Status: canonical five-level runtime behavior and V1 creative-publication enforcement implemented; production slot/evidence bindings remain fail-closed and the remaining portfolio requirements are not yet evidenced.
Implementation source commit: a3211ff29a4503cfa7862cd0c042f6a6fe2da901.
Work-package charter#
- Task ID:
REL-002 - Owner: shared autonomy-contract and Oshun creative-runtime maintainers; implementation owner: Codex
- Accountable product: V1–V9 autonomous-content portfolio
- Priority / milestone: P0 risk-tiered autonomy control plane
- Risk tier: critical human-authority, safety, rights, spend, and external side-effect risk
- Dependencies: content-slot registry, current evidence/proof records, human authority, creative review queue, and operator intervention API
- Affected schemas and slots: schema-version-1 autonomy scope, evidence, action-decision, and promotion-decision contracts; creative category and direction autonomy values; intervention and operator DTOs; generated coverage registry/inventory. No production slot was declared conformant by this slice.
- Rollout: legacy persisted
human-approvalvalues normalize toapproval; new values use the canonical ladder. Every V1 creative-director publication path now invokes the evaluator, but deployment must supply real slot-bound evidence before the director can run or policy adoption can be claimed. - Rollback: revert source commit
a3211ff29a4503cfa7862cd0c042f6a6fe2da901 - Evidence pack: this file
Requirement disposition#
| Ledger requirement | Implemented control surface | Closure state |
|---|---|---|
| Support Manual, Copilot, Approval, Review Window, and Full Auto | One canonical ordered vocabulary is used by the contracts, creative runtime, Agentic Studio intervention API, BFF route, operator DTO, and coverage artifacts. Runtime tests exercise all five levels. | Closed by implementation and tests |
| Bind autonomy to slot, action, risk, tenant, audience, and current evidence | evaluateAutonomyAction requires a canonical slot ID, tenant, audience policy, policy, risk tier, action, evaluation time, and freshness-bounded scoped evidence. The V1 director now resolves those inputs per category for every publication attempt. |
Release boundary integrated; remains open pending real production scopes/evidence |
| Require mandatory human/institutional authority for the ten enumerated risks | Exact authority evidence is required for consequential actions. Institutional authority covers named claims, minors, crisis/health, and scholarship; human authority covers identity, cloned voice, high-value spend, legal rights, and irreversible external effects. | Release boundary integrated; remains open pending real risk tagging and authority evidence |
| Lower autonomy automatically on regression | Reliability, cost, and provider regressions cap at Approval; safety caps at Manual; drift and human-outcome regressions cap at Copilot. The V1 adapter durably emits every automatic lowering before denying the current release. | Controller seam integrated; remains open pending a persisted production sink/evidence feed |
| Raise only after repeated-trial and human evidence; prohibit unsafe local overrides | Promotion is one level at a time and requires all six current health records, a predeclared repeated trial meeting its minimum count, and current human evidence. Policy and risk ceilings remain non-overridable. | Evaluator implemented; remains open pending live promotion-path invocation and evidence |
Only the first ledger checkbox is closed. The remaining four describe effective portfolio enforcement, not merely a reusable evaluator, and therefore remain open under the ledger completion law.
Canonical contract#
libs/contracts/src/content-coverage/autonomy-ladder.ts defines the exact
ordered levels:
manual < copilot < approval < review-window < full-auto
An action decision is bound to the canonical content-slot ID, tenant reference, audience-policy reference, autonomy-policy reference, risk tier, concrete action, evaluation time, and current evidence. Requesting actor identity is retained for audit but cannot change the verdict. The action vocabulary covers validation, research, suggestion, transformation of a human draft, selected node generation, candidate generation, release, spend, external effects, and policy change.
The evaluator applies these fail-closed controls:
- low, moderate, and high/critical risk cap autonomy at Full Auto, Review Window, and Approval respectively;
- the ten mandatory authority risks cap consequential actions at Approval and require the exact human or institutional evidence class;
- Review Window and Full Auto require current passing reliability, safety, cost, drift, provider, and human-outcome health evidence;
- an actual regression can lower a configured level even when that level would otherwise need no health bundle;
- equal-timestamp conflicting health evidence resolves to the non-passing observation, never whichever record happens to appear first;
- Approval release requires exact approval evidence; Review Window release requires explicit approval or an expired unvetoed window; Full Auto release requires current health evidence;
- Copilot transformation and selected-node generation require a human-owned draft; and
- a local action override can only retain or lower configured authority.
evaluateAutonomyPromotion permits exactly one-level promotion and requires the
six current health evidence classes, a predeclared repeated-trial record meeting
policy count, and current human evidence. Mandatory authority and configured
policy ceilings still apply. Contract tests reject stale, mismatched,
conflicting, incomplete, over-broad, and unsafe promotion inputs.
Live creative-runtime semantics#
The Oshun creative director, queue, steering layer, and publication retry path now use the same five values:
| Effective level | Enforced creative behavior |
|---|---|
| Manual | Excluded from agent ideation/production; validation-only disposition; no producer call, candidate, queue submission, or spend |
| Copilot | May derive/select assistance from a human-owned draft; assist-only disposition; no producer call, complete candidate, release gate, queue submission, or spend |
| Approval | Complete candidate enters review and can publish only after explicit human approval |
| Review Window | Complete candidate enters a veto window and can publish after explicit approval or an expired unvetoed window |
| Full Auto | Complete candidate may enter the automatic release path, subject to the existing release gates |
Approved and retried items are re-authorized against the category's current
autonomy at publication time through evaluateAutonomyAction. This single
publication boundary covers the approval loop, review-window expiry, full-auto
release, and retries. It resolves the category's canonical slot/action scope,
configured and local-override levels, current health/risk evidence, and
approval/window evidence. Missing or malformed scope/evidence fails closed.
Every automatic verdict below the explicit local/policy-capped baseline is sent
to the binding's durable lowering sink before the current release is denied; an
intentional local lowering is not misrecorded as a regression.
Lowering a category to Manual or Copilot blocks publication; Approval requires
an explicit approve transition; Review Window accepts explicit approval or
window-expired-auto-approve; and Full Auto requires all six current passing
health records. Policy-blocked items remain approved and expose a publish
failure instead of disappearing or being marked released. Assist-only
dispositions neither consume a pending tweak nor record production spend.
The queue independently rejects complete agent-produced candidates at Manual or Copilot, so a caller cannot bypass the director. Missing signal, category, champion, or runtime policy data fails loudly instead of relying on unsafe non-null assumptions. The BFF constructs no director unless the canonical authority binding is present alongside its producer, publisher, judge, semantic, calibration, and human-launch bindings.
Migration and operator surface#
Durable creative category, direction, and queue restoration normalizes the
legacy human-approval value to canonical approval. Seed registry schema v1
continues to accept that legacy source spelling for backward compatibility, but
generated output and all current sources use canonical values.
Agentic Studio's intervention workspace, the BFF autonomy route, and the web operator DTO expose the same five-level vocabulary. The BFF validates runtime inputs, permits at most a one-level raise, and rejects unknown strings. The operator console kept its existing restrained utility layout; the mandatory frontend workflow guided this slice toward precise operator vocabulary and Playwright coverage rather than an unrelated visual redesign.
The coverage seed registry and inventory were regenerated. Their larger diff also repairs stale source hashes caused by source changes already present in the workspace; the JSON was generator-produced, not hand-edited.
Automated verification#
Tested source: a3211ff29a4503cfa7862cd0c042f6a6fe2da901 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 Playwright 1.60.0. The
protected lockfile blob remained a37431842e330b3a06b030947c25bebdc72bc2b1.
Verification ran on 2026-07-17 UTC. Every command below exited 0.
pnpm verify:contracts:autonomy-ladder
29/29 tests passed; contract TypeScript and ESLint passed
pnpm verify:content-coverage:seed-registry
10/10 tests passed; deterministic output check, script/contracts TypeScript,
and ESLint passed
pnpm verify:content-coverage:inventory
10/10 tests passed; deterministic output check, script/contracts TypeScript,
and ESLint passed
(cd libs/oshun/creative-autonomy &&
pnpm exec vitest run --config vitest.config.ts --maxWorkers=1)
18 files, 274/274 tests passed
(cd libs/oshun/creative-autonomy && pnpm exec tsc --noEmit)
passed
(cd libs/oshun/agentic-studio &&
pnpm exec vitest run --config vitest.config.ts --maxWorkers=1)
26 files, 348/348 tests passed
(cd libs/oshun/agentic-studio &&
pnpm exec vitest run --config vitest.config.ts \
src/interventions/intervention-api.test.ts --maxWorkers=1)
1 file, 20/20 tests passed
(cd libs/oshun/agentic-studio && pnpm exec tsc --noEmit)
passed
pnpm exec vitest run --config apps/oshun/bff/vitest.config.ts \
apps/oshun/bff/src/agentic/autonomy-route.test.ts \
apps/oshun/bff/src/agentic/autonomy-bindings/autonomy-bindings.test.ts \
apps/oshun/bff/src/agentic/autonomy-bindings/plane-persistence-privacy.test.ts \
apps/oshun/bff/src/agentic/autonomy-bindings/autonomy-plane-e2e.test.ts \
--maxWorkers=1
4 files, 54/54 tests passed
pnpm --filter @oshun/bff typecheck:autonomy
passed
pnpm --filter @oshun/bff build
esbuild production bundle completed successfully
pnpm --filter @oshun/web typecheck
passed
(cd apps/oshun/web && E2E_WEB_ONLY=1 pnpm e2e -- \
operator-autonomy-console.spec.ts --project=chromium --workers=1)
5/5 tests passed, including keyboard/focus and mobile-overflow coverage
pnpm exec eslint <changed non-web source and test paths> --max-warnings=0
passed
(cd apps/oshun/web &&
pnpm exec eslint <changed web source and test paths> --max-warnings=0)
passed
git diff --check
passed
The commit hook's staged ESLint, Prettier, V2 asset, and V5 documentation gates
also passed. Its broad dependency typecheck was bypassed with the documented
SKIP_TYPECHECK=1 control only after every task-owning package had passed its
targeted typecheck and relevant test suite.
The host has 15 GiB RAM and was also running another workspace agent/Nx daemon. Per repository resource policy, verification was sequential and targeted; no unapproved broad Nx sweep or Next production build ran. The affected web path was instead covered by its package typecheck and real Chromium Playwright flow.
V1 publication-boundary integration verification#
The integration in the task-scoped commit containing this evidence was verified on 2026-07-27 UTC on the same 15 GiB host. Checks ran sequentially with one Vitest worker.
pnpm exec vitest run --config libs/contracts/vitest.config.ts \
libs/contracts/src/content-coverage/autonomy-ladder.spec.ts --maxWorkers=1
1 file, 29/29 tests passed
pnpm exec vitest run --config \
libs/oshun/creative-autonomy/vitest.config.ts \
libs/oshun/creative-autonomy/src/director/autonomy-authority.spec.ts \
libs/oshun/creative-autonomy/src/director/director.test.ts --maxWorkers=1
2 files, 46/46 tests passed
(cd libs/oshun/creative-autonomy &&
pnpm exec vitest run --config vitest.config.ts --maxWorkers=1)
19 files, 289/289 tests passed
pnpm exec vitest run --config apps/oshun/bff/vitest.config.ts \
apps/oshun/bff/src/agentic/autonomy-route.test.ts \
apps/oshun/bff/src/agentic/autonomy-bindings/autonomy-plane-e2e.test.ts \
apps/oshun/bff/src/agentic/autonomy-bindings/autonomy-bindings.test.ts \
--maxWorkers=1
3 files, 56/56 tests passed
pnpm exec tsc --noEmit -p libs/contracts/tsconfig.lib.json
passed
pnpm exec tsc --noEmit -p libs/oshun/creative-autonomy/tsconfig.json
passed
pnpm --filter @oshun/bff typecheck:autonomy remains blocked by a pre-existing
unrelated shared-tracing error at libs/shared/tracing/src/genai.ts:365
("v7-flagship-revision" is not assignable to the narrower tracing version
type). None of this task's files caused that diagnostic. The focused BFF
compile-through Vitest suites above passed. No user-facing UI changed, so
Playwright and the frontend workflow were not applicable to this backend
release-authority slice.
Remaining evidence required for the open requirements#
The V1 creative-director release boundary now invokes the canonical evaluator,
but the four current scheduled categories do not yet have approved production
slot scopes supplying canonical slot IDs, tenant and audience policy references,
risk tagging, policy references, and current evidence bundles. The server
therefore does not bind autonomyAuthority, and the director correctly remains
null rather than publishing without evidence. The live BFF raise route still
relies on its existing human-authority SLO drills and reviews, not this
contract's repeated-trial plus six-health-record promotion bundle.
No deployed content action currently supplies all mandatory risk classifications and exact current authority evidence to the new boundary. Regression observations are not connected to a production evidence feed or persisted live auto-lowering store. Consequently there is no deployed verdict, production placement receipt, provider run, human cohort, or continuous policy-state evidence to close the remaining four boxes. Database migrations, provider probes, media/Unreal checks, and live rollout evidence were inapplicable to this contract/runtime slice. Those integrations and their real freshness-bounded evidence remain required before portfolio enforcement can be claimed.