# V1 Domain Workbenches — audit & inventory toolkit

Deterministic evidence for the foundational (BASE / DEP) sections of
[`V1_DOMAIN_WORKBENCHES_TODOS_2026-07-23.md`](../../V1_DOMAIN_WORKBENCHES_TODOS_2026-07-23.md),
covering the seven initiatives (Isis, Metis, Yemaya, Veritas, Euterpe, Aja,
Bellona) plus the shared workbench platform.

## One command

```bash
pnpm inventory:v1-workbenches          # regenerate every artifact below
pnpm verify:inventory:v1-workbenches   # drift-check all + capability gate + full node --test suite
```

CI runs the verify on every PR touching the tooling/evidence and on push to main
(`.github/workflows/v1-workbenches-inventory.yml`).

## Artifacts

| Ledger          | Generator (`scripts/v1-workbenches/`)        | Output                                                                        | What it proves                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |
| --------------- | -------------------------------------------- | ----------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **BASE.1**      | `generate-inventory.mjs`                     | `inventory/{manifest,packages,applications,routes,tests,reconciliation}.json` | 473 projects, 48 apps, 125 framework-aware routes, 4.8k tests; 0 unexplained records; Studio-Isis fail-closed proof                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| **BASE.3**      | `generate-library-graph.mjs`                 | `inventory/library-graph.json`, `inventory/library-public-symbols.json`       | inbound consumers, outbound (cross-domain + deep) deps, 96k public symbols, reuse/retire/refactor disposition proposals                                                                                                                                                                                                                                                                                                                                                                                                                                    |
| **BASE.2.a**    | — (design doc)                               | `surface-classification-schema.md`                                            | 6 mutually-exclusive surface classes + sub-state fields + repo-grounded examples                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
| **DEP.1**       | `generate-dependency-graph.mjs`              | `dependency-graph.json`                                                       | §4 sequencing + §5 integrations as a validated DAG; cycle detector + unknown-ID validator + critical path                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| **BASE.9**      | `generate-capability-inventory.mjs`          | `capability-inventory{,.manifest,.documented-sample}.json`                    | 6,414 atomic capabilities across surface/operation/api-operation/event/command/job/entity kinds, with provenance                                                                                                                                                                                                                                                                                                                                                                                                                                           |
| **BASE.13**     | `capability-diff-gate.mjs`                   | — (gate)                                                                      | fails on any undispositioned added/removed/changed capability, naming the exact source                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
| **BASE.5.a/b**  | `generate-registry-reconciliation.mjs`       | `registry-reconciliation.json`                                                | domains.json ↔ ledger §1 ↔ DOMAINS.md assertion table + field diff; surfaces the Euterpe ratification + project-count contradictions                                                                                                                                                                                                                                                                                                                                                                                                                       |
| **BASE.5.e**    | `generate-studio-boundary-matrix.mjs`        | `studio-boundary-matrix.json`                                                 | Studio/Yemaya boundary: 31 AAA-only segments doc↔code agree, fail-closed allow-list, no contemplative regression                                                                                                                                                                                                                                                                                                                                                                                                                                           |
| **S0.1**        | `generate-tara-import-graph.mjs`             | `tara-import-graph.json`                                                      | Tara (`@oshun/tara-workbench`) export→consumer graph: 209 exports, 7 production consumers (all BFF), test edges segregated, 151 unused exports / 0 deep imports / 0 cycles surfaced                                                                                                                                                                                                                                                                                                                                                                        |
| **S0.2.a/b**    | `generate-tara-component-inventory.mjs`      | `tara-component-inventory.json`                                               | Tara web component extraction matrix: 29 components (28 owned + shared shell) role-tagged and route-linked, with props/state/hooks/reads/mutations/tokens/tests columns; 24 read + 27 mutation endpoints; 25 components lack a co-located test                                                                                                                                                                                                                                                                                                             |
| **S0.3**        | `generate-tara-bff-matrix.mjs`               | `tara-bff-matrix.json`                                                        | Tara BFF route-to-capability matrix: 70 routes (all guarded + fail-closed) with method/path/middleware, per-handler store+domain+audit+external-effect deps, boundary controls (auth/validation/idempotency/concurrency), durable store + fail-closed provider classification, and route→capability + test coverage (11 untested, 7 auth-only mutations surfaced)                                                                                                                                                                                          |
| **S0.8.a**      | `generate-tara-mutation-effect-matrix.mjs`   | `tara-mutation-effect-matrix.json`                                            | Mutation-effect matrix: all 42 Tara BFF mutations classified across 8 effect types (persistence-write 32, event/audit 34, provider 4, publish 3, blob 2, notification 2, job 0). 8 cross an external boundary (the S0.8 transaction-boundary candidates); every write mutation audits                                                                                                                                                                                                                                                                      |
| **S0.4.a/f**    | `generate-tara-studio-authoring-overlap.mjs` | `tara-studio-authoring-overlap.json`                                          | Tara ↔ @oshun/studio-authoring overlap: studio-authoring inventoried (315 exports across 9 categories, 0 UI components/hooks/stores — a pure logic library), 0 duplicate exported contract names with Tara, and an acyclic dependency (Tara → studio-authoring)                                                                                                                                                                                                                                                                                            |
| **S0.5.a–d**    | `generate-tara-contract-inventory.mjs`       | `tara-contract-inventory.json`                                                | Tara shared-contract inventory: 98 Zod schemas (44 enums) + 98 inferred types + 22 persistence objectContract rows, each row's source-to-Prisma lineage (0 source-less rows), a categorized consumer graph (14 files: bff/domain-lib/web), and the contracts↔domain-lib direction (contracts is the lower layer)                                                                                                                                                                                                                                           |
| **S0.11.a–c/f** | `deletion-threshold-gate.mjs`                | `deletion-threshold.json`                                                     | Deletion-threshold gate: real-consumer definition + per-export consumer report + >=2-consumer eligibility over the import graph. 0/209 Tara exports clear the bar today (56 single-, 153 zero-consumer) — none extractable as-is. Enforces (exit 1) once a shared kit is added to SHARED_KIT_PACKAGES                                                                                                                                                                                                                                                      |
| **S0.8.b/c**    | `generate-tara-transaction-boundaries.mjs`   | `tara-transaction-boundaries.json`                                            | Transaction-boundary decision table derived from the S0.8.a mutation-effect matrix via the standard outbox rule: 34 local ACID transactions (persistence+audit only) + 8 transactional-outbox mutations (external provider/publish/blob/notification), 0 sagas; plus the command/event/correlation identity contract                                                                                                                                                                                                                                       |
| **S0.9.a**      | `generate-design-system-comparison.mjs`      | `design-system-comparison.json`                                               | Design-system comparison of all 5 systems: token owners (oshun 600 / yemaya 67 defs) vs a token consumer that defines none locally (isis, 87 uses) vs non-token (euterpe inline styles, bellona minimal). Proves the 5 do not share a styling approach — a shared token contract can't be adopted uniformly                                                                                                                                                                                                                                                |
| **S0.9.b**      | `generate-ui-composition-boundary.mjs`       | `ui-composition-boundary.json`                                                | UI-composition ownership boundary derived from the S0.6 layering + S0.9.a divergence: 5 kit-owned concerns (neutral shell/layout/interaction primitives + the token CONTRACT + the theme-injection seam) vs 5 host/domain-owned (vocabulary, IA, composition, specialist visualization, and the token VALUES — which S0.9.a proves diverge and cannot be shared). Contract-vs-values is the forcing rule                                                                                                                                                   |
| **STD.8.a**     | `generate-workbench-api-register.mjs`        | `workbench-api-register.json`                                                 | Operation-level register of every workbench API across transports: 1,053 operations — 755 HTTP (640 administrative `/v1/admin/{isis,aja,yemaya,bellona,studio}`, 115 public studio-authoring incl. 70 Tara + Metis/Veritas/Aja adapters) + 298 bridge (294 Yemaya Electron IPC incl. the Bellona bridge, 4 Euterpe Tauri commands). streaming/internal/webhook honestly 0 (no SSE producers; webhook-config APIs are administrative, not receivers); Hathor excluded as a non-V1-phase. This is the foundation STD.8.b-h attach OWASP API risk controls to |
| **S0.10.a**     | `generate-tara-migration-surface.mjs`        | `tara-migration-surface.json`                                                 | Migration surface register: consolidates the six inventories into the 8 S0.10.a dimensions (public imports/routes/state stores/contracts/persisted data/events/generated artifacts/consumers), each with source provenance, plus pre-migration risk flags (151 unused exports, 11 untested routes, 25 untested UI components)                                                                                                                                                                                                                              |

| **Y0.6.a/b · V0.2.a · E0.3.a · A0.3.a · B0.3.a** |
`generate-domain-package-inventories.mjs` | `domain-package-inventories.json` |
Consolidated package inventory + capability catalog for the 5 remaining domains
(254 pkgs): yemaya 59 / veritas 64 / euterpe 49 / aja 41 / bellona 41, each
package with exports/consumers/deps/disposition/tests + capabilities linked to
exact source locations. Covers each domain's `.a` enumerate + yemaya's `.b`
extract | | **B0.2.a-c** | `generate-bellona-inventory.mjs` |
`bellona-inventory.json` | Bellona DCC-bridge console inventory: 4-route control
room + 82 /studio/bellona/_ pages (a), 10 runtime apps classified by transport
(4 DCC bridges = WebSocket, remote/build/render; websocket/grpc/http/ipc) (b), 6
CLI commands + 41 libs (c). The bridges use WebSocket/IPC, not HTTP routes | |
**A6.1.a** | `generate-aja-storage-inventory.mjs` | `aja-storage-inventory.json`
| Aja storage map (A6.1.a): of 7 storage kinds, 4 present
(cache/in-memory/file/blob), 3 absent (table/index/event). **Finding**: Aja has
**0 relational tables** — a compute domain with no relational persistence
(unlike Isis/Veritas/Yemaya). Storage is the 6-module `pipeline-cache` subsystem
(pose-estimation-cache/retargeting-cache/lifting-cache/…) + in-memory Map/LRU +
motion file I/O (fbx/bvh/npz). Completes the cross-domain storage comparison:
relational (Isis/Veritas/Yemaya) / desktop-file (Euterpe) / compute-cache (Aja)
| | **A0.14.a** | `generate-aja-code-census.mjs` | `aja-code-census.json` | Aja
launch code census (reconciliation inputs): 41 libs packages + 7 app projects,
445 code files + 43 tests, 98 Studio routes, 3 classified services, 7 CLI
commands, 7 formats, clean @oshun/embodiment-aja boundary, 52 docs, 7 deploy
targets — consolidated from fresh git counts + committed A0.2/A0.4/A0.7 evidence
| | **I2.4.a–f** | `generate-workflow-selection.mjs` |
`workflow-selection.json` + `workflow-selection/*.ts` | Isis workflow-template
selection — a COMPILING reference. 4-stage chain:
capability/IO/profile/target/dependency query (a) →
tier/host/entitlement/tenant-policy/classification exposure (b) →
lifecycle+provider+resource+budget+rights+safety eligibility (c) → deterministic
ranking policy→compat→quality→latency→cost→capacity with a **stable id/version
tie-break** (d). 19 deliberate rejections cover every code; a deprecated
workflow inside its window stays eligible. **(e)** a curated tier gets a
`CandidateSummary` type that **structurally lacks** graph/nodes/model/provider —
a leak is a TYPE ERROR — plus a serialized-view scan for internal identifiers.
**(f)** an explicit choice is honored only if it survives every filter, so a
directly-supplied forbidden id is REJECTED; dispatch RE-RUNS the chain, so a
workflow revoked between planning and dispatch is refused (race exercised).
Operator tier still allowed — tier-based, not blanket. I2.4.g (emit + audit,
end-to-end) runtime → parent unchecked. Compile + selection suite gated; 11
tests | | | **I2.3 (parent + a–i)** | `generate-reference-validation.mjs` |
`reference-validation.json` + `reference-validation/*.ts` | Isis reference-media
validation — a COMPILING reference. 8 checks over an IMMUTABLE stored reference:
the schema has **no `url` field**, so a mutable URL-only dependency isn't
expressible (a); the **sniffed** kind is authoritative over the declared one
(b); rights/consent incl. freshness + revocation (c); the checksum check
**recomputes the hash from the stored bytes** so a spoofed manifest hash — or
one flipped byte — is quarantined (d); dimension (e), timing (f),
codec/bit-depth (g) limits; every required scan must exist, PASS, be fresh, and
be **bound to the same hash** (h). **PARENT BAR MET (i)**: the provider is
reachable only via a **call-counted seam**, and 17 deliberate bad references are
each asserted alone AND mixed with a valid one = **34 zero-provider-call
assertions**; refusals name the reference, the check, and one of 5 issue kinds.
Compile + validation suite gated; 11 tests | | | **I2.2 (parent + a–h)** |
`generate-typed-controls.mjs` | `typed-controls.json` + `typed-controls/*.ts` |
Isis typed authoring controls + **hidden-control lint** — a COMPILING reference.
7 control blocks: positive intent w/ stable vocabulary + namespaced domain
blocks (a); exclusions carrying category/scope/**severity/source** so a
code-floor exclusion is distinguishable from a user preference (b); ordered
reference controls w/
version/hash/region-mask/influence-schedule/preprocessing/rights/consent +
explicit conflict behavior (c); typed spatial (d), temporal (e), audio (f),
delivery (g). **PARENT BAR MET**: prose lives only in a `notes` field declared
NON-AUTHORITATIVE, and the lint (h) — 9 detectors × all 5 surfaces
(compiler/ui/contract/import/export, incl. notes itself) — reports any control
expressed in prose, distinguishing a genuinely **HIDDEN** control (typed field
EMPTY → lint FAILS) from a harmless **DUPLICATE** (typed → reported only),
quoting the offending text. 7 deliberate hidden cases
(aspect/resolution/fps/duration/LUFS/exclusions/references) each caught with the
bypassed control named. Compile + controls suite gated; 11 tests | | | **I2.1
(parent + a–g)** | `generate-prompt-spec-compiler.mjs` |
`prompt-spec-compiler.json` + `prompt-spec/*.ts` | Isis deterministic
brief→PromptSpec compiler — a COMPILING reference. Versioned spec w/
system/policy SEPARATED from user intent and **no free-form prompt/instruction
field**, so untrusted subject/reference text can only land in named DATA fields
(`fenced: true` on every reference, a/e). Input validated + canonicalized before
mapping (b). Only approved versioned templates/vocabularies/mappings resolve,
with EVERY decision traced by source —
brief/default/derived/omitted/unsupported, so a default is visible AS a default
(c). Serialization host-independent: recursively sorted keys, NFC, **no
Date/Math.random/localeCompare** (d). **Tier restriction (e)**:
workflow/model/provider detail only for an authorized tier — and since a trace
entry is part of the spec, the **trace itself redacts** restricted identities.
Verified per-golden: **5/6 goldens are unauthorized-tier and carry ZERO
restricted identities**, the 1 authorized golden legitimately does. Failures
return 3 structured error kinds over 10 negative briefs and **never degrade to
prompt concatenation** (f). 6 canonical briefs compile to FROZEN goldens and
recompile byte-identically (g) = the parent's bar. Compile + compiler suite
gated; 12 tests | | | **I1.15 (parent + a–l)** |
`generate-asset-contract-fixtures.mjs` | `asset-contract-fixtures.json` +
`asset-contracts/*.ts` | Isis canonical asset contracts — a COMPILING reference.
**11 asset kinds**
(image/portrait/texture/video/audio/foley/voice/mesh/rig/splat/scene, a–k) as
ONE discriminated union with **branded units**
(Pixels/Seconds/Hertz/LUFS/Degrees/Millimeters/Ratio01) + typed refs
(AssetRef/EvidenceRef) — a raw number can never occupy a unit position. **5
fixture classes per kind = 55 fixtures**
(minimal/full/boundary/invalid/version-skew). **PARENT BAR MET**: every valid
fixture parses through the PUBLIC contract and ROUND-TRIPS byte-stably; **no
`any`, no index-signature payload bag** (comment-stripped source scan); an
unexpected field is REFUSED ("no unknown-payload bag"); a missing required field
is refused BY NAME; version skew is an explicit refusal, never a silent
migration. Every invalid fixture names its specific field. Coverage gate (l)
complete for all 11 kinds. Compile + fixture suite gated; 12 tests | | |
**I1.14.a–g** | `generate-provenance-aggregation.mjs` |
`provenance-aggregation.json` + `provenance-aggregation/*.ts` | Isis provenance
aggregation for nested workflows — a COMPILING reference. 5 run-node kinds w/
stable ordering + branch + **retry identity** (attempt 2 is a distinct node, a).
Stage provenance validated BEFORE aggregation — unknown artifact refs, hash
mismatches, missing provider/tool/model/actor, out-of-graph runs all rejected
(b). **AGGREGATION NEVER FLATTENS (c)**: all 9 dispositions survive — the same
artifact reused AND duplicated stays 2 distinct rows;
failed/retried/compensated/optional/manual-edited each stay visible; SELECTED
outputs distinguishable from those merely considered. Assertions carry
field-level source; a stage disagreement becomes a CONFLICT naming every source
(never a silently-picked winner); redacted recorded not dropped (d). Manifest
binds to final bytes + child manifests BY HASH, never inlined (e). **HEADLINE
(f): a TRUSTED PARENT CANNOT HIDE A CHILD FAILURE** — aggregate trust = worst
child status, parent claim preserved for audit + override flagged. Lineage
queries cycle-safe w/ unauthorized nodes REDACTED BY NAME (g). I1.14.h
(rebuild/round-trip report) runtime → parent unchecked. Compile + aggregation
suite gated; 13 tests | | | **I1.13.a–e/g/h** |
`generate-artifact-promotion.mjs` | `artifact-promotion.json` + `promotion/*.ts`
| Isis artifact promotion across dev/test/staging/production — a COMPILING
reference. 4 environments each w/ its own store/region/encryption/credential
scope (a, scopes never shared). **ONE admission function serves all 9
consumption channels**
(credential/url/api/provider/storage/cache/search/job/**direct-id**) — NO side
entrance, so a direct-id or cache lookup is refused exactly like an API read
(h). **Development isolation (b)**: a dev artifact is refused by every non-dev
environment through every channel. Promotion is SEQUENTIAL and mints a NEW
identity carrying the same verified bytes + lineage — in-place relabeling and
env-skipping refused. Gates grow toward prod: 11 test (c) / 9 staging (d) / 9
production incl. humanSignoff (e), each unmet prereq NAMED. Concurrent source
mutation caught on every channel via observed-hash comparison. All 7 lifecycle
ops (reject/quarantine/rollback/demote/revoke/recall/expire) leave the artifact
INADMISSIBLE everywhere (g). I1.13.f (byte copy + read-back) runtime → parent
unchecked. Compile + promotion suite gated; 12 tests | | | **I1.12.a–g** |
`generate-replay-eligibility.mjs` | `replay-eligibility.json` + `replay/*.ts` |
Isis deterministic replay eligibility — a COMPILING reference. 4-class taxonomy
(byte-deterministic / semantically-reproducible / provider-best-effort /
nonreplayable, the last ALWAYS with a reason). **The class is DERIVED from the
weakest input, never asserted upward** — an incomplete environment manifest caps
it below byte-deterministic even on a perfect provider; declared provider
nondeterminism caps at best-effort; missing versions/artifacts ⇒ nonreplayable.
Completeness over 15 version keys (b — a null is MISSING, never "latest"),
artifact access/hash/lineage (c), 12 environment keys (d); provider support
declared not assumed (e). **KEY SEPARATION (f): technical reproducibility is NOT
permission** — a perfectly reproducible run is ineligible when rights are
revoked / consent expired / policy changed / budget exhausted. Preview (g)
reports both independently across 6 problem kinds and leaves unknown cost null.
I1.12.h (execute + compare bytes) runtime → parent unchecked. Compile + replay
suite gated; 11 tests | | | **I1.11 (parent + a–h)** |
`generate-compatibility-constraints.mjs` | `compatibility-constraints.json` +
`compatibility/*.ts` | Isis compatibility constraints + planner solver — a
COMPILING reference. 6 component surfaces (model a / adapter b / LoRA c /
checkpoint d / node e / workflow f) consulted by ONE pure solver (g) returning
EXACT issues in 4 kinds (conflict / unsupported / missing-metadata /
version-skew), each naming the precise disagreeing components + field with
concrete alternatives. **NO IMPLICIT CONVERSION**: undeclared precision,
unscanned checkpoint/node, socket dataType/unit mismatch, unknown extension are
refused — never coerced or ignored; absent metadata BLOCKS (never assumed
compatible). Multi-component: adapter stacking, multi-LoRA target-module
overlap, version skew. 12-row fixture matrix (h) covers
pairwise/multi-component/version-skew/missing-metadata/unknown-extension/provider-capability-drift/post-plan-change
— **every incompatible plan is non-dispatchable**, satisfying the parent (fail
planning before dispatch). Compile + compat suite gated; 10 tests | | |
**I1.10.a–i** | `generate-release-gate-composition.mjs` |
`release-gate-composition.json` + `release-gates/*.ts` | Isis release-gate
policy composition — a COMPILING reference. 8 gate classes each w/ a declared
policy (safety a / provenance b / watermark c / quality d / policy e / rights f
/ shape g / review h). **I1.10.i**: one DETERMINISTIC, TOTAL aggregation over 9
outcome states x required/optional x class → 3 decisions
(release/block/review-required), enumerated across all **36 truth-table rows**.
**FAIL-CLOSED**: a required blocker/unavailable/stale/unsupported/partial gate
BLOCKS (quality `missingCountsAsPass:false` — no missing-to-pass).
**NON-OVERRIDABLE**: safety+review can never be waived (attempted waiver
blocks), review prohibits automation sign-off, a waiver w/o waiverRef is not a
waiver. Optional failures warn but never block; pending review ⇒
review-required, never auto-release. I1.10.j (version binding + audit,
integration) runtime → parent unchecked. Compile + release suite gated; 10 tests
| | | **I1.9.a–f** | `generate-quality-profile-schema.mjs` |
`quality-profile-schema.json` + `quality-profile/*.ts` | Isis quality profiles —
a COMPILING reference. Profile scoped to asset-kind + delivery target (a),
measure registry w/ unit/direction/evaluator/non-value states (b), target
thresholds (c). **KEY SAFETY INVARIANT (d)**: safety/integrity floors are
CODE-OWNED + non-overridable — `applyFloor` lets config only TIGHTEN (a
loosening config is clamped to the floor and FLAGGED, never silently accepted).
Layered merge global→…→project under a TOTAL precedence w/ provenance + no
ambiguity (e); a loosening layer is overridden by the code floor. Versioning
bumps + invalidates changed measures + retains raw (f). I1.9.g (calibration on
corpora + domain approvals) = runtime+human → parent unchecked. Compile +
quality suite gated; 9 tests | | | **I1.8.a–f** |
`generate-variant-group-schema.mjs` | `variant-group-schema.json` +
`variant-group/*.ts` | Isis variant groups + comparison membership — a COMPILING
reference. Group (a) scoped to tenant/project/exact-brief holds members w/
source provenance (b). Membership (c) REJECTS cross-tenant + cross-brief (unless
explicit fork); compatibility (d) REJECTS incompatible asset kinds unless a
typed cross-kind adapter. Lifecycle ops (e) IMMUTABLE — add/remove/freeze return
a NEW version w/ append-only history, enforce optimistic expectedVersion +
member limit, never mutate input. Frozen comparison (f) pins revision/hash +
score/gate versions, represents unavailable/stale/redacted members HONESTLY.
I1.8.g (decision integration tests) runtime → parent unchecked. Compile +
variant suite gated; 9 tests | | | **I1.7.a/b/d/e/f/g** |
`generate-lineage-schema.mjs` | `lineage-schema.json` + `lineage/*.ts` | Isis
output/ingredient lineage — a COMPILING directed-graph reference (pairs with
I1.6 outputs). 10 node types (source…release, a) + 9 typed ordered edges
(uses…publishes, b) w/ role/weight/mask/time/range/part; each edge carries
run/stage/actor/model/policy/rights provenance (e). Pure traversal computes
upstream/downstream impact/recall — cycle-safe + depth/fan-out bounded (f).
Integrity rejects dangling/cross-tenant/cross-brief/hash-mismatch and reports
inaccessible nodes HONESTLY (g). Composites preserve DUPLICATE uses without
collapsing equal hashes (d: same source twice at distinct layers). I1.7.c
(transactional edge-creation integration tests) is runtime → parent unchecked.
Compile + lineage suite gated; 10 tests | | | **I1.6.a–j** |
`generate-execution-record-schema.mjs` | `execution-record-schema.json` +
`execution-record/*.ts` | Isis immutable execution record (pairs with the I1.5
plan) — a COMPILING reference. 9 facets
(identity/provenance/parameters/inputs/policy/provider/cost/outputs/timing) +
append-only corrections. Pins EXACT plan+brief revisions (a) + content hashes
(immutability). **Honest seams**: cost.actual nullable — variance without a
settled actual REJECTED (g, no fabricated cost); provider request/response are
secure REFS not inlined secrets (f); null output ref requires a no-output reason
(h); a failed policy decision cannot be dispatch-eligible (e). **I1.6.j**
corrections are NEW linked records keeping the original unchanged + legal hold.
1 valid + 7 invalid fixtures; compile + record suite gated; 9 tests | | |
**I1.5.a–i** | `generate-generation-plan-schema.mjs` |
`generation-plan-schema.json` + `generation-plan/*.ts` | Isis resolved
generation plan (distinct from the brief) — a COMPILING reference that IMPORTS
the I1.4 brief. 7 plan facets + 3 DECISION records (workflow/model/provider,
each recording chosen + rejected alternatives WITH reasons — the "every reason"
requirement). **I1.5.b** brief→plan provenance mapping
(user/default/derived/constrained/unsupported; an over-limit count is
constrained to 8, not silently accepted). **I1.5.a/i** immutable plan
(planHash + replan that keeps the original unchanged + lineage link). **I1.5.h**
validation before dispatch: 1 valid + 6 invalid plan fixtures. Compile + plan
suite (validate/resolve/replan via `tsx check.ts`) gated in verify; 8 tests | |
| **I1.4.a–l** | `generate-generation-brief-schema.mjs` |
`generation-brief-schema.json` + `generation-brief/*.ts` | Isis generation brief
— a COMPILING 12-facet schema. 11 facet interfaces (a–k:
identity/intent/audience/assetKind/locale/references/rights/safety/quality/budget/delivery) +
a cross-field validator suite (l), strict `tsc --noEmit`. **I1.4.g** rights are
evidence-backed CLAIMS (`verified` rejected without an `evidenceRef` — client
truth not accepted). **I1.4.j** budget carries no fabricated cost (nullable
ceiling). **I1.4.l** invariant suite: 8 pure validators + 1 valid / 8 invalid
fixtures, each violating exactly one invariant — run for real via `tsx check.ts`
(accepts valid, rejects each invalid). assetKind grounds to the Isis
GenerationType contract (I0.5.a). Compile + invariant-suite gated in verify; 9
tests | | | **I1.3.b/c/d/e** | `generate-isis-endpoint-dispatch-model.mjs` |
`isis-endpoint-dispatch-model.json` | Isis endpoint dispatch model built on the
WELL-TYPED I1.3.a machine. **I1.3.b** lifecycle state machine (8 states
unconfigured→…→retired, 11 authorized transitions; only `ready` is
dispatch-eligible). **I1.3.c** health contract: 7 probe states (4 realized:
healthy/degraded/unhealthy/unknown + 3 contract additions:
unavailable/incompatible/stale) × 6 detail fields. **I1.3.d** dispatch input
schema (circuit + throttle/quota/capacity/regional/license/dependency), SEPARATE
from health. **I1.3.e** a DETERMINISTIC, TOTAL eligibility decision over all
**1344** lifecycle×health×circuit×maintenance×capability×capacity combos (4
eligible; rest refused with 10 priority-ordered reasons, all reachable) — the
"truth table tests" are pure-function known-value tests. I1.3.f/g need the
running service → parent I1.3 unchecked. 14 tests | | | **I1.2.b** |
`generate-isis-model-identity-boundary.mjs` |
`isis-model-identity-boundary.json` | Isis model identity vs immutable-version
boundary review (I1.2.b), built on I0.5.c entity fields + I1.2.a lifecycle
finding. Partitions 9 dimensions: MODEL IDENTITY (mutable) =
ownership/license/risk/supported-task/lifecycle; IMMUTABLE VERSION (frozen) =
checkpoint/artifact/evaluation/provider-compatibility. Each modeled dimension
maps to real ModelCard/ModelRegistry/ModelVersion fields (validated — drift
gate). **Honest gaps**: license/risk/provider-compatibility are NOT modeled
today; lifecycle is weakly formalized (no ModelStatus enum). I1.2.c–g demand
executable tests → parent I1.2 unchecked. 9 tests | | | **I1.1.b/c/d/g** |
`generate-isis-workflow-state-diagram.mjs` | `isis-workflow-state-diagram.json`
| Isis workflow state-machine VERIFICATION built on I1.1.a. **I1.1.b**
code-derived mermaid state diagram (DRAFT→PUBLISHED→DEPRECATED→ARCHIVED).
**I1.1.c** disposition table reconciling 7 canonical lifecycle states → 4
present (draft/active/deprecated/archived), 3 ABSENT (review/approved/revoked —
no review sub-flow). **I1.1.d** transition matrix: 8/12 implemented with
auth-preHandler guards, 4 absent (submit/review/approve/revoke); verb-synonyms
captured (edit=update, version=createVersion). **I1.1.g** drift gate — reads the
I1.1.a census so a new state/transition in code breaks --check + coherence until
the diagram updates. I1.1.e/f (negative + integration tests) need the running
service → parent I1.1 stays unchecked. 9 tests | | | **S0.10.b–f** |
`generate-migration-plan.mjs` | `migration-plan.json` + `docs/adr/ADR-S5-*.md` |
ADR-S5 incremental migration — Tara adopts the kit WITHOUT a flag day.
**S0.10.b** additive coexistence seams
(adapter/alias/version-negotiation/no-breaking-change); @oshun/tara-workbench
keeps serving all 70 routes. **S0.10.c** 13 migration slices = the S0.2 neutral
contracts, ordered by risk (tier1 EvidenceGroundingPanel/SourceLibrary = Veritas
2nd consumer clears S0.11 first; tier2 0-signal neutral; tier3 mixed-split).
**S0.10.d** enable→verify→coexist→remove per slice, NO dual-write (UI extraction
over the same Tara store, S0.8). **S0.10.e** mixed-version journey: 70 routes +
1 store + deep-links UNCHANGED (component-internal swap). **S0.10.f** retirement
milestone-gated on S0.11 (≥2 consumers) w/ @oshun/bff legacy consumer; absolute
dates = ratification input. Coherence-gated vs S0.2/S0.8/S0.10.a; 9 tests | | |
**S0.9.c/d/e/f** | `generate-workbench-kit-theme-contract.mjs` |
`workbench-kit-theme-contract.json` + `docs/adr/ADR-S4-*.md` +
`workbench-kit-theme/*.css` | ADR-S4 UI composition + theming, built on S0.9.a/b
(token contract kit-owned, values host-owned; 5 systems diverge). **S0.9.c**
9-category token contract (`--wk-*`): 6 mapped to real oshun `--l-*` tokens, 3
(spacing/density/focus) contract-introduces. **S0.9.d** theming isolation
prototype (`tokens.css`): scoped to `[data-workbench-theme]`, 0 bare-`:root`
leaks, `--wk-*` names not classes, single provider, reduced-motion +
forced-colors handled. **S0.9.e** 4 archetype compositions (`archetypes.css`:
calm-authoring/operator/creative-production/control-room) over the SAME contract
with divergent density/motion/contrast — no visual clone. **S0.9.f** a11y
matrix: 6 conditions (theme/density×2/zoom/reduced-motion/forced-colors) each →
a mechanism present in the prototype. Grounded vs S0.9.a comparison; honest that
live cross-host screenshots are runtime; 9 tests | | | **S0.8.d/e/f** |
`generate-event-audit-boundaries.mjs` | `event-audit-boundaries.json` +
`docs/adr/ADR-S3-*.md` | ADR-S3 event/audit/idempotency semantics, built on
S0.8.a/b/c (34 local-ACID + 8 transactional-outbox, 0 saga). **S0.8.d** delivery
semantics (6): at-least-once, per-aggregate ordering,
dedup-by-externalOperationId, durable replay, versioned compatibility,
dead-letter poison. **S0.8.e** audit semantics (6): atomic-in-tx,
refusal/failure capture, append-only immutable, redact-at-write,
retention-policy, evidence linkage. **S0.8.f** the 7 failure scenarios
diagrammed in ADR-S3 (mermaid): before-commit / after-commit-before-dispatch /
duplicate-delivery / worker-loss / external-success-local-timeout /
compensation-failure (NOT APPLICABLE — 0 sagas) / replay. Coherence-gated vs the
committed S0.8.b/c evidence (8 outbox, 0 saga, identity contract); 8 tests | | |
**S0.7.a–g** | `generate-workbench-kit-extension-model.mjs` |
`workbench-kit-extension-model.json` + `docs/adr/ADR-S2-*.md` +
`workbench-kit-extension-model/*.ts` | ADR-S2 workbench-kit extension model: 11
variability dimensions → 5 mechanisms (generic type param / discriminated
interface / plugin capability / composition slot / domain adapter, S0.7.a/b); 6
kit-owned invariants a plugin may not override
(authz/audit/idempotency/honest-capability/no-escape-hatch/deletion-threshold,
S0.7.c); capability-negotiation via typed `CapabilityResult` not-configured
(S0.7.d); **two COMPILING example plugins** — content-authoring (Tara-like, 0
meditation types leaked, S0.7.e) + operations
(jobs/queues/resources/host-targets, S0.7.f) — instantiate the same generic
contract with disjoint domain types; 3 rejected abstractions (single-use
generic, escape hatch, kit-level scoring, S0.7.g). The reference type-checks
under strict `tsc --noEmit` (gated in verify); 6/11 dimensions ground to an S0.2
neutral contract. Coherence-gated + 9 tests (incl. the live compile) | | |
**S0.4.b/c/d/e** | `generate-studio-authoring-overlap-selection.mjs` |
`studio-authoring-overlap-selection.json` +
`tara-studio-authoring-overlap-selection.md` | Recorded S0.4 selection over the
`@oshun/studio-authoring` overlap. Recomputes from the two committed
inventories: **0 exact-name duplicates**; **117/209** Tara exports share a
concern token (source/manifest/gate/review/…) but these are concern overlaps,
not impl duplicates. Selection over the 9 SA concern categories: **4 reuse**
(versioning/collaboration/asset-metadata/localization — neutral, each with a
migration path, S0.4.c) / **0 consolidate** (S0.4.d — no debt: 0 exact dups +
clean one-way acyclic layer) / **5 separate**
(creator-roles/editorial-lifecycle/taxonomy/authoring-blocks/templates —
meditation-specialized, S0.4.e). **Decisive finding: the two packages are
already correctly factored.** Coherence-gated (every category has a disposition,
reuse⇒migration path, consolidate consistent with 0-dup fact, layer one-way; 11
tests) | | | **S0.2.c/d/e/f** | `generate-component-extraction-matrix.mjs` |
`component-extraction-matrix.json` + `tara-component-extraction-matrix.md` |
Recorded S0.2 classification of Tara's 29 studio components (grounded in
S0.2.a/b inventory): **4 neutral candidates** / **13 Tara-owned** / **11 mixed
(9 split at a seam, 2 reject)** / 1 already-shared shell. The 4 neutral + 9
split rows yield **13 extractable neutral contracts**
(AssignmentPanel/CommentThread/ReviewDecisionPanel/PipelineFunnel/EvidenceGroundingPanel/GateRunnerPanel/CatalogBrowser/SchedulingCalendar/BundlePublisher/SourceLibrary/SourceInspector/TaxonomyEditor/AgentAutonomyBudget),
each with an authoring + ops-console two-archetype sketch (S0.2.f). **Seam = the
interaction CONTRACT over a domain-supplied adapter**, never the wired
component. Coherence-gated: partition total+disjoint, high-signal⇒Tara-owned,
zero-signal⇒neutral, every mixed has split\|reject, every contract has both
archetypes (11 tests) | | **Y6.1.a** | `generate-yemaya-storage-inventory.mjs`|
`yemaya-storage-inventory.json` | Yemaya storage inventory (Y6.1.a): 38
relational aggregates over 4 PrismaClients with **120 DB indexes**
(@@index/@@unique, 3.2/aggregate — index-heavy for a creative studio), 20
Asset-blob storage files, 13 Redis-cache files, 4 event entities
(Activity/Notification/Webhook/WebhookDelivery). **Architecture finding**: 0
repository/manager classes — service-direct (consistent with Isis/Veritas).
Backups ops-level (not code-modeled). Y6.1.b builds on it | | **Y6.7.a** |
`generate-yemaya-ipc-registry.mjs` | `yemaya-ipc-registry.json` | Yemaya desktop
IPC registry (Y6.7.a) with **directional orphan detection**: 280 registered
channels across 28 main modules (277 invoke/bidirectional, 3 send/one-way).
Splits renderer→main requests (75), renderer event listeners (23), and
main→renderer emissions (24) by direction — so 9 request channels
invoked-but-unregistered (bellona:connect, dialog:message, …) are correctly
flagged as the "unregistered channels" Y6.7.a eliminates, while
`bellona:engine-_`events are NOT mislabeled broken. 19 dead-listener candidates. Y6.7.b/c build on it | | **Y0.16.a**    |`generate-yemaya-code-census.mjs`     |`yemaya-code-census.json`                                                 | Yemaya launch census (12 artifact classes, each with a **source anchor**): 77 documents, 79 studio pages, 39 app routes (yemaya-route-inventory), 38 entities/406 fields (yemaya-data-inventory), 21 state enums (yemaya-state-vocabulary), 27,337 package exports across 59 packages / 891 capabilities (domain-package-inventories), 67 workers, 1 CLI, 4 integration modules (path-scoped), 11 case subsystems (yemaya-case-inventory), 62 desktop files (yemaya-desktop-partition), 1,225 tests. Alias normalization + dispositions are Y0.16.b+ | | **M8.15.a**    |`generate-metis-api-surface.mjs`      |`metis-api-surface.json`                                                  | Metis canonical API surface (M8.15.a) from committed evidence + the OpenAPI spec: 183 operations across 11 route families, audience tiers (106 privileged / 37 open / 40 authenticated), 20 distinct consumers, and the canonical schemas broken down by type — 421 response / 71 request / 54 base / 8 event / 3 error / 1 stream. **Finding**: the API is response-DTO-heavy with thin error/event/stream typing (12 total) — a canonical-schema gap for the M8.15.b drift gate. Single service owner | | **M7.1.a**     |`generate-metis-themis-integration.mjs`| `metis-themis-integration.json`                                           | Metis↔Themis integration/boundary inventory (M7.1.a): Themis exposes 71 public packages; Metis integrates exactly 3 (@themis/arbitration/justice/policy) through one canonical wiring module (16 typed`MetisThemis*` contracts, 2 enforcement consts). **Boundary finding**: 3 public bare-package imports, **0 private/deep imports** (`@themis/x/src\|dist\|internal`) — the boundary is clean, Metis reaches no Themis internals. M7.1.b+ build on it | | **M1.1.a**     | `generate-metis-contract-inventory.mjs`| `metis-contract-inventory.json`                                           | Metis contract inventory (M1.1.a) with duplicate detection, from the committed OpenAPI spec + shared TS contracts + DB entities (each read once): 558 OpenAPI component schemas collapsing to 518 logical shapes (429 Response / 68 Request / 57 base). **Duplicate-shape finding**: 40 logical shapes have >1 OpenAPI variant (the Response/Request duplication M1.1 targets); 0 cross-representation duplicates. 7 shared-contract files (115 Zod schemas / 108 types), 14 SQLAlchemy entities. M1.1.b+ build on it | | **M0.14.a**    |`generate-metis-code-census.mjs`      |`metis-code-census.json`                                                  | Metis launch census (11 artifact classes, each with a **source anchor**): 51 relevant documents, 39 app routes, 183 OpenAPI operations / 174 with response schemas, 558 schemas, 6,651 package exports across 28 packages, 14 entities / 141 fields / 13 tables, 296 job/event entries across 63 files, 547 app/library tests, 906 implementation files (180 py + 717 ts + 9 js), 5 unique apps + 28 libs. Atomic reconciliation follows in M0.14.b-f | | **B0.14.a**    |`generate-bellona-code-census.mjs`    |`bellona-code-census.json`                                                 | Bellona launch code census (13 artifact classes, each with a stable ID + package-derived owner): 41 libs packages + 23 app projects, 952 code files + 630 tests, 4 control-room + 82 studio routes, 10 runtime apps (dcc-bridge 4 / remote 3 / build 2 / render 1) over grpc/http/ipc/websocket, 6 CLI commands, contract/schema + workflow/plugin/bridge files, 51 docs, and domain-specific deploy targets (unreal/blender/remote-mac/cloud-agent smokes). Atomic decomposition + dispositions + signed owner review are B0.14.b-e | | **A0.6.a**     |`generate-aja-stage-ledger.mjs`       |`aja-stage-ledger.json`                                                    | Aja motion-pipeline stage ledger, derived from the typed`motion-pipeline-sdk`contracts (not imposed taxonomy): 11 canonical`PipelineStage`s (ingestion→delivery, each mapped to its named-reference modules; validation/pose-estimation/retargeting have dedicated libs, the rest run inline), typed inputs (7 `InputConstraints`) / outputs (6 `OutputSpec`) / failures (`StageRetryConfig`+ failed/skipped + failFast), resources (5 job priorities, per-stage+global timeouts, parallel flag), 6`pipeline-cache`modules, and 20 profile combinations (5 domains × 4 quality presets, 8 feature toggles). Canonical DAG + skip data are A0.6.b-c | | **I0.9.a**     |`generate-isis-workflow-class-register.mjs`|`isis-workflow-class-register.json`                                   | Isis workflow class register from canonical artifacts (the codebase's own classification, not imposed): the committed`workflow-taxonomy-registry.json` (66 workflow mappings — registry declares 64, both recorded — across 11 in-code categories, 37 asset families, 5 output profiles, 38 control modalities) + the generated Prisma enums (`WorkflowCategory`12,`GenerationType`15,`WorkflowEngine`5,`ModelType`7 /`ModelFormat`8 = models,`OutputFileType`11) + 30 ai-providers + llm registry + 7 workflow-class modules. Honest coverage check of I0.9.a's 9 proposed classes: 8/9 represented, **training** is a real gap. Class schema (ID/version/capability/cost/lifecycle) is I0.9.b | | **V6.1.a**     |`generate-veritas-storage-inventory.mjs`| `veritas-storage-inventory.json`                                          | Veritas storage inventory (V6.1.a): of 7 named layers, 6 present (tables=67 Prisma aggregates/4 clients, blobs=MediaAsset object storage, indexes=ArticleEmbedding vector, caches=Redis, queues=PushNotificationOutbox/BullMQ, CMS) and 1 absent (documents). **Finding**: Veritas is a rich multi-store server platform (relational-only, no document store) — contrast Euterpe's desktop-file model. Precise per-layer API/model patterns. V6.1.b+ build on it | | **V0.3.a**     |`generate-veritas-contract-diff.mjs`  |`veritas-contract-diff.json`                                               | Veritas contract diff (V0.3.a) — field-level diff of the two typed representations of each entity: the generated Prisma DB schema vs the Zod domain schema, for story(Article)/source(Source)/claim(Claim)/evidence(ClaimEvidence)/review(ClaimVerification). 60 shared fields, 16 Prisma-only (ids/timestamps/soft-delete/counters), 34 Zod-only. Surfaces the real structural mismatch — Zod **nests** (`ownership`, `bias`) while Prisma **flattens** (`ownershipName`, `biasConfidence`, …) — the exact semantic mismatch V0.3.b reconciles; nullability captured per Zod field. `release`has no schema pair (honestly noted) | | **V0.10.a**    |`generate-veritas-data-catalog.mjs`   |`veritas-data-catalog.json`                                                | Veritas data catalog from the generated Prisma client: 67 named data classes (57 primary entities + 10 satellite forms — model-context/embedding, derived render/variant/revision/translation/normalized, log outbox/delivery, tagged from model naming) against a 45-enum classification vocabulary. Of V0.10.a's 7 named form types, derived/model-context/log are entity-modeled while indexed/cached/backup/export are honestly recorded as infrastructure (not Prisma entities). Owner = the veritas package; the sensitivity/compartment classification is V0.10.b | | **Y0.12.a**    |`generate-yemaya-data-inventory.mjs`  |`yemaya-data-inventory.json`                                              | Yemaya data inventory from the generated Prisma client: 38 named entities / 406 scalar fields. Of Y0.12.a's 11 named data forms, 6 are entity-modeled (entities, fields, Asset file/blob storage, 4 log entities AuditLog/Activity/UsageRecord/AssetUsage, 4 event entities Activity/Notification/Webhook/WebhookDelivery) and 5 are honestly recorded as infrastructure not modeled as entities (prompts, search, exports, backups, local-caches). Classification labels + per-category owner are Y0.12.b | | **Y0.10.a**    |`generate-yemaya-state-vocabulary.mjs`|`yemaya-state-vocabulary.json`                                             | Yemaya state vocabulary inventory from the generated Prisma enums: 21 typed enums / 115 labels (89 distinct), with cross-enum label **collision detection** — 21 collisions (ACTIVE across ProjectStatus/SubscriptionStatus/UserStatus, DRAFT across AssetStatus/InvoiceStatus/ProjectStatus, ADMIN/FREE/OTHER ×3) — plus per-enum owner and lifecycle-stage keyword coverage. Honest finding: of Y0.10.a's 8 stages, content/production/release/consumer are typed while **study/case/review/render have no typed state vocabulary** (verified). The study-content state machine is Y0.10.b | | **I0.5.c**     |`generate-isis-model-registry.mjs`    |`isis-model-registry.json`                                                  | Isis model registry inventory (I0.5.c model rows) from typed schema: 7 model kinds (`ModelType`: CHECKPOINT/LORA/CONTROLNET/VAE/EMBEDDING/UPSCALER/CUSTOM) in 8 `ModelFormat`s, managed by 3 Prisma entities (ModelCard/ModelRegistry/ModelVersion, 52 scalar control fields parsed via `*MinAggregateOutputType`, categorized by control dimension). All 8 of I0.5.c's named controls map to typed evidence (model/checkpoint/adapter/LoRA→ModelType, version→ModelVersion, intake→storage/upload, evaluation→isVerified+hash provenance, selection→usage/download/star ranking); none absent. Generation controls = I0.5.a, training = I0.5.d | | **I1.3.a**     | `generate-isis-endpoint-state-machine.mjs`|`isis-endpoint-state-machine.json`                                    | Isis provider-endpoint state-machine source mapping (I1.3.a) from`libs/isis/ai-providers/src/provider-endpoints`(files read once): 4 typed health states (healthy/degraded/unhealthy/unknown), 9 provider kinds, 7 regions, 6 typed config interfaces (endpoint/cost/concurrency/retry/secret-rotation/registry), subsystems registry/health-monitor/router/dead-letter/queue-metrics. **Positive finding**: unlike the weakly-formalized model machine, this is well-typed — all 9/9 I1.3.a dimensions map to typed evidence; 16 UI components, 3 test files. I1.3.b+ build on it | | **I6.1.a**     |`generate-isis-persistence-inventory.mjs`|`isis-persistence-inventory.json`                                     | Isis persistence current-state inventory (I6.1.a): 24 Prisma aggregates over 1 schema, 9 persistence services. **Architecture finding**: 0 repository classes (Isis uses service-direct Prisma access, no repository layer); **3 PrismaClient instantiations** in the persistence layer — relevant to I6.1's shadow-store concern for the ownership ADR. 4 files use transactions, 7/24 aggregates are tenant-scoped (organizationId), 4 carry soft-delete + a`RetentionPolicy`aggregate; backup is ops-level (not modeled in code). I6.1.b+ build on it | | **I1.2.a**     |`generate-isis-model-state-machine.mjs`|`isis-model-state-machine.json`                                          | Isis model/model-version state-machine source mapping (I1.2.a) from`model.service.ts`(read once): 26 service operations (create/edit/delete + intake-upload/verification/delivery/audit/health). **Verification finding**: unlike the workflow machine, the model lifecycle is *weakly formalized* — NO typed`ModelStatus`enum, the intended states appear only as scattered literals across 28 files with no canonical definition, and status transitions ride on`update()`(no explicit approve/deprecate methods); 7 ModelType + 8 ModelFormat contracts, 39 model UI components, but only 2 model-service test files (thin). I1.2.b/c build on it | | **I1.1.a**     |`generate-isis-workflow-state-machine.mjs`|`isis-workflow-state-machine.json`                                    | Isis workflow-template state-machine source mapping (I1.1.a) from current code: 4`WorkflowStatus`states + 4 visibility levels, persisted on Workflow.status / WorkflowVersion, realized in`apps/isis/workflow-registry/.../workflow.service.ts`. **Verification finding**: 8/12 intended transitions have a code signal (core lifecycle create/edit/version/activate/deprecate/archive/restore/publish, via verb-synonym detection edit→update, version→createVersion), while the review sub-flow (submit/review/approve/revoke) is honestly reported absent. 62 tests, 16 operator-UI components, 7 event sources, guard preHandlers present. Diagram/reconciliation/verification are I1.1.b-f | | **I0.5.b**     | `generate-isis-workflow-controls.mjs` |`isis-workflow-controls.json`                                               | Isis workflow controls inventory (I0.5.b workflow rows) from typed schema + the comfyui surface: 5 graph engines (`WorkflowEngine`), 12 categories, 4 lifecycle states, 4 visibility levels; 4 Prisma entities (Workflow/Template/Version/Star); 9 `3d-comfyui-nodes`node definitions + 25 comfyui-sdk node types + a node-validator (5 exports) + 3 execution modes (always/on_change/manual). All 8 named controls map to typed evidence, none absent (graph→WorkflowEngine, template→WorkflowTemplate, node→3d-comfyui-nodes, parameter→InputSlot/OutputSlot, version→WorkflowVersion, import→changelog, validation→node-validator, execution→run-stats+ExecutionMode) | | **I0.5.d**     |`generate-isis-training-controls.mjs` |`isis-training-controls.json`                                              | Isis training controls inventory (I0.5.d training rows) from the`lora-training-surface` typed contracts: a 9-state training lifecycle (`TRAINING*RUN_STATES`: pending-rights-attestation→…→promoted/rejected/aborted), 5 dataset rights assertions (owns-or-licensed/no-pii/no-csam/…), 4 evaluation metrics (fid/clip-score/psnr/ssim), 38 typed control fields (rank/epochs/learningRate/costEstimateCents/checkpointStep/…) across 4 training libs + a multi-GPU resource provider. All 8 of I0.5.d's controls map to typed evidence, none absent (dataset→rights, training→lifecycle, fine-tune→LoRA-rank, checkpoint→CheckpointSample, evaluation→metrics, promotion→promoted-state, resource→multi-gpu, cost→costEstimateCents) | | **I0.5.a**     | `generate-isis-generation-controls.mjs`| `isis-generation-controls.json`                                           | Isis generation controls inventory (I0.5.a generation rows) from typed schema: 15`GenerationType`kinds; the`GenerationJob`run entity (32 fields bucketed by dimension — brief=prompt/negativePrompt/inputUrl, run-lifecycle, job-execution=queue/worker/gpu/cost) with`JobStatus`/`JobPriority`; the `GeneratedOutput`entity (29 fields) with`OutputFileType`(11)/`OutputStatus`(5); contract paths via 4 client resources (generation/models/outputs/workflows) + 30 ai-providers. brief/run/output/contract/provider/job map to typed evidence; **variant** has no dedicated entity (output cardinality per jobId) — recorded honestly. Completes the I0.5.a-d control family | | **A0.4.a**     | `generate-aja-embodiment-boundary.mjs`|`aja-embodiment-boundary.json`                                            | @oshun/embodiment-aja contract boundary: 3 declared export subpaths (. / adapter / canonical-adapter, 76 public symbols), 3 consumers, and 0 private/deep-import violations — the boundary is clean (every consumer import uses a declared public subpath), the A0.4 invariant | | **A0.7.a**     |`generate-aja-format-registry.mjs`    |`aja-format-registry.json`                                                 | Aja motion interchange format registry: 7 formats (fbx/gltf/usd/bvh/alembic/mocap-import/clip-database) over 10 extensions, each with import/export-capable symbols, validators, version constants, binary signatures; 5 round-trip formats; 58 consumers across 30 projects | | **A0.2.a-c**   |`generate-aja-inventory.mjs`          |`aja-inventory.json`
| Aja motion-capture console inventory (no web app): 98 /studio/aja/* Studio
routes + 3 motion services (svc-motion-ai 9 HTTP routes, svc-motion-pipeline 25,
svc-reference-video library — classified) + a CLI with 7 commands. Requiring a
leading`/`filters Map.get() false positives | | **E0.14.a**
|`generate-euterpe-code-census.mjs` |`euterpe-code-census.json` | Euterpe launch
code census (reconciliation inputs): 49 libs packages + 2 app projects, 918 code
files + 497 tests, 3 routes over 148 DAW modules + 58 components, dual
Next.js+Tauri host (4 native commands, 5 rust files), 52 docs, DAW deploy
targets (build:native/build:worklet) — from git counts + committed E0.1/E0.2
evidence | | **E0.1.a** |`generate-euterpe-host-inventory.mjs`
|`euterpe-host-inventory.json` | Euterpe host/entry-point inventory: a DUAL host
— Next.js web frontend (3 routes) wrapped by a Tauri desktop shell ("Euterpe
Studio", app.oshun.euterpe-studio) with 4 native #[command] handlers
(load/save*project, export_stems, reveal_in_file_manager) + capabilities, single
Nx project | | **E7.1.a** |`generate-euterpe-storage-inventory.mjs`|
`euterpe-storage-inventory.json` | Euterpe storage inventory (E7.1.a): of 9
named storage layers, 5 present (local=localStorage 4w/4r, desktop=4 Tauri
commands save/load/export/reveal, blob, realtime=yjs/CRDT, distributor) and **4
absent** (OPFS, database, index, cache). **Finding**: Euterpe is
desktop-file-centric (Tauri fs) + localStorage for client state, with NO
browser-persistent storage. Precise API-signature patterns (loose ones gave
false positives); 14 project-store modules. E7.1.b+ build on it | | **E0.2.a/b**
|`generate-euterpe-inventory.mjs` |`euterpe-inventory.json` | Euterpe DAW
route + module inventory: 3 Next.js routes (/, /daw, /realtime) over 148
engine/logic modules + 58 components across 6 subsystems (daw alone = 135
modules + 150 tests), with worklet/worker/Tauri-bridge signal detection (0
in-app worklets/workers — audio engine is external) | | **V0.1.c**
|`generate-veritas-collision-ledger.mjs`|`veritas-collision-ledger.json` |
Veritas route collision ledger (cross-surface from V0.1.a): 0 duplicate page
routes, 0 protected/public boundary violations (public reader is locale-scoped,
never reaches a protected /veritas/* path — boundary intact), 0 CMS↔BFF
overlaps, 0 dead paths. Split-ownership/legacy-alias deferred to V0.1.d | |
**V0.14.a** |`generate-veritas-code-census.mjs` |`veritas-code-census.json` |
Veritas launch census: 64 libs packages + 13 app subsystems (9
worker/agent/media), 1270 code files + 156 tests, 58 routes across
studio/reader/CMS/BFF, 9 state enums (boundary clean), 79 docs, 28 deploy
targets (web+mobile+workers) — from git counts + committed V0.1/V0.4 evidence |
| **V0.4.a**
|`generate-veritas-state-inventory.mjs`|`veritas-state-inventory.json` | Veritas
editorial state-value inventory: 9 state-vocabulary Zod enums across 9 distinct
vocabularies (kebab-case + lowercase), with the recurring values
(draft/pending/completed each in 4 enums) that V0.4 must reconcile into one
canonical vocabulary | | **V0.1.a**
|`generate-veritas-route-inventory.mjs`|`veritas-route-inventory.json` | Veritas
newsroom route inventory across 4 heterogeneous surfaces: 11 Oshun-Studio
/veritas/_ authoring pages + 3 veritas-web public-reader pages
([locale]-prefixed) + 28 veritas-cms Fastify service routes (headless, no
pages) + 2 BFF modules (16 registrations). Authoring vs reader vs service split
surfaced | | **Y0.7.a** |`generate-yemaya-case-inventory.mjs`
|`yemaya-case-inventory.json` | Yemaya interactive-case pipeline subsystem
inventory: 11 case-_ subsystems
(compiler/director/engine/verifier/eval/localization/contracts/…, 585 symbols,
111 code files, 37 tests) with per-subsystem exports/consumers/deps/disposition
and the intra-pipeline dependency graph (21 edges, acyclic — case-contracts is
the foundation) | | **Y0.2.a**
|`generate-yemaya-desktop-partition.mjs`|`yemaya-desktop-partition.json` |
Yemaya Electron desktop code partition: 62 files by process (main 36 / renderer
22 / preload / shared / score-editor), the IPC surface (~297 main handler
channels ↔ ~102 renderer calls via 1 contextBridge namespace), and renderer→main
leakage detection (0 — a clean boundary) | | **Y0.1.a**
|`generate-yemaya-route-inventory.mjs` |`yemaya-route-inventory.json` | Yemaya
Studio route inventory: 33 web (React-Router, RouteSuspense pageName + nested
Layout) + 6 Electron desktop renderer routes, each with
path/element/dynamic/test. Web pages lazy-load via a barrel, so API is captured
at app level (8 /api/\_ web endpoints, 18 desktop IPC channels) | | **M0.6.a/b**
|`generate-metis-package-inventory.mjs`|`metis-package-inventory.json` | Metis
package inventory + capability catalog: 28 packages and all 6,651 public
capabilities mapped to exact source anchors, audiences, owner contracts, API
families, parent tasks, and 28 reviewed count/hash-pinned dispositions. Missing,
extra, changed, unanchored, broken-reference, and unmapped audits are all zero |
| **M0.5.a**
|`generate-metis-persistence-inventory.mjs`|`metis-persistence-inventory.json` |
Metis persistence inventory: 2-migration linear Alembic chain (001→002, single
head), 13 current tables + 9 index ops (parsed from upgrade() only, not the
downgrade rollback), async pooled engine via get*db, direct-session data access
(no repository layer) | | **M0.4.a** |`generate-metis-data-model.mjs`
|`metis-data-model.json` | Metis data-model inventory: 14 SQLAlchemy entities
(141 fields, 16 FKs, 14 relationships) with per-field type/nullability/index/FK,
grouped by M0.4 data category (learner-data 5, assessment-evidence 3,
authoring-data 3, identity, institutional-config) | | **M0.7**
|`generate-metis-api-inventory.mjs` |`metis-api-inventory.json` | Metis API
route/flow ledger: Python-AST inventory of 183 canonical service operations + 2
Manim render routes, 30 registrations across 6 BFF modules, and 39 frontend
pages (254 rows total). Every row maps authority/owner/store/provider/contract,
8 boundary controls, live/fixture state, consumers/tests, and a reviewed
disposition (233 refactor / 19 facade / 2 reuse); all traceability drift gates
are zero | | **M0.8** |`generate-metis-content-hierarchy.mts`
|`metis-content-hierarchy.json` | Canonical strict hierarchy contract for all 13
kinds from source collection through prerequisite: 10 ordered/cardinality/ reuse
containment rules; exact revision/hash pins for stimulus, rubric, alignment, and
prerequisite edges; never-inline answer refs; 8 lifecycle relationship policies;
4 valid + 6 invalid goldens; 13-row existing-shape reconciliation; all 6 M0.8
controls present | | **M0.9** |`generate-metis-authoring-maturity.mts`
|`metis-authoring-maturity.json` | Independent Metis authoring maturity model:
27 current vocabularies across all 9 authoring/review/publication/learner/
attempt/mastery/completion/grade/integrity categories with declaring and
canonical owners; 10 maturity states and 14 guarded transition policies; 6
protected state owners; 30 normalized label collisions, 27 API/UI mappings, and
7 migration risks; all 6 M0.9 controls present with zero drift | | **M0.10**
|`generate-metis-threat-model.mts` |`metis-threat-model.json` | Metis threat
model and mitigation/test traceability: 78 owned context items across all 9
required dimensions; 60 explicit cases in 7 threat categories; 23 controls with
exact implementation/test anchors and honest implemented/partial/
contract-defined status; 7 incident playbooks and 7 derived launch gates. All 9
M0.10 controls are present with zero drift; 6 gates remain blocked and academic
integrity conditional, preserving the documented launch gaps instead of a false
green | | **M0.11.a/c/d/e** |`generate-metis-regional-policy.mts`
|`metis-regional-policy.json` | Metis regional education/privacy engineering
review packet: 38 owned applicability items across all 7 required dimensions and
8 launch cells; all 11 obligation prompts per cell; 16 nonwaivable policy rules
across 9 domains; platform→region→institution monotonic merge; 7 lifecycle
workflows with 31 downstream propagation actions across 8 processor classes; and
8 exact implementation/test anchors. All technical drift is zero and 4/6 M0.11
properties are complete. M0.11.b/f and the parent intentionally remain open: all
cells fail closed pending a real named counsel/privacy reviewer, reviewed
version, decision, and effective date | | **M0.12.a-f**
|`generate-metis-success-metrics.mts` |`metis-success-metrics.json` | Proposed
Metis success-measure catalog with 32 owned metrics across authoring quality,
accessibility, psychometrics, time-to-publish, interoperability, and learner
outcomes; 12 exact source/test anchors; and 4 consented aggregate outcome
hypotheses. Every metric defines formula/denominator, freshness, privacy,
countermetrics, cadence, proposed target, and prohibited uses. Flow measures are
withheld when correctness/release/accessibility guardrails fail; small groups
are suppressed; causal and individual high-stakes use are prohibited. All
technical drift is zero and 6/7 M0.12 properties are complete. M0.12.g and the
parent remain open pending real accountable-owner approvals and an effective
date | | **M0.13.a-e** |`generate-metis-design-doc.mts`
|`V1/METIS_COURSEWARE_WORKBENCH_DESIGN_2026-07-23.md` | Deterministic Metis
courseware workbench design synthesized from 33 content-addressed evidence,
document, and code sources: 11 explicit live/partial/fixture/proposed current-
state rows; all 6 audiences and 9 required roles; hierarchy, maturity, data,
API/job/event and standards boundaries; a restrained evidence-first UI thesis; 6
governed journeys; responsive/accessibility/offline/localization, trust,
evaluation, observability, and operations rules; a reversible 7-phase rollout;
and 7 owned unresolved ADRs. M0.13.a-e are technically complete. M0.13.f and the
parent remain open with approvals=0 and no fabricated named domain/security/
privacy/accessibility/engineering/product review | | **M0.14.b-f**
|`generate-metis-feature-reconciliation.mjs`
|`metis-feature-reconciliation.json` + `metis-feature-reconciliation.rows.jsonl`
| Deterministic atomic Metis feature reconciliation grounded in a
repository-reviewed alias decision map. It retains 10,021 source identities and
semantics from 1,747 document features, 39 routes, 183 OpenAPI operations, 558
schemas, 6,651 package exports, 296 job/event entries, and 547 tests; maps them
bidirectionally to 30 owned canonical capabilities; and assigns every row a
disposition, parent/child tasks, dependencies, and exact proof. All 23
capability families named by M0.14.c-e have implementation rows. The gate
resolves 1,466 collisions deterministically, explicitly defers 111
broad/future/unsupported assertions, and proves zero source/canonical orphans.
M0.14.g and the parent remain open because owner approval/effective date are not
fabricated | | **M0.1.b**
|`generate-metis-audience-classification.mjs`|`metis-audience-classification.json`
| Metis surface audience classification: all 39 routes assigned one of the 6
canonical audiences
(authoring/institutional-setup/educator-operations/learner-delivery/observer/operator-adjudication)
via auditable host+path rules, no blank cells. Owner/ratification stay M0.1.c-f
(human) | | **M0.1.a** |`generate-metis-route-inventory.mjs`
|`metis-route-inventory.json` | Metis route inventory across 4 fragmented
surfaces: metis-admin (15) + metis-web (13) + oshun /metis (10) +
/operator/metis (1) Next.js pages, each with source/host/mount/path, API
consumers (20 distinct /api/metis|admin|auth), nav links, dynamic-segment flag,
and test. /domains/metis has no tracked pages | | **I0.6.a**
|`generate-isis-package-inventory.mjs` |`isis-package-inventory.json` | Isis
package capability inventory: 76 libs/isis/* packages consolidated from the
BASE.3 graph + BASE.9 capabilities — per package
exports/consumers/deps/disposition, attributed atomic capabilities (496
entity+event) and test files (1138), plus a name-derived role signal. 29
no-consumer (retire-candidate) packages surfaced | | **I0.3.a**
|`generate-isis-page-inventory.mjs` |`isis-page-inventory.json` | Isis operator
web page inventory: 26 pages source-linked to their App.tsx route + permission
role, with isis-client API operations (118 distinct client.<ns>.<method>),
react-query query/mutation usage, local state, hooks, and test. Surfaces 2
unrouted page components + 6 pages without a co-located test | | **I0.1.d**
|`generate-isis-effective-route-matrix.mjs`|`isis-effective-route-matrix.json` |
Isis effective-route matrix (from the route register + studio-boundary rule):
all 192 routes classed by effective behavior — 88 studio routes gated-concealed
(source-mounted-but-unreachable via the empty allow-list), 77 live operator
APIs, 24 permission-gated, 2 public. No blank cells | | **I0.1.a/b**
|`generate-isis-route-register.mjs` |`isis-route-register.json` | Isis route
register across 3 surfaces: 27 operator-SPA routes (apps/isis/web, per-route
permission), 88 /studio/isis/\* Studio pages, 77 admin-isis-\_ BFF modules.
STUDIO_ISIS allow-list is empty ⇒ all 87 Studio segments fail-closed (AAA-only);
56 are blocked but absent from the documented 31-segment AAA catalog |

| **S0.6.c/d/e** | `generate-workbench-kit-boundary.mjs` |
`workbench-kit-boundary.json` | ADR-S1 staged-hybrid boundary: responsibility
map (concern→layer), 6-layer normative dependency rules (15 allowed downward /
21 forbidden upward+same-layer), and a cycle-simulation boundary proof —
proposed graph acyclic + cross-checked against the S0.4 current graph. Named
approval (S0.6.f) stays human |

## Determinism contract

Every generator is driven by `git ls-files` / `git grep` / committed artifacts,
emits sorted output with **no timestamps**, and supports `--check` (fail on
drift). The BASE.1 manifest pins the SHA to the last scoped-source commit so
snapshot/docs-only commits never spuriously drift. All outputs are
prettier-ignored (generator-owned).

## Ledger status (foundational sections)

Done with committed evidence: **BASE.1** (a–g), **BASE.2.a**, **BASE.3**
(a,b,c), **BASE.5** (a,b,e), **BASE.9** (a,b,c,d,h,i), **BASE.13**
(a,b,c,d,e,g,h), **DEP.1** (a–g), **S0.1** (parent + a–g), **S0.2** (a,b),
**S0.3** (parent + a–f), **S0.4** (a,f), **S0.5** (a,b,c,d), **S0.11**
(a,b,c,f), **S0.6** (a,b,c,d,e — ADR-S1 staged-hybrid adopted pending S0.6.f
named ratification), **S0.8** (a,b,c), **S0.9** (a), **S0.10** (a). Phase I:
**I0.1** (a,b,d), **I0.3** (a), **I0.6** (a). Phase M: **M0.1** (a,b), **M0.4**
(a), **M0.5** (a), **M0.6** (a,b), **M0.7** (a). Phase Y: **Y0.1** (a), **Y0.2**
(a), **Y0.7** (a). Phase V: **V0.1** (a,c), **V0.4** (a), **V0.14** (a). Phase
E: **E0.1** (a), **E0.2** (a,b), **E0.14** (a). Phase A: **A0.2** (a,b,c),
**A0.4** (a), **A0.7** (a), **A0.14** (a). Phase B: **B0.2** (a,b,c). All seven
domains now have a committed P0 baseline inventory. Package maps: **Y0.6**
(a,b), **V0.2** (a), **E0.3** (a), **A0.3** (a), **B0.3** (a).

S0.2.c–f (domain-neutral vs meditation-owned classification, seam splits,
two-archetype validation) remain human design decisions; the inventory carries a
per-component `meditationVocabularySignal` as raw input to that review, not a
verdict, so the S0.2 parent stays unchecked.

Deliberately left unchecked (honest scope): items needing **runtime execution**
(BASE.2.b–g, BASE.6/7/8 walkthroughs & probes), **human governance** (BASE.4.f,
BASE.5 approvals, BASE.11.h, BASE.15, DEP.2 ownership), or **semantic/large
work** (BASE.9.e state-machines/policies, BASE.9.f feature-docs, BASE.9.g
compound-split review, BASE.10 dedup, BASE.11 disposition, BASE.13.f
documentation gating, BASE.14 bidirectional map). The per-item ledger comments
record exactly what evidence exists and what remains.

## Owner-routed findings surfaced (see `inventory/reconciliation.json`)

- 11 stale `lilith-*` cross-domain name leaks in `libs/aja/*` project.json
  files.
- 3 duplicate npm package names (`@veritas/{analytics,cms,notifications}` across
  `apps/` and `libs/`).
- 191 libraries with no static external consumers (retire-candidates); 12
  refactor-candidates; `@oshun/documentation` has a missing declared entry.
- **S0.1 (Tara extraction source):** of `@oshun/tara-workbench`'s 209 exported
  symbols, only 56 are consumed in production — all by the single `@oshun/bff`
  project — and 151 have no production or test consumer at all (see
  `tara-import-graph.json.findings.unusedExports`). No web surface imports the
  domain library directly (it goes through the BFF); no deep imports past the
  barrel and no intra-package cycles exist today. This large single-consumer,
  mostly-unused surface is the primary input to the S0.11 deletion threshold and
  the S0.6/S0.7 extraction ADRs. | | **I2.5.a–h** |
  `generate-model-selection.mjs` | `model-selection.json` +
  `model-selection/*.ts` | Isis governed model selection — a COMPILING
  reference. Capability query over
  workflow/node/task/modality/asset/profile/IO/provider/runtime/hardware (a);
  governance requires an approved-active lifecycle, an attested source, a SIGNED
  artifact, a security scan BOUND TO THE ARTIFACT HASH, and a current evaluation
  (b); compatibility covers
  precision/format/context/adapter/endpoint/resource/target (c); licensing is
  evaluated for THIS use — commercial/training/derivative/identity, territory,
  attribution, restricted content classes (d); admission covers
  promotion/region/DATA RESIDENCY/entitlement/quota/capacity/budget (e); ranking
  is deterministic by policy→quality→reproducibility→latency→cost→capacity with
  a stable tie-break (f). All four states the parent names — quarantined /
  incompatible / unlicensed / inadmissible — are distinct typed codes with
  deliberate fixtures (19 refusals total). Curated tiers get an
  `AudienceSafeModel` that structurally lacks
  `internalWeightsUri`/`providerAccountId`/`trainingCorpus` (g); dispatch
  revalidates and BLOCKS a stale pinned artifact hash plus post-selection
  revocation/de-licensing (h). Serving readiness is a separate stage and every
  one of the 11 declared reason codes is reachable — an unconfigured integration
  is `not-configured` (an honest absence) and a draining endpoint is
  `unavailable` (transient), and the suite asserts neither is ever reported as a
  governance refusal | | | **I2.6.a–g** | `generate-modifier-selection.mjs` |
  `modifier-selection.json` + `modifier-selection/*.ts` | Isis LoRA/modifier
  selection confined to the 2 authorized hosts (aaa, operator) — a COMPILING
  reference. Registry entry carries base-model linkage, target modules, rank,
  trigger words, strength range, license, safety and lifecycle (a); host
  authorization resolves FIRST and independently, then entitlement and
  tenant/project scope — a foreign modifier marked `shared` stays visible, so
  the rule discriminates (b); 16 deliberate refusals across 13 reachable codes,
  and a likeness modifier is refused for missing consent even when no identity
  use is declared. COMBINATION rules decide what per-item checks cannot — stack
  depth, cumulative influence, per-module contention, exclusive-group collision
  — and an out-of-range weight is refused, never silently clamped (c).
  Effective-influence diagnostics fold weight × a trapezoid-integrated schedule
  × mask coverage and name all 4 ways a control can be set yet inert (d).
  Dispatch revalidates exact versions/hashes, blocks a stale reference and a
  planning→dispatch quarantine race (e). CURATED LEAK AUDIT: the curated types
  declare none of the 11 fields a modifier could travel through, plus a
  field-by-field scan with a negative control (f); NO EXISTENCE ORACLE — real
  and fictional IDs give byte-identical responses across all 4 entry points and
  the call-counted registry seam records zero calls (g) | | | **I2.7.a–g** |
  `generate-advanced-controls.mjs` | `advanced-controls.json` +
  `advanced-controls/*.ts` | Isis sampler/advanced node controls confined to 2
  authorized hosts — a COMPILING reference. Catalog with
  versions/units/ranges/dependencies/risk/impact (a); a 2-host × 4-verb
  permission matrix where a curated grant CANNOT BE WRITTEN
  (PermissionGrant.host is typed to the advanced hosts) and an exact row
  replaces its wildcard (b); ControlValue is a closed union of 5 scalar kinds —
  no object/json/expression/Record — so arbitrary JSON or code has no
  representation, effective values carry provenance across policy > user >
  preset > catalog > code-default, an invalid value falls back to the code
  default with its reason, an off-grid value is refused not rounded, and impact
  is computed (c); custom-node security gates
  catalog/signature-bound-to-digest/publisher/license/review-freshness/all 3
  sandbox capabilities/resources with 12 deliberate failures (d); dispatch
  revalidates and blocks a de-catalogued node race (e); PublicRunSummary
  declares none of the 9 control fields and lower-tier errors collapse to one
  generic message (f); and the adversarial suite closes all 5 bypass routes —
  direct, alias, nested/dotted, unknown field, and mass assignment with a real
  own `__proto__` key (the fixture is JSON.parse'd because a literal would
  create no own key) — because the guard iterates the allow-list rather than the
  request's keys (g) | | | **I2.8.a–g** | `generate-best-of-n.mjs` |
  `best-of-n.json` + `best-of-n/*.ts` | Isis best-of-N variant requests — a
  COMPILING reference. Typed contract with counts/diversity/selection/seed
  policy/hard maximums (a); DETERMINISTIC expansion deriving
  candidate/run/output identities from the request and index so two requests
  collide on neither identity nor derived seed, with shared request/brief
  lineage (b); an estimate itemizing all 6 cost components for the complete set,
  latency computed wave-wise not summed (c); ATOMIC reservation across 7 pools —
  every pool checked before any is debited, and the refusal names EVERY short
  pool with requested/available/short — plus contention coverage and a refused
  double release (d); a confirmation BOUND to a digest of the estimate it
  approved, so a token for 4 candidates cannot authorize 8 and a post-approval
  price move invalidates it, with all 4 threshold kinds proven reachable (e);
  execution through a worker seam so peak concurrency is measured, with bounded
  retry, partial failure, and cancellation that preserves never-started
  candidates and credits their share back exactly (f); and reconciliation where
  NULL IS NOT ZERO — an unreported cost stays null with no variance and the
  aggregate states how many never reported (g) | | | **I2.9.a–f** |
  `generate-variant-grouping.mjs` | `variant-grouping.json` +
  `variant-grouping/*.ts` | Isis manual + generated variant grouping — a
  COMPILING reference. Stable set identity, TOTAL ordering (order then id, so
  tied members never swap), roles, parent/revision links, source discriminator
  (a). PARENT BAR AT THE TYPE LEVEL: `ManualVariant` declares none of the 12
  generation-provenance fields, so inventing a run/model/seed/score for an
  upload is a compile error, and `Variant` is a discriminated union so no
  consumer can forget manual variants exist (b). A generated variant needs a
  COMPLETE immutable lineage and its measurements pass through unchanged
  carrying `provider:<scorer>` / `gate:<gate>` attribution (c). Absence survives
  aggregation — mean over measured only, null when nothing was measured,
  unmeasured count reported — and no payload may claim a score for a manual
  variant, invent a measurement, overwrite one, or declare an unrun gate passed,
  while restating a real value is allowed (d). Compatibility covers
  kind/profile/role/capacity/duplicates/parents with mixed-source sets forbidden
  by default and permitted by policy (e). All 8 decisions are recorded: history
  is appended never replaced, each entry carries members before and after, so
  any past state is reconstructible and a removed member still has a history (f)
  | | | **I2.10.a–e, h** | `generate-comparison.mjs` | `comparison.json` +
  `comparison/*.ts` | Isis side-by-side / A/B comparison — a COMPILING
  reference. Sessions pin every source to an exact revision + content hash (a).
  Image sync lives in NORMALIZED coordinates, not pixels, so two images of
  different pixel sizes share one fit scale; the suite proves the viewport
  transforms are exact inverses, that zoom leaves its focal point fixed
  off-centre, and that a difference view refuses mismatched dimensions rather
  than resampling; deliberate unlinking is first-class (b). Video maps each
  follower on its OWN frame grid and exercises all 4 duration-mismatch policies
  — `reject`/`anchor-start` return no time past the shorter clip rather than
  holding a meaningless frame (c). Audio NEVER normalizes away a measured
  difference: matching is off by default so a 6 LU gap survives, and when on the
  offset is reported while the measurement is carried unchanged (d). 3D
  reconciles units and axes — the same object at 2m/y-up and 200cm/z-up
  reconciles identically while a genuinely smaller object stays 200x smaller —
  with real trigonometry for framing (e). Decisions bind to revision AND content
  hash, so a same-numbered revision with changed bytes is stale, plus optimistic
  concurrency (h). **I2.10.f/g deferred**: their bars name an accessible audit
  of a rendered surface; the data models (NonvisualPanel, RestorableState) are
  built and asserted so the runtime audit has something to run against | | |
  **I2.11.a–g** | `generate-decision-records.mjs` | `decision-records.json` +
  `decision-records/*.ts` | Isis decision records for
  select/reject/park/regenerate — a COMPILING reference. Records are IMMUTABLE
  with a BACKWARD-only supersedes link, so a later outcome has no writable
  history to rewrite (a). PARENT BAR enforced twice: at write time a selection
  must account for every compared candidate as chosen or nonselected, and at
  read time as a REACHABILITY invariant asserted over a four-decision chain
  where all outputs stay linked while their dispositions change (b). Rejection
  needs a typed code and DELETES NOTHING — no arm carries any of 5 removal
  fields (c). A park needs an owner, a hold reason and a closed trigger, and
  restoring leaves the park visible as superseded (d). A changeless regeneration
  is refused as a retry, and the prior output is preserved (e). Authority is an
  explicit table, staleness binds to the content hash, and a new decision must
  supersede exactly the current head or the history forks (f). The projection
  rebuilds the chain from links so a shuffled log projects identically, surfaces
  superseded records, and exports an audit with exactly one current row (g) | |
  | **I2.12.a–h** | `generate-version-diff.mjs` | `version-diff.json` +
  `version-diff/*.ts` | Isis prompt/version diff across 6 layers — a COMPILING
  reference. PARENT BAR: every change is attributed to exactly one layer, proven
  by fixtures that each move a single thing (brief edit → only brief; moved
  compiler default → only compiled-spec; seed → only execution; policy bump →
  only policy). Cosmetic edits are NOT changes — reflowed prose, a reordered
  keyword list, and 1m vs 1000mm all report unchanged with the reformatting
  stated — while an unknown unit is never assumed convertible (b). Compiled
  values carry all 5 origins so a moved default is attributed to the compiler
  with "no brief field was edited" (c). NO CAUSATION is expressible: no causal
  field, and both generator and suite scan emitted details for 6 causal phrases
  (f). The one certain attribution is the policy-only verdict — identical bytes,
  different verdict — with a deliberately NON-VACUOUS negative case that flips a
  verdict AND changes the artifact and must yield no claim (g). Redaction is a
  STATE not an omission: the row stays, values are withheld, restricted strings
  are scanned for, and the change count is undistorted (d/e). UI/API parity is
  structural — the render model only groups rows the export already carries —
  and all 6 states have distinct announceable labels (h) | | | **I2.13.a–g** |
  `generate-brief-presets.mjs` | `brief-presets.json` + `brief-presets/*.ts` |
  Isis reusable brief presets with inheritance and code floors — a COMPILING
  reference. PARENT BAR: floors are re-applied AFTER EVERY STEP (each preset in
  the chain, then the caller's overrides), because an end-only check can be
  satisfied by a chain whose intermediate state was loose. A leaf that loosens
  all 7 floor domains at once is clamped on every one, as is a user override.
  Floors are DIRECTIONAL — a minimum safety score floors upward while a maximum
  retention floors downward, and the suite proves the two disagree about the
  same number. A clamp is never silent: it records the attempt, the clamped
  value, and WHICH STEP tried it, and that per-step attribution is itself the
  proof floors ran per step (d). Inheritance is single-parent, exactly
  version-pinned with cycle/depth detection, and a stale pin is distinguished
  from a missing preset (b). All 4 merge rules work over scalars/lists/maps, and
  `lock` blocks descendants AND user overrides (c). The preview shows all 5
  value states, with a clamped override reading as FLOORED rather than
  overridden (e). Versions are immutable and a pinned brief resolves identically
  after a new version ships; migration produces a new explicit pin (f). All 4
  scopes x 7 actions are authorized (g) | | | **I2.14.a–g** |
  `generate-portable-brief.mjs` | `portable-brief.json` + `portable-brief/*.ts`
  | Isis portable brief import/export — a COMPILING reference. A versioned
  envelope with manifest, canonical spec, compatibility, provenance summary and
  an integrity hash over the canonical body (a). SECRETS CANNOT TRAVEL BECAUSE
  THERE IS NOWHERE TO PUT THEM: the portable schema declares none of the 8
  secret categories, the internal fixture carries one distinctive string per
  category, and the suite scans the serialized export for every one with a
  negative control; a withheld field is NAMED in the manifest rather than
  vanishing (b). A reference is a declared STATE across all 5 kinds, never a
  link — no url/path/href field exists and the suite asserts no `http` string
  appears (c). Hostile import fails closed on all 8 limits including UNKNOWN
  FIELDS REFUSED not ignored, depth measured on the parsed structure, and a
  decompression bomb refused on its declared ratio before expansion (d).
  Validation covers
  version/integrity/audience/compatibility/rights/references/floors (e); all 6
  preview kinds precede a commit gated BY the preview and by authorization (f).
  Round-trip is byte-semantic against a FROZEN 512-char golden generated from
  the code, keys sorted so the hash cannot depend on insertion order, and an
  omission NEVER becomes a default — key sets are compared across the round-trip
  to prove it (g) | | | **I2.15.a–m** | `generate-continuity.mjs` |
  `continuity.json` + `continuity/*.ts` | Isis continuity profiles across all 9
  dimensions — a COMPILING reference. Every drift metric is a NAMED quantity in
  a stated unit chosen because the obvious one is wrong: colour temperature in
  MIRED not kelvin (the suite proves a 500K step differs in mired between warm
  and cool), intensity/aperture in STOPS with the square relationship so
  f/2√2→f/4 is exactly 1 stop while the rounded f/2.8 marking gives 1.029, CIE76
  ΔE for colour, Jaccard for sets, cosine for identity, and angular drift that
  WRAPS (350°→10° is 20°). The 180-DEGREE RULE is a 2D cross-product sign, with
  a camera starting ON the line correctly not counted as crossing (i). Each
  dimension enforces what it actually gets wrong — consent for likeness
  references (b), wardrobe damage that cannot reverse (c), declared prop
  transitions (d), a mandatory key light (f), palette contrast + intentional
  exceptions (g), a declared aspect matching its sensor (h), a cycle that must
  divide its path (j). UNAVAILABLE IS NOT ZERO: an unmeasured dimension reports
  null value, null uncertainty and `unavailable`, and diagnostics carry the
  unavailable count (l). An override changes STATUS ONLY — 7 guards, and
  revoking restores the original finding (m). PARENT BAR: all 5 consumer kinds
  pin an EXACT revision, published revisions are never overwritten, and unlinked
  consumers are a checkable invariant (k). **I2.15.n deferred**: its bar is an
  accessible end-to-end test of a rendered surface | | | **I2.16.a–m** |
  `generate-conditioning.mjs` | `conditioning.json` + `conditioning/*.ts` | Isis
  conditioning stacks across all 11 kinds — a COMPILING reference. THE
  LOAD-BEARING RULE IS THAT A CONVENTION IS NEVER ASSUMED: a depth and a
  disparity map look identical and are reciprocals, OpenGL and DirectX normals
  differ only by the green channel's sign, and optical flow can point either
  way. All 8 conventions are DECLARED per layer, and a mismatch is either
  converted by one of 9 NAMED transforms recorded in the provenance, or REFUSED.
  The refusals carry the substance — disparity needs a baseline (z = f·B/d has
  no B), normalized depth has no scale to recover, world-space normals need the
  mesh, and REVERSING OPTICAL FLOW IS NOT A NEGATION so it is refused without an
  occlusion mask and marked LOSSY when permitted. Per-kind checks catch what
  each kind actually gets wrong: skeleton keypoint counts, clip inside range,
  edge hysteresis + detector version, a segmentation label refused across a
  vocabulary version change, a camera orientation on tangent normals, style
  rights + declared protected attributes, mask edit history, camera-path bounds,
  and a flow direction contradicting its frame order. IDENTITY CONSENT IS
  REQUIRED BY THE TYPE (h). Multi-control validation covers all 7 stack checks
  with an explicit `max` blend exempted from the conflict rule (l). UI/API
  parity is structural — both project one resolved stack — and the parity
  fixture deliberately carries TWO layers of one kind, since with one per kind a
  UI that dropped rows would still have matched (m) | | | **I2.17.a–l** |
  `generate-sequence.mjs` | `sequence.json` + `sequence/*.ts` | Isis
  shot/storyboard/animatic/sequence authoring — a COMPILING reference. PARENT
  BAR: a patch rerender NEVER widens scope silently. Invalidation propagates
  only along DECLARED edges (explicit dependency, transition handle overlap,
  shared continuity profile), each recording which change reached it and by
  which route: a lens change on 1 of 5 shots invalidates exactly that shot and
  the one built from it, a storyboard note re-renders NOTHING, a continuity
  change never touches geometry passes, and a dissolve neighbour re-renders only
  the overlapping frames (j). Unaffected shots keep their EXACT identity — the
  invariant compares OBJECT identity, so a structurally identical re-mint is
  still caught, because re-minting untouched work voids approvals granted
  against it (k). Drop-frame timecode is implemented against the standard
  identities (107892-frame hour vs 108000 non-drop, 17982 for ten minutes, frame
  1800 = 00:01:00;02), refuses the two dropped labels per minute, exempts every
  tenth minute, and round-trips an hour of frames in both conventions with zero
  failures; retime changes record length without changing the source range, and
  a dissolve genuinely shortens the sequence (c). Rational order keys compared
  by cross-multiplication let two editors insert into one gap without
  renumbering (b). Data passes must be linear-raw since grading a measurement
  corrupts it (i); missing media is surfaced, distinct from a deliberate
  placeholder (g); and the editorial round-trip preserves stable ids, handles
  and media hashes (l) | | | **I2.18.a–m** | `generate-project3d.mjs` |
  `project3d.json` + `project3d/*.ts` | Isis advanced 3D project authoring
  across all 10 stage kinds — a COMPILING reference. PARENT BAR: a stage NEVER
  mutates its input; it writes a new layer over an immutable base, so
  cancel/fail leaves the source and every prior stage untouched — a destructive
  stage would make cancellation a lie. A failed stage may not present an output,
  a late cancel is refused, every stage names actor+tool+version, and
  measurements are nullable so an unfinished stage reports no duration rather
  than zero (l/m). THE GEOMETRY CHECKS ARE EXACT INVARIANTS: V−E+F compared
  against 2·shells−2·genus catches a torus mislabelled genus 0 (applied only to
  closed manifolds); SH counts are (n+1)² so degree 3 needs 16; UDIM is
  1001+u+10v and round-trips over the full grid; skin weights must sum to 1 and
  the refusal says which way it fails — collapse toward vs explode away from the
  origin (g–j). A DATA CHANNEL IS NOT A PICTURE: roughness/normal/metalness
  tagged sRGB is refused because it gamma-shifts every value (e). DELIGHTING IS
  AN ESTIMATE: the original is always retained, overwriting it is refused, and
  an un-evaluated estimate is unusable because its accuracy is unknown — not
  zero (f). An UNSCALED reconstruction is a real state, not 1:1, while lidar is
  metric by construction (b). Packaging checks a declared per-format capability
  matrix and names each consequence (k) | | | **I2.19.a–m, o** |
  `generate-audio.mjs` | `audio.json` + `audio/*.ts` | Isis audio/voice/music
  authoring across 8 clip kinds — a COMPILING reference. LOUDNESS IS MEASURED,
  NEVER ASSUMED: ITU-R BS.1770 as specified (-0.691 offset, LFE weighted ZERO,
  surrounds 1.41) with the TWO-STAGE GATE, verified against the standard's own
  calibration — one weighted channel reads exactly -3.01 LUFS, doubling the
  power adds exactly 3.0103 dB, and gating DISCARDS rather than averages so 90
  silent blocks leave the answer identical to the 10 loud ones alone. A SAMPLE
  PEAK IS NOT A TRUE PEAK: without an oversampled reading (4x minimum)
  compliance is UNKNOWN, and an unknown peak makes the whole verdict unmeasured,
  so nothing ships compliant on a guess (m). CONSENT IS REQUIRED BY THE TYPE and
  REVOCATION BEATS A LIVE TERM — the sharpest case is consent revoked while its
  term still runs, refused as revoked and specifically not as expired (g). STEMS
  ARE A ROUTING VIEW, NOT A BAKE: derived from the bus graph on demand, with a
  track reaching no stem or two stems both caught (j). Timing is in SAMPLES with
  frames and beats derived, so a tempo change cannot silently move a clip (l).
  Per-kind rules catch a looping SFX that ducks at each loop point, foley
  contacts outside the picture, an unreviewed dub, a speaker changing voice
  mid-conversation, and an undeclared overlap. The bundle carries all 9 required
  kinds and cannot ship an unmeasured compliance (o). **I2.19.n deferred**:
  playback/accessibility of a rendered surface | | | **I2.20.a–l** |
  `generate-workflow-graph.mjs` | `workflow-graph.json` + `workflow-graph/*.ts`
  | Isis operator workflow graph authoring — a COMPILING reference. The parent
  bar names three attacks and each has a real defence. GRAPH INJECTION:
  ParameterValue is a closed union of 6 scalar/reference shapes with no
  expression, template, script or object member, and the guard iterates the
  CATALOG's declared parameters — never the payload's keys — closing code,
  nested objects, prototype keys, unknown fields and mass assignment
  structurally; a near-miss spelling is REFUSED as an alias rather than
  resolved, and the hostile fixture is JSON.parse'd so `__proto__` is a real own
  key (l). INCOMPATIBLE NODE: a port carries its full type across 7 axes beyond
  modality — shape, unit, time base, coordinate space, colour space,
  cardinality, optionality — so an sRGB output feeding a linear input is
  refused, and once the modality disagrees the rest is not reported as noise
  (c). SUPPLY CHAIN: the lock is TRANSITIVE or it is nothing — the closure is
  walked to its leaves, a shallow lock is refused, and an entry nothing reaches
  is refused as an unreviewed artifact riding along; every artifact is signed
  over its own bytes and scanned against the digest it locks (e). The dry run
  reaches the provider through a call-counted seam with ZERO dispatches asserted
  on both the clean and failing paths (h); promotion passes through environments
  in order (j); and a rollback changes the selection only, so every past run
  keeps the exact version it consumed (k) | | **I3 — provider execution** |
  **I3.1.a–f** | `generate-executor.mjs` | `executor.json` + `executor/*.ts` |
  The V1 executor port and its composition — a COMPILING reference. The port
  NAMES NO PROVIDER: no client handle, endpoint or options bag, and `execute`
  returns a RESULT rather than throwing because a refusal is an outcome, kept
  distinct from a failure (a). THE DEFAULT IS `not_configured` and stays there:
  only 2 environments may bind a real executor, and the guard is DEFENCE IN
  DEPTH — `register` refuses to bind AND `resolveExecutor` refuses to hand one
  back. The suite exercises the second guard directly against a hand-built
  composition holding a binding, or it would be dead code nothing proves; in
  local/test/dev the call-counted executor records ZERO calls and the
  read-counted secret provider ZERO reads (b/f). CREDENTIALS REMAIN EXTERNAL:
  ExecutionRequest has no credential field, the value is resolved at execution
  time for one call, and the scan looks for actual secret VALUES across
  payload/snapshot/plan/logs — a pattern match would miss a secret that doesn't
  look like one, while a REFERENCE travels freely (c). Revalidation runs before
  the secret is read, so a stale plan never triggers a fetch; all 9 checks are
  exercised (d). The event mapping carries the honesty: a refusal emits NO
  started event, a cancellation reports the usage it consumed rather than
  zeroing it, a refusal reports null usage not zero, and a provider's own
  refusal is never reported as our not_configured (e) | | | **I3.2.a–f** |
  `generate-adapter.mjs` | `adapter.json` + `adapter/*.ts` | Isis RunPod/ComfyUI
  capability negotiation — a COMPILING reference. IDENTITY IS NEVER A DISPLAY
  NAME: a ComfyUI title is user-editable and a custom pack can shadow a core
  node under the same title, so identity is class type + module + version; a
  model's identity is its content hash, not its filename; a worker's capability
  is its MEASURED VRAM (two fixtures share the marketing name `RTX 4090` and
  carry 24GB vs 8GB). The generator asserts none of the 3 display-only fields is
  ever compared or matched against (c). A CACHED CAPABILITY IS LABELLED: on
  timeout/oversize/schema failure the last known state may return, marked
  stale + fromCache + its age, so a caller needing freshness refuses; a document
  that arrives intact but old is stale WITHOUT being from cache (b). PARENT BAR:
  all 15 refusal paths assert ZERO submissions against a call-counted queue seam
  — missing node, module mismatch, wrong model hash, unverified/quarantined
  model, adapter base mismatch, unsupported control/format/transport, exceeded
  limit, insufficient VRAM, all-drained endpoint (d/e). A capability change
  between plan and dispatch forces a REPLAN, never a silent substitution — the
  race is exercised by swapping in a node that shares the planned one's display
  title (f) | | | **I3.3.a–f** | `generate-routing.mjs` | `routing.json` +
  `routing/*.ts` | Isis provider routing — a COMPILING reference. A STALE
  CAPACITY SNAPSHOT IS NOT AVAILABLE CAPACITY: queue depth and free VRAM move
  fastest, so a snapshot past its freshness window is UNKNOWN capacity and
  excludes the endpoint — the fixture proving it deliberately has ample
  headroom, so it is excluded purely on age. Reserving DEBITS the snapshot so a
  second request sees what the first took (c). RESIDENCY BEATS EVERY RANKING
  SIGNAL: the US endpoint is faster, cheaper AND higher priority, and is refused
  anyway, because residency is a legal constraint; half-open is distinguished
  from open as a deliberate probe state (b). Ranking is deterministic across 6
  tiers using the COMPLETE cost including egress — the fixture is cheaper to
  compute and dearer in total, so ranking on the compute rate picks the wrong
  one — with an order-independent tie-break (d). The audit records what was
  CONSIDERED as well as what won, with audience-scoped rejection reasons (e). AT
  DISPATCH ROUTING MAY NEVER SUBSTITUTE ANOTHER REGION OR MODEL: the policy
  chooses fail or replan and there is no third option, asserted at both
  settings, and a blocked dispatch never names the live faster cheaper
  alternative region as a fallback (f) | | | **I3.4.a–f** |
  `generate-breaker.mjs` | `breaker.json` + `breaker/*.ts` | Isis circuit
  breakers and retry/backoff — a COMPILING reference. A BREAKER KEYED PER
  REQUEST NEVER TRIPS: a request id, tenant, idempotency key or prompt digest in
  a scope key gives every call a fresh empty history, so no threshold is ever
  reached; the 7 scope dimensions are a closed union those 4 fields cannot join,
  and the separator-collision fixture is shown to genuinely collide unescaped
  (a). A FAILURE WE CAUSED IS NOT EVIDENCE THE PROVIDER IS UNHEALTHY: counting
  validation errors, policy denials, cancellations or budget exhaustion lets one
  caller open the circuit for every other tenant. Retryability is an ALLOW-LIST
  from the canonical versioned policy, so a class added later counts toward the
  breaker and still earns no retry, and the policy itself is validated for
  unruled, ambiguous and unwatched-but-retryable classes (b). A THRESHOLD
  WITHOUT A MINIMUM SAMPLE OPENS ON THE FIRST FAILURE: the suite holds a circuit
  closed at exactly the threshold rate on four samples and opens it on the
  fifth. Half-open is a FRESH TRIAL, so one counting failure reopens where the
  same failure leaves a closed circuit closed, an inconclusive failure reopens
  nothing, cooldown doubles per consecutive open, an anchor in the future does
  not make a cooldown look finished, and probe slots are held under
  compare-and-set because a lost update lets every caller think it holds the
  only probe (c). NEVER RETRY EARLIER THAN THE PROVIDER ASKED: Retry-After is a
  floor and a hint beyond the cap ends the chain rather than being quietly
  shortened; full jitter comes from an INJECTED source, the curve is capped, and
  the attempt limit and deadline refuse with distinct codes (d). Cancellation,
  policy denial, validation failure, budget exhaustion, an unresumable stage, a
  committed external effect and a missing idempotency key each refuse against a
  CALL-COUNTED seam showing exactly one attempt — an already-open circuit shows
  zero (e). A RESTART MUST NOT RESET AN OPEN CIRCUIT TO CLOSED: the record
  carries state, cooldown and backoff but NOT the rolling window, restore
  re-anchors onto the new monotonic origin, and a backwards wall clock counts as
  no elapsed time; metrics are built from the 10 declared labels while being
  handed the whole context, so a secret-bearing provider body reaches a bounded
  FNV-1a digest and never a label (f) | | | **I3.5.a–g** |
  `generate-failure-taxonomy.mjs` | `failure-taxonomy.json` +
  `failure-taxonomy/*.ts` | Isis failure taxonomy — a COMPILING reference. A
  STABLE CODE IS A CONTRACT; A MESSAGE IS NOT: 30 codes across 8 categories (the
  7 the parent names plus caller input, so nothing falls through to a generic
  bucket), with terminal, retryable and ownership as three independent axes and
  the message-key namespace derived from ownership so a UI never blames a user
  for a GPU that fell over. The validator is pointed at a deliberately broken
  table, so every rule it states is exercised by a violation (a). A PROVIDER'S
  REFUSAL REASON IS THE PROVIDER'S, NOT OURS TO REPUBLISH: the mapping reads the
  status, their code and one allow-listed header — never the body, which in the
  fixture quotes the prompt, names an internal classifier and echoes a bearer
  token; the body is retained behind an incident reference, and an unmapped
  refusal gets its own code that pages (b). A DENIAL WE MADE CITES ITS RULE; A
  REFUSAL THEY MADE DOES NOT PRETEND TO — version, rule id and evidence are
  required rather than optional, and a provider refusal carries a null rule id
  instead of an invented one (c). A QUOTA IS NOT A QUEUE: timing is derived per
  class, and an unknown reset is a null with a stated reason rather than a
  plausible number (d). A TIMEOUT IS NOT A CANCELLATION: 6 interruptions, 6
  codes; a connect timeout sent nothing and is safe to retry, a read timeout was
  delivered and may already be billed so it needs an idempotency key, and a
  reclaimed spot worker does not page while one that vanished does (e). A 200 IS
  NOT A SUCCESS: status, content type, DECLARED size (checked before anything is
  parsed), manifest, required outputs, emptiness and hashes — with the expected
  hash from OUR plan, so the fixture that agrees with its own manifest to the
  character is refused anyway (f). THE CORRELATION ID IS THE ONLY THREAD ACROSS
  THE BOUNDARY and it actually pulls; it identifies a REQUEST, not an
  occurrence, so one request that failed many ways retrieves the whole chain (g)
  | | | **I3.6.a–g** | `generate-durable-jobs.mjs` | `durable-jobs.json` +
  `durable-jobs/*.ts` | Isis durable generation jobs — a COMPILING reference.
  Identity rests on a REAL SHA-256, implemented in full and asserted against the
  FIPS 180-4 vectors plus the 55/56/64-byte padding boundaries and a mixed
  two/three/four-byte UTF-8 string; the canonical form sorts keys so a reordered
  but identical request digests the same (a). AN IDEMPOTENCY KEY THAT IGNORES
  THE INPUTS IS A COLLISION: the same key with a different brief — or a
  different reservation — is a CONFLICT, because returning the earlier job hands
  that caller output made from something it never sent (b). A LEASE WITHOUT A
  FENCE IS A SUGGESTION: a worker stalled in a GC wakes holding a lease that
  looks perfectly valid to itself, so expiry cannot refuse its write and only a
  monotonic fence can. The suite hand-builds exactly that lease, the fence is
  checked BEFORE expiry, and the store applies the same rule to direct appends
  (c). A CHECKPOINT THAT DOES NOT RECORD ITS INPUTS CANNOT BE REUSED: 5
  staleness verdicts, each proven by a plan differing in exactly one thing, and
  an edited record fails its own digest (d). RESUME IS OPT-IN: only a stage
  declared resumable AND idempotent with a still-matching checkpoint picks up; a
  stale checkpoint is DISCARDED, and a stage that already published gets a
  terminal failure because a second attempt would publish twice (e). A LATE
  EVENT MUST NOT REGRESS STATE AND A DUPLICATE MUST NOT BILL TWICE — ordering
  comes from the fence and a per-attempt sequence, and usage is accumulated per
  (stage, attempt, sequence) (f). CURRENT STATE IS A PROJECTION OF AN
  APPEND-ONLY LOG: refused writes never reach the log, replay matches the live
  state at every prefix, and a log with a repeated event is refused rather than
  quietly absorbed (g) | | | **I3.7.a–f** | `generate-cancellation.mjs` |
  `cancellation.json` + `cancellation/*.ts` | Isis cancellation propagation — a
  COMPILING reference. Authorisation is an ALLOW-LIST of 4 of the 5 roles, and a
  cancellation without a reason is refused because it is an action somebody
  took. CANCELLING SOMETHING THAT ALREADY FINISHED IS NOT A FAILURE AND IS ALSO
  NOT A CANCELLATION: the job stays succeeded, stays billed and keeps its
  artifacts (a). THE CANCEL FLAG IS READ INSIDE THE SAME COMPARE-AND-SET THAT
  GRANTS THE CLAIM — the suite runs that race with a worker holding a read from
  before the request — and the mark and the reservation release are ONE step,
  because releasing afterwards leaves a window where the job is cancelled and
  the capacity is still held. A RESERVATION RELEASED TWICE IS CAPACITY INVENTED
  FROM NOTHING, so release is idempotent and says which of the two happened (b).
  "WE ASKED" IS NOT "IT STOPPED": of the 5 things a provider's cancel endpoint
  can have done, exactly one means the work stopped; a provider with no cancel
  endpoint is never signalled and reports UNSUPPORTED, attempts are bounded by
  both a budget and a deadline, and anything short of an acknowledgement settles
  as cancel-failed (c). A LATE ARTIFACT IS EVIDENCE, NOT AN OUTPUT: quarantined,
  retained, typed so it can never be promoted, absent from the outputs, and
  unable to move the job out of cancelled (d). WORK DONE BEFORE THE CANCEL IS
  STILL WORK DONE — consumed units are billed and the remainder released even
  when the provider will not confirm, and until it does the figure is explicitly
  NOT final (e). The projection keeps cancelling, cancelled and cancel-failed
  apart with their own message keys, over an immutable history (f) | | |
  **I3.8.a–f** | `generate-cache-eligibility.mjs` | `cache-eligibility.json` +
  `cache-eligibility/*.ts` | Isis deduplication and cache eligibility — a
  COMPILING reference. A CACHE KEY THAT OMITS ONE OUTPUT-AFFECTING INPUT SERVES
  THE WRONG PICTURE — not a stale one, a different one, confidently. The key is
  built by walking a declared list of 11 components; the suite changes each ONE
  AT A TIME and requires the key to move every single time, and requires the 4
  non-affecting fields never to move it. A golden key pins the algorithm, and
  the digest is the SAME SHA-256 the durable-jobs reference pins to the FIPS
  vectors, because a second hash implementation is a second place to be wrong
  (b). FOUR DIFFERENT QUESTIONS, AND ONE FLAG CANNOT ANSWER THEM: in-flight
  coalescing shares an execution that has not happened yet, which is fine even
  for a nondeterministic workflow, while handing back a PAST output is a
  different promise that a nondeterministic run can never keep (a). A MUTABLE
  DEPENDENCY IS NOT CACHEABLE EVEN IF NOTHING HAS CHANGED YET; all 7 exclusions
  are exercised by fixtures differing in exactly one property (c). A FOLLOWER IS
  NOT AUTHORIZED BY THE LEADER'S PERMISSION, the execution stops when the LAST
  reference drops, and the cost split always adds back up to the total (d).
  "FORBIDDEN" CONFIRMS EXISTENCE; A PRIVATE ENTRY MUST READ AS ABSENT — a lookup
  answers hit or miss and nothing else, and a hidden entry is proven
  indistinguishable from an absent one (e). A CACHE ENTRY WAS AUTHORIZED WHEN IT
  WAS WRITTEN, NOT WHEN IT IS READ: authorisation, expiry, policy, target,
  evaluator, consent, rights and the artifact BYTES are revalidated on every
  lookup, a spent or corrupted entry is invalidated for everyone, and a caller
  running an older policy simply misses rather than emptying the cache (f) |
