Eve library

V1–V9 Autonomous Content SOTA Gap-Closure TODO Ledger — 2026-07-15

Tracker

Revision 1

On this page

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

0. Purpose, authority, and completion law#

This is the executable remediation ledger for every actionable gap, recommendation, acceptance slice, release requirement, and validated follow-up finding in:

  • V1_V9_AUTONOMOUS_CONTENT_SOTA_AUDIT_2026-07-15.md;
  • V1_V9_SOTA_AUDIT_REVIEW_2026-07-15.md;
  • the repository evidence paths cited by those documents.

Baseline repository tip when this ledger was created: 07186e3737a64637e6227bbed19de0d4a7d1c516.

This file is deliberately more granular than the audit. It converts every gap into implementation, integration, verification, evidence, and operational tasks. It does not mark a seam complete because an interface, fixture, prompt, plan, fallback, provider request, or passing shape test exists.

Implementation is the closure target. Relabeling a missing capability, removing it from a contract, narrowing the advertised scope, or documenting it as absent may be required for interim honesty, but none of those actions closes the corresponding checkbox. Deterministic fallbacks and human escalation may remain as resilience controls; the real primary capability must still be built, invoked, and proved end to end.

0.1 Validated corrections applied while deriving this ledger#

  • V3's stored launch artifact reports 30/41 required commands passing and 24/39 named gates passing, with gaReady: false; the 29/40 figure in the canonical audit and upstream V3 audit is wrong.
  • The V3 artifact's generatedAt proves the consumed evidence is stale. It does not prove that nobody ran the verifier without rewriting evidence.
  • No in-repository operator UI consumer of /v1/agentic/autonomy/* was found. This ledger does not claim an external private client is impossible; it requires repository and deployed-consumer evidence.
  • V1 does record and persist CycleReport.realTokenUsage, but that value is partial: it principally reflects planner usage and does not provide complete generator, research-loop, critic, judge, tool, media-provider, GPU, or monetary accounting. The gap is incomplete end-to-end measured cost accounting, not total absence of token measurement.
  • The repository declares a canonical single-Hetzner-box V1 topology. A live topology probe is still required before claiming the deployed environment exactly matches that declaration.

0.2 Checkbox semantics#

  • [ ] means unproved or incomplete at this baseline.
  • [x] is allowed only after every nested requirement is satisfied by current, committed evidence.
  • A parent item stays open while any child is open.
  • Provider-, data-, engine-, cohort-, credential-, or production-gated work stays open until the real dependency is exercised. A fail-loud seam is honest engineering but is not closure of the external behavior.
  • Historical test counts and reports are not proof of current completion.
  • A task that changes source, policy, prompts, schemas, or generated artifacts must invalidate and rerun all dependent evidence.

0.3 Evidence pack required for every completed work package#

  • Record owner, implementation commit, affected content slots, and linked task IDs.
  • Record exact verification commands, repository commit, host/OS, relevant dependency and model/provider versions, start/end timestamps, and exit codes.
  • Attach value-asserting unit, integration, contract, adversarial, and E2E results appropriate to the changed boundary; explain any inapplicable avenue.
  • Attach artifact IDs/hashes, source/canon/policy/rubric versions, release verdicts, runtime placement evidence, and screenshots/media probes where an artifact or user flow changed.
  • Demonstrate at least one negative/adversarial fixture that makes the new gate fail for the intended reason.
  • Record human reviewer identity/role and signed decision where the policy requires human authority.
  • Record remaining external blockers without substituting mocks, fixtures, metadata, or deterministic fallbacks for real results.
  • Update generated capability/status views from machine evidence; do not hand-edit a green claim.

Before implementation begins, each parent work package must also record an owner, accountable product, priority, dependencies, risk tier, target milestone, affected schemas/slots, rollout strategy, rollback strategy, and evidence-pack location. Splitting work into issue-tracker children must preserve the IDs and completion conditions in this ledger.

Historical ledgers such as AGENTIC_CONTENT_AUTONOMY_GAP_CLOSURE_TODOS_2026-06-17.md and V1_SOTA_AUDIT_REMEDIATION_TODOS_2026-07-02.md remain implementation history. Their checked boxes may be linked as evidence, but are not automatically carried forward when this audit found the invoked end-to-end behavior still open.

0.4 Dependency order#

text
Audit truth and evidence freshness
  -> coverage registry + typed proof vocabulary
  -> common durable run/event/intervention spine
  -> invoked tool authority, cost accounting, and human control
  -> product-specific real vertical slices
  -> calibrated evaluation and live learning
  -> scaled coverage, autonomy, and continuous recertification

Frontier P2 work must not displace open P0 control-plane or proof gaps.

1. P0-A — Audit integrity, truth sources, and reproducibility#

AUD-001 — Correct the V3 launch-gate record everywhere#

  • Change V1_V9_AUTONOMOUS_CONTENT_SOTA_AUDIT_2026-07-15.md from 29/40 to 30/41 required commands and 24/39 named gates, preserving red and gaReady: false.
  • Correct V3_AUTONOMOUS_CONTENT_AGENTS_AUDIT_2026-07-14.md and every generated HTML or derived view that repeats 29/40 as current state.
  • Add a regression check that reads V3/launch/launch-readiness-release-gate.json rather than duplicating hand-maintained counts.
  • Verify prose and machine artifacts agree at the same commit.

Closure evidence: evidence/v1-v9/AUD-001.md; implementation commit d223d3e65696e8bd1532bf6a0c5c12987f050263.

AUD-002 — Make evidence freshness an enforceable property#

  • Add generatedAt, generator version, source commit, command identity, host identity, and evidence schema version to every generated gate artifact.
  • Define a freshness SLO per evidence class; do not use one universal age.
  • Make a consumer gate red when required evidence is older than its SLO, missing a source commit, or was generated against a different contract.
  • Distinguish “verifier was executed” from “current evidence was written and published.”
  • Rerun pnpm verify:v3:launch-readiness on the designated capable host, write fresh evidence through the generator's supported path, and retain the exact command/log.
  • Add stale, future-dated, wrong-commit, and tampered-evidence fixtures.

Closure evidence: evidence/v1-v9/AUD-002.md; final generated-gate portfolio commit d0f3dabf7ee7af0b94bcd6a4ad1d048efccded5b.

AUD-003 — Make the canonical audit verification table reproducible#

  • Add the exact command for every count in audit section 1.4.
  • Record included/excluded package boundaries, especially the composition of the 176 V9 tests.
  • Record the tested HEAD, lockfile hash, runtime versions, OS/host, RAM, swap, and relevant environment gates.
  • Store bounded logs or immutable CI artifact links for every result.
  • Mark historical host resource observations as observations, not timeless repository properties.
  • Add a script that regenerates the verification table from evidence records and fails on manual count drift.

Closure evidence: evidence/v1-v9/AUD-003.md; tested source commit 6f806b6f6cef12d42f4597518ebbe878699504b8; published record commit 16ac905031fb99d65e8725be87c86fc0508c15c2.

AUD-004 — Qualify host-dependent Unreal instructions#

  • Name the Linux UE5.5 execution host/toolchain in every closure instruction that depends on /root/workspace/UnrealEngine-5.5.
  • Document the equivalent remote-job path for developers on macOS or hosts without Unreal.
  • Record engine version, project, commandlet, run-as user, target platform, and output artifact in UE evidence.
  • Prevent a host without UE from converting “not run here” into “UE is unavailable to the project.”

Every product instruction marked [UE-AUD-004] is bound to the designated ue55-linux self-hosted profile and the executed-evidence contract in evidence/v1-v9/AUD-004/unreal-evidence-runbook.md. Missing local prerequisites produce only not-run-on-this-host; they never satisfy or fail the remote product gate.

Closure evidence: evidence/v1-v9/AUD-004.md; committed contract source ee25c59ab561b8e35df0d62fffbb4650966e905f; historical local status not-run-on-this-host at 2026-07-15T14:57:59.116Z. Current 2026-08-02 source-bound evidence records the engine and ueagent as present, exact V2/V3 project rebuild blockers, and bounded V4/V5/V7 executions without weakening any remaining provider, human, operational, or release gate.

AUD-005 — Eliminate machine-adjacent documentation overclaims#

  • Correct the runs-route.ts comment that claims the executed envelope contains tool calls, evidence, and cost before that behavior exists.
  • Correct case-pipeline/provenance.ts comments that describe a locator-derived hash as an asset-byte hash.
  • Add enforcement-seam comments and generated schema docs to the same drift checks as umbrella documentation where feasible.
  • Add static assertions or tests that examples/comments naming populated fields are exercised by a real path.

Closure evidence: evidence/v1-v9/AUD-005.md; implementation source commits 13602f854cee508abcc3246e340c732e2324c8bb and dc51ad2a613333881200f8d054363367fcfbbaac; refreshed canonical record commit ae73dc3b0f57b56ae836a5da1366ee64214c02ca.

AUD-006 — Generate capability truth instead of maintaining parallel prose#

  • Designate the future coverage registry and evidence store as the canonical machine truth.
  • Generate V_SERIES.md, CONTENT_GENERATION_CAPABILITY_MATRIX.md, product status views, and audit summaries from that truth.
  • Preserve manual analysis only in explicitly labeled commentary fields.
  • Fail CI when prose claims implemented/live/green while the registry or evidence records missing, stale, blocked, seam-only, or red.
  • Include source comments, TODO ledgers, test counts, provider probes, artifacts, placement receipts, and release records in contradiction detection.

Closure evidence: evidence/v1-v9/AUD-006.md; canonical registry source commit fbc29f40f565b8561d17e91544c7f0b91f6d5839; refreshed verification-record publication commit 977b335bbcb04a409c466c45d8c804e2505dd509 (tested source 3bfb0b285e670bbaa25c2995c04494efdf45761d).

AUD-007 — Apply test-coherence sampling to audit evidence#

  • Sample every credited suite for domain-value assertions rather than only schema, snapshot, mock-call, or happy-path shape checks.
  • Classify tests as unit/contract/integration/real-provider/engine/E2E/live and prevent one class from impersonating another.
  • Add mutation or deliberately broken fixtures to prove load-bearing tests fail when the claimed behavior is removed.
  • Publish uncovered claims and weakly covered assertions beside pass counts.

Closure evidence: evidence/v1-v9/AUD-007.md; coherence implementation source commit d76869b2ab861c4958f89c7447c2236c1ea7920f; refreshed verification-record publication commit 977b335bbcb04a409c466c45d8c804e2505dd509.

AUD-008 — Reconcile the adversarial review into canonical truth#

  • Fold every validated follow-up addition into the canonical audit rather than leaving two competing truth documents.
  • Qualify V1 human control as a real tested component with no discovered in-repository operator UI; keep the evidence-ladder score distinct from end-user operability.
  • Amend the follow-up spend section to acknowledge persisted partial realTokenUsage and state the actual missing full-cost boundary.
  • Phrase V3 freshness, external UI possibility, and declared/live topology claims at the evidence strength actually proved.
  • Link the follow-up review as adversarial provenance and record which findings were accepted, corrected, or rejected.
  • Regenerate derived HTML/status documents and run contradiction checks.

Closure evidence: evidence/v1-v9/AUD-008.md; reconciliation implementation source commit 597c63f958a48a70ca96413c15b54ad65537f3da; post-merge tested source 3bfb0b285e670bbaa25c2995c04494efdf45761d; refreshed verification-record publication commit 977b335bbcb04a409c466c45d8c804e2505dd509.

2. P0-B — Canonical content coverage graph#

COV-001 — Define and version the canonical content-slot schema#

  • Implement fields for slot ID, product, surface, content kind, modalities, locale, audience, required quantity, freshness SLO, source policy, autonomy policy, pipeline, gate suite, rights policy, accessibility requirements, placement target, owner, and risk tier.
  • Implement slot states: missing, planned, running, review, released, stale, invalidated, blocked, and withdrawn.
  • Require active artifact refs, latest run ref, blocker refs, next-required timestamp, and evidence freshness state.
  • Version and migrate the schema; reject unknown fields that would weaken policy silently.
  • Define stable identifiers and ownership rules so imports are idempotent.

Closure evidence: evidence/v1-v9/COV-001.md; implementation source commit 2ada10c7b9831b18d9217bc64a15352e3e254a3f; post-merge tested source 6f806b6f6cef12d42f4597518ebbe878699504b8; canonical verification-record publication commit 16ac905031fb99d65e8725be87c86fc0508c15c2.

COV-002 — Seed the registry mechanically from existing repository material#

  • Extract the frozen 28 V1 generation surfaces as initial surface rows.
  • Extract the four live V1 autonomy categories as current scheduled slots, without treating four categories as full coverage.
  • Extract every V1–V9 launch requirement and unchecked/blocked item from version TODO ledgers.
  • Extract V8/V9 gate contracts, product-specific required artifacts, and placement targets.
  • Reconcile duplicates and contradictions through reviewed migration rules; do not silently discard either source.
  • Publish registered/unregistered counts on day one and assign an owner to every unresolved row.

Closure evidence: evidence/v1-v9/COV-002.md; deterministic registry: content-coverage/seed-registry.v1.json; day-one report: content-coverage/seed-registry-report.md; implementation source commit 229a9056a334ecfacf9253c4751d5c26c791b851; post-merge tested source c3ddbd66c498d43401b111549ba020fbb9657537; refreshed canonical-record publication commit a88b6458c6790f3f84bc099210c9696c0b146dee.

COV-003 — Model the full dependency and invalidation graph#

  • Represent source, translation, consented voice, generation, watermark, caption, provenance, cook/render, placement, review, and freshness edges.
  • Model canon/entity/source/claim/character/world/asset relationships across products while keeping product views bounded.
  • Propagate source, canon, policy, rights, model, rubric, provider, and placement invalidation to dependent artifacts and approvals.
  • Detect cycles, orphan artifacts, unknown shipped content, and dangling evidence references.
  • Schedule missing or stale leaf work rather than attractive but unneeded content.

Closure evidence: evidence/v1-v9/COV-003.md; strict portfolio dependency contract: libs/contracts/src/content-coverage/dependency-graph.ts; implementation source commit 91ff4e9c1c453d16be7f200c1ac426c8b59e25a6; first remote-integrated tested source 8967eec7fce866f83e05b0d0856ff62df4fe6d96; publication metadata commit a93dd7b000339df8bdaa60d0944e88ab98defb49.

COV-004 — Inventory locale, modality, accessibility, and manual content#

  • Record the actual current locale count for every live autonomy category and product surface.
  • Register untranslated or single-locale behavior explicitly as a gap.
  • Record required alt text, captions, transcripts, audio description, keyboard/controller access, reading level, and assistive metadata per content kind.
  • Bring human-authored MOTD, patch notes, replacement artifacts, and other intentionally manual content into the same lifecycle and lineage ledger.
  • Report required/current/missing/stale/blocked quantities by product, surface, modality, locale, source policy, and accessibility requirement.

Closure evidence: evidence/v1-v9/COV-004.md; strict inventory contract: libs/contracts/src/content-coverage/inventory-registry.ts; deterministic registry/report: content-coverage/content-inventory.v1.json and content-coverage/content-inventory-report.md; implementation source commit 968c65790ff87e457c8cb7ccc8aff38a6ad38cae; remote-integrated tested source f0665ab54fdf1f62bea260ba51f023bc45c2f52e; publication metadata commit e64c8584ff73f564c3a942b4f52865efe5578040.

COV-005 — Enforce coverage closure#

  • Block release for unregistered slots, wrong source/autonomy policy, insufficient quantity, missing placement, stale content, or silent unknown shipped artifacts.
  • Require 100% of launch requirements to be registered.
  • Require 100% of active artifacts to link release, placement, gate, rights, accessibility, and lineage evidence.
  • Make percentage of shipped artifacts absent from the registry observable with target zero.
  • Add contract, migration, import, reconciliation, dependency, invalidation, and adversarial unknown-content tests.

Closure evidence: evidence/v1-v9/COV-005.md; strict release-closure contract: libs/contracts/src/content-coverage/release-closure.ts; implementation source commit 34aa7cc35c224fa369774f503f7a99e324895013; first remote-integrated tested source 93404a57474a99c8f8f40993c43d16c67251b268; publication metadata commit bd93d4fdea6ad0e743f08de1570bdf3604c473ca.

3. P0-C — Common durable run, event, and publication spine#

RUN-001 — Define one versioned Run/Step/Attempt/Artifact/Intervention model#

  • Define immutable run identity, parent/fork lineage, tenant, actor, purpose, content slot, risk tier, and autonomy level.
  • Define versioned plans and per-step dependency/state transitions.
  • Define attempt identity, retry reason, deterministic seed, and idempotency key.
  • Define tool request, approval, resolved grant, invocation, result, side-effect, and reconciliation events.
  • Define sources, canon, policies, prompts, models, rubrics, providers, artifacts, gates, interventions, cost, latency, token, GPU, and environmental measurements.
  • Define publication, placement, canary, rollback, withdrawal, revocation, invalidation, and recertification states.

Closure evidence: evidence/v1-v9/RUN-001.md; strict common run protocol: libs/contracts/src/run-protocol/run-model.ts; implementation and first remote-integrated tested source commit 84752f8c11923679953650e84438d335ea62a2a6; publication metadata commit 6ada35547cea8e16deb92941c345d5fa92d63d55.

RUN-002 — Implement durable append, checkpoints, migrations, and replay#

  • Persist every state transition before externally visible publish or tool effects.
  • Add checkpoint schema versions and forward/backward migration tests.
  • Resume on another process using a stable run/thread identity.
  • Allocate sequence and attempt IDs atomically in the store, not in process memory.
  • Reject duplicate/nonmonotonic events and preserve deterministic replay.
  • Compact/archive history without losing active lineage, approvals, legal records, or learning attribution.
  • Test process death at every transition and recover without fabricating a completed state.

Closure evidence: evidence/v1-v9/RUN-002.md; portable durability protocol: libs/contracts/src/run-protocol/durable-run-protocol.ts; crash-recoverable reference adapter: libs/contracts/src/run-protocol/file-durable-run-store.ts; first remote-integrated tested complete source commit 754a4ec55c812ecffc9c244169301a435501083f; publication metadata commit a38ba5ca50954e474d2363ac396693d6aa541a99.

RUN-003 — Make side effects idempotent and reconcilable#

  • Derive effect keys from run, step, attempt, tool, target, and artifact lineage.
  • Record intent before execution and authoritative provider result after execution.
  • Reconcile timeout/unknown outcomes with the provider before retry.
  • Prevent owner-controlled reuse of a run ID from repeating a mutating effect.
  • Implement exactly-once publication where possible and safe idempotent retry otherwise.
  • Add duplicate delivery, lost acknowledgment, timeout, partial success, restart, failover, and concurrent-worker tests.

Closure evidence: evidence/v1-v9/RUN-003.md; portable provider-effect protocol: libs/contracts/src/run-protocol/durable-effect-protocol.ts; shared-filesystem reference adapter: libs/contracts/src/run-protocol/file-durable-run-store.ts; implementation and first remote-integrated tested source commit aee2879220a22a0bc4402c5828a13acb135059a3; publication metadata commit 585ee7acd82e529ff54073ddced526e73e133f76.

RUN-004 — Adapt existing runtimes instead of creating competing standards#

  • Map V1 AgentRun and creative-autonomy events into the common model.
  • Map V8 StageJournal events and case-gate evidence into the common model.
  • Map V9 dependency/governance/agent-plan records into the common model.
  • Map V6 cognition run events and child authority into the common model.
  • Publish adapter conformance tests that prove semantic equivalence and identify unsupported fields explicitly.
  • Deprecate parallel envelopes only after migration, replay, and rollback evidence passes.
    • Boundary enforced: no parallel envelope was deprecated because product migration, replay, rollback, and zero-loss evidence is not present.

Closure evidence: evidence/v1-v9/RUN-004.md; shared conformance contract: libs/contracts/src/run-protocol/adapter-conformance.ts; V1, V8, V9, and V6 source adapters live beside their authoritative product envelopes; implementation and first remote-integrated tested source commit 0f68a5e5a6a72cef4f32fbc3e96db8e732c5c00c; publication metadata commit a70125182c315d6560c9a9fddefa2bb364476f17.

RUN-005 — Make approval and execution one atomic lifecycle#

  • Bind submitted, approved, executing, paused, tool-wait, review, completed, rejected, discarded, cancelled, killed, rolled-back, and withdrawn states into one state machine.
  • Version approvals against plan, source, artifact, policy, and gate hashes.
  • Invalidate stale approvals automatically after any dependent mutation.
  • Prevent approval-to-execution races and execution without the required current approval.
  • Persist final outputs, failures, rationale, tool/evidence ledger, and actual cost before marking completion.

Closure evidence: evidence/v1-v9/RUN-005.md; shared atomic lifecycle: libs/contracts/src/run-protocol/atomic-run-lifecycle.ts; durable transaction adapter: libs/contracts/src/run-protocol/file-durable-run-store.ts; implementation and first remote-integrated tested source commit f20ba9c007fdd175ed45faea2893a1768bdf0982; publication metadata commit c61768e3fb7178880f082bf2694dc6e11ea6512a.

4. P0-D — Human authority and co-authoring control plane#

HITL-001 — Build a human-usable operator surface for the V1 autonomy API#

  • Implement authenticated queue, run, artifact, provenance, source, budget, gate, and intervention views in an operator web console.
  • Wire approve, reject, veto, tweak, node-addressed tweak, standing direction, feedback, withdrawal, and rollback to the real endpoints.
  • Display stale evidence/approval warnings and distinguish held, blocked, failed, review-window, and released states.
  • Add accessible keyboard/screen-reader flows, confirmation for destructive actions, and clear irreversible-effect warnings.
  • Add deep Playwright coverage for every action, authorization failure, concurrent edit, stale version, reload, reconnect, and error recovery.
  • Prove the deployed console—not only an HTTP test—can operate the live review loop.

Closure evidence: evidence/v1-v9/HITL-001.md; shipped operator route: apps/oshun/web/src/app/operator/autonomy/page.tsx; control surface: apps/oshun/web/src/components/operator/AutonomyOperatorConsole.tsx; retained browser-to-real-BFF proof: apps/oshun/web/e2e/operator-autonomy-live-bff.spec.ts; implementation commit 06fe5f10396c0f52e22ccb92bfbb8f3369cb8401; live browser proof commit 7244cd392183bc26c1fa4f18b65b61ede1ada8a4; publication metadata commit 6dd993da7e47c8b52085bedc55be10d8fad174dd; canonical verification refresh 6e9fe8f8238db551b04bdae39277beef4b615506.

HITL-002 — Implement the complete intervention API#

  • Observe plan, current step, hidden/pending tool effects, artifacts, sources, budgets, gates, and lineage in real time.
  • Pause, resume, cancel, kill, choose branch, fork, rewind, retry with changes, and roll back publication.
  • Comment at run/artifact/node/rubric dimension and attach references.
  • Edit brief, plan, source set, tool arguments, and artifact nodes.
  • Claim a step, author from blank, replace an artifact, and merge a human branch.
  • Approve, reject, veto, defer, reassign, require extra review, grant a scoped exception, and change future autonomy level.
  • Preserve non-overridable safety, rights, child, crisis, identity, and irreversible-action laws.

Closure evidence: evidence/v1-v9/HITL-002.md; canonical implementation: libs/oshun/agentic-studio/src/interventions/intervention-api.ts; complete focused suite 20/20 and full Agentic Studio suite 309/309; source commits c6d53c2c25170209834d072b520ffbac2411d04b and cfa056db69aea4170c8211f3a23e08956252b0a8; closure evidence commit 3f21f7e7b183b1bcf209f5aad97aa3f494b18f2a.

HITL-003 — Make intervention transactional and provenance-bearing#

  • Claim a versioned step with optimistic concurrency.
  • Apply human edits to the canonical artifact/plan, not a sidecar note.
  • Invalidate dependent judgments, approvals, signatures, caches, and placement receipts.
  • Rerun the same required gates on human, agent, and mixed artifacts.
  • Record human identity, role, reason, inputs, edit diff, branch merge, and signed mixed lineage.
  • Resume from a durable checkpoint on another process without repeating effects.
  • Add restart, race, stale approval, duplicate effect, unauthorized edit, abandoned claim, merge conflict, and rollback tests.

Closure evidence: evidence/v1-v9/HITL-003.md; canonical transaction protocol: libs/contracts/src/run-protocol/transactional-intervention.ts; affected run-protocol regression 67/67 across 6/6 files; source commit 10f8aa6505384232602c879de91572643f7a3e5d; closure evidence commit 2d67fe50c01fdc2bc23804ca1a048423eab85a63.

HITL-004 — Compose existing authority primitives#

  • Bind Agentic Studio pause/resume/cancel/kill/fork/retry/operator-step functions to the V1 BFF and common store.
  • Bind scoped, expiring Lilith overrides with revocation and recurring-use review.
  • Bind Yemaya emergency stop, takeover, and checkpoint semantics.
  • Bind Studio Authoring drafts, blocks, branches, comments, replacements, checklists, and lifecycle to autonomous artifacts.
  • Prove human-authored replacements enter the exact provenance, quality, release, placement, and learning path used by agent output.

Closure evidence: evidence/v1-v9/HITL-004.md; canonical V1 bindings: apps/oshun/bff/src/agentic/run-lifecycle-store.ts and apps/oshun/bff/src/agentic/autonomous-authoring-store.ts; V1 AgentRun HTTP integration 8/8, authority restart 6/6, autonomous authoring 7/7, Agentic Studio runs 45/45, creative autonomy 255/255, Studio Authoring 138/138, Yemaya 19/19, and Lilith override governance 8/8; source commits 85f4f126f0c79a4e811e42813484aa9e780ca29d and 432ef426e4a27e890bb8d3826ae7ce6e502b99ea; closure evidence commit f35f30e6755351477bb26a1b65a1e1d6b192e1c9.

HITL-005 — Define measurable human-authority SLOs#

  • Measure takeover acknowledgment and checkpoint-resume latency.
  • Measure human minutes, comments, edit distance, replacement, regeneration, veto, fork, rollback, and appeal rates.
  • Measure the percentage of human edits incorporated correctly.
  • Detect recurring overrides and require policy review.
  • Require 100% success in takeover, restart, duplicate-delivery, stale-approval, authorization, and rollback drills before raising autonomy.

Closure evidence: evidence/v1-v9/HITL-005.md; canonical durable SLO and gate: apps/oshun/bff/src/agentic/human-authority-slo-store.ts; authority SLO 4/4, autonomy route 8/8, AgentRun lifecycle 7/7, V1 lifecycle HTTP 8/8, autonomy bindings 28/28, bound-plane restart/end-to-end 6/6, AgentRun HTTP 14/14, OpenAPI drift 1/1, and OpenAPI coverage 3/3; source commit 1de8d8948b84bfd16d21ad9475851a2e4d7fa7aa; closure evidence commit 183f670ea269a43ecaa496936b9c223b21ba18b1.

5. P0-E — Invoked tool authority, runtime enforcement, and cost#

AUTH-001 — Implement one portfolio capability broker#

  • Resolve least-privilege grants by agent, actor, tenant, run, purpose, resource, action, trust zone, expiry, and risk tier.
  • Enforce scope specificity, delegation subset, revocation cascades, and approval requirements.
  • Prohibit token passthrough; use audience/resource-bound short-lived credentials from secure storage.
  • Enforce egress allowlists, DNS/IP revalidation, SSRF defenses, secret isolation, filesystem/process/network sandboxes, and content trust zones.
  • Audit grant request, policy decision, approval, use, result, denial, revocation, and attempted bypass immutably.
  • Add direct, confused-deputy, cross-tenant, expired, revoked, encoded, redirect, DNS-rebinding, secret-exfiltration, and sandbox-escape tests.

Closure evidence: evidence/v1-v9/AUTH-001.md; canonical portfolio broker: libs/oshun/agentic-studio/src/grants/portfolio-capability-broker.ts; focused adversarial broker 12/12, complete Agentic Studio 323/323 across 22/22 files, and shared HTTP-client 95/95 across 2/2 files; official Agentic Studio and shared HTTP-client typechecks, dependency-aware build, and package lints passing; source commit f85485f3266720454243e01e312eb0e0c7954937; closure evidence commit cf8f5590b83e68da9e8dde4d9252de66c1412ec2.

AUTH-002 — Bind declared controls into the V1 executed route#

  • Invoke the general grant resolver for every planned tool call.
  • Invoke egress, SSRF, secret-scope, and isolation enforcement before dispatch.
  • Append attempted and executed tool calls, arguments, results, errors, evidence, effects, and resolved grants to the returned/persisted run.
  • Consume the actual budget after every call, including failed, partial, retried, and denied attempts according to policy.
  • Update the cost ledger and stage totals before the next admission check.
  • Recheck kill/pause/cancel and grant revocation at every tool boundary.
  • Prove a multi-step plan cannot reuse the same pre-call balance.

Closure evidence: evidence/v1-v9/AUTH-002.md; authoritative boundary: libs/oshun/agentic-studio/src/runs/governed-orchestrator.ts; V1 executed route: apps/oshun/bff/src/agentic/runs-route.ts; adversarial boundary 7/7, complete Agentic Studio 330/330 across 23/23 files, focused BFF route/lifecycle/catalog/governance 50/50 across 5/5 files, and V1 lifecycle/autonomy 16/16 across 2/2 files; official Agentic Studio and BFF typechecks, production builds, exact changed-source lint, and machine-adjacent drift gates passing; source commit e659629be3beab421e254a6082171410393547b1; closure evidence commit 4bb405a66c8298a2a9fa1522188cd8ba841d1385.

AUTH-003 — Implement the governed V9 agent DAG executor#

  • Implement model invocation, DAG scheduling, dependency resolution, tool execution, grants, budgets, checkpoints, evidence per node, intervention, idempotency, and transactional publication through the common spine.
  • Keep planning-only behavior, if useful, as an explicitly separate mode; it cannot satisfy or replace executable-runtime requirements.
  • Bind agent registry model/tool/scope declarations to actual runtime enforcement and persisted results.
  • Add conformance and adversarial E2E tests proving a real governed run from plan through effects, evidence, human intervention, and completion.

Closure evidence: evidence/v1-v9/AUTH-003.md; authoritative runtime: apps/metis/service/src/metis/services/agent_dag_executor.py; executable common-run adapter: libs/metis/api-client/src/execution-common-run-adapter.ts; complete Metis agent-runtime regression 61/61 across 10/10 files, authoritative executor conformance/adversarial 13/13, planning/execution common adapters 8/8, shared run-protocol 67/67 across 6/6 files, strict MyPy and TypeScript checks, OpenAPI/codegen drift, exact lint/format, and targeted production builds passing; source commit b8107e5841c48b100ceca1f116608d8c39f97ecf; closure evidence commit recorded as ac77031d3c631c1e1e7bf58792ee45edf9854a35.

COST-001 — Complete end-to-end measured usage and monetary accounting#

  • Preserve the existing real cycle-token record but label its current coverage accurately.
  • Instrument planner, generator, research loop, structured-output retry, critic, judge sample, reconciliation, media provider, TTS/ASR, translation, embedding, tool, render, GPU, storage, and egress usage.
  • Attribute usage to tenant, slot, run, step, attempt, artifact, model, provider, region, and quality tier.
  • Convert usage to money using a versioned price table and retain the raw provider usage/receipt for reconciliation.
  • Record estimated versus actual cost and feed measured error back into estimatedCostUnits and routing.
  • Add per-call, per-run, per-cycle, daily, monthly, tenant, provider, and portfolio hard caps with reserved headroom for recovery.
  • Stop, degrade, or require approval before exceeding policy; do not merely report after the fact.
  • Reconcile internal ledgers against provider invoices and alert on drift.

Closure evidence: evidence/v1-v9/COST-001.md; authoritative accounting: libs/oshun/creative-orchestrator/src/usage-accounting.ts; production V1 composition: apps/oshun/bff/src/agentic/autonomy-bindings/usage-accounting.ts; current regressions 524/524 across 40 unique files, both accounting-library typechecks, the BFF zero-error ratchet, package and exact changed-source lint, formatting, and the direct BFF production bundle passing. The refresh also proves that raw narration governance signals remain blocked without exact proof records. Original source commit a34012d3da4750c86d8f7c0346a52f64caf4defa; current persistence, restart, and authority hardening 42dac7a43ac9362213e15f08d0843e991930f3c2, c5aa83ecd11a5dfc7611258cd3777027f1543a13, and e3480d2eb708ebdb9374045908a68efc48ed819d; current test remediation 73d9e3b8aae38781713effad2203304614fca1b4; historical closure evidence identity e9306497762fb729531d250610497d70e79157f9.

COST-002 — Make economics part of quality and release evidence#

  • Report cost per attempted, completed, gate-passing, accepted, released, and outcome-successful artifact.
  • Include retries, failed candidates, judge panels, hidden tool calls, and unreleased media in cost.
  • Add cost/quality/latency Pareto views and prohibit optimizing cost by weakening hard gates.
  • Test cap races, simultaneous workers, provider-reported corrections, currency/price changes, missing usage, and unknown final cost.

Closure evidence: evidence/v1-v9/COST-002.md; authoritative economics module: libs/oshun/creative-autonomy/src/cost-economics.ts; operator evidence route: GET /v1/agentic/autonomy/economics; current regressions 488/488 across 36 unique files, both owning-library typechecks, the BFF zero-error ratchet, zero-error package/exact lint, TypeScript formatting, and the direct BFF production bundle passing. Original source commit fce8641f0487b82bebd8052c7886efef8658bc2b; current proof, human-quality, persistence, restart, and scoped-authority hardening is admitted in the pack; historical closure evidence identity f338a1fdb6ca222674f5fb255cc0b571d5e9136b.

6. P0-F — Prompt trust, safety, grounding, and proof semantics#

PROMPT-001 — Create a shared typed trust-zone renderer#

  • Define origin/trust metadata for user, source, retrieval, canon, world, memory, prior model output, critique, tool result, artifact, and policy fields.
  • Extract the CQJ delimiter, sanitizer, length bounds, and explicit non-instruction semantics into a shared primitive.
  • Escape delimiter-like and encoded content without destroying evidence identity.
  • Keep system policy, tool authority, source scope, rights, and release law outside model-controlled text.
  • Record rendered prompt hash, renderer version, trust labels, truncation, source refs, model/provider version, and raw response under privacy policy.

Closure evidence: evidence/v1-v9/PROMPT-001.md; authoritative shared package: libs/shared/prompt-trust; CQJ reference migration: libs/shared/content-quality-judge/src/judge-engine.ts; regressions 296/296 across 36/36 files, all three targeted typechecks, prompt-trust lint, exact-source lint/format, pnpm lock validation, and post-merge affected checks passing; inherited CQJ package lint baseline documented in evidence; source commit 93d6d841e72aca0be8448bc41d2cc2a31f81d06e; closure evidence commit 4fd02bbb58dee2fb24b2db9d5d1c7fb0ceadb5df.

PROMPT-002 — Migrate every identified model boundary#

  • V1 ideation: fence observed signals and recent-topic material.
  • V1 planner/text generation: fence briefs, dependencies, sources, and constraints; supply source-bound citation IDs.
  • V1 critique/revise: fence artifacts, prior drafts, rubrics, feedback, and quoted evidence.
  • V1 CQJ self-refine/Reflexion: fence prior output and remembered weak phrases; prevent memory instructions from acquiring policy authority.
  • V2 live commentary: fence raw cues while preserving the latency budget.
  • V2 post-match review: fence generated commentary and make empty model findings insufficient to clear deterministic safety checks.
  • V2 Hathor: fence persona, history, world, nearby entities, and user text.
  • V6 cognition/Clio: fence steward input, retrieved context, memories, prior output, event text, and critique.
  • V7 Forge/Mawu: fence brief, realm context, source metadata, and backend output; do not trust static regex as a sandbox.
  • V8 proposer/assets: fence canon/constraints and validate surface fields semantically.
  • V9 content/tutor: fence source bundles, previews, learner state, orchestration fields, prior output, and feedback.

Closure evidence: evidence/v1-v9/PROMPT-002.md; all 11 invoked boundary groups migrated; 296/296 behavior tests across 29/29 files, 14 scoped TypeScript checks, Clio/Isis dependency builds, lint/format, Rust/Python checks, and frozen merged-lock verification passing; source commit fd6ff0b7131df0f5625d80309f294dcf7379dd5b; source publication tip 7da910b22a356e2393bd4fc4be706b9b1a65119f; closure evidence commit bd9c0079efac4d966d994e00e2557f09320880e6.

PROMPT-003 — Build the portfolio prompt-injection evaluation suite#

  • Cover direct, indirect, cross-turn, retrieval-poisoning, source-poisoning, generated-output, critic-feedback, tool-result, multimodal, encoded, delimiter, memory-poisoning, and cross-item attacks.
  • Assert attacks cannot expand tool authority, source scope, rights, autonomy, budget, or release permissions.
  • Assert citations, quotes, and tool arguments are checked against original typed data after generation.
  • Run the suite at every migrated boundary and on every model/provider change.
  • Retain attacks as a held-out regression set separate from prompt examples.

Closure evidence: evidence/v1-v9/PROMPT-003.md; 13 held-out attack families executed at all 20 registered implementations (260 boundary/attack executions); 232/232 TypeScript tests across 5/5 files plus complete Python/Rust corpus runners, seven generated-output integrity negatives, TypeScript/ESLint/Ruff/Cargo/Prettier/frozen-lock checks, and the final non-cacheable daemon-free Nx aggregate target passing; always-on push/PR workflow has no path filter; source commit d698ae1c53c0a4895b29fe7e4eeeabed70fcf07f; source publication tip ee77b83a26ecf1fbeaf6077f78d582d7b722161c; closure evidence commit 16881fc6a603e16d2c19f7fa608e0c722087c2e0.

SAFETY-001 — Bind safety to the invoked product paths#

  • Invoke Hathor's supplied safety configuration/filter in dialogue generation and fail closed when required safety evaluation is unavailable.
  • Derive Hathor confidence from calibrated evidence with an explicit abstention state; remove the literal 0.85 without removing the required confidence capability.
  • Implement real memory retrieval/use attribution and populate memoriesTriggered; never return an unconditional fabricated integration result.
  • Require a successful V9 safety/trust-zone record before tutor generation, agent planning, or tool execution where policy applies.
  • Upgrade V9's literal threat lexicon with policy-defined classifiers and adversarial evaluation while retaining deterministic hard rules.
  • Preserve V8's fail-closed scanner requirement; add real multimodal safety providers and benchmark the offline keyword floor honestly.
  • Bind V7 computed/dynamic capability calls to a real sandbox instead of relying on literal-call regex.

Closure evidence: evidence/v1-v9/SAFETY-001.md; 419/419 unique current assertions passed across Hathor, Metis, Ariadne/Yemaya, Moremi, Forge Assist, and the V6/V7 prompt boundary, plus three explicit Loom skips. Targeted TypeScript, build, core zero-warning ESLint, broader Yemaya safety zero-error/12-warning lint, Prettier, Rust format/strict Clippy/locked Cargo, Ruff, Mypy, and compatible OpenAPI/generated-contract drift gates passed. Original source commits are fee5bba3101af670798333cf71844f7fda779d74, c06014ca70aa4e296689770d634827b281071e16, bbc441f42a04b7518f832a2b4d903a8c7dde680e, and 39c039a5713606982ca1b2f3b4c1e6f783fe82f6; current static-gate remediation and admitted source tip 6299a9fc53213fcd0dd02ebd2a6596ea3021f50a; current evidence refresh c9344f150c9023ee3004471acb14c2e893aef3a2; historical closure evidence 9453041294efd6888274ca640666276879527427.

PROOF-001 — Define non-substitutable proof types#

  • Represent separately: format validity, identifier existence, retrieval, passage match, claim entailment, source approval, rights/license receipt, consent, human approval, byte-bound integrity, authenticated signature, C2PA verification, runtime placement, and observed outcome.
  • Prevent weaker proof types from satisfying stronger gates through schema, type, policy, and test enforcement.
  • Require every release verdict and report to name exactly which proof type passed and link its evidence.
  • Add adversarial fixtures for citation-shaped strings, existing but irrelevant IDs, unsupported passages, stale rights, locator hashes, unkeyed digests, and structural quality proxies.

Closure evidence: evidence/v1-v9/PROOF-001.md; 1,195/1,195 focused and package tests passed across the canonical proof contract, publication/run protocol, shared release gates, Isis, BFF provider flows, content quality, creative autonomy, V3/V8/V9 release gates, integration flows, and downstream Metis/Yemaya/Calliope adapters, with task-scoped typecheck, lint, and build gates passing; source commit 73ee66c5077cd2ea00c7280e530ed0091072ea1b; current remediations 602de4516e8e19af1a7ca10824344c68141e59c3 and 9f82832200aba015187521d47c63d33c657c8894; historical closure evidence commit cfb7cb334b9ebc3ce2ffc03dc2f583d3d21ec686.

PROOF-002 — Close citation and semantic-grounding gaps#

  • Replace V1 Sophia's declared-citation-only default with source retrieval, passage capture, and claim-level entailment.
  • Supply V1 writers with bounded source bundles and stable citation IDs rather than asking the model to invent references.
  • Make V2 live commentary verify factual claims against authoritative match state within the latency/fallback budget.
  • Make V6 Clio verify free-prose assertions against cited event facts, not only citation membership.
  • Make V7 Mawu verify semantic source-to-output grounding, not merely source metadata and attribution.
  • Make V9 citation sufficiency fail when factual output has no supporting citations; do not allow vacuous no-citation success.
  • Supply V9 generation with approved source bundles and require retrieved passage/entailment proof.

Closure evidence: evidence/v1-v9/PROOF-002.md; 1,929/1,929 unique assertions passed across 68 files covering the canonical semantic-grounding contract, Iris/V2, Clio, Mawu, V9 lesson gates, full Metis verification and LLM packages, BFF autonomy flows, prompt-injection boundaries, and creative variant preservation, with task-scoped typecheck, lint, formatting, alignment, and Iris build gates passing. Original source commit 40d160fb7cf8941522a290682a781c80b2f2f9a6; current remediation 7faca4cc67cb27534228053f104f533763ec67d4; current test compatibility and verification tip 734b7bd04105fc3e8c762c9c1451390f3ef0614e; historical closure evidence identity 09c1ea147df55a6eea5083b095eb8541f14743d1.

  • Replace V1's automatic first-party/provider rights satisfaction with artifact-specific license, model, dataset, likeness, voice, and consent receipts.
  • Preserve hashScope and signed through V8 aggregate provenance.
  • Bind V8 media to downloaded asset bytes, not url + seed, before claiming byte integrity.
  • Add an authenticated V8 release-report signature with verified signer identity. A separately named unkeyed digest may remain only as a supplemental checksum, never as a substitute.
  • Generate and cryptographically verify real C2PA claims including creation, actions, ingredients, software agents, transformations, and human edits.
  • Add key identity, rotation, revocation, timestamp, verification-policy, offline-verification, and substituted-byte tests.
  • Require mixed human/agent lineage and invalidation after edits.

Closure evidence: evidence/v1-v9/PROOF-003.md; 236/236 unique executed assertions passed across 30 files, with three Loom skips reported separately, covering canonical artifact rights, V1 autonomy boundaries, real C2PA export, V8 case bundles, Loom, authenticated release verification, aggregate provenance, case contracts/evaluation, Daedalus, and content signing. Task-scoped typechecks, the BFF zero-error ratchet, lint, formatting, and direct C2PA/signing builds passed; the pack explicitly excludes an aggregate verifier build stopped by unrelated @iris/agents-core dependency errors. Original source commit 2968665255d826439aa8ca355aaa397f9b605182; current provenance, quality, and exact-gate hardening 54f183278714144ac35bdbca2e0aac4838a1a496, 2a8aea8f2107616e6bf0a87ac879dd3051db61de, and 4d0707691584e1053a15521764c7efb32a41ec21; current fixture remediation 84d1c2f65cb672a4943d5c8ed8a1cdc68c9d534c; historical closure evidence identity 0b0d101e20f9f1ad19ebfede5e611f3e49ab316f.

MEDIA-001 — Close the current media-evaluation contract gaps#

  • Add poster, thumbnail, or keyframe output to every video producer contract consumed by the media critic.
  • Make missing required frames a typed held/blocked result with actionable evidence rather than an unhandled dead path.
  • Add V8 asset best-of-N and perceptual/period/canon quality evaluation.
  • Preserve losing candidate, seed, provider, prompt, and score evidence.
  • Add image, video, audio, 3D, and gameplay artifact-specific probes and human comparison described in EVAL-005.

Closure evidence: evidence/v1-v9/MEDIA-001.md; 4,351/4,351 non-duplicated package and focused tests passed across canonical media evaluation, V8 case contracts/assets and Loom, creative critique/revision, and V1 BFF media/autonomy boundaries, with all task-scoped typecheck, lint, formatting, alignment, real video-probe, BFF ratchet, and direct BFF build gates passing; source commit f8046ba3a573f082aa24e95573cd30de8a5c31d6; closure evidence commit 573cce758a2d33cb5caaee8c0e824da09a0eb966.

7. P0-G/P1 — Evaluation, gold data, quality, and learning#

EVAL-001 — Implement one unified evaluation record and harness#

  • Define versioned task, trial, transcript, trajectory, artifact, outcome, grader, score, uncertainty, cost, latency, and failure records.
  • Support deterministic code graders, solvers, model graders, provider probes, engine/runtime checks, expert raters, target-user raters, and live outcome graders without conflating them.
  • Separate capability, regression, adversarial, safety, reliability, subgroup, cost, and live-canary suites.
  • Repeat nondeterministic trials and report sample size, seeds, model and provider versions, temperature, pass@k, pass^k, confidence intervals, and failure clusters.
  • Slice results by product, slot, modality, locale, audience, model, provider, prompt, rubric, policy, autonomy tier, and risk tier.
  • Fail loud when required graders, dependencies, source bundles, engines, providers, or human cohorts are absent.

Closure evidence: evidence/v1-v9/EVAL-001.md; 5,876/5,876 unique package assertions passed across 350 files in canonical contracts and Nous core, including the unchanged 23-test schema, statistics, preflight, repeated-trial, slicing, and adversarial subset plus the central-runner adapter. Both production typechecks, zero-warning exact lint, formatting, and uncached targeted builds passed; the pack explicitly excludes unrelated broad contracts and Nous spec typing debt. Original source commit 078e72990b264fb2a48d644026ab9f3e1aa30c75; historical equivalent source identity 2ce7dca60b2e2dfe9a5484884efff1195a47ab45; historical closure evidence identity 1d99982533499f26e878a21ee73d1d09c953baab.

EVAL-002 — Acquire licensed human-gold and preference corpora#

  • Build best-authored positive sets for every content family; do not use average legacy content as the excellence bar.
  • Add known-bad adversarial examples for every hard release gate.
  • Add borderline/high-disagreement items and culturally/linguistically diverse examples.
  • Include real provider, render, cook, in-engine, and placed outputs.
  • Hold out release items never used for prompt/rubric/policy optimization.
  • Version rater instructions, expertise, locale, conflict policy, and adjudication.
  • Record licenses, consent, provenance, retention, access, deletion, and contamination analysis.
  • Acquire the missing V5 authored-case corpus for V8 and the real learner/ teacher cohorts for V9 as P0 data work.

Progress evidence: evidence/v1-v9/EVAL-002.md; the versioned licensed-corpus, readiness, holdout/contamination, rights/consent, rater/cohort, and purpose-split admission boundaries are implemented and passed 4,390/4,390 owning-package tests. The machine-readable acquisition registry deliberately records zero acquired V5 cases and zero acquired V9 cohorts; all checklist items remain open until real contracted/consented human work replaces those external blockers. Repository source commit 238ebf35f23959be140663567563a5401550e67e; progress evidence commit 1d31a0c0cee03406e51ff3b11f451d6127a2a3bf.

EVAL-003 — Calibrate model judges against humans#

  • Use heterogeneous provider/model families where policy calls for a panel; record degraded diversity explicitly.
  • Blind candidate identity and run pairwise comparisons in both orders.
  • Measure position, verbosity, style, self-family, reference leakage, contamination, and judge-pleasing bias.
  • Report human-judge rank correlation, confusion by dimension, calibration error, inter-rater reliability, ICC, disagreement, and subgroup behavior.
  • Route high disagreement, low confidence, and out-of-distribution items to humans.
  • Keep bag-of-words diversity and V8's batch-dependent PCA/G5 scores as diagnostics until calibrated against semantic and human judgments.
  • Replace bag-of-words candidate-diversity admission/selection with a validated semantic measure where it makes a release or search decision; retain lexical diversity only as a labeled feature.
  • Require held-out champion/challenger regression before changing model, rubric, examples, weights, or thresholds.

Progress evidence: evidence/v1-v9/EVAL-003.md; the three repository-enforceable diagnostic/semantic/change-control requirements are complete and passed 4,440/4,440 non-duplicated owning-package tests. The first five calibration items remain open until real licensed EVAL-002 human gold, independent raters, blinded comparisons, bias probes, metrics, and routed human decisions exist. Repository source commit 47438840440d658eeb7f6b4185fcb1ed68f384fb; progress evidence commit 57556761768af611d15633205cc35ba3d9346c81.

EVAL-004 — Establish human-quality launch criteria#

  • Predeclare target audience, best-authored baseline, sampling, exclusions, statistical test, minimum effect, and stopping rule.
  • Require non-inferiority with lower 95% confidence bound above 0.50 before claiming authored parity.
  • Require sustained target-audience win rate above 0.60 and lower bound above 0.55 before using “amazing.”
  • Measure human minutes and edit distance so rubber-stamp approval cannot masquerade as efficiency; target at least 80% acceptance without full regeneration only after quality holds.
  • Report tails and subgroup distributions; prohibit a high mean from hiding catastrophic outliers.
  • Keep hard safety, rights, consent, provenance, accessibility, and runtime gates non-compensatory regardless of preference score.

Progress evidence: evidence/v1-v9/EVAL-004.md. The three repository-enforceable statistical/non-compensatory requirements are complete and passed 4,486/4,486 non-duplicated owning-package tests. The first, fourth, and fifth items remain open until a real preregistered target-audience study supplies qualified participants, a rights-cleared best-authored baseline, measured human editing effort, independent quality holds, tails, subgroups, and zero catastrophic outliers. Test fixtures are not launch evidence. Repository source commit 401b5496b6a00aa1a2f3a2f3d53caef16a60a14a; required BFF compatibility commit a432f8f8915431e1062218e598760f9896a15d34; progress evidence commit 9400e4a8975b3c7419dda2bb5e719b28996cf3b1.

EVAL-005 — Build domain-appropriate experiential batteries#

  • Image: composition, prompt/source alignment, identity, typography, aesthetics, cultural/brand correctness, artifact detection, accessibility, and blinded human preference.
  • Video: temporal continuity, direction, edit rhythm, lip sync, story comprehension, flicker/artifacts, captioning, and viewer preference.
  • Speech/music: intelligibility, WER where applicable, speaker/consent match, timing, loudness, silence/artifacts, musical structure, originality/fingerprint safety, and expert/listener preference.
  • 3D/game: topology, UV/rigging, collision, LOD/Nanite, cook, performance, compatibility, playability, novelty, accessibility, fun, and player comprehension.
  • Education: source faithfulness, misconception repair, near/far transfer, delayed retention, calibration, teacher judgment, learner preference, and controlled efficacy.
  • Treat external benchmarks as diagnostic inputs, never sole release authority.

Closure evidence: evidence/v1-v9/EVAL-005.md; MEDIA-001's published image/video/audio/3D/gameplay batteries and 4,351/4,351 verification are combined with the new exact-artifact education battery, controlled target- learner study contract, V9 G2/G3 enforcement, fail-closed lesson providers, and signed Prometheus provenance. The integrated EVAL-005 source passed 4,133/4,133 non-duplicated owning-package tests plus five typechecks, contracts emitted/Nx builds, lint, formatting, and diff checks. Production remains blocked without real evaluator/study evidence; fixtures are not claimed as live outcomes. MEDIA-001 source/evidence commits f8046ba3a573f082aa24e95573cd30de8a5c31d6 / 573cce758a2d33cb5caaee8c0e824da09a0eb966; EVAL-005 source commit 5ebde0db9f81c8d3010e9ea0a0185d6cc7311f43; closure evidence commit 9a1e240288e9aeb7778c205faf1b172c004191e9.

LEARN-001 — Implement a closed-loop, outcome-grounded improvement system#

  • Collect explicit preference, edits, replacements, vetoes, gate failures, complaints, corrections, task outcomes, and approved contextual engagement signals.
  • Attribute every signal to slot, run, step, artifact, prompt, model, provider, rubric, policy, source, intervention, and cohort.
  • Sample hard, novel, high-impact, high-disagreement, drifted, and failed trajectories for diagnosis.
  • Produce versioned natural-language diagnoses and proposed prompt, rubric, tool, routing, policy, or workflow changes.
  • Evaluate proposals on held-out capability, regression, safety, reliability, cost, latency, and subgroup suites.
  • Require human approval for policy/rubric/autonomy changes.
  • Promote through offline, shadow, canary, and champion/challenger stages; roll back automatically on regression.
  • Invalidate and recertify dependent artifacts after relevant changes.
  • Never optimize raw clicks/retention directly; constrain them by user intent, safety, explicit outcomes, and preference evidence.

Closure evidence: evidence/v1-v9/LEARN-001.md; the canonical learning plane now admits all nine signal families with exact slot/run/step/artifact/prompt/ model/provider/rubric/policy/source/intervention/cohort attribution, samples six hard-case classes deterministically, versions diagnoses/proposals, requires seven non-compensatory held-out and monitored dimensions, enforces signed human approval for policy/rubric/autonomy changes, and permits only the ordered offline → shadow → canary → statistically ready paired champion/challenger → production path. Regression auto-rolls back; production invalidates dependent artifacts until exact-version recertification; clicks/retention remain diagnostic-only. 4,706/4,706 non-duplicated owning-package tests passed, with final exact-source 15/15 focused reruns, three typechecks, two emitted builds, lint, formatting, and diff checks. No live telemetry deployment or promoted fixture is claimed. Source commit b1210f90706db87b94c82b0f1b2f78c6a4d629b2; closure evidence commit 9f88c864fe82dfad635a40557ef3269e7f87a3b5.

LEARN-002 — Add safe reflective prompt/policy optimization#

  • Run GEPA-like or equivalent reflective optimization only in a sandbox.
  • Optimize a predeclared quality/cost/latency/safety Pareto frontier.
  • Prevent the optimizer from editing non-overridable law, grant policy, release policy, production prompts, or rubrics directly.
  • Use held-out and contamination-controlled evaluations.
  • Require human-reviewed diffs, signed promotion, canary monitoring, and automatic rollback.

Closure evidence: evidence/v1-v9/LEARN-002.md; reflective search now runs only against attested isolated sandbox prompt, routing, or workflow drafts, with no network, secret, or production-write authority. Requests predeclare distinct training/held-out suites, contamination proof, and hard quality/cost/latency/ safety bounds; the runtime computes and restore-verifies the actual non-dominated frontier. Law, grant/release policy, production prompts, and rubrics are structurally forbidden targets. Only an exact Pareto candidate with a signed human diff review enters LEARN-001's seven-dimension held-out and ordered offline → shadow → canary → champion/challenger → production lifecycle; every forward stage is human-signed, regression is system-monitor-signed and automatically rolled back, and production requires dependent-artifact recertification. 4,447/4,447 non-duplicated owning-package tests passed, with final exact-source 16/16 focused reruns, three typechecks, two emitted builds, lint, formatting, and diff checks. No live optimizer, deployed sandbox, production promotion, or monitor result is claimed. Source commit 6a0d06f31676a7cb3a0dadfd968540e894769840; closure evidence commit 6bb46db48e4f4394c0679745ec009ac43bda2201.

8. P1 — Observability, privacy, operations, resilience, and release law#

OBS-001 — Standardize cross-version GenAI tracing#

  • Adopt OpenTelemetry GenAI semantic conventions for agent, workflow, generation, retrieval, tool, model, provider, token, and conversation operations.
  • Extend traces with slot, artifact, source, canon, gate, rights, policy, intervention, publication, placement, canary, and invalidation IDs.
  • Correlate V1, V6, V8, V9, product runtime, provider, and engine events through the common run spine.
  • Export real OTLP telemetry from V6 and every production path currently exposing only local/in-process observability.
  • Detect missing spans, broken parentage, duplicate effects, orphan artifacts, and untraceable released content.

Closure evidence: evidence/v1-v9/OBS-001.md; strict metadata-only GenAI contracts and shared helpers now cover agent, workflow, generation, retrieval, tool, model/provider, token, and conversation semantics plus every named Oshun lineage ID. Exact adapter-derived common-run identity prevents V1/V6/V8/V9 product substitution, and the live V6 cognition path joins product runtime, provider generations, and tool-engine effects in one correctly parented trace. Shared production tracing now fails closed without service/collector identity; V6 welfare retains its honest in-memory draft while an explicit command performs an awaited real OTLP/HTTP metric export and returns a bounded receipt. The graph auditor detects all five mandated failure classes. 4,403/4,403 non-duplicated owning-package tests passed, with final post-rebase exact-source 41/41 tests, six passing applicable typechecks, four emitted builds, lint, declaration validation, formatting, and diff checks. No deployed collector or production trace/alert result is claimed. Source commit f2a69ad905f3704ccc83e80b460dd8ee4ee30dad; closure evidence commit 986b2e263538a3b1eb7ec37f1a32641bec4d1a76.

PRIV-001 — Minimize and govern sensitive prompt/run data#

  • Default traces to metadata, hashes, typed references, and bounded redacted previews rather than full prompt/source bodies.
  • Define collection, purpose, encryption, access, tenant isolation, retention, deletion, legal hold, audit, and incident policy for prompts, sources, likeness, voice, learner data, private memories, human feedback, and standing direction.
  • Apply the policy to V6 full-prompt run events, V9 learner/source context, and V1 review/learning snapshots containing artifacts and human feedback.
  • Provide privileged, audited break-glass retrieval only where replay or incident response truly requires bodies.
  • Test redaction bypass, cross-tenant access, export/deletion, backup retention, and incident evidence handling.

Closure evidence: evidence/v1-v9/PRIV-001.md; the canonical policy/controller covers every named sensitive class, purpose, encryption/access/isolation, retention/deletion/hold, audit, incident, subject, backup, and one-time dual-approved break-glass control. V6 run events and V9 learner/source context are metadata-only; V9 caches are tenant-partitioned; and V1's actual durable review/learning body is tenant-bound AES-256-GCM ciphertext with authenticated inventory and fail-closed key configuration. Owning packages passed 4,350/4,350 non-duplicated tests, including final post-rebase exact-source 32/32, all applicable scoped/full-ratchet typechecks, emitted/production builds, strict lint, formatting, and diff checks. No deployed KMS, production vault, backup worker, or production privacy receipt is claimed. Source commit e155aa7f2ef852e550028506b0dba38513821809; closure evidence commit ed8aa03040a645714c6e14affad162d1abb37570.

OPS-001 — Verify the declared and live V1 production topology#

  • Generate a repository topology record from the canonical Hetzner compose, Caddy, GHCR/SSH workflow, data volumes, cycle driver, and backup paths.
  • Probe the authorized deployed environment and record whether production actually matches the declared single-box topology; identify external GPU, storage, DNS, monitoring, and provider dependencies.
  • Enumerate shared failure domains for BFF, director, PostgreSQL, Redis, MinIO, run state, review state, backups, and edge routing.
  • Define RPO/RTO, availability, capacity, maintenance, and recovery targets per state class.
  • Implement a second failure domain for critical runtime and durable state, with automated or operator-tested failover; whole-box recovery remains a required defense-in-depth control, not a substitute for resilience.

Progress evidence: evidence/v1-v9/OPS-001.md and the drift-sensitive evidence/v1-v9/OPS-001/repository-topology.json. The declared single-box record, seven shared failure domains, and nine state-class target profiles are verified by 8/8 adversarial tests plus source hashes and CI. The live probe remains unchecked because no authorized production target or monitoring plane was available, and the checked-in production Caddy site is disabled. The second failure domain remains unchecked because the canonical topology has no replica, promotion, fencing, or traffic-failover implementation; same-host dumps and an optional whole-box backup are not counted as resilience. Source commit f6571b0089edb0fd374c4cf8a780ae9a7f1ea0dc.

OPS-002 — Prove backup, restore, disaster recovery, and reconciliation#

  • Automate encrypted off-box backups for PostgreSQL, object/media stores, run/event/checkpoint state, configuration, and required secrets/keys.
  • Verify checksums, retention, immutability, key recovery, and restore compatibility continuously.
  • Run a clean-host/box-loss restore drill, not only an in-place database restore.
  • Reconcile restored publication, provider effects, queues, approvals, invoices, artifacts, and placement state before resuming autonomy.
  • Prove no already-completed mutating effect repeats after restore.
  • Record measured RPO/RTO and unresolved data loss as release evidence.

Progress evidence: evidence/v1-v9/OPS-002.md. Canonical compose now includes a dedicated encrypted restic worker covering PostgreSQL, Redis, MinIO API exports, file-backed run/snapshot state, configuration/secrets, and deployment inputs; backup credentials are isolated from application containers. Clean-host restore and seven-way restored/external reconciliation are implemented and fenced from autonomy, with 12/12 adversarial tests plus a real container build and compose validation. The remaining boxes stay unchecked because no authorized remote repository, provider retention/object-lock proof, independently recovered key, deployed receipt, real replacement-host drill, live external inventory, or measured production RPO/RTO was available. Source commit f6571b0089edb0fd374c4cf8a780ae9a7f1ea0dc.

OPS-003 — Build the operator metrics and SLO dashboard#

  • Coverage: required/current/missing/stale/blocked by version, surface, modality, locale, source policy; quantity, age tails, dependency backlog, invalidation backlog, and unregistered shipped percentage.
  • Human work: untouched release by autonomy tier; minutes, comments, edit distance, replacements, regenerations, approvals, vetoes, tweaks, takeovers, forks, rollbacks, recurring overrides, appeals, and edit incorporation.
  • Quality: authored/champion pairwise results, hard-gate escapes, pass@k, pass^k, judge calibration/bias/variance/disagreement, novelty/repetition, canon/style drift, and domain outcomes.
  • Operations: completion/placement, stage latency/cost, retries, timeouts, failover, duplicate prevention, restart recovery, trace/provenance completeness, kill/incident/withdrawal/rollback time, canary regression, and recertification backlog.
  • Show distributions and tail percentiles by product/locale/audience, not only aggregate means.
  • Link every dashboard number to queryable source records and freshness.

Progress evidence: evidence/v1-v9/OPS-003.md; implementation commit 2db7b53ba10b1a48eb21b22dca9d70aa0604a7f0. The authenticated Analytics ledger now defines all 54 required metrics, exact provenance/freshness queries, and product/locale/audience distributions and tails. The four metric-family outcome boxes remain unchecked because the repository inventory only supplies partial coverage observations; no authorized live V1–V9 human-work, quality, or operations streams were available, and missing evidence renders unknown rather than zero.

REL-001 — Implement portfolio release gates R0–R8#

  • R0 Coverage: registration, source/autonomy policy, quantity, freshness, and placement.
  • R1 Ground truth: source, canon, solver/simulation, factual input, consistency, freshness, and verification.
  • R2 Artifact completeness: modalities, locales, accessibility, dependencies, assets, cook/render, and runtime placement.
  • R3 Quality: calibrated floors, uncertainty, disagreement, and required human-baseline evidence.
  • R4 Safety/integrity: child/crisis, culture, academic/game fairness, injection, abuse, and policy-specific safety.
  • R5 Rights/provenance: consent, identity/likeness/voice, license, fingerprint, ingredients/actions, watermark, authenticated signature, attribution, and verification.
  • R6 Runtime quality: accessibility, latency, measured cost, performance, compatibility, replay, render/cook, and reliability SLOs.
  • R7 Human authority: current approval/review window, valid override, no stale approval, and available takeover/rollback.
  • R8 Live canary: target outcome, incidents, drift, subgroup behavior, operations, rollback, and recovery.
  • Require typed evidence refs for every verdict; a boolean is insufficient.
  • Require an adversarial fixture that makes each gate fail for its intended reason; a gate unable to fail is itself red.

Progress evidence: evidence/v1-v9/REL-001.md; canonical contract/evaluator commit df265d96fbdd339d87f3bee8c0db2599f2261ea2. The implementation defines all 68 required controls, blocks boolean-only or stale/unlinked/wrong proof, and self-certifies an isolated intended failure for every gate. R0–R8 remain open until real V1–V9 release paths invoke the evaluator with current source-bound evidence and publish deployed release/placement verdicts; tests and fixtures do not prove live portfolio conformance.

REL-002 — Enforce the five-level autonomy ladder per slot and action#

  • Support Manual, Copilot, Approval, Review Window, and Full Auto states.
  • Bind autonomy to content slot, action, risk, tenant, audience, and current evidence rather than merely an agent identity.
  • Require mandatory human/institutional authority for named claims, minors, crisis/health, real-person identity, cloned voice, scholarship authority, high-value spend, legal rights, and irreversible external effects.
  • Lower autonomy automatically on reliability, safety, cost, drift, provider, or human-outcome regression.
  • Raise autonomy only after predeclared repeated-trial and human evidence; local overrides cannot raise it past non-overridable policy.

Progress evidence: evidence/v1-v9/REL-002.md; implementation commit a3211ff29a4503cfa7862cd0c042f6a6fe2da901. The canonical five-state contract, creative runtime, intervention API, BFF, operator DTO, and durable legacy-value normalization are implemented and verified. The V1 creative director now invokes the canonical scoped evaluator at every publication boundary and fails closed unless a real slot/evidence/lowering binding is present. The remaining four boxes stay open until production supplies current source-bound slot/risk/tenant/audience/policy/authority/health evidence and persist automatic regression lowering and repeated-trial promotion decisions.

REL-003 — Enforce continuous recertification#

  • Invalidate affected evidence when source, canon, model, prompt, rubric, rights, consent, policy, provider, code, engine, or placement changes.
  • Schedule selective regeneration/review through the coverage graph.
  • Prevent stale signatures, approvals, benchmark scores, or provider probes from releasing changed artifacts.
  • Track recertification latency/backlog and block when an SLO expires.
  • Run periodic incident, override, drift, rollback, and recovery exercises.

Progress evidence: evidence/v1-v9/REL-003.md; implementation commit ac8606760bd6d9ed222a754611b9b8a944f25ab0. The canonical contract now covers all twelve change roots, graph-propagated prior-evidence invalidation, selective ready-leaf scheduling, independent release-target binding, four exact current proof classes, backlog/SLO metrics, and five exercise cadences. All five boxes remain open until real V1–V9 release paths persist and invoke the lifecycle with current production records and executed operational drills. The 2026-07-27 V9 integration adds the first product consumer: the canonical evaluator can issue an authenticated, short-lived decision only after a passing lifecycle evaluation, and Metis signed publication persists and revalidates that exact candidate-bound admission at both release and learner delivery. Forged, wrong-artifact, future, expired, and post-restart-stale admissions block. This is repository integration evidence, not a production recertification receipt.

9. Product-level P0 acceptance slices and local defect closure#

Every product slice below must use real providers or actual runtime/engine outputs, not contract fixtures. Every slice must include a takeover drill, best-authored baseline, signed lineage, placement, telemetry, canary, and rollback. Product-specific gates remain in force in addition to REL-001.

V1-001 — Expand V1 from four scheduled categories to accountable coverage#

  • Map all 28 documented generation surfaces to actual required content slots, owners, modalities, locales, quantities, freshness, accessibility, source/autonomy policy, and placement.
  • Identify surfaces that are tools/access paths rather than content requirements and link them without counting them as filled artifacts.
  • Register every remaining V1 launch content requirement and blocker.
  • Schedule missing/stale work through the canonical director.
  • Prove zero unknown shipped V1 content at the selected launch scope.

Progress evidence: evidence/v1-v9/V1-001.md; implementation is the task-scoped commit containing that evidence. All 28 frozen generation surfaces now have explicit empty content-slot mappings because they are tool/access capabilities, not artifacts. The authoritative selected-scope manifest registers the four live autonomy categories as complete canonical slots, records 12 owned source/rights/provider blockers, and produces a canonical director plan with 12 ready prerequisite jobs and four deferred generation jobs. The first four boxes are closed for the selected scope. The zero-unknown shipped-content box remains open because no independently authorized complete runtime observation exists.

V1-002 — Close V1 runtime and publication defects#

  • Complete AUTH-002, COST-001, HITL-001 through HITL-004, RUN-003, and RUN-005 on the actual V1 path.
  • Make every operationally required store durable and recoverable; keep only recomputable caches ephemeral, with tests proving their loss cannot drop authority, evidence, lineage, approvals, effects, or learning state.
  • Guarantee publication idempotency and approval freshness across restart.
  • Provide video frame outputs to media critique.
  • Replace declared-citation and automatic-rights proxies through PROOF-002 and PROOF-003.
  • Apply privacy/retention policy to V1 review, learning, full artifacts, feedback, and standing direction.

Progress evidence: evidence/v1-v9/V1-002.md; restart-safe publication and approval implementation commit 7e519c01c2cadfa34c7459bd7f44781b355fbbf8. All six boxes close. The autonomy plane now acknowledges ordered durable writes, rolls back a latest failed mutation, binds approval to exact plan/source/ artifact/policy/gate hashes, rejects stale evidence, and reconciles duplicate publication, takedown, and rollback delivery at one stable placement across restart. The all-store durability box is closed by the complete 1,937-root operational inventory and retained-ephemeral loss-proof gate in 80873ce591e73471090129de651a049605bab858 and 7e7cd0662d2891018d02e6a19ce736fe58f18a1f.

The disconnected Isis reference-bank admin authority is retired through 3ec98e462988da49aa6d223d27c55204c8443559. Repository-wide tracing disproved the store's claim that other BFF code consumed its underlying workflow bank: only the admin route and tests imported the singleton, while planning, keyframe, canon, lecture, generation, and publication paths received no binding. Production no longer presents a replica-local, capped, restart-volatile map as canonical identity/style/voice/motif state. The in-memory workflow adapter is available only through explicit test/development injection; a deployment must bind an awaited owning authority, otherwise all four valid operations fail closed with 503 reference_bank_authority_not_configured after authentication, scope, path, query, and schema validation. Focused tests passed 17/17 before and after remote integration, scoped static checks passed, the BFF type ratchet added no errors beyond the five current-main diagnostics, and a rebuilt production artifact passed all 32 migrations plus PostgreSQL/Redis startup smoke after rebase. This does not claim a production reference bank; a future binding must be the durable/shared authority actually consumed by generation and publication. The broader all-store box remains open.

The disconnected Isis scene/production/lecture ledger authority is retired through b6a090c7922f38e9db307e6ec653f89d28c53cea. Repository-wide tracing found all 17 admin operations—including model-backed lecture authoring and tweak routing—terminated in one capped, restart-volatile singleton that no planner, generator, renderer, assembler, publisher, web surface, or other BFF module consumed. Production now requires one awaited deploy-bound ledger authority across all three document families. Otherwise valid reads, writes, digests, deletion, authoring, and tweak routes return 503 ledger_authority_not_configured after auth and input validation; the lecture author checks this boundary before invoking its model, so it cannot spend a provider call and then acknowledge only process-local storage. The real workflow adapter remains explicitly injectable for tests/development. Focused tests passed 46/46 before and after remote integration, scoped static checks passed, the type ratchet added no errors beyond the five current-main diagnostics, and a rebuilt artifact passed all 32 migrations and PostgreSQL/Redis startup smoke after rebase. This does not claim a production ledger plane; a future binding must be the canonical durable/shared authority actually connected to planning, generation, assembly, and publication. The broader all-store box remains open.

The orphan Year in Practice activity-summary state is now code-enforced as an ephemeral/recomputable disposition through 8ed2347d86e9a352bdd5d87ca6a885f5a06d7575. Repository-wide tracing found no production mutation producer or first-party web/mobile consumer; only the route test seeded the former per-member snapshot map. Production no longer hydrates or serves that orphan snapshot key. It uses a frozen stateless source whose exact empty projection drives the canonical retrospective builder's quiet-period artifact, while rich synthetic quarters require an explicitly injected fixture store. A deploy-bound canonical activity projector can be injected later without restoring a shadow write plane. Focused route tests passed 5/5 before and after remote integration, including default-empty, member-isolation, exact-rich-retrospective, auth, and period-validation proof; scoped static checks and the type ratchet passed within the five unrelated current-main diagnostics, and a rebuilt artifact passed all 32 migrations plus PostgreSQL/Redis startup smoke after rebase. No activity authority, evidence, approval, effect, lineage, or learning state can be lost because production owns none in this seam. The live cross-domain activity projector remains future work, and the broader all-store box remains open.

The orphan Nyx Sky Tonight snapshot is now retired through 72e97fd5fdf8f14b3c1b0214e5d259e9424f1ca4. Repository-wide tracing found that only its route tests populated member follows, saved sites, forecasts, quiet hours, or push history; production had no scheduler, event consumer, notification bridge, or owning-service adapter, and no first-party web/mobile consumer called the decision route. Production no longer hydrates or writes the fixture-owned nyx-sky-tonight snapshot. The route now accepts one atomic, async-capable canonical source and, after auth plus complete query/instant validation, fails 503 sky_tonight_source_not_configured when no deploy binding exists. It does not fabricate event_not_followed, site_not_saved, forecast-quality, quiet-hour, deduplication, or nightly-cap knowledge. The mutable store remains an explicitly constructed test/development adapter, and all rich decisions still flow through the canonical Nyx domain library. Focused decision and app-composition tests passed 16/16; scoped static checks and the type ratchet passed within the same five unrelated current-main diagnostics; and the rebuilt artifact passed syntax validation, all 32 migrations, and PostgreSQL/Redis production startup smoke after remote integration. No authority, evidence, approval, acknowledged effect, lineage, or learning state can be lost because production owns none through this seam. A real cross-plane source and acknowledged push-history writer remain future work, and the broader all-store box remains open.

The orphan privacy-receipts snapshot is now retired through 24e010e7632b38d49c5188452b428f32aa5db31b. Repository-wide tracing found that only the route test recorded consent transitions, DSAR lifecycle rows, deletion executions, export bundles, operator accesses, retention purges, current consents, or audit-coverage attestations into the former map; no production privacy workflow or first-party web/mobile/admin consumer used it. Production no longer hydrates or writes the fixture-owned privacy-records snapshot. The authenticated route now accepts one atomic, async-capable canonical projection and, after limit plus cursor validation, fails 503 privacy_receipts_source_not_configured when unbound. It does not fabricate an empty receipt history, unknown/complete audit coverage, or standing assurances. The mutable store remains an explicitly constructed test/development adapter, while receipt projection, stable cursoring, and assurance logic remain owned by @oshun/privacy. Focused route and app-composition tests passed 10/10; scoped static checks and the type ratchet passed within the same five unrelated current-main diagnostics; and the rebuilt artifact passed syntax validation, all 32 migrations, and PostgreSQL/Redis production startup smoke. No canonical consent, DSAR, erasure, export, access-audit, retention, authority, or evidence state can be lost because production owns none through this shadow seam. A real cross-plane receipt projector remains future work, and the broader all-store box remains open.

The follow/premiere read-through boundary is now fail-closed through 978fb8b1fe98ee72ac3ebfa5ea33f46112143730. Blocklists, notification preferences, and weekly delivery counts are removed from the durable follow authority and live only in an explicitly injected synthetic fixture for tests. Production must bind one async-capable trust/notification source: follow creation fails 503 follow_read_through_source_not_configured rather than assuming no block, and premiere fan-out fails the same way rather than assuming no quiet hours or prior deliveries. Fan-out also rejects missing, duplicate, or unexpected per-follower projections with 503 follow_read_through_source_incomplete. The source does not weaken the already durable, restart-safe, DSAR/erasure-fenced follow, premiere, allocation, and lifecycle state. Relevant tests passed 61/61, focused tests passed 18/18, the real-PostgreSQL restart/erasure integration passed 1/1, scoped static checks and the type ratchet passed within the five unrelated current-main diagnostics, and a rebuilt artifact passed all 32 migrations plus PostgreSQL/Redis production startup smoke before and after remote integration. Canonical block, preference, and delivery projectors remain future work, and the broader all-store box remains open.

The Studio workspace-telemetry shadow buffer is retired through 81969832b2d2421779e501172d6a0159610c701d. The former bespoke ingest route acknowledged writes into one capped process-local array, silently swallowed forwarder failures, and exposed no production collector binding; only tests read its snapshot. Current web producers now emit canonical analytics envelopes through the shared authenticated /v1/analytics/events transport. The legacy payload route remains as a compatibility adapter, but it attributes the verified subject and returns success only after the already restart-safe, seven-day, DSAR/erasure-connected analytics store commits; failed persistence returns 503 analytics_persistence_unavailable without publishing readable state. The global buffer and mutable forwarder are removed. Hathor's documented but missing workspace mount/unmount emission is now real, and a no-retry Chromium journey proves it reaches only the canonical endpoint. Focused BFF, web, and transport tests passed 47/47 before and after remote integration, scoped static checks passed, the web ratchet is clean, the BFF ratchet reproduces only the same five unrelated current-main diagnostics, and a rebuilt artifact passed syntax validation, all 32 migrations, and PostgreSQL/Redis startup smoke twice. The canonical analytics store's existing restart, failed-write, race, exact-subject erasure, and bounded-retention proofs now cover this producer. The broader all-store box remains open.

Telegram authentication replay and webhook-rate state are now shared and fail-closed through 58f74752b0a6dd0b4d602b3ef92211dcca88739a. The former module-global replay cache and webhook limiter reset on restart and split their security decisions across replicas. Production now uses atomic Redis replay claims and one shared fixed-window counter; raw Telegram query/user/chat identifiers are hashed out of Redis key names. Missing Redis, Redis failures, malformed counter state, or invalid TTL state return 503 telegram_security_state_unavailable before a session, link mutation, or bot effect is acknowledged. Explicit in-memory adapters remain for development and tests. Focused BFF and messaging tests passed 39/39, the messaging library typecheck and scoped static checks passed, the BFF ratchet stayed within the same five unrelated current-main diagnostics, and the exact production artifact passed syntax validation, all 32 migrations, and PostgreSQL/Redis startup smoke. The disposable catalog was removed. This closes the two Telegram seams found by the expanded state scan, not the parent all-store inventory.

The weekly-digest input seam is now explicitly dispositioned through 28e6582b8da2d424e82a6c14adac59676b2688b7. Repository-wide producer tracing found that only tests populated the former global map. Production now uses a frozen stateless reader whose exact empty projection yields the composer's do-not-send verdict, while the only mutable implementation is an explicitly injected, test-named fixture. Focused loss proof shows clearing that fixture can only suppress a synthetic preview; it cannot drop source authority, evidence, lineage, approval, acknowledged effects, delivery state, or learning state. This is not a claim that the live product assembles or sends a weekly digest; the future read-through projector and delivery authority remain separate work. The broader all-store box remains open.

The tenant Living Scenes governance seam is now accurately dispositioned through ea3169fe8a0de109d11047548f14c4ac2cc4b65f. The former process-local map had no mutation producer: it mixed one deploy-time policy with two sample audit rows and the admin rendered those fixtures beside controls that did not write. Production now uses a stateless, version-controlled policy catalog with explicit read-only provenance, while tenant-scoped audit rows come only from the existing durable Living Scene share authority. The strict admin loader refuses ambiguous provenance or partial evidence; the panel removes fake edit/revoke affordances and renders an explicit zero-event state. Focused BFF tests passed 32/32, admin tests passed 10/10, real PostgreSQL restart recovery passed 1/1, the focused Playwright/axe flow passed 1/1, both type gates and production builds passed, and the migrated production artifact smoke gate passed. Operator-editable policy, operator revoke authority, and the durable snapshot's per-process queue remain outside this claim. The broader all-store box remains open.

The duplicate legacy takedown/repeat-infringer authority is now retired through add4cec8f42e1bd12193f9084d58ba20bc4d74de. Repository-wide tracing found no first-party consumer beyond its own tests, while its nine registered endpoints could still acknowledge process-local legal cases, uploader counter-notices, simulated removals/notifications, arbitrary strikes, and simulated account actions. The two volatile stores are deleted. The legacy route shapes remain authenticated and scope-gated but now return explicit 410 Gone migration tombstones pointing to the existing durable dmcaCaseStore authority; no payload can create shadow legal state or claim an effect. Focused compatibility and canonical-store tests passed 9/9, the post-rebase route proof passed 3/3, the zero-backlog BFF type ratchet, strict static gates, final production build, and fresh-15-migration artifact smoke passed. Generic non-DMCA legacy payloads now have no claimed replacement, and provider removal/delivery remain bounded by the earlier canonical DMCA evidence. This closes two shadow stores, not the broader all-store box.

The on-call schedule seam is now explicitly dispositioned through 26a84182fa3072dae0995672fd3ffe4843308011. Repository-wide tracing found that only the admin route test called the former global store's rotation mutator; production derived its team list from version-controlled service and queue ownership manifests and always reported rotations unseeded. Production now uses a stateless manifest catalog with no mutation authority, detached results, and the same truthful rotation_targets_not_seeded blocker. Synthetic rotation scenarios live only in an injected test fixture. Focused route/catalog tests passed 13/13 before and after the non-overlapping remote integration, the BFF type backlog remained zero, strict static gates and the final production build passed, and a disposable database boot applied all 15 migrations before the artifact smoke passed and cleaned up. This does not claim a live responder or provider roster; that requires a separately governed durable or read-through authority. No user or privacy partition was made durable because loss of this static catalog cannot discard operational authority. The broader all-store box remains open.

The connected contextual-help registry is now explicitly dispositioned through cb247571dd183cb6cead4a4a265fc032a4e9413e. Its customer web component consumes the authenticated route, but repository-wide tracing found no production caller for any of the former singleton's four registration methods; only test cleanup called reset, and all help resources were hard-coded constructor definitions. Production now uses a stateless version-controlled catalog with no mutation authority and fresh nested values per resolution, so caller mutation or process replacement cannot alter the next response. Focused catalog/route tests passed 8/8 before and after the non-overlapping remote integration, the BFF type backlog remained zero, strict static gates and an identical production build passed, and disposable fresh-15-migration artifact smoke passed and cleaned up. This does not claim a live help CMS, publishing workflow, or remote video availability; those need a separately governed content authority. No user or privacy partition was made durable because deployed source recreates the exact definitions. The broader all-store box remains open.

The three scene metadata, recap-event, and audience-view projection maps are now explicitly dispositioned through 950617ce2dca9840238b38a594768b520871bd55. Repository-wide tracing found their only writers in the audience, recap, reaction, and tip route tests; no production share, session, analytics, service, web, mobile, or admin path could populate them, so the connected production routes already returned their missing-projection 404/empty behavior. Production now uses stateless readers with no seed/reset authority, while mutable projection fixtures are test-only and explicitly injected. The five focused suites passed 31/31 before and after the non-overlapping remote integration, the BFF type backlog remained zero, strict static gates and an identical production build passed, and disposable fresh-15-migration artifact smoke passed and cleaned up. This does not claim a live scene projection authority: any future connection must be durable or authoritative read-through and carry ownership, DSAR/erasure, retention, and audit controls. Existing durable appreciation and tip authorities are unchanged; their upstream scene-governance projection remains unavailable. The broader all-store box remains open.

The connected account-protection screening partition is now durable through 93484dede75567663f6606f3aaca372af9b7f2c3. Successful-login and authorized- admin ATO/credential-stuffing verdicts commit as normalized PostgreSQL rows; session-id uniqueness is enforced across replicas, production verifies the tables before traffic, and no production seed data remains. Raw device and country inputs are analysis-only and never enter persistence. Exact-subject session DSAR now includes the committed screens, while signed session erasure deletes directly identifying rows and leaves a digest-only resurrection fence. Startup and seven-day execution enforce the canonical 90-day session window while honoring active legal holds and category retention exceptions. Fresh- database migration, restart, replica duplicate, insert-versus-erasure, adjacent- subject, retention, type-ratchet, build, and production-artifact proofs are green. Login awaits the attempt but deliberately remains fail-open on analyzer or database failure; this slice does not claim step-up, denial, suspension, or revocation enforcement. The broader all-store box remains open.

The account-protection administrator and browser boundary was subsequently hardened through cafc735249c591f3cb355da7cf88f28879b47e44. Both read and write responses now carry the exact durable_account_protection_screening_history disposition and bind the normalized-PostgreSQL, authenticated-subject/admin-scope, 90-day hold-aware, request-only device/country, advisory-only, fail-open-login, fail-closed-admin, and signed-erasure/resurrection-fence semantics. Strict request and browser parsers reject unknown fields, inflated authority claims, malformed dashboard aggregates, and acknowledgements that are not bound to the submitted session, subject, and event. The rendered workspace identifies the ledger as advisory screening rather than enforcement and distinguishes its single live route from four documented-only paths. Focused BFF route tests passed 9/9, web unit/integration/route tests passed 8/8, and live desktop/mobile Playwright with applicable axe analysis passed 8/8 against the source BFF and a disposable PostgreSQL database with all 32 current migrations. The real restart/race/ erasure/retention durability test passed 1/1 after isolating its fixture rows; store and login-producer tests passed 13/13. Web typecheck and affected lint passed, while the 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. Production-artifact smoke passed against the migrated database; the final 36,097,671-byte BFF entry bundle has SHA-256 08e5db98d4ac3a3a1d570d8f6cbe48fdb579d228db2dd9a5862f00b50249bb98. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This follow-up does not add step-up, denial, suspension, revocation, or provider-side enforcement.

The connected Isis artifact-detection review-package partition is now durable through bce1e432032d95a8570ea989505b88db2674c53a. The former process-local map and four synthetic production reports are replaced by normalized PostgreSQL report and digest-only operator-lifecycle rows. Queue-item uniqueness closes cross-replica duplicates; the lifecycle row serializes insert against account erasure and permanently fences recreation. createdBy now comes only from the authenticated administrator, and a successful response is emitted only after the insert plus its returned dashboard projection complete in one transaction. Asset, provider, and model values remain bounded opaque operational identifiers, not inferred customer ownership.

Signed profile erasure pseudonymizes the exact operator while retaining minimized internal quality evidence; the canonical seven-year review_package executor then physically purges due rows unless an active legal hold or exact-category retention exception blocks deletion. Production starts empty and fails closed when the durable repository is unavailable. Fresh-17-migration, restart, cross-replica duplicate, insert-versus-erasure, acknowledgement rollback, adjacent-subject, retention, type-ratchet, production-artifact, and rendered Playwright/axe proofs passed. The browser flow also exposed and closed sub-44px form controls and now renders specific deleted-account and storage-unavailable guidance. The web production build compiled and completed TypeScript, then hit the unrelated current-main /profile/persona undefined.join prerender defect; that page is outside this slice. This closes the artifact-detection operational partition, not provider quality enforcement or the broader all-store box.

The connected Isis 3D-benchmarking review-package partition is now durable through b4623dfe0e8ed9a6e9e776260f054cf7e602a4a0. The former seeded process-local map is replaced by normalized PostgreSQL benchmark-case and digest-only operator-lifecycle rows, and production starts with honest empty history. Case-id uniqueness closes cross-replica duplicates; the lifecycle row serializes evaluation against exact operator erasure and permanently fences recreation. The administrator route derives evaluatedBy only from validated auth, rejects caller attribution, and acknowledges a result only after the case and returned dashboard projection succeed in one transaction. Both vectors are bounded, finite, equal-dimension measurements; persisted metrics are recalculated on read and the database independently constrains vector shape, metric ranges, and verdict consistency.

Signed profile erasure pseudonymizes the exact evaluator while retaining the minimized, non-exportable internal quality record. The canonical seven-year review_package executor later purges due cases unless an exact legal hold or category retention exception applies, while retaining the digest fence. Production fails closed on missing durability. Fresh-18-migration, restart, cross-replica duplicate, insert-versus-erasure, acknowledgement rollback, adjacent-subject, retention, type-ratchet, build, production-artifact, and live Chromium/axe proofs passed. The browser now reports honest empty, deleted- operator, invalid-vector, duplicate, and storage-unavailable states; controls meet the 44px target. The web production build compiled and typechecked, then hit the unrelated current-main /profile/persona undefined.join prerender defect. This closes the connected benchmarking partition, not benchmark-corpus governance, provider/model release enforcement, or the broader all-store box.

The connected Isis feedback-evaluation review-package partition is now durable through 6998eb4883bd8437c3bf5b414b7ee0fbb7551552. The former seeded process-local map is replaced by normalized PostgreSQL sample and digest-only operator-lifecycle rows, and production starts with honest empty history. Sample-id uniqueness closes cross-replica duplicates; the lifecycle row serializes evaluation against exact operator erasure and permanently fences recreation. The administrator route derives evaluator attribution only from validated auth, rejects caller attribution, and acknowledges a score only after the sample and returned aggregate projection succeed in one transaction. Predicted/reference text, ids, n-gram settings, timestamps, and metrics are bounded; persisted BLEU and ROUGE projections are recalculated on read.

Signed profile erasure pseudonymizes the evaluator while retaining minimized, non-exportable internal quality evidence. The canonical seven-year review_package executor later purges due samples unless an exact legal hold or category retention exception applies, while retaining the digest fence. Production fails closed on missing durability and contains no customer feedback or invented customer ownership: the UI permits only approved, de-identified internal review text and states that the diagnostics are not automatic release authority. Fresh-19-migration, restart, cross-replica duplicate, insert-versus-erasure, acknowledgement rollback, adjacent-subject, retention, type-ratchet, production-artifact, and live Chromium/axe proofs passed. The web production build compiled and typechecked, then hit the unrelated current-main /profile/persona undefined.join prerender defect. This closes the connected feedback-evaluation partition, not customer-feedback ingestion, model release governance, or the broader all-store box.

The connected Isis model-quality assessment review-package partition is now durable through d1bbb3ca0f14fa8d913239dd4f86887cea463b28. The former seeded process-local map is replaced by normalized PostgreSQL assessment and digest- only operator-lifecycle rows, and production starts with honest empty history. Assessment-id uniqueness closes cross-replica duplicates; the lifecycle row serializes insertion against exact operator erasure and permanently fences recreation. The administrator route derives assessor attribution only from validated auth, rejects caller authorship, and acknowledges a result only after the assessment transaction commits. Strict schemas bound every nested delivery signal and persisted projection; reads recalculate the versioned evaluator result and fail closed on corruption or drift.

Signed profile erasure pseudonymizes the assessor while retaining minimized, non-exportable internal quality evidence. The canonical seven-year review_package executor later purges due signals and projections unless an exact legal hold or category retention exception applies, while retaining the digest fence. Production fails closed on missing durability. Fresh-20-migration, restart, cross-replica duplicate, insert-versus-erasure, acknowledgement rollback, adjacent-subject, corruption, retention, type-ratchet, production- artifact, and live Chromium/axe proofs passed. The UI states that assessments are diagnostic only, do not authorize release, and must not contain customer ids or personal information. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no user approval was provided; the focused web typecheck and rendered browser journey passed. This closes the connected model-quality assessment partition, not provider/model release enforcement, remediation, or the broader all-store box.

The connected Isis texture-quality diagnostic review-package partition is now durable through b0728341aa7a065968e9f90eaf9d040501bb4732. The former seeded process-local map is replaced by normalized PostgreSQL batch, ordered-fixture, and digest-only operator-lifecycle rows, and production starts with honest empty history. Batch-id uniqueness closes cross-replica duplicates; the lifecycle row serializes evaluation, erasure, and concurrent purge while permanently fencing recreation. The administrator route derives evaluator attribution only from validated auth, rejects caller authorship, and acknowledges a result only after the batch plus its returned fixture projection commit. Strict schemas bound canonical ids, fixture count/order/uniqueness, scores, thresholds, timestamps, evaluator revision, and response shape. Every read reruns the versioned worst-fixture scorer and rejects projection or evaluator-digest drift.

Signed profile erasure pseudonymizes the evaluator while retaining minimized, non-exportable internal quality evidence. The canonical seven-year review_package executor later purges due batches and cascading fixture rows unless an exact legal hold or category retention exception applies, while retaining the digest fence. Production fails closed on missing durability. A fresh database applied all 22 migrations; focused schema tests passed 2/2, BFF store/route tests 12/12, web component/page tests 10/10, and real-PostgreSQL restart/replica/erasure/retention/corruption/rollback tests 2/2. Current-source desktop Chromium/axe passed 4/4, and an exact-production-bundle Pixel 7/axe journey passed 1/1 with 44px controls and viewport containment. Both app type ratchets remained at zero backlog; Prisma validation, persistence typecheck, strict lint/format, the 33.5 MiB BFF production bundle, and its production- artifact smoke gate passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no user approval was provided. The UI labels this as diagnostic-only internal evidence, prohibits customer ids/source assets/personal information, and does not claim release authority. This closes the connected texture-quality partition, not upstream asset scoring, provider/model release enforcement, or the broader all-store box.

The connected Isis content-safety diagnostic review-package partition is now durable through 181979d868d606715a4d71179f16e1df827efad2. The former seeded process-local map is replaced by normalized PostgreSQL screen, ordered-message, ordered-issue, and digest-only operator-lifecycle rows, and production starts with honest empty history. Screen-id uniqueness closes cross-replica duplicates; the lifecycle row serializes evaluation, exact operator erasure, and concurrent purge while permanently fencing recreation. The administrator route derives evaluator attribution only from validated auth, rejects caller authorship and unknown customer/source fields, and acknowledges a result only after the complete diagnostic transaction commits. Strict schemas bound ids, message count/order/content, aggregate content size, issue vocabulary/order, timestamps, scores, evaluator revision, and exact response shape. Every read reruns the versioned evaluator from the ordered source messages and rejects projection, issue, revision, or evaluator-digest drift.

Signed profile erasure pseudonymizes every row owned by the evaluator digest, including a row whose display attribution was corrupted, while retaining minimized, non-exportable internal review evidence. The canonical seven-year review_package executor later purges due screens and cascading message/issue rows unless an exact legal hold or category retention exception applies, while retaining the digest fence. Production fails closed on missing durability. A fresh database applied all 24 migrations; focused schema tests passed 2/2, BFF store/route tests 12/12, web component/page tests 10/10, and real-PostgreSQL restart/replica/erasure/retention/corruption/rollback tests 2/2. Current-source desktop and Pixel 7 Chromium/axe coverage passed 5/5 with 44px controls and viewport containment. Both app type ratchets remained at zero backlog; Prisma validation, persistence typecheck, strict lint/format, the 33.5 MiB BFF production bundle, and its production-artifact smoke gate passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no user approval was provided. The UI limits entry to approved de-identified internal fixtures, prohibits customer content, customer ids, source assets, and personal information, and explicitly disclaims live moderation and release authority. This closes the connected content-safety diagnostic partition, not runtime moderation, provider/model release enforcement, or the broader all-store box.

Subsequent all-store hardening through 05856d6c76007de3a08d1c0a349ce2f8f547d7f3 (published through e6d37a508a0ef4ab72df6798f00fa790a6f507a0) made AgentRun/human-authority state, generation claims and output effects, usage reservations/receipts, budget/quota admission, kill-switch authority, operator audit evidence, governed Studio state, the central per-operator admin workspace, and the customer deletion schedule/receipt ledger restart safe and acknowledgement bound. Authenticated analytics events now also hydrate from a required durable snapshot, acknowledge only committed writes, derive ownership from validated BFF auth rather than caller context, and hard-delete exact-owner rows through the canonical signed analytics_event service. The buffered web transport resolves the current token per attempt instead of capturing stale identity. Uploaded profile avatars now also acknowledge only ordered durable snapshots. The avatar_pack service no longer signs a false shared-catalog absence: it deletes exact-subject inline bytes/metadata and every tracked current or superseded S3/MinIO object version, persists a resurrection-preventing subject fence, and fails closed on missing or failed byte deletion. Upload and erasure share a per-subject lifecycle barrier. Canonical Arete journal rows and both Tara ritual-completion partitions now hard-delete by exact userId; the ritual service performs its session and completion-event deletes in one database transaction. Both erasers are omitted when the V1 database is unconfigured, so the canonical fanout fails closed rather than signing a false absence receipt. The shared admin barrier covers all 187 core mutation routes, rolls failed writes back, fails closed without a sink, and rejects same-operator races. Opaque Studio drafts now hard-delete through the canonical continuity_state service with an atomic durable prefix erase and a signed, fail-closed fanout receipt. The existing signed generated_artifact service now also erases authenticated subject-owned durable job requests/results, catalog/provenance/lineage, and gallery records alongside Isis metadata rows. Per-subject lifecycle serialization prevents stale dispatch after erasure and waits for an already-running provider before deleting its resulting control-plane copies. The deletion lifecycle persists each schedule before external erasure and each signed terminal receipt before acknowledgement, survives restart with ID continuity, rolls failed writes back, and prevents duplicate or cancel-resurrecting runners. It now creates jobs for all 16 canonical deletable artifacts. Owner-scoped library collections now acknowledge required file snapshots or an atomic Postgres dual-write, and the canonical collection eraser deletes the exact subject from both the stub and canonical representations behind a durable resurrection fence. A configured but unavailable database fails closed, and live isolated-Postgres race proof shows create-versus-delete cannot leave subject rows. The real saved-library store now acknowledges save, unsave, and cross-device sync only after an ordered durable snapshot, rolls failed writes back, and fails closed without its deploy-bound sink. The canonical saved_claim_passage eraser removes the exact subject's full LWW set and tombstones, persists a restart-safe fence, and rejects stale device resurrection under either save/delete race ordering. Member-authored Nisaba annotations and notebooks now acknowledge only ordered durable snapshots, roll failed writes back, rebuild annotation-derived highlight state on restart, and exclude the shared corpus/manuscript seeds from customer snapshots. The canonical notebook eraser removes the exact study workspace behind a durable fence, preserves adjacent members, and refuses stale recreation under either create/delete race ordering. The original request-scoped evidence_pack structural-absence disposition has now been superseded for the BFF by the real owner-keyed Metis academic-record partition described below. Its exact-subject eraser and DSAR collector replace the absence receipt whenever that partition is injected. The separate canonical V1 assessment evidence and verdict tables are now queried by the exact-learner DSAR adapter described below when OSHUN_V1_DATABASE_URL is configured; an unconfigured deployment retains the honest gaps rather than claiming duplicate coverage. Consumer profile/preferences, customer auth accounts, refresh-token families, recovery and verification state, and the device-session registry now acknowledge ordered durable snapshots, fail closed when their required sink is missing, and restore only committed candidates. Passwords and recovery codes are persisted only as salted scrypt verifiers; refresh and email-verification secrets are persisted only as SHA-256 lookup hashes. Exact-subject profile and session erasers remove the owning records, tokens, device sessions, step-up state, and audit rows behind separate durable resurrection fences. The real export collectors compose the profile/auth and customer/device-session views without inventing defaults or returning raw secrets. All current canonical deletion artifacts are therefore mapped, and the repository full-account lifecycle completes when every injected eraser attests. Generated-artifact stored-object hardening through c41b4661da7f29f51628d957f106f128811efc53 (published through c974990c8104a7d5bd3aa0230223bd2d38542047) now purges every version and delete marker for the exact configured S3/MinIO bucket/key retained by each Isis GeneratedOutput before deleting its metadata. Unknown buckets, region mismatches, missing adapters, provider errors, pagination defects, and a version that appears during verification fail the signed receipt while preserving retry metadata. The same all-version primitive closes historical-byte retention for profile-avatar objects, and a live versioned-MinIO test proves multi-version and delete-marker removal, idempotent retry, and adjacent-key isolation.

Reminder-continuity hardening through aab9c2634210178443a54d2d8212688171a1cc10 (currently published through af7fffdb9a6fcaff863a57b516c0ecefba0f82f5) now makes the reminder schedule, delivered-ID ledger, delivery-ownership index, and in-app inbox one ordered durable candidate: a failed save cannot publish memory, and a missing required sink fails closed before provider delivery. Exact-subject erasure removes all four views behind a durable resurrection fence, survives restart, serializes against the same-process reminder cycle, and preserves adjacent recipients. DSAR export reports the same owned schedule, ledger, and inbox state and returns empty views after deletion. Focused unit, restart, race, export, and live Postgres integration coverage is green. This closes the current reminder partition without claiming distributed multi-worker serialization or provider-level exactly-once delivery; an external provider call can still be replayed if the final delivery-ledger commit fails.

Nisaba-local continuity hardening through af7fffdb9a6fcaff863a57b516c0ecefba0f82f5 now extends the same ordered candidate/save/publish discipline to saved passage IDs and study reminders without duplicating shared corpus bodies. Notebook/annotation, saved-passage, and study-reminder erasure use independent durable fences, so selective category deletion preserves unrelated state while rejecting stale recreation. The signed saved_claim_passage and continuity_state services now compose both Nisaba partitions with their existing stores and fail closed on invalid counts or partial errors. DSAR export includes both saved representations, Nisaba study reminders, notebooks, and annotations, and becomes empty for each erased category. Unit, route/contract, race/restart, live Postgres, type-ratchet, and production-bundle checks are green.

Notification-state hardening through 067d5fff4634a189a751faa366f4ac0342073a61 now makes domain, persona, quiet-hours, digest, read, unread, dismissal, and bulk-read mutations publish only after ordered durable snapshots commit. Production requires and hydrates both notification stores before traffic; failed persistence cannot acknowledge volatile state. Independent exact-subject fences survive restart, serialize with mutation, preserve adjacent users, and reject resurrection. Notification preferences now compose into signed profile deletion and export, while read/dismiss state composes into signed continuity_state; invalid counts or partial partition failures fail the aggregate receipt closed. Focused route, unit, restart/race, DSAR/deletion, live Postgres, type-ratchet, ESLint, and production-bundle verification is green.

Personalization-state hardening through 90e6592487f2be1d36069c9be4789fcf1e9120c2 replaces fire-and-forget vector and active-persona snapshots with ordered candidate/save/publish mutations. The cross-process vector consumer, persona selection/clear routes, and admin bulk assignment now await persistence; production requires and hydrates both stores before traffic, and failed saves cannot publish volatile state. Independent exact-subject fences survive restart, serialize against stale vector/persona mutations, and preserve adjacent users. The signed personalization_vector service count-validates and composes both partitions, while DSAR keeps their existing personalization_vector and persona views and proves both empty after erasure. Focused unit, consumer/admin route, assistant, deletion lifecycle, DSAR, live Postgres, type-ratchet, ESLint, convention, and production-bundle checks are green.

Customer message-center hardening through 25f9acaf3fba4cc20859b2472b4a90589457e4bd makes visible and archived messages, read state, delivery receipts, channel/category opt-outs, and quiet-hours delayed dispatches one ordered durable snapshot. Customer/operator mutations, Trust & Safety decisions, and Metis appeal notifications await persistence; multi-receipt delivery is atomic, replay failures remain queued with durable attempt state, failed saves cannot publish candidates, and routes return 503 or post-erasure 409 as appropriate. Exact-subject fences survive restart, serialize against stale sends/schedules, and preserve adjacent users. The signed continuity_state service count-validates this partition, while DSAR exports complete history, policies, delayed work, and revision—including archived rows—and proves the section empty after erasure. Pure-engine, unit, route, producer, notification, deletion-lifecycle, DSAR, live Postgres, type-ratchet, ESLint, convention, and production-bundle checks are green.

Device push-token hardening through e6b90e13762cc94a8b8a0f70750248c9b4f14519 replaces fire-and-forget registration/removal with ordered candidate/save/publish mutations. Failed saves return 503 without changing live or restart-restored delivery targets; missing production persistence fails closed. Per-subject revisions and exact deletion fences survive restart, serialize both race orders against stale registration, and preserve adjacent users. The signed session service now count-validates auth sessions, device-session/audit records, and push tokens as independent partitions. Session DSAR exports complete device metadata and a SHA-256 fingerprint without exposing the raw delivery credential, then proves the section empty after erasure. Unit, route, deletion-lifecycle, DSAR, live Postgres, type-ratchet, ESLint, convention, and production-bundle checks are green.

Customer channel-binding hardening through c6e034e46dbe119409147fa6d3356fe3d2f0d274 replaces fire-and-forget email, SMS, and WhatsApp binding writes with ordered candidate/save/publish mutations. Request, confirmation, and step-up unbind routes await persistence; failures return 503 without exposing a new address, burning an attempt, or revoking a verified binding. Per-subject revisions and exact deletion fences survive restart, serialize both race orders against stale requests, and preserve adjacent users. The signed profile service independently count-validates consumer profiles, customer-auth profiles, notification preferences, and channel bindings. Profile DSAR exports binding and verification metadata while withholding the salted verification codeHash, then proves the section empty after erasure. Store, route, connected-consumer, deletion-lifecycle, DSAR, live Postgres, type-ratchet, ESLint, convention, and production-bundle checks are green.

Per-user Telegram hardening through 6262e3a94e09aded2f2b5ead93d035366a7fc2e1 replaces fire-and-forget bot capture, identity-link, voice, delivery-stop, and unlink writes with one ordered candidate/save/publish snapshot. Bot effects, authenticated link minting, login binding, webhook link completion, and the shared quiet-hours preference update all await persistence; failures return 503 without claiming a write. Separate saved-content and profile revisions/fences survive restart, preserve adjacent users and the other category under selective deletion, serialize both race orders, and reject stale recreation. Telegram notes/links independently count-validate under saved_claim_passage; link and voice/delivery metadata do the same under profile. DSAR exports capture provenance and profile metadata with a SHA-256 nonce fingerprint while withholding the raw pending-link nonce. Store, effects, route, connected profile/notification, deletion-lifecycle, DSAR, live Postgres, type-ratchet, ESLint, convention, and production-bundle checks are green.

Celestial ritual-schedule hardening through 844ee060a65e29a55ea6028959c2763bbdff621c replaces fire-and-forget accepted- trigger snapshots with one ordered candidate/save/publish stream. The route awaits persistence and returns 503 without publishing volatile state; missing production persistence fails closed. Per-subject revisions and exact deletion fences survive restart, serialize both schedule/delete race orders, reject stale recreation with 409, and preserve adjacent users. The signed continuity_state service count-validates the independent celestial_ritual_schedule partition. DSAR exports complete occasion, timing, provenance, quiet-hours displacement, acceptance, ownership, and revision metadata and proves the exact section empty after erasure. Store, route, deletion-lifecycle, DSAR, live Postgres, type-ratchet, ESLint, convention, and production-bundle checks are green.

Wearable-health hardening through 0656aabd9b7a3b57588f897be492a8b9c49a3b7a now commits a stable per-platform delivery intent before invoking HealthKit or Google Fit, retains provider IDs and acknowledgement-lost operation IDs across restart, retries only unfinished effects, and publishes session/HRV state only after ordered persistence succeeds. Exact-subject erasure requires idempotent provider confirmation before removing the local identities and installing its durable resurrection fence; the signed ritual_completion fanout runs that partition before its atomic canonical V1 database delete and fails closed on a missing bridge or invalid count. DSAR exports the wearable partition and now composes the exact-owner canonical Tara ritual-session and sitting-completion rows when OSHUN_V1_DATABASE_URL is configured; without that binding, it explicitly reports the canonical V1 ritual rows as a coverage gap. Unit, route, deletion, export, live Postgres, shell-library, type-ratchet, ESLint, formatting, build, and production-bundle checks are green. This closes the BFF store's durability/privacy composition, not live provider integration: the repository still has no production HealthKit/Google Fit adapter, so the production route remains fail-closed with 503 and provider-native verification remains open.

Morning-capture hardening through 989ce593373199adf3e752e1da89f59089a66e3c replaces same-day fire-and-forget self-report snapshots with one ordered candidate/save/publish stream. Failed or missing persistence returns 503 without publishing the capture; per-subject revisions and exact deletion fences survive restart, serialize both submit/delete race orders, reject stale recreation with 409, and preserve adjacent users. The signed journal_entry service now count-validates the independent BFF morning_capture partition before deleting canonical Arete rows. DSAR exports the full normalized capture, waking mood, dream note, crisis flag, timestamps, ownership, and revision while explicitly including exact-owner canonical V1 journal rows when OSHUN_V1_DATABASE_URL is configured; without it, the canonical journal gap remains explicit. Unit, route, deletion, export, live Postgres, type-ratchet, ESLint, formatting, and production-bundle checks are green.

Living Scene kept-artifact/share hardening through da8d927d9dc8cfe30d21500b8be0785287579043 now makes kept artifacts, exact C2PA-bound APNG bytes, share links/indexes, audit/takedown evidence, revisions, and deletion fences one strictly validated candidate/save/publish stream. Production startup requires and initializes the durable partition, restores and migrates it before traffic, and physically purges expired bounded-retention artifacts before serving; a supervised periodic executor continues that purge. Failed saves and purges cannot publish volatile gallery, viewer, media, or audit state. Exact-subject generated_artifact export and erasure now include this partition, with password verifiers withheld, adjacent owners preserved, operator audit identity pseudonymized where evidence must remain, and a restart-safe fence rejecting stale recreation. Fifteen-minute reissuable download grants are explicitly ephemeral and proven safe to lose across restart. The collector honestly leaves canonical Isis outputs and the separate generation-jobs/catalog/operator-gallery views as coverage gaps. Unit, full route, DSAR/erasure, live PostgreSQL, type-ratchet, ESLint, formatting, and production-bundle checks are green.

Generated-artifact DSAR completion through acc16f9080be7698099cba6752f8af1287db92ea now exports the exact subject's durable generation requests/results/governance, released-output catalog, provenance/lineage, and operator-gallery records/action facts. Immutable owner attribution defeats spoofed request fields, adjacent accounts are excluded, and internal operator identity is withheld from the customer bundle. Canonical Isis GeneratedOutput DSAR completion through 82497742474e5ae6354fa824c5e35a58a146bc2b now adds a lazy exact-owner reader at the deliberate privacy boundary. With explicit ISIS_DATABASE_URL configuration, the bundle includes canonical output metadata, provenance, and only lineage whose two endpoints are owned by the subject; recursive credential fields are redacted, BigInt sizes remain lossless decimal strings, adjacent outputs are excluded, and malformed or mismatched rows fail closed. A generic BFF DATABASE_URL cannot activate the reader. Without the explicit Isis binding, the collector continues to report canonical_isis_generated_outputs as an honest coverage gap; with it, the named generated-artifact DSAR gap is closed. Unit, route, migrated live-local-PostgreSQL, Prisma client/schema, privacy/database build and typecheck, zero-backlog BFF type-ratchet, strict lint, formatting, and resource-safe production-bundle checks are green. The broader all-store V1-002 inventory remains open.

Canonical shared-V1 DSAR completion through 234523f6c6323ab4a546828c2fa2adba37120247 closes the named database-read gaps for journal_entry, ritual_completion, and evidence_pack in deployments that bind OSHUN_V1_DATABASE_URL. One injected reader issues exact userId or learnerId queries against Arete journal entries, Tara ritual sessions and sitting-completion events, and Metis assessment-evidence packs and academic- integrity verdicts. Explicit field allowlists withhold internal tombstone actor identity, recursive JSON normalization redacts credential-shaped fields, adjacent and null-owner records are excluded, ordering is deterministic, and malformed or owner-mismatched rows fail the export without publishing a bundle. When the database is unconfigured, all four named canonical partitions remain honest coverage gaps. Unit, collector, authenticated-route, existing deletion- eraser, real-PostgreSQL five-table cleanup, Prisma schema, zero-backlog type- ratchet, strict lint/formatting/conventions, and exact production-bundle checks are green. No production database or production request is claimed, and this DSAR progress does not close the broader all-store inventory.

Calendar OAuth connection hardening through 20b4e99fa49118643894e5e8413fd20f195e7625 closes the local per-user connection partition. Production requires and authoritatively hydrates its durable snapshot; callback saves, disconnects, and exact-subject erasure serialize candidate state and publish only after an awaited commit. Profile DSAR exposes connection metadata but never the refresh credential. Signed profile erasure removes every local credential and in-flight CSRF state for the exact subject, preserves adjacent accounts, and persists a fence that rejects stale starts and saves across restart and both race orderings. A subject's final disconnect also survives restart as an active revision tombstone and reconnects monotonically. Unit/route/export/deletion tests, real-local-PostgreSQL restart proof, the zero-backlog BFF type ratchet, strict lint/formatting/conventions, and the direct production bundle are green. This does not revoke Google, Apple, or Microsoft grants or delete provider-side calendar data; provider-specific revocation and reconciliation remain open, as does the wider operational-store inventory. No production provider, database, or deletion request is claimed.

Arete member-continuity hardening through 43a87fdb6e03cea3e92f00f6250fd19550f547b1 closes the previously fire-and-forget weekly-review closure, living-offering decision, and coach decision stores. Production now requires and authoritatively hydrates all three existing snapshot keys before traffic. Mutations use one ordered candidate/save/publish authority, return bounded 503, 409, or 422 responses at the API boundary, and cannot acknowledge state a failed snapshot did not commit. Deterministic per-subject revisions and deletion fences survive restart, reject stale recreation, preserve adjacent members, and retain a coach clear tombstone so the next decision advances monotonically. Exact-owner DSAR exports the three partitions under continuity_state; the signed continuity fanout hard-deletes the exact subject across all three independently retriable stores and refuses a successful receipt on any partial error. These are lifecycle-bound product-history records, not compliance evidence: they have no time-based retention exception and are physically removed by signed account erasure. Unit, route, export/deletion, real-local-PostgreSQL restart, zero-backlog BFF type-ratchet, strict lint/formatting/conventions, and direct production-bundle checks are green. No UI, provider, schema, production database, or production request is claimed, and the all-store inventory remains open.

Nyx member-continuity hardening through 5abc400e018c6beff1bc0302a795bd000b642d27 closes the saved-celestial-object, sky-event-reminder, and authenticated object-search-count partitions behind the in-process Nyx adapter. All five mutation forms now serialize one complete candidate, await the existing snapshot key, and publish only committed state; production requires and authoritatively hydrates that authority before traffic. Deterministic subject revisions and erasure fences survive restart, serialize both mutation/deletion race orders, reject stale recreation, and preserve adjacent members. Exact-owner DSAR exports the three views as nyx_member_state under continuity_state; query text is never stored, only the per-member search-count aggregate. The signed continuity fanout physically removes the exact subject and count-validates the partition. Saved objects, reminders, and the aggregate are lifecycle-bound product state with no time-based retention or legal-hold exception; signed account erasure ends their retention. Store, adapter, route, DSAR/deletion, real-local-PostgreSQL restart, zero-backlog type-ratchet, strict lint/formatting, and direct production-bundle checks are green. No UI, external provider, schema migration, production database, or production request is claimed, and the all-store inventory remains open.

Authenticated client-state hardening through aedda7fe1ac5a05fdf90ea8b1ebe71c674535711, corrected and published through c0681992f5638c86ccf8d73b442d31da99e7b67d, closes the per-user JSON snapshot behind authenticated GET/PUT /v1/client-state/:stateKey. The BFF admits four allow-listed slots, while repository inventory demonstrates only workspace-layout as a current web producer; no UI producer is invented for the other three. PUT and retention-touching GET now share one ordered candidate/save/publish authority over the legacy snapshot key. Failed or missing persistence returns 503 without publishing a value or access clock; post-erasure writes return 409. Strict canonical JSON validation enforces the 64 KiB cap on UTF-8 bytes and rejects unsafe, cyclic, non-finite, deeply nested, duplicate, or fence-contradicting state. Deterministic subject revisions and deletion fences survive restart, serialize both write/deletion race orders, reject stale recreation, and preserve adjacent members. Exact-owner DSAR exports active slots as client_state under continuity_state; the signed fanout physically deletes and count-validates the exact subject. The store derives the canonical 30-day last-access window and daily purge cadence from OSHUN_RETENTION_RULES, durably advances access before serving a slot, refuses to revive expired values, purges at startup and on a supervised daily timer, and permits genuinely new state after inactivity expiry. Store, guarded-route, DSAR/deletion, retention, real-local-PostgreSQL restart/expiry, zero-backlog BFF type-ratchet, strict lint/formatting/conventions, and direct production-bundle checks are green. No rendered UI, mobile flow, schema migration, external provider, production database, or production request is claimed, and the all-store inventory remains open.

Authenticated Psyche dialogue-session hardening through ea6721e1d1bc0d92ff41f16e611729fae828fe2d closes the raw-turn store shared by the customer HTTP and WebSocket session flows. Create, turn, and end acknowledgements now await one ordered candidate/save/publish authority; production requires the sink, failed persistence publishes nothing, and controlled HTTP/WebSocket errors distinguish unavailable durability, an account-erasure fence, and an oversized UTF-8 turn. Authoritative boot hydration migrates the legacy snapshot and strictly rejects unknown, unsafe, duplicate, noncanonical, oversized, or fence-contradicting state. Deterministic revisions and exact-subject fences survive restart, serialize both turn/deletion race orders, reject stale recreation, and preserve adjacent accounts. DSAR exports the exact owner as psyche_dialogue_sessions under conversation_history; the signed eraser count-validates and physically removes that partition alongside the older transcript and recent-search stores. Startup and supervised daily purges derive the canonical 30-day creation window from OSHUN_RETENTION_RULES and honor exact active/pending-release legal holds plus active category retention exceptions. Store, HTTP, WebSocket, DSAR/deletion, retention, real-local-PostgreSQL restart/expiry/held-history, zero-backlog integrated BFF type-ratchet, strict lint/formatting/conventions, and direct production-bundle checks are green. This did not close the separate cross-process conversation.turn.recorded transcript store at that source boundary; the later shell-assistant transcript section below closes it. No rendered UI, mobile flow, schema migration, external provider, production database, or production request is claimed, and the all-store inventory remains open.

Shell-assistant and external transcript hardening through b8d4535215eb9f8abace1d556d02bc893461d572 closes that older raw-turn partition and its actual in-repository producer. The authenticated shell message route now awaits one atomic user/assistant exchange containing the final text returned to the member, including synthesis, memory enrichment, handoff, blocked, and crisis-intercepted paths; failed durability returns controlled 503, an erasure fence returns 409, and oversized raw content returns 413. External conversation.turn.recorded delivery now awaits the snapshot before event acknowledgement, dead-letters malformed/stale/fenced input, propagates operational failure for retry, and replays unacknowledged durable events at boot. One ordered candidate/save/publish authority covers single turns, atomic exchanges, expiry, and deletion. Schema-v1 recovery strictly migrates the legacy shape and validates deterministic revisions, ordered bounded turns, canonical timestamps/identities, size limits, and persistent exact-subject fences. Exact-owner DSAR emits conversation_turns under conversation_history; the signed eraser physically removes and count-validates it alongside recent search and Psyche dialogue. Startup and supervised daily purges implement the canonical legal-hold/retention-exception-aware 30-day creation window. Store, consumer, actual shell route, safety/synthesis, DSAR/deletion, retention, real-local-PostgreSQL plus Redis commit-before-ack/restart proof, zero-backlog BFF type-ratchet, strict lint/formatting/conventions, and direct production-bundle checks are green. No rendered UI, mobile flow, schema migration, external provider, production database, or production request is claimed. The separate process-local shell-session engine and the wider all-store inventory remain open.

Shell-assistant session-continuity hardening through 140a6bce686dac8d1c41fd811e7663f3c65b214a closes that separate process-local engine boundary. Customer-visible session identity, turn admission, bounded recent-intent routing, domain context, continuity/disclosure/memory/grounding posture, and pending handoff now commit through one ordered restart-safe checkpoint authority before an acknowledgement is returned. The minimized checkpoint deliberately excludes the full raw transcript already governed by the conversation-history partition, while bounded grounding/handoff content remains exported and erasable as assistant_session_checkpoints under continuity_state. Production rejects a missing sink; failed saves publish no candidate; optimistic per-session revisions reject concurrent stale updates; normal close awaits physical removal; and exact-subject deletion durably fences recreation while preserving adjacent members. Startup and supervised five-minute purges enforce the shell engine's exact 30-minute creation-time lifetime. This short-lived, non-evidentiary duplicate-free partition is intentionally not held beyond that product lifetime; legal-hold-aware raw dialogue remains in the separate transcript authority. Store, route, DSAR/deletion, local-PostgreSQL restart/expiry, shell-library and BFF type gates, strict lint/formatting/ conventions, and production-bundle checks are green. No rendered UI, mobile flow, schema migration, external provider, production database/request, model, or Unreal claim is made. The wider all-store inventory remains open.

Connected BFF Iris memory-authority hardening through a64a03415eda5b11a2edb283a7b2f2e2622326cf closes the previously volatile, fire-and-forget local memory_scope partition. Candidate snapshots now migrate, validate, serialize, and commit through one production-required durable authority before working or archival memory is published; a failed save exposes neither tier, while restart hydration, deterministic revisions, detached reads, and persistent exact-subject deletion fences prevent stale resurrection. Normal reads durably refresh last access, and startup plus supervised seven-day sweeps physically enforce the canonical 365-day inactivity boundary. Hard memory delete physically removes content; soft delete remains intentionally recoverable. The customer bundle now includes exact-owner local memory under memory_scope, with real inline JSON/portable/Markdown/CSV rendering, truthful byte counts and SHA-256 digests, and explicit failure rather than falsely labelling plaintext as encrypted. Signed account erasure requires the local physical deletion before accepting the independent canonical Iris receipt; adjacent users remain intact, and local failure overrides remote success. The shell assistant lazily rebuilds only this intentionally ephemeral bridge after durable session recovery and discards it if the authoritative session checkpoint fails. Store, route, DSAR/deletion, real-local-PostgreSQL restart/expiry/erasure, shell and memory library regressions, zero-backlog BFF type ratchet, strict static gates, and production-bundle verification are green. No rendered UI, mobile flow, schema migration, external provider, production database/request, model, Unreal, or performance claim is made. The canonical Iris service remains the independent signed-receipt authority, and the wider all-store inventory remains open.

Reminder-source projection hardening through c6a17f56f1a6c6942559972a4049a38afe2bf776 closes three additional connected BFF continuity partitions: per-member due assignments, upcoming sessions, and scheduled content drops. Their event consumers now commit one validated, production-required snapshot before acknowledging Redis delivery; failed saves publish nothing and leave delivery replayable, while boot replays unacknowledged events. Schema-v1 hydration strictly migrates the three legacy collection shapes, applies expiry before publication, and rejects malformed, duplicate, noncanonical, credential-bearing, over-capacity, unknown-version, or fence-contradicting state. Exact-subject erasure physically removes all three views and commits SHA-256-only resurrection fences, including zero-row deletes; adjacent members remain intact. Exact-owner DSAR emits the three partitions under continuity_state, and calendar reads now use the same subject-scoped authority and durably refresh last access. Startup and a supervised daily sweep enforce the canonical 30-day continuity-state window, anchored no earlier than the future occurrence so live intent is not deleted before it becomes terminal. Focused store, consumer, producer, calendar, export/erasure, race/restart, real-local-PostgreSQL, real-local-Redis, zero-backlog BFF type-ratchet, strict static, and production-bundle checks are green. These are BFF projections, not the owning assignment/session/content services: those services still must emit the per-recipient events, and no provider delivery or distributed exactly-once claim is made. No rendered UI, mobile, schema-migration, external-provider, production-database/request, model, Unreal, or performance claim is made. The wider all-store inventory remains open.

Metis academic-record hardening through 65200d9e043b3acefd6eef19743e2de7f7033e97 now makes Themis integrity verdicts, appeals, gradebook consequences/corrections, deletion fences, and notification outbox state one ordered operation ledger. Candidate state is replayed and persisted before either the canonical bridge or gradebook is published; failed saves cannot expose partial adjudication, decision, correction, or notification intent. Production requires and hydrates the durable partition before traffic. Tenant isolation, authenticated operator attribution, stable acknowledgement- lost notification replay, restart validation, and both write/delete race orders are covered. Exact-subject evidence_pack export and erasure now include this local partition: learner records are physically removed, retained evidence is pseudonymized/redacted when the deleted subject acted as operator or appellant, adjacent learners remain, and a SHA-256-only restart-safe fence rejects stale recreation. At that source boundary, the export included complete local verdict/ audit/appeal and gradebook history, declared a 730-day default with legal hold required, and explicitly said automatic scheduled purge was not implemented. That local scheduled-expiry gap is closed below. Canonical V1 assessment evidence and verdict tables now compose through the configured exact-learner DSAR adapter above and remain named gaps only when the shared V1 database is unconfigured. Focused unit/route/export/fanout tests, live PostgreSQL restart/deletion proof, zero-backlog BFF type ratchet, lint, formatting, conventions, and the exact production esbuild entry bundle are green. The dependency-expanded Nx build remains blocked by unrelated @oshun/audit-platform TypeScript-6 baseUrl deprecation debt.

Stripe provider-deletion hardening through 1ecef2d5c36ae810d5666b2434752736aa44d42d now requires confirmed deletion of the bound Stripe Customer before locally pseudonymizing retained Fiat Stripe rows and installing their durable resurrection fence. The live full-account fanout resolves the production Stripe client, uses one stable SHA-256 operation key, and reconciles acknowledgement loss only from Stripe's deleted-customer tombstone; an ambiguous 404, an active customer, a malformed success response, or missing provider configuration while a binding exists fails the signed partition closed. A checkout that never acquired a provider customer remains locally erasable. Stripe documents that Customer deletion immediately cancels active subscriptions; provider-retained invoices and the deleted-customer tombstone are therefore explicitly treated as provider compliance records, not claimed as a total provider-record purge. Focused client/store/route/export and fanout tests, real PostgreSQL restart/fence proof, zero-backlog BFF type ratchet, typecheck, lint, formatting, and the production entry bundle are green. This closes the Stripe Customer erasure boundary only. At that source boundary, Apple/Google provider dispositions and the shared legal-hold/retention-exception executor remained open.

Compliance-retention execution through 068f2a2619f01933181773302c7d2b3056040366 now gives each new Fiat Stripe account redaction a durable seven-year retention anchor and runs an awaited pre-traffic plus supervised periodic purge. The shared evaluator collects every persisted admin legal hold and retention exception, applies subject/category/ status/expiry matching, and emits bounded per-candidate outcomes for not-due, missing-anchor, hold-blocked, exception-blocked, or purged records. An eligible Stripe candidate physically loses its pseudonymous customer/subscription binding and pending-checkout attribution while retaining only the irreversible resurrection fence and purge timestamp; adjacent subjects remain intact. Legacy redactions without a trustworthy timestamp fail closed as missing-anchor rather than receiving an inferred deletion date, and a replayed account erasure can start a conservative new window. Focused executor/store and connected route/export/eraser tests, real PostgreSQL expiry/purge/restart proof, the zero-backlog BFF type ratchet, affected hook typechecks, lint, formatting, and the production entry bundle are green. At that source boundary, this connected only the Fiat Stripe compliance partition to the reusable executor. Customer billing, consent, export-request, financial tip, and support-case retention partitions plus Apple/Google provider dispositions remained open.

Customer-billing retention through 7c93da74220ddac485f90c4de72e4c71ea62c070 now connects the second real partition to that executor. New account redactions durably anchor the complete customer-billing pseudonym for 2,555 days; a due, unblocked purge removes the subscription, trial/pause/cancel/save/churn history, family-plan ownership or seat, scholar-verification evidence, Google Play purchase-token digest, Apple original-transaction binding, and every newly chain-attributed Apple replay ID without touching an adjacent account. The SHA-256 subject fence and purge timestamp remain. Pre-upgrade redactions without an anchor defer until a conservative erasure replay, and pre-upgrade Apple replay IDs without a safe owner-chain attribution remain global replay fences rather than being guessed or claimed as subject-purged. Real PostgreSQL purge/restart proof, 41 focused store/provider-route/executor tests, 48 connected route/export/eraser tests, strict lint and formatting, the zero-backlog BFF type ratchet, affected hook typechecks, and the production entry bundle are green.

Customer data-export request retention through 8beaa6c940b5971d8516874f3830a28273d53f73 now connects the third real partition to the shared executor. Full-account erasure rekeys request metadata to a SHA-256 pseudonym, removes live download URLs and free-text failure details, durably anchors a 2,555-day data_export_request window, and installs an irreversible resurrection fence. Once due and not blocked by the shared legal-hold or category-exception authority, physical purge removes the retained request rows while preserving only the fence, redaction timestamp, and purge timestamp. Legacy redactions without a trustworthy anchor defer; a conservative erasure replay may establish one. Strict hydration rejects corrupt timestamps, duplicate identities or export IDs, unrecognized pseudonyms, purged-row resurrection, and direct-subject resurrection. Adjacent accounts survive both redaction and purge. Focused store/executor/route/contract tests, the connected export/deletion integration flow, real PostgreSQL redaction/purge/restart proof, strict lint and formatting, contracts typecheck, the zero-backlog BFF type ratchet, and the production entry bundle are green. Short-lived export-bundle bytes remain governed separately by their existing generated-artifact expiry and deletion path; this slice does not misclassify them as seven-year request metadata.

Canonical consent retention through d862143ff8b79f127e93cfaa87c9a2c7d4bf979c now connects the fourth real partition to the shared executor. Full-account erasure rekeys the complete per-flow decision/evidence/history ledger to a SHA-256 pseudonym, replaces caller-supplied evidence labels, removes evidence URIs and notes plus subject or operator free text, preserves bounded system policy notes and canonical status/policy/verification/timestamp authority, durably anchors a 2,555-day consent_record window, and installs an irreversible resurrection fence. A due, unblocked purge physically removes every pseudonymous flow record while retaining only that fence and the redaction/purge clocks. Failed redaction or purge saves publish nothing. Legacy redactions without a trustworthy anchor defer; strict hydration rejects invalid retention clocks, unknown or purged pseudonyms, and direct-subject resurrection. Focused store/executor tests, 78 connected consent/residency/persona/mentor/export route tests, real PostgreSQL redaction/purge/restart proof, strict lint and formatting, the zero-backlog BFF type ratchet, the targeted two-project hook typecheck, and the production entry bundle are green.

Scene financial-tip retention through 5e22686cf704fbd3f89c009cffc839b3f3c484f1 now connects the fifth real partition to the shared executor under the 2,555-day billing_history window. Account erasure pseudonymizes viewer, creator, and authority-recorder identity and durably anchors the window. A due record remains blocked by the shared legal-hold or category-exception authority. Once unblocked, expiry atomically destroys every stored redacted:<subject-digest> link and replaces it with one random financial tombstone whose mapping to the retained digest is never stored. The shared transaction amount, currency, sequence, running accrual, settlement, 10-percent fee split, counsel memo, and adjacent counterparty identity remain exact; removing or rewriting those facts would corrupt the counterparty's financial evidence. The digest fence and redaction/purge clocks remain. Failed expiry saves publish nothing, legacy redactions without an anchor defer, and strict hydration rejects invalid clocks, unknown or purged pseudonyms, malformed tombstones, and direct-subject resurrection. Focused store/executor tests, 40 connected tip-route/DSAR/profile tests, real PostgreSQL redaction/expiry/restart proof, strict lint and formatting, both affected TypeScript checks, the zero-backlog BFF ratchet, and the production entry bundle are green.

Customer-feedback support-case retention through c8e99bf217937a9797c1cee9218d1955ec6d5d60 now connects the sixth real partition to the shared executor under the 1,825-day support_case window. Account erasure removes submitter/contact identity and caller-authored contact fields, replaces the exact account key with its SHA-256 pseudonym, and durably anchors the window. A due case remains blocked by the shared legal-hold or category-exception authority. Once unblocked, expiry atomically deletes every retained case for that pseudonym while preserving only its one-way recreation fence and redaction/purge clocks; adjacent customers remain intact. Failed saves publish nothing, legacy redactions without a trustworthy clock defer, and strict hydration rejects invalid clocks, rows remaining after purge, and direct subject resurrection. Sixty-one focused store/executor/eraser/route/export tests, real PostgreSQL redaction/expiry/restart proof, strict lint and formatting, both affected TypeScript checks, the zero-backlog BFF ratchet, and the production entry bundle are green. The two then-remaining Themis moderation-appeal and Living Scene abuse-report support_case partitions are closed below.

Themis moderation-appeal retention through abe4159c3b4de8f6bd716117d426d8a11a9fc146 now connects the seventh real partition to the same 1,825-day support_case executor. Account erasure removes appellant-authored narrative/evidence and direct appeal/action/statement identifiers, rekeys the categorical due-process record to its SHA-256 pseudonym, and durably anchors the window. At the exact due boundary, shared legal holds and support_case exceptions block deletion; once unblocked, an ordered save physically removes every appeal under the pseudonym while preserving only its one-way fence and redaction/purge clocks. Failed saves publish nothing, adjacent appellants remain exact, legacy redactions without an anchor defer, and strict hydration rejects invalid clocks, retained rows after purge, and direct-subject resurrection. Fifty-six focused store/executor/eraser/route/export tests, real PostgreSQL redaction/expiry/restart proof, strict lint and formatting, both affected TypeScript checks, the zero-backlog BFF ratchet, and direct plus dependency-excluded Nx production bundles are green. The final named Living Scene abuse-report support_case partition is closed below.

Living Scene abuse-report retention through 7544b773daa7c9aec87a29a28e7a95f093facd3e now connects the eighth real partition to the same 1,825-day support_case executor. Account erasure removes reporter-authored narrative and direct report identity, pseudonymizes reporter and action-actor identity, removes the erased actor's resolution note, durably anchors the window, and installs an irreversible recreation fence. At the exact due boundary, shared legal holds and support_case exceptions block expiry. Once unblocked, reporter-owned cases are physically deleted; actor-only links on an adjacent reporter's case are irreversibly replaced with a fresh random purged-actor tombstone whose mapping is never retained, while the adjacent reporter's categorical evidence remains exact. Failed saves publish nothing, legacy redactions without an anchor defer, and strict hydration rejects invalid clocks, retained subject links after purge, malformed reserved pseudonyms, and direct-subject resurrection. Eighty-three connected store/executor/eraser/ route/export tests, real PostgreSQL redaction/expiry/restart proof, strict lint and formatting, both affected TypeScript checks, the zero-backlog BFF ratchet, and direct plus dependency-excluded Nx production bundles are green. All three named support_case partitions are now connected; this does not close the broader all-store V1-002 gate.

Metis academic evidence retention through d01e4c2c315f3156b4abb4c1b26d39c36cc558fb now connects the ninth real partition to the shared executor under the 730-day evidence_pack window. Account erasure still physically removes learner-owned verdict, appeal, gradebook, and notification operations immediately. Only adjacent learners' categorical evidence remains when the erased subject acted as a reviewer, respondent, or appellant; those roles are pseudonymized and all narrative authored by that subject—including the duplicated correction rationale—is redacted. At the exact due boundary, shared legal holds and evidence_pack exceptions block expiry. Once unblocked, one ordered save replaces every retained subject link with a fresh random purged-subject tombstone whose mapping is never stored, while preserving adjacent learner evidence. The SHA-256 deletion fence and deletion/purge clocks remain. Failed saves publish nothing, legacy snapshots without a trustworthy anchor defer until a conservative erasure replay, and strict hydration rejects invalid clocks, duplicate or noncanonical retention metadata, unknown/deleted links after purge, malformed reserved pseudonyms, incomplete text redaction, and direct-subject resurrection. All 58 connected ledger/executor/route/export tests, the 11-test deletion-fanout suite, real PostgreSQL expiry/restart proof, strict lint and formatting, the zero-backlog BFF ratchet, both affected hook typechecks, and direct plus dependency-excluded Nx production bundles are green. The export now truthfully declares automatic local purge implemented. This closes the local Metis academic-record expiry adapter; the separately named canonical V1 evidence and verdict-table DSAR gaps are closed for configured deployments by 234523f6c6323ab4a546828c2fa2adba37120247. The broader all-store V1-002 gate remains open.

Apple's published account-deletion guidance says Apple-billed auto-renewal can continue after app-account deletion and directs the customer to Apple's subscription-management surface; it also requires an immediate-deletion option. Google documents cancellation/revocation operations, but both require the raw purchase token, while this service intentionally persists only its digest. The documented Apple App Store Server and Google Play purchase surfaces expose no merchant transaction-record deletion operation, so no provider-record deletion receipt is fabricated. Local identifiers are governed by the executor above; authenticated, secret-free subscription authority and provider-specific client guidance are now published in 270cfe67cfae36131ecbfccee0ee122f98f1e888. Web and mobile send Apple- and Google-billed members to canonical provider management surfaces, suppress misleading local plan controls, warn that store billing can continue after Oshun account deletion, and preserve immediate local deletion. The same slice adds an optional Google stop-renewals seam that accepts the raw token only in the authenticated request, hash-matches the durable owner, never persists or echoes the token, and fails loudly without a configured provider manager. No production Google client or live cancellation is claimed; native Maestro device execution also remains unproven because no booted device or installed dev client was available.

DMCA case-authority hardening through 9d8faf8bf6fe90b0f5fe865baa93fed55d5e5234 now replaces the BFF's process-local notice, ownership, counter-notice, restoration, and strike state with one ordered, production-required durable snapshot. Candidate mutations validate and save before publication; failed saves expose no partial statutory state, and strict startup hydration rejects malformed, duplicate, noncanonical, credential-bearing, over-capacity, or redaction/purge-contradicting snapshots. Authenticated uploader records now export under support_case without exposing third-party claimant identity. Full-account erasure pseudonymizes the uploader and strike ledger, removes ownership attribution, scrubs counter-notice identity and narrative, and installs a persistent resurrection fence while preserving claimant statutory evidence. The resulting five-year retention anchor now participates in shared legal-hold/retention-exception evaluation and physical purge. Focused connected tests passed 69/69, real-local-PostgreSQL recovery and purge proof passed 1/1, the zero-backlog BFF type ratchet and exact production bundle passed, and strict lint, formatting, diff, and convention gates passed. This proves durable case records only: no external content/provider removal, uploader/claimant notification delivery, current production caller for the ownership-registration seam, or unauthenticated claimant rights workflow is claimed. The process-local DSA decisions authority and broader all-store inventory remained open at that source boundary.

DSA decision-authority hardening through a839ad6ffd802a919135e6d445eb290a37a792cd now replaces the BFF's process-local statement-of-reasons, reportable-decision, appeal-aggregation, and monthly- active-recipient state with a bounded, production-required durable snapshot. The live statement route now requires an explicit affected-account key and atomically commits its statement/reporting pair before returning 201; failed saves return controlled 503, conflicts return 409, and deletion-fenced subjects return 410. A new admin route durably records only monotonic MAR attestations, while deployment-environment fallback parsing is strict. Full- account erasure treats these non-exportable internal records as canonical review_package evidence: it pseudonymizes the affected account, removes statement and appeal narrative, installs a persistent resurrection fence, and anchors a seven-year legal-hold/retention-exception-aware physical purge. Focused store/route/retention tests passed 32/32, real-local-PostgreSQL recovery, redaction, attestation, and purge proof passed 1/1, and the zero-backlog BFF type ratchet, strict lint/format/diff/convention gates, and exact production bundle passed. This does not claim that a statement was delivered to the affected recipient or submitted to an external transparency database. The local generic appeal-aggregation methods remain a test/future seam: the separately durable Themis appeal intake has no connected operator verdict/reversal adapter, so production-complete appeal outcome reporting is still open. The broader all-store inventory also remains open.

Authenticated Metis ingest-job hardening through 10c0ef948db6813c3d1db2b81ddc483c9d9dc2ef replaces the route's synchronous map mutation and fire-and-forget snapshot save with a bounded, schema-validated, production-required candidate/save/publish authority. The live route now awaits the ordered save before returning 202, returns controlled 503 on missing or failed persistence, and returns 410 after a restart-persistent account-erasure fence. Strict startup validation and legacy migration cover jobs, derived outlines, identities, timestamps, kind/status coupling, source minimization, ordering, duplicates, capacity, and fence contradictions. Raw pasted text is never stored; URL credentials/query/fragment and PDF path components are removed before persistence. Exact-owner jobs and outlines now export under saved_claim_passage and compose into its signed, count-validated full-account erasure while adjacent learners remain exact. Focused store, route, and export/erasure tests passed 136/136, live-local-PostgreSQL recovery and erasure proof passed 1/1, and strict lint/format/policy gates, the zero-backlog BFF type ratchet, and exact production bundle passed. URL fetching, PDF byte extraction, asynchronous processing, remote provider assets, the DSA-to-Themis verdict adapter, and the wider all-store inventory remain open.

Authenticated Metis tutor-memory hardening through 87c6cb62d615504adfc37ecfeb114ca0d2145da8 replaces the live route's in-memory- first mastery/session mutation and fire-and-forget snapshot callback with one serialized candidate/save/publish authority. Production now requires and hydrates the sink before traffic; failed or missing persistence returns controlled 503 without exposing an uncommitted posterior, turn sequences are monotonic, and strict bounded recovery rebuilds every derived memory and validates the complete memory/state/diff graph. The legacy four-surface snapshot migrates and saves before traffic, and restore replaces rather than overlays state. Exact-owner session memories, current mastery, and the complete diff log now export under continuity_state, purge after the canonical 30-day activity window at startup and daily, and compose into count-validated signed account erasure with a restart-persistent zero-row fence while adjacent learners remain exact. Focused store, route, consumer, export/erasure, and retention tests passed across all 182 affected cases; live-local-PostgreSQL recovery and erasure proof passed 1/1; and strict lint/format/policy gates, the zero-backlog BFF type ratchet, and exact production bundle passed. No new tutor UI, voice/avatar runtime, remote Metis database, or production request is claimed. The DSA-to- Themis verdict adapter, provider-native erasure inventory, and wider all-store inventory remain open.

Metis BYO-material operator-decision hardening through e417f95c0ee568dd9b2a7865d3f095d5e1e789b1 replaces the live admin route's in-memory-first mutation and detached snapshot save with a bounded, schema-v1, production-required candidate/save/publish authority. The route now returns 201 only after the complete review verdict and rationale commit; missing or failed primary persistence returns controlled 503 without a row or audit mirror. Its searchable admin-audit event is explicitly secondary: after the review commits, the route awaits the durable audit append when supported and then commits its correlation in a second snapshot; a mirror/link failure leaves the already-committed verdict authoritative with a null correlation. Strict request/query schemas, deterministic snapshots, legacy migration, authoritative replace-on-restore, defensive reads, concurrency serialization, random UUID identities, and corruption/capacity/fence validation close the local recovery boundary. The rows are canonical non-exportable review_package governance records. Signed account erasure pseudonymizes the exact operator reference and free-text rationale, installs a persistent zero-row resurrection fence, and connects the record to the shared seven-year legal-hold/retention-exception- aware physical purge while preserving adjacent operators. Focused store, route, audit-adapter, retention, and profile-erasure tests passed 41/41; real- local-PostgreSQL restart/redaction proof passed 1/1; and the existing real HTTP/operator-UI Playwright journey passed 7/7 against an isolated current- source BFF. The zero-backlog BFF type ratchet, strict static/policy gates, and exact production bundle also passed. This does not connect the fixture-backed operator queue to the real ingest/course-build transition, attribute a reviewed source to a learner, make internal review packages customer-exportable, or provide a cross-store atomic transaction with the secondary audit table. No production request or remote provider was exercised. The DSA-to-Themis verdict adapter, provider-native erasure inventory, and wider all-store inventory remain open.

General operator-inbox decision hardening through 9609569b9ec8883296953100a7aaad0b9c819244 replaces the shared admin route's in-memory-first mutation and detached snapshot save with a production-required, schema-v1 candidate/save/publish authority. Record and reopen responses now acknowledge only the committed primary review row. A failed primary save returns controlled 503 without publishing state or an audit mirror; secondary audit append/link failure cannot roll back or misreport a committed verdict. Strict request and recovery schemas, deterministic snapshots, bounded legacy migration, replace-on-restore hydration, defensive reads, serialized per-process writes, random UUID identities, corruption/capacity checks, and persistent redaction fences close the local recovery boundary. Canonical internal, non-exportable review_package rows retain for seven years subject to legal holds and retention exceptions. Signed account erasure domain-pseudonymizes the exact reviewer, removes their rationale and authored assignment, scrubs exact assignee-value matches in adjacent rows, and connects due records to physical purge while preserving unrelated operators. Focused store, route, audit, retention, and profile-erasure tests passed 43/43; real-local-PostgreSQL restart/redaction proof passed 1/1; real HTTP Playwright passed 6/6; and the existing rendered operator journey passed 3/3 against an isolated current-source BFF. The zero-backlog BFF type ratchet, strict static/policy gates, and exact production bundle also passed. This does not replace the fixture-backed queue with canonical incident/citation/republication transitions, turn its free-form assignee display label into an operator-directory identity, provide a database transaction across the snapshot and audit tables, or provide distributed multi-instance snapshot coordination. No production request or remote provider was exercised. The DSA-to-Themis verdict adapter, provider-native erasure inventory, and wider all-store inventory remain open.

Themis moderation-appeal outcome integration through 2a7f20afcb6f18de1cfef779af5a2d0cdd710934 now connects the authenticated, durable customer intake to an admin-scoped, three-reviewer operator lifecycle and the durable DSA appeal aggregate. Original, triage, and deciding reviewers must be distinct; only accepted intake can advance; and an account-matched, appealable DSA decision for the exact action must exist before the primary verdict commits. The Themis snapshot remains the primary outcome authority: its acknowledged decision sets a durable pending marker, the DSA projection commits second, and a final Themis acknowledgement marks it reported. Exact replay, startup reconciliation after both stores hydrate, and deletion-time reconciliation before both subject redactions close the crash window without claiming a cross-store transaction. Customer readback exposes the verdict and rationale but removes internal reviewer identities. Focused store/route/DSA tests passed 52/52, three real-local-PostgreSQL recovery suites passed 3/3, and an isolated real-HTTP Playwright flow passed 2/2. The zero-backlog BFF type ratchet, strict static/policy gates, and exact production bundle also passed. This closes the previously named Themis-to-DSA appeal-outcome reporting seam. It does not claim a push notification, append-only operator audit mirror, external provider-side content restoration, statement delivery, submission to an external transparency database, distributed multi-instance compare-and-swap, or an atomic transaction across the two snapshot documents. Provider-native erasure inventory and the wider all-store inventory remain open.

Shared operator-incident authority hardening through f74af22b7e296edf23827c6155d5ebeed1d5d376 replaces process-visible mutation followed by a detached snapshot save with a schema-v1 candidate/save/publish queue. Acknowledge, assign, escalate, snooze, resolve, and test-reset responses now await the primary snapshot; failed saves return controlled 503 without publishing incident state or a secondary audit event. Production fails closed until the snapshot sink is wired, current restore is strict and replace-only, validated partial legacy documents migrate over the fixed six-item catalog, and terminal resolution now rejects every later lifecycle verb. The connected citation-drift terminal route also awaits this incident commit before recording its separate audit/detail rows. Signed account erasure pseudonymizes exact operator actors and acknowledgements, removes authored lifecycle detail and exact assignee references, installs a persistent resurrection fence, and retains the minimized internal review_package for seven years subject to legal holds and retention exceptions before irreversible identity purge. Focused incident, citation-route, profile-erasure, and retention tests passed 68/68; real-local- PostgreSQL recovery/redaction/purge proof passed 1/1; and the existing rendered incident-triage journey passed 13/13 against an isolated current-source BFF and a fresh schema-only database. The zero-backlog BFF type ratchet and source/static gates passed, and esbuild emitted the expected bundle. Runnable artifact closure through f52037994aeb2229960e8c23d67e44da6b31663f replaces the brittle CLI bundle with an ESM build authority that supplies bundled CommonJS with an ESM-safe require, resolves each imported sharp version through a direct workspace dependency that owns that exact native package, and loads c2pa-node through its declared @oshun/c2pa-export runtime owner. The build fails if an imported sharp version has no direct BFF provider. A checked-in production-artifact smoke gate now requires durable Postgres and Redis, boots the emitted dist/server.js with production configuration on an isolated port, verifies the authentic /healthz response, and requires graceful exit. That gate passed against a fresh schema-only local database, which was removed afterward. Citation-drift review-authority closure through bf5cc04b0dfef3753ba1aca48762cf883e2a286e replaces both remaining detached resolution/classification saves with strict schema-v1 candidate/save/publish stores that production requires and hydrates before traffic. Terminal decisions use a durable pending intent, commit the exact shared incident authority, then durably complete; exact retry and startup reconciliation close the inter-store crash window without claiming an atomic cross-document transaction. Source classifications commit before their secondary audit mirror. Failed saves return controlled 503 without publishing volatile review detail. Signed account erasure reconciles pending terminal work first, pseudonymizes exact reviewers, removes authored rationale and audit correlation, persists independent resurrection fences, and connects both internal review_package partitions to seven-year legal-hold/retention-exception-aware purge. Focused unit/privacy tests passed 45/45, real-local-PostgreSQL crash/restart/redaction/purge proof passed 1/1, and the two affected rendered Playwright flows passed 2/2 against an isolated current-source BFF and fresh migrated database. Assignee remains a free-form label rather than an operator-directory mapping, and the wider all-store inventory remains open.

Tenant SSO configuration hardening through 257b390ac439866669a9952bde1c2be57d7f4975 replaces process-visible mutation followed by a detached snapshot save with a strict schema-v1 candidate/save/publish queue. Create, import, update, and delete routes await the primary snapshot; a failed save returns a stable 503, leaves the prior configuration live, and does not append a success audit event. Deployable boot requires and hydrates the snapshot sink before traffic, current restore is strict and replace-only, bounded legacy snapshots migrate into canonical order, and concurrent whole-snapshot mutations serialize per process. Focused store, route, and probe tests passed 47/47; real-local-PostgreSQL restart, canonical-migration, CRUD, and final-state proof passed 1/1; and the existing real-HTTP tenant SSO CRUD/probe journey passed 8/8 against an isolated current-source BFF and fresh migrated database. The zero-backlog BFF type ratchet, lint, build, and emitted-production-artifact smoke gate also passed. This store contains tenant federation configuration rather than a user profile or customer-authored content, so it is not added to user account-erasure fanout. The wider all-store inventory remains open.

Tenant invitation, membership, and seat-capacity hardening through 747edeab5215be2349c16430a8b2e87b62974cb0 and 92bb24d911954b874ea75f0b1e897847332dd14c replaces detached saves with one strict schema-v2 candidate/save/publish queue. Invite, direct-add, accept, revoke, expiry, capacity, bulk-auth provisioning, and signed account erasure now await the primary snapshot; failed writes return a controlled 503 and publish neither new roster state nor success audit. Deployable boot requires and hydrates the Postgres-backed sink before traffic. Invite, member, and pool state restore together, bounded v1/legacy documents migrate conservatively, and accepted invitations carry an internal exact subject link that never enters the public view. Profile export and signed erasure therefore still find the right accepted invite when the current account email differs from the invited email; erasure removes the subject's memberships and matching invites, releases seats, and installs an exact-user resurrection fence. New invite/member success audits retain operational record ids and categorical state while omitting raw target email and user id. Focused store tests passed 22/22, five connected route/bulk/read/privacy suites passed 66/66, real-local-PostgreSQL migration/restart/erasure proof passed 1/1, and the two affected real-HTTP Chromium journeys passed 12/12 against an isolated exact-source BFF and fresh migrated database. The zero-backlog BFF type ratchet, lint, production bundle, and emitted-artifact smoke gate also passed. Serialization remains per process, accepted-email equality is not enforced, terminal invite records have no general time-based purge, historical audit PII is not retroactively rewritten, and acceptance still has no dedicated audit event. This closes the tenant invitation/member/seat snapshot partition within that boundary, not the wider all-store inventory.

Tenant organization-onboarding hardening through fb0d40ead8b36badad3f04284096fda68b1a3403 replaces detached snapshot saves and process-first publication with one strict schema-v1 candidate/save/publish authority. New tenants commit once as active; the store does not claim an external provisioning effect that does not exist. Deployable boot requires and hydrates the Postgres-backed sink, strictly replaces memory, canonically migrates legacy rows, and reconciles crash-left provisioning rows before traffic. The durable record preserves exact tenant contract, hierarchy, residency/region, billing, seat-allocation, isolation, and lifecycle fields. Failed writes return stable 503 without publishing a tenant. Public detail omits creator identity; profile export includes the exact operator-created tenant view, while signed account erasure pseudonymizes creator provenance and persists a one-way resurrection fence without deleting the tenant organization. Focused route/export/store tests passed 140/140, real local PostgreSQL restart/redaction proof passed 1/1, the BFF type ratchet stayed at zero, and the official 33.0 MiB artifact passed startup and graceful shutdown against a fresh database with all 15 migrations plus Redis. There is no shipped UI caller for the onboarding endpoint, so no new browser flow is claimed. This closes the organization-onboarding snapshot partition within its per-process boundary, not the wider all-store inventory.

Tenant OneRoster apply/commit hardening through de1c284a6ffd457b5f6ef020882191c32fc14884 replaces process-first roster mutation plus a detached save with one strict schema-v1 candidate/save/publish authority. Admin apply, externally triggered pull, and roster bulk-operation rows now await the primary snapshot before returning success or recording a success audit. Failed persistence returns stable 503, publishes neither roster nor replay-prevention commit, and leaves the mutation queue usable. Deployable boot requires and hydrates the Postgres-backed sink before traffic; bounded legacy documents migrate canonically and remove caller-supplied unknown record fields. Conflicts do not fabricate empty tenant rosters, duplicate incoming user/enrollment identities are rejected, and pull-conflict audit rows retain only a count rather than raw username-bearing conflict text. Focused store/route/bulk/dry-run tests passed 35/35, real local PostgreSQL restart proof passed 1/1, and both affected real-HTTP Playwright files passed 8/8 against an isolated exact-source BFF and fresh database with all 15 migrations. A second live process hydrated two tenants and three commits from the same database and rejected a persisted replay with 409. The BFF type ratchet remained at zero, lint passed, and the production entry bundled at 33.6 MiB. OneRoster external ids and usernames are institution roster authority, but the store has no exact customer-auth subject link; user-account erasure/export therefore cannot infer ownership from usernames and is not falsely claimed. Tenant offboarding, distributed replica compare-and-swap, stale-record removal, real SIS fetching, and scheduler ownership remain open. This closes the local OneRoster apply/commit snapshot partition within those boundaries, not the wider all-store inventory.

Admin Telegram editorial-channel binding hardening through d5b3e8c1a2b38cc31688200f4d0682be778502e0 replaces process-first mutation and detached snapshot saves with one strict schema-v1 candidate/save/publish authority. Bind, crisis-suppression, and takedown routes now await the ordered primary snapshot; failed writes return stable 503, publish no volatile state, and leave the queue usable. Deployable boot requires and hydrates the Postgres-backed sink before traffic, writes canonical seeds on first boot, migrates and minimizes legacy rows once, and rejects malformed or unsupported documents. Domain/pipeline mismatches and case-insensitive handle duplicates are rejected. Focused store/route tests passed 22/22, real local PostgreSQL restart proof passed 1/1, and the affected admin Chromium journey passed 2/2 against an isolated exact-source BFF and fresh database with all 15 migrations. A second BFF process recovered a separately persisted binding through real HTTP before its acknowledged cleanup. BFF/admin type gates, lint, formatting, and the 33.1 MiB production entry bundle passed. This is shared operator configuration with no customer-auth subject or Telegram credential; no account DSAR/erasure claim is made. The snapshot is not a durable rationale/audit ledger, does not invoke a Telegram provider, and serializes only per BFF process. This closes the local admin Telegram binding partition within those boundaries, not the wider all-store inventory.

Operator persona-lifecycle authority hardening through d145bc035f60448c688e79d458b9f71178587905 now commits the complete status and validator-produced immutable signoff envelope before a transition is visible or acknowledged. Production requires and hydrates the schema-v1 partition before traffic; failed saves return stable 503, publish no status or success audit, and leave the per-process mutation queue usable. Strict recovery bounds the snapshot to 10,000 personas, 10,000 audit events per persona, and 16 MiB; rejects unknown fields, duplicates, broken audit chains, status/timestamp drift, and future versions; migrates/minimizes legacy documents once; reconciles newer catalog seeds; and does not rewrite a current complete document. The route then awaits the separately durable admin search/explorer projection when available and truthfully reports it as degraded if that secondary append fails after the primary commit. Focused store/route tests passed 76/76, real PostgreSQL recovery passed 1/1 on a fresh connection, exact-bundle real-HTTP Playwright passed 5/5, and the unchanged operator console passed 7/7 browser/DOM/accessibility cases on a fresh 15-migration database. A second exact-bundle process recovered the same approved-for-test state, 31 audit entries, and update timestamp. The BFF type ratchet remained at zero, changed-source lint/format passed, and the production entry bundled at 33.2 MiB. Serialization remains per process with no distributed compare-and-swap; the separate admin projection is not cross-store atomic; no customer-subject mapping, DSAR/erasure, time-based audit purge, provider effect, native, Unreal, or model/media claim is made. This closes the local operator persona-lifecycle partition within those boundaries, not the wider all-store inventory.

Payment-time chargeback-screening authority hardening through bc7cf14ab3811627b8520265014cf3bcda7d9968 replaces fire-and-forget snapshot writes with a strict schema-v1 candidate/save/publish queue. The Telegram invoice trigger now awaits the screening commit before the record can appear in the admin console; a failed save publishes no screening and leaves the invoice response additive/fail-safe. Production requires and hydrates the sink before traffic. Recovery bounds 100,000 screens, 100,000 redaction anchors, and 64 MiB; rejects unknown fields, duplicate/noncanonical ids, invalid risk/numeric/text state, privacy contradictions, and future versions; and rewrites valid legacy { screens } rows once. Exact-id billing-receipt export now includes raw screens. Profile erasure pseudonymizes both user and transaction ids, persists a no-resurrection fence, and connects the pseudonymous rows to the canonical 2,555-day billing-history retention executor with legal-hold and exception blocking. Focused store/Telegram/admin/export tests passed 70/70; connected privacy/contract tests passed 16/16; real PostgreSQL screening, restart, redaction, purge, and fence recovery passed 1/1; and the exact-bundle Playwright journey passed 2/2 on a fresh database with all 15 migrations. Production smoke, the zero-backlog BFF type ratchet, lint/format, and the 33.2 MiB bundle passed. A second exact-bundle process recovered the same deterministic transaction, subject, amount, risk band, and timestamp. The queue remains per process with no distributed compare-and-swap; the Telegram trigger's caller-supplied user id is not proven to be a customer-auth identity; the screen is a proposed-payment risk assessment rather than provider settlement or chargeback outcome; and no non-Telegram rail, external fraud provider, native, Unreal, or model/media claim is made. This closes the local chargeback-screening snapshot partition within those boundaries, not the wider all-store inventory.

Crypto-invoice authority hardening through b8f31c4d344131bf7c155175ba62fc1051cba853 replaces process-visible, fire-and-forget issuance/confirmation writes with a strict schema-v1 detached- candidate/save/publish queue. Production requires and hydrates the partition before retention or traffic; failed issuance returns stable 503 with no invoice, and failed settlement exposes no status, receipt, or support disposition. Settlement facts, the signed Ed25519 receipt, and support flag now commit atomically. Durable row facts replace the restart-lost replay guard: exact transaction/block/time redelivery is idempotent across process restart, while changed facts return 409. The separately durable customer-auth grant is reconciled through an acknowledgement-bound invoice marker and safe replay. Recovery bounds 100,000 invoices, 100,000 redaction anchors, and 64 MiB; rejects unknown/future/malformed/noncanonical/privacy-inconsistent state; and rewrites valid legacy { invoices } documents once. Exact authenticated-purchaser billing export now includes the ledger. Profile erasure pseudonymizes the purchaser, removes optional deep links, persists a no-resurrection fence, and connects financial evidence to the canonical 2,555-day hold/exception-aware billing_history purge. Expiry is a pure projection; only stale unconfirmed intents prune, and confirmed evidence is never discarded for capacity. Focused store/quote/settlement/export tests passed 83/83, real route tests 11/11, connected privacy/contract tests 53/53, real PostgreSQL recovery/redaction/purge passed 1/1, and the exact-bundle browser journey passed 11/11. The zero-backlog BFF type ratchet, lint/format, 33.2 MiB build, and fresh-database production smoke with all 15 migrations and 15 durable snapshot rows passed. A second exact-bundle process recovered the same confirmed invoice, transaction, block, receipt, and audit-key facts through real HTTP. The whole-snapshot queue remains per process with no distributed compare-and-swap; invoice/customer-auth updates are reconciled rather than atomic; provisioning/watching remain deploy-bound; historical receipt verification lacks a key registry; anonymous invoice-id capability reads and late-payment product policy remain explicit boundaries; and no provider deletion/refund/dispute, rate-oracle, live-chain, native, Unreal, or model/media claim is made. This closes the local crypto-invoice snapshot partition within those boundaries, not the wider all-store inventory.

This is progress evidence only: the completed all-store disposition, deletion/privacy fanout for other newly durable user-scoped partitions, and any provider-native remote asset without a configured bucket/key deletion identity remain open. The previously named customer message-center and newly discovered device push-token, customer channel-binding, per-user Telegram, celestial ritual-schedule, BFF wearable-health, morning-capture, and Living Scene kept- artifact/share partitions are closed. The tenant SSO configuration partition is also closed within its stated per-process snapshot boundary. The tenant invitation, membership, and seat-capacity partition is now closed within the same per-process snapshot boundary and its stated retention/audit limitations. The tenant organization-onboarding authority is also closed within its stated single-process, local-provenance, and no-external-provisioning boundary. The local OneRoster apply/commit authority is now also closed within its stated per-process, institution-identity, additive-reconciliation, and externally triggered-pull boundary. The admin Telegram editorial-channel binding authority is closed within its per-process configuration-only boundary. The operator persona-lifecycle authority is closed within its per-process governance-state and non-atomic secondary-projection boundary. The payment-time chargeback screening authority is closed within its per-process, exact-identifier, Telegram-only, proposed-payment-risk boundary. Any other unclassified user-scoped operational partition still requires disposition. The crypto- invoice ledger is now also closed within its per-process snapshot, non-atomic entitlement-reconciliation, deploy-bound provider, and hold-aware financial- retention boundaries.

Model-governance license/admission authority hardening through source chain 250f2fcc66a629eb1210d8aea3baeeb5c84d3d1c and published correction 3f30e0aa302ab2a720b55516c47db80c385d497b now persists uploaded license declarations and complete minimized commercial/IP decisions together. Upload plus initial admission is one acknowledgement-bound candidate; failed saves publish neither side and return controlled 503; production requires and hydrates the authority before traffic. The strict schema-v1 store bounds 10,000 models, 100,000 decisions, and 64 MiB, fails loudly on invalid recovery, canonicalizes legacy state, avoids PostgreSQL JSONB key-order rewrites, and never persists raw prompt/output descriptions. Focused store and route tests passed 37/37, real PostgreSQL passed 1/1, the zero-backlog BFF type ratchet, lint/format, and a 33.3 MiB production build passed. A fresh-database smoke applied all 15 migrations and observed 16 snapshot rows; signed production-auth Playwright seed/recover runs proved version 2, one model, and one complete decision survive a clean exact-bundle restart without a boot rewrite or prompt email residue. The queue remains per process with no distributed CAS or atomic model/deployment transaction; license declarations remain attestations; no connected visual admin workflow, automatic archive, operator-subject privacy mapping, external rights provider, native, Unreal, or media-quality claim is made. This closes that local snapshot partition within those boundaries, not the wider all-store inventory.

Authenticated Iris desktop-to-mobile handoff authority hardening through 93d0a355339e7e42db8a14d5bd97da3ee6b7ab6c replaces process-first mutation and detached snapshot saves with a strict schema-v1 candidate/save/publish queue. Production requires and hydrates the sink before traffic; PUT, retention-touching GET, consume DELETE, retention purge, and exact-subject erasure acknowledge only after durable commit. Recovery is canonical and bounded, migrates the validated legacy document once, rejects malformed/future or resurrection state, enforces exact-contract expiry plus 30-day continuity retention, exports the exact subject, and composes signed erasure with a digest- only recreation fence. The shared mobile JSON client now omits content type for bodyless DELETE so Fastify reaches the consume route. Focused BFF tests passed 54/54, mobile tests passed 26/26, real PostgreSQL restart/export/erase proof passed 1/1, both type gates and targeted lint/format/docs validation passed, and the production entry bundled at 33.3 MiB. A fresh database applied all 15 migrations and initialized 17 snapshot rows. Three clean exact-bundle processes with signed production auth proved seed at version 4, recover/consume at version 7, and consumed absence at version 8 while preserving an adjacent subject. The mailbox remains a per-process, whole-snapshot latest record with no distributed CAS, device target/receipt, awaited native exactly-once consume, persisted mobile-to-web return, merge protocol, or bounded-fence archive policy. This closes that local handoff partition within those boundaries, not the wider all- store inventory.

Authenticated Studio authoring-record hardening through 6d1b0d3983a81b64cad39eaf8744a92426ddbc95 closes four previously process-local world-configuration, story-graph, quest, and timeline draft partitions. Their authenticated GET, POST, and DELETE paths now share strict schema-v1 candidate/save/publish authorities, require and hydrate Postgres-backed sinks before production traffic, and return controlled persistence, erasure-fence, and capacity failures without publishing volatile candidates. Exact-owner continuity_state export includes complete records plus server creation/access timestamps; signed continuity erasure physically removes the exact subject from all four snapshots, preserves adjacent owners, persists four one-way SHA-256 fences, and rejects stale recreation across restart. Startup and supervised daily purges enforce the canonical 30-day inactivity window, with authenticated reads durably refreshing access. Focused route/store/export/fanout tests passed 91/91, real local PostgreSQL restart/export/erase proof passed 1/1, the BFF type ratchet remained at zero, lint/format/conventions passed, and the rebased production entry bundled at 33.3 MiB. A fresh database applied all 15 migrations, and three clean exact-bundle processes with signed production auth and signed Redis deletion fanout proved seed, downloadable DSAR bytes, cross-owner 403, restart recovery, completed erasure, adjacent preservation, four durable fences, and second-restart stale-write rejection across 43 snapshot rows. The APIs have no demonstrated first-party web consumer; each whole-snapshot queue remains per process with no distributed compare-and-swap or cross-partition transaction; bounded fences have no archive/identity-reuse policy. This closes those four local partitions within those boundaries, not the wider all-store inventory.

Authenticated Mentor Presence rendered-performance hardening through 1b7533f36846528641f8a1da36c6b9eb26d1a5e7 replaces the process-local APNG map with a strict schema-v1 candidate/save/publish authority. Production now requires and hydrates the Postgres-backed sink before traffic; a rendered-media POST returns 201 only after its exact APNG bytes and SHA-256 commit, and media reads require the exact authenticated owner, return cross-owner 404, and use private immutable caching. Recovery is canonical and bounded, and exact-owner generated_artifact export includes the media bytes and metadata. Signed generated-artifact erasure physically removes the exact subject, preserves adjacent owners, persists a one-way SHA-256 resurrection fence, and rejects stale recreation across restart. Startup and supervised weekly purges enforce the canonical 365-day creation-time window while active legal holds and category retention exceptions defer physical deletion; expired held media is not served. Focused store/route/export/fanout tests passed 76/76, real PostgreSQL restart/hold/erase/fence proof passed 1/1, the BFF type ratchet remained at zero, lint/format/conventions passed, and the rebased production entry reproduced at 34,960,210 bytes with SHA-256 6a28093c8b731dda242167904b0e32297fee19262d0f18bbf3bb27a3f391eb26. A fresh database applied all 15 migrations, and two clean exact-bundle processes proved committed creation, exact 332,722-byte owner media, adjacent-owner 404, private cache policy, and byte/ETag-identical restart recovery. The store remains a per-process whole-snapshot queue with no distributed compare-and-swap, sharding, multi-region ordering, object-store offload, listing/revocation UI, or production provenance-signer/embodiment proof. This closes that rendered-performance partition within those boundaries, not the wider all-store inventory.

Authenticated Lilith Studio Tara scene-publish hardening through 343b87fe359270d1ad756050f2f67a1ffebd52a4 replaces the unowned process-local marker map and resettable sequence with a strict schema-v1 owner authority. The editor now submits the exact manifest JSON already rendered in its workspace; a 201 waits for the complete receipt and manifest to commit, and a stable domain-separated identity makes exact retries idempotent. Production requires and hydrates the Postgres-backed sink before traffic. Published-list reads are exact-owner only and expose the manifest SHA-256 without returning raw editorial JSON. Exact-owner generated_artifact DSAR includes those original bytes, while signed erasure physically removes the subject, preserves adjacent owners, persists a one-way SHA-256 resurrection fence, and rejects stale retries after restart. Startup and supervised weekly purges enforce the canonical 365-day creation-time window; active legal holds or category exceptions defer physical deletion, but expired held receipts remain absent from ordinary reads and DSAR. Focused store, route, export, and erasure tests passed 53/53; real PostgreSQL restart/hold/erase/fence proof passed 1/1; BFF and web type ratchets remained at zero; and the existing first-party Playwright publish journey passed 5/5 against the exact production BFF bundle. Two clean processes recovered the same stable receipt and manifest digest, and the rebased bundle reproduced at 34,982,194 bytes with SHA-256 68c46219f13f2f7dafb25632f82924ef0cb2146d44085569469c22aebcfaeba4. This is a durable receipt, not an approved customer release: no editorial-review identity, audit event, supersede/withdraw workflow, or /tara//lilith consumer is claimed. The bounded whole-snapshot queue remains per process with no distributed compare-and-swap. This closes this local partition within those boundaries, not the wider all-store inventory.

Authenticated Isis resource-recommendation hardening through e1a7ae32087b1b17cd3d37cef9df7a3afd791dd7 separates the immutable, explicitly estimate-only fallback catalog from an initially empty normalized PostgreSQL history. Recommendation ids are cross-replica idempotency keys; each read recomputes and verifies the exact catalog projection, revision, attribution, ordering, and aggregate. POST derives operator identity from auth, rejects caller attribution and all customer/source/content fields, and acknowledges only after commit. Exact-subject profile erasure pseudonymizes seven-year internal review_package rows behind a permanent digest fence; the shared legal-hold/retention-exception-aware executor physically purges due history. Focused BFF tests passed 12/12, persistence schema tests 2/2, real PostgreSQL restart/race/erasure/hold/corruption proof 2/2, web tests 11/11, and isolated desktop/mobile Playwright plus axe passed 8/8. BFF/web/persistence typechecks, lint, format, Prisma validation/generation stability, all 27 fresh migrations, direct production-entry compilation, and current-source runtime readback also passed. The surface records configured cost-governance recommendations only; it does not claim current provider pricing, model switching, budget enforcement, runtime fallback, release authority, or customer-owned records. This closes that local partition within those boundaries, not the wider all-store inventory.

Authenticated Isis intelligent-routing hardening through 1f90b287735cf6087e64aec76a999b4317891036 now treats the same immutable isis-resource-recommendation-catalog-v1 fallback ladder as configured, estimate-only governance input and stores only observed operator routing decisions in an initially empty normalized PostgreSQL history. Routing ids are cross-replica idempotency keys; every read validates and recomputes the exact catalog projection, revision, newest-first order, attribution, and aggregate. POST derives routedBy from the authenticated principal, rejects caller attribution and customer/source/content fields, and acknowledges only after the complete transaction commits. Exact-subject profile erasure pseudonymizes the seven-year internal review_package history behind a permanent digest fence; the shared legal-hold/retention-exception-aware executor physically purges due rows. Focused BFF tests passed 12/12, persistence schema tests 2/2, real PostgreSQL restart/replica/race/erasure/hold/corruption/rollback proof passed 2/2, web tests passed 10/10, and isolated desktop/mobile Playwright plus axe passed 8/8. BFF/web/persistence typechecks, strict lint and formatting, all 28 fresh migrations, and a 35,880,273-byte direct production-entry compilation also passed. The primary Studio route is mounted; four adjacent destinations remain explicitly documented contracts with no mounted page. This surface does not claim live provider price or health, model execution, budget enforcement, runtime fallback, release authority, or customer-owned routing records. This closes that local partition within those boundaries, not the wider all-store inventory.

Authenticated Isis cost-tracking hardening through 8bab273a22d48801748efc21c7ed062166824288, published and finally verified through ae490aa4fb500ad0c6e471999eb5ed70b2f34727, now separates an immutable 12-model configured pricing catalog (isis-token-budget-pricing-v1) from an initially empty normalized PostgreSQL history. Strict opaque record ids are cross-replica idempotency keys; every read validates and recomputes catalog revision, rates, costs, attribution, newest-first ordering, and aggregate totals. POST derives recordedBy from the authenticated principal, rejects caller attribution, customer/source/content fields, precomputed values, and unknown inputs, and acknowledges only after the complete transaction commits. Exact-subject profile erasure pseudonymizes the seven-year internal review_package history behind a permanent digest fence; the shared legal-hold/retention-exception-aware executor physically purges due rows. Focused BFF tests passed 14/14, persistence schema tests 2/2, real PostgreSQL restart/replica/race/erasure/hold/corruption/rollback proof passed 2/2, web tests passed 9/9, and isolated desktop/mobile Playwright plus axe passed 8/8 against all 30 fresh migrations. BFF/web/persistence typechecks, strict lint and formatting, Prisma validation/generation stability, and a 35,987,898-byte direct production-entry bundle with SHA-256 4e021928bfb668ed28f8220f7c8bc51c07e49d0ee31a4b7a43ac8b802f476fa2 also passed. The primary Studio route is mounted; four adjacent destinations remain explicitly documented contracts with no mounted page. This surface records configured estimates only; it does not claim live provider prices, invoices, health, execution, budget enforcement, runtime fallback, release authority, or customer-owned records. This closes that local partition within those boundaries, not the wider all-store inventory.

Authenticated Isis cost-optimization hardening through 9dc92a2905b182d4e4bcca11393ca1b0b0e6d43e removes the fabricated seeded run history and process-local retention of prompt-derived previews. The real deterministic optimizer is now explicitly ephemeral_recomputable, with persistence none, request-lifetime retention, immutable catalog revision isis-token-budget-optimizer-v1, and no authoritative state. GET always returns an empty history; POST validates a strict machine-token correlation id, exact catalog technique, bounded content and optional controls, rejects unknown fields, emits Cache-Control: no-store, and returns the result without retaining input, output, identity, or metrics. Repeat ids are safe because they correlate a request rather than identify a write. The Studio surface strictly validates the no-retention contract, keeps only the latest current-tab preview, visibly loses it on reload, and labels all four adjacent destinations as documented contracts without implying mounted pages or authority. Focused BFF tests passed 8/8, web tests passed 9/9, and isolated desktop/mobile Playwright plus axe passed 7/7 against all 32 fresh migrations. BFF/web typecheck ratchets, strict lint, formatting, affected Nx typechecks, and a 35,985,040-byte direct production-entry bundle with SHA-256 8e0d832b830ecd320e6f12283f3301e1fcf39a4cb82bb6879d595764df11c31a also passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim live provider pricing or health, model execution, budget enforcement, durable optimization history, learning, fallback, release authority, or customer-owned records. This closes the former local map within that explicit ephemeral boundary, not the wider all-store inventory.

Authenticated Isis prompt-template preview hardening through b876a5ce0e393ef732ea32a7af899e9daf9cef5a removes two fabricated seeded templates and process-local retention of operator prompt source, variables, rendered output, and metrics. The real prompt engine is now explicitly ephemeral_recomputable, with persistence none, request-lifetime retention, engine revision isis-prompt-template-preview-v1, and no authoritative state. GET always returns an empty history; POST validates a strict machine-token correlation key, bounded name and content, and a bounded JSON variable graph, rejects unknown identity/retention fields and prototype traversal, emits Cache-Control: no-store, and returns the preview without retaining any input or output. Accepted nested objects are normalized to null-prototype records. Repeat keys are safe because they correlate a request rather than identify a write. The Studio surface strictly validates the no-retention contract, keeps only the latest current-tab preview, visibly loses it on reload, and labels all four adjacent destinations as documented contracts without implying pages or authority. Focused BFF tests passed 8/8 on the published tree, web tests passed 9/9, and isolated desktop/mobile Playwright plus axe passed 8/8 against all 32 fresh migrations; the post-input-hardening live subset passed 2/2. BFF/web typecheck ratchets, strict source/test/E2E lint, formatting, affected Nx typechecks, and a 36,093,443-byte direct production-entry bundle with SHA-256 c3e0456b478310ecadca18ffcd3fdb68d2af4d62ce7c7cddc2d35b67a9e46f34 also passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim a reusable-template repository, durable prompt history, approvals, lineage, provider execution, learning, release authority, or customer-owned records. This closes the former local map within that explicit ephemeral boundary, not the wider all-store inventory.

Authenticated Isis managed-model registry hardening through 5a47c333dd7a5117920a21966b16035dcdd9ef4f removes five fabricated approved model records and their invented license, hash, recipe, moderation, and release-like metadata. Repository tracing found no production intake/admission producer or persistence adapter, so the real boundary is now explicitly unbound_empty_projection, with persistence none, producer none, retention not_applicable, authority binding not_configured, and no authoritative state. Every valid browse is empty across new instances; the authenticated route rejects unknown or malformed query fields and emits Cache-Control: no-store. The Studio parser fails closed if the service claims any record, asset type, cursor, nonzero total, or authority; the UI leads with the missing authority, reports zero admitted models, and hides filters and pagination until a real registry is bound. Focused BFF tests passed 6/6, web tests passed 8/8, and isolated desktop/mobile Playwright plus axe passed 8/8 against all 32 fresh migrations. Both app typecheck ratchets returned zero backlog; strict source, test, and E2E lint, formatting, affected Nx typechecks, syntax validation, and a 36,333,874-byte production-entry bundle with SHA-256 b2040f174d504c828362f7686a353530f4c1ea8e29685b2bc5c6207c0bde6ecc also passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim model intake, admission review, license verification, artifact hash validation, provider availability, recipes, approvals, customer-owned records, or release authority. This closes the fabricated local catalog within that explicit unbound boundary, not the wider all-store inventory.

Authenticated Isis Civitai intake hardening through a2e8e0d69d1fb4bc667c10594b884b27b85f747a removes a fabricated search candidate, queued candidate, reviewer history, license/hash/lineage claims, and process-local approval pipeline. Repository tracing found no production search or intake producer, reviewer directory, durable admission store, or downstream registry binding. GET now exposes an exact no-store unbound_empty_projection: zero queue entries, null operator, zero search candidates, persistence and producer none, authority binding not_configured, decisions unavailable, and no authoritative state. The authenticated decision endpoint preserves the Isis scope gate but returns 503 intake_authority_not_configured for every authorized body and cannot mutate or acknowledge approval, rejection, takedown, or promotion. The Studio parser rejects any candidate, reviewer, history, or authority claim; its restrained boundary view hides all search, filter, metadata, and decision controls. Focused BFF tests passed 7/7, web tests passed 8/8, and isolated desktop/mobile Playwright plus axe passed 8/8 against all 32 fresh migrations. Both app typecheck ratchets returned zero backlog; strict source/test/E2E lint, formatting, affected Nx typechecks, syntax validation, and a 36,326,165-byte production-entry bundle with SHA-256 9506d12ce42e4d39a1cc640d00673dd40945826631c682c65c2797c5d9e5ab82 also passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim remote discovery, candidate intake, reviewer identity, rights or license verification, hash validation, moderation, promotion recipes, model registration, provider effects, or release authority. This closes the fabricated local intake/review seam within that explicit unbound boundary, not the wider all-store inventory.

Authenticated Isis model-merging hardening through 8dca9a9e33996b791085b663b59d46ddcffbd0a6 removes five fabricated model components, three fixture sets, an invented operator, process-local saved plans, and a provenance-shaped acknowledgment with no execution effect. Repository tracing found no production component catalog, fixture provider, artifact writer, merge executor, durable plan store, or downstream binding. GET now exposes an exact no-store unbound_empty_projection: null operator, zero components, zero fixtures, zero plans, persistence and producer none, authority binding not_configured, plans and execution unavailable, and no authoritative state. The authenticated plan endpoint preserves the Isis scope gate but returns 503 model_merging_authority_not_configured for every authorized body and cannot mutate or acknowledge a plan. The Studio parser rejects any component, fixture, saved-plan, or authority claim; its restrained boundary view hides all selection, ratio, fixture, planning, and execution controls. Focused BFF tests passed 7/7, web tests passed 8/8, and isolated desktop/mobile Playwright with applicable axe coverage passed 7/7 against all 32 fresh migrations. Both app typecheck ratchets returned zero backlog; strict source/test/E2E lint, formatting, affected Nx typechecks, syntax validation, and a 36,319,322-byte production-entry bundle with SHA-256 6d7e67f6fc4469f1aff629bfd3f7aeb7432e1cec8a46afd99a5539ca72c2d869 also passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim model admission, component integrity, fixture provenance, merge execution, artifact writing, quality comparison, provider effects, approval history, or release authority. This closes the fabricated local merge seam within that explicit unbound boundary, not the wider all-store inventory.

Authenticated Isis voice-cloning hardening through 73d0f24553b3fb596c97d0374973b35542b0490d removes two fabricated workflows, named consent acknowledgements, identity-verification timestamps, biometric sample and script hashes, quality and abuse-risk scores, watermark claims, reviewer signoffs, lifecycle history, and a released production profile. Repository tracing found no production consent ledger, identity verifier, biometric sample store, scoring producer, signer directory, profile registry, provider executor, or downstream TTS release binding. GET now exposes an exact no-store unbound_empty_projection: zero records, persistence and producer none, authority binding not_configured, every workflow, transition, identity, biometric, scoring, signoff, and profile capability unavailable, and no authoritative state. The authenticated create and transition endpoints preserve the Isis scope gate but return 503 voice_cloning_authority_not_configured for every authorized body and cannot retain or acknowledge consent-shaped or biometric data. The Studio parser rejects any workflow, consent, biometric, signoff, profile, or authority claim; its restrained boundary view hides all subject, consent, identity, sample, scoring, signoff, registration, release, and revocation controls. Focused BFF tests passed 7/7, web tests passed 8/8, and isolated desktop/mobile Playwright with applicable axe coverage passed 7/7 against all 32 fresh migrations. Both app typecheck ratchets returned zero backlog; strict source/test/E2E lint, formatting, affected Nx typechecks, post-rebase BFF typechecking, syntax validation, and a 36,342,561-byte production-entry bundle with SHA-256 523786b2c0cbbf2606d11309f27a1ee2220d7be14e5c589bcfc494d0f23054f7 also passed. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim consent validity, identity verification, biometric custody, sample integrity, scoring, watermark or abuse-risk verification, reviewer authority, profile registration, provider effects, revocation propagation, or release authority. This closes the fabricated local voice-workflow seam within that explicit unbound boundary, not the wider all-store inventory.

Authenticated Isis ComfyUI-node hardening through facf0d71effd58f6ccbbcac060d653fd106ce14a replaces the hand-authored 10-node subset and removes invented workflow-usage, deprecation, migration-alternative, and process-local mutation claims. Repository tracing found a canonical source-defined registry in libs/yemaya/comfyui-integration/src/node-registry/node-registry.ts, but no runtime discovery binding, workflow-usage index, durable deprecation authority, or operational mutation store. GET now exposes an exact no-store source_defined_projection recomputed from all 44 canonical definitions, with source-control persistence, source-defined schema authority, runtime discovery, workflow usage, deprecation decisions, and mutations unavailable, authoritative schema state true, and authoritative operational state false. The authenticated deprecation compatibility endpoint preserves the Isis scope gate but returns 503 comfy_node_governance_authority_not_configured for every authorized body and cannot mutate or acknowledge a governance decision. The Studio parser requires exactly 44 unique full schemas and the exact disposition, rejects any operational or authority claim, renders the complete catalog read-only, and hides all search, usage, filtering, deprecation, and mutation controls. Focused BFF tests passed 7/7, the canonical registry tests passed 293/293, web tests passed 8/8, and isolated desktop/mobile Playwright with applicable axe coverage passed 7/7 against all 32 fresh migrations. Both app typecheck ratchets returned zero backlog; strict source/test/E2E lint, formatting, affected Nx typechecks, bundle syntax validation, and production-artifact startup smoke validation also passed. The verified BFF entry bundle is 35,781,696 bytes with SHA-256 775c044a8fa26844557639b87c4ad1709592c5a3ec248737f78f3ee779933aa6. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim runtime installation, provider availability, workflow use, execution, deprecation approval, migration readiness, governance history, customer-owned state, or release authority. This closes the fabricated local operational layer while preserving the real source-defined schema authority, not the wider all-store inventory.

Authenticated Isis RunPod-operations hardening through f640bac1278ed6d3d55f59338024e81e9fb3b2d0 removes the fabricated two-endpoint inventory and health rows, tenant cost and budget records, queued jobs, failover history, operator identity, key identifiers, completed rotation history, and process-local rotation state. Repository tracing found no bound endpoint-inventory, runtime-health, billing, queue-telemetry, failover-history, or key-governance producer. The existing RunPod generation executor can submit and poll generation work from deployment configuration, but is not operational inventory, billing, queue, failover, key-custody, or rotation authority. GET now returns an exact no-store unbound_empty_projection: every collection is empty, queue capture time and runtime/operator/time scalars are null, persistence and producer are none, authority binding is not_configured, every operational capability and mutation is unavailable, and authoritative state is false. Both authenticated secret-rotation compatibility endpoints preserve the Isis scope gate but return 503 runpod_operations_authority_not_configured for every authorized body and cannot create or transition rotation state. The Studio parser accepts only that exact empty context and disposition, rejects fabricated operational or authority claims, presents the boundary and route map read-only, and hides endpoint, filtering, budget, queue, failover, and rotation controls. Focused BFF tests passed 7/7, web tests passed 8/8, and isolated desktop/mobile Playwright with applicable axe coverage passed 7/7 against all 32 fresh migrations. Both app typecheck ratchets returned zero backlog; strict source/test/E2E lint, formatting, affected Nx typechecks, bundle syntax validation, and production-artifact startup smoke validation also passed. The verified BFF entry bundle is 35,773,380 bytes with SHA-256 609b6b6928e2988b1eba1fa5cf9502429740766d3682498d7464cd9765722b1c. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim endpoint inventory, provider availability, runtime health, cost or budget records, queue telemetry, failover history, key custody, rotation history, audit history, customer-owned records, or release authority. This closes the fabricated local RunPod operational projection within its explicit unbound boundary, not the wider all-store inventory.

Authenticated Isis LoRA-training hardening through 91db1ea6e142874dd5298e5f796d8615f4708300 removes the fabricated two-run queue, tenant/operator and cost configuration, dataset-rights attestations, checkpoint previews and metrics, lifecycle histories, promotion record, and process-local create/transition state. Repository tracing found pure lifecycle, planning, and orchestration models in @isis/lora-training-surface, libs/nous/training, and libs/calliope/visage, but no bound training-run ledger, dataset-rights authority, training executor, checkpoint collector, evaluation authority, or promotion authority behind this BFF surface. GET now returns an exact no-store unbound_empty_projection with zero runs, persistence and producer none, authority binding not_configured, run inventory, dataset-rights authority, execution, checkpoint telemetry, evaluation, promotion, lifecycle history, and mutations unavailable, and authoritative state false. Both authenticated lifecycle compatibility POST paths preserve the Isis scope gate but return 503 lora_training_authority_not_configured for every authorized body and cannot create or transition a run. The Studio and admin parsers independently require the exact empty payload and disposition, reject any run or authority claim, render only a restrained read-only boundary, and hide create, search, filter, attestation, abort, evaluation, and promotion controls. Focused BFF tests passed 13/13, web tests passed 10/10, and admin tests passed 28/28. Isolated desktop/mobile Studio Playwright with applicable axe coverage passed 7/7, and isolated admin Playwright with axe coverage passed 1/1; the Studio stack used all 32 fresh Oshun migrations and both fresh Isis migrations. The BFF and web typecheck ratchets returned zero backlog, the admin full typecheck passed, and the commit hook's four-project affected typecheck passed. Strict source/test/E2E lint, formatting, bundle syntax validation, and migrated production-artifact startup/API smoke validation also passed. The verified BFF entry bundle is 35,898,498 bytes with SHA-256 3047c712912d7322cccc91b6972076faf40a915e76133e6bcd843f240b3164c6. The web and admin production builds were not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim a training queue, dataset ownership or rights decisions, training execution, checkpoint or cost telemetry, evaluation results, promotion or managed-model registration, audit history, customer-owned records, or release authority. This closes the fabricated local LoRA-training operational projection within its explicit unbound boundary, not the wider all-store inventory.

Authenticated Isis suspicious-activity hardening through 422b1047f38f01824d94a9c98cae6c6c32966f68 removes the two seeded transactions, process-local mutable map, and manual-screening second authority. Repository tracing found an existing durable chargeback-screening snapshot ledger produced at Telegram invoice payment time, acknowledged by the billing receipt retention boundary and already covered by subject export and exact- subject redaction. GET now exposes only a no-store, read-only projection of that ledger with exact disposition durable_read_only_projection, persistence durable_snapshot_store, producer telegram_invoice_payment_time, retention billing_receipt_2555_days, authority binding chargeback_screening_financial_records, screening inventory and payment-time ingestion available, scoring revision unrecorded, factor evidence, manual screening, provider settlement, and chargeback outcome unavailable, mutations unavailable, and authoritative state true. The authenticated compatibility POST preserves the Isis scope gate but returns 503 suspicious_activity_manual_screening_not_configured for every authorized body and cannot create or change screening state. The Studio parser requires the exact disposition, screening schema, stable newest-first transaction order, unique transaction identifiers, and recomputed aggregate counts and mean risk; it rejects fabricated claims and renders only a restrained read-only boundary without manual-screening controls. Focused BFF route tests passed 5/5, direct web tests passed 8/8, and the embedded parent anomaly/GPU workspace tests passed 8/8. Isolated desktop/mobile Playwright with applicable axe coverage passed 8/8 against all 32 fresh Oshun migrations and both fresh Isis migrations. The BFF and web typecheck ratchets returned zero backlog; strict changed-source, test/E2E lint, formatting, bundle syntax validation, and migrated production- artifact lifecycle and exact signed API smoke validation also passed. The verified BFF entry bundle is 35,907,385 bytes with SHA-256 356cc25c44f05fa9b88fadba40e57b65f5abbeed4c8024695ec7ff88f68aea3b. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim manual or non-Telegram-rail screening, a recorded scoring revision, factor evidence, provider settlement, chargeback outcome, or release authority. This closes the fabricated local suspicious-activity projection by binding it to the existing payment-time financial-record authority, not the wider all-store inventory.

Authenticated Isis creator-monetization hardening through fe91672dc8decf95dc425d3f82d17463c1762b0c preserves the real deterministic @isis/3d-marketplace-ops collaborator split calculator while removing its false publication, payout-activation, governance-history, audit-artifact, and release-signoff presentation. Repository tracing found no plan ledger, collaborator-identity authority, terms-acceptance authority, rights authority, listing publisher, payout activator, or settlement provider behind this surface. GET and POST now return exact no-store contracts with disposition ephemeral_recomputable_planner, persistence none, producer request_time_pure_function, retention request_only, authority binding not_configured, source-defined role catalog and split planning available, plan persistence, rights authority, listing publication, payout activation, settlement, and mutations unavailable, collaborator identity unverified, terms acceptance self-asserted and unverified, and authoritative state false. The planner accepts only an exact bounded collaborator schema, rejects duplicate identifiers, extra fields or body authority, overlong identifiers or names, out-of-range percentages, and more than 100 collaborators, and returns no server-side record. The Studio starts with an empty local draft, requires the exact catalog, disposition, plan envelope, normalized-share schema, and unique identifiers, independently recomputes totals, status, warnings, and suggested lead, and labels the result as calculator guidance rather than payout or publication approval. Its route map now distinguishes the one mounted planner from four documented-only shapes and removes fabricated scene, export, revision, governance, and release capabilities. Focused BFF tests passed 8/8 and focused web tests passed 7/7 without React act warnings. All nine isolated Playwright product flows passed with applicable rendered-page axe coverage, including narrow mobile layout, authorization, loading and network failures, fabricated catalog and internally inconsistent plan rejection, live valid, manual-review and blocked calculations, and exact direct no-store API checks; the stack used all 32 fresh Oshun migrations and both fresh Isis migrations. Both app typecheck ratchets returned zero backlog, the commit hook's three-project affected typecheck passed, and strict source/test/E2E lint, formatting, bundle syntax validation, migrated production-artifact lifecycle smoke, and exact signed production API smoke validation also passed. The final rebased BFF entry bundle is 35,918,872 bytes with SHA-256 3a9672e7bf82c5f0468019e50e5bbc9c0888e11394f8dcc867bd5577358ec7b1. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim verified collaborator identity or terms, rights decisions, durable plans, listing publication, payout activation, settlement or payout outcomes, history, audit artifacts, customer-owned records, or release authority. This closes the overclaimed creator-monetization calculator as an explicit ephemeral/recomputable planner, not the wider all-store inventory.

Authenticated Isis budget-management hardening through b50ce7136ae4aedb376ebc3bc1a07a4e62712bc6 removes the four fabricated tenant budgets, hard-coded May 2026 cost window, invented caps, spend and projections, synthetic status rollup, and resettable process-local store. Repository tracing found a real pure @isis/runpod-surface budget evaluator, but no tenant budget- envelope authority, spend ingestion binding, projection producer, alert delivery, or kill-switch enforcement behind this route. GET now returns an exact no-store unbound_empty_projection with zero tenants and status counts, persistence, producer and retention none, authority binding not_configured, evaluator source_defined_only, budget inventory, budget-envelope authority, spend ingestion, projections, alert delivery, kill-switch enforcement, and mutations unavailable, and authoritative state false. Independent BFF app instances return the same empty contract and cannot resurrect the former seed rows. The Studio parser requires that exact empty payload and disposition, rejects seeded tenants, status claims, or fabricated enforcement authority, renders a restrained read-only capability boundary, and hides filtering, alerting, and enforcement controls. Its route map distinguishes the one mounted boundary from four documented-only shapes and removes fabricated scene, export, revision, governance, audit, and release capabilities. Focused BFF tests passed 3/3 and focused web tests passed 6/6. Isolated desktop/narrow-mobile Playwright with applicable rendered-page axe coverage passed 8/8 against all 32 fresh Oshun migrations and both fresh Isis migrations, including exact direct no-store API validation, authorization, seeded/fabricated response rejection, and loading, server, and network failure behavior. Both app typecheck ratchets returned zero backlog, the commit hook's three-project affected typecheck passed, and strict source/test/E2E lint, formatting, bundle syntax validation, migrated production-artifact lifecycle smoke, and exact signed production API smoke validation also passed. The final rebased BFF entry bundle is 35,917,006 bytes with SHA-256 a2acc62b771376d060bba0e4131c6c4f4d5bce006fd27f9c4ba30dc2a00bfed5. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim tenant budgets or envelopes, current spend, forecasts, alert delivery, kill-switch state or enforcement, provider billing, history, customer-owned records, or release authority. This closes the fabricated local budget- management projection within its explicit unbound boundary, not the wider all- store inventory.

Authenticated Isis 3D-retention-policy hardening through d59fef9a63a7474696d09d87c18ed584df349a71 removes the three fabricated artifact records, fixed May 2026 evaluation clock, process-local evaluation map, duplicate-record semantics, and false inventory-scan and deletion-operations presentation. Repository tracing found a real deterministic @isis/model-governance-3d policy builder and evaluator, but no bound artifact inventory or ownership authority, tenant policy owner, legal-hold authority, customer-export or benchmark evidence producer, deletion worker, or audit ledger behind this route. GET now returns the source-defined policy and artifact kinds with an exact empty records array and no-store disposition ephemeral_recomputable_evaluator, persistence none, producer request_time_pure_function, retention request_only, authority binding not_configured, evaluator source_defined_only, all operational authorities and mutations unavailable, caller export and benchmark facts self-asserted and unverified, and authoritative state false. POST strictly validates caller- supplied artifact facts and an optional evaluation clock, recomputes a preview without retaining it, and safely permits repeated correlation identifiers; the authenticated scan compatibility route returns exact 409 scan_unavailable and cannot imply an inventory or deletion pass. The Studio parser requires that exact empty context and disposition, rejects seeded records or fabricated authority, keeps only the latest request-local preview, and labels the source-defined rules as evaluator inputs rather than an approved tenant policy. Its route map distinguishes the one mounted evaluator from four documented-only shapes and removes governance-history, audit, and release claims. Focused BFF tests passed 5/5, focused web tests passed 8/8, and the domain evaluator tests passed 4/4. Nine isolated desktop/narrow-mobile Playwright flows passed with applicable rendered-page axe coverage against all 32 fresh Oshun migrations and both fresh Isis migrations, including exact live API contracts, deterministic repeat evaluation across instances, authorization, caller-assertion labeling, seeded/fabricated response rejection, malformed-preview and scan failure, and loading, server, and network failure behavior. Both app typecheck ratchets returned zero backlog, the commit hook's three-project affected typecheck passed, and strict source/test/E2E lint, formatting, bundle syntax validation, migrated production-artifact lifecycle smoke, and exact signed production API smoke validation also passed. The final rebased BFF entry bundle is 36,067,940 bytes with SHA-256 f16c4c8413048a073d73496ce44abd231b65aea42620ce73afa705cfae003082. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim an artifact inventory or ownership proof, approved tenant policy, legal-hold authority, verified export or benchmark facts, scan or deletion execution, audit evidence, retained decisions, customer-owned records, or release authority. This closes the fabricated local retention-policy store as an explicit request-only evaluator, not the wider all-store inventory.

Authenticated Isis pipeline-class and 3D-provenance hardening through e4cd1eed1f03a9a12371b70517919bfc197e0c39 removes two fabricated approved pipeline classes, their fixed May 2026 timestamps and rehearsal claims, the process-local authoring registry, the process-local sealed-bundle map, and the caller-asserted output, source, model, watermark, chain, and sealing authority. Repository tracing found real pure builders, validators, hashes, step/cost vocabularies, provider keys, and modes in @isis/three-d-pipelines, but no bound tenant pipeline-class catalog, approval authority, rehearsal-evidence producer, runtime registry, execution binding, output inventory or ownership authority, source/model/watermark evidence producer, chain observer, or bundle-sealing authority behind these compatibility routes. Both GETs now return exact no-store unbound_empty_projection contracts with empty operational collections, persistence and producer none, retention not_applicable, authority binding not_configured, source-defined vocabulary only, every operational capability and mutation unavailable or unverified, and authoritative state false. Both authenticated compatibility POSTs preserve the Isis scope gate but return exact 503 authority-not-configured contracts for every authorized body and cannot create a class or seal a bundle. The 3D route additionally identifies the real append-only generation-provenance ledger as separate_not_bound; that distinct ledger is not borrowed as evidence for this route. Both Studio parsers accept only the exact empty contracts, reject seeded rows, bundles, unknown fields, and fabricated authority, render restrained read-only boundary worksheets, hide all authoring and sealing controls, and distinguish one mounted route from four documented-only shapes. The provenance page links to the separate canonical generation-ledger inspector without embedding its generated-artifact panel. Focused BFF tests passed 6/6, focused web tests passed 17/17, and the underlying 3D-pipeline library tests passed 23/23. Nine isolated Chromium flows passed with applicable rendered-page axe coverage before and after rebasing over upstream shell/navigation changes; the fresh stack applied all 32 Oshun migrations and both Isis migrations and covered exact live GET/POST contracts, unsigned-token rejection in production smoke, authorization, malformed/hostile responses, loading/server/network failures, narrow-mobile overflow and 44-pixel affected controls, and anonymous redirect. Both app typecheck ratchets returned zero backlog, the commit hook's three-project affected typecheck passed, and strict source/test/E2E lint, formatting, migrated production-artifact startup, lifecycle, and exact signed API smoke validation also passed. The final BFF entry bundle is 36,056,997 bytes with SHA-256 09f737250a2a0787317513f7554e8841be9eaf88a07e5bdfcf9b7a18742eeb63. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim an approved pipeline-class inventory, rehearsal evidence, runtime registration or execution, output ownership, observed source/model/watermark evidence, chain observation, sealed 3D bundles, customer-owned records, or release authority. The separate generated-artifact provenance panel still used on other output and operator surfaces was not changed by this slice and remains part of the wider all-store/UI inventory. This closes the two fabricated local compatibility stores within their explicit unbound boundaries, not the wider provenance or pipeline-execution program.

Authenticated Isis output-lineage hardening through 62acc45ecdbbdeccbd8173ac6393398ab439a082 removes five fabricated output identifiers, four fixed May 2026 cross-domain edges, the process-local graph, and the false claim that generation, citation, lesson, passage, and scene events record those relationships upstream. Repository tracing found normalized generated_outputs and lineage_edges models with owner, organization, project, expiry, deletion, and edge-type fields in the Isis Prisma schema, but the production generation service never writes either model; only the development seed writes one sample output and it writes no edge. The generation service instead persists job output JSON and a separate append-only provenance ledger. @isis/output-gallery supplies a cycle-checked in-memory DAG and a pure branch helper, but no application producer binds that helper to durable, tenant-safe lineage authority. The GET now returns an exact no-store unbound_empty_projection with empty outputs and edges, persistence and producer none, retention not_applicable, authority binding not_configured, database schema declared_not_producer_bound, pure graph library source_defined_only, unavailable inventory, edge observation, descendant reach and takedown cascade, unverified ownership, separate unbound canonical generation ledger, unavailable mutations, and authoritative state false. Authenticated node lookup and POST compatibility routes return exact no-store 503 output_lineage_authority_not_configured contracts and cannot turn caller assertions into outputs, edges, neighborhoods, or cascade reach. The Studio parser accepts only the exact empty contract, rejects fabricated outputs, edges, missing disposition fields, and unknown authority, renders a restrained read-only boundary worksheet, hides output selectors and takedown controls, distinguishes one mounted route from four documented-only shapes, links to the separate canonical generation ledger, and provides 44-pixel navigation targets at narrow-mobile width. Focused BFF tests passed 3/3 and focused web tests passed 5/5. Seven isolated Chromium flows passed with applicable rendered-page axe coverage against a clean stack after all 32 Oshun and both Isis migrations; coverage proves the exact live GET, node and POST contracts, Studio hard boundary, scope denial, fabricated-response rejection, loading/server/network failure states, honest route language, narrow-mobile overflow and touch targets. Both app typecheck ratchets returned zero backlog, the commit hook's three-project affected typecheck passed, and focused lint, formatting, source build, and post-rebase production-artifact startup and exact HS256 signed API smoke also passed. The BFF build target now invokes its existing native-dependency bridge script instead of the stale raw esbuild command that could not start its bundled pg, Sharp, and C2PA dependencies. The final post-rebase BFF entry bundle is 36,063,968 bytes with SHA-256 b7e61fdc968d167e558bf2820625d069a0213d3726d589368fa8fc7035021831. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim observed lineage, tenant-safe output ownership, descendant reach, takedown execution, customer-owned records, retention or erasure coverage, or release authority. This closes the fabricated local output-lineage store within its explicit unbound boundary, not the wider provenance, output-catalog, cross-product lineage, or takedown-cascade program.

Authenticated Isis model-comparison hardening through 3f2e1f2a46d2db16cb77c95c89222b95e04abf8c removes two fixed development diff measurements and the false assertion that an offline comparison job produced them. It also removes Generation Gallery's separate fabrication of pixel, frame, waveform, and mesh measurements from output-cost differences. Repository tracing found a durable output-gallery snapshot populated by the generation output catalog and the real pure @isis/output-gallery layout and kind-compatibility builder, but no asset-byte reader, metric producer, metric persistence, or measurement authority. The authenticated POST now returns an exact no-store durable_catalog_recomputable_layout contract: selected records come from the durable gallery, layout is recomputed per request, retention and authority inherit from that catalog, source-defined kind compatibility remains active, and metric production, metric persistence, asset bytes, mutations, and authoritative measurements are explicitly unavailable. The Studio parser accepts only that exact disposition, the exact requested slot ordering, and an empty diff-pair list; it rejects fabricated measurements and unknown fields, renders a restrained boundary worksheet, links to the durable Generation Gallery, and labels the four unmounted compatibility shapes as documented contracts rather than live IAM, capture, rollback, or approval capabilities. Focused BFF tests passed 6/6, focused web tests passed 22/22, and the underlying output-gallery library tests passed 28/28. Seven isolated Chromium flows passed with applicable rendered-page axe coverage against a clean stack after all 32 Oshun and both Isis migrations; coverage proves the exact live no-store API, Studio hard boundary, durable-record layout, fabricated-measurement rejection, scope and network failures, and narrow-mobile overflow and 44-pixel controls. The web typecheck ratchet returned zero backlog. The full BFF compiler reported no model-comparison errors but remains red on two unrelated pre-existing autonomous-content errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts; the commit used the hook's documented typecheck bypass after its staged lint, formatting, architecture ratchets, and docs validation passed. Focused lint, formatting, source build, and post-rebase production- artifact startup and exact HS256 signed API smoke passed; the production-empty catalog returns no-store unknown_output, and unsigned development tokens are rejected. The final BFF entry bundle is 36,065,656 bytes with SHA-256 6cc1f10e0385f6edf5a54e60a238d036ea16249c77b19fdbfde5a0dc0371edf3. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim measured media differences, asset-byte access, durable metric history, measurement authority, customer-owned comparison records, or release authority. This closes the seeded and cost-derived model-comparison metrics within an honest durable-catalog/recomputable-layout boundary, not the wider media-measurement program.

Authenticated Isis workflow-template hardening through 2a55a0f177 replaces the BFF route's claim of a live workflow-template catalog with an exact source-defined/recomputable inspector boundary. Repository tracing found real Prisma workflow_templates schema and seed data plus a separate @isis/workflow-registry service that defaults to PostgreSQL, runs its own migrations, lists durable templates, creates workflows from them, increments usage, and records audit events. This Oshun BFF has no discovery URL, service-to-service credential, or adapter for that service, however, and its three prompt strings are unrelated local DSL examples. The GET and inspector POST now return exact no-store source_defined_examples_recomputable_inspection dispositions: persistence is none, examples are illustrative rather than operational, parsing is source-defined and request-recomputed, the durable registry is separate_not_bound, runtime execution and mutations are unavailable, and authoritative state is false. Generated timestamps were removed. The Studio parser accepts only the exact disposition and exact top-level response shapes, rejects fabricated durable-registry authority and unknown fields, labels the examples as illustrative, renders the boundary capabilities, links to the separate Workflow Registry surface, and identifies one mounted route and four documented-only compatibility shapes without claiming IAM, audit exports, rollback, collaboration, policy decisions, or release signoff. Focused BFF tests passed 4/4, focused web tests passed 7/7, and the real workflow DSL library passed 89/89. Eleven isolated Chromium flows passed with applicable rendered- page axe coverage against a clean stack after all 32 Oshun and both Isis migrations; coverage proves exact live GET/POST no-store contracts, source- example selection and parsing, pending/network/server/malformed states, fabricated-registry rejection, scope denial, anonymous redirect, and 360-pixel overflow and 44-pixel controls. The web typecheck ratchet returned zero backlog. The full BFF compiler reported no workflow-template errors but remains red on the same two unrelated pre-existing autonomous-content errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts; the commit used the hook's documented typecheck bypass after its staged lint, formatting, stub scan, architecture ratchets, and docs validation passed. Focused lint, formatting, source build, and exact HS256 signed production API smoke passed; unsigned development tokens are rejected. The final BFF entry bundle is 36,066,374 bytes with SHA-256 5173cd76b34ed45c1e757f1e2c7f9cecd6c25d091ff175d4b46509f7adc75515. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim a bound durable template catalog, registry ownership or retention, workflow publication, execution, audit history, governance decisions, customer-owned records, or release authority. This closes the BFF's seeded DSL-example surface within its source-defined/recomputable boundary, not the separate durable workflow-registry program.

Authenticated Isis workflow-definition and workflow-versioning hardening through ed01b90a1b replaces two live-registry and approval-adjacent presentations with exact source-catalog boundaries. Repository tracing confirmed that @isis/workflows exposes a hard-coded WORKFLOW_REGISTRY described by its own source as static, precomputed metadata; its filters and statistics are pure, and the version surface recomputes parsed semver, category-relative maxima, and comparisons for each request. A separate PostgreSQL-backed @isis/workflow-registry service exists, but these Oshun BFF routes have no service discovery, credential, or adapter binding to it. Definitions now return an exact no-store source_defined_workflow_metadata disposition: persistence is source control, retention is repository history, GPU/time estimates are source- declared rather than observed, the durable registry and runtime execution are not bound, approval state is not observed, mutations are absent, and authoritative state is false. Versioning returns an exact no-store source_defined_recomputable_version_analysis disposition with source-defined schema compatibility, request-recomputed analysis, an unbound durable registry, no observed release history, no mutations, and no authoritative state. Both routes removed generated timestamps. Both Studio clients require exact top- level, disposition, catalog/statistics, parsed-version, and comparison shapes; unknown or legacy authority-inflated payloads fail closed. The UI identifies declared estimates and source-catalog maxima instead of observed telemetry, deployed releases, or update recommendations; renders the non-authoritative boundary; links the separate Workflow Registry surface; and marks one route as mounted and four as documented contracts without claiming IAM, exports, audit history, rollback, policy decisions, or signoff. Focused BFF tests passed 9/9, focused web tests passed 16/16, and 15 isolated Chromium flows passed with the repository fixture's applicable rendered-page axe checks against the real BFF and the existing fresh database stack after all 32 Oshun and both Isis migrations. Browser coverage proves the Studio hard block, live signed catalog, source filters, semver comparison, loading/empty/network/server/malformed failure states, scope denial, anonymous redirect, exact route status, and the separate-registry handoff. Focused lint, formatting, the web typecheck, and the BFF source build passed. The expanded BFF compiler completed only after raising its default 4 GiB heap, reported no errors in either changed route, and remains red on a large unrelated cross-workspace backlog led by Arete API Fastify typing and Yemaya marketplace type-only/DOM errors; the source commit used the hook's documented typecheck bypass after its staged ratchets passed. Exact signed production-artifact GET smoke returned 200 and no-store for both routes, while an unsigned development token was rejected with 401. The final BFF entry bundle is 36,067,442 bytes with SHA-256 b7643c801740ac87559cffced39dfaf6b0a06b48e6d95c2755af9d0a42fb6e6a. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. These surfaces do not claim durable workflow ownership, observed runtime measurements, deployment or release state, approval history, mutation authority, rollback, or publication authority. This closes the two static-catalog projections within their source-defined and recomputable boundaries, not the separate durable workflow-registry program.

Authenticated Isis workflow-portability hardening through 8cf5f9a1db replaces deploy-target and previously-published-class implications with an exact caller- supplied, request-recomputed analysis boundary. Repository tracing confirmed that @isis/comfyui-factory's runPortabilityCheck is a pure deterministic function: the candidate class, declared dependencies, available node types, available model hashes, and prior cost class all arrive in the request. It does not discover a deployment environment, resolve an installed inventory, read a managed-class catalog, or retrieve release history. GET and POST now return the same exact no-store caller_supplied_recomputable_portability_analysis disposition: persistence is none, the producer is request input, retention is not applicable, authority is not configured, target inventory is caller- supplied rather than discovered, candidate/prior classes are not resolved from registry or release history, analysis is recomputed, runtime execution is not bound, mutations are absent, and authoritative state is false. Generated timestamps and the generic POST ok wrapper were removed. Request parsing now rejects unknown top-level, candidate, dependency, and model-reference fields, including fabricated deployment or approval assertions. The Studio client requires exact catalog, disposition, report, cost-drift, and successful-response shapes; labels availability and prior class fields as caller-supplied; renders the non-authoritative boundary; and presents the technical portable boolean as “supplied inputs satisfied,” explicitly not deployment certification. The aggregate Workflows page now describes itself as a composed inspector instead of a control plane and marks one route mounted plus five documented-only route shapes without claiming execution records, diagnostics, parity evidence, governance decisions, or signoff. Focused BFF tests passed 6/6, focused web tests passed 9/9, and the canonical ComfyUI factory tests passed 12/12. Six isolated Chromium flows passed with applicable rendered-page axe coverage against the real BFF and existing fresh database stack after all 32 Oshun and both Isis migrations; coverage proves the Studio hard block, exact live catalog, real recomputation POST and input payload, malformed authority-inflated contract rejection, scope denial, anonymous redirect, and route status. Web typecheck, focused formatting/lint, and the BFF source build passed; two narrow lint exceptions document the existing app-to-canonical-Isis-kernel imports. The expanded BFF compiler remains red on the unrelated cross-workspace backlog recorded in the preceding slice, with no workflow-portability error observed; the source commit used the hook's documented typecheck bypass after staged ratchets passed. Exact signed production-artifact GET/POST smoke returned 200 and no-store. The final BFF entry bundle is 36,069,110 bytes with SHA-256 0848aefcb9f8b2ea3e4b97e69c6bc30596040e6f482fcdbc86dc579db8361fbf. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim discovered environment compatibility, installed inventory, durable analysis history, managed-catalog resolution, release drift, deployment certification, execution, mutation, or publication authority. This closes the portability calculator within its caller-supplied/recomputable boundary, not a live portability-certification program.

Authenticated Isis integration-event hardening through 35dbb52b3b replaces event-publication implications with an exact caller-supplied, request-recomputed schema-validation boundary. Repository tracing confirmed that the Oshun store only calls @isis/job-envelope's source-defined parser and defaulting helper. It does not call an event bus, reserve an idempotency key, inspect consumer state, persist a validation result, or prove delivery or acceptance. GET and POST now return the same exact no-store caller_supplied_recomputable_envelope_validation disposition: persistence is none, the producer is request input, retention is not applicable, authority is not configured, the schema is source-defined, validation is recomputed per request, the event bus is not bound, publication is absent, the defaulted idempotency value is not reserved, mutations are absent, and authoritative state is false. Generated timestamps and the generic POST ok wrapper were removed. Request parsing rejects unknown top-level fields, including fabricated publication assertions. The Studio client requires exact catalog, disposition, normalized-envelope, validation-result, and successful-response shapes, plus consistent valid/error invariants; legacy and authority-inflated payloads fail closed. The UI calls a conforming result schema-conforming input rather than a valid or published event, renders the non-authoritative boundary, and marks one route mounted plus four documented-only route shapes without claiming scene entitlements, exports, audit artifacts, revisions, rollback, policy decisions, or release signoff. Focused BFF tests passed 7/7, focused web tests passed 10/10, and the canonical job-envelope target passed from the Nx local cache for its 25-test suite. Seven isolated Chromium flows passed with applicable rendered-page axe coverage against the real BFF and existing fresh database stack after all 32 Oshun and both Isis migrations; coverage proves the Studio hard block, exact live GET/POST no-store contracts, real defaulting and invalid-schema paths, malformed authority-inflated contract rejection, scope denial, anonymous redirect, and exact route status. Web typecheck, focused formatting/lint, the BFF source build, and the repository production-artifact smoke passed. The expanded three-project commit-hook typecheck passed the web and prompt-injection projects; the BFF ratchet reported no integration-event error but remains red on two unrelated pre-existing errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts, so the source commit used the hook's documented typecheck bypass after its staged ratchets passed. Exact HS256-signed production-artifact GET/POST smoke returned 200 and no-store with only the declared response keys and publication: none. The final BFF entry bundle is 36,070,189 bytes with SHA-256 f59051fa4fa1d4e55e9e337893bc895fd257e365be9cbd782f102c01f17104fe. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim event publication, delivery, consumer acceptance, bus observation, idempotency reservation, durable validation history, customer-owned records, or release authority. This closes the integration-event validator within its caller- supplied/recomputable boundary, not an event-publication or delivery program.

Authenticated Isis resource-recommendation authority hardening through 1f9ad008e3 exposes the existing durable boundary without implying live routing or release control. Repository tracing confirmed PostgreSQL recommendation and operator-lifecycle tables, production startup binding, authenticated operator attribution, durable recommendation-id idempotency, cross-replica serialization, projection drift checks, redaction with a resurrection fence, and the canonical finite review-package retention partition. GET and successful POST now return the same exact no-store durable_operator_review_packages_recomputable_projection disposition: persistence is PostgreSQL, the producer is an authenticated operator decision, retention uses the canonical review-package window, the fallback catalog is source-defined deployed configuration, prices are configured estimates rather than live provider observations, stored inputs are recomputed and drift-checked, routing execution is not bound, release authority is absent, and only recommendation creation is exposed. Generated timestamps and the ambiguous success ok wrapper were removed. The strict Studio parser requires the exact disposition and catalog, projection, attribution, ordering, aggregate, and acknowledgement shapes; authority-inflated or corrupted responses fail closed. The UI distinguishes authoritative durable recommendation history from non-authoritative routing or pricing claims and marks one mounted route plus four documented contracts. Focused BFF/store tests passed 12/12, focused web tests passed 11/11, and real-PostgreSQL durability tests passed 2/2 across restart, replica races, erasure, retention holds/exceptions, corruption, and transaction rollback. Eight Playwright flows passed across Chromium and mobile Chrome with applicable axe coverage, proving durable reload, attribution, duplicate/spoof/ customer-field/off-chain rejection, failure states, route status, narrow layout, scope denial, and anonymous redirect. Web typecheck, focused lint/formatting, BFF source build, repository production-artifact smoke, and exact HS256-signed production GET/POST smoke passed. The source commit used the documented BFF typecheck bypass for the same two unrelated pre-existing ratchet errors recorded above. The final BFF entry bundle is 36,075,880 bytes with SHA-256 c79bebd64bf01d0721a64c18a9cb9299946cf996ff8a443bf8f0283bfacfde99. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim live provider prices, automatic routing, routing execution, customer recommendation exports, policy signoff, or release authority. This closes the durable resource-recommendation history within its stated boundary, not an intelligent routing or release-governance program.

Authenticated Isis intelligent-routing authority hardening through 11836e0084 exposes the existing durable decision-package boundary without claiming provider execution or live traffic control. Repository tracing confirmed normalized PostgreSQL routing-record and operator-lifecycle tables, production startup binding, authenticated operator attribution, durable route-id idempotency, cross-replica serialization, stored-input recomputation with projection-drift rejection, redaction with a resurrection fence, and the canonical finite review-package retention partition. GET and successful POST now return the same exact no-store durable_operator_routing_decision_packages_recomputable_projection disposition: persistence is PostgreSQL, the producer is an authenticated operator decision, retention uses the canonical review-package window, the catalog is source-defined deployed configuration, pricing is configured rather than live, selection is recomputed and drift-checked, provider execution is not bound, traffic routing and release authority are absent, and only decision- package creation is exposed. Generated timestamps and the ambiguous success ok wrapper were removed. The strict Studio parser requires the exact disposition, catalog, projection, attribution, ordering, aggregate, and acknowledgement shapes, including an exact echo of submitted decision inputs; authority-inflated, corrupted, or mismatched responses fail closed. The UI distinguishes authoritative durable decision history from non-authoritative provider execution, traffic routing, pricing, and release claims, and visibly marks one mounted route plus four documented contracts. Focused BFF/store tests passed 12/12, focused web tests passed 11/11, and real-PostgreSQL durability tests passed 2/2 across restart, replica races, erasure, retention holds and exceptions, projection and attribution corruption, and transaction rollback. Eight Playwright flows passed across Chromium and mobile Chrome with applicable axe coverage, proving the Studio hard block, exact live GET/POST contracts, durable reload and attribution, duplicate/spoof/customer-field/off-chain rejection, malformed and network failure states, explicit route status, narrow layout, scope denial, and anonymous redirect. Web typecheck, focused lint/formatting, BFF source build, repository production-artifact smoke, and exact HS256-signed production GET/POST smoke passed. The source commit used the documented BFF typecheck bypass for the same two unrelated pre-existing ratchet errors recorded above. The final BFF entry bundle is 36,081,550 bytes with SHA-256 0d3774f0e5bc590788587b4ae21a93a4a2b5948400c9a92d2d9736af17939b8e. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim live provider prices or availability, provider invocation, traffic handoff, policy enforcement, customer routing exports, release signoff, or release authority. This closes durable intelligent-routing decision history within its stated boundary, not an automatic traffic-routing or release- governance program.

Authenticated Isis job-submission hardening through ff78a836e9 replaces a submission/control-plane presentation with the exact request-local job-envelope builder that is actually bound. Repository tracing confirmed that this route only generates an unreserved UUID and calls @isis/job-envelope's source-defined builder and validator. It does not enqueue a generation job, reserve idempotency, persist state, invoke a provider, publish an artifact, or grant release authority; the separate /v1/generation/jobs surface remains the durable queue authority. GET and successful POST now return the same exact no-store request_local_non_authoritative_envelope_builder disposition: persistence is none, the producer is request input, retention is request-only, authority binding is access-scope-only and is not embedded in the envelope, the schema is source-defined, validation is recomputed, the job identifier is a server-generated unreserved UUID, the default idempotency value is not reserved, queue submission and mutation, provider execution, publication, and all mutations are absent, and authoritative state is false. Generated timestamps and the ambiguous success ok wrapper were removed. Request parsing rejects unknown fields and protected authority fields such as caller-supplied job identifiers, statuses, actor identities, queue timestamps, and provider execution. The strict Studio parser requires exact disposition, catalog, and envelope shapes, verifies the submitted type, priority, and input echo, and requires the default idempotency key to equal the returned UUID. The UI labels the result as a request-local envelope rather than a submitted job, exposes the non-authoritative boundary, and marks one mounted route plus four documented contracts without claiming scenes, exports, revision history, governance, or release controls. Focused BFF tests passed 6/6 and focused web tests passed 7/7. Ten isolated Chromium flows passed across desktop and mobile with applicable rendered-page axe coverage against the real BFF after all 32 Oshun migrations, proving the Studio hard block, exact live GET/POST contracts, source catalog, strict invalid and protected-field rejection, malformed-success and authority-inflated response rejection, loading/server/network failures, scope denial, anonymous redirect, narrow-layout overflow, and touch targets. Web typecheck, focused lint/formatting, BFF source build, repository production- artifact smoke, and exact HS256-signed production GET/POST smoke passed. The source commit used the documented BFF typecheck bypass for the same two unrelated pre-existing ratchet errors recorded above. The final BFF entry bundle is 36,083,259 bytes with SHA-256 c09f04d9cebb907cba69fa9043656df5a1b616734292243682ff9977d6d458f8. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This surface does not claim a queued job, idempotency reservation, durable submission history, provider execution, artifact publication, customer-owned records, or release authority. This closes the request-local job-envelope builder within its stated boundary, not the separate durable job-management or generation-execution program.

Authenticated Isis job-management hardening through 9cf4b29919 exposes and constrains that separate durable generation queue without implying a complete provider operations program. Repository tracing confirmed production startup hydrates a PostgreSQL snapshot, every processor claims durable running state before provider dispatch, release evaluation is mandatory and fail-closed, and restart recovery moves stranded running work to needs_review. HTTP enqueue now binds submitter and tenant attribution from the authenticated token, rejects caller-supplied ownership/authority fields, and cannot create a customer-owned record. Tenant-bearing operators can list, read, process, and retry only their tenant's jobs; an operator token without a tenant remains the explicit global administration boundary. GET, enqueue, process, read, and retry emit no-store responses carrying the same exact durable_operator_generation_queue disposition. Processing accepts only an exact bounded max of 1–25. Provider exceptions after dispatch become needs_review and cannot be automatically retried; only a durable failed/provider_not_configured pre-dispatch result can be re-queued. The strict Studio parser rejects malformed, additional, or authority-inflated response fields. The UI states that provider availability is not live-observed, requires an explicit external-effect/spend acknowledgement before a bounded provider tick, exposes retry only for the safe pre-dispatch class, and labels one mounted route plus four documented contracts without inventing governance, export, revision, or release surfaces. Focused BFF tests passed 37/37 and focused web tests passed 10/10. Ten isolated Playwright flows passed across desktop and mobile Chromium with applicable rendered-page axe coverage against the real compiled BFF and freshly migrated PostgreSQL, including exact contracts, tenant attribution, bounded execution, safe retry, malformed acknowledgement rejection, scope denial, anonymous redirect, narrow layout, and touch targets. Web and BFF direct typechecks, focused lint/format, the BFF source build, and the repository production-artifact verifier passed. The expanded commit-hook typecheck remained red only on the same two unrelated clean-tree BFF ratchet errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts, so the source commit used the hook's documented scoped bypass after the affected checks passed. Exact HS256-signed production-artifact smoke proved enqueue, server-bound tenant/operator attribution, bounded processing, pre-dispatch failure classification, safe retry, persistence, exact response keys, and no-store headers. The final BFF entry bundle is 36,086,985 bytes with SHA-256 9e5f8bd5b8536c90e252d37dc65b2f4c026dd9f527e22e8317f1e63dd2f4b61d. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. The queue's configured executor catalog is not a live provider-health observation; provider execution is an explicit real external effect. General retention and action audit remain unconfigured, and exact-subject DSAR export/erasure applies only to jobs carrying explicit subject ownership. This closes the durable operator queue and its safe-retry boundary, not live provider availability, provider-side idempotency, automatic dispatch, action-audit, retention, customer ownership, artifact publication, or release-governance completion.

Authenticated Isis audio-generation hardening through 42b97fcae8 replaces an audio-generation/operations presentation with the exact request-local loudness calculator that is actually bound. Repository tracing confirmed that the route only calls @isis/audio-generation's source-defined computeFrameRms over caller-supplied samples and derives peak and mean RMS. It does not generate audio, decode or upload media, invoke a provider, persist an analysis, produce an artifact, export evidence, create revisions, record governance, or grant release authority. GET and successful POST now return the same exact no-store request_local_deterministic_loudness_analysis disposition: persistence is none, the producer is caller-supplied normalized samples, retention is request- only, authority binding is access-scope-only, analysis is the source-defined frame-RMS function, generation is unavailable, provider execution, artifact output, and mutations are absent, and authoritative state is false. The generated timestamp and ambiguous ok wrapper were removed. POST accepts only the exact samples and frameSize fields, caps work at 16,384 finite normalized PCM samples in [-1, 1], and requires a frame size from one through the sample count. The result binds sample count and frame size to its RMS envelope. The strict Studio parser requires exact disposition, metric, and analysis shapes, recomputes frame-count, peak, and mean invariants, and rejects extra authority fields or mathematically inconsistent acknowledgements. The restrained worksheet leads with the non-authoritative boundary, uses 44px controls, and marks one mounted route plus five documented-only route shapes without implying scenes, exports, revisions, parity evidence, governance, or signoff. Focused BFF tests passed 6/6 and focused web tests passed 10/10. Eight isolated Chromium flows passed with applicable rendered-page axe coverage against the real BFF and freshly migrated PostgreSQL, proving the Studio hard block, exact live GET/POST/400 contracts, deterministic RMS result, normalized-input rejection, inflated catalog and malformed acknowledgement rejection, route honesty, scope denial, narrow-layout containment, and touch targets. BFF and web direct typechecks, focused lint/formatting, the BFF source build, repository production- artifact smoke, and exact HS256-signed production GET/POST/400 smoke passed. The source commit used the documented typecheck bypass only because the expanded BFF ratchet still reports the same two unrelated clean-tree errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts; the affected direct typechecks passed. The final BFF entry bundle is 36,088,734 bytes with SHA-256 39570b28cce2aabf9cbb772e7e01e67737196b764a8c5682ddcac41108c351fd. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes the request-local deterministic RMS analyzer within its stated boundary, not audio generation, provider/media execution, durable analysis history, artifact or evidence export, governance, or release authority.

Authenticated Isis music-generation hardening through a487770cca replaces an admission- and provider-adjacent guardrail console with the exact request-local evaluation boundary that is actually bound. Repository tracing confirmed that the Studio route only calls @isis/music-generation's source-defined checkGuardrails and prompt-string heuristics over caller-supplied metadata. A separate /v1/isis/music/generate endpoint owns entitlement evaluation and generation-executor dispatch; this administrator worksheet has no adapter to that route and does not resolve provider health, enqueue work, persist a decision, execute a provider, or create audio. GET and successful POST now return an exact no-store request_local_non_authoritative_guardrail_evaluation disposition: persistence is none, the producer is caller-supplied request metadata, retention is request-only, authority binding is access-scope-only, evaluation uses source-defined guardrails and prompt heuristics, generation is unavailable, queue submission, provider execution, artifact output, and mutations are absent, and authoritative state is false. Provider names are explicitly source-defined identifiers rather than availability observations. License and consent IDs are caller-supplied heuristic inputs and are not verified here.

POST now accepts only four exact top-level fields and six exact request fields. It bounds prompt and evidence-identifier lengths, MIME length, work to 4,096 finite intensity points in [0, 1], positive finite duration and tempo, and a non-negative safe-integer file size. Unknown authority fields, invalid classes, empty curves, non-finite values, and out-of-domain values fail with 400. The successful result is named locallyValid, not admitted, and is nested beside the same non-authoritative disposition. The strict Studio parser requires exact catalog, disposition, guardrail, issue, and successful-response shapes; checks that the workflow class matches the submitted class; recomputes the relationship between issues, the content verdict, and locallyValid; and rejects additional authority fields. The restrained worksheet leads with the non-generation boundary, distinguishes provider identifiers from availability, labels unverified evidence references, uses 44-pixel controls, contains the table on narrow screens, and marks one route mounted plus four documented-only shapes without inventing scenes, exports, revisions, governance, or signoff.

Focused BFF tests passed 10/10, focused web tests passed 8/8, and the underlying music-generation library passed 21/21. Eight isolated Chromium flows passed with applicable rendered-page axe coverage against the real BFF, proving the Studio hard block, exact live GET/POST/400 no-store contracts, local-valid and content-issue results, client-side intensity rejection, authority-inflated catalog and acknowledgement rejection, scope denial, route honesty, narrow layout, and touch targets. Web typecheck, focused lint/formatting, the BFF source build, repository production-artifact verifier, and exact HS256-signed production GET/POST/400 smoke passed. The expanded BFF ratchet reports no error in the changed music route and remains red only on the same two unrelated clean-tree errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts; the source commit used the hook's documented scoped bypass after the affected checks passed. The final BFF entry bundle is 36,090,225 bytes with SHA-256 df19f06f566c602037fa5b8d6966f0cd8c5eead71414bfc2f721abfb6b9bbf1c. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes the request-local music guardrail evaluator within its stated boundary, not generation admission, verified rights or consent, live provider availability, queue or provider execution, durable evaluation history, audio artifacts, governance, or release authority.

Authenticated Isis AI-video-generation hardening through 5bea5fb55e replaces a video-generation presentation with the exact request-local frame-count worksheet that is actually bound. Repository tracing confirmed that the admin route only calls the pure planFrameInterpolation timestamp-arithmetic helper. A separate generation-job pipeline resolves configured video providers and owns real text-to-video/image-to-video execution; this worksheet has no adapter to that pipeline. It does not receive image or frame bytes, synthesize frames, resolve provider availability, enqueue work, persist a plan, create an artifact, record governance, or grant release authority.

GET and successful POST now return the same exact no-store request_local_non_authoritative_frame_count_plan disposition: persistence is none, the producer is caller-supplied keyframe timestamps, retention is request-only, authority binding is access-scope-only, planning is deterministic timestamp arithmetic, video generation is unavailable, queue submission, provider execution, frame synthesis, artifact output, and mutations are absent, and authoritative state is false. The catalog publishes the precise formulas and bounded input contract rather than a generated timestamp or ambiguous ok wrapper. POST accepts only exact keyframes and targetFps fields; each keyframe accepts only id and timeMs. Work is capped at 4,096 keyframes, identifiers at 128 characters, timestamps at safe-integer milliseconds within 24 hours, and target FPS from 1 through 240. IDs must be nonempty and unique, and timestamps must be nonnegative and strictly increasing. Unknown authority fields, fractional or out-of-range values, duplicate IDs, and non-increasing times fail with 400.

The arithmetic now guarantees at least one output frame per interval because the interval's ending keyframe is real. This repairs the prior sub-frame-gap case where rounding could yield zero output frames and contradict the stated totalSynthesizedFrames === totalOutputFrames - keyframeCount invariant. The strict Studio parser requires exact catalog, disposition, planner, plan, interval, and summary shapes; independently recomputes every interval and total from the submitted timestamps; and rejects additional authority fields or a mathematically false acknowledgement. The restrained worksheet leads with the non-generation boundary, names the separate generation-job authority, labels results as estimates rather than media, uses a 44-pixel submit control, contains its table on narrow screens, and marks one route mounted plus four documented-only shapes without inventing scenes, exports, revisions, governance, or signoff.

Focused BFF tests passed 10/10 and focused web tests passed 10/10. Eight isolated Chromium flows passed with applicable rendered-page axe coverage against the real in-memory BFF, proving the Studio hard block, exact live GET/POST/400 no-store contracts, ordinary and sub-frame-gap arithmetic, local invalid-input rejection, authority-inflated catalog and acknowledgement rejection, route honesty, scope denial, narrow-layout containment, and touch targets. The managed Playwright launcher first reproduced the repository's known missing account_protection_subject_state table in its default database; the isolated run therefore used the established no-persistence BFF harness, while every worksheet API request still reached the real BFF route. Web typecheck, focused lint/formatting, the BFF source build, repository production-artifact verifier, and exact HS256-signed production GET/POST/400 smoke passed. The expanded BFF ratchet reports no error in the changed route or tests and remains red only on the same two unrelated clean-tree errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts; the source commit used the hook's documented scoped bypass after the affected checks passed. The final BFF entry bundle is 36,092,484 bytes with SHA-256 719e536d0a286d388bfc07581904b7d1a9b588f6e65f2c8bfdf0ed1332a0d6cb. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes the request-local frame-count planner within its stated boundary, not video generation, frame synthesis, live provider availability, queue or provider execution, durable plan history, video artifacts, governance, or release authority.

Authenticated Isis video-generation hardening through 95942c1f62 replaces a provider-running and release-ready presentation with the exact source-defined manifest inspector that is actually bound. Repository tracing confirmed that the shared admin route only reads @isis/job-envelope's static generation-type UI/output-package rules and calls its deterministic rule matcher. The separate generation-job pipeline resolves configured video providers and owns real text-to-video/image-to-video execution; the manifest inspector has no adapter to that pipeline and does not observe its health. It does not validate a full job envelope, evaluate entitlements or safety, admit or queue work, execute a provider, persist a check, create a video artifact, record governance, or grant release authority.

GET and successful POST now return the same exact no-store source_defined_recomputable_generation_manifest_compatibility disposition: persistence is none, the producer is the source-defined job-envelope manifest, retention is recomputed-per-request, authority binding is access-scope-only, the catalog is static UI/output-package rules, compatibility is a deterministic rule match, generation admission and queue submission are absent, provider availability is not observed, provider execution, artifact output, and mutations are absent, and authoritative state is false. The generated timestamp and ambiguous ok wrapper were removed. POST accepts only exact generationType and outputPackage fields; the package must contain exactly a source-defined type and an explicit boolean multiple. Unknown authority fields, omitted multiplicity, and unknown generation or package identifiers fail with 400. A successful response nests the echoed shape and boolean under check, never an admission verdict.

The strict Studio parser requires exact catalog, disposition, input-contract, manifest-entry, UI, output-rule, and successful-check shapes. It verifies the complete ordered set of 18 source generation identifiers and five package identifiers, binds the acknowledgement to the submitted generation/package shape, independently recomputes compatibility from the published entry, and rejects additional authority fields or a false result. The shared image and text-to-audio consumers were adjusted to the non-breaking nested check response; their broader boundary audits remain separate. The restrained video inspector leads with the non-generation boundary, distinguishes source compatibility from admission, names the separate execution authority, uses 44-pixel controls, and marks one route mounted plus four documented-only shapes without inventing scenes, exports, revisions, governance, or signoff.

Focused BFF tests passed 4/4, focused affected web tests passed 9/9, and the underlying job-envelope suite passed 18/18. Eight isolated Chromium flows passed with applicable rendered-page axe coverage against the real in-memory BFF, proving the Studio hard block, exact live GET/POST/400 no-store contracts, compatible and incompatible package shapes, authority-inflated catalog and acknowledgement rejection, route honesty, scope denial, narrow-layout containment, and touch targets. Web typecheck, focused lint/formatting, the BFF source build, repository production-artifact verifier, and exact HS256-signed production GET/POST/400 smoke passed. The expanded BFF ratchet reports no error in the changed route or tests and remains red only on the same two unrelated clean-tree errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts; the source commit used the hook's documented scoped bypass after the affected checks passed. The final BFF entry bundle is 36,094,004 bytes with SHA-256 9634f08ed8822d4eecd0078b760363a0dfea8e6ac9b1f8618154b11d8e928af9. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes the source-defined video-manifest compatibility inspector within its stated boundary, not job validation or admission, live provider availability, video generation, queue or provider execution, durable check history, video artifacts, governance, or release authority.

Authenticated Isis image-generation hardening through 88b4b3bae7 applies the same exact source-defined manifest boundary to the text-to-image consumer. The shared BFF contract and strict catalog/check parser were established and verified in the immediately preceding video-manifest slice at 95942c1f62; this slice closes the remaining image-specific presentation and consumer seam without changing or recharacterizing that backend authority. The image route now requires the complete ordered 18-generation-type and five-package catalog, the exact static UI and output rules, the exact non-authoritative disposition, and an exact compatibility acknowledgement bound to the submitted text-to-image package shape. A malformed successful response, incomplete catalog, extra authority field, false compatibility acknowledgement, or missing image entry fails closed instead of being rendered as an ordinary incompatible result.

The restrained image manifest inspector now leads with the actual boundary: the catalog is source-defined, checks are recomputed per request, persistence and generation admission are absent, provider availability is not observed, and provider execution and artifact output are absent. Compatible and incompatible results both state that no job was admitted. The text-to-image entry exposes its real prompt, negative-prompt, image-control, and output-package rules without presenting provider health or a generation workflow. Controls meet the 44-pixel target, the route remains contained at a narrow viewport, and the route map reports one mounted inspector plus four documented-only shapes; it does not invent scenes, exports, revisions, governance records, or signoff.

Focused image component, page, route-map, and integration tests passed 7/7. Eight isolated Chromium flows passed with applicable rendered-page axe coverage against the real in-memory BFF, proving the Studio hard block, exact live GET/POST/400 no-store contracts, compatible image and incompatible video package shapes, authority-inflated catalog and acknowledgement rejection, scope denial, route honesty, narrow-layout containment, and touch targets. The web typecheck and focused lint/formatting passed, and the commit hook repeated the web typecheck ratchet with zero app-owned backlog. No BFF source or bundle changed in this consumer-only slice; the shared route's focused tests, source build, production-artifact verification, signed production smoke, and bundle digest remain recorded in the immediately preceding manifest evidence. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes the source-defined image-manifest compatibility inspector within its stated boundary, not full job validation, entitlement or safety evaluation, generation admission, live provider availability, image generation, queue or provider execution, durable check history, image artifacts, governance, or release authority.

Authenticated Isis text-to-audio hardening through 05d4a9bfbc closes the last known Studio consumer of the shared source-defined generation manifest. As with the image consumer, the BFF route and strict shared catalog/check contract were already established at 95942c1f62; this consumer-only slice does not change or expand that backend authority. The text-to-audio route now requires the full ordered 18-generation-type and five-package catalog, exact static UI and output rules, exact non-authoritative disposition, and an exact acknowledgement bound to the submitted text-to-audio package shape. Its focused fixture now models the real audio output rule instead of inheriting an image placeholder. Partial catalogs, malformed successful responses, additional authority fields, and compatibility booleans that disagree with the source rule all fail closed.

The restrained text-to-audio manifest inspector leads with the same explicit boundary: persistence and generation admission are absent, provider availability is not observed, provider execution and artifact output are absent, and real execution belongs to the separate generation-job pipeline. Both compatible and incompatible outcomes state that no job was admitted. The route exposes prompt, image-control, and audio output-package rules without claiming synthesis, provider health, or an artifact. Controls and sibling navigation meet the 44-pixel target, the surface is contained on a narrow viewport, and the route map distinguishes one mounted inspector from four documented-only shapes rather than inventing scenes, exports, revisions, governance records, or signoff.

Focused text-to-audio component, page, route-map, and integration tests passed 7/7. Eleven isolated Chromium flows passed with applicable rendered-page axe coverage against the real in-memory BFF, proving the Studio hard block, exact live GET/POST/400 no-store contracts, compatible audio and incompatible image packages, loading and network-failure behavior, authority-inflated catalog and acknowledgement rejection, scope denial, anonymous redirect, route honesty, narrow-layout containment, and touch targets. The older browser session seeder could not create accounts against the deliberately no-persistence BFF, so the final run used the established explicit isolated-session harness; every manifest request still reached the real BFF route. Web typecheck, focused lint/formatting, and the commit-hook typecheck ratchet passed with zero app-owned backlog. No BFF source or bundle changed; its build, production-artifact, signed smoke, and digest evidence remain in the shared manifest slice. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes the source-defined text-to-audio compatibility inspector within its stated boundary, not full job validation, entitlement or safety evaluation, generation admission, live provider availability, audio synthesis, queue or provider execution, durable check history, audio artifacts, governance, or release authority.

Authenticated Isis 3D-result-management hardening through adcf24fcc9 replaces a generated-result acceptance and post-pipeline presentation with the exact request-local topology worksheet that is actually bound. Repository tracing confirmed that the admin route accepts only caller-supplied flat position and triangle-index arrays and calls the pure @isis/3d-post-pipeline topology helper. It receives no generated-result or artifact identity and has no adapter to generation, post-processing admission, mesh repair, persistence, export, publication, or provider execution.

GET and successful POST now return the same exact no-store request_local_non_authoritative_mesh_topology_analysis disposition: persistence is none, the producer is a caller-supplied indexed mesh, retention is request-only, authority binding is access-scope-only, analysis is the deterministic source-defined topology helper, pipeline admission and mesh repair are absent, generation is unavailable, provider execution, artifact output, and mutations are absent, and authoritative state is false. The generated timestamp and ambiguous ok wrapper were removed. POST accepts exactly positions and indices; extra admission fields fail with 400. Both arrays must be nonempty complete triples, coordinates must be finite with absolute value at most one billion, indices must be in-range nonnegative integers, and each array is capped at 30,000 values. This bounds the helper to at most 10,000 vertices and 10,000 submitted triangles per request.

Successful analysis is nested as a topology description rather than an acceptance verdict. The strict Studio parser requires exact catalog, disposition, input-contract, acknowledgement, analysis, and 11-field topology shapes. It binds vertex and triangle counts to the submitted arrays, requires nonnegative safe-integer counts and boolean predicates, enforces that a watertight result is also closed and manifold, and rejects additional authority fields or a mismatched response. The worksheet now leads with the request-only non-admission boundary, labels output as a topology description with no artifact admitted, uses a 44-pixel submit control, remains contained on a narrow viewport, and marks one route mounted plus four documented-only shapes without inventing scenes, exports, revisions, governance, or release signoff.

Focused BFF tests passed 7/7, focused web tests passed 5/5, and the underlying manifold/topology library tests passed 4/4. Eight isolated Chromium flows passed with applicable rendered-page axe coverage against the real in-memory BFF, proving the Studio hard block, exact live GET/POST/400 no-store contracts, closed-tetrahedron and open-triangle descriptions, authority-inflated catalog and acknowledgement rejection, route honesty, scope denial, narrow-layout containment, and touch targets. Web typecheck and focused lint/formatting passed. The BFF source build passed; its final entry bundle is 36,096,255 bytes with SHA-256 346cfe0a1dc5581c57025519a7b0eb2e2ffaa657b607bdf7567af387fe3ed65a. The scoped BFF ratchet required the documented 6 GiB heap allowance and reached only the same two unrelated clean-tree errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts; the source commit used the hook's documented typecheck-only bypass after all affected checks passed. Production-artifact startup again reached the known missing account_protection_subject_state table and therefore could not complete its health smoke; this changed route was exercised directly through the real source BFF in Playwright. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes the request-local mesh-topology analyzer within its stated boundary, not generated-result management, post-pipeline acceptance, mesh repair, durable analysis history, generation or provider execution, 3D artifacts, governance, or release authority.

Authenticated Isis Analytics and Reporting hardening through 9ac7f3cc0f0013229e6e041055da9f548e3c5bce replaces broad telemetry, reporting-history, export, governance, and release language with the exact two pure helpers that are actually bound. Repository tracing found only the administrator route and its tests calling @isis/operation-orchestrator: caller-supplied retry attempts are aggregated into descriptive counts and distributions, while caller-supplied backoff parameters produce a deterministic delay preview. No telemetry producer, operation executor, retry scheduler, database, report history, export artifact, jitter application, or release authority is connected.

GET, retry-stats POST, and backoff POST now return the same exact no-store request_local_non_authoritative_retry_analysis disposition. Persistence and mutations are absent, retention is request-only, authority is access-scope plus caller-supplied values, and authoritative state is false. Exact-key parsing rejects fabricated persistence or scheduling fields. Requests are capped at 1,000 attempts; operation references are canonical and bounded to 200 characters; attempt numbers, delays, and multipliers have safe finite bounds; and backoff attempts begin at one. The browser requires exact catalogs and semantically binds every aggregate count, distribution, average, and retry rate to the submitted attempts. It independently recomputes the deterministic backoff result and rejects a successful acknowledgement that does not match the submitted parameters.

The rendered Retry Analysis Preview leads with its request-only boundary, labels its JSON as an editable caller-supplied sample, states that the delay is not scheduled, meets the 44-pixel control floor, remains contained on a narrow viewport, and distinguishes one mounted route from four documented-only shapes. Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 8/8. Eight isolated Chromium flows passed with applicable axe coverage against the real source BFF and a disposable PostgreSQL database with all 32 migrations, proving the Studio hard block, exact live GET/two-POST/400 contracts, authority-inflated request and response rejection, scope denial, route honesty, touch targets, and mobile containment. Web typecheck, affected lint, formatting, and whitespace checks passed. The BFF build and production-artifact smoke passed against that database; the final 36,100,449-byte entry bundle has SHA-256 31d4c9edca2bae238ef1718d7cb013e9d616f8169b2c42b70001e25702a86f8e. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes only the request-local retry-analysis and backoff-preview boundary, not production analytics, telemetry, execution, scheduling, reporting history, export, governance, or release authority.

Authenticated Isis 3D Editing hardening through f7c89cc85db1f1641911a3384b00ed99af7ff13f separates the browser-local demo editor from the only server-side operation actually bound to the surface. Repository tracing confirmed that sculpt history, seeded geometry, previews, and browser downloads live only in the current tab. There is no project save, collaboration record, server upload, archive, generated-artifact identity, or downstream delivery. The BFF independently accepts caller-supplied indexed-mesh arrays and invokes the pure @isis/3d-post-pipeline manifold-repair helper; it does not receive the editor's mesh and deliberately returns no repaired geometry.

GET and successful POST now return the same exact no-store request_local_non_authoritative_mesh_repair_analysis disposition. It declares no persistence, request-only retention, access-scope-only authority binding, no browser-editor binding, no repaired-mesh output, no export or pipeline admission, no artifact output, no mutations, and non-authoritative state. POST accepts exactly positions, indices, and an optional exact repair-options object. Arrays are capped at 30,000 values, triples must be complete, indices must be safe and in range, and finite coordinates are bounded to one billion in absolute value. The response is a topology-before/after simulation with actions, metrics, warnings, blockers, and a topology-readiness description—not an export verdict.

The strict browser parser binds submitted vertex and triangle counts to the acknowledgement, validates both topology snapshots, every action and metric, and the relationship between watertightness, blockers, and topologyReady. It rejects authority-inflated catalogs or acknowledgements. The rendered workspace leads with the browser-local loss boundary, calls downloads browser downloads, states that analysis returns no repaired mesh or export admission, meets the 44-pixel control floor, remains contained on a narrow viewport, and marks one route mounted plus five documented-only shapes.

Focused BFF tests passed 7/7, focused web tests passed 18/18, and the underlying manifold-repair library tests passed 4/4. Eight isolated Chromium scenarios passed with retries disabled and applicable axe coverage against the real production BFF artifact and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, exact live GET/POST/400 no-store contracts, authority-inflation rejection, scope denial, route honesty, touch targets, mobile containment, a real sculpt stroke, and reload loss of that browser-only history. Web typecheck, affected lint, formatting, and whitespace checks passed. The production BFF build and migrated-database artifact smoke passed; the final 36,103,182-byte entry bundle has SHA-256 b7a3bdcb8fea7a939bcee418b9da37546c12d1d978568d72f84201dff33cf15e. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes only the browser-local editing and request-local repair- simulation boundary, not durable editing, collaboration, repaired-mesh output, server-side export, artifacts, pipeline admission, governance, or release authority.

Authenticated Isis 3D Generation hardening through 45f1a952fa76045ddafc3210f2ee60cf7431fe89 replaces generation, orchestration, export, history, governance, and release language with the exact pure helper that is actually bound. Repository tracing found that only the administrator route calls @isis/3d-product-parity's buildStructured3DPrompt: it assembles a string from caller-supplied subject, material, style, topology, poly-budget, export-target, and animation fields. It has no generation request, provider, queue, project, scene, revision history, export writer, artifact, audit record, approval, or release connection.

GET and successful POST now return the exact no-store request_local_non_authoritative_structured_prompt_preview disposition. It declares no persistence, request-only retention, access-scope-only authority, caller-supplied inputs, deterministic source-defined prompt assembly, and no generation admission, provider execution, queue submission, export admission, artifact output, mutations, or authoritative state. Exact-key parsing rejects generation and export authority fields. Subject text is bounded to 1,000 characters, optional text fields to 500, control characters and supplied empty values are rejected, and topology/export values must come from the exact source vocabularies.

The browser requires the exact catalog, disposition, and input contract. It independently reconstructs the canonical prompt and ordered field list from the submitted values, so mismatched output or a successful acknowledgement claiming generation or export authority fails closed. The rendered worksheet calls export target a prompt phrase, leads with its request/browser-tab loss boundary, uses 44-pixel controls, states that no 3D asset is generated or admitted, and marks the primary worksheet plus the separately mounted browser turntable viewer honestly while leaving five aspirational shapes documented-only.

Focused BFF tests passed 6/6, focused web tests passed 9/9, and the underlying prompt-builder tests passed 2/2. Eight isolated Chromium scenarios passed with retries disabled and applicable axe coverage against the real production BFF artifact and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, exact live GET/POST/400 no-store contracts, generation/export authority-inflation rejection, scope denial, route honesty, touch targets, mobile containment, and reload loss of edited prompt fields. Web typecheck, affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database production-artifact smoke passed; the final 36,105,735-byte entry bundle has SHA-256 5b1bbf47f6e10c3da822db25186e20ea09b5d9d213cb029d75399edd81537bf5. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes only the request-local structured-prompt preview, not 3D generation, provider or queue execution, project persistence, export, artifacts, governance, or release authority.

Authenticated Isis Auto-Rigging hardening through 55f36ce86cf9d0fb8f263b070e5c69bc4e5196e8 separates the large browser-local sample-session demo from the only server helper actually bound to the route. The demo's apparent landmark, correction, skin-weight, validation, and motion actions are React state lost on reload; they create no project, collaboration record, rig mutation, animation application, export, artifact, governance record, or release decision. The API receives only caller-supplied bone names and performs a deterministic source-defined membership comparison; it receives no mesh, detected skeleton, rig identity, or artifact identity.

GET and successful POST return the exact no-store request_local_non_authoritative_humanoid_bone_map_analysis disposition. It declares no persistence, request-only retention, access-scope-only authority, and no mesh binding, skeleton detection, rig mutation, retarget admission, animation application, export admission, artifact output, mutations, or authoritative state. Exact-key parsing rejects invented mutation fields. Input is bounded to 256 unique case-sensitive names of at most 100 characters, with whitespace and control-character ambiguity rejected.

The browser requires the exact contract and independently recomputes every presence row, category list, coverage value, completeness boolean, and summary count from the submitted names. Authority-inflated catalogs and semantically false successful acknowledgements fail closed. The restrained route leads with both loss/non-authority boundaries, labels one mounted route plus four documented-only shapes, meets the 44-pixel control floor, and replaces legacy fixed-minimum stage grids with contained responsive layouts at 390 pixels.

Focused BFF tests passed 5/5 and focused web component, integration, page, and route-map tests passed 27/27. Eight isolated Chromium scenarios passed with retries disabled and applicable axe coverage against the real built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, exact live GET/POST/400 no-store contracts, demo-state reload loss, catalog authority-inflation rejection, semantic acknowledgement rejection, scope denial, route honesty, touch targets, and descendant-level mobile containment. Web typecheck, affected lint/formatting, and whitespace checks passed. The BFF build and migrated-database production-artifact smoke passed; the final 36,107,416-byte entry bundle has SHA-256 0c9de0daca5094217454ce4a9901ae2c837e99d3107435135240720df5edfd88. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build was not run because this 15 GiB host is below the repository's 32 GiB hard gate and no approval was provided. This closes only the browser-local sample workflow and request-local bone-name analysis, not real skeleton detection, rig mutation, retargeting, animation application, project persistence, collaboration, export, artifacts, governance, or release authority.

Authenticated Isis InstantID hardening through 359b8063613466a8891e23dcb9d5aa4eb5609574 binds the shared ControlNet endpoint to its actual request-local string-normalization behavior. It receives only a caller-supplied domain and mode and invokes the pure source-defined job-envelope alias normalizer and legacy-name mapper. It receives no face, image, identity, conditioning input, job, provider result, or artifact identity and has no path to execution, generation, persistence, export, governance, or release.

GET and successful POST expose the exact no-store request_local_non_authoritative_control_mode_alias_analysis disposition: no persistence, request-only retention, access-scope-only authority, deterministic source-defined alias normalization, and no image binding, identity extraction, conditioning execution, job admission, provider execution, artifact output, mutations, or authoritative state. POST accepts exactly a catalog domain and a bounded trimmed control-free mode string; invented execution fields fail with 400.

The InstantID browser requires the full exact contract and independently recomputes the canonical and legacy names for every submitted alias. Partial or authority-inflated catalogs and successful mappings that disagree with the submitted string fail closed. The route leads with its string-only boundary, labels resolved output as a description with no job admitted, and distinguishes one mounted route from four documented-only shapes. Focused backend tests passed 6/6; the combined shared-consumer component/page/route-map suite passed 20/20; and 10 retry-free Chromium scenarios passed with applicable axe coverage against the exact built BFF and a disposable database with all 32 migrations. The BFF build and migrated-database production-artifact smoke passed; the final 36,108,864-byte bundle has SHA-256 50e4a00c3ff930b33eac2d338d49c8d3623891b49e30a263441f0da3822be5ad. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the InstantID alias-analysis consumer and shared backend disposition, not the IP-Adapter or ControlNet browser consumers and not identity extraction, conditioning, generation, persistence, export, artifacts, governance, or release authority.

Authenticated Isis IP-Adapter hardening through 8fa00ff094c0a38f74da1635c52d8b08d8dcb1cb closes the second browser consumer of the shared request-local mode-alias analyzer. The exact no-store backend contract was already established at 359b806361; this consumer-only slice does not expand its authority. The browser now uses the same shared exact catalog as InstantID and independently recomputes each IP-Adapter alias's canonical and legacy-workflow names. Partial or authority-inflated catalogs and successful acknowledgements that disagree with the submitted string fail closed.

The route leads with its actual boundary: it receives no reference image or embedding, extracts no style or identity, runs no conditioning, admits no job, and creates no provider result, artifact, export, governance record, or release decision. It labels resolved output as a string-alias description with no job admitted and distinguishes one mounted analyzer from four documented-only route shapes. The combined shared-consumer component/page/route-map suite passed 22/22, and 10 retry-free Chromium scenarios passed with applicable axe coverage against the exact previously built BFF and a disposable database with all 32 migrations. The web typecheck and affected lint/formatting passed. No BFF source or bundle changed, so its build, production-artifact smoke, and digest remain recorded in the preceding shared-contract evidence. The web production build remains prohibited by the below-32-GiB host gate. This closes only the IP-Adapter alias-analysis consumer, not the ControlNet browser consumer and not image or embedding analysis, conditioning, generation, persistence, export, artifacts, governance, or release authority.

Authenticated Isis ControlNet hardening through 0ebba001c2d5a38ef38081db1ee5dd69759ab1f2 completes the three mounted browser consumers of the shared request-local mode-alias analyzer. The backend contract remains the exact no-store boundary established at 359b806361; this consumer-only slice does not add execution authority. One shared browser module now defines the exact catalog and complete source-defined alias semantics for ControlNet, IP-Adapter, and InstantID. The ControlNet consumer independently recomputes every result from the selected domain and submitted mode; partial or authority-inflated catalogs, cross-domain mappings, false legacy names, and invented preprocessing fields fail closed.

The route states that no control image is inspected, no detector or preprocessor runs, no conditioning graph executes, no job is admitted, and no provider result, artifact, export, governance record, or release decision exists. Resolved output is a string-alias description with no admitted job, and the route map distinguishes one mounted analyzer from four documented-only shapes. The combined three-consumer component/page/route-map suite passed 24/24, and 10 retry-free Chromium scenarios passed with applicable axe coverage against the exact previously built BFF and a disposable database with all 32 migrations. Web typecheck and affected lint/formatting passed. No BFF source or bundle changed, so its build, production-artifact smoke, and digest remain recorded in the preceding shared-contract evidence. The web production build remains prohibited by the below-32-GiB host gate. This closes only the three consumers' request-local alias-analysis boundary, not control-image analysis, detection, preprocessing, conditioning, generation, persistence, export, artifacts, governance, or release authority.

Authenticated Isis Florence-2 hardening through 9d02afc4e146bd1eb215f225478ab0361debbe5e binds the route to the only producer and transformation actually present. The endpoint receives caller-supplied image dimensions and pixel-space boxes, quantizes each coordinate onto the Florence-compatible 0–999 location-token grid, decodes those bins, and reports round-trip error. No image or pixels enter the endpoint, and repository tracing found no object detection, visual grounding, captioning, Florence model inference, job admission, provider execution, persistence, artifact, export, governance, or release connection.

GET and successful POST now expose the exact no-store request_local_non_authoritative_florence_location_token_codec disposition. It declares caller-supplied input, request-only retention, access-scope-only authority, deterministic source-defined coordinate quantization, and no image binding, detection, grounding, captioning, inference, job, provider, artifact, mutation, persistence, or authoritative state. Exact-key parsing rejects invented image, prompt, or execution fields. Dimensions are bounded to 100,000 pixels, requests to 256 boxes, and trimmed control-free unique identifiers to 100 characters.

The browser requires the exact catalog and independently recomputes every bin, location-token string, decoded coordinate, per-box error, and summary from the submitted dimensions and boxes. Authority-inflated catalogs and successful responses with false disposition or result semantics fail closed. The rendered surface leads with the coordinate-only boundary, says explicitly that it does not run Florence-2, labels results as request-local encoding with no inference or job, distinguishes one mounted route from four documented-only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 12/12. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/POST/400 no-store contracts, scope denial, browser-side suppression of out-of-contract input, catalog and result authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database artifact smoke passed; the final 36,110,871-byte entry bundle has SHA-256 5d9d0a14ac3e74ea73c8d3d166538f7dd692184a2797906dc559206da204b3c9. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local Florence-compatible coordinate codec, not image understanding, detection, grounding, captioning, model inference, generation, persistence, export, artifacts, governance, or release authority.

Authenticated Isis AnimateDiff Lightning hardening through 6a944df69b2d4d6e2e0cb1a56dee24968ebcdd38 binds the route to its actual pure arithmetic helper. The endpoint receives only caller-supplied inference-step count, training-timestep count, and spacing convention. It computes an integer timestep list and presentation fractions. No prompt, video, latent, model or scheduler instance, AnimateDiff checkpoint, denoising loop, job, provider, persistence, artifact, export, governance, or release connection exists.

GET and successful POST now expose the exact no-store request_local_non_authoritative_diffusion_timestep_schedule_preview disposition. It declares caller-supplied parameters, request-only retention, access-scope-only authority, deterministic source-defined timestep arithmetic, and no model, prompt, or video binding, inference, denoising execution, job admission, provider execution, artifact output, mutations, persistence, or authoritative state. Exact-key parsing rejects invented model or execution fields. Training timesteps are bounded to 100,000 and the response-producing inference-step count to the smaller of 1,000 or the training horizon.

The browser requires the exact catalog and independently recomputes trailing, leading, and linspace timestep arithmetic, fractions, and summaries from the submitted values. Authority-inflated catalogs and successful responses with false execution claims or arithmetic fail closed. The surface leads with its integer-only boundary, states that it does not run AnimateDiff Lightning, labels output as a request-local preview with no inference or job, distinguishes one mounted route from four documented-only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 12/12. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/POST/400 no-store contracts, scope denial, browser-side suppression of out-of-contract input, catalog and result authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database artifact smoke passed; the final 36,112,658-byte entry bundle has SHA-256 9a87f2c110ea3bdee6287f5861580025ecd9f01302fd09c1d167c221b16165e0. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local diffusion timestep arithmetic preview, not model binding, AnimateDiff inference, denoising, video generation, persistence, export, artifacts, governance, or release authority.

Authenticated Isis Live Preview hardening through b943dce68cb1a29a453974fae1d44ecefcd9f8e9 binds the route to its actual pure preview-step arithmetic. The endpoint receives only caller-supplied total-step and preview-budget integers, then selects hypothetical emission step numbers and display fractions. It does not observe a generation, bind a render, decode a frame, stream media, execute denoising, admit a job, call a provider, persist state, emit an artifact, export, govern, or release anything.

GET and successful POST now expose the exact no-store request_local_non_authoritative_preview_frame_schedule_preview disposition. It declares caller-supplied values, request-only retention, access-scope-only authority, deterministic source-defined preview-step arithmetic, and no generation or render binding, frame decode, streaming, denoising execution, job admission, provider execution, artifact output, mutation, persistence, or authoritative state. Exact-key parsing rejects invented stream or execution fields, and both step and preview counts are bounded to 10,000.

The browser requires the exact catalog and independently recomputes every hypothetical emission step, fraction, gap, and summary. Authority-inflated catalogs and successful responses claiming decoded or streamed frames or false arithmetic fail closed. The surface leads with the integer-only boundary, states that it does not observe or stream a live generation, labels output as a request-local preview with no frames or job, distinguishes one mounted route from four documented-only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 12/12. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/POST/400 no-store contracts, scope denial, browser-side suppression of out-of-contract input, catalog and result authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database artifact smoke passed; the final 36,114,125-byte entry bundle has SHA-256 a0c12a52fe80a8659324294710b8cd14830fa891ed3c402615db3c8268c0aebe. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local preview-step arithmetic, not live generation observation, frame decoding, streaming, denoising control, persistence, export, artifacts, governance, or release authority.

Authenticated Isis Topology Verification hardening through 26ab3cfe4ce30491b09351ba2266f01fcefc7ed7 binds the route to its actual summary-rule helper. The endpoint receives only caller-supplied vertex, face, degenerate-face, non-manifold-edge, watertight, and bounding-volume values plus a caller-selected watertight requirement. It does not receive geometry, inspect a mesh, ingest a file, bind an asset, run a model, repair anything, admit a job, call a provider, persist state, emit an artifact, govern, or release anything.

The fabricated clean and degenerate example meshes are removed. GET and successful POST now expose the exact no-store request_local_non_authoritative_topology_summary_evaluation disposition. It declares caller-supplied summary values, request-only retention, access-scope-only authority, deterministic source-defined summary-rule evaluation, and no mesh binding, geometry inspection, file ingestion, model inference, repair execution, job admission, provider execution, artifact output, mutation, persistence, or authoritative state. Exact-key parsing rejects invented mesh URLs, geometry, repair, or execution fields. Counts are bounded to 10,000,000, bounding volume to 1,000,000,000,000, degenerate-face count cannot exceed face count, and the capability catalog identifies vertex count explicitly as carried but unevaluated by the source helper.

The browser requires the exact catalog and independently recomputes the five summary rules, issue order, validity flag, and helper summary from the submitted values. Authority-inflated catalogs and successful responses claiming geometry or asset verification or false rule output fail closed. The surface leads with the summary-only boundary, labels a passing response as supplied values passing five rules rather than a valid mesh, states that the response is not proof of a valid mesh or artifact, distinguishes one mounted route from four documented- only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 12/12. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/POST/400 no-store contracts, scope denial, browser-side suppression of incoherent input, catalog and result authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and production-artifact smoke passed; the final 36,116,803-byte entry bundle has SHA-256 a333ac55ecb23015d3d7a6030738ff344a3308d9dfb1cfacbf19585902a53909. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only request-local evaluation of caller-supplied topology summary values, not geometry inspection, mesh or asset verification, repair, persistence, export, artifacts, governance, or release authority.

Authenticated Isis Multi-Provider 3D Generation hardening through fd84b5cfa3b8b23042eb5a0596bc87625090c543 binds the route to its actual pure boolean-to-action helper. The endpoint receives exactly four caller-supplied defect flags and maps them to ordered reference-image action names plus the helper's static warning. It receives no image, detects no defect, processes no pixels, selects or dispatches no provider, generates no 3D content, admits no job, persists no state, and emits no artifact.

GET and successful POST now expose the exact no-store request_local_non_authoritative_reference_image_defect_flag_action_preview disposition. It declares caller-supplied flags, request-only retention, access-scope-only authority, deterministic source-defined boolean-to-action mapping, and no image binding, defect detection, image processing, provider selection or dispatch, 3D generation, job admission, artifact output, mutation, persistence, or authoritative state. All four booleans are required; missing, extra, and non-boolean fields fail closed.

The browser requires the exact catalog and independently recomputes action order and warning from the submitted flags. It initializes every assertion to false rather than inventing detected defects. Authority-inflated catalogs and successful responses claiming processing, dispatch, generation, jobs, or artifacts fail closed. The surface labels inputs as caller assertions, distinguishes one mounted route from four documented-only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 12/12. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/POST/400 no-store contracts, scope denial, catalog and result authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database production-artifact smoke passed; the final 36,121,295-byte entry bundle has SHA-256 032dd3d8e45a24bb3674127b98be960b25f7f49006955eeed60f7bae6538e13b. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local defect-flag action preview, not image analysis or processing, provider selection or dispatch, 3D generation, jobs, persistence, export, artifacts, governance, or release authority.

Authenticated Isis ComfyUI Integration hardening through 440cfde358798fa522c1b51df544200f293462fc binds the surface to the two pure validators it actually calls. One endpoint checks four caller-supplied node-metadata strings; the other checks or coerces one caller-supplied scalar widget value against an INT, FLOAT, BOOLEAN, or STRING declaration. Neither helper registers a node, receives a graph or workflow, contacts a ComfyUI service, executes work, admits a job, persists state, emits an artifact, governs, or releases anything.

GET and both successful POST routes expose the exact no-store request_local_non_authoritative_comfyui_sdk_validation_preview disposition. It declares caller-supplied projections, request-only retention, access-scope-only authority, deterministic source-defined metadata/value validation, and no node registration, graph or workflow binding, ComfyUI service binding, execution, job admission, artifact output, mutation, persistence, or authoritative state. Exact-key parsing rejects invented provider, registration, or execution fields; numeric bounds are finite, ordered, and limited to INT/FLOAT declarations.

The browser requires the exact catalog and independently recomputes all metadata issues, scalar coercion, range errors, and validity. Authority-inflated catalogs and successful responses claiming registration or service binding, or returning false validation output, fail closed. The surface labels both inputs as caller supplied, describes a valid response as a request-local preview rather than registration, distinguishes one mounted route from four documented-only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 12/12. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/two-POST/400 no-store contracts, scope denial, catalog and result authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database production-artifact smoke passed; the final 36,123,015-byte entry bundle has SHA-256 b555d84f4e741590e010b033c97a0501a9c707c67d1d63a38f5b1cde749cc30c. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only request-local SDK validation of caller-supplied metadata and one scalar value, not ComfyUI integration, node registration, graph or workflow execution, persistence, export, artifacts, governance, or release authority.

Authenticated Isis Provider Integration hardening through 77f2505d6497ff36ec1da03c31036527df4bed93 binds the surface to its actual pure weight-normalization helper. The endpoint receives only caller-supplied provider labels, weights, and hypothetical health booleans, then calculates percentage shares and a sorted label order. It does not bind a registry or credential, probe health, select or dispatch a provider, execute failover, admit a job, persist state, emit an artifact, govern, or release anything.

GET and successful POST expose the exact no-store request_local_non_authoritative_provider_weight_routing_preview disposition. It declares caller-supplied values, request-only retention, access-scope-only authority, deterministic source-defined normalization/sorting, and no provider registry, credential, health-probe, selection, dispatch, failover, job, artifact, mutation, persistence, or authoritative-state binding. Exact-key parsing rejects invented provider URLs and execution fields. Provider count, identifier length, and finite positive weights are bounded; identifiers must be unique.

The browser requires the exact catalog and independently recomputes every share, sorted label, first label, routability boolean, and summary. It canonicalizes JSON object keys while preserving array order, so semantic exactness does not falsely depend on member order. Authority-inflated catalogs and successful responses claiming health probes or dispatch, false arithmetic, and authority-bearing input fail closed. The surface uses generic labels rather than inventing live vendor state, calls health/primary/failover values hypothetical, distinguishes one mounted route from four documented-only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, input, page, and route-map tests passed 14/14. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/POST/400 no-store contracts, scope denial, browser-side suppression of out-of-contract input, catalog/result authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database production-artifact smoke passed; the final 36,124,172-byte entry bundle has SHA-256 6c104b012a22e54b7a37d9cba6bd3e99db3685e6c6a5d10db65b1d901a768c3d. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only request-local arithmetic over caller-supplied provider labels, weights, and hypothetical health flags, not provider integration, monitoring, selection, dispatch, failover, persistence, export, artifacts, governance, or release authority.

Authenticated Isis Provider Factory hardening through 688552b625beaebe0fde7aaf40ff02802e062338 binds the surface to what its GET route actually does: deterministically project the local and runpod endpoint-path and default-configuration constants exported by @isis/ai-providers. The route does not call the provider factory, instantiate a provider, bind credentials, contact a service, validate an endpoint, observe health, capacity, or pricing, dispatch work, admit a job, persist state, emit an artifact, govern, or release anything. The previous six-provider tests and “registered cloud providers” copy were stale and authority-inflating; the source module implements exactly two provider types and explicitly rejects the other named backends.

The exact no-store source_defined_non_authoritative_provider_configuration_catalog disposition declares source-defined production, recomputable retention, access-scope-only authority, deterministic configuration projection, and no provider instantiation, credential binding, connectivity, endpoint validation, health, capacity, or pricing observation, dispatch, job, artifact, mutation, persistence, or authoritative state. The response has no generated timestamp and returns defensive copies of only the two source-defined entries.

The browser requires exact catalog equality while preserving array order, so a successful response that adds a provider or claims any operational authority fails closed. The surface labels all values as source defaults or configured estimates, says explicitly that no live provider exists, distinguishes one mounted route from four documented-only shapes, uses 44-pixel navigation targets, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 3/3 and focused web component, integration, page, and route-map tests passed 7/7. Eight isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/no-store contract, direct auth and scope denial, invented-provider and authority-inflation rejection, route honesty, touch targets, and descendant-level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated-database production-artifact smoke passed; the final 36,124,899-byte entry bundle has SHA-256 98ad80e4e525d342cb7bbced56feb2ba2c9f59a6939f78fae2ca0f3c96e3cf05. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the source-defined provider configuration catalog, not a provider factory, provider registry, credentials, live service connectivity or observation, dispatch, jobs, persistence, exports, artifacts, governance, or release authority.

Authenticated Isis Unified Provider Architecture hardening through ef059667cea789d21beb254c543ed5a63e2b5ce2 binds the surface to the three pure arithmetic helpers it actually calls. The endpoint receives a caller-authored hypothetical capability, input counts, requirements, and normalized weights, then estimates cost and latency and calculates a weighted score. It does not bind a registry, provider identity, credential, or live metric; select or dispatch a provider; admit a job; persist state; emit an artifact; govern; or release anything.

GET and successful POST expose the exact no-store request_local_non_authoritative_provider_capability_scoring_preview disposition. It declares caller-supplied values, request-only retention, access-scope-only authority, deterministic source-defined arithmetic, and no provider registry, identity, credential, live observation, selection, dispatch, job, artifact, mutation, persistence, or authoritative-state binding. Exact-key parsing rejects invented provider identifiers and execution fields. Money, latency, prompt length, image/video counts, success rate, and each weight are finite and bounded; weights must sum to one, preventing unbounded string/array allocation and non-normalized score inflation.

The browser requires the exact catalog, validates bounds before POST, and independently recomputes estimated cost, estimated latency, every score dimension, and the weighted total. Authority-inflated catalogs and successful responses claiming selection or dispatch, or returning false arithmetic, fail closed. The surface calls all inputs hypothetical and caller supplied, distinguishes one mounted route from four documented-only shapes, uses 44-pixel controls, and contains every visible descendant at 390 pixels.

Focused BFF tests passed 6/6 and focused web component, integration, page, and route-map tests passed 8/8. Ten isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact live GET/POST/400 no-store contracts, scope denial, browser-side bound suppression, catalog/result authority-inflation rejection, route honesty, touch targets, and descendant- level mobile containment. Web typecheck reached zero app-own errors; affected lint, formatting, and whitespace checks passed. The BFF build and migrated- database production-artifact smoke passed; the final 36,127,531-byte entry bundle has SHA-256 45ad0fbc258236b7a4064ce9f48f94b49bfe55dfb94e6db4738383d6cbd83a3d. The 6 GiB-scoped BFF ratchet reached only the same two unrelated current-main errors in autonomous-content-metrics-store.ts and autonomy-route.test.ts. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only request-local arithmetic over a caller-supplied hypothetical capability, not unified provider architecture, provider integration, observation, selection, dispatch, failover, jobs, persistence, exports, artifacts, governance, or release authority.

Authenticated Isis AI Providers hardening through 4c1841549f8ae76ffb26fa2402a1f34196e5ff47 establishes that the route is a UI-only composition shell around three independently bounded surfaces. It has no composite API or state. The prior control-plane, diagnostics/trace, security/audit, and API-parity cards were unsupported authority claims and have been replaced by exact descriptions of the source configuration catalog, caller-supplied hypothetical weight-routing arithmetic, and caller-supplied hypothetical capability-scoring arithmetic.

Each embedded lane preserves its own exact fail-closed contract: source-defined defaults do not become provider instances, caller health flags do not become observations or failover, and caller capability values do not become live metrics or provider selection. The composition summary explicitly disclaims a registry, control plane, credentials, diagnostics, telemetry, security policy, parity validation, dispatch, jobs, persistence, artifacts, governance, and release authority. One composition route is mounted; five provider, operation, diagnostic, parity, and governance paths are documented shapes only.

Focused component, integration, page, and route-map tests passed 6/6. Seven isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, all three live authenticated child GETs, both request-local child POSTs, scope denial across all lanes, absence of an invented composite BFF API, one-h1 semantics, 44-pixel controls, and descendant-level mobile containment. The first mobile run exposed and fixed a long documented-route label overflow; the full zero-retry rerun passed 7/7. Focused lint, formatting, whitespace, and web typecheck passed with zero app-own backlog. No BFF source changed, so the immediately preceding migrated and smoke-tested 36,127,531-byte artifact remains byte-identical at SHA-256 45ad0fbc258236b7a4064ce9f48f94b49bfe55dfb94e6db4738383d6cbd83a3d. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the AI Providers composition boundary, not a provider control plane, registry, integration, observation, selection, dispatch, failover, jobs, persistence, exports, artifacts, governance, or release authority.

Authenticated Isis Conversational AI hardening through e9e87f07402c9d2408d35632cd85731f98a38c61 establishes the exact disposition request_local_non_authoritative_reasoning_step_graph_structure_preview. The legacy route name does not describe a conversational runtime: the only mounted capability performs deterministic dependency validation, topological ordering, dependency-depth calculation, and declared-token estimation over a caller- supplied reasoning-step graph. It does not execute a model, expose hidden model reasoning, bind a provider, or create a conversation, session, transcript, realtime transport, speech input, voice output, tool invocation, memory, artifact, governance record, release decision, mutation, or authoritative state.

The BFF now publishes an exact no-store disposition and input contract, echoes the normalized request in an exact result envelope, and rejects extra, missing, malformed, duplicate-dependency, and allocation-inflating fields. Bounds are 1–20 steps, identifiers 1–10,000, trimmed instructions 1–2,000 characters, per-step declared tokens 1–100,000, at most 19 unique dependencies per step, and at most 2,000,000 declared tokens total. The browser acknowledges the exact catalog and exact result keys, input echo, types, bounds, and token arithmetic; it suppresses malformed shorthand before POST and rejects authority-inflated successful responses. One composition route is mounted. Its scene, export, revision, and governance paths are documented shapes only.

Six focused BFF contract/adversarial tests and eight focused web component, page, and route-map tests passed. Nine isolated Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. They prove the Studio hard block, anonymous redirect, exact catalog and request-local POST, client allocation suppression, BFF scope denial, catalog/result authority inflation rejection, direct BFF auth/exact-key/bound/no-store behavior, one-h1 semantics, 44-pixel controls, and descendant-level 390-pixel containment. The first harness attempt correctly exposed that production mode rejects development bearer tokens; a second attempt exposed an interrupted stale Next dev process. After isolating those harness conditions, the complete zero-retry run passed 9/9, and the final rebuilt artifact was rerun 9/9.

Focused lint, formatting, whitespace, and web typecheck passed with zero web backlog. The BFF ratchet reported only two unrelated pre-existing errors in the autonomous-content V10 union and agentic autonomy test callback; it introduced no slice-owned error. The exact 36,132,520-byte artifact built successfully, passed the production health/startup/shutdown smoke against all migrations, and has SHA-256 720012b0633c7646b6bb53ea4878300d20f2c1da3372af2db6b688d3bd055bef. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the bounded structure-preview seam, not conversational AI, provider/model execution, runtime reasoning, dialogue, speech, sessions, transcripts, tools, memory, persistence, artifacts, governance, or release authority.

Authenticated Isis 3D Prompt Engineering hardening through 86dc7efef2138d829992d92a30b1499f8948f063 establishes the exact disposition ephemeral_recomputable with no persistence, request-lifetime retention, engine revision isis-prompt-template-preview-v1, and no authoritative state. GET always returns exact empty history; POST performs one request-lifetime render through the prompt engine. The browser retains an accepted preview only in current-tab React state, and reload intentionally clears it.

The BFF bounds template keys to 1–128 grammar-conforming characters, names to 160 characters, content and serialized variables to 20,000 characters each, top-level variables to 100 keys, nesting to depth 8, and aggregate values to 1,000 nodes. It rejects prototype-pollution keys, template references, and unknown identity or retention fields. The client now requires exact GET, POST, disposition, and 11-field record shapes, rejecting authority-inflated fields such as providerExecution or retained. One route is mounted; four scene, export, revision, and governance paths are documented shapes only.

Eight focused BFF contract/adversarial tests and eleven focused web component, integration, page, and route-map tests passed. Eight isolated desktop/mobile Chromium scenarios passed with one worker, retries disabled, and applicable axe coverage against the exact built BFF and a disposable PostgreSQL database with all 32 migrations. Focused lint, formatting, whitespace, and web typecheck passed with zero web backlog. The BFF ratchet remained limited to the same two unrelated pre-existing autonomous-content V10 union and agentic autonomy test callback errors. No BFF source changed, and the exact artifact remained byte- identical at 36,132,520 bytes with SHA-256 720012b0633c7646b6bb53ea4878300d20f2c1da3372af2db6b688d3bd055bef. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-lifetime recomputable prompt-preview seam, not a reusable template registry, history, revision, approval, provider execution, artifact/export, governance, or release authority.

Authenticated Isis Gaussian Splatting hardening through 67434db2d9b570f85ceb2475425eb65c77afdaa2 establishes a genuine hybrid with two exact non-authoritative boundaries. The review workspace performs real browser-local PLY/SPLAT/IGSPLAT parsing, LOD selection, compression roundtrips, mesh extraction, animation, and WebGPU/fallback profiling in the current tab; it creates no server-side scene, artifact, provider-execution, retained telemetry, approval, governance, or release record. Its admin lane now declares request_local_non_authoritative_storage_footprint_preview, persistence none, retention request_only, caller-supplied splat-count/SH-degree inputs, a deterministic source-defined four-byte-float layout, and authoritative false. The exact bounds are 1–50,000,000 splats and SH degree 0–4. GET returns only {generatedAt, planner, inputBounds, disposition} and POST returns only {ok, input, result, disposition}; both server and client reject missing, extra, malformed, partially populated, incorrectly echoed, arithmetically inconsistent, or authority-inflated data. The route inventory now distinguishes the one mounted calculator from five documented-contract-only capabilities.

Focused evidence passed 6 BFF route tests, 15 web component/integration tests, and 9/9 production-artifact Playwright scenarios with one worker and zero retries across desktop Chromium and mobile Chrome, including auth and scope denial, exact envelopes, no-store behavior, malformed/inflated suppression, reload loss, browser-local renderer interactions, responsive touch containment, and applicable axe checks. The walk applied all 32 persistence migrations to a disposable database. Web TypeScript and production-file lint completed with zero diagnostics. The default 4 GiB BFF typecheck exhausted its heap; a supervised 6 GiB run reached the current remote baseline's 872 unrelated monorepo diagnostics, with no Gaussian Splatting file in the diagnostic set, so that broad ratchet is not claimed green. The exact BFF artifact is 36,134,008 bytes with SHA-256 e46631d997bdb0286cc637b6b0a95f22177ec9e5dcdf5936804f3a1fdce20137 and passed production health/startup/shutdown smoke. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the browser-local review tooling and request-local storage-footprint preview, not a durable scene registry, history, collaboration, artifact/export, provider, telemetry, governance, approval, or release authority.

Authenticated Isis Texture Enhancement hardening through 1775f1497593f7c2a39aa6fe92e71b5f480df46e establishes the exact disposition durable_internal_texture_quality_review_packages_recomputable_projection. This is genuine durable internal diagnostic history over authenticated operator-supplied fixture scores: normalized PostgreSQL batch and ordered fixture rows, authenticated-subject attribution, a recomputed and drift-checked projection, durable batch-id idempotency, canonical review_package retention for 2,555 days, durable redaction with a resurrection fence, legal holds, and retention purge. It does not ingest a texture asset, run an upstream scorer or provider, create an artifact, or grant governance, approval, or release authority.

GET now returns exactly {disposition, generatedAt, defaultThreshold, batches} and POST returns exactly {ok, batch, disposition}; authenticated scoped success and failure responses are no-store. The browser requires exact disposition, batch, fixture, and report shapes and rejects missing, extra, malformed, authority-inflated, incorrectly echoed, or arithmetically inconsistent successful responses. The route map distinguishes its one mounted diagnostic route from four documented-contract- only asset, artifact, revision, and governance shapes.

Focused BFF route and store evidence passed 12/12, focused web component, integration, page, and route-map evidence passed 12/12, and the real-database durability integration passed 2/2 after applying all 32 migrations. Five isolated Playwright scenarios passed 5/5 with one worker and zero retries across desktop Chromium and mobile Chrome against the exact production BFF artifact, a disposable PostgreSQL database, isolated Redis, and a signed production JWT. They prove the hard route boundary, durable authenticated write/reload/duplicate behavior, spoof rejection, exact route inventory, visible disposition, 410/503 failure handling, touch targets, responsive containment, and applicable axe coverage. Web TypeScript, focused lint, formatting, and whitespace checks passed with zero diagnostics. A supervised 6 GiB BFF typecheck reached 871 unrelated monorepo diagnostics with no Texture Enhancement file in the diagnostic set, so that broad ratchet is not claimed green. The exact 36,134,947-byte BFF artifact has SHA-256 5677e688ff01ff65c385772d6c10a6853a8f004b2dfae5aebc559a5eb8bdf96f and passed production health/startup/shutdown smoke against a fresh migrated database. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only durable internal fixture-score diagnostic history, not upstream asset scoring, texture-source ingestion, provider execution, artifact output, governance, approval, or release authority.

Authenticated Isis Topaz Enhancement hardening through 293e112c26b4af068303d19bdf5a9c18f030a25f establishes the exact disposition request_local_non_authoritative_film_aware_upscaling_plan_preview. The mounted surface performs deterministic model/backend selection, grain analysis, and pixel arithmetic over caller-supplied frame metadata and a source-defined catalog. It has no persistence, retains nothing beyond the request, and does not ingest frame assets, execute a GPU/provider, create an artifact, record governance, or authorize release.

The BFF accepts exactly 1–256 frames with unique trimmed 1–128-character ids, integer dimensions up to 2048×1556, exact source/content-profile enums, and nine finite normalized metrics, plus factor 2 or 4 and bounded supported optional preferences. It rejects missing, extra, duplicate, non-finite, out-of-envelope, and authority-inflated data. GET publishes an exact nine-field catalog, bounds, and disposition envelope; POST publishes exactly {ok, input, summary, grainAnalysis, disposition} with normalized input echo. Both authenticated scoped routes are no-store. The browser independently requires those exact shapes and verifies echo, enums, grain consistency, normalized scores, and pixel arithmetic before rendering a result. One route is mounted; four scene, export, revision, and governance shapes are documented contracts only.

Focused BFF route evidence passed 6/6, and focused web component, integration, page, and route-map evidence passed 9/9. Four isolated Playwright scenarios passed 4/4 with one worker and zero retries across desktop Chromium and mobile Chrome against the exact production BFF artifact, a disposable PostgreSQL database with all 32 migrations, isolated Redis, and a signed production JWT. They prove the hard route boundary, direct exact/no-store GET and POST, authority-inflated request and response rejection, honest route inventory, visible disposition, 44-pixel controls, responsive containment, and applicable axe coverage. The first cold run exposed a development-hydration delay; the warmed rerun then exposed and fixed a real 23-pixel selector touch target before the complete zero-retry run passed. Web TypeScript, production-file lint, formatting, and whitespace checks passed. The affected BFF route lint now passes with the repository's narrow canonical-domain-adapter rationale. A supervised 6 GiB BFF typecheck remained at 871 unrelated monorepo diagnostics, with no Topaz file in the diagnostic set, so that broad ratchet is not claimed green. The exact 36,141,156-byte BFF artifact has SHA-256 256221f97c1e0ca0c4b1a40f11561731463b23af76e54b1c2569382a0b176a0c and passed production health/startup/shutdown smoke. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local metadata-plan preview, not frame-asset ingest, GPU or provider execution, media artifacts, durable history, governance, approval, or release authority.

Authenticated Isis RIFE Interpolation hardening through a5cd7fd7d4ab1d21e6473e491484a870d3b32fe5 establishes the exact disposition request_local_non_authoritative_rife_interpolation_plan_preview. The mounted surface applies a deterministic RIFE planner to caller-supplied source/target rates and a synthetic uniform frame count. It has no persistence, retains nothing beyond the request, and does not ingest actual frame assets, execute a GPU/provider, create an artifact, record governance, or authorize release.

GET returns exactly {generatedAt, sourceFps, targetFps, inputBounds, disposition}. POST accepts exactly {sourceFps, targetFps, frameCount} and returns exactly {ok, input, plan, disposition} on success. The source catalog bounds source fps to 24 or 30, target fps to 48, 60, or 120, and frame count to 2–10,000. Authenticated scoped responses are no-store. The browser independently requires the exact catalog, disposition, normalized input echo, ten-field plan, six source-defined model variants, four supported backends, and the pair/generated frame arithmetic before rendering. Missing, extra, malformed, incorrectly echoed, arithmetically inconsistent, or authority-inflated successes fail closed. The route map distinguishes the one mounted planner from four documented-contract-only shapes.

Focused BFF route evidence passed 6/6, and focused web component, integration, page, and route-map evidence passed 12/12. Eight isolated Playwright scenarios passed 8/8 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome against the exact production BFF artifact, a disposable PostgreSQL database with all 32 migrations, isolated Redis, and signed production JWTs. They prove the hard route boundary, direct exact/no-store catalog and plan contracts, exact conversion arithmetic, invalid and unsupported rejection, malformed/inflated suppression, pending and network states, honest route status, scope and anonymous denial, 44-pixel touch targets, responsive containment, and applicable axe coverage. Web TypeScript, production-file lint, formatting, and whitespace checks passed. A supervised 6 GiB BFF typecheck remained at 871 unrelated monorepo diagnostics, with no RIFE file in the diagnostic set, so that broad ratchet is not claimed green. The exact 36,142,820-byte BFF artifact has SHA-256 75bd50f0373f7b6ff049e840bd207a40287d9492c867875b147bc34442d7ece4 and passed production health/startup/shutdown smoke. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local synthetic-frame plan preview, not actual frame ingest, GPU/provider execution, media artifacts, durable history, governance, approval, or release authority.

Authenticated Isis NVIDIA RTX Processing hardening through 36300aec34ae24ccd3d2c824d7354ce78dbf4365 establishes the exact disposition request_local_non_authoritative_four_x_super_resolution_plan_preview. The mounted surface applies the source-defined deterministic 4× planner to caller-supplied frame metadata. It has no persistence, retains nothing beyond the request, and does not ingest media, execute NVIDIA RTX or another GPU/provider, create an artifact, record governance, or authorize release.

GET returns exactly {generatedAt, scale, models, backends, contentProfiles, inputBounds, disposition}. POST accepts exact bounded frame metadata plus supported optional preferences and returns exactly {ok, input, result, disposition} on success. The contract limits batches to 1–256 unique frames, trimmed 1–128-character ids, integer dimensions up to 1920×1080, six finite normalized metrics, exact source-defined enums, and an optional trimmed 1–160-character batch name. Authenticated scoped responses are no-store. The browser independently requires the exact catalog, disposition, normalized input echo, model/frame/tile/summary shapes, source/output dimensions, 4× pixel arithmetic, aggregate totals, frame count, model/backend consistency, bounded scores, and positive estimates before rendering. Missing, extra, malformed, incorrectly echoed, arithmetically inconsistent, or authority-inflated successes fail closed. The route map distinguishes the one mounted planner from four documented-contract-only shapes.

Focused BFF route evidence passed 6/6, and focused web component, integration, page, and route-map evidence passed 8/8. Four isolated Playwright scenarios passed 4/4 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome against the exact production BFF artifact, a disposable PostgreSQL database with all 32 migrations, isolated Redis, and signed production JWTs. They prove the hard route boundary, direct exact/no-store catalog and plan contracts, adversarial request/response rejection, honest route status, visible disposition, responsive containment, a live result, and applicable axe coverage. The first complete run exposed and fixed real 26.5-pixel desktop controls; the final run proves 44-pixel inputs, selects, and submission control. Web TypeScript, production-file lint, formatting, and whitespace checks passed. A supervised 6 GiB BFF typecheck produced 747 primary pre-existing TypeScript diagnostics across 876 output lines, with no NVIDIA RTX file in the diagnostic set, so that broad ratchet is not claimed green. The exact 36,145,691-byte BFF artifact has SHA-256 49ccd04564f0155dd9b9d101b9eb510af64f3cada60356b134fc23d4dd040cc3 and passed production health/startup/shutdown smoke. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local metadata-plan preview, not frame-asset ingest, GPU/provider execution, media artifacts, durable history, governance, approval, or release authority.

Authenticated Isis Reference Image Processing hardening through f5f31ce2366d2b0911680adb0d3534e8c1ee5b79 establishes the exact disposition request_local_non_authoritative_reference_aspect_ratio_bucket_preview. The mounted surface applies the source-defined deterministic Seedance aspect-ratio bucketer to caller-supplied width and height metadata. It has no persistence, retains nothing beyond the request, and does not ingest or decode an image, transform media, execute a provider or generation, create an artifact, record governance, or authorize release.

GET returns exactly {generatedAt, supportedAspectRatios, inputBounds, disposition}. POST accepts exactly {width, height} and returns exactly {ok, input, result, disposition} on success. Both dimensions must be integers from 1 through 16,384, and the exact supported catalog is 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Authenticated scoped responses are no-store. The browser independently requires the exact catalog, bounds, disposition, input echo, raw width/height ratio, and allowlisted snapped ratio before rendering. Missing, extra, malformed, incorrectly echoed, arithmetically inconsistent, or authority-inflated successes fail closed. The route map distinguishes the one mounted planner from four documented-contract-only shapes.

Focused BFF route evidence passed 5/5, and focused web component, integration, page, and route-map evidence passed 9/9. Four isolated Playwright scenarios passed 4/4 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome against the exact production BFF artifact, a disposable PostgreSQL database with all 32 migrations, isolated Redis, and signed production JWTs. They prove the hard route boundary, direct exact/no-store catalog and plan contracts, adversarial request/response rejection, honest route status, visible disposition, responsive containment, a live result, 44-pixel controls, and applicable axe coverage. The first cold run timed out while the development server compiled the page; the warmed complete run passed without a product error. Web TypeScript, production-file lint, formatting, and whitespace checks passed. A supervised 6 GiB BFF typecheck produced 747 primary pre-existing TypeScript diagnostics across 876 output lines, with no Reference Image Processing file in the diagnostic set, so that broad ratchet is not claimed green. The exact 36,147,231-byte BFF artifact has SHA-256 64cd86da38c940126372d42dea1bb3ab6b3837beb3de9732fb926953002f3f96 and passed production health/startup/shutdown smoke. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local dimension-bucket preview, not image-asset ingest, decoding, transformation, provider/generation execution, media artifacts, durable history, governance, approval, or release authority.

Authenticated Isis Video Production Workflows hardening through 8d37570cd1fe761274cc593adfa9e218348aae9d establishes the exact disposition request_local_non_authoritative_workflow_schema_validation_preview. The mounted surface applies the source-defined deterministic @isis/workflows schema validator to a caller-supplied workflow definition. It has no persistence, retains nothing beyond the request, and does not register or execute a workflow, invoke a provider, create an artifact, record governance, publish, or authorize release.

GET returns exactly {generatedAt, canonicalNodeFormat, requiredFields, inputBounds, disposition}. POST accepts exactly {workflow} and returns exactly {ok, input, result, disposition} on success. The workflow must be an object whose JSON serialization is 2–262,144 UTF-8 bytes and whose object-map nodes number 0–256. Authenticated scoped responses are no-store. The browser independently requires the exact catalog, bounds, disposition, normalized input byte/node summary, and internally consistent valid or invalid result before rendering. Missing, extra, malformed, incorrectly summarized, or authority- inflated successes fail closed. The route map distinguishes the one mounted validator from four documented-contract-only shapes.

Focused BFF route evidence passed 6/6, and focused web component, integration, page, and route-map evidence passed 10/10. Eleven isolated Playwright scenarios passed 11/11 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome against the exact production BFF artifact, a disposable PostgreSQL database with all 32 migrations, isolated Redis, and signed production JWTs. They prove the hard production gate, direct exact/no-store catalog and error contracts, valid and invalid live validation, input and authority-inflation rejection, honest route status, visible disposition, loading/error/pending/offline/auth states, responsive containment, 44-pixel controls, and applicable axe coverage. The first cold run passed 10/11; its catalog scenario exhausted the 60-second test budget after a 58-second Next development compile even though the logged BFF request returned 200 in 4 ms. The scenario now has an explicit 90-second cold-compile budget, and the complete warmed zero-retry run passed. Web TypeScript, production-file lint, strict commit-hook lint, formatting, and whitespace checks passed. A supervised 6 GiB BFF typecheck produced the unchanged 747 primary pre-existing TypeScript diagnostics across 876 output lines, with no Video Production Workflows file in the diagnostic set, so that broad ratchet is not claimed green. The exact 36,149,854-byte BFF artifact has SHA-256 7a25a0487be99f165d2438807ac33f2ed22ae84ad8cce7f9e8ae0ea6ecd64751 and passed production health/startup/shutdown smoke. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local schema-validation preview, not registration, workflow/provider execution, media artifacts, durable history, governance, approval, publication, or release authority.

Isis Anomaly Detection aggregator hardening through 6163d3c571e6a056b19f5883d58d9fe9e995f5bb removes the false cross-lane control- plane, threshold, remediation, diagnostics, parity, audit, governance, and release-signoff presentation. The route is now explicitly ui_only_non_authoritative_embedded_lane_composition: it mounts five independent detection/quality workspaces, delegates persistence and authority to each lane's exact contract, issues no anomaly-aggregator API request, retains no cross-lane state, and creates no incident, diagnostic, governance, approval, or release record.

The visible Composition Boundary identifies independent lane authority, the absence of an aggregator endpoint, the absence of cross-lane operations, and the absence of governance/release authority. The route map now distinguishes one mounted-composition route from five documented-contract shapes and no longer describes nonexistent signal triage, incidents, diagnostics, parity runs, audit records, or release workflows as mounted. Focused component, integration, page, and route-map tests passed 6/6. Isolated Playwright passed 2/2 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome with applicable axe coverage. It proves the explicit disposition, five embedded lanes, no direct aggregator endpoint request, honest route statuses, 44-pixel quick actions, and responsive containment. The first browser run returned the real production hard-block 404 and exposed a stale walkthrough assumption; the corrected localhost-only Isis lane entry then passed the full desktop/mobile matrix. Web TypeScript, production-file lint, strict commit-hook lint, formatting, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the UI composition seam, not cross-lane correlation, thresholding, signals, incidents, diagnostics, parity, remediation, audit/governance, approval, or release authority.

Isis Automated Quality Scoring hardening through 3fb616f4ce807ab91a654faa3a0aa8ef7d305b5f establishes the exact disposition deterministic_fixture_regression_read_model. The mounted admin surface runs the source-defined @isis/workflows golden-output perceptual regression suite over its built-in deterministic fixture dataset and reports its suite-local threshold verdict, metrics, and per-case breaches. It does not accept customer assets, persist runs, produce signed evidence, or authorize a production release.

The BFF exposes exactly one admin-scoped, fail-closed GET /v1/admin/isis/quality-scoring; refresh remains an idempotent GET and no POST compatibility write is registered. The browser independently requires version and timestamp fields, finite summary metrics, integer and partition count consistency, string-only failure arrays, and complete per-case fields before rendering. Malformed successes fail closed with an explicit retry. The route map distinguishes one mounted-read-model route from four documented-contract shapes and no longer presents nonexistent scene history, exports, revisions, rollback, governance, approval, or release signoff as mounted authority.

Focused web component, page, and route-map evidence passed 7/7. The real BFF route evidence passed 4/4, including authentication/scope denial, source-defined report execution, deterministic summary stability, and explicit absence of POST. The underlying golden-output engine passed 3/3. Isolated Playwright passed 4/4 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome with applicable axe coverage. It proves the direct production hard-block, localhost-only lane entry, visible fixture-only disposition, ready/filter/refresh flow, GET-only request discipline, denied and malformed fail-closed states, honest route statuses, 44-pixel controls, and responsive containment. The first browser pass completed 3/4 and exposed React development Strict Mode's duplicate effect loads; the corrected assertion proves the actual contract invariant—refresh increases reads and every request is GET—and the complete final run passed. Web TypeScript, production-file lint, strict commit-hook lint/convention/formatting ratchets, formatting, and whitespace checks passed. The commit hook's broad typecheck was explicitly skipped because the BFF has a known unrelated TypeScript baseline; focused BFF, engine, web unit, web typecheck, and browser checks are green. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the deterministic fixture-report read model, not customer-asset evaluation, durable run history, export or signed evidence, governance, approval, or release authority.

Isis Generation API aggregator hardening through acb87a44566c5114549611424eb121b701691868 removes the fabricated 36-workflow control plane, forced-failure toggles, client-only pass/fail runs, parity certification, 54 synthetic alert channels, 37 nonexistent runbook links, release actors, governance actions, and endpoint-operation claims. The route is now explicitly ui_only_non_authoritative_embedded_lane_composition: it mounts five independent client SDK, CLI, integration-event, content-safety, and resource-recommendation workspaces, delegates authority to each child contract, issues no Generation API aggregator request, and retains no shared operational state.

The visible Composition Boundary identifies independent lane authority, the absence of an aggregator endpoint, the absence of synthetic operations, and the absence of governance/release authority. The route map distinguishes one mounted-composition route from five documented-contract shapes and no longer presents nonexistent requests, jobs, evidence, parity, governance, approval, or release workflows as mounted. Focused component, integration, page, and route-map tests passed 5/5. Isolated Playwright passed 2/2 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome with applicable axe coverage. It proves the explicit disposition, five embedded lanes, no direct aggregator endpoint request, absence of synthetic workflow/parity/telemetry controls, honest route statuses, 44-pixel quick actions, and responsive containment. Web TypeScript, production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the UI composition seam, not generation requests, job execution, evidence records, parity certification, cross-lane telemetry, governance, approval, or release authority.

Isis GPU Worker aggregator hardening through dfe81bafdc3fc4613be10fc6d6fa2c54c50740a9 removes the false worker-control- plane, scheduling, execution, diagnostics, trace, permission/audit, parity, remediation, and release-assurance presentation. The route is now explicitly ui_only_non_authoritative_embedded_lane_composition: it mounts five independent Generation API, CLI, integration-event, resource-recommendation, and suspicious-activity workspaces, delegates authority to each child contract, issues no GPU Worker aggregator request, and retains no fleet, pool, queue, job, or other shared operational state.

The visible Composition Boundary identifies independent lane authority, the absence of an aggregator endpoint, the absence of worker operations, and the absence of governance/release authority. The route map distinguishes one mounted-composition route from five documented-contract shapes and no longer presents nonexistent jobs, pools, evidence, parity, governance, approval, or release workflows as mounted. Focused component, integration, page, and route- map tests passed 6/6. Isolated Playwright passed 2/2 with one worker and zero retries across desktop Chromium and Pixel 7 mobile Chrome with applicable axe coverage. It proves the explicit disposition, five embedded lanes, no direct aggregator endpoint request, honest route statuses, 44-pixel quick actions, and responsive containment. The first focused run had one over-broad text assertion that surfaced separate parity wording inside the nested CLI child; the assertion was correctly scoped to GPU Worker's own boundary, leaving that child debt visible for its own audit, and the complete rerun passed. Web TypeScript, production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the UI composition seam, not worker fleets, pools, queues, jobs, scheduling/execution, observability evidence, parity certification, governance, approval, or release authority.

Isis CLI aggregator hardening through e87be73e81db0e25a5b4a1c90e46912cdf3242f3 removes the false CLI control-plane, command-action, diagnostics/trace, permission/audit, parity/remediation, and release-assurance presentation. The route is now explicitly ui_only_non_authoritative_embedded_lane_composition: it mounts five independent CLI Surface, Client SDK, Integration Events, Account Protection, and Chargeback Prevention workspaces, delegates authority to each child contract, issues no CLI aggregator request, and retains no shared command, session, execution, parity, governance, or release state.

The visible Composition Boundary identifies independent lane authority, the absence of an aggregator endpoint, the absence of command operations, and the absence of governance/release authority. The route map distinguishes one mounted-composition from five documented-contract route shapes without inventing pages or authority. Six focused Vitest component, integration, and route-map tests pass. Two Playwright projects pass with zero retries across desktop Chromium and Pixel 7 mobile Chrome with applicable axe coverage. They prove the explicit disposition, four boundary cards, five embedded lanes, no direct /v1/admin/isis/cli request, honest route statuses, 44-pixel quick actions, and responsive containment. The first focused run had one substring assertion collision because the legitimate child endpoint /v1/admin/isis/cli-surface contains the aggregator path; the assertion was correctly changed to exact request-path equality, and the complete rerun passed. Web TypeScript, production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the UI composition seam, not shared command/session state, command execution, diagnostic or trace evidence, permission/audit controls, parity certification, remediation, governance, approval, or release authority.

Isis CLI Surface hardening through a08a88adbd7ccaeadded14f16bfd8e0f3b054ec3 establishes the exact disposition request_local_non_authoritative_cli_token_resolution_preview. The admin-scoped, fail-closed BFF deterministically classifies caller-supplied query tokens against a caller-supplied command registry by exact name, alias, unique prefix, ambiguity, or absence. It loads no canonical command catalog, executes no command, retains no session, and creates no diagnostic, trace, permission, audit, governance, approval, or release evidence.

The catalog and resolution envelopes now carry the exact disposition and reject malformed or authority-inflated success responses. The route accepts exactly commands and queries, caps requests at 256 commands, 32 aliases per command, 1,024 queries, and 128 characters per whitespace-free token, rejects extra execution-like fields, and explicitly returns cache-control: no-store. The visible Resolution Boundary identifies the caller-supplied registry, request-only computation, absence of command execution, and absence of operational authority. The route map distinguishes one mounted-preview from four documented-contract shapes without inventing pages or authority.

Six focused BFF route tests and ten focused web component, integration, live-response, page, and route-map tests pass. Two Playwright projects pass with zero retries across desktop Chromium and Pixel 7 mobile Chrome with applicable axe coverage. They prove the production hard-block, isolated fresh BFF contract, exact no-store envelopes, authority-inflated request rejection, deterministic six-query output, four visible boundary cards, honest route statuses, 44-pixel submit control, keyboard-accessible result scrolling, and responsive containment. Early browser attempts first exposed a stale shared BFF whose months-old route table returned 404 and then an isolated-harness auth-seeding mismatch; separating shared shell authentication from the fresh CLI route server corrected the harness without weakening assertions. The first valid desktop run passed while mobile axe found a serious scrollable-region-focusable defect in the results table; the region received a named keyboard focus target, and the complete zero-retry rerun passed. Web TypeScript, targeted production lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The BFF package typecheck remains non-green on its documented unrelated Arete/Yemaya monorepo baseline after a bounded 6 GiB heap run; the focused changed-route compile/runtime suite is green. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local token-resolution preview, not canonical catalog authority, command/session execution, diagnostic or audit evidence, permission controls, governance, approval, or release authority.

Isis Output Registry hardening through 6f5b5c5504f76c58c73ce72e024f782d31c18813 removes false registry control-plane, diagnostics, permission/audit, parity/remediation, publication, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: it composes five independent child workspaces under their own exact contracts, sends no Output Registry aggregator request, and retains no registry entry, package, artifact, provenance, publication, or release state.

The hard-coded GENERATED_ARTIFACT_PROVENANCE_RECORDS panel was removed from this operational wrapper because source-defined demonstration fixtures are not live or durable provenance, publication records, or evidence authority. Four boundary cards now state the independent-lane, no-endpoint, no-provenance/publication, and no-governance/release limits. The route map distinguishes the single mounted-composition from five documented-contract shapes that have no pages or corresponding authority.

Five focused component, route-map, and integration assertions passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production AAA hard-block, localhost-only lane entry, absence of the static provenance fixture and /v1/admin/isis/output-registry request, 44-pixel mobile navigation targets, responsive containment, route status, and applicable axe rules. Web TypeScript, targeted production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the honest UI composition seam, not a standalone registry, durable provenance source, publication workflow, evidence record, parity certification, governance, approval, or release authority.

Isis Web aggregator hardening through 413bc80691c0419f28af0d835a162364f668d164 removes the false Web control plane, runtime-control, diagnostics/trace, permission/audit, parity/remediation, evidence, rollout, governance, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: it composes five independent Generation API, GPU Worker, Output Registry, Client SDK, and CLI workspaces under their own exact contracts, sends no Isis Web aggregator request, and retains no session, workflow, release, evidence, parity, governance, or other shared operational state.

Four boundary cards state the independent-lane, no-endpoint, no-operations/ evidence, and no-governance/release limits. The route map distinguishes the single mounted-composition from five documented-contract session, release, evidence, parity, and governance shapes that have no pages or corresponding authority.

Five focused component, route-map, page, and integration assertions passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production AAA hard-block, localhost-only lane entry, five embedded lanes, exact absence of /v1/admin/isis/web, honest route statuses, 44-pixel mobile navigation targets, responsive containment, and applicable axe rules. The first isolated browser bootstrap stopped before test execution because a fresh BFF detected the unrelated missing pre-existing account_protection_subject_state table; the corrected fresh Web-only harness used the healthy shared shell-auth BFF while intercepting every child admin request. The first valid run passed mobile and exposed only an invalid test assumption that at least one child must fetch; removing that assumption preserved the exact no-aggregator invariant, and the complete rerun passed. Web TypeScript, targeted production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the honest UI composition seam, not a Web control plane, retained session store, runtime operation, diagnostic/evidence record, parity certification, governance, approval, or release authority.

Isis Database aggregator hardening through 8057c72c96e771767846d5a9f031676fb05b037f removes the false database control plane, cluster/snapshot, persistence-control, backup/restore, diagnostics/trace, permission/audit, parity/remediation, governance, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: it composes five independent Output Registry, Model Registry, Collection Management, Lineage Graph, and Retention Policies workspaces under their own exact contracts, sends no Isis Database aggregator request, and retains no cluster, snapshot, backup, restore, diagnostic, parity, governance, or other shared operational state.

This boundary does not erase the real @isis/database Prisma-backed backend substrate; it records that the Studio wrapper neither imports nor administers it. Four boundary cards state the independent-lane, no-endpoint, no-database-operations/evidence, and no-governance/release limits. The route map distinguishes the single mounted-composition from five documented-contract cluster, snapshot, diagnostics, parity, and governance shapes that have no pages or corresponding authority.

Seven focused component, route-map, page, and integration assertions passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production AAA hard-block, localhost-only lane entry, five embedded lanes, exact absence of /v1/admin/isis/database, honest route statuses, 44-pixel mobile navigation targets, responsive containment, and applicable axe rules. Web TypeScript, targeted production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the honest UI composition seam, not a Database control plane, cluster/snapshot store, backup/restore workflow, diagnostic/evidence record, parity certification, governance, approval, or release authority.

Isis Workflow Registry hardening through 050e5d67900c04110ab645cce0a4e74cd32c29f2 removes the false registry control plane, workflow/version lifecycle, orchestration/migration, diagnostics/trace, permission/audit, parity/remediation, evidence, governance, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: it composes five independent Generation API, GPU Worker, Output Registry, Web, and Integration Events workspaces under their own exact contracts, sends no Workflow Registry aggregator request, and retains no workflow, version, migration, evidence, parity, governance, or other shared operational state.

Four boundary cards state the independent-lane, no-endpoint, no-orchestration/evidence, and no-governance/release limits. The route map distinguishes the single mounted-composition from five documented-contract workflow, version, evidence, parity, and governance shapes that have no pages or corresponding authority.

Five focused component, route-map, page, and integration assertions passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production AAA hard-block, localhost-only lane entry, five embedded lanes, exact absence of /v1/admin/isis/workflow-registry, honest route statuses, 44-pixel mobile navigation targets, responsive containment, and applicable axe rules. The first valid browser run passed mobile while desktop rendered the correct deeply nested composition but exceeded the generic 60-second budget; an explicit 120-second budget for that one deep proof retained one worker and zero retries, and the complete rerun passed in under a minute. Web TypeScript, targeted production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the honest UI composition seam, not a Workflow Registry, retained workflow/version store, orchestration/migration workflow, diagnostic/evidence record, parity certification, governance, approval, or release authority.

Isis Event Publisher hardening through d43d8ecbf1b3c9083a6c1e8facc769966e1a0dd6 removes the false publisher control plane, topic/stream lifecycle, publish/retry/delivery, diagnostics/trace, permission/audit, parity/remediation, evidence/receipt, governance, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: it composes four independent Integration Events, Workflow Registry, Generation API, and Output Registry workspaces under their own exact contracts, sends no Event Publisher aggregator request, and retains no topic, stream, delivery, retry, receipt, evidence, parity, governance, or other shared operational state.

Four boundary cards state the independent-lane, no-endpoint, no-publication/delivery/evidence, and no-governance/release limits. The route map distinguishes the single mounted-composition from five documented-contract topic, stream, diagnostics, parity, and governance shapes that have no pages or corresponding authority.

Seven focused component, route-map, page, and integration assertions passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production AAA hard-block, localhost-only lane entry, four embedded lanes, exact absence of /v1/admin/isis/event-publisher, honest route statuses, 44-pixel mobile navigation targets, responsive containment, and applicable axe rules. Web TypeScript, targeted production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the honest UI composition seam, not an Event Publisher, retained topic/stream/delivery store, publish/retry workflow, receipt/evidence record, parity certification, governance, approval, or release authority.

Isis Outputs hardening through f7d487a866380cea73b4009b0af25ee60f3b9053 removes the false Outputs control plane, artifact/publication lifecycle, diagnostics/trace, permission/audit, parity/remediation, provenance/evidence, governance, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: it composes five independent Output Registry, Output Manifest, Collection Management, Retention Policies, and Integration Events workspaces under their own exact contracts, sends no Outputs aggregator request, and retains no aggregate artifact, publication, provenance, evidence, parity, governance, or other shared operational state.

Four boundary cards state the independent-lane, no-endpoint, no-publication/provenance, and no-governance/release limits. The route map distinguishes the single mounted-composition from five documented-contract artifact, release, diagnostics, parity, and governance shapes that have no pages or corresponding authority. The source-defined generated-artifact provenance fixture was removed from this operational wrapper because it was neither live nor durable evidence.

Seven focused component, route-map, page, and integration assertions passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production AAA hard-block, localhost-only lane entry, five embedded lanes, exact absence of /v1/admin/isis/outputs, absence of the static provenance fixture, honest route statuses, 44-pixel mobile navigation targets, responsive containment, and applicable axe rules. The deeply nested desktop composition completed in 1.7 minutes within its explicit 120-second test budget; mobile completed in 8.8 seconds. Web TypeScript, targeted production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the honest UI composition seam, not an Outputs control plane, aggregate artifact/publication store, provenance/evidence record, parity certification, governance, approval, or release authority.

Isis Client hardening through d96b17d57316842f1035a515e4ca260982cb8f34 removes the false Client control plane, session/credential lifecycle, delivery controls, diagnostics/trace, permission/audit, parity/remediation, governance, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: it composes four independent Client SDK, CLI, Generation API, and Integration Events workspaces under their own exact contracts, sends no Client aggregator request, and retains no aggregate session, credential, delivery, diagnostic, audit, parity, governance, or other shared operational state.

Four boundary cards state the independent-lane, no-endpoint, no-session/delivery, and no-governance/release limits. The route map distinguishes the single mounted-composition from five documented-contract session, release, diagnostics, parity, and governance shapes that have no pages or corresponding authority.

Seven focused component, route-map, page, and integration assertions passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production AAA hard-block, localhost-only lane entry, four embedded lanes, exact absence of /v1/admin/isis/client, honest route statuses, 44-pixel mobile navigation targets, responsive containment, and applicable axe rules. The first browser probe correctly returned the production 404 hard block and exposed a stale walkthrough assumption; the valid proof then asserted the hard block before using only the explicit localhost E2E bypass. Desktop completed in 36.8 seconds and mobile in 8.4 seconds. Web TypeScript, targeted production-file lint, strict commit-hook lint/convention/formatting ratchets, whitespace, and the zero-backlog affected web typecheck ratchet passed. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the honest UI composition seam, not a Client control plane, retained session/credential store, delivery workflow, diagnostic/audit record, parity certification, governance, approval, or release authority.

Isis Job Envelope hardening through 24096747101bbe52b177cc27c30c4fc972cc425e preserves the real admin-scoped, pure @isis/job-envelope field-normalization engine while removing false control-plane, dispatch/lifecycle, diagnostics/trace, permission/audit, parity/remediation, governance, and release claims. The wrapper now declares the exact request_local_non_authoritative_normalization_preview_with_embedded_lane_composition disposition: caller-supplied aliases are normalized to canonical tokens without retaining inputs or results, validating a complete envelope, enqueueing or dispatching work, retrying, remediating, certifying parity, creating evidence, governing, approving, or authorizing release. Five embedded lanes retain their own exact contracts and authority.

Seven focused component, route-map, page, and integration assertions passed, along with production lint, full web TypeScript, strict commit-hook ratchets, and whitespace checks. The focused Chromium authority proof passed with zero retries, proving the production hard block, localhost-only bypass, real normalization form, exact disposition, and one mounted-hybrid-tool plus five documented-contract route statuses. The original 462-line live-BFF suite was preserved and extended after a pre-push review caught an accidental replacement. Its hard-block and anonymous cases pass; its retained live catalog/normalize cases currently receive 404 from the shared BFF for all Isis admin routes, an environment route-state limitation recorded without weakening those tests. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local normalization and composition boundary, not a durable job-envelope store, dispatcher, diagnostic/evidence record, parity certification, governance, approval, or release authority.

Isis AI Video hardening through e913550e41ccdb2d5f5b262450e0bc973eb464a2 removes the false control-plane, project/shot lifecycle, render/diagnostic, permission/audit, parity/remediation, governance, and release claims. The route now declares the exact ui_only_non_authoritative_embedded_lane_composition disposition: five independent AI Video Generation, Video Generation, Video Production Workflows, AnimateDiff Lightning, and RIFE Interpolation lanes retain their own contracts, while the wrapper sends no /v1/admin/isis/ai-video request and retains no shared operational state. One mounted composition and five documented route shapes are labeled explicitly.

Five focused assertions, production lint, full web TypeScript, strict commit-hook ratchets, and whitespace checks passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production hard block, localhost-only entry, five lanes, endpoint absence, route statuses, 44-pixel mobile navigation targets, responsive containment, and applicable axe rules. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the UI composition seam, not an AI Video control plane, project/shot store, render or diagnostic evidence, parity certification, governance, approval, or release authority.

Isis ComfyUI SDK hardening through 3c133f64a9a7466cf4fd3d7dead75e0ebe020afc preserves the real admin-scoped semantic-version engine while removing false control-plane, package/release, distribution/diagnostic, permission/audit, parity/remediation, governance, and release claims. The wrapper now declares the exact request_local_non_authoritative_semver_evaluation_with_embedded_lane_composition disposition: caller-supplied versions are parsed, compared, range-tested, and bumped without retention, package publication, distribution, migration, certification, governance, approval, or release authority. Five embedded lanes retain their own contracts.

Seven focused assertions, production lint, full web TypeScript, strict commit-hook ratchets, and whitespace checks passed. Two zero-retry Playwright projects passed on desktop Chromium and Pixel 7, proving the production hard block, localhost-only entry, real semver form, exact disposition, one mounted-hybrid-tool, five documented-contract routes, 44-pixel mobile targets, responsive containment, and applicable axe rules. The web production build remains prohibited by the 15 GiB host's below-32-GiB resource gate. This closes only the request-local semver and composition boundary, not a package registry, distribution pipeline, diagnostic/evidence record, parity certification, governance, approval, or release authority.

This remains progress evidence only. The model-governance license/admission, Iris desktop-to-mobile handoff, four Studio authoring-record snapshot partitions, Mentor Presence rendered performances, the Tara scene-publish receipt authority, Isis resource-recommendation history, Isis intelligent- routing history, Isis cost-tracking history, Isis cost-optimization seam, Isis prompt-preview ephemeral/recomputable seam, and Isis managed-model registry unbound projection, Isis Civitai intake/review unbound projection, and Isis model-merging unbound projection, Isis voice-cloning unbound projection, and Isis ComfyUI node source-defined projection, and Isis RunPod-operations unbound projection, Isis LoRA-training unbound projection, and Isis suspicious-activity durable read-only projection, and Isis creator-monetization ephemeral/ recomputable planner, Isis budget-management unbound projection, and Isis 3D-retention request-only evaluator, Isis pipeline-class unbound projection, and Isis 3D-provenance unbound projection are now closed within their stated boundaries, as are the Isis output-lineage unbound projection and the Isis model-comparison durable-catalog/recomputable-layout boundary, plus the Isis workflow-template source-defined/recomputable inspector and the Isis workflow-definition source catalog and recomputable workflow-version analysis, plus the Isis Provider Factory source-defined configuration catalog. The Isis workflow-portability caller-supplied/recomputable calculator is also closed within its stated boundary, as is the Isis integration-event caller- supplied/recomputable schema validator and the Isis resource-recommendation durable review-package history, plus the Isis intelligent-routing durable decision-package history and the Isis job-submission request-local envelope builder, plus the Isis job-management durable operator queue and safe pre-dispatch retry boundary, and the Isis audio-generation request-local deterministic RMS analyzer, plus the Isis music-generation request-local non-authoritative guardrail evaluator, and the Isis AI-video request-local non-authoritative frame-count planner, plus the Isis video-generation source-defined/recomputable manifest inspector, plus the Isis image-generation source-defined/recomputable manifest inspector, plus the Isis text-to-audio source-defined/recomputable manifest inspector, plus the Isis 3D-result request- local non-authoritative topology analyzer, plus the Isis account-protection durable advisory screening ledger and its exact rendered admin boundary, plus the Isis Analytics and Reporting request-local non-authoritative retry-analysis and deterministic backoff-preview boundary, plus the Isis 3D Editing browser- local workspace and request-local non-authoritative repair-simulation boundary. The Isis 3D Generation request-local non-authoritative structured-prompt preview is also closed within its stated boundary, as is the Isis Auto-Rigging browser-local demo and request-local non-authoritative humanoid bone-map analysis. All three Isis ControlNet, IP-Adapter, and InstantID request-local non-authoritative mode-alias consumers are also closed within their stated boundaries, as is the Isis Florence-2 request-local non-authoritative location-token codec and the Isis AnimateDiff Lightning request-local non-authoritative timestep-arithmetic preview, plus the Isis Live Preview request-local non-authoritative preview-step arithmetic, plus the Isis Topology Verification request-local non-authoritative summary-rule evaluation, plus the Isis Multi-Provider 3D Generation request-local non-authoritative defect-flag action preview, plus the Isis ComfyUI Integration request-local non-authoritative SDK validation preview, plus the Isis Provider Integration request-local non-authoritative provider-weight routing preview, plus the Isis Unified Provider Architecture request-local non-authoritative capability- scoring preview, plus the Isis AI Providers UI-only bounded composition shell. The Isis Conversational AI request-local non-authoritative reasoning-step graph structure preview is also closed within its exact boundary, as is the Isis 3D Prompt Engineering request-lifetime recomputable prompt-preview seam, plus the Isis Gaussian Splatting browser-local review workspace and request-local non-authoritative storage-footprint preview, plus the Isis Texture Enhancement durable internal fixture-score diagnostic history, plus the Isis Topaz Enhancement request-local non-authoritative film-aware plan preview, plus the Isis RIFE Interpolation request-local non-authoritative synthetic-frame plan preview, plus the Isis NVIDIA RTX Processing request-local non-authoritative 4× super-resolution plan preview, plus the Isis Reference Image Processing request-local non-authoritative aspect-ratio bucket preview, plus the Isis Video Production Workflows request-local non-authoritative schema-validation preview. The Isis Anomaly Detection UI-only non-authoritative embedded-lane composition is also closed within its exact boundary, as is the Isis Automated Quality Scoring deterministic fixture-regression read model, plus the Isis Generation API UI-only non-authoritative embedded-lane composition and the Isis GPU Worker UI-only non-authoritative embedded-lane composition, plus the Isis CLI UI-only non-authoritative embedded-lane composition, plus the Isis CLI Surface request- local non-authoritative CLI token-resolution preview, plus the Isis Output Registry UI-only non-authoritative embedded-lane composition, plus the Isis Web UI-only non-authoritative embedded-lane composition, plus the Isis Database UI-only non-authoritative embedded-lane composition, plus the Isis Workflow Registry UI-only non-authoritative embedded-lane composition, plus the Isis Event Publisher UI-only non-authoritative embedded-lane composition. V1-002's Isis Outputs UI-only non-authoritative embedded-lane composition is also closed within its exact boundary, as is the Isis Client UI-only non-authoritative embedded-lane composition, plus the Isis Job Envelope request-local non-authoritative normalization preview and embedded-lane composition, plus the Isis AI Video UI-only non-authoritative embedded-lane composition, plus the Isis ComfyUI SDK request-local non-authoritative semver evaluator and embedded-lane composition.

The pre-authenticated OIDC pending-login authority is now closed through 99344f4c9be62aaeb8810c77874984325bc86d8d. Production no longer stores state, nonce, connection, redirect, and expiry in a process-local map. The deployable binds the route to a shared Redis authority with digest-only keys, an exact ten-minute TTL, collision-safe NX issuance, and one atomic read-and-delete consumption script; missing Redis fails both issuance and callback closed with an explicit 503. State survives issuer-process replacement, two replicas cannot consume it twice, expired/future/malformed records cannot mint a session, and raw browser state is absent from Redis keys. The in-memory implementation remains injectable only for direct development and tests, with expiry pruning and the same single-use semantics. This state is pre-account, short-lived, and contains no mapped user identity, so it has no user DSAR or long-term retention partition. Focused route/store coverage passed 9/9, a real two-client Redis test passed 1/1, the targeted production build and fresh-32-migration artifact smoke passed, and the disposable database was removed. Evidence: evidence/v1-v9/V1-002.md.

The tenant LMS pending-login and managed-connector control-plane partitions are now closed through fab8ac2b074e9aa31d003e3c9ad2757b0d3b7f34. LTI 1.3 state and nonce no longer depend on the issuing process: production uses digest-only Redis keys, fixed ten-minute PX expiry, collision-safe NX issuance, and one atomic read/delete consumption script, while missing or malformed Redis authority returns explicit 503 instead of falling back to memory. Managed LMS connector overrides and environment-connector tombstones now hydrate and write through the PostgreSQL snapshot authority; admin mutation responses publish only after an ordered durable write, failed writes remain invisible, malformed snapshots fail startup hydration, and production without PostgreSQL refuses the mutation with 503. Focused route/store coverage passed 34/34; real two-client Redis and PostgreSQL process-replacement tests passed 1/1 each; the exact production artifact built and passed startup against Redis plus a disposable database with all 32 migrations before that database was removed. The claim is bounded: the framework's separate LTI launch-session and SCORM runtime/learning session maps remain to be dispositioned, and cross-replica connector mutation serialization/read refresh is not claimed by this snapshot-backed single-writer path. Evidence: evidence/v1-v9/V1-002.md.

The two remaining LMS framework session maps are now honestly dispositioned through cf6a834163bb785f5df6f8b01629a5b6888aa74d. Repository-wide tracing found no production Deep Linking, NRPS, or AGS consumer for the normalized LTI launch session id; production therefore returns the verified request projection but retains neither the full claim set nor a process-local session after the response. The shared framework retains sessions only when an explicit consumer opts in. The SCORM HTTP seam did claim mutable learner progress, but no production durable runtime, authenticated learner-owner mapping, DSAR identity, or retention authority exists. Production now disables the framework's SCORM map and returns 503 tenant_scorm_runtime_store_unavailable before returning a launch token or acknowledging an update; direct framework tests retain an explicit in-memory adapter. Focused shared/BFF coverage passed 45/45, the shared library direct typecheck passed, the exact production artifact built and passed startup with Redis plus a fresh 32-migration database, and that database was removed. This closes the volatile-map classification, not a live SCORM product flow: implementing SCORM requires one owner-mapped durable authority with privacy, erasure, retention, concurrency, and recovery proof. Evidence: evidence/v1-v9/V1-002.md.

The Living Scenes render-budget map is now closed through 4e0876107bf0421f7ee20995e6636beb3594e601. The prior raw-user-keyed, process-local one-in-flight lease and six-per-minute counter reset on restart and could be bypassed across replicas. Production now injects one atomic Redis authority with digest-only, Redis-Cluster-co-slotted keys, ownership-checked release, a 60-second counter expiry, and a bounded 30-minute crash-recovery lease. Missing or malformed Redis fails the route closed with 503 render_budget_unavailable; process memory remains only an explicit direct- test adapter. This is short-lived pseudonymous operational protection, not artifact, approval, learning, or publication authority and not a durable DSAR partition. Focused route/gate coverage passed 10/10, a live two-client Redis race/restart test passed 1/1, the exact production artifact built and passed startup with Redis plus a fresh 32-migration database, and that database was removed. Evidence: evidence/v1-v9/V1-002.md.

The three orphan notification preference/digest maps are now closed through 6f696261d7cb2abeaa96004572c0773f5c25bf41. Repository-wide tracing found no runtime caller for the raw-user-keyed quiet-hours config, digest config, or last-delivery maps, and deleting both legacy modules left the same-tree production bundle byte-identical. The live authenticated preference routes already own quiet-hours and digest choices through the canonical PostgreSQL snapshot authority: writes publish only after ordered persistence, hydrate across process replacement, fail closed with 503, and compose into exact- subject erasure with a durable resurrection fence. Focused route/store coverage passed 25/25, real PostgreSQL restart/erasure proof passed 1/1, and the final integrated production artifact built and passed startup with Redis plus a fresh 32-migration database. No orphan last-delivery value was migrated because it had no producer or consumer; a future digest dispatcher still requires its own durable idempotent delivery/outbox authority. Evidence: evidence/v1-v9/V1-002.md.

The achievement read-refresh debounce map is now explicitly bounded and recomputable through fb671e9fd4910c3a66bb353433817ecd8b00c800. It no longer retains raw user ids or grows for the process lifetime: keys are SHA-256 subject digests, claims expire after 30 seconds, expired claims are pruned, and oldest claims are evicted at a hard 10,000-entry ceiling. The cache owns no progress or unlock record; clearing it can only repeat one live cross-domain refresh and idempotent achievement evaluation. Focused cache/route coverage passed 9/9, including total-cache-loss proof with an unchanged durable unlock count, and the final integrated production artifact passed build plus startup against Redis and a fresh 32-migration database. Evidence: evidence/v1-v9/V1-002.md.

The two maps inside createAgenticStudioGovernanceGate are now classified as constructor-local run evaluation state, not a production store. Repository-wide tracing found no BFF composition root, route, or worker caller; only the adapter's own test constructs it. Each call builds envelopes/meters from the supplied starting budget list and retains them only through that returned gate. Focused coverage passed 5/5 for kill-switch, meter exhaustion, terminal-run, and real-router behavior. This closes the maps in the inventory without claiming a live or restart-safe governance path: any future production caller must add durable run checkpoints and atomic shared budget authority (or prove a strictly non-resumable request boundary). Evidence: evidence/v1-v9/V1-002.md.

The module-global Veritas personal-retraction acknowledgement map is now closed through 44e733459bd13b60dbffea5b5c64dedc2430220c. The authenticated route now acknowledges only after an ordered schema-v1 PostgreSQL snapshot commit, recovers the suppression decision across process replacement, and fails reads/writes closed with a controlled 503 when production persistence is unavailable. The account-scoped profile partition is hard-bounded, contributes an exact-subject minimized DSAR record, and participates in signed profile erasure with a domain-separated pseudonymous resurrection fence. Focused store/route/export coverage passed 64/64, real PostgreSQL restart/export/ erasure/restart proof passed 1/1, and the final rebased artifact passed direct build, syntax validation, and startup against Redis plus a fresh 32-migration database. This closes the acknowledgement map, not the static fixture notice or the broader all-store inventory. Evidence: evidence/v1-v9/V1-002.md.

The Tara procedural ambient-WAV map is now explicitly bounded and recomputable through 6d185b90dc7d93ddf02af9e1e93e507067f15675. The cache no longer retains raw session ids, uses domain-separated SHA-256 session/duration keys, and enforces both a 32-entry LRU ceiling and a 64 MiB aggregate byte ceiling instead of permitting roughly 254 MiB through the former entry-only bound. Oversized buffers are served without retention. Focused cache and real-route coverage passed 11/11, including total singleton-cache loss followed by byte-identical real WAV regeneration. The final rebased artifact passed direct build, syntax validation, and startup against Redis plus a fresh 32-migration database. This cache owns no listening, completion, artifact, approval, publication, or learning record and is not a DSAR authority. Evidence: evidence/v1-v9/V1-002.md.

The V3 regional-waitlist module-global map and false browser confirmation are now closed through 29228d13b6. The public route no longer retains anonymous email addresses behind predictable record ids: valid requests return a no-store 503 regional_waitlist_not_configured disposition with emailStored: false, malformed requests do not echo input, and guessed legacy record ids always return 404. Launch decisions disclose that signups are not accepted. The image-led Lilith page now calls the real route and reports unavailable, network-failure, and unverifiable-success states without claiming an email or invite was saved. Focused BFF/web coverage passed 8/8; all three targeted Playwright flows and scoped axe passed, with one cold-start navigation passing on the configured retry; the V3 region verifier, targeted lint, clean web type ratchet, BFF build, and fresh 32-migration artifact smoke passed. The V3 product checkbox is honestly reopened: a real signup remains open until it has consent, delivery, durable owner mapping, privacy, retention, and recovery proof. Evidence: evidence/v1-v9/V1-002.md.

The authenticated V6 Steward continuity map is now closed through 8708bc2c3e4b3d5280d09eeaf3f64a70dd7cfc27. Production requires and hydrates a schema-validated PostgreSQL snapshot before traffic; objectives, squad comms, Crossroads counsel, homestead settings, visit invites, account preferences, and tier decisions use one ordered candidate/save/publish stream. Failed or missing persistence returns a controlled 503 without publishing the candidate. Subject keys and persistent erasure fences are domain-separated SHA-256 digests; exact-subject DSAR export and signed continuity erasure preserve adjacent users and reject post-restart resurrection with 409. Reads of static fixtures do not silently retain a user record, and every retained history plus the subject population and record bytes is hard-bounded. Focused store/export coverage passed 49/49, authenticated route coverage passed 6/6, real PostgreSQL restart/export/erasure/restart proof passed 1/1, and the final artifact passed direct build, syntax validation, and startup against Redis plus a fresh 32-migration database. The snapshot path deliberately claims ordered single-writer mutation, not cross-process mutation serialization; that replica-coordination seam and the broader all-store inventory remain open. Evidence: evidence/v1-v9/V1-002.md.

The admin reminder-producer module-global map and its false durable-batch contract are now closed through 3cba2649419a40e3c749692551651e6e2e3b5d6e. Repository-wide tracing found no live domain reader, scheduler, outbox handoff, delivery effect, or runtime consumer for the retained batches. The route is therefore explicitly a request-local, non-authoritative preview: it runs the real pure reminder producers over operator-submitted slices, returns 200 with retained: false and deliveryScheduled: false, sets cache-control: no-store, and discards the contact-bearing projection. It no longer returns 201, invents restart-safe batch idempotency with 409, or exposes process-local history. Strict allowlists, canonical timestamps, HTTPS-only links, channel/contact checks, field and recipient-byte bounds, and a 1,000-item combined ceiling constrain the preview. Focused route coverage passed 7/7; the BFF build, syntax check, and fresh 32-migration production-artifact smoke passed. This closes the volatile map classification only. G-MSG-1 remains open until a live producer reads authoritative domain state and durably schedules delivery with appropriate idempotency and recovery. Evidence: evidence/v1-v9/V1-002.md.

The shared authenticated route-response cache is now explicitly bounded and recomputable through d77d5b60418cd32a07d979779edc2242710703f6. Its 25 cache instances across 17 route modules previously removed an expired entry only when that exact key was read and had no entry ceiling, so varied tenant-, residency-, scope-, user-, route-, query-, and pagination-keyed projections could accumulate for the process lifetime. Every instance now prunes all expired entries on writes and size inspection, uses access-order LRU eviction at a default hard 1,000-entry ceiling, and clamps retention to between one second and an absolute five-minute maximum. Keys remain SHA-256-derived partitions. The cache owns no domain record or idempotency decision; clearing it yields a miss and recompute. Direct capacity, expiry, unsafe-configuration, and total-loss proof passed 7/7, and representative domain, notification, search-suggestion, Nisaba, and admin route suites passed 84/84 while preserving hit and partition behavior. The BFF build, syntax check, and fresh 32-migration artifact smoke passed. Evidence: evidence/v1-v9/V1-002.md.

The consent-gated Mentor Presence aggregate analytics map is now closed through a7e03be1803a7648b0e864e1dc4ae20118c6e12e. Production requires and hydrates a strict schema-v1 PostgreSQL snapshot before traffic; event mutations use one ordered candidate/save/publish stream, and failed or missing persistence returns a controlled 503 without incrementing the visible metric. Snapshot decoding rejects unknown fields, versions, buckets, duplicate categories, unsafe counts, oversized documents, revision/count mismatches, and fallback-rung/reason total mismatches. The store retains global counters only: no subject ids or raw event history. Unrecognized client fallback text is irreversibly folded into other instead of being retained or exposed. Intake is authenticated, consent-authorized, strictly shaped, length-bounded, and abuse-protected; summary reads are operator-scoped and no-store. Focused store and full route coverage passed 25/25, real PostgreSQL replacement recovery passed 1/1, and the final BFF artifact passed direct build, syntax validation, and startup with Redis plus a fresh 32-migration database. The counts remain client-reported and the whole-snapshot queue is process-local, so this does not claim distributed exactly-once measurement, cross-replica serialization, or production outcome validity. Evidence: evidence/v1-v9/V1-002.md.

The CTIA SMS suppression process map and previously orphaned outbound check are now closed within the local BFF boundary through 3be16ee88908562540297214e5411ba5380cff2d. Production requires and hydrates a strict schema-v1 PostgreSQL snapshot before traffic. STOP and START mutations use one ordered candidate/save/publish stream; the Twilio webhook does not emit confirmation TwiML until the complete decision commits, and returns a retryable 503 without publishing volatile state when persistence is missing or fails. The bounded snapshot retains a domain-separated recipient digest, canonical reason, provider receipt id, scope, and effective time, but no raw phone number. Authenticated START removes exactly that recipient while preserving adjacent suppressions across restart. The live reminder path now consults the hydrated authority before its Twilio seam, so a suppressed external SMS creates no provider request while the independently governed in-app reminder remains available. Focused store, webhook, reminder-route, scheduler, and delivery coverage passed 63/63; real PostgreSQL STOP/restart/START/restart recovery passed 1/1; the messaging library typecheck passed; and the exact final BFF artifact passed direct build, syntax validation, and startup with Redis plus a fresh 32-migration database. The whole-snapshot queue is process-local, historical provider deliveries cannot be recalled, and provider-native/carrier reconciliation plus an owned retention process remain open; no cross-replica serialization or universal messaging compliance claim is made. Evidence: evidence/v1-v9/V1-002.md.

The shared customer Trust & Safety report queue is now closed within the local BFF boundary through 934d697bc9d663a91648d2e3a0ac21baf890f492. The former module-global 200-row array acknowledged submissions and terminal moderation decisions before any durable boundary, disappeared on restart, and silently evicted older reports. Production now requires and hydrates a strict schema-v1 PostgreSQL snapshot before traffic. Intake, decision, subject redaction, and retention purge use one ordered candidate/save/publish stream; missing, failed, or capacity-exhausted persistence returns a controlled 503 without publishing the candidate, notifying the reporter, or recording the decision audit side effect. Conflicting terminal re-decisions return 409 instead of rewriting review history, and all report-bearing responses are no-store. The hard-bounded authority retains the full terminal reason, contributes reporter-owned records to the real support_case export, excludes confidential reports from target-side export, and participates in signed account deletion. Reporter, exact user-target, reviewer, and embedded exact-id text are minimized behind domain-separated persistent fences; the established legal-hold/retention-exception executor preserves redacted cases for five years and then removes reporter cases plus residual target/reviewer pseudonyms. Focused authority, route, and full export-bundle coverage passed 64/64; real PostgreSQL intake/decision/restart/redaction/restart recovery passed 1/1; and the final BFF artifact passed direct build, syntax validation, and startup with Redis plus a fresh 32-migration database. The snapshot queue is process-local, and notification plus the existing audit adapter remain post-primary-commit side effects rather than a distributed transaction; cross-replica serialization and the parent all-store inventory remain open. Evidence: evidence/v1-v9/V1-002.md.

The orphan Isis seed-data notice set is retired through 42b225eb86ca3a872a0f094155b4fa4cb4176de0. Repository-wide symbol and import tracing proved that its module-global “real record” marker and notice builder were imported only by their colocated unit test: no production route, store, server registration, or response schema could read or mutate them. Keeping the set therefore represented a stateful auto-drop contract that production could never execute. The unreachable helper and its self-referential test are deleted; the real admin integrations registry's independent durable seed-versus-real behavior remains covered 19/19. A fresh BFF build produced the exact same previously smoked artifact hash, proving the removed module was outside the deployable graph. The BFF ratchet reported only the same five unrelated current-main errors. This is a negative-reachability disposition, not a claim that every seeded operator console is connected to live intake. Evidence: evidence/v1-v9/V1-002.md.

The production-reachable mock-domain mutable map and fabricated-domain switch are closed through 1625bc1d8387edfce49d447d4f5be0704cd3b227 and 79aa5ce43e31bc3386466fa514b444c6f26e6662. The deploy-bound default adapter resolver previously honored USE_MOCK_ADAPTERS=true and OSHUN_VERITAS_ADAPTER=mock|outage even in production, allowing fixture journalism/domain data and process-local follow-topic state to appear real. Default app construction now rejects every such combination when either NODE_ENV or OSHUN_ENV is production. Dev/test follow state is factory-local, limited to 1,000 subjects and 100 topics per subject, and cannot bleed across app instances. Focused environment, actual-app-construction, isolation, bound, and existing Veritas route coverage passed 34/34; scoped lint passed under the rebased configuration; and the exact BFF artifact passed syntax validation and startup with Redis plus a fresh 32-migration database. Explicit adapter injection remains an intentional programmatic test seam, and mock state remains lossy by design outside production; no live Veritas provider or content-validity claim is made. Evidence: evidence/v1-v9/V1-002.md.

The orphan analytics warehouse event buffer is retired through de34f0c5d26a854e9687a2462604b75e3df65089. Production tracing found no mutation caller for its module-global array and no composition-root warehouse sink; the admin export route could only read an empty process singleton before failing on the absent sink. The real durable analytics intake cannot be losslessly substituted because warehouse rows require emit-time consent, pre-hashed member identity, tenant, and taxonomy surface that the privacy-owned intake intentionally does not retain in that form. The 47-line false authority is deleted. The endpoint remains as an authenticated, abuse-protected composition seam, but now requires an explicit authoritative event source and returns 503 warehouse_event_source_not_configured before touching a sink when none is wired. Injected source/sink taxonomy, consent-gate, digest, receipt, and error behavior remains covered. Focused route, boot, durable-intake regression, and warehouse-library coverage passed 34/34; the exact rebased BFF artifact passed syntax validation and startup with Redis plus a fresh 32-migration database. No warehouse source, sink, or live delivery is claimed; those remain open integration work. Evidence: evidence/v1-v9/V1-002.md.

The false admin residency-audit singleton authority is closed through e2d1cb280718b475ed05d3763ed496fa66eb91e2 and the async durable-source follow-up e9b47d94f5a6cf6115bb4df1d9d56f9864e140ac. Production tracing proved that the bounded canonical-event ring was imported only by its admin read route and explicit tests: no production residency guard, enforcer, publisher, worker, or composition root ever wrote it. The deployed endpoint therefore reported an authoritative-looking empty residency history after every start. The global instance is removed; the remaining bounded adapter is explicitly process-local and injection-only. App composition now accepts a synchronous or awaited canonical AdminResidencyAuditSource; after authentication, scope, and filter validation, an absent source returns 503 residency_audit_source_not_configured instead of an empty success. Focused route and region-aware enforcement coverage passed 46/46, including async-source reads and fail-loud ordering; scoped lint passed; and the exact artifact passed syntax validation and startup with Redis plus a fresh 32-migration database. No durable canonical audit repository or production residency-enforcement binding is claimed; both remain open integration work. Evidence: evidence/v1-v9/V1-002.md.

The process-local admin collaboration authorities are closed through 9d7754ca7996834161959bcfd49abcd54c1f931f. Explicit operator cross-links and shared entity-comment threads previously lived only in module-global maps even though their routes acknowledged mutation and investigation bundles described them as independently persisted. Both stores now use validated, bounded, versioned snapshots and ordered candidate/save/publish adapters over the existing PostgreSQL admin_store_snapshot authority. Startup hydrates both before traffic; routes and investigation-bundle assembly receive the same durable instances; and deployable reads or mutations fail with a controlled 503 when that authority is absent. Exact-subject export includes authored and referencing collaboration records. Signed deletion removes owned cross-links, redacts authored comment text and actor references, and persists digest-only resurrection fences. Focused store, route, investigation, export, deletion, and restart coverage passed 84/84; the real fresh-PostgreSQL durability suite passed 2/2 after all 32 migrations; scoped lint and formatting passed; and the exact rebased artifact passed syntax validation and production startup with fresh PostgreSQL plus Redis. This closes restart loss and false acknowledgement for the two connected stores, not multi-writer serializability for whole-snapshot documents or V1-002's parent all-store inventory. Evidence: evidence/v1-v9/V1-002.md.

The false Aja consent, content-security, and data-retention governance authorities are retired through e90c1e371b53782196ab648b725e5fb3a4f792d9. Repository-wide producer and consumer tracing found that the three module-global manager instances were reachable only from their own admin routes and tests: no upload/processing consumer, content delivery/decrypt/share enforcer, storage lifecycle worker, or deletion executor consumed their decisions. Production app composition no longer creates those volatile authorities. Authenticated catalogs explicitly report authority.configured: false; valid decisions return controlled, surface-specific 503 responses, while malformed or unauthorized requests still fail before authority disclosure. The underlying managers remain only as factory-local, explicitly injected test/development adapters. The three Studio pages render a shared “Authority not connected” boundary, name its exact reason, disable every field and mutation, and issue no optimistic request. Focused BFF coverage passed 22/22, focused web coverage passed 15/15, and authenticated Playwright coverage passed 3/3 with automatic WCAG 2.1 AA scans plus desktop and narrow-viewport overflow checks. Scoped lint, formatting, and selected walkthrough rendering passed; web typecheck is clean and the BFF ratchet contains only the same five unrelated current-main errors. The exact rebased artifact passed syntax validation and production startup with Redis plus a fresh 32-migration PostgreSQL database. This closes false acknowledgement and restart-ambiguous governance, not the future work to bind durable owning services with enforcement, subject export/deletion, legal holds, retention, and recovery. Evidence: evidence/v1-v9/V1-002.md.

The false admin restore-review authority is retired through bccaaba0bf8e9cf91a4a75560b3550e3ea466eb8. Repository-wide producer, consumer, import, and composition tracing found that the module-global AdminRestoreReviewStore was populated only by tests: the claimed background ingestion does not exist, no restore executor consumes its approvals, and the deployed instance had no audit publisher. Production therefore advertised an empty restore queue after every start and could never receive a real request; if populated indirectly, its approvals, evidence, and notes would disappear on restart without restoring an artifact. The singleton is removed. AdminRestoreReviewStore remains only as an explicitly injected deterministic test/development state-machine adapter. App composition now accepts a synchronous or awaited canonical AdminRestoreReviewAuthority; authenticated, scope-valid, payload-valid reads and decisions return 503 restore_review_authority_not_configured when it is absent. Authentication, scope, query, path, and action-schema failures retain their existing 401, 403, and 400 ordering. Explicit injection preserves the full transition and audit regression harness. Focused route, authorization, lifecycle, and audit coverage passed 41/41; scoped lint and formatting passed; and the BFF ratchet reported only the same five unrelated current-main errors. After integrating the moved remote tip, the exact artifact was rebuilt and passed syntax validation and production startup with Redis plus a fresh 32-migration PostgreSQL database. This closes a false process-local control plane, not the still-open work to connect a durable owning restore service, canonical intake, audited decisions, and an idempotent artifact-restoration executor. Evidence: evidence/v1-v9/V1-002.md.

The false admin moderation queue, appeal, and ban authority is retired through 9a37575bcb7a4fa2775579312e4d7655b3a1edf5. Repository-wide tracing proved that no production scanner ingested results into the module-global ModerationStore, and no authentication, content-delivery, account, or publication enforcer consumed its decisions or bans. The admin route itself could fabricate scan records, approve or reject them, adjudicate appeals, and issue bans that affected only one BFF process; the singleton disappeared on restart and separate replicas could disagree. The singleton is removed, and the real Aja moderation state machine remains only as an explicitly injected test/development adapter. App composition now accepts a synchronous or awaited AdminModerationAuthority. All eight valid queue, decision, appeal, ban, and status operations return 503 moderation_authority_not_configured when the owning service is absent, after authentication, scope, path, and payload validation. The affected Aja project was already an Nx library but lacked its required type:lib metadata; that tag is corrected without weakening the module-boundary rule. Focused route and library coverage passed 52/52, scoped lint/formatting and the library lint target passed, and the BFF ratchet retained only the same five unrelated current-main errors. After integrating the moved remote tip, the exact artifact was rebuilt and passed syntax validation and production startup with Redis plus a fresh 32-migration PostgreSQL database. This closes fabricated process-local moderation and enforcement claims, not the future work to connect canonical scanner intake, durable case/appeal authority, authenticated actor attribution, audited decisions, and real enforcement. Evidence: evidence/v1-v9/V1-002.md.

The repository-wide BFF operational-state inventory is now machine-enforced through 41d31b78c1ab181def5f2ff6debd72c5e3c455ff. An AST detector and checked-in registry disposition all 238 non-test module-scope constructed owners, operational factories, mutable collections, route caches, and mutable bindings; the drift test rejects additions, removals, moved lines, or missing rationale/evidence. The first review classified 110 durable authorities, 60 durable coupled projections, 43 loss-safe ephemeral bindings, 10 stateless catalogs, nine unbound projections, and six test/development adapters. It also removed implicit production Isis-generation and Psyche-embodiment stores and the false message-delivery transparency map: valid unbound production requests now return explicit 503s. Focused coverage passed 36/36, scoped static checks passed, the BFF ratchet retained only the same five unrelated current-main diagnostics, and the exact artifact passed all 32 migrations plus PostgreSQL/Redis production startup smoke. Evidence: evidence/v1-v9/V1-002.md.

Manual review then proved four of the original ephemeral dispositions wrong. Through ea0bd4e6e18f226641191b684eb0305089517ae5, achievements, routines, wearable, and desktop shell engines are no longer constructed as production authorities: their retained unlock/social/challenge, routine/execution/effect, device/preference/reminder/summary, and desktop-shell state now requires an explicit owning injection or fails closed with a domain-specific 503. Routine and wearable authentication remains ordered before availability; pure source-controlled catalogs remain readable. The corrected counts are 110 durable authorities, 60 coupled projections, 39 loss-safe ephemeral bindings, 10 stateless catalogs, 13 unbound projections, and six test/development adapters. Focused and contract coverage passed 18/18, post-integration inventory/boundary checks remained green, scoped static checks passed, the BFF ratchet returned to the same five unrelated diagnostics, and the rebuilt artifact passed a fresh 32-migration PostgreSQL/Redis startup smoke. Evidence: evidence/v1-v9/V1-002.md.

The legacy Nyx aggregate event-action map is corrected through 286f70747dfdd1a36d518f2ab308937ff806ee46. Registry review found that its saved/followed/reminder projection had no persistence or hydration seam and duplicated information already governed by the durable NyxMemberStateStore. Production no longer binds the compatibility singleton; authenticated and scope-valid aggregate routes return 503 nyx_event_action_authority_not_configured until a canonical projector is injected. The existing durable Nyx owner, export, signed erasure, fence, and recovery behavior remain unchanged. The registry now reports 109 durable authorities, 60 coupled projections, 39 loss-safe ephemeral bindings, 10 stateless catalogs, 13 unbound projections, and seven test/development adapters. Focused Nyx and inventory coverage passed 21/21, static and type ratchets stayed within the known baseline, and the exact artifact passed a fresh 32-migration PostgreSQL/Redis startup smoke. Evidence: evidence/v1-v9/V1-002.md.

V1-002's all-store box remains open until every remaining operational store has a durable or recomputable disposition, every required ephemeral-loss proof exists, and every newly durable user-scoped or provider-native partition has complete deletion, privacy, retention, and recovery evidence.

The detector itself is hardened through 80470f8f1ff49ee4738fdd9a389b7cb21c9ad000: mutation-method analysis now discovers uppercase and nonstandard module-scope collections instead of relying only on lower-camel/cache naming. That pass added the admin mutation checkpoint WeakMap, reviewed it as request-lifetime rollback state whose loss cannot acknowledge or erase an admin mutation, and raised the reviewed registry to 239 entries.

Startup-composition coverage is added through 1ff95429d25a551ce48ebbd703e56c7f63a7567b. The AST gate now also inventories operational constructors and factories in the deployable BFF's main composition, closing the blind spot around Redis idempotency, SSO/LTI login state, Telegram replay/rate state, abuse and generation gates, the durable admin snapshot/audit composition, Prisma rail/Ori state, and the Tara workbench. The 263 reviewed entries now comprise 123 durable authorities, 62 durable coupled projections, 44 loss-safe ephemeral bindings, 14 stateless catalogs/policies, 13 deliberately unbound production projections, and seven test/development adapters.

Cross-store verification passed 67/68 loss-safety tests (one explicitly skipped provider integration), 17/17 deletion-worker/eraser tests, 59/59 live-Redis privacy/export/deletion-fanout integrations, and 17/17 release/OpenAPI coverage-and-runtime-drift tests. The focused post-integration inventory suite passed 2/2, and the BFF ratchet reported only the same five unrelated current-main diagnostics. This is stronger closure evidence, but the checkbox remains open while factory-local and route-composition ownership roots outside main receive the same machine-enforced disposition review. Evidence: evidence/v1-v9/V1-002.md.

That deeper scan immediately found and retired a production calendar shadow through b00b985cc9621a4d3512054e5e411394b82cff4d. Configured Google calendar imports previously acknowledged imported events, reconciliation watermarks, and run evidence into MemoryCalendarSyncStore; a restart silently discarded all three. Production import/export now authenticates and checks admin scope first, then returns 503 calendar_sync_authority_not_configured unless deployable composition injects an owning CalendarSyncStore. Connector visibility remains available, and the memory adapter remains test/development-only. Focused calendar reconciliation and route coverage passed 15/15 before and after two remote integrations; inventory, lint, formatting, and bundle syntax passed; the BFF ratchet retained only the five known diagnostics; and the exact artifact passed all 32 migrations plus PostgreSQL/Redis production startup smoke. Evidence: evidence/v1-v9/V1-002.md.

Nested store/factory coverage is completed through c5277d7c0b47424270e70497659e3fabaca4e992. The detector now scans store-, repository-, registry-, queue-, gate-, limiter-, authority-, and state-shaped constructions inside factories, class fallbacks, hydration paths, and route composition while avoiding duplicate module/startup roots. All 282 additional live candidates were reviewed. The 545-entry registry now contains 146 durable authorities, 151 coupled durable projections, 88 loss-safe temporary/reconstructable values, 122 stateless catalog/adapter constructions, 25 test/development adapters, and 13 deliberately unbound production projections. Its post-integration drift/regression suite passed 3/3.

That pass also retired the process-local Aja ban console through 07762a54c5440a3fa69055754f6279a107b4c4ab. Production previously created a fresh BanManager per BFF app and acknowledged ban issue/check/revoke state that vanished on restart. The source-controlled ban-type/category catalog remains readable, but valid state operations now authenticate and validate first, then return 503 aja_content_moderation_authority_not_configured unless an owning authority is injected. The full Aja and generic moderation regression set passed 21/21, the post-integration Aja gate passed 8/8, static and type ratchets stayed within the known baseline, and the exact artifact passed a fresh 32-migration PostgreSQL/Redis production smoke. The all-store box remains open for the next owner-name expansion beyond store-shaped constructions. Evidence: evidence/v1-v9/V1-002.md.

The owner-name expansion found and retired the Rail TV watch-party shadow authority through 136c89c155a352b3065c21896cb70a997f7e9cba. Production previously constructed one process-local TvWatchPartyCoordinator, so session creation, phone pairing, snapshots, playback commands, and live-chat membership could be acknowledged by one replica and disappear or disagree after restart. All state-bearing paths still authenticate and validate their request boundary, then return 503 tv_watch_party_authority_not_configured unless deployable composition injects an owning coordinator. Direct tests inject the coordinator and retain the complete create/pair/poll/command/chat behavior. Focused coverage passed 4/4 before and after remote integration; scoped formatting, lint, diff, and inventory checks passed; the BFF ratchet contained only the same five unrelated current-main diagnostics; and the integrated production artifact passed syntax validation plus Redis startup against a fresh PostgreSQL database with all 32 migrations before that database was removed. This retires the false authority; it does not claim that a shared watch-party owner has been implemented. Evidence: evidence/v1-v9/V1-002.md.

Manager-, engine-, coordinator-, tracker-, service-, session-, framework-, governor-, learner-, sampler-, runtime-, and plane-shaped construction coverage is machine-enforced through ce41cd5b03170adabb973f81f77921d98d049684. All 88 newly discovered candidates were reviewed, including autonomy learning and accounting views coupled to the encrypted durable plane, Ori's file-durable agent runtime, LMS connector projections coupled to their snapshot owner, request-lifetime provider/model sessions, deterministic simulation engines, fail-closed shell adapters, and explicit development-only Aja/Psyche/moderation managers. The resulting 633-entry registry contains 148 durable authorities, 166 durable-coupled projections, 125 loss-safe ephemeral bindings, 138 stateless engines/catalogs, 39 test/development adapters, and 17 deliberately unbound production projections, with zero undispositioned entries.

The post-integration inventory gate passed 3/3. Representative recovery, fail-closed, and loss-boundary suites passed 86/86 across autonomy, account protection, academic integrity, group sessions, LMS, Ori, health probes, real-time music, Psyche, and Aja. Scoped formatting, lint, syntax, drift, and diff checks passed; the BFF ratchet reproduced only the five known unrelated diagnostics. The exact integrated artifact passed a fresh 32-migration PostgreSQL/Redis production smoke. This closes the named owner-family blind spot but not the parent box: other construction vocabularies and anonymous/local mutable collection roots still require the same machine-enforced review, and the existing user/provider deletion, privacy, retention, recovery, and ephemeral-loss conditions remain mandatory. Evidence: evidence/v1-v9/V1-002.md.

Persistence/control construction coverage is extended through e1cef3526a13e585726c092ddc547b3fef44160b. The AST gate now inventories sink-, breaker-, controller-, emitter-, collector-, accountant-, director-, broker-, solver-, evaluator-, blackboard-, aggregator-, planner-, bridge-, processor-, analyzer-, narrator-, reader-, resolver-, source-, and policy-shaped constructors/factories. All 161 additional candidates were reviewed. The 101 snapshot sinks and the autonomy accounting/director, academic bridge/emitter, and fiat settlement processor/emitter are explicitly coupled to their durable owners. Circuit breakers, metrics/probe collectors, capability brokers, external sources/resolvers, and narrators are loss-safe request/rebind state; deterministic policies, simulations, planners, and bridges are stateless; the one Aja manager policy construction remains test/development-only.

The resulting 794-entry registry contains 148 durable authorities, 275 durable-coupled projections, 154 loss-safe ephemeral bindings, 160 stateless engines/catalogs, 40 test/development adapters, and 17 deliberately unbound production projections, with no undispositioned candidate. The inventory gate passed 3/3, the shared snapshot primitive passed 10/10, and representative BFF sink/broker/emitter/processor/probe/recovery suites passed 98/98. Scoped static checks passed, and the BFF ratchet contained only the same five unrelated current-main findings. The pass changes no production runtime; the parent box remains open for client/provider/adapter and remaining local collection escape families plus the standing loss, privacy, deletion, retention, and recovery conditions. Evidence: evidence/v1-v9/V1-002.md.

The false recommendation-feedback acknowledgement boundary is closed in 9f7877877b8194db3d15a50d111eec9717c9be5f. Search and cross-domain telemetry no longer default to module-global console-only analytics clients; server events enter the bounded, owner-attributed durable analytics authority atomically. Recommendation feedback now returns success only after that authority commits, and returns a persistence-specific 503 without claiming a record when the production sink is unavailable. The focused analytics, recommendation, and search suite passed 49/49, the post-integration regression suite passed 12/12, static and inventory checks passed, and the BFF ratchet contained only the same five unrelated findings. The integrated production artifact passed all 32 migrations and an isolated PostgreSQL/Redis smoke. Evidence: evidence/v1-v9/V1-002.md.

Provider/client construction coverage is extended in 60f35d06ee65700c57bdb2606d3c6e9a1ca6ee50. The AST gate now includes provider-, adapter-, client-, signer-, executor-, classifier-, verifier-, watermarker-, and bus-shaped factories and constructors. All 139 newly exposed candidates were reviewed. Seven persistent/replaying Redis buses are durable authorities; the non-persistent deletion request/reply bus is loss-safe because signed attestations are its audit authority. External clients, providers, signers, classifiers, watermarkers, and executors are reconstructable bindings; deterministic adapters are stateless; three Psyche compatibility adapters remain test/development-only; and the analytics snapshot sink remains coupled to its durable owner.

The resulting 932-entry registry contains 155 durable authorities, 275 durable-coupled projections, 262 loss-safe bindings, 180 stateless engines/catalogs, 43 test/development adapters, and 17 deliberately unbound projections, with zero undispositioned candidates. The inventory gate passed 3/3, the shared snapshot primitive passed 10/10, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for anonymous or escaping local collection roots and the standing loss, privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Escaping closure collection coverage is added in 375773f3147294d446a2a59644f0bf4ab311fd1b. The AST gate now detects mutable maps, sets, arrays, and objects that survive through returned functions or returned API methods, while excluding immediate request-local callbacks. All 32 new roots were reviewed: 28 record/revision/deletion/pending/serialization collections are coupled to durable snapshot owners; the per-run governance meter, achievement request coalescer, and bounded response cache are loss-safe; and Nyx event-action state remains a non-production adapter behind a production fail-closed injection boundary. Regression checks explicitly reject two immediate-callback false positives.

The resulting 964-entry registry contains 155 durable authorities, 303 durable-coupled projections, 265 loss-safe bindings, 180 stateless engines/catalogs, 44 test/development adapters, and 17 deliberately unbound projections. The inventory gate passed 3/3, representative closure-owner and loss-boundary suites passed 135/135, the shared snapshot primitive passed 10/10, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for returned collection objects, class-field and anonymous collection roots, plus the standing loss, privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Class-field and returned collection coverage is extended in e80e47e07f67946a4d090eb38af804bf0ac065cf. The AST gate now detects mutated collection-valued class fields, named Map/Set roots returned by identity or object property, and anonymous Map/Set properties created directly inside returned objects. All 407 new candidates were reviewed: 337 record/index/revision/deletion/snapshot views are coupled to their durable owners; five metrics/debounce/media-cache fields are loss-safe; 24 validation, normalization, ranking, geometry, and source-controlled catalog values are stateless; and 41 fields or returned roots belong to explicit test/development adapters behind durable, empty-reader, or fail-closed production bindings.

The resulting 1,371-entry registry contains 155 durable authorities, 640 durable-coupled projections, 270 loss-safe bindings, 204 stateless engines/catalogs, 85 test/development adapters, and 17 deliberately unbound projections, with zero undispositioned candidates. The inventory gate passed 3/3, representative Iris/Isis/Lilith/living-scenes suites passed 47/47, the shared snapshot primitive passed 10/10, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for anonymous collections passed as call arguments, returned array/object roots, aliased field mutations, and the standing loss, privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Returned array/object coverage is extended in 9c5bd1b0f74bbb042762a6cfd639d50d36b4bea2. The AST gate now admits locally initialized arrays and empty objects only when the constructing function mutates and returns them, including member assignment on object roots. All 513 new candidates were reviewed: 55 candidate/clone/export/replay/snapshot and durable-derived collections are coupled to their owners, 454 one-shot validation/normalization/ranking/aggregation/parser/engine results are stateless, and four roots belong to explicit test/development adapters.

The resulting 1,884-entry registry contains 155 durable authorities, 695 durable-coupled projections, 270 loss-safe bindings, 658 stateless engines/catalogs, 89 test/development adapters, and 17 deliberately unbound projections, with zero undispositioned candidates. The inventory gate passed 3/3, representative audit/accessibility/provider-quality suites passed 23/23, the shared snapshot primitive passed 10/10, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for anonymous collections passed as call arguments, alias-mediated mutations, and the standing loss, privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Anonymous Map/Set call-argument coverage is extended in 5948456fae341e03efa9b90ca1f3c8bca7255ffb. The AST gate now assigns stable callee-and-argument identities to anonymous Map, Set, WeakMap, and WeakSet instances passed directly to calls or constructors, while the production-source filter excludes assertion-only test values. All 28 new roots were reviewed: 16 deletion fences, migrated snapshot components, and nested candidate collections are coupled to their durable publish/snapshot owners, and 12 validation, deduplication, cycle-guard, normalization, domain-status, and preview arguments are stateless.

The resulting 1,912-entry registry contains 155 durable authorities, 711 durable-coupled projections, 270 loss-safe bindings, 670 stateless engines/catalogs, 89 test/development adapters, and 17 deliberately unbound projections, with zero undispositioned candidates. The inventory gate passed 3/3, representative assistant-session/channel-binding suites passed 30/30, the shared snapshot primitive passed 10/10, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for anonymous array/object call arguments, alias-mediated mutations, and the standing loss, privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Retained anonymous array/object argument coverage is extended in 120bee8452cad86ae534f56b6a35d2dfbfb51024. A census found 913 literal array/empty-object call arguments, but 886 are ordinary schema, SQL, cache-key, promise, validation, or deterministic inputs with no retention evidence. The gate therefore admits only the statically provable retained subset: values stored through Map.set or Set.add. All 27 new roots were reviewed: eight nested durable-state arrays are coupled to their owners, 12 one-shot graph/grouping values are stateless, and seven belong to explicit test/development adapters.

The resulting 1,939-entry registry contains 155 durable authorities, 719 durable-coupled projections, 270 loss-safe bindings, 682 stateless engines/catalogs, 96 test/development adapters, and 17 deliberately unbound projections, with zero undispositioned candidates. The inventory gate passed 3/3, device-session durability passed 5/5, the shared snapshot primitive passed 10/10, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. Non-retaining literal arguments remain intentionally excluded by the machine rule rather than bulk-labelled as operational state. The parent box remains open for alias-mediated mutation/escape analysis and the standing loss, privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Simple alias-mediated mutation/escape coverage is added in 44ddd9b7d6a48515c7fe89079270434747cff61c. Local collection roots are now followed through chained lexical aliases for closure mutation, direct mutation, and return escape; collection-valued class fields are likewise followed through local aliases. A pure per-source analysis seam and synthetic regression prove both an aliased local Map and an aliased class field. The current production corpus adds no candidates under this rule, so the reviewed registry remains at 1,939 entries with zero undispositioned roots.

The inventory gate passed 4/4 including the synthetic alias cases, the shared snapshot primitive passed 10/10, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for reassigned, destructured, parameter, and interprocedural aliases and the standing loss, privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

The ISIS cost-optimization and prompt-template preview engines are correctly classified as stateless through ae2fbf78aad7b5a32e6076dd270502948448d6e0. Neither engine retains request inputs, previews, histories, approvals, evidence, effects, or learning state; their public disposition contracts and inventory entries now agree. Route tests prove deterministic recomputation and empty retained histories. The reviewed 1,939-entry registry now contains 155 durable authorities, 719 durable coupled projections, 268 loss-safe ephemeral roots, 684 stateless engines or catalogs, 96 test/development adapters, and 17 unbound production projections. The two route suites and inventory regression passed 20/20, the direct BFF production build and artifact syntax check passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for the remaining retained-ephemeral loss audit and the standing privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Per-subject generation and personalization serialization tails are explicitly loss-proven through 0d81dcc2a51ea634f69c405e9605f6596e0a20ca. Their inventory entries now cite concrete restart tests: generation preserves subject erasure and recovers uncertain durable provider claims to needs_review without redispatch, while personalization preserves adjacent vectors and deletion fences and rejects stale resurrection. The three focused suites passed 50/50, the 1,939-entry inventory remained drift-free, scoped static checks passed, and the BFF ratchet contained only the same five unrelated findings. Registry counts are unchanged. The parent box remains open for the remaining retained-ephemeral cache and process-binding audit and the standing privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

All 25 module-scoped route response caches are bound to a shared, executable loss proof through 41e2bba2445a613d7fca6ce12ddae4239247830c. Registry schema version 2 records the tested factory, exact owner count, required loss-safe disposition, guarantee, and evidence. The checker now rejects owner-count, disposition, or proof-metadata drift, with synthetic negative coverage for the first two. The bounded TTL/LRU primitive's total-loss test proves clearing all entries yields a miss and recomputation without dropping domain state, idempotency, user history, evidence, lineage, approval, effects, or learning. The focused suites passed 13/13, inventory and static checks passed, and the BFF ratchet contained only the same five unrelated findings. Registry counts are unchanged. The parent box remains open for the remaining retained-ephemeral class-field, closure, and process-binding audit and the standing privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Achievement refresh debouncing, ambient WAV caching, and BFF metrics are bound to exact multi-root loss proofs through 4b3955a19ce11a30f68da94c997b26e693f0467a. The selector covers each singleton and its internal retained fields without sweeping unrelated collections into the claim. Existing reset/total-loss tests prove fresh idempotent achievement refresh, byte-identical ambient recomputation, and an empty replacement metrics collector; their capacity and privacy checks remain bound as evidence. Missing exact IDs, duplicates, count drift, and disposition drift now fail the registry gate. Four focused suites passed 29/29, inventory and static checks passed, and the BFF ratchet contained only the same five unrelated findings. Registry counts are unchanged. The parent box remains open for the remaining retained-ephemeral closures and process bindings and the standing privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

The process-local BFF ID fallback counter is removed through 279e1c4befc38a7913f0c9af23ce33d8f8d85103. The Node BFF now always uses node:crypto.randomUUID, eliminating retained state and the fallback's restart collision window. The reviewed registry drops to 1,938 entries: 155 durable authorities, 719 durable coupled projections, 267 loss-safe ephemeral roots, 684 stateless engines or catalogs, 96 test/development adapters, and 17 unbound production projections. ID and inventory tests passed 10/10, the direct BFF production build and artifact syntax check passed, static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for the remaining retained-ephemeral closures and process bindings and the standing privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

The realtime-music process connection counter is removed through 52f50014a7a1ee110dc6a204914da33e24d5a6f4. Sessions now use mrt2--prefixed node:crypto.randomUUID values, eliminating retained sequence state and restart-collision risk. A regression proves 25 IDs are unique UUID v4 values. The registry now contains 1,937 entries: 155 durable authorities, 719 durable coupled projections, 266 loss-safe ephemeral roots, 684 stateless engines or catalogs, 96 test/development adapters, and 17 unbound production projections. Realtime and inventory tests passed 12/12 with only the optional real-engine case skipped, the direct production build and artifact syntax check passed, static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for the remaining retained-ephemeral closures and reconstructable process bindings and the standing privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

Aja's deterministic watermark codec and the unbound agentic governance adapter are correctly classified through d833fcec0c7f2b623490a7fc385cc690ddf7c9a4. The codec is stateless and its production route proves deterministic encoding plus lossless round trips. The governance meter exists only in a test/development adapter; an import-graph regression proves no non-test source imports or composes it. The 1,937-entry registry now contains 264 loss-safe ephemeral roots, 685 stateless engines/catalogs, and 97 test/development adapters, with other counts unchanged. The three focused suites passed 19/19, inventory and static checks passed, and the BFF ratchet contained only the same five unrelated findings. The parent box remains open for the remaining retained-ephemeral closures and reconstructable process bindings and the standing privacy, deletion, retention, and recovery requirements. Evidence: evidence/v1-v9/V1-002.md.

The final retained-ephemeral audit closes V1-002's all-store durability box through 80873ce591e73471090129de651a049605bab858 and 7e7cd0662d2891018d02e6a19ce736fe58f18a1f. Exact loss proofs now cover achievement refresh coalescing, request-scoped admin durability checkpoints, PostgreSQL adapter/connect bindings, internal route-cache entries, the fail-closed payments runtime binding, reconstructable C2PA signing, and the durable crisis-frame store binding. Together with the earlier route-cache, class-field, metrics, and serialization-tail proofs, every retained ephemeral root is covered. A final inventory invariant rejects any future module, closure, or class-field ephemeral root without a shared/exact proof. Focused runtime and recovery suites passed 108/108, including 29 live-PostgreSQL tests; the final inventory gate passed 10/10; inventory/static checks passed; and the BFF ratchet contained only the same five unrelated findings. All 1,937 candidates remain dispositioned: 155 durable authorities, 719 durable coupled projections, 264 loss-safe ephemeral roots, 685 stateless engines/catalogs, 97 test/development adapters, and 17 unbound fail-closed projections. Evidence: evidence/v1-v9/V1-002.md.

Post-closure V1 Domain Workbench hardening closes I6.5.c with the reusable snapshot/keyset contract in 023ac1ecfe, then extends the existing production output route and its OpenAPI/SDK/CLI/operator-UI/Chromium consumers through ab10c989ea, 1512613f0c, and 5949061dec. Both output reads now use allowlisted keyset orderings with unique id tie-breakers, bounded pages, and opaque HMAC-signed cursors under one deployment key. /outputs/page pins a scan-wide creation-time snapshot; /outputs explicitly promises forward continuation rather than a transaction snapshot. The derived workload census recognizes both paths. I6.5.a's missing record kinds and I6.5.f's broader deleted/changed-row and role-tier fixture set remain open. Evidence: evidence/v1-v9/V1-002.md and evidence/v1-workbenches/isis-workload-catalog.json.

V1-003 — Prove the V1 reference vertical at level 4#

  • Generate real provider-backed artifacts for a bounded representative set of text and media slots.
  • Exercise live operator pause/edit/replace/approve/rollback through the UI.
  • Calibrate judge panels on licensed human gold and measure provider/model diversity, panel weighting, and disagreement.
  • Attribute engagement and explicit human outcomes without optimizing raw engagement.
  • Run controlled A/B or champion/challenger evaluation against the best authored baseline.
  • Publish/place under canary, observe SLOs and costs, then roll back safely.

Fail-closed evidence-contract source commit df17d74948d2d2428f83797609f96a10587d2905 defines a single fresh-run gate for all six requirements and retains the current machine-readable readiness report in evidence/v1-v9/V1-003-reference-vertical-readiness.json. Its adversarial suite passed 9/9. The present report is honestly blocked: this environment has no detected live provider credential names or current provider-probe artifact, and it cannot fabricate licensed human gold, authenticated UI receipts, a live controlled experiment, canary observations, or an executed rollback drill. Consequently every V1-003 checkbox remains open. Full boundary and unblock procedure: evidence/v1-v9/V1-003.md.

V2-001 — Close live commentary and NPC execution#

  • Wire real commentary, translation, TTS, face/lipsync rendering, NPC dialogue, and service operations with bounded fallbacks.
  • Add authoritative match-state fact verification to the live latency path.
  • Keep deterministic commentary fallback labeled and measurable.
  • Fence live cue and post-match reviewer prompts; retain deterministic safety checks when the model returns no findings.
  • Complete Hathor SAFETY-001 work and validate persona, intent, action, memory, and confidence outputs.
  • Prove rollback hashing, latency tiers, consent, voice disclosure, captions, audio ducking, and accessibility in the real path.

The invoked gRPC commentary boundary is closed in source commits 89e16c774fd123b869b1de3989e38f988c03da8a and 79d3c53c0407136c67eeded1a49f3c9a266cdfdc, plus prompt-fencing commit 0172b9c22b4a3b69ca9e446a918c708bd61ac324 and Hathor/V2 safety-boundary commit 4436396cc24e26b19abe5a3ba4d898ae97df5146, plus fallback-bank integrity commit c0ad0010628dd9fadeba671320942216f05e7007, with evidence in evidence/v1-v9/V2-001.md. CueRequest carries an immutable authoritative match-state snapshot, the service composition injects the semantic verifier, and generated claims must be entailed within the same latency budget. Missing state is rejected, while verifier absence, contradiction, provider failure, and budget failure produce distinct labeled deterministic fallback reasons. The same service object records generated, fallback, and rejected counts, fallback reason counts, fallback rate, and cumulative latency buckets/sum/max; the server exposes that recorder to its operational composition. The focused wire suite passes 18 cases, including a real in-process gRPC server, entailed generation, contradiction fallback, missing-state rejection, provider outage, rollback-origin rejection, metrics assertions, and duplex streaming. The live cue and existing post-match writer/reviewer boundaries now both use the shared typed trust-zone renderer; delimiter-smuggling is neutralized, and exact prompt hash/label evidence is retained. Nine post-match tests independently prove deterministic prompt-control blocking even when the model reviewer reports zero findings. The V2 NPC adapter now imports Hathor's canonical response types, binds personality identity to the requested NPC, and validates persona, declared intent/emotion, bounded actions, retrieved-memory attribution, calibrated confidence, answer threshold, and abstention semantics before returning or caching dialogue. Cached dialogue is revalidated, and the cache schema rejects older weak entries. The focused Hathor suite passes 110 cases and the V2 adapter passes 9, including adversarial foreign-persona, unknown-intent, path-like action, fabricated-memory, confidence-drift, malformed-response, unsafe abstention, and stale-cache cases. No real model, translation, TTS, face/lipsync, live provider NPC, deployed service, consented voice, cooked media, or accessibility result is claimed. The local fallback-bank hash gap is closed: all four launch banks are source-controlled, exact-byte SHA-256 validated, mirrored into Unreal defaults, cue-complete, and selected deterministically from the full fallback identity. That local integrity proof does not establish production rollback, latency, consent, disclosure, caption rendering, ducking, accessibility, or in-engine playback, so the remaining V2-001 and provider/execution V2-002 bullets stay open.

Live-execution gate source commit 1d7142e76f29dc32fbf057329bb69584b423d00a now rejects controlled provider doubles, URI-shaped voice placeholders, package plans without immutable rendered outputs, and configuration-only presentation claims. A fresh passing run must jointly prove real grounded commentary, translation, TTS, face/lipsync, NPC service execution, exact rollback hashes, both latency tiers, active consent, rendered disclosure and captions, measured playback ducking, accessibility automation, observed costs/SLOs, and a verified operational rollback drill. Its adversarial suite passed 9/9; the retained report is blocked on all six checks, so neither remaining V2-001 checkbox is closed. The old missing-host observation is superseded by the 2026-08-02 machine preflight: UE 5.5.4, RunUAT.sh, ueagent, and the required EOS editor binary are present, but the retained V2 editor build ID differs from the current engine build ID. Qualified automation/cook therefore requires a resource-safe V2Editor rebuild and still lacks the real provider/runtime package. Evidence and unblock command: evidence/v1-v9/V2-001.md and evidence/v1-v9/V2-unreal-host-preflight.json.

V2-002 — Produce the V2 flagship evidence slice#

  • Use one real match to generate fact-grounded live and post-match commentary.
  • Produce localized captions, consented voice, face animation, and signed provenance. [UE-AUD-004] In-game playback must bind V2/ue/V2.uproject on ue55-linux and retain runtime media with the executed evidence.
  • [UE-AUD-004] Produce/cook Isis cosmetic, venue, and related executor assets where the experience requires them on ue55-linux, binding V2/ue/V2.uproject, RunUAT BuildCookRun, target Linux, run-as user ueagent, and the cooked archive in executed evidence.
  • Bring manual MOTD/patch-note lifecycle into the common ledger.
  • Exercise operator intervention under the actual latency budget.
  • Measure naturalness, factuality, usefulness, latency, and player/ commentator preference on a production-representative cohort.

Manual live-ops lifecycle reconciliation is closed with evidence in evidence/v1-v9/V2-002.md. The shared COV-004 registry records the three V2 MOTDs and one V2 patch-note release as source-digest-bound manual-content rows with stable lineage, locale/accessibility requirements, lifecycle state, quantities, explicit gaps, and ownership. The MOTDs remain honestly stale and the patch-note release remains review; neither is promoted to live placement. The refreshed registry passes all 10 deterministic/adversarial inventory tests, both owning V2 services pass 7 focused tests plus typecheck/lint/build, and both V2 static lifecycle checkers pass. Current source-digest refresh commit: 2d29a2fc3b3d12840ee308eb4d99cef9daec70f4. The other five V2-002 bullets remain open.

Flagship evidence-gate source commit b238631d1a2af477815a7ebfb6f2a36794b6bde7 binds those five open bullets to one fresh match package and enforces the exact [UE-AUD-004] project, host, BuildCookRun, Linux, ueagent, Isis-asset, and cooked-archive identities. It also requires grounded live/post-match receipts, consented and provenance-signed runtime media, actual UI intervention within latency bounds, and signed cohort results for all five named qualities. The combined V2 gate suite passes 17/17. The retained flagship report remains blocked on all five checks; no live match, UE execution, or cohort result is fabricated. The current host preflight distinguishes the available UE toolchain from the stale V2 editor build and does not claim Isis assets, a cook, or playback, so no additional checkbox is closed. Evidence: evidence/v1-v9/V2-002.md and evidence/v1-v9/V2-unreal-host-preflight.json.

V3-001 — Close the V3 generation center#

  • Integrate real music synthesis, lyric generation, TTS, motion capture or AI assist, world-asset generation, and UE playback providers; any UE execution follows [UE-AUD-004] with V3/ue/V3.uproject on ue55-linux.
  • Preserve fail-closed Isis/Saraswati governance, tenant/release policy, fingerprinting, consent, and real c2pa-rs provenance.
  • Collect real pose input and validate motion classification/quality against human labels.
  • Complete live V1 registration, operator sign-offs, and operational concert evidence.
  • Rerun fresh launch readiness and keep remaining external gates honestly red.

Evidence: evidence/v1-v9/V3-001.md. The 2026-07-27 aggregate run now routes all 40 verify:v3 sub-gates through scripts/verify-manifest.json plus the avatar-pipeline test, rather than consulting the obsolete per-script package.json entries. Adversarial routing and evidence-integrity coverage passes. The fresh machine artifact remains correctly unpublished and red: 30/41 constituent commands and 24/39 named gates pass, 11 constituent verifiers fail on known readiness gaps, and V3/V3_TODOS.md retains 39 waived external or operational tasks. Source commit: a31b8063bb50871550019911b4c11b68c9ef9997. The unsupported-region waitlist regression is also closed at source commit 1de5d44c99e085879381241933fea585631d87df: explicit consent, a bounded delivery contract, durable owner mapping, restart recovery, export, erasure, retention, and accessible Playwright coverage restore the same honest 30/41 and 24/39 state. The other four V3-001 bullets remain open.

V3-002 — Produce the V3 flagship evidence slice#

  • Produce one original track and one concert segment through real providers.
  • Pass source/rights/consent/fingerprint checks and measured mix/loudness gates.
  • Perform human musical/performance review and audience preference testing.
  • Generate and verify C2PA lineage across every transformation.
  • [UE-AUD-004] Stage and play the content through V3/ue/V3.uproject on ue55-linux; record engine 5.5.x, UnrealEditor-Cmd, run-as ueagent, the selected target, exact argv, log/report, and playback artifacts.
  • Exercise intervention, canary, rollback, and operational recovery.

Evidence: evidence/v1-v9/V3-002.md. Source commit 78ab6bd3456f33fa0dde419740539dfa6ea41011 adds a strict fresh-run contract covering real track and concert provider receipts, rights/consent and dual fingerprint clearance, measured LUFS/true peak, signed human and audience review, four-stage C2PA lineage, exact UE 5.5 Linux playback identity and artifacts, and intervention/canary/rollback/recovery receipts. The retained report is honestly blocked on all six checks because no coherent live run exists; no checkbox is closed. The 2026-07-27 missing-host observation is superseded by the 2026-08-02 source-bound preflight: UE 5.5.4, UnrealEditor-Cmd, RunUAT.sh, ueagent, the V3 project, and its enabled EOS editor binary are present, but V3 has no retained UnrealEditor.modules and therefore needs a current V3Editor build before qualified playback. The build was resource-safely deferred while another checkout held about 7 GiB RSS on this 15 GiB host. This changes the blocker from host-unavailable to rebuild-required; it supplies no provider, cook, playback, or operational evidence. Machine record: evidence/v1-v9/V3-unreal-host-preflight.json.

V4-001 — Close V4 commentary and Unreal execution#

  • Wire real Calliope prose, Psyche voice, localization, native-speaker review, bias review, watermarking, consent, and review queue operations.
  • Keep deterministic templates distinct from provider output in evidence.
  • [UE-AUD-004] Run targeted V4 automation against V4/ue/V4.uproject on ue55-linux with /root/workspace/UnrealEngine-5.5/Engine/Binaries/Linux/UnrealEditor-Cmd, commandlet UnrealEditor-Cmd, run-as user ueagent, and target LinuxEditor; developers without UE dispatch v4-pr-gate.yml and retain v4-ue55-linux-evidence.
  • [UE-AUD-004] Verify adaptive-director bounds, procgen connectivity/placement, replay, and commentary integration in the same recorded V4/ue/V4.uproject runtime.
  • [UE-AUD-004] Drive the final-polish gate only from fresh provider evidence and schema-valid v4-ue55-linux-evidence; a skipped local run is not a UE artifact.

Repository evidence boundary (updated 2026-07-29): V4/release/final-polish-execution-evidence.json and apps/v4/scripts/src/v4-execution-evidence.mjs now make the strict final-polish command fail closed on live Calliope/Psyche receipts, explicit template separation, consent/watermark/provenance, 12-locale native-speaker review, human bias-review queue operations, and one schema-valid UE5.5 runtime covering director, procgen, replay, and commentary. Structural preflight is explicitly separate so it can run before the self-hosted UE job. Source commit 0950fe9c0df44c816feabaafebd15c9dae9de52a adds the exact V4.IntegratedRuntime target and passes all four named tests in one authenticated UE 5.5.4 LinuxEditor session as ueagent, with zero test warnings/errors and exit code 0. The targeted-automation and integrated-runtime checkboxes are therefore closed; live provider, human-review, and composed final-polish evidence remain open. See evidence/v1-v9/V4-001.md.

V4-002 — Produce the V4 flagship evidence slice#

  • Use one real match to produce end-of-match and live-spectator commentary.
  • Bind provider output to factual events, consent, native/bias review, audio provenance, and runtime playback.
  • Exercise operator intervention and rollback.
  • Run blinded player/commentator comparison against authored commentary.
  • [UE-AUD-004] Retain the exact V4 commandlet argv, engine/project identity, target, run-as user, exit code, Automation log/report digests, and media evidence defined by oshun.unreal-evidence.v1.

Repository flagship boundary (2026-07-27): scripts/v4/verify-flagship-evidence.mjs now requires one real match with distinct end-of-match and spectator provider receipts, factual-event grounding, all V4-001 governance and UE evidence, actual intervention/canary/rollback/ recovery, blinded player and commentator cohorts, and exact hashed runtime media/log/report artifacts. The retained report is honestly blocked on all six checks because no coherent live package exists; no checkbox is closed. See evidence/v1-v9/V4-002.md.

V5-001 — Close V5 authored-world and asset gaps#

  • Implement and invoke the declared AI Director with bounded adaptation, policy, telemetry, and human control. [UE-AUD-004] In-engine outcomes must come from V5/ue/V5.uproject on ue55-linux.
  • [UE-AUD-004] Produce real cooked binary assets through V5/ue/V5.uproject, RunUAT BuildCookRun, target Linux, and run-as ueagent; do not count references or metadata as game-ready content.
  • Integrate real TTS, recap, STT, and translation providers.
  • Complete human director, cell, continuity, rights, and release receipts.
  • [UE-AUD-004] Verify deterministic procgen, replays, highlights, broadcast shots, persona adaptation, puzzle rotation, clearance, and provenance in V5/ue/V5.uproject on ue55-linux; record the LinuxEditor commandlet evidence separately from any Linux cook evidence.
  • Preserve HMAC/replay integrity and fail-closed provider seams.

Repository evidence boundary (2026-07-29): scripts/v5/verify-authored-world-evidence.mjs remains the stricter aggregate release contract. It requires live TTS/recap/STT/translation and decision provider receipts, signed human approvals, packaged .pak/.ucas/.utoc output with launch smoke, and external key-custody evidence in addition to the repository-local UE runtime and cook. Those provider, human, package/launch, and external-custody claims remain blocked; the aggregate contract was not weakened. See evidence/v1-v9/V5-001.md.

Repository-local AI Director increment (2026-07-27): V5Gameplay now contains a portable deterministic decision core and a UGameInstanceSubsystem adapter with exact policy bounds, monotonic observation/replay protection, PvP, Trial, Speedrun, scripted-phase, and player-disable exclusion, adjustment rate limits, decision telemetry, and signed-receipt pause/override/rollback/release controls. The core passes a real C++20 warnings-as-errors ASan/UBSan compile-and-run harness; static integration checks prove that the UE adapter invokes that core and retains decision/control evidence. The qualified UE run now builds and invokes the subsystem against real policy assets and exercises bounded adaptation plus authority/replay behavior. This closes the repository-local director row without claiming a live external decision-provider receipt or a real human operator approval.

Repository-local replay-integrity increment (2026-07-27): the prior unkeyed SHA-1 replay digest omitted event kind and rounded float inputs. V5Core now provides one portable SHA-256/HMAC-SHA256 core shared with online JWT crypto, while V5Persistence signs an exact binary canonical packet binding all frame and security fields, expiry, key ID, and a single-use nonce. Runtime acceptance fails closed without receipt-bearing external key custody and durable atomic nonce consumption, clears transient key bytes, and audits every attempt. The portable core passes published vectors under C++20 ASan/UBSan, and adversarial UE automation covers event/precision tampering and nonce replay. The qualified UE run now exercises the runtime attack path and fail-closed authority seams, closing the repository integrity row. No live external key-custody or nonce authority receipt is claimed.

Qualified repository-local execution (2026-07-29): source commit cb562888f78ab1bba71d2e39b5ec12993f0f7047 ran V5/ue/V5.uproject on UE 5.5.4 as ueagent. One separate LinuxEditor invocation passed the exact 14 selected tests with zero warning-bearing, failed, or not-run tests, covering the director, deterministic procgen, replay, highlights, broadcast, persona, puzzle, clearance/provenance verification semantics, and HMAC/replay integrity. A separate non-iterative Linux RunUAT BuildCookRun cooked 448 packages and retained a 158,312,903-byte cook tree with SHA-256 1c9aa45c53b17fada652afe5baced5ef709e5ec3e31f241535e7cb6abc1fe609, including real cooked BootMap, director policy, and safety-policy packages. This closes four repository-local rows. The live-provider and human director/cell/continuity/rights/release rows remain open. Cook-only evidence does not claim a packaged executable or launch smoke; the stricter aggregate readiness report therefore remains blocked. See evidence/v1-v9/V5-001.md and evidence/v1-v9/V5-001-unreal-evidence.json.

V5-002 — Build the authored V5 case corpus needed by V8#

  • Convert all 25 intended cases into structured canon, culprit, means/motive/opportunity, clue, red-herring, deduction-graph, timing, asset, prose, VO, and outcome records.
  • Bind source rights, author identity, locale, tone/style metadata, and playable artifact refs.
  • Validate every graph for uniqueness, completeness, fair play, reference integrity, and runtime availability.
  • Reserve held-out cases for V8 regression/tone comparison and contamination control.

Commit d29e714026 establishes the fail-closed V5 authored-case corpus boundary without converting inventory metadata into human-gold content. The repository audit found 25 generic Vice Squad mission rows, 20 live-service pack skeletons, and 8 period cold-case retrieval skeletons, but zero complete case truth records, playable case assets, recorded case VO files, case-bound rights receipts, or held-out cases. The canonical corpus manifest therefore declares acquisition-required, contains zero cases, and inventories the three skeleton sources and their missing record classes.

The new strict validator requires exactly five cases for each of Patrol, Traffic, Homicide, Vice, and Arson; at least five contamination-controlled holdouts; human author identity and source/asset rights receipts; primary-locale prose, tone/style, and recorded VO; resolvable playable/runtime evidence; and complete canon, culprit, means/motive/opportunity, clue, red-herring, timing, outcome, and deduction records. It independently rejects duplicate IDs, unavailable clue artifacts, non-eliminable red herrings, cyclic or multi-solution graphs, solutions that omit required player-visible clues, missing files, and holdout leakage. Audit mode truthfully reports 0/25 and ready: false; release-mode --strict exits non-zero. The synthetic positive and adversarial validator tests plus the adjacent V5 mission/cold-case tests passed 7/7.

All V5-002 product boxes remain open. Completing them requires human authors, rights/identity receipts, prose and VO production, real playable Unreal artifacts, play evidence, and a sealed held-out split; none of those external content-acquisition facts can be manufactured from the existing numbered skeletons.

V5-003 — Produce the V5 flagship evidence slice#

  • Generate one playable side quest or daily detective experience and one replay-highlight bundle.
  • [UE-AUD-004] Cook, place, and validate both through V5/ue/V5.uproject on ue55-linux; record RunUAT BuildCookRun, target Linux, run-as ueagent, and the cooked archive digest.
  • Complete human receipts and signed provenance.
  • Compare experience quality against the best authored baseline.
  • Exercise canary, telemetry-driven defect discovery, revision, and rollback.

Repository flagship boundary (2026-07-27): scripts/v5/verify-flagship-evidence.mjs now requires the ready V5-001 foundation, two distinct cooked and runtime-completed outputs, exact Linux BuildCookRun archive identity, human-signed provenance, blinded authored- baseline comparison, and actual telemetry-led canary/revision/rollback receipts. The retained report is blocked on all six checks because no coherent live package exists; no checkbox is closed. See evidence/v1-v9/V5-003.md.

V6-001 — Make cognition durable, private, and interruptible#

  • Replace in-process-only run, budget, failure, kill/quarantine, and child state with durable common-spine checkpoints.
  • Preserve tier token/iteration limits, child-authority subset rules, kill scopes, and quarantine thresholds across restart and concurrent workers.
  • Add operator pause/edit/claim/resume/takeover after process loss.
  • Apply PROMPT-001/PROMPT-002 trust fencing to steward input, retrieved context, memory, story state, prior output, and critique.
  • Replace full prompt logging with PRIV-001 metadata/redaction policy (evidence/v1-v9/PRIV-001.md).
  • Add semantic faithfulness verification for Clio prose against event facts, beyond citation-ID membership.

Evidence: evidence/v1-v9/V6-001.md, evidence/v1-v9/V6-001-PROMPT.md, and evidence/v1-v9/V6-001-DURABILITY.md. Clio now requires an exact-prose SHA-256 receipt, complete claim-extraction attestation, verbatim narrative anchoring, bounded Ori-log retrieval proof, and passed passage-match plus claim-entailment proof for every factual claim. Stale receipts, incomplete extraction, unanchored claims, unknown citations, and non-entailment fail closed. The V10 Ori narrator supplies the production Moirai verifier seam, and the V6 drift gate now requires the adversarial narration suite. The shared V6 cognition gateway also renders steward input, persona, retrieved context, memory, story state, prior output, and critique as separately labelled untrusted zones under typed system, tool, and task authority; the Moirai mount forwards story state unchanged, and traces retain only its SHA-256 reference. The canonical durable runtime now checkpoints run, budget, failure, child, kill, and quarantine state into the RUN-001/RUN-002 common spine; reconstructible limit/authority bindings and optimistic child reservation preserve constraints across restart and concurrent workers. Ori Presence consumes that runtime in production and fails closed without its backed-up store path. The same store now commits operator pause, optimistic claim or takeover, versioned plan-step edit, dependency-bound reapproval, and resume through the HITL-003 intervention journal and RUN-005 lifecycle. The admin BFF derives tenant and actor from authenticated context, requires an agentic-operator scope plus caller-stable command identity/time, exposes only control metadata, and fails closed without the durable authority. Independent BFF/store instances complete the whole recovery sequence; competing claims serialize, retries are idempotent, edits abandon stale claims and invalidate approval, preapproval resume is rejected, and terminal runs cannot be reanimated. Source commits: 9362dd74cbadded485be7f036d224fb4617f55ea, 20cb046ad033ee9cb3df230735d196240a6f0328, d33a98f2cf35549c1eedcfbdf38906ddaabdd7da, 98ebbf320a11dd40cee3fad479a62e5811dd2461, and 50a3a76031f41f8485434694f7a887ae47254659. Clio passes 16 focused tests, the durable Iris runtime passes 223 tests, the focused transactional-intervention contract passes 13 tests, the BFF recovery/runtime surface passes 7 tests, the portfolio prompt-injection suite passes 232 tests, owning direct typechecks and the BFF production build pass, and the V6 Chronicle checker passes. This closes the durable post-loss control plane; it does not claim live provider calibration or traffic, multi-host load testing, a browser consumer for these new recovery routes, or V6-002 live in-run co-authoring and mixed-lineage publication.

V6-002 — Close V6 live operational seams#

  • Deploy and exercise WebTransport/SFU, vendor ASR/TTS, real V1 promotion, OTLP export, and live Themis dispatch.
  • Measure template fallback frequency and label fallback output to users and evidence; reduce fallback only through real provider reliability.
  • Bind human authoring to in-run takeover and mixed-lineage publication.
  • Test durable memory continuity, welfare/safety policy escalation, voice consent, latency, failover, and incident response.

Evidence: evidence/v1-v9/V6-002-FALLBACK.md. The cognition stack now derives a bounded fallback-frequency report from typed output provenance, and the Tier-2 web localization panel and diagnostics disclose the current 6/6 (100.0%) template fallback rate for English, Spanish, and Yoruba. Only real source: model output can lower the measured rate; presentation changes cannot. Source commit: c100c9f70350c9014abe5b8b3ad269ceebf7fe2b. Owning cognition, prompt-trust, Clio, and web unit suites, four direct TypeScript checks, the production web build, and the complete 10-test desktop plus 10-test mobile Playwright projects pass. This does not claim a reduced production rate or any of the remaining live-provider, in-run co-authoring, continuity, consent, latency, failover, safety, or incident-response bullets.

Operational-resilience evidence: evidence/v1-v9/V6-002-OPERATIONAL-RESILIENCE.md. The canonical assistant memory bridge and BFF Iris authority preserve consent-scoped memory through a fresh PostgreSQL-backed process reconstruction; the same exercise proves retention expiry, exact-subject erasure, adjacent-subject isolation, and post-erasure fencing. Lilith welfare thresholds trigger review and a safety-transfer departure, while the live local crisis cascade publishes through Redis and durably projects non-overridable synthesis and memory-write halts to every contemplative surface. TTS use remains bound to an active, tenant/instructor/purpose-specific voice lock. The real-time session suite enforces ASR, first-token, first-audio, avatar-frame, pipeline, and transport latency targets, and preserves identity, transcript, interruption, and memory continuity across reconnect, provider failover, and text fallback. Operator incident acknowledgement, escalation, assignment, resolution, redaction, and purge survive PostgreSQL-backed restart. The owning suites pass 575 tests, including three live local PostgreSQL/Redis integration cases. This closes the repository-owned operational exercise; it does not claim production traffic or the still-open live-provider/deployment exercise above.

Co-authoring evidence: evidence/v1-v9/V6-002-COAUTHORING.md. Human authoring now requires an active, operator-owned, checkpoint-valid Yemaya takeover whose durable AgentRun payload names the exact autonomous production run. The same binding must remain active at branch merge and publication, is embedded inside the signed replacement content, and travels with the persisted mixed-lineage publication record. The release route also requires current approval, source freshness, a gate-passing exact variant, a verified mixed-lineage signature, and the configured publisher's real reference. Source commit: 9ddf36de1b5a492ef9c3bf4a36b7476921361acc. Owning and broader BFF suites pass 84 tests, the common transactional contract passes 13 tests, the autonomy TypeScript check passes, and the direct production build passes. This proves the production code path with controlled seams; it does not claim external publisher traffic or the remaining V6-002 operational exercises.

Live-operations boundary (2026-07-27): scripts/v6/verify-live-operational-evidence.mjs now requires deployed WebTransport/SFU receipts, live consent-bound ASR/TTS with failover, signed production V1 promotion and readback, OTLP backend visibility, and acknowledged live Themis disposition. The retained report is blocked on all five checks because no coherent external exercise exists; the remaining checkbox stays open. See evidence/v1-v9/V6-002-LIVE-OPERATIONS.md.

V6-003 — Produce the V6 flagship evidence slice#

  • Have a persistent companion complete a multi-session dialogue, story, and biography arc with live speech and durable memory.
  • Exercise operator takeover, direct co-authoring, policy escalation, restart recovery, and rollback.
  • Produce signed mixed lineage and runtime placement evidence.
  • Run longitudinal human assessment of coherence, continuity, safety, latency, voice quality, trust, and emotional value.

Readiness evidence: evidence/v1-v9/V6-003-READINESS.md. Source commit 2233026d33e818b77303af1ebc4b1154bc326d8f adds the canonical fail-closed flagship manifest and validator. The gate requires at least three ordered live speech sessions bound to one companion, steward, and tenant; artifact-hashed transcript, story, biography, ASR, TTS, audio, and durable-memory receipts; later-session retrieval of earlier-session memory; takeover, co-authoring, policy escalation, restart recovery, rollback, mixed-lineage, and placement receipts; plus signed multi-assessor observations across all seven required dimensions over at least seven days. Six positive and adversarial tests pass. The canonical audit honestly reports 0 sessions, 0 assessments, and ready: false; strict mode exits non-zero. All four V6-003 product boxes remain open pending consented participants, live providers, deployed placement, and longitudinal assessment.

V7-001 — Enforce Forge and Mawu at the executable boundary#

  • Execute generated Forge code in a real capability sandbox that mediates literal and computed/dynamic calls.
  • Recheck capability policy after every refinement and at actual invocation.
  • Bind grant, egress, secret, resource, time, and process limits through AUTH-001.
  • Replace Mawu source-metadata-only evidence with semantic source-to-output grounding.
  • Keep the procedural reference backend labeled as procedural rather than model-generated.

Closure evidence: evidence/v1-v9/V7-001-MAWU-AND-FORGE-READINESS.md. Mawu now fails closed unless an injected semantic verifier supplies source-to-output claims and assessments that pass the canonical retrieval, passage-match, and claim-entailment proof bundle; its local reference backend is explicitly identified as reference-procedural-backend in artifact provenance, while absence of a product backend returns backend-not-configured with no artifact. The owning Mawu suite passes 23/23 tests. Forge's source-bound Moremi path now requires a unique, live, source-bound AUTH-001 use for every declared capability. Each generation and refinement re-enters static policy, AUTH-001, and sandbox admission; each Wasmtime host callback rechecks the capability binding and its 15-second lifetime while the store enforces grant-derived fuel, memory, and wall-time ceilings with zero egress, secrets, or processes. Forge now defines a deliberately restricted executable language: constant string bindings plus top-level oshun.invoke statements with strict JSON payloads. Moremi rejects every other statement, resolves literal and constant-computed calls, rejects unresolved runtime-dynamic calls, embeds the complete accepted program and canonical payload bytes in a core Wasmtime module, and audits the exact payload hashes at the host boundary. Complete Moremi 82/82, Forge Assist 25/25, and V6/V7 prompt boundary 67/67 suites pass. Complete-program source commit: 16d3588f7824d44cd31dbe6b9a54cf1722bd4f4d.

V7-002 — Close V7 headline backend gaps#

  • Integrate and measure real RL training and match-based meta discovery.
  • Integrate real ML terrain generation and reference-image vision.
  • Integrate product text-to-3D, texture, world, and NPC generation backends.
  • Integrate live Themis dispatch and full-engine parity checks.
  • Retain procedural/statistical cores as measurable baselines and fallbacks, not substitutes for the advertised neural experience.
  • Add model/data/license/provenance, cost, safety, quality, and failover evidence for every backend.

RL/meta closure evidence: evidence/v1-v9/V7-002-HEADLINE-BACKENDS.md. Crucible Agents now runs a bounded CPU one-step tabular Q-learning baseline against real crucible-core matches, persists learned Q parameters, derives its reward curve from completed matches, and evaluates the selected policy across 50 held-out side-balanced matches. Match-based meta discovery executes every candidate on both team sides, uses measured wins to select each generation's parents, and binds the run to a SHA-256 result digest. The fixture completes 6,000 training frames and 48 evolutionary matches; Crucible Core 11/11, Crucible Agents 12/12, and the Python seam 4/4 pass. PPO, distributed GPU, self-play, and the Python execution path remain explicitly unsupported rather than inheriting the bounded baseline's evidence. Source commit: bdf6e7113ce0aca52e0e5c3081c04226637f0465.

Terrain/vision closure evidence: evidence/v1-v9/V7-002-HEADLINE-BACKENDS.md. Loom now has two distinct, executable CPU learning paths. loom-inverse measures actual RGB pixels and trains an 8→12→4 MLP on 16 images, measuring four held-out images at MAE 0.032510392. loom-terrain-ml trains a 10→12→1 MLP on six sketch/genome heightmap tiles (384 pixels), measuring two held-out tiles (128 pixels) at MAE 0.013766719, then generates heightmaps from the learned parameters. Both reject duplicate train/validation identities, validate retained parameter digests, and reproduce the same metric/model on repeat. Both owning suites pass 7/7 with format and warning-denied clippy gates. These are bounded local MLP baselines, not CNN/diffusion or evidence for Earthbender, TerraFusion, Mesa, production corpora, or product-scale quality. Source commit: 8e887a83cc2fe5382164f8ec3f9fe73d6bcc95e0.

Product-generation closure evidence: evidence/v1-v9/V7-002-HEADLINE-BACKENDS.md. Mawu now exposes one asynchronous, fail-loud product execution path for text-to-3d, text-to-texture, text-to-world, and text-to-npc. Each tool has a separately configured HTTP route and provider identity; requests carry fenced prompt/audit material, and bounded JSON responses must match the request, tool, model, provider, media, and tool-specific payload contract. Successful output re-enters semantic grounding, attribution, provenance, content hashing, and output-trust checks. Denied policy does not dispatch; absent configuration and structured provider failures produce no artifact. The owning suite passes 27/27, the V6/V7 injection boundary passes 67/67, and the targeted library typecheck, lint, and Nx build pass. No live external endpoint or credentials were available, so credentialed output, model/data/license, quality, cost, and failover claims remain open under the per-backend evidence task. Source commit: 53f1c31d29d9f3317c5d4f42633fcf55b6db0bfd.

Themis/parity integration progress: evidence/v1-v9/V7-002-HEADLINE-BACKENDS.md. Eunomia now has a concrete, fail-loud HTTP dispatcher in addition to its route constructor. It requires secure complete configuration, bounds requests and responses, never dispatches rejected lifecycle input, and accepts delivery only through a schema- and identity-bound receipt. Crucible parity references now bind a supported runtime, engine build, executable SHA-256, run/artifact identities, UTC completion, and the complete scenario/config digest; parity requires exact state-hash equality plus a ≤1% margin delta. The owning Eunomia suite passes 18/18, its app typechecks/lint/build pass, and Crucible Core passes 11/11 plus format and warning-denied clippy. The combined checkbox remains open because this host had no configured Themis endpoint/credentials and no local Unreal Engine 5.5 executable, so no external delivery receipt or independent full-engine result exists. Source commit: a95341090610bfef0010fdc9c2d6e5aed62a44fb.

Baseline/fallback governance evidence: evidence/v1-v9/V7-002-HEADLINE-BACKENDS.md. The bounded RL and local MLP paths retain measured learned parameters and held-out metrics while named unsupported backends fail loud. Terrain's named neural/product paths remain procedural_stand_in, Mawu's local implementation remains reference-procedural-backend, Loom Agent identifies its planners as rule-based with target-versus-measured reward semantics, and Crucible labels design priors and uncomputed balance results. Loom Diffgen schema v2 now labels prompt-hash guidance and trigonometric erosion as procedural, records that no trained model or perceptual backend is configured, executes bounded optimization instead of returning hardcoded benchmark quality, and derives its winner from measured final loss. The owning suite passes 5/5 plus format and warning-denied clippy. These results establish measurable baselines, not CLIP, neural erosion, diffusion, or product quality. Source commit: 171d5d624624df6f9f287f80702cb76a477d24bd.

Per-backend operational-evidence enforcement progress: evidence/v1-v9/V7-002-HEADLINE-BACKENDS.md. The new fail-closed catalog inventories all ten V7 headline execution boundaries and requires exactly one fresh, content-digest-bound record for each. It fixes every boundary's implementation class and requires exact implementation/model/data/output-license identity, run/artifact provenance, measured cost/latency, passing safety and threshold-derived quality results, and an exercised class-appropriate recovery path with zero data loss. Model-backed paths cannot mark undisclosed data or licenses not applicable. The owning suite passes 35/35, strict production and test lint, production typecheck, and the targeted Nx build. The checkbox remains open honestly: validation fixtures are synthetic, and no retained production/provider/full-engine evidence catalog was available. Admission contract source commit: 31f4ee1b6e7cf3e673230c45d6f6f516b52abc58.

V7-003 — Produce the V7 flagship evidence slice#

  • Generate one bounded world region with terrain, assets, NPCs, economy, quests, and required localization/accessibility.
  • [UE-AUD-004] Cook and place it through V7/ue/Mawu.uproject on ue55-linux; record RunUAT BuildCookRun, target Linux, run-as ueagent, and the cooked archive digest.
  • Simulate, play, review, sign, and canary the complete bundle.
  • Capture measured failures from players/simulation/operations and revise the region through the governed run spine.
  • Compare procedural, model-generated, and best-authored baselines.

Bounded-region evidence: evidence/v1-v9/V7-003-FLAGSHIP-REGION.md. The deterministic region:mawu-market-quarter:v1 procedural fixture composes all four governed Mawu reference artifacts into a 128 m square region with 16 terrain cells, 4 placed assets, 3 scheduled NPC roles, 5 zero-sum ledger transactions, 2 dependency-checked/reward-bound quests, and 15 required strings complete across English, Spanish, and French. Its content contract also requires captions, non-color markers, keyboard/gamepad parity, 200% text scaling, reduced motion, and positional-audio indicators. Complete regeneration produced digest 32e93d6c7b7ef1cbbf208aa62f1a2a643bd2561bca54fc34986005a10809f89f and zero evaluator violations. A deterministic 32-session simulation then measured the world-layer asset as unreachable in all sessions: 96/128 navigation successes, 64/64 quest completions from structured target bindings, 160 balanced ledger replays, and zero imbalance events. The admitted revision moved that exact asset to reachable terrain through common run run_b161fd0cb90ef0a2e9ac851d99797eee241f0b92; the sealed after-region c04723a73696bb98601032e1e6d45d2e7b12bfe0db7b8ba585ae08e91d1da42a replayed at 128/128 navigation successes with no failures. The common run is explicitly system-authored and full-auto; it does not manufacture human approval. The full owning suite passes 45/45, the common adapter/trace slice passes 8/8, and strict lint, both production typechecks, both targeted builds, deterministic replay, and adversarial nested-evidence validation pass. This closes the bounded manifest and measured simulation-revision tasks only: it is explicitly a procedural reference baseline, not a UE cook, product-model result, playable realm, human review, signature, or canary. Source commits: a21a5593d90a3ff3f0a85d9bf6045ea2ea9cd30e and d8d564fa633c12e5fbbc7b23479676487d622a7b.

Flagship admission boundary (updated 2026-07-29): scripts/v7/verify-flagship-evidence.mjs now requires the exact UE 5.5 Linux cook and runtime placement, full-engine simulation plus human play/review and signed canary evidence, and one blinded procedural/model/best-authored comparison. Executed run v7-003-4d4ff1a6efc4-20260729T221647Z, source commit 4d4ff1a6efc43f033804779fdd752329f5dc9844, passed all 12 Mawu Automation tests, RunUAT BuildCookRun for Linux as ueagent, and runtime placement of the exact digest-bound 3-layer/23-primitive region from a 125,741,681-byte staged archive with SHA-256 a446bb3fc0b83b6812a25d70596f36e7cf4bdda9baef92edc96e5e1b026deb14. The retained package therefore passes unrealCookAndPlacement; that checkbox is closed. The package remains blocked on the human simulate/play/review/sign/canary and three-way baseline checks, so those two boxes stay open without inheriting claims from automated placement or the procedural simulation. See evidence/v1-v9/V7-003-FLAGSHIP-READINESS.md.

V8-001 — Restore current contract/document drift integrity#

  • Run node libs/yemaya/case-contracts/tools/gen-v5-target.mjs against the current V5 headers and inspect the generated semantic diff. (2026-07-17: diff was cosmetic only — regenerated header emits /* eslint-disable */ where the committed file had a blank line; enum/type content identical.)
  • Rerun node scripts/verify-v8-docs-drift.mjs and all dependent contract, case-engine, compiler, and verifier tests. [UE-AUD-004] V8 has no standalone .uproject; any UE-facing test must first record an approved owning project rather than silently selecting another version's runtime. (2026-07-17: gate passed; case-contracts 17/17, case-engine 20/20, case-compiler 10/10, case-verifier 26/26. No UE-facing test was run.)
  • Commit the regenerated contract only when it matches intended V5 runtime semantics; do not green the gate by accepting an unreviewed drift. (2026-07-17: committed — semantics unchanged.)
  • Store current command, HEAD, host, generator version, and output evidence. (2026-07-17: command node libs/yemaya/case-contracts/tools/gen-v5-target.mjs at HEAD a0bf7555000895ff3df5d3f68bfda5588533c4be on Elikems-MacBook-Air.local; verifier output "contracts in sync with V5 headers; package graph intact"; details in the fix commit body.)

Closure evidence: evidence/v1-v9/V8-001.md.

V8-002 — Make journal and HITL state multi-process safe#

  • Allocate StageJournal sequence and attempt numbers atomically in the durable store.
  • Support restart and concurrent journal writers without duplicates, nonmonotonic order, or lost events.
  • Replace process-local HITL IDs/counters with durable allocation.
  • Apply edit-and-approve changes to a versioned case transactionally.
  • Invalidate dependent solver, quality, provenance, approval, and release evidence after edits.
  • Rerun required gates, sign mixed lineage, and resume publication through one edit-apply-reverify-resume transaction.
  • Add restart, race, duplicate, stale-review, conflicting-edit, and replay tests.

StageJournal allocation hardening through d430228c4eb9f95c0dd656e5f49d9de205987227 removes the process-local sequence and attempt maps. JournalStore.appendAllocated is now the single allocation and persistence boundary, and StageJournal publishes only after that durable boundary returns the stored event. The in-memory implementation prevents duplicate (caseId, seed, stage, attempt) keys, continues sequence and attempt allocation across replacement StageJournal instances sharing the same store, and no longer persists an event a second time when the event bus publishes it.

Commit bb75ba13357740ad56ffec86bfeebaa8af0434f2 completes the durable boundary with a production PostgresJournalStore, canonical @oshun/database client, and migrations 202607230001 and 202607230002. Allocation and insert execute in one transaction under a transaction-scoped advisory lock keyed by case and seed. Database primary/unique constraints reject duplicate (case_id, seed, stage, attempt) and (case_id, seed, seq) keys, while check constraints enforce the journal stage/status vocabulary. Production dependency construction now fails closed unless V8_JOURNAL_DATABASE_URL or YEMAYA_DATABASE_URL is configured, and every mint handled by one server uses the same durable journal.

The focused 30-test pipeline/adapter run and the complete yemaya-case-contracts and yemaya-case-pipeline test, lint, and typecheck targets passed for the allocation-contract slice. For the Postgres slice, v8-loom-service test, lint, and typecheck and all three yemaya-case-contracts targets passed. A live Postgres run passed all three adapter cases: two independent pools concurrently persisted 60 events with exact sequences 1..60 and attempts 0..59; a closed and recreated pool then resumed at sequence 61 and attempt 60; an explicit duplicate attempt was rejected; and ordered replay returned all 61 rows without duplication or loss. The migration CLI applied both migrations and a repeated run was idempotent. These checks establish database-backed restart and concurrent-writer behavior; they do not claim that the separately injected live event bus provides cross-process delivery.

Commit 44df7eff1b7eabcaf4d113e569eb54ae7bb1adc0 replaces the process-local HITL counter in production with PostgresHitlQueue and migration 202607230003. PostgreSQL owns the monotonic allocation sequence; allocation and queue insert commit in one transaction. Resolution is an atomic compare-and-set from pending to resolved, so racing reviewer decisions cannot both succeed. The queue persists ground truth, verifier-report reference, context, submitter, status, timestamps, and the final decision, and production server composition shares the durable queue across mints while continuing to expose the in-memory queue only on the explicit offline path.

The complete v8-loom-service and yemaya-case-pipeline test, lint, and typecheck targets passed. The live Postgres HITL run passed both cases: two independent pools concurrently allocated 60 unique IDs spanning sequence values 0..59; a closed and recreated pool resumed at 60; and two simultaneous decisions produced exactly one success, one rejection, one resolved record, and 60 still-pending records. The migration applied successfully, its second run was idempotent, database constraint inspection passed, and the temporary live test table and sequence were removed.

Commit 5f1368dc804923338ecae92cf0fcb3bf7451bcd5 adds the immutable PostgresCaseReviewStore and migration 202607230004. Production mint results initialize revision 1 in the durable store. editAndApprove locks both the pending review and current case lineage, requires the review's expected revision to remain current, applies a prototype-safe merge patch, writes exactly one immutable successor revision, invalidates the solver, quality, provenance, approval, and release evidence ledger, and resolves the review in the same transaction. Direct PostgresHitlQueue.resolve now rejects edit-and-approve, preventing callers from resolving the review without changing the versioned case.

The complete v8-loom-service test, lint, and typecheck targets passed, as did the live Postgres transaction suite. The live proof created revision 2 while retaining revision 1 byte-for-byte, invalidated all five evidence classes, and atomically resolved its review. A stale revision-1 review rolled back and remained pending. Two independent pools then raced competing revision-2 edits: exactly one committed revision 3 and resolved its review, while the loser rolled back and remained pending. A closed and recreated pool replayed revisions 1,2,3 in order. Together with the earlier journal duplicate/replay and HITL restart/race cases, this covers every test category in the final V8-002 test box. Migration application, idempotency, constraints/indexes, and temporary-table cleanup also passed.

Commit 80078c9c6ebf1ba96ba20c07f51053444ee9e73c completes V8-002 with PostgresCaseReviewStore.editApplyReverifyResume and migration 202607230005. One database transaction locks the pending review and current revision, applies the edit as a held revision with all five dependent evidence classes invalidated, reruns verification, authenticates the Ed25519 release envelope and timestamp, binds the mixed human/agent lineage to the edited content and asset-manifest hashes, writes the published successor revision, and resolves the review. Any failed gate, lineage mismatch, stale revision, or signature failure rolls back the revision writes and review resolution together.

The complete v8-loom-service test, lint, and typecheck targets passed. The live Postgres and real-cryptography proof produced held revision 4 and published revision 5 with fresh valid solver, quality, provenance, approval, and release evidence; authenticated the release report; and retained the human edit in the signed lineage. A deliberately corrupted release signature then rolled back the whole attempted transition, leaving revision 5 current and the review pending. Migration 202607230005 applied successfully, a second application was idempotent, lifecycle and authenticated-report constraints were present, and temporary live-test tables were removed.

Closure evidence: evidence/v1-v9/V8-002.md.

V8-003 — Correct provenance and asset-quality semantics#

  • Preserve locator versus byte hash scope and signed status in aggregate records and UI.
  • Download/hash authoritative asset bytes and verify real C2PA claims before upgrading proof strength.
  • Implement and verify an authenticated case-report signature; retain any unkeyed digest only as a separately named supplemental checksum.
  • Add seeded best-of-N, perceptual, period, canon, identity, typography, and accessibility evaluation to asset generation.
  • Retain every candidate, seed, prompt, model/provider, selection score, rights receipt, and rejected/held reason.

Commit 54f183278714144ac35bdbca2e0aac4838a1a496 established the real V8 asset-verification and selection path. The verifier only accepts allowlisted HTTPS or loopback locators, downloads bounded authoritative bytes without redirects, hashes those bytes, validates identity-bearing C2PA claims through verifyC2paExport, requires creation action and software-agent evidence, and then upgrades provenance to hashScope: asset-bytes and signed: true. Generated assets use deterministic seeded best-of-N selection; every candidate is evaluated against exact perceptual, period, and canon evidence plus the modality-specific media bundle, whose required dimensions cover identity, typography, and accessibility where applicable. Missing, malformed, mis-scoped, or sub-threshold evidence holds the candidate instead of selecting an arbitrary fallback.

Commit 2968665255d826439aa8ca355aaa397f9b605182 supplies the authenticated release boundary: an Ed25519 case-report signature, signer identity and active key policy, independently signed timestamp token, exact case/seed/manifest and lineage bindings, rotation and revocation handling, and offline verification. The unkeyed report value remains separately and accurately named checksumSha256; tests establish that it is not the authenticated signature and that substitutions, stale timestamps, revoked keys, and lineage mutations fail.

Commit b6e028375daf40e4ca373879229240dc59d65759 completes governed retention. Candidate records now preserve the exact artifact rights bundle alongside seed, prompt, provider, model, source locator, artifact digest, media evaluation, selection scores, selected state, and every failure or hold reason. The selection gate validates active output, model, and dataset receipts against the exact candidate subject. Durable case revisions retain the full aggregate provenance record through initial save, human edit, and re-verification; reject aggregate/manifest substitution; and expose a fail-closed authenticated GET /v8/cases/:caseId/provenance projection for both held and published cases.

Commit 453a67708c213eb486ca293abb126d558a920787 adds the operator surface at /operator/v8/provenance. Its server loader validates the Loom response and publication invariants before rendering. The restrained evidence ledger keeps locator digest versus verified asset-byte scope and unsigned versus C2PA-signed state explicit, and exposes the candidate trajectory, scores, prompt, provider, model, rights-receipt identities, rejected/held reasons, release authentication, and lineage without collapsing those meanings into a generic “verified” badge.

The complete yemaya-case-contracts, yemaya-case-assets, and v8-loom-service owning test, lint, and typecheck targets passed; the focused asset selection/provenance run passed all 26 cases. The live Postgres review suite passed both durable provenance cases. The web app typecheck passed with zero app-owned backlog, lint passed with zero errors (25 pre-existing warnings), and the focused loader/component Vitest run passed all 6 tests. Targeted Playwright passed both Chromium desktop and mobile projects, including protected route authentication, locator/byte and signature semantics, candidate and rights evidence, selection interaction, mobile overflow, and the shared serious/critical accessibility scan.

Closure evidence: evidence/v1-v9/V8-003.md.

V8-004 — Calibrate formal validity separately from experience quality#

  • Preserve TruthWeaver/ClueSmith/DPLL/Minos uniqueness, completeness, fair-play, balance, and player-available-proof gates as hard correctness.
  • Fence canon input and strengthen optional LLM surface-field schema and semantic validation.
  • Treat Theseus and deterministic G5/PCA outputs as diagnostics until calibrated against authored cases and real players.
  • Build licensed authored prose/VO/tone gold, including the structured V5 corpus from V5-002.
  • Measure clue comprehension, deduction success, frustration, fairness, tone, delight, replayability, and player preference.

The deterministic case core continues to derive clue and deduction state from the solve-first ground truth, and the Minos verifier retains uniqueness, completeness, fair-play, balance, and player-available proof as independent release gates. Optional LLM enrichment remains confined to narrow surface fields: llm-proposer.ts places canon/specification/constraints in explicitly untrusted prompt zones, parses a strict proposal shape, and rejects culprit, identifier, tag, or semantic substitutions before they can affect the deterministic kernel. The owning engine and verifier suites cover poisoned canon input, invalid proposal semantics, ambiguous or incomplete cases, withheld essential clues, and compiled cases whose runtime deduction graph is no longer solvable.

Commit d482d553e78fcf8cf9ec0a5ddea47a1d1a9a9a7f establishes the two non-substitutable experience-quality evidence classes. Authored-gold readiness requires the generic rights, consent, rater, holdout, and contamination governance contract plus at least 25 complete structured V5 items, each with a human-authored or human-edited case record, prose, VO, and tone profile. Real-player readiness recomputes clue comprehension, deduction success, frustration, fairness, tone, delight, replayability, and player preference from raw, consented, target-audience observations across independent preregistered waves. Exact candidate digest, authored-corpus identity, locale, instrument, randomization, blinding, exclusion, approval, timing, and per-dimension sample thresholds all fail closed.

Commit aa054059a4d3e60c5863ee7e965dd1ac4b7c67a7 wires that evidence into the Loom publication boundary. The service hashes the exact compiled draft, asks the governed provider for evidence bound to that reference and digest, recomputes the study decision, and forces G5 false when the provider or evidence is missing, throws, substitutes the authored-gold manifest, or fails any dimension. Theseus, the structural heuristic, and PCA therefore remain diagnostic-only until both external evidence classes pass; the production composition does not install synthetic evidence. The v8-loom-service typecheck, complete test target, focused pipeline/adversarial run (20 tests), and lint target passed.

The final two boxes intentionally remain open. No sealed, licensed 25-case V5-002 corpus or real-player observation set exists in the repository, and test fixtures are explicitly synthetic and test-only. Those facts cannot be manufactured by an implementation change; V8 publication now exposes them as hard evidence-acquisition blockers instead of treating heuristic scores as proof of player experience.

V8-005 — Produce the V8 first level-4 reference slice#

  • Generate cases through real model/media providers under the solve-first truth kernel, the authoritative eight-gate ReleaseDecision, and the separate seven-gate platform content harness.
  • [UE-AUD-004] Place, cook, and play them with real media and accessibility in the owning V5/ue/V5.uproject, which enables V8_Ariadne_CaseClient through ../../V8/ue; record the ue55-linux editor, Linux cook, archive digest, placement, play, and accessibility evidence.
  • Exercise zero-human route and full human takeover/edit route.
  • Compare against held-out best-authored cases with predeclared statistics.
  • Run provider failure, journal restart, live-service reliability, canary, incident, rollback, and recertification drills.
  • Obtain GA sign-off only after authored regression/tone data and player outcomes satisfy policy.

Commit 977d459e7a37e3fb785058d969770298105008e2 adds a fail-closed V8 level-4 reference-slice decision to yemaya-case-eval. It requires four distinct non-fixture production provider roles (case-surface LLM, media generation, semantic safety, and media quality), each bound to the exact release artifact with request, response, credential-owner, version/model, rights, and output-digest evidence. It separately requires the named V5 Unreal project, UE5.5/Linux/ueagent, editor and cook commands, a cooked archive digest, placement, play, accessibility, and a policy-minimum set of unique real-media asset digests.

The same evaluator requires exact-artifact executions of both the zero-human authoring and full human takeover/edit routes; the latter must retain review, edited-artifact, and mixed-lineage evidence. Provider failure, journal restart, live-service reliability, canary, incident response, rollback, and recertification must each have a unique planned injection, observation, and recovery record. Reliability success rate is recomputed from raw request and success counts, p95 is thresholded, evidence freshness is enforced, and GA approval must follow every evidence timestamp and remain unexpired. Tests reject fixtures, artifact substitution, TypeScript-literal/status spoofing, duplicate media, missing takeover/rollback, undersampling, low success rate, excessive latency, and premature approval. The focused suite passed all 6 tests; the full owning typecheck and test targets passed, and lint passed with only the 10 pre-existing warnings in unrelated case-eval files.

All V8-005 product boxes remain open: the repository does not contain the required real provider executions, cooked/playable real-media archive, complete zero-human and takeover runs, preregistered authored comparison, operational drills, or post-evidence GA approval. The evaluator makes those external proofs non-substitutable instead of converting implementation or fixtures into GA evidence.

Concurrent admission boundary (2026-07-27): scripts/v8/verify-flagship-evidence.ts now retains one machine-readable package over the canonical reference-slice evaluator. It additionally requires exact-artifact G1–G7 execution receipts and a preregistered, held-out, blinded, randomized, independently rated best-authored comparison whose lower 95% candidate-preference bound exceeds 0.50. The retained package is blocked on all six checks, so every checkbox above remains open. See evidence/v1-v9/V8-005-FLAGSHIP-READINESS.md.

Gate-contract documentation audit (2026-07-27): the code-grounded V8 architecture and feature companions now distinguish the authoritative G1–G8 ReleaseDecision from the separate G1–G7 platform content harness. scripts/verify-v8-docs-drift.mjs enforces both memberships and rejects the stale seven-gate pipeline wording. The TypeScript and authoritative Python generated-pack validators now require the exact G1–G8 set, reject missing G8 and substituted gate names, and reject malformed gate containers. This closes the generated-manifest admission loophole. The authenticated release-report contract now consumes the same exact gate type, and both signing and offline verification reject partial, substituted, malformed, inherited-only, or non-green gate maps. This closes the corresponding signed-envelope loophole; it does not supply any of the external production evidence required above.

V9-001 — Connect generation, verification, adaptation, and publication#

  • Trace one production path from approved source package through outline, rich generation, claim verification, pedagogy/STEM/process gates, adaptive planning, experience rendering, review, signing, promotion, and learner delivery.
  • Make the P0 verification gate mandatory for ordinary generation; optional best-of-N/refine paths must not be the only safe path.
  • Make a second invalid rich-generation response fail closed or route to explicit review rather than return it as-is.
  • Label the Python source forge's deterministic outline/lesson templates as templates until real generation is invoked; do not count structural percentage gates as factual, pedagogical, or experiential quality.
  • Implement real source-grounded lesson generation on the primary path; retain deterministic templates only as labeled availability fallbacks.
  • Supply approved source bundles and replace citation-shape validation with retrieval and entailment.
  • Emit transactional evidence per stage/node and prove TypeScript verification evidence governs Python promotion.

Commit 45d3467e715a88841c8e9f2cfbe14ad7768105dd makes the ordinary ContentGenerator.generateLesson() entry point invoke the verified generation path. It now requires a non-empty approved source bundle and a configured P0 verification gate, and returns a lesson only for an exact pass decision. Persistent block and needs-human decisions throw LessonReleaseBlockedError with the exact bound verification result; the latter is explicitly recorded as released: false. Non-release authoring remains available only through the clearly named generateLessonDraft() entry point. Tests prove that missing gate configuration does not invoke the provider, that pass releases, and that persistent block and needs-human retain their evidence and do not reach ordinary learner release. The full metis-verification:test target, both owning library typechecks, and the focused verification/generation suites passed.

The rich-generation dispatcher also validates the stricter retry and throws GeneratedContentValidationError when the second response remains invalid; there is no return-as-is branch. Its regression test exercises two malformed responses and requires the typed failure. Commit 41dfbcbfeadf236f84520ac6a0837a9adb3219c2 then aligns the complete metis-llm-client test corpus with that fail-closed source policy: simulated lessons no longer invent bibliographies, source references are emitted only from supplied bundle IDs, content type is explicit request metadata, and prompt version assertions inspect the actual task message. The five repaired files passed 496 tests; the complete owning test target and library typecheck passed, and lint completed with 0 errors and 95 pre-existing warnings.

Commit 8dd9d9c7f0d0a0d475b1d28419fee7f8a0fcaecf makes the Python source forge's actual contract truthful without pretending that deterministic scaffolding became model-authored content. Outline, generation, and lesson responses now carry immutable authoring_mode: deterministic_template, artifact_status: template_draft, release_eligible: false, and quality_evidence_scope: structure_only fields. Rendered lesson text begins with a non-release template disclosure, so the limitation survives metadata stripping. API summaries, persisted-response defaults, structured logs, and the canonical OpenAPI artifact expose the same semantics. The focused service and contract suite passed 24 tests, downstream evaluation/refresh/benchmark coverage passed 13 tests, Ruff and strict MyPy passed, and canonical OpenAPI drift verification passed.

Commit 6c1ce4df9bba1e305c9d1082d1bc8c8d3a8c9cee adds the real OpenAI-compatible lesson-authoring path ahead of that fallback. It accepts only an approved, high-stakes-allowed package with nonempty package-scoped retrieval, fences specification and source evidence as untrusted data, parses strict JSON, and independently rejects citation IDs outside the retrieved set or missing inline citation markers. Exact provider/model, prompt/request/response digests, retrieved result IDs, citations, and token counts survive persistence. Model-authored lessons remain generated_draft, retrieval_grounded_unverified, and non-release; model/template aggregates are explicitly mixed. Provider or evidence failure records a labeled fallback reason, while allow_template_fallback: false fails closed and the API returns 503. Seven direct model-path tests, 84 source-subsystem tests with 2 expected skips, 26 focused contract tests, Ruff, strict MyPy, and OpenAPI/client freshness pass. No live provider execution is claimed. Evidence: evidence/v1-v9/V9-001-SOURCE-GROUNDED-GENERATION.md.

Commit 5fbf04e9aec333287b642952bfd0fe00f098ddf5 makes TypeScript P0 verification authoritative at the Python source-forge promotion boundary. @metis/verification now emits an exact-content-bound metis-p0-v1 receipt only when all required factuality, faithfulness, citation, code, process, and pedagogical verifiers pass. The trusted verifier signs the stable receipt digest with HMAC-SHA256. Python recomputes the same cross-language material, verifies the signature and exact lesson coverage, and rejects non-pass, unconfigured, incomplete, stale, mixed/template, or digest-mismatched inputs. It snapshots the approved manifest, extraction, outline, model generation, and passing evaluation; binds their dependencies and hashes as stage/node evidence; rechecks every snapshot immediately before an fsynced atomic promotion-ledger replace; and cleans temporary state on failure. TypeScript and Python assert the same known content, receipt, and signature digests. The complete verification suite passed 176 tests, the source/OpenAPI suite passed 109 tests with 2 expected extraction skips, strict focused typechecks and lint passed, and generated OpenAPI/client freshness passed. Evidence: evidence/v1-v9/V9-001-TRANSACTIONAL-PROMOTION.md.

The remaining V9-001 boxes stay open. These slices do not supply production approved source bundles or calibrated live evaluator evidence.

The authenticated-promotion-read increment documented in evidence/v1-v9/V9-001-AUTHENTICATED-PROMOTION-READS.md hardens the durable handoff without claiming that full path. Promotion records now carry a domain-separated receipt digest and HMAC over the complete canonical response. Every read authenticates that outer receipt, rechecks all persisted prerequisite digests, nested TypeScript receipt signatures, exact lesson coverage, canonical stage/dependency/storage metadata, top-level digest bindings, and the deterministic promotion ID before returning the record. Tests reject signed metadata mutation, signature forgery, post-promotion artifact mutation, and a correctly re-signed but noncanonical graph. The two compound V9-001 boxes remain open; at that increment they still required external production evidence and the adaptive rendering, review, signing, and delivery integration.

The signed-publication increment documented in evidence/v1-v9/V9-001-SIGNED-PUBLICATION-DELIVERY.md now supplies that repository-owned integration. An authenticated promotion can become a content-addressed adaptive candidate with escaped static rendering, complete assessment and study payloads, exact P0 receipt bindings, explicit cold-start abstention, and an owner-bound audience. Administrator-only inspection and idempotent review produce a signed decision; only an exact signed approval can be atomically promoted to a signed release. Every learner read reauthenticates the complete chain, enforces audience membership, and durably archives a signed delivery receipt before returning content. The first compound box remains open: fixtures and authenticated test roles are not evidence of an authorized production source, a live provider run, a real administrator review, a real learner delivery, or production operations for the path.

The REL-003 admission increment recorded in evidence/v1-v9/REL-003.md now places continuous recertification directly on that software path. The canonical TypeScript evaluator is the only issuer: blocked backlog, stale proof, missing exercise, or target mismatch cannot yield a signed admission. Metis independently authenticates the decision, exact candidate digest, graph/fingerprint/slot bindings, and expiry before release, persists it inside the immutable signed release, and repeats authentication and freshness checks on every learner read. The compound production-path box remains open because no fixture admission is a production graph transition, real evidence set, operational drill, administrator decision, or learner delivery.

The retrieval/entailment hardening increment documented in evidence/v1-v9/V9-001-RETRIEVAL-ENTAILMENT.md closes a remaining exact-evidence gap without claiming external source acquisition. Canonical textual retrieval proofs now bind the SHA-256 of the exact UTF-8 passage bytes; all current semantic-grounding producers record that binding, the shared evaluator rejects reference-preserving passage substitution, and V9 rejects a substituted approved-source passage before any provider call or derived proof. The ordinary release path still requires complete retrieval, passage-match, and semantic claim-entailment coverage. The compound source-bundle checkbox remains open until authorized production bundles and calibrated live evaluator evidence exist.

The source-authority hardening increment documented in evidence/v1-v9/V9-001-SOURCE-AUTHORITY.md closes a second repository-local laundering path without claiming that test fixtures are production sources. Every V9 generation passage now requires three distinct, current, source-bound records before any provider call: exact-byte retrieval, an affirmative source approval, and an active rights/license receipt containing the exact scope:v9-lesson-generation grant, with both authority records evaluated under the exact v9.metis.approved-source-generation.v1 policy. Missing, malformed, cross-source, future-dated, expired, revoked, wrong-policy, or insufficiently scoped authority fails closed, duplicate authority proof IDs are rejected, and the authority receipts are removed before source text is serialized into the model prompt. The compound source-bundle checkbox remains open because no production approval decisions, licenses, acquired passages, live provider output, or calibrated live entailment report were supplied.

V9-002 — Replace heuristic mastery with evidence-grounded learning state#

  • Stop treating absence of lexical misconception markers as a correct tutoring attempt.
  • Stop increasing assessment confidence after failed evidence unless a calibrated model explicitly supports it.
  • Feed graded observations into the BKT/GKT path or preserve/abstain when evidence is insufficient.
  • On missing evidence pack, preserve state or abstain; never infer correct merely because prior mastery was high.
  • Stop or explicitly skip dependent adaptive-loop stages after an upstream failure; do not record an error and continue with synthetic/default evidence that mutates mastery.
  • Define how BKT mastery, graph propagation, FSRS retention, assessment, tutor confidence, and human corrections combine without double counting.
  • Validate parameters and calibration on real learner sequences and report subgroup/misconception behavior.

Commit 533877026580d66c87fe090eda00fe684c2fdaf6 removes the two outcome-inference shortcuts from the persisted Python mastery path. Tutoring turns are now recorded as explicitly ungraded observations; neutral prose and lexical uncertainty signals cannot change mastery probability/level, confidence, attempts, correct-attempt counts, streaks, or confirmed misconceptions. Lexical signals remain reviewable only in the evidence log. Failed assessment evidence can no longer raise mastery, and confidence decreases rather than receiving the previous unconditional +0.08.

Regression tests begin from established mastery and prove exact numeric/counter preservation for uncertain and neutral tutoring turns, and non-increase for failed assessment mastery/confidence. The focused mastery suite passed 7 tests; the adaptive-learning, graph-learning, tutoring-orchestrator, and assessment calibration suites passed 10 tests; Ruff and strict MyPy passed. At that commit, the remaining V9-002 boxes stayed open because the slice did not claim BKT/GKT/FSRS unification or calibration on real learner sequences.

Commit 182dc5be10acd4e60df94d1cf62849893bfad441 closes the matching TypeScript adaptive-loop failure modes. Both the graph-aware AdaptiveKnowledgeEngine updater and the pure BKT updater consume only a successful AssessmentEvidencePack; when it is absent they return abstained: true, preserve the prior band, and do not mutate BKT beliefs, graph neighbors, or FSRS scheduling. Tests prove a later first graded observation still begins at the untouched BKT prior and that no FSRS review card was created.

The loop now has explicit stage dependencies. An evidence-build failure skips mastery, personalization, integrity, and next-item selection with typed dependency-failed outcomes, reports progression as blocked, and still runs telemetry so the failure chain is observable. Skips are counted separately from runner failures and cannot be mislabeled as a latency/backpressure event. The focused loop/updater suite passed 25 tests, the full metis-tutoring test target and owning typecheck passed, and lint completed with 0 errors and 2 pre-existing warnings.

Commit fea4eb9c0fd5f6e1470f216ccf5dd7ec4a5d8a76 closes the signal-unification and double-counting gap. The graph-aware engine now has one identified evidence ledger shared by direct BKT mastery, graph propagation, and FSRS retention: each assessment evidence ID updates all three exactly once, and retries are idempotent no-ops. Tutor confidence is retained as provenance but is not a second mastery or retention input.

A human correction names the original evidence ID, replaces its grade, and deterministically replays the ordered authoritative history. It therefore produces the same mastery graph and FSRS card as if the corrected grade had been authoritative initially, rather than appending a contradictory second learner turn. Unknown/mismatched corrections abstain without mutation. The pure-BKT fallback uses the same evidence-ID and correction-by-replay authority rule per learner/concept.

Regression coverage proves retry idempotency across mastery, graph neighbors, and FSRS; correction equivalence to an initially authoritative grade; tutor confidence non-influence; rejection of unknown correction targets; and the matching pure-BKT replay behavior. The full metis-adaptive and metis-tutoring test targets passed. Both production typechecks and the tutoring spec typecheck passed; the adaptive spec typecheck remains blocked by five pre-existing unused declarations outside this change. Owning lint passed with zero errors and only the existing 16 adaptive and 2 tutoring warnings. Real-sequence calibration remains the sole open V9-002 gate.

V9-003 — Bind tutor and agent safety/authority#

  • Fence retrieved previews, learner/course/mastery state, orchestration fields, integrity context, and source text.
  • Require successful source approval and safety/trust-zone evaluation on the tutor's invoked path.
  • Prevent safety/integrity handoffs from silently bypassing required generation or evidence transitions.
  • Complete AUTH-003 for the agent runtime and require grants/tool isolation at execution if an executor is built.
  • Add indirect-injection, learner-data, source-poisoning, academic-integrity, unsafe-advice, cross-tenant, and tool-escalation tests.

V9-003 is closed by the already-completed cross-cutting packages, reconciled here against their invoked V9 paths. PROMPT-002 fences the tutor topic/mode, learner/mastery and orchestration context, curriculum-integrity policy, retrieved source previews, conversation turns, and prior model output with the shared trust-zone renderer. SAFETY-001 requires a current persisted tenant/input/policy-bound safety screening before provider generation and returns a recorded learner-visible degradation when admission fails. Required source sessions reject a package outside the approved usage/review posture before retrieval or generation.

Safety, curriculum-policy, confidence-abstention, integrity-limited, and human handoff branches retain explicit message metadata and evidence state rather than silently presenting a bypass as ordinary generation. AUTH-003 supplies the executable DAG boundary: runtime registry bindings, grants, permissions, tenant/data scopes, human gates, tool isolation, idempotency, and transactional publication are enforced rather than copied from a plan.

The task-specific reconciliation pack is evidence/v1-v9/V9-003.md; its authoritative constituent evidence remains evidence/v1-v9/PROMPT-002.md, evidence/v1-v9/PROMPT-003.md, evidence/v1-v9/SAFETY-001.md, and evidence/v1-v9/AUTH-003.md. Current reconciliation verification passed 36 focused Python tests across trust rendering, the invoked tutoring contract, mastery abstention, safety classification, and governed DAG execution, plus 20 TypeScript adversarial-regression tests. These cover indirect/source poisoning, learner-data fencing, academic-integrity and unsafe-tutor steering, cross-tenant tool/effect denial, and undeclared authority escalation. Commit a05226b6d0e950456b9491048b5bfb8e06af0845 also aligns the invoked tutoring integration test with the V9-002 ungraded-evidence contract. Ruff lint/format passed for that repaired test; the Metis agents production typecheck continues to expose the pre-existing unrelated AgentContext.id error in core/lifecycle.ts.

V9-004 — Complete the rendered learner and teacher experience#

  • Complete Ask a Wonder, lesson player, mastery view, explorable, teacher correction, source inspection, and operator review flows.
  • Complete mobile/offline behavior, localization at scale, embodied face, V2–V8 game bridge, and creator payout paths where required by launch scope.
  • Add accessibility automation and deep Playwright/mobile E2E coverage for the critical learner and teacher flows.
  • Ensure correction invalidates stale lesson/mastery/release evidence and produces a revised path.

Repository-actionable rendered-surface closure is published in commit 3cfa8c72df0330142c54d5bccf0633733a87e1e6. Existing Metis tutoring, course player, Kalika explorable, mastery/recommendation, source-rights, and operator inspection surfaces are now joined by an explicit teacher/operator correction cascade. A request_changes or reject decision demotes the affected course from its release state, invalidates the learner lesson sequence, mastery and recommendation snapshot, and publication/release-gate evidence, and returns a typed revised_path_required record with the next revision and gate actions. The admin review workspace renders that state in place with accessible status, invalidated-evidence, and required-revised-path sections. The same change repairs the review service's repository-root resolution so canonical tutor, prompt, routing, avatar, and consent artifacts resolve from the actual workspace root.

Focused verification passed 10 service review-gate tests, 35 admin unit tests, admin lint and zero-new-error typecheck, Ruff lint, OpenAPI export/type drift, and 16 Chromium Playwright flows across tutoring, course/player/explorable, mastery, source inspection, operator inspection, correction, and four accessibility-hardening paths. The correction flow's serious-impact axe scan also exposed and closed a selected-review contrast defect. Both mobile render/model suites pass 11/11 tests, and a Maestro smoke flow covers playable, assessment, offline-download, degraded tutor, and grounded-source surfaces. The latter has not been executed in this environment because no device or installed Metis native build is attached; Firefox and WebKit browser binaries are also absent. Therefore mobile/offline/localization/embodiment/bridge/payout launch completion and the full cross-browser/mobile E2E signoff remain open rather than being inferred from repository fixtures. Strict MyPy on the two service modules retains 78 pre-existing broad str-to-Literal findings in the legacy admin service; the new correction-action finding was corrected.

The mobile execution blocker was re-audited on 2026-07-25 rather than carried forward by assumption. The original software-TCG attempt reached a true Android boot but streamed Expo Go installs destabilized the package service. Recovery execution on 2026-07-27 used a non-streamed Expo Go 57.0.2 install, ahead-of-time compiled Maestro 2.5.1 drivers, a warmed Metro runtime, and an API 35 Android 15 AVD with /dev/kvm still absent. The checked-in Expo Go recovery flow passed 1/1 in 1,312 seconds with zero failures, covering learner home/resume, Course Play and transcript, formative assessment, Downloads/offline-bundle entry, degraded Tutor/text action status, and grounded-source inspection. This closes the mobile-E2E part of the accessibility/deep-automation checkbox. It is not airplane-mode, restart-persistence, localization-scale, embodiment, game-bridge, payout, or packaged-release proof, so the separate launch-completion checkbox remains open. Exact evidence: evidence/v1-v9/V9-004-MOBILE-E2E-EXECUTION-ATTEMPT.md and evidence/v1-v9/V9-004-mobile-e2e-junit.xml.

Mobile offline persistence was advanced on 2026-07-27 with a versioned, bounded AsyncStorage queue, live NetInfo transitions, serialized writes, fail-closed snapshot validation, restart restoration, and event-identity-safe delivery retries. The learner UI now reports the honest connectivity phase, bundle revision, and pending count without claiming server sync while the authenticated transport is unconfigured. Focused automated coverage exercises offline recording, restart recovery, concurrent actions, corruption and queue bounds, transport failure, acknowledgement-checkpoint retry, and reconnection flush. A checked-in Android airplane-mode/restart flow was also added. Its live execution was infrastructure-blocked before application launch on both available API 35 software-emulated AVDs because system_server repeatedly restarted and removed Android's package/activity/settings services. The composite launch checkbox remains open for authenticated server delivery and conflicts, a stable device pass, release packaging, localization, embodiment, game bridges, and payouts. Exact evidence: evidence/v1-v9/V9-004-OFFLINE-PERSISTENCE.md.

Offline sync admission was hardened on 2026-07-27 at Metis's existing authenticated durable xAPI boundary. A statement UUID is now bound to the exact learner, authority, verb, object, result, context/cmi5 registration, and client timestamp: exact retries return the original durable statement, while changed content fails with HTTP 409. A savepoint/readback path applies the same semantic check when concurrent workers race at the unique constraint. The request contract also rejects attachments until their bytes and hashes can be durably retained instead of silently dropping them. The composite checkbox remains open because mobile auth/transport wiring, live delivery, production conflict policy, and the other launch surfaces are not present. Exact evidence: evidence/v1-v9/V9-004-OFFLINE-SYNC-INTEGRITY.md.

Offline client identity was hardened on 2026-07-27 with cryptographically secure Expo UUIDs that remain stable across persistence and delivery retries. Pending snapshots using the former process-local progress_* identifiers are migrated and durably checkpointed before network subscription or dispatch. Malformed, repeated, or case-variant-colliding identifiers fail closed without mutating the queue, while a failed migration checkpoint preserves and reports the original pending state without sending it. The composite checkbox remains open because the mobile client still lacks a Metis session/token provider and authenticated transport, live device delivery, and the other launch surfaces. Exact evidence: evidence/v1-v9/V9-004-OFFLINE-EVENT-IDENTITY.md.

Authenticated mobile sync wiring was completed on 2026-07-27. Metis Mobile now uses the service's login, verified-profile, refresh, and xAPI contracts; stores native token pairs in Expo SecureStore; coalesces refreshes; retries one rejected access token; and sends the durable event UUID as the xAPI statement UUID. Offline snapshot schema v2 binds new events to the verified learner, blocks cross-account queue mixing, and requires an explicit visible claim before ownerless v1 events can be sent. The server contract test accepts the exact mobile statement and rejects a student-supplied different learner ID. Focused rendered mobile-web QA also closed a narrow sync-rail wrap. The composite checkbox remains open because no credentialed live device/deployed-statement result is available, and packaged release, localization, embodiment, game bridges, and payouts remain unproved. Exact evidence: evidence/v1-v9/V9-004-AUTHENTICATED-MOBILE-SYNC.md.

Mobile localization infrastructure was advanced on 2026-07-27 with typed English, Spanish, and French critical-shell dictionaries, device-locale negotiation, and a durable validated learner override. Non-English chrome retains an explicit English-course-content disclosure rather than laundering untranslated lesson prose as localized. Unit, render, lint, typecheck, export, and narrow-viewport persistence checks passed, and a restart-persistence Maestro flow is checked in. The composite checkbox remains open because no reviewed translation corpus, localized generated output/media/accessibility pack, stable native execution, embodiment, game bridge, or payout proof exists. Exact evidence: evidence/v1-v9/V9-004-MOBILE-LOCALIZATION.md.

The governed mobile tutor handoff was completed on 2026-07-27. Both the avatar-request and explicit text path now create an authenticated, approved-source tutoring session before recording progress. Mobile accepts only an owner-matched response with a consistent fallback chain and approved, disclosed, consented representation; denial cannot emit xAPI progress. The surface renders the server-selected mode, disclosure, and consent instead of claiming a static voice fallback. The API also returns structured source versus representation denials for its existing fail-closed policy boundary. The composite checkbox remains open because no real avatar frames, voice socket, fidelity/lip-sync measurement, consent-withdrawal drill, stable native pass, game bridge, payout, or production source/provider proof exists. Exact evidence: evidence/v1-v9/V9-004-MOBILE-TUTOR-GOVERNANCE.md.

The V2 portion of the V2–V8 game bridge was implemented on 2026-07-27 as an off-rollback Combat Trials/Mission curriculum adapter. It composes authoritative Shakti ruleset/technique references with Metis assessment selection, adaptive difficulty, and authored tutoring hints. Adaptation abstains without calibrated assessor evidence; duplicate, expired, mismatched, or unverifiable evidence blocks the whole export. Live combat frames, rollback state, outcomes, and gameplay telemetry are rejected as learning evidence, while V2 remains the sole runtime owner. Focused typecheck, 8 contract tests, lint, and bundle passed. This does not prove a cooked V2 consumer, full V2 curriculum, or V3–V8 bridges, so the composite checkbox remains open. Exact evidence: evidence/v1-v9/V9-004-V2-COMBAT-TRAINING-BRIDGE.md.

The V3 bridge followed on 2026-07-27 with governed canonical-package admission for the Commons Metis Lecture Hall. It reads the exact released publication, lesson, and learner-bound assessment-evidence records through @metis/api-client; requires export authorization, operator approval, rights-cleared/provenance-verified sources, approved grounding, accessible and checksummed projected media, visible disclosures, and private review posture; and strips raw learner artifacts, excerpts, hashes, notes, and integrity detail from the world manifest. Render-frame and rollback calls fail before I/O. Because the canonical client exposes no notebook mutation, V3 is expressly forbidden from calling its static note refs persisted. Focused typecheck, 10 contract tests, lint, and build passed. No cooked/runtime consumer or notebook write receipt exists, and V4–V8 remain open. Exact evidence: evidence/v1-v9/V9-004-V3-METIS-LECTURE-HALL-BRIDGE.md.

The V4 bridge followed on 2026-07-27 with a shared fail-closed Metis release-admission contract and a V4-specific Training Range cook manifest. It is locked to the authoritative five-cell, 90-minute curriculum JSON; every cell must bind one exact released Metis lesson with approved rights, provenance, grounding, disclosure, accessibility, fallback, and media evidence. V4 remains the sole owner of training execution and completion, and the bridge rejects live frames, rollback, gameplay telemetry, matchmaking, AI Director, combat outcome, and competitive-ranking inputs before provider reads. This is not a cooked Unreal consumer or in-engine rendering receipt, and V5–V8 remain open. Exact evidence: evidence/v1-v9/V9-004-V4-METIS-TRAINING-BRIDGE.md.

The V5 bridge followed on 2026-07-27 for the authoritative six-bay Training Range manifest. Each verification bay must bind one exact released Metis lesson, while V5 retains sole authority over combat, vehicles, signs, AI, timing, live metrics, and completion. The bridge rejects live metrics, gameplay and trial outcomes, leaderboard data, authoritative tuning, AI Director state, and rollback inputs before provider reads. This is authoring/cook admission, not a cooked Unreal consumer; V6–V8 remain open. Exact evidence: evidence/v1-v9/V9-004-V5-METIS-TRAINING-BRIDGE.md.

The V6 bridge followed on 2026-07-27 as a player-education-only stewardship boundary bound to the two exact, fully-passing learning-by-example eval cases. Released Metis lessons can teach a steward but cannot become synthetic agent observations, invoke agent learning, or create value, priority, behavior, bond, or memory mutations. Sensitive eval learning inputs are stripped. This is not a cooked Unreal consumer; V7–V8 remain open. Exact evidence: evidence/v1-v9/V9-004-V6-METIS-STEWARD-EDUCATION-BRIDGE.md.

The V7 bridge followed on 2026-07-27 as creator education bound to five exact Sekhmet and Abundantia source-code authorities: signed supply-chain evaluation, sandbox-capability verification, clean-scan distribution gating, the reviewed publishing lifecycle, and console-authorized catalog publication. Released Metis lessons explain each proof without accepting executable or decision payloads and without authority to scan, attest, grant capabilities, moderate, publish, authorize consoles, govern, or pay creators. This is not a rendered V7 client or live publication; V8 remains open. Exact evidence: evidence/v1-v9/V9-004-V7-METIS-CREATOR-EDUCATION-BRIDGE.md.

The V8 bridge followed on 2026-07-27 as player learning behind Clew's authoritative minimal proof-tree selector and the exact all-eight-gate ReleaseDecision. It combines one already-released solve-path clue with one exact released Metis reasoning lesson, while stripping ground truth, culprit, conclusion, deduction path, unreleased clues, release evidence, and player state. It has no evidence, interrogation, accusation, case generation/repair, release, canon, director, or player-model authority. This is not a rendered V8 client, but it closes the repository-actionable canonical-provider seam for V2–V8. Exact evidence: evidence/v1-v9/V9-004-V8-METIS-DETECTIVE-LEARNING-BRIDGE.md.

Cross-browser execution was extended on 2026-07-27 with Playwright 1.60.0, Firefox 150.0.2, and WebKit 26.4. The complete learner project passed 36/36 in Firefox with one worker and zero retries. WebKit passed 35/36 with no failures; the authenticated service-worker warm-cache case is explicitly skipped because the Playwright WebKit runtime drops the HttpOnly session from controlled requests and redirects every protected shell to login. The admin correction and governance decision flow separately passed 1/1 in both Firefox and WebKit with zero retries. The run closed real cross-engine navigation races in the assessment, authentication, and connected learner-surface tests; deferred PWA warming until auth resolution; and aligned /study and /saved with both the warm and authenticated-navigation cache policies. The Firefox warm-cache flow proves all seven canonical learner routes. The WebKit runner limitation remains explicit. Together with the subsequently passed Android Maestro learner flow, the cross-browser suites close the V9-004 accessibility/deep-automation checkbox, while the broader mobile/offline launch-completion checkbox remains open. Exact evidence: evidence/v1-v9/V9-004-CROSS-BROWSER-E2E.md and evidence/v1-v9/V9-004-MOBILE-E2E-EXECUTION-ATTEMPT.md.

V9-005 — Produce the V9 flagship evidence slice#

  • Enroll a real consented learner cohort and independent evaluator.
  • Have learners ask wonders, complete adaptive lessons/explorables, and receive evidence-grounded revised paths.
  • Measure immediate learning, misconception repair, transfer, delayed mastery, week-four retention, usability, trust, and teacher/learner preference against a controlled baseline.
  • Exercise teacher/operator inspection, correction, takeover, withdrawal, and rollback.
  • Run closed beta, mobile/offline, localization, cost, safety, privacy, scale, canary, and incident gates before efficacy claims.

Flagship admission boundary (2026-07-27): scripts/v9/verify-flagship-evidence.mjs now composes approved powered cohort governance, exact-candidate learner journeys, four-week controlled outcomes, production intervention exercises, and all nine closed-beta release gates. The retained package is blocked on every check and deliberately leaves sample thresholds unset until an approved power analysis exists. See evidence/v1-v9/V9-005-FLAGSHIP-READINESS.md.

10. P1 completion packages not owned by one product#

P1-001 — Durable side effects and recovery certification#

  • Certify RUN-002, RUN-003, RUN-005, HITL-003, and OPS-002 across every consequential pipeline.
  • Require 100% pass on checkpoint/restart, duplicate-delivery, stale-approval, provider-timeout, kill-switch, and publication-rollback drills with lineage preserved and no repeated effects.

Certification infrastructure is published in commit 249779de5035201d35ebc380cf2e49496da74d05 and documented in evidence/v1-v9/P1-001.md. The shared run protocol now requires a closed-world authoritative pipeline manifest, independent complete runtime observation, and fresh staging/production evidence for all six drills on every exact pipeline revision and source commit. Unknown or unobserved pipelines, missing or stale drills, hermetic-only evidence, revision/commit mismatch, broken lineage, missing receipts, and any repeated effect fail closed. The contract passes 12/12 focused adversarial cases and the complete targeted run-protocol gate passes 68/68 tests across 6/6 files. Both package checkboxes remain open because the repository has no owner-reviewed complete deployed-pipeline inventory or fresh deployed six-drill receipt matrix, and OPS-002 still lacks production whole-host restore/provider-reconciliation evidence.

P1-002 — Cross-product canon and asset lineage#

  • Implement the shared entity/source/claim/character/world/asset graph from COV-003.
  • Provide bounded product/tenant views, authorization, invalidation, provenance, and audit queries.
  • Prove a source/canon/right change reaches every dependent V1–V9 artifact without leaking unrelated tenant data.

Closed in source commits 278cea79a362bc9d2b8ee9d35b20103093262ff8 and eae00e927ee2ba25f3f6c8e8528867b92ee9ef94, with evidence in evidence/v1-v9/P1-002.md. The COV-003 graph is now held behind a durable, tenant-authorized BFF authority that restores before serving queries, persists graph and audit mutations before publication, supplies redacted product views and product-bounded provenance, and plans audited invalidation. Six focused service cases prove restart recovery, cross-tenant read/write denial, bounded provenance, optimistic revisions, idempotent audit identities, fail-closed persistence, and source/canon/rights propagation to all nine V1–V9 shipped dependents. This closes the repository implementation and deterministic cross-product proof; it does not claim that production tenants have populated live graph records or executed artifact withdrawals.

P1-003 — Provider redundancy and quality-aware routing#

  • Define at least one tested failover strategy for every critical provider class.
  • Route by calibrated quality, capability, cost, latency, locale, rights, safety, and regional availability rather than static preference alone.
  • Preserve model/provider identity and degraded-mode labeling in artifacts and user experience.
  • Run provider outage, quality regression, quota, billing, and regional failover drills.

Closed in source commit 266a0ae959b713e87d9a833d41c4b80d8d2c5f4c, with evidence in evidence/v1-v9/P1-003.md. A shared fail-closed certification contract now covers ten critical provider classes, requires independent provider and failure-domain redundancy, filters on calibrated/fresh quality, capability, cost, latency, locale, rights, safety, quota, billing, and regional availability, and preserves endpoint/provider/model identity in every decision. Every incident route carries a mandatory degraded-mode customer label. The focused gate passes 21 adversarial cases, including a 20-provider canonical portfolio and all 50 required class/scenario drills. This closes the executable repository contract and hermetic drill matrix; it does not claim deployed provider inventory, live traffic failover, or provider-issued operational receipts.

P1-004 — Continuous portfolio health governance#

  • Review incidents, overrides, rights expirations, subgroup outcomes, quality distributions, cost anomalies, and recertification backlog on a defined cadence.
  • Assign owners and deadlines to every blocked/stale slot.
  • Publish model, rubric, prompt, policy, provider, and data cards with known limits and last calibration date.
  • Prevent “proposed” consolidation architecture from being reported as implemented until conformance evidence exists.

Governance infrastructure is published in commit 5c942b51a2b5bc1f949ba73d7be2782b70f83d54 and documented in evidence/v1-v9/P1-004.md. The fail-closed certification contract requires current reviews for all seven agenda axes, owner/deadline accountability for every blocked or stale item, published current cards for all six required card families, and conformance evidence plus verification for any architecture reported as implemented. It passes 24/24 focused adversarial cases. The first three boxes remain open because no authoritative current portfolio review set, complete blocked/stale work inventory, or published six-family card catalog was available. The fourth box is closed because proposed-as-implemented reporting now deterministically blocks certification, while implemented claims without conformance evidence or verification also fail.

11. P2 — Frontier improvements after P0/P1 control closure#

P2-001 — Dynamic specialist-agent teams#

  • Define independent breadth-heavy tasks appropriate for parallel workers: research, ideation, specialist critique, localization, and adversarial evaluation.
  • Keep solver-coupled case construction, shared world mutation, publication transactions, and safety decisions deterministic or strongly serialized.
  • Run controlled comparisons against simpler workflows and require measured quality/latency/cost gain before adoption.
  • Cap worker count, tokens, tools, recursion, and wall time; preserve one accountable coordinator and common evidence trace.

Governance infrastructure is published in commit d49f50c1317b133c6c4b3ddae25c3f73eae30192 and documented in evidence/v1-v9/P2-001.md. The shared adoption gate enumerates the five parallel-safe task classes, forces all four coupled/mutating/transactional/safety classes and every shared-state writer to serialize, and requires hard worker/token/tool/recursion/wall-time caps plus one coordinator and evidence trace. Fifteen focused cases pass. Adoption also requires a calibrated minimum-30-sample comparison with strict quality, p95-latency, and mean-cost improvement over the simple baseline. That comparison box remains open because no authoritative measured product experiment was available; the passing fixture proves the gate, not the gain.

P2-002 — Learned quality/cost/latency routing#

  • Train only on attributed, calibrated production outcomes.
  • Retain hard exploration floors, safety constraints, provider diversity, and rollback.
  • Evaluate offline, shadow, and canary performance against deterministic routing with subgroup and tail analysis.

Adoption infrastructure is published in commit ea922ae92ca78b377776ea064ebfe8d63e8f54a7 and documented in evidence/v1-v9/P2-002.md. The strict outcome schema accepts only attributed production observations with provider/model/route identity and calibration evidence. Policy validation requires a nonzero exploration floor, at least two providers, hard safety references, deterministic rollback artifact/trigger, and baseline identity. The gate requires one passing offline, shadow, and canary comparison with baseline binding, at least two subgroups, and p95/p99 tail analysis. Eight focused cases pass. Training and evaluation boxes remain open because no authorized calibrated production outcome corpus or executed staged experiment was available; fixtures prove rejection and adoption semantics only.

P2-003 — Hierarchical taste models#

  • Model product, tenant, locale, expert cohort, and user preferences without collapsing minority taste into a global majority.
  • Use privacy-preserving aggregation, minimum-data thresholds, uncertainty, opt-out/deletion, and anti-manipulation controls.
  • Keep safety, rights, factuality, and accessibility outside taste optimization.

Closed in source commit 64d87fdf18c2a464f0c7486e27564fa392f52c70, with evidence in evidence/v1-v9/P2-003.md. The contract preserves the exact product→tenant→locale→expert-cohort→user hierarchy and validates immediate parentage instead of flattening preferences. Differential-privacy bounds, per-subject contribution caps, aggregate/user sample floors, uncertainty fallback, opt-out, deletion, and manipulation detection are mandatory. Insufficient, uncertain, opted-out, deleted, or manipulated nodes fall back to their broader parent rather than being optimized. Safety, rights, factuality, and accessibility remain separately referenced hard constraints. Fifteen focused adversarial cases pass.

P2-004 — World-model, RL, and embodied generation#

  • Begin only after V7/V9 real simulation, outcome, safety, and cost harnesses exist.
  • Define environment ground truth, reward decomposition, off-policy evaluation, reward-hacking tests, containment, human oversight, and rollback before training.
  • Compare against procedural/deterministic and authored baselines.

Pre-training governance is published in commit 5d87475436b9bf35b489dc62fc018684b1a8e4e5 and documented in evidence/v1-v9/P2-004.md. Training fails closed unless V7 and V9 each provide one real passing staging/production simulation, outcome, safety, and cost harness receipt. Plans require environment ground truth, normalized decomposed rewards, off-policy evaluation, reward-hacking tests, containment, human oversight, and rollback. Adoption separately requires passing, safety-clean procedural, deterministic, and authored comparisons. Fourteen focused cases pass. The comparison box remains open because no authorized executed experiment was available; no world-model/RL training or adoption is claimed.

P2-005 — Distilled specialist judges and generators#

  • Establish fidelity to a calibrated reference system before distillation.
  • Measure capability, calibration, bias, uncertainty, safety, cost, latency, and subgroup regression.
  • Route out-of-distribution/low-confidence cases back to the reference system or humans.

Adoption infrastructure is published in commit ad0bd4308f9c5b84e16f347cf0793b175ae778af and documented in evidence/v1-v9/P2-005.md. The gate binds a distilled candidate to a calibrated reference and rejects fidelity, calibration, bias, uncertainty, safety, or subgroup regression while requiring capability, cost, and p95-latency measurements. An explicit OOD detector, confidence floor, and unique ordered reference-system/human fallbacks are mandatory; OOD or low-confidence cases cannot stay on the distilled specialist. Eleven focused cases pass. Fidelity and measurement boxes remain open because no actual distilled model or calibrated evaluation corpus was available; fixtures prove the gate only.

Final repository actionability audit#

The 2026-08-02 audit in evidence/v1-v9/REMAINING-ACTIONABILITY-AUDIT.md re-read every unchecked entry against repository tip 139c19d99a2d96237cee1d0bc1a759ac3df7becf and the qualified host. It retains the V4, V7, and V5 rows closed by authenticated UE execution and supersedes the old V2 and V3 missing-host observations with exact project-build preflights. Of 145 remaining unchecked entries, 8 are standing evidence-pack requirements, 125 across 36 task packages require external authority or state, and 12 are dependent portfolio completion gates. No remaining checkbox can be truthfully completed by local implementation or fixtures alone. The row-level evidence/v1-v9/remaining-actionability-manifest.json binds every exact requirement to its blocker reason, required external evidence, required authority, and hashed supporting pack. scripts/verify-remaining-actionability-audit.mjs and scripts/audit/remaining-actionability-manifest.mjs fail closed on ledger, classification, profile, or supporting-evidence drift so any change forces a new manual review.

12. Source-to-task completeness crosswalk#

12.1 Canonical P0 gap register#

Audit gap Closing task families
P0-01 canonical coverage contract COV-001 through COV-005, V1-001
P0-02 common durable run protocol RUN-001 through RUN-005
P0-03 human takeover/co-authoring HITL-001 through HITL-005
P0-04 no outcome proof for “amazing” EVAL-002 through EVAL-005, all V*-005/flagship packages
P0-05 provider/render/cook gaps V2-001 through V9-005 product packages
P0-06 capability truth drift AUD-001 through AUD-008, COV-002, COV-005
P0-07 missing human-gold data EVAL-002, V5-002, V8-004, V9-005
P0-08 portfolio tool authority AUTH-001 through AUTH-003
P0-09 unenforced runtime controls AUTH-002, AUTH-003, COST-001, COST-002, RUN-003, RUN-005
P0-10 inconsistent prompt trust PROMPT-001 through PROMPT-003, SAFETY-001
P0-11 proof conflation PROOF-001 through PROOF-003, MEDIA-001, REL-001

12.2 Canonical P1 and P2 gap register#

Audit gap Closing task families
P1-01 unified evaluation harness EVAL-001
P1-02 judge calibration/anti-overfitting EVAL-003
P1-03 portfolio learning loop LEARN-001
P1-04 safe prompt/policy evolution LEARN-002
P1-05 cross-product canon/asset lineage COV-003, P1-002
P1-06 multimodal experiential evaluation MEDIA-001, EVAL-005
P1-07 cross-version observability OBS-001, PRIV-001, OPS-003
P1-08 durable side effects/recovery RUN-002 through RUN-005, OPS-001, OPS-002, P1-001
P1-09 risk-tiered autonomy REL-002
P1-10 continuous recertification REL-003, P1-004
P2 dynamic teams P2-001
P2 learned routers P2-002
P2 hierarchical taste P2-003
P2 world-model/RL/embodied generation P2-004
P2 distilled specialists P2-005

12.3 Validated adversarial-review deltas#

Follow-up finding Ledger disposition
V3 29/40 is wrong AUD-001 uses 30/41 required and 24/39 named
V3 evidence is stale AUD-002 adds evidence freshness without claiming an unrecorded rerun never happened
Host-ambiguous UE advice AUD-004 and V4-001 name the Linux UE5.5 host/path
Verification commands/HEAD absent AUD-003
No in-repo V1 operator UI consumer HITL-001; wording remains repository-scoped until deployment evidence exists
Live spend not governed COST-001/COST-002 preserve existing partial cycle-token evidence and close missing full monetary accounting
runs-route comment overclaims AUD-005 and AUTH-002
Hathor memoriesTriggered fabricated SAFETY-001
Coverage registry can be seeded mechanically COV-002
Declared V1 single-box failure domain OPS-001/OPS-002, subject to live topology probe
Locale coverage absent COV-004 and product slices
Generated-artifact accessibility unclear COV-004, REL-001 R2/R6, product slices
V1 snapshot retention/access unclear PRIV-001
Test coherence not sampled AUD-007
V8 drift is a mechanical current blocker V8-001, with generated semantic diff review and dependent tests
Follow-up remained separate from canonical report AUD-008

12.4 Prompt-boundary coverage#

Boundary Closing tasks
V1 ideation, planner, writer, critique/revise, CQJ memory PROMPT-001 through PROMPT-003, PROOF-002
V1 media/video MEDIA-001, PROOF-003
V2 live/post-match commentary PROMPT-002, PROMPT-003, V2-001
V2 Hathor PROMPT-002, SAFETY-001, V2-001
V3/V4/V5 provider seams product packages; remain open until real provider output exists
V6 cognition/Clio PROMPT-002, PRIV-001, V6-001
V7 Forge/Mawu PROMPT-002, AUTH-001, SAFETY-001, V7-001
V8 proposer/assets PROMPT-002, MEDIA-001, V8-003, V8-004
V9 content/tutor PROMPT-002, PROOF-002, SAFETY-001, V9-001 through V9-003
V9 planner/runtime AUTH-003, V9-003

12.5 Product acceptance-slice coverage#

Product Required package
V1 V1-001 through V1-003
V2 V2-001 and V2-002
V3 V3-001 and V3-002
V4 V4-001 and V4-002
V5 V5-001 through V5-003
V6 V6-001 through V6-003
V7 V7-001 through V7-003
V8 V8-001 through V8-005
V9 V9-001 through V9-005

12.6 SOTA comparison and operating-program coverage#

Audit consideration Closing task families
Fixed workflows versus dynamic agents and environmental ground truth RUN-001 through RUN-005, REL-001, P2-001
Multi-agent breadth, coordination cost, and serialization boundaries P2-001
Iterative improvement, search, and safe reflective optimization LEARN-001, LEARN-002
Repeated-trial evaluation, grader diversity, pass@k/pass^k, and judge bias EVAL-001 through EVAL-004
Multimodal and experiential quality MEDIA-001, EVAL-005, product slices
Durable human interrupts and idempotent resume HITL-001 through HITL-005, RUN-002, RUN-003
Cross-version tracing and privacy OBS-001, PRIV-001, OPS-003
Security, authority, and provenance AUTH-001 through AUTH-003, PROOF-001 through PROOF-003
Target coverage graph and common run protocol COV-001 through COV-005, RUN-001 through RUN-005
Source policy and five-level autonomy ladder REL-002
Three independent gate classes and launch thresholds REL-001, EVAL-004, section 13 completion gates
Gold-set design and learning loop EVAL-002, LEARN-001, LEARN-002
Phase A portfolio truth AUD, COV, and RUN task families
Phase B composed human authority HITL, AUTH, and PROOF task families
Phase C one real vertical per version all V1–V9 product packages
Phase D calibrated automated improvement EVAL and LEARN task families
Phase E scaled coverage and recertification COV-005, P1-003, P1-004, REL-003
Portfolio R0–R8 release gates REL-001
Coverage, human-work, quality, and operations metrics OPS-003

13. Portfolio completion gates for this ledger#

The ledger is not complete until all of the following are proved concurrently:

  • Every task above is checked with a current evidence pack; no provider, data, engine, UI, human, cohort, deployment, or operational seam is counted as filled by a contract or fixture.
  • 100% of required V1–V9 slots are registered and either current/released or carry an explicit owned blocker; zero shipped artifacts are unknown.
  • All hard safety, rights, consent, provenance, formal validity, accessibility, runtime, and authority gates pass with typed evidence.
  • Critical deterministic tasks achieve pass^5 at or above 0.95 under the predeclared reliability protocol; weaker results lower autonomy or block.
  • Each flagship slice is non-inferior to its best-authored baseline with the lower 95% preference bound above 0.50; any “amazing” claim additionally sustains win rate above 0.60 with lower bound above 0.55.
  • Every released artifact traces to slot, run, steps, attempts, tools, sources, interventions, gates, model/provider versions, measured spend, publication, and placement.
  • Human operators can observe, pause, edit, claim, replace, approve, veto, resume after restart, roll back, and withdraw through the deployed control plane without bypassing policy or repeating effects.
  • Restart, concurrent-worker, duplicate-delivery, stale-approval, provider-timeout, grant-revocation, kill-switch, publication rollback, and whole-host restore drills pass with 100% lineage preservation and no duplicate side effects.
  • Judge panels are calibrated to licensed human gold; high disagreement and out-of-distribution cases abstain or route to humans.
  • Source, canon, model, prompt, rubric, policy, rights, consent, provider, code, engine, and placement changes trigger selective invalidation and recertification.
  • Production canaries demonstrate target quality, safety, reliability, accessibility, cost, subgroup, and outcome SLOs for all nine flagship slices, with tested automatic or operator rollback.
  • The capability registry, generated documentation, audit evidence, and deployed observations agree at one identified commit and freshness window.

Only after these gates hold may the portfolio claim complete, production- observed, human-governed SOTA autonomous content coverage. Until then, status must remain scoped to the specific components and vertical slices actually proved.

Concurrent admission boundary (2026-07-27): scripts/audit/verify-portfolio-completion-evidence.mjs now requires all 12 completion claims in one fresh, non-fixture staging/production package at one source commit. Each gate has claim-specific invariants and an immutable evidence artifact; a generic pass cannot satisfy it. The retained package is blocked on all 12 checks, so every checkbox above remains open. See evidence/v1-v9/PORTFOLIO-COMPLETION.md.