Status: canonical contract, fail-closed evaluator, typed verdict evidence, and runtime adversarial failability implemented; product release-path invocation and real R0–R8 evidence remain open.
Implementation source commit: df265d96fbdd339d87f3bee8c0db2599f2261ea2.
Work-package charter#
- Task ID:
REL-001 - Owner: shared release-gate and contract maintainers; implementation owner: Codex
- Accountable product: V1–V9 autonomous-content portfolio
- Priority / milestone: P0 release control plane
- Risk tier: critical release integrity, safety, rights, human authority, and live-canary risk
- Dependencies:
PROOF-001exact non-substitutable proof vocabulary and the existing@oshun/content-release-gatesboundary - Affected schemas and slots: schema-version-1 portfolio evidence bundle, 68 required control records, nine gate verdicts, and nine adversarial fixture verdicts; no shipped content slot or artifact was changed
- Rollout: additive export from the canonical contracts and shared
release-gate packages; product release paths must supply a complete current
bundle and receive
releaseAllowed=truebefore claiming adoption - Rollback: revert source commit
df265d96fbdd339d87f3bee8c0db2599f2261ea2 - Evidence pack: this file
Requirement disposition#
| Requirement | Implemented control surface | Closure state |
|---|---|---|
| R0 Coverage | Six controls: registration, source policy, autonomy policy, quantity, freshness, and runtime placement | Evaluator implemented; remains open pending invoked product evidence |
| R1 Ground truth | Eight controls: source, canon, solver, simulation, factual input, consistency, freshness, and verification | Evaluator implemented; remains open pending invoked product evidence |
| R2 Artifact completeness | Eight controls: modalities, locales, accessibility, dependencies, assets, cook, render, and runtime placement | Evaluator implemented; remains open pending invoked product evidence |
| R3 Quality | Four controls: calibrated floors, uncertainty, disagreement, and human baseline | Evaluator implemented; remains open pending invoked product evidence |
| R4 Safety/integrity | Eight controls: child, crisis, cultural, academic-fairness, game-fairness, injection, abuse, and policy-specific safety | Evaluator implemented; remains open pending invoked product evidence |
| R5 Rights/provenance | Twelve controls: consent, identity, likeness, voice, license, fingerprint, ingredients, actions, watermark, authenticated signature, attribution, and verification | Evaluator implemented; remains open pending invoked product evidence |
| R6 Runtime quality | Nine controls: accessibility, latency, measured cost, performance, compatibility, replay, render, cook, and reliability | Evaluator implemented; remains open pending invoked product evidence |
| R7 Human authority | Six controls: current approval, review window, valid override, approval freshness, takeover, and rollback | Evaluator implemented; remains open pending invoked product evidence |
| R8 Live canary | Seven controls: target outcome, incidents, drift, subgroup behavior, operations, rollback, and recovery | Evaluator implemented; remains open pending invoked product evidence |
| Typed evidence for every verdict | Strict typed source-record, source-query, proof-record, required-control, integrity-error, and adversarial-fixture references; a passing status without a proof blocks | Closed by implementation and tests |
| Gate-specific adversarial failure | Every passing bundle is re-evaluated nine times with one configured control forced to fail; release remains blocked unless each target gate blocks for control-reported-failure while the other eight gates remain passing |
Closed by implementation and tests |
Canonical contract and fail-closed evaluation#
libs/contracts/src/common/portfolio-release-gates.ts owns the R0–R8 and
68-control catalogs, exact required proof type per control, source provenance,
evidence-bundle, verdict, adversarial-fixture, and final-report schemas. The
report schema rejects missing, duplicate, or misassigned gates and controls;
gate/control status contradictions; passing controls without satisfied exact
proof verdicts and typed source/query/proof references; proof verdicts bound to
the wrong subject, policy, or control; mismatched proof-record references;
misconfigured adversarial fixtures; and a top-level status or releaseAllowed
value that contradicts the underlying verdicts.
libs/shared/content-release-gates/src/portfolio-release-gates.ts evaluates the
contract. It blocks on:
- malformed bundles, duplicate controls, or missing required control records;
- failed or unknown control observations;
- future-dated or freshness-SLO-expired source records;
- absent proofs, including a bare positive boolean;
- non-exact proof type, subject, or policy;
- failed, unavailable, revoked, future-dated, or expired proofs;
- proof records that do not link the source evidence; and
- proof verification times that do not equal the source observation time.
Every valid control also passes through the canonical
evaluateProofRequirements policy from PROOF-001. A gate passes only if all
of its controls pass. A malformed input returns typed integrity blocks for all
68 controls and all nine gates; it does not omit a verdict or coerce missing
evidence to success.
Runtime adversarial certification#
A fully passing core bundle is not immediately releasable. The evaluator clones
it once per gate and forces that gate's declared adversarial control to
failed. Each self-test passes only when:
- the target gate becomes blocked;
- the target control reports exactly
control-reported-failure; and - all eight non-target gates remain passing.
Any self-test that cannot produce that isolated intended failure is red and
keeps releaseAllowed=false. Core-blocked or integrity-blocked bundles report
all fixtures as not-run; they never use an adversarial result to disguise a
primary evidence failure.
Automated verification#
Tested source: df265d96fbdd339d87f3bee8c0db2599f2261ea2 on host dev-server,
Linux 6.8.0-110-generic x86_64. Toolchain: Node 24.14.1, pnpm 10.25.0,
TypeScript 6.0.3, Vitest 4.1.7, and ESLint 10.4.0. The protected lockfile blob
remained a37431842e330b3a06b030947c25bebdc72bc2b1.
Verification ran on 2026-07-17 UTC. The final contract run began at 22:35:06 and the final shared-package run began at 22:35:17; every command below exited 0.
pnpm exec vitest run --config libs/contracts/vitest.config.ts src/common/proof-types.test.ts src/common/portfolio-release-gates.test.ts
2 files, 14/14 tests passed
pnpm exec vitest run libs/shared/content-release-gates/src
3 files, 47/47 tests passed
portfolio evaluator: 26/26
existing shared release/consent regression: 21/21
pnpm exec tsc --noEmit -p libs/contracts/tsconfig.lib.json
passed
pnpm exec tsc --noEmit -p libs/shared/content-release-gates/tsconfig.json
passed
pnpm exec eslint <changed contract/shared source and test paths>
zero errors; repository lint config ignores test files
git diff --check
passed
The 26 evaluator cases include a complete 68-control pass, one intended failure for each of R0–R8, failed/unknown controls, absent proof, missing control, future/stale sources, wrong proof type/subject/policy, non-passing/future/expired proofs, unlinked proof evidence, mismatched observation time, duplicate and malformed bundles, and invalid evaluation time. Contract cases assert the exact 68-control catalog and reject forged report structure and release state.
The commit hook's staged ESLint, Prettier, V2 asset, and V5 documentation gates
passed. Its broad dependency typecheck was bypassed with the documented
SKIP_TYPECHECK=1 control only after both task-owning packages had passed
direct typecheck and their complete relevant tests. This host has 15 GiB RAM,
below the repository's 32 GiB gate for an unapproved broad Nx sweep.
Inapplicable verification and remaining evidence#
No customer UI, API route, persistence schema, content artifact, model/provider,
media pipeline, or runtime placement changed. Playwright, accessibility,
screenshots, database migrations, provider probes, and engine checks are
therefore inapplicable to this additive contract/evaluator slice. No human
release decision was required to add the boundary; the evaluator nevertheless
requires exact human_approval and observed takeover/rollback evidence for R7.
The nine R0–R8 ledger boxes remain open under the ledger's completion law. No V1–V9 production release path currently supplies this 68-control bundle, no deployed release verdict or placement receipt invokes it, and no real current source/canon/solver/simulation/rights/safety/runtime/human/canary evidence set was available. Unit, contract, and adversarial fixtures prove enforcement behavior, not live portfolio conformance. Product release integrations must fail closed on this report and publish real freshness-bounded evidence before those boxes can be checked.