# AUD-003 evidence pack — reproducible canonical audit verification

## Work-package charter

- **Task IDs:** `AUD-003`
- **Owner:** Oshun portfolio audit-integrity maintainer; implementation owner:
  Codex
- **Accountable product:** V1–V10 autonomous-content portfolio
- **Priority:** P0
- **Target milestone:** 2026-07-15 SOTA gap-closure program
- **Risk tier:** High audit-reproducibility integrity
- **Dependencies:** pnpm lockfile, targeted Vitest configurations, V8 drift
  verifier, Git history, host resource observation, canonical audit Markdown
- **Affected schemas and content slots:** New versioned audit-verification
  manifest and record schemas; audit evidence only, with no shipped content slot
  or customer flow change
- **Rollout strategy:** Commit the command/boundary manifest and tested record
  generator first; run the targeted checks sequentially from that clean source
  commit; commit bounded logs, a machine record, and the generated audit table;
  enforce record/table coherence in CI.
- **Rollback strategy:** Revert the generator, manifest, record, and generated
  table together. Never retain a hand-edited count after removing its evidence
  consumer.
- **Evidence-pack location:** This file

## Completed implementation

1. `scripts/audit/v1-v10-verification-manifest.json` is the versioned command
   and boundary contract. It names every included and excluded package/file, the
   relevant environment gates, and the invalidation paths that make a retained
   record stale.
2. `scripts/audit/run-v1-v10-verification-record.mjs` runs every configuration
   sequentially with Vitest `--maxWorkers=1`, records the clean source commit,
   lockfile, runtime, host, resource observations, command identities, exit
   codes, output byte counts/digests, and bounded excerpts, then generates audit
   section 1.4 through the repository's filename-specific Prettier policy.
3. The verifier rejects manifest/runner drift, unavailable or non-ancestor
   tested commits, wrong source lockfiles, changed tested boundaries,
   substituted commands, red required checks, mutated logs, unformatted records,
   duplicate section markers, and hand-edited audit counts.
4. `.github/workflows/audit-verification-record.yml` runs the verifier with full
   Git history whenever the record, generator, audit, lockfile, or any tested
   boundary changes.

## Published record

- **Tested source commit:** `b856ddd2fe3ff0f7ead8751554c21d7d8a010a20`
- **Run interval:** `2026-07-22T11:13:14.314Z` through
  `2026-07-22T11:21:42.073Z`
- **Record:** `evidence/v1-v9/AUD-003/verification-record.json`
- **Record SHA-256:**
  `df83cb685f3f3f20fcd922b59866fde26f09e29a09ae93d9afd63f5e1aa859c4`
- **Manifest SHA-256:**
  `0f0dbf5a9b891ab60dededaf255e70cc3abbffa0fa459b54742d6df627a516c7`
- **Runner/manifest contract SHA-256:**
  `ec2281c7c638b60aa7ed7c889f81ada87ef2ebab3b72c1d7d921ff889cce452e`
- **pnpm lockfile SHA-256:**
  `f0061e1a5951c82f1f8d5297d08bc6ed2e777107c8d6c35cb87a820c91b71b22`

| Check                  | Authenticated result                                                                                                                        |
| ---------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- |
| V1 creative autonomy   | 284/284 tests passed                                                                                                                        |
| V1 Agentic Studio      | 348/348 tests passed                                                                                                                        |
| V1 studio authoring    | 138/138 tests passed                                                                                                                        |
| V1 BFF autonomy slice  | 66/66 tests passed; the redirected JSON report was required to be freshly rewritten and its 22,841 bytes were hashed                        |
| V8 explicit boundary   | 404/404 tests passed across exactly 18 configurations; `canon-enforcement` remained explicitly excluded                                     |
| V8 docs/contract drift | Passed with valid evidence; the machine-adjacent V8 targets and their generated counterparts agree                                          |
| V9 explicit boundary   | 188/188 tests passed across exactly 13 configurations                                                                                       |
| V10 explicit boundary  | 782/785 tests passed with 3 explicitly pending and 0 failed across exactly 19 configurations: 3 apps, 15 libraries, and canonical contracts |
| Report hygiene         | Both recorded formatting/whitespace commands passed; the generated Markdown and JSON independently pass the locked Prettier configuration   |

All nine checks have committed bounded logs under
`evidence/v1-v9/AUD-003/logs/`. The record authenticates each log by exact byte
length and SHA-256 while retaining full stdout/stderr byte counts and hashes.
The bounded logs total roughly 55 KiB; raw unbounded runner output is not
promoted as durable evidence.

## Recorded environment and claim boundary

- Node.js `v24.15.0`, pnpm `10.25.0`, Vitest `4.1.7`, git `2.43.0`, Linux
  `6.8.0-134-generic` x64, four AMD EPYC-Milan CPUs.
- Supervision observations remained above the repository's safety floors
  throughout the sequential run.
- Secret-bearing gates are recorded by presence only; no secret value is
  persisted. The recorded run had no provider API key or database gate enabled.
- These host and resource values are historical observations at the recorded
  timestamps. They are not timeless repository properties or proof about a
  different host.
- This pack proves the stated targeted unit/integration configurations and the
  passing V8 drift result. It does not claim a broad workspace gate, production
  deployment, provider call, human review, learner efficacy, media quality, or
  certification beyond those boundaries.

## Adversarial coverage

`node --test scripts/audit/run-v1-v10-verification-record.test.mjs` passes eight
tests covering the exact V8/V9/V10 composition, complete Vitest counters,
rejection of prose-only/negative summaries, stale redirected reports,
redirected-report byte authentication, Linux process names containing spaces,
duplicate checks, substituted V9/V10 composition, one-byte bounded-log mutation,
manual table edits, and secret-value non-persistence.

## Verification

- `pnpm generate:audit:v1-v10-verification-record` — passed from the clean
  tested source; all required checks green.
- `pnpm verify:audit:v1-v10-verification-record` — passed; 8/8 adversarial tests
  plus command, boundary, log, lockfile, host, record, and generated-table
  coherence.
- `node --check scripts/audit/run-v1-v10-verification-record.mjs` — passed.
- `pnpm eslint scripts/audit/run-v1-v10-verification-record.mjs scripts/audit/run-v1-v10-verification-record.test.mjs`
  — passed.
- `pnpm prettier --check V1_V9_AUTONOMOUS_CONTENT_SOTA_AUDIT_2026-07-15.md evidence/v1-v9/AUD-003/verification-record.json evidence/v1-v9/AUD-003.md`
  — passed before this closure update and rerun after it.
- `git diff --check` — passed.
