolchain/ plugin version and isolated project fixture. Evidence: environment manifest.
    • B4.8.b Import/open the exact artifact/package through the public adapter and capture warnings/errors, duration, and created object refs. Evidence: import trace.
    • B4.8.c Build/cook/package where required and launch/load in the target runtime with bounded health/crash/log checks. Evidence: runtime trace.
    • B4.8.d Inspect expected objects, dependencies, materials, animation, scripts/assets, and entry scene/state after load. Evidence: smoke assertions.
    • B4.8.e Classify unavailable/incompatible versus asset/build/runtime failure and clean the fixture. Evidence: state matrix.
  • B4.9 Implement round-trip comparison through source→USD→target→USD where supported. Evidence: target matrix.
    • B4.9.a Lock source bytes/revision, source adapter, USD profile, target adapter/host, mappings, configuration, and declared loss. Evidence: round-trip manifest.
    • B4.9.b Convert source to canonical USD, import/write target state, and export/read back to a separately identified canonical USD stage. Evidence: execution trace.
    • B4.9.c Compare source, first USD, target inspection, and returned USD across every applicable typed measure. Evidence: four-way report.
    • B4.9.d Attribute deviations to source conversion, target import, target export, declared loss, or nondeterminism where possible. Evidence: diagnostic attribution.
    • B4.9.e Publish supported/unavailable/unsupported pair cells and prevent one target's evidence from satisfying another. Evidence: target matrix.
  • B4.10 Persist raw measures separately from threshold verdicts. Evidence: policy recomputation.
    • B4.10.a Persist immutable measure set with exact source/target revisions, algorithm/version/config, units/basis, raw values, ranges, uncertainty, and artifact report. Evidence: measure schema.
    • B4.10.b Persist verdict with exact profile/version, thresholds/rules, measurement refs, status/reasons, freshness, and reviewer/waiver refs. Evidence: verdict schema.
    • B4.10.c Re-evaluate unchanged measures under new policy and retain both historical and current verdicts. Evidence: recomputation test.
  • B4.11 Version adapters, DCC/engine versions, validators, profiles, thresholds, render settings, and corpus. Evidence: result provenance.
    • B4.11.a Register immutable version/checksum/compatibility records for every adapter, host/plugin, validator, profile, threshold set, render configuration, and fixture corpus. Evidence: registries.
    • B4.11.b Persist the complete version tuple, runtime/environment, and selected fixture/ground-truth revisions on results. Evidence: provenance query.
    • B4.11.c Detect revoked, stale, expired, unsupported, or mutually incompatible dependencies and enumerate affected evidence/releases. Evidence: impact tests.
  • B4.12 Add unavailable-host, unsupported-feature, low-confidence, insufficient-comparison, and validator-failure states. Evidence: none renders as pass.
    • B4.12.a Define typed non-pass statuses, reason codes, diagnostic fields, retryability, owner, permitted action, and release effect. Evidence: state contract.
    • B4.12.b Emit unavailable-host and validator-failure for infrastructure/ tool inability rather than asset quality failure. Evidence: outage tests.
    • B4.12.c Emit unsupported-feature, insufficient-comparison, and low- confidence with exact affected features/measurements. Evidence: applicability tests.
    • B4.12.d Prove API, UI, automation, waiver, and release paths never coerce a non-pass state to pass. Evidence: fail-closed suite.
  • B4.13 Build representation preview with synchronized source/USD/target camera and selection. Evidence: accessible comparison.
    • B4.13.a Load authorized immutable source/USD/target representations and resolve comparable object mappings, cameras, units/axes, and time. Evidence: preview contract.
    • B4.13.b Synchronize camera, framing, transport/time, variant/layer, object/component selection, isolation, and overlays across views. Evidence: interaction tests.
    • B4.13.c Support side-by-side, overlay/difference, A/B, and target switching while retaining exact revision/profile identity. Evidence: comparison journey.
    • B4.13.d Provide semantic hierarchy/property/measurement/difference tables and keyboard controls equivalent to visual selection. Evidence: accessibility parity.
    • B4.13.e Surface unavailable renderer/host, loading, stale mapping, partial assets, and unsupported previews honestly. Evidence: state matrix.
  • B4.14 Build layer/dependency/metadata/measurement diffs with semantic tables. Evidence: non-visual parity.
    • B4.14.a Select exact comparable revisions and compute stable path/ identity mappings with ambiguity surfaced. Evidence: diff setup tests.
    • B4.14.b Diff layer stack/opinions/composition/variants and dependency add/remove/change/resolve/license state. Evidence: structural fixtures.
    • B4.14.c Diff names/namespaces/metadata/custom properties and typed raw measurement sets/verdicts. Evidence: property fixtures.
    • B4.14.d Present filterable add/remove/change/move/rename/conflict/ unavailable rows with before/after, source, impact, and exact location. Evidence: semantic table.
    • B4.14.e Synchronize table focus with visual object/property/layer and prove every visual difference is represented non-visually. Evidence: parity journey.
  • B4.15 Build image/render diff with calibrated settings and explicit nondeterminism. Evidence: known render corpus.
    • B4.15.a Lock scene/artifact/camera/frame, renderer/version, hardware, OCIO, resolution, sampling, lights, random seeds, and output transforms. Evidence: render manifest.
    • B4.15.b Render source/reference and target with deterministic settings where supported and record stochastic/runtime variation otherwise. Evidence: repeated-render study.
    • B4.15.c Align images and compute raw absolute/relative/color/perceptual/ edge/region metrics with masks and uncertainty. Evidence: analytical image fixtures.
    • B4.15.d Calibrate tolerances against the known corpus and distinguish structural conversion loss from renderer noise. Evidence: calibration report.
    • B4.15.e Provide accessible image descriptions/region tables and exact settings/limitations with verdict. Evidence: comparison report.
  • B4.16 Build exception/waiver with exact loss, scope, target, expiry, and approver. Evidence: incompatible release cannot reuse waiver.
    • B4.16.a Define waivable and prohibited gates/losses plus required role, independent approver, rationale, mitigation, expiry, and evidence. Evidence: waiver policy.
    • B4.16.b Bind a request to exact source/target/artifact/profile/measure/ verdict/loss/target-purpose revisions and bounded objects/features. Evidence: waiver schema.
    • B4.16.c Review, approve/reject/condition, sign, audit, notify, and surface the exception in preview/package/release evidence. Evidence: workflow test.
    • B4.16.d Invalidate on any material input, target/version, measurement, policy, scope, purpose, or expiry change. Evidence: reuse negative test.

B5 — Artifact library, builds, evidence, and release#

  • B5.1 Build artifact library search by asset type, representation, version, source/target, rights, quality, lineage, status, project, and owner. Evidence: permission-aware query tests.
    • B5.1.a Define the searchable projection and approved fields from authoritative artifact/representation/rights/quality/lineage/lifecycle records. Evidence: index contract.
    • B5.1.b Enforce tenant/project/object/property/rights authorization before query, facets, counts, suggestions, snippets, and previews. Evidence: inference tests.
    • B5.1.c Implement exact/full-text search, filters, facets, sort, pagination, and version/source-target/dependency queries. Evidence: query fixtures.
    • B5.1.d Reconcile index state after new version, supersession, quarantine, revoke, retirement, restore, and deletion. Evidence: lifecycle tests.
  • B5.2 Build exact and structural dedupe with non-destructive merge decisions. Evidence: retained aliases and lineage.
    • B5.2.a Detect exact duplicates from normalized bytes/checksums while preserving distinct ownership, rights, and source records. Evidence: exact fixtures.
    • B5.2.b Detect structural candidates from versioned hierarchy/geometry/ material/rig/dependency signatures with compatibility and uncertainty. Evidence: structural fixtures.
    • B5.2.c Compare representations, versions, rights/licenses, quality, lineage, dependencies, releases, and exact differences. Evidence: review journey.
    • B5.2.d Record keep-both, relate, alias, or consolidate-reference decisions without deleting immutable artifacts or provenance. Evidence: non-destructive tests.
  • B5.3 Build artifact version, branch, supersession, deprecation, quarantine, retirement, and restore. Evidence: lifecycle tests.
    • B5.3.a Define immutable versions, mutable branch pointers, parent/ ancestry, merge provenance, semantic diff, and optimistic concurrency. Evidence: version fixtures.
    • B5.3.b Define distinct superseded/deprecated/quarantined/retired/ deleted/restored effects on visibility, selection, execution, release, dependencies, and retention. Evidence: state machine.
    • B5.3.c Authorize transitions with reason, effective time, replacement, impact preview, audit, and downstream propagation. Evidence: transition tests.
    • B5.3.d Restore only eligible retained bytes/metadata into a new current revision/state and revalidate dependencies/rights/quality. Evidence: restore journey.
  • B5.4 Build collections, kits, dependencies, related assets, and release sets through versioned curation. Evidence: bundle golden.
    • B5.4.a Define typed versioned curation operations for membership, ordering, roles, kits, dependency pins, related edges, metadata, and release-set state. Evidence: bundle schema.
    • B5.4.b Validate refs/versions, cycles, compatibility, rights/license, quality, lifecycle, permissions, and release completeness. Evidence: preflight tests.
    • B5.4.c Review and apply an exact bundle atomically/idempotently with concurrency, per-operation result, audit, and compensation. Evidence: bundle golden.
    • B5.4.d Read back collection/kit/dependency/release projections and prohibit direct production writes outside the applier. Evidence: authority tests.
  • B5.5 Build build-template registry, typed parameters, lifecycle, and compatibility. Evidence: template tests.
    • B5.5.a Register immutable template versions with owner/purpose, target/ profile ranges, toolchain/dependency locks, typed parameters, resources, stages, gates, and output contracts. Evidence: registry tests.
    • B5.5.b Validate defaults/bounds/conditional parameters, secret refs, command allowlists, host/target compatibility, and reproducibility. Evidence: schema fixtures.
    • B5.5.c Review, approve, activate, deprecate, supersede, revoke, and migrate templates with immutable historical use. Evidence: lifecycle journey.
    • B5.5.d Compute affected builds/releases and block selection of incompatible, revoked, or stale templates. Evidence: impact tests.
  • B5.6 Build batch/build scheduling with host, worker, license, storage, render, and reviewer capacity. Evidence: reservation conflict tests.
    • B5.6.a Model resource pools/capabilities/locality, calendars/ maintenance, quotas, licenses/tokens, storage/egress, reviewers, and time zones. Evidence: capacity schema.
    • B5.6.b Estimate exact per-stage demand, duration, dependencies, affinity, deadlines, cost, and alternative compatible resources. Evidence: scheduling inputs.
    • B5.6.c Create tentative holds and committed fenced reservations transactionally across all required resource types. Evidence: reservation tests.
    • B5.6.d Detect overlap, version/capability, quota, license, storage, dependency, setup, and reviewer conflicts. Evidence: conflict matrix.
    • B5.6.e Reschedule/cancel/expire/release reservations and reconcile queued/running work without leaks or double allocation. Evidence: lifecycle tests.
  • B5.7 Build batch preview with eligibility, losses, cost, capacity, dependencies, and gates. Evidence: mixed-result journey.
    • B5.7.a Resolve a stable artifact/query snapshot, requested template/ parameters/targets, caller authority, and current policy. Evidence: batch contract.
    • B5.7.b Evaluate per-item version/rights/lifecycle/dependency/format/ host/profile/gate compatibility and expected semantic loss. Evidence: eligibility fixtures.
    • B5.7.c Estimate per-item and aggregate stage/resource/license/storage/ duration/cost/capacity with uncertainty. Evidence: estimate tests.
    • B5.7.d Present included/skipped/blocked/warning items with exact reasons, dependencies, losses, gates, and schedule options. Evidence: preview journey.
    • B5.7.e Execute only an explicitly approved exact preview and report per-item progress/partial/failure/compensation. Evidence: mixed-result test.
  • B5.8 Build release candidate locked to exact artifact, layer, dependency, template, toolchain, host, adapter, validator, gate, and approval versions. Evidence: mutation invalidation.
    • B5.8.a Resolve and lock the full transitive artifact/layer/dependency/ mapping/template/parameter graph with versions/checksums. Evidence: candidate manifest.
    • B5.8.b Lock toolchain/environment/host/plugin/adapter/validator/profile/ threshold/render/build target versions. Evidence: execution lock.
    • B5.8.c Lock raw measurements/verdicts, rights/licenses, security/SBOM, approvals/waivers, limitations, and target plan. Evidence: assurance lock.
    • B5.8.d Detect any changed, stale, revoked, expired, missing, quarantined, or incompatible lock dependency and invalidate readiness. Evidence: mutation tests.
    • B5.8.e Rebuild a new candidate rather than updating a locked one. Evidence: immutability test.
  • B5.9 Build tamper-evident release-evidence assembly. Evidence: deterministic hash and signature validation.
    • B5.9.a Define canonical manifest ordering/serialization and include candidate locks, artifacts, commands, builds, gates, approvals, limitations, targets, and read-back. Evidence: manifest schema.
    • B5.9.b Resolve every evidence ref to immutable bytes/checksum or signed external record and reject missing/ambiguous refs. Evidence: assembly tests.
    • B5.9.c Build deterministically, compute manifest/Merkle or equivalent integrity hashes, sign with authorized release key, and timestamp. Evidence: repeated-build hash test.
    • B5.9.d Independently verify schema, refs/checksums, signature/key status, authorization, and candidate identity. Evidence: verifier suite.
    • B5.9.e Preserve superseded/revoked evidence and publish status without allowing artifact replacement under the same ref. Evidence: tamper tests.
  • B5.10 Build target export/deploy with plan, execute, read-back, and compensate. Evidence: partial failure recovery.
    • B5.10.a Build a read-only target plan with exact candidate, destination, files/objects/actions, prerequisites, permissions, conflicts, cost, downtime, and rollback/compensation. Evidence: plan fixture.
    • B5.10.b Revalidate candidate/target/credentials/capacity immediately before execution and authorize the exact plan revision. Evidence: stale-plan tests.
    • B5.10.c Execute typed idempotent steps with checkpoints, target revisions, bounded logs, progress, cancel, and immutable attempt audit. Evidence: deployment tests.
    • B5.10.d Read back target objects/files/versions/checksums/state through an independent inspection path. Evidence: verification report.
    • B5.10.e On partial failure, stop safely and retry/roll back/compensate per step while exposing unresolved target state. Evidence: recovery journey.
  • B5.11 Build target-specific post-deploy smoke and fidelity checks. Evidence: transport success alone cannot release.
    • B5.11.a Select the exact target/version/profile and its required import/ load/build/runtime/scenario/fidelity check registry. Evidence: target test plan.
    • B5.11.b Execute target-native open/import/load and dependency/ configuration/entry-state smoke checks. Evidence: smoke results.
    • B5.11.c Execute target-specific structural, semantic, visual/runtime, performance, and security checks with raw measurements. Evidence: fidelity results.
    • B5.11.d Distinguish unavailable infrastructure from deployment or artifact failure and prevent transport-only success from release. Evidence: state tests.
    • B5.11.e Attach exact target read-back and current verdicts to release evidence. Evidence: evidence linkage.
  • B5.12 Build rollback, supersession, revoke, quarantine, and recall. Evidence: dependent projects and targets reconcile.
    • B5.12.a Define distinct source/library/build/release/target/consumer effects, authorities, prerequisites, and terminal criteria for each action. Evidence: lifecycle policy.
    • B5.12.b Traverse exact artifacts/versions/dependencies/builds/packages/ deployments/projects/targets/consumers and preview blast radius. Evidence: impact graph.
    • B5.12.c Execute the authorized saga with fencing, checkpoints, per- dependency result, retry/compensation, notifications, and audit. Evidence: propagation tests.
    • B5.12.d Verify target and dependent-project state by read-back and keep partial/unreachable recipients unresolved. Evidence: reconciliation report.
    • B5.12.e Restore or supersede only through a new validated candidate and approval path. Evidence: recovery tests.
  • B5.13 Build artifact retention and purge with dependency/legal-hold preview. Evidence: referenced artifacts remain.
    • B5.13.a Calculate retention/expiry eligibility from artifact class, project/rights/license, release/evidence policy, deletion request, and legal hold. Evidence: policy tests.
    • B5.13.b Traverse versions, branches, dependencies, packages, releases, targets, caches, backups, and audit/evidence refs before purge. Evidence: dependency preview.
    • B5.13.c Classify deletable, retained, tombstoned, anonymized, blocked, and separately approved objects with exact reasons. Evidence: purge plan.
    • B5.13.d Execute authorized bounded purge with checkpoints, per-store confirmation, immutable tombstone/evidence, and retry. Evidence: purge journey.
    • B5.13.e Prove referenced or legally held bytes remain and unreferenced eligible bytes become inaccessible across active stores/caches. Evidence: read-back tests.
  • B5.14 Build release notifications and audit export with safe redaction. Evidence: authorization tests.
    • B5.14.a Define notification event types, minimal payloads, recipients/ subscriptions, severity, locale, dedupe, and escalation. Evidence: notification contract.
    • B5.14.b Deliver release/failed/revoked/recalled/partial state through governed channels with retries, preference/policy handling, and audit. Evidence: delivery tests.
    • B5.14.c Define audit export scope, query filters, exact time/object refs, schema/version, pagination/chunking, signature, retention, and purpose. Evidence: export contract.
    • B5.14.d Authorize object/property visibility and redact secrets, credentials, local paths, content, sensitive metadata, and unauthorized existence. Evidence: redaction matrix.
    • B5.14.e Verify recipient/download identity, expiry, checksums, and reproducible query manifest. Evidence: authorization tests.
  • B5.15 Build UI-governed asset, representation, schema, material, rig, animation, target, build, fidelity, and catalog taxonomies over code-owned floors. Evidence: versioned review, alias, deprecation, and migration journey.
    • B5.15.a Define independent taxonomy schemas, owners, code-owned invariants/floors, allowed relations, and consumers. Evidence: registry.
    • B5.15.b Author draft terms with stable IDs, localized labels/ definitions, aliases, mappings, constraints, and migration notes. Evidence: authoring tests.
    • B5.15.c Review, approve, publish, deprecate, supersede, retire, and restore versions without reinterpreting historical records. Evidence: lifecycle journey.
    • B5.15.d Reject collisions, cycles, incompatible mappings, floor violations, and referenced-term deletion. Evidence: validation suite.
    • B5.15.e Migrate refs only through previewed, authorized, resumable jobs with per-record result and rollback/compensation. Evidence: migration tests.
  • B5.16 Build taxonomy/profile impact over templates, sessions, artifacts, validators, search, builds, targets, and releases. Evidence: read-only preview precedes an authorized migration job.
    • B5.16.a Build a complete versioned reference index from taxonomy/profile concepts to templates, active sessions, artifacts, validators, search, builds, targets, releases, docs, and clients. Evidence: reference audit.
    • B5.16.b Preview changed eligibility/mapping/validation/search/runtime/ release behavior, incompatible refs, active-operation risk, and cost without mutation. Evidence: impact report.
    • B5.16.c Require authorized approval of exact source/destination versions and preview revision before scheduling. Evidence: stale-preview test.
    • B5.16.d Execute/checkpoint/resume/retry/partially fail/compensate the migration and read back all projections. Evidence: migration journey.

B5A — Specialized production, runtime, and automation operations#

  • B5A.1 Complete per-host typed command registries for Blender, Maya, Houdini, 3ds Max, Unreal, Unity, Godot, and DaVinci, including selection, scene/entity, asset/import/export, material, rig/animation, camera/light, render/build, undo/transaction, diagnostics, capability, and version behavior. Evidence: no generic "bridge works" proof substitutes for command-by-command host conformance.
    • B5A.1.a Define the shared command-definition/conformance schema with exact host/plugin versions, typed inputs/results, prerequisites, permissions, side effects, undo/compensation, limits, and diagnostics. Evidence: registry schema.
    • B5A.1.b Complete and validate Blender and Maya command registries across every named command category. Evidence: independent host matrices.
    • B5A.1.c Complete and validate Houdini and 3ds Max command registries across every named command category. Evidence: independent host matrices.
    • B5A.1.d Complete and validate Unreal and Unity command registries across every named command category. Evidence: independent host matrices.
    • B5A.1.e Complete and validate Godot and DaVinci command registries across applicable command categories with explicit unsupported cells. Evidence: independent host matrices.
    • B5A.1.f Test each host-command-version row for success, invalid input, unsupported capability, authorization, stale revision, idempotency, cancellation, compensation, crash/reconnect, and audit. Evidence: command-by-command conformance.
    • B5A.1.g Prohibit generic bridge health from satisfying any untested command row. Evidence: release coverage gate.
  • B5A.2 Build GPU render-queue operations for scene/version intake, renderer/profile, frame ranges, passes, dependencies, color, resources, licenses, priority/fairness, preview, cost, progress, pause/cancel/retry, checkpoint, output QC, review, and publish. Evidence: partial-frame and worker/host-loss recovery.
    • B5A.2.a Lock scene/camera/layer/variant/dependency revisions and select renderer/version/profile, frame/range/handles, passes, outputs, OCIO, and destination. Evidence: render manifest.
    • B5A.2.b Validate dependencies, rights, renderer/plugin compatibility, target/output rules, and pre-render scene/color gates. Evidence: preflight suite.
    • B5A.2.c Estimate and preview frames/tasks, GPU/CPU/memory/storage/ bandwidth/licenses, priority/fairness, schedule, duration, and cost. Evidence: queue plan.
    • B5A.2.d Reserve capacity and dispatch idempotent frame/pass tasks with fenced leases, checkpoint/resume, bounded retries, and backpressure. Evidence: scheduler tests.
    • B5A.2.e Stream bounded progress/thumbnails/logs and support authorized pause/resume/cancel/retry/reprioritize without hidden late output. Evidence: operations journey.
    • B5A.2.f Aggregate ordered frames/passes, detect missing/duplicate/corrupt output, verify color/naming/format/checksums, and run calibrated QC. Evidence: output-quality suite.
    • B5A.2.g Review/approve/publish exact outputs with artifact lineage and target read-back. Evidence: release journey.
    • B5A.2.h Recover worker/host/GPU/license/storage loss and partial frames/ passes without false completion. Evidence: fault drill.
  • B5A.3 Build gameplay-system operations for input/action maps, save/load schemas and migrations, inventory/items, combat/damage, AI behavior/ navigation, validation scenarios, target builds, runtime telemetry, and rollback. Evidence: deterministic or explicitly stochastic engine tests for every supported target/version.
    • B5A.3.a Register versioned gameplay schemas/configuration for input/ actions, saves, inventory/items, combat/damage, behavior/AI, navigation, and state with one authoritative source. Evidence: gameplay registry.
    • B5A.3.b Author changes through typed commands with exact project/object revisions, semantic diff, validation, authorization, and undo/ compensation. Evidence: mutation tests.
    • B5A.3.c Implement save upgrade/downgrade/rollback fixtures and validate integrity, compatibility, data preservation, and failure recovery. Evidence: save migration matrix.
    • B5A.3.d Define allowlisted deterministic test scenarios with seeded or explicitly stochastic setup, actions, timing, assertions, cleanup, and telemetry. Evidence: scenario schema.
    • B5A.3.e Build/package/deploy every supported engine target/version and execute input, save, inventory, combat, AI, navigation, and state scenarios independently. Evidence: target matrix.
    • B5A.3.f Capture bounded runtime correctness/performance telemetry and target read-back tied to source/build revisions. Evidence: result manifests.
    • B5A.3.g Roll back build/config/save migrations and prove prior target state remains viable. Evidence: rollback drill.
  • B5A.4 Build virtual-production operations for camera/lens tracking, calibration, LED-wall configuration, ICVFX layers/frusta, genlock/ timecode, playback, take/recording, health, latency, color, failover, emergency stop, and dailies handoff. Evidence: stage-safe simulation and supported-device/host matrix.
    • B5A.4.a Enroll and map stage, camera/tracker/lens, wall/processor, render nodes, clocks, recording, and safety devices to supported profiles. Evidence: device inventory.
    • B5A.4.b Calibrate camera/tracker/lens/world/wall/color systems with versioned measurements, uncertainty, validity, drift, and approval. Evidence: calibration suite.
    • B5A.4.c Configure and preview exact content/layer/variant, camera, frustum, wall/color, genlock/timecode, playback, recording, and resource plan. Evidence: session plan.
    • B5A.4.d Start/supervise/stop playback and takes with synchronized state, health/latency/color/sync warnings, bounded controls, and immutable take identity. Evidence: supervision journey.
    • B5A.4.e Enforce device-safe interlocks, failover/fallback, emergency stop, late command fencing, and incident evidence. Evidence: stage-safe simulation.
    • B5A.4.f Package recorded media/metadata/calibration/timecode/content lineage and hand off dailies through governed editorial seams. Evidence: dailies read-back.
    • B5A.4.g Validate each supported device/host/version combination independently with unavailable cells explicit. Evidence: matrix.
  • B5A.5 Build XR operations for shared capabilities plus visionOS, Meta Quest, and every ratified platform: project settings, spatial assets, interactions/input, comfort/accessibility, device/runtime compatibility, performance budgets, package/sign, install, launch, smoke, telemetry, and rollback. Evidence: platform-specific validation; no inherited desktop build proof.
    • B5A.5.a Register each XR platform/device/OS/runtime/SDK profile, capabilities, entitlements, project/package/signing requirements, performance/comfort/accessibility budgets, and test devices. Evidence: XR registry.
    • B5A.5.b Validate and configure project settings, render path, permissions, spatial assets/anchors, input/actions, interaction/locomotion, and accessibility per exact profile. Evidence: project tests.
    • B5A.5.c Preview device/runtime compatibility, unsupported features, entitlements, assets, build settings, secrets/signing, cost, and target limitations. Evidence: deployment plan.
    • B5A.5.d Build/sign/package, install, launch, and inspect exact versions on visionOS, Meta Quest, and every other ratified platform separately. Evidence: platform journeys.
    • B5A.5.e Run platform-native input/tracking/spatial/interaction, comfort, accessibility, frame-time/memory/thermal, lifecycle, and offline smoke scenarios. Evidence: validation matrix.
    • B5A.5.f Collect bounded telemetry/read-back and roll back/uninstall/ restore prior target state. Evidence: rollback tests.
    • B5A.5.g Prohibit desktop or another XR platform's evidence from satisfying an untested profile. Evidence: independence gate.
  • B5A.6 Build motion-capture operations for vendor streams, clocks, skeleton mapping, subject/consent refs, retargeting, recording/playback, format conversion, animation export, health, dropouts, and Aja handoff. Evidence: consent/provenance and live network fault journeys.
    • B5A.6.a Register vendor/device/firmware/protocol/plugin profiles, modalities, skeleton/channel schemas, clock/sync, capability, and network/security requirements. Evidence: mocap registry.
    • B5A.6.b Enroll session/devices/subjects with current consent/purpose, operator, location, calibration, skeleton mapping, target rig, and Aja handoff scope. Evidence: admissibility tests.
    • B5A.6.c Connect and monitor signed streams, clock skew, packet order/ loss, latency, confidence, health, dropouts, reconnect, and modality degradation. Evidence: live supervision suite.
    • B5A.6.d Record immutable raw and normalized streams with timecode, subject-safe refs, configuration, mappings, checksums, and interruption markers; support synchronized playback. Evidence: recording tests.
    • B5A.6.e Retarget/convert/export through versioned mappings and formats with loss preview, quality/consent gates, provenance, and read-back. Evidence: conversion journey.
    • B5A.6.f Handoff to Aja only through its canonical contract with idempotency, health/degraded states, ownership, and recall propagation. Evidence: seam tests.
    • B5A.6.g Inject network/device/clock/host loss, consent revocation, and late packets; prove explicit partial state and safe recovery. Evidence: fault journeys.
  • B5A.7 Build MetaHuman operations for mesh/identity import, face rig, body/customization, skeleton/material compatibility, Live Link, animation, rights/likeness, versioning, performance, package, and Unreal validation. Evidence: identity/consent and supported-Unreal-version gates.
    • B5A.7.a Register supported Unreal/MetaHuman/plugin/DNA/rig/Live Link versions and exact host/tool compatibility. Evidence: profile registry.
    • B5A.7.b Admit mesh/identity/reference sources only with current rights, likeness/subject consent, intended use, retention, and protected-data controls. Evidence: identity gate.
    • B5A.7.c Import/create immutable MetaHuman identity, mesh, DNA, face/body rig, skeleton, materials, clothing/customization, and dependency versions with lineage. Evidence: asset manifest.
    • B5A.7.d Validate geometry/deformation, skeleton/animation/blend shapes, material/clothing, target platform, and current Unreal compatibility. Evidence: compatibility suite.
    • B5A.7.e Configure/supervise Live Link identity/source/subjects, channels/timecode, confidence, recording, animation transfer, dropouts, and revoke. Evidence: live journey.
    • B5A.7.f Measure editor/runtime frame time, memory, LOD, animation/ rendering quality, and target budgets. Evidence: performance report.
    • B5A.7.g Package/build/open/read back in every supported Unreal version and propagate rights/consent/version quarantine or recall. Evidence: release gates.
  • B5A.8 Build DaVinci operations for project/timeline exchange, media/ proxy, edit/conform, color/OCIO, markers, audio, render/deliver, relink, version, and round-trip diagnostics. Evidence: locked timeline and media-hash fixtures.
    • B5A.8.a Register supported Resolve/version/platform/project/timeline/ interchange/color/audio/render profiles and plugin capabilities. Evidence: compatibility registry.
    • B5A.8.b Lock project/timeline/bin/clip/media revisions, hashes, timecode/rates, proxies, edit decisions, markers, audio, and OCIO/color context before exchange. Evidence: timeline manifest.
    • B5A.8.c Import/export/conform through typed commands with mapping, handles, transitions/effects, speed changes, multicam, grades, audio, and unsupported-loss preview. Evidence: conform fixtures.
    • B5A.8.d Relink originals/proxies by approved identity/hash/path mapping and reject ambiguous, missing, or changed media. Evidence: relink tests.
    • B5A.8.e Render/deliver exact timeline/ranges/presets with color/audio/ subtitle/data-level metadata, checksums, review, and read-back. Evidence: delivery journey.
    • B5A.8.f Round-trip and compare timeline structure, timing, media hashes, markers, color/OCIO, audio, and declared loss. Evidence: diagnostic matrix.
  • B5A.9 Build Houdini operations for HDA definition/version/parameter, dependency, cook/cache, geometry/USD, simulation, output, farm execution, diagnostics, and round-trip. Evidence: deterministic seed/toolchain or explicit nondeterminism and license/capacity handling.
    • B5A.9.a Register Houdini/Engine/plugin/renderer/HDA versions, platforms, licenses, farm/runtime compatibility, and supported node/format capabilities. Evidence: profile registry.
    • B5A.9.b Ingest and verify HDA definition/library/version/signature, typed parameters/defaults/expressions, dependencies, permissions, and source provenance. Evidence: HDA contract.
    • B5A.9.c Plan cook/simulation with exact input geometry/USD, parameter revision, frames, seeds, caches, environment/toolchain, resources, licenses, outputs, cost, and nondeterminism. Evidence: cook plan.
    • B5A.9.d Execute locally or on farm with fenced work items, checkpoint/ cache, progress, pause/cancel/retry, worker loss, and license backpressure. Evidence: execution tests.
    • B5A.9.e Validate output geometry/USD/simulation fields, dependencies, frames, checksums, known values, diagnostics, and declared variation. Evidence: output QC.
    • B5A.9.f Round-trip supported inputs/outputs and prove deterministic replay by checksum/measure or document bounded nondeterminism. Evidence: replay report.
  • B5A.10 Build audio/video asset operations for ingest, transcode, proxy, color/audio metadata, sync/timecode, renditions, engine/DCC attachment, review, package, and target validation. Evidence: A/V fidelity and rights lineage survive conversions.
    • B5A.10.a Ingest immutable A/V bytes with type/profile, checksum, source, rights/license/consent, streams, timecode, language, and provenance. Evidence: intake tests.
    • B5A.10.b Inspect codec/container/profile, resolution/rate/color/HDR, audio channels/rate/loudness/layout, captions/metadata, duration, and corruption. Evidence: media report.
    • B5A.10.c Plan and execute versioned transcode/proxy/rendition operations with exact tools/config, sync/timecode policy, color/audio transforms, loss, resources, and lineage. Evidence: conversion fixtures.
    • B5A.10.d Attach exact renditions to engine/DCC scene/timeline/assets through portable dependencies and typed commands. Evidence: attachment tests.
    • B5A.10.e Review synchronized picture/audio/captions/timecode and raw fidelity measurements with accessible alternatives. Evidence: review journey.
    • B5A.10.f Package and read back in each target, verifying bytes/streams/ metadata/sync/color/audio and rights lineage. Evidence: target matrix.
  • B5A.11 Build automatic cross-domain ingestion and lore-to-engine/ text-to-3D compilation for Yemaya, Isis, Aja, Hathor, and other ratified producers through public contracts. Evidence: source entities, generated ingredients, consent/rights, compiler/toolchain, mapping decisions, artifacts, and target read-back remain traceable.
    • B5A.11.a Register each producer contract/version/capability, authoritative source entities/revisions, artifact types, rights/consent/canon fields, events, and ownership boundary. Evidence: producer matrix.
    • B5A.11.b Admit signed authorized immutable manifests idempotently and reject stale, revoked, unsupported, cross-tenant, or incomplete inputs. Evidence: intake contract tests.
    • B5A.11.c Normalize source entities and generated ingredients into versioned intermediate/target mappings with explicit alternatives, omissions, defaults, and human decisions. Evidence: mapping records.
    • B5A.11.d Compile through pinned compiler/toolchain/templates/adapters/ models with seeds/config/environment, staged outputs, validation, and deterministic or disclosed-nondeterministic lineage. Evidence: compilation trace.
    • B5A.11.e Package/deploy and read back target entities/assets/state while retaining source IDs/revisions, canon, rights/consent, generated regions, mappings, and artifact checksums. Evidence: target journey.
    • B5A.11.f Propagate producer correction/revocation/recall and reconcile every compiled derivative/target. Evidence: downstream closure test.
  • B5A.12 Build governed agent/MCP remote-control planning with natural-language intent translated into typed allowlisted commands, exact target/revision, dry-run diff, risk/cost/blast radius, required approval, least-privilege session, step execution, stop conditions, and immutable tool-call evidence. Evidence: there is no raw shell or unreviewed autonomous destructive path.
    • B5A.12.a Define supported agent/MCP identities, clients, model/tool versions, scopes, tenants, budgets, data policy, and prohibited capabilities. Evidence: agent policy.
    • B5A.12.b Parse intent into an immutable structured goal, assumptions, exact target/object/revision, constraints, success criteria, and uncertainty; require clarification on material ambiguity. Evidence: intent tests.
    • B5A.12.c Compile only to versioned typed allowlisted commands with explicit dependencies, expected revisions, permissions, idempotency, deadlines, compensation, and no shell/eval escape. Evidence: planner contract.
    • B5A.12.d Dry-run every step and present semantic diff, affected objects/ paths, fidelity/security/release risk, cost/spend, capacity, blast radius, irreversible effects, and alternatives. Evidence: plan preview.
    • B5A.12.e Determine required human approvals/step-up/dual control and bind them to the exact immutable plan revision. Evidence: approval tests.
    • B5A.12.f Open a time/resource/command/object/path-bounded least- privilege host session and revalidate preconditions before each step. Evidence: grant tests.
    • B5A.12.g Execute sequential/parallel steps under declared dependencies, budgets, stop conditions, live status, cancellation, fencing, and human redirect checkpoints. Evidence: execution journey.
    • B5A.12.h Persist intent, plan/revisions, model/tool versions, approvals, normalized tool calls/results, diffs, costs, decisions, and final read- back with redaction. Evidence: immutable audit.
  • B5A.13 Build agent pause, redirect, cancel, budget/spend kill, organization policy, host quarantine, replay, and partial-result reconciliation. Evidence: agents cannot waive fidelity/security/release gates or disguise a failed/partial command as complete.
    • B5A.13.a Define control authority and semantics for pause/resume, redirect/replan, cancel, spend kill, policy stop, and host quarantine. Evidence: control state machine.
    • B5A.13.b Enforce organization/tenant/project/model/tool/host/data/ network/time/token/cost policies on plan admission and before each step. Evidence: policy tests.
    • B5A.13.c Pause at safe boundaries, redirect only through a new diff/ risk/approval plan revision, and cancel with typed compensation/unknown outcomes. Evidence: intervention tests.
    • B5A.13.d Stop dispatch immediately on budget/spend limit or quarantine, fence late results, release reservations, and surface incurred cost/ side effects. Evidence: kill-switch tests.
    • B5A.13.e Replay only eligible idempotent steps from locked evidence and require new approval when inputs/policy/risk changed. Evidence: replay tests.
    • B5A.13.f Reconcile every step/target by read-back and report succeeded, failed, cancelled, skipped, compensated, partial, or unknown separately. Evidence: partial-result journey.
    • B5A.13.g Enforce fidelity/security/rights/release gates outside agent authority and reject attempts to waive or relabel them. Evidence: adversarial suite.
  • B5A.14 Complete package-builder and export operations for engine/DCC/XR/ VP targets, native dependencies, plugins, content, configuration, licenses, SBOM, provenance, signatures, install/upgrade/uninstall, target smoke, and rollback. Evidence: reproducible release-evidence bundle per supported target.
    • B5A.14.a Register target-specific package formats/layouts, toolchains, native/runtime dependencies, plugins, content/config, permissions, signing/notarization, install lifecycle, and validators. Evidence: target profiles.
    • B5A.14.b Resolve and lock exact source artifacts, generated/native dependencies, plugins, configs, licenses, rights, target, toolchain, and build environment. Evidence: dependency manifest.
    • B5A.14.c Scan/validate components and generate normalized licenses/ notices, SBOM, vulnerability results, provenance attestations, and policy verdicts. Evidence: supply-chain bundle.
    • B5A.14.d Build/package reproducibly where supported, compute per-file/ package checksums, sign/notarize, and independently verify signatures. Evidence: reproducibility test.
    • B5A.14.e Install fresh, upgrade supported prior versions, preserve/ migrate state, repair, uninstall cleanly, and handle dependency/ permission failures. Evidence: lifecycle matrix.
    • B5A.14.f Launch/open and run target-specific smoke/fidelity/performance/ security checks with read-back. Evidence: target journey.
    • B5A.14.g Roll back failed install/upgrade/release and prove prior target viability with no orphaned partial state. Evidence: rollback drill.
    • B5A.14.h Assemble one immutable release-evidence bundle per exact target profile/version; prohibit inherited proof. Evidence: coverage gate.
  • B5A.15 Complete event-handler/publisher operations for asset, session, command, interchange, render, build, deployment, runtime, and recall events with schema/version registry, subscriptions, replay, dead letter, redaction, and consumer status. Evidence: cross-domain event conformance and no duplicate side effects.
    • B5A.15.a Register versioned minimal event schemas with stable event/ aggregate IDs, revisions, ordering key, time, tenant, trace/causation, data classification, and compatibility. Evidence: event registry.
    • B5A.15.b Publish asset/session/command/interchange/render/build/ deployment/runtime/recall events transactionally from authoritative state with no protected payload. Evidence: outbox tests.
    • B5A.15.c Create authorized subscriptions with event/filter/tenant scope, endpoint/topic, secret, delivery ordering, rate, retention, and lifecycle. Evidence: subscription tests.
    • B5A.15.d Sign and deliver with backoff, replay protection, idempotency, bounded payload/logging, secret rotation, and explicit consumer acknowledgement/status. Evidence: delivery suite.
    • B5A.15.e Dead-letter exhausted/invalid events with diagnostics and authorize bounded replay by exact range/version while preserving event identity. Evidence: replay journey.
    • B5A.15.f Reconcile consumer projections and prove duplicates/reordering/ retries do not duplicate commands, builds, deployments, releases, or recalls. Evidence: conformance matrix.
  • B5A.16 Complete TypeScript, Python, C++, CLI, remote-extension, and host plugin developer surfaces for capability discovery, auth/session, commands, assets, jobs, events, examples, diagnostics, compatibility, versioning, and deprecation. Evidence: generated public contracts and external-client conformance.
    • B5A.16.a Generate shared schemas/types and implement consistent endpoint, authentication, session, version negotiation, timeout/retry, idempotency, error, logging/redaction, and telemetry foundations. Evidence: core parity tests.
    • B5A.16.b Implement capability/format/host discovery and auth/enrollment/ session lifecycle across TypeScript, Python, C++, CLI, remote extension, and host plugin. Evidence: discovery/session matrix.
    • B5A.16.c Implement typed command plan/execute/status/cancel/result, artifact transfer, job/batch, validation/release, and event subscription across applicable surfaces. Evidence: operation matrix.
    • B5A.16.d Provide executable examples for connect, inspect, interchange, build/render, release/read-back, failure recovery, and revoke without embedded secrets or unsafe commands. Evidence: example tests.
    • B5A.16.e Provide bounded diagnostics, support bundle, compatibility inspection, feature detection, and honest unsupported/degraded states. Evidence: diagnostic tests.
    • B5A.16.f Publish versions/changelogs/migration guides/deprecation notices and enforce compatibility windows in clients/servers. Evidence: lifecycle tests.
    • B5A.16.g Run external conformance clients for every supported language/ surface and compare observable behavior against generated contracts. Evidence: public-client report.

B6 — Persistence, APIs, fleet, workers, and integration#

  • B6.1 Decide authoritative stores for hosts, sessions, commands, jobs, artifacts, checks, templates, builds, and releases. Evidence: ownership ADR.
    • B6.1.a Inventory current databases, artifact/evidence stores, indexes, caches, queues, logs, files, and target state for every named entity. Evidence: storage map.
    • B6.1.b Assign one authoritative store/owner per entity/property plus projections, consistency, retention, mutation, and recovery boundaries. Evidence: ownership matrix.
    • B6.1.c Resolve duplicate/shadow authorities, especially artifact bytes/ metadata and release evidence, with migration/compatibility plan. Evidence: decision record.
    • B6.1.d Trace complete lifecycle and downstream reconciliation for each authority. Evidence: approved ADR.
  • B6.2 Add missing schemas and safe migrations without shadowing artifact store or release-evidence data. Evidence: rehearsal.
    • B6.2.a Specify explicit normalized tables/columns/types, keys, constraints, indexes, tenancy, state invariants, audit/outbox refs, and external artifact/evidence refs. Evidence: schema review.
    • B6.2.b Write additive expand migrations and bounded resumable backfills safe under concurrent service/host traffic. Evidence: migration tests.
    • B6.2.c Operate old/new versions through explicit translation or dual behavior without copying authoritative artifact/evidence bytes into shadow stores. Evidence: compatibility suite.
    • B6.2.d Rehearse production-size upgrade, backfill, lock/latency impact, interruption, rollback/roll-forward, and contract cleanup. Evidence: rehearsal report.
  • B6.3 Implement immutable artifacts/checksums and transaction-bound metadata/audit/outbox writes. Evidence: rollback tests.
    • B6.3.a Stage artifact bytes, verify size/type/security/checksum, and atomically promote a content-addressed immutable ref. Evidence: blob tests.
    • B6.3.b Persist metadata/version/dependency/provenance/lifecycle records with exact artifact refs and prohibit ref content replacement. Evidence: immutability tests.
    • B6.3.c Write state mutation, immutable audit event, and transactional outbox event in one database transaction. Evidence: atomicity tests.
    • B6.3.d Reconcile staged/promoted orphan bytes and outbox retries after rollback/crash without phantom metadata or duplicate side effects. Evidence: fault tests.
    • B6.3.e Independently rehash artifact bytes at validation/release/read- back boundaries. Evidence: corruption test.
  • B6.4 Implement host/session/command APIs with mutual auth, idempotency, authorization, deadline, and audit. Evidence: contract suite.
    • B6.4.a Implement enrollment/host identity, capability/health, maintenance/drain/quarantine, and revoke endpoints with mutual authentication. Evidence: host API tests.
    • B6.4.b Implement session open/heartbeat/reconnect/suspend/resume/close/ terminate endpoints with scoped grants and fencing epochs. Evidence: session API tests.
    • B6.4.c Implement typed command preview/submit/status/stream/cancel/ reconcile endpoints against the command registry. Evidence: command API tests.
    • B6.4.d Enforce tenant/actor/host/object/property/path/capability authorization, expected revision, idempotency, and deadlines on every path. Evidence: negative matrix.
    • B6.4.e Persist redacted audit/outbox events and filter lists/errors/ expansions so unauthorized existence cannot leak. Evidence: contract suite.
  • B6.5 Implement interchange/job/check/build/release APIs with optimistic concurrency. Evidence: stale revision tests.
    • B6.5.a Implement interchange plan/create/configure/run/status/rerun/ cancel/result endpoints with immutable job/DAG revisions. Evidence: interchange contracts.
    • B6.5.b Implement measurement/check/verdict/review/waiver endpoints with raw-result separation and freshness. Evidence: check contracts.
    • B6.5.c Implement template/build preview/run/status/cancel/artifact and release candidate/evidence/approve/deploy/recall endpoints. Evidence: build/release contracts.
    • B6.5.d Require expected revision/ETag and idempotency on mutations and return typed conflict/stale/partial/unavailable errors. Evidence: concurrency suite.
    • B6.5.e Authorize object/property/target/approval authority and emit transactional audit/outbox state. Evidence: security tests.
  • B6.6 Implement remote gateway backpressure, connection limits, and circuit breakers. Evidence: load/fault tests.
    • B6.6.a Define per-environment/tenant/host/IP/session limits for handshakes, connections, streams, commands, bytes, rates, buffers, and idle/duration. Evidence: gateway policy.
    • B6.6.b Enforce bounded admission queues, token/leaky-bucket limits, prioritization, fair sharing, and explicit retry-after/degraded states. Evidence: load tests.
    • B6.6.c Propagate streaming backpressure end-to-end and terminate abusive or non-reading peers without unbounded memory. Evidence: slow-peer tests.
    • B6.6.d Trip half-open/closed circuit breakers by dependency/host/profile with jittered recovery and no cross-tenant cascade. Evidence: fault tests.
    • B6.6.e Expose bounded saturation/rejection/breaker telemetry and preserve in-flight command reconciliation. Evidence: overload drill.
  • B6.7 Implement worker/host capability scheduling and anti-starvation. Evidence: queue tests.
    • B6.7.a Persist exact work requirements for host/DCC/plugin/tool/model/ target versions, capabilities, resources/licenses, locality, deadlines, affinity, and safety scope. Evidence: requirement schema.
    • B6.7.b Match only live compatible enrolled workers/hosts and reserve CPU/GPU/memory/storage/license/session capacity atomically. Evidence: matching tests.
    • B6.7.c Apply tenant/class weights, priorities, aging, quotas, concurrency, reservations, and starvation bounds. Evidence: fairness policy.
    • B6.7.d Dispatch with fenced lease/heartbeat and reconcile loss, capability change, maintenance, quarantine, or late completion. Evidence: lifecycle tests.
    • B6.7.e Prove urgent work progresses without starving ordinary work and incompatible work never dispatches. Evidence: queue report.
  • B6.8 Implement artifact upload/download with resumable transfer, checksum, range, authorization, and egress limits. Evidence: interruption and abuse tests.
    • B6.8.a Create an authorized transfer session bound to tenant/project, artifact/version or upload intent, size/type/checksum, purpose, expiry, rate/egress quota, and idempotency. Evidence: transfer schema.
    • B6.8.b Upload numbered checksum-protected chunks with overlap/gap/order/ duplicate handling, bounded storage, expiry, and resume cursor. Evidence: upload interruption tests.
    • B6.8.c Assemble, rehash, scan, and atomically promote only exact complete content; clean abandoned/invalid sessions. Evidence: completion tests.
    • B6.8.d Download authorized immutable ranges/chunks with signed short- lived grants, conditional revision/checksum, bandwidth/egress limits, and resume. Evidence: download tests.
    • B6.8.e Reject cross-tenant refs, enumeration, range abuse, checksum/ size/type mismatch, replay, quota bypass, and revoked artifacts. Evidence: abuse suite.
  • B6.9 Implement permission-aware artifact, host, command, and release search. Evidence: secrets/paths/unreleased names do not leak.
    • B6.9.a Define separate approved index projections and searchable fields for artifacts, hosts, commands, and releases. Evidence: field registry.
    • B6.9.b Exclude secrets/credentials, raw command arguments/log content, local paths, protected metadata, and unauthorized unreleased names. Evidence: index inspection.
    • B6.9.c Enforce tenant/project/host/object/property/rights/release authorization before query, suggestions, filters, facets, counts, and snippets. Evidence: inference tests.
    • B6.9.d Reconcile indexes on permission/lifecycle/version/quarantine/ recall/deletion changes. Evidence: propagation tests.
  • B6.10 Implement signed deep links between Oshun Studio, Bellona control room, Yemaya, and target systems. Evidence: tenant/expiry/replay tests.
    • B6.10.a Register allowed source-destination route/context pairs, audiences, purposes, return paths, and prohibited embedded data/ authority. Evidence: link registry.
    • B6.10.b Mint signed bounded tenant/project/object/revision/action-hint/ expiry/nonce links with issuer/audience/key version. Evidence: signing tests.
    • B6.10.c Validate signature/audience/tenant/expiry/replay/current authorization/object existence/revision/lifecycle at destination. Evidence: negative suite.
    • B6.10.d Render safe forbidden/expired/missing/stale states without existence leakage and never execute a mutation from link context alone. Evidence: route journey.
  • B6.11 Implement target adapter health/capability probes. Evidence: UI uses live state.
    • B6.11.a Define bounded probe contracts for adapter/service/target identity/version, auth, reachability, capabilities/limits, dependencies, capacity/license, latency, and freshness. Evidence: probe schema.
    • B6.11.b Run least-privilege non-mutating probes with timeouts/rate limits/circuit breakers and safe credentials. Evidence: probe tests.
    • B6.11.c Compute healthy/degraded/unavailable/incompatible/ unauthenticated/stale states with component reasons. Evidence: state fixtures.
    • B6.11.d Cache only within declared freshness, emit bounded telemetry, and drive scheduler/API/UI eligibility from the same projection. Evidence: live-state integration test.
  • B6.12 Implement two-tenant/project/host isolation across API, gateway, artifact, search, queue, logs, and export. Evidence: integration suite.
    • B6.12.a Create colliding tenant/project/host IDs, names, checksums, paths, artifacts, commands, jobs, and releases in isolated fixtures. Evidence: isolation corpus.
    • B6.12.b Prove API/list/error/gateway/session/command authorization and connection/resource isolation. Evidence: control-plane suite.
    • B6.12.c Prove artifact/ref/signed URL/cache/search/queue/worker/log/ telemetry/event isolation including count and timing inference. Evidence: data-plane suite.
    • B6.12.d Prove build/export/release/recall/backup/restore/admin operations cannot cross boundaries without explicit governed authority. Evidence: integration suite.
  • B6.13 Generate OpenAPI, remote protocol, SDK, and host consumer contracts. Evidence: drift gate.
    • B6.13.a Generate OpenAPI with stable operations, schemas/errors, auth, pagination, idempotency/concurrency, events, examples, and deprecations. Evidence: spec validation.
    • B6.13.b Generate remote gateway/host protocol message and state-machine contracts with negotiation/capability versions. Evidence: protocol artifacts.
    • B6.13.c Generate SDK types/clients and host plugin/extension consumer types/fixtures for every supported language/surface. Evidence: consumer builds.
    • B6.13.d Compare generated/checked-in server and consumer artifacts and block undocumented breaking drift. Evidence: CI drift gate.
  • B6.14 Implement compatibility telemetry for host/plugin/adapter/target versions. Evidence: unsupported or expiring versions visible.
    • B6.14.a Emit bounded version/capability/support-window observations from enrollment, session, command, job, build, and target probes. Evidence: telemetry schema.
    • B6.14.b Join observations to the approved compatibility registry without logging host secrets, local paths, project names, or unbounded versions. Evidence: privacy tests.
    • B6.14.c Surface supported/degraded/expiring/deprecated/incompatible/ unknown distributions and affected operations/releases. Evidence: compatibility views.
    • B6.14.d Alert owners before expiry and link to upgrade/drain/quarantine/ exception workflows with outcome tracking. Evidence: expiry drill.

B7 — Control-room and Studio experience#

  • B7.1 Refactor the existing control-room hub around sessions, host health, interchange failures, consistency gates, builds, cost, and releases. Evidence: live-data design review.
    • B7.1.a Define operator questions, critical states, prioritized actions, role visibility, and canonical drill-down routes. Evidence: task map.
    • B7.1.b Replace fixture/duplicated summaries with authorized live projections for sessions, hosts, interchange/gates, builds, cost, and releases. Evidence: data contract tests.
    • B7.1.c Compose a restrained task-led workspace with one operational focus, clear hierarchy, and no decorative hero or generic card mosaic. Evidence: frontend review.
    • B7.1.d Preserve tenant/project/time/filter context through drill-down and render loading/empty/stale/partial/unavailable/failed states. Evidence: Playwright journey.
  • B7.2 Build host fleet, session, command, ingest, interchange, artifact, QA, template, batch, schedule, release, and audit routes. Evidence: route integrity.
    • B7.2.a Define canonical route IDs/hierarchy/params/loaders/permissions/ breadcrumbs/titles/deep links and ownership. Evidence: route registry.
    • B7.2.b Implement host/session/command routes with exact enrolled host, grant, revision, and live-state context. Evidence: control route tests.
    • B7.2.c Implement ingest/interchange/artifact/QA routes with immutable source/job/output/measurement context. Evidence: asset route tests.
    • B7.2.d Implement template/batch/schedule/release/audit routes with operation authority and protected evidence boundaries. Evidence: operations route tests.
    • B7.2.e Verify direct load/refresh/history, signed link, not-found, forbidden, stale revision, and permission change on every route. Evidence: Playwright integrity suite.
  • B7.3 Build live bridge-session view with identity, DCC state, grants, heartbeat, latency, load, command history, and termination. Evidence: fault journey.
    • B7.3.a Load immutable session/host/actor/project/purpose/grant/ negotiated-version identity and subscribe to resumable live updates. Evidence: session view contract.
    • B7.3.b Display heartbeat/freshness, latency/clock, DCC/project/modal/ dirty state, resources/load/licenses, and maintenance/quarantine. Evidence: live-state fixtures.
    • B7.3.c Show exact grants, expiry, permitted command/path/resource scope, active reservations, and redacted command/attempt/result history. Evidence: least-privilege view.
    • B7.3.d Preview and authorize suspend/drain/terminate/quarantine actions with in-flight blast radius and step-up. Evidence: control journey.
    • B7.3.e Reconcile disconnect/reconnect, event loss/reorder, expiry, command uncertainty, and late result without false healthy/success. Evidence: fault journey.
  • B7.4 Build interchange DAG with stage status, host/worker, attempts, logs, artifacts, measurements, cache, cost, and rerun. Evidence: live job journey.
    • B7.4.a Load immutable job/DAG/plan and subscribe to ordered resumable stage/attempt updates with snapshot reconciliation. Evidence: live-data tests.
    • B7.4.b Render dependencies, stage/attempt state, host/worker/resources, progress/timing, cache, cost, artifacts, measures, and target results. Evidence: DAG fixtures.
    • B7.4.c Inspect exact inputs/config/toolchain/logs/checkpoints/findings/ outputs/failure/lineage for a selected stage. Evidence: inspector journey.
    • B7.4.d Provide equivalent ordered/dependency semantic table and keyboard selection; do not require graph manipulation. Evidence: accessibility parity.
    • B7.4.e Preview downstream invalidation/loss/cost/capacity and enqueue an authorized exact rerun revision. Evidence: rerun journey.
  • B7.5 Build safe command composer only for allowlisted typed commands. Evidence: raw command shell is absent.
    • B7.5.a Filter commands by current host/plugin/version, negotiated capabilities, session grants, actor authority, project/object context, and host state. Evidence: command eligibility tests.
    • B7.5.b Generate forms from typed schemas with units/bounds/enums/object pickers/path virtual roots/defaults and property-level authorization. Evidence: form tests.
    • B7.5.c Resolve exact target/revision and preview normalized command, semantic diff, dependencies, side effects, cost/risk, reversibility, and required approval. Evidence: dry-run journey.
    • B7.5.d Require step-up/approval where defined and submit idempotently with deadline, expected revision, and immutable audit. Evidence: execution tests.
    • B7.5.e Display progress/result/partial/unknown/cancel/compensation and target read-back. Evidence: command fault journey.
    • B7.5.f Prove no raw shell, script, eval, console, unbounded argument, or arbitrary path/network field exists. Evidence: security assertion.
  • B7.6 Build source/USD/target preview and diff with semantic parity. Evidence: keyboard/screen-reader journey.
    • B7.6.a Load authorized exact source/USD/target representations and map comparable hierarchy/objects/properties/cameras/time. Evidence: preview setup tests.
    • B7.6.b Synchronize camera/framing, variant/layer, selection/isolation, transport, and overlays across available visual previews. Evidence: interaction journey.
    • B7.6.c Display structural, material, rig, animation, metadata, dependency, measurement, and calibrated image differences at exact locations. Evidence: diff fixtures.
    • B7.6.d Provide tree/table/list alternatives with before/after values, units, sources, status, and keyboard focus synchronized to visuals. Evidence: semantic parity.
    • B7.6.e Render unavailable host/renderer, unsupported preview, stale mapping, partial data, and nondeterminism honestly. Evidence: state matrix.
  • B7.7 Build consistency gate panel with raw measures, policy, freshness, remediation, rerun, and authorized waiver. Evidence: state matrix.
    • B7.7.a Show exact source/target revisions, fidelity profile/version, validators, raw measures/units/ranges/uncertainty, and artifact reports. Evidence: evidence view.
    • B7.7.b Show separate policy rules/thresholds/verdict/reasons, freshness/staleness dependencies, limitations, and non-pass states. Evidence: verdict fixtures.
    • B7.7.c Link failures to exact objects/properties/frames/files and allowlisted remediation or rerun plan with impact preview. Evidence: remediation journey.
    • B7.7.d Request/review/approve/reject eligible waivers with exact scope, loss, target, purpose, mitigation, expiry, and approver. Evidence: waiver tests.
    • B7.7.e Prove stale/unavailable/unsupported/low-confidence/validator- failure or expired waiver never renders or releases as pass. Evidence: state matrix.
  • B7.8 Build artifact lineage and dependency inspector. Evidence: impact preview supports quarantine/retire.
    • B7.8.a Traverse authoritative source/version/representation/layer/ dependency/command/job/build/release/target/consumer edges with exact revisions. Evidence: lineage query.
    • B7.8.b Present upstream/downstream hierarchy/graph plus semantic table, edge types, checksums, rights/license, validation, lifecycle, and unresolved refs. Evidence: inspector parity.
    • B7.8.c Filter by relation/depth/status/target and inspect path-level transformation/evidence without leaking unauthorized nodes. Evidence: authorization tests.
    • B7.8.d Preview quarantine/retire/supersede/revoke/recall blast radius, active work, releases, targets, blocked actions, and remediation. Evidence: impact journey.
  • B7.9 Build batch/build/release operations with preview, progress, partial result, retry, cancel, and read-back. Evidence: mixed-result journey.
    • B7.9.a Resolve a stable selected/query item set, exact template/ parameters/targets, caller authority, and current policy. Evidence: operation contract.
    • B7.9.b Preview per-item eligibility/loss/dependency/gates and aggregate resources/licenses/capacity/schedule/cost/target impact. Evidence: preview tests.
    • B7.9.c Authorize the exact preview revision and execute idempotently with per-item/stage progress, attempts, artifacts, and audit. Evidence: execution tests.
    • B7.9.d Pause/cancel/retry only eligible scope and show succeeded/failed/ skipped/cancelled/partial/unknown/compensated separately. Evidence: mixed-result journey.
    • B7.9.e Read back builds/packages/targets/releases and keep unresolved partial state visible. Evidence: reconciliation report.
  • B7.10 Build mobile/tablet incident, session terminate, review, approve, and release supervision. Evidence: step-up controls.
    • B7.10.a Prioritize incident queue, host/session health, command/job failure, evidence review, approve/reject, terminate/quarantine, and release/recall tasks for compact viewports. Evidence: mobile task map.
    • B7.10.b Provide semantic summaries/tables/thumbnails and exact evidence links without requiring dense graph/3D/full-scene editing. Evidence: compact journey.
    • B7.10.c Implement touch/keyboard targets, reflow/zoom/orientation/safe areas, reduced motion, interruption, offline/reconnect, and focus recovery. Evidence: responsive accessibility suite.
    • B7.10.d Enforce exact revision/impact, step-up/dual control, reason, confirmation, immutable audit, and read-back for critical actions. Evidence: mobile security tests.
  • B7.11 Build governed Studio facade showing permitted summaries and deep links, not a duplicate console. Evidence: parity/ownership test.
    • B7.11.a Define permitted summary fields/states/actions and prohibited host command, secret, path, raw log, protected artifact, and admin data. Evidence: facade contract.
    • B7.11.b Project authorized live job/release/target summaries with source owner, freshness, limitations, and canonical Bellona identity. Evidence: projection tests.
    • B7.11.c Route detailed inspection/mutation through signed expiring deep links and reauthorize in the control room. Evidence: link journey.
    • B7.11.d Prove no duplicate mutation endpoints/state or divergent status exists in Studio and parity reconciles from authoritative events. Evidence: ownership test.
  • B7.12 Build command palette and saved views filtered by role. Evidence: unauthorized commands never appear or execute.
    • B7.12.a Register stable navigation/operation IDs, keywords, context requirements, route/target resolvers, capabilities, step-up, and availability. Evidence: command registry.
    • B7.12.b Filter palette, shortcuts, recent items, suggestions, and saved view content by current role/tenant/project/host/object authority. Evidence: visibility tests.
    • B7.12.c Create/update/share/delete saved queries/layouts with owner/ scope/version/permissions and safe defaults. Evidence: saved-view tests.
    • B7.12.d Reauthorize resolved target/revision at navigation or execution and reject stale permission/context. Evidence: execution negative suite.
  • B7.13 Build honest disconnected-host, incompatible-version, missing-DCC, unavailable-validator, stale, partial, fixture, and failed states. Evidence: state matrix.
    • B7.13.a Define each state, source signal, precedence, reason code, diagnostic detail, owner, permitted action, retryability, and release effect. Evidence: state contract.
    • B7.13.b Render disconnected/missing/incompatible host dependencies and unavailable validators with last-known time and safe remediation. Evidence: dependency matrix.
    • B7.13.c Distinguish stale/partial data and mixed per-item/per-stage/ target outcomes from empty/success. Evidence: partial-state tests.
    • B7.13.d Persistently label fixtures/demos and prohibit production mutation or evidence claims. Evidence: fixture tests.
    • B7.13.e Preserve typed failure diagnostics, retry/compensation, and unknown/late result without optimistic completion. Evidence: fault journey.
  • B7.14 Apply existing theming/i18n/mobile/accessibility work without introducing decorative dashboard clutter. Evidence: frontend conformance review.
    • B7.14.a Inventory/reuse existing design tokens, components, themes, locale/formatting infrastructure, responsive patterns, and accessibility primitives. Evidence: reuse map.
    • B7.14.b Implement restrained visual hierarchy, typography, spacing, color, motion, and operational density around task priority rather than ornamental cards/gradients/hero copy. Evidence: design review.
    • B7.14.c Externalize/localize strings and test expansion, plural/ date/number/time-zone formatting, RTL where supported, and technical identifiers. Evidence: i18n suite.
    • B7.14.d Validate wide, compact, and single-column layouts, touch/ keyboard, zoom/reflow, themes/contrast, reduced motion, and semantic alternatives. Evidence: Playwright conformance suite.
  • B7.15 Add permission-filtered control-room routes or governed deep links for render, gameplay, virtual production, XR, mocap, MetaHuman, DaVinci, Houdini, audio/video, cross-domain compilation, agent/MCP plans, package-builder, events, and developer compatibility. Evidence: every retained specialized capability is reachable by its authorized audience and no duplicate mutation UI exists.
    • B7.15.a Map each specialized capability to its authoritative existing or new control-room route, permitted facade/deep-link entry, owner, and required roles. Evidence: reachability matrix.
    • B7.15.b Add task-led routes for render/gameplay/VP/XR/mocap/MetaHuman operations with profile-specific states/evidence and no generic inheritance. Evidence: route journeys.
    • B7.15.c Add task-led routes for DaVinci/Houdini/A-V/cross-domain compilation/package operations with exact source-target lineage. Evidence: route journeys.
    • B7.15.d Add task-led agent-plan/event/developer compatibility routes with risk/approval, delivery status, and version matrices. Evidence: route journeys.
    • B7.15.e Filter navigation/routes/data/actions by authorization and feature/profile availability; reauthorize on load/action. Evidence: role tests.
    • B7.15.f Prove every authorized capability is reachable and every mutation has exactly one authoritative UI/service surface. Evidence: zero-orphan/duplication audit.

B8 — Per-host compatibility and release matrix#

  • B8.1 Validate Blender host install, auth, session, pull, push, import, export, round-trip, and revoke. Evidence: supported-version matrix.
    • B8.1.a Install/upgrade/rollback/uninstall the signed Bellona plugin on every supported Blender/OS version and verify files/preferences/ dependency cleanup. Evidence: install matrix.
    • B8.1.b Enroll/authenticate/rotate/revoke the Blender host and exercise scoped session open/heartbeat/reconnect/close. Evidence: security journey.
    • B8.1.c Pull/push/import/export representative Blender scenes/assets through typed commands with authorization, cancellation, undo, and crash recovery. Evidence: command traces.
    • B8.1.d Round-trip Blender geometry/material/rig/animation/camera/ dependencies and compare known fidelity values/declared loss. Evidence: version matrix.
  • B8.2 Validate Maya host install, auth, session, pull, push, import, export, round-trip, and revoke. Evidence: supported-version matrix.
    • B8.2.a Install/upgrade/rollback/uninstall the signed Bellona plugin on every supported Maya/OS version and verify modules/preferences/ dependency cleanup. Evidence: install matrix.
    • B8.2.b Enroll/authenticate/rotate/revoke the Maya host and exercise scoped session open/heartbeat/reconnect/close. Evidence: security journey.
    • B8.2.c Pull/push/import/export representative Maya scenes/assets through typed commands with authorization, cancellation, undo, and crash recovery. Evidence: command traces.
    • B8.2.d Round-trip Maya geometry/material/rig/animation/camera/ dependencies and compare known fidelity values/declared loss. Evidence: version matrix.
  • B8.3 Validate Houdini host install, auth, session, pull, push, import, export, round-trip, and revoke. Evidence: supported-version matrix.
    • B8.3.a Install/upgrade/rollback/uninstall the signed Bellona package on every supported Houdini/OS version and verify package paths/assets/ dependency cleanup. Evidence: install matrix.
    • B8.3.b Enroll/authenticate/rotate/revoke the Houdini host and exercise scoped session open/heartbeat/reconnect/close. Evidence: security journey.
    • B8.3.c Pull/push/import/export and cook representative Houdini scenes/ HDAs/USD through typed commands with cancel/checkpoint/license/crash recovery. Evidence: command traces.
    • B8.3.d Round-trip geometry/material/rig/animation/camera/dependencies/ HDA outputs and compare fidelity/nondeterminism. Evidence: version matrix.
  • B8.4 Validate 3ds Max host install, auth, session, pull, push, import, export, round-trip, and revoke. Evidence: supported-version matrix.
    • B8.4.a Install/upgrade/rollback/uninstall the signed Bellona plugin on every supported 3ds Max/Windows version and verify files/settings/ dependency cleanup. Evidence: install matrix.
    • B8.4.b Enroll/authenticate/rotate/revoke the 3ds Max host and exercise scoped session open/heartbeat/reconnect/close. Evidence: security journey.
    • B8.4.c Pull/push/import/export representative 3ds Max scenes/assets through typed commands with authorization, cancellation, undo, and crash recovery. Evidence: command traces.
    • B8.4.d Round-trip 3ds Max geometry/material/rig/animation/camera/ dependencies and compare known fidelity values/declared loss. Evidence: version matrix.
  • B8.5 Validate Unreal host install, auth, session, pull, push, import, export, round-trip, and revoke using the local UE5 path where applicable. Evidence: supported-version matrix.
    • B8.5.a Build/install/enable/upgrade/rollback/uninstall the signed Bellona plugin on every supported Unreal/platform version, including local UE 5.5 where applicable. Evidence: install matrix.
    • B8.5.b Enroll/authenticate/rotate/revoke editor/on-box hosts and exercise scoped session open/heartbeat/reconnect/close without editor deadlock. Evidence: security journey.
    • B8.5.c Pull/push/import/export/cook representative Unreal assets/maps through typed commands with transactions, cancel, compile, and crash recovery. Evidence: command traces.
    • B8.5.d Round-trip Unreal geometry/material/skeleton/animation/camera/ metadata/dependencies and validate editor/runtime load. Evidence: version matrix.
  • B8.6 Validate Unity host install, auth, session, pull, push, import, export, round-trip, and revoke. Evidence: supported-version matrix.
    • B8.6.a Install/upgrade/rollback/uninstall the signed Bellona package on every supported Unity/platform version and verify package/project/ dependency cleanup. Evidence: install matrix.
    • B8.6.b Enroll/authenticate/rotate/revoke the Unity host and exercise scoped session open/heartbeat/reconnect/close across domain reload. Evidence: security journey.
    • B8.6.c Pull/push/import/export/build representative Unity assets/scenes through typed commands with undo, cancel, compile, and crash recovery. Evidence: command traces.
    • B8.6.d Round-trip geometry/material/rig/animation/camera/metadata/ dependencies and validate editor/player load. Evidence: version matrix.
  • B8.7 Validate Godot host install, auth, session, pull, push, import, export, round-trip, and revoke. Evidence: supported-version matrix.
    • B8.7.a Install/upgrade/rollback/uninstall the signed Bellona add-on on every supported Godot/platform version and verify project/settings/ dependency cleanup. Evidence: install matrix.
    • B8.7.b Enroll/authenticate/rotate/revoke the Godot host and exercise scoped session open/heartbeat/reconnect/close. Evidence: security journey.
    • B8.7.c Pull/push/import/export/build representative Godot resources/ scenes through typed commands with undo, cancel, reimport, and crash recovery. Evidence: command traces.
    • B8.7.d Round-trip Godot geometry/material/skeleton/animation/camera/ metadata/dependencies and validate editor/runtime load. Evidence: version matrix.
  • B8.8 Validate host offline/reconnect without duplicate commands or partial hidden state. Evidence: network fault tests.
    • B8.8.a Inject disconnect before dispatch, after acceptance, during execution/streaming/commit, and after host completion before receipt. Evidence: fault matrix.
    • B8.8.b Assert gateway/session liveness, command uncertainty, user state, reservations, and downstream job state become explicit. Evidence: state tests.
    • B8.8.c Reauthenticate and reconcile session epoch, command IDs/results, host object revisions, artifacts, and checkpoints on reconnect. Evidence: reconciliation tests.
    • B8.8.d Retry/resume/compensate only safe work and prove no duplicate command side effect or hidden partial state. Evidence: network fault report.
  • B8.9 Validate DCC crash and restart recovery. Evidence: session/job state reconciles.
    • B8.9.a Crash each supported host during idle, dirty scene, import/ export, mutation, save, render/build, and result commit. Evidence: crash matrix.
    • B8.9.b Assert session/command/job/lease state, checkpoints, temp files, artifact promotion, and user-visible diagnostics are safe. Evidence: invariant tests.
    • B8.9.c Restart host/plugin, reauthenticate, inspect project/object/file state, and correlate recovered command journals/checkpoints. Evidence: restart journey.
    • B8.9.d Resume/retry/compensate or require human resolution according to command contract with no false success. Evidence: reconciliation report.
  • B8.10 Validate plugin upgrade/downgrade and server version skew. Evidence: compatibility behavior.
    • B8.10.a Test every supported plugin/server/protocol/host version combination and boundary outside the compatibility window. Evidence: skew matrix.
    • B8.10.b Upgrade with active/idle sessions and verify drain, state preservation, negotiation, health, and rollback. Evidence: upgrade journey.
    • B8.10.c Downgrade only where approved and verify schema/state/config/ artifact compatibility or explicit rejection. Evidence: downgrade tests.
    • B8.10.d Surface degraded/expiring/incompatible status before command dispatch and never silently select an unsafe protocol. Evidence: compatibility behavior.
  • B8.11 Validate unit, axis, color, material, rig, animation, camera, metadata, and unsupported-feature behavior for every supported pair. Evidence: pairwise coverage matrix.
    • B8.11.a Generate the exact supported source-host/version to target-host/ version/profile pair matrix with declared unsupported cells. Evidence: coverage registry.
    • B8.11.b Run analytical unit/axis/transform/hierarchy/camera and known color/OCIO fixtures per pair. Evidence: foundational results.
    • B8.11.c Run material/texture/shader and skeleton/skinning/blend-shape/ animation fixtures per pair. Evidence: asset results.
    • B8.11.d Run metadata/naming/dependency/variant/scene and target import/ runtime fixtures per pair. Evidence: scene results.
    • B8.11.e Assert every unsupported/lossy/substituted feature produces exact diagnostics, preview/acceptance, and never silent pass. Evidence: behavior matrix.
  • B8.12 Validate large asset, large scene, many dependencies, and slow link. Evidence: performance and backpressure.
    • B8.12.a Define representative maximum-supported bytes, geometry/ texture/animation, object/layer count, dependency fan-out/depth, and network latency/bandwidth/loss fixtures. Evidence: workload plan.
    • B8.12.b Measure ingest/scan/resolve/convert/validate/preview/transfer/ import/build/read-back latency, throughput, memory, storage, and cost. Evidence: performance report.
    • B8.12.c Exercise resumable chunking, incremental conversion, caching, pagination/lazy preview, scheduling, quotas, and end-to-end backpressure. Evidence: behavior tests.
    • B8.12.d Inject interruption/cancel/resource exhaustion and prove bounded memory/storage, cleanup, recovery, fairness, and no corrupt partial release. Evidence: overload report.
  • B8.13 Validate path/resolver behavior across supported operating systems. Evidence: portable package fixtures.
    • B8.13.a Cover Windows drives/UNC/device/reserved names/case, macOS case/ Unicode, and Linux roots/symlinks/case with supported host combinations. Evidence: OS path corpus.
    • B8.13.b Test separators/dot segments/encoding/normalization/length/ collisions/symlinks/junctions/mounts and virtual-root containment. Evidence: security results.
    • B8.13.c Test resolver contexts, environment independence, relative/ virtual mappings, external refs, packages, and dependency checksums on every OS. Evidence: resolver matrix.
    • B8.13.d Move packages among clean supported OS/host environments and prove resolution/read-back without source-machine paths. Evidence: portability fixtures.
  • B8.14 Publish known limitations and loss profiles per host/target pair. Evidence: UI and generated documentation derive from the same registry.
    • B8.14.a Record each pair's unsupported/partial/substituted/baked/ nondeterministic features, bounds, symptoms, mitigation, and affected versions in the compatibility registry. Evidence: limitation rows.
    • B8.14.b Link each limitation to source tests, raw measurements, adapter/ host issue, owner, status, and expiry/review date. Evidence: provenance audit.
    • B8.14.c Generate user documentation, conversion preview diagnostics, API/SDK discovery, and control-room explanations from the same versioned data. Evidence: generation parity.
    • B8.14.d Block unacknowledged material loss and test that docs/UI update with registry changes. Evidence: stale-content gate.
  • B8.15 Validate the complete typed command registry for each supported host/version/plugin rather than only connect/pull/push. Evidence: command-capability matrix includes success, unsupported, authorization, invalid revision, cancellation, undo/compensation, crash, and audit.
    • B8.15.a Enumerate every command ID/version/category and every supported host/version/plugin tuple; declare required, optional, and unsupported cells. Evidence: command-capability matrix.
    • B8.15.b Generate schema-valid boundary and invalid argument cases for each supported command, including path/object/property/resource limits. Evidence: command fixtures.
    • B8.15.c Test success/idempotent replay/result/read-back and explicit unsupported/incompatible behavior per cell. Evidence: functional conformance.
    • B8.15.d Test unauthorized property/scope, stale revision/precondition, deadline/cancel, duplicate, session expiry, and host quarantine per applicable cell. Evidence: control conformance.
    • B8.15.e Test undo/compensation/partial/unknown outcome and host/plugin crash/reconnect reconciliation per mutating cell. Evidence: recovery conformance.
    • B8.15.f Verify redacted immutable audit and coverage gate; prohibit connect/pull/push from standing in for untested commands. Evidence: release audit.
  • B8.16 Validate renderers/farms, virtual-production devices, XR devices/runtimes, mocap vendors, MetaHuman/Live Link, Resolve, Houdini, engine runtime/gameplay, and package/install targets independently. Evidence: specialized pairwise matrix with declared unavailable cells.
    • B8.16.a Register exact producer/control-room/host/plugin/device/runtime/ target version tuples and required specialized capabilities. Evidence: specialized matrix.
    • B8.16.b Validate each renderer/farm and VP device/stage combination for auth, commands, resources, safety, output, failure, and read-back. Evidence: independent rows.
    • B8.16.c Validate each XR platform/device/runtime and mocap vendor/device/ protocol combination for live/deploy/quality/failure behavior. Evidence: independent rows.
    • B8.16.d Validate each MetaHuman/Live Link, Resolve, and Houdini version combination for identity/media/procedural workflows and round-trip. Evidence: independent rows.
    • B8.16.e Validate each engine runtime/gameplay and package/install target for scenario/build/install/upgrade/rollback/read-back. Evidence: independent rows.
    • B8.16.f Declare unavailable/unsupported cells with reason/owner/plan and prevent any specialized profile from inheriting generic bridge or another target's evidence. Evidence: independence gate.

B9 — Bellona observability, verification, and launch#

  • B9.1 Instrument host, session, command, transfer, interchange, cache, validation, build, release, and recall traces/metrics. Evidence: bounded schema.
    • B9.1.a Define bounded span/event/metric names, attributes, units, cardinality, sampling, retention, and owners for every subsystem. Evidence: telemetry registry.
    • B9.1.b Propagate tenant-safe trace/session/command/job/stage/artifact/ build/release correlation across gateway, hosts, workers, adapters, and targets. Evidence: traced fixture.
    • B9.1.c Instrument success, latency, traffic/load, retries/backpressure, partial/unknown outcomes, typed failures, gates, cost, and recall closure. Evidence: signal tests.
    • B9.1.d Redact credentials, tokens, local paths, asset/project names, command payloads, content, free text, and unbounded identifiers. Evidence: telemetry privacy scan.
  • B9.2 Define SLOs for connection, command, transfer, queue, conversion, validation, build, target availability, and release verification. Evidence: error-budget policy.
    • B9.2.a Define user-visible indicators, populations, windows, sources, exclusions, and measurement quality for each named operation. Evidence: SLI catalog.
    • B9.2.b Ratify latency, availability, freshness, correctness, fidelity, durability, and capacity objectives by support tier/profile. Evidence: target matrix.
    • B9.2.c Build dashboards and multi-window burn alerts with actionable host/adapter/target drill-down and bounded labels. Evidence: alert tests.
    • B9.2.d Define error-budget release freeze, exception, remediation, recovery, and retrospective policy. Evidence: policy exercise.
  • B9.3 Run protocol, contract, migration, and version-skew suites. Evidence: host/client/server matrix.
    • B9.3.a Test current/supported prior gateway-host protocol negotiation, messages, state machines, capabilities, errors, and unknown fields. Evidence: protocol matrix.
    • B9.3.b Test API/event/SDK/host-plugin contracts across server/client/ consumer version combinations. Evidence: contract matrix.
    • B9.3.c Test schema migration upgrade/backfill/mixed-version/rollback/ roll-forward/cleanup at representative scale. Evidence: migration report.
    • B9.3.d Test host/plugin/adapter/target version boundaries, deprecation/expiry, downgrade, and fail-closed unsupported behavior. Evidence: skew suite.
  • B9.4 Run known-value USD, geometry, transform, material, rig, animation, camera, dependency, and fidelity tests. Evidence: gate registry coverage.
    • B9.4.a Register independently derived fixtures with source, generator, expected raw values/tolerances, profile, rights, and checksums. Evidence: fixture registry.
    • B9.4.b Cover USD composition/schema/resolver/dependency and geometry/ topology/transform/unit/axis measures. Evidence: foundational report.
    • B9.4.c Cover material/texture/color/render and skeleton/skinning/blend- shape/animation measures. Evidence: asset report.
    • B9.4.d Cover camera/light/scene/metadata/dependency, target import/ runtime, and round-trip fidelity separately per profile. Evidence: target report.
    • B9.4.e Prove every production measure/gate/profile/version maps to current known-value proof and non-pass states cannot satisfy coverage. Evidence: gate audit.
  • B9.5 Run usdchecker and pinned custom-validator goldens. Evidence: current reports.
    • B9.5.a Pin OpenUSD/usdchecker and custom validator code/config/ dependencies/profiles with checksums. Evidence: validator manifest.
    • B9.5.b Run valid/boundary/invalid/malicious USD stages/packages in hermetic resolver contexts and compare raw reports to goldens. Evidence: golden suite.
    • B9.5.c Verify typed parser mapping, severity/location/rule identity, truncation, unknown output, and validator failure/unavailability. Evidence: parser tests.
    • B9.5.d Publish exact current report artifacts/versions and block stale validator evidence at release. Evidence: freshness gate.
  • B9.6 Run real-database, artifact, gateway, host, queue, cache, build, target, and search integration tests. Evidence: durable round-trip.
    • B9.6.a Stand up production-like isolated database/artifact/evidence/ gateway/queue/cache/search/build/target services plus supported host adapters. Evidence: environment manifest.
    • B9.6.b Exercise enrollment/session/command, intake/interchange/cache/ validation, build/release/deploy/read-back, and search/audit lifecycle. Evidence: durable round-trip.
    • B9.6.c Exercise duplicate/reordered messages, concurrency, restart, stale cache/index, partial target, and cleanup/reconciliation. Evidence: fault cases.
    • B9.6.d Verify checksums, immutable lineage, tenancy, authorization, audit/outbox, and target truth at every boundary. Evidence: invariant report.
  • B9.7 Run full-journey Playwright: host connect to verified release. Evidence: real or hermetic supported host adapter.
    • B9.7.a Provision a supported real or hermetic host/plugin/version with a known rights-cleared asset/scene, credentials, target, and expected measures. Evidence: fixture manifest.
    • B9.7.b Enroll/approve/connect the host, inspect health/capabilities, open scoped session, and execute typed pull/ingest commands. Evidence: Playwright trace.
    • B9.7.c Preview conversion/loss, run interchange, inspect source/USD/ target diff and raw/current gates, and approve exact evidence. Evidence: review assertions.
    • B9.7.d Build candidate/evidence, deploy, run target smoke/fidelity, independently read back, and assert verified release. Evidence: journey report.
  • B9.8 Run full-journey Playwright: host failure, reconnect, rerun, compare, waive/reject, and release. Evidence: exact state.
    • B9.8.a Inject disconnect/crash during a deterministic command or stage and assert session/command/job/attempt/partial/unknown state. Evidence: failure trace.
    • B9.8.b Reconnect/restart, reconcile host/project/object/artifact state, and rerun/resume from the safe boundary. Evidence: recovery assertions.
    • B9.8.c Compare exact failed/prior/new revisions and inspect raw fidelity loss with semantic parity. Evidence: comparison trace.
    • B9.8.d Exercise eligible scoped waiver and non-waivable rejection, expiry/staleness, and named approval. Evidence: decision assertions.
    • B9.8.e Release only the eligible new revision and prove failed/stale/ unwaived outputs cannot publish. Evidence: exact-state report.
  • B9.9 Run full-journey Playwright: artifact quarantine/recall and downstream reconciliation. Evidence: target verification.
    • B9.9.a Release a fixture artifact into library, dependent project/build/ package, and supported target with traceable consumer refs. Evidence: setup trace.
    • B9.9.b Initiate quarantine or recall from an exact finding/revision and assert blast-radius preview, authority, step-up, and audit. Evidence: action trace.
    • B9.9.c Assert new use/build/release is blocked and active/derived/ deployed state transitions per policy with explicit partial results. Evidence: propagation assertions.
    • B9.9.d Retry unreachable recipients and independently read back every target/consumer until closure or visible unresolved state. Evidence: reconciliation report.
  • B9.10 Run WCAG 2.2 AA, keyboard, announcement, non-drag, and semantic preview parity suites. Evidence: themes/viewports/locales.
    • B9.10.a Run automated WCAG checks across authenticated roles/routes, themes, locales, and wide/compact/single-column viewports. Evidence: accessibility report.
    • B9.10.b Complete keyboard-only host/session/command, DAG, preview/diff, gate/waiver, batch/schedule, release/recall journeys. Evidence: keyboard traces.
    • B9.10.c Verify names/roles/values, focus order/restoration, live announcements, errors, step-up dialogs, and stale/partial/failure state. Evidence: assistive-technology matrix.
    • B9.10.d Verify non-drag alternatives, 200–400% zoom/reflow, touch targets, color independence, reduced motion, and semantic graph/3D/image parity. Evidence: conformance suite.
  • B9.11 Run fleet, concurrent session, command, transfer, large scene, conversion, batch, build, and search load tests. Evidence: backpressure and SLO.
    • B9.11.a Define representative/maximum host fleet, tenants, session/ command mixes, artifact/scene sizes, dependencies, network, batch/build, search corpus, resources, and budgets. Evidence: load plan.
    • B9.11.b Measure enrollment/heartbeat/session/gateway/command streaming and artifact transfer under sustained/burst/slow/abusive clients. Evidence: control-plane report.
    • B9.11.c Measure scan/resolve/convert/cache/validate/preview/search, batch/scheduling/build/package/deploy/read-back under load. Evidence: data-plane report.
    • B9.11.d Verify quotas/fairness/backpressure/circuit breakers/cancel/ recovery, bounded memory/storage/logs, correctness, cost, and SLO. Evidence: overload report.
  • B9.12 Run command injection, host spoof, plugin tamper, malicious asset/USD/archive, path, SSRF, secret leak, artifact tamper, auth, tenant, and release bypass tests. Evidence: zero blockers.
    • B9.12.a Test command/argument/schema confusion, shell/script/eval/ console escape, expected-revision bypass, and unsafe compensation. Evidence: command security suite.
    • B9.12.b Test enrollment/host/session spoof, credential replay/theft, attestation/signature bypass, plugin downgrade/tamper, and compromised host results. Evidence: identity/supply-chain suite.
    • B9.12.c Test malicious files/archives/USD/resolvers/dependencies, traversal/symlink/Unicode paths, SSRF/rebinding/redirect, and parser/ decompression abuse. Evidence: content security suite.
    • B9.12.d Test secret/path/content leakage through APIs/UI/search/logs/ traces/events/support bundles and two-tenant isolation. Evidence: disclosure suite.
    • B9.12.e Test artifact/evidence/signature/target read-back tamper and gate/waiver/approval/release/recall bypass. Evidence: release security suite with zero blockers.
  • B9.13 Run gateway/host/worker loss, artifact corruption, cache loss, backup/restore, release rollback, and recall drills. Evidence: RPO/RTO.
    • B9.13.a Inject gateway/host/worker process/network/resource loss during sessions, commands, transfer, conversion, build, and commit. Evidence: compute drill.
    • B9.13.b Corrupt/lose staged, authoritative, cached, packaged, and deployed artifact bytes and verify checksum detection/quarantine/ recomputation/recovery. Evidence: artifact drill.
    • B9.13.c Restore databases/artifacts/evidence/config/keys/indexes/queues from backup and reconcile sessions/jobs/builds/search/audit/targets. Evidence: restore report.
    • B9.13.d Roll back failed release and execute quarantine/recall during active downstream use with per-target read-back. Evidence: release drill.
    • B9.13.e Measure achieved RPO/RTO, data loss, duplicate work, target inconsistency, and unresolved consumers against objectives. Evidence: scorecard.
  • B9.14 Publish host install, bridge, interchange, fidelity, build, release, security incident, and recovery runbooks. Evidence: game-day.
    • B9.14.a Assign owner/audience/prerequisites/triggers/authority/ diagnostics/actions/safety/escalation/rollback to each runbook. Evidence: template audit.
    • B9.14.b Write per-host install/update/revoke and bridge/session/command/ gateway operational runbooks. Evidence: reviewed docs.
    • B9.14.c Write intake/interchange/dependency/fidelity/validator/waiver and build/render/package/release/target runbooks. Evidence: reviewed docs.
    • B9.14.d Write security/privacy/supply-chain incident, quarantine/recall, backup/restore/disaster/dependency-outage runbooks. Evidence: reviewed docs.
    • B9.14.e Execute representative game-days, remediate deviations, and revalidate every command/link/owner. Evidence: exercise report.
  • B9.15 Produce engineering-readiness evidence per supported host/target pair. Evidence: incomplete pairs cannot inherit another pair's proof.
    • B9.15.a Build one evidence bundle per exact host/plugin/OS/server/ protocol/source-target/profile tuple with owners and support window. Evidence: pair registry.
    • B9.15.b Map requirements to design/code/contracts/migrations/known- value/fidelity/security/accessibility/performance/recovery/runbook proof with exact versions. Evidence: traceability index.
    • B9.15.c Include command-by-command, conversion/round-trip, build/deploy, target smoke/read-back, limitations, waivers, and approvals per pair. Evidence: readiness bundle.
    • B9.15.d Surface failed/stale/missing/unavailable evidence explicitly and block the pair without borrowing another pair's proof. Evidence: independence gate.
    • B9.15.e Reproduce and verify bundles/checksums/links from authoritative evidence. Evidence: clean-room generation.
  • B9.16 Obtain named DCC/engine, pipeline, security, accessibility, engineering, and product approvals. Evidence: human signoff.
    • B9.16.a Define approver scope/competence/independence, required evidence, supported host/target coverage, decision options, conditions, and expiry. Evidence: approval policy.
    • B9.16.b Present the exact release candidate/readiness bundles, compatibility matrix, limitations, risks, waivers, and unresolved items. Evidence: review packet.
    • B9.16.c Capture named signed approve/reject/conditional decisions with role, time, exact revisions/pairs, rationale, and conditions. Evidence: signoff records.
    • B9.16.d Invalidate/renew approvals after material code/config/profile/ host/target/evidence/risk changes. Evidence: staleness tests.
  • B9.17 Run distinct full journeys for render queue, gameplay systems, virtual-production stage, XR deployment, mocap stream/record/retarget, MetaHuman, DaVinci conform, Houdini cook, audio/video conversion, cross-domain/lore compilation, agent/MCP dry-run approval, package install/rollback, events, and SDK clients. Evidence: no generic bridge or build journey substitutes for specialized proof.
    • B9.17.a Run independent render-queue and gameplay-system journeys with specialized resources, failures, quality/scenarios, target read-back, and rollback. Evidence: production traces.
    • B9.17.b Run independent VP-stage, each XR-platform deployment, and each mocap vendor live/record/retarget journey with device safety, consent, faults, and target read-back. Evidence: device traces.
    • B9.17.c Run independent MetaHuman/Live Link, DaVinci conform/delivery, Houdini cook/farm, and A/V conversion/package journeys. Evidence: creative traces.
    • B9.17.d Run cross-domain/lore compilation through producer contracts, mappings, build/deploy/read-back, and correction/recall propagation. Evidence: integration trace.
    • B9.17.e Run agent/MCP intent-to-plan/dry-run/approval/execution/ intervention/read-back and adversarial boundaries. Evidence: agent trace.
    • B9.17.f Run package fresh install/upgrade/uninstall/rollback, event publish/deliver/replay/dead-letter, and every external SDK/CLI/plugin client. Evidence: ecosystem traces.
    • B9.17.g Prohibit generic bridge/build success from satisfying any missing specialized journey. Evidence: coverage gate.
  • B9.18 Run agent/remote-control adversarial suites for ambiguous intent, prompt/tool injection, stale selection/revision, hidden destructive side effect, approval confusion, privilege expansion, organization/spend bypass, unsafe path/network target, late result, host compromise, and audit tamper. Evidence: fail-closed thresholds and incident drills.
    • B9.18.a Test ambiguous/conflicting/incomplete intent, wrong context, hallucinated object, prompt/content/tool-output injection, and unsupported goal; require safe clarification/refusal. Evidence: intent suite.
    • B9.18.b Test stale selection/revision, TOCTOU between dry-run/approval/ execution, hidden/described-mismatched side effects, irreversible actions, and approval scope/confusion. Evidence: plan integrity suite.
    • B9.18.c Test role/session/command/path/property privilege expansion, cross-tenant/org policy evasion, step-up/dual-control bypass, and chained-tool authority laundering. Evidence: authorization suite.
    • B9.18.d Test token/time/cost/spend/capacity budget evasion, parallel fan-out, retry loops, late results, cancellation races, and partial completion disguised as success. Evidence: resource/control suite.
    • B9.18.e Test traversal/SSRF/rebinding/redirect/credential exfiltration, malicious host/plugin compromise, falsified results/read-back, and quarantine escape. Evidence: execution security suite.
    • B9.18.f Test audit suppression/redaction abuse/reordering/deletion/ tamper and ensure immutable intent/plan/approval/tool/result/target reconstruction. Evidence: audit integrity suite.
    • B9.18.g Ratify fail-closed thresholds, run pause/kill/quarantine/recall/ recovery incident drills, and close every release blocker. Evidence: adversarial release report.
  • B9.19 EVE-SEAM-EXIT: require "required Eve reads and writes admitted with source and behavior evidence, or an exact task-0.8 explicit non-goal proves zero operations; a named defer preserves ownership but never passes completion" before this phase exits. Evidence: source-derived totality record and passing verifier.
    • B9.19.a Re-run the route, bridge, state-signal, and complete domain-code envelope census; review every changed candidate instead of carrying a stale phase disposition forward. Evidence: current ratchets.
    • B9.19.b Bind the exact ratified task-0.8 workflow and requirement set; classify it as a required seam or an exact source-backed non-goal. Evidence: charter-bound disposition set.
    • B9.19.c For a required seam, admit both read and write operations with byte-pinned source and behavior evidence. A named defer retains its owner and unblock condition but fails completion. Evidence: operation proofs.
    • B9.19.d Permit a zero-operation boundary only for an exact ratified task-0.8 explicit non-goal; absence or unavailability is not a non-goal. Evidence: source-bound non-goal proof.
    • B9.19.e Run the totality verifier, every retained gate-case CLI, and the injected durable-candidate control; retain red and green receipts. Evidence: release gate logs.

5. Cross-domain integration completion#

  • X.1 Prove Isis-generated assets enter Yemaya projects with immutable generation and rights provenance. Evidence: cross-domain contract, integration test, and traced journey.
    • X.1.a Define the Isis export/Yemaya intake contract for immutable asset/ generation revisions, ingredients, model/tool/config/seed, rights/ license, disclosures, checksums, and lifecycle. Evidence: contract.
    • X.1.b Authorize and transfer the minimal tenant/project-scoped manifest idempotently without granting Yemaya Isis generation authority. Evidence: seam tests.
    • X.1.c Create a Yemaya project reference/import revision that retains canonical Isis identity and provenance. Evidence: integration test.
    • X.1.d Propagate Isis supersession/revoke/quarantine/recall and verify Yemaya project/derivative state by read-back. Evidence: traced journey.
  • X.2 Prove Yemaya-promoted assets enter Oshun output gallery through the editorial boundary. Evidence: cross-domain contract, integration test, and traced journey.
    • X.2.a Define the Yemaya promotion/gallery-ingest contract with exact asset/project/release revisions, editorial decision, rights, quality, provenance, renditions, locale, and lifecycle. Evidence: contract.
    • X.2.b Require authorized editorial promotion and current gallery eligibility without granting the gallery Yemaya mutation authority. Evidence: boundary tests.
    • X.2.c Materialize permission-aware gallery projection/renditions and verify displayed/downloaded bytes/metadata against the manifest. Evidence: integration journey.
    • X.2.d Propagate supersession/unpublish/revoke/recall and reconcile search/cache/gallery/consumer state. Evidence: read-back test.
  • X.3 Prove Aja motion assets can enter Yemaya/Bellona without bypassing consent, watermark, or validation. Evidence: cross-domain contract, integration test, and traced journey.
    • X.3.a Define Aja motion export contract with exact artifact/skeleton/ mapping/profile/quality/consent/rights/watermark/provenance revisions and intended consumers. Evidence: contract.
    • X.3.b Enforce current consent, quality, compatibility, disclosure, and watermark gates at Aja handoff and Yemaya/Bellona intake. Evidence: negative tests.
    • X.3.c Import/reference through public Yemaya and Bellona contracts while preserving Aja authority and immutable lineage. Evidence: integrations.
    • X.3.d Propagate subject/consent/quality/artifact revoke or recall to projects/builds/packages/targets and verify closure. Evidence: traced journey.
  • X.4 Prove Bellona can package Yemaya/Aja/Isis artifacts without mutating their authoritative lineage. Evidence: cross-domain contract, integration test, and traced journey.
    • X.4.a Define a common immutable producer-artifact reference contract with domain authority, exact revision/checksum, dependencies, rights/ consent/canon/quality, and recall endpoint. Evidence: contract.
    • X.4.b Resolve and lock producer refs through public APIs without copying or rewriting authoritative provenance records. Evidence: resolver tests.
    • X.4.c Build Bellona package manifests/representations with conversion lineage pointing back to each producer identity. Evidence: integration test.
    • X.4.d Read back package/target and prove source authority/provenance unchanged; propagate producer recall. Evidence: traced journey.
  • X.5 Prove Metis media uses curated Isis and Yemaya surfaces rather than raw generation machinery. Evidence: cross-domain contract, integration test, and traced journey.
    • X.5.a Define the curated-media discovery/selection contract with approved asset/release/rendition, audience/locale, rights/consent, accessibility, correctness, and provenance. Evidence: contract.
    • X.5.b Restrict Metis identity/capabilities to curated search/preview/use and prohibit raw generation/model/job/configuration endpoints. Evidence: least-privilege tests.
    • X.5.c Select and attach exact eligible Isis/Yemaya media revisions to a Metis lesson through public seams. Evidence: integration journey.
    • X.5.d Propagate media correction/unpublish/revoke/recall and verify learner rendering/catalog/cache reconciliation. Evidence: read-back.
  • X.6 Prove Metis embodied instruction calls Aja only through the canonical adapter and handles degraded health. Evidence: cross-domain contract, integration test, and traced journey.
    • X.6.a Define/version the canonical health/capability, demonstration, overlay, asynchronous handoff/status/result/cancel contracts. Evidence: adapter spec.
    • X.6.b Prevent Metis imports/credentials/calls to Aja internal jobs, models, media, stores, or privileged APIs. Evidence: architecture and auth tests.
    • X.6.c Run demonstration/overlay/handoff journeys with idempotency, correlation, consent/quality/provenance, and intended-use boundaries. Evidence: integration trace.
    • X.6.d Handle healthy/degraded/unavailable/incompatible/stale/late result states without fabricating guidance or success. Evidence: failure journey.
  • X.7 Prove Veritas and Metis grounding share generic mechanics while retaining distinct source/claim policies. Evidence: cross-domain contract, integration test, and traced journey.
    • X.7.a Define shared source ingestion, chunking, retrieval, citation, claim-evidence, revision, correction, audit, and evaluator primitives. Evidence: generic contracts.
    • X.7.b Define separate Veritas publication/source-risk and Metis pedagogy/curriculum/learner claim policies, roles, thresholds, and outputs. Evidence: policy contracts.
    • X.7.c Run identical mechanical fixtures through both domains and prove shared behavior where intended. Evidence: integration comparison.
    • X.7.d Run policy-divergent fixtures and prove neither domain inherits the other's admissibility, verdict, review, or release decision. Evidence: boundary journey.
  • X.8 Prove Euterpe-generated or edited audio carries rights/provenance when consumed by Isis, Yemaya, Veritas, or Metis. Evidence: cross-domain contract, integration test, and traced journey.
    • X.8.a Define Euterpe audio handoff with source/voice/performer/model/ edit revisions, rights/license/consent, generated/edited ranges, disclosure, quality, renditions, and checksum. Evidence: contract.
    • X.8.b Enforce consumer-specific intended use, locale, accessibility, quality, rights, and consent at each intake. Evidence: eligibility tests.
    • X.8.c Run independent Isis, Yemaya, Veritas, and Metis consumption journeys retaining canonical Euterpe lineage. Evidence: integration matrix.
    • X.8.d Propagate source/voice/performer rights or consent revoke and verify every derivative/publication/lesson/project. Evidence: recall trace.
  • X.9 Prove identity, tenant, workspace, locale, and trace context survive every registered cross-domain handoff. Evidence: cross-domain contract, integration test, and traced journey.
    • X.9.a Register every handoff with allowed identity/service actor, tenant/project/workspace mapping, locale/time zone, trace/baggage, and data-classification rules. Evidence: context registry.
    • X.9.b Propagate authenticated actor/on-behalf-of and least-privilege service identity without forwarding reusable credentials. Evidence: identity tests.
    • X.9.c Map tenant/workspace/object scopes explicitly and reject missing, ambiguous, stale, or cross-tenant context. Evidence: isolation tests.
    • X.9.d Preserve locale/time zone and tenant-safe trace/causation through async queues/webhooks/retries while bounding/redacting baggage. Evidence: traced matrix.
  • X.10 Prove cross-domain retries are idempotent and do not duplicate jobs, assets, reviews, releases, grades, or notifications. Evidence: cross-domain contract, integration test, and traced journey.
    • X.10.a Define operation-specific idempotency scope/key/request hash, retention window, response replay, conflict, and retry-safe boundary. Evidence: contract.
    • X.10.b Persist producer outbox and consumer inbox/dedupe state transactionally with domain mutation. Evidence: atomicity tests.
    • X.10.c Inject duplicate, delayed, reordered, lost-ack, timeout, crash, and concurrent retry at every handoff. Evidence: fault matrix.
    • X.10.d Verify authoritative counts/identities and downstream read-back show one job/asset/review/release/grade/notification side effect. Evidence: traced journey.
  • X.11 Prove source, consent, rights, model, canon, or artifact revocation reaches every registered downstream consumer. Evidence: cross-domain contract, integration test, and traced journey.
    • X.11.a Maintain a complete versioned consumer/dependency registry from each revocable authority to derivatives, caches, indexes, packages, releases, channels, targets, and external consumers. Evidence: registry.
    • X.11.b Emit signed minimal revocation events with authority, exact subject/revision/scope, effective time, reason class, replacement, and correlation. Evidence: event contract.
    • X.11.c Compute impact and execute fail-closed quarantine/unpublish/ recall/recompute/review actions per consumer with checkpoint/retry. Evidence: propagation tests.
    • X.11.d Verify every consumer/target by read-back and retain unreachable/ failed recipients as unresolved incidents. Evidence: closure report.
  • X.12 Prove cross-domain search and notifications do not leak protected titles, counts, source identities, learner data, biometric data, or unreleased media. Evidence: cross-domain contract, integration test, and traced journey.
    • X.12.a Define allowed search/notification fields and prohibited titles/ counts/identity/learner/biometric/unreleased/content metadata by domain/role/purpose. Evidence: data policy.
    • X.12.b Build authorization-filtered minimal projections before indexing or notification fan-out. Evidence: projection tests.
    • X.12.c Test query/suggestion/facet/count/snippet/timing/error and notification recipient/subject/body/link/preview inference attacks. Evidence: leakage suite.
    • X.12.d Propagate permission/release/consent/revocation changes to indexes, caches, digests, queued notifications, and external delivery. Evidence: reconciliation journey.
  • X.13 Prove deep links enforce current authorization, tenant, expiry, and object existence instead of trusting embedded context. Evidence: cross-domain contract, integration test, and traced journey.
    • X.13.a Register allowed issuer/audience/route/purpose/context fields, maximum lifetime, return behavior, and prohibited secrets/authority. Evidence: link policy.
    • X.13.b Mint signed links with tenant/project/object/revision, audience, purpose, expiry, nonce, and key version. Evidence: signature tests.
    • X.13.c Validate signature/audience/tenant/expiry/replay plus current actor authorization, object existence/revision, and lifecycle at load. Evidence: negative matrix.
    • X.13.d Render safe forbidden/expired/missing/stale outcomes without existence leakage and require fresh authorization for any mutation. Evidence: browser journey.
  • X.14 Prove audit and trace records reconstruct each cross-domain journey without storing protected content. Evidence: cross-domain contract, integration test, and traced journey.
    • X.14.a Define common correlation/causation/resource revision/event time/ actor/service/domain/action/outcome fields and domain-owned audit event mappings. Evidence: audit contract.
    • X.14.b Apply structured allowlists/redaction/hashing to content, source/ subject identity, biometrics, learner data, free text, secrets, URLs, and paths before export. Evidence: redaction tests.
    • X.14.c Preserve immutable ordered evidence refs through synchronous and asynchronous calls, queues, webhooks, retries, compensation, and recall. Evidence: trace continuity test.
    • X.14.d Reconstruct representative successful/failed/partial/retried/ revoked journeys using refs and authorized drill-down, not copied protected content. Evidence: investigation exercise.
  • X.15 Prove version negotiation and compatibility windows across all producer/consumer pairs. Evidence: compatibility matrix.
    • X.15.a Register every producer/consumer contract/profile/event/SDK pair with supported versions, capabilities, constraints, owner, environment, and support dates. Evidence: matrix.
    • X.15.b Negotiate the highest mutually approved version/capability deterministically and bind it to the handoff. Evidence: selection tests.
    • X.15.c Test current/prior/next unknown fields, optional capability, deprecation, expiry, downgrade, and unsupported combinations per pair. Evidence: skew suite.
    • X.15.d Surface expiring/incompatible consumers and block silent coercion or use outside the compatibility window. Evidence: operational gate.
  • X.16 Run end-to-end failure drills across provider, queue, storage, search, host, target, and publisher dependencies. Evidence: no silent success or irrecoverable partial state.
    • X.16.a Define failure injection points and expected state/ownership/ retry/compensation/recovery for each dependency and handoff. Evidence: drill plan.
    • X.16.b Inject provider and queue timeout/unavailable/duplicate/reorder/ poison/dead-letter cases during representative journeys. Evidence: messaging report.
    • X.16.c Inject database/blob/cache/search/index corruption/loss/staleness and reconcile authoritative state. Evidence: storage/search report.
    • X.16.d Inject host/worker/target/publisher outage, partial commit, unknown outcome, and late result. Evidence: external dependency report.
    • X.16.e Verify UI/API/events/audit expose exact failed/partial/degraded state and recovery/read-back reaches closure within objectives. Evidence: end-to-end scorecard.
  • X.17 Prove Metis lecture, diagram, simulation, voice, and other teaching media use Isis/Yemaya/Euterpe provider and production seams with source/correctness/accessibility/consent lineage intact. Evidence: contract, traced journey, revocation, and learner-render read-back.
    • X.17.a Define separate lecture/diagram/simulation/image/video/audio/ voice handoff profiles with source/claim/correctness/accessibility/ consent/rights/provenance requirements. Evidence: contracts.
    • X.17.b Route generation/production only through authorized Isis/Yemaya/ Euterpe public seams and retain Metis pedagogical/editorial authority. Evidence: boundary tests.
    • X.17.c Run each media type from source-linked request through candidate/ review/validation/lesson attachment and accessible learner render. Evidence: traced matrix.
    • X.17.d Propagate source correction/withdrawal and media/voice/consent revocation into lessons/catalogs/caches and verify learner read-back. Evidence: recall journey.
  • X.18 Prove Veritas podcast, video, live, social, and multilingual assets use governed Euterpe/Yemaya/Isis production seams without granting those systems protected-source or publication authority. Evidence: least-privilege contract and correction/retraction journey.
    • X.18.a Define separate podcast/video/live/social/multilingual production requests containing only approved script/claim/source-safe refs, publication intent, locale, rights, and accessibility requirements. Evidence: contracts.
    • X.18.b Enforce least-privilege provider identities/views that cannot access protected sources or decide publication/correction. Evidence: authorization tests.
    • X.18.c Run each asset type through Euterpe/Yemaya/Isis production, Veritas review/gates, channel packaging/publication, and target read- back. Evidence: traced matrix.
    • X.18.d Propagate Veritas correction/retraction and provider asset/voice/ consent recall without transferring editorial authority. Evidence: lifecycle journey.
  • X.19 Prove Yemaya's engine/DCC, virtual-production, XR, mocap, render, and build actions use Bellona public contracts and Aja handoffs without duplicating bridge/runtime authority. Evidence: host-specific traced journeys and version/failure behavior.
    • X.19.a Define Yemaya-to-Bellona project/asset/action/target requests and Bellona status/result/evidence contracts with exact revisions and ownership. Evidence: seam spec.
    • X.19.b Define Aja motion/capture handoff contracts and prevent Yemaya or Bellona from duplicating Aja consent/quality authority. Evidence: boundary tests.
    • X.19.c Run engine/DCC, VP, XR, mocap, render, and build journeys through public typed commands per supported host/profile. Evidence: traces.
    • X.19.d Exercise version incompatibility, host/device/resource loss, partial target, cancel/rollback, and read-back without duplicate runtime control. Evidence: fault matrix.
  • X.20 Prove Bellona cross-domain and lore-to-engine compilation preserves Hathor/Yemaya/Isis/Aja source entity IDs, canon, generated ingredients, consent, rights, revisions, mappings, and target artifacts. Evidence: deterministic evidence bundle and downstream read-back.
    • X.20.a Define producer-specific source manifests with stable entity/ revision/canon/ingredient/consent/rights/artifact identities and correction/revoke hooks. Evidence: contracts.
    • X.20.b Record every normalized intermediate and source-to-target entity/property/asset mapping, default, substitution, omission, and human decision. Evidence: mapping ledger.
    • X.20.c Compile with pinned toolchain/templates/adapters/models/config/ seeds/environment and preserve deterministic or disclosed nondeterministic step provenance. Evidence: compilation trace.
    • X.20.d Assemble signed evidence and read back target entities/assets/ state to verify all source IDs/revisions/canon/rights/mappings. Evidence: deterministic bundle.
    • X.20.e Propagate source/canon/consent/rights/artifact changes through recompilation/quarantine/recall and verify target closure. Evidence: lifecycle test.
  • X.21 Prove Veritas correction/retraction and Metis source withdrawal propagate into generated audio/video/lecture/diagram/simulation, newsletters/social/B2B, catalogs, and other registered derivatives. Evidence: complete consumer registry, saga, and verified read-back.
    • X.21.a Register every derivative type/channel/catalog/external consumer with producer source/claim refs, authoritative mutation endpoint, correction action, read-back, and owner. Evidence: zero-orphan registry.
    • X.21.b Emit versioned correction/retraction/withdrawal events with exact scope/effective time/replacement/reason/correlation and no protected payload. Evidence: event contract.
    • X.21.c Traverse generated media/teaching assets/publications/social/B2B/ catalog derivatives and preview required update/unpublish/recompute/ notification. Evidence: impact report.
    • X.21.d Execute a resumable idempotent propagation saga with per-consumer retry/compensation and explicit unreachable/partial state. Evidence: saga tests.
    • X.21.e Verify every registered target by read-back and keep incomplete closure as an incident. Evidence: reconciliation report.
  • X.22 Prove Euterpe and Aja biometric/voice/performer/subject consent revocation reaches Isis-generated derivatives, Yemaya projects, Bellona packages/builds, Metis lessons, and Veritas media. Evidence: impact preview and fail-closed recall journey.
    • X.22.a Maintain protected dependency edges from Euterpe voice/ performer and Aja subject/source consent grants to all registered derivatives/uses without exposing identity. Evidence: dependency audit.
    • X.22.b Ingest revocation with authenticated authority, exact grant/ subject-safe scope, effective time, purpose/region, replacement, and legal-hold semantics. Evidence: revocation contract.
    • X.22.c Preview affected Isis assets, Yemaya projects, Bellona artifacts/ packages/builds/targets, Metis lessons, and Veritas publications with release risk. Evidence: impact view.
    • X.22.d Immediately block new use/export/release and execute quarantine/ unpublish/recall/recompute across domains with retry. Evidence: fail- closed journey.
    • X.22.e Verify every store/cache/index/channel/target by read-back while preserving protected audit/legal-hold evidence. Evidence: closure report.
  • X.23 Prove marketplaces, plugins/models, external SDK clients, webhooks, remote agents, DCC hosts, devices, distributors, and channel adapters use common identity/tenant/version/audit principles without sharing secrets or authority implicitly. Evidence: external-boundary contract and adversarial matrix.
    • X.23.a Define common external principal/device/workload identity, enrollment/authentication, tenant/project scope, capability grant, version negotiation, secret ownership, audit, and revoke primitives. Evidence: boundary standard.
    • X.23.b Define specialized profiles for marketplaces/models/plugins, SDK/webhook clients, agents/MCP, DCC/devices, and distributors/channels without collapsing their risks or authority. Evidence: profile matrix.
    • X.23.c Prove credentials/secrets are audience/environment/tenant/ purpose-bounded, independently rotated, never forwarded, and cannot confer another integration's authority. Evidence: secret tests.
    • X.23.d Test spoof/replay/version downgrade, tenant confusion, privilege expansion, tool/content injection, supply-chain tamper, duplicate side effect, audit tamper, and revoke per profile. Evidence: adversarial matrix.
    • X.23.e Reconcile lifecycle/compatibility/security status across all external boundaries and surface unsupported/expired actors. Evidence: operational report.
  • X.24 Prove domain-specific taxonomies remain interoperable through versioned mappings rather than collapsed into a universal lossy enum. Evidence: round-trip fixtures for assets, media, rights, people/roles, locale, status, quality, and release concepts.
    • X.24.a Define independently owned/versioned domain taxonomies with stable IDs, meanings, aliases, constraints, lifecycle, and code-owned floors. Evidence: taxonomy registry.
    • X.24.b Define pairwise/contextual mapping records with source/ destination versions, exact/broader/narrower/compound/unmapped/ prohibited relation, conditions, and owner. Evidence: mapping schema.
    • X.24.c Require explicit unmapped/ambiguous/loss states and human/policy decision instead of a universal fallback enum. Evidence: negative tests.
    • X.24.d Round-trip asset/media/rights/people-role/locale/status/quality/ release fixtures and verify preserved identity or declared loss. Evidence: fixture matrix.
    • X.24.e Preview and migrate mapping/taxonomy version changes across contracts/search/data/gates/releases with compatibility windows. Evidence: lifecycle journey.

6. Rollout waves and final definition of done#

Recommended waves:

  1. Baseline, standards, scope decisions, and shared kit.
  2. Tara migration proof.
  3. Metis and Veritas, because their V1 authoring scope and grounding reuse are the clearest.
  4. Isis integration across the existing three audience hosts.
  5. Yemaya AAA-suite reconciliation plus study/case additions inside the existing Studio.
  6. Aja and Bellona operations consoles, with supported-host matrices.
  7. Euterpe integration and release profiles in its existing Studio.
  8. Cross-domain revocation, recovery, performance, and launch exercises.
  • DONE.1 Every checkbox claimed complete links to authoritative current evidence. Evidence: automated missing-evidence audit.
    • DONE.1.a Parse every parent/child checkbox and require exactly one non-placeholder evidence statement or immutable evidence reference. Evidence: ledger schema audit.
    • DONE.1.b Resolve referenced files/URLs/records, revisions/checksums, owners, dates, and environment against authoritative current sources. Evidence: evidence-link audit.
    • DONE.1.c Mark stale, missing, unverifiable, superseded, revoked, or unauthorized evidence as non-complete. Evidence: negative fixtures.
    • DONE.1.d Block a completed parent while any required child/evidence is incomplete. Evidence: hierarchy audit.
  • DONE.2 Every discovered package, route, app, contract, migration, job, gate, publisher, and critical journey has an explicit disposition. Evidence: zero unclassified inventory rows.
    • DONE.2.a Regenerate normalized inventories from repository, deployment, contract/schema, migration, job/event, gate, publisher, and test sources. Evidence: inventory manifests.
    • DONE.2.b Match each stable inventory row to exactly one retained/ refactored/migrated/replaced/removed/external/deferred disposition with owner/rationale/destination. Evidence: mapping audit.
    • DONE.2.c Resolve duplicate, conflicting, obsolete, and newly discovered rows through domain-owner review. Evidence: disposition log.
    • DONE.2.d Prove no unclassified or orphaned rows and no ledger disposition lacks a source item or explicit net-new scope. Evidence: zero-orphan audit.
  • DONE.3 Every capability in Section 3 is implemented per applicable domain or has an approved decision record with compensating control. Evidence: final release-evidence index and requirement audit.
    • DONE.3.a Map each Section 3 capability to every applicable domain/ profile/host/target with owner and applicability rationale. Evidence: coverage matrix.
    • DONE.3.b Link implemented cells to exact design/code/schema/migration/ test/operation evidence. Evidence: traceability audit.
    • DONE.3.c Link non-implemented cells to named approved decision, risk, compensating control, validation, expiry, and remediation owner. Evidence: exception audit.
    • DONE.3.d Block completion on missing, inherited, stale, or inapplicable- without-rationale cells. Evidence: final requirement gate.
  • DONE.4 Every domain-specific algorithm has independently verified known-value tests. Evidence: final release-evidence index and requirement audit.
    • DONE.4.a Register every production algorithm/version/profile and its inputs, outputs, units, applicability, owner, and release consumers. Evidence: algorithm registry.
    • DONE.4.b Provide analytically, independently, or externally derived valid/boundary/invalid fixtures with expected values/tolerances and checksums. Evidence: fixture manifests.
    • DONE.4.c Run current implementations against fixtures in supported environments and retain raw results/version provenance. Evidence: test reports.
    • DONE.4.d Prove no algorithm inherits another algorithm/profile's proof and unavailable evaluation cannot pass. Evidence: coverage gate.
  • DONE.5 Every model-based evaluator has current calibration, uncertainty, subgroup, threshold, drift, and human-review evidence. Evidence: final release-evidence index and requirement audit.
    • DONE.5.a Register exact model/evaluator/config/preprocessing/corpus/ labels/splits/intended-use versions and protected-data approvals. Evidence: evaluator registry.
    • DONE.5.b Measure performance, calibration, uncertainty/abstention, subgroup error, out-of-distribution behavior, and human disagreement on current held-out data. Evidence: evaluation report.
    • DONE.5.c Ratify pass/fail/review thresholds from declared costs and prohibit low-confidence/unavailable from passing. Evidence: threshold record.
    • DONE.5.d Verify drift monitoring, expiry, replacement/canary/rollback, and historical replay. Evidence: operational proof.
  • DONE.6 Every provider or external host is live in the approved environment or fails loudly as not configured/unavailable. Evidence: final release-evidence index and requirement audit.
    • DONE.6.a Register each required provider/host/target/environment with approved identity, endpoint, version/profile, credentials owner, capability, and health contract. Evidence: dependency registry.
    • DONE.6.b Run least-privilege live capability/health and representative operation probes in the approved environment. Evidence: probe reports.
    • DONE.6.c Exercise missing credentials/configuration, outage, timeout, incompatible version, degraded, and revoked states. Evidence: failure matrix.
    • DONE.6.d Prove APIs/UI/jobs/releases show explicit unavailable/not- configured state and never fixture or false success. Evidence: honesty gate.
  • DONE.7 Every mutation path has auth, object/property authorization, idempotency/concurrency, audit, tenancy, and negative tests. Evidence: final release-evidence index and requirement audit.
    • DONE.7.a Inventory every UI/API/job/CLI/SDK/webhook/agent/host/database mutation and its authoritative surface. Evidence: mutation registry.
    • DONE.7.b Test authentication, role/capability, tenant/project, object, property, purpose/consent, step-up/dual control, and secret boundaries. Evidence: authorization matrix.
    • DONE.7.c Test idempotency, expected revision/optimistic concurrency, duplicate/reorder/retry, TOCTOU, cancellation, and late result. Evidence: race suite.
    • DONE.7.d Verify immutable redacted audit/outbox, target read-back, and two-tenant isolation for success/failure/partial/compensation. Evidence: integration proof.
    • DONE.7.e Resolve or block every untested/duplicate/bypass mutation path. Evidence: zero-gap gate.
  • DONE.8 Every release/export path recomputes current gates, locks exact revisions, writes deterministically where possible, and verifies by target read-back. Evidence: final release-evidence index and requirement audit.
    • DONE.8.a Inventory every release/export/publish/deploy/package/channel path with target/profile/version and authoritative gate owner. Evidence: release registry.
    • DONE.8.b Recompute current quality, rights/consent, safety/integrity, security, compatibility, approval, and target gates immediately before execution. Evidence: freshness tests.
    • DONE.8.c Lock exact source/dependency/config/toolchain/model/adapter/ validator/profile/approval/target revisions and checksums. Evidence: candidate manifests.
    • DONE.8.d Produce deterministic bytes/manifests/signatures where possible or declare and bound nondeterminism. Evidence: reproducibility reports.
    • DONE.8.e Independently read back target contents/state/identity/ checksums/semantics and block transport-only success. Evidence: target verification matrix.
  • DONE.9 Every revocation/correction/recall path reaches all registered downstream dependencies and proves closure by read-back. Evidence: final release-evidence index and requirement audit.
    • DONE.9.a Register every revocable source/consent/right/model/canon/ artifact/claim/release authority and downstream derivative/consumer/ target edge. Evidence: dependency registry.
    • DONE.9.b Test exact impact preview, authority, effective scope/time, replacement, legal hold, and fail-closed new-use block. Evidence: control tests.
    • DONE.9.c Execute resumable idempotent propagation with checkpoint, retry, compensation, notification, and explicit unreachable/partial state. Evidence: saga matrix.
    • DONE.9.d Read back every store/cache/index/package/channel/target/ external consumer and keep unresolved closure as an incident. Evidence: reconciliation proof.
  • DONE.10 Every critical browser journey has deep Playwright coverage against live or explicitly labeled hermetic dependencies. Evidence: final release-evidence index and requirement audit.
    • DONE.10.a Inventory critical role/domain/profile/host/target journeys, mutations, failures, mobile paths, and accessibility alternatives. Evidence: browser coverage matrix.
    • DONE.10.b Use live approved dependencies or explicit production- equivalent hermetic adapters with persistent fixture labeling and contract parity. Evidence: environment manifests.
    • DONE.10.c Assert authoritative backend/target state, exact revisions, evidence/lineage, authorization, idempotency, and read-back—not only screenshots or DOM. Evidence: Playwright traces.
    • DONE.10.d Cover loading/empty/stale/partial/degraded/unavailable/failure/ retry/cancel/recovery and retain diagnostics/artifacts. Evidence: fault coverage.
    • DONE.10.e Eliminate or approve every critical unautomated gap with owner, reason, equivalent proof, and expiry. Evidence: final audit.
  • DONE.11 Every applicable route meets WCAG 2.2 AA with keyboard, announcement, non-drag, zoom, theme, viewport, locale, and semantic alternative coverage. Evidence: final release-evidence index and requirement audit.
    • DONE.11.a Inventory authenticated/public routes, roles, states, complex widgets, visualizations, media, and critical actions by applicability. Evidence: accessibility matrix.
    • DONE.11.b Run automated WCAG checks across themes, locales, and wide/ compact/single-column viewports. Evidence: automated reports.
    • DONE.11.c Complete keyboard/screen-reader journeys with focus, announcements, error recovery, step-up, status, and dynamic updates. Evidence: assistive-technology traces.
    • DONE.11.d Verify non-drag controls, 200–400% zoom/reflow, touch targets, color independence, reduced motion, captions/transcripts, and semantic graph/3D/canvas/media parity. Evidence: manual/automated results.
    • DONE.11.e Close or explicitly block every severity/applicability gap; no domain/profile inherits another's evidence. Evidence: release gate.
  • DONE.12 Every interactive and asynchronous path meets its ratified performance, reliability, capacity, and cost budgets. Evidence: final release-evidence index and requirement audit.
    • DONE.12.a Inventory interactive routes/actions and async API/queue/ worker/provider/host/build/publisher paths with populations and owners. Evidence: budget registry.
    • DONE.12.b Ratify latency/responsiveness/throughput/availability/ correctness/durability/capacity/cost indicators and budgets by tier. Evidence: SLO records.
    • DONE.12.c Run representative/worst-supported sustained, burst, large- object, slow dependency, and failure/recovery tests. Evidence: load reports.
    • DONE.12.d Verify backpressure/quotas/fairness/degradation/cancellation, bounded resource use, tenant isolation, data integrity, and cost. Evidence: overload proof.
    • DONE.12.e Link operational telemetry/error budgets and block unresolved regressions or unmeasured critical paths. Evidence: final gate.
  • DONE.13 Every supported standards/profile/host/target pair has its own conformance evidence and declared limitations. Evidence: final release-evidence index and requirement audit.
    • DONE.13.a Generate the exact supported standards/profile/producer/ consumer/host/plugin/OS/target/version pair matrix with owners and support windows. Evidence: compatibility registry.
    • DONE.13.b Run pair-specific contract, known-value/fidelity, security, accessibility, performance, failure/recovery, and read-back tests. Evidence: conformance bundles.
    • DONE.13.c Record unsupported/partial/lossy/nondeterministic limitations, mitigations, UI/docs/API discovery, and expiry per pair. Evidence: limitations registry.
    • DONE.13.d Declare unavailable cells explicitly and block one pair/ profile/host/target from inheriting another's proof. Evidence: independence audit.
  • DONE.14 Security, privacy, safety, rights, integrity, accessibility, and supply-chain scans have zero unresolved release blockers. Evidence: final release-evidence index and requirement audit.
    • DONE.14.a Define current threat/data/safety/rights/accessibility/supply- chain scopes, severity policy, scanner/tool/config versions, owners, and exception rules. Evidence: assurance plan.
    • DONE.14.b Run code/dependency/container/IaC/secret/license/SBOM/ provenance, application/host/agent adversarial, privacy leakage, safety/ claim, rights/consent, integrity, and accessibility assessments. Evidence: current reports.
    • DONE.14.c Deduplicate/triage findings, prove remediation with targeted regression, and record residual limitations. Evidence: closure records.
    • DONE.14.d Verify zero unresolved release blockers and no expired, overbroad, conflicting, or self-approved exception. Evidence: release gate.
  • DONE.15 Backup, restore, replay, rollback, revocation, dependency outage, and disaster-recovery drills meet ratified RPO/RTO. Evidence: final release-evidence index and requirement audit.
    • DONE.15.a Inventory authoritative databases/blobs/evidence/config/keys/ indexes/queues/targets and ratify RPO/RTO, retention, restore order, and owners. Evidence: recovery plan.
    • DONE.15.b Execute backup integrity and clean-environment restore with checksums, keys/config, permissions, queues/indexes, and target reconciliation. Evidence: restore report.
    • DONE.15.c Execute event/job replay, release/deployment rollback, source/consent/correction revocation, and provider/host/target outage drills. Evidence: drill matrix.
    • DONE.15.d Measure data loss, downtime, duplicate/partial work, consumer/ target inconsistency, and achieved RPO/RTO. Evidence: scorecard.
    • DONE.15.e Remediate and rerun every missed objective or unresolved recovery path. Evidence: closure proof.
  • DONE.16 Documentation, generated API references, compatibility matrices, user help, runbooks, and architecture diagrams match the shipped system. Evidence: final release-evidence index and requirement audit.
    • DONE.16.a Inventory required operator/developer/user/policy/runbook/ architecture/compatibility documents with owners, audience, source, and release version. Evidence: documentation registry.
    • DONE.16.b Regenerate API/event/SDK/CLI/host contracts and compatibility/ limitation matrices from authoritative sources. Evidence: drift-free artifacts.
    • DONE.16.c Verify commands, routes, permissions, examples, screenshots, links, diagrams, versions, defaults, failure/recovery, and support contacts against the release candidate. Evidence: doc tests.
    • DONE.16.d Run representative user/operator/developer tasks from docs in clean environments and correct discrepancies. Evidence: walkthroughs.
    • DONE.16.e Block stale, contradictory, inaccessible, unlocalized where required, or orphaned documentation. Evidence: final docs gate.
  • DONE.17 Feature flags, fixtures, compatibility shims, temporary migrations, deprecated routes, and honesty debt each have an owner and removal date. Evidence: final release-evidence index and requirement audit.
    • DONE.17.a Inventory every flag, fixture/demo path, shim/adapter, temporary dual read/write/backfill, deprecated route/API/schema, and misleading/placeholder state. Evidence: debt registry.
    • DONE.17.b Record purpose, exact scope, owner, introduced date, dependencies, risk, telemetry, removal criterion/date, and rollback. Evidence: complete rows.
    • DONE.17.c Prove production fixtures are visibly labeled/non-mutating, flags fail safely, shims honor compatibility, and deprecations notify consumers. Evidence: behavior tests.
    • DONE.17.d Create tracked removal work and alert/escalate overdue or ownerless entries; block release blockers and expired shims. Evidence: governance gate.
  • DONE.18 Engineering-readiness packets are current and reproducible from repository evidence. Evidence: final release-evidence index and requirement audit.
    • DONE.18.a Define packet schema for release identity, scope, traceability, contracts/migrations, verification, assurance, SLO/performance, recovery, runbooks, limitations, approvals, and provenance. Evidence: schema.
    • DONE.18.b Assemble packets from immutable repository/test/build/target artifacts and approved external evidence with exact revisions/ checksums. Evidence: generated packets.
    • DONE.18.c Validate links/checksums/freshness/coverage/signatures and surface missing/failed/stale/waived/unavailable proof. Evidence: packet audit.
    • DONE.18.d Reproduce packets from a clean checkout/environment and obtain identical canonical hashes except declared nondeterministic fields. Evidence: reproducibility test.
  • DONE.19 Named human governance approvals exist for each audience, host, standard profile, and supported integration. Evidence: final release-evidence index and requirement audit.
    • DONE.19.a Register required approver roles/competence/independence, scope, evidence prerequisites, decision options, conditions, expiry, and invalidation triggers. Evidence: approval policy.
    • DONE.19.b Generate exact audience/host/profile/integration coverage matrix and attach the immutable release/readiness evidence presented. Evidence: review packets.
    • DONE.19.c Capture named signed approve/reject/conditional decisions with role, time, exact revision/scope, rationale, conditions, and conflicts. Evidence: approval records.
    • DONE.19.d Block missing/inherited/expired/stale/conflicted approvals and reapprove after material change. Evidence: governance gate.
  • DONE.20 The release commit is present on origin/main; no local-only commit or branch divergence remains. Evidence: final release-evidence index and requirement audit.
    • DONE.20.a Verify applicable checks pass on the exact staged diff and no unrelated/user changes are included. Evidence: verification/staging record.
    • DONE.20.b Commit with a task-scoped message and record commit identity, tree, authorship, verification, and release-evidence refs. Evidence: local commit.
    • DONE.20.c Fetch and integrate current origin/main, resolve conflicts, and rerun affected verification when remote history moved. Evidence: integration record.
    • DONE.20.d Push the exact release commit to origin/main and verify local HEAD equals the fetched remote branch tip. Evidence: remote parity check.
    • DONE.20.e Verify no relevant uncommitted change, local-only commit, or branch divergence remains. Evidence: final Git status.
  • DONE.21 The atomic capability inventory is bidirectionally complete: every feature asserted or implemented by an app, route, component, package, contract, API/job/CLI, schema, event, test, inventory, deep dive, walkthrough, proposal, or release spec has a reviewed disposition and ledger mapping, and every ledger feature maps back to authoritative or explicitly net-new scope. Evidence: deterministic zero-orphan audit and seven named domain-owner approvals.
    • DONE.21.a Enumerate and normalize feature assertions from every named repository and documentation source with stable source location, symbol/route/contract ID, domain, owner, and checksum. Evidence: deterministic source inventory.
    • DONE.21.b Decompose compound assertions into independently reviewable capabilities without generating generic or keyword-derived task text. Evidence: semantic review record.
    • DONE.21.c Map every source capability to one or more exact ledger parents/children and a reviewed retain/refactor/migrate/replace/remove/ external/defer/net-new disposition. Evidence: forward mapping.
    • DONE.21.d Map every ledger capability back to authoritative source items or an explicitly approved net-new requirement with rationale/ owner/proof. Evidence: reverse mapping.
    • DONE.21.e Detect and resolve unmapped/orphaned/duplicate/conflicting/ obsolete/ambiguous items deterministically and preserve reviewed exceptions. Evidence: zero-orphan audit.
    • DONE.21.f Obtain named Isis, Metis, Yemaya, Veritas, Euterpe, Aja, and Bellona owner approval of their inventories, dispositions, boundaries, and unresolved limitations. Evidence: seven signoffs.

This ledger is complete only when the evidence proves the full system. A large amount of existing code, a green narrow test, an attractive demo, or the absence of an obvious TODO is not proof of completion.