# AUD-006 evidence pack — generated capability truth

## Work-package charter

- **Task ID:** `AUD-006`
- **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 portfolio-truth and release-claim integrity
- **Dependencies:** AUD-002 evidence freshness, AUD-003 canonical verification
  record, AUD-004 Unreal observation semantics, AUD-005 enforcement comments, V3
  launch release record, repository proof sources
- **Affected schemas and content slots:** New `oshun.capability-truth.v1`
  registry schema and derived capability/status documents. This is the
  capability/status seed; COV-001 through COV-005 still own the full
  content-slot inventory and release controller.
- **Rollout strategy:** Seed conservative product/capability states from typed
  evidence; generate every managed status view; reject claim promotion and
  generated drift in CI; retain missing provider, placement, and release records
  as negative evidence.
- **Rollback strategy:** Revert the schema, registry, generator, managed views,
  audit summary, package scripts, and both workflow changes together. Never
  restore hand-maintained green prose after removing the machine source.
- **Evidence-pack location:** This file and `evidence/v1-v9/AUD-006/`

## Canonical machine truth

### 2026-07-22 scope amendment

The active registry now covers all ten V1–V10 products, 22 capability rows, and
38 typed evidence sources. The most recently published canonical runner is
`pnpm generate:audit:v1-v10-verification-record`; its verifier is
`pnpm verify:audit:v1-v10-verification-record`. That authenticated record
captured V1 at 284/284, 348/348, 138/138, and 66/66; V8 at 404/404 across 18
configurations with its docs/contract drift check passing; V9 at 188/188 across
13 configurations; and V10 at 782/785 with 3 pending and 0 failed across 19
configurations. Its freshness-limited results are not represented as current.
Publication identities and counts below retain the initial 2026-07-15 evidence
window as history.

1. `scripts/audit/capability-truth-registry.json` is the designated registry;
   `evidence/v1-v9`, `V3/launch`, and the RunPod probe-artifact location are its
   declared evidence roots. The versioned JSON schema rejects unknown weakening
   fields.
2. The registry contains all ten products, 22 capability rows, 38 typed evidence
   sources, explicit claim states, evidence references, proving boundaries, and
   manual analysis only in fields named `commentary`.
3. The evidence adapters cover every required contradiction input:
   `source-comment`, `todo-ledger`, `test-count`, `provider-probe`, `artifact`,
   `placement-receipt`, and `release-record`.
4. `implemented` requires clean passing repository evidence. `live` additionally
   requires a passing provider probe and placement receipt. `green` additionally
   requires a passing release record and rejects missing, seam-only, blocked,
   red, or stale evidence.
5. Test-count evidence has a seven-day SLO; the V3 launch release record has its
   existing 24-hour SLO. Aged evidence becomes `stale`, and future-dated
   evidence is invalid.

## Generated views and current result

- `V_SERIES.md` and `CONTENT_GENERATION_CAPABILITY_MATRIX.md` are now wholly
  generated status views rather than parallel hand-maintained truth.
- `evidence/v1-v9/AUD-006/product-status.generated.md` publishes the V1–V10
  evidence/state table.
- `evidence/v1-v9/AUD-006/audit-summary.generated.md` publishes claim counts and
  all 38 observed contradiction inputs.
- Canonical audit section 2.1 is generated from the same registry and points to
  both detailed views.
- Current product result is deliberately non-green: three `seam-only`, five
  `blocked`, one `red`, and one `stale`. Current capability result is nine
  `implemented`, six `seam-only`, three `blocked`, and four `missing`.

### 2026-08-02 recertification refresh

- The first canonical drift check against source baseline
  `681ff61b91ad5c057d90fee13e3e489a3e13f978` failed because V8 still claimed
  `implemented` after its AUD-003 test and drift evidence exceeded the enforced
  seven-day freshness SLO. The registry now reports V8 as `stale`; no passing
  result was extended by prose.
- Current source-bound evidence replaces the historical 2026-07-15 host
  observation in every active claim: V2 reports its exact stale editor-build
  preflight, V3 reports its missing editor-module build blocker, V4 and V5
  report their recent authenticated Unreal executions, and V7 reports its
  cooked-region runtime-placement receipt. The old AUD-004 record remains
  available only as historical contradiction input.
- The V7 receipt supports the bounded level-placement capability for that exact
  governed region. It does not promote the missing generic Bellona authoring
  execution, live providers, human review, or release evidence.
- The broad AUD-003 verification record was not regenerated on this host. The
  required resource gate found a competing verification worker using about 7 GiB
  RSS on a 15 GiB host, so expired broad evidence remains explicitly `stale` and
  this refresh uses narrow audit-integrity checks only.
- After explicit approval for the product-scoped broad gate, the V3 launch
  verifier executed all 41 declared constituent commands against published
  source `fabc9ddc30`. The fresh record remains honestly red at 30/41 commands
  and 24/39 named gates with `gaReady: false`; 11 constituent failures and 39
  waived local tasks still require external certification hardware, deployed
  fleets and stores, staffed operations, live catalogs, or measured production
  thresholds. The global generated-evidence freshness inventory now passes
  without promoting V3.
- `pnpm verify:audit:v1-v10-verification-record` still passes its eight
  adversarial unit tests, then correctly rejects the retained record because
  authenticated source boundaries changed after its tested commit. Its generated
  audit renderer also retains a static historical V8-red sentence despite the
  record's later passing drift verdict; correcting that evidence-bound runner
  and section requires the deferred full recertification rather than a hand
  edit.
- The targeted ISIS producer run regenerated all 416 certification records
  sequentially against source commit `08b0aa047d0b424bf7dbad8fffc21da9a4a810c0`.
  At that boundary their semantic payloads and verdicts did not change: 358
  remained passing and 58 remained honestly failed. A subsequent SDK contract
  parity repair preserved conditional control-map parameters in four generated
  SDK reports, and 13 affected Matrix C certifications were regenerated against
  implementation merge `8ee7cf891bb0868f0235eff8763a31d92a83cbc4` with all five
  surfaces and all negative regression scenarios passing. A second SDK parity
  repair preserved pose/depth/segmentation maps and NPC control strengths in
  four additional reports; four dependent certifications were regenerated
  against implementation merge `0d4460ae3a7f7b0e830b0244a0b1d9f2086e1964` and
  passed. Prop-pack-general implementation `8a0590f12e` then added typed,
  budgeted, quality-gated, SDK, native web, and runtime-graph exposure for pose,
  normal, and lineart controls; certification-title validator correction
  `b01704cdc4` removed the copied character-sheet label. `MXC-0004`, `MXC-0052`,
  `MXC-0084`, and `MXC-0148` now pass all five delivery surfaces and every
  negative regression scenario against that final source tip. Architecture
  kitbash implementation `6fd96937c5`, published through implementation merge
  `cf21c781fb`, then added the same typed, budgeted, quality-gated, SDK, native
  web, RunPod, and six-lane runtime-graph coverage while preserving the existing
  mask-region guide. `MXC-0016`, `MXC-0064`, `MXC-0096`, and `MXC-0160` now pass
  all five delivery surfaces and every negative regression scenario against that
  remote-backed source tip. Environment-layout implementation `99c5a4fea4`,
  published through merge `0ab8b446f4`, added typed, budgeted, quality-gated,
  SDK, native web, RunPod, and six-lane runtime-graph exposure for pose, normal,
  canny, and lineart controls. The environment pose certificate title and its
  validator were corrected in `9c3aef3f0f` before publication. `MXC-0003`,
  `MXC-0035`, `MXC-0051`, and `MXC-0083` now pass all five delivery surfaces and
  every negative regression scenario against that remote-backed source tip.
  Vehicle-turntable implementation `72f8bb56dd` added typed, budgeted,
  quality-gated, SDK, native web, and six-lane runtime-graph exposure for pose,
  normal, and lineart controls, and corrected the copied vehicle-pose
  certificate title before publication. `MXC-0005`, `MXC-0053`, and `MXC-0085`
  now pass all five delivery surfaces and every negative regression scenario
  against that remote-backed source tip. Material PBR implementation
  `1b8f65a148`, published through merge `3d53013ece`, added typed, budgeted,
  quality-gated, SDK, native web, RunPod, and six-lane runtime-graph exposure
  for pose, normal, canny, and lineart texture-generation controls. It also
  corrected the copied pose-certificate title before publication. `MXC-0013`,
  `MXC-0045`, `MXC-0061`, and `MXC-0093` now pass all five delivery surfaces and
  every negative regression scenario against that remote-backed source tip.
  Character-sheet implementation `52578366a4`, published through merge
  `1ca80f2e4d`, then added typed, budgeted, quality-gated, SDK, native web,
  RunPod, and six-lane runtime-graph exposure for normal, canny, and lineart
  controls; all six control strengths are explicit across the API, SDK, UI, and
  graph input contract. The canonical requirement coverage also now includes the
  existing image-conditioned IP-Adapter style reference path. `MXC-0033`,
  `MXC-0049`, `MXC-0081`, and `MXC-0097` pass all five delivery surfaces and
  every negative regression scenario against that remote-backed source tip. NPC
  crowd implementation `17b3191a01`, published through merge `e3ae0c0129`, added
  typed, budgeted, quality-gated, SDK, native web, RunPod, and six-lane
  runtime-graph exposure for normal, canny, and lineart controls while retaining
  the family-specific mask-region guide. `MXC-0047`, `MXC-0063`, and `MXC-0095`
  pass all five delivery surfaces and every negative regression scenario against
  that remote-backed source tip. Terrain-heightmap implementation `38d26f94f2`,
  published through merge `648e72311d`, added typed, budgeted, quality-gated,
  SDK, native web, RunPod, and six-lane runtime-graph exposure for pose, normal,
  canny, and lineart reconstruction controls while retaining depth,
  segmentation, and the family-specific mask-region guide. Certification-title
  correction `de9ba33c03` also hardened the pose validator against copied target
  headings. `MXC-0012`, `MXC-0060`, `MXC-0092`, and `MXC-0156` pass all five
  delivery surfaces and every negative regression scenario against that
  remote-backed source tip. VFX control-intent correction `ac3ec5be6a` then
  reconciled the canonical requirement and all 18 derived family reports with
  the already implemented pose API, SDK, native UI, runtime graph, RunPod, and
  E2E surfaces. `MXC-0008` passes all five delivery surfaces and every negative
  regression scenario against that remote-backed source tip. UI-icon
  control-intent correction `6c6b1d66b6` likewise reconciled the canonical
  requirement and all 18 derived family reports with the existing pose API, SDK,
  native UI, runtime graph, RunPod, and E2E surfaces. `MXC-0009` passes all five
  delivery surfaces and every negative regression scenario against that
  remote-backed source tip. Final normal-control implementation `044817a17a`,
  published through merge `c4b8b1f003`, added typed, budgeted, quality-gated,
  SDK, native web, RunPod, and six-lane runtime-graph exposure for storyboard,
  cinematic, marketing-art, social-cutdown, and audio-reactive families. The
  five-family Playwright matrix proves normal-only validation and submission for
  every affected workflow. `MXC-0086`, `MXC-0087`, `MXC-0090`, `MXC-0091`, and
  `MXC-0094` pass all five delivery surfaces and every negative regression
  scenario against that remote-backed source tip. The current ISIS portfolio is
  therefore 416 passing and 0 failed; all corresponding `ISIS_TODOS_2.md` claims
  are closed from generated evidence rather than prose assertion.
  `pnpm verify:evidence:generated-gates:isis` authenticates all 416 fresh
  records, enforces exact checklist/verdict alignment, and proves agreement with
  every legacy validator. The default global inventory still rejects exactly the
  expired V3 launch-readiness record; it is not hidden or refreshed by the
  ISIS-only scope.
- Focused recertification passed the eleven capability-registry tests, exact
  five-view generation drift check, Unreal closure contract, and V2 preflight
  record verification. Targeted ESLint covered the generator, registry tests,
  and Unreal verifier; targeted Prettier and `git diff --check` covered every
  changed file. The Unreal verifier now normalizes whitespace before checking
  required prose, so canonical Markdown wrapping cannot create a false failure.

## Historical 2026-07-15 publication identity

- **Canonical registry/generator implementation commits:**
  `5ff2bef37ac51a949a405dafa1dcbea8a1440789` and
  `fbc29f40f565b8561d17e91544c7f0b91f6d5839`; AUD-007 generated-section
  compatibility commit `d76869b2ab861c4958f89c7447c2236c1ea7920f`; AUD-008 V1
  commentary qualification commit `597c63f958a48a70ca96413c15b54ad65537f3da`
- **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 live provider call is claimed.

| Artifact                                             | SHA-256                                                            |
| ---------------------------------------------------- | ------------------------------------------------------------------ |
| `scripts/audit/capability-truth.schema.json`         | `68ef9c5a9ebec997e726b9d3ae2bc85fce76b1140602c5363e6d127e842f9bee` |
| `scripts/audit/capability-truth-registry.json`       | `4a817d6f272d0434be48411b8ef9cc31fbcdc6ee076732136f172f6b36a4dfa5` |
| `scripts/audit/generate-capability-truth.mjs`        | `8f54ebca63ff97e36b04da7ca1f839f47000c88e25643ff66ce449fd3e708bdc` |
| `scripts/audit/capability-truth-registry.test.mjs`   | `4ba09baf2c6324fe31bfe976dd3d87d6d0cb0e0a52d77d2de2ca773e6891e6f1` |
| `.github/workflows/capability-truth-drift.yml`       | `8e9df2eaaacdd5a770566d5568ace5ee06e0854b4edb94550c6acd8ed7276b90` |
| `V_SERIES.md`                                        | `809cae149a0fe78775c5d235d66b32e380d3954440401cf3b35320e248924900` |
| `CONTENT_GENERATION_CAPABILITY_MATRIX.md`            | `ebb70997311166b542c6718043d69dec77f0d3bbf31c10a2989ea12cdf2dbe2c` |
| `evidence/v1-v9/AUD-006/product-status.generated.md` | `57ae4d8d3240d77cef346bcf33252d58deded5a5488fa114d536a553637c56bc` |
| `evidence/v1-v9/AUD-006/audit-summary.generated.md`  | `d495720265b83edad4bf1ee791b41c1ccdbcebaa2d938e42615e73b721fa5b92` |

## Adversarial and automated verification

- **Final focused verification window:** `2026-07-15T16:17:48.513Z` through
  `2026-07-15T16:18:24.240Z`; every command exited `0` on the recorded host
  unless the canonical record explicitly describes its retained V8 product-gate
  exit `1`.
- `pnpm verify:audit:capability-truth` — passed seven adversarial tests and the
  live workspace drift check. The tests reject absent evidence kinds, unlabelled
  analysis, over-promotion to implemented/live/green, unsupported
  red/stale/blocked/missing states, duplicate audit sections, a one-byte view
  mutation, aged evidence, and future-dated evidence.
- JSON Schema Draft 2020-12 validation of the canonical registry — passed.
- Targeted ESLint with `--no-ignore --max-warnings=0` for the generator and
  tests — passed.
- Targeted Prettier across the schema, registry, scripts, workflows, managed
  views, package manifest, and canonical audit — passed.
- Both affected GitHub workflows parsed as YAML — passed.
- `pnpm verify:audit:machine-adjacent-docs` — passed 3/3 tests and the workspace
  contract check.
- `pnpm generate:audit:v1-v9-verification-record` — passed sequentially from the
  clean source commit with `--maxWorkers=1`. The authenticated result has
  351/351 V8 tests and 176/176 V9 tests passing, while retaining the known V8
  generated-contract drift result as measured exit 1.
- `pnpm verify:audit:v1-v9-verification-record` — passed 8/8 adversarial tests
  and record/table/log/source coherence.
- `git diff --check` — passed.

## Claim boundary and inapplicable avenues

- This task does not claim a complete content-slot registry, active placement
  controller, provider deployment, current Unreal run, learner efficacy,
  customer UI change, or release certification. Those require later COV and
  product work packages.
- Provider-probe, V2/V6 placement, and V9 live-learner release evidence is
  intentionally registered as missing. In the historical 2026-07-15 window V3
  and V8 were red; V8's later authenticated bounded pass is now stale. Source/
  test files establish at most `implemented`; they cannot impersonate live or
  green proof.
- No customer UI or media artifact changed, so Playwright, mobile automation,
  screenshots, accessibility scans, and media probes are inapplicable.
- No policy in this work package requires a human signature. No human approval
  is manufactured or implied.
