Status: canonical change-bound planning, evidence invalidation, selective scheduling, release guard, backlog/SLO metrics, and exercise-cadence contract implemented; V9 signed publication now consumes authenticated, short-lived admissions at release and delivery; no production recertification run or operational evidence is claimed.
Implementation source commit: ac8606760bd6d9ed222a754611b9b8a944f25ab0.
Work-package charter#
- Task ID:
REL-003 - Owner: shared content-coverage and release-control maintainers; implementation owner: Codex
- Accountable product: V1–V9 autonomous-content portfolio
- Priority / milestone: P1 continuous portfolio health control plane
- Risk tier: critical stale-authority, stale-quality, dependency, placement, and recovery-readiness risk
- Dependencies: COV-003 dependency/invalidation graph and ready-frontier scheduler; COV-005 release closure; current signature, approval, benchmark, provider-probe, backlog, and exercise records
- Affected schemas and slots: schema-version-1 continuous-recertification policy, plan, backlog, evidence, exercise, release-input, blocker, metric, and decision contracts; COV-003 node/edge/invalidation vocabulary. No production slot or artifact was declared recertified by this slice.
- Rollout: publish through
@oshun/contracts/content-coverage; a release path must supply its independently computed current input fingerprint, graph revision, affected slots, and live policy-bound records. The evaluator blocks on any omission or mismatch. - Rollback: revert source commit
ac8606760bd6d9ed222a754611b9b8a944f25ab0 - Evidence pack: this file
Requirement disposition#
| Ledger requirement | Implemented control surface | Closure state |
|---|---|---|
| Invalidate affected evidence on source, canon, model, prompt, rubric, rights, consent, policy, provider, code, engine, or placement change | COV-003 now exposes all twelve exact change roots. The plan verifies current change evidence, propagates through blocking edges, revokes approvals, withdraws placement/shipped state, and enumerates every affected prior node/edge evidence reference for stale disposition. | Contract implemented; remains open pending persisted V1–V9 invocation and real invalidation receipts |
| Schedule selective regeneration/review through the coverage graph | The planner projects the invalidation into the next graph revision and invokes the COV-003 ready-frontier scheduler. Only a required leaf with current prerequisites is scheduled; blocked descendants are explicitly deferred. | Contract implemented; remains open pending worker/queue adoption and completed real work |
| Prevent stale signatures, approvals, benchmark scores, or provider probes from releasing changed artifacts | The release guard requires one current passing record of each class, bound to the change event, active policy, next graph revision, current input fingerprint, and exact affected node/slot sets. An independently supplied release target must match. V9 publication consumes the resulting signed admission at release and every delivery. | One repository release path integrated; remains open pending production records and remaining V1–V8 paths |
| Track recertification latency/backlog and block when an SLO expires | Exactly one owner-bound work record is required per affected node. The decision reports total/completed/open/blocked/overdue/late-completed counts and maximum completion latency; open overdue work and policy capacity excess block release. | Contract implemented; remains open pending persisted backlog records, service metrics, and a live SLO |
| Run periodic incident, override, drift, rollback, and recovery exercises | The policy defines a separate cadence for all five exact exercise classes. The guard blocks missing, failed, future-dated, policy-mismatched, or overdue evidence, and resolves equal-time conflict to failure. | Contract implemented; remains open pending five real current exercises and retained operational evidence |
No REL-003 ledger checkbox is closed by this slice. Each requirement is phrased as effective portfolio behavior or operational proof; unit fixtures and a reusable evaluator do not satisfy that completion law.
V9 signed-admission integration — 2026-07-27#
issueContinuousRecertificationAdmission is the only issuer added by this
increment. It first reruns assertContinuousRecertificationReleaseAllowed; a
blocked backlog, stale proof, missing exercise, forged plan, or release-target
mismatch therefore cannot be converted into an admission. A passing decision is
bound to the decision and plan IDs, policy, graph revision, current input
fingerprint, exact affected slots, exact artifact SHA-256, evaluation time, and
a maximum seven-day expiry. The domain-separated receipt and HMAC-SHA256
signature use a fixed newline encoding shared with the Python consumer.
Metis publication requires that admission in SourcePublicationReleaseRequest.
The service authenticates it with a key separate from promotion and publication
signing, requires its release reference and artifact digest to equal the exact
canonical candidate, rejects future or expired decisions, and embeds it in the
immutable signed release. The admission receipt also contributes to the release
ID, allowing a newly recertified release without overwriting the old audit
record. deliver() repeats admission authentication and freshness checks after
loading the immutable archive; an expired release stops serving after process
restart rather than inheriting authority from the earlier publish call.
Focused adversarial coverage proves that blocked lifecycle state cannot issue an admission; altered artifacts, foreign keys, malformed TTLs, and short secrets fail at the issuer; and missing consumer configuration, forged signatures, wrong-artifact receipts, future decisions, expired decisions, and expiry after restart all fail closed. The OpenAPI document and generated TypeScript client carry the new required request and persisted release fields.
Verification on 2026-07-27 UTC passed:
pnpm verify:contracts:continuous-recertification
2 files, 88/88 tests passed; ESLint and full contracts typecheck passed
PYTHONPATH=apps/metis/service/src pytest -q \
apps/metis/service/tests/test_source_publication_service.py
18/18 passed
PYTHONPATH=apps/metis/service/src pytest -q \
apps/metis/service/tests/test_source_*.py
114 passed, 2 expected extraction skips
focused publication/promotion/OpenAPI/export suite
50/50 passed
Ruff, Ruff formatting, strict MyPy, and pnpm metis:contracts:check
passed
Change-bound invalidation and selective scheduling#
libs/contracts/src/content-coverage/dependency-graph.ts extends the canonical
graph with prompt, consent, code, and engine node/edge roots, bringing the exact
invalidation vocabulary to the twelve REL-003 change classes. Change events
cannot predate the graph snapshot, and their evidence must exist, be available,
be observed no later than the event, and remain unexpired.
planContinuousRecertification composes that graph engine rather than creating
a competing dependency model. It:
- proves that the previous and current input fingerprints differ;
- recomputes the complete blocking-edge invalidation plan;
- marks prior evidence attached to the changed root, dependent nodes, and causal edges for invalidation while preserving the new change observation;
- projects affected nodes, approvals, placement, and shipped state into the next graph revision;
- runs the existing required-leaf scheduler over that projection; and
- emits a deterministic plan ID bound to the event, graph revisions, fingerprints, affected slots/nodes/evidence, and scheduling result.
The planner rejects a change with no dependent node, no canonical affected slot, or no prior evidence to invalidate. A caller cannot turn a contextual relationship into a release dependency; propagation remains limited to the blocking graph defined by COV-003.
Fail-closed release decision#
evaluateContinuousRecertificationRelease independently regenerates the plan
from the baseline graph and rejects any forged plan ID, changed effect set, or
changed scheduling result. The release target separately supplies the current
graph ID/revision, input fingerprint, and slot set; matching values repeated
only inside the plan are not trusted as proof of the artifact being released.
Each affected node requires one backlog record bound to the change, policy, current fingerprint, owner, opening time, completion time, and completion evidence. Missing, duplicate, foreign, open, blocked, future-completed, overdue, or excess-capacity work blocks. Completed-after-SLO work remains visible in the metrics even after remediation, preserving the latency breach as an audit fact.
Release also requires exactly one current record for each of:
signature | approval | benchmark-score | provider-probe
Every record binds the change event, policy, graph ID/revision, current input fingerprint, and exact affected node/slot coverage. Pre-change, future-dated, expired, class-SLO-expired, failed, revoked, duplicate, wrong-policy, wrong-revision, wrong-fingerprint, or incomplete-coverage evidence blocks.
The five operational exercise classes are independent:
incident | override | drift | rollback | recovery
Each must have current passing portfolio evidence under the active policy and its own cadence. A future record never substitutes for a current result; when the latest records share a timestamp, a failure wins over a pass.
Automated verification#
Tested source: ac8606760bd6d9ed222a754611b9b8a944f25ab0 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, ESLint 10.4.0, and Nx 22.7.4. The protected
lockfile blob remained a37431842e330b3a06b030947c25bebdc72bc2b1.
Verification ran on 2026-07-18 UTC. Every completed command below exited 0.
pnpm verify:contracts:continuous-recertification
2 files, 85/85 tests passed
targeted ESLint passed
complete contracts TypeScript check passed
pnpm verify:contracts:content-release-closure
1 file, 21/21 tests passed
targeted ESLint passed
complete contracts TypeScript check passed
NX_DAEMON=false pnpm exec nx run @oshun/contracts:build --parallel=1
contracts TypeScript build passed
pnpm exec prettier --check <six changed implementation paths>
passed
git diff --check
passed
The 85 focused tests comprise all twelve change roots plus the prior COV-003 graph/invalidation/scheduling suite. Adversarial cases cover unchanged fingerprints; pre-snapshot events; absent, future, or revoked change evidence; selective leaf/deferred work; missing/duplicate/open/blocked/overdue/foreign backlog; invalid completion; release-target mismatch; forged plans; graph integrity failure; and every required proof and exercise class. Proof cases exercise pre-change, future, expired, freshness-SLO-expired, failed, revoked, duplicate, wrong-policy, wrong-revision, wrong-fingerprint, and incomplete coverage states.
The implementation commit hook's staged ESLint, Prettier, V2 asset, and V5
documentation gates passed. Its broad inferred dependency typecheck was skipped
with the documented SKIP_TYPECHECK=1 control only after the focused verifier,
full contracts typecheck, adjacent release-closure verifier, and contracts build
were green.
The host has 15 GiB RAM and another workspace agent/Nx daemon was active. Per repository resource policy, verification remained targeted, sequential, and single-project. The contracts build peaked while roughly 10 GiB remained available; no full-workspace sweep ran.
Remaining evidence required for closure#
V9 now persists and revalidates the authenticated output of the decision at its release and delivery boundaries. It does not yet persist the underlying plan, advance a production dependency-graph revision, atomically mark enumerated evidence stale, or submit the selective work frontier to a durable worker. V1–V8 release paths do not yet consume admissions. The current seed still has no complete canonical product slot population, so fixtures cannot be relabeled as deployed graph records.
No authorized production signature, approval, benchmark score, provider probe, backlog record, SLO alert, release verdict, placement receipt, or five-exercise portfolio set was available. There was therefore no real human reviewer decision, provider call, runtime placement, incident drill, override drill, drift drill, rollback drill, or recovery drill to record. Those production records, durable control-plane operations, and remaining release integrations remain mandatory before any REL-003 box can be checked.