Eve library

AUD-008 evidence pack — adversarial-review reconciliation

Evidence

Revision 1

On this page

Work-package charter#

  • Task ID: AUD-008
  • Owner: Oshun portfolio audit-integrity maintainer; implementation owner: Codex
  • Accountable product: V1–V9 autonomous-content portfolio
  • Priority: P0
  • Target milestone: 2026-07-15 SOTA gap-closure program
  • Risk tier: High canonical-truth and release-claim integrity
  • Dependencies: AUD-001 through AUD-007, the canonical audit, the retained adversarial review, and the authenticated AUD-003 verification record
  • Affected schemas and content slots: New oshun.adversarial-review-reconciliation.v1 audit schema, generated Markdown and HTML disposition views, and canonical audit section 2.3; no shipped customer content slot changed
  • Rollout strategy: Preserve the review as provenance, bind every validated review section to an explicit canonical marker and disposition, prohibit superseded absolute claims, and enforce generated/status/UI drift in CI
  • Rollback strategy: Revert the manifest, schema, generator, generated views, canonical/review edits, package scripts, workflows, and browser test together. Do not restore the review as a competing current truth source.
  • Evidence-pack location: This file and evidence/v1-v9/AUD-008/

Reconciled canonical truth#

  1. scripts/audit/adversarial-review-reconciliation.json enumerates exactly the eleven validated review sections, §3.1 through §3.4 and §4.1 through §4.7. Each has one stable ID, one canonical marker, one review provenance marker, one disposition, and linked closure or follow-up IDs. The current result is seven accepted, four corrected, and zero rejected.
  2. V1 human control remains credited as a real tested API component. The score is explicitly not evidence of end-user operability: repository-wide consumer searches found no in-repository operator UI calling /v1/agentic/autonomy/*. An external private client remains possible but unevidenced by repository or deployed-consumer proof.
  3. CycleReport.realTokenUsage is retained as persisted partial planner-side evidence. Complete generator, research-loop, critic, judge, tool, media-provider, GPU, monetary, artifact-level, feedback, and hard full-path spend accounting remains open.
  4. V3's retained launch record was stale when reviewed, which did not prove the verifier was never run without rewriting evidence. AUD-001 refreshed the record at 2026-07-15T12:30:30.172Z and AUD-002 enforces its 24-hour SLO.
  5. The repository declares a single-Hetzner-box production design. The failure domain follows if deployment matches that design; this audit did not perform a live topology probe and does not promote the declaration to deployed fact.
  6. The canonical audit now includes the coverage-seed, Hathor memoriesTriggered, locale/accessibility, V1 snapshot-policy, test-coherence, host-profile, and verification-reproducibility follow-ups. Open product work remains owned by the linked task IDs rather than being implied complete by reconciliation.

Generated contract and operational view#

  • scripts/audit/generate-adversarial-review-reconciliation.mjs rejects an unknown or duplicate finding, a wrong review-section assignment, missing source markers, malformed disposition/follow-up data, a prohibited superseded claim, or generated drift.
  • evidence/v1-v9/AUD-008/reconciliation-status.generated.md is the compact machine-derived ledger. The HTML sibling presents the same eleven rows and four summary counts with text labels, source links, a keyboard skip link, reduced-motion support, and an internally scrollable mobile table without document-level horizontal overflow.
  • The operational HTML follows the repository's frontend guidance: restrained utility-table composition, one accent, explicit statuses that do not depend on color, no card mosaic, and no marketing treatment.
  • .github/workflows/adversarial-review-reconciliation.yml runs the contract verifier and the focused desktop/mobile Chromium test. The canonical audit verification workflow also runs the reconciliation drift gate.

Publication identity#

  • Reconciliation implementation source commit: 597c63f958a48a70ca96413c15b54ad65537f3da
  • Authenticated record tested source commit: 3bfb0b285e670bbaa25c2995c04494efdf45761d
  • Canonical verification-record publication commit: 977b335bbcb04a409c466c45d8c804e2505dd509
  • Verification-record run interval: 2026-07-15T17:21:30.977Z through 2026-07-15T17:26:12.076Z
  • Host/runtime: dev-server, Linux 6.8.0-110-generic x64, Node.js v24.14.1, pnpm 10.25.0, Vitest 4.1.7, git 2.43.0
  • Provider/model gates: no OpenAI, Anthropic, RunPod, or database credential was present; no real-provider, deployed-client, or live-topology proof is claimed
Artifact SHA-256
scripts/audit/adversarial-review-reconciliation.schema.json 15e32b42eeb29ce3bfd9c4ef4e596a7c120b83a29f567ee12d5c216c66f26f63
scripts/audit/adversarial-review-reconciliation.json 3ee5413c4375da1a693a889e97362e22535c81c2dca2634c946b1169a7de8218
scripts/audit/generate-adversarial-review-reconciliation.mjs 81610fa51b9a1c30e20a7e3b4ab151d4e752beadfe9d244f19dabd0e3f199611
scripts/audit/adversarial-review-reconciliation.test.mjs 2c02f39be88b08e4ebe40dc95c974fc5ade89ef7f8225b01b23dd73f18a2e8ab
.github/workflows/adversarial-review-reconciliation.yml 9022492cb3f05e580a9eb8ed1fee7abb537db0c2d1bfc8f295088b870e0aff71
evidence/v1-v9/AUD-008/reconciliation-status.generated.md 82c31cda916c73a91cbd8bcd98fcaf5fb29473ec16c9b91e36687e8ce04ffbda
evidence/v1-v9/AUD-008/reconciliation-status.generated.html 2a46140aa018265bc1074225e67a0bc43a859b13fd20b5e9cc4bd99f369fac9d
V1_V9_SOTA_AUDIT_REVIEW_2026-07-15.md 43ad38e2f3960fdc06ded8d03e656f67e1302adf6359e5d6712e671a40615295
V1_V9_AUTONOMOUS_CONTENT_SOTA_AUDIT_2026-07-15.md 858708dda86c5592d4dd7aa44881eacc53f3884c5b70550f9b077d9614917a92
evidence/v1-v9/AUD-003/verification-record.json 0356552d4184dec3e27eff215d3f447f70248cda3cf3334708b353aa83ab5689

Automated, adversarial, and visual verification#

  • pnpm verify:audit:review-reconciliation passed 7/7 adversarial tests and workspace drift. The suite proves exact §3.1–§4.7 coverage, rejects wrong section assignment and missing canonical markers, rejects the superseded absolute spend claim, verifies disposition counts and output boundaries, and protects the canonical insertion boundary.
  • JSON Schema Draft 2020-12 validation of the reconciliation manifest passed. Targeted ESLint with --no-ignore --max-warnings=0, targeted Prettier, JavaScript syntax checks, both workflow YAML parses, and git diff --check passed.
  • Focused Playwright ran the generated HTML in Chromium at desktop 1440×900 and Pixel 7 mobile dimensions: 2/2 tests passed. It asserted all eleven ledger rows, 7/4/0/11 counts, four corrected labels, source links, keyboard skip behavior, table-scroll containment, zero document horizontal overflow, and no browser errors.
  • Settled desktop and mobile screenshots were visually inspected. The hierarchy, status labels, contrast, table containment, and narrow-screen layout were readable and intentional; the first animation-state capture was not used as final visual evidence.
  • pnpm generate:audit:v1-v9-verification-record completed sequentially from the clean source commit with Vitest --maxWorkers=1: V1 recorded 248/248, 289/289, 138/138, and 40/40; V8 recorded 351/351; V9 recorded 176/176. The known V8 generated-contract drift remained measured red with exit 1.
  • pnpm verify:audit:v1-v9-verification-record passed 8/8 adversarial tests and record/table/command/boundary/log/source coherence after AUD-008 moved outside AUD-003's generated ownership boundary.
  • pnpm verify:audit:capability-truth passed 7/7 tests and all five generated views; pnpm verify:audit:test-coherence passed 6/6 tests and its generated view; pnpm verify:audit:machine-adjacent-docs passed 3/3 tests and workspace drift; the Unreal closure contract passed.
  • The record run was supervised below 30-second intervals. Available memory stayed near 12 GiB with about 30 GiB free swap; no safety floor was approached.

Claim boundary and inapplicable avenues#

  • Reconciliation closes competing audit truth, not the linked product gaps. Accepted/corrected means the issue is represented accurately in canonical truth; it does not mean COST, HITL, COV, OPS, PRIV, PROMPT, or product work is implemented.
  • Repository search proves only that no in-repository consumer was discovered. It cannot prove that an unobserved external client does not exist.
  • Repository topology declarations establish a design and recovery risk, not a live deployment observation. No production probe, provider call, customer session, learner result, engine execution, or human sign-off is claimed.
  • The generated HTML is audit status UI, not a V1 operator console. Mobile app automation, production E2E, media-quality evaluation, and Unreal execution are outside this documentation-only work package.