# Phase 182 Human Video Implementation Tracker

Last updated: 2026-09-12

This tracker is the operational execution record for Phase 182 (controlled human
video generation). The canonical checklist stays `TODOS/phase-182.md`; this file
carries ownership, dependency, status, changed files, verification, provider and
hardware evidence, cost, and blockers for every `182.C` item.

Nothing in this file may be used to mark a checkbox. A `[x]` in
`TODOS/phase-182.md` is the sole completion signal; a row here that says
`release-enabled` while the checklist still says `[ ]` means the row is wrong.

## Stage Order and the Downstream Prohibition (`182.C.01.02`)

Work runs top to bottom. The stages group into four gates, and a gate may not be
entered until every non-deferred item in the previous gate is checked in
`TODOS/phase-182.md` or recorded here as `blocked-*` with an owner, a reason,
and an unblock path.

| Gate                     | Stages                | What it establishes                                                                                    |
| ------------------------ | --------------------- | ------------------------------------------------------------------------------------------------------ |
| G1 Honest foundation     | `182.C.01`–`182.C.10` | Audited reality, one canonical contract, authorized ingestion, shot plans, live capability registry    |
| G2 Vertical slice        | `182.C.11`–`182.C.21` | One promoted route end to end: transport, storage, real evaluation, bounded retry, fail-closed release |
| G3 Creator-ready product | `182.C.22`–`182.C.30` | Long form, Studio/API, consent/security/provenance, admin review, observability                        |
| G4 Release               | `182.C.31`–`182.C.40` | V1 convergence, research track, corpus, bakeoff, deep verification, rollout, final evidence            |

Four prohibitions, each of which has a named upstream:

1. **No release-path item may be checked before its measurement exists.**
   Nothing in `182.C.28` (release proof), `182.C.38` (rollout), or `182.C.40`
   (final evidence) may be checked until `182.C.05` (contracts), `182.C.16`
   (durable storage), `182.C.17`–`182.C.20` (real evaluators), and `182.C.26`
   (consent) are complete or explicitly deferred with an owner and expiry.
2. **No provider submission item may be checked before consent.** Nothing in
   `182.C.11`–`182.C.14` may be marked `live-probed` until `182.C.26.01`–
   `182.C.26.02` bind an authorization record to the reference digests that the
   probe uploads.
3. **No quality claim may be checked before durable storage.** Base64 payloads
   cannot be measured at production scale, so `182.C.17`–`182.C.21` depend on
   `182.C.11.08` and `182.C.16.02`.
4. **No Studio or mobile item may be checked before the surface decision.**
   `182.C.23` and `182.C.24` depend on `182.C.24.05` and the release-train
   record in `182.C.01.11`.

## Status Vocabulary (`182.C.01.03`)

Statuses are ordered. A row may only claim a status when every weaker status is
also true. A plan-only module, a typed seam, or a mocked provider response never
gets past `implemented`.

| Status                  | Means                                                                                                                                                                                   | Minimum evidence in the row                                                      |
| ----------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- |
| `not-started`           | No code or document exists for this item.                                                                                                                                               | —                                                                                |
| `implemented`           | Code or documentation is merged and unit-tested in isolation. It may have no runtime caller. **A schema, a port interface, an adapter with no live call, or a plan object stops here.** | Changed files + a passing targeted test command                                  |
| `wired`                 | A real runtime path reaches it — a BFF route, a worker stage, a Studio action — with no test-only construction in between.                                                              | The caller's file:line + an integration test that exercises the real path        |
| `live-probed`           | It ran against the real provider, hardware, or model at least once. Provider marketing, a fixture replay, or a job-accepted response is not a probe.                                    | Dated redacted request/response fixture + output artifact digest (`182.C.01.04`) |
| `benchmark-passed`      | Its cohort met the documented thresholds at the documented sample size with confidence intervals, on held-out data.                                                                     | Bakeoff run ID + cohort size + CI + evaluator versions (`182.C.34`)              |
| `release-enabled`       | It is switched on for a real audience behind the flag set from `182.C.38.01`, with a passing release proof.                                                                             | Flag name + release-proof digest                                                 |
| `blocked-local`         | Correct and specified, but this hardware cannot run it. Not a failure; a hosting dependency.                                                                                            | The hosting route and the item that unblocks it (`182.C.01.10`)                  |
| `blocked-vendor`        | Needs a funded external account that does not exist yet.                                                                                                                                | The decision item that unblocks it                                               |
| `blocked-release-train` | Correct, but its surface is out of the current release scope.                                                                                                                           | The release that restores it (`182.C.01.11`)                                     |

## Provider Capability Evidence Rule (`182.C.01.04`)

Every claimed provider capability in this tracker, in code comments, in the
capability registry, and in `CONTENT_GENERATION_CAPABILITY_MATRIX.md` must link
both of:

1. A **dated, redacted request/response fixture** committed under
   `docs/domains/isis/human-video/evidence/<yyyy-mm-dd>/<model>/`, with API
   keys, signed URLs, provider download tokens, biometric bytes, and full
   private scripts removed, and the redaction performed by the shared redactor
   rather than by hand.
2. A **real generated artifact digest** — the SHA-256 of the bytes that actually
   came back, plus duration, resolution, codec, and audio-track presence read
   from `ffprobe`, not from the provider's metadata.

The catalog advertising a field is not evidence that the field works. A model's
`allowed_passthrough_parameters` listing `audio` is a permission to probe, not a
measured capability; until the probe fixture exists the capability row stays
`implemented` and the router must fail closed.

## Quality Claim Evidence Rule (`182.C.01.05`)

Every quality number — sync confidence, offset, WER, identity similarity,
artifact rate, pass rate — must be computed from the **exact bytes that were
released**, after the final mux and transcode, and must record:

- the digest of the measured file, which must equal the released asset's digest;
- the evaluator name, version, and weight digest;
- the frame and audio window the number came from;
- the coverage fraction (how much of the clip was measurable at all).

A provider preview, a candidate before transcode, a metadata summary, a progress
payload, or a score supplied by any party other than our own evaluator run is
not a measurement and may not populate a metric field. Modules that accept an
externally supplied score (`182.C.31.05`) are quarantined until they execute a
pinned evaluator themselves.

## Revalidation Cadence for Mutable Provider Facts (`182.C.01.06`)

The OpenRouter catalog, pricing, durations, option names, and retention terms
all change without notice. Each fact class has an owner and a maximum age; a
fact older than its maximum quarantines every route that depends on it.

| Fact class                                                                                                              | Source                                             | Max age | On change                                                                      |
| ----------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------- | ------- | ------------------------------------------------------------------------------ |
| Model existence, `canonical_slug`, durations, resolutions, aspect ratios, sizes, frame images, `generate_audio`, `seed` | `GET /api/v1/videos/models`                        | 24 h    | Diff and snapshot (`182.C.10.03`); removed model quarantines its routes        |
| `allowed_passthrough_parameters`                                                                                        | same                                               | 24 h    | A removed passthrough name fails the adapter closed immediately                |
| `pricing_skus`                                                                                                          | same                                               | 24 h    | Re-run the normalizer; a new SKU shape blocks admission estimates until mapped |
| Live behavior of a passthrough (does `audio` actually condition the render?)                                            | Canary probe                                       | 30 days | Demote the lane to `implemented` and re-probe                                  |
| Benchmarked cohort quality                                                                                              | Bakeoff (`182.C.34`)                               | 30 days | Re-run bounded benchmark; demote on regression (`182.C.34.09`)                 |
| Reference limits, retention, region, data handling                                                                      | **Manual dated record** — the catalog carries none | 90 days | Expired manual record fails the route closed                                   |
| Endpoint shapes, webhook signature scheme, event names                                                                  | Official guide                                     | 90 days | Contract tests re-derived from the guide                                       |

Revalidation runs in CI on the daily schedule and on every deploy; a fetch
failure quarantines rather than falls back to the cached snapshot once the max
age is exceeded.

## Exceptions, Waivers, and Disabled Evaluators (`182.C.01.07`)

Every temporary exception is a row here. No exception is open-ended: an expired
row fails its gate closed on the next run, and nothing removes a row except the
work that makes it unnecessary.

| ID  | Kind | Scope | Reason | Owner | Opened | Expires | Unblock path |
| --- | ---- | ----- | ------ | ----- | ------ | ------- | ------------ |
| —   | —    | —     | —      | —     | —      | —       | —            |

Kinds are: `model-waiver` (a specific model may skip a soft gate),
`evaluator-disabled` (a gate is not running), `manual-review` (a cohort requires
human sign-off in place of an automated gate), and `temporary-exception`
(anything else). Consent, exact-speech, safety, and severe identity/sync gates
are non-compensable and may never appear here (`182.C.21.03`).

## Verification Command Registry (`182.C.01.08`)

Commands are added as packages are created; an item is not checked until its row
names the exact command that was run. Nx is bypassed where worktree duplicates
break it, per `CLAUDE.md`.

| Class               | Command                                                                                                     | Status                   |
| ------------------- | ----------------------------------------------------------------------------------------------------------- | ------------------------ |
| Markdown format     | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md TODOS/phase-182.md`     | Active                   |
| Catalog snapshot    | `curl -sS https://openrouter.ai/api/v1/videos/models`                                                       | Active (public, no auth) |
| Typecheck (package) | `npx tsc --noEmit -p <package>/tsconfig.json`                                                               | Added per package        |
| Unit (package)      | `npx vitest run <file>`                                                                                     | Added per package        |
| Lint                | `npx eslint <paths>`                                                                                        | Added per package        |
| Build (library)     | `npx tsup` from the library directory                                                                       | Added per library        |
| Playwright (web)    | `apps/oshun/web/e2e-inspect/`, `workers=1`, chromium only                                                   | Added at `182.C.23`      |
| Provider canary     | Recorded per probe with its bounded spend                                                                   | Added at `182.C.11`      |
| Evaluator weights   | `~/.cache/oshun-human-video/` — sha256 per file in `evidence/2026-09-10-cpu-sync-and-alignment/README.md`   | Active (CPU, this box)   |
| Python evaluators   | venv with `--system-site-packages` (torch 2.14 cpu, faster-whisper, transformers, onnxruntime preinstalled) | Active                   |
| Release gate        | Recorded at `182.C.28`                                                                                      | Pending                  |

## Model Binding and Cost (`182.C.01.09`)

Every LLM-shaped step this phase adds — brief-to-shot planning, prompt
compilation, semantic-substitution checks, review triage — and every test,
harness, probe, and local stack binds the cheapest capable model through
OpenRouter, per `CLAUDE.md`:

```
OSHUN_ASSISTANT_PROVIDER=openrouter
OSHUN_ASSISTANT_OPENROUTER_MODEL=deepseek/deepseek-v4-flash-0731
OPENROUTER_PROVIDER_SORT=price
OPENROUTER_PROVIDER_QUANTIZATIONS=fp8   # when measuring
```

This binding is not negotiable per-item and overrides any recipe in this phase
that names a stronger model for a "quality" leg. Video generation is the
exception in kind but not in discipline: video canaries bind **tiny bounded
spend** with the model, duration, resolution, and expected cost written into the
row before the call is made.

Video pricing is not one per-second rate. The normalization of the heterogeneous
`pricing_skus` shapes — per second by resolution, per second with or without
audio, per generation mode, per video token, per reference image, per input
image, per megapixel-second, and per-generation minimums — is published in
`docs/agents/model-cost-openrouter.md` and is the source for the admission-time
estimate in `182.C.14.06`.

## GPU, ASR, and Local Memory Constraints (`182.C.01.10`)

The phase was drafted against a 16 GB MacBook Air with 2 GB of swap that had
frozen three times under stacked load. The worktree that implements it runs on
Linux, 15 GiB RAM, 4 cores, no GPU. Two consequences bind this phase:

**`blocked-local` is a first-class status, not a failure — and it is earned per
item, never inherited (rewritten 2026-09-10).** The original text here said
`182.C.17`, `182.C.19`, `182.C.20`, and `182.C.32` "need GPU inference this
machine cannot host". Measured on this box on 2026-09-10, that is false for the
evaluator side: SyncNet v2 (the `182.C.17.01` baseline, MIT) runs a forward pass
at about 1 s of CPU per second of 25 fps video, a wav2vec2 CTC forced aligner
aligns 7 s of audio in 1.3 s with a maximum raw onset error of 81 ms, SFace
produces identity embeddings, and YuNet produces face boxes and head pose — see
`evidence/2026-09-10-cpu-sync-and-alignment/`,
`evidence/2026-09-10-cpu-identity-embedding/`, and
`evidence/2026-09-10-cpu-reference-quality/`. Evaluators are therefore built and
probed on CPU first; a **RunPod GPU worker through `libs/isis/runpod-surface`
and `apps/oshun/bff/src/generation/runpod-comfy-client.ts`** is a production
throughput decision taken at `182.C.17.13` from the numbers measured at
`182.C.17.12`. The only genuinely GPU-bound rows are self-hosted video
_generation_ candidates (`182.C.32.03`–`.08`), because no CUDA device exists
here (`nvidia-smi` absent, torch is the `+cpu` build). Wherever an evaluator is
unavailable at runtime, its seam fails loud rather than returning a default
score. Model weights are staged outside every worktree in
`~/.cache/oshun-human-video/` with digests recorded in the evidence; Python
dependencies go in a venv and never touch the pnpm lockfile.

**`182.C.18` (exact-speech evaluation) split on 2026-09-10.** The vendor
decision at `182.C.02.12` is made and funded — OpenRouter audio-input models,
live-probed against known ground truth — so transcription is no longer blocked
on anything. What remains blocked is narrower and different in kind: word-level
**forced alignment**, which `182.C.18.02` needs and which the probe measured
that no OpenRouter audio model provides (three recognizers disagreed by up to
1013 ms on one clip, 8.1× the ITU reject boundary). That residual is no longer
blocked at all: a wav2vec2 CTC forced aligner was measured on this CPU the same
day (1.3 s for a 7 s clip, max 81 ms raw onset error against a known-onset
fixture; `evidence/2026-09-10-cpu-sync-and-alignment/`). Whisper's
cross-attention timestamps were measured in the same probe and are disqualified
alongside the chat models' (600 ms first-word error, merged words).

**`blocked-local` must name what was attempted and what failed (2026-09-10).**
Two corrections in one day came from the same cause: a constraint sentence
became a row-level status without the item being tried. `182.C.06.06` was
recorded as needing a GPU worker when ffmpeg and a 227 KB Apache-2.0 model
handle most of it, and all ten `182.C.19` rows carried one copy-pasted "No GPU
host" blocker when four are unblocked, one (`.10`) was never blocked at all, and
the genuinely hard ones (`.08`, `.09`) are blocked on **corpora**, not compute.
A blocker repeated across a group of items is a claim about none of them. Any
row claiming `blocked-local` states the specific thing that was run and the
specific way it failed.

**Correction to the hardware assumption (2026-09-10).** The phase constraints
describe a 16 GB MacBook Air on which "SyncNet-class evaluators, face tracking,
and ASR are hardware-blocked locally". This worktree actually runs on Linux with
15 GiB RAM and 4 cores, and CPU probes showed face detection, five-landmark head
pose (SQPnP, real degrees), blur and exposure, SFace identity embeddings, a
SyncNet v2 forward pass, and CTC forced alignment all run here — see the three
`evidence/2026-09-10-cpu-*` records. Python packages also do not touch the pnpm
lockfile, so the `CLAUDE.md` install caution does not extend to them.
`blocked-local` must therefore be justified per item by what was actually tried,
not inherited from that description. (An earlier version of this paragraph still
said `182.C.19` was "genuinely blocked because no CPU model here produces an
ArcFace-class embedding"; SFace does, and that sentence was itself the
untested-inheritance error it warned about.)

**Memory checkpoint before every heavy invocation.** Any local `ffmpeg`,
`ffprobe`, evaluator, Playwright, `tsc`, or build run in this phase is preceded
by `free -m` on this Linux box (`memory_pressure | tail -2` on the Mac); below
about 3 GB available (35% free on the Mac) the stack is torn down before the
run, not alongside it. Dev servers die with the session and are restarted at
every phase boundary.

## Release Train and Surface Decisions (`182.C.01.11`)

Target release train, read from `libs/oshun/navigation/src/release-scope.ts`
rather than assumed:

- **V1.0** — the release currently shipping (`LILITH_CURRENT_RELEASE`). Web and
  PWA only, four rooms: Tara, Nyx, Arete, Nisaba. Phase 182's creator flow
  targets this train on the web surface.
- **V1.1** — adds the native iOS/Android apps
  (`LILITH_NATIVE_APPS_RELEASE_NAME`, gated by `LILITH_NATIVE_APPS_AVAILABLE`).
  Every `182.C.24` mobile item lands against this train.
- **V1.2** — restores Veritas and Metis (`V1_DEFERRED_DOMAIN_RELEASE`,
  `V1_DEFERRED_DOMAIN_IDS`). No console may advertise, link to, count, or
  compose those rooms before then.

**Veritas-touching items are V1.2-gated** and carry status
`blocked-release-train`: `182.C.13.02` (wire or retire the Veritas unified
provider), `182.C.31.07` (consolidate Veritas avatar/video providers), and
`182.C.39.08` (the Veritas half of the migration guide). Their libraries,
contracts, and code may be written and tested inside the deferral — what is
forbidden is a surface that presents them.

**Desktop surface decision (from `182.C.24.05`): out of scope for Phase 182.**
No Oshun studio desktop app exists, and the V1 line ships no desktop surface at
all — V1.0 is web + PWA, V1.1 adds native mobile, V1.2 restores two rooms. The
existing Electron surfaces (`apps/yemaya/studio-desktop`, `apps/lilith/desktop`,
`apps/iris/desktop`) belong to other products and would each need the full
server contract, capability registry, consent, quality decision, and
release-proof stack duplicated to host this flow. Phase 182 ships **web (V1.0)
and mobile (V1.1)** only. This decision is recorded here as the precondition for
Studio work and is formalized in `182.C.24.05`; reversing it requires naming a
host app and re-opening `182.C.24`.

---

## Progress Tables

One row per `182.C` item. `Changed Files`, `Verification`,
`Provider/Hardware Evidence`, and `Cost` stay `—` until the item is worked; they
are filled in the same commit that flips the checkbox in `TODOS/phase-182.md`.

### 182.C.01 Phase Control, Tracker, and Evidence Rules

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner       | Dependency  | Status      | Changed Files                                                                                                   | Verification                                                                         | Provider/Hardware Evidence                                                       | Cost                      | Blocker                                                       |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | ----------- | ----------- | ----------- | --------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------- | ------------------------- | ------------------------------------------------------------- |
| 182.C.01.01 | Create `docs/domains/isis/human-video/implementation-tracker.md` with one row per `182.C` item, owner, dependency …   | Release Eng | —           | implemented | `docs/domains/isis/human-video/implementation-tracker.md`                                                       | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.02 | Record the stage order and prohibit checking downstream release work before contracts, consent, real evaluation …     | Release Eng | —           | implemented | `implementation-tracker.md` § Stage Order and the Downstream Prohibition                                        | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.03 | Define “implemented,” “wired,” “live-probed,” “benchmark-passed,” “release-enabled,” and “blocked” so plan-only …     | Release Eng | —           | implemented | `implementation-tracker.md` § Status Vocabulary                                                                 | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.04 | Require every claimed provider capability to link a dated, redacted request/response fixture and a real generated …   | Release Eng | —           | implemented | `implementation-tracker.md` § Provider Capability Evidence Rule                                                 | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.05 | Require every quality claim to link measurements computed from the exact released bytes …                             | Release Eng | —           | implemented | `implementation-tracker.md` § Quality Claim Evidence Rule                                                       | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.06 | Define a revalidation cadence for mutable provider models, pricing, duration limits, reference limits, option names … | Release Eng | —           | implemented | `implementation-tracker.md` § Revalidation Cadence                                                              | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.07 | Add explicit owner and expiry fields for every temporary exception, disabled evaluator, manual review …               | Release Eng | —           | implemented | `implementation-tracker.md` § Exceptions, Waivers, and Disabled Evaluators                                      | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.08 | Record targeted lint, typecheck, unit, integration, provider-sandbox, Playwright, accessibility, performance …        | Release Eng | —           | implemented | `implementation-tracker.md` § Verification Command Registry                                                     | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | Not applicable; documentation-only.                                              | —                         | None                                                          |
| 182.C.01.09 | Bind every test, harness, probe, and local stack to the cheap OpenRouter model per `CLAUDE.md` …                      | Release Eng | —           | implemented | `implementation-tracker.md` § Model Binding and Cost; `docs/agents/model-cost-openrouter.md` § Video generation | `curl -sS https://openrouter.ai/api/v1/videos/models` (200, 28 models, 2026-09-09)   | Live public catalog fetch, 26,578 bytes, 2026-09-09                              | $0 (catalog read is free) | None                                                          |
| 182.C.01.10 | Add a `blocked-local` tracker status for GPU-bound evaluator and ASR work, record the hosting route for each …        | Release Eng | —           | implemented | `implementation-tracker.md` § GPU, ASR, and Local Memory Constraints                                            | `pnpm exec prettier --check docs/domains/isis/human-video/implementation-tracker.md` | `memory_pressure` 58% free at authoring; RunPod route named, not yet provisioned | —                         | Hosting route is named but unprovisioned until `182.C.17.13`. |
| 182.C.01.11 | Record the target release train, the V1.2 gating of every Veritas-touching item …                                     | Release Eng | 182.C.24.05 | implemented | `implementation-tracker.md` § Release Train and Surface Decisions                                               | `libs/oshun/navigation/src/release-scope.ts` read for V1.0/V1.1/V1.2 facts           | Not applicable; documentation-only.                                              | —                         | Desktop decision recorded here; formalized at `182.C.24.05`.  |

### 182.C.02 Repository Capability Audit and Gap Ledger

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner | Dependency  | Status      | Changed Files                                                                                                           | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Provider/Hardware Evidence                                                   | Cost | Blocker                                   |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | ----- | ----------- | ----------- | ----------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------- | ---- | ----------------------------------------- |
| 182.C.02.01 | Audit the curated video card types and validation path; record every field needed for references, dialogue, audio …   | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | Read `libs/isis/curated-cards/src/card-types.ts` (228 lines) and `card-validators.ts` (354 lines); confirmed two video card kinds, set-level `CARD_INPUT_FIELD_ALLOWLIST` enforcement (`card-validators.ts:61-79`), enum checks in `validateSharedVideoParams` (`card-validators.ts:283-311`), and the `https:`/`data:image/` reference constraint (`card-validators.ts:337-342`); gap table records all eleven missing human-video fields                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Doc audit — no provider call                                                 | —    | None                                      |
| 182.C.02.02 | Audit the BFF video executor, provider environment, router, queue, catalog, storage, governance, and release gates …  | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`, `apps/oshun/bff/src/generation/video-provider-env.ts`              | Re-read all seven files; confirmed fail-closed `provider_not_configured` (`video-executor.ts:160,179`) and empty-output refusal (`:130-131`), F-07 `unsigned_urls` download (`openrouter-video-provider.ts:261`), F-08 `generate_audio` never set (zero hits across `src/generation/`), Marsaglia–Tsang Beta–Thompson sampler (`provider-quality-router.ts:59-83,153`), fail-closed governance compose (`video-governance.ts:105-116`), six-state enum (`jobs-route.ts:59-68`). Found F-06 fixed in the header but **surviving in the `adaptVideoProviderGenerate` docstring** — corrected. `npx vitest run src/generation/video-provider-env.test.ts` → 20 passed                                                                                                                                                                                                                                                                                                                                                                                                                       | Doc audit — no provider call; route stays `implemented`, never `live-probed` | —    | None                                      |
| 182.C.02.03 | Audit `@isis/ai-video`, `@isis/ai-providers`, and `@isis/seedance-provider` for duplicate clients …                   | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`, `libs/isis/seedance-provider/vitest.config.ts`                     | F-01 fix confirmed in source (`lip-sync-validation.ts` 811->735 lines; fabricated helpers gone, `SeedanceLipSyncServiceNotConfiguredError` thrown at `:157`); F-03 counts re-measured exactly (117 `.js`, 118 `.d.ts`, all ten artifacts still beside source in `video-generation/`); F-02 six abstractions re-counted (`src`-scoped: ai-video 22, yemaya 18) and both stale id vocabularies re-read (`wan-provider.ts:49` `alibaba/wan-2-6-t2v` vs `model-registry.ts:45` `wan-ai/wan2.6-t2v`); F-11 caveat present at `:657-663`. Added the two aspects the first pass omitted: **F-12** four incompatible `VideoGenerationRequest` contracts, **F-13** reusable workflow primitives. Adversarial grep over `lip-sync-validation.ts` -> one hit, describing the fail-loud behaviour. `npx vitest run src/lip-sync-validation.test.ts` -> 3 passed (needed a new `@oshun/math` alias; the suite could not load standalone before)                                                                                                                                                       | Doc audit - no provider call                                                 | —    | None                                      |
| 182.C.02.04 | Audit identity, scene, shot, wardrobe, and temporal consistency modules and map them to the new human-video contract. | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | Read all ten `libs/isis/workflows/src/consistency/` modules. Corrected the first pass, which called them "bookkeeping over supplied scores": `identity-drift.ts` computes real cosine similarity over embedding vectors (`:112-131`) and mean euclidean landmark displacement (`:134-152`), and `temporal-quality-qa.ts` computes weighted pairwise frame metrics. The real boundary is feature EXTRACTION - nothing here derives an embedding or a landmark from a decoded frame, which is why `182.C.19` still needs the GPU-hosted extractor and may reuse these assessors. `npx vitest run src/consistency/identity-drift.spec.ts src/consistency/temporal-quality-qa.spec.ts` -> 8 passed                                                                                                                                                                                                                                                                                                                                                                                           | Doc audit - no provider call                                                 | —    | None                                      |
| 182.C.02.05 | Audit every `@nous/portrait-animation` export and label it executable, adapter-backed, planning-only …                | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | Re-counted the library: 23 non-spec files = 22 modules + `index.ts`, 12,248 non-spec lines (first pass said ~11k). The first pass labelled only 20 of the 22 - read and labelled the two it omitted: `portrait-animation-library.ts` (metadata-only, 17 lines of Nx identity constants) and `portrait-animation-streaming-output-pipeline.ts` (planning-only; its HTTP matches are spec-URL constants at `:5-11` and a WHIP signalling literal, not requests). Re-ran the no-inference scan: only two hits, both plan DATA - a `.safetensors` checkpoint path string and a `--flag-do-torch-compile` argv entry in an argument array the module never executes. Zero executable, zero adapter-backed stands.                                                                                                                                                                                                                                                                                                                                                                             | Doc audit - no provider call                                                 | —    | None                                      |
| 182.C.02.06 | Audit `apps/psyche/avatar-engine` for entry point, dependencies, model weights, license, GPU/CPU behavior …           | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | Re-verified every F-04 claim against the tree: 157 files, no `main.py`, no `scripts/` package, `tests/` holds only `.gitkeep`, `smplx`/`insightface` in `pyproject.toml`, `model-license-review.ts:6` names avatar-engine weights with no record filed, `backend.ts` speaks HTTP at `:399`/`:611`. **Corrected the `train` row**: the first pass reported `trainer.py` missing; it exists (692 lines - `AvatarDataset`, `BlendshapeModel(nn.Module)`, `AvatarTrainer`, `DistributedTrainer`) but defines only classes, so `python -m avatar_engine.training.trainer` is a silent no-op rather than an import error. Retire-not-harden recommendation for `182.C.31.06` unchanged.                                                                                                                                                                                                                                                                                                                                                                                                        | Doc audit - no GPU run (hardware-blocked)                                    | —    | None                                      |
| 182.C.02.07 | Audit Veritas HeyGen, Tavus, D-ID, Runway, and unified providers; identify which are actually constructed and …       | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | Re-traced reachability from `main.ts`/`server.ts`/`index.ts` rather than from the barrel. HeyGen confirmed live (`main.ts:138,231`). **Corrected the first pass**, which reported Runway, OpusClip and Shotstack as never constructed: they ARE constructed (`services/broll.ts:188`, `services/youtube-shorts.ts:469`, `services/video-rendering.ts:243`, `services/video-pipeline.ts:879`) inside four service modules that have zero importers repo-wide. Tavus/D-ID are constructed only in their own lazy singletons and `unified.ts:247`. Conclusion unchanged, reason corrected - wiring Runway or Shotstack means adopting an unimported service module, not just constructing a client.                                                                                                                                                                                                                                                                                                                                                                                         | Doc audit - no provider call                                                 | —    | V1.2-gated (`182.C.13.02`, `182.C.31.07`) |
| 182.C.02.08 | Audit V1 persona/avatar/voice documentation against current code and create a fact table of delivered, partial …      | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | Read the sources behind every row instead of their file names. Two corrections: `voice-consent-store.ts` is a pure EVALUATOR (`evaluateVoiceConsent:68`) wired at `routes/admin-studio-voice-consent.ts:99` that persists nothing - evaluation delivered, custody missing; `persona-lifecycle-store.ts` genuinely persists via `@oshun/persistence/durable-snapshot-store` with a fail-closed path at `:110`, so packs are delivered, not partial. Also found a SECOND `evaluateVoiceConsent` in `libs/euterpe/provenance/src/rights-consent/rights-consent.ts:164` with an incompatible signature and a wall-clock rather than logical-sequence clock - `182.C.26` must pick one. Viseme table verified: 39 ARPABET phonemes -> 11 visemes.                                                                                                                                                                                                                                                                                                                                             | Doc audit - no provider call                                                 | —    | None                                      |
| 182.C.02.09 | Update `CONTENT_GENERATION_CAPABILITY_MATRIX.md` only from verified findings …                                        | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | No status hand-edited, and none warranted: the matrix is rendered from `scripts/audit/capability-truth-registry.json` under a do-not-hand-edit banner, this pass produced no live generation or encoder receipt, and the matrix has no lip-sync or human-video row for the F-01 fix to falsify. Re-verified the two rows the audit relies on - `generation-video` is `missing` (matrix `:46`) and `temporal.test.ts` asserts exact MSE values (`[100,400]`, mean 250) plus a typed throw, so it survives a stub test. New **F-14**: the `generation-video` registry entry cites only `runpod-provider-probe` while the deployable lane is fal/OpenRouter, so `182.C.11` must register an OpenRouter probe ref rather than inherit the RunPod one.                                                                                                                                                                                                                                                                                                                                        | Doc audit - no provider call                                                 | —    | None                                      |
| 182.C.02.10 | Create a deletion/consolidation list across the five provider abstractions that already exist (`@isis/ai-video` …     | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | Eight dispositions covering all six abstractions plus avatar-engine and portrait-animation; each rests on a finding verified in `02.03`-`02.07`. Spot-checked the one claim the list makes about the deployable path: the BFF does import LTX from the SDK of record (`video-provider-env.ts:37`). Extended the list with the part it was missing - consolidation is a CONTRACT problem first (F-12: one request/result contract, `ProviderVideoRequest` as the base), `ai-video` cannot retire until `video-workflow-manager.ts` (F-13) is lifted out, and the F-03 build artifacts get deleted with the ADR rather than ahead of it.                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Doc audit - no provider call                                                 | —    | None                                      |
| 182.C.02.11 | Audit the BFF generation primitives named in the anchor list (job envelope and status enum, budget/quota gates …      | Isis  | 182.C.01    | implemented | `docs/domains/isis/human-video/capability-audit.md`                                                                     | All fourteen extend-or-replace rows re-checked against the tree; every line count in the table matches exactly (415/323/411/61/208/167/167/154/221/445/521/218/95/289). F-10 verified at source: `release-proof.test-fixtures.ts:12-20` lists exactly the nine named gates and none covers speech, sync, speaker attribution, identity or likeness consent; `proof-types.ts` `PROOF_TYPES` is exactly fourteen and already includes `consent`, with `ConsentProofSchema:181` and `ObservedOutcomeProofSchema:295` available to carry the five new gates. `dub-faithfulness.ts` confirmed fail-loud (`:132-133`) with a real credential-free token-cosine (`:76`). Decision stands: extend all fourteen, rebuild none.                                                                                                                                                                                                                                                                                                                                                                    | Doc audit - no provider call                                                 | —    | None                                      |
| 182.C.02.12 | Decide and fund the ASR vendor for exact-speech evaluation (OpenAI has no credits and there is no Deepgram account …  | Isis  | 182.C.02.02 | live-probed | `docs/domains/isis/human-video/evidence/2026-09-10-openrouter-asr/` (new), `libs/contracts/src/human-video/dialogue.ts` | DECIDED AND FUNDED 2026-09-10: OpenRouter audio-input models. Settled by a round trip with KNOWN GROUND TRUTH rather than by reading model cards - `openai/gpt-audio-mini` synthesized a known sentence (streaming + pcm16 are mandatory for audio output; wrapped to WAV with ffmpeg), then three candidates transcribed it. All three were correct: voxtral-small ($0.000306), gemini-3.1-flash-lite ($0.000070, the cheap default per CLAUDE.md), gpt-audio-mini ($0.000066). 46 audio-input models are on the account, so the `182.C.18.07` ensemble is a routing choice. The SAME probe disqualified word timings: the three disagreed by up to 1013 ms on onsets for one 3.95 s clip - 8.1x the 125 ms ITU reject boundary - so they are generated, not force-aligned, and cannot carry `182.C.18.02`. Probe also surfaced a real defect: "$42" for spoken "forty two dollars" scored CWER>0 on a PERFECT transcription; normalizer now expands currency/percent symbols, with a second fix for "$42." at sentence end being misread as a decimal. Total probe spend under $0.002. | Live provider probe; dated fixture + 3.95 s artifact committed               | —    | None - decided and funded                 |

### 182.C.03 Product Contract, Use Cases, and Honest Quality Promise

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner          | Dependency | Status      | Changed Files                                             | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Provider/Hardware Evidence                                              | Cost | Blocker                                                              |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | -------------- | ---------- | ----------- | --------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------- | ---- | -------------------------------------------------------------------- |
| 182.C.03.01 | Define supported modes: text-to-scene with generated cast, reference-to-video, exact-dialogue performance …           | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md` (new) | Seven modes defined, each with its promise, mandatory inputs, lanes and the reason it is a mode rather than a parameter. Lane assignments taken from the dated catalog snapshot, not assumed: exact-dialogue is restricted to the two models with an `audio` passthrough (wan-2.7, wan-2.6); continuation and edit are direct-lane only because OpenRouter exposes no extend or edit operation - verified the in-repo modules exist (`video-extension.ts:10,24`, `video-inpainting.ts:21,56`, `character-replacement.ts`); hailuo-3 and aleph-2 recorded as probe candidates, not lanes. Mode is carried on the request and never inferred; the router picks a model within a mode.                                                                                                                                                         | Doc - lane claims trace to the 2026-09-09 catalog snapshot, none probed | —    | None                                                                 |
| 182.C.03.02 | Define active-person benchmark scenarios: walking, cooking, presenting, driving while stationary/safely staged …      | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md`       | Ten scenarios (S1-S10), each admitted only because it breaks something a talking head does not: scale change, chin occlusion, off-axis mouth geometry, side lighting, motion blur, non-speech open mouth, low face-pixel regime, hand-face overlap, prop/wardrobe continuity, mid-utterance angle change - each mapped to the evaluator that must catch it. Same script and audio across all ten so the variable is motion and framing only; multi-character variants reported separately, never averaged in. Scenarios are cohort keys for the existing Beta-Thompson router extended per `182.C.34.07`, not a pass/fail list, and each must fit its lane duration bound (2-10s wan-2.7, 5-10s wan-2.6) or be authored as a multi-shot plan.                                                                                               | Doc - no runs; sample sizes and CIs belong to `182.C.34`                | —    | None                                                                 |
| 182.C.03.03 | Define out-of-scope behavior for launch: live impersonation, unauthorized public figures, unsafe stunts …             | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md`       | Nine boundaries, split into REFUSED (policy) and UNSUPPORTED (capability) so a safety line cannot drift into a roadmap item. Refusals bind to the existing taxonomy rather than a parallel policy - verified `nonconsented-likeness`, `voice-clone-abuse`, `deepfake` and `watermark-removal` are already in `HARD_BLOCK_KEYS` (`policy-taxonomy/categories.ts:191-200`). Two capability rules written out: off-screen speech may ship only as labelled narration and its frames score `unknown`, never `pass`; crowd dialogue is refused at validation because no evaluator can attribute speech to an unenrolled face. A refusal never degrades into a nearby product.                                                                                                                                                                    | Doc - policy binding only, no provider call                             | —    | None                                                                 |
| 182.C.03.04 | Define exact-dialogue semantics: canonical script, approved pronunciation, allowed disfluencies, expected language …  | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md`       | Seven parts defined: digest-pinned canonical script (an edit invalidates candidates rather than re-scoring them), pronunciation lexicon as approval reported per token, disfluencies default none, per-turn language with code-switch spans, a split word-error target (content-word error rate gated at 0, function-word WER reported only), timing tolerance on two clocks, and nonverbal vocalizations permitted in approved audio but never as model inventions. Sync tolerances taken from the published broadcast standards rather than invented: ATSC IS-191 +15/-45 ms as the release target, ITU-R BT.1359-1 +45/-125 ms detectability as the hard reject, `needs_review` in between (reuses the existing job status). Explicitly records that every gate here is unmeasurable until `182.C.02.12` funds ASR.                      | Doc - thresholds cite ITU-R BT.1359-1 and ATSC IS-191; no runs          | —    | Measurement blocked on `182.C.02.12` (ASR vendor)                    |
| 182.C.03.05 | Define “perfect delivered sync” as thresholded audiovisual alignment plus exact speech, correct speaker attribution … | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md`       | Defined as a conjunction of six measured conditions (C1-C6), not a weighted score: LSE-C>=6 / LSE-D<=8 plus the ATSC offset window, CWER=0, 100% speaker attribution, ArcFace cosine>=0.35 with no critical drift, >=85% visible-mouth coverage on speech-active frames, FVD<=120 / head-pose MAE<=8deg / zero blank-frozen findings. Numbers inherited from the defaults verified in `portrait-animation-quality-metrics-pipeline.ts:291-304`, NOT invented - and recorded as thresholds with no producer today (F-09), to be recalibrated on our own corpus at `182.C.34` before they gate anything. `unknown` never counts as `pass`; per-shot then per-delivery, no averaging.                                                                                                                                                          | Doc - thresholds traced to in-repo defaults; no evaluator exists yet    | —    | Producers blocked: `182.C.17`/`.19`/`.20` (GPU), `182.C.02.12` (ASR) |
| 182.C.03.06 | Define separate quality tiers for draft, standard, production, and hero delivery …                                    | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md`       | Four tiers with explicit candidate counts (1/2/4/6) and retry budgets (0/1/2/3), all sharing the SAME six release conditions - a tier buys work, never a waived gate. Generalizes the rule the hero path already states at `video-hero-gate.ts:12-14`. `draft` is explicitly not releasable and cannot be promoted by relabelling. A retry is allowed only against a named failed condition and must change something addressed to it, otherwise it counts as a candidate. Budget and quota gates admit per tier before any provider call; exhausting the budget ends `blocked`/`needs_review` with the failed conditions listed, never a best-of-a-bad-set delivery or a silent downgrade. Hero additionally requires recorded human sign-off.                                                                                             | Doc - no provider call                                                  | —    | None                                                                 |
| 182.C.03.07 | Define a deterministic failure message for requests no currently healthy provider can satisfy …                       | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md`       | Failure contract defined: stable `human_video_unsatisfiable` code plus ten enumerated reasons, each carrying the required capabilities and what each healthy lane actually offers. Code style matches the existing loud snake-cased errors on the video path (`provider_not_configured`, `video_provider_missing_output_url`, `openrouter_videos_empty_response`) rather than adding a dialect, and maps onto the existing `failed`/`blocked` job states. Four prohibited substitutions written out - no mode downgrade, no dialogue downgrade, no silent tier downgrade, no fabricated success. `capability_unprobed` is a deliberate reason so an unproved capability is distinguishable from a provider fault in the job record. Includes `zdr_tenant_ineligible`, since a ZDR-enforcing tenant gets no OpenRouter video routing at all. | Doc - no provider call                                                  | —    | None                                                                 |
| 182.C.03.08 | Define creator-visible limitations for side profiles, occlusion, very fast speech, singing, shouting, whispering …    | Isis + Product | 182.C.02   | implemented | `docs/domains/isis/human-video/product-contract.md`       | Ten limitations, each with the reason it is hard and the concrete system response - measure, disclose, or refuse. Two are refusals in exact-dialogue mode (singing, whispering) and stay refusals until an evaluator can score them. Occlusion, blur and tiny faces resolve to `unknown` spans counted against the C5 coverage budget rather than scored as passes; off-axis spans route to `needs_review` with angle named as the cause so an angle artefact is not recorded as bad sync. Surfaced in three places: pre-flight notice, per-condition delivered evidence (e.g. "87% mouth coverage", not "sync passed"), and the published per-cohort pass rates.                                                                                                                                                                           | Doc - no provider call                                                  | —    | None                                                                 |

### 182.C.04 Architecture Decisions and Ownership

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner     | Dependency | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                            | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Provider/Hardware Evidence                                      | Cost | Blocker |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | ----------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------- | ---- | ------- |
| 182.C.04.01 | Write an ADR naming one canonical human-video request, provider capability, job state, measurement …                   | Contracts | 182.C.03   | implemented | `docs/domains/isis/adr/ADR-0006-canonical-human-video-contracts.md` (new), `docs/domains/isis/adr/README.md`                                                                                                                                                                                                                                                                                                                                             | ADR-0006 accepted, following the existing Isis ADR format and added to the index. Names one contract per concern: request (`HumanVideoRequest`, based on the BFF `ProviderVideoRequest` - narrowest and the only one on the deployable path, with `durationSeconds` as the single duration representation), capability (discovered per lane, `unprobed` until a fixture exists), job state (extend the existing six-state enum, add stages not statuses), measurement (tri-state per condition with `unknown` never coerced to `pass`), and release proof (extend `ProofRecordSchema` with the five missing gates). Consequences record the SDK-of-record choice, the ai-video migrate-then-retire sequencing behind F-13, the yemaya id quarantine, and the F-03 artifact deletion tied to the migration.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Doc - decisions trace to audited file:line evidence             | —    | None    |
| 182.C.04.02 | Decide which types belong in shared contracts versus Isis implementation libraries versus the Oshun BFF boundary.      | Contracts | 182.C.03   | implemented | `docs/domains/isis/adr/ADR-0006-canonical-human-video-contracts.md` (section 6)                                                                                                                                                                                                                                                                                                                                                                          | Three homes with an explicit test - who must AGREE on the type, not who wrote it first. `@oshun/contracts` takes anything two deployables must agree on or that appears in a release proof; `@isis/*` takes provider- and algorithm-shaped types; the BFF keeps boundary-only types. Three rules follow: provider passthrough parameter names never enter shared contracts (shared carries the capability vocabulary instead, since the vendor list is a dated mutable fact), everything in a release proof is shared by definition (four readers), and credential/biometric types stay in the BFF because a shared type is importable by a browser bundle - which `182.C.04.04` forbids.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Doc - no provider call                                          | —    | None    |
| 182.C.04.03 | Decide whether `@isis/ai-video` or `@isis/ai-providers/video-generation` is the provider SDK of record and …           | Contracts | 182.C.03   | complete    | `libs/isis/ai-video/` (deleted), `apps/oshun/bff/src/generation/human-video-workflow.ts` (new) + spec, `libs/yemaya/remote-film-capture/src/isis-ai-video-types.ts` (new), `libs/isis/video-enhancement/src/lib/lens-physics.ts` (relocated) + spec, `docs/domains/isis/adr/ADR-0006-canonical-human-video-contracts.md`, `docs/domains/isis/human-video/capability-audit.md`, `eslint.config.js`, `tools/conventions/conventions-ratchet-baseline.json` | RETIRED IN CODE, not only in decision. Four ADR-0006 steps, four commits. **1** The F-15 shim was narrowed to the library's exact unions rather than deleted: pointing the web tsconfig at the real source pulls the runtime graph into an app typecheck the shim exists to keep small, and re-exporting inside `declare module` is TS2439, which under `skipLibCheck` silently resolves the surface to `any` (measured before choosing). A probe project mirroring `apps/oshun/web/tsconfig.json` confirmed all six previously-accepted values are now rejected. **2** F-13 lifted to `apps/oshun/bff/src/generation/human-video-workflow.ts`: priority queue, concurrency ceiling, `enqueueBatch`, shot fan-out, and a template executor, re-authored so every step names a `HumanVideoStage` and every advance is checked with `canTransition`; no default handlers (unbound stage throws); no polling. `WORKFLOW_TEMPLATES` left behind, four templates authored against the stage order. 30 tests. **3** The single consumer took local ownership of its eight types (`Plate` prefix) instead of `@oshun/contracts/human-video` - that contract is for controlled human video and this consumer generates establishing shots and plates, and the vocabularies disagree (`2160p` vs `4k`). Shim, both web aliases, the library alias, and the workspace dep removed; 3 tests pass. **4** Library deleted (27 files) with the ten F-03 build artifacts under `ai-providers/video-generation/`, the eslint bridge entry, and seven ratchet-baseline paths. **F-16 found while deleting:** `advanced/` was never inventoried. `lens-physics.ts` is real optics with no provider coupling - relocated to `@isis/video-enhancement`, 7 tests pass there. `ensemble-hdr.ts` fabricates VMAF from provider metadata, tone-maps by renaming a file extension, reports the configured peak nits as `measuredPeakNits`, and returns an `.exr` path it never wrote, all behind optional seams both of its tests inject - F-01's shape, so it was deleted with the provider stack rather than kept. User decision on the split, 2026-09-10. | Doc + code — no provider call                                   | —    | None    |
| 182.C.04.04 | Keep provider credentials, webhooks, downloads, and biometric uploads at server boundaries …                           | Contracts | 182.C.03   | implemented | `docs/domains/isis/adr/ADR-0006-canonical-human-video-contracts.md` (server-boundary section)                                                                                                                                                                                                                                                                                                                                                            | Four material classes with an enforcement point each - credentials, webhook receipt and HMAC verification, Bearer-authed provider downloads, biometric uploads. Verified the boundary currently holds rather than asserting it: no `NEXT_PUBLIC_*` name carries a provider key, no `.tsx` under `apps/oshun/web/src` reads a provider secret, and the single client component that touches `process.env` (`config/runtime-context.tsx`) is limited to nine `NEXT_PUBLIC_*` keys, none of them a credential. Recorded as an invariant to preserve, and tied back to the section-6 rule that keeps credential and biometric types out of shared contracts on reachability grounds.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | Static verification of the web bundle surface; no provider call | —    | None    |
| 182.C.04.05 | Define ports for asset ingestion, speech production, shot planning, capability discovery, generation, evaluation …     | Contracts | 182.C.03   | implemented | `docs/domains/isis/adr/ADR-0006-canonical-human-video-contracts.md` (ports section)                                                                                                                                                                                                                                                                                                                                                                      | Ten ports P1-P10, each naming the existing seam it extends rather than a greenfield interface, and each with its unconfigured behaviour stated - all copying the `DubFaithfulnessNotConfiguredError` fail-loud pattern at `dub-faithfulness.ts:132-133`. Verified the named seams exist (`generated-artifact-object-store.ts`, `narration-executor.ts`, `video-executor.ts:43`, `video-probe.ts`, `audio-probe.ts`, `video-governance.ts`, `media-releasability.ts`, `release-gate.ts`). Four cross-cutting rules: a port is a capability boundary not a layer (three implementations of P5 without a lowest-common-denominator payload), no port returns a value it did not compute (P6 returns `unknown`, the F-01 lesson in architecture form), P7 correction is disclosed by construction, and ports inject at the deployable boundary so credentials stay in one place.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Doc - no provider call                                          | —    | None    |
| 182.C.04.06 | Define durable event ownership and compatibility rules for BFF, workers, Studio clients, Veritas, and admin tools.     | Contracts | 182.C.03   | implemented | `docs/domains/isis/adr/ADR-0006-canonical-human-video-contracts.md` (events section)                                                                                                                                                                                                                                                                                                                                                                     | Extends the existing Isis event family on the shared envelope rather than adding a bus - verified `events/envelope.ts` already enumerates `isis` as an `EventSource` and carries correlation/causation/trace ids, and `events/isis.ts` already defines the six job events plus `IsisAssetGenerated`. One producer per event type, with Studio clients and Veritas explicitly owning nothing. Six compatibility rules: additive-only, consumers tolerate unknown enum members, the status enum never grows (stages ride inside the six-state envelope so old consumers stay correct), digests never payloads (keeps biometric material out of the event log), measurements immutable and versioned by evaluator so drift is visible, and breaking changes ship as a new side-by-side type.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Doc - no provider call                                          | —    | None    |
| 182.C.04.07 | Define how direct providers, OpenRouter, Comfy/RunPod, and future self-hosted models implement the same capabilities … | Contracts | 182.C.03   | implemented | `docs/domains/isis/adr/ADR-0006-canonical-human-video-contracts.md` (backend parity section)                                                                                                                                                                                                                                                                                                                                                             | Four backend families reconciled without an intersection payload - which is explicitly rejected, since it would delete exact-dialogue (only two lanes accept audio), edit and extend (direct only), and cap duration at the weakest bound. Three mechanisms instead: a capability vocabulary the request declares in our terms (never vendor parameter names), typed per-lane options validated at the adapter and never widened into the shared request (the `CARD_INPUT_FIELD_ALLOWLIST` discipline one layer down), and capability gaps that fail with a specific reason rather than degrade. Grounded in the real surfaces: verified `comfy-video-provider.ts` already implements the same `VideoProviderClient` seam and documents its ~5s TI2V-5B single-shot cap with segment chaining in `comfy-video-longform.ts`.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Doc - no provider call                                          | —    | None    |
| 182.C.04.08 | Produce a dependency diagram and data-flow threat model covering every copy of face images, voice audio …              | Contracts | 182.C.03   | implemented | `docs/domains/isis/human-video/data-flow-threat-model.md` (new)                                                                                                                                                                                                                                                                                                                                                                                          | Mermaid dependency/data-flow diagram plus a ten-row COPY inventory - the point being that a threat model covering only the happy path misses the copies, which is where likeness and voice actually leak. Twelve threats T1-T12 with mitigations, each tied to an existing enforcement point or a named `182.C` item. Two limits stated honestly rather than engineered around: the provider-side copy is outside our control and revocation cannot recall it (which the product must say at consent time), and OpenRouter video is not ZDR-eligible at all so a ZDR tenant is refused at admission rather than warned. Face embeddings are classified as biometric derivatives, not anonymous numbers.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Doc - no provider call                                          | —    | None    |

### 182.C.05 Canonical Request, Response, and State Contracts

**Recorded dependency (permitted by the stage-order rule).** `182.C.05.01`
assembles the request from the pieces defined in `182.C.05.02`–`182.C.05.08`, so
those are implemented first and `182.C.05.01` is checked last among them. Every
other item in this section keeps its listed order.

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner     | Dependency | Status      | Changed Files                                                                                                                                                                                                                                                                   | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Provider/Hardware Evidence                                              | Cost | Blocker |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------- | ---- | ------- |
| 182.C.05.01 | Add a versioned `HumanVideoGenerationRequest` schema with tenant/project, mode, creative brief, references, cast …    | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/request.ts` (new), `request.spec.ts` (new)                                                                                                                                                                                                      | Every field the task lists is present and versioned: tenant/project, mode, creative brief, references, cast, dialogue (canonical script), audio (per-segment asset+digest), shots, controls, output, quality policy, budget (via the policy ceilings), consent, callback. Based on the BFF `ProviderVideoRequest` per ADR-0006, with `durationSeconds` as the ONLY duration representation - a test asserts no `duration` key survives. `validateHumanVideoRequest` enforces the cross-part rules the schema cannot: text-to-scene refuses references and likenesses, likeness modes require a cast, exact-dialogue requires a script and a releasable tier, two speakers are routed to multi-character mode rather than scored as one, every segment must be delivered in exactly one shot, approved audio must fit the shot durations, and a wide framing is refused when the mode promises identity. `requirementsForRequest` DERIVES the hard requirements from mode and references so a caller cannot under-declare past the router. 175 tests pass across the seven contract modules. | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.02 | Add typed reference roles: identity, face-angle, full-body, wardrobe, pose, composition, first-frame, last-frame …    | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/references.ts` (new), `references.spec.ts` (new)                                                                                                                                                                                                | All fourteen roles present and verified by count: identity, face-angle, full-body, wardrobe, pose, composition, first-frame, last-frame, motion, performance, environment, prop, voice, source-video. The role drives two structural rules rather than a naming convention: `PERSON_BEARING_ROLES` makes subjectId+authorizationId a PARSE error to omit (threat T1 enforced by the type, not by a runtime check somebody forgets), and `ROLE_MEDIA_KIND` rejects a role/kind mismatch. `npx vitest run src/human-video/references.spec.ts` -> 23 passed; `npx tsc --noEmit -p tsconfig.lib.json` -> 0 errors in human-video (15 pre-existing TS2315 Uint8Array errors elsewhere, untouched).                                                                                                                                                                                                                                                                                                                                                                                               | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.03 | Allow multiple references per role with priority, subject binding, temporal range, strength, notes …                  | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/references.ts`, `references.spec.ts`                                                                                                                                                                                                            | Multiple references per role supported and tested (three face-angles for one subject). Carries priority (integer ORDERING within a role, deliberately not a weight - `strength` in [0,1] is the weight), subject binding, temporalRange (half-open, rejected on still roles), notes (never sent to a provider), and provider-safe transform metadata with the post-transform digest. `submittedDigests()` returns what the provider actually received, which is what `182.C.26` binds authorization to. The transform list deliberately excludes anything that would change the depicted person.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.04 | Add cast and speaker IDs so every dialogue segment, reference, blocking instruction …                                 | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/cast.ts` (new), `cast.spec.ts` (new)                                                                                                                                                                                                            | One `castId` namespace, deliberately not a separate speaker-id space - two id spaces for one person is how attribution bugs get written. `CastMemberSchema` enforces the consent story structurally: an `authorized-likeness` requires an authorizationId AND a `generated` member is refused one, so a real likeness cannot be laundered through the generated path. `checkCastReferenceBinding` cross-checks cast against references and distinguishes the two real failures - a reference naming a non-cast subject (identity results would have no owner) versus an authorized likeness with no reference (the identity gate could only report `unknown`); a generated member with no references is correct and is not reported. `requiresSpeakerAttribution` drives the C3 gate. `npx vitest run src/human-video/` -> 37 passed.                                                                                                                                                                                                                                                       | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.05 | Add canonical dialogue segments with exact text, locale, pronunciation lexicon, emotion, delivery, start constraint … | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/dialogue.ts` (new), `dialogue.spec.ts` (new)                                                                                                                                                                                                    | All nine required fields present - exact text, locale (BCP-47 with code-switch spans), pronunciation lexicon, emotion, delivery, start constraint (discriminated union), overlap rule, audio asset + digest. The shared normalizer lives here on purpose: two normalizers would let a take pass the promise side and fail the measurement side. Real domain logic, not a shape - contraction expansion, an English cardinal expander so "42" and "forty-two" compare equal, and a REFUSAL for locales it has no rules for (returns normalized:false rather than applying English rules to Portuguese and inventing differences). Writing the tests surfaced a genuine defect: `not`/`no` in the function-word list made "we will ship" and "we will not ship" identical under C2, so negations are now content words with a test pinning it. `npx vitest run src/human-video/` -> 82 passed; adversarial scan clean.                                                                                                                                                                        | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.06 | Add provider-neutral controls for duration, aspect, resolution, FPS, camera, lens, framing, shot size, motion …       | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/controls.ts` (new), `controls.spec.ts` (new)                                                                                                                                                                                                    | All nineteen control categories present in OUR vocabulary, never a vendor parameter name (ADR-0006 section 6): duration, aspect, resolution, fps, camera (grip vocabulary - pan/tilt/dolly/truck/pedestal/crane/orbit/handheld/steadicam), lens (35mm-equivalent mm + f-number), framing, shot size, motion, blocking, expression, gaze, gesture, pace, lighting (rembrandt/butterfly/split/rim/high-key/low-key), style, negative constraints, seed, continuity. Two product-contract rules are COMPUTED rather than left as prose: `shotSizeSupportsFaceEvaluation` refuses wide framings when the mode promises identity or sync (the tiny-faces limitation - cheaper to refuse than to generate and score `unknown`), and `checkControlCoherence` binds per-cast direction and gaze targets to real cast members. Blocking is deliberately prose, not coordinates we could not verify against a rendered frame. Fixed a zod-v4 trap found by the tests: `.default({})` does not re-parse, so nested defaults were silently dropped. 100 tests pass.                                     | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.07 | Add `requiredCapabilities` distinct from preferences so the router can fail closed instead of silently discarding …   | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/capabilities.ts` (new), `capabilities.spec.ts` (new)                                                                                                                                                                                            | Requirements and preferences are separate types with separate consequences: an unmet REQUIREMENT fails the match with one of the twelve `182.C.03.07` reasons, an unmet PREFERENCE is reported in `unmetPreferences` and never fatal. `matchCapabilities` fails closed by construction - a lane is eligible only when every requirement is affirmatively satisfied, so there is no code path that drops a requirement and routes anyway. Each of the five silent downgrades named in the task has a test proving it fails instead: native audio, multi-reference count, multi-speaker, extension, and resolution. Probe discipline is enforced separately from advertisement (`probed` is a subset of `capabilities`), so an advertised but unprobed capability yields `capability_unprobed` unless the caller IS the probe. `primaryReason` reports the most common cause deterministically rather than whichever lane was checked last. 122 tests pass; adversarial scan clean.                                                                                                           | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.08 | Add a quality policy with metric thresholds, candidate and retry budgets, allowed fallback classes …                  | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/quality-policy.ts` (new), `quality-policy.spec.ts` (new)                                                                                                                                                                                        | All five parts present: metric thresholds (inherited in-repo defaults plus the ATSC/ITU sync windows, with the review band falling out of the gap between them), candidate and retry budgets per tier (1/2/4/6 and 0/1/2/3), allowed fallback classes, human-review rules, and max spend/latency. The tier rule is structural, with a test: every tier gets the IDENTICAL thresholds, so a tier buys work and never a waived gate. `FALLBACK_FOR_CONDITION` stops a retry being a second roll of the dice - a re-download cannot fix bad sync, a reseed cannot fix a paraphrase - and `evaluateRetry` refuses a retry that names no failed condition (that is a candidate, not a retry). `exhaustionOutcome` can only return `blocked` or `needs_review`, never a delivery. 149 tests pass; adversarial scan clean.                                                                                                                                                                                                                                                                         | Contract + unit tests; thresholds still uncalibrated pending `182.C.34` | —    | None    |
| 182.C.05.09 | Add a response contract containing durable asset URLs, provider/model/version, seed, timing, cost, measurements …     | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/response.ts` (new), `response.spec.ts` (new)                                                                                                                                                                                                    | All eleven required elements present: durable asset URLs (ours, never a provider URL carrying our token), provider/model/version (dated canonical slug, not a floating alias), seed, timing, cost, measurements, warnings, candidate lineage, provenance, release state. The tri-state measurement is enforced by the schema, not by convention - an `unknown` outcome may NOT carry an observed value and MUST carry a reason, and a decided outcome must carry the value it was decided on, so "not measured" can never be recorded as "measured as zero". `evaluateConditions` implements the conjunction and keeps three distinct not-pass states: measured failure, `unknown`, and MISSING (a pipeline bug, reported separately from an honest limit) - and no measurements at all is `unknown`, never a pass. `selectCandidate` only ever returns a fully-passing candidate. `totalCandidateCost` bills every candidate including rejects, and returns null on mixed currencies rather than summing into one dishonest number. 23 tests pass; adversarial scan clean.                 | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.10 | Extend the existing `GenerationJobStatus` in `jobs-route.ts` (queued, running, complete, failed, blocked …            | Contracts | 182.C.04   | wired       | `libs/contracts/src/human-video/job-state.ts` (new), `job-state.spec.ts` (new), `index.ts` (new), `libs/contracts/package.json`, `tsconfig.base.json`, `apps/oshun/bff/src/generation/jobs-route.ts`                                                                            | All fifteen stages defined and each mapped onto one of the six durable statuses, exactly as the task asks and without a parallel enum: `blocked` and `needs_review` keep coming from the release decision (`statusForStageAndGate`), never from progress. The status enum is NOT grown - a test asserts all ten in-flight stages read as `running`, so a consumer built against the six stays correct. Transition rules encode two real constraints: corrective processing must return through `evaluating` (a corrected take is re-measured, never trusted straight to encode), and nothing may skip evaluation on the way to release. Wired for real: `GenerationJob.stage?: HumanVideoStage` added to the BFF envelope (optional, so existing PostgreSQL snapshots stay readable), with the `@oshun/contracts/human-video` subpath export and tsconfig path added. `npx vitest run src/human-video/` -> 214 passed; `npx vitest run src/generation/jobs-route` in the BFF -> 37 passed.                                                                                                  | Contract + unit tests + BFF envelope; no provider call                  | —    | None    |
| 182.C.05.11 | Version schemas with forward/backward compatibility tests and reject unknown security- or consent-critical fields.    | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/versioning.ts` (new), `versioning.spec.ts` (new), `index.ts`                                                                                                                                                                                    | Both halves done and tested in both directions. BACKWARD: a new build reading an older payload fills every omitted optional, including nested control defaults. FORWARD: an unknown field outside the critical set is stripped and tolerated (top-level and nested), while a wrong VALUE still fails - new is tolerated, invalid is not. The consent-critical set (`consent`, `callback`, `cast[]`, `references[]`) is REFUSED rather than stripped, because a dropped consent field reads to the caller as an accepted one - a client sending `bypassConsent` or `authorizationOverride` gets a typed `UnknownConsentFieldError` naming every offending path, not a silent strip and a success. Detection runs on the RAW payload before zod strips the evidence, and the version check runs first so an unreadable payload fails with a version error rather than a field error. 232 tests pass across the module; adversarial scan clean.                                                                                                                                                | Contract + unit tests; no provider call                                 | —    | None    |
| 182.C.05.12 | Generate OpenAPI/JSON Schema/SDK types from the canonical source and prohibit hand-copied client types.               | Contracts | 182.C.04   | implemented | `libs/contracts/src/human-video/schema-registry.ts` (new), `schema-publication.spec.ts` (new), `libs/contracts/scripts/generate-human-video-schemas.ts` (new), `libs/contracts/schemas/human-video/*` (28 generated), `libs/contracts/package.json`, `.github/workflows/ci.yml` | Follows the established study-schema pattern: one registry, a generator emitting 26 JSON Schemas (draft 2020-12) plus an OpenAPI 3.1 components document and a manifest, and a `--check` drift mode wired into CI beside the existing study checks. Drift is what makes the registry canonical in practice - a schema cannot change without regenerating, so a generated SDK can never quietly describe last month contract. The hand-copy prohibition is enforced by a test that git-greps the repo for re-declarations. It guards FOUR names, not all 26, and the reason is measured rather than assumed: twelve published names are ordinary words unrelated domains already use (`Candidate` 22 declarations, `Provenance` 8, `Shot` 8, `ReferenceAsset` 4, and eight others), so guarding them would flag a Veritas camera `Shot` as a hand-copied human-video type. The four kept are the ones an SDK client would actually copy; the rest rely on the drift check. 241 tests pass.                                                                                                   | Generated artifacts committed; drift check green                        | —    | None    |

### 182.C.06 Reference Asset Ingestion and Validation

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner       | Dependency | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | Cost | Blocker                                                                                                                                                                                                                                                                                                                                                                                                                          |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | ----------- | ---------- | ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.06.01 | Build resumable, signed image/audio/video upload sessions with tenant-scoped object keys, quotas, MIME sniffing …     | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-upload-session.ts` (the port is ASYNC now, and the digest map is gone), `human-video-upload-bytes.ts` (new: the sink port, a filesystem implementation, a memory one for tests), `human-video-upload-store.ts` (new: the Prisma adapter), `human-video-upload-bytes.spec.ts`, `human-video-upload-store.spec.ts`, `human-video-upload-vertical.spec.ts` (all new), `human-video-asset-handlers.ts` (three new handlers), `human-video-api-surface.ts` (three new operations, 32 total), `human-video-api-semantics.ts` (`abandon-upload-session` is a safety retraction), `human-video-api-routes.ts` (an encapsulated octet-stream parser), `human-video-runtime.ts` (the binding), `human-video-sdk-generator.ts` (binary bodies), `sdks/human-video-ts/client.ts` (regenerated), `libs/oshun/persistence/prisma/migrations/20260912010000_human_video_upload_session/migration.sql` (new), `schema.prisma`, `prisma-renderer.ts` | DELIVERED 2026-09-12, AND THE THING IT FIXED IS WORSE THAN THE THING IT ADDED. The running SHA-256 lived in a module-level Map and `completeUploadSession` read it with a FALLBACK: `createHash('sha256').digest('hex')` — THE DIGEST OF AN EMPTY FILE — whenever the Map had no entry, which is every resume after a restart and every completion on a second worker. `182.C.26` binds an AUTHORIZATION to that value, so every such record would have bound to the same constant and ANY TWO SUBJECTS WOULD HAVE APPEARED TO CONSENT TO THE SAME BYTES. The digest is now computed at completion BY READING THE STORED OBJECT BACK: process-independent, and it hashes what a later reader will actually get, so a sink that lost a chunk gives a mismatch rather than a confident wrong answer. A SESSION WAS NOT AN UPLOAD: the surface declared `POST /upload-sessions` and a `GET` that answers a `nextOffset` — send the next chunk here — AND THERE WAS NOWHERE TO SEND IT. Three operations added, and the chunk's body IS the chunk: `application/octet-stream`, parsed in the human-video routes' OWN ENCAPSULATED Fastify scope so the parser is not imposed on the rest of the BFF, and declared on the endpoint so the generated SDK sets the header it will be parsed under — a generated client that cannot call the endpoint it was generated from is the exact drift the generator exists to prevent. The offset is IN THE PATH rather than a query parameter, so the SDK carries it. THE PORT WAS SYNCHRONOUS, which is why nothing ever implemented it; it is async now, and that is most of the diff. THE RESUME OFFSET COMES FROM THE BYTES, NOT THE ROW: a row ahead of the object makes a client resume past the end and leave a hole in a file whose length still comes out right, so the sink is asked on every append and the row is repaired from it, and the row is written BEFORE the bytes so a crash leaves it behind rather than ahead. ABANDONING DESTROYS THE BYTES and returns a manifest rather than a 204, because an abandoned upload left staged is a copy of somebody's face that no record points at — outside every retention rule and every erasure sweep. BINDING NEEDS BOTH HALVES: a session table with no staging directory would accept an open, answer a `nextOffset` and drop every chunk into nothing, which a client cannot tell from success, so the runtime binds neither and the boot line says which half is missing. There is NO DEFAULT staging directory — an implicit one takes biometric bytes somewhere nobody chose.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | 5370 human-video BFF tests green. 10 end-to-end tests over a real Fastify instance and a real filesystem sink, including the one that matters: a completion on a store rebuilt from scratch — a restart, a second worker — still digests correctly, and the test asserts the answer is NOT the empty-file hash. 17 sink tests, run against BOTH implementations, covering the sparse-hole refusal, path escape, and refusing to digest an absent object. 13 store tests including that the tenant byte footprint is SUMMED IN THE DATABASE rather than loaded. The bounded-memory test from 182.C.16.02 now runs against a real filesystem sink, because an in-memory one holds every byte by design and the test would have passed while measuring the opposite of its claim. The migration was applied to a probe database and every CHECK constraint was fired individually — including the one that refuses a completed session with no digest, which is the row the old defect produced. Scratch typecheck and eslint clean; SDK regenerated and its drift guard passes. | —    | The sink stages to local disk and `finalize` is a no-op. A resumable upload straight to S3 is a multipart upload with a 5 MiB floor on every part but the last, which browser chunk sizes do not obey, so staging first is the shape that works — but the handover to object storage is DECLARED AND NOT WRITTEN. A deployment with more than one BFF node needs shared storage under OSHUN_HUMAN_VIDEO_UPLOAD_ROOT until it is. |
| 182.C.06.02 | Validate dimensions, duration, codecs, sample rate, channels, bit depth, frame rate, orientation, color profile …     | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-asset-validation.ts` (new), `human-video-asset-validation.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Every property the task lists is checked from a real ffprobe stream read, never from the declared content type: dimensions, duration, codecs (per-kind allowlists), sample rate, channels, bit depth, frame rate, orientation, color profile, alpha, and corruption. Split so the rules are testable without a binary - `mapFfprobeOutput` is a pure mapping of ffprobe JSON (tested against real-shaped documents) and `validateAssetProbe` is a pure policy check that returns EVERY finding rather than the first, so a creator fixing an upload sees all defects in one pass. Two rules that are easy to get backwards are handled explicitly: rotation is applied before any size rule (a 1080x1920 stream tagged rotate=90 is a 1920x1080 picture, and a face detector fed the unrotated frame sees a sideways face), and alpha is REFUSED because every lane composites onto opaque frames so transparency silently becomes black. Corruption detection is the ffprobe failure itself - a container ffprobe cannot parse is one no downstream stage can parse either. Missing values report `*_unknown` rather than defaulting to zero. 34 tests pass; adversarial scan clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Unit-tested against captured ffprobe document shapes; no binary in the test path                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | —    | None                                                                                                                                                                                                                                                                                                                                                                                                                             |
| 182.C.06.03 | Strip unneeded EXIF/GPS/device metadata while retaining color and orientation data required for correct processing.   | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-metadata-strip.ts` (new), `human-video-metadata-strip.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Strips at the CONTAINER level with no re-encode, so pixel data is bit-identical and the operation cannot alter the likeness it protects (a test pins the JPEG scan data and the PNG IDAT as unchanged). JPEG: every APPn and COM segment is dropped - JFIF, EXIF, XMP, comments - except APP2/ICC. PNG: tEXt/zTXt/iTXt/tIME/eXIf/pHYs dropped, iCCP/gAMA/cHRM/sRGB/sBIT kept. The two things usually stripped by ACCIDENT are handled deliberately: the ICC profile survives (dropping it reinterprets pixels as sRGB and shifts skin tones - a change to the very thing the identity gate measures), and the EXIF orientation is RETURNED to the caller before being removed, since a stripped orientation nobody recorded turns a portrait photo sideways downstream. Includes a minimal TIFF reader for tag 0x0112 handling both byte orders, returning null on anything malformed rather than inventing a rotation. Idempotent, and refuses an unsupported container rather than passing an unstripped file through as clean. 22 tests pass; no new dependency added.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Unit-tested against synthesized JPEG/PNG byte streams                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | —    | None                                                                                                                                                                                                                                                                                                                                                                                                                             |
| 182.C.06.04 | Scan uploads for malware, decompression bombs, malformed containers, adversarial payloads …                           | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-upload-scan.ts` (new), `human-video-upload-scan.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | All five hazards covered, with the split that matters made explicit. DECIDABLE FROM THE BYTES, computed exactly here: decompression bombs (pixel ceiling plus a decoded-to-compressed expansion ratio, applied to header dimensions BEFORE anything decodes, and multiplied by frame count for video; the ratio is skipped below 1 KB where it is meaningless, so a 64x64 icon is not a false positive) and trailing payloads (data after JPEG FFD9 or PNG IEND - a polyglot that is an image to one parser and a ZIP to another; refused rather than trimmed, since trimming would silently alter an uploaded asset). STRUCTURAL: malformed containers caught here before a decoder sees them, and again by ffprobe at `182.C.06.02`. INJECTED FAIL-LOUD SEAMS: malware scanning and policy content, both following the `dub-faithfulness.ts` pattern - an unbound seam THROWS, because returning `clean: true` with no scanner would be a verdict indistinguishable from a real one and false. Reuses the existing `ImageSafetyClassifier` rather than adding a second moderation path. Scanner id and definitions version are recorded even on a clean verdict so it is auditable. 24 tests pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Unit-tested with injected seams; no live scanner bound in this environment                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | —    | Malware scanner and moderation credentials must be bound at deploy                                                                                                                                                                                                                                                                                                                                                               |
| 182.C.06.05 | Detect faces/people and reject ambiguous identity packs until the creator maps each face to a cast member.            | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-identity-pack.ts` (new), `human-video-identity-pack.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-10-openrouter-face-detection/` (new), plus the adapter 2026-09-12: `apps/oshun/bff/src/generation/human-video-pack-store.ts` (new), `human-video-pack-store.spec.ts` (new), `human-video-runtime.ts` (the binding), `libs/oshun/persistence/prisma/migrations/20260912020000_human_video_identity_pack/migration.sql` (new), `schema.prisma`, `prisma-renderer.ts`                                                                                                                                                                                                                                                                                                                                                                                                                                              | Unblocked 2026-09-10 by the OpenRouter key, but the probe changed the DESIGN rather than just supplying a detector. Two vision models given the same generated office photo returned 2 and 6 faces - one counted the foreground subjects, the other counted the background workers too - and both reported confidence. Inspecting the image confirmed both readings are correct; the question was under-specified, not the models. So an exact face count is an artefact of model and prompt, while "is this pack unambiguous?" is robust: the gate keys on AMBIGUITY, with a test asserting the same verdict from either reading (2 or 6 or 12 faces). More than one face holds the reference for the creator to map, which is a question rather than a rejection; zero faces is a hard finding since an identity reference with no face cannot feed the identity gate. Mapping validation covers unknown face, unknown reference, non-cast member and conflicting mappings. The detector is an INJECTED PORT that is unconfigured by default and throws when unbound - binding it to a hosted model sends reference faces to a third party at ingestion (copy 4, unrecallable), which is a deployment decision rather than a code one. Probe cost $0.0002-0.0003 per detection. 14 tests pass. BOUND 2026-09-12, AND BINDING NEEDS BOTH HALVES: a pack store with no face detector would accept a pack nothing examined, and since `evaluateIdentityPack` REJECTS a reference with no face in it, an absent detector answering "no faces" would tell a creator their photograph has nobody in it — so the runtime binds NEITHER and the boot line says which half is missing. THE LISTING IS KEYSET ON `(created_at, id)` AND THE INDEX CARRIES THE SAME PAIR: ordering on `created_at` alone SKIPS A ROW whenever two packs share a millisecond, and the page comes back short by one, forever, with nothing reporting it. `limit + 1` rows rather than a COUNT over the tenant's whole history. A ROW THIS BUILD CANNOT READ IS REFUSED rather than guessed at — an unknown state read as `accepted` lets a job run against an unexamined reference set — and an EMPTY reference array is refused as firmly as a missing one, matching the migration's own CHECK.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Live provider probe; fixture + counts committed as evidence 11 store tests over a delegate double, covering the keyset pair, the `limit + 1`, and the four unreadable-row refusals. 4 runtime tests for the two-halves binding. The migration was applied to a probe database and every CHECK fired individually — WHICH IS HOW THE `cast_ids` ONE WAS CAUGHT: `array_length('{}', 1)` is NULL, not 0, and a CHECK whose expression is NULL is SATISFIED, so the constraint accepted exactly the row it existed to refuse until COALESCE was added.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —    | The detector is a local CPU process; a hosted one is still a deployment privacy decision.                                                                                                                                                                                                                                                                                                                                        |
| 182.C.06.06 | Compute face visibility, pose/yaw/pitch/roll, blur, exposure, occlusion, crop, resolution, full-body visibility …     | Media Infra | 182.C.05   | complete    | `apps/oshun/bff/src/generation/human-video-reference-quality.ts`, `human-video-reference-quality.spec.ts`, `tools/human-video/body-geometry.py` (new), `tools/human-video/face-occlusion.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-cpu-occlusion-and-body/` (new), plus the adapter 2026-09-12: `apps/oshun/bff/src/generation/human-video-face-geometry-adapter.ts` (new) and `human-video-face-geometry-adapter.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | COMPLETE. The seven metrics measured earlier the same day stand (ffmpeg blur/exposure, YuNet box + SQPnP pose, dHash duplicates); the two that were outstanding are now measured too, on this box's CPU - `docs/domains/isis/human-video/evidence/2026-09-10-cpu-occlusion-and-body/`. **Occlusion** comes from MediaPipe's multiclass selfie segmenter (Apache-2.0, 16 MB): within the YuNet face box, face-skin share against clothes-plus-accessory share. Four images derived from the committed probe photo by single ffmpeg commands: clear 0.835 face-skin / 0.000 obstruction, lower-face blocked 0.434 / 0.487, eye band blocked 0.568 / 0.266, a real hand crop composited over the face 0.248 / 0.281 - while the second, untouched face in the same photo held 0.733-0.840 through every variant, which is the control. Thresholds sit in the measured gaps (0.65 face-skin, 0.15 obstruction). Body-skin share is recorded but NOT thresholded: a clear face carries 0.04-0.06 of it at the neck and the hand case reached only 0.095, one reading being no basis for a rule. **Full-body visibility** comes from MediaPipe Pose Landmarker Lite (Apache-2.0, 5.5 MB), which is also the person detector this item asked for - it found exactly the two foreground subjects in a photo with five detectable faces. A region counts as shown only when presence clears 0.5 AND the predicted coordinates fall inside the frame, because the model reports coordinates for landmarks it places below the bottom edge (ankles at y=1.21, presence 0.045). Cropping the same subject to head-and-shoulders drops the hips from 0.997 to 0.001. **A named failed attempt is recorded**: MediaPipe's own per-landmark `visibility` was tried first and does not respond to occlusion - a black block over a mouth moved it 0.998 to 0.999, over a wrist 0.754 to 0.758 - so `body-geometry.py` does not report that field at all. The head-height estimate is kept as the labelled fallback (`fullBody.source` is `measured` or `estimated`) and only a measurement raises the warning, since most identity references are portraits. 46 tests pass. Not claimed: occlusion of the BODY - the scorer's occlusion issue is scoped to the face box. THE TOOL IS BOUND 2026-09-12. Its own docstring said it was "used by the BFF's reference-quality scoring through an injected provider port" and NOTHING WAS INJECTED; this is the child process, with an argv array so a path can never become a command. ZERO FACES AND A DETECTOR THAT DID NOT RUN ARE THE SAME SHAPE AND MUST NOT BE THE SAME ANSWER: every process failure, unparseable output, empty result and timeout THROWS, because `evaluateIdentityPack` rejects a reference with no face and a crashed interpreter would refuse a perfectly good photograph. AND A FACE ID HAS TO SURVIVE A SECOND RUN, WHICH THE TYPE DOES NOT SAY: `DetectedFace.faceId` is documented as stable within ONE run, but the awaiting-mapping flow re-runs the detector after the creator answers, and YuNet orders by confidence which can flip between near-equal faces — so the id is derived from WHERE the face is, quantised to whole percent of the frame, and two faces quantising to one id are refused rather than deduplicated. | CPU on this machine; measurements committed as evidence 19 adapter tests, 17 against an injected runner covering every failure mode and 2 RUNNING YUNET FOR REAL against a committed evidence image — measured: five faces found, the same five ids on a second run, all distinct. Skipped with a stated reason where the weights are not staged, so a box with no model asserts nothing rather than passing silently.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | —    | None                                                                                                                                                                                                                                                                                                                                                                                                                             |
| 182.C.06.07 | Normalize provider-bound assets to accepted formats without silently changing aspect, speech timing …                 | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-normalization.ts` (new), `human-video-normalization.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | All four "without silently changing" invariants are enforced by construction, not by review. ASPECT: scaling is uniform on the longest edge, dimensions round to even on BOTH axes, and a resulting ratio drift over 0.5% throws rather than shipping a reframed reference - no crop or pad is emittable. SPEECH TIMING: resampling is duration-preserving and says so; `atempo`/`asetrate` are on the forbidden list. CHANNEL ALIGNMENT: a downmix THROWS unless explicitly authorized, because dropping a channel can drop a speaker, while an upmix proceeds (no content lost) - and the authorization is recorded in the plan. IDENTITY: the emittable filter set is enumerated (`scale`, `format`, `aformat`) and every identity-affecting filter (unsharp, denoise, eq, curves, colorbalance) is banned. `assertPlanInvariants` re-checks the produced argv as a second line of defence, so a future caller assembling its own arguments still cannot smuggle a crop past it - with a test that it does not false-positive on a path containing "padded". Audio is stream-copied when nothing about it changed, avoiding a needless generation loss on approved speech. 24 tests pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Unit-tested; ffmpeg execution is injected, argv is asserted                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —    | None                                                                                                                                                                                                                                                                                                                                                                                                                             |
| 182.C.06.08 | Generate low-risk thumbnails/proxies separately from biometric originals and enforce the same authorization on both.  | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-derivatives.ts` (new), `human-video-derivatives.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Both halves of the task, with the trap named in the module header: "low-risk" here means low BANDWIDTH, never low authorization - a 128px thumbnail of a face is still that face. SEPARATE STORAGE: derivatives get their own top-level prefix (`human-video-derived/`), so a bulk grant written against the originals prefix cannot accidentally span them, with per-kind key separation so one derivative cannot overwrite another. SAME AUTHORIZATION: `decideDerivativeAccess` resolves THROUGH to the source rather than trusting the derivative - a function rather than a convention, because the failure it prevents is silent (a revoked likeness still visible as a thumbnail looks exactly like correct behaviour). Cross-tenant access is refused even with a matching authorization id, and an orphaned derivative is refused rather than served. Derivation from a biometric asset with no authorization throws. Erasure is inherited: `erasureTargetsFor` computes the original plus its derivatives so no caller has to remember they exist, and does not sweep another asset. Per-kind size bounds refuse a full-size copy masquerading as a thumbnail. 19 tests pass. **2026-09-10:** a real type error in the spec fixture (`Partial<Derivative>` overrides spread into `registerDerivative`'s `RegisterDerivativeInput`, papered over with `as never`) surfaced once the BFF typecheck was re-run against `tsconfig.typecheck.json` rather than `tsconfig.json`, which sets `noCheck: true`. Fixed by typing the fixture's overrides as `Partial<RegisterDerivativeInput>`.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Unit-tested; no object store in the test path                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | —    | None                                                                                                                                                                                                                                                                                                                                                                                                                             |
| 182.C.06.09 | Support authorized HTTPS pull with SSRF protection, allowlisted redirects, size/time limits, content validation …     | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-asset-pull.ts` (new), `human-video-asset-pull.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Every guard the task names, with the SSRF check done on RESOLVED ADDRESSES rather than hostnames - a hostname allowlist cannot see `internal.example.com` resolving to 10.0.0.5, which is the actual attack. Full private-range table including 169.254.0.0/16 (cloud metadata, the range that hands out credentials), CGNAT, TEST-NETs, multicast and reserved; IPv6 loopback, ULA, link-local and IPv4-mapped forms (`::ffff:10.0.0.5` is refused, not just the bare spelling). Ranges are matched by parsed 32-bit value against a CIDR mask, not string prefixes. Redirects are followed MANUALLY and re-validated per hop - delegating them to the HTTP client is the single most common bypass, and a test drives the exact scenario (permitted first hop 302ing to the metadata endpoint). A literal IP in the URL is checked without consulting DNS at all. Size is bounded on the STREAMED count as well as the declared one, since a Content-Length can lie, and the declared check runs before the body is read at all. Writing the timeout test surfaced a real gap: the deadline was only checked BETWEEN steps, so a hanging origin was unbounded - now the resolver and the fetch are each raced against the remaining budget. 54 tests pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Unit-tested with an injected fetch and resolver; no live network                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | —    | None                                                                                                                                                                                                                                                                                                                                                                                                                             |
| 182.C.06.10 | Record an immutable asset lineage from original through crop, normalization, provider upload, candidate, correction … | Media Infra | 182.C.05   | implemented | `apps/oshun/bff/src/generation/human-video-asset-lineage.ts` (new), `human-video-asset-lineage.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | All nine stages the task names, from original through crop, normalization, provider upload, candidate, correction and release. IMMUTABILITY IS ENFORCED, not documented: each entry carries a chain digest over its own fields plus its predecessor chain digest, so an edited producer, a swapped byte digest, a moved timestamp, a removed entry or a reordered pair all fail `verifyLineage` - each has its own test. Append returns a NEW lineage rather than mutating, stage order is checked (nothing runs backwards, nothing repeats, nothing follows `released`), and skipping stages that did not happen is allowed. Three queries make the record useful rather than decorative: `providerSubmittedDigest` returns the POST-transform digest the provider actually received (what `182.C.26` binds authorization to, not the uploaded one), `releasedDigest`, and `measurementMatchesRelease`, which enforces `182.C.01.05` by rejecting a measurement of the pre-watermark candidate and returning false when nothing has been released at all rather than granting unearned assurance. 23 tests pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Unit-tested; tamper cases exercised explicitly                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | —    | None                                                                                                                                                                                                                                                                                                                                                                                                                             |

### 182.C.07 Identity Pack and Character Continuity

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner | Dependency | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                            | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Provider/Hardware Evidence                 | Cost                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Blocker                                                                |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------- | --- | -------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.07.01 | Extend the existing identity-profile contract into a versioned character pack with subject ID, authorized references … | Isis  | 182.C.06   | complete    | `libs/contracts/src/human-video/character-pack.ts` (new), `character-pack.spec.ts` (new), `schema-registry.ts`, `schema-publication.spec.ts`, `index.ts`, `libs/contracts/schemas/human-video/` (4 new schemas + manifest + openapi), `libs/isis/workflows/src/consistency/identity-profiles.ts`, `identity-profiles.spec.ts`                                                                                            | `CharacterPackSchema` in `@oshun/contracts/human-video` carries all seven fields the item lists - subjectId (the same id space as `castId` and `ReferenceAsset.subjectId`), authorizedReferenceIds, appearanceTraits, allowedTransformations, wardrobe, voice binding, and a consent record with its version - plus `packVersion`/`revision`/`supersedesRevision`, so a finished project can be re-rendered against the pack it used. The rule the file exists for: an IMMUTABLE trait can never be an allowed transformation. That is enforced by the enum itself rather than a refine, because a refine cannot be expressed in JSON Schema and the published artifact would have told SDK clients that `face-structure` is a legal value - caught by inspecting the generated schema, and the enum carries a custom error so a consent review reads "identifies the person and can never be an allowed transformation" instead of "invalid option". Four further rules that are relationships, not shapes, live in `checkCharacterPackBinding`: a reference must exist in the request, be person-bearing, depict THIS subject, and carry the pack's own authorization; a pack may not attach to a `generated` cast member (the same reasoning `cast.ts` applies to authorizationId); wardrobe and voice references must have the matching role; an expired authorization is reported. Self-contradiction is refused too: more than one wardrobe with `wardrobe` not listed as changeable. EXTENDS rather than replaces the Isis identity profile: `IdentityProfileSnapshot` gains `characterPackId` + `characterPackRevision`, and half a join (either without the other) throws. Embeddings and adapter weights stay in Isis per ADR-0006 §6. Published through the `182.C.05.12` registry: 4 new JSON Schemas, OpenAPI components regenerated, the diff purely additive (329 insertions, 0 deletions, no existing schema changed). `CharacterPack` added to `GUARDED_TYPE_NAMES` after confirming no other declaration of that name exists in the repo. 277 contract tests pass (30 new); 11 identity-profile tests pass (7 new); tsc reports zero errors in the touched files.                                                                                                   | Contract work — no provider call           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None                                                                   |
| 182.C.07.02 | Build guidance for a useful reference set: clean face, three-quarter/profile angles, full body, expression range …     | Isis  | 182.C.06   | complete    | `apps/oshun/bff/src/generation/human-video-reference-guidance.ts` (new), `human-video-reference-guidance.spec.ts` (new), `human-video-reference-quality.ts` (role-aware scoring), `human-video-reference-quality.spec.ts`                                                                                                                                                                                                | `assessReferenceSet` scores a whole PACK rather than one photo, over six dimensions, each TRI-STATE - `satisfied` when a measurement says so, `missing` when a measurement says it is not, `not-measured` when the input was never bound. Unknown is never coerced to pass, matching ADR-0006's rule for delivery measurements. It reads the per-reference verdicts `182.C.06.06` already produced instead of re-deriving them, so the two cannot disagree. **A real defect surfaced by writing this and fixed with it:** the per-reference scorer applied `maxYawDegrees` 35 and `minFaceRelativeArea` to EVERY reference, so it rejected exactly the three-quarter/profile and full-body references this item tells creators to provide. `scoreReferenceQuality` now takes the reference role: an `identity` reference is bounded at 35 degrees as before, a `face-angle` reference is bounded at 60 - the side-profile limitation the product contract already publishes (`182.C.03.08`), reused rather than invented so a reference is not accepted at an angle delivery would flag - and a small face in a `full-body` reference is a warning rather than a rejection. An unlabelled reference is still scored as `identity`, the strictest reading. Thresholds are derived, not picked: the three-quarter band is [15, `maxFaceAngleYawDegrees`], one source for "too far round"; the lighting-range requirement is a quarter of the exposure band the scorer itself accepts, and the fraction is stated as the choice it is. Two dimensions report `not-measured` with the reason in the finding: EXPRESSION RANGE (blendshapes are available and `face_landmarker.task` is staged, but the threshold separating two expressions needs the `182.C.34` corpus, and a number invented before it would be a constant wearing a measurement's name) and WARDROBE without a character pack (intended wardrobe is a creator statement the pack carries, read from it when one is passed). 20 guidance tests plus 5 regression tests for the role fix; 96 tests pass across the three human-video BFF specs; tsc reports zero errors in the touched files.                                                                                                                            | Contract + scoring work — no provider call | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None                                                                   |
| 182.C.07.03 | Detect mixed identities, face swaps, age mismatches, and severe reference contradictions before spending provider …    | Isis  | 182.C.06   | implemented | `apps/oshun/bff/src/generation/human-video-identity-consistency.ts` (new), `human-video-identity-consistency.spec.ts` (new), `tools/human-video/face-embedding.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-cpu-identity-consistency/` (new) , `human-video-face-swap.ts` (new 2026-09-11), `human-video-face-swap.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-face-swap/` (new) | COMPLETED 2026-09-11; the checkbox is now [x]. Both detections that were outstanding are resolved and neither by hand-waving. FACE SWAP IS NOW DETECTED, by identity-attribute inconsistency rather than by a manipulation detector - which is still unstaged and would still be the thing that catches the rest. It works because a swap is not a change of person, it is a change of FACE on a person who is otherwise still there. Measured on a REAL swap built from the committed frame: subject A's face region composited into subject B's face box with a feathered elliptical mask so the edges do not read as a tear. The swapped face embeds as A at 0.737 and as B at 0.381 while the HAIR around it measures lightness 27.06 - B's hair (26.67), not A's (9.80). Neither reading is wrong alone and together they are impossible. BOTH HALVES ARE REQUIRED: a face matching nobody is an unknown face and hair matching nobody is a haircut, and the spec asserts both non-cases. A third signal fell out - the swapped face's SKIN measured 49.41, between A's 45.49 and B's 70.59, because a feathered composite mixes the two - and it is reported SEPARATELY, since a blend and a clean paste leave different traces. Blind spots are named in code rather than implied: a whole-head replacement where the hair travels with the face, a swap between lookalikes, a fully synthesised person, and any manipulation leaving face and surround the same person; a clean result from a consistency check is not proof of no manipulation. The comparison is gated to ONE FRAME by assertSameFrame, because hair and skin are colours and 182.C.19.02 measured a stop of exposure moving skin tone 58 degrees of ITA. AGE MISMATCH IS RESOLVED AS HUMAN REVIEW BY DECISION, not for want of a model: 182.C.19.02 recorded on 2026-09-10 that estimating age from face geometry has a documented record of skewing by ethnicity, so a number here would be worse than none. That is a different state from 'pending a model' and UNMEASURED_FINDINGS now says which, pointing the reviewer at the pack's declared apparent-age-band from 182.C.07.01.                                                                                                                    |                                            | EARLIER (2026-09-10): PARTIAL - the checkbox STAYS `[ ]` because two of the four detections this item names are not delivered, and both are named in the result rather than left silent. **Delivered: mixed identities (mechanism) and reference contradictions (complete).** `human-video-identity-consistency.ts` compares every pair of references over an injected embedding port; `tools/human-video/face-embedding.py` (new) is the CPU backend - YuNet for the box and landmarks, SFace for the 128-d vector, both Apache-2.0. Measured on the committed probe photo (`evidence/2026-09-10-cpu-identity-consistency/`): two subjects at 0.261, each subject against a degraded copy of itself at 0.739 and 0.819, cross pairs at 0.146-0.177. **The cutoff is deliberately absent.** The earlier probe found a stranger pair at 0.422, above the inherited 0.35 gate, so a threshold picked from five clean pairs would repeat that mistake. The module takes the calibration as an input, throws when one fitted for a different model is passed, and returns `uncalibrated` rather than a verdict when none is bound. It uses a two-sided band (same above / different below / uncertain in between routed to a human) because a single cutoff is what turned 0.422 into an accept. Real operating points come from `182.C.19.08` on the `182.C.33` corpus. Contradictions need no model and are complete: references naming two different subjects, one image submitted under two roles, a `face-angle` reference that is front-on, a `full-body` reference whose measured frame stops above the ankles. `safeToSpend` is true only when a calibrated measurement says one person AND nothing contradicts. **Not delivered: face swaps and age mismatches.** No manipulation detector and no age estimator on this stack; an uncalibrated deepfake score invites belief it has not earned, and a declared `apparent-age-band` cannot be checked against an age nothing measures. Attribute models are scheduled at `182.C.19.02`, which the `02.03` audit already labelled model-and-corpus work. Both are returned as `not-measured` findings carrying their reason. 30 tests; embeddings are NOT committed as evidence (biometric, `182.C.19.10`) - only cosines. tsc clean on the touched files. | CPU probe on this machine; cosines committed as evidence               | —   | Face-swap and age detections need models this stack does not have (`182.C.19.02`); the identity cutoff needs the `182.C.33` corpus (`182.C.19.08`) |
| 182.C.07.04 | Support immutable identity traits and mutable shot traits so camera, makeup, hairstyle, wardrobe, and scene changes …  | Isis  | 182.C.06   | complete    | `libs/contracts/src/human-video/shot-traits.ts` (new), `shot-traits.spec.ts` (new), `request.ts` (`ShotSchema.traitChanges`), `index.ts`, `schema-registry.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                                                                                                                                      | `shot-traits.ts` in `@oshun/contracts/human-video` makes a change a DECISION rather than drift, and the direction is the design: **a trait not declared for a shot is held**. The other default - the model may vary anything the pack permits - would make an intentional wardrobe change and identity drift indistinguishable after the fact, since both would read as "something the pack allowed". `ShotTraitChange` names a cast member, a stylable trait, and what it changes to. An immutable trait cannot even be named: the element schema is the MUTABLE enum, so `face-structure` fails at parse. A `wardrobe` change must name an approved `wardrobeId` from the pack and no other trait may carry one. `checkShotTraitChanges` applies each person's OWN pack - a production may be free to restyle one character's hair and not another's - and reports four kinds: `transformation-not-allowed` (naming the pack and revision that refused it), `unknown-wardrobe`, `wrong-subject` (no pack covers them), `duplicate-change`. `heldTraitsForShot` derives what the prompt compiler must hold and the evaluator must not flag: every immutable trait plus every stylable one this shot did not declare - including traits the pack PERMITS but this shot left alone. Derived rather than declared, so a trait cannot be held in one place and changed in another. Wired into the request: `ShotSchema` carries `traitChanges`, defaulting to empty, so an existing request holds everything. Published through the `182.C.05.12` registry; the regenerated diff is additive only (245 insertions, 0 deletions; `Shot` and the request gain a field, `ShotTraitChange` is new). 298 contract tests pass (21 new); tsc reports zero errors in the new files.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Contract work — no provider call           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None                                                                   |
| 182.C.07.05 | Bind each person to stable cast tokens and provider-specific reference ordering without leaking provider …             | Isis  | 182.C.06   | complete    | `libs/contracts/src/human-video/cast-tokens.ts` (new), `cast-tokens.spec.ts` (new), `index.ts`, `apps/oshun/bff/src/generation/human-video-reference-ordering.ts` (new), `human-video-reference-ordering.spec.ts` (new)                                                                                                                                                                                                  | Two halves, and the second is the one the item is really about. **Cast tokens** (`cast-tokens.ts`): a person is shown to the creator by display name and referred to in every compiled prompt by a token DERIVED from `castId`, not stored beside it - a stored token is a second identifier that drifts, and drift here is speech attributed to the wrong mouth, which is exactly the second id space `cast.ts` refuses. Deriving also makes it stable across shots, retries and re-renders, which `182.C.07.07` needs. A collision (two ids differing only in punctuation) THROWS rather than being suffixed, because a suffix would silently merge two people in every downstream binding. `findLeakedDisplayNames` detects a real name in provider-bound DIRECTION and deliberately does not rewrite anything: a script legitimately contains spoken names, and rewriting dialogue is the one thing this phase must never do. Tokens are stripped before matching so correct direction (`@cast-ama walks…`) does not report itself - a check that fires on correct input is a check people turn off. **Reference ordering** (`human-video-reference-ordering.ts`): the creator states a role and a priority in our vocabulary; this turns that into the ordered list one lane receives. Role precedence puts identity first (losing it loses the person), then the frame-constraining and angle/body roles, then scene-shaping roles a model can plausibly invent. Priority orders within a role and the reference id is the final tie-break, so a retry sends the same list rather than a differently-ordered request. Two rules keep it honest: a dropped reference is REPORTED with the reason (silent truncation is how a wardrobe reference vanishes and the take returns in the wrong clothes with nothing in the record), and an unprobed `maxImageReferences` (`null`) THROWS rather than guessing a limit. Non-image references are separated rather than counted against the image limit. "Without leaking provider semantics into the UI" is asserted, not just intended: a test checks the creator-facing output contains no vendor name, parameter name, or model slug. 312 contract tests (14 new) and 11 ordering tests pass; tsc clean on both new files.         | Contract + adapter work — no provider call | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None                                                                   |
| 182.C.07.06 | Integrate existing scene, shot, wardrobe, identity-drift, temporal-drift, and golden-output modules behind the …       | Isis  | 182.C.06   | complete    | `apps/oshun/bff/src/generation/human-video-continuity-bridge.ts` (new), `human-video-continuity-bridge.spec.ts` (new)                                                                                                                                                                                                                                                                                                    | All six module families named by the item are now reachable from canonical objects, in one direction: canonical in, canonical out. The `182.C.02.04` audit found them real but UNREACHED - each took its own input shape and returned its own verdict shape, so nothing in the delivery path could ask them a question or record their answer. **identity-drift** and **temporal-drift** map onto canonical `ConditionMeasurement`s (`identity-preservation`, `no-critical-artifacts`), scored against the canonical `MetricThresholds` rather than each module's own severity bands - one number with two threshold vocabularies is how a take passes one gate and fails the other with nobody able to say which is authoritative. **Three fabrication guards**, each because the alternative passes something from nothing: no baseline embedding is `unknown`, not a pass; embeddings from two different models are `unknown`, not a comparison; and a temporal sequence whose frames carry no embedding, histogram or motion scores 0 in the Isis module - which reads as perfectly stable - so coverage is checked first and partial coverage is `unknown` rather than a verdict over the pairs that happened to have data. **scene / shot / wardrobe** profiles are seeded from the canonical request and their projections returned: environment named by its own reference (falling back to the shot id so two shots do not silently share a scene identity they never established), lighting setup, camera angle + shot size, lens, camera move + intensity, first/last-frame anchors, and the ONE wardrobe a shot wears - the declared change (`182.C.07.04`) or the pack's first, never all of them at once. **golden-output** needed a decoder, not a metric: the suite computes its own MAE, PSNR, SSIM and edge drift from pixels. `GoldenFrameDecoder` is an injected port with no default that throws when unbound, `buildGoldenFrameArgs` gives the ffmpeg invocation, and both sides are decoded to the same dimensions so a rendition is comparable with its golden rather than skipped for a size mismatch. Writing the tests caught the real contract: the suite validates RGBA at four bytes per pixel, not grayscale. 27 tests; tsc clean on the touched file. | Adapter work — no provider call            | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None                                                                   |
| 182.C.07.07 | Define cross-shot continuity handoff using reference packs, seeds where meaningful, selected terminal frames …         | Isis  | 182.C.06   | complete    | `libs/contracts/src/human-video/continuity-handoff.ts` (new), `continuity-handoff.spec.ts` (new), `index.ts`, `schema-registry.ts`, `job-state.ts`, `quality-policy.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                                                                                                                             | `continuity-handoff.ts` names five carriers - reference pack (with its REVISION), seed, selected terminal frame, source clip, and the shot state from `182.C.07.06` - and enforces the item's own "not seed reuse alone" rather than only recording it. A seed reproduces a SAMPLING PATH: change the prompt and the same seed gives a different image, change the model and it means nothing at all. So a handoff carrying only a seed is refused whatever it claims to preserve, a seed with no lane fails at parse (a seed is meaningless outside the lane it was drawn in), and a seed carried into a DIFFERENT lane is reported as not portable instead of being passed to a model that reads the number differently. `ASPECT_REQUIREMENTS` maps each of the nine `CONTINUITY_ASPECTS` to the carriers that can actually hold it: identity needs a pack revision - a terminal frame shows a face, it does not authorize one - while set, lighting and props can ride on a terminal frame, a source clip, or shot state. A promise nothing carries is reported, and a test walks every aspect to prove none is silently accepted. The handoff is also checked against the NEXT shot's real reference set: a terminal frame that is not a first-frame reference there, or a source clip that is not a source-video reference, would seed nothing. Two pack revisions for one subject fail at parse. Published through the registry (additive: one new schema, 127 insertions, nothing changed). 331 contract tests pass (19 new). Drive-by, same defect, same phase: `job-state.ts` and `quality-policy.ts` had `Object.freeze([...])` widening literal unions to `string[]`, which was two of the contracts library's pre-existing type errors; `as const` inside the freeze fixes both, and human-video now typechecks clean (library total 17 -> 15, the rest in tts/v10/study).                                                                                                                                                                                                                                                                                                                                                                                                | Contract work — no provider call           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None                                                                   |
| 182.C.07.08 | Support creator-approved reference updates with versioned invalidation, re-consent, benchmark impact, and …            | Isis  | 182.C.06   | complete    | `libs/contracts/src/human-video/pack-revision.ts` (new), `pack-revision.spec.ts` (new), `index.ts`                                                                                                                                                                                                                                                                                                                       | `pack-revision.ts` answers the four questions an approved reference update raises. `assessRevisionImpact` takes two FULL packs rather than a caller-computed diff, because a diff computed upstream can be wrong in exactly the direction that skips a re-consent. **Invalidation**: dropping a reference invalidates the embeddings derived from it - the custody store is keyed by reference (`182.C.19.10`), so this is the exact set, not a search to get wrong. Replacing the CONSENT invalidates every embedding under the pack, not only the dropped ones: the old authorization is what permitted deriving them and the new one has not covered those derivations. **Re-consent** fires on a changed `authorizationId` OR a changed consent `version`. **Benchmark impact**: any change to the reference set, to the traits or transformations the evaluator must hold, or to wardrobe/voice marks the previous pass rate as a statement about a different pack, with the reason recorded. An identical revision reports no change and no staleness. **Reproducible old projects** is a property of the data model - revisions are monotone and never edited (`182.C.07.01`) - so this adds the honest exception: `unreproducible-by-erasure`. A withdrawn consent erases the references and a project pinned to that revision can no longer be rendered, which is the INTENDED outcome; a system that re-rendered it from whatever references remain would be re-depicting somebody who withdrew. Erasure outranks a missing revision because it is the stronger fact. 16 tests; 347 human-video contract tests pass; tsc reports zero errors across human-video.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Contract work — no provider call           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None                                                                   |
| 182.C.07.09 | Define identity-score thresholds by framing and visibility; do not penalize shots where a face is legitimately …       | Isis  | 182.C.06   | complete    | `libs/contracts/src/human-video/identity-thresholds.ts` (new), `identity-thresholds.spec.ts` (new), `index.ts`, `schema-registry.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                                                                                                                                                                | `identity-thresholds.ts` encodes both halves of the item. **By framing:** the cutoff is per shot size, because a cosine measured on a close-up and one measured on a medium-long shot are not the same measurement - the second has a fraction of the face pixels and the same person scores lower through no fault of the render, so one number either fails honest wide shots or passes dishonest close ones. **By visibility:** measurability is settled BEFORE any comparison, so a shot that could never have been scored never produces a `fail`. Five legitimate reasons, each a shot the creator asked for rather than a defect: framing wider than `WIDEST_FACE_EVALUABLE_SHOT_SIZE` (reused from `controls.ts`, not redefined), a covered face, no face detected, a face below the area the model was fitted down to, and a head turned past the fitted yaw. All route to human review; none counts against release. **Per embedding model:** the table names its model, version and corpus; a table fitted for another model THROWS (that carry-across is what produced the 0.422 false accept against the inherited 0.35 gate on 2026-09-10); with no table bound the outcome is `uncalibrated`, and a shot size nobody fitted is `uncalibrated` too rather than borrowing a neighbouring size's number, which would be a different measurement. **No default table ships**, deliberately: a plausible-looking default is how an unfitted number reaches a release gate. `182.C.19.08` fits the real ones on the `182.C.33` corpus. 15 tests; 362 human-video contract tests pass; published additively (one new schema, 71 insertions); tsc clean across human-video.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | Contract work — no provider call           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | None (the numbers themselves are `182.C.19.08`, as the task text says) |

### 182.C.08 Script, Voice, and Exact Speech Preparation

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                         | Owner       | Dependency | Status   | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Cost | Blocker                                                                           |
| ----------- | ---------------------------------------------------------------------------------------------------------------- | ----------- | ---------- | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---- | --------------------------------------------------------------------------------- |
| 182.C.08.01 | Add a script editor contract that preserves exact text, punctuation intent, stage directions, nonverbal sounds … | Isis        | 182.C.07   | complete | `libs/contracts/src/human-video/script-editing.ts` (new), `script-editing.spec.ts` (new), `index.ts`, `schema-registry.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | `script-editing.ts` holds the document a creator works in, beside the canonical script `dialogue.ts` already defines. Its three element kinds are separate VARIANTS of a discriminated union rather than lines with a flag, which is what makes the central rule structural: a stage direction cannot reach `compileSpokenSegments`, because it is not the same type. That matters because a stage direction read aloud by a synthesiser produces a take no evaluator flags - the audio matches the text it was given, exactly. Nonverbal sounds are separated for the mirror reason: a laugh compared against a transcript reports as an insertion. **Punctuation intent**: a marked pause must sit on an actual punctuation character, checked at parse, so intent cannot drift off its mark. **Locked spans**: bounded, non-overlapping (two owners would each believe they hold the same text), and enforced by `applyEdit`, which refuses an edit that changes locked text and names who locked it and why. Locks are re-anchored by TEXT rather than by index - an edit earlier in the line shifts every later index while leaving approved words untouched, so index-based locks would be unusable - and an edit that makes the locked words appear twice is refused rather than re-anchored to a guess. A pause whose punctuation an edit removed is dropped rather than left pointing at whatever character inherited its index. 18 tests, including one that asserts no stage-direction or nonverbal text appears in the compiled spoken output. 380 human-video contract tests pass; published additively (227 insertions, nothing changed); tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Contract work — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | —    | None                                                                              |
| 182.C.08.02 | Add a pronunciation lexicon supporting phonetic spelling, IPA/provider phonemes where available, names …         | Isis        | 182.C.07   | complete | `libs/contracts/src/human-video/pronunciation.ts` (new), `pronunciation.spec.ts` (new), `index.ts`, `schema-registry.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | `pronunciation.ts` extends the per-script lexicon `dialogue.ts` already carries into the vocabulary this item names, because most people who need a name said correctly cannot write IPA. Four representations, resolved in a stated order: per-provider phonemes, portable IPA/ARPABET, a phonetic respelling ("KWAH-me"), and an exemplar recording. Entry kinds cover word, name, place, acronym, number and foreign-word; `sayAs` (letters/word/digits/cardinal/ordinal/year/currency/telephone) is accepted only for acronyms and numbers, where it means something; a foreign word carries its own locale. **Nothing converts between representations.** A provider's phoneme string is never handed to a different provider - phoneme sets disagree about symbols, and a string written for one engine mispronounces on another while looking deliberate. `toCanonicalLexiconEntry` converts only entries that already carry phonemes and REPORTS the rest with the reason: turning a respelling into IPA is a grapheme-to-phoneme model, not a mapping table, and an invented phoneme string reaches a synthesiser looking exactly as authoritative as a real one. An entry carrying no pronunciation at all fails at parse ("says nothing"), phonemes must declare their notation, and two entries for one provider are refused. **Project-level reuse** is a merge in which the script wins - it is the more specific statement - and a token the two disagree about is REPORTED rather than resolved quietly; silently preferring either is how a name gets said two ways across a series with nothing to look at. 21 tests; 401 human-video contract tests pass; published additively (126 insertions); tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Contract work — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | —    | None                                                                              |
| 182.C.08.03 | Support creator-uploaded speech, consented voice synthesis (reusing the narration executor/provider-env TTS …    | Isis        | 182.C.07   | complete | `apps/oshun/bff/src/generation/human-video-speech-source.ts` (new), `human-video-speech-source.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | `human-video-speech-source.ts` covers all three sources and REUSES the existing path rather than rebuilding it: `NarrationProviderGenerate` from `narration-executor.ts` is the synthesis port and `SynthesisDisclosure` from `synthesis-disclosure.ts` is the disclosure port, both injected and both fail-loud when unbound. **The rule that shapes the module:** the approved speech asset is the exact timing source and the DISCLOSED delivery asset is not. `synthesis-disclosure.ts` PREPENDS spoken words, so every word sits later in the disclosed file by exactly the lead-in; alignment or sync scoring computed against it would fail a take by that offset with nothing in the render at fault. Two assets are returned, the lead-in is MEASURED as the difference between two probes rather than assumed, `assertMeasuredAgainstApprovedSpeech` throws when a downstream measurement carries the disclosed digest, and `offsetInDelivery` maps a speech offset onto the delivered timeline for a review UI. Duration on the approved asset comes from the ffprobe path in `audio-probe.ts`, never from the provider's response - a test supplies a provider claiming 99 s and asserts the measured 3.2 s is what lands. **Disclosure follows the voice, not the vendor relationship:** a provider's catalogue voice is still synthesis and goes through the same audible gate; a creator upload does not. Consent is evaluated with the existing `evaluateVoiceConsent` BEFORE anything is synthesised or stored, so a denied use costs nothing and leaves nothing behind - and a creator upload is consented too, because a recording of a real person is that person's voice whether or not a model was involved. 18 tests; tsc clean on the touched files.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Injected ports — no provider call in tests                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | —    | None                                                                              |
| 182.C.08.04 | Integrate the existing voice-consent and persona lifecycle stores; reject expired, revoked, scope-mismatched …   | Isis        | 182.C.07   | complete | `apps/oshun/bff/src/generation/human-video-voice-use-gate.ts` (new), `human-video-voice-use-gate.spec.ts` (new), `human-video-speech-source.ts` (now gated), `human-video-speech-source.spec.ts`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | `human-video-voice-use-gate.ts` asks THREE existing authorities together, because each answers a different question and a use needs all three to say yes: the voice-consent ledger (`studio/voice-consent-store.ts`) for expired, revoked and scope-mismatched; the persona lifecycle (`admin/persona-lifecycle-store.ts`) for whether the persona was ever reviewed; and clone verification for whether the person confirmed this model sounds like them - consenting to be HEARD is not confirming that a particular clone is you, and `isis/voice-cloning-store.ts` is an unbound projection with no verification ledger behind it, so the id is required explicitly rather than assumed. **Every reason is collected, not the first.** A creator whose consent expired AND whose persona is still in review is told both, or they fix one, retry, and are refused again. **`approved-for-test` permits a test render and refuses a release.** The lifecycle vocabulary already draws that line and collapsing it to "approved" is how a rehearsal voice reaches an audience; `deprecated`/`retired` are refused outright as past tense. An unbound lifecycle authority THROWS rather than defaulting to approved. Wired into `182.C.08.03`: `prepareApprovedSpeech` now runs this gate instead of consent alone, before anything is synthesised or stored, so a refused use costs nothing. 16 gate tests plus the 18 speech-source tests still passing; tsc clean on all four files.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Injected authorities — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | —    | None                                                                              |
| 182.C.08.05 | Add voice preview and approval before expensive video generation, including word transcript, duration …          | Isis        | 182.C.07   | complete | `apps/oshun/bff/src/generation/human-video-speech-preview.ts` (new), `human-video-speech-preview.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | `human-video-speech-preview.ts` puts the cheap step before the expensive one: almost every reason a take is thrown away - a name said wrong, a line four seconds over the shot, the wrong read - is audible before a frame is rendered. The preview carries all six things the item names: the script text, the locale, the requested emotion and delivery, the measured duration, a pace in words per minute computed from that duration, and the lexicon entries that applied with what each resolved to. **The approval binds to the DIGEST of the audio that was heard.** Re-synthesise the line, change a lexicon entry, adjust the pace, and the bytes change, the digest changes, and the approval stops applying - `assertSpeechApprovedForVideo` then names precisely which case it is, because "no approval" and "approved for audio that has since changed" mean different things to a creator, and the second is what happens right after somebody fixes a pronunciation. The creator LISTENS to the disclosed delivery (what an audience hears, disclosure included) while the approval binds to the SPEECH digest, which is what the render is synchronised against (`182.C.08.03`). The transcript shows what a recognizer actually heard, since script and audio differing is the entire point of listening; with no recognizer bound it reports `unavailable` with a reason and is NEVER filled in from the script, which would make the check a tautology that passes on every take. 15 tests; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Injected recognizer — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | —    | None                                                                              |
| 182.C.08.06 | Normalize loudness, peak, silence, sample rate, channels, and leading/trailing padding without shifting word …   | Media Infra | 182.C.07   | complete | `apps/oshun/bff/src/generation/human-video-speech-normalization.ts` (new), `human-video-speech-normalization.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-10-speech-normalization/` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | `human-video-speech-normalization.ts` handles all six: loudness, true peak, sample rate, channels, and leading/trailing padding. It is deliberately NOT `human-video-normalization.ts` (`182.C.06.07`), which forbids every filter used here - that one runs on provider-bound assets where nothing may change, this one runs EARLIER, on speech that has not yet become the approved timing source. The ordering rule is explicit: normalize, then approve; normalizing an approved asset changes its digest and `182.C.08.05`'s digest-bound approval is what stops that shipping silently. **"Without shifting word timing" is measured, not asserted.** Gain, resample and channel steps must leave the length alone; trims and pads change it by exactly their own amount; `verifySpeechNormalization` compares the measured duration against that arithmetic and THROWS beyond codec frame rounding (0.05 s, two MP3 frames). Head and tail are tracked separately because only the head moves a word. **Two bugs the unit tests could not catch, both found by running the emitted chain through real ffmpeg** (`evidence/2026-09-10-speech-normalization/`): the planned TAIL trim was never emitted (`atrim=start=` only), so a correct clip would have thrown the timing error the module exists to raise; and `loudnorm` ran BEFORE the stereo-to-mono downmix, landing -19.0 LUFS against a -16 target, because summing correlated channels changes the level the filter just set. With the order corrected the same clip lands -15.9 LUFS and 3.800 s, exactly the planned 5.0 - 0.8 - 0.4. Both now have tests, one asserting filter ORDER rather than presence. `LOUDNESS_TOLERANCE_LU` is 1 on measured grounds: single-pass `loudnorm` landed within 0.1 LU on 3.5 s, 4 s and 20 s clips, so 1 sits outside the observed error and inside the 3 LU a wrong order produces - a tolerance wide enough to hide that bug would not be worth having. 24 tests; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Real ffmpeg runs on this machine; measurements committed                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | —    | None                                                                              |
| 182.C.08.07 | Produce word- and phoneme-level alignment with confidence, uncertainty, pause boundaries, viseme hints …         | Isis        | 182.C.07   | complete | `libs/contracts/src/human-video/alignment.ts` (new), `alignment.spec.ts` (new), `index.ts`, `schema-registry.ts`, `tools/human-video/forced-align.py` (new), `apps/oshun/bff/src/generation/human-video-alignment-adapter.ts` (new) + spec + fixtures, `docs/domains/isis/human-video/evidence/2026-09-10-forced-alignment/` (new), plus the adapters 2026-09-12: `apps/oshun/bff/src/generation/human-video-audio-adapters.ts` (new), `human-video-audio-adapters.spec.ts` (new), `human-video-audio-store.ts` (new), `human-video-audio-store.spec.ts` (new), `human-video-runtime.ts` (the binding), `libs/oshun/persistence/prisma/migrations/20260912030000_human_video_audio_take/migration.sql` (new), `schema.prisma`, `prisma-renderer.ts` | All six things the item lists, and the aligner is committed rather than described. `tools/human-video/forced-align.py` does FORCED alignment (the words are given; the question is where they sit) with a CTC trellis over the transcript, not recognition - on the committed fixture faster-whisper's word timings were off by up to 600 ms, snapped the first word to 0.0 and merged two words into one token. **It reproduces the earlier session's measurement to the tenth of a millisecond**: onset errors +81.2, +65.2, +16.6, +17.5, +61.9, +34.6 ms against `align-accuracy.json`'s [81.15, 65.15, 16.61, 17.46, 61.89, 34.68]. That match matters - the tool committed here IS the measurement, not a description of one. Phonemes come from a second CTC pass (`wav2vec2-lv-60-espeak-cv-ft`, Apache-2.0) and are real IPA: `forty` aligned as f/oːɹ/ɾ/i including the American flap. **A bug only the ground truth could catch:** the first version reported every word after the first ~700 ms early, because the word SEPARATOR token is emitted across the whole silent gap and was being counted as part of the following word. The output looked entirely plausible. The separator now belongs to no word, and the evidence records both tables. Contract: `alignment.ts` requires every timing to say whether it was `measured` or `estimated` (an approximate phoneme span and a measured one are both numbers in seconds, and conflating them is how an approximation gates a release), carries `uncertaintySeconds` as a NUMBER - 0.081 for onsets, the largest measured error - and makes pauses first-class rather than "the gap between word n and n+1", since a breath is where a lip-sync scorer must not expect mouth movement. Words may not overlap, phonemes may not overlap or escape their word, and word indices must be consecutive or the join back to the script breaks. Adapter: assigns a phone to a word by MIDPOINT and clamps it, because the two models are independent and disagree at edges; drops phones matching no word with a count; gives a viseme hint only where English IPA maps through to the renderer's existing 39-ARPABET table, and `undefined` (never `rest`, a closed mouth) elsewhere. Its fixtures are the real aligner output, not hand-written. 17 contract tests + 14 adapter tests; 418 human-video contract tests pass; tsc clean. BOUND 2026-09-12. `forced-align.py`'s own docstring said it was "used by the BFF through an injected aligner port" and NOTHING WAS INJECTED; `human-video-audio-adapters.ts` is the child process. THE ALIGNER MEASURES A DERIVED COPY AND SAYS SO: the tool takes a mono 16 kHz WAV and a take is whatever the creator recorded, so ffmpeg makes a copy and the aligner reads THAT. Resampling does not move anything in time, so the timings are still about the take — the copy is not the take, nothing binds to it, and it is deleted. THE DURATION IS PROBED FROM THE ORIGINAL for the opposite reason: a transcode can move it by a frame at the boundaries, and 182.C.35.09 measured a +20 ms AAC padding that a real provider render then confirmed. AND `format.duration` IS NOT THE AUDIO STREAM'S: ffprobe reports the LONGEST stream, so a take that arrived inside an mp4 with a video track would come back longer than the words it contains — the probe asks for streams and takes the audio one. NEITHER TOOL MAY ANSWER "NOTHING": a probe returning 0 s reads as `silent_audio`, a finding about the RECORDING, when the cause was a missing interpreter; an aligner returning no words makes every word unaligned, which blocks approval with a reason about the creator's SCRIPT. Every failure throws. | CPU aligner run on this machine; output committed as evidence 16 adapter tests, 14 against an injected runner and 2 RUNNING BOTH TOOLS FOR REAL against the committed clip in `evidence/2026-09-10-cpu-sync-and-alignment/` — every onset within 150 ms of the ground truth committed beside it. ONSETS ONLY: the tool's docstring reports word ENDS running 96-141 ms early because CTC emits where the evidence is decisive rather than where the sound stops, and it reports the bias rather than correcting it, so asserting on ends would be asserting on a known uncorrected error. 11 store tests. 5 runtime tests for the three-part binding. | —    | None                                                                              |
| 182.C.08.08 | Split long speech only at safe phrase/breath boundaries and preserve sample-accurate offsets for later …         | Media Infra | 182.C.07   | complete | `apps/oshun/bff/src/generation/human-video-speech-splitting.ts` (new), `human-video-speech-splitting.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | `human-video-speech-splitting.ts` cuts long speech only where the aligner (`182.C.08.07`) found a pause, and REFUSES when the window a lane offers contains none - a cut inside a word is audible however well the pieces are joined, and it destroys what the sync evaluator needs, since a word starting in one chunk and ending in another has an onset in neither. The refusal distinguishes "no pause in this window" from "one word is longer than the whole window", which are different problems for a creator. **Offsets are SAMPLES, not seconds.** Seconds are floating point and a few additions later the pieces no longer meet; a sample index at a known rate is an integer and stitching integers is exact. `buildChunkCutArgs` uses `atrim=start_sample=…:end_sample=…` rather than `-ss`. `assertChunksReassemble` proves the pieces cover the original exactly once - tests catch a one-sample gap and a one-sample overlap by name - and `assertEveryWordDelivered` proves each word appears in exactly one chunk, in order. The cut lands as LATE inside a pause as the window allows rather than at the pause's middle: late cutting fills each chunk and makes the fewest joins. Testing (not reading) turned up two cases a middle-of-pause rule refuses and this one handles: a pause that STRADDLES the window limit, where the limit itself falls in silence, and a pause that BEGAN before the current chunk but still has silence left. 16 tests; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Pure computation over alignment — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | —    | None                                                                              |
| 182.C.08.09 | Validate script-to-audio exactness before rendering and require explicit creator approval for any …              | Isis        | 182.C.07   | complete | `apps/oshun/bff/src/generation/human-video-script-exactness.ts` (new), `human-video-script-exactness.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | `human-video-script-exactness.ts` compares what the speech SAYS against what the script APPROVED, before a frame exists, using the shared normalizer from `dialogue.ts` - so "the same words" means one thing across this check, the release gate and the dub-faithfulness seam. It is a Levenshtein diff, not a set comparison: "we open eight at" is a reordering, and a set comparison would call it identical. A spoken "forty two dollars" matches a written "42 dollars" through the shared numeral expansion. Three kinds of difference, treated differently because they mean different things: a MISSING word is a broken promise; a SUBSTITUTION is usually a mispronunciation or a synthesiser's own rewrite and needs a person; a DISFLUENCY ("um", "uh") is a delivery choice, and `CanonicalScript.disfluenciesPermitted` already existed for exactly that - when set they are recorded but not blocking, because permitted is not the same as invisible. Content words are marked as such, matching `182.C.03.05` C2's stricter treatment of them. **Every deviation needs an explicit approval, keyed to the audio digest.** Approving a substitution for one take does not carry to a re-synthesis, for the same reason the voice approval in `182.C.08.05` binds to a digest, and a test asserts that. **Unmeasured is not exact.** With no recognizer bound, or a locale the shared normalizer has no rules for (it refuses to apply English contraction rules to Twi), the verdict is `unknown` and `assertScriptExactBeforeRender` refuses to render as firmly as it does for a real deviation. Treating "we could not check" as "it is fine" is the failure this phase is written against. 19 tests; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Injected recognizer — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | —    | None                                                                              |
| 182.C.08.10 | Support Ghanaian English, Akan/Twi, Ga, Ewe, and code-switching fixtures with names and local …                  | Isis        | 182.C.07   | complete | `libs/contracts/src/human-video/locales.ts` (new), `locale-fixtures.ts` (new), `locales.spec.ts` (new), `index.ts`, `schema-registry.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | `locales.ts` + `locale-fixtures.ts` cover Ghanaian English, Akan/Twi, Ga and Ewe, with a code-switched line whose span is asserted to land on the right characters. **The rule the file exists for: a locale nobody probed is not a locale a provider supports.** Every TTS vendor lists dozens of languages and that list is a marketing surface, not a measurement, so lane support is one of FOUR states with `unprobed` as the DEFAULT - and `unprobed` BLOCKS an exact-dialogue request while only warning a prompt-native one, because words rendered with the wrong phonology are not those words. Any state other than `unprobed` needs the date somebody listened, and `approximated` must say what is wrong with it or a creator cannot judge whether it is good enough. `extendedLatinCharacters` is the pre-flight for the failure mode that matters here: a lane that strips unknown characters turns `Ɛte sɛn?` into `te sn?` - it changes the word and reports success. Ga and Ewe add ŋ, ɖ, ƒ, ʋ. **The fixtures refuse to be a benchmark they have not earned.** Each carries `reviewedBySpeaker`, which is FALSE for every non-English line, and `scorableFixtures()` returns only the reviewed ones. Scoring a lane's Twi on text nobody who speaks Twi has checked measures our spelling, not the lane - the fixtures exercise the pipeline (orthography survival, code-switch spans, names reaching the lexicon) and say so. 22 tests; 440 human-video contract tests pass; published additively; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Contract + fixtures — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | —    | Native-speaker review of the Twi/Ga/Ewe fixtures before any of them scores a lane |
| 182.C.08.11 | Keep prompt-native dialogue as a separately labeled mode with mandatory post-render ASR and no exact-word …      | Isis        | 182.C.07   | complete | `apps/oshun/bff/src/generation/human-video-dialogue-mode.ts` (new), `human-video-dialogue-mode.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | `human-video-dialogue-mode.ts` keeps prompt-native and exact-dialogue from blurring, and they blur in ONE direction: a prompt-native take that happens to say roughly the right words looks exactly like an exact-dialogue take that worked. Three rules. **Labelled**: `buildDialogueModeLabel` produces the mode, whether exact words were promised, and creator-facing wording ("the wording in this video was generated by the model and was not scripted"), so nothing downstream infers it from the absence of supplied audio. **No word promise**: the `exact-speech` condition is reported explicitly as `unknown` WITH A REASON rather than omitted - an omitted condition reads as satisfied to anything that iterates what it was given, so "we never checked" would become "it passed" by silence. **Post-render ASR is mandatory anyway**: `assertPostRenderTranscript` refuses a take without one in EVERY mode. The tempting shortcut is to skip ASR where no particular words were promised, and that reasoning is backwards - it is exactly the mode where nobody knows what was said, and captions, the safety scan, and a creator entitled to know whether the model put a product claim in their character's mouth all need it. The label carries the transcript reference, because a label saying the words were not guaranteed is only useful beside the words that were actually said. 11 tests; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Policy + gate — no provider call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | —    | None                                                                              |

### 182.C.09 Shot Planning and Performance Direction

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                     | Owner | Dependency | Status   | Changed Files                                                                                                                                                                                                                         | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Provider/Hardware Evidence               | Cost | Blocker |
| ----------- | ------------------------------------------------------------------------------------------------------------ | ----- | ---------- | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------- | ---- | ------- |
| 182.C.09.01 | Build a shot-plan schema containing cast, speaker, action, blocking, beat timing, camera, lens, framing …    | Isis  | 182.C.08   | complete | `libs/contracts/src/human-video/shot-plan.ts` (new), `shot-plan.spec.ts` (new), `index.ts`, `schema-registry.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                 | `shot-plan.ts` carries every field the item lists - cast, speakers, dialogue segments, blocking, beats, camera/lens/framing/lighting/performance (through `CreativeControls`), props, environment, continuity handoff (`182.C.07.07`) and trait changes (`182.C.07.04`). It COMPOSES rather than redefines: a second vocabulary for camera or wardrobe would be a second answer to the same question. Everything a lane is eventually told comes from here, so this is where internal contradictions are caught, and the checks are relationships rather than shapes: a speaker who is not in the shot, blocking for somebody who is not in it, a beat delivering a segment the shot does not carry, one segment delivered by two beats, a beat that runs past the end of its shot, blocking that acts on a prop the shot has not got, one line delivered in two shots ("a line is said once"), non-consecutive ordering, a handoff that claims to come from another shot or points at one the plan does not contain. **Two deliberate absences.** No coordinate system: blocking uses theatre marks (upstage-left, frame-right) and prose, because `controls.ts` already argues that a number nobody can verify against the rendered frame is a number that looks measured and is not. No absolute beat clock: a beat is timed within its own shot, since plan-wide times would need recomputing whenever a duration changed and the copy that did not get recomputed is the one that ships. 23 tests; 463 human-video contract tests pass; published additively; tsc clean.                                                                                                                                                                                                                                                                                                                                                          | Contract work — no provider call         | —    | None    |
| 182.C.09.02 | Translate a creator brief into editable shots without changing locked dialogue or consented identities.      | Isis  | 182.C.08   | complete | `apps/oshun/bff/src/generation/human-video-shot-planner.ts` (new), `human-video-shot-planner.spec.ts` (new)                                                                                                                           | `human-video-shot-planner.ts` turns a brief into an editable plan, and is built so it CANNOT damage the two things that must survive translation. **Locked dialogue**: the proposer port is given segment IDS and speakers and nothing else - a test asserts the payload does not contain the line text - so it cannot paraphrase words it never holds. **Consented identities**: speakers are DERIVED from the script on this side of the boundary, never taken from the proposal, so a proposer cannot move a line into a different mouth because it is never asked whose line it is; a cast id the cast does not contain is refused rather than filtered. A proposal is validated before anything is built: a line that would go unsaid, a line delivered twice, a segment the script does not contain, an empty proposal. A proposer that quietly drops a line produces a video missing a sentence and one that repeats a line produces a person saying the same thing twice - both look like plans until somebody watches. `assertPlanPreservesScriptAndCast` runs after a creator EDITS too, which is where a line most easily lands in the wrong shot: an editor moving a segment between shots has no reason to know whose line it was. With no proposer bound there is still a real plan - one shot per speaker turn, which is the shape a two-hander takes anyway and needs no model. 13 tests. **Correction 2026-09-10:** the original "tsc clean" here was vacuous — the scratch project extended `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so it type-checked nothing. Re-run against `tsconfig.typecheck.json` it had two real errors: a `mark is string` type predicate over a `PositionMark` literal union, and the `mark.replace` that followed it. Fixed by dropping the predicate for an explicit loop; the 13 tests were green before and after, which is the point — vitest does not typecheck. | Injected proposer — no provider call     | —    | None    |
| 182.C.09.03 | Estimate speech duration before planning and prevent shot durations that cannot contain the approved audio … | Isis  | 182.C.08   | complete | `apps/oshun/bff/src/generation/human-video-speech-fit.ts` (new), `human-video-speech-fit.spec.ts` (new)                                                                                                                               | `human-video-speech-fit.ts` keeps two facts apart on purpose. The ESTIMATE reuses `estimateNarrationDurationSeconds` from `audio-probe.ts`, inheriting its own "a wide sanity band, not an exact prediction" caveat rather than quietly dropping it, and is what plans a shot before anybody has heard the line. The MEASUREMENT replaces it the moment approved audio exists, because a shot that fits the estimate and not the audio fits nothing. The verdict says WHICH it used, and one estimated line makes the whole verdict an estimate - a total is only as sound as its weakest part. **Transition handles are the part people forget.** An editor cutting or dissolving needs frames on both sides that are not carrying a word, and a shot cut exactly to the length of its dialogue has none, so the first frame of a dissolve lands on a syllable. Handles are added to the REQUIREMENT, not to the shot, so a plan that does not fit is reported rather than silently lengthened; consecutive lines also get air between them. `assertSpeechFits({requireMeasured:true})` is what the submission path sets: planning may proceed on an estimate, but spending a generation on a fit nobody measured against the actual audio is spending it on a guess. That refusal is its own error naming the unmeasured lines, not a shortfall error reporting 0 s short. 13 tests; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Pure computation — no provider call      | —    | None    |
| 182.C.09.04 | Mark visible-speech windows, off-screen speech, reaction shots, cutaways, overlaps, and deliberate …         | Isis  | 182.C.08   | complete | `libs/contracts/src/human-video/speech-windows.ts` (new), `speech-windows.spec.ts` (new), `shot-plan.ts` (role + windows), `shot-plan.spec.ts`, `index.ts`, `schema-registry.ts`, `libs/contracts/schemas/human-video/` (regenerated) | `speech-windows.ts` tells the evaluator what to expect and when, and `PlannedShot` now carries `role` and `speechWindows`. The reason it exists: a sync scorer that does not know a line is off-screen scores the mouth it can find and FAILS the take; one that does not know a character deliberately turns away reads the turn as a hidden mouth; one that does not know two people speak at once attributes both voices to whichever face it locked onto first. In each case the take is fine and the measurement is wrong - and a wrong measurement that fails good work is worse than none, because somebody re-renders until the scorer happens to agree. Five visibilities (on-camera, off-screen, deliberately-occluded, back-to-camera, extreme-profile) and six shot roles. **A deliberate occlusion needs a written REASON**, because "the mouth was hidden" is what a bad take looks like too, so the difference between direction and defect is recorded before the render by the person who wanted it - and `unmeasurableWindows` hands that reason to the reviewer `182.C.07.09` routes it to. Overlaps BETWEEN speakers are allowed and expected - people talk over each other, which is why the speaker-attribution condition exists - while an overlap within ONE speaker is refused, since one mouth cannot deliver two lines at once. `shotRequiresSpeakerAttribution` is deliberately named apart from `cast.ts`'s `requiresSpeakerAttribution`: that one says the condition applies to the REQUEST, this one says which shots actually put two visible mouths on screen at once. The collision was caught by tsc, not by review. `syncCoverageRatio` reports how much of a shot's speech a measurement can cover at all. 17 window tests + 4 new shot-plan tests; 484 human-video contract tests pass; published additively; tsc clean.                                                                         | Contract work — no provider call         | —    | None    |
| 182.C.09.05 | Add blocking primitives for walking, sitting, standing, turning, object interaction, position marks, path …  | Isis  | 182.C.08   | complete | `libs/contracts/src/human-video/shot-plan.ts` (viaMarks, facing, validation), `shot-plan.spec.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                | All nine primitives the item names are in `BlockingInstruction`, and a test walks the list: walking, sitting, standing and turning are actions; object interaction is the `reach`/`hand-over`/`pick-up`/`put-down`/`gesture-to` set with a required `targetId`; position marks are the theatre vocabulary; speed is a five-step scale; the endpoint is `toMark`. **Path was the one genuinely missing** (`182.C.09.01` shipped the rest) and is now `viaMarks`: endpoints alone under-describe a move, since "upstage-left to frame-right" is satisfied by a straight line behind the counter and by a walk around the front of it, and those are two different shots. Marks rather than a curve, for the same reason the module refuses coordinates - a number nobody can verify against the rendered frame looks measured and is not. Four contradictions are refused: a turn that does not say where it leaves somebody facing (otherwise it is only a note that they moved), a route on an action that does not travel, a route repeating a mark without going anywhere in between, and a walk at a standstill. 33 shot-plan tests (6 new); 490 human-video contract tests pass; published additively; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Contract work — no provider call         | —    | None    |
| 182.C.09.06 | Add performance primitives for emotion, intensity, posture, gesture frequency, eye line, attention target …  | Isis  | 182.C.08   | complete | `libs/contracts/src/human-video/controls.ts` (posture, emotionIntensity, conversationalTurn, breath, pauseStyle), `controls.spec.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                             | `PerformanceDirection` in `controls.ts` already carried emotion, gesture frequency, eye line and attention target; this adds the four the item names that were missing - posture, emotion intensity, conversational turn and breath - plus a pause STYLE. A test walks the item's nine so they cannot quietly become five. Three separations are the point. **Posture is not motion**: how the body is held and how much it moves are different directions, and a still, leaning-in performance is a real one. **Intensity is not emotion**: "concerned, restrained" and "concerned, strong" are the same expression played two ways, and collapsing them loses the note a director actually gives. **Pause STYLE is not pause PLACEMENT**: where a pause falls is authored in the script (`182.C.08.01` marks it on punctuation) and measured in the alignment (`182.C.08.07`); this says only how the performer holds one, and deliberately has no field that could relocate an authored beat - a performance setting that moved one would change the delivery the creator approved. `conversationalTurn` is optional rather than defaulted, because a single-speaker shot has no exchange to take a turn in. 25 controls tests (7 new); 497 human-video contract tests pass; published additively; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Contract work — no provider call         | —    | None    |
| 182.C.09.07 | Add camera primitives for static, pan, tilt, dolly, truck, crane, orbit, handheld, tracking, rack focus …    | Isis  | 182.C.08   | complete | `libs/contracts/src/human-video/controls.ts` (tracking, RackFocus), `controls.spec.ts`, `libs/contracts/schemas/human-video/` (regenerated)                                                                                           | `controls.ts` already had static, pan, tilt, dolly, truck, pedestal, crane, orbit, zoom, handheld and steadicam-follow, plus focal length, depth of field and move intensity. This adds the two the item names that were missing, and a test walks its list of thirteen so none can quietly go. **`tracking`** is named apart from `steadicam-follow` on purpose: that one names the rig and its look, this names the camera's RELATIONSHIP to a moving subject, and a lane can support one without the other. **Rack focus is a LENS action, not a camera move** - the camera does not go anywhere during a focus pull - so it sits on `LensControl` with from/to focus targets, ids where the target is a person or a prop, and an offset within the shot rather than an absolute time, for the reason `shot-plan.ts` gives about clocks. Two contradictions are refused: a pull that starts and ends on the same thing (not a pull), and focusing on "cast" or "prop" without saying which one. A pull between two different people is allowed, which is the common case in a two-hander. 32 controls tests (7 new); 504 human-video contract tests pass; published additively; tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Contract work — no provider call         | —    | None    |
| 182.C.09.08 | Add feasibility validation for contradictory actions, impossible timing, too-small speaking faces, unsafe …  | Isis  | 182.C.08   | complete | `apps/oshun/bff/src/generation/human-video-shot-feasibility.ts` (new), `human-video-shot-feasibility.spec.ts` (new)                                                                                                                   | `human-video-shot-feasibility.ts` checks five kinds of impossible before a generation is paid for. Four are decidable from the plan; the fifth says it is not. **Contradictory actions** are actually contradictory rather than unlikely: a turn towards another person in a shot with nobody else in it, a focus pull racking onto somebody who is not in the shot, a line delivered back-to-camera by somebody the performance direction has looking down the lens. A cast member who exits AND has an on-camera line is a WARNING, not a block - the order might be fine and a human should look. **Impossible timing**: the `182.C.09.03` speech-fit shortfall, beats adding up to more than the shot, a focus pull starting at or after the shot ends. **Too-small speaking faces**: an on-camera line in a framing wider than `WIDEST_FACE_EVALUABLE_SHOT_SIZE` is blocked with the reason that the take could never pass the sync condition - better said before the render than after. **Provider limits**: duration, aspect and resolution against the lane. **Unsafe activity is NOT decided here.** A keyword list pretending to be a classifier is worse than nothing - it passes anything phrased carefully and blocks "shoot the scene" - so the classifier is injected and, with none bound, the finding is `unsafe-activity-not-screened` at severity `unscreened`: an honest gap that is reported and does not block, rather than a green light. What gets screened is props and environment as well as blocking verbs, because "a kitchen knife on the counter" is where a scene's problem often is. 17 tests; tsc clean.                                                                                                                                                                                                                                                                                            | Injected classifier — no provider call   | —    | None    |
| 182.C.09.09 | Compile provider-neutral direction into model-specific prompts/options and store the compilation for …       | Isis  | 182.C.08   | complete | `apps/oshun/bff/src/generation/human-video-prompt-compiler.ts` (new), `human-video-prompt-compiler.spec.ts` (new)                                                                                                                     | `human-video-prompt-compiler.ts` turns the neutral vocabulary into the string and options one model takes, and three properties matter more than the wording. **No real name leaves**: the prompt refers to people by the cast tokens from `182.C.07.05`, and the result is CHECKED with `findLeakedDisplayNames` before it is returned - a test puts a display name in an environment description and asserts the compiler throws. A creator's colleague's name in a provider payload is a disclosure nothing can recall. **Deterministic**: the same shot, lane and compiler version give byte-identical output, and the compilation id is derived from the direction digest rather than a clock, so compiling the same shot in June and February gives the SAME id. A retry that changes the prompt is not a retry of the same shot, and nobody could tell afterwards which version failed. **Stored**: the record carries the prompt, both option sets, the direction digest, the prompt digest, the model slug and the compiler version, so "why did this shot look like that" is answerable six months later - which is what `182.C.01.05` is for. `sameCompilation` compares digests rather than text, so a stored record can be checked without holding the prompt. Vendor field names live in an injected `LaneAdapter`, per ADR-0006: a passthrough parameter list is a property of one model on one platform on one date, and freezing it into every consumer's build is how a contract stops being canonical. The digest deliberately covers only what changes the render. 13 tests; tsc clean.                                                                                                                                                                                                                                                                                                                            | Injected lane adapter — no provider call | —    | None    |
| 182.C.09.10 | Provide manual shot locking so retries can change only the failed dimension—sync, identity, motion …         | Isis  | 182.C.08   | complete | `apps/oshun/bff/src/generation/human-video-shot-locks.ts` (new), `human-video-shot-locks.spec.ts` (new)                                                                                                                               | `human-video-shot-locks.ts` makes a retry vary ONE dimension and carry the rest verbatim. The problem it solves: a take fails on sync, the framing and wardrobe and performance were all right, and the obvious retry re-rolls everything - so the second take fails on something new, and nobody can tell which change caused which outcome. Nine lockable dimensions, each read from where it actually lives. The comparison is dimension-by-dimension rather than a deep equality over the whole shot, because a whole-shot diff cannot say WHICH approved choice a retry disturbed, and that sentence is the value of the module. Two distinct refusals: a LOCKED dimension changed (a creator's decision overridden) and an UNEXPECTED dimension changed (a retry that will teach nobody anything). Neither is corrected - a correction hides that something tried to change it. `DEFAULT_VARIABLE_DIMENSIONS` maps each release condition to what a retry should vary, and is explicitly a starting point rather than a finding: which dimension actually fixes a sync failure is empirical and `182.C.34` answers it with a corpus. It is overridable, and one entry is worth naming - `exact-speech` varies NOTHING about the picture, because re-rendering cannot fix wrong words and that retry belongs to the speech path. `describeRetry` tells a creator plainly when everything useful is locked or when a retry could not help at all. 16 tests; tsc clean. This completes 182.C.09.                                                                                                                                                                                                                                                                                                                                                                                                                                    | Pure computation — no provider call      | —    | None    |

### 182.C.10 Live Provider Capability Registry

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner     | Dependency | Status   | Changed Files                                                                                                                                                                                           | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                | Cost                            | Blocker |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------- | ------- |
| 182.C.10.01 | Define a capability vocabulary covering text/image/video/ audio inputs, first/last frames, reference counts/types …    | Oshun BFF | 182.C.05   | complete | `libs/contracts/src/human-video/capability-vocabulary.ts` (new), `capability-vocabulary.spec.ts` (new), `index.ts`, `docs/domains/isis/human-video/evidence/2026-09-10-openrouter-video-catalog/` (new) | `capability-vocabulary.ts` answers two questions and refuses to blur them: what a lane can be asked for, and where the answer legitimately comes from. 31 entries, each with a `source` — 15 `catalog`, 3 `transport`, 2 `passthrough`, 7 `probe`, 4 `manual-record`. The provenance column is the whole point. Read against the live catalog on 2026-09-10 (28 models, 17 keys per record), the catalog is generous about format and **silent about every entry a creator actually asks for**: no record in the response carries a reference-count limit, an edit or extend flag, a retention window, or any statement that a model accepts supplied audio. The only trace of supplied audio is the parameter NAME `audio` in `allowed_passthrough_parameters` on `alibaba/wan-2.7` and `-2.6` — a knob exists, and that is all a name can say — so `accepts-supplied-audio` is `passthrough`-sourced and probe-gated, and `exact-script-lane` is `derived` rather than stored, because a stored flag is exactly how a prompt-native lane ends up labelled exact. Two measured details drive the design. First, a present field is often a null field: `supported_sizes` null for 9 of 28, `generate_audio` and `supported_frame_images` for 4, `seed` and `supported_durations` for 3 — so catalog entries carry `fallbackWhenAbsent: 'probe'`, since null is the catalog declining to state a bound, not the model refusing the feature. `upscale_factor` and `creativity` are the two exceptions where null IS the answer (27 of 28; only `flux-video-upscale` has them). Second, `requiresProbe` is deliberately wider than the probe-sourced set: an advertised `seed` is a claim about a parameter and the creator's promise is about a result, so 21 of 31 entries are probe-gated and the 10 exempt ones are envelope facts (`model-version`, `pricing-basis`, `provider-options`, `callback-delivery`) and terms no generation could settle. `unmappedCatalogFields()` is the anti-drift device: the catalog client will call it on every fetch and a test calls it on the 2026-09-10 key list, so a `supported_reference_count` field OpenRouter adds later becomes a failing test rather than a silently ignored capability. Names reuse `VIDEO_CAPABILITIES` verbatim — one name per concept — and a test fails if the two lists drift. `npx vitest run src/human-video` in `libs/contracts`: 24 files, 523 tests pass (19 new). Typecheck via a scratch project over the two new files (`libs/contracts/tsconfig.json` is a solution-style config with `files: []` and checks nothing, so a `-p tsconfig.json` run is vacuously clean): 0 errors. The compile-time name proof was verified by mutation — renaming `multi-subject` to `multi-subjects` produces `Type 'boolean' is not assignable to type 'never'` before any test runs.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Live `GET /api/v1/videos/models` 2026-09-10: 200, 26578 bytes, 28 models, raw response and full field/null analysis committed under `evidence/2026-09-10-openrouter-video-catalog/`. Six lane rows in the phase snapshot re-verified field by field: all unchanged.                                                       | $0.00 (catalog GET is unbilled) | None    |
| 182.C.10.02 | Implement an OpenRouter model-catalog client with strict schema validation, ETag/cache handling, timeout, retry …      | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-catalog-client.ts` (new), `human-video-catalog-client.spec.ts` (new)                                                                                         | `human-video-catalog-client.ts` reads `GET /api/v1/videos/models`, and four measurements against the live endpoint on 2026-09-10 contradict what the task text assumes. **There is no ETag.** No `Last-Modified`, no `Cache-Control` either, and a bogus `If-None-Match` is ignored - the server answers 200 with the whole 26578-byte body. So conditional revalidation cannot be the change detector and the SHA-256 of the exact response bytes is; two consecutive reads hashed identically, so the hash is a sound validator rather than a source of false churn. The `If-None-Match`/304 path is implemented anyway because honouring a validator the server later adds costs nothing, and the header comment says plainly that it is dead code today rather than implying it saves bandwidth. **The endpoint is public**: with no `Authorization` header, and again with a deliberately invalid key, it returned 200 and the full catalog - so the key is sent when configured and its absence is not a failure, because refusing to refresh the registry over a missing secret would be a self-inflicted outage. Fail-closed lands in two places. A record missing any of `id`, `canonical_slug`, `pricing_skus` or `allowed_passthrough_parameters` is QUARANTINED with its issues rather than repaired with defaults, so its routes drop out while the other 27 models keep working; absent and null capability bounds parse to the same `null`, because null is the catalog declining to state a bound and the vocabulary already sends both to a probe. And a body that is not `{ data: [...] }`, or one that yields zero usable records, throws - an empty catalog served as success reads downstream as "every lane is gone" while reporting health. Unknown keys are KEPT (`z.looseObject`) and run through `unmappedCatalogFields`, so a `supported_reference_count` field OpenRouter adds later is reported instead of stripped. Retry is bounded and status-aware (408/429/5xx and network errors only, never other 4xx), full-jitter backoff from 500 ms capped at 8 s, and `Retry-After` is honoured in both documented forms - but a request to wait longer than 30 s ends the loop with a message saying so rather than holding a caller for ten minutes. Timeout is an `AbortController` and an aborted request reports `timeout`, not a generic network failure. A cache entry inside the 24 h TTL short-circuits the network entirely; past it, a failed refresh serves the cached copy LABELLED `stale-cache` with `stale: true` up to 7 days, and beyond that refuses rather than routing on it - whether to route on stale capability is `182.C.10.04`'s decision, not this module's. `fetch`, `sleep`, `random` and `now` are all injected, so the backoff assertions are exact rather than approximate. 28 tests, one of which parses the real committed 28-model response from `evidence/2026-09-10-openrouter-video-catalog/` and asserts the nulls it actually contains. Full BFF human-video suite 661 green; targeted tsc against `tsconfig.typecheck.json` clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Live probes 2026-09-10: no `ETag`/`Cache-Control`/`Last-Modified` on the 200; `If-None-Match` with a bogus value returned 200 and the full body; no-auth and invalid-key both returned 200 and the full catalog; two reads byte-identical (sha256 c8ffc953...). Test doubles at the fetch boundary; no billed generation. | $0.00 (catalog GET is unbilled) | None    |
| 182.C.10.03 | Persist dated catalog snapshots and diffs so removed or changed capabilities automatically quarantine affected routes. | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-catalog-snapshots.ts` (new), `human-video-catalog-snapshots.spec.ts` (new)                                                                                   | One rule makes this useful rather than noisy: **narrowing quarantines, widening does not**. A model that gains 4K is a strictly better lane and nothing stops; a model that loses `last_frame`, drops from 30 s to 10 s, flips `seed` from true to null, or loses the `audio` passthrough has invalidated whatever was approved on that basis. The subtle case the code is explicit about: a stated bound becoming `null` is a LOSS, not "no constraint" - `supported_durations` going from `[4..30]` to null is the catalog declining to state a bound, and a route admitted because 30 s was in range no longer has evidence for it. Reading null as an empty set would have called that a widening. A moved `canonical_slug` is its own cause (`model_version_changed`): every advertised bound can be identical and it is still a different model behind the same id, which is precisely the silent substitution this section exists to catch - and it is checked against the CURRENT record rather than only against today's diff, so a route pinned two snapshots ago is still caught on a quiet day. `watchedFieldsFor` resolves a route's declared capability dependencies to catalog fields THROUGH the 182.C.10.01 vocabulary, so a route needing `accepts-final-frame` automatically watches `supported_frame_images`, and a route needing `accepts-supplied-audio` (passthrough-sourced) or `max-image-references` (probe-sourced) correctly watches nothing here - a probe result is not invalidated by a catalog edit, and that seam belongs to 182.C.10.04. Pricing is split: a removed SKU takes the basis a cost estimate was built on and quarantines; a repriced SKU is reported and does not. Snapshots are one JSON file per distinct catalog, named `<date>-<hash12>` so the directory sorts chronologically and a human can read any day's catalog without the app; an unchanged daily refresh returns `persisted: false` instead of filling the store with identical files. The first refresh quarantines nothing and returns a NULL diff rather than inventing an empty baseline that would make every route look verified. 31 tests, including a diff of the real committed 28-model catalog against itself (no changes) and against a copy with a model removed. Targeted tsc against `tsconfig.typecheck.json` clean, verified by mutation.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Real 2026-09-10 catalog used as diff input; file store exercised against a real temp directory. Pure computation plus local filesystem - no provider call.                                                                                                                                                                | $0.00                           | None    |
| 182.C.10.04 | Combine advertised capability with live-probe status, benchmark promotion, regional availability, policy status …      | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-lane-eligibility.ts` (new), `human-video-lane-eligibility.spec.ts` (new)                                                                                     | Six inputs decide and every one can veto: what the catalog advertises today, whether a live probe confirmed it against THIS dated slug, benchmark promotion for the tier being asked for, regional availability, policy status, and health plus any 182.C.10.03 quarantine. The line that makes the phase's rule structural is in `grantCapability`: an entry the 182.C.10.01 vocabulary marks `requiresProbe` is DENIED while advertised-only, so "the catalog says it accepts audio" can never become "this lane does exact dialogue". Probes are bound to the dated `canonical_slug` they ran against - a probe of `wan-2.7-20260801` is not evidence about `-20260901`, it is evidence about a model no longer being served - which catches the silent substitution from the opposite side to the snapshot diff. Two states are deliberately NOT collapsed into "no", because collapsing them loses the only thing an operator can act on. `region_unrecorded` is not `region_unavailable`: nobody has written down where the lane may serve, and reading silence as "allowed everywhere" is how a likeness gets processed in a jurisdiction the authorization never covered, so it blocks and says which of the two it is. `probe_inconclusive` is not `probe_failed`: a probe that could not be read is not evidence of absence, and filing it as a failure would bury a lane that probably works. `advertisedCapabilities` derives the claim set from a real catalog record - `generate_audio: true` advertises both the native track and the switch to turn it off, a null field advertises nothing - and the passthrough name map is kept to `audio`/`video`/`videos`, the three measured on 2026-09-10. `keyframes` is deliberately NOT read as `supports-edit`: a parameter by that name could equally be first/last-frame control, and guessing is the fabrication this phase forbids. Newest probe wins in both directions (a re-probe exists to supersede), probes expire at 90 days, and `advertisedButUnproven` reports the probe backlog separately so Studio shows it as unverified rather than as a feature. Blockers are collected whole, not short-circuited, because "no lane is available" is not actionable while "two quarantined, one unhealthy, one never probed for supplied audio" is. 23 tests, driven off the real 28-model catalog for every advertisement assertion - one of which corrected a wrong assumption of mine mid-run: `alibaba/wan-2.7` does enumerate `supported_sizes`; the 9 records with a null size list are others. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Advertisement tests read the committed 2026-09-10 catalog directly. Probe, benchmark, health and region inputs are typed records; no provider call and no billed generation.                                                                                                                                              | $0.00                           | None    |
| 182.C.10.05 | Record the dated `canonical_slug` as the model version, a normalized pricing basis derived from the heterogeneous …    | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-model-record.ts` (new), `human-video-model-record.spec.ts` (new)                                                                                             | The catalog half and the manual half, kept visibly apart. `canonical_slug` is pinned as the model VERSION and the `id` is treated as an alias, because the id is what silently changes underneath an approved take. Duration bounds come from `supported_durations` and are NULL when the catalog states none - `heygen/avatar-iv` has no durations because its length follows the script - never 0 and never unlimited; output formats carry an `unstated` flag for the same reason. The pricing normalizer is the substance. Measured on the real catalog there are **37 distinct SKU keys across 28 models in six different units**, and the unit is carried by the KEY PREFIX with no field to read it from: `duration_seconds*` is USD per second while `cents_per_second_*` and `cents_per_video_output_second_*` are cents, plus `cents_per_megapixel_second_*`, `video_tokens*`, `cents_per_image_input`, `reference_images` and a `minimum_cents_per_generation` FLOOR that is not a rate at all. The dollars-versus-cents split is not a guess: `runway/aleph-2` publishes `cents_per_second_output: 28` beside `minimum_cents_per_generation: 56`, coherent as $0.28/s with a $0.56 two-second floor and absurd as $56, and `google/veo-3.1`'s `duration_seconds_with_audio: 0.40` is Google's own published per-second rate. Facets are peeled in a LOOP rather than a fixed order, because the catalog writes them both ways - `duration_seconds_with_audio_4k` puts the tier last, `video_tokens_4k_with_video_input` puts it in the middle - and a fixed-order parse silently fails on one of them. All 37 keys parse; a key that does not is surfaced as unrecognised and never folded into the nearest known unit, since mistaking cents for dollars is a 100x error in a number a creator acts on. **What cannot be estimated is the point.** Every Seedance model prices `video_tokens`, and the token count comes from the vendor's own unpublished tokenizer over resolution, frame rate and duration - so `estimateCost` returns `estimable: false, reason: 'token_priced'` rather than a number, and the megapixel-second upscaler returns `area_priced`. An estimate nobody can derive is worse than none, because it is shown to someone deciding whether to press render. SKU selection scores specificity and treats a contradiction as a refusal rather than a near miss: a `_without_audio` SKU cannot price a request that wants audio at any score. The four manual records (reference limits, data handling, retention, geography) are dated, expire at 180 days, and every field has a "not established" state that is not permissive - `maxImageReferences: null` is unprobed, not unlimited; empty `geographicRestrictions` is unrecorded, not worldwide. 20 tests, all pricing assertions driven off real records; the estimate rounds to micro-dollars to drop IEEE-754 noise (0.28x10 = 2.8000000000000003). Targeted tsc clean. **Correction 2026-09-10:** this row said geography, data handling and retention are things no catalog carries. True of `GET /api/v1/videos/models`, which is what it was measured against - but `GET /api/v1/providers` (105 records) carries `headquarters`, `datacenters`, `privacy_policy_url` and `terms_of_service_url`, so two of the four manual fields now have a machine-readable starting point. Neither states a retention window or a training policy, so a person still has to read the documents and `ManualModelRecord` is unchanged; what changed is that `sourceUrl` and the geographic input no longer start from nothing. Recorded in `evidence/2026-09-10-openrouter-video-transport/`. | All 37 live SKU keys parsed with 0 unrecognised; all 28 models normalize with 0 unrecognised keys. Pure computation over the committed 2026-09-10 catalog - no provider call.                                                                                                                                             | $0.00                           | None    |
| 182.C.10.06 | Add an operator override with scope, owner, reason, expiry, audit log, and automatic rollback …                        | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-operator-override.ts` (new), `human-video-operator-override.spec.ts` (new), `human-video-lane-eligibility.ts` (structured `cause` on a block)                | An override exists because a registry that can only say no gets worked around OUTSIDE the system, and a workaround leaves no record. Four properties make that safe rather than a hole. **Expiry is mandatory and capped at 30 days** - there is no permanent override, because a permanent exception is a policy change and belongs in the policy status; rollback is therefore automatic, since an expired override simply stops applying and nothing has to remember to remove it. **Some things cannot be overridden at all**: a `probe_failed` denial (a measurement, not a bureaucratic obstacle - overriding it ships something known broken), a `not_advertised` denial (there is nothing to enable), `region_unavailable` (a vendor does not start serving a region because we said so), a quarantine caused by `model_removed` (nothing at the other end of the route), and a `suspended` policy status (the one status that exists to stop exactly this). The last two are recognised from a STRUCTURED `cause` field added to `EligibilityBlock` rather than by pattern-matching the block's prose, which would have been a defect waiting for a reworded message. **The original decision is preserved**: `applyOverrides` never edits `blockers`, it returns `remainingBlockers` and `applied` beside the untouched decision, so `decision.eligible` (did it ship cleanly) and `eligibleUnderOverride` (did it ship at all) are different questions with different fields, and `releaseEvidenceFor` cannot report a lane as clean when it was not. **Refusals are audited beside successes** - an override someone reached for and could not use is a signal about the registry, and dropping it would leave the loudest cases invisible. Two smaller rules that matter: a reason under 20 characters is refused because a ticket number is not a reason, and a capability grant clears `capability_denied` only when EVERY denial on the lane has been granted, so a partial exception cannot ship whole. `lapsedSince` reports what rolled back in a window so an operator learns it from a report rather than from a failed render. 23 tests. Targeted tsc clean; the eligibility spec re-run green after the `cause` field.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Pure computation over typed records; no provider call and no billed generation.                                                                                                                                                                                                                                           | $0.00                           | None    |
| 182.C.10.07 | Expose a sanitized capability view to Studio for live UI enablement, estimates …                                       | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-studio-capability-view.ts` (new), `human-video-studio-capability-view.spec.ts` (new), `human-video-lane-eligibility.ts` (`availableCapabilities`)            | Sanitization here is STRUCTURAL, not a filter: the view is built by naming the handful of fields that may cross, with no spread of an internal object anywhere in the file, so a field added to a lane registration or a probe record later cannot leak by default. A filter-based sanitizer fails open the day someone adds a field; this one fails closed, and the test asserts it by serializing a view built from a lane carrying sentinel internals. Three things are deliberately coarser than internally. Health becomes `temporarily-unavailable` and nothing more - a creator cannot act on "7 consecutive failures since 11:04", and the count is an operational detail about our infrastructure. An override becomes "enabled temporarily, until <date>" with the OWNER AND REASON withheld: 182.C.10.06's requirement is that release evidence never hides an override, and that record keeps the full text; a creator-facing surface is not the place for a staff name. And vendor parameter names never appear - Studio speaks our capability vocabulary per ADR-0006 §6, because a UI built against `audio` breaks the day the vendor renames it and leaks which provider is behind a lane. What is NOT softened is the `unverified` state: advertised, unconfirmed, shown as neither enabled nor hidden, which is the entire point of the registry. A defect found by writing the tests: reading `decision.granted` for control state showed a probed, working capability as unavailable whenever the current request did not happen to require it, so 182.C.10.04 gained `availableCapabilities` ("what can this lane do") beside `granted` ("did this request's requirements hold") and the view reads the former. Estimates come through `estimateCost` and keep its refusals in plain words - a token-priced model says the cost cannot be predicted before rendering, an area-priced one says it depends on the source. Six warning kinds, including unrecorded reference limits and unrecorded retention terms, so the two manual-record gaps are visible where a creator makes the decision. 19 tests. Targeted tsc caught a fixture that vitest could not: the override spec's `LaneEligibility` literal was missing the new field and 23 tests still passed.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Views built from the real 2026-09-10 records (wan-2.7, veo-3.1, seedance-2.5, flux-video-upscale, heygen/avatar-iv, wan-3.0). No provider call.                                                                                                                                                                           | $0.00                           | None    |
| 182.C.10.08 | Test catalog additions, removals, partial fields, unknown options, stale caches, incompatible changes …                | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-catalog-scenarios.spec.ts` (new)                                                                                                                             | The unit specs check each module against its own contract; this drives the whole pipeline the way production will - fetch, snapshot, diff, quarantine, eligibility, override, Studio view - against the REAL 2026-09-10 catalog as the baseline, mutated one scenario at a time, asserting the outcome a creator or operator would see rather than an intermediate value. All seven named scenarios: an ADDITION reports the new model, quarantines nothing, and still denies its advertised capability as `unprobed` (with the companion case that a model which never offered it says `not_advertised`, a distinction a creator reads differently). A REMOVAL quarantines the route, turns the Studio lane to `temporarily-unavailable`, and cannot be rescued by an override. PARTIAL FIELDS splits in two: nulls are accepted, produce a `capability_narrowed` quarantine and a null max duration - a record that states less is not a broken record - while a record that loses `pricing_skus` is quarantined and the other 27 keep working. UNKNOWN OPTIONS reports a new catalog field as unmapped without any downstream reading it as a capability, and an unfamiliar passthrough name (`lipsync_strength`) advertises nothing. A STALE CACHE serves the last good catalog through an outage, labels it, and surfaces `capability-data-stale` in the Studio view, then refuses past 7 days. An INCOMPATIBLE CHANGE quarantines a pinned route when the dated slug moves AND stops the old slug's probes counting, while a narrowing can be taken on by an operator - on a clock that expires on its own. TOTAL OUTAGE fails loudly with nothing cached, refuses a 200 carrying zero usable models or an HTML error page, and leaves the last good snapshot intact for the next refresh. 16 tests; full BFF human-video suite 793 green across 36 files. This completes 182.C.10.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Every scenario is a mutation of the committed 2026-09-10 response, so the baseline is a real catalog rather than a fixture. Test doubles at the fetch boundary; no provider call.                                                                                                                                         | $0.00                           | None    |

### 182.C.11 Complete the OpenRouter Video Transport

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner     | Dependency | Status   | Changed Files                                                                                                                                                                                                                                                                                                                     | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Provider/Hardware Evidence                                                                                                                                                                                                                                                   | Cost                                        | Blocker |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------- | ------- |
| 182.C.11.01 | Extend the create payload with image-only `input_references` (the documented shape is …                                | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-openrouter-payload.ts` (new), `human-video-openrouter-payload.spec.ts` (new), `human-video-reference-ordering.ts` (exported `IMAGE_REFERENCE_ROLES`, new `no-passthrough-on-route` drop reason), `docs/domains/isis/human-video/evidence/2026-09-10-openrouter-video-transport/` (new) | The transport has exactly ONE documented reference field - `input_references`, entries `{ type: 'image_url', image_url }` - and it is image-only. There is no audio reference type and no video reference type; an earlier draft of this phase assumed otherwise and was corrected against the live API on 2026-09-09, and this module is where that correction is enforced rather than remembered. Audio and video therefore have exactly one route: the model-specific `provider` passthrough object, and only where the model's own `allowed_passthrough_parameters` names a parameter. Measured 2026-09-10: `audio` on `alibaba/wan-2.7` and `-2.6`; `video`, `videos`, `images`, `last_image` on `wan-2.7`; nothing of the kind on any Seedance, Veo, Kling, Sora or HeyGen entry. Failing closed is the point: when a request needs a class the model has no parameter for, it THROWS, because the alternative is a payload that looks valid provider-side, renders, and comes back without the creator's approved speech in it - a silent downgrade from exact dialogue to prompt-native, which 182.C.03.07 forbids. When the class was not required it is dropped with a reported reason instead. A second refusal the catalog forced: `wan-2.7` names a SINGULAR `audio`, so two approved takes for one shot have no honest destination - sending an array through a singular parameter guesses at the vendor's parser and keeping the first would drop approved audio silently - so it throws. Ordering comes from 182.C.07.05 rather than being restated, and `IMAGE_REFERENCE_ROLES` is now exported from there so the two modules cannot disagree about what an image is; a disagreement would show up as a reference quietly going nowhere. First- and last-frame references are kept OUT of `input_references` because `frame_images` takes precedence when both are present, so sending the same image twice is at best redundant - they are returned as frame candidates for 182.C.11.02, and a test asserts no signed URL is even resolved for them yet. The URL resolver is injected because signing is a boundary concern and the result is a credential-bearing URL 182.C.11.10 forbids logging. Payloads are deterministic: a test builds the same request from a reversed input list and compares the serialized result, because a retry that reorders is a different request. 14 tests. Targeted tsc clean. **Correction, same day:** fetching the official guide for 182.C.11.02 showed two shapes wrong here. `image_url` is an OBJECT `{ url }`, not a bare string - the phase snapshot's `{ type: 'image_url', image_url }` reads naturally as a string and is not one. And the `provider` object is NOT a flat map: the documented shape is `{ options: { '<provider-slug>': { parameters: { ... } } } }`, and the guide states only the options for the MATCHED provider are forwarded - so a flat `{ audio: ... }` is accepted by the API and forwarded to nobody, which is the same silent downgrade this module exists to prevent, reached by a shape error instead of a routing decision. Both corrected. The slug turns out not to be derivable from the model id either: of the ten video model-id prefixes, `google`, `bytedance`, `x-ai` and `kwaivgi` do not appear in `GET /api/v1/providers` (the guide's own example serves `google/veo-3.1` through `google-vertex`, and `google-ai-studio` also exists), so it is a dated manual record and a new `ProviderSlugUnknownError` fails closed when a passthrough parameter is needed and no slug is recorded. 17 tests. Evidence: `evidence/2026-09-10-openrouter-video-transport/`. **Second correction, same day:** the conclusion that the provider slug is a manual record was the wrong conclusion from the right observation. `GET /api/v1/models/{id}/endpoints` answers it live and unbilled, and the full table for all 28 video models is now committed as `provider-slugs.json` with a `resolveProviderEndpoints` client that fails closed. **15 of the 28 do not match their own id prefix** - more than half - and the decisive pair is `alibaba`: `wan-2.7` is served by `atlas-cloud` while `wan-3.0` is served by `alibaba`, with nothing in the model id to distinguish them. | Passthrough name sets in the tests are the live 2026-09-10 values for wan-2.7, wan-2.6 and seedance-2.5. Injected URL resolver; no provider call.                                                                                                                            | $0.00                                       | None    |
| 182.C.11.02 | Support first- and last-frame `frame_images` independently of general identity/scene references.                       | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-openrouter-payload.ts` (frame images), `human-video-openrouter-payload.spec.ts`                                                                                                                                                                                                        | Frame images are independent of identity/scene references BY CONSTRUCTION: a frame never occupies an image-reference slot, so a lane's `maxImageReferences` limit can never cost a creator their first or last frame - a test drops three references to the limit and asserts the frame survives. Entries carry the documented `frame_type` (`first_frame`/`last_frame`) alongside `image_url: { url }`, and first is emitted before last whatever order they arrived in, so a retry is the same payload. Two refusals. An unsupported slot THROWS rather than dropping the frame: a last-frame reference is a hard constraint on where the shot ENDS and a take that ignores it is a different shot, not a slightly worse one - and 9 of the 28 models on 2026-09-10 are first-frame-only, so this is the common case rather than the exotic one. Two references claiming the same slot also throw, because picking one for the creator is a choice we would not be able to explain later. The third and most consequential rule came from the guide's exact wording, read on 2026-09-10: _"If both fields are provided, `frame_images` takes precedence and the request is treated as image-to-video."_ The MODE changes - the identity references are not degraded, they are not read at all - so `assembleImageFields` refuses a payload carrying both unless the caller passes `acknowledgeImageToVideo`, making the mode change a decision somebody took rather than a side effect of two features both working. When it IS taken deliberately, `input_references` is omitted rather than sent-and-ignored, so the stored request evidence does not say something untrue about the render. 27 tests in the file (10 new). Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | `supported_frame_images` values and the both-fields precedence rule read from the official guide 2026-09-10; extracts committed under `evidence/2026-09-10-openrouter-video-transport/`. No provider call.                                                                   | $0.00                                       | None    |
| 182.C.11.03 | Support `generate_audio` (it defaults to true on audio-capable models, so exact-audio requests must set it …           | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-openrouter-create.ts` (new), `human-video-openrouter-create.spec.ts` (new), `human-video-openrouter-payload.ts` (flat `providerParameters` for merging)                                                                                                                                | Every field is checked against the model's own catalog record and every problem is collected rather than thrown one at a time, so a caller fixing a request sees all of it at once instead of discovering the next issue after each round trip - a test drives six simultaneous problems against `openai/sora-2-pro`. **`generate_audio` needed the most care.** The guide says it defaults to TRUE for models that support audio output, so an exact-audio request that merely omits the field gets a model-generated track beside or instead of the creator's approved speech. Omission is not neutral there; it is a request for the wrong thing. So a supplied-audio request must set `generate_audio: false` EXPLICITLY and omitting it is an error rather than a default we quietly fill in - filling it in would have been the convenient choice and would have meant the one field deciding whether dialogue is exact was never a decision anyone made. Setting it true alongside supplied speech is refused outright, and on a model advertising `generate_audio: false` or `null` the flag is rejected as claiming something the model does not do or does not state. **Seed is the mirror image**: advertised per model (true on 16 of 28, false on 9, null on 3) and silently ignored where unsupported, after which continuity work believes it has a reproducible take and does not - so a seed on an unadvertised model is rejected rather than dropped, with `seed_unstated` kept distinct from `seed_unsupported`. `size` and `resolution`+`aspect_ratio` are documented as interchangeable, and sending both is refused rather than resolved, because the two can disagree and picking a winner would make the payload say something the caller did not. `callback_url` must be HTTPS and parseable - a callback in the clear carries a signed job notification. The `provider` object is checked name by name against `allowed_passthrough_parameters` with EVERY disallowed name reported, and needs the recorded provider slug once any parameter is used. Optional fields are omitted rather than sent as `undefined`, so the request stored as evidence is exactly what went over the wire. 23 tests, all driven off real catalog records (wan-2.7, heygen/avatar-iv, sora-2-pro, hailuo-2.3, hailuo-3-max, grok-imagine-video). Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Validation is against the committed 2026-09-10 records; the `generate_audio` default and the size/resolution interchangeability are from the official guide, extracted under `evidence/2026-09-10-openrouter-video-transport/`. No provider call.                            | $0.00                                       | None    |
| 182.C.11.04 | Validate each payload against the selected live capability record and reject unsupported fields before provider …      | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-openrouter-create.ts` (submission guard), `human-video-openrouter-create.spec.ts`                                                                                                                                                                                                      | `validateCreatePayload` deliberately re-derives its own view of the payload rather than trusting the builder: 182.C.11.03 validates an INPUT and this validates the OBJECT that will be serialized, so a payload assembled elsewhere, mutated by a caller, or replayed from a stored record passes the same gate. An undocumented top-level field is REJECTED rather than forwarded, and the reason is the interesting one - OpenRouter would very likely ignore it, and that is the problem: a misspelled `generate_audio` gets ignored into a model-generated track over the creator's approved speech while the stored request shows a field that never did anything. Frame entries are checked for the documented `{ type, image_url: { url }, frame_type }` shape AND for a slot the model actually offers; reference entries for the same object shape - which now also catches the exact bug 182.C.11.01 shipped this morning, `image_url` as a bare string, from the other side. Both image fields together are rejected whoever assembled them. The `provider` object is checked for the nested `{ options: { '<slug>': { parameters } } }` shape, because a flat one is accepted by the API and forwarded to nobody; more than one slug is reported too, since only the matched provider's options travel and the rest are silently inert. Every problem is collected, and `assertSubmittable` throws the whole list. 34 tests in the file (11 new). Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Documented field list taken from the guide's request-parameter table, 2026-09-10; slot and passthrough checks run against the committed catalog records. No provider call.                                                                                                   | $0.00                                       | None    |
| 182.C.11.05 | Preserve OpenRouter request ID, generation ID, polling URL, model, resolved upstream provider where available, usage … | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-transport-record.ts` (new), `human-video-transport-record.spec.ts` (new)                                                                                                                                                                                                               | Six months after a take ships the questions are which model rendered it, what it cost, who upstream served it, and what the provider said when it failed - none of which the video file answers, so they are captured at the moment of the exchange or not at all. The resolved upstream provider comes from `X-Provider-Name`, and the way we know that header exists is worth recording: the catalog response advertised `access-control-expose-headers: X-Generation-Id, X-Provider-Name, request-id, cf-ray` on 2026-09-10. That is the CORS expose list - it names the headers the API intends callers to read and is NOT a promise that every response carries every one - so each is captured as nullable and a missing header is recorded as absent rather than filled in. Three refusals to guess. An unrecognised status is preserved VERBATIM under `unknown` instead of being coerced to the nearest known one, because a status we have not seen is exactly what an incident needs read back accurately - and it is not assumed terminal, so a poll loop keeps its bound. A later response that omits an identity never erases what an earlier one told us: `generation_id` appears on the poll and not on the create, and `X-Provider-Name` may be present on one response and absent on the next. And a create response with no `id` THROWS rather than yielding a record with an empty one - an un-pollable job that looks started is worse than a failed create, because nothing will resolve it and the cost is already incurred. Evidence is stored redacted per 182.C.11.10: query strings are stripped from every URL (a download URL's query is where a token lives, verified against the documented `content?index=0` form), any field whose NAME reads like a credential is replaced by substring match so `provider_api_key` and `downloadToken` both go without being listed first, data URIs are cut at the comma, and long strings truncated - but unknown fields are KEPT, because a field OpenRouter adds is the most useful thing in an incident and an allow-list redactor would throw it away. 19 tests, driven off the guide's own documented 202 and poll bodies. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Body shapes from the guide's response-format section; header names from the live `access-control-expose-headers` value observed 2026-09-10. No billed generation - capture is exercised against the documented bodies.                                                       | $0.00                                       | None    |
| 182.C.11.06 | Implement callback/webhook receipt that verifies the HMAC-SHA256 `X-OpenRouter-Signature` …                            | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-webhook-receiver.ts` (new), `human-video-webhook-receiver.spec.ts` (new), `evidence/2026-09-10-openrouter-video-transport/README.md` (webhook contract)                                                                                                                                | Built against the guide's own verification steps, read 2026-09-10: extract `t` and `v1` from `X-OpenRouter-Signature: t=…,v1=…`, build `{timestamp},{raw_request_body}` joined with a COMMA, HMAC-SHA256 it with the workspace signing secret, compare hex. Two rules make it safe rather than convenient. **An unverified delivery never advances a job.** The guide is explicit that the signature header appears only when a signing secret is configured, so an unconfigured receiver cannot tell OpenRouter's callback from anyone else's POST - so instead of trusting it the receipt says `pollInstead`, because the job's real state is knowable from the API and one request is cheaper than a wrong belief about a render. That applies to every rejection, not only signature failures. **The RAW body is what gets verified**: the signature covers exact bytes, and a body parsed and re-serialized is a different string, so the signature is checked before the parser ever runs - a test signs compact JSON, delivers the pretty-printed equivalent, and asserts the mismatch. The replay window is the guide's 300 s, closed in BOTH directions: a future timestamp is the same window pointed forward at a skewed clock. Comparison is timing-safe and length-checked, so a short hash returns a mismatch rather than throwing. All four event types parse, and the sentence the receiver is built around is honoured: _"generation_id and model in data may be null when a job fails before those values are assigned"_ - so both are optional and a null is carried through as null rather than treated as malformed, because an early validation failure is exactly the delivery you most want to keep. Idempotency prefers OpenRouter's own `X-OpenRouter-Idempotency-Key` (`<job_id>-<status>`) and derives the same value from the body when the header is absent, so a header-less retry still dedupes; a DIFFERENT status on the same job is correctly not a duplicate. `pollFallbackDue` covers the case with no signal at all - a callback that never arrives leaves a job pending forever, and the absence is the problem - so the poller takes over past a deadline measured from the last delivery, or from submission when there has been none. 21 tests, including real HMACs computed with node:crypto against the guide's documented payloads. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Webhook payloads, header names, signature scheme and the null-`generation_id` sentence transcribed from the official guide 2026-09-10 into `evidence/2026-09-10-openrouter-video-transport/README.md`. Signatures in the tests are computed, not fixtures. No provider call. | $0.00                                       | None    |
| 182.C.11.07 | Honor retry/backoff headers, bound polling, classify retryable versus permanent errors …                               | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-poll-policy.ts` (new), `human-video-poll-policy.spec.ts` (new), `evidence/2026-09-10-openrouter-video-transport/README.md` (cancellation probe)                                                                                                                                        | Retry disposition differs by CALL, not just by status code, and that asymmetry is the substance: a **create** retries only what is clearly retryable, because a repeated create can bill a second generation - losing a request costs a request, duplicating one costs a render - while **poll** and **download** also retry the ambiguous cases, because the job is already paid for and abandoning the poll abandons the output. An unrecognised status stays `unknown` rather than being sorted into the nearest bucket, and `shouldRetry` decides what that means per call. **Cancellation turns out not to be exposed**, and the evidence is in the shape of the 404s rather than their status: probed 2026-09-10, `GET /api/v1/videos/{id}` on a missing job answers the API's JSON envelope `{"error":{"message":"Job … not found","code":404}}` while `DELETE`, `POST`, and `POST …/cancel` on the same path all answer a plain-text `404 Not Found` - a JSON error means the route exists and the job does not, plain text means there is no such route. The guide documents none either; `cancelled` is only a webhook event, so a job can BE cancelled without us being able to cancel one. `requestCancellation` throws a dated `CancellationNotExposedError` carrying that evidence rather than being a missing function a caller assumes something about. **Exhausting our poll budget is not the job failing** - the guide's own troubleshooting says generation can take several minutes and to keep polling - so `budget_exhausted` comes back with `jobStillLive: true` and nothing is written into the record; four of the five poll outcomes decline to state anything about the job, and only `terminal` reports a verdict because the provider gave one. A permanent 404 marks the job gone; any other permanent failure marks it still live, because being unable to LEARN about a job is not the same as it having stopped. `expired` is terminal, with the guide's own wording. `Retry-After` is honoured over our curve when present (0 occurrences in the video guide, so never depended on) and a wait beyond 60 s stands the loop down rather than holding a caller. 20 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Cancellation probe (4 requests, unbilled) and the troubleshooting quotation recorded in `evidence/2026-09-10-openrouter-video-transport/README.md`. Error-envelope test uses the exact body the live API returned.                                                           | $0.00 (probes are unbilled 404s)            | None    |
| 182.C.11.08 | Replace base64 MP4 data URIs with streamed authenticated download to quarantine object storage …                       | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-output-download.ts` (new), `human-video-output-download.spec.ts` (new)                                                                                                                                                                                                                 | What this replaces: the existing OpenRouter video adapter fetches the clip and projects it to a `data:video/mp4;base64,…` URI - the whole video in memory twice, once as bytes and once as base64 a third larger, then carried as a string through every layer above. A 20-second 1080p take is tens of megabytes; the base64 of it is not a URI. **The security point is the one worth reading twice.** The download URL comes from the PROVIDER (`unsigned_urls` in the poll response) and the guide is explicit that fetching it requires our Bearer key - so we would be attaching our API key to a URL a remote service chose. The host is therefore checked against an allow-list BEFORE the Authorization header is attached, and a URL that fails is refused rather than fetched header-less: a URL we cannot authenticate to is not a URL we should follow. A test asserts that an attacker host leaves the quarantine store completely untouched. Everything else follows from "the bytes are untrusted until measured": quarantine first under a prefix nothing serves from, PROMOTE LATER as a separate call the caller makes only after 182.C.11.09 has probed the file - so a take that turns out to be an HTML error page never occupies the key a release serves from - digest incrementally while streaming, because hashing afterwards means holding the file to hash it and that is the thing being removed, and bound the size TWICE, on the declared `Content-Length` before the first byte and on the running count while streaming, because a declared length is a claim. Every failure path aborts the writer first, so a partial object never sits in the store waiting to be mistaken for a complete one - asserted for oversize, empty, mid-stream disconnect and digest mismatch. An HTML content-type served with a 200 is refused; `application/octet-stream` is accepted, because providers do send mp4 that way. The stored `sourceUrl` drops its query, where a token would live. 18 tests, one of which asserts the result contains neither `data:` nor `base64`. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Download URL form and the Bearer requirement from the guide's "Downloading the Video" section, 2026-09-10. Injected fetch and quarantine store; no bytes are fetched from a provider and no generation is billed.                                                            | $0.00                                       | None    |
| 182.C.11.09 | Extend the existing ffprobe/ffmpeg measurement in `video-probe.ts` to reject missing, truncated, HTML/error …          | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-output-validation.ts` (new), `human-video-output-validation.spec.ts` (new)                                                                                                                                                                                                             | `video-probe.ts` measures duration, resolution, sampled-frame safety and luma - and ASSUMES the file is a video. This is the check that runs first and decides whether that assumption holds, because the failure modes a provider download actually produces are not "the video is poor" but "this is an HTML error page with an .mp4 name" and "the connection dropped at 13 of 100 frames and the header still says four seconds". Every threshold was MEASURED on this box on 2026-09-10 against files built for the purpose. **HTML vs truncated:** ffprobe reports the SAME failure for both - `moov atom not found / Invalid data found` - so it cannot tell a 502 page from a cut-off download. The leading bytes can: a real mp4 opens `00 00 00 20 66 74 79 70` (`ftyp` at offset 4) and KEEPS it when truncated, while the error page opens `<html>`. So the container is sniffed from the first bytes before ffprobe runs and the two get different codes. **Truncation is invisible to duration:** a 4-second `+faststart` clip cut to 8 KB still reports `duration: 4.000000` from the moov at the front while decoding yields 13 frames - 0.52 s of the four it claims - so truncation is caught by comparing DECODED frames against the declared duration, not by reading the duration. **Silence is invisible to stream layout:** a silent AAC track is a normal audio stream with the same `nb_read_frames` as a real one, and only the level differs; measured, a 440 Hz sine came back at `max_volume: -17.7 dB` and an `anullsrc` track at **-91.0 dB**, the AAC encoder's floor rather than true -inf, which is why the threshold is -60 dB and not "equals -inf". A missing audio stream and a silent one are DIFFERENT rejections, because a lane that returned no track failed differently from one that returned a track of nothing and the retry differs too. One more measured trap: `volumedetect` writes to stderr at info level, so the binary is run WITHOUT `-v error` - suppressing it is what makes the measurement silently return nothing. Every rejection is collected rather than stopping at the first. 16 tests, all against files ffmpeg builds in `beforeAll` and the real binaries measure - a mocked probe would assert nothing about any of these numbers. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Real on-box ffmpeg 7.x / ffprobe. Fixtures: 2 s testsrc+sine, testsrc+anullsrc, testsrc with `-an`, a `+faststart` 4 s clip cut to 8 000 bytes, an HTML body and a JSON error body named `.mp4`. Measured values quoted above are from this run.                             | $0.00                                       | None    |
| 182.C.11.10 | Store only redacted payloads; never log API keys, signed reference URLs, biometric bytes, full private scripts …       | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-redaction.ts` (new), `human-video-redaction.spec.ts` (new), `human-video-transport-record.ts` (delegates its redactor here)                                                                                                                                                            | Five classes must never reach a log line: API keys, signed reference URLs, biometric bytes, full private scripts, and provider download tokens. FOUR of them must not reach stored evidence either, and the fifth is why this has two profiles rather than one rule. `stored-evidence` KEEPS the script - 182.C.01.05 exists so that "why did this shot look like that" is answerable six months later, and a provenance record with the prompt removed cannot answer it; the script is the creator's, inside our boundary, under their consent record. `log` does not: a log line is copied into aggregators, alerts and screenshots, and a creator's unreleased dialogue has no business in any of them, so the script becomes its length and a 16-hex digest - enough to confirm two runs used the same script, useless to a reader. A DIGEST rather than a truncation, because the first eighty characters of a script are still the script. Signed URLs lose their query in both profiles - that is where a token lives, for the reference URLs we mint and the `unsigned_urls` OpenRouter returns alike - and data URIs are replaced by media type, byte count and digest, because a base64 blob in a stored record is the biometric bytes a second time. Field names match by SUBSTRING so `provider_api_key`, `downloadToken` and `x_signature` are caught the first time they appear rather than after the incident that revealed them, and credential PATTERNS (`Bearer …`, `sk-…`, `whsec-…`) are scrubbed even from free text nobody named. `findSecretMarkers` reports WHERE a marker survived rather than that one did, and `assertRedacted` throws with every finding - cheap enough to run before a write. 182.C.11.05's response-side redactor now DELEGATES here rather than being a second implementation: two redactors drift, and the one that drifts is the one nobody tests. That refactor improved one behaviour - the old evidence form `data:video/mp4;base64,[redacted]` still literally contained `data:` and `;base64,`, which this module's own leak guard treats as markers; it is now a summary object, and the 11.05 test records why it changed. 17 tests, driven off a payload carrying one of each of the five classes. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Pure computation; no provider call. The leak guard is asserted against a payload built to break it.                                                                                                                                                                          | $0.00                                       | None    |
| 182.C.11.11 | Add contract tests from official examples plus real, redacted canary fixtures for create, poll, callback, download …   | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-transport-contract.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-10-canary-exchanges/` (new), `human-video-output-validation.ts` (`selectPrimaryVideoStream`)                                                                                                        | Two sources, and the distinction is the point. The OFFICIAL EXAMPLES are the guide's own request and response bodies transcribed verbatim - a module that cannot read them is wrong whatever the API does today - and the CANARY is a real billed exchange recorded the same day: `x-ai/grok-imagine-video`, 1 second, 480p, no people in the prompt, create → 3 polls → download, ~18 s end to end, **$0.05** (user approved a ~$5 cap; running tally in the evidence README). **The canary checked the pricing normalizer against money**: `cents_per_video_output_second_480p: 5` × 1 s predicted $0.05 and `usage.cost` came back 0.05. It also confirmed the download needs the key - without `Authorization` the content URL answers **401**, which is exactly what 182.C.11.08's host allow-list protects - and that no `Content-Length` is sent, so the streaming byte count is the only size bound that fires. **Two corrections it forced.** First, the headers the CORS list advertises are NOT sent on this route: neither `request-id`, `X-Generation-Id` nor `X-Provider-Name` appeared on the create or any poll, only `content-type` and `cf-ray`. No code changed because 182.C.11.05 already captured each as nullable, but that row now rests on a measurement instead of an inference. Second, and this one was a real defect: the clip came back with THREE streams - h264 848x480 at 24 fps with 25 frames, an aac track, and an `mjpeg` stream of a SINGLE FRAME at the same dimensions, an embedded thumbnail. 182.C.11.09 took the first `codec_type: "video"` entry, which happened to be the h264 one; ordered the other way it would have called the take `video_codec_unsupported` or read one frame of a one-second clip as a truncation. Fixed by `selectPrimaryVideoStream`, and the contract test reverses the stream order to prove it. One more thing worth keeping: `x-ai/grok-imagine-video` carries `generate_audio: null` and the take came back WITH an aac track - null really is "the catalog declines to state", which is what the vocabulary's `fallbackWhenAbsent: 'probe'` was built for, now confirmed rather than reasoned. Two contract tests assert the shapes an earlier draft assumed - a string `image_url` and a flat `provider` - are now rejected. **Timeout has no canary and the file says so**: a real `expired` job needs one to sit past a TTL the provider does not publish, so that contract rests on the guide's envelope plus our own bound, and the two are asserted to be different outcomes. 22 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Real billed exchange committed redacted at `evidence/2026-09-10-canary-exchanges/` (JSON only - the clip's sha256, byte count and full ffprobe output are kept, the 588 KB of bytes are not). Official examples transcribed from the guide the same day.                     | $0.05 (running total $0.05 of ~$5 approved) | None    |
| 182.C.11.12 | Add integration tests proving temporary provider URLs can expire after durable ingestion without breaking the …        | Oshun BFF | 182.C.10   | complete | `apps/oshun/bff/src/generation/human-video-durable-ingestion.spec.ts` (new), `evidence/2026-09-10-canary-exchanges/README.md` (expiry observation)                                                                                                                                                                                | Runs against a REAL filesystem store rather than an in-memory double: quarantine is a temp directory, promotion is an atomic rename, and serving reads the bytes back off disk. It needs no external service so it is a plain spec rather than one of the repo's `.integration.` files, but the property is an integration property - after ingestion, nothing in the release path touches the provider again. The measurement that shapes it: the canary's `unsigned_urls` entry was STILL answering 200 with all 588 899 bytes five minutes after the job completed, so **no expiry was observed** and the TTL is not published. That is exactly why the property is proved structurally rather than by waiting - an expiry we have not seen is not one we can rely on not happening. Serving takes a durable key and no URL and no fetch, so it cannot reach the provider even by accident; the test asserts that across 404, 410, 401 and 403, and that the provider was called exactly ONCE, during ingestion. A separate case shows the release surviving even when the whole provider host would now be refused by 182.C.11.08's allow-list. Two more properties worth having: a URL that dies mid-ingestion fails the INGESTION and leaves nothing promoted, so nothing half-ingested can be served; and because the durable key is content-addressed by digest, re-ingesting identical bytes lands on the identical key rather than forking the released asset. The stored `sourceUrl` keeps the path for provenance and drops the query, the only part that could carry a token. 7 tests. This completes 182.C.11. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Real temp-directory store and rename; expiry observation recorded against the live canary job at `evidence/2026-09-10-canary-exchanges/README.md`. No additional spend.                                                                                                      | $0.00 (running total $0.05 of ~$5 approved) | None    |

### 182.C.12 Model-Specific OpenRouter Adapters

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner     | Dependency | Status   | Changed Files                                                                                                                                                       | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Provider/Hardware Evidence                                                                                                                                                                                                    | Cost  | Blocker |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----- | ------- |
| 182.C.12.01 | Implement a Wan-family adapter with `alibaba/wan-2.7` as the primary route (first/last frame, 2–10 s …                 | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-lane-adapters.ts` (new), `human-video-lane-adapters.spec.ts` (new)                                                       | ADR-0006 §6 says the vendor vocabulary stops at the edge; this is the edge. Above it a request asks for `supplied-speech` and `negative-direction`, below it one model calls those `audio` and `negative_prompt` and another has no name for either. The rule that keeps the table honest is the phase's own: **map each control only where the live catalog names it** - so a mapping is never trusted on its own. Every passthrough send is checked against the model's `allowed_passthrough_parameters` from the record read that day, and a parameter the record does not name is NOT sent whatever this file says; a test removes `audio` from a copy of the wan-2.7 record and asserts the control becomes ignored (or, if required, a refusal). A table that has drifted ahead of the catalog produces a report, not a payload the API quietly drops. The other half is what happens when a control cannot be carried: a REQUIRED one fails the request, because if a creator's approved speech cannot reach the model then rendering anyway produces a take with the wrong words and no record of why; an OPTIONAL one is reported in `ignoredOptionalControls` with the reason, never silently dropped, because a negative prompt that vanished is the difference between a creator thinking the model ignored their note and knowing we never sent it. Fallback is deliberately narrow: `selectAndCompile` moves to wan-2.6 only on a required control the primary cannot carry, never to save an optional one - a fallback that costs a last frame to save a negative prompt is a worse trade than the one the creator was making - and it reports `fallbackCost` in words rather than pretending the two models are the same. The mapping tables are the measured difference between them: 2.7 names `ratio`, `videos`, `prompt_extend` and takes a last frame; 2.6 names `size`, `shot_type`, `enable_prompt_expansion`, first frame only, 5-10 s. A test asserts each model maps nothing the OTHER one has. Consumed controls come back in vocabulary order rather than caller order, so the report is identical for an identical request. 16 tests, every mapping cross-checked against the committed 2026-09-10 records. Targeted tsc clean. **Correction, same day:** this shipped `providerSlug: 'alibaba'` for both Wan models, taken from the id prefix. The endpoints API says **`atlas-cloud`**. Keyed under `alibaba`, wan-2.7's audio passthrough would have been accepted by the API and forwarded to NOBODY - a take back without the creator's approved speech and a payload that looked perfectly valid, which is precisely the failure 182.C.11.01 fails closed on. Corrected, and the spec now asserts each model's slug against the recorded table rather than a literal. | Every passthrough name in both tables is asserted present in that model's live catalog record; the provider slug `alibaba` is from `GET /api/v1/providers`. No provider call.                                                 | $0.00 | None    |
| 182.C.12.02 | Implement a Seedance 1.5 Pro-family adapter for prompt-native audiovisual generation only (first/last frames, 4–12 s … | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-lane-adapters.ts` (`SEEDANCE_15_FAMILY`, `refusedModes` typed by `HumanVideoMode`), `human-video-lane-adapters.spec.ts`  | A prompt-native audiovisual lane and nothing more. It renders sound - the catalog says `generate_audio: true` - but it has **no audio input on OpenRouter**: `allowed_passthrough_parameters` is exactly `watermark` and `req_key`, and the transport has no audio reference type. So the words in a take from this model are the model's GUESS at the script, not the script. **The item's own wording hides a trap and the adapter refuses it.** 182.C.12.02 describes multi-character dialogue as a strength of this family, and it is - but `multi-character-scene` is one of the contract's `EXACT_DIALOGUE_MODES`, a mode that promises the approved words for each speaker. A model that never receives them cannot serve it. So BOTH exact-dialogue modes are refused, not just the obvious one, and the refusal says where multi-character work belongs instead: `text-to-scene` or `reference-to-video`, where the dialogue is REPORTED rather than promised. `refusedModes` is now keyed by `HumanVideoMode` rather than by string, so a typo'd mode name fails to compile instead of quietly never refusing - the worst possible failure for a table whose whole job is to say no. Camera direction and multilingual prompts are real strengths of this family and they travel in the compiled prompt (182.C.09.09); there is no parameter for either, so there is no mapping for either, and a test asserts none was invented. The provider slug is `seed`, from the endpoints API - not `bytedance`, which is what the id says. 23 tests in the file (7 new). Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Passthrough list, `generate_audio` and durations from the committed 2026-09-10 record; slug from `GET /api/v1/models/bytedance/seedance-1-5-pro/endpoints`. No provider call.                                                 | $0.00 | None    |
| 182.C.12.03 | Implement a Seedance 2.5-family adapter for what OpenRouter exposes (image-only `input_references`, first/last frame … | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-lane-adapters.ts` (`SEEDANCE_25_FAMILY`, `availableElsewhere`), `human-video-lane-adapters.spec.ts`                      | The longest-take lane OpenRouter exposes - 4 to 30 seconds where wan-2.7 stops at 10, asserted against both records - with image-only `input_references`, first and last frame, 480p/720p, `generate_audio`, seed, and exactly the three passthrough names the record carries (`watermark`, `req_key`, `output_format`; the 1.5 model has no `output_format`). **What it cannot do HERE is the interesting half.** Audio and video references, edit and extend are real Seedance capabilities and none is reachable through OpenRouter - the catalog names no parameter for any of them - so the adapter fails closed on each AND says where they are available. That distinction earned a new ignored-reason, `available_on_another_route`: "Seedance 2.5 cannot edit" is wrong while "Seedance 2.5 cannot edit through OpenRouter, and the direct client can" is right, and a creator told the first thing gives up on something they could have had. The refusals name `@isis/seedance-provider` and 182.C.13.01 by hand rather than gesturing at "another route". Being the long-take lane and being the audio-input lane are different jobs and this is not the second one, so both exact-dialogue modes are refused for the same reason as on Seedance 1.5. Slug `seed`, from the endpoints API. 30 tests in the file (7 new). Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | Durations, resolutions and the three passthrough names from the committed 2026-09-10 record, cross-checked against wan-2.7 and seedance-1-5-pro. No provider call.                                                            | $0.00 | None    |
| 182.C.12.04 | Implement a Veo 3.1-family adapter (`google/veo-3.1`, `-fast`, `-lite`) for prompt-native audio, first/last frames …   | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-lane-adapters.ts` (`VEO_FAMILY`, consent gating), `human-video-lane-adapters.spec.ts`                                    | All three models on the `google-vertex` slug - which the official guide's own example uses and which the endpoints API confirms - in quality order, 4/6/8 s, 720p through 4K on 3.1 and `-fast`, 720p/1080p on `-lite`. **`personGeneration` is not a setting, it is a CLAIM.** Sending `personGeneration: "allow"` asserts to Google that we may depict people in this render, and that assertion is only ours to make when a consent record says so. So the adapter gained a consent gate: a control listed in `requiresConsentFor` with no record, or with an inactive one, is REFUSED - and never as an ignorable optional, because the claim is the problem rather than the requirement. `aspectRatio` is deliberately NOT mapped even though the record names it: the transport already has `aspect_ratio` and 182.C.11.03 refuses a payload that sets a dimension two ways, so mapping the passthrough as well would manufacture exactly that conflict. No audio input, so both exact-dialogue modes are refused - prompt-native audio is not the approved script. The fallback notes say the thing that matters and is easy to get wrong: `-fast` and `-lite` are DIFFERENT MODELS with their own benchmarks, not cheaper settings of 3.1, so a promotion for 3.1 does not cover them and 182.C.10.04 blocks the fallback until each has its own; `-lite` additionally renders no 4K at all, which the test asserts against the two catalog records rather than taking on trust. 38 tests in the file (7 new). Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Resolutions, durations and the five passthrough names from the committed 2026-09-10 records for all three models; slug from `GET /api/v1/models/google/veo-3.1/endpoints` and from the guide's own example. No provider call. | $0.00 | None    |
| 182.C.12.05 | Implement HeyGen Avatar IV as an explicitly avatar-scoped adapter supporting script plus passthrough `voice_id` …      | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-lane-adapters.ts` (`HEYGEN_AVATAR_FAMILY`), `human-video-lane-adapters.spec.ts`                                          | An avatar lane and nothing else, which is why SIX of the seven modes are refused - only `talking-head-avatar` is served. All nine passthrough names the record carries are mapped and no others, and the absences are asserted against the record rather than trusted: no durations (the length follows the script), no frame images, `seed: false`, `generate_audio: false`, so there is no duration, frame, seed or native-audio mapping. **The nuance worth stating is the one an adapter could easily get wrong.** This route can produce the EXACT WORDS - the script is the prompt - and it still cannot serve `exact-dialogue-performance`, because that mode promises the creator's approved AUDIO, the person's own recorded performance, and supplied audio is not exposed on OpenRouter for this model. Exact words in a synthesized voice is a different product from the person saying them, and a creator who asked for the second must not be handed the first; the refusal says so in those words. `voice_id` is a claim in exactly the sense Veo's `personGeneration` is - it asserts we may speak in a particular person's voice - so it is listed in `requiresConsentFor` and refused without an active voice-consent record, missing or inactive alike. Supplied audio is reported as `available_on_another_route`, because HeyGen's own API accepts it and OpenRouter's route does not - the distinction 182.C.13.03 exists to close. 45 tests in the file (7 new). This completes the five adapters of 182.C.12.01-05. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | All nine passthrough names, the null durations and frames, and `seed`/`generate_audio` false read from the committed 2026-09-10 record; slug from the endpoints API. No provider call.                                        | $0.00 | None    |
| 182.C.12.06 | Ensure every adapter returns an honest `consumedControls` and `ignoredOptionalControls` report …                       | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-lane-adapters.ts` (`LANE_ADAPTERS`, `findReportViolations`, `assertReportIsHonest`), `human-video-lane-adapters.spec.ts` | The report is the only place a creator learns that something they asked for did not reach the model, so a report that quietly loses a control is WORSE than no report at all - it reads as confirmation. Four things must hold and `findReportViolations` checks all four: every control asked for appears exactly once across the two lists; nothing appears in both; nothing is reported consumed that was never asked for; and no required control is in the ignored list, ever. The last is enforced by construction - `compileForModel` throws instead - and this is the assertion that it actually did. Verified as a PROPERTY rather than per adapter: the spec asks every one of the 8 models across all 5 families for all 26 controls at once and asserts the two lists together cover the request exactly, then walks every (model, control) pair asking for each control as REQUIRED and asserts it either lands in `consumed` or throws - never in `ignored`. A third sweep asserts no consumed passthrough is absent from that model's live record, which closes the loop back to 182.C.12.01's rule. Three negative tests prove the checker bites: a report with a control dropped, one with a control in both lists, and one claiming a control nobody asked for. `assertReportIsHonest` is cheap enough to run on every compilation rather than only in tests. `LANE_ADAPTERS` is the registry the property sweeps read, and a test asserts every model it can route to is in the live catalog. 52 tests in the file (7 new). Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Property sweep over 8 models x 26 controls against the committed 2026-09-10 records. No provider call.                                                                                                                        | $0.00 | None    |
| 182.C.12.07 | Add golden payload tests and rejection tests per adapter, including unknown live-schema changes and removed model …    | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-adapter-goldens.spec.ts` (new), `apps/oshun/bff/src/generation/__goldens__/human-video-adapters/goldens.json` (new)      | **The golden FILE is the deliverable, not the spec.** It holds the exact create payload each of the eight models produces for one canonical request - the same request for all of them - and its value is that changing a mapping shows up as a reviewable DIFF rather than as a passing test. The file says so, because a golden that is only ever regenerated proves nothing. Reading it is instructive on its own: wan-2.7 carries `audio`, `videos`, `negative_prompt`, `prompt_extend` and `ratio` under `atlas-cloud`; wan-2.6 carries `size` and `shot_type` instead; the Seedance models carry two or three names and nothing else; heygen carries all nine and has **no `duration`, no `seed` and no `generate_audio` at all**, which is the catalog's null durations and false flags showing through into a payload. Every golden is additionally run through 182.C.11.04's pre-submission guard and 182.C.12.06's honesty check, so the goldens are not merely stable, they are valid. Rejections cover what the item names. A live schema that STOPS naming a parameter drops the control to `parameter_not_in_catalog` when it was optional and fails the request when it was required. A removed model version refuses with `not in the current catalog` rather than routing to it. A golden validated against a record whose durations and resolutions have narrowed under it produces exactly `duration_unsupported` and `resolution_unsupported`. A consent-bearing control with no record refuses on EVERY adapter that declares one, swept rather than spot-checked. And a drifted table's unknown passthrough is caught by the payload guard. 19 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Goldens generated from the committed 2026-09-10 catalog and committed for review; the spec rebuilds and compares rather than regenerating. No provider call.                                                                  | $0.00 | None    |
| 182.C.12.08 | Keep adapter selection data-driven so successor versions can be canaried without changing the canonical creator …      | Oshun BFF | 182.C.11   | complete | `apps/oshun/bff/src/generation/human-video-lane-adapters.ts` (`selectWithCanary`, `inCanaryShare`, `withCanary`), `human-video-lane-adapters.spec.ts`               | A canary is DATA: adding one is an edit to the adapter table and nothing above it changes, because a creator's request names a mode and a set of controls and never a model - the contract is invisible to the trial by construction. `withCanary` returns a new adapter so a canary can be introduced from configuration rather than by editing a constant, which is what "data-driven" has to mean in practice. **Assignment is hashed from a stable key, never randomized, and that is the whole design.** A retry has to reach the SAME model as the take it is retrying; otherwise the second attempt is a comparison between two models rather than a second attempt, and neither the creator nor the evidence can say which change mattered. `Math.random()` there would quietly make every retry a different experiment. The split is checked empirically over 2000 keys - a 0.5 share lands 900-1100 and a 0.1 share lands 140-260 - rather than asserted. Two fall-through rules keep a trial from costing a take: a canary that cannot carry a required control, and a canary that has left the catalog, both fall back to the normal preference order with the reason recorded, because being in the share is not a reason to lose a take. And a canary is a different model with its own eligibility - 182.C.10.04 gates it on its own probes and its own benchmark promotion, so "same family" buys it nothing. **No canary is configured on any family**, and a test asserts that: the mechanism ships, the trial does not, because starting one is a routing decision somebody makes rather than a default. 60 tests in the file (8 new). This completes 182.C.12. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Deterministic hashing verified over 2000 synthetic keys. Pure computation; no provider call.                                                                                                                                  | $0.00 | None    |

### 182.C.13 Direct, Fallback, and Avatar Provider Lanes

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                             | Owner | Dependency         | Status   | Changed Files                                                                                                                                                                                                                            | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Provider/Hardware Evidence                                                                                                                                                                                                                                                                    | Cost                              | Blocker                                                                                                                                                                                                                                                                                                                                                 |
| ----------- | -------------------------------------------------------------------------------------------------------------------- | ----- | ------------------ | -------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.13.01 | Bring the direct fal/RunPod/Comfy video adapters and the existing `@isis/seedance-provider` fal, PiAPI …             | Isis  | 182.C.10           | complete | `apps/oshun/bff/src/generation/human-video-direct-lanes.ts` (new), `human-video-direct-lanes.spec.ts` (new)                                                                                                                              | Two things were measured before anything was written. **`@isis/seedance-provider` really does implement what OpenRouter cannot reach**: `video-extension.ts`, `video-inpainting.ts`, `object-removal.ts`, `recreation.ts`, `audio-reference.ts` and `reference-assets.ts` all exist, export the cited symbols, and are re-exported from the library index. And **none of them is reachable from the BFF**: `resolveVideoProviderGenerate` resolves exactly three things - a self-hosted ComfyUI Wan endpoint, the OpenRouter provider, and a fal LTX provider - chosen by comparing `OSHUN_VIDEO_PROVIDER` against two strings, with no branch that reaches a Seedance client. So there is no special-case executor behaviour to remove here; there is an ABSENCE to record, and recording it is what this module does. Every declaration carries an explicit `executor` status, a lane marked `declared-not-wired` is REFUSED by `decideDirectLane` before any capability is considered, and the refusal carries the reason - which turns 182.C.13.02's rule (never leave unreachable adapters presented as runtime capabilities) into a mechanism rather than a warning. A lane that declares exactly what you need and cannot be called is the most dangerous entry in a registry, because every other signal about it looks green. `toCatalogRecord` projects a declaration into the record shape the eligibility engine and the adapters already read, so ONE engine serves both lanes; everything a catalog would state and a declaration cannot is `null`, which the vocabulary already routes to a probe, and `canonical_slug` carries the declaration DATE so a lane declared in September and unrevisited in March is as visible as a model version that moved. **The citation test earned its place immediately**: the declarations name a module and an exported symbol per capability, the spec reads the other library's own source and index to check each one, and writing it caught an invented `buildSeedanceReferenceAssetPlan` where the real export is `prepareSeedanceReferenceAssets`. Verified by mutation. A declaration citing a function nobody wrote is the same lie as a capability nobody implemented, and it is the lie a registry is least able to catch on its own. 11 tests. Targeted tsc clean. | Module and symbol existence read from `libs/isis/seedance-provider/src/`; the unreachability read from `resolveVideoProviderGenerate` in `video-provider-env.ts`. No provider call.                                                                                                           | $0.00                             | Wiring an executor for the direct lanes is not done and is deliberately not claimed - the declarations are `declared-not-wired` and refuse. That wiring is what 182.C.13.03-13.08 build on.                                                                                                                                                             |
| 182.C.13.02 | (V1.2-gated per `release-scope.ts`) Wire the Veritas unified provider deliberately or retire it …                    | Isis  | 182.C.01.11 (V1.2) | partial  | `apps/oshun/bff/src/generation/human-video-avatar-boundary.spec.ts` (new)                                                                                                                                                                | Audited 2026-09-10, and the audit changes the question. **Veritas is deferred to V1.2 by `release-scope.ts`** - `libs/oshun/navigation/src/release-scope.ts` lists it among the deferred domains and the BFF middleware blocks `/veritas/...` and `/domains/veritas/...` - so its video providers are gated by an explicit, deliberate mechanism. "Wire deliberately or retire" already has a third answer the item did not enumerate: gated on purpose. And `providers/unified.ts` IS used inside its own app, by `video-pipeline`, `video-rendering` and `article-video`; retiring it would break Veritas rather than tidy it. What the audit DID find is narrower and is not this phase's to decide: `apps/veritas/video/src/providers/did.ts` is re-exported by the providers barrel and called by nothing, while a SECOND D-ID client exists at `libs/veritas/video-production/src/clients/d-id.ts` - two implementations, one of them dead, inside a release-gated domain. Which survives is a Veritas question. What IS this phase's to decide is that none of it can leak into human video, and that is now enforced rather than merely true: the registry offers exactly ONE avatar route (`heygen/avatar-iv` through OpenRouter, reachable, on a resolved slug), a test asserts the serialized registry contains no `tavus`, `d-id`, `did` or `veritas` string anywhere, and another asserts every `declared-not-wired` direct lane stays out of the routable set - the general form of the rule the item is asking for. 6 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | Deferral read from `release-scope.ts` and the BFF middleware; provider reachability read by following imports across `apps/veritas/video/src`. No provider call.                                                                                                                              | $0.00                             | LEFT UNCHECKED. The human-video side is enforced, but the item asks to wire or retire the Veritas provider and the live question there - whether `apps/veritas/video/src/providers/did.ts` or `libs/veritas/video-production/src/clients/d-id.ts` survives - belongs to the Veritas owner, inside a V1.2-gated domain this phase should not reach into. |
| 182.C.13.03 | Separate active-scene, avatar, video-edit, dubbing, and corrective-sync capability classes so routing cannot cross … | Isis  | 182.C.10           | complete | `apps/oshun/bff/src/generation/human-video-capability-classes.ts` (new), `human-video-capability-classes.spec.ts` (new), `human-video-lane-adapters.ts` (`serves`, four families' missing refusals), `human-video-lane-adapters.spec.ts` | Capabilities answer "can this model take a last frame". Classes answer a question no list of capabilities can: WHAT COMES OUT. An avatar lane and an active-scene lane match on every capability a request names - duration, resolution, references, seed - and produce products a creator would never confuse: one is a person performing in a place, the other a head against a background. So crossing a class is not a degraded result, it is the wrong product, and the creator finds out by watching. That is why it is a refusal and why the message names both products rather than saying no. **The consistency check found a real gap immediately.** `serves` and the hand-written `refusedModes` are written in different places and must agree, so `findClassInconsistencies` cross-checks them - and it reported that `wan`, `seedance-1.5` and `veo-3.1` never refused `video-edit` or `talking-head-avatar`, and `seedance-2.5` never refused `talking-head-avatar`, despite all four serving only `active-scene`. Four families would have silently accepted a request for a different product. Fixed in all four, with the class boundary now existing in the routing rather than only on paper. Two of the five classes have no mode yet: `dubbing` and `corrective-sync` operate on footage that already exists and `HUMAN_VIDEO_MODES` has no entry for either, so `modesForClass` returns empty for them and `classesWithoutModes` names them - a class with no mode is a visible gap while an undeclared class is an invisible one. One judgment call is stated rather than buried: `continuation-extension` is classed `active-scene`, not `video-edit`, because it consumes a take and produces NEW performance, and new performance needs the authorization a first take needed. The reverse direction is deliberately allowed - a lane may serve a class and still refuse one of its modes for a capability reason, which is what every "no audio input" refusal is - and a test asserts that is not reported as an inconsistency. 13 tests here; full BFF human-video suite 1113 green across 52 files. Targeted tsc clean.                                                                                                                                                                             | Pure computation over the adapter registry and the contract's mode list. No provider call.                                                                                                                                                                                                    | $0.00                             | None                                                                                                                                                                                                                                                                                                                                                    |
| 182.C.13.04 | Define provider credential/region health checks and deployment readiness independent of TypeScript module import …   | Isis  | 182.C.10           | complete | `apps/oshun/bff/src/generation/human-video-provider-readiness.ts` (new), `human-video-provider-readiness.spec.ts` (new), `evidence/2026-09-10-openrouter-video-transport/README.md` (credential checks)                                  | The failure this prevents is quiet and common: a deploy reports a provider available because the module imported and an environment variable was non-empty. Neither fact TOUCHES the provider - the key can be revoked, the budget exhausted, the account restricted to regions the tenant is not in, and every one of those deploys green. So signals are split by what they can PROVE, and `module-loaded` and `credential-present` are marked as proving nothing. A lane whose only passing signals are those two comes back **`unknown`**, never `ready`, and `readinessFromImportAndEnvOnly` exists so that honest answer has a name - anything reporting `ready` on that evidence is reporting something it does not know. Two live, UNBILLED checks make the real signal possible, measured 2026-09-10: `GET /api/v1/key` returns limit, `limit_remaining`, `is_free_tier`, `expires_at` and usage, and `GET /api/v1/credits` returns credits and usage; a revoked key answers 401 `"User not found."` while no key answers 401 with a DIFFERENT message, which is why the two are separate signals rather than one. `could not verify (HTTP 503)` is kept distinct from `the provider rejected the credential (401)`, because those must not read the same in a deploy log. The `label` field is a partial key and never enters a result - a test runs 182.C.11.10's leak guard over one built from the real body shape. Region reuses 182.C.10.04's rule rather than restating it loosely: an unrecorded region is `not-checked`, never `pass`, and a lane cannot reach `ready` on one - silence is not permission, and a readiness check is exactly where defaulting to yes would be tempting. `degraded` exists for the honest middle: the provider answers and we still cannot promise what it does. 16 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | `GET /api/v1/key` and `GET /api/v1/credits` probed live with a good key, a revoked key and no key, all unbilled; shapes and 401 messages recorded in the transport evidence README. The credits delta (20.968672686 → 21.018672686) is a third confirmation of the canary's $0.05.            | $0.00 (all three probes unbilled) | None                                                                                                                                                                                                                                                                                                                                                    |
| 182.C.13.05 | Add a self-hosted-provider port for future audio-conditioned video models without claiming parity before inference … | Isis  | 182.C.10           | complete | `apps/oshun/bff/src/generation/human-video-selfhosted-port.ts` (new), `human-video-selfhosted-port.spec.ts` (new)                                                                                                                        | The repo ALREADY self-hosts video - `ComfyVideoProvider` runs Wan on a RunPod ComfyUI endpoint and is the default when configured - which makes this the easiest item in the phase to file as done incorrectly. Grepped 2026-09-10, that provider takes a prompt and an optional image and has **no audio parameter anywhere**, so it is a self-hosted VIDEO port and not an implementation of the audio-conditioned one. The note saying so is re-checked against the file itself by a test rather than trusted. The port defined here is for the thing that does not exist yet: a self-hosted model that takes a creator's approved SPEECH and renders a performance to it - the one capability the entire exact-dialogue lane rests on, and today a rented `audio` passthrough on two Alibaba models through one vendor. The item's own qualification is the design: **no parity claim before inference and benchmarks exist.** So an implementation carries an evidence record and `parityStatus` walks four states - `no-implementation`, `no-inference`, `inference-only`, `benchmarked` - and only the last may be described as comparable to a hosted lane. `assertParityClaimAllowed` refuses the other three, most pointedly the middle one: "it runs on our hardware" is not "it is as good as the thing we were paying for", and that is exactly where the two get confused. A benchmark must name the lane it was compared against, because parity with nothing in particular is not a claim anyone can check. The unconfigured port is an honest fail-loud seam rather than a stub - every method throws with the reason, and the refusal itself explains that the exact-dialogue lane is rented - because returning a plausible empty result here is precisely what the phase's bright line forbids. Current state is data, not a comment: `adapterName: null`, so Studio and the release evidence read the same answer this file gives. 9 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                               | The no-audio claim about `ComfyVideoProvider` is asserted against its source in the spec. Nothing is implemented, so nothing was run; no provider call.                                                                                                                                       | $0.00                             | None                                                                                                                                                                                                                                                                                                                                                    |
| 182.C.13.06 | Record each provider's licensing, acceptable-use, commercial-use, biometric, public-figure, minors, retention …      | Isis  | 182.C.10           | complete | `apps/oshun/bff/src/generation/human-video-provider-constraints.ts` (new), `human-video-provider-constraints.spec.ts` (new), `evidence/2026-09-10-openrouter-video-transport/provider-constraint-sources.json` (new)                     | Eight constraints per provider - licensing, acceptable use, commercial use, biometric data, public figures, minors, retention, deletion - across all ten providers behind the video catalog, and **not one of them is filled in**. That is the honest state and the module is built so it stays visible rather than becoming a blank that reads as permission. Reading a vendor's terms and reducing them to "commercial use: allowed" is a LEGAL CONCLUSION; fetching a page and summarizing it would produce that conclusion with nothing behind it, and a creator would route a real person's likeness on the strength of it. So every constraint starts `unread`, `assertConstraintsPermit` refuses on `unread` exactly as on `forbidden` - silence is not permission - and the two stay distinct in the record because one can be resolved by someone spending an hour with a document and the other cannot. `unclear` is a fifth verdict for the case a human DID read and could not tell, which is a different and more alarming answer than not having looked. What is machine-readable IS filled in, from `GET /api/v1/providers` and the endpoints API: headquarters, datacenters, terms and privacy URLs, and which of the 28 models each provider serves - a test asserts the ten records account for all 28 between them. **Two findings from that pass matter on their own.** `atlas-cloud` publishes NO terms-of-service URL at all, and it is the provider serving `alibaba/wan-2.7` and `-2.6` - the entire exact-dialogue lane - so there is nothing to read yet rather than something unread. And datacenters cross jurisdictions the model id does not hint at: `alibaba` runs SG and CN, `minimax` is SG-headquartered with US datacenters, `seed` is BytePlus in SG, while AtlasCloud records none, so where a wan-2.7 render actually runs is not recorded anywhere. 12 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                          | `GET /api/v1/providers` (105 records) and `GET /api/v1/models/{id}/endpoints` read 2026-09-10; the ten-provider extract is committed as `provider-constraint-sources.json` and the spec asserts the module matches it field for field. No terms were read, and none are claimed to have been. | $0.00                             | Every constraint is `unread` by design. Filling them needs a person reading ten vendors' terms; the gate refuses until then, and `atlas-cloud` has no terms URL to read.                                                                                                                                                                                |
| 182.C.13.07 | Prohibit a provider fallback that weakens consent, retention, required inputs, exact-dialogue, resolution …          | Isis  | 182.C.10           | complete | `apps/oshun/bff/src/generation/human-video-fallback-safety.ts` (new), `human-video-fallback-safety.spec.ts` (new)                                                                                                                        | Falling back is a good idea that fails one specific way: the second lane is chosen because the first was unavailable, and nobody re-asks the questions the first lane had already answered. The take comes back at 720p instead of 1080p, or from a provider whose retention nobody has read, or with the words approximated instead of spoken - and it COMES BACK, which is what makes it hard to notice. Six dimensions are compared before a fallback is allowed and any one coming back weaker refuses. **The rule that matters as much: unknown refuses too.** A fallback whose safety cannot be compared is not a safe fallback, and treating "we could not tell" as "fine" is how every one of these failures actually happens - the same rule as `region_unrecorded` in 182.C.10.04 and `unread` in 182.C.13.06, in a third domain. Two comparisons are counter-intuitive and are handled explicitly. **Retention: SHORTER is stronger** - a fallback that keeps a creator's likeness for longer is a weakening even though the number goes up. And **required inputs are compared against the REQUEST, not the lanes**, because the same fallback is safe for one request and not for another: a lane that cannot carry supplied speech is fine for a request that never asked for it. Resolution ordering runs over the tiers the catalog actually publishes, with `768p` placed between 720p and 1080p because `minimax/hailuo-3-max` says so, and a tier not in the list compares as UNKNOWN rather than being guessed into position. `stronger` is permitted and reported rather than passed over in silence, because a fallback that improves something is still a change the release record should show - the creator asked for the first lane. 13 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Pure comparison over recorded guarantees; the resolution ordering is the catalog's own tier set. No provider call.                                                                                                                                                                            | $0.00                             | None                                                                                                                                                                                                                                                                                                                                                    |
| 182.C.13.08 | Add a provider kill switch that stops new submission while allowing safe ingestion/evaluation of already completed … | Isis  | 182.C.10           | complete | `apps/oshun/bff/src/generation/human-video-kill-switch.ts` (new), `human-video-kill-switch.spec.ts` (new)                                                                                                                                | The naive kill switch turns a provider off and every job against it dies with it. That punishes creators for an operational decision they had no part in, and it throws away money already spent - a job submitted five minutes before the switch may already have rendered, and the take may be perfect. So the switch is scoped to the ONE operation that starts new work. Submitting is blocked; polling, downloading, validating, evaluating and promoting all continue, because each only reads or measures something that already exists. If the output IS the danger - a safety incident rather than a cost or reliability one - 182.C.11.09's validation and the release gates are what stop the take, not this. **A retry counts as a submission**: it is a new job, it costs money, and it goes to the provider we just decided to stop using. The allow list is written as a whitelist of what a switch STOPS rather than a blacklist of what it permits, so adding an operation to the system defaults to continuing through a kill switch - the right default for a read-only operation, and a decision worth seeing, which is why `SUBMISSION_OPERATIONS` is exported. Scope is provider, lane or global, engagement is time-bounded in both directions (not yet engaged and already released both pass), and releasing is a DATED ACT rather than a deletion. `laneStopped` takes a stopped lane out of routing rather than letting it refuse at the last step, because a refusal that arrives after a creator has set up a shot arrives at the worst moment. The audit records `allowed-in-flight` as deliberately as `blocked-submission` - "why did this render after we turned it off" is exactly what someone asks later, and it deserves an answer already written down. Kept distinct from a quarantine on purpose: a quarantine is a CONSEQUENCE the catalog produced, a kill switch is a DECISION with a person's name on it, and the two should never be read from the same field. 13 tests; full BFF human-video suite 1176 green across 57 files. This completes 182.C.13 except 13.02, left open for the Veritas owner. Targeted tsc clean.                                                                                                                                                            | Pure computation over typed records. No provider call.                                                                                                                                                                                                                                        | $0.00                             | None                                                                                                                                                                                                                                                                                                                                                    |

### 182.C.14 Quality-, Capability-, Policy-, and Cost-Aware Routing

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner     | Dependency  | Status   | Changed Files                                                                                                                                                                                            | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Provider/Hardware Evidence                                                                                                                                                                                                                        | Cost  | Blocker |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | --------- | ----------- | -------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----- | ------- |
| 182.C.14.01 | Filter first by hard capabilities, consent/policy, geography, retention, tenant allowance, and live health.           | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-route-filter.ts` (new), `human-video-route-filter.spec.ts` (new)                                                                                              | Ranking is a comparison between routes that are all ACCEPTABLE. The moment a score can rescue a route that failed a hard constraint, the constraint is not hard - and the failure is invisible, because a high score is exactly what a good outcome looks like. So this runs first, returns a SET rather than an order, and nothing in it reads a quality number; the returned candidate carries four identity fields and no score to rank on. Six gates cover the six dimensions the item names, and `DIMENSION_OWNERS` says which gate owns which - because the honest failure mode of a composed filter is a dimension everybody assumed somebody else checked, and a test walks the mapping. The ORDER is about the message rather than about speed, because a creator or an operator reads the FIRST refusal: an operator's stop comes first (nothing else about the route matters and it is not the route's fault), then the wrong PRODUCT (no capability, score or budget changes that), then the tenant (a decision about the account), then **terms not established - deliberately ahead of capabilities, because whether it CAN is a less important question than whether it MAY** - then eligibility, then readiness. `degraded` refuses, since degraded is not ready. Where every candidate fails, `dominantGate` reports the most common reason with ties broken toward the more fundamental gate, so the answer is the same whichever order the candidates were checked in - asserted by running the same pair reversed. Gate wording is creator-facing and names no provider or model, asserted against a vendor list. 12 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Composition over the existing gates; pure computation. No provider call.                                                                                                                                                                          | $0.00 | None    |
| 182.C.14.02 | Rank eligible routes by benchmarked lip sync, exact speech, identity, motion, visual quality, locale …                | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-route-ranking.ts` (new), `human-video-route-ranking.spec.ts` (new), `provider-quality-router.ts` (`ProviderRouteKind` gains a `human-video:${string}` branch) | The existing router already does the hard part well - Beta-Thompson over each candidate's measured release rate, favouring the proven route without starving a newcomer, deterministic given an injected RNG - and it is keyed by an opaque string on BOTH axes. So the extension is one added branch on `ProviderRouteKind`: the cohort travels in the kind (`human-video:<cohort>`) and the model id travels as the provider id. Nothing about the four original media kinds changes, and a test asserts that. Why per-cohort at all: a model's release rate on a single-speaker close-up says very little about its rate on a two-person wide shot, and one shared number would average those into something true of NEITHER - a test drives the same model to 100% in one cohort and 0% in another to show the separation is real. **The rule that keeps the ten signals honest: a signal with no measurement contributes nothing and is reported as absent.** Not a default, not a midpoint, not "assume average" - those all put a number where there is no evidence, and it then competes with real ones. The weighted mean re-normalizes over the AVAILABLE weights, so a route measured on sync and identity is scored on sync and identity: neither penalised for the eight nobody measured nor credited for them. A test asserts a two-signal and a one-signal route with the same measured quality score the SAME, with the difference reported in `availableSignals` and `totalSamples` rather than folded into the number - which is what 182.C.14.04 will govern. Expected cost is DERIVED rather than declared: measured cost per attempt over measured release rate gives cost per released take (20 cents an attempt at one-in-four is 80 cents), and it is `null` for a route never tried or never released, because "never released" is unmeasured rather than infinitely expensive. Latency and cost normalize against the other routes in the cohort; a lone route scores 1 on them, which is harmless because a single-candidate cohort is decided by the filter rather than the ranking. Default weights put sync and exact speech above cost and latency - a beautiful shot of the wrong words is a failed take - and sum to 1. 15 tests. Targeted tsc clean; the router's own 5 tests still pass.                                                   | Deterministic seeded RNG for the Thompson draws; measured statistics fed through the real router. No provider call.                                                                                                                               | $0.00 | None    |
| 182.C.14.03 | Maintain separate rankings for active single-speaker, active multi-speaker, profile/occlusion, high-motion …          | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-cohorts.ts` (new), `human-video-cohorts.spec.ts` (new)                                                                                                        | The item names nine things to keep separate and the obvious reading - cross all nine - is the one that fails: six shot kinds x four locales x three resolution bands x three duration bands is **216 cells**, and a measurement corpus that would take a year to fill has roughly zero samples in each. A router that ranks on empty cells ranks on noise. So they are not all the same kind of thing. The SHOT COHORT is what the camera is looking at and is mutually exclusive; locale, resolution band and duration band are FACETS that modify it. Statistics are read from the most specific cell that has evidence and the answer says WHICH LEVEL it came from, so a number borrowed from the shot-only cohort is never presented as a locale-specific measurement. **The ordering of the fallback chain is a claim, and it is stated rather than assumed**: duration drops first (a model that holds identity for eight seconds usually holds it for twelve), resolution next, and **locale drops LAST** - a lane that renders Twi mouth shapes badly renders them badly at every resolution and every length, and pooling that away is how a locale failure becomes invisible. A test proves the ordering bites by giving the shot-only cell thirty times the evidence and asserting the locale cell is still used. Two more judgments are encoded and tested: an edit is an edit however it is framed, and a HIGH-MOTION two-hander is filed under motion rather than multi-speaker, because motion is what breaks identity and pooling a chase with a seated conversation would hide the failure a creator would notice. Bands sit where the catalog's own ranges end rather than on round numbers - `sd`/`hd`/`uhd` over its six published tiers, and `long` exists because `seedance-2.5` and `wan-3.0` reach 30 s. A descriptor with no locale collapses two chain levels onto one key and the duplicate is dropped, so a caller cannot read the same cell twice and think it agreed with itself. `resolveCohort` returns the broadest cell WITH ITS REAL SAMPLE COUNT when nothing qualifies, reported as thin rather than dressed up - deciding what to do about that is 182.C.14.04's job. 15 tests. Targeted tsc clean.                                                                                                                                    | Bands derived from the committed 2026-09-10 catalog's tier and duration ranges. Pure computation.                                                                                                                                                 | $0.00 | None    |
| 182.C.14.04 | Treat confidence intervals and sample size as routing data; prevent a tiny benchmark from displacing a proven model.  | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-route-confidence.ts` (new), `human-video-route-confidence.spec.ts` (new)                                                                                      | The failure this prevents has a shape you can write down: a new model is tried twice, releases twice, measures 100%, and beats a 500-trial incumbent at 85% on every point-estimate ranking - on two observations, one of which might have been a shot with no face in it. So a rate is never used as a NUMBER, it is used as an INTERVAL, and the comparison is between lower bounds. Wilson was chosen over the normal approximation precisely because it does not collapse at the extremes - the normal approximation gives 2-of-2 the interval [1,1], which IS the bug rather than a description of it. Computed: **2/2 measures 1.000 and its lower bound is 0.342**, against the incumbent's 0.816. **The example that shows the real boundary is better than the obvious one, and writing the test found it.** My first attempt assumed 30-of-30 would be refused; it is not - 30 consecutive releases put the lower bound at 0.887, genuinely above 0.816, and the gate correctly allows it. The case that matters is 26-of-30: it MEASURES 0.867, above the incumbent's 0.850, and its interval reaches down to 0.703. Measured better, known worse, refused. That is the whole item in one comparison, and it is a statistical test rather than a rule against newcomers. Two separate gates because they answer different questions: `hasEnoughData` asks whether a route may be RANKED (10 trials), `canDisplace` asks whether it may take the top spot from something proven (30 trials, and only against an incumbent with 100+). When no incumbent is proven the rule relaxes to a plain lower-bound comparison, because refusing both would leave the router unable to prefer anything. **Exploration is not sacrificed**: the Beta-Thompson draw still routes traffic to a newcomer, which is how it accumulates the samples this gate asks for - this governs the RANKING, not the exploration, and a system that refuses to try anything new never gets data on anything new. An unmeasured route is ordered last with `rankable: false` rather than dropped, because dropping it would make it look identical to a bad one, and `heldBelowIncumbent` flags the route that measured higher so nobody has to rediscover why the "better" model is second. 15 tests. Targeted tsc clean.                                                                     | Wilson bounds computed independently before the tests were written: 2/2 → 0.3424, 26/30 → 0.7032, 30/30 → 0.8865, 425/500 → 0.8160, 195/200 → 0.9428. Pure computation.                                                                           | $0.00 | None    |
| 182.C.14.05 | Support draft/standard/production/hero policies with different candidate counts, evaluators, providers, budgets …     | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-tier-routing.ts` (new), `human-video-tier-routing.spec.ts` (new)                                                                                              | `TIER_POLICIES` in the contracts already answers FIVE of the seven things this item names - candidates per shot, retry budget, allowed fallbacks, human review, spend and latency budgets - and it stays the authority for them. This adds the two it does not carry, EVALUATORS and PROVIDERS, and composes the rest at call time rather than copying it: a second table with the same numbers would drift, and the one that drifted would be the one nobody was reading. A test walks all four tiers asserting every shared field equals the contract's, so a change there appears here without anyone remembering to mirror it. Evaluators escalate and never shrink, and draft runs NONE - it is `releaseEligible: false` in the contract, so the full evaluator set would spend GPU time answering a question nobody asked. Locale pronunciation is hero-only. Two routing rules are new and stated plainly. **A hero take is not where you trial a model**: 182.C.12.08's canary is allowed at draft, standard and production and REFUSED at hero, because the tier that costs the most and gets the most human attention is the worst place to learn something about a successor version. And **draft may route to a `degraded` lane** - it exists to preview and cannot ship - while every other tier requires `ready`, because a take that might ship must come from a lane somebody measured; `unknown` and `not-ready` route at NO tier, asserted across all four. Benchmark promotion must REACH the tier rather than merely exist: promoted-to-standard does not serve production. Refusal order puts the capability class first, so the most fundamental reason is the reported one. 12 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Composition over the contract's own tier table; pure computation. No provider call.                                                                                                                                                               | $0.00 | None    |
| 182.C.14.06 | Estimate end-to-end cost before admission, including TTS, multiple candidates, provider retries, downloads …          | Oshun BFF | 182.C.01.09 | complete | `apps/oshun/bff/src/generation/human-video-cost-estimate.ts` (new), `human-video-cost-estimate.spec.ts` (new)                                                                                            | `docs/agents/model-cost-openrouter.md` already publishes the normalizer this item points at, written 2026-09-09, and 182.C.10.05 implements it. Reading the two against each other is reassuring rather than redundant: the document's unit rule ("a key beginning `cents_` is denominated in US cents; every other key is denominated in US dollars"), its six SKU families, its warning that families A and D order their qualifiers DIFFERENTLY so a fixed-order key builder silently misses half the catalog, and its conclusion that token-priced models cannot be estimated pre-flight - all of it matches what the parser independently does. So the generation leg is called rather than re-derived. What this adds is the other eight legs, because a render is the visible cost and rarely the whole one. **The estimate is a FLOOR and says so.** Almost none of those rates has been measured on this system, and a total assembled from guesses is a number somebody budgets against - so a component with no measured rate contributes NOTHING, is named in `unknownComponents`, and the result is `floorUsd` rather than `totalUsd` with `complete: false`. A floor is honest and useful (if it already exceeds the ceiling the request cannot be admitted whatever the unknowns turn out to be) and cannot be mistaken for a forecast. The generation leg multiplies through candidates AND expected attempts, so six candidates at a 50% release rate is priced at twelve renders - the number that actually leaves the account - and when the release rate is unmeasured the multiplier is unknown and the leg becomes unknown with it, rather than silently assuming one attempt. Retries are reported as a component and contribute zero, because they are inside that multiplier and adding them again would double-count. The published rule for token-priced routes is ENFORCED rather than restated: a ceiling that needs a pre-flight number refuses a Seedance route, while an after-the-fact ceiling admits it with the unknowns recorded. `floor-over-ceiling` is checked before `estimate-incomplete`, so the refusal that is safe to make on incomplete information wins. The seven unmeasured rates ship as data with every `usdPerUnit` null, so the gap is countable and filling one in is an edit to a table. 13 tests. Targeted tsc clean. | Generation leg priced from the committed 2026-09-10 catalog through 182.C.10.05; veo-3.1 at $0.40/s x 8 s x 2.5 renders checked against the normalizer. Cross-read against `docs/agents/model-cost-openrouter.md` (2026-09-09). No provider call. | $0.00 | None    |
| 182.C.14.07 | Re-route on capability regression, quality drift, outage, throttling, or budget pressure while preserving immutable … | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-reroute.ts` (new), `human-video-reroute.spec.ts` (new)                                                                                                        | Five things make a route stop being the right one - the catalog took a capability away, quality drifted, the provider went down, it started throttling, the budget ran short - and none of them is a reason to deliver something DIFFERENT. That is the whole difficulty: a re-route happens mid-flight, usually automatically, usually while somebody is waiting, which is exactly the moment where "1080p was not available so we used 720p" gets decided by a timeout rather than by a person. So the intent is frozen at admission, digested, and re-checked on the far side of every move; a proposal whose digest differs is not a re-route but a different request wearing the same id, and the refusal NAMES the field that moved. The digest sorts its arrays and fixes its field order, so an intent assembled by a different code path still compares equal - a re-route must not fail because someone rebuilt the same request. The intent check runs BEFORE any lane is compared, because a good lane does not rescue a proposal that already changed the request. **Budget pressure never buys a downgrade.** It is the trigger most likely to be handled by quietly dropping to a cheaper, worse lane, because that feels like the helpful thing to do - and it is the one case where the creator would certainly have wanted to be asked. A cheaper lane that is not weaker is a fine answer; when none exists this refuses with `budget-pressure-would-downgrade` and says it is a budget decision for a person rather than a routing one. It also refuses to move for budget reasons when either cost is UNKNOWN, because that is a guess with a budget-shaped excuse - a test shows the same alternative being accepted for an outage and refused for budget pressure. The safety comparison itself is 182.C.13.07's, unchanged: weaker refuses and uncomparable refuses. 16 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                    | Pure computation over recorded guarantees and a hashed intent. No provider call.                                                                                                                                                                  | $0.00 | None    |
| 182.C.14.08 | Return a machine-readable explanation of selected route, rejected routes, tradeoffs …                                 | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-route-explanation.ts` (new), `human-video-route-explanation.spec.ts` (new)                                                                                    | Everything upstream already decided honestly; this is where those decisions become ONE artifact somebody can read six months later, when the question is "why did this shot come from that model" and nobody involved is still in the room. Four things go in and **the third is the one that usually gets left out**. Selected and rejected routes are obvious. TRADEOFFS are the choices made silently and correctly - statistics borrowed from a broader cohort, a fallback that was equal but not identical, a score standing on two of ten signals, a canary that served the take, a cost that is only a floor. None is a failure and every one is something a reader would want before trusting the ranking, so they are DERIVED from what the pipeline already recorded rather than passed in - a caller cannot forget to mention one. The fourth is the optional controls that did not reach the model, carried through from 182.C.12.06's report untouched, because a creator asked for those and the report is the only place that fact exists. Two rules about contents: model ids and dated VERSIONS are carried (provenance is useless without them) and credentials, signed URLs and scripts are not, with 182.C.11.10's leak guard run over the artifact in the tests; and `toCreatorSummary` names no vendor, model or lane at all - asserted against a term list - because a creator's question is "why did I get this" and the answer to that is never "AtlasCloud". The whole artifact round-trips through JSON unchanged. When nothing could be routed, the dominant gate's creator-facing wording becomes the headline, and the distinct reasons become the notes. 18 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Pure composition over the upstream decisions; JSON round-trip and leak-guard asserted. No provider call.                                                                                                                                          | $0.00 | None    |
| 182.C.14.09 | Add shadow evaluation and canary weights for new models; never promote straight from catalog discovery to customer …  | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-promotion-ladder.ts` (new), `human-video-promotion-ladder.spec.ts` (new)                                                                                      | A model appearing in the catalog is a fact about OpenRouter's inventory and nothing else - nobody has sent it a request, nobody has watched its output, and it has never rendered a face this system holds an authorization for. The failure this prevents is the short path from that fact to "the default route for customer work", which is a single configuration change away at any moment and looks entirely reasonable while you are making it. Five rungs, none skippable: `discovered` → `probed` → `shadow` → `canary` → `default`, and every skip is refused, not only the long one. **Shadow is the rung people skip, and it costs nothing to skip until it does.** Its whole point is that a model can be measured on real requests without a creator ever receiving its output, so a model that turns out to blur faces at 1080p is discovered by US rather than by them. A shadow run whose output was served is not a shadow run - `shadowOutputServed` makes that a recorded FACT rather than a convention, because the convention is exactly what erodes when a deadline arrives, and advancing on evidence carrying it is refused with its own reason. Each rung's entry evidence is the previous rung's exhaust: nothing here asks anyone to write a justification, it asks whether the runs happened - a passing probe before shadow, 20 shadow runs before serving anyone, 50 canary takes before becoming the default. The last rung additionally applies 182.C.14.04's displacement test, because becoming the default means displacing something: a canary at 57-of-60 is refused against an incumbent at 480-of-500 and allowed against one at 350-of-500, on interval lower bounds rather than rates. `SHADOW_RUN_POLICY` is stated as data - renders real requests, output never reaches a creator, counts against budget, is measured - because it is the definition most likely to be quietly redefined. 13 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                          | Pure computation; the displacement test reuses 182.C.14.04's Wilson bounds. No provider call.                                                                                                                                                     | $0.00 | None    |
| 182.C.14.10 | Test that the router never picks Avatar IV for an active- scene request or a prompt-only model for exact-dialogue …   | Oshun BFF | 182.C.12    | complete | `apps/oshun/bff/src/generation/human-video-routing-invariants.spec.ts` (new)                                                                                                                             | Both mistakes are named in the item because both are PLAUSIBLE: Avatar IV matches a great many capabilities an active-scene request asks for, and a prompt-only model matches every single one an exact-dialogue request asks for except the one that matters. A router ranking on capability overlap would pick either cheerfully. So each is asserted at FIVE independent layers rather than at the one that happens to catch it today - a single guard is one refactor away from being bypassed, a property holding at five layers is not. For Avatar IV: the class gate, the adapter's named mode refusals, the composed filter, the tier gate, and the catalog record itself (null frame images, null durations, `seed: false` - it is not a shot, it is a script read aloud). For prompt-only models: the family's mode refusals, the adapter refusing `supplied-speech` as a required control, eligibility refusing to grant `accepts-supplied-audio`, **a stray probe record claiming it passed being refused anyway** because advertisement is checked before probes and a probe cannot manufacture a capability the model does not have, and finally that only the wan family can serve exact dialogue at all - which is then cross-checked against the catalog, since wan is also the only family whose models name an `audio` passthrough. Two closing sweeps state the invariants generally rather than by example: no adapter accepts a mode outside the classes it serves (either it serves the class or it refuses the mode, never neither), and no family without an audio passthrough serves an exact-dialogue mode. 14 tests; full BFF human-video suite 1319 green across 67 files. This completes 182.C.14. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Every layer asserted against the real registry and the committed 2026-09-10 catalog. No provider call.                                                                                                                                            | $0.00 | None    |

### 182.C.15 Durable Job Orchestration

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner     | Dependency | Status   | Changed Files                                                                                                                                                                                                                                                                                 | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Provider/Hardware Evidence                                                                                                                                                                                | Cost  | Blocker |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----- | ------- |
| 182.C.15.01 | Extend the generation job envelope with request version, project, shot, candidate, attempt, provider correlation …     | Oshun BFF | 182.C.05   | complete | `libs/contracts/src/human-video/job-envelope.ts` (new), `job-envelope.spec.ts` (new), `index.ts`, `schema-registry.ts`, `schemas/human-video/` (3 new published artifacts)                                                                                                                    | Everything a job needs to be picked up by a process that did not start it. Two fields carry most of the weight. **`policySnapshot` is frozen at admission**: a job admitted under a production policy stays admitted under that policy even if the tier table changes an hour later, because the alternative is a job whose rules moved while it was running - and the creator agreed to the first set. It is a copy rather than a reference for exactly that reason, and carries only the fields that change what the job DOES. **`idempotencyKey` is DERIVED, not chosen**, and the ATTEMPT NUMBER is part of the identity it is derived from - which is the subtle part. A retry is not a duplicate of the thing it is retrying, so a key omitting the attempt would silently swallow every retry as a redelivery; a test asserts attempt 1 and attempt 2 differ. The schema itself REFUSES an envelope whose key it did not derive, so a hand-set key cannot collide with another job's or drift from its own attempt - `superRefine` recomputes and compares. The key also carries the identity in plain text beside the hash so a human reading a queue can see which job a row belongs to. Two more invariants live in the schema rather than in a caller: a candidate index outside the admitted `candidatesPerShot` is rejected, and an attempt beyond `retryBudget + 1` is rejected - the frozen snapshot bounds the job it belongs to. Asset digests are content addresses rather than handles, because a handle can be re-pointed and "the audio we approved" has to mean the bytes somebody listened to. Provider correlation starts entirely null, which is what the real 2026-09-10 exchange returned. `nextAttempt` drops the previous job id and output (they belong to the attempt that failed), keeps the approved audio and the frozen policy, and refuses to exceed the budget. FNV-1a rather than a crypto digest because this package stays node-free; the key is a deduplication token, not a security boundary. 17 tests; whole human-video contract suite 540 green. Published as three new schema artifacts.                                                                                                                                                                                                                                                                                                                                                     | Pure contract; published JSON Schema regenerated and the drift check passes. No provider call.                                                                                                            | $0.00 | None    |
| 182.C.15.02 | Implement a durable workflow/saga for validation, consent, audio, planning, submission, polling/callback, download …   | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-saga.ts` (new), `human-video-saga.spec.ts` (new)                                                                                                                                                                                                   | `human-video-workflow.ts` already runs the stages forward with concurrency, priority and a fan-out. What a SAGA adds is the other direction - what happens to the work already done when a later step fails - and the usual saga assumption is that every step has an inverse. **Here three do not, and pretending otherwise is the failure this exists to prevent.** Submission spends money the moment the provider accepts it, and 182.C.11.07 probed for a cancellation endpoint and found none, so a saga claiming to roll back a submission is describing something it cannot do; it can stop consuming the result and record the spend. Polling can be stopped and the RENDER cannot - it is happening on the provider's hardware and by the time we know we do not want it, it exists. And release is public: unpublishing is a new act with its own record, not an undo. So `compensationPlan` returns TWO lists rather than one - the compensations it can actually run, newest first because a compensation usually depends on the state the later steps left, and the `unrecoverable` entries naming what was spent or done that no compensation reaches, with the total. An operator reading a failed job sees both, and the second list is the one that costs something. Resumption reads the DURABLE LOG rather than memory, because a process that restarted has none - which is the whole reason the log exists - and it resumes from the first MISSING step rather than the highest completed one, so a gap (something ran that should not have, or a write was lost) is not papered over; `outOfOrderSteps` names the anomaly instead. Two steps survive a failed shot on purpose, approved audio and the consent check: one costs money and the other costs a person's time, which is what 182.C.15.09 builds on. 21 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Pure computation over a step log; the cancellation claim rests on the 2026-09-10 probe. No provider call.                                                                                                 | $0.00 | None    |
| 182.C.15.03 | Guarantee idempotent provider submission and callback handling across process restart, queue redelivery, timeout …     | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-idempotency.ts` (new), `human-video-idempotency.spec.ts` (new), `docs/.../evidence/2026-09-10-canary-exchanges/idempotency-duplicate-create.json` (new)                                                                                            | The task says GUARANTEE idempotent provider submission, so the first question was whether the provider offers one. It does not: two byte-identical creates carrying the same `Idempotency-Key` header, sent seconds apart, returned two job ids, rendered twice and billed twice ($0.10, credits delta confirmed). Three doors closed with it - `GET /api/v1/videos` is a plain-text 404 so you cannot list your own jobs; `/generation?id=` and `/credits` both still reported nothing for a generation that had already returned a cost, so neither can answer 'did my submission land' while the question is live; and a `user` field in the create body does not reach the generation record's `external_user`. So the defence is local and has to be written before the call. TWO durable writes, not one: a `claimed` row (we intend to submit) and then an `in-flight` row (bytes are on the wire). That second write is the whole design - it is what separates the two crash windows. A worker that died on `claimed` provably sent nothing and another worker may retake it; a worker that died on `in-flight` may have spent money, and no lease expiry makes that safe, so the row becomes `orphaned`, which is a STATE and not an error. `claimSubmission` is a compare-and-set insert, which is the single-flight answer to a redelivered queue message and to two workers racing - the loser gets `duplicate` and the row, because a redelivery is the queue doing its job. Ambiguity is the default and only two things settle it: a connection that never opened, or the API's own validator - measured, since four invalid creates returned 400 zod envelopes having created nothing. 429 is deliberately NOT on that list. The one channel that returns a value we chose is `callback_url`, honoured per request, so the claim key goes in its path and an orphan can announce itself: `adoptOrphan` reattaches the job id a delivery names, and a SECOND job id for one attempt is a conflict left for a person rather than a field to overwrite. Callbacks apply at most once and never contradict a settled job. `recoverAfterRestart` consults no memory, because a restarted process has none. 34 tests. Targeted tsc clean.                                                                                                                                                                                                                                        | `idempotency-duplicate-create.json` plus the canary README: two job ids, `usage.cost` 0.05 each, `total_usage` 21.018672686 -> 21.118672686. Route probes and the four unbilled 400s in the same record.  | $0.10 | None    |
| 182.C.15.04 | Persist state transitions transactionally with an outbox so Studio and admin cannot observe impossible job sequences.  | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-transition-outbox.ts` (new), `human-video-transition-outbox.spec.ts` (new)                                                                                                                                                                         | The failure being prevented is a dual write: the stage lives in one store and the event telling Studio about it goes somewhere else, and a process dying between them loses whichever one it did not do yet. State first loses the event and Studio sits on a stale stage forever; event first publishes a stage the job was NEVER IN, which is not a delay but a UI stating something false about somebody's render. So the event is written into the same transaction as the state row, as a row, and a relay publishes it after - the pattern `libs/yemaya/study-workspace` already runs (`inTransaction` + `enqueueDomainEvent`, idempotent on the event id), followed deliberately rather than reinvented, and not imported because that port is bound to Yemaya's aggregates and ADR-0006 keeps isis generation state in the BFF. Three things have to hold for 'cannot observe an impossible sequence' to mean anything, and each is a separate mechanism. (1) An illegal transition writes NOTHING: `canTransition` is checked before anything is staged, so there is no window in which a bad stage exists. (2) The relay never overtakes itself on one job - `claimOutboxBatch` releases only the lowest undispatched sequence per job and holds the next while one is leased, so at-least-once delivery cannot become out-of-order delivery; different jobs still go in parallel. A quarantined entry stops its own job's whole stream rather than letting later events jump it, and `blockedJobs` names those for a person, because the state changes DID happen and nobody is being told. (3) A consumer that misses one still cannot go backwards: the per-job sequence is a cursor, an older event is ignored, and a forward jump is applied while the skipped sequences are RECORDED - the projection knows the current stage and knows it cannot narrate how the job got there. Status is computed once at write time from stage and gate together, so the admin's six statuses and Studio's stages are two readings of one number. `findSequenceViolations` is the audit that reads a delivered stream the way an auditor would. 24 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                            | Pure computation over an in-memory store whose transaction stages writes and applies them only on a clean return; the rollback test abandons a transaction mid-way and asserts both stores are untouched. | $0.00 | None    |
| 182.C.15.05 | Add cancellation semantics for queued, submitted, downloading, evaluating, retrying, and review states …               | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-cancellation.ts` (new), `human-video-cancellation.spec.ts` (new)                                                                                                                                                                                   | The task asks for cancellation semantics 'including what upstream spend cannot be recovered', and on this API that qualifier is most of the subject. 182.C.11.07 probed for a cancellation endpoint and found none, so once a job is submitted cancelling has exactly ONE honest meaning: we stop, the provider does not. The render continues on their hardware, finishes, and is billed. A cancel button implying otherwise lies to whoever pressed it, which is why every decision here carries the money and separates a settled charge from a quote. What changes by stage is not whether we can stop - we always can - but what stopping saves: before submission everything is saved and nothing lost; after submission the render is gone whatever we do; `retrying` is the one mid-pipeline stage where cancelling still saves real money, because the attempt about to be sent is not sent; at review the only thing cancellation prevents is release. `preparing-audio` is named separately from 'queued' because a speech provider may already have been paid even though no video exists. A model priced per token cannot be quoted before it runs (182.C.10.05), so those jobs go in `unknownSpend` and are NEVER folded into a total that would look authoritative and be invented. Output disposition turns on the reason and not on the money: a cancelled job's file is normally the evidence behind its charge and is kept, but a withdrawn consent means we no longer have permission to hold a depiction of that person, and permission is not a billing question. Finally, because the render was never stopped its webhook still arrives - `handleLateTerminal` records the charge, discards the output, and leaves the stage `cancelled`, so a cancelled shot cannot come back to life through its own callback. 16 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Rests on the 2026-09-10 cancellation probe recorded in the transport evidence README; no new provider call.                                                                                               | $0.00 | None    |
| 182.C.15.06 | Add per-tenant/provider/model concurrency, rate, quota, spend, and GPU controls with fair scheduling and priority …    | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-admission-controls.ts` (new), `human-video-admission-controls.spec.ts` (new)                                                                                                                                                                       | The task says EXTEND the two existing gates, and reading them first is what made the extension small. `generation-budget-gate.ts` already holds a durable per-day USD ceiling with atomic idempotent reservations and `generation-quota-gate.ts` does the same for per-user volume; both are off unless configured and both fail closed. What neither has is a DIMENSION - the budget window key is `day:<date>` for the whole deployment - so a $6 hero render on one model cannot be bounded differently from a hundred drafts on another. A dimension is just a different window key over the same atomic primitive, so nothing here reimplements accounting: `windowKeysFor` builds the keys and `reserveSpendAcrossScopes` reserves against several at once. Multi-scope reservation is a small saga, since the scopes are independent atomic operations: take them narrowest-first (so a tenant at its own limit never touches the global window) and hand back what was already taken when a later scope refuses - a partial reservation would quietly shrink a tenant's ceiling every time a provider window was full. The release needs NO new store method: settling to zero moves the window back by exactly the estimate, which is what the existing settle path already does for a job cheaper than its quote. Three controls are added locally because they are not per-day accumulators: concurrency is a live count with leases (a dead worker must not hold a tenant's only slot forever, and re-acquiring under the same holder refreshes rather than double-counting); rate is a sliding window, which is what a provider's own 429 counts; and GPU is DECLARED AND NOT WIRED - 182.C.13.05 records the self-hosted port has no adapter, so a GPU lane is refused with that reason rather than queued against capacity that does not exist. Fair scheduling is the other half, because priority alone starves: deficit round-robin across tenants within a class, charged in DOLLARS so one 4K hero shot does not cost the same share as one draft, and one class promotion per ageing interval with NO ceiling - a bounded ladder still leaves the bottom class able to starve forever. Admission runs cheapest-first and undoes everything it took on refusal. 28 tests, several of which run against the real `InMemoryGenerationBudgetStore` rather than a double. Targeted tsc clean.                                                                               | Uses the existing budget store implementation directly, so the reserve/settle arithmetic is the shipped one; refusal tests read the spend back out of it to prove nothing was left held.                  | $0.00 | None    |
| 182.C.15.07 | Classify validation, consent, capability, provider, transport, corrupt-output, quality, safety, budget, cancellation … | Oshun BFF | 182.C.05   | complete | `libs/contracts/src/human-video/failure-codes.ts` (new), `failure-codes.spec.ts` (new), `schema-registry.ts` + `index.ts` (modified), 2 new published artifacts (47 total), `apps/oshun/bff/src/generation/human-video-failure-mapping.ts` (new), `human-video-failure-mapping.spec.ts` (new) | A failure code is part of the CONTRACT - a client branches on it, an operator searches for it, a runbook names it - so the vocabulary lives in the published library where the drift check makes renaming one a visible breaking change, and the BFF holds only the mapping from the error surfaces it actually raises. 57 codes across the eleven families the task names, each `hv.<family>.<specific>`, each carrying a retry disposition, who has to act, an HTTP status, whether the job is over, and ONE sentence a creator may read. Two rules are enforced by test rather than by convention: the `internal` family never blames the request (that is what keeps a bug from being reported to a creator as their mistake, and all four internal codes share one sentence so a bug leaks nothing), and no creator message anywhere names a vendor - ADR-0006 section 6 stops provider vocabulary at the edge, and 'atlas-cloud refused' teaches a creator nothing while leaking a routing decision. `retry: 'never'` on `hv.transport.submission_ambiguous` is the 182.C.15.03 measurement written into the taxonomy: a repeated create bills a second render. Each BFF mapping is a RECORD keyed by the source union rather than a switch with a default, so a new refusal code in 11.08 or a new rejection in 11.09 fails the typecheck until somebody decides what it means. One judgement is stated rather than buried: a provider 400 maps to `provider.rejected_request` because the provider did refuse, but `assertSubmittable` validated that payload first, so `isOurPayloadBug` names it as OUR compiler disagreeing with their validator and not a creator's mistake. 13 + 16 tests. Both sides typechecked; schema publication drift check green.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | 47 published artifacts regenerated from `libs/contracts/`; `schema-publication.spec.ts` green (9 tests).                                                                                                  | $0.00 | None    |
| 182.C.15.08 | Add bounded retries with jitter, attempt lineage, circuit breakers, poison-job quarantine …                            | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-retry-policy.ts` (new), `human-video-retry-policy.spec.ts` (new)                                                                                                                                                                                   | Four of the five things asked for already have a home, and the work was using them rather than growing a second copy: the backoff curve and its jitter come from `@oshun/resilience` (`backoffDelay`), the breaker is that library's `CircuitBreaker` wrapped rather than reimplemented - the conventions ratchet exists to keep both singular - 'no infinite provider polling' is already settled by 11.07's poll budget, and the attempt counter is bounded by the job envelope. What is new is the JUDGEMENT about which failures deserve which mechanism, which is the part that gets this wrong most often. A circuit breaker measures whether a lane is BROKEN, not whether it is producing work we like, so transport, provider and corrupt-output failures count toward it and QUALITY FAILURES DO NOT: a lane that renders promptly and misses the identity threshold is available, and tripping its circuit would take a working lane out of service over a judgement that belongs to routing. Budget, consent, validation, cancellation and our own internal errors cannot be a lane's fault at all, and two more are excluded by name - a content-policy refusal (every lane would refuse it) and an ambiguous submission (a network event already handled by refusing to resubmit). Without that classifier one malformed request rejected four times would open the circuit on a lane that is working perfectly. Poison is the other new judgement: if the same NON-PROVIDER failure follows a shot across two different lanes, or one of our invariants fires twice, the next attempt is not the one that works - something about this job breaks whatever renders it, and retrying spends money to learn nothing. Quarantine stops the retries, keeps the lineage and its running cost, and never re-enters the queue without a named person. The retry curve is video-scaled (5 s first, 5 min ceiling, full jitter) because the library default of 100 ms is right for a request that usually succeeds on the second try, not for minutes of somebody else's GPU. `MAX_JOB_WALL_CLOCK_MS` is the outer bound 11.07's budget must sit below, and `checkPollBound` walks a schedule to prove it terminates - a delay that shrinks to zero or a budget past the wall clock shows up as `bounded: false` rather than as a worker polling a dead job until someone notices the bill. 20 tests, one of which runs the SHIPPED poll curve through the audit. Targeted tsc clean. | Exercises the real `@oshun/resilience` CircuitBreaker with an injected clock, and the real `nextPollDelayMs` curve from 11.07.                                                                            | $0.00 | None    |
| 182.C.15.09 | Allow failed shots to resume without regenerating approved audio, references, or already released shots.               | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-resume.ts` (new), `human-video-resume.spec.ts` (new)                                                                                                                                                                                               | 182.C.15.02 recorded WHICH saga steps survive a failed shot; this is the rule that decides, per run, whether they actually do. It is not a lookup, because 'we already have it' is the wrong test and it is the tempting one: an approved take of yesterday's script is still a file, still approved, and still THE WRONG WORDS. So every reuse is digest-first - 182.C.15.01 put content addresses on the envelope for exactly this, since a handle can be re-pointed and 'the audio somebody approved' has to mean the bytes somebody listened to. Four things can be carried forward and each fails differently. Approved audio is the one most worth keeping (a provider charge plus a person's sign-off) and survives a digest match AND LIVE CONSENT - a voice take of someone who has withdrawn permission is not reusable however recent it is. Prepared references survive a digest match unless the pack revision that authorised them dropped or invalidated them, which is why this takes 182.C.06's `RevisionImpact` rather than recomputing a guess; a moved revision with NO impact supplied regenerates rather than assuming, and one that `requiresReConsent` blocks reuse entirely. The shot plan and the alignment survive only while the script digest holds, because they describe those words at those times. And a released shot is not resumed at all - it is public, re-rendering would replace something an audience already has, and withdrawing it is a separate act with its own record. Savings are reported PER ARTEFACT rather than as a total, because the audio figure is a real provider charge and the rest is our own compute; adding them would make a saving look bigger than the money it represents. 17 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | Pure computation over digests, consent state and a revision impact; no provider call.                                                                                                                     | $0.00 | None    |
| 182.C.15.10 | Add progress events that reflect real stages and avoid fake percentages where provider progress is unavailable.        | Oshun BFF | 182.C.05   | complete | `apps/oshun/bff/src/generation/human-video-progress.ts` (new), `human-video-progress.spec.ts` (new)                                                                                                                                                                                           | The task says avoid fake percentages 'where provider progress is unavailable', and on this API it is unavailable exactly where it would matter most: the poll response carries a `status` of `pending`, `in_progress`, `completed` or `failed` and nothing else - no percentage, no frame counter, no queue position (measured 2026-09-10). So `Progress` is a UNION rather than a number. `counted` is produced only where something real is being counted, and `uncounted` carries what is actually known - how long this has been going, and whether that is long compared with MEASURED history. An elapsed time is not progress and is not presented as one; it is simply the honest thing to show someone waiting. `progressEvent` REFUSES counted units for a stage whose source is not countable rather than believing the caller, because that is the exact mistake this exists to prevent and it is easiest to make from outside. Download is the case that looks countable and is not: 182.C.11.12 recorded no `Content-Length` on the content response, so bytes-so-far has no denominator and is reported as an observation with a null total rather than a bar filling at an invented rate. Project level IS countable and is the one place a fraction is honest - a shot is discrete and a finished shot is a fact, so it counts TERMINAL SHOTS rather than stages: a shot at `rendering` is not 60% done, it is not done. `estimateRemaining` returns null until there are 20 finished runs on the lane to take a median from and carries the sample size with the answer, because one canary is a data point and '2 minutes remaining' derived from a single 23-second render is a guess wearing a clock. `findProgressViolations` audits a delivered stream for the three ways one lies: a fraction where nothing is counted, a fraction that goes backwards, and a denominator that moves so the bar appears to advance while nothing happens. 15 tests. Targeted tsc clean.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Rests on the 2026-09-10 poll-response and download-header measurements in the transport and canary evidence.                                                                                              | $0.00 | None    |

### 182.C.16 Media Storage, Processing, and Delivery

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner       | Dependency | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Cost                                                        | Blocker                |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | ----------- | ---------- | ----------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------- | ---------------------- | --- | ---- |
| 182.C.16.01 | Define quarantine, intermediate, candidate, approved, released, and deleted object namespaces with lifecycle and …     | Media Infra | 182.C.15   | wired       | `apps/oshun/bff/src/generation/human-video-object-namespaces.ts` (new), `human-video-object-namespaces.spec.ts` (new), `human-video-output-download.ts`, `human-video-output-download.spec.ts`, `human-video-durable-ingestion.spec.ts`                                                                                                                                                                                                                                                                                       | Six namespaces, each with its own non-prefixing top-level prefix so a prefix-scoped bucket grant covers exactly one of them — the invariant a policy cannot state about itself, asserted pairwise in the spec. The namespace, not a database column, decides the readers, whether a signed URL exists at all, and the retention: `quarantine` is pipeline-only with no URL because unmeasured bytes and an HTML error page are indistinguishable until something has looked; `delivery` reads `released` and nothing else; the evaluator reads `released` too, because `182.C.01.05` wants a score computed from the bytes a viewer gets. Lifecycle moves are enforced rather than documented — `quarantine → approved` and `candidate → released` are refused as stage-skips, `intermediate` is a leaf that leaves only by deletion, and a move whose digest changed is refused as a re-encode masquerading as a relocation. Derivation is a separate relation from transition and no namespace derives from `quarantine`: a proxy inherits nothing of the check its source never had. Two defects found and fixed while wiring it: the quarantine key was keyed by candidate index alone, so a timeout-triggered retry racing the attempt it replaced shared a key and the dead attempt's abort would delete the live object (attempt is now part of the key); and `promoteOutput` accepted its destination key from the caller, which could land measured bytes in any namespace — it now derives the destination through `planTransition`. `durableKeyFor` is gone: a probed take is a candidate, not a release, and the old `human-video/` prefix said otherwise.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | n/a (no provider call)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | —                                                           | None                   |
| 182.C.16.02 | Stream uploads/downloads through bounded memory; prohibit base64 video payloads in jobs, databases, logs …             | Media Infra | 182.C.15   | wired       | `apps/oshun/bff/src/generation/human-video-inline-media-guard.ts` (new), `human-video-inline-media-guard.spec.ts` (new), `human-video-redaction.ts`, `human-video-transition-outbox.ts`                                                                                                                                                                                                                                                                                                                                       | Detection is by SNIFFING a bounded prefix, not by length, and that is the design decision the item turns on: a 4 KB base64 string is routinely a JWT, a detached signature, or a serialized key, so a length rule would produce false positives, false positives produce an exception list, and the exception list is where a clip eventually hides. Container signatures are matched instead, reusing the upload path's sniffer, so a JWT, a signature, a hex digest, and an ffmpeg argument blob all pass while an MP4 is caught in four shapes: a data URI, a bare `b64_json`-style field, a `Buffer` that went through `JSON.stringify`, and a raw byte container. A `data:text/plain` wrapper around MP4 bytes is caught by the bytes, because the declared type is a claim. Two responses by source: payloads WE assembled are refused with the JSON path and size (`applyTransition` guards both halves of its write), while a PROVIDER's payload is summarized to type/size/digest by the redaction path - losing the evidence would be the worse trade. One defect found and fixed: the `log` redaction profile bounded no string length at all, so a bare-base64 clip in an unnamed provider field reached a log line whole; both profiles now bound it and the log bounds it tighter. Bounded memory is proved deterministically rather than by measuring RSS: every chunk is yielded from ONE buffer overwritten between yields, so a consumer that retained chunk references would hash N copies of the last chunk and produce a different digest - the digest matching IS the no-retention proof, on the download path (32 MiB) and the upload path (8 MiB).                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | n/a (no provider call)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | —                                                           | None                   |
| 182.C.16.03 | Probe and normalize container, codec, dimensions, FPS, duration, time base, rotation, pixel format, color space …      | Media Infra | 182.C.15   | implemented | `apps/oshun/bff/src/generation/human-video-media-facts.ts` (new), `human-video-media-facts.spec.ts` (new), `human-video-media-conform.ts` (new), `human-video-media-conform.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                    | All fourteen named fields come off one `ffprobe -show_format -show_streams` plus `-show_entries stream_side_data=rotation`, with loudness from a separate opt-in `ebur128` decode. Two measurements on this box on 2026-09-10 drove the design. FIRST: ffprobe OMITS `color_space`, `color_transfer`, `color_primaries` and `color_range` when the file does not carry them - an unflagged libx264 mp4 has none of the four, the same encode with `-colorspace bt709` has all four - so untagged is reported as `null`, never defaulted to BT.709, because that default is exactly how a delivery ends up washed out. SECOND: ffprobe signs a display matrix the opposite way from how a person says it (a clip that displays 90 deg clockwise reads back as `rotation: -90`), so the sign is flipped once here and `displayWidth`/`displayHeight` are derived - a caller reading `width`/`height` alone concludes 16:9 about a 9:16 phone video. The rotated fixture had to be made by patching the `tkhd` matrix directly, because ffmpeg 6.1 on this box will not write one (`-metadata:s:v rotate=` is ignored on output, `-display_rotation` rotates the pixels instead); ffprobe reads the patched file back as real side data, which the test asserts rather than assumes. The conform half keeps TAG separate from CONVERT: writing tags onto an untagged file is legitimate and is recorded as an assumption, while a file genuinely tagged BT.601 gets a real `colorspace` filter, because re-tagging without converting is how skin tone shifts between two shots of the same person. Loudness is role-gated - refused outright on the approved speech asset (the evaluators align against its level and timing), skipped for an evaluation proxy, and for a delivery rendition applied as a COMPUTED STATIC GAIN capped by the measured true peak rather than a dynamics processor. Proof is a round trip: plan, run the built argv through real ffmpeg, re-probe the output, and assert the encoder wrote what the plan claimed - 30 fps, 90 kHz timescale, bt709 on all four, 48 kHz stereo, integrated loudness within a decibel of the computed target, duration unchanged, and the source's `comment` tag gone. | ffmpeg/ffprobe 6.1.1 on the Linux dev box; fixtures built per run                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —                                                           | None                   |
| 182.C.16.04 | Preserve the untouched provider output for audit while deriving evaluation proxies and delivery renditions from …      | Media Infra | 182.C.15   | implemented | `apps/oshun/bff/src/generation/human-video-derivation.ts` (new), `human-video-derivation.spec.ts` (new), `human-video-object-namespaces.ts`, `human-video-object-namespaces.spec.ts`                                                                                                                                                                                                                                                                                                                                          | The provider output is registered ONCE as an audit master and never written again: `registerMaster` is idempotent on identical bytes (a redelivered webhook and a resumed job both land there legitimately) and REFUSES a second registration with different ones, because the way an audit copy gets quietly replaced is not malice, it is a retry that reused an identity. Everything else is a derivative addressed by the MASTER's digest rather than its own - so a proxy's key says what it is a proxy OF and cannot be re-pointed at another take, and a rendition's says which approved master it was transcoded from. The recipe digest covers the tool, its VERSION, and the exact argv, so the same arguments on two ffmpeg builds are two derivatives rather than one silently overwriting the other; a proxy whose recipe moved under a stable key would break `182.C.01.05` quietly. `verifyAgainstMaster` takes a FRESHLY MEASURED digest rather than the record's own copy, which is what makes 'the provider output is preserved untouched' answerable instead of asserted - the spec writes a real file, hashes it, edits it, and the verification is what notices. `auditTrailFor` refuses to attest to a derivative that was planned and never produced. One defect fixed in `182.C.16.01` on the way: `intermediate` was flagged `contentAddressed: false` while its keys ARE digest-scoped, so `parseObjectKey` returned null for every proxy; the flag is now `digestAddressed: 'own'                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | 'source'                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | 'none'`, which is the distinction that was being conflated. | n/a (no provider call) | —   | None |
| 182.C.16.05 | Generate thumbnails, contact sheets, scrub proxies, waveforms, captions, transcript overlays for review …              | Media Infra | 182.C.15   | implemented | `apps/oshun/bff/src/generation/human-video-review-artifacts.ts` (new), `human-video-review-artifacts.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                           | Timecode is the part that is quietly hard and it is where the tests concentrate. A reviewer's note at `00:00:12:07` has to mean ONE frame, and NTSC makes that awkward: 29.97 counted at 30 drifts ~3.6 s per hour, which drop-frame corrects by skipping two LABELS every minute except every tenth - a labelling convention, not a change to the frames, and conflating the two is how a note lands on the wrong shot. Both counts are implemented against the standard checkpoints (frame 1800 to `00:01:00;02`, 17982 to `00:10:00;00`, 107892 to `01:00:00;00`), the drift is asserted rather than described (the same hour counted straight gives `00:59:56:12`), and every frame across two drop points round-trips. The base is CHOSEN from the measured rate by proximity, because ffprobe reports `30000/1001` and never the number 29.97; 23.976 counts straight because no standard drop-frame convention defines it. Two design decisions carry the rest. Captions come from the APPROVED alignment rather than from ASR of the delivered video, so a take that dropped a word shows the caption with nothing under it - captioning from the output would make the review agree with the take by construction. And the transcript overlay is a SIDECAR bound to frame ranges, never burned in, because text rendered into the pixels changes what the identity and sync evaluators measure and what a viewer would be served. Every ffmpeg recipe (poster, contact sheet, scrub proxy, waveform) is executed against a real 1280x720 take and the outputs re-probed for the planned dimensions, all inside the existing `DERIVATIVE_LIMITS` bounds; a waveform for a take with no audio is refused rather than drawn as a flat line, because a flat line looks like a silent take and an absent stream is a different fault with a different retry.                                                                                                                                                                                                                                                                                                                                                                 | ffmpeg/ffprobe 6.1.1 on the Linux dev box                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                           | None                   |
| 182.C.16.06 | Implement deterministic muxing so approved speech audio remains sample-aligned after normalization, stitching …        | Media Infra | 182.C.15   | implemented | `apps/oshun/bff/src/generation/human-video-mux.ts` (new), `human-video-mux.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                     | Tested with a signal whose position is known to the sample: 48 kHz of digital silence carrying ONE full-scale click at sample 24000, muxed against a 25 fps render, decoded back, and the click's index re-measured. THE MODULE'S OWN TEST FOUND A DEFECT IN THE MODULE. `-avoid_negative_ts make_zero` looks like exactly the right flag for deterministic timestamps and is how the mp4 edit list gets destroyed: the AAC encoder's priming delay is carried as a NEGATIVE start timestamp, and `make_zero` rewrites it to zero, turning the priming samples into real leading audio. Isolated on this box 2026-09-10 - plain, `-muxdelay 0 -muxpreload 0`, `+bitexact`, and `-avoid_negative_ts auto` all return the click to sample 24000; `make_zero` alone returned it at 25024, a silent 21.3 ms lip-sync error introduced by the step whose job is not to introduce one. It is now a forbidden flag/value PAIR (the flag alone is fine) with the measurement in its refusal message. Two intents follow from a second measurement: PCM in Matroska round-trips bit-exact (click at 24000, value 32767, 96000 frames) while AAC in MP4 keeps the click at 24000 but changes its value to 24482 and adds 256 samples of trailing padding - alignment survives lossy encoding, sample identity does not - so `exact-audio` refuses a lossy codec outright and `delivery` accepts the loss while holding alignment to the same standard. `-shortest` is banned because it truncates the shorter input, i.e. silently drops the end of somebody's approved speech; a length mismatch is refused so an operator sees it. `aresample=async` is banned by construction rather than by test, because it does nothing when timestamps are clean and only acts on the take where it matters. Verification decodes both files and compares sample counts and the PCM digest: short is always a refusal, long past the measured padding bound is a refusal, and a lossy file cannot be called the exact-audio deliverable. Stitching refuses a seam where rate, layout, codec, geometry or fps differ rather than absorbing it by re-encoding, since absorbing it is what retimes everything after the seam.                          | ffmpeg/ffprobe 6.1.1 on the Linux dev box                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                           | None                   |
| 182.C.16.07 | Add HLS/DASH or signed progressive delivery as appropriate, range requests, CDN cache controls, expiry …               | Media Infra | 182.C.15   | implemented | `apps/oshun/bff/src/generation/human-video-delivery.ts` (new), `human-video-delivery.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                           | The signature covers KEY, TENANT, AUDIENCE and EXPIRY together, and the tests are written as attacks rather than round trips: edit the tenant, upgrade the audience, push out the expiry, move the key to another rendition, present it as another tenant, sign with another secret. Signing the key alone would let a candidate link be replayed with a different tenant in the query; omitting the audience would let a six-hour delivery link be reused as a fifteen-minute creator review link or the reverse. The signature is checked BEFORE the clock, because reporting `expired` for a forgery tells an attacker the shape was right and only the clock was wrong. TTLs are not accepted from the caller - they go through the namespace's own `signedUrlTtlFor`, so a six-hour link to a candidate cannot be minted by asking for one, and quarantine has no URL at all. Cache-control resolves a real tension: a released rendition is content-addressed so `immutable` is truthful, but the object is somebody's face and a shared cache holding it past the signature is the authorization outliving itself at the edge - so max-age is CLAMPED to what is left of the signature and everything below `released` is `private, no-store`. Range handling covers the suffix form (`bytes=-500` is the LAST 500 bytes, and reading it as 0-500 serves the wrong part of the file with a 206 that looks like success), refuses multipart rather than answering half of it, clamps an over-long end, and returns a 416 naming the true size so a client can correct itself. HLS is chosen only when it earns its complexity (over a minute, or more than one rendition a player could switch between); every segment URI is individually signed with the PLAYLIST's expiry, because a VOD playlist is fetched once and shorter-lived segments would start failing mid-take with nothing to re-fetch.                                                                                                                                                                                                                                                                                                                     | n/a (no provider call)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | —                                                           | None                   |
| 182.C.16.08 | Verify decoded duration and A/V timestamps after every transcode; reject drift introduced by variable frame rate or …  | Media Infra | 182.C.15   | implemented | `apps/oshun/bff/src/generation/human-video-timestamp-verify.ts` (new), `human-video-timestamp-verify.spec.ts` (new), `human-video-media-conform.ts`, `human-video-media-conform.spec.ts`                                                                                                                                                                                                                                                                                                                                      | The central measurement, on this box 2026-09-10: the same 29.97 fps source encoded twice, differing ONLY in `-video_track_timescale`. At 1/90000 every frame interval is exactly 3003 ticks (299 of them, one value). At 1/1000 they are 33 ticks 189 times and 34 ticks 110 times. Nothing about the picture changed - the stream became VARIABLE FRAME RATE because 1000 ticks per second cannot represent 30000/1001 (a frame is 33.3667 ticks and a container stores integers), and ffmpeg emits no error for it. That is 1 ms of jitter re-randomised every frame, and a lip-sync evaluator has no way to tell it apart from the take being slightly out, so the container's error becomes the model's score. Exactness is therefore COMPUTED (`frameDurationTicks`) rather than eyeballed, and `planConform` now asserts the profile's own timescale/rate pairing before anything is encoded. A second thing the tests corrected: 90000 does NOT represent 23.976 or 59.94 - both fall exactly half a tick short - which is easy to assume it does because 90000 is the standard mp4 timescale; representing 24000/1001 needs a timescale divisible by 24000 and 59.94 needs one divisible by 60000, and both are asserted with the working values. Timestamps are sorted before differencing because packets arrive in DECODE order and a B-frame stream legitimately presents them out of it - treating decode order as presentation order reports negative intervals on a good file, which is the false alarm that gets a check switched off. The before/after comparison measures decoded counts rather than declared duration (a `+faststart` clip cut in half still reports its original length) and compares the A/V START OFFSET, which is the failure durations cannot see: a pass that shifts audio against video by a frame produces two files with identical lengths and different sync.                                                                                                                                                                                                                                                                                                                       | ffmpeg/ffprobe 6.1.1 on the Linux dev box                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                           | None                   |
| 182.C.16.09 | Make deletion propagate to derivatives, provider copies where supported, CDN, caches, search indexes …                 | Media Infra | 182.C.15   | wired       | `apps/oshun/bff/src/generation/human-video-erasure.ts`, `human-video-erasure.spec.ts`, `human-video-object-store.ts` (new), `human-video-object-store.spec.ts` (new), `human-video-object-store.integration.spec.ts` (new), `human-video-erasure-runner.ts` (new), `human-video-erasure-runner.spec.ts` (new), `human-video-destruction-store.ts` (new), `human-video-destruction-store.spec.ts` (new), `human-video-runtime.ts`, `libs/oshun/persistence/prisma/migrations/20260912090000_human_video_destruction_manifest/` | The SET is computed here rather than assembled by each caller, the same way `human-video-derivatives.ts` does for a reference asset - master, every proxy, every delivery rendition, quarantine objects from every attempt that ever ran (an attempt that died mid-flight left bytes nothing else references), search documents, edge paths, provider copies. A deletion reaching four of those and reporting success is worse than one that fails loudly, because the subject has been told their likeness is gone. Two ordering decisions carry the item. The edge is purged TWICE: purging only before origin deletion leaves a window where a request re-fills the cache from an origin that still has the bytes, and purging only after leaves a window where the edge serves what origin can no longer describe - so purge, delete, purge, and the second pass catches a re-fill triggered between the first two. And the tombstone is written LAST, so a crash halfway leaves the job looking un-erased and an idempotent retry redoes everything; writing it first would leave a job that looks finished with bytes still on disk, which is the exact failure the item exists to prevent. The tombstone is checked rather than trusted: `assertTombstoneCarriesNoLikeness` reuses `182.C.16.02`'s sniffing detector so a base64 thumbnail kept 'for the audit record' is caught in every shape, and refuses a numeric run long enough to be a face embedding - a vector identifies a person as surely as a photograph. A lane with no deletion API produces a RETENTION DISCLOSURE obligation instead of a deletion step, because claiming a provider deletion nobody performed is the specific lie this phase forbids. `verifyErasure` refuses a target nobody reported on rather than treating silence as success, since an unverified target and a missed one are indistinguishable from there. Composed into the existing signed-deletion fan-out through `generated-artifact-erasure.ts`'s `additionalPartitions` shape rather than as a second path beside it.                                                                                                                                                     | 182 tests across the seven affected specs, plus 5 against a real versioned MinIO bucket (`human-video-object-store.integration.spec.ts`, excluded from the default run and driven through `vitest.integration.config.ts`). Typechecked against a scratch config extending `apps/oshun/bff/tsconfig.typecheck.json` with a deliberate control error, which reports while the touched files do not; eslint and prettier clean. RUN LIVE 2026-09-12 against the dev Postgres and a versioned MinIO: `DELETE /assets/:id` destroyed a take's master and the proxy no table records, left the sibling take's objects and the job's quarantine alone, and left NO SURVIVING VERSION; the withdrawal covered every take, the job's orphaned attempt, and the uploaded references and audio. Two measurements shaped the adapter. A HEAD CANNOT CONFIRM A DELETION - `HeadObject` on an absent key and on an absent BUCKET come back byte-identical (`Error`, message `NotFound`, 404, no `Code`, because a HEAD has no body for the SDK to parse a code out of), and one means the bytes are gone while the other means this deployment looked somewhere they are not; so `resolves()` is a LIST of the exact key, which is a GET and does carry a body. AND A PLAIN `DeleteObject` ON A VERSIONED BUCKET LEAVES EVERY BYTE VERSION behind a delete marker - 404 to every reader, fully present to anyone who can list versions - measured at `deleted 3 version(s)` on a key an ordinary delete had already removed. Evidence: `evidence/2026-09-12-erasure-bound-live/`. | —                                                           | None                   |
| 182.C.16.10 | Load-test large references, long clips, concurrent downloads, and evaluator fan-out without exhausting BFF memory or … | Media Infra | 182.C.15   | implemented | `apps/oshun/bff/src/generation/human-video-load-budget.ts` (new), `human-video-load.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                            | The load question here is not requests per second, it is what runs out first, and the answer is not the BFF's heap. `182.C.16.02` proved the download and upload paths retain no chunk, so twenty concurrent downloads cost twenty chunk buffers rather than twenty videos. What actually runs out is TEMP DISK (four concurrent conforms of a 200 MB take is a gigabyte of scratch nothing in the JS heap accounts for) and CHILD PROCESS memory (ffmpeg's is ffmpeg's, invisible to `process.memoryUsage()`), so the budget counts those two and the concurrency. It REFUSES rather than queueing without bound, because an unbounded queue converts a capacity problem into a timeout an hour later in a different component; and a reservation larger than the entire budget is a PERMANENT failure rather than a wait, because waiting for capacity that can never exist is a hang with a nicer name. Waiters are served in arrival order so a stream of small jobs cannot starve a large one - head-of-line blocking is the honest trade, and it is asserted. A double release is refused, since silently inflating capacity is the exhaustion this exists to prevent, arriving later and elsewhere. The disk estimate is a documented approximation (source size x 2.5, exported so the reserver and the refusal-reader use the same number) because an output size is not knowable before the transcode runs and reserving nothing until the bytes exist reserves after the disk is gone. The load spec is real: 384 MiB of streamed traffic through the actual download path across six concurrent downloads, proved held-nothing by the reused-buffer digest trick rather than by RSS sampling; four concurrent ffmpeg conforms of a 30 s take with the fourth waiting, with reserved and actual scratch both sampled against the budget; and an evaluator fan-out bounded at two in flight while keeping result order. 13 tests, 13 s.                                                                                                                                                                                                                                                                                | ffmpeg 6.1.1 on the Linux dev box; 4 cores, ~12 GB free                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | —                                                           | None                   |

### 182.C.17 Real Audiovisual Lip-Sync Evaluation Runtime

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner | Dependency  | Status         | Changed Files                                                                                                                                                                                                              | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Provider/Hardware Evidence                                                                                               | Cost | Blocker                                                                   |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | ----- | ----------- | -------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------ | ---- | ------------------------------------------------------------------------- |
| 182.C.17.01 | Select and license at least two complementary sync evaluators—one SyncNet/LSE-class baseline and one current robust … | Isis  | 182.C.16    | implemented    | `apps/oshun/bff/src/generation/human-video-sync-evaluators.ts` (new), `human-video-sync-evaluators.spec.ts` (new), `docs/domains/isis/human-video/sync-evaluator-selection.md` (new)                                       | Two evaluators selected to FAIL DIFFERENTLY rather than to agree, because two measures sharing a failure mode are one measure with extra cost. `syncnet-v2` (MIT, Joon Son Chung; licence text read off disk at `~/.cache/oshun-human-video/LICENSE.md`, not from a repository badge) is the learned LSE-C/LSE-D baseline - strong exactly where geometry is weak, since it never has to locate a landmark, and blind on profile and small faces WITHOUT lowering its own confidence. `lip-envelope-correlation` (Apache-2.0 via MediaPipe face_landmarker) is the geometric complement - it catches directly the two failures an embedding scores only indirectly (a mouth that never moves while words play, a mouth moving through silence) and reports its own landmark confidence, so its degradation is visible. A window neither can measure is COVERAGE, not a pass. The third-slot question is answered from a record rather than re-litigated: AV-HuBERT is CC-BY-NC, non-commercial, disqualified for a paid product whatever its accuracy - a licensing fact, not a performance judgement; a Wav2Lip-class discriminator shares SyncNet's lineage so its agreement would be one measurement reported twice; SparseSync needs its licence and CPU throughput measured here first. `assertIndependentAgreement` enforces both rules and takes an injected registry lookup so the correlated-lineage branch is exercised against a registry that HAS such a pair - a rule nothing tests is a rule that stops working the day it is needed. The calibration protocol is executable rather than prose: `assertCalibrationDesign` refuses an unblinded design first (a rater who sees the score is not rating the video), then a stratum nobody sampled, then a cell under 30 clips (below which a proportion cannot separate a 90% pass rate from an 80% one), then fewer than 3 raters (below which there is no agreement statistic), then Krippendorff alpha under 0.667. The spec verifies every artifact digest against the weights actually staged on this box, so a registry that has drifted from what the machine loads fails rather than documenting a file nobody has.                                                                                                                          | Weights verified on disk: syncnet_v2.model 961e8696…, SyncNetModel.py 41b85aed…, face_landmarker.task 64184e22…          | —    | None                                                                      |
| 182.C.17.02 | Implement reproducible model acquisition with pinned weights/digests, license record, device capability checks …      | Isis  | 182.C.16    | implemented    | `apps/oshun/bff/src/generation/human-video-model-acquisition.ts` (new), `human-video-model-acquisition.spec.ts` (new), `human-video-sync-evaluators.ts`                                                                    | THE DIGEST IS THE AUTHORITY, NOT THE URL. Bytes are hashed while streaming and checked before the file gets the name a loader looks for - it lands at a `.partial` path and is renamed only when both digest and size match, so a URL that starts returning something else (a moved repository, a redirect to a login page, an HTML error with a 200) fails here instead of becoming the model. That matters more than usual because one artifact is CODE: `SyncNetModel.py` is a Python module the evaluator imports, so fetching it is a supply-chain step and the pin is the only thing between a changed upstream and arbitrary code in the evaluator process. Its registry URL was corrected to the RAW GitHub url and verified against the network on 2026-09-10: 3342 bytes, sha256 `41b85aed…`, matching the staged file exactly (a blob url returns an HTML page). Hosts are allow-listed before any request, the same way `182.C.11.08` checks a provider download url before attaching an API key. A body longer than the pin is stopped mid-stream rather than allowed to fill the disk. A MISSING EVALUATOR FAILS CLOSED: `assertEvaluatorAvailable` throws with the file, the measured digest, the pinned digest and the url, because the tempting alternative - skip sync scoring and let everything else proceed - produces a release that was never measured reported as one that passed. Existence is not presence: a file with the right name and the wrong contents is `corrupt`, and is re-fetched rather than left for the next startup to make the same decision about. Device capability is checked against the registration at startup, so a GPU-only evaluator on a CPU host is refused rather than discovered mid-job. The spec runs against the real staged 54 MB checkpoint.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | SyncNetModel.py raw URL verified against network 2026-09-10 (3342 B, 41b85aed…); staged weights verified by the spec     | —    | None                                                                      |
| 182.C.17.03 | Decode real frames and audio, detect/track faces, locate mouth regions, resample consistently …                       | Isis  | 182.C.16    | implemented    | `tools/human-video/evaluator/prepare_evaluator_inputs.py` (new), `apps/oshun/bff/src/generation/human-video-evaluator-inputs.ts` (new), `human-video-evaluator-inputs.spec.ts` (new)                                       | Runs for real against a real face: the committed probe fixture moved through frame by a zoompan at 30 fps, so detection tracks something that changes and the conform to 25 is exercised rather than skipped. Measured on this box - 100 frames decoded, a face found and tracked in all 100, 100 mouth crops, 400 MFCC frames, 2.3 s wall for a 4 s clip. Four decisions carry it. FRAMES ARE CONFORMED TO 25 fps AND THE RESAMPLING IS REPORTED, because SyncNet's five-frame window is 200 ms only at 25 and conforming duplicates or drops frames - the manifest carries the source rate, the decoded count and the count the source would have produced, so a caller can tell how much of what it scores is interpolation. THE MOUTH IS LOCATED FROM YuNet'S OWN LANDMARKS (two of its five points are the mouth corners) rather than from a face-box fraction, which survives head roll; the fraction approach fails silently - the crop drifts off the mouth, the score drops, and nothing says the crop was wrong rather than the sync. A FRAME WITH NO FACE IS A GAP, NOT A SKIP, and coverage is computed from frames CROPPED rather than frames a face was seen in, because a face at the edge of frame whose crop falls outside it is presence without measurability and `182.C.17.06` treats the two differently. MFCCs COME FROM `python_speech_features` at its documented defaults - the library SyncNet's reference implementation uses - because a reimplementation would be a second thing to be wrong, and wrong quietly. Every crop carries a source-time second, which is what converts an offset reported in windows into milliseconds against a creator's take. `assertWindowGeometry` refuses a manifest where the video and audio windows no longer coincide, since a worker moved to 30 fps without changing the window would silently score 167 ms of video against 200 ms of audio. Crop scale is a PARAMETER, not a constant: it is not yet calibrated against SyncNet's training framing and `182.C.17.04` must fit it.                                                                                                                                                                                                                                                           | Real face fixture; YuNet 0.959 mean detection score over 100 frames; venv interpreter at ~/.cache/oshun-human-video/venv | —    | None                                                                      |
| 182.C.17.04 | Calculate per-frame/window sync confidence and offset, not only one clip-level LSE-C/LSE-D value.                     | Isis  | 182.C.16    | live-probed    | `tools/human-video/evaluator/score_syncnet.py` (new), `tools/human-video/evaluator/selftest_scoring.py` (new), `apps/oshun/bff/src/generation/human-video-sync-scoring.ts` (new), `human-video-sync-scoring.spec.ts` (new) | SyncNet v2 loaded strictly from the pinned 54 MB checkpoint into the pinned module and run over the real mouth crops from `182.C.17.03` - 96 windows in 7 s on 4 CPU threads. THE CASE FOR PER-WINDOW SCORING WAS MEASURED, NOT ARGUED. On a real clip of a frozen face with a pure tone - a take with no alignment anywhere - the clip-level figures came back as offset 0 frames, confidence 0.28, while per-window confidence had a median of 0.032 and per-window offsets scattered from -9 to +9 with a standard deviation of 5.6. Averaging 96 scattered offsets produced a confident-looking '0 ms, in sync' for a clip that was never in sync with anything. The offset arithmetic is proved independently of the model: `selftest_scoring.py` builds embedding series where one IS the other shifted by K and asserts the scorer returns exactly K, per window and clip-level, for seven shifts positive and negative - a check that survives a weights change and that catches a sign flip in `vshift - idx` or an off-by-one in the padding, which would otherwise show as a slightly worse score nobody notices. TWO MEASURED PROPERTIES OF THE CONFIDENCE NUMBER that any threshold has to account for. (1) It has a FLOOR from the search width: unrelated random embeddings score 7.26 against 44.2 for an exact alignment, because the minimum of 21 draws sits below their median whether or not anything aligns - so confidences computed at different `vshift` are not comparable and `vshift` is part of the record. (2) It CONFLATES 'nothing aligned' with 'nothing moving': the frozen face scored 0.032, far BELOW the random-noise floor, because a constant lip series has no spread for the median to sit above - so low confidence alone cannot tell a bad take from a still one, and `182.C.17.07` must separate them. Windows whose minimum sits at the search edge are reported rather than clamped, because clamping turns 'we could not see far enough' into a confident answer. Nothing here is a threshold.                                                                                                                                                                                                                                                                   | SyncNet v2 961e8696… loaded strict=True, 96 windows scored on a real face clip; known-shift recovery exact for 7 shifts  | —    | None                                                                      |
| 182.C.17.05 | Bind speech-active windows to tracked cast members and score only the expected speaker while detecting false mouth …  | Isis  | 182.C.16    | implemented    | `tools/human-video/evaluator/speech_and_motion.py` (new), `apps/oshun/bff/src/generation/human-video-speaker-binding.ts` (new), `human-video-speaker-binding.spec.ts` (new)                                                | The shot plan says who speaks when and only that track is scored; everyone else is asked a different question - were they moving their mouth while somebody else talked. Speech activity is the MFCC's own energy coefficient thresholded RELATIVE to the take's quiet floor, because an absolute dBFS cut calls a quiet take silent throughout and a noisy one speech throughout. TWO DEFECTS WERE FOUND BY BUILDING THE FIXTURE. FIRST, the obvious baseline is wrong: normalising motion by a track's own overall median makes the exact case this item exists for invisible, because a listener moving through the WHOLE line has a ratio of about 1 everywhere - their constant motion IS their median. The first fixture did precisely that and the detector reported nothing. The baseline is now the track's motion while NOBODY speaks, a track with no silence is reported UNMEASURABLE rather than cleared, and a track moving as much at rest is reported as CONSTANT MOTION rather than speech-correlated. SECOND, a pixel-difference motion measure has a COMPRESSION FLOOR: the static half of the test frame is pixel-identical in the source and still measured up to 8.4 units of change, with spikes appearing before any speech - h264 rate control reallocating bits across the whole frame when the other half starts flickering. That is what the five-frame run-length guard is for; the spikes run one or two frames. Measured separation on the gated fixture: p95 motion 38.9 on the track that moves during speech against 5.17 on the one that never moves, both medians near 0.002. Speech runs nobody is planned to be speaking in are REFUSED (a wrong plan, or dialogue nobody approved - both need a person), as are two speakers assigned to the same moment. Windows are scored only where the bound speaker's track has speech actually playing, because scoring silence measures a mouth against nothing and the low confidence is indistinguishable from a real failure. The sensitivity is provisional and says so.                                                                                                                                                                                                                                                      | Gated two-face fixture; 27/100 frames speech-active against a 25-frame gate; motion p95 38.9 vs 5.17                     | —    | None                                                                      |
| 182.C.17.06 | Treat missing/tiny/occluded/profile faces as coverage data, not an automatic pass …                                   | Isis  | 182.C.16    | implemented    | `apps/oshun/bff/src/generation/human-video-measurability.ts` (new), `human-video-measurability.spec.ts` (new), `tools/human-video/evaluator/prepare_evaluator_inputs.py`                                                   | The failure guarded is the quietest one in the evaluation stack: a take where the face is too small, turned away or simply absent produces NO sync findings, and 'no findings' reaches a release gate looking exactly like 'no problems' - a shot of the back of somebody's head passes every check there is. Every window therefore gets one of THREE verdicts, never two: measurable; EXPECTED coverage (unmeasurable and the plan said so - a wide establishing shot has a small face, that is the shot working, and it still cannot carry a sync claim); or UNEXPECTED coverage (unmeasurable and the plan said otherwise - a planned close-up that came back too small to measure is a shot that did not deliver, which needs a reroute rather than an 'accept, unmeasured'). That third verdict is why the rules are shot-plan-aware: the same forty-pixel face is a correct wide shot and a broken close-up, and a threshold that cannot tell them apart either fails every wide shot or passes every broken close-up. The FLOOR does not vary by shot size - an evaluator cannot score a twenty-pixel mouth whatever the plan called for, and below about 0.02 of the frame diagonal the crop is upsampled from fewer pixels than the first convolution needs, so the score measures the resampler. The pipeline now records per-frame geometry: relative face height, inter-ocular distance against the frame diagonal (resolution independent in a way raw pixels are not), a documented yaw PROXY (nose displacement from the eye midpoint over inter-ocular distance - a cheap signal that says 'this face is turned enough that a mouth crop is not the front of a mouth', explicitly not the SQPnP solve `182.C.06.06` uses), and whether the box touches the frame edge. A window is measurable only when EVERY frame in it is, since the evaluator sees all five and one that vanishes mid-window makes the number about the vanishing. `assertMeasurableEnough` and `assertNoCoverageCountedAsPass` are the rules the item exists for: the second catches a caller reporting 96 windows in sync when only 40 were measurable. Verified against real geometry - a real large frontal face measures 0.10 of the diagonal with a yaw proxy near 0.01 and classifies measurable at every shot size. | Real fixture geometry: interocular 0.102 of diagonal, yaw proxy 0.010-0.016, relative height 0.52-0.57                   | —    | None                                                                      |
| 182.C.17.07 | Detect global offset, local drift, intermittent desync, frozen mouth, excessive mouth motion during silence …         | Isis  | 182.C.16    | implemented    | `apps/oshun/bff/src/generation/human-video-desync-patterns.ts` (new), `human-video-desync-patterns.spec.ts` (new)                                                                                                          | 'Out of sync' is not one finding, and treating it as one produces the wrong response every time: a take uniformly 120 ms late needs a mux fix, a drifting one needs a rate conform, one fine except for a second needs a retry of that shot, a frozen mouth needs a different model. Same score, four actions. Each detector is therefore built on its own terms and, critically, tested against the OTHER five patterns' series - the value of separate detectors is entirely in their not firing on each other. GLOBAL OFFSET is the median, reported only when the SPREAD is small, because an average of scattered offsets is not a global offset - exactly the mistake `182.C.17.04` measured the clip-level figure making on a frozen face. LOCAL DRIFT is a least-squares slope in ms per second, reported only when the line explains enough variance; without that condition every noisy series produces a slope and a rate conform gets applied to a take whose problem is jitter. A perfectly flat series returns R-squared 0 rather than 1, since it has no variance to explain and calling it a perfect fit would report every steady take as a confident zero-slope drift. INTERMITTENT DESYNC measures against the clip's OWN median rather than zero, so a take with a global offset does not read as intermittently desynced end to end. FROZEN MOUTH uses motion, which is the case `182.C.17.04` showed confidence cannot express - a constant lip series and a badly desynced one both score low. MOTION IN SILENCE is its mirror. EDIT BOUNDARY is a real two-segment change-point search over every interior split rather than a threshold on adjacent differences (which fires on one noisy window), accepted only when the step exceeds twice the scatter it is splitting, because a split always exists. Unmeasurable windows are excluded from all six.                                                                                                                                                                                                                                                                                                                                                                                                                                | n/a (pure analysis over the measured series)                                                                             | —    | None                                                                      |
| 182.C.17.08 | Calibrate thresholds against blinded human ratings across languages, skin tones, facial hair, head poses …            | Isis  | 182.C.33    | blocked-corpus | —                                                                                                                                                                                                                          | NOT locally actionable, and the blocker is specific rather than a shrug. Calibration needs blinded human ratings over a stratified corpus - `182.C.17.01`'s executable protocol requires 8 strata x 30 clips x 5 raters at Krippendorff alpha >= 0.667, and `assertCalibrationDesign` refuses anything less. Neither half exists here: the corpus is `182.C.33.01`-`.04` (consented, licensed, balanced across complexion, facial hair, head pose, language including Ghanaian English/Akan/Twi/Ga/Ewe, and compression), and the raters are people. Fitting thresholds to invented ratings would be the exact fabrication this phase's bright line forbids, so no fitter was built against data that does not exist. WHAT IS READY FOR IT: every threshold the evaluators use is already a named, exported, provisional constant with the reason it was chosen written beside it - `PROVISIONAL_THRESHOLDS` (six desync patterns), `MEASURABILITY_FLOOR` and `EXPECTED_INTEROCULAR`, `PROVISIONAL_FALSE_MOTION_RATIO`, `PROVISIONAL_MIN_MEASURABLE_FRACTION` - so calibration is a matter of re-fitting values, not of finding where they live. Two measured facts the calibration must account for, both recorded in `182.C.17.04`: SyncNet confidence has a floor set by the search width (7.26 for unrelated series against 44.2 for an exact alignment at vshift 10), so confidences from different widths are not comparable and must be calibrated per width; and confidence conflates 'nothing aligned' with 'nothing moving' (a frozen face scored 0.032, BELOW the noise floor), so a threshold on confidence alone cannot separate a bad take from a still one.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Unblocks when `182.C.33.01`-`.04` deliver a rated corpus                                                                 | —    | Needs the 182.C.33 rated corpus and human raters; neither exists locally. |
| 182.C.17.09 | Return structured evidence with distributions, worst windows, annotated review frames, model/version, confidence …    | Isis  | 182.C.16    | implemented    | `apps/oshun/bff/src/generation/human-video-sync-evidence.ts` (new), `human-video-sync-evidence.spec.ts` (new)                                                                                                              | A reviewer decides about somebody's likeness in a few minutes, and a single number is not enough - not because it is imprecise but because it cannot be CHECKED. Five things, each present because leaving it out makes a specific decision impossible. THE DISTRIBUTION, NOT THE AVERAGE: `182.C.17.04` measured a clip-level '0 ms, confidence 0.28' on a take with no alignment anywhere while the per-window offsets scattered across plus or minus 9 frames - a reviewer shown the average would have approved it. THE WORST WINDOWS WITH TIMECODES, ranked by confidence ascending so the windows the evaluator was least sure of come first, since those are where a human adds most; a reviewer who cannot find the bad second cannot check the finding. COVERAGE AS ITS OWN NUMBER, because 60 measurable windows of 96 is a different document from 96 of 96 with the same scores and a reviewer must not have to infer that. THE MODEL, ITS WEIGHTS DIGEST AND ITS SEARCH WIDTH - the width belongs in the identity because `182.C.17.04` measured that confidence has a floor which moves with it, so the same take scores differently at a different vshift for a reason that is not the take; `comparable`/`incomparableBecause` name every difference at once rather than the first. A FAILURE REASON IN WORDS, not a code, because a reviewer's job is to agree or disagree with a claim and a code is not a claim - each sentence names the pattern, where it is in timecode, and how big it is. Annotated frames take the MIDDLE frame of a window, since its first frame is also the last of the window before it and is the least characteristic moment of the finding, and they carry an object key rather than bytes (`182.C.16.02`). Two refusals: evidence over nothing measurable, and a 'passed' verdict carrying findings - a document that says a take is fine while listing what is wrong teaches a reviewer to stop reading the list.                                                                                                                                                                                                                                                                                                                                              | n/a (assembles the measured series)                                                                                      | —    | None                                                                      |
| 182.C.17.10 | Require the evaluator to process the exact final muxed bytes before release; pre-transcode success is insufficient.   | Isis  | 182.C.16    | implemented    | `apps/oshun/bff/src/generation/human-video-final-bytes-gate.ts` (new), `human-video-final-bytes-gate.spec.ts` (new)                                                                                                        | Not a theoretical gap - this phase measured THREE ways a pre-transcode score goes stale, and the refusal message cites them. `182.C.16.06` found `-avoid_negative_ts make_zero` moving the audio 1024 samples (21.3 ms) during the MUX, so a take measured perfectly in sync before that step ships 21 ms out with a measurement saying it is fine. `182.C.16.08` found a 1 kHz timescale turning a constant-rate source into a variable-rate output during the CONFORM, with jitter re-randomised every frame and no ffmpeg error anywhere. `182.C.16.03` found an untagged source re-tagged rather than converted changing how every frame is displayed. All three are invisible to an earlier measurement and all three change what a viewer sees. The gate therefore refuses: a measurement whose digest is not the releasing artifact's; a measurement of an evaluation PROXY (a downscaled copy made for a machine, refused before the digest is even compared, so a proxy that somehow hashed to the release is still a proxy); a measurement of an earlier STAGE, whatever it scored - the score is not the question, the file is. Comparison is on the DIGEST rather than the key, because a key can be re-pointed and a file at a key replaced while a digest is the bytes. A multi-track release needs EVERY track to have measured the delivery: three measuring the delivery and one measuring the candidate is a release with an unmeasured speaker in it. `MEASUREMENT_INVALIDATING_STEPS` is offered as a convenience so a pipeline can re-measure before reaching the gate, and is explicitly NOT the rule - a step that changed the bytes without appearing on the list still fails, because the gate compares digests rather than trusting a list.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | n/a (gate over recorded digests)                                                                                         | —    | None                                                                      |
| 182.C.17.11 | Add adversarial tests where metadata says aligned but audio is shifted, stretched, replaced, silent …                 | Isis  | 182.C.16    | implemented    | `apps/oshun/bff/src/generation/human-video-adversarial-sync.spec.ts` (new), `human-video-desync-patterns.ts`                                                                                                               | Six clips built to LIE - well-formed containers, right-length streams, ffprobe content, and the sound not belonging to the picture: audio shifted (`adelay=1200`), stretched (`atempo=2.0`), replaced with a different take's tone, present but silent, routed to the wrong speaker, and the honest baseline as the control. The control matters: a suite where everything fails proves only that the detector fires, so the honest clip is asserted to produce NO frozen-mouth and NO motion-in-silence findings. Measured separation: on the honest clip speech occupies frames 24-50 and mouth motion 25-51, 26 of 27 frames overlapping; delay the audio 1.2 s and speech moves to 54-80 with the motion where it was - overlap ZERO, and both halves of the lie are visible (speech over a still mouth AND a mouth moving in silence), which is what makes the finding legible. WHICH detector catches them is stated rather than glossed, and asserted so it cannot quietly stop being true: this fixture is a real face that FLICKERS, not one that talks, so SyncNet's embedding has no mouth-to-phoneme correspondence to score and its confidence stays at the 0.032 floor measured in `182.C.17.04` - everything here is caught by the speech-against-motion pathway from `182.C.17.05`, and a suite claiming the embedding caught them would be claiming what this fixture cannot show. A test asserts the honest and shifted clips have IDENTICAL motion series and different speech series, which is exactly that claim in executable form. Silence produces motion-in-silence and explicitly NOT frozen-mouth, because the two findings are different faults and are not interchangeable. Also adds `buildAnalysisWindows`, which joins the three series on the window grid - speech true when ANY frame in the window has it (a window straddling the start of a line has speech in it), motion the window's MEDIAN so one compression spike does not carry a window.                                                                                                                                                                                                                                                                                                                            | Real adversarial clips through the full pipeline; honest overlap 26/27 against shifted 0/27                              | —    | None                                                                      |
| 182.C.17.12 | Benchmark evaluator throughput and schedule GPU/CPU work without starving render workloads.                           | Isis  | 182.C.16    | live-probed    | `apps/oshun/bff/src/generation/human-video-evaluator-throughput.ts` (new), `human-video-evaluator-throughput.spec.ts` (new)                                                                                                | Measured on this box over real clips of a real face through the real pipeline, at TWO lengths to confirm linearity. Per video-second: prepare (decode/YuNet/crop/MFCC) 0.73 s, SyncNet scoring 0.91 s, speech-and-motion 0.05 s, TOTAL ~1.70 s = 0.59x realtime; 16.63 s for a 10 s clip and 34.16 s for a 20 s clip, so capacity is multiplication rather than a curve fit. THE SCHEDULING FINDING IS THE SURPRISE: running several evaluators at once buys almost nothing and TURNS OVER inside the measured range - 0.42 video-s per wall-s at one, 0.50 at two (+19%), 0.52 at three (+24%), and 0.48 at FOUR, which is worse than three. Each process already saturates the cores with its own threads (torch at 4, plus ffmpeg and OpenCV), so more processes contend rather than parallelise. The policy follows from the curve rather than from taste: one evaluator at a time, three cores reserved for render/transcode/request path, because 24% more work is not worth three-quarters of the box. THREADS are capped alongside processes because they are the same resource - one process with every thread starves the box exactly as thoroughly as four processes with one thread each, and a policy counting only processes would have a hole the size of the box. `estimateCapacity` uses the MEASURED throughput at the chosen concurrency rather than dividing the single-process figure, which would make a capacity plan wrong by 2.5x at concurrency three. `marginalThroughputGain` returns null past the measured range rather than extrapolating, since a line through a curve that already turns over predicts the wrong sign.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | 4 cores, no GPU, 2026-09-10; 10 s and 20 s clips; concurrency swept 1-4                                                  | —    | None                                                                      |
| 182.C.17.13 | Decide and provision where evaluators run (a RunPod GPU worker through `libs/isis/runpod-surface`/ …                  | Isis  | 182.C.17.12 | implemented    | `docs/domains/isis/human-video/evaluator-hosting-decision.md` (new)                                                                                                                                                        | DECIDED: CPU inside the BFF worker, one evaluator at a time, is the production path; no GPU worker is provisioned now. Taken from the `182.C.17.12` numbers rather than from a preference. THREE REASONS. (1) The throughput is sufficient: a ten-second take evaluates in about 24 s while generation itself takes a provider tens of seconds to minutes, so evaluation is a fraction of a step that is already slower than it is. (2) A GPU worker would buy latency the pipeline cannot spend - shaving twenty seconds off a step inside a multi-minute job changes nothing a creator perceives, while adding a network hop, a queue, a second place for weights to go stale and a second thing to be down. (3) The scaling lever is HORIZONTAL: the measured curve says one box gains almost nothing from concurrency, so more takes means more workers, which the deployment already knows how to make. THREE NAMED REVERSAL THRESHOLDS rather than a judgement call later: sustained demand past the point where adding workers costs more than a GPU host (~25 video-minutes per hour per worker); an evaluator that genuinely needs a GPU (the current two run on CPU because that was MEASURED, not assumed); and long-form takes, where 0.59x realtime means a ten-minute take costs seventeen minutes of CPU and `182.C.22` faces a different calculation. The route is already in the repo when it reverses - RunPod through `libs/isis/runpod-surface/`. THE CPU PATH IS KEPT REGARDLESS, and not as a fallback: it is what CI runs, and every measurement in `182.C.17.03`-`.11` including the adversarial suite was produced by it. An evaluation stack that can only be exercised on hardware CI does not have is a stack whose regressions are found in production.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Decision record from the measured curve                                                                                  | —    | None                                                                      |

### 182.C.18 Exact Speech, Pronunciation, and Audio Quality Evaluation

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner | Dependency  | Status         | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Cost                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Blocker                                                                                                                                                                                                   |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | ----- | ----------- | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --- | --------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.18.01 | Run independent ASR on delivered audio (extending the fail-loud transcriber seam in `dub-faithfulness.ts` …            | Isis  | 182.C.17    | implemented    | `apps/oshun/bff/src/generation/human-video-script-fidelity.ts` (new), `human-video-script-fidelity.spec.ts` (new), `human-video-delivered-asr.ts` (new), `human-video-delivered-asr.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | The spec REPLAYS the real transcripts recorded when the vendors were probed on 2026-09-10 - three models, one known sentence - and asserts each scores ZERO content-word errors. That is not decoration: the probe found a real defect where two of the three wrote the spoken 'forty two dollars' as '$42', correct speech in a different orthography, and the content-word normalizer scored a perfect transcription as a blocked delivery. A test asserts BOTH orthographies are present in the fixture, so the check is only meaningful if both shapes are still there. Normalization runs through the contract's own `normalizeForComparison` before alignment; aligning raw strings would report a substitution on every number in every script. Content-word errors are counted SEPARATELY from function-word errors because missing 'the' and missing a person's name are both one substitution and are not the same event - `182.C.03.05` C2 gates on content only, since a recognizer dropping 'the' is a recognizer artifact far more often than a missing word. A locale the normalizer has no rules for is REFUSED rather than scored on raw tokens, which would report artifacts of not normalizing. Semantic substitutions separate a wrong number or a different word from a spelling variant ('colour'/'color' is one word, 'Thursday'/'Tuesday' is a different day), and the variant rule is stricter on short words where one edit is usually a different word ('tie'/'lie'). Speaker turns compare as a SEQUENCE, since two speakers saying their lines in the wrong order is a take where the conversation does not happen and a set comparison calls it correct. The transcriber refuses to be the generating model OR a model from the same vendor - a self-check always passes, and one that always passes is worse than no check because it produces evidence with a number on it. An unavailable transcriber makes exactness UNMEASURABLE rather than fine.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Recorded probe replayed: voxtral $0.000306, gemini-flash-lite $0.0000698, gpt-audio-mini $0.0000657                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | A live billed ASR call was not made; the recorded 2026-09-10 fixtures cover the response shape.                                                                                                           |
| 182.C.18.02 | Align delivered words to approved-audio words and flag timing stretch, truncation, re-synthesis, unexpected dialogue … | MLOps | 182.C.18.01 | implemented    | `apps/oshun/bff/src/generation/human-video-word-alignment.ts` (new), `human-video-word-alignment.spec.ts` (new), `tools/human-video/evaluator/ctc_align.py` (new), `tools/human-video/evaluator/calibrate_aligner.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-cpu-sync-and-alignment/aligner-calibration.json` (new), `.../README.md` (corrected), plus the adapters 2026-09-12: `apps/oshun/bff/src/generation/human-video-audio-adapters.ts` (new), `human-video-audio-adapters.spec.ts` (new), `human-video-audio-store.ts` (new), `human-video-audio-store.spec.ts` (new), `human-video-runtime.ts` (the binding), `libs/oshun/persistence/prisma/migrations/20260912030000_human_video_audio_take/migration.sql` (new), `schema.prisma`, `prisma-renderer.ts` | The first step the item names was done first and it CORRECTED THE EVIDENCE PAGE. That page claimed CTC's systematic late bias meant 'a calibrated constant correction brings it inside the ATSC window'. Measured on the committed known-onset fixture: median bias +48.288 ms, so the correction is -48.288 ms, and the residual scatter that remains is +/-32.87 ms (worst lead -31.68, worst lag +32.87). ATSC IS-191 allows 15 ms of LEAD, so the corrected aligner is OUTSIDE ATSC; it is inside ITU-R BT.1359-1 (+45/-125) with about 30% margin. The README now says that, and a test asserts insideAtsc === false so the optimistic version cannot come back. Writing the aligner surfaced a second real defect, in my own first draft of the CTC trellis: STAYING on a token was charged the token's probability instead of the BLANK's, which forces a non-blank emission every frame and therefore cannot represent silence - it compressed the fixture's 600 ms inter-word gaps to 80 ms and snapped the first word to 0.0, precisely the failure Whisper was disqualified for. Fixed, and the aligner then reproduced the recorded onsets exactly (+81, +65, +17, +17, +62, +35 ms). That +/-32.87 ms is the instrument's noise floor and every detector is expressed as a MULTIPLE of it rather than as an absolute figure somebody chose, so a re-measured aligner loosens the gates instead of silently claiming a precision it no longer has. Stretch is separated from OFFSET by a least-squares fit: slope away from 1 is a time scale and every word drifts further, slope 1 with an intercept is a mux or delivery constant, and they need different repairs. Measured while building it: alternating noise at twice the residual manufactures 1.05% of apparent slope on a six-word take, past a flat 1% tolerance, on a take with NO stretch in it - so the stretch gate is the slope's own standard error (sigma/sqrt(Sxx)), which shrinks with take length and word count, and the flat 1% is only a floor. Truncation suppresses the per-word comparisons, because once the words no longer correspond a per-word fault is an artifact of the misalignment. Pauses compare as GAPS, so a whole-take offset does not read as every pause having changed. Word ENDS are never used: they measure 96-141 ms early, consistently. The spec runs the REAL aligner against the committed fixture and re-derives the constant, so a drifting aligner fails the test instead of the constant quietly becoming wrong. BOUND 2026-09-12. `forced-align.py`'s own docstring said it was "used by the BFF through an injected aligner port" and NOTHING WAS INJECTED; `human-video-audio-adapters.ts` is the child process. THE ALIGNER MEASURES A DERIVED COPY AND SAYS SO: the tool takes a mono 16 kHz WAV and a take is whatever the creator recorded, so ffmpeg makes a copy and the aligner reads THAT. Resampling does not move anything in time, so the timings are still about the take — the copy is not the take, nothing binds to it, and it is deleted. THE DURATION IS PROBED FROM THE ORIGINAL for the opposite reason: a transcode can move it by a frame at the boundaries, and 182.C.35.09 measured a +20 ms AAC padding that a real provider render then confirmed. AND `format.duration` IS NOT THE AUDIO STREAM'S: ffprobe reports the LONGEST stream, so a take that arrived inside an mp4 with a video track would come back longer than the words it contains — the probe asks for streams and takes the audio one. NEITHER TOOL MAY ANSWER "NOTHING": a probe returning 0 s reads as `silent_audio`, a finding about the RECORDING, when the cause was a missing interpreter; an aligner returning no words makes every word unaligned, which blocks approval with a reason about the creator's SCRIPT. Every failure throws. | Real CTC alignment run in the spec (8.7 s, CPU): correction -48.288 ms, residual max 32.866 ms, insideAtsc false, insideItu true; `aligner-calibration.json`. 23 tests pass. No provider spend. 16 adapter tests, 14 against an injected runner and 2 RUNNING BOTH TOOLS FOR REAL against the committed clip in `evidence/2026-09-10-cpu-sync-and-alignment/` — every onset within 150 ms of the ground truth committed beside it. ONSETS ONLY: the tool's docstring reports word ENDS running 96-141 ms early because CTC emits where the evidence is decisive rather than where the sound stops, and it reports the bias rather than correcting it, so asserting on ends would be asserting on a known uncorrected error. 11 store tests. 5 runtime tests for the three-part binding. | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | None                                                                                                                                                                                                      |
| 182.C.18.03 | Evaluate pronunciation for names, acronyms, numbers, lexicon entries, and code-switched phrases with locale-aware …    | MLOps | 182.C.02.12 | implemented    | `apps/oshun/bff/src/generation/human-video-pronunciation.ts` (new), `human-video-pronunciation.spec.ts` (new), `tools/human-video/evaluator/pronunciation_gop.py` (new), `.../evidence/2026-09-10-cpu-sync-and-alignment/pronunciation-gop.json` (new), `.../README.md`, `tools/human-video/evaluator/ctc_align.py` (bug fix)                                                                                                                                                                                                                                                                                                                                                                                                                                                        | 182.C.18.01 checks the letters the recognizer WROTE, which is exactly where a pronunciation fault hides: a recognizer's language model repairs a mangled brand name into the spelling it expects and the transcript comes back clean. So this asks whether the ACOUSTICS support the canonical form, using Witt & Young goodness of pronunciation over the same CTC posteriors - GOP = log P(canonical                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | emission frame) - max_s log P(s                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | emission frame), at most 0, where 0 means the model would itself have chosen that character. Measured on two clips: every correctly transcribed word on BOTH scored exactly 0.0, and substituting one word for a phonetic near neighbour against the same audio dropped it to -9.62 (costs->coasts), -11.17 (forty->fourteen), -15.32 (tuesday->thursday), -20.18 (engine->enjoying) while every other word stayed at exactly 0.0, so the fault localises to a word and to a character in it. THE FINDING THAT CONSTRAINS USE: the measure SATURATES. Two voices, two sentences, all at exactly 0.0 - there is no headroom above zero to rank a perfect pronunciation against an accented but acceptable one. It is a detector, not a grade, and both fixtures are clean synthetic English so they do not establish a threshold for real human speech at all. The floor ships at -3.0 labelled provisional, placed nearer the correct end deliberately because a false 'mispronounced' on a real accent rejects a take a human would accept, and `assertThresholdCalibrated` REFUSES so it cannot reach a release proof reading as calibrated; calibration is blocked on the 182.C.33 rated corpus, same blocker as 182.C.17.08. Locale is the other half: wav2vec2-base-960h is LibriSpeech English, so a Yoruba name it has no symbols for scores far below any floor when said perfectly - those are reported UNMEASURABLE with minGop null (null rather than a sentinel, because a sentinel gets averaged) and routed to 182.C.18.07. Every declared term gets a finding including the unmeasurable ones, since a term absent from the report is a term nobody checked and the report is read as coverage; `assertEnoughMeasured` refuses a report where most terms could not be measured, because zero faults from checking nothing is the failure this module exists to prevent. Acronyms carry accepted spoken forms ('FBI' as 'f b i'), a multi-word term scores by its WORST word rather than its average, and lookup goes through the contract's normalizer so '42' finds the aligned 'forty two'. THIS WORK FOUND A DEFECT IN THE ALIGNER COMMITTED HOURS EARLIER (182.C.18.02): tokenize silently dropped characters outside the model vocabulary, so aligning 'costs 40 dollars' against audio saying 'costs forty dollars' dropped both digits, put '40' 463 ms late, and LOST 'dollars' from the output entirely. It now refuses with unalignable_character, and merge_words groups by the tokenizer's own labels instead of counting transcript characters in parallel - the two walks agreeing was the assumption that broke. The 18.02 calibration onsets are identical after the rewrite and its 23 tests still pass. | Real GOP scoring run in the spec (CPU): thursday minGop -15.3229 reproduced, all other words exactly 0.0; `pronunciation-gop.json`. 19 tests pass, 68 across the 18.x family. No provider spend.          | —   | Threshold is provisional, not calibrated: needs the 182.C.33 rated corpus and real accented speech. `assertThresholdCalibrated` refuses until then. |
| 182.C.18.04 | Detect clipping, silence, low SNR, pumping, phasing, duplicate audio, channel inversion, excessive noise …             | MLOps | 182.C.02.12 | implemented    | `apps/oshun/bff/src/generation/human-video-audio-defects.ts` (new), `human-video-audio-defects.spec.ts` (new), `tools/human-video/evaluator/audio_defects.py` (new), `tools/human-video/evaluator/make_defect_fixtures.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-audio-defects/` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Nine defects, measured against twelve clips derived from the committed clean fixture by arithmetic on REAL SAMPLES - clipping is gain into the ceiling, phasing is one channel delayed 0.375 ms, the upsample is a real anti-aliased decimation and interpolation. A fixture that merely sets the flag a detector reads proves the detector reads the flag. The clips are not committed; a seeded generator reproduces all twelve byte-identically (verified by SHA-256) and 25 kB of measurements is a better record than a dozen wavs nobody can diff. Eight of the nine separate cleanly: clipping 0-sample runs vs a 15-sample run, silence 0.599 s word gaps vs a 1.200 s hole, SNR 81.3 dB vs 11.9 and 3.0, duplicate 0.261 vs 1.000, inversion correlation -1.000 with the mono sum down 73.7 dB, phasing a -6 sample delay at correlation 1.000 with its first null at 1333 Hz, band limit 8000 Hz vs 4426. THE CLEAN CLIP PASSES EVERY GATE - the most important assertion in the spec, because a detector that fires on the good take teaches people to ignore it. THREE DESIGNS WERE BUILT AND MEASURED AND FAILED, and all three are recorded so nobody rebuilds them. (1) Duplicate detection on the ENVELOPE scored the clean clip 0.813 and the duplicated one 0.795 - rhythmic speech correlates with itself at the word period and beats a real duplicate; rebuilt on the waveform. (2) Phasing band-by-band reported 53 'notches' spaced at the grouping width rather than the comb's real nulls, because speech harmonic structure puts plenty of bins below -6 dB by chance; rebuilt as inter-channel cross-correlation, which is the standard measurement. (3) PUMPING - three designs against a real 20:1 limiter fixture and its uncompressed control, and NONE separated them: envelope-vs-floor correlation 0.495 against 0.536, gap ramps scattered in both, max gap climb 1.17 dB against 2.87 dB which is BACKWARDS. Design 1 also had the physics sign wrong (a compressor ducks during speech and recovers after, so pumping is the background swelling in the GAPS). So pumping is measured and NOT gated: PUMPING_VALIDATED is false, assertPumpingGateable refuses, gateAudio will not fail a take on it, and the numbers still reach a human through reportedNotGated because a finding suppressed silently is a finding nobody sees. Time-stretch is likewise not claimed - a phase vocoder's transient smearing has no signature separable from a soft consonant, the RESAMPLE half IS detected, and the speaking-rate change a stretch causes is 182.C.18.02's slope. SNR is the percentile-split estimate, never called a true SNR, and every finding says so in its own text. Thresholds live in TypeScript and quantities in Python, because a measurement that has already applied a threshold cannot be re-thresholded when the threshold turns out wrong - and one of them did.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | 12 fixtures regenerated and re-analysed inside the spec (1.8 s): peak, SNR, duplication and band limit all reproduced; `evidence/2026-09-10-audio-defects/measurements.json`. 22 tests pass. No provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Pumping is measured but not validated: no design separated a real compressed fixture from its control. Needs the 182.C.33 rated corpus. Time-stretch artifacts are caught by 182.C.18.02 slope, not here. |
| 182.C.18.05 | Verify loudness and true peak against delivery profiles without hiding provider speech defects through aggressive …    | MLOps | 182.C.02.12 | implemented    | `apps/oshun/bff/src/generation/human-video-loudness.ts` (new), `human-video-loudness.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-10-loudness-conformance/` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | The measuring already existed (human-video-media-facts.ts runs ebur128=peak=true). What this adds is the second half of the item's sentence, and that half is not a footnote - it is measured and it is worse than it sounds. THE CLIPPED FIXTURE ARRIVES WITH A 15-SAMPLE RUN OF SAMPLES PINNED AT FULL SCALE AND A TRUE PEAK OF +0.9 dBTP; AFTER loudnorm=I=-16:TP=-1 THE CLIPPING DETECTOR MEASURES A LONGEST RUN OF ZERO. The damage is untouched - scaling down does not restore waveform tops a converter discarded - but nothing downstream can see it, so a conform step launders a clipped take into one that passes the clipping gate. Hence assertDefectsMeasuredFirst, which refuses a 182.C.18.04 report produced after ANY processing step, including one labelled 'as-received' that lists steps (the history wins over the label). LOUDNESS_HIDES names clipping as hidden and the other seven as surviving, WITH the measurement quoted: the very-noisy fixture measured 2.98 dB SNR before and 3.06 after, because scaling moves speech and noise together. Saying which defects survive matters as much as saying which do not - a rule that cries wolf about all nine gets turned off. planLoudness reports the gain needed AND whether the target is reachable by gain alone, because gain is transparent and limiting is not; assertConformableWithoutLimiting refuses a take that only conforms through limiting, since the profile is a delivery specification and not a licence to reshape a bad take until it fits. Every fixture fails that check, which is the right answer: they are quiet enough that the gain to reach -16 LUFS breaches the ceiling, so they go back rather than into a limiter. assertNotAlreadyOverProcessed catches a take that ARRIVES squashed at a provisional 13 dB crest floor - under every clean measurement, over the clipped fixture's 11.6, and under our own conform output's 17.3 so it never rejects what this pipeline produced. Dynamic range is the cost every take pays: the clean fixture's crest fell 20.8 -> 17.3 dB and its LRA 1.9 -> 1.6 LU purely from meeting the target. True peak is refused when unmeasured rather than substituted with sample peak, because the difference is where inter-sample overs live.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | 6 clips measured with ebur128 before and after loudnorm, and re-analysed for defects; `evidence/2026-09-10-loudness-conformance/measurements.json`. 21 tests pass. No provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | None                                                                                                                                                                                                      |
| 182.C.18.06 | Define exact-audio mode, exact-words mode, and creative- dialogue mode separately …                                    | MLOps | 182.C.02.12 | implemented    | `apps/oshun/bff/src/generation/human-video-dialogue-fidelity-mode.ts` (new), `human-video-dialogue-fidelity-mode.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | human-video-dialogue-mode.ts already split takes two ways - a mode either promises the approved words or it does not - and that split is right and not enough, because it puts two very different promises on the same side. exact-audio means the approved RECORDING plays, sample for sample. exact-words means the model SPOKE the approved script: the words are guaranteed and the voice, the pronunciation and the timing are the model's. Collapsing them into 'exact' tells someone who needed their own recording that they got it. Each definition carries what it does NOT promise, written out, because the part that surprises people is never in the promise - exact-words explicitly disclaims the voice, pronunciation (182.C.18.03), timing (182.C.18.02) and nonverbal sound. THE ACHIEVED MODE IS THE POINT: what was requested and what was delivered are different questions and the delivery can only ever be equal or WORSE. An exact-audio request whose provider re-encoded the track achieves exact-words at best - the samples are not ours any more - and if the words also changed it achieves creative-dialogue. An exact-words request whose delivered audio says something else is not a failed exact-words take that can be labelled as one; it is a take with unscripted words in it. Downgrade-only by construction, and assertNotUpgraded refuses a claim stronger than what was asked for, since that is a claim whose conditions nobody measured. Timing faults deliberately do NOT downgrade: exact-words promises the WORDS and a stretched take still said them, so downgrading for stretch would make the label mean something its definition does not say - but unexpected-dialogue DOES downgrade, because words absent from the script are exactly what exact-words rules out. Unmeasured is refused rather than treated as achieved, with one deliberate exception: a bit-exact take needs no ASR to prove its words, because the words are in the samples we approved and asking a recognizer to confirm our own recording would be theatre. A transcript is mandatory in EVERY mode including creative-dialogue - it is tempting to skip ASR for the mode that promises no particular words and that reasoning is backwards. The release proof carries both modes and a one-sentence disclosure written for the person who asked for exact-audio and got exact-words, who needs to be told rather than to find it by comparing two fields.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | 22 tests pass. Pure policy over the 18.01-18.05 measurements; no new measurement and no provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | None                                                                                                                                                                                                      |
| 182.C.18.07 | Add multilingual ASR ensembles or human review fallback where a single recognizer is not validated for the requested … | MLOps | 182.C.02.12 | implemented    | `apps/oshun/bff/src/generation/human-video-asr-ensemble.ts` (new), `human-video-asr-ensemble.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | The honest state of recognizer validation is thin and this records it as thin: three models probed on 2026-09-10 against ONE English sentence with known ground truth, and RECOGNIZER_VALIDATION carries `sentences: 1` because 'validated' without a corpus size behind it is a word. MIN_SENTENCES_FOR_SINGLE_RECOGNIZER is 20, which nothing meets, so TODAY EVERY TAKE GOES TO AN ENSEMBLE - the honest reading of one probe sentence rather than a claim that one sentence validated anything. HOW AGREEMENT IS COMPUTED IS DECIDED BY THE PROBE: all three models heard the sentence perfectly, and on RAW STRINGS they split two-to-one ('$42' twice, '42 dollars' once). A string-voting ensemble reports a disagreement on a clip every model got right and escalates a correct take to a human. So agreement runs on NORMALIZED CONTENT WORDS through the contract's own normalizer, where all three are identical - and a function-word difference is ignored, since a recognizer dropping 'the' is a recognizer artifact far more often than a missing word. THE HARD STOP is that NORMALIZABLE_LOCALE_PREFIXES is ['en']. For a non-English take no recognizer has been probed AND the ensemble's agreement cannot be scored at all, so human review is not a fallback there - it is the only path, until either the normalizer gains rules for the language or a recognizer is probed against known ground truth in it. planRecognition checks scoreability FIRST, before recognizer count, because an ensemble in an unscoreable language is worse than no ensemble: it spends money to produce a number nobody can interpret. A disagreement is never resolved by silently taking the majority - the dissenter is named, because which model disagreed is what tells someone whether the take is bad or the recognizer is - and the human review request carries EVERY transcript including the dissenting one, since a reviewer handed only the majority text is being asked to rubber-stamp a decision already made. For an unscoreable language the reviewer is told their transcript IS the measurement. The ensemble is ordered cheapest-first and priced from the probe's own measured costs.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | The 3 recorded probe transcripts replayed as an ensemble: 2 distinct raw strings, 0 content-word errors across all 3 pairs. 18 tests pass. No new provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Non-English is human-review-only because the contract normalizer is English-only. Lifting it needs either per-language normalizer rules or a probed recognizer with known ground truth in that language.  |
| 182.C.18.08 | Test fast/slow speech, whisper, shouting, laughter, singing, emotional delivery, background music, accent variation …  | MLOps | 182.C.02.12 | blocked-corpus | `apps/oshun/bff/src/generation/human-video-delivery-conditions.ts` (new), `human-video-delivery-conditions.spec.ts` (new), `tools/human-video/evaluator/make_delivery_fixtures.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-delivery-conditions/` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | DELIBERATELY LEFT UNCHECKED. The item names nine phrases, which expand to ten conditions in code because 'fast/slow speech' is two and they behave differently. FOUR were measured with real fixtures and real evaluators; SIX could not be built at all. The four built are TRANSFORMATIONS of a signal - a tempo change, a mix, an offset copy - and each produces audio that genuinely has the property. The six are different PHONATION or different SPEAKERS: whispered speech is unvoiced, shouting changes formants and not only level, laughter and singing are not speech, and an accent is another person. Filtering a normal voice into 'a whisper' produces a quiet normal voice, and testing a detector against that proves the detector reads the filter. All six need recorded speech - the 182.C.33 corpus - and each is listed with the fixture it waits for. THE FOUR THAT WERE MEASURED FOUND TWO REAL HOLES IN WORK ALREADY SHIPPED. (1) TIME STRETCH PRODUCES FALSE MISPRONUNCIATION FINDINGS: the fast and slow clips contain identical words from an identical speaker, only the tempo changed, and 182.C.18.03's GOP falls to -4.99 at 1.25x and -10.20 at 0.80x, past its -3.0 floor - a phase vocoder's transient smearing looks like bad pronunciation to a frame-level acoustic model. assertGopMeasurable now REFUSES a pronunciation score on time-stretched audio; refuses rather than discounts, because a discounted wrong number is still a wrong number. The aligner held there (68.4 ms at 1.6x, well inside ITU +45/-125), so timing claims are still allowed. (2) OVERLAPPING SPEECH BREAKS THE ALIGNER SILENTLY: a second voice starting 0.5 s in gives a 513.8 ms onset error - four times the ITU reject boundary - and the aligner reports SUCCESS, because forced alignment is given the words and asked where they are and will always find them somewhere. At 0.5 s the GOP is exactly 0.00, so the take looks perfectly clean while its timing is half a second wrong; a clean-looking evaluator beside a broken one is how a bad take gets through. Both evaluators now refuse on it. Not a hole but worth knowing: a harmonic bed 5 dB under the speech leaves the aligner and the GOP COMPLETELY unaffected while the percentile-split SNR reads 16.8 dB and trips the 182.C.18.04 low-snr gate at 20 - music under speech is not noise and that estimator cannot tell the difference, so the finding carries a caveat rather than being suppressed. coverageReport() reports 4 of 10 and never rounds up: a report presenting 'the measured conditions all pass' as coverage is the exact claim this module exists to prevent.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | 9 fixtures built from the committed clip and run through the real aligner, GOP scorer and defect analyzer; `evidence/2026-09-10-delivery-conditions/measurements.json`. 17 tests pass. No provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | BLOCKED on the 182.C.33 rated corpus for whisper, shouting, laughter, singing, emotional delivery and accent variation. Checkbox stays unchecked at 4 of 10 conditions, same treatment as 182.C.17.08.    |

### 182.C.19 Identity, Anatomy, and Visual Integrity Evaluation

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner | Dependency | Status         | Changed Files                                                                                                                                                                                                                                                                                                                 | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Provider/Hardware Evidence                                                                                                                                                                                                                          | Cost | Blocker                                                                                                                                                                                                                                       |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | -------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.19.01 | Measure subject-specific identity similarity over visible frames and report temporal drift …                          | MLOps | 182.C.16   | implemented    | `libs/isis/workflows/src/consistency/identity-trajectory.ts` (new), `identity-trajectory.spec.ts` (new)                                                                                                                                                                                                                       | Built on `identity-drift.ts` rather than duplicating its cosine, adding the temporal layer the task asks for. The verdict is driven by the SUSTAINED WORST WINDOW, never the peak - a test drives the exact failure the task names, a take with one 0.99 frame and fifteen bad ones, and it fails. Also reports drift slope (a take whose identity decays is a different defect from one uniformly mediocre), the below-threshold ratio, and coverage. Coverage is first class: unmeasured frames are `null` with a reason, never dropped, and low coverage yields `unknown` rather than `pass` - a take nobody could measure has not passed. Per-frame similarity is the max over the enrolled references (a pack holds several angles); multi-subject takes report per subject and never average, so one preserved cast member cannot hide another. Threshold default is deliberately 0.5, NOT the inherited 0.35 that the SFace probe showed accepts two different people at 0.422. 23 tests pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Unit-tested; real crypto, no mocks                                                                                                                                                                                                                  | —    | None                                                                                                                                                                                                                                          |
| 182.C.19.02 | Evaluate face shape, complexion, age presentation, hair, distinctive features, body proportions, wardrobe …           | MLOps | 182.C.16   | implemented    | `apps/oshun/bff/src/generation/human-video-appearance.ts` (new), `human-video-appearance.spec.ts` (new), `tools/human-video/evaluator/appearance_attributes.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-appearance-attributes/` (new)                                                                       | Eight attributes, and they do not share a threshold. THE MEASUREMENT THAT DECIDES THE WHOLE DESIGN: complexion is dominated by lighting. On the committed two-subject frame the two foreground people differ by 63.7 degrees of Individual Typology Angle (-11.6 and +52.1) - visibly different people - and A ONE-STOP EXPOSURE CHANGE MOVES ONE OF THEM BY 58.2 DEGREES. Almost the entire between-person gap from nothing but the light; a modest white-balance shift moves it 10.0. So a complexion comparison across frames lit differently is a lighting comparison wearing a complexion's name, and assertLightingComparable refuses it. FACE SHAPE, by contrast, SURVIVES: across both white-balance shifts and both one-stop exposure changes, width-over-height spanned 1.92%, jaw width 0.6%, interocular 1.3%. Geometry survives light and colour does not, which is why they are gated differently - and blocking geometry across mismatched lighting would discard a working measurement to punish a different one, so it is not blocked. The lighting tolerance is calibrated from a ten-point sweep: about 2.4 degrees of ITA per unit of median luminance and 90-175 per unit of grey-world red gain, so matching luminance within 2 and red gain within 0.03 holds drift near 5 degrees - and inside those bounds ~5 degrees of any observed drift is STILL the light, so a complexion finding has to clear it. The 1.9231% face-shape spread is carried rounded UP to 1.93, because understating a noise floor is the direction that makes a gate fire on the light. TWO ATTRIBUTES ARE NOT MEASURED AT ALL and are routed to a person rather than approximated: age presentation, because no age model is staged and estimating age from face geometry has a documented record of skewing by ethnicity (a number would be worse than none); and distinctive features, because which scar or mole matters is a per-subject declaration and DISCOVERING them is a different and more invasive thing than checking a declared one. Two more measured facts shape the API: a subject's complexion measurement DISAPPEARED under brightening (unavailable at +0.1, +0.25 and +1.0 stops while present at +0.5), so 'not measured in this frame' is its own outcome never folded into 'no change' and never satisfying or violating a permitted-changes check; and of six detected faces only the two foreground subjects (relative height 0.163, 0.158) yielded anything, the four bystanders (0.033-0.068) producing neither geometry nor colour, which matches the 182.C.17 measurability floor. Colour uses median CIELAB rather than mean, since a specular highlight is a handful of extreme pixels a mean walks toward. The dense mesh is used ONLY for gross geometry, never asymmetry or lip curvature - 182.C.19.05 disqualified those.                                                    | The committed 2-subject frame through 5 lighting variants and a 10-point sweep, with YuNet, face-landmarker, selfie segmenter and pose-landmarker; `evidence/2026-09-10-appearance-attributes/measurements.json`. 28 tests pass. No provider spend. | —    | Thresholds are provisional; calibration against a labelled corpus is 182.C.19.08. Age presentation and distinctive features are human-review by design, not by blockage.                                                                      |
| 182.C.19.03 | Track identity through profile turns, partial occlusion, camera motion, lighting changes, cuts, interactions …        | MLOps | 182.C.16   | implemented    | `apps/oshun/bff/src/generation/human-video-identity-tracking.ts` (new), `human-video-identity-tracking.spec.ts` (new), `tools/human-video/evaluator/identity_tracking.py` (new), `tools/human-video/evaluator/make_tracking_sequences.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-identity-tracking/` (new) | Seven conditions, and under all of them one question: does the track that started on this person still belong to them. TWO FAILURES PRODUCE A WRONG ANSWER AND THEY NEED OPPOSITE RESPONSES, so they are measured apart. ASSOCIATION failure fragments - more tracks than people - and is SAFE, since a broken track loses continuity and whoever picks it up starts again. IDENTITY failure SWAPS - the track continues and the person changes - and is dangerous, because the take reports the wrong person with full confidence. Six 24-frame sequences were built from the committed two-subject frame by real geometry and photometry (a moving crop is a camera move, a black rectangle is an occluder, an exposure ramp is a lighting change, a splice is a cut, two real face crops crossing is an interaction) and tracked with YuNet + SFace, associating by IoU ALONE on purpose - mixing the embedding into association would make the tracker paper over the very failure being measured. RESULTS: camera pan held identity (0 swaps, min cosine 0.889) while association fragmented into 49 tracks, because a pan moves a small face further than its own width per frame; zoom held both; lighting held through a -1 to +1 stop ramp and COST 0.25 of cosine (0.643 against 0.889-0.945), so a threshold fitted under even lighting is tighter than a lit scene can support; the cut correctly BROKE tracks rather than continuing across one; the crossing held with one track splitting at the overlap, the safe failure. OCCLUSION IS THE ONE THAT BREAKS: 2 identity swaps and a minimum cosine of 0.099, as the embedding drifted under the occluder until the nearest enrolled subject was the OTHER person and the track carried on with them - so identity assignment is refused under occlusion rather than trusted. THE FINDING THAT GOVERNS EVERY ASSIGNMENT: 47 of the 49 pan tracks matched NOBODY (highest mean cosine 0.243) and every one was still handed a subject name, because nearest-neighbour always has a nearest. An assignment with no bound calibration names somebody for every face in frame; same conclusion 182.C.07.03 reached, same remedy - return unassigned, three-way with an `uncertain` band between the calibration's two operating points, and let 182.C.19.08 fit the cutoff. Enrolling all six detected faces instead of the two above the 0.10 relative-height floor changed swap counts in BOTH directions (occlusion 2->5, zoom 4->2) because unusable bystander embeddings absorb ambiguous frames, so an enrolment below the floor is a phantom identity and is refused. Profile turns are NOT TESTED and not claimed: rotating a face away from camera is a 3D transformation of a head and a still gives no geometry to rotate; a sheared frontal face is a warped frontal face. Embeddings are biometric and never committed - only the cosines. | 6 sequences x 24 frames through real YuNet + SFace tracking; `evidence/2026-09-10-identity-tracking/measurements.json`. 23 tests pass. No provider spend.                                                                                           | —    | Profile turns need 182.C.33 recorded footage. The identity cutoff itself is 182.C.19.08; this module refuses without a bound calibration rather than guessing.                                                                                |
| 182.C.19.04 | Detect face merging, identity swaps, duplicate people, flicker, sudden age/skin-tone shifts …                         | MLOps | 182.C.16   | implemented    | `apps/oshun/bff/src/generation/human-video-visual-integrity.ts` (new), `human-video-visual-integrity.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-10-visual-integrity/` (new)                                                                                                                              | Six detections, and what each one's SIGNATURE is was measured on real face crops rather than assumed. THE FINDING THAT MATTERS MOST IS NOT ONE OF THE SIX: blur destroys IDENTITY long before it destroys DETECTION. At Gaussian sigma 8 the detector is 0.879 confident there is a face and that face's cosine against its own unblurred self has collapsed to 0.033 - random; at sigma 12 it is still 0.792 confident and the self-cosine is -0.103, NEGATIVELY correlated. A pipeline that gates on detection score feeds noise into identity comparison and gets a confident wrong answer, so assertEmbeddingUsable runs before all six detections and takes a REFERENCE cosine rather than the detection score. FACE MERGING: a 50/50 alpha blend of two real faces embeds at 0.341 and 0.320 - neither person and EQUALLY unlike both - against 0.882/0.251 and 0.210/0.828 at the unblended ends. So the signature is a low best cosine AND a near-zero margin, in conjunction: a low cosine alone is an unknown face and a small margin alone is two strangers who look alike, and the spec asserts both non-cases. DUPLICATE PEOPLE: the same crop twice in one frame gives two SIMULTANEOUS detections at mutual cosine 0.9565 against 0.3243 for two different people; simultaneity is what makes it a fault, since the same person across two frames is the point of tracking. FLICKER: reproduced at the detection boundary (blur sigma 12 detected / 16 not) flipping on 11 of 11 transitions, and it did NOT reproduce by size - YuNet held a 16-pixel face and never dropped one between 56 and 40 pixel patches - so flicker here is a marginal-QUALITY phenomenon, not a marginal-size one, and it is measured as a flip RATE so a long sequence does not accumulate flips unfairly. SKIN-TONE SHIFT: a +40.9 degree ITA move under CONSTANT lighting left the identity cosine at 0.954, so SFace is essentially blind to skin tone and an identity check will NOT catch one - independent detections, both needed - and a tone comparison across mismatched lighting is refused because 182.C.19.02 measured one stop of exposure moving a subject 58.2 degrees. IDENTITY SWAPS are detected under 182.C.19.03. INHERITED MOUTH OR VOICE is split honestly: the mouth half is already built in human-video-speaker-binding.ts, and the voice half needs speaker embeddings or diarisation which is not staged, so it is reported not-measured. AGE SHIFT is likewise not measured. Both gaps travel IN the report rather than in a comment, because a report listing clean detections reads as clean whether or not two more were never attempted.                                                                                                                                                                                                                                              | Alpha-blend sweep, duplicate composite, blur sweep and tone shift on real face crops with YuNet + SFace; `evidence/2026-09-10-visual-integrity/measurements.json`. 27 tests pass. No provider spend.                                                | —    | Age shift needs an age model (deliberately absent, see 182.C.19.02). The voice half of "inherited mouth or voice" needs speaker embeddings, not staged.                                                                                       |
| 182.C.19.05 | Detect malformed teeth/tongue/lips, jaw discontinuity, asymmetry spikes, face boundary tearing, extra limbs/digits …  | MLOps | 182.C.16   | blocked-corpus | `apps/oshun/bff/src/generation/human-video-frame-integrity.ts` (new), `human-video-frame-integrity.spec.ts` (new), `tools/human-video/evaluator/frame_continuity.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-frame-continuity/` (new)                                                                       | DELIBERATELY LEFT UNCHECKED at 1 of 7. ONE ARTIFACT IS NOW DETECTED and FOUR candidate approaches were measured and failed, and the failures are the more useful half - somebody will otherwise build them again. WHAT WORKS: face boundary tearing, by adjacent-line correlation. A tear is a DISPLACEMENT, so at the band's edges one row no longer continues into the next. On the committed three-panel probe the torn panel's worst adjacent-row correlation is 0.292 with 2 rows under 0.7, against 0.796 and 0 for BOTH the clean and the doubled-mouth panels, and it names row 168 - exactly where the band was displaced. A vertical tear built the same way scores 0.224 on the COLUMN axis at column 170 while its rows stay at 0.558, so the axis with the lower minimum names the orientation. The 0.7 threshold sits between 0.796 and 0.292 with room both sides, and TWO discontinuous lines are required rather than one, because a displaced band has two edges and a single bad line is as likely to be a hard cut inside the frame - a title card edge, a split screen. WHAT DOES NOT WORK, all measured: (1) SEGMENTATION CONSISTENCY, which is the ITEM'S OWN suggested candidate - all three panels give one connected face-skin component, a largest share of 1.000 and zero holes, and boundary roughness moves 33% for the tear but only 5% for the doubled mouth, inside the difference between two ordinary frames. (2) COPY-MOVE BLOCK MATCHING, on the idea that a doubled mouth is a pasted copy - all three panels return an IDENTICAL best distant-block correlation of 0.971 on an IDENTICAL coordinate pair and none has a pair over 0.99, so it is not a pixel copy and block matching cannot see it. (3) and (4) dense face-mesh asymmetry and lip curvature, disqualified earlier the same day when the mesh scored the doubled mouth as MORE REGULAR than the clean face. So a malformed mouth remains undetected after four attempts and the module says so: assertDetectable refuses and carries every disqualification into the refusal message. Jaw discontinuity is kept separate from a soft malformation on purpose - a hard displacement would show in continuity and a soft one would not, and conflating them would let a passing continuity check read as coverage. coverageReport() reports 1 of 7 and never rounds up: a report saying 'no artifacts found' over six undetected artifact classes is the claim this module exists to prevent.                                                                                                                                                                                                                                                                                                                                                                                                                  | Line continuity, segmentation consistency and copy-move all run on the committed 3-panel probe plus a vertical tear; `evidence/2026-09-10-frame-continuity/measurements.json`. 15 tests pass. No provider spend.                                    | —    | BLOCKED on 182.C.33 for a trained artifact detector: four hand-built approaches have now been measured and rejected for the malformed-mouth case. Fused props and impossible contact need scene understanding, which nothing staged provides. |
| 182.C.19.06 | Evaluate motion smoothness, temporal consistency, optical flow outliers, frame duplication, freeze, judder, warping … | MLOps | 182.C.16   | implemented    | `apps/oshun/bff/src/generation/human-video-motion-quality.ts` (new), `human-video-motion-quality.spec.ts` (new), `tools/human-video/evaluator/motion_quality.py` (new), `tools/human-video/evaluator/make_motion_sequences.py` (new), `docs/domains/isis/human-video/evidence/2026-09-10-motion-quality/` (new)               | Eight faults, seven separating cleanly on seven 16-frame sequences built from the committed frame by real geometry - a moving crop window, a copied buffer, a remap - with dense Farneback flow. TWO FINDINGS WHERE THE OBVIOUS APPROACH POINTS THE WRONG WAY, and both changed the design. (1) JUDDER HAS MORE IDENTICAL FRAMES THAN A FREEZE: 9 identical pairs against the freeze's 5, because 3:2 pulldown repeats frames all the way through while a freeze holds one. A detector counting identical pairs would call judder the worse freeze. What separates them is the longest CONSECUTIVE run - 2 for judder, 5 for the freeze - and a freeze is in turn not a STATIC shot, which has an even longer run of 15; the discriminator there is the motion AROUND the run, flow standard deviation 8.578 against 0.000. (2) WARPING IS INVISIBLE UNDER A PAN. Comparing each pixel's flow to the FRAME's median was built first and fails: a 6-pixel warp raised global incoherence 2.6% and LOWERED the outlier fraction, because a subject and a background at different depths already disagree with the frame median. Comparing to a LOCAL neighbourhood mean is the right measure and it is decisive on a still camera - 0.00041 with no warp, 0.01078 at 2 pixels (26x), 0.03325 at 6 pixels (81x) - and swamped under a pan (1.170 against 1.215, 4%). So mean flow above 1.0 disqualifies a warp claim, detectWarping REFUSES on a moving shot rather than returning an uninterpretable number, and that refusal becomes an entry in the report's `unmeasured` list, because a report with six clean detections and a silently skipped seventh reads as seven clean detections. The rest: smoothness is the SECOND difference of flow magnitude so a fast steady move is smooth (pan max acceleration 1.025 against 18.48 for freeze and judder); background instability needs the two regions measured apart (11.4x against the pan's 0.88) or a moving camera looks like an unstable background; and flow outliers use an ABSOLUTE threshold - a pixel moving more than a pixel differently from its neighbours - because a multiple of the frame's own median flagged 42-45% of every sequence when it was tried.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | 7 sequences x 16 frames through dense Farneback flow, plus a 4-point warp sweep on a static camera; `evidence/2026-09-10-motion-quality/measurements.json`. 19 tests pass. No provider spend.                                                       | —    | Warping is measurable only below 1.0 mean flow; above that it is refused rather than guessed. Lifting that needs motion compensation before the flow comparison.                                                                              |
| 182.C.19.07 | Evaluate camera/control adherence, action completion, blocking, gaze, emotion, gesture, framing …                     | MLOps | 182.C.16   | implemented    | `apps/oshun/bff/src/generation/human-video-direction-adherence.ts` (new), `human-video-direction-adherence.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-10-direction-adherence/` (new)                                                                                                                     | Eight aspects: five are geometry and are measured, three are JUDGEMENTS and are refused rather than dressed up as measurements. The noise floor came from the SAME real face in five framings - full, 0.7x, 1.4x, a tight crop and a 30-pixel shift - so every difference between readings is noise and nothing else. TWO FINDINGS SHAPE THE API. (1) BLENDSHAPES ARE NOT EQUALLY RELIABLE: the smile blendshape moves 0.014 across pure framing changes and browDown moves 0.138, ten times as much, so an emotion inference leaning on brow position would be reading the framing. Each blendshape carries its own measured stability, a claim smaller than that blendshape's noise is refused, and one whose stability was never measured is ALSO refused - an unmeasured stability is not a good one. (2) GAZE HAS A LARGE PER-FACE BIAS AND SMALL NOISE: on a face looking straight at the camera the two eyes read 0.430 and 0.615, a 0.185 asymmetry, STABLE across all five framings (spreads 0.023 and 0.010). It is a property of the face and the perspective, not of where the person is looking, so gaze measures CHANGE well - the bias cancels between two readings - and absolute direction not at all without a per-subject baseline, which absoluteGazeDirection refuses to proceed without. Head pose is repeatable to about 1.5 degrees in yaw and pitch and 0.3 in roll: good enough to gate a profile turn, not a subtle glance. A pan or tilt is decidable from the median flow's dominant axis, remembering the IMAGE moves opposite to the camera - getting that sign wrong turns a compliant take into a violation, so it is written down rather than left to the reader - but a PUSH-IN OR PULL-OUT IS NOT: a zoom moves pixels outward from the centre and the median of that is near zero, the same as a static shot, so assertMoveDecidable refuses rather than guessing. Framing is plain geometry (shot size from the face's share of frame height, headroom, distance from the thirds). The three judgements - emotion, gesture, action completion - travel in EVERY report, because a report showing framing and camera adherence clean reads as 'the direction was followed' unless it also says nobody checked the rest.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | One face x 5 framings through face_landmarker (478 landmarks, 52 blendshapes, transform matrix); `evidence/2026-09-10-direction-adherence/measurements.json`. 24 tests pass. No provider spend.                                                     | —    | Emotion, gesture and action completion are human judgements by design. Push-in/pull-out adherence needs flow divergence, a measurement not built here.                                                                                        |
| 182.C.19.08 | Calibrate identity and artifact thresholds by visibility, shot size, motion blur …                                    | MLOps | 182.C.16   | blocked-corpus | `apps/oshun/bff/src/generation/human-video-threshold-calibration.ts` (new), `human-video-threshold-calibration.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-10-threshold-calibration/` (new)                                                                                                               | STILL UNCHECKED - fitting the operating point needs the 182.C.33 labelled corpus and that has not changed. What does NOT need it, and is now measured and shipped, is how far each condition MOVES the score: one real face embedded with SFace and measured AGAINST ITSELF under each condition, with two different people from the same frame giving the scale at 0.2817. THE FINDING: TWO CONDITIONS PUT A FACE BELOW THE DIFFERENT-PERSON BASELINE AGAINST ITSELF. At 31 pixels of motion blur the self-cosine is 0.2509 and under an edge-preserving stylization it is 0.2674, both under 0.2817 - a person's own face resembles itself LESS than two strangers resemble each other. No cutoff separates that, because a threshold low enough to admit the blurred self-match admits every stranger too; it is not 'the threshold is not yet fitted', it is 'the decision is impossible', and assertDecidable refuses rather than returning one. Conditions do not compensate either - a face at 31px of blur is undecidable however well lit - so the WORST condition present governs. Everywhere else the measured retention is what a fitted cutoff must be adjusted by: 95% at -1 stop, 76% at +1 stop, 90% at a face 14% of frame height falling to 24% at 4.4%, 87% at 5px of blur falling to 4% at 21px, 88% for grayscale falling to 48% for posterization. Lighting being the mild one is worth saying because 182.C.19.03 measured a lighting RAMP costing 0.25 of cosine - the difference is that the ramp compounded with framing change across a sequence, and a stop on fixed framing does not. adjustForConditions REFUSES a base that was never fitted on labels: scaling an unfitted number by a measured factor produces a number that LOOKS calibrated and is not, which is exactly what the inherited 0.35 gate was when SFace put two strangers at 0.422. AGAINST_UNIVERSAL_CUTOFF carries five independent measurements from this session, each with its evidence directory, because a reader asked to accept per-condition calibration deserves to see why rather than take the item's word.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | 22 condition levels x one real face against itself, SFace cosines; `evidence/2026-09-10-threshold-calibration/measurements.json`. 19 tests pass. No provider spend.                                                                                 | —    | BLOCKED on 182.C.33 for the operating point itself. The per-condition adjustment factors are measured and shipped; the absolute cutoff they adjust is not.                                                                                    |
| 182.C.19.09 | Add fairness slices for dark skin, varied lighting, facial hair, head coverings, makeup, glasses, age …               | MLOps | 182.C.16   | blocked-corpus | `apps/oshun/bff/src/generation/human-video-fairness-slices.ts` (new), `human-video-fairness-slices.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-fairness-availability/` (new)                                                                                                                           | STILL UNCHECKED - populating eight slices needs subjects, which is 182.C.33. What is built is the machinery that makes a corpus report anything useful, plus ONE measurement showing why it has to exist. TWO SUBJECTS from the committed frame, differing by 63.7 degrees of ITA, through nine exposure levels: both were DETECTED in all nine and both had their face GEOMETRY measured in all nine, and what differed was the COMPLEXION measurement, which depends on the face-skin segment having enough pixels - the lighter subject lost it in four conditions and the darker one in one. AGGREGATES HIDE THIS: pooled, complexion succeeded 13 of 18 times, 72%, a number describing neither subject (they were 89% and 56%), so assertSliced refuses an aggregate carrying no slices. AVAILABILITY IS THE THING TO SLICE, not just accuracy: every gate in this phase distinguishes 'measured and passed' from 'not measurable', so a subject whose attribute is more often unmeasurable gets LESS scrutiny, not more - a fairness fault with no wrong answer in it anywhere, invisible to any report that only tracks pass rates, which is why availability and passRateWhereMeasured are reported APART (a bucket where nothing was measurable has a pass rate of nothing-over-nothing, printing as a blank or a 100% depending on who wrote the formatter). AND THE DIRECTION WAS NOT THE EXPECTED ONE: the commonly-cited failure is dark skin under-detected in low light, and here the LIGHTER subject lost more measurements under BRIGHTENING. Two subjects prove nothing about a population and SEED_OBSERVATION.isFairnessEvaluation is false in code; MIN_SUBJECTS_PER_BUCKET is 30, deliberately larger than anything available, so no bucket is powered today and assertBucketPowered refuses a rate from any of them. Skin tone and lighting are MEASURED slices; the rest are DECLARED, because a facial-hair or makeup classifier would be a second uncalibrated model and age is declared and never estimated for the reason 182.C.19.02 records. Facial or motor differences are declared by the SUBJECT and by nobody else, and the note says why the slice matters: every artifact detector in this phase treats asymmetry and irregular motion as faults. A report is never complete while any slice is missing OR has no powered bucket - a report of clean slices, none powered, is a report of nothing having been measured.                                                                                                                                                                                                                                                                                                                                                                                                                                                      | 2 subjects x 9 exposure levels through the appearance analyzer; `evidence/2026-09-11-fairness-availability/measurements.json`. 22 tests pass. No provider spend.                                                                                    | —    | BLOCKED on 182.C.33 subjects. Six of eight slices are declared rather than measured by design; no bucket is powered today and the module refuses to report a rate from any.                                                                   |
| 182.C.19.10 | Store only necessary embeddings with encryption, tenant isolation, deletion linkage, access audit …                   | MLOps | 182.C.16   | implemented    | `apps/oshun/bff/src/generation/human-video-embedding-store.ts` (new), `human-video-embedding-store.spec.ts` (new)                                                                                                                                                                                                             | All five properties implemented rather than promised. ENCRYPTION: AES-256-GCM; a tampered ciphertext fails its auth tag instead of yielding garbage. TENANT ISOLATION: per-tenant HKDF key from one master secret, and a test bypasses the scope guard entirely by rewriting the stored tenant - decryption still fails, so only the cryptography is load-bearing. PROHIBITED REUSE: project, subject and purpose are bound as AAD, so cross-project reuse is not a policy branch a caller can skip - it fails arithmetically, and a test proves that with the guard bypassed too. DELETION LINKAGE: every record carries its authorizationId, so revocation names a consent and gets exactly its records back, with nothing readable afterwards. ACCESS AUDIT: store, read, delete AND refusal, with the actor - refusals are logged because that is the entry an investigation actually wants. MINIMIZATION: per-subject-per-project cap, and the record has no field for a source image. 24 tests pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Unit-tested; real crypto, no mocks                                                                                                                                                                                                                  | —    | None                                                                                                                                                                                                                                          |

### 182.C.20 Active-Scene and Multi-Speaker Evaluation

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner | Dependency            | Status         | Changed Files                                                                                                                 | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Cost                                                                           | Blocker                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | ----- | --------------------- | -------------- | ----------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.20.01 | Build active-scene tests with walking, turning, hand/prop interaction, sitting/standing, camera tracking …             | MLOps | 182.C.17, 182.C.26.02 | blocked-corpus | `apps/oshun/bff/src/generation/human-video-active-scene-cases.ts` (new), `human-video-active-scene-cases.spec.ts` (new)       | "WHILE SPEAKING" IS THE LOAD-BEARING HALF OF THIS ITEM AND THE EASIEST TO DROP. A walking case where the subject walks, stops and THEN delivers the line tests walking and tests speech and tests NEITHER OF THE THINGS THIS PHASE IS ABOUT, because every frame the evaluator cares about is one where motion and speech coincide - the clip looks right, the label is accurate, and the cell it fills is empty. So each case carries the overlap between its speech window and its condition window, measured OVER THE SPEECH (ten seconds of walking with one second spoken over it is a one-second test, whatever the walk duration suggests). AND A CONDITION IS ONLY A TEST IF IT CAN FAIL DIFFERENTLY FROM THE OTHERS: walking and camera-tracking both arrive at the evaluator as MOTION BLUR, so in the same band they are ONE TEST WEARING TWO NAMES - the six conditions stress only FOUR distinct mechanisms, and the matrix is narrower than it looks. That matters because 182.C.34.07 prices a cohort at 299 clean renders and nobody pays that twice for the same measurement by accident. THE CELL NOBODY COULD FILL IS THE ONE THAT GETS QUOTED AS COVERED, so empty cells are listed BY NAME rather than folded into a fraction - a coverage figure over the cells that have clips in them is arithmetic about the corpus, not about the system. Shot sizes are 182.C.17.06 own five, not a second list, because a case at a size the evaluator cannot score is a cell that can never be filled. The full 6x5 grid prices at $17,940 in renders alone (30 cells x 299 x 8 s x $0.25/s), which is the argument for all-pairs over the cross product.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | 14 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. The cell count is asserted as ACTIVE_CONDITIONS x `SHOT_SIZES` from 182.C.17.06 rather than typed, the $17,940 is recomputed in the test from `CLEAN_RENDERS_PER_COHORT` and `USD_PER_DELIVERED_SECOND`, the four-mechanism count is derived from the map, and the walk-then-talk case is asserted at EXACTLY 0% concurrency before the refusal fires.                                                                                                           | No provider spend; the $17,940 figure is the UNSPENT cost of filling the grid. | BLOCKED on 182.C.33 subjects and clips - THE CHECKBOX STAYS UNCHECKED. What exists is the machinery that says whether a corpus covers the conditions; 29 of the 30 cells are empty because there is no corpus, and nothing in this file can substitute for one. Same position 182.C.19.09 took for fairness slices. The concurrency floor (50% of the line) is PROVISIONAL and catches the sequenced case rather than characterising how much is enough. |
| 182.C.20.02 | Measure sync and identity during motion rather than only cherry-picking stationary frontal frames.                     | MLOps | 182.C.17, 182.C.26.02 | implemented    | `apps/oshun/bff/src/generation/human-video-motion-sampling.ts` (new), `human-video-motion-sampling.spec.ts` (new)             | THE MEASURABILITY FILTER IS THE CHERRY-PICK. 182.C.17.06 excludes windows where the face is too small, turned away or absent, and it is right to - "no findings" reaching a release gate looking like "no problems" is the worst failure in the stack. But the windows it excludes are exactly the ones with movement in them, so what survives is SYSTEMATICALLY THE EASY HALF AND NOBODY HAD TO CHERRY-PICK ANYTHING FOR THAT TO HAPPEN. The number is not "sync", it is "SYNC GIVEN THE FACE WAS FRONTAL AND STILL", and the two get written down identically. THE BIAS RUNS ONE WAY, WHICH IS THE ONLY USEFUL THING ABOUT IT: excluded frames are harder, never easier, so the aggregate is an UPPER BOUND ON THE CLIP RATHER THAN AN ESTIMATE OF IT - and a bound is a real result while a mean presented as an estimate is not. Worked on a 20-frame clip where half the frames carry movement and the filter passes none of the fast band: the same band scores read 0.91 over the scored frames and 0.75 under the clip own motion mix - SIXTEEN POINTS OF DIFFERENCE FROM THE WEIGHTS ALONE - and with the fast band unmeasured there is nothing to weight at all. SO SCORES GO IN BANDS, NOT IN A MEAN, because a single number over a skewed sample cannot be corrected afterwards - the information needed to correct it is what the exclusion threw away. The band boundaries are 182.C.19.08 measurements reached through 182.C.09.09 thresholds (21 px weak, 31 px undecidable) rather than round numbers. AND PAST 31 PX IDENTITY IS NOT LOW, IT IS UNDECIDABLE: a face scored 0.251 against its own clean self, under the 0.282 two-different-people baseline, so recording "identity failed" there states a finding the pixels cannot support - the same fail-versus-unmeasurable distinction 182.C.17.06 drew for sync, arriving from the identity side.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | 14 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. Band boundaries are asserted EQUAL to `WARNING_THRESHOLDS.motionBlurWeakPx`/`motionBlurUndecidablePx` rather than retyped, so a change to 182.C.19.08 moves the bands instead of leaving them behind; the naive 0.908 and clip-weighted 0.750 are both asserted to three places in the test; and the unmeasured-band case is asserted to return null rather than a number.                                                                                       | No provider spend; no model call.                                              | NO REAL CLIP HAS BEEN THROUGH THIS. Inputs are per-frame motion blur and per-band scores, which come from the evaluator wiring that does not exist yet. The bound cannot be turned into an estimate by anything here - closing that gap needs measurements in the moving bands, which is 182.C.20.01/.07 challenge material and 182.C.33 corpus.                                                                                                         |
| 182.C.20.03 | Detect whether the correct person's mouth moves for every dialogue segment and whether non-speakers remain naturally … | MLOps | 182.C.17, 182.C.26.02 | implemented    | `apps/oshun/bff/src/generation/human-video-speaker-attribution.ts` (new), `human-video-speaker-attribution.spec.ts` (new)     | THE MOUTH THAT MOVED MOST IS NOT THE MOUTH THAT SPOKE. Chewing, smiling and a yawn all move a mouth, so the measurement is CORRELATION WITH THE SPEECH ENVELOPE, not amplitude - demonstrated with a distractor whose mouth moves more than three times as much as the speaker and correlates at 0.07, which loses. AN ARGMAX ALWAYS RETURNS SOMEBODY, and two distinct failures wear the same face without a floor and a margin: nobody in frame is speaking, and two people both look like they are. Both refuse rather than answer. THE UNCOMFORTABLE PART IS THAT THIS METRIC REWARDS THE DEFECT THE ITEM ALSO ASKS US TO CATCH: a model rendering non-speakers as still images produces the CLEANEST POSSIBLE ATTRIBUTION, because nothing else in frame competes - measured here, a reacting listener correlating at 0.35 cuts the margin to 0.65 while a frozen one leaves it at 1.00. A perfect margin over a motionless cast is therefore TWO findings, and the one that reads as confidence is caused by the one that reads as a frozen cast. TWO CORRECTNESS FIXES THE TESTS FORCED. (1) A CONSTANT SERIES HAS NO CORRELATION AND ZERO IS THE WRONG WAY TO SAY SO - and `variance === 0` is the wrong guard, because a tracker reporting identical openness every frame leaves floating-point residuals and the exact test let a motionless face through at -1e-16, reading as "uncorrelated" and hiding exactly the face this module exists to find. The guard is a tolerance. (2) A NEGATIVELY-CORRELATED FACE MUST NOT INFLATE CONFIDENCE: the competing hypothesis is "that face is the speaker" and a negative correlation is no evidence for it, so the second-best is clamped at zero before the subtraction - left signed, a listener moving AGAINST the audio reported more confidence than no competitor at all.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | 18 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. Every fixture is a deterministic series and every claim is recomputed in the test: the distractor motion ratio, its 0.07 correlation, the listener 0.35, and the frozen-versus-reacting margin comparison. Both correctness fixes came from failing assertions rather than from review. Floor 0.5, margin 0.15 and non-speaker motion 0.01 are exported as PROVISIONAL - 182.C.17.08 is the calibration that replaces them and it is blocked on 182.C.33 corpus. | No provider spend; no model call.                                              | NO REAL CLIP HAS BEEN THROUGH THIS. The inputs are a per-face mouth-openness series and a speech envelope; producing them from delivered frames is the evaluator wiring, and the thresholds stay provisional until rated material exists (182.C.17.08, blocked-corpus). Non-speaker liveness is measured on MOUTH motion only - blink and head movement would strengthen it and need those signals from the tracker.                                     |
| 182.C.20.04 | Validate conversational turn-taking, interruptions, overlaps, reaction beats, eyelines …                               | MLOps | 182.C.17, 182.C.26.02 | implemented    | `apps/oshun/bff/src/generation/human-video-conversation-dynamics.ts` (new), `human-video-conversation-dynamics.spec.ts` (new) | OVERLAP BREAKS THE ATTRIBUTION 182.C.20.03 JUST BUILT, AND IT IS RIGHT TO. Two people speaking at once means one envelope and two mouths moving with it, so the correlation identifies BOTH and the attribution returns `ambiguous` - reproduced in the test by running the real `attribute()` on two identical mouth series. That is the correct answer to a question nobody should have asked: OVERLAP IS NOT AN AMBIGUITY, IT IS A SCENE THE SINGLE-ENVELOPE METHOD CANNOT EVALUATE, and filing it as a defect makes the ambiguous rate RISE WITH HOW NATURALLY PEOPLE TALK. The disposition turns on whether per-speaker audio exists, which is a fact rather than a judgement. TURN-TAKING IS IN THE GAPS AND THE VARIANCE MATTERS MORE THAN THE MEAN: two four-turn exchanges with IDENTICAL 300 ms mean gaps, one at zero spread and one at 406 ms including an interruption, and only the spread separates a conversation from a metronome. A REACTION IS A LAG, WHICH IS WHAT SEPARATES IT FROM A SECOND SPEAKER: 20.03 found this metric family rewards a frozen cast, and the opposite temptation is to accept any listener motion as liveness - but motion TIME-LOCKED TO THE LINE AS IT IS SAID is a mouth moving with the audio (a second speaker, a dubbing artefact, or an attribution about to go wrong). Cross-correlation recovers the +3 frame delay of a real reaction and the 0 of a simultaneous one. AND EYELINE CANNOT BE CHECKED IN THE SHOT IT BREAKS: two characters facing each other look in OPPOSITE screen directions, so a shot of A looking camera-left is correct or an error depending entirely on B shot - demonstrated by a violating pair where EACH SHOT ALONE PASSES and the pair does not. The convention is declared once per pair; an undeclared pair makes every per-shot check pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | 16 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. The overlap block calls 182.C.20.03 real `attribute()` rather than asserting what it would say; the equal-mean/different-spread pair is asserted to ten places; the reaction lag is recovered by cross-correlation over +/-6 frames using the same `correlation` function; and the eyeline block asserts the single-shot check PASSES on the violating shot before the cross-shot check fails it.                                                                | No provider spend; no model call.                                              | NO REAL CLIP HAS BEEN THROUGH THIS. Turns, per-face motion and per-shot screen direction all come from the evaluator wiring that does not exist. The gap-spread floor (80 ms) is PROVISIONAL and catches only the degenerate case; fitting it to natural speech needs 182.C.33 material. Per-speaker audio is the thing that would make overlap evaluable at all, and no route supplies it today.                                                        |
| 182.C.20.05 | Reject scenes where a speaking face is persistently hidden contrary to the shot plan or where cuts disguise failed …   | MLOps | 182.C.17, 182.C.26.02 | implemented    | `apps/oshun/bff/src/generation/human-video-scene-rejection.ts` (new), `human-video-scene-rejection.spec.ts` (new)             | Two evasions, and they are the same shape: a take that produces no FAILING measurement because it produced NO measurement. PERSISTENT HIDING is the scene-level question 182.C.20.06's per-window judgement does not ask - one unevaluable window is a moment, a cast member hidden through most of their dialogue is a take nobody can check. Both the FRACTION and the longest consecutive RUN gate, because six scattered seconds is a busy scene and six consecutive seconds is a person who never faced the camera while talking; hiding the PLAN asked for is not counted at all. CUTS DISGUISING FAILED SYNC is the half with a number that is ARITHMETIC rather than judgement, and it comes from the evaluator's own constants: score_syncnet.py scores over WINDOW_FRAMES=5 consecutive frames and searches DEFAULT_VSHIFT=10 frames either side at TARGET_FPS=25, so finding an offset needs 5 + 2x10 = 25 consecutive frames of ONE CONTINUOUS TRACK - one full second. A segment shorter than that cannot have its offset searched: it does not fail, it produces windows at the EDGE of the search range, which 182.C.17's windowsAtSearchEdge already flags and which a caller reading only the verdict never sees. So a take chopped into sub-second segments is one where every window passed because none COULD fail, and it looks identical to a well-cut scene; assertScorableLongEnough refuses it and names the arithmetic in the message. MIN_SCORABLE_SECONDS is computed from the three constants rather than typed, so a change to the evaluator moves it instead of leaving a stale number behind. The second signature - WHERE the cuts land - is REPORTED AND NOT GATED: cuts are part of filmmaking and a rule against them would be absurd, and while cuts landing systematically at the moment sync degrades is suspicious, a director cutting on a reaction beat produces the same pattern and nobody here has separated the two, so cutAlignment carries `gated: false` and the reason. judgeScene collects EVERY rejection rather than the first, because a take hidden against plan AND chopped too fine has two problems and fixing one leaves the other.                                                                                                                                                                                                                                                                                                                                            | 17 tests pass. Policy derived from the shipped evaluator constants; no new measurement and no provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Bounded provider spend per clip (`182.C.14.06` estimate)                       | The cut-alignment signature is reported, not gated: a director cutting on a reaction produces the same pattern and separating them needs rated material (182.C.33).                                                                                                                                                                                                                                                                                      |
| 182.C.20.06 | Define when off-screen dialogue is legitimate and require the plan/release evidence to distinguish it from …           | MLOps | 182.C.17, 182.C.26.02 | implemented    | `apps/oshun/bff/src/generation/human-video-offscreen-dialogue.ts` (new), `human-video-offscreen-dialogue.spec.ts` (new)       | 182.C.09.04 already gave the PLAN a vocabulary for this - per-window visibility, on-camera as the only scorable one, a written reason required for the visibilities that look like defects. What it left open is the half this item names: A PLAN IS A CLAIM ABOUT A DELIVERY and nobody was checking it against one. Four outcomes now, and only one is a pass. SCORED: planned on camera, delivered on camera, sync measured. LEGITIMATE-OFF-PLAN: planned off-screen or occluded, delivery agrees, sync not measured and should not be - neither a pass nor a failure. UNEVALUABLE-DEFECT: planned on camera, delivery has no measurable mouth - THE DANGEROUS ONE, because it looks exactly like the legitimate case and the spec asserts that both produce no sync number while only one is a fault. PLAN-VIOLATED: the two disagree in either direction, which may be fine footage of the WRONG shot - a different problem from bad footage of the right one, so merging it with the defect case would send both to the same queue. OBSERVED visibility is MEASURED rather than taken from the plan: face present or not, head yaw past 60 degrees an extreme profile and past 120 a back, boundaries two orders of magnitude clear of the 1.55-degree pose noise 182.C.19.07 measured, which is what makes the plan checkable at all. THE ARITHMETIC IS WHERE THIS GOES WRONG IN PRACTICE: a take with nine legitimate off-screen windows and one scored window that passed is NOT '100% sync pass', it is one window measured, so coverageOf counts excluded windows in the DENOMINATOR and assertSyncCoverage refuses a thin scored fraction. That gate is adversarial by design - a plan that puts every line off screen is a plan that cannot be checked, and a generator optimised against a gate that ignored coverage would learn to produce exactly that. The reason rule is enforced AGAIN at delivery, because a plan can be edited after a render and a reason written afterwards to explain a bad take is what the rule exists to prevent. The release evidence carries `syncAbsentBecause` per window: a proof that simply omits a sync number for nine windows is indistinguishable from one where nine windows failed.                                                                                                                                                                                                                                                                                              | 24 tests pass. Pure policy over the 182.C.09.04 contract and the 182.C.19.07 pose measurement; no new measurement and no provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Bounded provider spend per clip (`182.C.14.06` estimate)                       | None                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
| 182.C.20.07 | Add challenge cases for side/back turns, masks, hands near the face, eating/drinking, wind/hair occlusion …            | MLOps | 182.C.17, 182.C.26.02 | blocked-corpus | `apps/oshun/bff/src/generation/human-video-challenge-cases.ts` (new), `human-video-challenge-cases.spec.ts` (new)             | A CHALLENGE SUITE WHERE EVERY EXPECTED ANSWER IS "REJECT" IS PASSED BY A SYSTEM THAT REJECTS EVERYTHING - and that is exactly what this list invites, because masks, back turns, wind and low light all read as things that should fail. The suite then has NO DISCRIMINATING POWER AND LOOKS MORE RIGOROUS THE LONGER IT GETS. Each challenge declares the answer correct FOR IT, and a suite with no acceptance in it is refused; the ten split 4 pass / 1 fail / 5 decline. FOR HALF OF THEM THE RIGHT ANSWER IS NEITHER PASS NOR FAIL: a back turn has no visible speech and a mask covers the mouth, so DECLINING IS THE CORRECT OUTCOME (182.C.17.06, 182.C.20.06), and recording one as a failure rejects footage for failing a test that could not run on it. Side and back turns are split on 182.C.20.06 OWN YAW BOUNDARIES (60 and 120 degrees) and get OPPOSITE expectations, which is the reason to split them. THREE OF THEM ATTACK THE EVALUATOR RATHER THAN THE TAKE - fast camera moves, low light and shallow focus can leave a take a human would accept and a measurement nobody can make, so the pass criterion there is THAT THE SYSTEM KNOWS IT CANNOT TELL. Intermittent occlusion (wind, a passing hand) deliberately stays SCORABLE, because treating it as unevaluable would excuse a real failure. AND COMPRESSION IS NOT IN THE WORLD, IT IS IN OUR PIPELINE: it is damage we do on the way out, so it is applied AFTER evaluation on the bytes that ship - applied before, it charges the model for what the encoder did.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | 15 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. The reject-only suite is BUILT by filtering the real spec table rather than hand-written, so the refusal fires on the actual expectations; turn boundaries are asserted against `PROVISIONAL_YAW_BOUNDARIES` from 182.C.20.06 rather than retyped; and the evaluator-attack set is asserted to be exactly the three whose expected outcome is decline.                                                                                                           | No provider spend; no model call.                                              | BLOCKED on 182.C.33 clips - THE CHECKBOX STAYS UNCHECKED. The cases and their correct answers are defined; shooting or generating ten challenge conditions is corpus work, and the compression case additionally needs the 182.C.16.09 final-byte path to run against. Same position as 182.C.20.01 and 182.C.19.09.                                                                                                                                     |
| 182.C.20.08 | Require manual review until each active/multi-speaker cohort reaches its sample-size and agreement targets.            | MLOps | 182.C.17, 182.C.26.02 | implemented    | `apps/oshun/bff/src/generation/human-video-cohort-review.ts` (new), `human-video-cohort-review.spec.ts` (new)                 | 182.C.17 already set the floors for calibrating an evaluator - three raters minimum for an agreement statistic to exist at all, thirty clips per stratum cell, Krippendorff's alpha 0.667 - and this applies them PER COHORT, which is what the item is about. A cohort is a combination of what an active multi-speaker scene actually varies in: speaker count, motion, shot size, occlusion. 3 x 4 x 3 x 4 = 144 combinations, which is the honest size of the problem and the number the coverage summary divides by. THREE STATES, AND THE THIRD IS THE ONE THAT GETS MISSED. AUTOMATED: enough rated clips and agreeing raters. UNDER-POWERED: clips exist but too few, or the raters disagreed - and the blockers say WHICH, because more clips fixes one and explicitly will NOT fix the other (a low alpha means the raters are not measuring the same thing, so the cohort or the rating scale has to change). NEVER SEEN: no clip of this cohort has EVER been rated, which is no data rather than no failures, and it is the state a new shot type lands in silently - returned as its own verdict rather than folded into under-powered, because a cohort with zero clips reads as CLEAN in any report that only counts failures. The refusal codes separate the three causes because they go to three different people: more clips is data collection, disagreement is a protocol problem, a never-seen cohort is a scoping surprise. coverageSummary counts never-seen cohorts IN THE DENOMINATOR - a programme that rated three cohorts thoroughly and never looked at the other 141 has covered 2% of the space, and any summary dividing by 'cohorts we have data for' would report 100%. An uncomputed alpha is treated as not a passing one. And a cohort is NOT covered by its neighbours: NON_TRANSFERABLE_DIMENSIONS names each dimension with the measurement behind it - motion (182.C.19.08: a face at 21px of blur retains 4% of its identity margin and at 31px scores below the different-person baseline against itself), occlusion (182.C.19.03: the one condition that SWAPS a track onto the other person while camera motion and lighting held), shot size (margin 90% at a face 14% of frame height falling to 24% at 4.4%), and speaker count (182.C.18.08: overlapping speech giving a 514 ms onset error while the aligner reported success). An undeclared cohort is refused, since no cell of the calibration covers it and it would be gated by whichever operating point happened to be loaded. | 14 tests pass. Policy over the 182.C.17 calibration floors and the 182.C.18/19 measurements; no new measurement and no provider spend.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | Bounded provider spend per clip (`182.C.14.06` estimate)                       | No cohort is automated today: the 182.C.33 corpus is what populates the cells. The module reports 0 of 144 rather than treating an empty space as clean.                                                                                                                                                                                                                                                                                                 |

### 182.C.21 Candidate Generation, Selection, Retry, and Correction

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                             | Owner     | Dependency | Status      | Changed Files                                                                                                         | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Provider/Hardware Evidence                                                                                     | Cost | Blocker                                                                                                             |
| ----------- | -------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | ----------- | --------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- | ---- | ------------------------------------------------------------------------------------------------------------------- |
| 182.C.21.01 | Generate multiple bounded candidates for production/hero requests and retain a shared request/shot/audio lineage.    | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-candidate-lineage.ts` (new), `human-video-candidate-lineage.spec.ts` (new) | 182.C.21.01 and .10 are THE SAME STRUCTURE FROM TWO ENDS and were built together: .01 asks for several bounded candidates with a shared request/shot/audio lineage, .10 asks that regenerating one shot leaves unrelated approved work alone, and both answer the question of which candidates belong together. COMPARABILITY IS WHY THE LINEAGE EXISTS, not provenance for its own sake. Two candidates are comparable only when they share request, shot AND audio - and the AUDIO is the one people forget: 182.C.18.02 measures word timing AGAINST the approved audio and 182.C.17 scores sync against the audio that was muxed, so two candidates from different audio have sync numbers that are not on the same scale and ranking them compares two measurements of two things. assertComparable refuses, and says WHICH field differs because the fixes differ. comparisonSets groups a mixed pile into sets a creator can actually compare, because presenting takes from different audio side by side invites exactly that comparison. BOUNDED MEANS BOUNDED BY ALL THREE: a cap on candidate COUNT does not bound spend (candidates differ in cost by lane), and a cap on SPEND does not bound latency (a cheap lane can be slow), so assertWithinBounds takes count, spend and elapsed and the first one hit stops the fan-out. WHAT A REGENERATION REACHES is the trap in .10: invalidation follows the DEPENDENCY EDGES, not the shot list. The safe-looking rule 'regenerating a shot invalidates that shot' is WRONG IN BOTH DIRECTIONS - it misses a neighbour that took a continuity handoff from the regenerated shot's final frame, and it over-reaches when nothing depends on the shot at all. A change to the AUDIO invalidates every shot that used it however far apart they are in the timeline, and a change to the SCRIPT invalidates the audio and therefore everything downstream. invalidatedBy walks the graph and `survives` reports the other half, which is what .10 is actually about. assertApprovalsAccepted is a CONFIRMATION rather than a prohibition - sometimes the script really does have to change and three approved shots really do have to go - so the caller passes the approvals it accepts losing and anything not on that list stops the change: somebody clicking 'regenerate this shot' should not discover afterwards that it took two approved neighbours with it.                                                                                                                                                                                                                                                                                                                          | 20 tests pass. Policy and graph arithmetic; no new measurement and no provider spend.                          | —    | None                                                                                                                |
| 182.C.21.02 | Rank candidates with a Pareto-aware policy so visual beauty cannot compensate for failed words, wrong speaker …      | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-candidate-ranking.ts` (new), `human-video-candidate-ranking.spec.ts` (new) | 182.C.21.02 and .03 are ONE MECHANISM and were built together: .03 asks for hard gates and soft preferences defined separately with exact speech, consent, safety and severe sync/identity failures NON-COMPENSABLE, and .02 asks the ranking to honour it - which a ranking that scores everything and sorts cannot do, however the weights are chosen. THE DISTINCTION THAT MAKES IT WORK: a hard gate EXCLUDES, it does not score low. A low score can be beaten by a high score somewhere else - that is what scoring is for - and an exclusion cannot be beaten by anything, which is what non-compensable means. So a candidate perfect on every soft dimension that failed exact speech is NOT RANKED LAST, it is not ranked, and the spec asserts exactly that case. A failed gate and an UNMEASURED one are excluded separately, because one candidate was checked and was wrong and the other was never checked - both excluded, only one a defect in the take. AND WITHIN THE SOFT DIMENSIONS, ANY WEIGHTED SUM IS COMPENSATION: scoring sharpness 0.4 plus smoothness 0.6 lets a sharp candidate outrank a smoother one, which may be right and is a PRODUCT decision rather than an arithmetic one. So the soft dimensions reduce by PARETO DOMINANCE - dropped only when another is at least as good on every dimension and better on one - and what survives is the front, returned with its trade-offs intact rather than collapsed into a number that looks decided. Picking inside the front REFUSES without a stated order, and the order is lexicographic so no later dimension can outweigh an earlier one: the same non-compensability rule applied one level down, and written down rather than hidden. Domination on an incomplete comparison is refused (conservative: both stay on the front). visible-mouth-coverage and no-critical-artifacts are deliberately left SOFT - coverage is a measurability fact that 182.C.20.06 handles, and 'critical' in artifacts is a severity judgement 182.C.19.05 showed we cannot make for six of seven classes; a hard gate on an unmeasurable thing fires on the measurement rather than on the take. assertGateClassification refuses to promote a soft quantity into the non-compensable set or demote a hard one out of it. The ranking proof carries EVERY exclusion: a proof saying 'candidate 3 selected from 4' hides whether the other three were merely worse or whether two had no consent.                                                                                                                                                                                                                                                                              | 18 tests pass. Policy over the existing RELEASE_CONDITIONS contract; no new measurement and no provider spend. | —    | None                                                                                                                |
| 182.C.21.03 | Define hard gates versus soft preferences and make exact speech, consent, safety …                                   | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-candidate-ranking.ts` (new), `human-video-candidate-ranking.spec.ts` (new) | 182.C.21.02 and .03 are ONE MECHANISM and were built together: .03 asks for hard gates and soft preferences defined separately with exact speech, consent, safety and severe sync/identity failures NON-COMPENSABLE, and .02 asks the ranking to honour it - which a ranking that scores everything and sorts cannot do, however the weights are chosen. THE DISTINCTION THAT MAKES IT WORK: a hard gate EXCLUDES, it does not score low. A low score can be beaten by a high score somewhere else - that is what scoring is for - and an exclusion cannot be beaten by anything, which is what non-compensable means. So a candidate perfect on every soft dimension that failed exact speech is NOT RANKED LAST, it is not ranked, and the spec asserts exactly that case. A failed gate and an UNMEASURED one are excluded separately, because one candidate was checked and was wrong and the other was never checked - both excluded, only one a defect in the take. AND WITHIN THE SOFT DIMENSIONS, ANY WEIGHTED SUM IS COMPENSATION: scoring sharpness 0.4 plus smoothness 0.6 lets a sharp candidate outrank a smoother one, which may be right and is a PRODUCT decision rather than an arithmetic one. So the soft dimensions reduce by PARETO DOMINANCE - dropped only when another is at least as good on every dimension and better on one - and what survives is the front, returned with its trade-offs intact rather than collapsed into a number that looks decided. Picking inside the front REFUSES without a stated order, and the order is lexicographic so no later dimension can outweigh an earlier one: the same non-compensability rule applied one level down, and written down rather than hidden. Domination on an incomplete comparison is refused (conservative: both stay on the front). visible-mouth-coverage and no-critical-artifacts are deliberately left SOFT - coverage is a measurability fact that 182.C.20.06 handles, and 'critical' in artifacts is a severity judgement 182.C.19.05 showed we cannot make for six of seven classes; a hard gate on an unmeasurable thing fires on the measurement rather than on the take. assertGateClassification refuses to promote a soft quantity into the non-compensable set or demote a hard one out of it. The ranking proof carries EVERY exclusion: a proof saying 'candidate 3 selected from 4' hides whether the other three were merely worse or whether two had no consent.                                                                                                                                                                                                                                                                              | 18 tests pass. Policy over the existing RELEASE_CONDITIONS contract; no new measurement and no provider spend. | —    | None                                                                                                                |
| 182.C.21.04 | Classify failures into prompt/control, provider transport, speech, sync, speaker attribution, identity, anatomy …    | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-failure-response.ts` (new), `human-video-failure-response.spec.ts` (new)   | 182.C.21.04, .05 and .06 are ONE PIPELINE and were built together: a class chooses a strategy, a strategy costs money and time, and the limits decide when to stop - so splitting them would put the strategy table somewhere it could disagree with the classes. THESE ARE NOT THE CONTRACT'S FAILURE_FAMILIES, and the spec asserts zero overlap. The families classify why a JOB failed (validation, consent, provider, budget); these classify why a DELIVERED TAKE failed EVALUATION, which is a different question - a take that arrived fine and says the wrong words is `speech`, and its job family is nothing at all because the job succeeded. Collapsing them would lose the distinction between 'the provider broke' and 'the provider worked and the result is wrong'. Each of the eleven classes carries what CHANGES on a retry, in order, because retrying identically after a deterministic failure spends money to reproduce it - and levers are never pulled twice, with the run stopping instead. The orderings encode real reasoning: a SPEECH failure binds the approved audio harder BEFORE asking the same model to read more carefully, and a SPEAKER-ATTRIBUTION failure SUBDIVIDES THE SHOT first, so one person speaks per segment and the ambiguity the model got wrong is gone. THREE CLASSES ARE NOT RETRYABLE AT ALL: policy, because a safety or consent refusal is a decision rather than a flake and a retry that happens to pass is a retry that found a hole; continuity, because a shot contradicting an approved neighbour needs the neighbour or the plan to change and that is a person's call; and encoding, because the bytes are wrong and a re-render produces different wrong bytes - the fix is in the mux, not the model. THE RULE .06 IS REALLY ABOUT: when the limits are hit the honest output is the BEST UNRELEASED candidate plus its diagnostics, and the failure mode the item names is specific and tempting - a run nearly out of budget quietly relaxing a threshold so the last candidate passes. assertThresholdsUnchanged compares what a verdict was measured under against what the policy declared and refuses ANY drift, in either direction, including a threshold that appeared from nowhere: a gate that moved during a run is not a gate. The report's `released` is a literal false rather than a boolean so a caller cannot read a truthy field and ship, and bestUnreleased carries the MEASURED VALUES beside the thresholds - 'sync failed' sends somebody to re-render, and 'sync measured 5.2 against a threshold of 6' tells them how far off it was and whether the threshold is the thing to argue with.                                                             | 24 tests pass. Policy over the existing contract; no new measurement and no provider spend.                    | —    | None                                                                                                                |
| 182.C.21.05 | Retry selectively with changed seed, clarified prompt, reference ordering, shot subdivision …                        | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-failure-response.ts` (new), `human-video-failure-response.spec.ts` (new)   | 182.C.21.04, .05 and .06 are ONE PIPELINE and were built together: a class chooses a strategy, a strategy costs money and time, and the limits decide when to stop - so splitting them would put the strategy table somewhere it could disagree with the classes. THESE ARE NOT THE CONTRACT'S FAILURE_FAMILIES, and the spec asserts zero overlap. The families classify why a JOB failed (validation, consent, provider, budget); these classify why a DELIVERED TAKE failed EVALUATION, which is a different question - a take that arrived fine and says the wrong words is `speech`, and its job family is nothing at all because the job succeeded. Collapsing them would lose the distinction between 'the provider broke' and 'the provider worked and the result is wrong'. Each of the eleven classes carries what CHANGES on a retry, in order, because retrying identically after a deterministic failure spends money to reproduce it - and levers are never pulled twice, with the run stopping instead. The orderings encode real reasoning: a SPEECH failure binds the approved audio harder BEFORE asking the same model to read more carefully, and a SPEAKER-ATTRIBUTION failure SUBDIVIDES THE SHOT first, so one person speaks per segment and the ambiguity the model got wrong is gone. THREE CLASSES ARE NOT RETRYABLE AT ALL: policy, because a safety or consent refusal is a decision rather than a flake and a retry that happens to pass is a retry that found a hole; continuity, because a shot contradicting an approved neighbour needs the neighbour or the plan to change and that is a person's call; and encoding, because the bytes are wrong and a re-render produces different wrong bytes - the fix is in the mux, not the model. THE RULE .06 IS REALLY ABOUT: when the limits are hit the honest output is the BEST UNRELEASED candidate plus its diagnostics, and the failure mode the item names is specific and tempting - a run nearly out of budget quietly relaxing a threshold so the last candidate passes. assertThresholdsUnchanged compares what a verdict was measured under against what the policy declared and refuses ANY drift, in either direction, including a threshold that appeared from nowhere: a gate that moved during a run is not a gate. The report's `released` is a literal false rather than a boolean so a caller cannot read a truthy field and ship, and bestUnreleased carries the MEASURED VALUES beside the thresholds - 'sync failed' sends somebody to re-render, and 'sync measured 5.2 against a threshold of 6' tells them how far off it was and whether the threshold is the thing to argue with.                                                             | 24 tests pass. Policy over the existing contract; no new measurement and no provider spend.                    | —    | None                                                                                                                |
| 182.C.21.06 | Stop retries at spend/latency limits and return the best unreleased diagnostics rather than silently lowering …      | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-failure-response.ts` (new), `human-video-failure-response.spec.ts` (new)   | 182.C.21.04, .05 and .06 are ONE PIPELINE and were built together: a class chooses a strategy, a strategy costs money and time, and the limits decide when to stop - so splitting them would put the strategy table somewhere it could disagree with the classes. THESE ARE NOT THE CONTRACT'S FAILURE_FAMILIES, and the spec asserts zero overlap. The families classify why a JOB failed (validation, consent, provider, budget); these classify why a DELIVERED TAKE failed EVALUATION, which is a different question - a take that arrived fine and says the wrong words is `speech`, and its job family is nothing at all because the job succeeded. Collapsing them would lose the distinction between 'the provider broke' and 'the provider worked and the result is wrong'. Each of the eleven classes carries what CHANGES on a retry, in order, because retrying identically after a deterministic failure spends money to reproduce it - and levers are never pulled twice, with the run stopping instead. The orderings encode real reasoning: a SPEECH failure binds the approved audio harder BEFORE asking the same model to read more carefully, and a SPEAKER-ATTRIBUTION failure SUBDIVIDES THE SHOT first, so one person speaks per segment and the ambiguity the model got wrong is gone. THREE CLASSES ARE NOT RETRYABLE AT ALL: policy, because a safety or consent refusal is a decision rather than a flake and a retry that happens to pass is a retry that found a hole; continuity, because a shot contradicting an approved neighbour needs the neighbour or the plan to change and that is a person's call; and encoding, because the bytes are wrong and a re-render produces different wrong bytes - the fix is in the mux, not the model. THE RULE .06 IS REALLY ABOUT: when the limits are hit the honest output is the BEST UNRELEASED candidate plus its diagnostics, and the failure mode the item names is specific and tempting - a run nearly out of budget quietly relaxing a threshold so the last candidate passes. assertThresholdsUnchanged compares what a verdict was measured under against what the policy declared and refuses ANY drift, in either direction, including a threshold that appeared from nowhere: a gate that moved during a run is not a gate. The report's `released` is a literal false rather than a boolean so a caller cannot read a truthy field and ship, and bestUnreleased carries the MEASURED VALUES beside the thresholds - 'sync failed' sends somebody to re-render, and 'sync measured 5.2 against a threshold of 6' tells them how far off it was and whether the threshold is the thing to argue with.                                                             | 24 tests pass. Policy over the existing contract; no new measurement and no provider spend.                    | —    | None                                                                                                                |
| 182.C.21.07 | Implement optional corrective lip-sync/video-edit providers behind an explicit fallback policy for otherwise …       | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-correction.ts` (new), `human-video-correction.spec.ts` (new)               | 182.C.21.07, .08 and .09 are ONE MECHANISM and were built together, because they are only coherent together: a correction that is opt-in but not re-gated is a hole, and one that is re-gated but keeps no before/after is a pass nobody can audit. WHY A CORRECTION IS DANGEROUS RATHER THAN MERELY OPTIONAL: a corrective lip-sync provider repaints the mouth region, which is precisely where this phase's other evaluators look. 182.C.19.04 measured blur leaving a detector 0.879 confident while the face scored 0.033 against its own unblurred self - a smoothing correction can do the same to identity while every face still detects. 182.C.19.05 measured face boundary tearing as a row-continuity break, and a composited mouth has a boundary by construction. 182.C.07.03 measured a face composited onto another head reading 0.737 to its source and 0.381 to the target - a correction IS a composite. So identity or texture damage is EXPECTED and must be measured rather than assumed absent, which is why .08 exists and why .09 keeps the before. THE RULE THAT MAKES .08 REAL: a gate that PASSED before the correction is STALE, not inherited. Correction changes pixels, so a pass measured on the old pixels describes a file that is no longer the deliverable, and assertAllGatesRemeasured refuses any pre-correction verdict - inheriting the easy passes is exactly what a tired implementation does. All eight named gates must carry a post-correction verdict, PROVENANCE included, for a reason easy to miss: a correction produces new bytes, so the content hash and the signature over it are both stale and a provenance record pointing at the uncorrected file describes a file nobody will play. .07's 'never an invisible default' is enforced as TWO separate requirements - the policy must be opted into before the render AND the creator must be told a correction happened, because a take corrected under a standing policy nobody remembers enabling is still an invisible default to the person watching it. 'Otherwise excellent' is enforced rather than assumed: only sync is correctable, and a candidate failing several gates is refused, since correcting the mouth of a bad take produces a bad take with a repaired mouth. .09's change report says WHERE to look rather than pass or fail - frames touched, the heavily changed ones, and whether the correction was wider than the caller expected, because a lip-sync corrector that repainted every frame of a ten-second take did something other than fix the mouth. Permission is checked FIRST, before anything is measured: measuring an unauthorised correction produces a record that looks like a considered decision. | 18 tests pass. Policy over the 182.C.07/19 measurements; no new measurement and no provider spend.             | —    | No corrective provider is integrated yet; this is the policy and the re-gating every integration must pass through. |
| 182.C.21.08 | Re-run identity, anatomy, motion, exact-speech, sync, encoding, safety, and provenance gates after any correction.   | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-correction.ts` (new), `human-video-correction.spec.ts` (new)               | 182.C.21.07, .08 and .09 are ONE MECHANISM and were built together, because they are only coherent together: a correction that is opt-in but not re-gated is a hole, and one that is re-gated but keeps no before/after is a pass nobody can audit. WHY A CORRECTION IS DANGEROUS RATHER THAN MERELY OPTIONAL: a corrective lip-sync provider repaints the mouth region, which is precisely where this phase's other evaluators look. 182.C.19.04 measured blur leaving a detector 0.879 confident while the face scored 0.033 against its own unblurred self - a smoothing correction can do the same to identity while every face still detects. 182.C.19.05 measured face boundary tearing as a row-continuity break, and a composited mouth has a boundary by construction. 182.C.07.03 measured a face composited onto another head reading 0.737 to its source and 0.381 to the target - a correction IS a composite. So identity or texture damage is EXPECTED and must be measured rather than assumed absent, which is why .08 exists and why .09 keeps the before. THE RULE THAT MAKES .08 REAL: a gate that PASSED before the correction is STALE, not inherited. Correction changes pixels, so a pass measured on the old pixels describes a file that is no longer the deliverable, and assertAllGatesRemeasured refuses any pre-correction verdict - inheriting the easy passes is exactly what a tired implementation does. All eight named gates must carry a post-correction verdict, PROVENANCE included, for a reason easy to miss: a correction produces new bytes, so the content hash and the signature over it are both stale and a provenance record pointing at the uncorrected file describes a file nobody will play. .07's 'never an invisible default' is enforced as TWO separate requirements - the policy must be opted into before the render AND the creator must be told a correction happened, because a take corrected under a standing policy nobody remembers enabling is still an invisible default to the person watching it. 'Otherwise excellent' is enforced rather than assumed: only sync is correctable, and a candidate failing several gates is refused, since correcting the mouth of a bad take produces a bad take with a repaired mouth. .09's change report says WHERE to look rather than pass or fail - frames touched, the heavily changed ones, and whether the correction was wider than the caller expected, because a lip-sync corrector that repainted every frame of a ten-second take did something other than fix the mouth. Permission is checked FIRST, before anything is measured: measuring an unauthorised correction produces a record that looks like a considered decision. | 18 tests pass. Policy over the 182.C.07/19 measurements; no new measurement and no provider spend.             | —    | No corrective provider is integrated yet; this is the policy and the re-gating every integration must pass through. |
| 182.C.21.09 | Preserve before/after candidates and record which frames changed so reviewers can detect identity or texture damage. | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-correction.ts` (new), `human-video-correction.spec.ts` (new)               | 182.C.21.07, .08 and .09 are ONE MECHANISM and were built together, because they are only coherent together: a correction that is opt-in but not re-gated is a hole, and one that is re-gated but keeps no before/after is a pass nobody can audit. WHY A CORRECTION IS DANGEROUS RATHER THAN MERELY OPTIONAL: a corrective lip-sync provider repaints the mouth region, which is precisely where this phase's other evaluators look. 182.C.19.04 measured blur leaving a detector 0.879 confident while the face scored 0.033 against its own unblurred self - a smoothing correction can do the same to identity while every face still detects. 182.C.19.05 measured face boundary tearing as a row-continuity break, and a composited mouth has a boundary by construction. 182.C.07.03 measured a face composited onto another head reading 0.737 to its source and 0.381 to the target - a correction IS a composite. So identity or texture damage is EXPECTED and must be measured rather than assumed absent, which is why .08 exists and why .09 keeps the before. THE RULE THAT MAKES .08 REAL: a gate that PASSED before the correction is STALE, not inherited. Correction changes pixels, so a pass measured on the old pixels describes a file that is no longer the deliverable, and assertAllGatesRemeasured refuses any pre-correction verdict - inheriting the easy passes is exactly what a tired implementation does. All eight named gates must carry a post-correction verdict, PROVENANCE included, for a reason easy to miss: a correction produces new bytes, so the content hash and the signature over it are both stale and a provenance record pointing at the uncorrected file describes a file nobody will play. .07's 'never an invisible default' is enforced as TWO separate requirements - the policy must be opted into before the render AND the creator must be told a correction happened, because a take corrected under a standing policy nobody remembers enabling is still an invisible default to the person watching it. 'Otherwise excellent' is enforced rather than assumed: only sync is correctable, and a candidate failing several gates is refused, since correcting the mouth of a bad take produces a bad take with a repaired mouth. .09's change report says WHERE to look rather than pass or fail - frames touched, the heavily changed ones, and whether the correction was wider than the caller expected, because a lip-sync corrector that repainted every frame of a ten-second take did something other than fix the mouth. Permission is checked FIRST, before anything is measured: measuring an unauthorised correction produces a record that looks like a considered decision. | 18 tests pass. Policy over the 182.C.07/19 measurements; no new measurement and no provider spend.             | —    | No corrective provider is integrated yet; this is the policy and the re-gating every integration must pass through. |
| 182.C.21.10 | Let creators compare takes and regenerate one shot/segment without invalidating unrelated approved work.             | Oshun BFF | 182.C.20   | implemented | `apps/oshun/bff/src/generation/human-video-candidate-lineage.ts` (new), `human-video-candidate-lineage.spec.ts` (new) | 182.C.21.01 and .10 are THE SAME STRUCTURE FROM TWO ENDS and were built together: .01 asks for several bounded candidates with a shared request/shot/audio lineage, .10 asks that regenerating one shot leaves unrelated approved work alone, and both answer the question of which candidates belong together. COMPARABILITY IS WHY THE LINEAGE EXISTS, not provenance for its own sake. Two candidates are comparable only when they share request, shot AND audio - and the AUDIO is the one people forget: 182.C.18.02 measures word timing AGAINST the approved audio and 182.C.17 scores sync against the audio that was muxed, so two candidates from different audio have sync numbers that are not on the same scale and ranking them compares two measurements of two things. assertComparable refuses, and says WHICH field differs because the fixes differ. comparisonSets groups a mixed pile into sets a creator can actually compare, because presenting takes from different audio side by side invites exactly that comparison. BOUNDED MEANS BOUNDED BY ALL THREE: a cap on candidate COUNT does not bound spend (candidates differ in cost by lane), and a cap on SPEND does not bound latency (a cheap lane can be slow), so assertWithinBounds takes count, spend and elapsed and the first one hit stops the fan-out. WHAT A REGENERATION REACHES is the trap in .10: invalidation follows the DEPENDENCY EDGES, not the shot list. The safe-looking rule 'regenerating a shot invalidates that shot' is WRONG IN BOTH DIRECTIONS - it misses a neighbour that took a continuity handoff from the regenerated shot's final frame, and it over-reaches when nothing depends on the shot at all. A change to the AUDIO invalidates every shot that used it however far apart they are in the timeline, and a change to the SCRIPT invalidates the audio and therefore everything downstream. invalidatedBy walks the graph and `survives` reports the other half, which is what .10 is actually about. assertApprovalsAccepted is a CONFIRMATION rather than a prohibition - sometimes the script really does have to change and three approved shots really do have to go - so the caller passes the approvals it accepts losing and anything not on that list stops the change: somebody clicking 'regenerate this shot' should not discover afterwards that it took two approved neighbours with it.                                                                                                                                                                                                                                                                                                                          | 20 tests pass. Policy and graph arithmetic; no new measurement and no provider spend.                          | —    | None                                                                                                                |

### 182.C.22 Long-Form, Multi-Shot, Edit, and Extension Workflows

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                             | Owner | Dependency | Status      | Changed Files                                                                                                                                                                                             | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Provider/Hardware Evidence                                                                                                                                | Cost | Blocker                                                                                                                                                                 |
| ----------- | -------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | ----------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------- | ---- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.22.01 | Segment long scripts by sentence, performance beat, provider duration limit, continuity, and safe audio boundaries.  | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-cut-safety.ts` (new), `human-video-cut-safety.spec.ts` (new)                                                                                                   | 182.C.22.01 and .04 are ONE QUESTION - where can a boundary go - and were built together. THE FINDING THAT DECIDES IT: YOU CANNOT CUT AT AN ALIGNED WORD END. 182.C.18.02 measured word ONSETS carrying a +48.3 ms systematic bias that a constant corrects, leaving +/-32.87 ms of residual, and word ENDS 96 to 141 ms EARLY consistently, because CTC emits a token before the sound finishes. So a cut at an aligned word end lands up to 141 ms INSIDE the word and takes the end of it - the phoneme trim this item forbids, arrived at by doing the obvious thing. The spec asserts that a cut 100 ms AFTER an aligned end is still refused. WHAT IS SAFE INSTEAD is a measured silence gap, and THE MARGIN IS NOT A GUESS: 32.87 + 141 = 174 ms of uncertainty about where speech actually stops, so a gap must exceed TWICE that before a cut fits with room either side, and the boundary is the gap's MIDPOINT rather than an edge because the edges are exactly where the uncertainty is. Segmentation takes the LAST safe boundary under a provider's duration limit rather than the nearest to it, because a boundary 20 ms under the limit is one provider rounding away from being over it; when no gap falls in range the segmenter REFUSES and says the script cannot be cut for this provider without re-recording or a longer limit, rather than cutting mid-word. DUPLICATING WORDS is the cross-fade failure: an overlap inside a gap duplicates silence nobody hears, and one reaching a word replays it - 182.C.18.04 measured that signature at waveform autocorrelation 1.000 against 0.261 for clean speech, so it WOULD be caught downstream, after the render, having been paid for. A/V TIMESTAMP CONTINUITY is compared in INTEGER TICKS rather than floating seconds: a container stores integer ticks, that is what a demuxer compares, and a seconds comparison fails on representation alone - 3.5 + 1/90000 does not subtract back to exactly 1/90000, and a correct one-tick join was refused until this was fixed. The tolerance is one tick (11 microseconds at 90 kHz), because a millisecond tolerance would pass a 90-tick gap.                                                                                                                                                                                                                                                      | 21 tests pass, over the 182.C.18.02 calibration this repo measured. No new measurement and no provider spend.                                             | —    | Sentence and performance-beat segmentation take the boundaries this provides; the beat detection itself is a script-side concern (182.C.22.02 ledger).                  |
| 182.C.22.02 | Maintain cast, wardrobe, props, environment, time of day, camera grammar, voice …                                    | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-continuity-ledger.ts` (new), `human-video-continuity-ledger.spec.ts` (new)                                                                                     | 182.C.22.02 and .03 both carry state between shots and were built together. 182.C.09 already gave a handoff its five carriers and said which one actually carries identity; what was missing is the running LEDGER of what each approved shot established, so the next shot is checked against it rather than against a memory of the brief. THE DISTINCTION THAT MAKES THE LEDGER USEFUL is that the eight aspects are NOT equally checkable, and pretending otherwise produces a ledger reporting contradictions it cannot support. DECLARED aspects (cast, wardrobe, props, camera grammar) are what the plan SAID, and a contradiction between two declarations is a planning error decidable with no model at all. MEASURED aspects (voice, speech timing) are what the delivery DID, so a contradiction is a measurement with an ERROR BAR: the aligner's +/-32.87 ms residual is the floor, and a 20 ms drift is a claim about the aligner rather than about the take. AND TWO ASPECTS ARE NEITHER - environment and time of day: the declaration is checkable against itself and whether the PIXELS show the same place is not decidable from anything staged, so assertVerificationHonest REFUSES a `verified: true` on them, because such a row reads as somebody having looked. Contradictions are allowed when DECLARED as intended - 182.C.07.04 already established that wardrobe and scene changes are often deliberate - and refused otherwise, because a change nobody declared looks exactly like one the director asked for. Only an APPROVED shot writes to the ledger: one recording what was ATTEMPTED would have the next shot matching a take somebody rejected. THE HANDOFF SIDE is a routing question with the trap already in the contract: A SEED IS A WEAK CARRIER AND ONLY MEANS ANYTHING WITHIN ONE LANE, so planHandoff DROPS it across a lane change and records why - carrying it would look like continuity and be nothing. Carriers a lane cannot accept are dropped with reasons, and dropping the reference pack says plainly that nothing else on the list replaces it. assertHandoffCarriesSomething refuses a plan whose only survivors are weak or non-structural, because it would otherwise report a carrier list that looks non-empty while carrying no identity at all.                                                                                                          | 22 tests pass, over the 182.C.09 carrier contract and the 182.C.18.02 aligner residual. No new measurement and no provider spend.                         | —    | Environment and time-of-day continuity are declared and never verified against pixels; nothing staged can do that (182.C.19.02).                                        |
| 182.C.22.03 | Support selected terminal-frame, final-frame, reference- video, or extension handoff based on provider capability.   | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-continuity-ledger.ts` (new), `human-video-continuity-ledger.spec.ts` (new)                                                                                     | 182.C.22.02 and .03 both carry state between shots and were built together. 182.C.09 already gave a handoff its five carriers and said which one actually carries identity; what was missing is the running LEDGER of what each approved shot established, so the next shot is checked against it rather than against a memory of the brief. THE DISTINCTION THAT MAKES THE LEDGER USEFUL is that the eight aspects are NOT equally checkable, and pretending otherwise produces a ledger reporting contradictions it cannot support. DECLARED aspects (cast, wardrobe, props, camera grammar) are what the plan SAID, and a contradiction between two declarations is a planning error decidable with no model at all. MEASURED aspects (voice, speech timing) are what the delivery DID, so a contradiction is a measurement with an ERROR BAR: the aligner's +/-32.87 ms residual is the floor, and a 20 ms drift is a claim about the aligner rather than about the take. AND TWO ASPECTS ARE NEITHER - environment and time of day: the declaration is checkable against itself and whether the PIXELS show the same place is not decidable from anything staged, so assertVerificationHonest REFUSES a `verified: true` on them, because such a row reads as somebody having looked. Contradictions are allowed when DECLARED as intended - 182.C.07.04 already established that wardrobe and scene changes are often deliberate - and refused otherwise, because a change nobody declared looks exactly like one the director asked for. Only an APPROVED shot writes to the ledger: one recording what was ATTEMPTED would have the next shot matching a take somebody rejected. THE HANDOFF SIDE is a routing question with the trap already in the contract: A SEED IS A WEAK CARRIER AND ONLY MEANS ANYTHING WITHIN ONE LANE, so planHandoff DROPS it across a lane change and records why - carrying it would look like continuity and be nothing. Carriers a lane cannot accept are dropped with reasons, and dropping the reference pack says plainly that nothing else on the list replaces it. assertHandoffCarriesSomething refuses a plan whose only survivors are weak or non-structural, because it would otherwise report a carrier list that looks non-empty while carrying no identity at all.                                                                                                          | 22 tests pass, over the 182.C.09 carrier contract and the 182.C.18.02 aligner residual. No new measurement and no provider spend.                         | —    | Environment and time-of-day continuity are declared and never verified against pixels; nothing staged can do that (182.C.19.02).                                        |
| 182.C.22.04 | Add edit handles and transitions without trimming phonemes, duplicating words …                                      | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-cut-safety.ts` (new), `human-video-cut-safety.spec.ts` (new)                                                                                                   | 182.C.22.01 and .04 are ONE QUESTION - where can a boundary go - and were built together. THE FINDING THAT DECIDES IT: YOU CANNOT CUT AT AN ALIGNED WORD END. 182.C.18.02 measured word ONSETS carrying a +48.3 ms systematic bias that a constant corrects, leaving +/-32.87 ms of residual, and word ENDS 96 to 141 ms EARLY consistently, because CTC emits a token before the sound finishes. So a cut at an aligned word end lands up to 141 ms INSIDE the word and takes the end of it - the phoneme trim this item forbids, arrived at by doing the obvious thing. The spec asserts that a cut 100 ms AFTER an aligned end is still refused. WHAT IS SAFE INSTEAD is a measured silence gap, and THE MARGIN IS NOT A GUESS: 32.87 + 141 = 174 ms of uncertainty about where speech actually stops, so a gap must exceed TWICE that before a cut fits with room either side, and the boundary is the gap's MIDPOINT rather than an edge because the edges are exactly where the uncertainty is. Segmentation takes the LAST safe boundary under a provider's duration limit rather than the nearest to it, because a boundary 20 ms under the limit is one provider rounding away from being over it; when no gap falls in range the segmenter REFUSES and says the script cannot be cut for this provider without re-recording or a longer limit, rather than cutting mid-word. DUPLICATING WORDS is the cross-fade failure: an overlap inside a gap duplicates silence nobody hears, and one reaching a word replays it - 182.C.18.04 measured that signature at waveform autocorrelation 1.000 against 0.261 for clean speech, so it WOULD be caught downstream, after the render, having been paid for. A/V TIMESTAMP CONTINUITY is compared in INTEGER TICKS rather than floating seconds: a container stores integer ticks, that is what a demuxer compares, and a seconds comparison fails on representation alone - 3.5 + 1/90000 does not subtract back to exactly 1/90000, and a correct one-tick join was refused until this was fixed. The tolerance is one tick (11 microseconds at 90 kHz), because a millisecond tolerance would pass a 90-tick gap.                                                                                                                                                                                                                                                      | 21 tests pass, over the 182.C.18.02 calibration this repo measured. No new measurement and no provider spend.                                             | —    | Sentence and performance-beat segmentation take the boundaries this provides; the beat detection itself is a script-side concern (182.C.22.02 ledger).                  |
| 182.C.22.05 | Reconcile provider clips with CFR/VFR, time base, color, resolution, grain, motion cadence …                         | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-clip-reconciliation.ts` (new), `human-video-clip-reconciliation.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-clip-reconciliation/` (new) | The item lists seven differences and they do NOT behave the same way - the split is the whole content. FORMAT properties conform away: frame rate, time base, pixel format, colour tagging and resolution are all 182.C.16 planConform targets. PERCEPTUAL properties DO NOT, and they are what makes a cut visible. MEASURED: a clean pan and the same pan with grain, conformed through h264/yuv420p/30fps, read flat-region noise 0.854 against 3.368 AFTERWARDS - a ratio of 3.94, essentially unchanged from the 3.82 before. The conform normalised the container and the codec and removed NONE of the difference. AND HOW GRAIN IS MEASURED MATTERS, because the obvious way does not work: a WHOLE-FRAME high-frequency residual reads 13.078 for the clean clip and 13.597 for the grainy one, a ratio of 1.04, because image detail (hair, fabric) dominates it; measured in the flattest 10% of 16x16 patches, where there is no detail to dominate, the same pair reads 1.021 and 3.896. That negative result is carried in GRAIN_MEASUREMENT_NOTE so nobody rebuilds it. MOTION CADENCE likewise survives: a 3:2 (24-in-30) clip gives 9 identical frame pairs with a longest run of 2 against 0 for a native clip - the judder signature 182.C.19.06 measured - and the duplicated frames ARE the clip, so conforming to 30 fps re-encodes them rather than removing them. So reconcile() returns format gaps and perceptual gaps as SEPARATE lists, plus a third list of perceptual properties nobody measured, and conformSufficient() answers the question a planner actually asks before routing two shots to two providers. assertPerceptuallyReconciled is NOT a prohibition on cutting between providers - a caller can accept the gap or apply a matching pass - it is a refusal to let a CONFORM stand in for that decision, which is what 'reconcile' quietly becomes when the format gaps are the only ones anybody looked at. Unmeasured is refused FIRST, because a property nobody measured is not one that matched and it could be either.                                                                                                                                                                                                                                                                                                                                                       | 3 sequences measured before and after a real h264 conform; `evidence/2026-09-11-clip-reconciliation/measurements.json`. 13 tests pass. No provider spend. | —    | The grain ratio and ambience thresholds are provisional - nothing has fitted them against a viewer. CFR/VFR itself is handled by 182.C.16 timestamp verification.       |
| 182.C.22.06 | Apply room tone, music, effects, and mix automation after dialogue approval while preserving the speech reference …  | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-assembly-evaluation.ts` (new), `human-video-assembly-evaluation.spec.ts` (new)                                                                                 | 182.C.22.06 and .07 share ONE PRINCIPLE, the same one 182.C.21.08 applies to corrections: A VERDICT IS ABOUT THE ARTIFACT IT WAS MEASURED ON. Mixing produces a new artifact and so does assembling, and neither inherits the approvals of its parts - for reasons that are MEASURED rather than cautious. .06: 182.C.18.08 measured a harmonic bed 5 dB under the speech leaving the CTC aligner and the pronunciation measure COMPLETELY unaffected (52.9 ms onset error at every depth, minGop exactly 0.0) while the percentile-split SNR estimator read 16.8 dB and tripped the 20 dB gate. So the mixed master is NOT re-measurable on the same thresholds: sync still works on it and SNR does not. The speech reference must survive as its OWN asset and assertSpeechReferencePreserved refuses a mix that overwrote it, because the reference is the only thing the approved sync numbers describe. The other half is 182.C.18.05, which measured loudness normalisation ERASING a 15-sample clipping run entirely - so the defect report must come from BEFORE the mix, and assertDefectsPreMix refuses one produced from the master. GATES_INVALIDATED_BY_MIX is carried BESIDE GATES_SURVIVING_MIX, with loudness invalidated in the OTHER direction (the mix is what sets delivered loudness, so a pre-mix verdict describes a file nobody ships); a proof listing only survivors reads as a complete set and the reader has no way to know SNR was measured on different audio. .07: THREE THINGS FALL THROUGH THE GAP BETWEEN SHOTS. Edit boundaries are inside NO shot, so every per-shot pass can hold while every join is broken. The master's encoding and loudness are properties of the whole file - ten shots each at -16 LUFS do not assemble to -16 LUFS unless their durations happen to be equal. And the adversarial one: 182.C.20.05 measured that segments shorter than the evaluator's 1.0 s search span produce windows that all pass because none could FAIL, so a sequence of short shots each passing is that evasion assembled - and the assembly is where it becomes visible, because one short shot is ordinary and twelve in a row is a pattern. assertSequenceEvaluated reports every hole rather than the first, because closing one leaves the others, and it refuses a multi-shot assembly with NO boundary recorded at all: the joins exist whether or not anything measured them. | 21 tests pass, over the 182.C.18.05 and 182.C.18.08 measurements this repo made. No new measurement and no provider spend.                                | —    | The per-boundary evaluation itself is 182.C.22.05 (reconciling clips); this is the rule that a boundary must be evaluated and the shape check that catches the evasion. |
| 182.C.22.07 | Evaluate each shot, every edit boundary, and the assembled master; a per-shot pass must not automatically pass the … | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-assembly-evaluation.ts` (new), `human-video-assembly-evaluation.spec.ts` (new)                                                                                 | 182.C.22.06 and .07 share ONE PRINCIPLE, the same one 182.C.21.08 applies to corrections: A VERDICT IS ABOUT THE ARTIFACT IT WAS MEASURED ON. Mixing produces a new artifact and so does assembling, and neither inherits the approvals of its parts - for reasons that are MEASURED rather than cautious. .06: 182.C.18.08 measured a harmonic bed 5 dB under the speech leaving the CTC aligner and the pronunciation measure COMPLETELY unaffected (52.9 ms onset error at every depth, minGop exactly 0.0) while the percentile-split SNR estimator read 16.8 dB and tripped the 20 dB gate. So the mixed master is NOT re-measurable on the same thresholds: sync still works on it and SNR does not. The speech reference must survive as its OWN asset and assertSpeechReferencePreserved refuses a mix that overwrote it, because the reference is the only thing the approved sync numbers describe. The other half is 182.C.18.05, which measured loudness normalisation ERASING a 15-sample clipping run entirely - so the defect report must come from BEFORE the mix, and assertDefectsPreMix refuses one produced from the master. GATES_INVALIDATED_BY_MIX is carried BESIDE GATES_SURVIVING_MIX, with loudness invalidated in the OTHER direction (the mix is what sets delivered loudness, so a pre-mix verdict describes a file nobody ships); a proof listing only survivors reads as a complete set and the reader has no way to know SNR was measured on different audio. .07: THREE THINGS FALL THROUGH THE GAP BETWEEN SHOTS. Edit boundaries are inside NO shot, so every per-shot pass can hold while every join is broken. The master's encoding and loudness are properties of the whole file - ten shots each at -16 LUFS do not assemble to -16 LUFS unless their durations happen to be equal. And the adversarial one: 182.C.20.05 measured that segments shorter than the evaluator's 1.0 s search span produce windows that all pass because none could FAIL, so a sequence of short shots each passing is that evasion assembled - and the assembly is where it becomes visible, because one short shot is ordinary and twelve in a row is a pattern. assertSequenceEvaluated reports every hole rather than the first, because closing one leaves the others, and it refuses a multi-shot assembly with NO boundary recorded at all: the joins exist whether or not anything measured them. | 21 tests pass, over the 182.C.18.05 and 182.C.18.08 measurements this repo made. No new measurement and no provider spend.                                | —    | The per-boundary evaluation itself is 182.C.22.05 (reconciling clips); this is the rule that a boundary must be evaluated and the shape check that catches the evasion. |
| 182.C.22.08 | Support extension/edit requests that lock approved frames and dialogue regions and clearly identify provider-side …  | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-edit-scope.ts` (new), `human-video-edit-scope.spec.ts` (new)                                                                                                   | THE THING THAT MAKES .08 HARD IS NOT THE LOCK - it is that A LOCK IS A CLAIM ABOUT WHAT SOMEBODY ELSE WILL NOT TOUCH. We send a request saying 'leave frames 0-120 and the dialogue between 2.1 and 5.4 seconds alone' and the PROVIDER decides what it actually does; nothing here prevents an inpainting model from repainting a locked frame. So a lock is not enforced, it is DECLARED and then VERIFIED, reusing the before/after frame diff 182.C.21.09 already built for corrections and pointing it at a different question. assertLocksVerified refuses an edit result carrying no post-edit comparison, for the same reason 182.C.21.08 refuses an inherited gate verdict: a lock nobody verified is a lock nobody has. ANY change counts as a violation rather than a large one - a provider that re-encoded a locked frame into slightly different pixels has re-encoded a frame somebody approved. THE SCOPE IS STATED IN BOTH DIRECTIONS, which is what 'clearly identify provider-side inpainting or regeneration scope' actually needs: not a label on the request but the MEASURED scope after the fact, because a provider that says it inpainted a mask and in fact re-rendered the whole clip has told the truth about its INTENT and not about its OUTPUT, and only the diff separates them. The ceilings are deliberately loose - an inpainting model legitimately spills at the blend edge - because the point is catching the re-render, not bounding the spill. .09 IS WHAT MAKES ANY OF THIS POSSIBLE: a lock is verified against a previous version, so the previous version has to exist, and versionsOf / latestApproved / alternateTakes give resume, versioning and alternate takes over one history. AND ROLLBACK IS NOT JUST THE ARTIFACT - the half that gets forgotten. Rolling a shot back to v2 while the continuity ledger still records what v3 established leaves the ledger claiming something the restored shot does not establish, and the NEXT shot is then checked against a fiction. rollback() returns the ledger state to restore alongside the asset and names which aspects revert, and assertRollbackComplete refuses an application that moved only the file or left downstream approvals standing against a version that no longer exists.                                                                                                                                | 19 tests pass, reusing the 182.C.21.09 frame-diff shape. No new measurement and no provider spend.                                                        | —    | Locks are detected after the fact, never enforced: no provider API here offers a guarantee, and the module says so rather than implying one.                            |
| 182.C.22.09 | Provide shot-level resume, versioning, alternate takes, approvals, and rollback.                                     | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-edit-scope.ts` (new), `human-video-edit-scope.spec.ts` (new)                                                                                                   | THE THING THAT MAKES .08 HARD IS NOT THE LOCK - it is that A LOCK IS A CLAIM ABOUT WHAT SOMEBODY ELSE WILL NOT TOUCH. We send a request saying 'leave frames 0-120 and the dialogue between 2.1 and 5.4 seconds alone' and the PROVIDER decides what it actually does; nothing here prevents an inpainting model from repainting a locked frame. So a lock is not enforced, it is DECLARED and then VERIFIED, reusing the before/after frame diff 182.C.21.09 already built for corrections and pointing it at a different question. assertLocksVerified refuses an edit result carrying no post-edit comparison, for the same reason 182.C.21.08 refuses an inherited gate verdict: a lock nobody verified is a lock nobody has. ANY change counts as a violation rather than a large one - a provider that re-encoded a locked frame into slightly different pixels has re-encoded a frame somebody approved. THE SCOPE IS STATED IN BOTH DIRECTIONS, which is what 'clearly identify provider-side inpainting or regeneration scope' actually needs: not a label on the request but the MEASURED scope after the fact, because a provider that says it inpainted a mask and in fact re-rendered the whole clip has told the truth about its INTENT and not about its OUTPUT, and only the diff separates them. The ceilings are deliberately loose - an inpainting model legitimately spills at the blend edge - because the point is catching the re-render, not bounding the spill. .09 IS WHAT MAKES ANY OF THIS POSSIBLE: a lock is verified against a previous version, so the previous version has to exist, and versionsOf / latestApproved / alternateTakes give resume, versioning and alternate takes over one history. AND ROLLBACK IS NOT JUST THE ARTIFACT - the half that gets forgotten. Rolling a shot back to v2 while the continuity ledger still records what v3 established leaves the ledger claiming something the restored shot does not establish, and the NEXT shot is then checked against a fiction. rollback() returns the ledger state to restore alongside the asset and names which aspects revert, and assertRollbackComplete refuses an application that moved only the file or left downstream approvals standing against a version that no longer exists.                                                                                                                                | 19 tests pass, reusing the 182.C.21.09 frame-diff shape. No new measurement and no provider spend.                                                        | —    | Locks are detected after the fact, never enforced: no provider API here offers a guarantee, and the module says so rather than implying one.                            |
| 182.C.22.10 | Benchmark continuity decay and determine maximum reliable sequence length per route before enabling automatic …      | Isis  | 182.C.21   | implemented | `apps/oshun/bff/src/generation/human-video-continuity-decay.ts` (new), `human-video-continuity-decay.spec.ts` (new)                                                                                       | The SHAPE of this measurement is what matters, and it is not the shape most gates in this phase have. DECAY IS A CURVE, NOT A THRESHOLD: continuity does not fail at shot 7, it degrades, and the question is where the degradation crosses a tolerance - so a single pass/fail per sequence throws away the information the benchmark exists to produce. What is measured is identity as a function of shot index, plus the rate the continuity ledger picks up contradictions. AGAINST THE FIRST SHOT, NOT THE PREVIOUS ONE, and this is the part that decides whether the benchmark measures anything: shot 8 compared against shot 7 can look perfect at every step while shot 8 and shot 1 are different people - that is what drift IS, and comparing neighbours is exactly how it becomes invisible. assertComparedAgainstFirst REFUSES a neighbour-compared benchmark, because its curve is flat by construction. AND 'THE BENCHMARK ENDED' IS NOT A MAXIMUM: a run of six shots that never crossed the tolerance says those six are fine and says NOTHING about the seventh, so it is returned as its own `limitedBy` value rather than as a number - reporting it as a maximum would make the length of the EXPERIMENT a property of the ROUTE. A compounding estimate is provided for a route with a per-handoff retention figure (182.C.19.08 measured retention per condition, and a chain applies its own degradation at each link, so 0.9 per link is 0.9^n), and it is LABELLED an estimate with the assumption it rests on - independent, equal degradation per link - which no measurement here supports. THE STATE TODAY IS THAT NO ROUTE HAS A BENCHMARK, so automatic long-form is off for every route and assertAutoLongFormAllowed refuses; that is the correct reading of having measured nothing, and the alternative is a default length somebody guessed that would be indistinguishable in the code from a measured one. decayCoverage reports it rather than inferring from an absence of failures: a route nobody ran a long sequence on has no failures, which is not the same as holding.                                                                                                                                                                                                                                                                                                      | 14 tests pass. Benchmark shape and gate; the benchmark runs themselves need generated sequences (spend) and no route has one.                             | —    | No route is benchmarked, so automatic long-form is off everywhere. Running the benchmarks needs real multi-shot sequences per route (spend).                            |

### 182.C.23 Studio Creator Experience

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                                | Owner      | Dependency  | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Cost                                                                               | Blocker                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |
| ----------- | ----------------------------------------------------------------------------------------------------------------------- | ---------- | ----------- | ----------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.23.01 | Design a dedicated “Directed Human Video” flow distinct from simple text-to-video and talking-head avatar cards.        | Studio Web | 182.C.21    | implemented | `docs/domains/isis/human-video/studio-flow-design.md` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | THE THREE FLOWS ARE NOT A DIFFICULTY LADDER, THEY DIFFER IN WHAT THEY PROMISE - text-to-video promises neither the words nor the person, a talking-head avatar promises the script spoken by AN AVATAR, and this one promises THE APPROVED AUDIO EXACTLY, SPOKEN BY THIS AUTHORIZED PERSON. 182.C.13.03 already refuses to ROUTE across that boundary; the flow job is to refuse the DRIFT - a creator who starts in one and gradually acquires the other promise by filling in fields. SO IT IS A DIFFERENT CARD KIND, NOT A MODE TOGGLE: there can be no control that turns a text-to-video card into this one, BECAUSE THE THING THAT WOULD HAVE TO APPEAR IS SOMEBODY CONSENT. THE STRUCTURAL DIFFERENCE IS THAT THIS FLOW WAITS FOR ANOTHER PERSON: every other generation card is a form the creator fills and submits, and this one has a gate in the middle THAT IS NOT THEIRS TO CLOSE. Three consequences a wizard would design away: the WAITING STATE IS NOT AN ERROR ("ready, waiting for Ama" is a legitimate resting place, possibly for days); PROGRESS IS NOT LINEAR (adding a reference after enrolment is a NEW enrolment, so the surface shows what is COVERED rather than what is uploaded); and WITHDRAWAL CAN ARRIVE AT ANY POINT, including after generation. A spinner here produces the support ticket "it is stuck". AND YOU CANNOT START THIS FLOW FROM A PROMPT - the first input is a cast, because starting from an idea and attaching a face later is PRECISELY HOW AN UNAUTHORIZED LIKENESS ENTERS: the prompt is written, the work feels underway, and consent becomes the thing standing between somebody and a finished piece. A creator with no authorized cast has exactly one action: invite somebody. Five regions in the order the DATA forces, four never-do rules, and the four DIFFERENT empty states a new project has.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | The promise table is grounded in 182.C.13.03 capability classes and 182.C.31.09 flow guarantees; the ordering constraint (audio approval fixes the durations the storyboard must fit) from creator-workflow.md; the enrolment-is-not-an-edit rule from 182.C.26.02; the pre-submission refusals from the shipped 182.C.23.06; and acceptance-as-a-deciding-operation from 182.C.25.03. Both cross-document links resolve. Prettier-formatted. No code changed.                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | No provider spend; no model call.                                                  | A DESIGN, NOT A SURFACE. No Studio route exists for this flow, and the BFF routes it would call are unmounted. The component inventory and interaction detail of regions 2-4 are 182.C.23.02-.05 and stay open; 182.C.23.13 (frontend-skill convention plus the four states) is the item that must be confirmed BEFORE implementation.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
| 182.C.23.02 | Add drag/drop reference sets with role, cast member, priority, quality feedback, consent state, crop …                  | Studio Web | 182.C.21    | implemented | `apps/oshun/web/src/components/isis/human-video/ReferenceSet.tsx`, `ReferenceSet.spec.tsx`, and now the region that feeds it: `apps/oshun/web/src/app/studio/human-video/page.tsx` (new), `apps/oshun/web/src/components/isis/human-video/HumanVideoStudioRoute.tsx` (new), `HumanVideoStudio.tsx` (new), `projectState.ts` (new), `projectPort.ts` (new), plus the four specs, plus the draft API: `apps/oshun/bff/src/generation/human-video-project-draft.ts` (new), `human-video-project-handlers.ts` (new), `human-video-api-surface.ts` (3 endpoints), `libs/contracts/src/human-video/failure-codes.ts` (`hv.validation.draft_conflict`), `sdks/human-video-ts/client.ts` (regenerated), plus the ingestion half 2026-09-12: `apps/oshun/web/src/components/isis/human-video/uploadPort.ts` (new), `uploadPort.spec.ts` (new), `projectState.ts` (the `reference/add` edit), `projectPort.ts` (`uploadSessionId` through the round trip), `ReferenceSet.tsx`, `HumanVideoStudio.tsx`, `HumanVideoStudioRoute.tsx` and their specs, `apps/oshun/web/src/lib/api-client.ts` (bytes are sent as bytes), `apps/oshun/bff/src/generation/human-video-project-draft.ts` (`DraftReference.uploadSessionId`) | THE CAST REGION REFERENCE SET, and FIVE of the seven things the item names are not the control they sound like. DRAG AND DROP IS AN AFFORDANCE, NOT A CONTROL - `frontend-convention.md` states the rule, so the drop zone is decoration over a REAL file input and reordering is buttons with the step in the aria-label, because a control that exists only as a drag is a control some people cannot use. "WHICH CAST MEMBER" IS THE WRONG QUESTION FOR A TWO-FACE REFERENCE: the detector has already found the faces and what is missing is WHICH FACE is the cast member, so the control there is a FACE PICKER and it appears only on an ambiguous reference; a dropdown of cast names records an answer to a question nobody asked. PRIORITY ONLY MATTERS WHEN A LANE TAKES FEWER REFERENCES THAN YOU GAVE, so the note says what it decides and says it is NOT a quality ranking - a creator who believes it is deletes the ones at the bottom - and says "order decides nothing" outright when every route takes them all. A CROP IS APPLIED AT USE TIME AND NEVER TO THE STORED REFERENCE, because an authorization binds to the DIGEST and cropping the file produces different bytes the permission would silently stop covering. CONSENT IS DISPLAYED AND NOT EDITABLE - there is no control for it at all, asserted by the test - and a reference waiting on enrolment renders as `LPending`. AND AN ELIGIBILITY INDICATOR IS ABOUT TODAY ROUTES: it counts the routes that can take this reference and names the constraint, and the word "unsupported" appears nowhere, because it reads as a permanent property of the image. THE LAST GAP CLOSED 2026-09-12. Adding was disabled with the reason attached because the surface had no endpoint to send a chunk to; `182.C.06.01` built one, and this is the client half. A REFERENCE IS NOT ADDED UNTIL THE BYTES HAVE STOPPED: the digest an authorization binds to exists only at completion, so a row written when the file was chosen would name an upload whose bytes can still change under it. THE ANSWER IS PER FILE, NOT PER DROP - six photographs, the fourth a screenshot of an error page, and refusing the whole drop throws away five good uploads while saying nothing about which was wrong. THE DRAFT COULD NOT POINT AT ITS OWN BYTES: `DraftReference` carried a role, a cast member and a crop and NO UPLOAD, so a reference was a row depicting nothing and no pack could be built from it; `uploadSessionId` is optional rather than required-and-nullable because that is what an older stored document actually contains. PROGRESS IS WHAT THE SERVER SAYS IT KEPT, never what the client sent, and a chunk that is accepted while storing nothing STOPS the loop rather than spinning against a service that is answering. AND THE WEB API CLIENT WAS JSON.stringify-ING EVERY BODY: a Uint8Array became `{"0":137,...}`, several times the size, which no byte parser can read - this is the first caller in the app to send a body that is not JSON. | 18 component tests (a new one: a disabled drop zone swallows neither the file nor the reason); 136 tests across the flow; chromium at 1280/768/400 reports 0 axe violations and no undersized target. 13 upload-port tests (the multi-chunk walk, the resume from a repeated open, the per-chunk idempotency key, the four service failures kept apart, and the two shapes that must NOT count as an uploaded reference: no session id, and a completion with no digest). 4 route tests including that one bad file does not lose the good ones. 193 web tests green; 5410 BFF tests green. Typechecked against `tsconfig.typecheck.json` on both apps with a live control error; eslint clean.                                                                                                                                                                                                                                     | —                                                                                  | The pack is not declared from these references yet — that is the next adapter, and it needs the cast mapping the region already collects. And a real session is still refused by the surface: nothing mints an `hv:` scope, which is an entitlement decision.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.23.03 | Add script/speaker editing, pronunciation tools, voice selection/upload, TTS preview, word timing …                     | Studio Web | 182.C.21    | partial     | `ScriptAndVoice.tsx`, `ScriptAndVoice.spec.tsx`, and now the region that feeds it: `apps/oshun/web/src/app/studio/human-video/page.tsx` (new), `apps/oshun/web/src/components/isis/human-video/HumanVideoStudioRoute.tsx` (new), `HumanVideoStudio.tsx` (new), `projectState.ts` (new), `projectPort.ts` (new), plus the four specs, plus the draft API: `apps/oshun/bff/src/generation/human-video-project-draft.ts` (new), `human-video-project-handlers.ts` (new), `human-video-api-surface.ts` (3 endpoints), `libs/contracts/src/human-video/failure-codes.ts` (`hv.validation.draft_conflict`), `sdks/human-video-ts/client.ts` (regenerated), plus the take read and the approval 2026-09-12: `apps/oshun/web/src/components/isis/human-video/audioPort.ts` (new), `audioPort.spec.ts` (new), `HumanVideoStudio.tsx`, `HumanVideoStudioRoute.tsx`, `projectState.ts` (the `audio/approve` edit REMOVED), `apps/oshun/web/src/lib/api-client.ts` (`onHeaders`)                                                                                                                                                                                                                                        | THE SCRIPT AND VOICE REGIONS. A ONE-WORD SCRIPT EDIT VOIDS THE AUDIO AND THE EDITOR SAYS SO BEFORE THE EDIT - `182.C.23.11` established the asymmetry (a reference change leaves the audio alone; a script change does not), and an editor that accepts the keystroke and reports the damage in a toast HAS ALREADY DONE IT, so a line with work under it is read-only until the creator takes a deliberate act that names what it costs. "APPROVED" IS A PROPERTY OF BYTES, NOT OF A LINE: `assertSpeechApprovedForVideo` separates "no approval" from "approved for DIFFERENT audio", and on a screen both look like a ticked line, so the third state is rendered explicitly and the tick is never stale - this is exactly what happens after somebody fixes a pronunciation and re-synthesises. THE LEAD-IN A CREATOR HEARS IS NOT PART OF WHAT IS MEASURED: the approval binds to `speechDigestSha256`, the SPEECH and not the disclosed delivery, so the screen says the disclosure seconds are neither in the duration the storyboard is fitted to nor in what sync is scored against. A TRANSCRIPT THAT IS UNAVAILABLE IS NOT ONE THAT MATCHED, rendered through the three-outcome primitive. A PRONUNCIATION IS A DELIVERY INSTRUCTION, NOT A SCRIPT EDIT - respelling a name in the script would change what the delivered-word check compares against, because the script is the promise. And a CREATOR UPLOAD IS CONSENTED TOO: there is no path that skips the authorization, because a recording of a real person is that person voice whether or not a model was involved. Pace is stated as computed rather than chosen, and word timings that were ESTIMATED are named individually. THE APPROVAL IS A RECORD NOW, AND UNTIL 2026-09-12 THE CONTROL ONLY EDITED LOCAL STATE. The draft carries no approval field, so the click set something that VANISHED ON RELOAD - the "looks like it worked" outcome this control's own comment names as the worst of the three options, sitting in the code the comment is attached to. `approveAudioTake` sends the real thing with an `If-Match` carrying the tag the take was READ at, because approving is a DECIDING operation: between the read and the decision the aligner can have been re-run, and `182.C.21.08` calls an approval inherited across a re-measurement an inherited verdict. A 412 says LOOK AGAIN rather than retry, because retrying with the same tag fails forever. The take is RE-READ afterwards rather than painted optimistically: the service decides the approver, the time, and whether the approval was allowed at all. AND THE `audio/approve` EDIT IS GONE - its absence is the point, because a reducer that can set an approval is a reducer through which a local claim can re-enter. The ETag reaches the client through a new `onHeaders` callback on the shared API client, which is additive and names the headers a caller wants rather than handing over all of them.                                                                        | 18 component tests (a new one: approving is refused, not silently ignored, when nothing can record it); the invalidation warning is computed from the live project rather than stored. BROWSER-VERIFIED END TO END on the real dev server against a stub service: create loads the project, a script edit saves once per burst (three bursts, three PATCHes), and a concurrent change returns the conflict with the creator's typing intact. 14 audio-port tests including the three approval ones (the tag is sent, the take is read back, and a 412 says look rather than retry). 2 Studio tests rewritten: the approval goes to the SERVICE and nothing is written into the draft, and a take that was not read from the service offers no approval at all because there is no version to decide against. 215 web tests green. Two failures in `src/lib` are PRE-EXISTING and unrelated - reproduced with these changes stashed. | —                                                                                  | STILL UNCHECKED, and the one gap is a TTS preview — which is a DECISION rather than effort, measured 2026-09-12. `human-video-speech-source.ts` already reuses the narration path rather than rebuilding one, and `narration-provider-env.ts` offers exactly two: ElevenLabs (billed, needs a key and a spend decision) and Kokoro THROUGH COMFYUI (`resolveComfyVoiceProvider`), which needs a running Comfy server this box does not have. The Kokoro-82M WEIGHTS are cached under `~/.cache/huggingface` and the `kokoro`, `soundfile` and `misaki` PACKAGES ARE NOT INSTALLED in the human-video venv, so a local CPU synthesiser is a new tool plus a new provider plus a python dependency on a shared box. Three routes, all of which cost somebody something; picking one is not this item's call. Everything else in the item works as of 2026-09-12: script and speaker editing, pronunciation, voice upload, word timing, and explicit approval against the version the take was read at. |
| 182.C.23.04 | Add storyboard/timeline controls for shots, actions, blocking, camera, expression, gaze, gesture, props, dialogue …     | Studio Web | 182.C.21    | implemented | `DirectionTimeline.tsx`, `DirectionTimeline.spec.tsx`, and now the region that feeds it: `apps/oshun/web/src/app/studio/human-video/page.tsx` (new), `apps/oshun/web/src/components/isis/human-video/HumanVideoStudioRoute.tsx` (new), `HumanVideoStudio.tsx` (new), `projectState.ts` (new), `projectPort.ts` (new), plus the four specs, plus the draft API: `apps/oshun/bff/src/generation/human-video-project-draft.ts` (new), `human-video-project-handlers.ts` (new), `human-video-api-surface.ts` (3 endpoints), `libs/contracts/src/human-video/failure-codes.ts` (`hv.validation.draft_conflict`), `sdks/human-video-ts/client.ts` (regenerated), plus the take read 2026-09-12: `apps/oshun/web/src/components/isis/human-video/audioPort.ts` (new), `audioPort.spec.ts` (new), `projectState.ts` (the `unrecorded` reason), `HumanVideoStudioRoute.tsx` and their specs                                                                                                                                                                                                                                                                                                                          | THE DIRECTION REGION. Nine of the ten things the item names are direction; the tenth decides whether the take can be CHECKED. A PLAN THAT PUTS EVERY LINE OFF-CAMERA IS A PLAN THAT CANNOT BE CHECKED - sync is scorable in exactly ONE visibility, and `coverageOf` divides by the TOTAL rather than by the scorable ones precisely so such a plan does not score perfectly, so the running coverage sits IN the timeline and moves as the creator works: meeting that floor only after the render means learning it from a rejection. THE REASON IS REQUIRED AT THE MOMENT OF THE CHOICE, because `assertReasonGiven` is re-enforced at delivery for a stated reason - a plan can be edited after a render, and a reason written afterwards to explain a bad take is what the rule exists to prevent - so the field appears on `deliberately-occluded` and `back-to-camera` and nowhere else. APPROVAL ALREADY FIXED THE DURATIONS, so a window past the approved audio is not a shot that is slightly wrong but one that CANNOT EXIST, refused in the control rather than collected for a later pass. A SILENT SHOT ROLE WITH A SPEECH WINDOW IS TWO INSTRUCTIONS THAT CANNOT BOTH BE FOLLOWED. Problems are RETURNED rather than thrown, because a half-written plan is the normal state and stopping at the first hides the second. And EVERY BOUNDARY TAKES A NUMBER AS WELL AS A DRAG: numeric entry bounded by the approved audio, plus a nudge whose 0.05s step is in the aria-label, and each non-scorable visibility says "sync not measurable" in the option itself. THE BLOCK IS GONE 2026-09-12: `GET /audio/:audioId` is bound and `audioPort.ts` reads it, so a take's duration and approval reach the region instead of the region waiting on a read nobody could do. BUILDING IT FOUND A DEFECT IN THE CHECK THAT GUARDS THE WHOLE IDEA: `segmentSpeaksCurrentScript` is `spokenLines.every(...)`, and `every` ON AN EMPTY ARRAY IS TRUE — so a segment with NO RECORD of what its bytes say passed the very test that exists to catch a script edit outrunning the audio, which is the failure the field's own comment describes. It is a fourth not-fixed reason now, `unrecorded`, kept apart from `superseded` because the instructions differ: a superseded take needs re-recording and an unrecorded one needs something to READ it, and telling a creator to re-record a perfectly good take costs them a studio session. Filling `spokenLines` from the script would have been the tempting shortcut and is exactly the failure: the check would compare the script against itself and pass forever.                                                                                                                                                                                                                                                                                                                                                                                                                     | 16 component tests (a new one: with no floor from the server the coverage is shown as a count and the verdict is withheld); the approved duration reaching the timeline is `fixed` or the region is blocked. 11 audio-port tests, including the THREE answers kept apart (a take that came back, a take the service has never heard of, and a read that failed — the middle one is a fact about the segment and the last about the service). 4 new projectState tests for the vacuous-truth fix, including that the region tells the creator NOT to re-record. 3 route tests including that the take reads do not hold the first paint. 211 web tests green; typecheck and eslint clean with a live control error.                                                                                                                                                                                                                  | —                                                                                  | None for this item. Approving a take from this surface is 182.C.23.03's clause and needs the ETag the API client cannot currently see.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
| 182.C.23.05 | Provide simple presets with an expandable advanced panel; do not expose raw provider JSON as the primary control …      | Studio Web | 182.C.21    | implemented | `apps/oshun/bff/src/generation/human-video-presets.ts` (new), `human-video-presets.spec.ts` (new), `human-video-preset-handlers.ts` (new), `human-video-api-surface.ts` (`get-preset-catalogue`, the 29th operation), `human-video-handlers.ts`, `human-video-handlers.spec.ts`, `human-video-runtime.ts` (boot check + boot line), `human-video-runtime.spec.ts`; `apps/oshun/web/src/components/isis/human-video/presetPort.ts` (new), `presetPort.spec.ts` (new), `PresetPanel.tsx`, `PresetPanel.spec.tsx`, `projectState.ts` (the `presets` region), `projectState.spec.ts`, `HumanVideoStudio.tsx`, `HumanVideoStudioRoute.tsx` and their specs, `apps/oshun/web/e2e/human-video-studio.spec.ts`; `docs/domains/isis/human-video/api-guide.md`, `frontend-convention.md`                                                                                                                                                                                                                                                                                                                                                                                                                              | THE CATALOGUE IS SERVED NOW, AND SERVING IT RESTATED THE ITEM'S PROHIBITION AS A RULE. Raw provider JSON is forbidden NOT BECAUSE IT IS JSON but because nothing downstream knows what it asked for, so no delivered take can be checked against it — and written that way the rule reaches past the textarea it was aimed at: A NAMED DROPDOWN NOTHING CAN SCORE IS THE SAME FAILURE IN NICER CLOTHING. So every setting declares the `DirectionAspect` it moves, taken from the ten scores in `human-video-direction-adherence.ts` rather than invented, and `assertPresetCatalogueCoherent` REFUSES a setting whose aspect is one of the three `JUDGEMENT_ASPECTS`. THOSE THREE ARE SERVED AS DELIBERATELY ABSENT WITH THEIR MEASURED REASONS: 52 blendshapes come back from a face and mapping them onto "warm" rather than "amused" is not a measurement, so a mood dropdown is exactly the forbidden control arrived at politely; omitting them silently is worse, because an absent control reads as an oversight and the fix somebody reaches for next quarter is a free-text box. A PRESET'S AVAILABILITY IS NOT A PROPERTY OF THE PRESET: how many routes can serve it is one request against the lane registry at one moment, which `182.C.23.06` answers just before submit, so the catalogue carries NO count and `eligibleRoutes` is NULL on the wire — `0` would be a claim made from a registry nobody read, and it DISABLES the control. AND A PRESET CANNOT CLAIM ITS OWN SCORABILITY: `syncScorable` is derived from the preset's own speech visibility against `SYNC_SCORABLE_VISIBILITIES`, so `voiceover-over-action` states that a take from it has NO lip-sync score — not a poor one. THE BIGGEST FINDING WAS STRUCTURAL: the panel lived inside the DIRECTION region, which is BLOCKED until an approved audio duration exists, so the presets were invisible on the first screen a creator sees — gated on an upstream they do not depend on. Presets is its own region now, derived from the catalogue read alone and READY on a project with no cast, script or audio. The catalogue is a PROP rather than a project field for a second measured reason: as a field, a catalogue landing a beat after the project counted as a NEWER VERSION of it, and the draft reducer refuses a reseed over unsent edits — so a creator who had started typing was told their work blocked a refresh they never asked for.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | 18 BFF tests (the shipped table is coherent AND each of the five rules is driven by a case that fires it, asserting the refusal CODE rather than its prose); 25 panel tests; 10 web port tests; 5 region-derivation tests including the one that matters — presets is `ready` on a project where cast and direction are not; 176 web unit tests green overall; 8 BFF handler tests including the endpoint serving 200 on a deployment that bound NOTHING, and 401/403 proving it is still gated. Scratch typechecks clean on both apps; eslint clean. Chromium e2e re-run on a real stack with the new region.                                                                                                                                                                                                                                                                                                                      | —                                                                                  | None for this item. The three declined settings stay declined until something can score them; that is 182.C.17/19 work, not this one.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
| 182.C.23.06 | Show current eligible routes, capability limitations, retention warning, expected candidate count, price range …        | Studio Web | 182.C.21    | done        | `apps/oshun/bff/src/generation/human-video-pre-submission-summary.ts` (new), `human-video-pre-submission-summary.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | The item lists six things and TWO OF THEM CANNOT BE STATED THE WAY THE WORD SUGGESTS; the other four are assembly over 182.C.10.07's per-lane view. A PRICE RANGE OVER LANES SOME OF WHICH CANNOT BE PRICED IS NOT A RANGE: 182.C.10.05 found lanes priced per token and per frame-area that cannot be quoted before the request runs, so "$1.20 to $3.40" over three of five eligible routes is true about three and READ AS A CEILING by everybody who sees it, and a creator who sees a ceiling plans against it. The numbers still ship - they are true about the routes that priced - but unpricedLanes is a first-class field, isBound is separate from the numbers, and assertRangeBounds refuses at the point the figure is USED as a cap rather than where it is computed. LATENCY IS NOT AN ESTIMATE UNTIL THERE IS HISTORY: 182.C.15.10 fixed 20 finished runs and gave the reason, so laneLatency returns null under it, and assertLatencyNotBorrowed refuses the aggregation mistake - four routes with no history and one with forty produce a tempting "about 3 minutes" for the set, which is a measurement of one route presented as a property of five. All measured or none are both honest; the MIXTURE is what lies. THE EXPECTED TAKE COUNT IS THE SMALLEST BOUND, NOT THE REQUESTED NUMBER: 182.C.21.01 stops a fan-out at the first of count, spend and elapsed time, so "you will get 4" when the spend cap stops it at 2 is a number that was never going to happen; expectedCandidates returns the BINDING CONSTRAINT alongside the figure and returns null rather than a guess when a route cannot be priced, because rounding to the requested number quietly assumes the cap never binds. RETENTION SURVIVES AGGREGATION: four routes with recorded terms and one without is not "terms recorded" - the creator may be routed to the fifth, and what a provider keeps of a person's face is not a thing to discover afterwards; assertRetentionDisclosed refuses a summary that rounded to the best case. An unverified capability is listed as a LIMITATION and not as support, preserving the distinction 182.C.10.07 exists to keep.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | 16 tests pass; eslint clean. buildSummary keeps the two honest refusals as FIELDS rather than throwing - a summary that throws shows nothing - with the asserts exported for the places the numbers become bounds. TYPECHECK CORRECTED 2026-09-11: the original "targeted tsc" here extended `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so that run exited 0 having checked NOTHING. Re-run against `tsconfig.typecheck.json` with a deliberate control error included to prove the harness reports. Re-verified after the fix: 0 errors in this module and its spec, against 973 pre-existing errors elsewhere in the repo, so the filter is real rather than an empty run.                                                                                                                                                                                                                                       | No provider spend.                                                                 | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| 182.C.23.07 | Warn about small speaking faces, excessive motion, conflicting references, unsupported language …                       | Studio Web | 182.C.21    | done        | `apps/oshun/bff/src/generation/human-video-pre-generation-warnings.ts` (new), `human-video-pre-generation-warnings.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | The six warnings the item names, each carrying the measurement that produced it rather than an adjective. THE SPLIT THAT MATTERS: two of the six are not warnings at all. A face under 7% of frame height and a take shorter than one second do not produce a WORSE measurement, they produce NO measurement - 182.C.19.02 detected six faces on the committed frame and only the two above 10% of frame height yielded geometry or a complexion, and 182.C.20.05's evaluator needs 5 window frames plus 10 either side at 25 fps or every window returns at the edge of the search range and passes because none can fail. Telling somebody "quality may be reduced" when the truth is "nothing will be checked" invites them to click through, so those come back `blocking` and assertMeasurable refuses the submission - 182.C.21.03 makes sync non-compensable, so no other quality rescues it. Motion blur splits the same way from the same table: 21px retains 4% of the margin above the 0.2817 two-different-people baseline (advisory), 31px scores 0.2509 against itself which is BELOW that baseline (blocking, because no cutoff separates it). ORDERED BY WHAT THE FIX COSTS, not by severity - reframe, reshoot, fix the pack, re-record, re-plan - because somebody deciding whether to spend is better served by the cheapest fix first than by a ranked list of how bad each problem is. assertActionable refuses any warning shipped without a remedy AND a cited measurement, so a future "this might not work well" cannot be added; a list of unactionable warnings trains people to click through the ones that matter. The reference-conflict threshold names its own false negative: 0.4 sits between 0.261 for two subjects and 0.739/0.819 for degraded self-matches, but an earlier probe put two STRANGERS at 0.422, so it catches a gross mismatch and not a close one. A null fact stays silent rather than warning, because "not known" is not "bad".                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | 14 tests pass; eslint clean. Every cited figure re-verified against `evidence/2026-09-10-threshold-calibration/measurements.json` and `evidence/2026-09-10-appearance-attributes/measurements.json` before shipping. TYPECHECK CORRECTED 2026-09-11: the original "targeted tsc" here extended `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so that run exited 0 having checked NOTHING. Re-run against `tsconfig.typecheck.json` with a deliberate control error included to prove the harness reports. Re-verified after the fix: 0 errors in this module and its spec, against 973 pre-existing errors elsewhere in the repo, so the filter is real rather than an empty run.                                                                                                                                                                                                                                     | No provider spend - every number reused from measurements already made this phase. | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| 182.C.23.08 | Show honest stage progress and allow safe cancellation, background completion, notification …                           | Studio Web | 182.C.21    | done        | `apps/oshun/bff/src/generation/human-video-studio-session.ts` (new), `human-video-studio-session.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | The studio-side view over three things already built - 182.C.15.10 (no invented percentage), 182.C.15.05 (cancellation, and the absent cancel endpoint), 182.C.12.02 (idempotent submission) - and building it surfaced TWO THINGS IN NONE OF THEM. FIRST: "COMPLETE" AND "READY" ARE DIFFERENT EVENTS. The provider answers completed when the render finishes, at which point NOTHING has been evaluated - not sync, not exact speech, not identity. A notification sent then tells a creator a take passed checks that have not run, and some of those takes fail every one of them. So notification is bound to stages with something for a person to DO; assertNotifiable refuses rendering, downloading, evaluating and encoding, and PREMATURE_NOTIFICATION_REASONS keeps the WHY as data rather than a comment, because completed is the most obvious event in the pipeline and the one that means least to the person waiting. The wording follows: awaiting-review deliberately does not say "ready" - something is there to look at, and whether it is good is exactly what the creator is being asked. SECOND: AN ELAPSED TIME IS ONLY HONEST ANCHORED SERVER-SIDE. The obvious client notes when the browser first saw a stage and counts up; reload forty minutes into a render and it restarts at zero, showing "3 seconds" for a job running since breakfast - worse than showing nothing, because it READS as progress. buildView takes stageEnteredAt from the job record and assertServerAnchored refuses a view that wants to time the stage itself. assertBarHonest closes the other half of 182.C.15.10 at the RENDER boundary: the temptation is to build an honest uncounted event and then fill a bar from the elapsed number anyway. A RELOAD MUST NOT RE-SUBMIT: a fresh page naturally generates a fresh idempotency key, and a second key is a second render and a second charge, so assertResumeNotResubmit refuses one that does not match the job. Background completion genuinely works here (server-side polling plus webhook) and the corollary is what reaches the creator - closing the tab does not stop the SPEND either, because there is no cancel endpoint to call. cancelAffordanceFor derives its label and consequence from 182.C.15.05's own providerKeepsRunning flag rather than forming a second opinion about which stages are post-submission, with a test asserting the two agree across all 15 stages.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | 17 tests pass; eslint clean. Missed notifications on resume are filtered through the same notifiability rule, so a tab that missed downloading and evaluating is told about neither - it cost the creator nothing to miss them. TYPECHECK CORRECTED 2026-09-11: the original "targeted tsc" here extended `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so that run exited 0 having checked NOTHING. Re-run against `tsconfig.typecheck.json` with a deliberate control error included to prove the harness reports. Re-verified after the fix: 0 errors in this module and its spec, against 973 pre-existing errors elsewhere in the repo, so the filter is real rather than an empty run.                                                                                                                                                                                                                          | No provider spend.                                                                 | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| 182.C.23.09 | Build a take-comparison review with synchronized playback, A/B view, per-metric explanation, worst-window navigation …  | Studio Web | 182.C.21    | done        | `apps/oshun/bff/src/generation/human-video-take-comparison.ts` (new), `human-video-take-comparison.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Seven affordances; most are presentation over 182.C.17.09 (worst windows, annotated frames, timecodes) and 182.C.21.09/21.10 (the correct and regenerate actions). TWO ARE NOT PRESENTATION, and both hide inside the word "synchronized". SYNCHRONIZED PLAYBACK BY TIMESTAMP IS WRONG FOR THE NORMAL CASE: two takes of one line differ in their word timings - that is most of what differs between them - so scrubbing both to 00:00:04.200 puts one on "never" and the other still on "would", and the reviewer compares two different moments while calling it a comparison. alignPlayheads anchors on the shared WORD from 182.C.18.02 and interpolates INSIDE it rather than snapping to the onset, because snapping makes the second player stutter through a long word and a reviewer reads stutter as a defect in the take. assertPlayheadsAlignable refuses when the two takes do not say the same words - without a correspondence, pairing by timestamp puts two unrelated frames side by side under a label saying they match, which is worse than showing them unlinked because the reviewer believes the pairing. AN A/B NEEDS TWO COMPARABILITIES FROM TWO DIFFERENT PLACES, and they fail independently: 182.C.21.01 (same request, shot and AUDIO - different audio means sync numbers measured against different references) and 182.C.17.09 (same evaluator, weights, search width, fps and crop - confidence has a floor that moves with the search width). WORST-WINDOW NAVIGATION TAKES THE UNION OF BOTH TAKES: navigating A's worst windows shows where A is weak and says nothing about B, and B is the other half of the decision; stops from B are mapped back into A's timeline through the same word anchor. PER-METRIC EXPLANATION is where this phase's honesty either survives or quietly dies, because an explanation is exactly where a number picks up a verdict it never earned - explainMetric returns not-decidable for a reading with no fitted threshold and names the item the fitting waits on, and assertThresholdAttributed refuses a threshold nobody can say where they got. actionConsequences shows a CORRECTION as a re-measurement rather than a small adjustment: 182.C.21.09 invalidates eight gates, so most of what the reviewer just read stops being true.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | 18 tests pass. TYPECHECK CORRECTION: this session had been extending `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so every earlier "targeted tsc clean" in these rows was vacuous. Re-run against `tsconfig.typecheck.json` (`noCheck: false`) with a deliberate control error included to prove the harness reports. It caught a real bug the 35 green tests had not: `CandidateLineage` takes `audioDigest`/`scriptDigest`/`attempt` and two spec fixtures wrote `audioAssetId`, so the "refuses two takes made from different audio" test was comparing two identical keys and never exercised the refusal. Both fixtures corrected; the test now genuinely fails the comparison.                                                                                                                                                                                                                                 | No provider spend.                                                                 | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| 182.C.23.10 | Require creator confirmation before accepting a warning- eligible take and forbid override of consent, exact-speech …   | Studio Web | 182.C.21    | done        | `apps/oshun/bff/src/generation/human-video-acceptance.ts` (new), `human-video-acceptance.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | THE HIDDEN WORD IN THE ITEM IS "SEVERE", because it implies a MILD tier that could be overridden, and whether one exists had to be answered per gate rather than assumed. SYNC HAS A REAL ONE AND WE DID NOT INVENT IT: ATSC IS-191 gives the production tolerance (audio leading 15ms, lagging 45ms) and ITU-R BT.1359-1 the detectability boundary (45ms / 125ms); between them a take is out of production tolerance and still under what a viewer perceives, which is a genuine mild tier fitted by two standards bodies on real viewers. IDENTITY HAS NO TIER AND SAYING OTHERWISE WOULD BE A FABRICATION: grading needs a calibrated axis, quality-policy.ts inherits 0.35 from an ArcFace pipeline, and the 2026-09-10 probe put two DIFFERENT people at 0.422 with SFace - a false accept ABOVE the gate - while 182.C.19.08 leaves the operating point unfitted until the 182.C.33 corpus exists. On an axis whose zero is unknown "mild" is a word, not a measurement, so classifyIdentitySeverity REFUSES rather than returning a tier and every identity failure is non-overridable. That is stricter than the item asks, in the safe direction. Consent, exact speech, safety and speaker attribution have no axis at all - consent is a permission, and there is no such thing as a mildly different word. NEVER_OVERRIDABLE is derived from HARD_GATES rather than retyped, so a gate added to 182.C.21.03 cannot become quietly overridable here by being forgotten, and overrideRefusalReason gives six DISTINCT reasons because a creator told "non-overridable" six times learns nothing. THE CONFIRMATION BINDS TO WHAT WAS SHOWN: it carries a sha256 of the exact findings screen (order-independent, so a re-render does not invalidate a live confirmation) and assertConfirmationCurrent refuses a stale one - a take re-measured after the click has findings the creator never saw, and carrying the acceptance across approves something nobody looked at, the same reason 182.C.21.08 refuses an inherited verdict. A take is warning-eligible when a condition was NOT MEASURED, not only when one warned, because 182.C.20.07 reports 10 delivery conditions of which 4 are measured; assertUnmeasuredDisclosed refuses a screen that rendered a silent condition as passing, since a confirmation obtained that way is consent to a different take. Hard gates are checked BEFORE the confirmation is asked for, so a creator is never asked to confirm something that will be refused anyway - that teaches them the confirmation is a formality.                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | 17 tests pass; eslint clean. One test asserts the two sync windows equal the quality-policy defaults, so a 182.C.34 re-fit cannot leave this tier grading against a boundary nobody uses. TYPECHECK CORRECTED 2026-09-11: the original "targeted tsc" here extended `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so that run exited 0 having checked NOTHING. Re-run against `tsconfig.typecheck.json` with a deliberate control error included to prove the harness reports. Re-verified after the fix: 0 errors in this module and its spec, against 973 pre-existing errors elsewhere in the repo, so the filter is real rather than an empty run.                                                                                                                                                                                                                                                                | No provider spend.                                                                 | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| 182.C.23.11 | Support shot locking, variant naming, project history, reference/script/audio version changes …                         | Studio Web | 182.C.21    | done        | `apps/oshun/bff/src/generation/human-video-project-history.ts` (new), `human-video-project-history.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | A LOCK PRESERVES BYTES AND DOES NOT PRESERVE VERDICTS - that is the whole content of this item and it is not in 182.C.21.10, which already walks the edges. Lock shot 4 and change the audio: the file survives, which is what the lock is for and it works, but shot 4's sync score was measured AGAINST THE OLD AUDIO and its exact-speech verdict against the old transcript. Those numbers were never properties of the file; they were properties of the file NEXT TO something that changed, so the lock cannot keep them. A preview therefore has THREE outcomes and not two - regenerated, KEPT BUT VOID, untouched - and the middle one is what a UI would otherwise render as "locked, fine", which is precisely the inherited-verdict failure 182.C.21.08 refuses. assertVoidLocksAcknowledged makes somebody say so per shot. The same split runs through approvals: approvalsLost (regenerate it) and approvalsVoidedDespiteLock (re-measure it) are separate fields because both stop being approved and they need different work, and merged they look like one problem. VERDICTS_VOIDED_BY is derived from which measurement each verdict is a fact ABOUT rather than chosen - a reference change cannot void a sync score, because sync is a fact about the picture next to the audio and the audio did not move. THE ENTRY POINTS CARRY THE FINDING A CREATOR NEEDS: a REFERENCE change does not invalidate the audio, because voice comes from the script and the voice model rather than from a face, so a reference swap keeps every second of approved audio while a one-word script change rebuilds the audio and everything after it. Those two actions feel the same size in a UI and are not. A VARIANT NAME IS NOT AN IDENTITY: it binds to a node AND a version, because "warmer take" after a regeneration points at different bytes than the creator watched, and renameVariant is inert on the 182.C.21.01 comparability key - a claim the spec checks against comparabilityKey rather than a promise in a comment. The same label on two shots is fine ("wide"); on two takes of one shot at one version it is refused, because a creator choosing between them has nothing to choose by. History stores the PREVIEW a change was applied under, not the event, because "audio changed at 14:02" cannot answer why a particular shot is stale; whyStale returns the whole chain that reached a node rather than the last event. previewCost counts work (shots to rebuild, approvals lost, approvals to regain, verdicts to re-measure) and deliberately does NOT return money - the per-lane rate is not known until the request is built, so a currency figure would be invented.                                                                                                                                                                                                                                                                                                                                | 17 tests pass; eslint clean. One test drove out a real semantic: changing the audio loses the audio's OWN approval while not appearing in `regenerated`, because the creator is authoring it rather than having it rebuilt - the test expectation was corrected, not the code. TYPECHECK CORRECTED 2026-09-11: the original "targeted tsc" here extended `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so that run exited 0 having checked NOTHING. Re-run against `tsconfig.typecheck.json` with a deliberate control error included to prove the harness reports. Re-verified after the fix: 0 errors in this module and its spec, against 973 pre-existing errors elsewhere in the repo, so the filter is real rather than an empty run.                                                                                                                                                                           | No provider spend.                                                                 | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| 182.C.23.12 | Add export controls for aspect/rendition, clean/captioned master, audio stems where licensed, captions, transcript …    | Studio Web | 182.C.21    | done        | `apps/oshun/bff/src/generation/human-video-export-controls.ts` (new), `human-video-export-controls.spec.ts` (new), `libs/contracts/src/human-video/request.ts` (exports the `OutputRendition` type)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Seven export controls; five are selection and TWO ARE GATES, because of things this phase measured. A RENDITION IS NOT THE THING THAT WAS GATED: every verdict here was measured on the master - sync on its frames, identity on its face, artifacts on its pixels - and a derived rendition is a different picture. SCALING usually carries them across and "usually" is a MEASUREMENT rather than a hope: 182.C.19.08 embedded one real face at seven crop sizes, and the margin above the two-different-people baseline is 100% at 256px of FACE height, 79% at 96px, 61% at 64px, then falls off a cliff to 24% at 48px and 14% at 36px. So the floor is 64px of face height, in ABSOLUTE pixels because that is what was measured. A 1080p master whose speaker fills 14% of frame height keeps a 151px face; at 720p it is 101px, still clear; a 6% face at 720p is 43px and the identity verdict does not carry down. REFRAMING never carries them, and the face size is not even computed for a crop - a 9:16 crop of a 16:9 two-shot may not contain the speaking face at all, and a sync verdict about a mouth no longer in frame is a fact about a different video. This is 182.C.21.08's inherited-verdict failure wearing an aspect ratio. It is NOT a refusal to export the crop: a creator may have it, what they may not have is the crop wearing gate results measured on something else. A CLEAN MASTER IS NOT A CAPTIONED ONE WITH THE CAPTIONS OFF - burned captions are pixels and there is no un-burning them, so with no pre-burn source retained the clean master is IMPOSSIBLE rather than degraded, and "we can make you a slightly worse one" would be false. STEMS ARE A DIFFERENT DISTRIBUTION, NOT A DIFFERENT FILE FORMAT: an isolated voice track of a synthesized performance is the voice alone, re-usable over anything, and 182.C.08.04 scopes consent to a USE - so an isolated-voice-distribution scope is required and its absence is a refusal rather than a default. planExport reports EVERY omission at once instead of throwing on the first, because a creator asking for stems, a transcript and a clean master deserves to hear all three rather than fix them one refusal at a time; the asserts stay exported for the places a single control is enforced. Presets are a rendition plus a caption choice and NOTHING else - one that also set bitrate, colour and loudness would be four decisions behind one label, and 182.C.16 already conforms those from the delivery profile.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | 20 tests pass; eslint clean; the contracts request spec (26 tests) still passes after the type export. One test pins the preset shape to exactly four keys, so a future preset cannot quietly acquire a second place to set an encoding parameter. TYPECHECK CORRECTED 2026-09-11: the original "targeted tsc" here extended `apps/oshun/bff/tsconfig.json`, which sets `noCheck: true`, so that run exited 0 having checked NOTHING. Re-run against `tsconfig.typecheck.json` with a deliberate control error included to prove the harness reports. Re-verified after the fix: 0 errors in this module and its spec, against 973 pre-existing errors elsewhere in the repo, so the filter is real rather than an empty run.                                                                                                                                                                                                       | No provider spend.                                                                 | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| 182.C.23.13 | Follow the `frontend-skill` design convention (confirm it resolves in the working environment or link the guidance it … | Studio Web | 182.C.01.11 | implemented | `docs/domains/isis/human-video/frontend-convention.md`, `apps/oshun/web/src/design-system/lilith/waiting-and-measured.tsx`, `tokens.ts` (`MIN_TARGET_PX`), `apps/oshun/web/src/app/studio/human-video/page.tsx` (new), `apps/oshun/web/src/components/isis/human-video/HumanVideoStudioRoute.tsx` (new), `HumanVideoStudio.tsx` (new), `projectState.ts` (new), `projectPort.ts` (new), plus the four specs                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | THE FOUR STATES ARE FIVE, AND THE FOUR EMPTIES ARE SEVEN. Direction before approval is not loading, not failed, not review and NOT empty — an empty state hands over the next action and the next action is in another region, so `blocked` names that region and links there. And a subject nobody has asked yet is not a wait: a project can sit for days saying “waiting for Ama” when Ama was never sent anything.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Chromium against the real stylesheet, five surfaces covering all seven region states: 0 axe violations (wcag2a/2aa/21a/21aa) at 1280, 768 and 400 px, no horizontal overflow, every one of the 25 region frames carrying its state as a word, three distinct computed shapes surviving a full colour flatten. FIFTEEN CONTROLS MEASURED UNDER 24 px (WCAG 2.2 SC 2.5.8) — the exact “without fine pointer input” clause, invisible to 156 passing jsdom tests; `MIN_TARGET_PX` is now the floor and the re-measure reports none.                                                                                                                                                                                                                                                                                                                                                                                                    | No provider spend; no model call.                                                  | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |

### 182.C.24 Mobile and Desktop Workflow Parity

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                            | Owner         | Dependency  | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                          | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Cost                              | Blocker                                                                                                                                                                                                                                                                                                                                                                    |
| ----------- | ------------------------------------------------------------------------------------------------------------------- | ------------- | ----------- | ----------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.24.01 | Decide which authoring controls are full-parity on mobile and which use a review/approval-focused experience.       | Studio Mobile | 182.C.24.05 | implemented | `docs/domains/isis/human-video/mobile-parity.md` (new)                                                                                                                                                                                                                                                                                                                                                                 | THE ITEM TITLE MISSES THE STRONGEST REASON TO HAVE A MOBILE SURFACE AT ALL: ENROLMENT IS NOT AN AUTHORING CONTROL. THE SUBJECT IS NOT THE CUSTOMER - the person whose face this is may have no Studio account, no reason to open a laptop, and no relationship with us beyond agreeing to this - but they do have the phone the reference photographs were probably taken on. Their flow is short, high-stakes and entirely mobile-shaped, and 182.C.26.02 makes it a SEQUENCE OF SMALL FINAL ACTS RATHER THAN AN EDITOR, because the proof binds to the digests of exact assets and adding a reference later is a NEW enrolment rather than an edit. So mobile carries TWO flows and only one is in the item title. THE SPLIT RULE IS NOT SIMPLE-VERSUS-COMPLEX but two questions: is the phone BETTER at this (camera, microphone and video player are phone-native; a storyboard timeline is not, and the reason is not screen size but that 182.C.23.04 controls are continuous and spatial with a coarse pointer over an occluded target), and is it a DECIDING operation (approving audio, accepting a take, releasing, withdrawing all carry If-Match) - a deciding operation is not HARDER on a phone, it is RISKIER, because THE THING THAT JUSTIFIES IT IS THE PART THAT FITS WORST ON THE SCREEN. Four rules follow. AN APPROVE BUTTON WITHOUT THE MEASUREMENTS IS A RUBBER STAMP WITH A PHONE ATTACHED - mobile may carry an approval only where it carries the measurements AND the windows nobody could measure, because "no findings" reads as "no problems" faster on a small screen than anywhere else. NON-WAIVABLE IS NON-WAIVABLE ON A PHONE. REVIEW-ONLY IS A REAL STATE, not a cramped editor producing worse decisions than no editor. AND THE SUBJECT FLOW IS NOT BEHIND THE CREATOR LOGIN, because withdrawal is in force from the instant it is made and cannot require an app the subject never installed. | Per-stage table built from the ten stages in creator-workflow.md, each with its treatment and reason; four full-parity, four review-only, one not-on-mobile, plus the subject enrolment flow. Grounded in the modules that constrain it: 182.C.25.03 If-Match on deciding operations, 182.C.26.02 digest-bound proof (verification methods read from `VERIFICATION_METHODS` - attestation and signature based, so no live-capture requirement is claimed), 182.C.17.06 unmeasured windows, and the twelve non-waivable gates. All three cross-document links resolve. Prettier-formatted. No code changed.                                                                                                              | No provider spend; no model call. | A SPLIT, NOT AN IMPLEMENTATION. No mobile surface exists for this flow and the web surface does not either. 182.C.24.02 (resumable uploads, previews, deep links, recovery), .03 (leakage), .04 (accessibility) and .06 (automation) follow from this split rather than being choices inside it, and all four stay open.                                                   |
| 182.C.24.02 | Implement resumable background uploads, constrained-device previews, notification deep links …                      | Studio Mobile | 182.C.24.05 | implemented | `apps/oshun/mobile/src/humanVideo/uploadRuntime.ts` (new), `src/native/humanVideoCapabilities.ts` (new), `app/_layout.tsx` (listener + relaunch recovery), `HumanVideoReviewScreen.tsx` (the preview rule), plus two specs, plus the correction 2026-09-12: `apps/oshun/mobile/src/native/humanVideoCapabilities.ts`, `humanVideoCapabilities.spec.ts`, `scripts/human-video/verify-mobile-upload-transport.mjs` (new) | THE KEY IS WRITTEN BEFORE THE SESSION IS OPENED. The worst case is then a key with no session, which the server can resolve; the other order gives a session with no key, which nothing can — and a key regenerated on relaunch opens a SECOND session holding a quota slot nobody will write to. `sessionType: 'background'` is the part the item is actually asking for: a foreground task dies with the process, and every upload then takes the resume path rather than the interrupted ones. The sidecar holding the key lives in the DOCUMENT directory, because a cache purge mid-upload is exactly the case it exists for. THE TRANSPORT HAD DRIFTED IN FOUR PLACES AT ONCE, corrected 2026-09-12, and NOTHING COULD HAVE CAUGHT IT: it was written against endpoints that did not exist, so no test could drive them and no device was attached. It opened a session with `declaredBytes` where the surface reads `bytes` (400 on EVERY open); it posted chunks to `/chunks` where the route is `/chunks/:offset` (404 on EVERY chunk); it carried the offset in a `content-range-start` header the server does not read; and it sent no `Content-Type: application/octet-stream`, which is the only type the chunk route's parser is registered for. A FIFTH THING IS NOW REFUSED RATHER THAN SENT: `UploadTask` hands the OS a FILE, not a byte range, so a resume at 4 MB would append the file's FIRST byte at 4 MB - the server's offset check refuses it, and if it did not the stored object would be the file twice over with a hole in the middle. The transport throws with the offset named, which is an honest limitation until somebody can test a tail upload on a device.                                                                                                                                                                                                                                     | 128 tests across 9 suites under the app's own jest; scratch typecheck with a live control reports only the control error; eslint 0 errors. The expo modules are named in one file outside the guarded paths, which the 182.C.24.03 lint rule enforces. `scripts/human-video/verify-mobile-upload-transport.mjs` reads the paths and the create handler's body field OUT OF THE BFF SOURCE rather than holding its own copy — the reason the SDK's drift guard exists — and runs with no device. PROVED BY BREAKING EACH CHECK IN TURN: the chunk path, the body field name and the content type each fail the verifier when changed and pass when restored. 15 capability tests and 113 mobile human-video tests green. | —                                 | Known caller gap, stated rather than hidden: `startOrResumeUpload` has no screen calling it, because the screen that would upload on a phone is enrolment (182.C.26) rather than anything in §24.                                                                                                                                                                          |
| 182.C.24.03 | Prevent biometric reference/audio leakage through camera roll thumbnails, OS share sheets, logs, clipboard …        | Studio Mobile | 182.C.24.05 | implemented | `apps/oshun/mobile/src/humanVideo/osContainment.ts` (new), `osContainment.spec.ts` (new), `HumanVideoReviewScreen.tsx` (its caller), `eslint.config.js` (the enforcement)                                                                                                                                                                                                                                              | THE POLICY EXISTED AND NOTHING CALLED IT, SO 'PREVENT' NEEDED A CONSTRUCTION RATHER THAN A CONVENTION. Every path out of the human-video mobile surface — share sheet, camera roll, clipboard, log, disk cache — goes through one layer that asks `biometricContainment` first, and an eslint rule makes a direct import of any of those modules an error inside the human-video paths. AN ABSENT CAPABILITY REFUSES RATHER THAN NO-OPS: this build has no clipboard, sharing or media-library module, and a silent no-op tells the caller the copy was made and starts working unreviewed the day the module lands. The cache location is chosen from the material, because the caller is the one place that does not know a biometric file in the documents directory survives the withdrawal that was supposed to end it.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | 104 tests across 7 suites under the app's own jest; scratch typecheck with a live control reports only the control error (it caught two real defects the tests could not: an interface extending a union, and a helper typed wide enough to drop a write's location); eslint 0 errors, and a probe file importing `expo-file-system` fails with the rule's message.                                                                                                                                                                                                                                                                                                                                                     | —                                 | None for this item. The five channels are contained; a sixth (the OS preview) is reported unavailable always, which is the finding the policy module was built around.                                                                                                                                                                                                     |
| 182.C.24.04 | Add accessible touch targets, screen-reader labels, captioned previews, transcript navigation, reduced motion …     | Studio Mobile | 182.C.24.05 | implemented | `apps/oshun/mobile/src/humanVideo/HumanVideoReviewScreen.tsx` (new), `reviewSource.ts` (new), `app/human-video-review.tsx` (new), plus the three specs                                                                                                                                                                                                                                                                 | THE CHECKER WAS NEVER CALLED, SO THE SCREEN NOW RENDERS FROM THE SPEC IT READS. `checkControls` knew the six rules and nothing used them; a screen laid out separately and audited afterwards drifts on the first change. The touch target comes off `heightDp` and the label off `accessibilityLabel`, so a control that breaks a rule cannot be drawn without the spec saying so. Three findings from the wiring: reduced motion must not stop the take (the take is the content under review — chrome only, and nothing autoplays either way); a measured value with no fitted operating point is a FOURTH outcome, because `not-measurable` denies a measurement that happened and pass/fail invents the threshold 182.C.19.08 leaves unfitted; an unread alignment is not an empty one, and both render as a blank caption block.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | 90 tests under the app's own jest (6 suites), scratch typecheck with a live control reports only the control error, eslint 0 errors. Captions come from `GET /audio/:audioId`'s alignment — the read added the same day — and a word the aligner could not place stays untimed rather than being given a cue.                                                                                                                                                                                                                                                                                                                                                                                                           | —                                 | None for this item. The route shows the honest failure until a candidate store is bound, and the caller passes the audio id because the alignment travels with the audio rather than the candidate.                                                                                                                                                                        |
| 182.C.24.05 | Decide whether a desktop surface is in scope and which app hosts it (no Oshun studio desktop app exists …           | Studio Mobile | 182.C.23    | implemented | `docs/domains/isis/adr/ADR-0007-human-video-authoring-surfaces.md` (new)                                                                                                                                                                                                                                                                                                                                               | ANSWERED NO, AND THE DECISIVE ARGUMENT IS THE BIOMETRIC COPY RATHER THAN THE ENGINEERING COST. All three named candidates exist and were inspected: `apps/yemaya/studio-desktop` is ELECTRON with 63 source files and a film/game production domain, `apps/lilith/desktop` is TAURI 2 with 14, `apps/iris/desktop` is Tauri with 19. The release train ships V1.0 web+PWA, V1.1 native mobile, V1.2 Veritas/Metis - NO DESKTOP SURFACE AT ANY OF THE THREE. Four reasons, in order of weight. (1) The most sensitive artifact here is a photograph of a real person; 182.C.24.03 already fights leakage through camera roll, share sheets, logs and caches on mobile, and a desktop host adds a general-purpose filesystem, an unmanaged disk and a crash reporter - A THIRD SURFACE IS A THIRD PLACE A LIKENESS LANDS, for a capability whose FIRST surface does not exist yet, and 182.C.26 already established that what leaves our storage cannot be recalled. (2) CHOOSING A HOST WOULD SETTLE A REPO-WIDE QUESTION THIS PHASE HAS NO STANDING TO SETTLE: two desktop runtimes are live (Electron in Yemaya, Tauri in Lilith and Iris) and picking one makes Phase 182 the precedent. (3) The item own requirement - same server contracts, capability registry, consent, quality decisions and release proof as web - means ANY host duplicates the whole stack, and none of the three has any of it, so the cheapest honest desktop surface is a web view of the web surface. (4) Nothing is demonstrable anywhere yet. The host is NAMED ANYWAY (yemaya/studio-desktop, the only video-domain candidate) so a reversal starts from a recommendation rather than a blank page.                                                                                                                                                                                                                                                 | All three candidate apps inspected on disk: runtime identified from `forge.config.ts`/`src-tauri`, source-file counts taken by find, purpose from each README. Release-train facts read from `libs/oshun/navigation/src/release-scope.ts` (`LILITH_NATIVE_APPS_RELEASE_NAME` = V1.1, `V1_DEFERRED_DOMAIN_RELEASE` = V1.2). Recorded as `docs/domains/isis/adr/ADR-0007-human-video-authoring-surfaces.md`, matching the ADR-0006 format. Prettier-formatted. No code changed.                                                                                                                                                                                                                                           | No provider spend; no model call. | NONE - this decision UNBLOCKS 182.C.23 and 182.C.24 under the tracker prohibition 4, which required the surface decision before any Studio item. Reversing it needs a named owner, a target release, and an answer to the biometric-copy argument, not simply a decision to build it. Whether mobile is full-parity authoring or review-focused stays open as 182.C.24.01. |
| 182.C.24.06 | Add platform-appropriate automation for critical upload, author, submit, cancel, review, approve, and export flows. | Studio Mobile | 182.C.24.05 | partial     | `apps/oshun/mobile/e2e/flows/shell/human-video-review.yaml` (new), `scripts/human-video/verify-mobile-review-automation.mjs` (new)                                                                                                                                                                                                                                                                                     | THE REVIEW FLOW IS THE ONE THIS PLATFORM HAS. 182.C.24.01 decided the phone is review-and-approval, so the flow drives the act a phone is for; the assertions are the ones only a device can make — that nothing is playing before somebody asks, that the outcome WORDS are on screen, and that a refused share leaves no control to tap. A flow tapping through screens that do not exist would pass by finding nothing, which is also why the verifier exists.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | The verifier runs on a machine with no device: it maps every test id the flow names onto the screen, including the ones built from the control spec by template, and was proved by renaming a control — it fails, and passes again when restored. The flow itself needs a device or emulator and was not executed here.                                                                                                                                                                                                                                                                                                                                                                                                 | —                                 | STILL UNCHECKED: the item names upload and generation flows too. Generation is a web surface, and the upload a phone makes is enrolment (182.C.26), which has no screen — so those two cannot be driven until that screen exists.                                                                                                                                          |

### 182.C.25 Public API, OpenAPI, SDK, and Events

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner     | Dependency | Status | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Cost                              | Blocker                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | ------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.25.01 | Add versioned endpoints for upload sessions, identity packs, voice/audio approval, human-video jobs, candidates …      | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-api-surface.ts` (new), `human-video-api-surface.spec.ts` (new), `human-video-api-routes.ts` (new), `human-video-api-routes.spec.ts` (new), `human-video-api-semantics.ts` (corrected), `apps/oshun/bff/src/generation/human-video-api-store.ts` (new), `human-video-api-store.spec.ts` (new), `apps/oshun/bff/src/generation/human-video-asset-handlers.ts` (new), `human-video-asset-handlers.spec.ts` (new), `human-video-api-routes.ts` (corrected: HandlerContext carries the idempotency key, HandlerRefusal added, gate-4 not-found recoded), `libs/contracts/src/human-video/failure-codes.ts` (`hv.validation.unknown_resource` added), `libs/contracts/schemas/human-video/{HumanVideoFailureCode,FailureReport,openapi.components}.json` (regenerated), `apps/oshun/bff/src/generation/human-video-audio-handlers.ts` (new), `human-video-audio-handlers.spec.ts` (new), `human-video-decision-subjects.ts` (new), `human-video-job-handlers.ts` (new), `human-video-job-handlers.spec.ts` (new), `human-video-pre-generation-warnings.ts` (corrected: `hasFrontalReference` is now nullable), `human-video-candidate-handlers.ts` (new), `human-video-candidate-handlers.spec.ts` (new), `human-video-review-handlers.ts` (new), `human-video-review-handlers.spec.ts` (new), `human-video-decision-subjects.ts` (extended to five deciding operations), `human-video-candidate-handlers.ts` (corrected: the check name is a closed vocabulary), `human-video-job-handlers.ts` (corrected: a shot records its continuity handoff), `human-video-release-handlers.ts` (new), `human-video-release-handlers.spec.ts` (new), `human-video-handlers.ts` (new), `human-video-handlers.spec.ts` (new), `human-video-project-draft.ts` (new), `human-video-project-handlers.ts` (new), `human-video-revocation.ts` (`authoring-drafts` copy location), `human-video-audio-handlers.ts` (`get-audio`), `human-video-draft-store.ts` (new), `libs/oshun/persistence/src/prisma-renderer.ts` + `prisma/schema.prisma` + `prisma/migrations/20260911190000_human_video_project_draft/migration.sql` (new), `human-video-runtime.ts` (new), `apps/oshun/bff/src/server.ts` (mounted) | FIRST PIECE OF THE WIRING PROGRAM: an implementation of `HumanVideoApiStore`, which nothing in the repository had. THE WHOLE FILE TURNS ON A CASE THE PORT SIGNATURE CANNOT EXPRESS. `readDecisionSubject` returning `null` means "NO SUCH RESOURCE" and the route turns it into A 404 (human-video-api-routes.ts:254-259), so a store that catches a database error and returns null TELLS THE CALLER THEIR JOB DOES NOT EXIST - the same truthful-looking false statement the 503 exists to prevent, arriving through the one path the existing guard does not cover: the guard handles UNCONFIGURED, and this handles CONFIGURED BUT UNABLE TO ANSWER. The rule is one line: an error propagates, it never becomes null. AND THE TWO NULLS ARE NOT THE SAME - `readIdempotency` returning null is ordinary and correct (no prior call under this key that I still remember, which is what a first use looks like), while `readDecisionSubject` returning null is A CLAIM ABOUT THE WORLD. The same rule on the idempotency side has a different consequence, spelled out in the test: null there classifies as `first-use`, and A FIRST USE OF A KEY THAT ALREADY RAN SENDS THE REQUEST AGAIN. FORGETTING IS PART OF THE CONTRACT: records expire, and the window must be LONGER than any client retry horizon, because a client retrying inside its own policy that meets a forgotten key pays twice - 72 hours, with expiry treated as absence of memory rather than as a replay. AND THE UPDATE PATH NEVER TOUCHES THE DIGEST, because that digest is what produces 182.C.25.03 422 on a key reused with a different body; overwriting it would erase the evidence that refuses the reuse. SECOND PIECE: THE FIRST FIVE HANDLERS (upload sessions and identity packs). THE FIVE GATES NEVER LOOK AT THE TENANT ON THE ROW - gate 2 checks a scope every tenant client holds, and `UploadSessionStore.get(sessionId)` takes an id and no tenant at all, so the tenant comparison lives in the handler and its refusal is BYTE-IDENTICAL to the one a missing id gets (asserted, not asserted-about): a 403 on somebody else session confirms the id exists, which is the one fact worth hiding. A PACK WAITING ON A MAPPING HAS NOT FAILED, AND A 4xx SAYS IT HAS - `evaluateIdentityPack` has THREE outcomes, and a reference with two faces in it is a QUESTION FOR THE CREATOR that a 422 turns into an accusation plus an invitation to retry a body that can never succeed; that pack is created as `awaiting-mapping` with the reference named, which is the pending state `frontend-convention.md` recorded as missing. THE OTHER DIRECTION DOES NOT SYMMETRISE: a pack REJECTED at creation (no face) is fixed by replacing the reference, i.e. a different pack, so storing it leaves a row that LISTS LIKE AN ASSET and can never be used - it is refused with every finding named instead. WITHOUT A CURSOR SECRET, PAGE ONE IS THE LIE: the first page still builds, and its absent `nextCursor` states that twenty-five of four hundred packs are all of them, so the whole listing is refused rather than its second page. A REFERENCE WITHOUT A DIGEST CANNOT BE AUTHORIZED - an authorization binds a subject to digests, `digestSha256` is written at completion, and a pack over a still-open session is authorized for whatever arrives next. Two smaller ones: the route CONSUMED the idempotency key at gate 3 and did not pass it on while `openUploadSession` dedupes on the same key, so a fabricated key made every retry a new session holding one of sixteen quota slots nobody would write to; and the catalog had NO NOT-FOUND CODE, so the existing 404 carried "The request could not be read", which is false - `hv.validation.unknown_resource` (404) now says it, merged with not-yours on purpose. THIRD PIECE: THE AUDIO HANDLERS, and the first DECIDING operation to become reachable. A CLIENT-SUPPLIED ALIGNMENT IS A CLAIM AND EVERYTHING DOWNSTREAM QUOTES IT AS A MEASUREMENT - a sync score computed against word timings nobody measured reads exactly like a real one, so the alignment is produced here by a bound aligner and NO ALIGNER MEANS NO TAKE. THE LAST WORD END IS NOT THE FILE DURATION: they differ by the trailing silence, so the probe is a REQUIRED port rather than an arithmetic convenience, and the test asserts the two numbers differ. A WORD THE ALIGNER COULD NOT PLACE IS NOT A WORD WITH NO PROBLEMS - `adaptAlignerOutput` correctly leaves it out and names it, which is also the shape that loses it, because nothing about a SHORT ALIGNMENT reads as wrong; unplaced words are carried on the record and block approval of an exact-dialogue take, since a word with no measured window cannot be checked for visible speech. THE RECOGNIZER ABSENCE IS NOT FATAL AT REGISTRATION AND IS FATAL AT APPROVAL: `unknown` is honest for a take whose mode never promised exact speech, and refuses as firmly as a deviation for one that did. A DEVIATION SIGN-OFF NAMES AN ID STABLE ONLY WITHIN ONE (SEGMENT, DIGEST) PAIR, so one carried over from another alignment names a finding that is not on this take and is refused rather than ignored. Approval also refuses ESTIMATED timings (`isFullyMeasured`), and re-approving the same evidence DOES NOT RESTAMP `approvedAt` - the decision was made when it was made. `human-video-decision-subjects.ts` is the third thing the mount needed: it answers `approve-audio` and THROWS for the eighteen unwired operations, because returning null there would tell a caller their candidate does not exist when the truth is that nobody here can answer. FOURTH PIECE: THE JOB HANDLERS, where the workflow ordering stops being advice. AN UNREACHABLE CONSENT STORE IS NOT AN ABSENT CONSENT: `hv.consent.missing` is a TERMINAL 403 that blames the PERSON, so saying it because a store timed out sends somebody to ask a subject for a signature they may already have given, and no retry corrects it - an unbound authorization port is a 503 instead. "NO WARNINGS" AND "NOTHING MEASURED" ARE THE SAME EMPTY LIST: `warningsFor` correctly returns nothing for null facts, which is indistinguishable from a clean request, so the job carries `unmeasuredFacts` NAMING EACH CHECK THAT DID NOT RUN. That also forced a correction in the warnings module - `hasFrontalReference` was a plain boolean, and this edge can establish NEITHER value: true suppresses a real warning and false raises one nobody measured, so it is now nullable like every other fact. A REGISTERED TAKE IS NOT AN APPROVED ONE (the sync verdict is a fact about the audio the creator accepted), a lapsed permission is`hv.consent.expired`while one that never covered this is`hv.consent.scope_exceeded`(one is renewed, the other renegotiated), and CANCELLING RECOVERS NOTHING - the response reports avoided and unrecoverable SEPARATELY, lists unquotable spend by id rather than folding it into a total, and names the provider jobs that keep running. The cancellation vocabulary has NO value for "the creator changed their mind", so `reason` is REQUIRED rather than defaulted: the reason decides whether the artifact survives, and the cheapest default to pick is the one that keeps a depiction somebody withdrew. FIFTH PIECE: THE CANDIDATE READS, and the one response shape that matters most. AN EMPTY FINDINGS LIST IS THE MOST DANGEROUS ANSWER THIS API CAN SEND - a shot of the back of somebody head produces NO SYNC FINDINGS, and "no findings" is indistinguishable from "nothing wrong" to every client ever written. So `measured` and `notMeasured` are SEPARATE FIELDS with the coverage summary beside them, and a candidate nobody has evaluated returns NULL for each rather than an empty array, because an empty array is a claim that somebody looked. An unmeasured check may also carry `remedy: null`, which is a real answer: an identity check on a shot with no face is not waiting on anything, and inventing a remedy is advice that cannot be followed. THE ARITHMETIC IS CHECKED ON THE WAY OUT: `assertNoCoverageCountedAsPass` runs on the READ, so a stored record that scored more windows than were measurable is refused as `hv.internal.invariant_violated` rather than served with a note - a note gets read past, and the internal family never blames the request. TWO TAKES OF ONE LINE ARE NOT AUTOMATICALLY COMPARABLE, so the listing returns `comparisonSets` rather than one flat array: sync and word timing are both measured AGAINST the audio, and a flat list invites the side-by-side `assertComparable` refuses. The nested listing READS THE JOB FIRST, because scoping by job id alone answers for another tenant job to anybody who guesses one. SIXTH PIECE: THE FOUR REVIEW DECISIONS. THE FINDINGS SCREEN IS BUILT SERVER-SIDE FROM THE STORED MEASUREMENT - a creator confirms by returning a DIGEST OF THE SCREEN THEY SAW, so if the screen also came from the request the comparison is CIRCULAR and a screen that omitted a failure verifies perfectly. AND A CONFIRMATION THAT WAS SENT IS CHECKED EVEN WHEN NONE WAS NEEDED: `accept()` verifies the digest only for a WARNING-ELIGIBLE screen, which is the right rule for whether one is REQUIRED and the wrong one for whether a supplied one is TRUE - a clean take approved against a stale digest means the creator was looking at a screen this take no longer has. WHO CONFIRMED COMES FROM THE TOKEN, because a body-supplied creator id attributes an acceptance to a colleague. A TAKE NOBODY EVALUATED IS NOT A TAKE THAT PASSED - with no measurement the screen has no findings, nothing is warning-eligible, no hard gate is listed as failed, and `accept()` waves it through - so approval requires one and REJECTION DOES NOT, because setting a take aside is not a claim about its quality. A HARD GATE THE MEASUREMENT NEVER MENTIONS is counted as not cleared, which a list built only from measured-plus-unmeasured misses entirely; so is one whose `passed` is NULL, because `182.C.19.08` leaves identity unfitted and a number with no verdict has not cleared a gate. A CORRECTION REMOVES THE MEASUREMENT rather than flagging it, because a flagged number is the one the next step picks up. AND A REGENERATION IS PLANNED OVER THE WORK GRAPH: the first call REFUSES by naming every approval it would destroy, the second carries the losses the creator accepted. Two corrections fell out of it - the measurement check name is now a CLOSED VOCABULARY (a free-text name never matches a hard gate, the failed list comes back empty, and every take becomes acceptable), and a shot now RECORDS its continuity handoff at submission, because inferring it from shot order invents dependencies and omitting it misses the neighbour that actually has one. SEVENTH PIECE: THE LAST FIVE OPERATIONS, THREE OF WHICH ARE RETRACTIONS - and that word does the work. A RETRACTION IS NEVER BLOCKED BY A PRECONDITION, and the state check that looks like diligence ("this take is not released") is "your withdrawal failed, please refresh" wearing a different hat, so A WITHDRAWAL OF SOMETHING UNPUBLISHED SUCCEEDS: the outcome asked for is that it is not published, and that is already true. DELETING TWICE RETURNS THE SAME MANIFEST from a ledger rather than a 404, because a 404 on the retry makes a successful deletion look like a failure and whoever reads it that way stops retrying. A TAKEDOWN NEVER REPORTS COMPLETE WHILE A COPY IS OUT OF REACH - four of the eight copy locations are not ours, the unreachable list SHIPS IN THE RESPONSE, and `complete` is decided by the REACH rather than by the erasure report, because every target being gone from our side says nothing about a signed link already handed to somebody. A RELEASE PROOF IS BOUND TO THE FINAL BYTES: a candidate id survives a re-encode and the bytes do not, so the caller names the digest and it is checked against the one that was measured. THE TABLE IS NOW COMPLETE: `human-video-handlers.ts` merges the six groups, `assertHandlersMatchSurface` PASSES for the first time since the surface was declared, and all 23 routes register on a real Fastify instance. Assembling the table is NOT mounting it - `server.ts` still mounts nothing, because a surface whose every call fails closed is worth serving only once one adapter behind it is real. EXTENDED 2026-09-11 to twenty-six operations: the surface had no read for the thing it authors, so a project could be listed and not opened. The three project operations keep the authoring work between sessions, which this flow needs because it waits days on a subject. Their concurrency control is a version in the BODY rather than an If-Match, because the precondition path is for deciding operations and an autosave decides nothing. AND THE DRAFT CREATED AN OBLIGATION IT THEN MET: a project that was never submitted has no job, so the withdrawal path's job enumeration missed exactly the case where somebody withdrew before anything was made. `authoring-drafts` is now a copy location; a withdrawal removes the cast entry, the references and the recorded voice, keeps the creator's words with the speaker cleared, and never blocks on an unbound draft store — it reports the drafts as not reached instead. AND A DECIDING OPERATION WAS UNREACHABLE TWICE: `approve-audio` requires an If-Match and the only place that tag appeared was the submission response, so a creator who closed the tab could not approve at all. `GET /audio/:audioId` (twenty-seven operations) is what makes the approval reachable from a page that did not submit it — which in a flow that waits days on a subject is the ordinary case, not an edge one. THE FIRST REAL ADAPTER: two tables (`human_video_idempotency`, `human_video_project_draft`) with the projections derived on write, a GIN index on the cast so a withdrawal can find an unsubmitted project, and a WRITE THAT IS A COMPARE-AND-SET — the version check happens in the handler, so an unconditional upsert would overwrite whoever landed in between and report success. MOUNTED 2026-09-11. The surface is registered unconditionally and fails closed; the boot line names every port that will answer 503, so an operator meeting one can tell a misconfigured database from an operation with no adapter yet. A generated Prisma client that predates the two tables is treated exactly like no database — the delegate is checked for the methods it must have, because binding a missing one produces `undefined.findUnique` on the first request. | 13 tests pass; eslint clean; typechecked against a scratch config extending tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the store does not - AND THAT TYPECHECK CAUGHT A REAL BUG THE PASSING TESTS DID NOT: a fake DecisionSubject in the spec carried `evidenceVersion` where the contract has `evidenceDigest`. Both classification paths run through the REAL `classifyIdempotency` rather than a restatement (first-use, replay, and the 422 on a reused key), tenant isolation is asserted through the row identity rather than a filter, and the sweep is checked to remove exactly the expired row. The delegate is an injected port in the shape the repo Prisma delegates already take, so the module pulls in no infrastructure. Handlers: 21 tests pass (103 in the five neighbouring API specs still pass, plus 46 in contracts after the catalog addition and schema regeneration); eslint 0 errors; scratch typecheck over the module, its spec, the routes and the store reports ONLY the deliberate \_\_control.ts error. The typecheck again caught what the green tests did not - `.catch((error) => error as HandlerRefusal)` widens the result to "refusal or success", so a call that WRONGLY SUCCEEDED would have read as a refusal with every field undefined; the captures now go through a helper that fails if the call resolves. The face detector, the stores and the cursor secret are injected ports, so the module pulls in no infrastructure. Audio handlers: 21 tests pass; eslint 0 errors; scratch typecheck reports only the deliberate control error. The refusal tests assert MESSAGE TEXT, so the reasoning cannot be deleted while the code still passes - including the substituted word ("eight" was said as "nine"), the unplaced-word wording, and the stale deviation id. Job handlers: 24 tests pass, and the 49 in the three modules that consume the widened warnings type still pass; eslint 0 errors; scratch typecheck reports only the deliberate control error - AND IT CAUGHT A SILENT BUG AGAIN: `localeNormalizable` was computed by comparing the normalizer result against an empty STRING, but that field is a BOOLEAN, so the comparison was always true and the unsupported-language warning could never fire for any locale. It now asks `isNormalizableLocale` directly. Candidate handlers: 13 tests pass; eslint 0 errors; scratch typecheck reports only the deliberate control error. The tests assert the null-versus-empty distinction directly, and that the invariant refusal stays in the internal family. Review handlers: 23 tests pass; eslint 0 errors; scratch typecheck reports only the deliberate control error. The spec asserts the exact gap in the decision-subject reader rather than describing it: the set difference against `DECIDING_OPERATIONS` must equal `accept-with-warnings, release`, so wiring one without updating the list fails. Release handlers: 19 tests pass, and 203 across the ten wiring specs together; eslint 0 errors; scratch typecheck reports only the deliberate control error - it caught a fake AuditMaster carrying `requestedAt` where `ProviderAttribution` has `upstreamProvider`. The assembly spec registers every declared route on a real Fastify instance and collects them through an `onRoute` hook rather than reading the printed radix tree, which splits shared prefixes so that `cancellation` prints as two nodes and a substring check there proves nothing. Three requests against the fully-unbound deployment confirm 503 on a read, 503 on a submission, and 401 before any store is reached. The end-to-end spec that asserted 'is not wired into the server' now asserts the mount and the REMAINING gap (jobs, candidates and releases have no adapter, so the twelve-step saga still has no production caller); create/read/save/conflict/tenant-isolation are exercised over a real Fastify instance. VERIFIED AGAINST A REAL POSTGRES: the migration applies, a project created over HTTP lands as a row whose cast_ids projection is written from the document, a stale save answers 409, and a draft carrying a consent field is refused with its sentence. The committed Prisma client is NOT regenerated in this commit — running `prisma generate` rewrites unrelated V10Stage models and adds ProductGraph models the committed copy already predates, so a deployment regenerates it rather than inheriting churn nobody authored. MOUNTING FINISHED 2026-09-12: the boot line ends "answering 503 until bound: nothing" (`evidence/2026-09-12-fourteen-ports-bound/`). Binding the last port, `releases`, found a FOURTEENTH nobody had counted: the route reads an If-Match subject before any deciding handler runs, the runtime passed `decisionSubjects: null`, and all six deciding operations - including `approve-candidate` - answered 500 inside `human_video_decision_subject_unreadable`, a refusal that was right and loud and that nobody was reading while the boot line reported candidates and reviews as bound. They were, for READING. A reader for five of the six had existed since the review store landed, with `release` deliberately omitted because its own port answered 503; what was missing was anything PASSING it, behind a stale comment saying no entity store existed. `accept-with-warnings` stays unwired for a reason no adapter changes: it is declared and has no endpoint. `decisions` is reported as its own entry because it cuts ACROSS the ports. The release proof set is DERIVED from what a candidate stores rather than written as rows, for the reason the review queue is, and `human_video_release`'s CHECK requires all sixteen gates so a row cannot exist for a take that passed fifteen. | No provider spend; no model call. | ALL TWENTY-THREE HANDLERS EXIST and the boot assertion passes. What remains is BINDING, not building: no durable adapter implements `UploadSessionStore`, `IdentityPackStore`, `AudioTakeStore`, `JobStore`, `CandidateStore`, `ReviewStore`, `ReleaseStore`, `ReleaseProofStore`, `ErasureRunner`, `DestructionLedger`, `RevocationInputs`, `ProjectStore`, `LikenessAuthorizationStore` or the `IdempotencyDelegate`, and no aligner, recognizer or audio probe is bound - every one of them fails closed with a 503 rather than guessing. `createDecisionSubjectReader` answers five of the six deciding operations; `release` still throws by design. `server.ts` mounts none of it, which is why `api-guide.md` still says no response in it was captured from a running server. |
| 182.C.25.02 | Add cursor pagination and authorization for projects, jobs, candidates, assets, and evaluations.                       | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-list-access.ts` (new), `human-video-list-access.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Pagination and authorization are ONE item because doing either separately gets the other wrong, and there are FOUR PLACES THE OBVIOUS IMPLEMENTATION LEAKS. FILTER BEFORE PAGINATING, NOT AFTER: fetching fifty rows and dropping the ones the caller may not see is the natural way to write it and it tells the caller exactly what it was meant to hide - a page of fifty that returns three says there are forty-seven rows here you cannot see. buildQuery puts the caller's reach INTO the query, and assertFilteredInQuery is exported so a store adapter cannot quietly go back to post-filtering, since the leak is invisible from outside precisely because the response looks fine. NO TOTAL COUNT: the same leak wearing a number - a count over the unfiltered set is a census of what the caller cannot reach, and a correctly filtered one costs a second full scan per page. buildPage takes limit+1 rows and uses the extra purely to answer hasMore, never returning it. AN UNSIGNED CURSOR IS AN OFFSET THE CLIENT CAN EDIT - base64 is not opacity, and a cursor decoding to {lastId, lastCreatedAt} is hand-craftable to start a scan anywhere including inside another tenant's range, so it is HMAC-signed over the tenant and the resource. AND A CURSOR MUST NOT OUTLIVE THE PERMISSION THAT MINTED IT: the trap in signing it is that a signed cursor carrying its own scope becomes a CAPABILITY that keeps working after the grant is revoked, so the scope inside is COMPARED against the caller's present grants rather than trusted - the signature proves where the scan was, and current authorization decides whether it may continue. Keyset rather than offset because this list mutates under the reader (jobs change stage constantly, and an offset page two taken after a row moved skips or repeats rows), ordered on (createdAt, id) where the id tiebreak is not decoration - two rows created in the same millisecond can swap between pages and the one that swaps out is never seen. A deleted anchor row falls back to the ordering rather than restarting at the top, so the reader is not silently handed rows they already saw. ONE THING SPECIFIC TO THIS DOMAIN: listing assets is listing FACES, so the asset scope is separate from project access rather than implied by it - 182.C.19.10 treats an embedding as a biometric derivative and a reference image is what it came from, so somebody who may see a project's jobs has not thereby been granted the faces those jobs were built from.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | 22 tests pass; eslint clean. Negative cases cover a hand-crafted cursor, four malformed ones, a cursor lifted from another tenant, one minted for a different listing, and one whose scope was revoked.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | No provider spend.                | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| 182.C.25.03 | Define idempotency, optimistic concurrency, request size, rate, quota, retry, and stable error semantics.              | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-api-semantics.ts` (new), `human-video-api-semantics.spec.ts` (new), `human-video-api-semantics.ts` (corrected), `human-video-api-store.ts` (corrected), `human-video-api-routes.ts` (corrected)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Seven semantics; most of the machinery already existed (182.C.15.07 defines every failure code with its retry disposition, blame and HTTP status; 182.C.15.06 holds per-tenant admission; 182.C.15.08 bounds retries) and what was missing is the HTTP boundary, where THREE OF THE SEVEN are wrong in the obvious implementation. AN IDEMPOTENCY KEY IS NOT IDEMPOTENT UNLESS THE REQUEST IS THE SAME - the failure everybody ships is a retry under the same key with a changed body returning the FIRST response, so the change is silently discarded, the client believes it took effect, and nothing records that two different requests shared a key. The key is bound to a canonical digest and a mismatch is 422, never a replay. Canonicalization sorts object keys (a client that serializes differently on retry has not changed its request) and PRESERVES array order (reordered shots are a different video). An unfinished first request returns in-flight rather than a replay, because there is no response to hand back and an empty success is how a client concludes a job exists that does not. OPTIMISTIC CONCURRENCY BELONGS ON THE THING BEING DECIDED, NOT THE ROW: a row version catches two people editing a job and MISSES the case that matters - approving a candidate against an evaluation that has since been re-run, where the row never changed and the EVIDENCE did. The ETag carries both, and 428 (precondition required) is kept separate from 412 because a client that never sends If-Match has no concurrency control at all and otherwise looks like it works. Same argument as 182.C.21.08 and 182.C.23.10. RETRY-AFTER FOLLOWS THE CODE, NOT THE STATUS: hv.transport.submission_ambiguous is a 5xx that must NOT be retried (the provider may have accepted it) and hv.budget.rate_limited is a 429 that must be, so retryAdvice reads 182.C.15.07's disposition and assertRetryAfterHonest REFUSES to attach a delay to a never/after_change failure - not a harmless extra header, because a well-behaved client obeys it, sleeps and re-sends the identical request, turning a settled consent refusal into a polling loop. Request size is PER FIELD, because one 30MB body may be a single reference image or sixty small ones and those are different requests to serve. Rate and quota carry the RESET TIME (a client that cannot schedule against it retries on a timer it invented, which is how a rate limit becomes a thundering herd) and the quota SCOPE (a project refusal is actionable, a tenant one is not). The error envelope takes its message from the contract's creatorMessage rather than the throw site, so one code cannot acquire six wordings across six handlers - which is how clients end up matching on message text and breaking when somebody improves a sentence. CORRECTED 2026-09-11 while declaring the 182.C.25.01 surface: the deciding-operations list originally held withdraw-consent and delete-asset, and that was WRONG. A precondition guards ADMISSION and never a SAFETY RETRACTION - the guard exists so a decision cannot be recorded against evidence that has since been re-measured, which applies to approving a take and to rejecting one alike, but for pulling a published take back, withdrawing a consent or destroying an asset the failure mode reverses: a stale ETag would BLOCK the retraction, and "your withdrawal failed, please refresh and try again" is not something to say to a person asking that their face stop being used. Those three moved to SAFETY_RETRACTIONS, which is named rather than merely absent so the omission reads as a decision, and the two lists are asserted to partition. CORRECTED 2026-09-11 while wiring the handlers, on two counts the original left open. AN IDEMPOTENCY KEY IS NOT IDEMPOTENT UNLESS IT IS CLAIMED BEFORE THE WORK: the record was written AFTER the handler returned, so two identical requests arriving together both found nothing, both ran, and both billed - which is what a client retry on a slow response looks like. The claim is now a `create` THAT CAN LOSE, because an upsert always wins and so can never report that somebody else got there first. The one case that is not a competitor is an EXPIRED row, which is an absence of memory; refusing to reserve over one would make a key unusable for good after a single abandoned call, so it is deleted and the claim retried exactly once, and the unique-violation check is narrow (P2002/23505) because treating a dropped connection as "taken" answers 409 about a request nothing is serving. AND A REPLAY HANDS BACK THE RESPONSE, NOT A POINTER TO IT: the replay answered `{replayedFrom: id}` where the first call answered a document, so a client reading a field off it got `undefined` ON THE RETRY PATH - the path that runs when something has already gone wrong. The response now travels in the record, and one over the 64 KiB replay limit is NOT KEPT and makes the retry REFUSE, because a truncated document handed back as the original is the failure the mechanism exists to prevent. THIRD STATE: `abandoned`. A call that ended without a describable outcome is neither `in-flight` (nothing is coming, and waiting is the wrong advice) nor absent (which lets the request run again and pay twice). A refusal that BLAMES THE REQUEST releases the key instead, because nothing was attempted and holding it would strand a client on a typo for the retention window.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | 27 tests pass, including a sweep asserting every one of the 62 contract failure codes yields an envelope with a 4xx/5xx status, a creator-safe message carrying no vendor name or stack, and a Retry-After consistent with its disposition. eslint clean. Typecheck run against `tsconfig.typecheck.json` WITH A DELIBERATE CONTROL ERROR included and confirmed reported: 0 errors in this module and its spec, against 974 pre-existing elsewhere in the repo. One spec assertion had encoded the error (it required withdraw-consent to be guarded) and was rewritten to assert the corrected split; 28 tests now pass. Idempotency correction: 34 semantics tests, 20 store tests and 22 route tests pass, the route ones over a real Fastify instance - including two concurrent identical submissions where the handler runs ONCE and the loser is answered from the winner's record, a request-blamed refusal whose key is released and then reused successfully, an unknown outcome that answers 409 rather than re-running (asserting the handler ran once), and an over-limit response that refuses to replay. Scratch typecheck reports only the deliberate control error.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | No provider spend.                | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| 182.C.25.04 | Generate SDK clients and examples for exact-audio active scenes, multi-reference cast, multi-speaker dialogue …        | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-sdk-generator.ts` (new), `human-video-sdk-generator.spec.ts` (new), `sdks/human-video-ts/client.ts` (generated), `sdks/human-video-ts/examples/*.ts` (5 new), `sdks/human-video-ts/README.md` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | The item says GENERATE and that word is the whole design: a hand-written client is a seventh consumer nobody wrote a contract for, drifting from the server a commit at a time with the drift invisible until somebody ELSE'S integration breaks. 182.C.25.01 already declares every endpoint with its method, path, scope and semantics and serves exactly that table, so emitting the client from the same table makes three things one thing. THE COMMITTED CLIENT IS CHECKED AGAINST THE GENERATOR, NOT JUST PRODUCED BY IT - a generator nobody re-runs is a hand-written client with extra steps, so the spec regenerates and compares BYTE-IDENTICALLY against the file on disk. Adding an endpoint without regenerating fails a test rather than shipping an SDK that cannot call it. THREE SEMANTICS THE SURFACE KNOWS AND A NAIVE GENERATOR DROPS: idempotency (every mutation sends a key BY DEFAULT rather than leaving it to the caller, because the caller who forgets pays for a second render on a timeout, and 182.C.25.03 binds the key to the body so a generated retry is safe by construction); PRECONDITIONS IN THE TYPE (a deciding method takes ifMatch as a REQUIRED parameter - not a runtime check, because a client that discovers at runtime it needed an ETag has already made the call); and RETRACTIONS TAKE NO ifMatch AT ALL, so nobody can pass a stale one and have a takedown refused. The five examples are the five the item names rather than five that were easy, and assertExamplesResolve refuses one calling a removed endpoint - a documentation bug that otherwise compiles. Placed in `sdks/` per that directory's own README (external-facing clients live there, not under libs/), which needs no Nx scaffolding.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | 17 tests pass; eslint clean. THE SDK TYPECHECK PROVED THE DESIGN CLAIM EMPIRICALLY rather than asserting it: a standalone tsc over `sdks/human-video-ts/` is clean, and a deliberately broken example that omits the ETag on a deciding call fails with "Property 'ifMatch' is missing in type ... but required" - so a client that forgets a precondition does not compile. Both that run and the BFF-side one were confirmed live with a deliberate control error first.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | No provider spend.                | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| 182.C.25.05 | Sign outbound webhooks, rotate secrets, prevent replay, retry with bounded delivery, expose delivery logs …            | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-outbound-webhooks.ts` (new), `human-video-outbound-webhooks.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Six things, and the BACKOFF WAS ALREADY SETTLED - studio/webhook-delivery-store.ts classifies a response and schedules the next attempt, so decideDelivery calls it rather than growing a second copy (a second backoff implementation is how one of the two quietly becomes gentler). FOUR OF THE REMAINING FIVE HAVE A WRONG OBVIOUS ANSWER. ROTATION WITHOUT AN OVERLAP IS AN OUTAGE: swap a secret atomically and every receiver that has not updated starts rejecting, including deliveries already in flight, so both signatures are sent and either verifies - which is what lets the receiver move on ITS schedule rather than ours. AND AN OVERLAP WITHOUT AN END IS TWO LIVE SECRETS FOREVER: assertRotationCompleted refuses an old secret with no expiry, or one left live past it, because a permanent overlap is not a rotation - it is a second live secret nobody is tracking, and the old one was rotated away for a reason. SIGNING THE BODY ALONE LETS A CAPTURED DELIVERY BE REPLAYED FOREVER: the timestamp goes INSIDE the signed material, which is what 182.C.11.06 already does inbound reading OpenRouter's own scheme, and the EVENT ID goes in too so a replay cannot keep its signature while claiming to be a different event. The window is checked in both directions, because a clock skewed forward is the same replay window pointed backwards. A DELIVERY LOG IS WHERE A REDACTED EVENT GETS UN-REDACTED: "log the full event so support can see what was withheld" puts the script the integrator never received into a log with the integrator's endpoint beside it - a WORSE disclosure than the one 182.C.25.07 prevented, because it also records who it was for. The log holds a sha256 of the delivered body and the audience it was projected for. AND A DEAD ENDPOINT IS A SECURITY PROBLEM, NOT A NUISANCE: a URL failing for three days is somebody's decommissioned server, domains lapse and get re-registered, and a host that stopped resolving can start resolving again for somebody else - so continuing to POST a person's likeness there is the risk rather than the failed requests. Disablement triggers on 50 consecutive failures OR 72 hours failing (a URL nobody retried much can still have been dead for three days), and the reason names that risk and asks the operator to confirm the endpoint is still theirs. decideDelivery checks the endpoint FIRST, because retrying against a disabled one is exactly what disabling was meant to stop.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | 26 tests pass; eslint clean. Verification includes a receiver-side verify so the spec proves BOTH secrets validate during a rotation, and negative cases for a late replay, a backwards-skewed clock, a swapped event id, a tampered body and four malformed headers (none of which throw).                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | No provider spend.                | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| 182.C.25.06 | Define events for every durable job/review/release/deletion state with tenant, project, request, shot, candidate …     | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-events.ts` (new), `human-video-events.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | 182.C.15.04 already carries stage transitions out transactionally through an outbox, and TWO THINGS ARE MISSING for a public stream. NOT EVERY DURABLE STATE IS A STAGE TRANSITION: a job sits in awaiting-review while a reviewer approves candidate A and rejects candidate B - two durable decisions, and the stage never moved, so a stream built on transitions emits NOTHING. Release is the same shape and DELETION is worse, because destroying a reference image may change no job stage at all and is the single most important event to emit, since every consumer holding a derived copy has to drop it. 12 of the 15 event types do not correspond to a transition. CORRELATION IS NOT METADATA - without it an event cannot be joined, and CORRELATION_PURPOSE keeps the WHY as data because the refusal quotes it: somebody told "correlation incomplete" learns nothing. The field that earns its place least obviously is ATTEMPT: without it two rendering->downloading events for one job are indistinguishable from a duplicate delivery, so a consumer dedupes the second render away - and 182.C.15.08 RETRIES, which makes that the normal case rather than the odd one. isDuplicateOf makes that concrete. THE TRACE MUST BE CAPTURED AT TRANSITION TIME: the outbox exists so the publish happens LATER in another process, so a relay stamping the current trace stamps its own - the batch that shipped the event rather than the request that caused it - which is WORSE than no trace because it looks like one and joins to the wrong span. A DELETION EVENT CARRIES THE LEAST OF ANY EVENT: an allowlist of seven keys, because a denylist that misses one fails OPEN and the thing it leaks is the content of something somebody asked to be destroyed, to every subscriber and every log that retains events. That is the one place in the surface where the failure is unrecoverable, so it is structural rather than filtered.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | 18 tests pass; eslint clean. Typecheck against `tsconfig.typecheck.json` with a deliberate control error confirmed reported: 0 errors in this module and its spec, against 974 pre-existing elsewhere.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | No provider spend.                | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| 182.C.25.07 | Redact private script/reference/audio/provider data from events and separate creator-safe measurement summaries from … | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-event-redaction.ts` (new), `human-video-event-redaction.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Two jobs that look like one: keeping private content out of events is about CONTENT, separating a creator-safe summary from forensic evidence is about AUDIENCE, and a single "redact" pass conflates them - which is how a field that is fine for the tenant who owns the script ends up on a webhook to a third party. SO THERE ARE THREE AUDIENCES, NOT TWO: owner (their own content, it is theirs), INTEGRATOR (somebody the tenant pointed a webhook at - they may see that a job finished and how it scored, and not the script, because subscribing an endpoint is not consenting to publish a person's likeness to it), and internal. The middle one is the audience everybody forgets and the one an event stream actually has. AN ALLOWLIST, BECAUSE A DENYLIST FAILS OPEN: projectFor NAMES what may cross, so a field added to an internal record later is absent from every projection - including internal - until somebody classifies it, because the point is to NOTICE it rather than route it. The transcript is classified with the script, since redacting one and shipping the other redacts nothing. The face embedding reaches NOBODY outside, including the owner, because there is nothing an owner can do with it and every copy is a liability (182.C.19.10). A MEASUREMENT SUMMARY IS NOT A ROUNDED FORENSIC RECORD: the tempting implementation ships the evidence with the paths stripped, and it fails because the forensic record is forensic in its SHAPE - a per-window offset series with timecodes reconstructs the performance, where the speaker paused and which words were hard. summarise BUILDS a five-field object instead, and assertNotForensicShape refuses a per-item series by SHAPE rather than by field name, because the next series added will have a different name and the same problem. AND WITHHOLDING IS VISIBLE: a projection reports what it kept back by name and never by value, because an event silently missing a field is indistinguishable from one where the field was absent and a consumer cannot tell "no findings" from "findings you may not see". assertNoPrivateContent checks the OUTPUT rather than trusting the builder, which is the thing most likely to be changed by somebody adding a field in a hurry.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | 20 tests pass; eslint clean. One test drove a real improvement: three content classes carried reasons too thin to act on (audioBytes was "somebody's voice"), and the reasons were deepened rather than the assertion lowered. Typecheck against `tsconfig.typecheck.json` with a control error confirmed reported.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | No provider spend.                | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
| 182.C.25.08 | Add consumer-driven contract tests for Studio, mobile, Veritas, admin, analytics, and external SDK clients.            | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-consumer-contracts.ts` (new), `human-video-consumer-contracts.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | THE POINT OF A CONSUMER-DRIVEN CONTRACT IS NOT THAT THE PROVIDER IS TESTED AGAIN FROM ANOTHER ANGLE - it is that the provider learns what it is FREE TO CHANGE. A contract that re-asserts the whole response is a provider test wearing a different name: it passes, it feels thorough, and it locks every field including the ones nobody reads, so assertContractIsNarrow refuses a consumer claiming more than two thirds of a document (a judgement, stated as one so nobody reads it as fitted). All six contracts are DECLARED in the spec in each consumer's own terms - what it calls, what it reads, what it branches on - and changeBudget then answers the question the whole exercise exists for: candidate.durationSeconds and release.releaseId are read by nobody and can be removed. whoBreaks NAMES THE CONSUMERS rather than answering yes or no, because "this breaks analytics and the SDK" is a conversation and "breaking change" is a wall. THREE KINDS OF CHANGE, ONLY TWO OF WHICH A SCHEMA CAN SEE: adding is safe, removing breaks exactly the consumers that named it, and CHANGING WHAT A FIELD MEANS breaks consumers while every schema check and every type stays green - measurableFraction moving from "of windows" to "of seconds" is the same name, the same type, the same range and a different number, so a contract pins an INTERPRETATION and detectMeaningDrift compares interpretations rather than types. AND A CONTRACT CAN BE UNSATISFIABLE BEFORE ANY TEST RUNS: 182.C.25.07 gives each consumer an audience, so analytics (an integrator) declaring it reads scriptText is broken by construction, and the SDK cannot read the per-window series while Veritas (internal) can. Worth catching at the contract rather than at the first empty field in production, because absent reads as "there is none" rather than as "you may not have this". NOTED 2026-09-12, AND IT IS A REAL LIMIT OF WHAT THIS ITEM CHECKS: the six contracts are SPEC FIXTURES, and the machinery checks a contract's DECLARATION against the provider. Nothing checks a consumer's CODE against its declaration - so two hand-written clients had drifted from the surface and these tests were green throughout. The mobile client opened sessions with the wrong body field, posted chunks to a route that does not exist, and carried the offset in a header the server ignores; the admin sent two query parameters the service does not read. `scripts/human-video/verify-client-api-calls.mjs` closes the loop for ALL SEVEN hand-written clients - web, admin and mobile - by checking the CODE against the surface directly, which is stronger than code->contract->surface. The generated TypeScript SDK is deliberately absent from it: it already has a byte-identical drift guard and cannot drift. The item is not re-opened: it asked for consumer-driven contract tests and it built them; this records what they do not reach.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | 23 tests pass; eslint clean. The six contracts are real declarations checked against the live surface and failure-code vocabulary, so a removed endpoint or a retired code fails them. Typecheck against `tsconfig.typecheck.json` with a control error confirmed reported.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | No provider spend.                | None The contracts are fixtures rather than declarations any client is held to, so a consumer's code can still drift from its own contract - closed for all seven hand-written clients by `verify-client-api-calls.mjs`.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |

### 182.C.26 Consent, Likeness, Voice, and Impersonation Safety

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner          | Dependency | Status | Changed Files                                                                                                                   | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Cost                                                                                                                | Blocker |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | -------------- | ---------- | ------ | ------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------- | ------- |
| 182.C.26.01 | Define explicit likeness and voice authorization scopes by subject, tenant, project, purpose, media type, geography … | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-likeness-authorization.ts` (new), `human-video-likeness-authorization.spec.ts` (new) | NOTHING HERE STORES A CONSENT, which is what the item asks for: evaluateVoiceConsent still decides whether a voice record is live, @aja/consent-management still holds the enrolment record and its withdrawal, and the libs/oshun/trust-safety taxonomy still classifies content. This is the SCOPE VOCABULARY those three lack and the resolver that composes them, and LivenessInputs is deliberately an INPUT rather than something computed here - re-deriving either store's answer would be the fourth consent model the item forbids. THE NINE DIMENSIONS ARE NOT THE SAME KIND OF THING AND TREATING THEM UNIFORMLY IS THE BUG: media type, provider class and transformations are ALLOWLISTS, geography and duration are BOUNDS, tenant and project are HIERARCHICAL, purpose is a controlled vocabulary. Which matters because of one rule that has to hold across all of them - AN UNSPECIFIED DIMENSION IS A REFUSAL, NEVER A WILDCARD. A scope that forgot to say which geographies is not a global scope, and the empty list has to mean "none" everywhere or it means "all" somewhere, and the somewhere will be whichever dimension was added last by whoever did not know the rule existed. The single exception is transformations, where empty legitimately means the likeness unaltered. PROVIDER CLASS IS THE DIMENSION NOBODY THINKS OF AS CONSENT: agreeing that a video be made is not agreeing that your face be sent to any vendor who will render it, and the contract already records that OpenRouter video is not zero-data-retention eligible at all - so "any provider" and "only providers that retain nothing" are materially different permissions for the same person, and only one is what most people picture. Classes are named by WHAT THE VENDOR DOES rather than by vendor name, because a person consents to a property of the handling and vendor names change under them. AND ALLOWED TRANSFORMATIONS IS WHERE IMPERSONATION LIVES: consent to appear is not consent to be de-aged, put in a different room, or have your mouth reanimated to words you did not say, and every one of those is technically "your likeness in a video". IMPERSONATION_RISK names the five that change what a person appears to have DONE OR SAID, and the refusal message says so. A project grant that names something its tenant grant never had is refused as a new permission wearing a narrow name. And liveness is checked BEFORE scope, so a withdrawn consent does not produce an argument about geography - the subject's answer is the answer.                                  | 28 tests pass; eslint clean. One test asserts the provider-class vocabulary contains no vendor names, so the "by property, not by name" decision cannot erode.                                                                                                                                                                                                                                                                                                                       | No provider spend.                                                                                                  | None    |
| 182.C.26.02 | Require proof/attestation at reference and voice enrollment, bind it to asset digests …                               | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-enrolment-proof.ts` (new), `human-video-enrolment-proof.spec.ts` (new)               | Three requirements, and THE MIDDLE ONE IS THE WHOLE ITEM. A PROOF THAT IS NOT BOUND TO THE ASSETS PROVES NOTHING ABOUT THEM: the common shape is a consent record with a signature and a timestamp, and separately a set of uploaded images, with nothing connecting them - swap the images afterwards and the record still says "this person agreed", still verifies, and is now attached to photographs the person never saw. IT WAS NEVER FALSE; IT WAS NEVER ABOUT ANYTHING. So the proof covers a canonical digest over the asset SET rather than each item, with the COUNT inside the hash, because binding each digest individually still lets somebody append a sixth asset and claim the proof covers it. Ordering is normalised, so the order assets happened to be uploaded in is not part of the proof. Adding a reference later is a NEW ENROLMENT requiring new proof, because "they agreed to these four photographs" cannot be stretched to five by appending. THE VERIFICATION METHOD IS PART OF THE PROOF, NOT METADATA ABOUT IT: a checkbox and a countersigned release are both "consent" and are not the same evidence, and a record saying only consented:true has thrown away what a reviewer needs - the five methods are ORDERED so 182.C.26.08 can require "at least this strong", which only works because the method travels with the record. AND THE CONSENT VERSION HAS TO REACH THE JOB FROM THE ENROLMENT: consent text changes, and a job running today under a pack enrolled last year was authorized by the text as it read LAST YEAR, which is the text any dispute is about. Stamping the current version would be a small quiet falsification of exactly the record that matters - easy to write by accident, because the current version is the one in scope at the call site, and impossible to notice afterwards because the stamped record looks entirely correct. acceptEnrolment checks the BINDING FIRST, because a proof that is not about these assets makes every later question moot.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | 20 tests pass; eslint clean. Coverage includes an appended asset, a SWAPPED asset with the count unchanged, a reversed upload order (which must not change the digest), a lapsed proof, and a stamp carrying the version in force today.                                                                                                                                                                                                                                             | No provider spend.                                                                                                  | None    |
| 182.C.26.03 | Handle third-party/public-figure likeness, parody/satire, political content, sexual content, fraud, harassment …      | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-content-policy.ts` (new), `human-video-content-policy.spec.ts` (new)                 | WHAT THIS MODULE IS NOT, stated in its header: it is not legal advice and it does not encode the law of any jurisdiction. It is a PLACE TO PUT those decisions, with the shape of each fixed so a lawyer's answer can be entered without redesigning anything. THREE OF THE SEVEN CLASSES ARE NOT JURISDICTIONAL AT ALL - fraud, harassment and deceptive impersonation are prohibited everywhere this operates, and modelling them as varying by region invites a lookup that returns "allowed" for somewhere when there is no somewhere. They stay IN the table so the table is the whole list, because a policy surface that omits them implies they live elsewhere and somebody will go looking. THE OTHERS VARY IN DIFFERENT DIRECTIONS: public-figure protection is WEAKER in some places and STRONGER in others, so a rule written either way is wrong somewhere and there is no safe global default - publicFigurePosture refuses rather than assuming. Parody is a defence in some systems and not a category in others, and is the class most often claimed by things that are not it. Political content attracts DISCLOSURE REQUIREMENTS rather than prohibitions in most places, so the common failure is not generating something forbidden - it is shipping something lawful WITHOUT THE LABEL THAT MADE IT LAWFUL. AND THE JURISDICTION IS NOT ONE PLACE: the tenant is somewhere, the SUBJECT is somewhere, and the audience is somewhere, routinely three different somewheres - an engine taking "the jurisdiction" has already decided which one matters without saying so. The subject's region is the one most often left out, because the tenant has no particular reason to record it and it is the person whose likeness it is. The strictest posture across all three governs, and the detail says explicitly that this is NOT legally exhaustive - it is chosen because being wrong permissively is the direction that hurts the depicted person, and this system cannot adjudicate a conflict of laws. AN UNSET POSTURE REFUSES IN BOTH DIRECTIONS: permissive-by-default is unsafe, and prohibited-by-default silently blocks lawful work WHILE LOOKING RESPONSIBLE, which is worse than an error because nobody investigates a refusal that seems careful. emptyRegionPolicy pre-fills nothing, since a guessed posture is indistinguishable from a researched one once it is in the table and the guess is what somebody would rely on. coverage reports how much is still unset, because an empty table and a completed one look identical from outside until somebody hits a refusal. | 22 tests pass; eslint clean. The three prohibited-everywhere classes are asserted refused even when all three applicable regions are set to permitted, so no configuration reaches them.                                                                                                                                                                                                                                                                                             | No provider spend. No legal position is asserted; every region posture ships UNSET for a qualified person to enter. | None    |
| 182.C.26.04 | Add stricter minor safeguards, guardian authorization, age-uncertainty escalation, prohibited scenarios, retention …  | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-minor-safeguards.ts` (new), `human-video-minor-safeguards.spec.ts` (new)             | THE LOAD-BEARING WORD IN THE ITEM IS "AGE-UNCERTAINTY", AND IT CHANGES THE SHAPE OF EVERYTHING ELSE. 182.C.19.02 already recorded the governing finding: estimating age from face geometry has a documented record of SKEWING BY ETHNICITY, and this phase declined to ship an estimator for that reason - which means its errors would not fall randomly, they would fall on particular children. So THERE IS NO AGE DETECTOR AND THERE IS NOT GOING TO BE ONE, and the system must behave correctly WITHOUT an answer. assertNotFaceEstimated keeps that as an explicit refusal rather than an omission, because "we could just run an age model on the reference photo" is a reasonable-sounding suggestion and the reason against it is not obvious unless somebody wrote it down. THE ONLY SAFE DEFAULT UNDER UNCERTAINTY IS THE STRICT ONE, and the argument is an ASYMMETRY rather than a preference: wrongly applying minor safeguards to an adult costs friction, and wrongly treating a child as an adult costs the thing the safeguards exist to prevent - those are not comparable, so the default cannot be chosen by weighing them evenly. `uncertain` is its own band, and a SELF-DECLARED ADULT DOES NOT RESOLVE TO ADULT: a checkbox saying "I am over 18" is not evidence about the person in the photograph at all, it is evidence about whoever clicked it. GUARDIAN AUTHORIZATION IS NOT A SECOND CONSENT, IT IS THE ONLY ONE - and the minor's ASSENT is recorded alongside it and separately required, because a child who does not want to be in the video should not be in the video whatever the guardian signed; a single consented flag cannot express that. PROHIBITED SCENARIOS ARE NOT GATED, REVIEWABLE OR OVERRIDABLE, because a reviewable prohibition is a strongly worded default - and the list is SHORT ON PURPOSE, since a long one invites the reading that everything not on it is fine. Speech reanimation and voice cloning are refused by the TRANSFORMATION as well as by the scenario name, because the scenario is what somebody typed and the transformation is what the pipeline will actually do. Retention is capped at 90 days against an adult 365, and access narrows from a held role to a written list.                                                                                                                                                                                                                                                                                                                                                       | 30 tests pass; eslint clean. The uncertain band is exercised through every control - authorization, prohibition and retention - so it cannot be strict in one place and permissive in another.                                                                                                                                                                                                                                                                                       | No provider spend.                                                                                                  | None    |
| 182.C.26.05 | Make consent revocation stop new work immediately, cancel eligible in-flight jobs, block unreleased candidates …      | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-revocation.ts` (new), `human-video-revocation.spec.ts` (new)                         | Four things have to happen and they are ONE item because doing three of them is indistinguishable from doing none - the person's face is still being used. THE FOUR REACH DIFFERENT DISTANCES AND ONLY ONE IS INSTANT. NEW WORK stops immediately and completely: the only one with no caveat, checked at EVERY STAGE BOUNDARY rather than once at submission, because the race that matters is a job that passed its consent check a second before the withdrawal and there was no way to have checked later. IN-FLIGHT JOBS are "eligible" for cancellation and that word carries 182.C.15.05: there is no cancel endpoint upstream, so we stop waiting while the provider's render finishes on their hardware and is billed. Cancelling does NOT stop the person's face being processed by the vendor - it stops the RESULT reaching anybody, and saying "cancelled" without saying that tells the subject something untrue. ALREADY-RELEASED work is a TAKEDOWN, not a deletion, and is the one this system cannot complete alone: a signed link handed out works until it expires, a CDN holds copies, a viewer may have downloaded it. 182.C.16.07 clamps the cache to the signature's own lifetime, which BOUNDS it - and a bound is not a retraction. COPY_REACHABILITY classifies all eight copy locations by whether this system can actually destroy them, and takedownReach reports the unreachable ones rather than omitting them, because a report listing only successes READS AS COMPLETION to the one person entitled to know it is not. assertNotOverclaimed refuses a "complete" claim while copies remain: telling a subject their likeness has been removed when it is still served is worse than telling them nothing, because it ends the conversation. The zero-retention provider class is the single case where an unreachable location is genuinely empty - which is why provider class is a consent dimension in 182.C.26.01. The subject summary is written for the person who withdrew, so it states what did NOT stop.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | 21 tests pass; eslint clean. A REAL BUG was found and fixed by a failing test: the disposition was classified on stage TERMINALITY, which made `releasable` - whose render has already happened and been billed - fall through to "nothing had been submitted", a false thing to tell a person asking what stopped. Re-classified on whether provider-render is sunk in 182.C.15.05's own table, and a test now pins the rule across all 15 stages so terminality cannot creep back. | No provider spend.                                                                                                  | None    |
| 182.C.26.06 | Prevent voice/face packs from being shared across tenants, projects, purposes …                                       | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-pack-isolation.ts` (new), `human-video-pack-isolation.spec.ts` (new)                 | 182.C.26.01 says what a scope permits; this is what makes the scope mean anything. THE FOUR BOUNDARIES ARE NOT EQUALLY LEAKY AND THE LEAKS ARE NOT WHERE PEOPLE LOOK - nobody accidentally passes another tenant's pack id, because the first WHERE tenant_id anybody writes catches that. The real ones: COPYING ("duplicate this project to start the next campaign" carries the pack, and afterwards the copy is legitimately in the new project and EVERY READ-TIME CHECK PASSES), so assertCopyAllowed checks the ACT rather than the read. THE DERIVED ARTEFACT: 182.C.19.10 records that an embedding outlives the image it came from and is what makes re-identification possible, so a pack that may not cross while its embedding may is not isolation - it is isolation of the least useful copy. A COLLABORATOR WHO LEAVES: a grant recorded when the pack was shared does not end when the person does, so membership is read at USE time; what they already downloaded is 182.C.26.05's takedown problem and is not pretended otherwise. PURPOSE has NO TECHNICAL EDGE AT ALL - nothing about a pack changes when a marketing project starts making political content - so it can only be enforced by re-checking the authorization on every use, with the purpose of THIS job rather than the one the pack was enrolled under (the enrolment purpose is what happened to be true on the day; the authorization is the authority). AND SHARING IS NOT A PROPERTY OF THE PACK: a sharedWith list makes the PACK the authority on who may see a person's face, and it is not - the person's authorization is. deriveShare computes a share from the authorization and refuses to mint one it does not support, which means narrowing the authorization narrows the shares with no cleanup step for anybody to forget, and a share derived from a superseded authorization simply stops surviving.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | 20 tests pass; eslint clean. Coverage includes all four derived-artefact kinds crossing, a share that outlives its authorization, and the enrolment-purpose-versus-authorization case in both directions.                                                                                                                                                                                                                                                                            | No provider spend.                                                                                                  | None    |
| 182.C.26.07 | Run image, audio, prompt, script, and produced-video safety checks at the appropriate stages …                        | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-safety-stages.ts` (new), `human-video-safety-stages.spec.ts` (new)                   | Five surfaces and one sentence that carries the item: POST-GENERATION SAFETY IS MANDATORY. WHY THAT SENTENCE IS THERE AND WHY IT IS NOT OBVIOUS - every pre-generation check is a check on an INPUT, and THE OUTPUT IS NOT A FUNCTION OF THE INPUTS ALONE. A safe reference photo, a safe script and a safe prompt can produce a video that is none of those things: the model contributes, and what it contributes is exactly what nobody screened. A pipeline that screens all four inputs thoroughly and skips the output has screened everything except the thing it publishes. assertPostGenerationRan refuses a release without it AND REFUSES IT ON THE GROUNDS THAT THE INPUTS PASSED, because that is the argument that gets made. THE FIVE SURFACES ARE NOT INTERCHANGEABLE AND TWO CANNOT BE CHECKED WHERE PEOPLE PUT THE CHECK: a reference image screened after enrolment has already been stored, embedded and sent to a provider, so refusing it then does not undo any of that - it is not a late check, it is a different activity, and assertCheckedInTime refuses it. Audio is the same before approval. The PROMPT is an intention and screening intentions says nothing about outputs. The SCRIPT is screened for what the words DO when a real person appears to say them - defamation, a false confession, a fabricated endorsement - rather than only for slurs. The PRODUCED VIDEO is the only surface where the thing screened and the thing shipped are the same thing, and it is the one most pipelines make optional because it is the most expensive. AND A SKIPPED CHECK IS NOT A PASSED CHECK: every outcome is three-valued, and assertOutcomeHonest refuses a pass that names no stage it was checked at - a not-run recorded as a pass is how an unscreened video ships with a green tick beside it. The gate keeps `failed` and `unscreened` as separate lists, because a failure is a decision somebody made about this content and an unscreened surface is a question nobody asked.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | 22 tests pass; eslint clean. Every surface is asserted screenable at its own latest useful stage, and each carries a stated blind spot so a reviewer knows what the check does not cover.                                                                                                                                                                                                                                                                                            | No provider spend.                                                                                                  | None    |
| 182.C.26.08 | Add step-up authentication and human review for high-risk likeness, political, financial, medical, legal …            | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-step-up-review.ts` (new), `human-video-step-up-review.spec.ts` (new)                 | TWO CONTROLS ANSWERING DIFFERENT QUESTIONS, AND COLLAPSING THEM IS THE MISTAKE THIS FILE PREVENTS. Step-up authentication answers "is the person making this request who the account says they are" and defends against a COMPROMISED SESSION. Human review answers "should this be made at all" and defends against the LEGITIMATE ACCOUNT HOLDER. A pipeline with strong step-up and no review is well defended against somebody else's account being used and undefended against the owner - and most abuse of a likeness system is the second. assertReviewed reports review_substituted_for_step_up specifically when the actor DID authenticate strongly, because that is exactly when the substitution argument gets made, and the message says a stronger login does not make a fabricated endorsement safer. The reviewer may not be the requester, since the control defends against the account holder. RISK IS A PROPERTY OF THE COMBINATION, NOT OF ANY SINGLE FIELD: a political script is not high risk on its own, and a political script performed by a real identifiable person who did not write it is - so classify multiplies the topic against 182.C.26.01's IMPERSONATION_RISK transformations and against whether the subject is a real person, rather than reading the topic alone. The same transformation on a synthetic subject classifies as nothing. AND MASS DISTRIBUTION IS THE ONE THAT ARRIVES AFTER THE DECISION: reach is a property of what somebody does with the file afterwards, not of the content, so it is the only class that cannot gate generation - it gates RELEASE, and assertReviewAtRightPoint refuses a review attached to the wrong stage, because a requirement attached to the wrong stage is one that never runs. decide checks step-up BEFORE review, so a compromised session never reaches a human reviewer's queue - sending it there spends a person's attention on a request that was never legitimate. Step-up strength is taken as the STRONGEST across the classes present, reusing 182.C.26.02's method ordering.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | 22 tests pass; eslint clean. Coverage includes the same transformation on a real versus a synthetic subject, a topic with and without a real person carrying it, and a review attached to the wrong stage.                                                                                                                                                                                                                                                                           | No provider spend.                                                                                                  | None    |
| 182.C.26.09 | Maintain appeal, takedown, abuse-report, incident, evidence-preservation, and subject-notification workflows.         | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-appeals.ts` (new), `human-video-appeals.spec.ts` (new)                               | Six workflows sharing one property that decides the design: EVERY ONE IS STARTED BY SOMEBODY WHO IS NOT THE TENANT. An appeal comes from a creator we refused; a takedown and an abuse report come from the DEPICTED PERSON or somebody acting for them, WHO VERY OFTEN HAS NO ACCOUNT HERE AT ALL. That is the requirement everybody misses - a takedown route behind a login is one the person whose face it is cannot use, and they are the one it exists for, because they were never the customer. assertOpenToUnauthenticated requires a CONTACT FOR THE REPLY rather than an account, which are different things and only the second is illegitimate to insist on. EVIDENCE PRESERVATION AND DELETION PULL IN OPPOSITE DIRECTIONS AND BOTH ARE RIGHT: 182.C.26.05 destroys every copy it can reach and an incident needs the evidence kept, and resolving that with "keep everything during an incident" QUIETLY CONVERTS A REPORT OF MISUSE INTO A REASON TO RETAIN THE MISUSED LIKENESS INDEFINITELY - the opposite of what the reporter wanted. So a hold NAMES THE EXACT ARTEFACTS, states why, carries an expiry, and does not suspend deletion of anything it did not name: planDeletionUnderHolds still deletes the rest, because an incident about one video is not a reason to keep a person's whole reference pack. A hold with no expiry is a retention policy wearing an incident's name, and assertHoldsRelevant refuses a deletion blocked only by holds naming none of its artefacts. An appeal heard by its own decider is a second look by the same eyes. AND NOTIFYING THE SUBJECT IS NOT NOTIFYING THE REPORTER: they are often different people, the subject may not know a report was made, and telling them is usually right and OCCASIONALLY HARMFUL - a notice to somebody whose depicted content was reported by a person they are hiding from discloses that person's activity. So it is a decision with a recorded reason in EITHER direction, because a silent default is the failure whichever way it points. Acknowledgement is tracked separately from resolution, because a takedown that will take a week is bearable if somebody said so on day one and unbearable in silence, and the depicted person gets the shorter target.                                                                                                                                                                                                                                                                                                                                              | 26 tests pass; eslint clean. Coverage includes an expired hold being ignored, a hold naming none of the artefacts being deleted, and both directions of the notification decision.                                                                                                                                                                                                                                                                                                   | No provider spend.                                                                                                  | None    |
| 182.C.26.10 | Test consent expiry/revocation races, collaborator removal, job retry after revocation, provider fallback …           | Trust & Safety | 182.C.06   | done   | `apps/oshun/bff/src/generation/human-video-consent-races.spec.ts` (new)                                                         | Tests rather than a module, because the item asks to TEST and because each of the six is a QUESTION ABOUT HOW THE PIECES INTERACT - 26.01's scopes, 26.02's proofs, 26.05's revocation and 26.06's isolation each behave correctly alone, and the failures live in the seams. EXPIRY IS TREATED EXACTLY AS REVOCATION, with no separate path, because two code paths is how one of them ends up missing a check; and a use is evaluated at the time it HAPPENS rather than at submission, so a job started a second before expiry does not finish under it. JOB RETRY AFTER REVOCATION is refused at the retrying boundary rather than inheriting its first approval - a retry that inherits the original check is the exact shape of "we already checked this job". TWO OF THE SIX HAVE NO CLEAN ANSWER AND THE TESTS SAY SO RATHER THAN ASSERTING A COMFORTABLE ONE. PROVIDER FALLBACK can move a render to a vendor in a class the subject never agreed to, and THE ROUTING LAYER DOES NOT KNOW THAT, because provider class is a CONSENT dimension rather than a capability one - a fallback chosen for availability can land outside what a person permitted, and the class also decides whether a later deletion can ever be complete. CACHED RELEASE URLS OUTLIVE A REVOCATION BY DESIGN: 182.C.16.07 clamps the cache to the signature's own lifetime, which BOUNDS the exposure and does not end it, so the test pins the bound rather than pretending it is zero. Collaborator removal stops the next use immediately and does NOT pretend to recall what they already hold. Derived assets cannot cross a boundary their source may not, are destroyed by a revocation because they are ours to destroy, and invalidate an enrolment proof when the set changes under it. And shares disappear when the authorization narrows, with no cleanup step for anybody to forget.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | 17 tests pass across all six named races. One iterates every one of the 15 stages to prove the revocation refusal has no gap, and the cached-URL test asserts the refusal message says a bound is not a retraction.                                                                                                                                                                                                                                                                  | No provider spend.                                                                                                  | None    |

### 182.C.27 Privacy, Security, and Provider Data Governance

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner          | Dependency | Status | Changed Files                                                                                                             | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Cost               | Blocker                                                                                                                                                                                                                               |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | -------------- | ---------- | ------ | ------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.27.01 | Classify face images, face embeddings, voice recordings, voice embeddings, scripts, videos, and evaluator artifacts … | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-data-classes.ts` (new), `human-video-data-classes.spec.ts` (new)               | Seven classes, and THREE THINGS ABOUT THE LIST ARE LOAD-BEARING. THE CLASSES HAVE DIFFERENT RISK HALF-LIVES SO RETENTION IS NOT ONE POLICY: a script's risk is roughly constant, and A FACE EMBEDDING'S RISK GROWS, because re-identification gets easier as other datasets accumulate - a vector near-useless when written becomes a lookup key later WITHOUT ANYBODY TOUCHING IT. So both embedding classes are held 90 days against their sources' 365, and assertDerivedNotLonger refuses the inverse: a derivative outliving its source is the exact shape of 182.C.19.10's problem, where the source is deleted, the record says the likeness is gone, and the thing that identifies the person is still there. MINIMUM AND MAXIMUM RETENTION PULL OPPOSITE WAYS AND CAN BE UNSATISFIABLE: the item asks for MINIMUM retention, which is the unusual direction - a minimum exists because an audit or dispute needs the record, a maximum because the artefact is somebody's face, and THE TWO ARE WRITTEN BY DIFFERENT PEOPLE FOR DIFFERENT REASONS WITH NOBODY CHECKING THEY ARE COMPATIBLE. assertRetentionSatisfiable refuses rather than silently picking one, and the message names what each choice would cost: "we destroyed the evidence" or "we kept the face". AND THE LEGAL BASIS IS NOT A LABEL - IT DECIDES WHAT A WITHDRAWAL DOES. Consent-based processing stops on withdrawal; contract and legitimate-interest processing does not, so effectOfWithdrawal answers PER CLASS and says plainly that "withdraw at any time" without that caveat is misleading. assertContinuationConsistent catches the contradiction the record itself makes visible - basis consent, consent withdrawn, processing continuing - which is easy to reach by accident because the withdrawal and the processing are handled by different code. THE EVALUATOR ARTIFACT IS THE CLASS THAT GETS FORGOTTEN and it is classified biometric: a frame extracted so a reviewer could look at a mouth is the same likeness as the reference image, stored under a name that does not sound like one. And an UNCLASSIFIED artefact is refused, because it has no retention and no basis - it is kept forever under nothing, which is what a new feature produces by default since the classification table is not the file anybody edits when adding one.                                                                                                                                                                                                                                                                                                                                                                        | 22 tests pass; eslint clean. The script is deliberately the one class kept LONGER (365-day minimum), because destroying the record of what a person was made to appear to say protects nobody except whoever wrote it.                                                                                                                                                                                                                                                                                                                         | No provider spend. | None                                                                                                                                                                                                                                  |
| 182.C.27.02 | Document OpenRouter asynchronous-video retention and non-ZDR status in policy, UI, tenant controls …                  | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-zdr-posture.ts` (new), `human-video-zdr-posture.spec.ts` (new)                 | The FACT was already recorded in the capability vocabulary and enforced by the eligibility check; what this adds is the other four surfaces, WHICH FAIL DIFFERENTLY. ROUTING already refuses, and the only addition is that the refusal must not be mistaken for a transient one - zdr_tenant_ineligible is not "no capacity right now", so a client that retries it retries forever. THE UI FAILURE IS THAT THE CONTROL AND ITS CONSEQUENCE LIVE APART: a tenant toggling ZDR on in settings learns what it cost at their NEXT SUBMISSION, in a different screen, possibly days later, as an error - so previewToggle computes it BEFORE the toggle and states it in routes lost rather than in policy language. THE TENANT-CONTROL FAILURE IS THE INTERESTING ONE: ENABLING ZDR DOES NOT UN-SEND WHAT WAS ALREADY SENT. A tenant enabling it today has data at the provider from yesterday; the toggle changes future routing and CANNOT CHANGE HISTORY, and a settings page reading "zero data retention: on" without that is making a claim about the past that is not true - the same overclaim 182.C.26.05 refuses for takedowns. assertNoRetroactiveClaim takes the CLAIM rather than the setting, because a settings page, a compliance export and an answer to a customer can all make it by omission. AND THE EMPTY ROUTE LIST NEEDS AN EXPLANATION: a ZDR tenant sees zero eligible routes, and an empty list with no reason READS AS A BUG - which is how somebody ends up turning the setting off to make the product work, the exact opposite of what it was for. The explanation says it is the setting working rather than a fault, names how many routes turning it off would restore, and calls that a decision about the tenant's data rather than a workaround. The policy disclosure is ASSEMBLED from the same ZDR_FACT the routing uses, because a policy document becomes wrong not by being written badly but by being written once, separately from the behaviour.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | 19 tests pass; eslint clean. One asserts the policy text is identical to the routing fact rather than a paraphrase of it, so the two cannot drift.                                                                                                                                                                                                                                                                                                                                                                                             | No provider spend. | None                                                                                                                                                                                                                                  |
| 182.C.27.03 | Encrypt all biometric/source media in transit and at rest with tenant isolation, scoped signed URLs, short TTLs …     | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-key-custody.ts` (new), `human-video-key-custody.spec.ts` (new)                 | MOST OF THIS ITEM WAS ALREADY BUILT and is composed rather than duplicated: 182.C.16.03 holds the object namespaces with per-namespace signed-URL rules and TTLs, 182.C.16.07 signs over key+tenant+audience+expiry together and clamps any cache to what is left of the signature, and 182.C.19.10 encrypts embeddings with AES-256-GCM under an HKDF-derived per-tenant key with tenant, project, subject and purpose bound in as AAD. WHAT WAS NOT BUILT IS KEY ROTATION, and that is where the reasoning is. ROTATING A KEY THAT ENCRYPTED DATA DOES NOTHING UNLESS THE DATA MOVES: "we rotate quarterly" usually means a new key for new data and the old key kept so the old data still opens - THAT IS NOT A ROTATION, IT IS AN ACCUMULATION, and the old key is still live, still a target, and now nobody is watching it because the dashboard says the rotation completed. Envelope encryption is the standard fix, and the CONSEQUENCE is what matters here: a rotation is complete when THE LAST DEK STOPS BEING WRAPPED by the old KEK, which is a computed fact rather than a date on a calendar. assertRetired refuses to retire a KEK that still wraps something, and the message names which of the two outcomes actually happens - the data becomes unopenable, or the key is quietly kept alive, and it is always the second. A SCHEDULED AND A COMPROMISE ROTATION ARE DIFFERENT OPERATIONS: a scheduled one may re-wrap lazily because the old key is not believed to be in anybody else's hands, and a compromise one may not - A RECORD NOBODY TOUCHES IS EXACTLY THE ONE AN ATTACKER HAS TIME FOR - so plan refuses to treat a compromised key as a schedule. AND TENANT ISOLATION MAKES ROTATION PARTIAL: keys are derived per tenant, so "rotate the key" is a thousand rotations, and assertNotRoundedUp refuses a global "rotated" while any tenant remains, naming the worst-affected one - such a report is wrong about the only tenant it matters for. What must be encrypted is DERIVED from 182.C.27.01's biometric flag rather than listed again, so adding a data class cannot quietly add an unencrypted one; transit encryption applies to every class including the script, because it is what a person was made to appear to say.                                                                                                                                                                                                                                                                                                                                                                                                                                                   | 25 tests pass; eslint clean. custodyReport flags a live-KEK count above two as an accumulation rather than a rotation, which is the state the module exists to make visible.                                                                                                                                                                                                                                                                                                                                                                   | No provider spend. | None                                                                                                                                                                                                                                  |
| 182.C.27.04 | Prohibit provider access to permanent object URLs; issue single-purpose expiring references where remote pull is …    | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-provider-references.ts` (new), `human-video-provider-references.spec.ts` (new) | THE DANGEROUS URL IS NOT THE ONE THAT LASTS FOREVER - IT IS THE ONE THAT OUTLIVES ITS PURPOSE. "Permanent" is easy to ban and nobody ships it deliberately; what ships is a URL with a generous TTL because somebody did not know how long the provider would take, AND A GENEROUS TTL IS A PERMANENT URL MEASURED IN HOURS. So the bound is not a duration at all - it is the JOB: referenceFor derives the expiry from the job's own deadline, capped at six hours, and assertBoundToJob refuses a reference whose ttlDerivedFrom is anything but the deadline. The check is separate from the minting because the dangerous path is a reference built somewhere ELSE - a helper, a fixture, or a provider adapter that made its own URL. ONE REFERENCE, ONE ASSET, ONE JOB: a prefix is refused outright, because a provider holding a reference to the reference SET could enumerate a person's whole pack from one job's grant, and reuse across jobs is refused because such a reference has become a credential. The audience is a specific provider, so another provider presenting the same URL is rejected. AND THE PART THAT IS NOT A CONTROL AT ALL: once the provider has fetched the bytes, the reference expiring does NOTHING - they have the image. So this bounds what can be FETCHED and for how long, and it is not a retention control, which is exactly what a short TTL reads as to anybody who does not look closely. assertNoRetentionClaim refuses that claim, and the message names the true statement that means nothing: "references expire after six hours" is true while saying nothing about what they still hold. Retention is the provider class in 182.C.26.01 and the ZDR posture in 182.C.27.02. The fallback to a remote pull is stated as legitimate rather than as a failure, because refusing to work with a pull-only provider is a product decision and this is not the place to make it.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | 24 tests pass; eslint clean. One pins a canonicalization property rather than a character: the signing payload joins on NUL, and the test constructs two references differing only in where the field boundary falls, so a formatter normalising that byte away would be caught.                                                                                                                                                                                                                                                               | No provider spend. | None                                                                                                                                                                                                                                  |
| 182.C.27.05 | Add least-privilege service identities for ingest, render, evaluation, review, delivery …                             | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-service-identities.ts` (new), `human-video-service-identities.spec.ts` (new)   | THE ROLES ARE NOT A LADDER, AND MODELLING THEM AS ONE IS THE WHOLE MISTAKE. The intuitive picture grades the six from weakest (ingest) to strongest (deletion), and it is wrong in both directions. DELETION IS THE NARROWEST IDENTITY, NOT THE WIDEST: it needs the right to destroy and NOT the right to read, and a deletion identity that can read every namespace is an exfiltration identity with a reassuring name - the one nobody questions, because deleting sounds like the dangerous verb. REVIEW READS FACES AND MUST NEVER WRITE ONE (a reviewer with write on candidate can put a frame back after a refusal). DELIVERY WRITES A RELEASED ASSET AND MUST NEVER READ A SOURCE: it serves the public edge, so its reach into source media is zero, not small. So privilege is a SET and LEAST_PRIVILEGE is written as one; assertNoLadder refuses a level, because the moment a number exists "deletion needs level 6" follows and the read grant arrives with it. PROMOTION OUT OF QUARANTINE IS A SERVER-SIDE COPY, NOT A READ - the distinction that keeps quarantine, which holds every unscreened upload, readable by NO service identity at all. Modelling it as read-then-write is what would hand ingest every tenant's unscreened bytes for an operation where they never enter its process; 182.C.06.01 sniffs and digests on the way in and quarantine allows no signed URL, so nothing needs them back. THE GRANT THAT MATTERS IS THE ONE NOBODY ASKS FOR: every stage will ask for source read because a failing job looks like a permissions bug, but EMBEDDING READ is the one that counts - only evaluation needs it, and a render identity holding it can exfiltrate a biometric identifier without ever touching a face image, a copy that survives the deletion of everything it came from. A SHARED IDENTITY IS THE UNION OF ITS PERMISSIONS, which is what was actually granted: assertNotShared names that union rather than the sharing, because nobody would have approved it as a single request. AND AN AUDIT WITHOUT REFUSALS IS AN ACCESS LOG OF THINGS THAT WERE ALLOWED - authorize decides and logs in ONE function so a refusal cannot take a path that skips the sink, and it refuses an access with no PURPOSE, because the same identity reading the same object for a different reason is the actual event and a log of (actor, object) cannot show it. Break-glass is kept exceptional rather than banned: the failure is not that it exists, it is that an open-ended grant with no incident behind it has become the way the job gets done.                                                                                                                            | 35 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error present, which reports while the new files do not. privilegeReport asserts that exactly one of six namespaces is readable by no identity, which is the number a ladder drives to zero. A runtime failure caught the import NAMESPACE_POLICY where the module exports NAMESPACE_RULES - vitest does not typecheck, so the two report tests failed on undefined before tsc ever ran.                              | No provider spend. | The grants are a declared model, not an IAM policy applied to a running cluster; binding them to real service accounts is deployment work.                                                                                            |
| 182.C.27.06 | Threat-model SSRF, webhook spoofing, IDOR, signed-URL leak, queue tampering, prompt/reference injection …             | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-threat-model.ts` (new), `human-video-threat-model.spec.ts` (new)               | A THREAT MODEL IS ONLY WORTH WRITING IF IT DISTINGUISHES MITIGATED FROM ONLY-BOUNDED FROM NOT MITIGATED, and most do not - a table where every row says "addressed by input validation" has recorded a hope. So every entry names the SPECIFIC control by item number (asserted by a test), and MitigationKind separates prevented, bounded, detected-after and none, because writing a detection and a prevention both as "mitigated" is how somebody concludes the risk is closed. FIVE ARE GENUINELY PREVENTED and cite where: SSRF by 182.C.16.02 pulling assets to our own storage so there is no fetch to redirect; webhook spoofing by 182.C.11.06 refusing to advance a job on an unverified delivery INCLUDING when no secret is configured; IDOR by 182.C.25.02 putting the reach into the query and signing the cursor over the tenant; queue tampering by 182.C.15.04's transactional outbox with per-job head-of-line release; biometric exfiltration by 182.C.19.10's AAD binding plus 182.C.26.06's derived-artefact rule. THE THREE USEFUL ROWS ARE THE OTHERS. SIGNED-URL LEAK is BOUNDED rather than prevented, and mitigated in an unobvious place - not by making the URL hard to guess but by the TTL and 182.C.16.07's cache clamp, and a reader who thinks the defence is entropy will happily lengthen the TTL. PROMPT AND REFERENCE INJECTION is DETECTED-AFTER: there is no validator that separates a photo of a person from a photo of a person plus a payload, so 182.C.26.07's post-generation screening is the honest answer and it is a detection, not a prevention. MODEL OUTPUT POISONING is UNMITIGATED and the reason is structural: the output IS the product, so there is no reference to compare against - A SUBTLY WRONG VIDEO IS A VIDEO, and none of the gates would notice a steganographic watermark or a frame differing in a way nobody thought to measure. assertUnmitigatedAcknowledged refuses a release that has not acknowledged it - NOT a refusal to ship with a known gap, which is a decision somebody may make, but a refusal to ship without anybody having said they know. And malicious containers explicitly does not claim to stop a decoder zero-day.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | 19 tests pass; eslint clean. One asserts every non-none control cites a 182.C item number, and the summary states that a fully-prevented model is one to distrust - it usually means the hard threats were written down in a form that could be answered.                                                                                                                                                                                                                                                                                      | No provider spend. | Model output poisoning is recorded as unmitigated and must be acknowledged at release rather than resolved.                                                                                                                           |
| 182.C.27.07 | Add data residency and provider allow/deny controls for regulated or enterprise tenants …                             | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-residency.ts` (new), `human-video-residency.spec.ts` (new)                     | RESIDENCY IS NOT ONE FACT - IT IS THREE, AND THEY ARE ROUTINELY CONFLATED: where the bytes are STORED, where the inference is PROCESSED, and where the request TRANSITS. A provider can truthfully say "your data is stored in the EU" while running the GPUs on another continent, because storage and inference are different facilities with different economics, and a control that checks only storage passes exactly that route. assertClaimComplete refuses a partial claim and names which parts were answered, because A PARTIAL CLAIM IS WORSE THAN NO CLAIM - it reads as an answer. THE OBLIGATION FOLLOWS THE SUBJECT, NOT THE TENANT. The naive lookup reads the tenant setting, and the tenant is the company that bought the seat: a US studio filming an EU actor has an EU obligation the tenant setting is silent about. obligationFor takes both and returns the intersection, which can legitimately be EMPTY - a tenant restricted to the US with an EU subject has no compliant region, and that is a finding rather than a misconfiguration. AN EMPTY ALLOW-LIST IS NOT "ALLOW EVERYTHING": `allow.length > 0 && !allow.includes(p)` is the shape this bug always takes, skipping the guard exactly when nothing has been approved. Deny is evaluated first and unconditionally; an unconfigured allow-list denies and says so. THE THING THAT DEFEATS RESIDENCY IS THE FALLBACK - a compliant primary with an unconstrained fallback is a compliant route with a non-compliant one behind it, waiting for the primary to be busy - so planRoute runs the fallback set through the SAME predicate. FAILING CLOSED IS EASY; FAILING CLOSED LEGIBLY IS THE ITEM: 182.C.27.02 already measured that an empty route set with no reason reads as a product fault and somebody turns the control off to make it work, so the two empty cases are distinguished by their opposite remedies - "no lane is configured at all" wants configuration, "every lane was excluded" wants a compliant provider and NOT turning the requirement off. AND DERIVED DATA CROSSES BORDERS TOO: an embedding computed from an EU face and stored elsewhere is a transfer of biometric data, and "the video never left" is true and irrelevant because the identifier did. The requirement itself ships UNSET, the same call 182.C.26.10 made for jurisdictional posture, and assertRequirementSet refuses a route checked against nothing - the silent failure where nothing configured means nothing excluded means every route passed and the check reported green. Regions are ROUTING regions, not jurisdictions, and assertNoLegalityClaim refuses the sales answer that in-region routing makes a transfer lawful. | 33 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test drives the empty-intersection case (US tenant, EU subject) and asserts the note calls it a real answer; another asserts the two empty-route explanations differ, since collapsing them is the failure mode.                                                                                                                                                 | No provider spend. | Every tenant ships with the requirement UNSET; no jurisdiction posture is encoded, and provider residency claims are inputs this module validates for completeness rather than facts it can verify.                                   |
| 182.C.27.08 | Verify provider deletion/export capabilities and record unavoidable retention gaps in the job's consent and release … | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-retention-gaps.ts` (new), `human-video-retention-gaps.spec.ts` (new)           | 182.C.16.09 already records a lane with no deletion API as a disclosure obligation; this is the harder half - VERIFYING the capability rather than reading it, and putting what cannot be closed in front of the person rather than in a runbook. A CAPABILITY THAT WAS NEVER EXERCISED IS A CLAIM. The documentation says DELETE /v1/jobs/{id}; it does not say the endpoint exists on THIS account, that it returns anything but 501, or what it removes. So capability is four states rather than a boolean - verified, documented, refused, unknown - and assertNotAssumedFromDocs refuses the one substitution that matters, which is silent and reasonable-looking: the docs say the endpoint exists, so the plan includes the step, so the report says the copy was deleted, and nobody ever called it. AND "DELETED" IS NOT WHAT A SUCCESSFUL DELETE PROVES. A 204 proves the API accepted the call; a follow-up GET returning 404 proves the REFERENCE is gone; neither proves the bytes are, and from outside somebody else's infrastructure there is usually no way to prove that at all. The strongest honest outcome is accepted-and-unreferenceable, and describeOutcome does not contain the word "deleted" for ANY of the four states - a test asserts that across all of them. EXPORT IS NOT THE MIRROR OF DELETION, AND THE INVERSION IS THE USEFUL PART: export tells you WHAT THEY HAVE, which is the only thing that makes a gap measurable, so a lane that can delete but cannot export produces UNMEASURABLE rather than none. That sounds like the better posture and is the worse one for evidence. AN UNVERIFIABLE GAP IS NOT A ZERO GAP - the default when nothing is known is a recorded gap of unknown size, because absence of evidence arriving as evidence of absence is how a compliance answer ends up confidently wrong. AND A GAP FOUND AFTER CONSENT CANNOT BE BACKFILLED INTO IT: consent evidence records what the person agreed to AT THE TIME THEY AGREED, so a gap discovered afterwards goes into the RELEASE evidence and makes the consent stale - a re-consent, not a backfill - which is the same call 182.C.27.02 made about toggling ZDR. assertGapsInEvidence checks what the evidence CONTAINS rather than what the pipeline intended to write, because a gap recorded in a runbook is not a disclosure and the person it concerns never sees the runbook. The report ends by naming a zero gap count as the number to distrust: it usually means export was never verified, so nothing could be enumerated to find.                                                                                                                                                    | 28 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test loops every DeletionEvidence state and asserts the word "deleted" appears in none of the descriptions, so the fabrication cannot re-enter through a reworded message.                                                                                                                                                                                       | No provider spend. | No lane ships with deletion VERIFIED - verification means running the call against a real account, which is deployment work; every lane is therefore documented or unknown, and every gap is unmeasurable or unknown until it is run. |
| 182.C.27.09 | Add subject-access/export/deletion workflows that resolve originals, derivatives, embeddings, provider IDs …          | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-subject-requests.ts` (new), `human-video-subject-requests.spec.ts` (new)       | A SUBJECT IS NOT A TENANT, AND EVERY INDEX IN THIS SYSTEM IS BUILT THE OTHER WAY ROUND - so the lookup problem comes first and it is the one that is quietly failed. A tenant knows which packs belong to which subject because they enrolled them; we know it only through the AUTHORIZATION, so anything enrolled under a subject reference nobody linked is invisible to this search while being perfectly visible to whoever holds it. reachOf reports the reach the search actually had and assertReachStated refuses a response that presents a partial search as a complete one, which is the quiet failure of every subject-access implementation: it returns what it found and says "here is your data" rather than "here is what we could not look in". THE SIX CATEGORIES ANSWER DIFFERENTLY AND TWO OF THEM ARE NOT DATA. Originals and derivatives are ours - returned and deleted. EMBEDDINGS ARE OURS AND ARE NOT EXPORTABLE: a vector means nothing to the person and everything to whoever else gets the export, so including it hands a biometric identifier to whoever the email reaches; its existence is disclosed, its contents are not, and it is deleted. PROVIDER IDS ARE NOT DATA ABOUT THE PERSON AT ALL - they are pointers into somebody else's system, listed so the person can see WHERE their likeness went without pretending we can delete what is behind them. RELEASED ASSETS are out: 182.C.26.05 settled that a takedown is not a deletion, and assertReleasedNotClaimedDeleted refuses reporting one as deleted because that ends the conversation the subject would otherwise keep having. AUDIT EXCEPTIONS is the category that exists to be honest about a REFUSAL - a legal hold or a 182.C.27.01 minimum retention keeps something the person asked to remove, and assertExceptionsNamed refuses the tempting simplification of returning the shorter list, because it produces a cleaner answer and a false one and the person it misleads is the only one entitled to know. AND AN EXPORT IS A DISCLOSURE IN ITS OWN RIGHT: assembling everything about one person into a single file and sending it somewhere is the exact shape of the harm the rest of this phase prevents, so the requester is verified against the SUBJECT rather than against the tenant that holds the data.                                                                                                                                                                                                                                                                                                                                                                                          | 32 tests pass; eslint clean; targeted tsc over apps/oshun/bff/tsconfig.typecheck.json reports the deliberate \_\_control.ts error and nothing in the new files. One test asserts no category returns on access what it withholds from export, so a direction split cannot appear without somebody writing down why.                                                                                                                                                                                                                            | No provider spend. | The search reaches a person only through their authorizations; a subject reference nobody linked is not findable and the answer says so rather than omitting it.                                                                      |
| 182.C.27.10 | Run application, media-parser, object-storage, webhook, and authorization security testing before public release.     | Trust & Safety | 182.C.26   | done   | `apps/oshun/bff/src/generation/human-video-security-tests.ts` (new), `human-video-security-tests.spec.ts` (new)           | A SECURITY TEST THAT PASSES BECAUSE THE FEATURE IS NOT WIRED UP YET HAS NOT PASSED - the whole trap of testing a pre-release system. There is no deployed object store, so every object-storage test is green; no webhook endpoint receiving traffic, so every spoofing test is green; and a suite run in that state reports a clean bill of health for a system with NO ATTACK SURFACE, indistinguishable from one taken after the surface exists. So a check records what it EXERCISED, one that found nothing to attack reports `no-surface` rather than a pass, and assertNoSurfaceNotPass refuses the collapse. THE FIVE AREAS ARE NOT FIVE SUITES, THEY ARE FIVE KINDS OF EVIDENCE. Application and authorization are logic properties and ARE exercised here. MEDIA-PARSER is the one assumed to need a corpus and a deployment, and it does not: a parser takes untrusted bytes and malformed bytes can be generated, so the spec REALLY FUZZES sniffMediaType and findInlineMedia on this machine - 2000 deterministically generated byte sequences, every proper prefix of nine container signatures, a RIFF container with an unknown form, an ftyp polyglot, HTML (which is what a failed download actually is), a 5000-deep object, a 5000-element array and a cyclic one. The prefix test pins each signature's DISCRIMINATING LENGTH in both directions - nothing is recognised below it and the format is recognised at it - so an "optimisation" that shortens a check fails here rather than in an upload. MP4 is 8 rather than 12 because ftyp at offset 4 already decides it; RIFF is 12 because RIFF alone says only "some RIFF container". AUTHORIZATION drives decodeCursor with every single-character edit of the signature (43 positions, not one example, so a length-only or prefix-only comparison would be caught), a re-encoded body, a stripped signature, a foreign tenant, a different secret, and a caller whose scope was revoked. OBJECT-STORAGE and WEBHOOK are needs-deployment, which is a different word from passed and a different word from skipped, and the decoder-CVE check is needs-corpus because the sniffer reads magic bytes and NEVER DECODES - fuzzing it says nothing about libavcodec, and claiming otherwise would be the exact overclaim this item is about. assertReleaseCovered refuses calling an unmeasured area tested, which is not a refusal to ship with a gap: that is somebody's decision.                                                                                                                                                                                                                                                             | 29 tests pass and the fuzzing is real rather than described: the run drives 2000 generated byte sequences, 45+ signature prefixes and 43 signature bit-edits through the actual modules. eslint clean. Typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not - and that typecheck caught seven real errors the 29 green tests had not: the Caller fixtures were missing projectIds, so the authorization block had been exercising an incomplete caller. | No provider spend. | Object-storage and webhook testing have NO executed check and releaseReadiness names both; decoder fuzzing needs the 182.C.33 corpus. None of the three is a passing area.                                                            |

### 182.C.28 Provenance, Disclosure, Watermarking, and Release Proof

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner     | Dependency | Status | Changed Files                                                                                                                                                                                                                                                                       | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Cost                                   | Blocker                                                                                                                                                                                                                                 |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | --------- | ---------- | ------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.28.01 | Extend video C2PA/provenance to cover source references, approved script/audio digests, provider/model/version, seed … | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-provenance.ts` (new), `human-video-provenance.spec.ts` (new)                                                                                                                                                                             | Built on the repo's EXISTING real Ed25519 signer (audio-c2pa-signer.ts), which already binds the SHA-256 of exact bytes; this decides what goes IN one for a human video. THE MANIFEST BINDS THE BYTES AND EVERYTHING ELSE IS A CLAIM ABOUT THEM. The media hash is checkable by anybody holding the file; "the approved script had digest X" is a sentence we signed. Both survive the SAME signature check, so a verifier reporting "provenance verified" makes a reader believe all of it was confirmed when one part was - and exactly one of the ten assertions is BOUND. So each declares itself bound (recomputable from the asset) or ATTESTED (true only if the issuer is honest), verificationSummary reports the two separately, and assertAttestedNotReportedVerified refuses the verifier UI that a default implementation produces: ten rows, one tick. A SIGNATURE PROVES WHO SAID IT, NEVER THAT IT IS TRUE. A SEED IS NOT A REPRODUCTION RECIPE: recording one implies somebody could regenerate this, and 182.C.14 measured that providers do not pin model weights, so the same seed against the same model id can produce different frames next month. The seed is recorded as an identifier of the ATTEMPT, with the reason in the assertion data, and assertNoReproducibilityClaim refuses the promise. A PARTIAL EDIT CHAIN IS WORSE THAN NO EDIT CHAIN - a manifest listing [A, B] for an asset that went through [A, B, C] ASSERTS completeness while an absent list claims nothing - so assertChainContinuous refuses a step whose parent is not the previous step's output, refuses an empty chain (even a native one-pass generation is a step), and refuses a chain whose last output is not the bytes being signed, which is the case where something happened after the chain stopped recording. AND A MANIFEST TRAVELS WITH THE FILE, SO THE REVIEWER'S NAME WOULD TOO: human approval is the assertion everyone writes as approvedBy: "j.smith", and a C2PA manifest is public by design - it ships inside the asset - so that publishes an employee's identity with every video, permanently, to everyone the video reaches. It carries a decision reference and a PER-TENANT SALTED digest instead; without the salt a public manifest becomes a correlation handle for how many videos one reviewer signed off across the industry. The synthetic-media disclosure assertion is present unconditionally and says in its own data that it is not a tenant setting.                                                                                                                          | 31 tests pass, and the signing is real rather than described: the spec generates an Ed25519 key pair, drives the repo signer, and the round trip verifies, detects altered media bytes (media hash mismatch) and detects an EDITED ASSERTION (canonical hash mismatch) - which is the point of signing the claims rather than only the bytes. eslint reports 0 errors (3 non-null-assertion warnings, the file's prevailing style). Typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One correction during the work: the approver-digest separator was written as a literal NUL byte, re-introducing the grep-blindness fixed earlier the same day; it is now String.fromCharCode(0).                                                                                            | No provider spend.                     | Nine of the ten assertions are ATTESTED - a verifier can confirm who issued them and that nobody edited them, and cannot check them against anything.                                                                                   |
| 182.C.28.02 | Apply visible and/or imperceptible synthetic-media disclosure according to policy, geography, tenant …                 | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-disclosure.ts` (new), `human-video-disclosure.spec.ts` (new)                                                                                                                                                                             | THE IMPERCEPTIBLE MARK DOES NOT SURVIVE DELIVERY, AND THAT IS MEASURED RATHER THAN FEARED. video-watermark.ts recorded it on this repo's own encoder: a lossy re-encode ANNIHILATES the spatial spread-spectrum watermark - H.264 4:2:0 at CRF 4, already a very high quality setting, drops coverage from 1.0 to about 0.13 against the gate's 0.99 floor, because chroma subsampling, DCT quantization and the deblocking filter each erase part of a sub-pixel pattern. Only lossless holds it, at 25-30x the file size. SO THE USUAL ORDERING IS BACKWARDS: imperceptible marking is not the discreet default with visible as the crude fallback - it is available ONLY on the lossless master, and for every rendition anybody actually streams the only disclosure that survives the pipeline is the VISIBLE one. availableOn derives that from RENDITION_TRANSFORM rather than from a setting, and assertDisclosureSurvives refuses selecting it for a re-encoded rendition, because the failure is otherwise silent: the watermarker runs, the log says a mark was applied, and the shipped bytes carry a pattern measuring 0.13. A DISCLOSURE POLICY IS A UNION OF OBLIGATIONS, NEVER A PREFERENCE - geography, channel and the provider's licence each IMPOSE, a tenant can only ADD, and the settings page offering "disclosure: off" is offering to break somebody else's requirement; assertTenantCannotWeaken names which source imposed what. "THE PROVIDER ALREADY WATERMARKS IT" IS NOT A REASON TO SKIP OURS: their mark identifies their model, ours says this video depicts a real person whose likeness was used under an authorization, and one does not stand in for the other - while removing theirs is a licence breach rather than an optimisation. AND A MARK THE PLATFORM CROPS AWAY IS NOT A DISCLOSURE: social-vertical is a reframe, a corner badge placed for 16:9 is outside the 9:16 frame, and that is the rendition that reaches the most people - so placement is resolved per rendition and a corner on a reframe is refused. Finally, assertImperceptibilityNotClaimedMeasured keeps the two apart: COVERAGE is measured on the re-decoded delivered bytes; whether a viewer notices the pattern is a perceptual claim nothing here tested, and the two arrive under the same word.                                                                                                                                                                                                                                                                                                    | 31 tests pass; eslint clean (0 errors); typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not; both files confirmed free of literal control bytes. One test asserts the measured constants directly (1.0 lossless, ~0.13 lossy, below the 0.99 floor) so the argument cannot drift from the measurement it rests on, and another loops every rendition asserting safePlacementFor survives its own transform.                                                                                                                                                                                                                                                                                                                                                                          | No provider spend.                     | No jurisdiction is encoded: geography obligations are inputs, consistent with 182.C.26.10 shipping every posture UNSET. Imperceptible marking is possible on 1 of 4 renditions and that is a property of the codec, not a gap to close. |
| 182.C.28.03 | Verify provenance survives supported transcodes and expose a verifier result alongside each downloadable rendition.    | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-provenance-transport.ts` (new), `human-video-provenance-transport.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-provenance-transcode/` (new)                                                                        | A MANIFEST THAT SURVIVES A TRANSCODE REPORTS TAMPERING - measured, not argued. evidence/2026-09-11-provenance-transcode/ runs it on ffmpeg 6.1.1: the `comment` tag carrying 8000 characters survives a full re-encode AND a scale to 160x120, intact, and the transcoded file is of course not byte-identical to its source. So the manifest riding along still binds the OLD bytes and a verifier returns "media hash mismatch" - THE SAME ANSWER IT GIVES FOR A MALICIOUSLY EDITED FILE. Carrying provenance through a transcode does not preserve provenance, it MANUFACTURES A TAMPER SIGNAL, and it does so on the renditions most people download. So "survives" is the wrong goal and the split is on the one measured axis: a REMUX is byte-identical (same SHA-256, measured) so the manifest stays valid and is CARRIED; a TRANSCODE needs its own manifest over the new bytes whose edit assertion names the parent hash, and the inherited one is REPLACED. assertMismatchNotCalledTampering refuses the verifier report that accuses the holder: from the file alone a verifier cannot tell an edit from a re-encode, so the honest verdict names the pipeline failure - this was never reissued - while a mismatch after a REMUX stays reportable as tampering, because there the bytes should not have changed. Two transport findings shape the rest. A CUSTOM METADATA KEY IS SILENTLY DROPPED by the MP4 muxer: c2pa_manifest=... never reached the file and nothing errored, so the write succeeds, the read returns nothing, and the gap is visible only if somebody checks. And THE EMBEDDED COPY IS A CONVENIENCE, NOT THE AUTHORITY - one flag (-map_metadata -1) removes it and that flag is in most sanitising pipelines, a container change alters the key case (comment becomes COMMENT after mp4 to webm), and the recognised field is one anything else may overwrite. The sidecar decides. Finally the verifier surface: ONE BADGE ABOVE FOUR DOWNLOAD BUTTONS is a claim about whichever file the reader is about to click, so each rendition carries its own result and assertResultPerRendition refuses a single verdict standing for several.                                                                                                                                                                                                                                                                                                                                                                                                                                               | 30 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The measurements are REAL ffmpeg runs, not assumptions: evidence/2026-09-11-provenance-transcode/measure.sh reproduces all eight results and results.json is its raw output. The first spec block pins every constant, so the reasoning cannot drift from the run it rests on.                                                                                                                                                                                                                                                                                                                                                                                                                     | No provider spend (local ffmpeg only). | Measured on ffmpeg 6.1.1 and the MP4/WebM muxers; another encoder or a CDN that rewrites containers could differ, and the sidecar is the authority precisely because the embedded copy is one flag from gone.                           |
| 182.C.28.04 | Produce a canonical release proof binding exact final asset digest to consent, safety, speech, sync, identity …        | Oshun BFF | 182.C.21   | done   | `libs/contracts/src/common/proof-types.ts` (threshold_measurement added), `proof-types.test.ts` (extended), `apps/oshun/bff/src/generation/release-proof.test-fixtures.ts` (nine gates to sixteen), `human-video-release-proof.ts` (new), `human-video-release-proof.spec.ts` (new) | The item is explicit that this EXTENDS ProofRecordSchema and the nine gates rather than defining a second proof record, and the reason is worth stating: two records for one artifact means two answers to "was this released correctly", and the one a reader finds first is whichever code path happened to write it. So EXACTLY ONE PROOF TYPE WAS ADDED TO THE CONTRACT, and it is the one the existing fourteen cannot say. observed_outcome records THAT something was observed; it cannot say a value cleared its threshold by a hair rather than by an order of magnitude, and it cannot say the instrument could not resolve the difference at all. This whole phase turns on that distinction, so threshold_measurement carries measuredValue, unit, threshold, comparison AND NOISE FLOOR - and its rule lives in the SCHEMA rather than in a caller: a passed threshold measurement must satisfy its own comparison, and a value nearer the threshold than the noise floor CANNOT BE PASSED AT ALL, because a measurement inside its own noise floor is not a measurement that succeeded - it is the absence of one, and recording it as a pass is how an unmeasurable gate ships green. Three new schema tests pin that: a value on the wrong side of its threshold is refused, 44 ms against a 45 ms threshold with an 8 ms floor is refused, and a below-noise-floor status is refused however good the number looks. THE SEVEN ADDED GATES are the ones a human likeness needs on top of the nine (which cover everything a generated asset needs and nothing specific to a person being in it): consent-evidence, exact-speech, audiovisual-sync, identity-preservation, encoding-conformance, provider-lineage and final-bytes-binding. FOUR take threshold_measurement because each is a number against a threshold with an instrument floor; the other THREE REUSE consent, authenticated_signature and byte_bound_integrity, because a human-video copy of an existing proof type is the second-record problem in miniature. assertNotSubstituted refuses observed_outcome standing in - the substitution is tempting precisely because it is already in the fixtures and already passes. AND THE PROOF IS ABOUT ONE DIGEST: assertBoundToFinalBytes refuses a subject naming a job or a candidate, because a proof about "the job" is satisfied by whichever rendition happened to be measured. The fixture file was EXTENDED in place from nine gates to sixteen and its nine-deep nested ternary refactored to a lookup, since sixteen gates over eight payload shapes is one branch too many to read. | 23 new tests plus 16 in the extended contract spec (up from 11), all passing. The last spec block drives the REAL extended fixture through the REAL ProofRecordSchema and evaluateProofRequirements, so "extended, not duplicated" is checked rather than asserted, and it verifies every threshold measurement clears its own noise floor. All 53 existing fixture consumers (release-gate, provider-measurement, video-executor, image-executor) still pass against the 16-gate set. libs/contracts: 146 of 147 test files pass; the one failure is contract-prisma-migrations.test.ts, which fails identically with these changes stashed out (the installed Prisma CLI rejects `--to-schema`) - verified by stashing, re-running and restoring. eslint clean; typechecked with a deliberate \_\_control.ts error, which reports while the new and extended files do not. | No provider spend.                     | The gate table and the fixture are the shape of the record; wiring a real pipeline to emit these sixteen proofs per job is 182.C.28.05 onward.                                                                                          |
| 182.C.28.05 | Fail closed if required measurements, evaluator versions, consent evidence, provider lineage …                         | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-release-fail-closed.ts` (new), `human-video-release-fail-closed.spec.ts` (new)                                                                                                                                                           | "MISSING" IS FOUR DIFFERENT THINGS AND ONLY ONE OF THEM IS A BUG: never-attempted (a step was skipped), attempted-failed (the evaluator ran and crashed), not-measurable (it ran and the input could not support a measurement - a face too small, audio too short) and refused (a person or a policy declined). ALL FOUR FAIL CLOSED, so a caller needing only a yes/no can stop at the boolean - but the four REMEDIES differ, and a gate that collapses them to "missing" sends every one to the same wrong place. A retry on a NEVER-ATTEMPTED gate runs the same pipeline and skips it again; a retry on a NOT-MEASURABLE gate runs the same evaluator over the same unusable input and fails the same way, FOREVER; a refusal is not a fault and nothing technical resolves it. AND NOT-MEASURABLE IS THE ONE THAT ARGUES ITS WAY TO A PASS, because somebody can truthfully say nothing was wrong - nothing was RIGHT either, the measurement did not happen, and the reason it did not is a property of the input this release is about. 182.C.19.08 measured where that bites: 14% of the identity margin survives at a 36 px face, so an unmeasurable identity gate on a small face is not a neutral result, it is the specific case the floor exists for. A STALE EVALUATOR VERSION IS MISSING EVIDENCE, NOT OLD EVIDENCE: a measurement made by evaluator 1.2 against thresholds fitted for 1.4 is A NUMBER FROM A DIFFERENT INSTRUMENT, and the failure is silent - the value is present, in range, and looks nothing like a problem. Also refused: a measurement with no evaluator version at all. AND THE DEFAULT HAS TO BE REFUSAL RATHER THAN ABSENCE OF APPROVAL. The common shape is `if (anyGateFailed) block`, which passes when the gate list is EMPTY, when a gate name is misspelled, and when an evaluator returns undefined - three ways to release with no evidence at all. decideRelease starts from a refusal and walks ALL_GATES, so a gate nobody mentioned is an absence rather than a silence, and an entirely empty submission produces sixteen absences rather than a release. Evidence marked present that ALSO carries an absence reason is treated as absent: it has been filled in by something that did not want to block.                                                                                                                                                                                                                                                                                                                                                             | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test asserts the four reasons map to four DISTINCT actions, so collapsing two of them fails here; another drives the empty submission and asserts it produces one absence per gate rather than a pass.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | No provider spend.                     | None.                                                                                                                                                                                                                                   |
| 182.C.28.06 | Distinguish native one-pass generation, provider edit, corrective lip sync, audio replacement, manual edit …           | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-edit-history.ts` (new), `human-video-edit-history.spec.ts` (new)                                                                                                                                                                         | THE SIX DIFFER IN WHAT THEY CAN UNDO, NOT IN HOW THEY WERE MADE. A creator reading "provider edit" learns nothing they can act on; the question they have is whether they can go back, and the answer splits the six in a way the names do not. A PROVIDER EDIT IS NOT REVERSIBLE - the model produced new pixels from a prompt, the previous state exists only if we stored it, and nothing about the operation guarantees that - while a re-encode, a lip sync and a hand edit all revert by pointing at an artifact that still exists, and a native generation has nothing to go back TO. assertUndoOffered refuses showing the button before checking the previous artifact was retained, because THE BUTTON IS THE PROMISE and the failure would otherwise land on the creator at the moment they act on it. WHAT EACH KIND VOIDS IS NOT GUESSABLE FROM ITS NAME EITHER. A re-encode cannot change who is in the frame and neither can an AUDIO REPLACEMENT - which voids exact-speech and sync COMPLETELY while leaving identity untouched, the pair people get backwards because "I only changed the audio" sounds small. And a CORRECTIVE LIP SYNC does void identity, which is the one people skip: re-timing a mouth alters the face region, which is exactly what the identity measurement looks at. A hand edit voids everything, which is why it is used least. AND A RE-ENCODE THAT CHANGES NOTHING VISIBLE STILL CHANGES THE DIGEST - 182.C.28.03 measured why that matters, so assertReEncodesShown refuses hiding them: they look like noise and they are what explains why the release proof had to be reissued, and hidden, the reissue becomes an unexplained event. Finally THE LABEL SAYS WHAT CHANGED, NOT WHICH TOOL RAN: the creator is deciding whether to re-approve, so "corrective lip sync" becomes "Mouth movement re-timed to the audio" and assertLabelled refuses an internal kind reaching a person, offering the label they should have seen.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | 30 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test asserts every kind voids final-bytes-binding (no edit leaves the digest alone) and another asserts the exact set that voids identity, so a future edit to the table that gets the audio-replacement case wrong fails here. A first attempt asserted labels contain no hyphen, which was wrong - "re-timed" is a word - and now asserts the property meant: a label is never the kebab-case kind and always reads as a sentence.                                                                                                                                                                                                                                                           | No provider spend.                     | Whether a provider edit is revertible depends on retention this module cannot decide; it refuses to offer the undo rather than assuming either way.                                                                                     |
| 182.C.28.07 | Prevent a candidate or pre-transcode URL from being shared as a released artifact before the final gate passes.        | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-premature-sharing.ts` (new), `human-video-premature-sharing.spec.ts` (new)                                                                                                                                                               | 182.C.16.01 already closes the TECHNICAL path - only the released namespace answers a CDN request and candidate allows no public signed URL - so this is about the other path, which is the one that actually happens. THE LEAK IS NOT A BUG, IT IS A PERSON PASTING A LINK: a reviewer opens a candidate, copies the URL out of the address bar and sends it to the client because the client is waiting and the transcode is twenty minutes out. NOTHING MALFUNCTIONED - the candidate URL worked because it had to work for the reviewer, and it is now in an email thread that will outlive the job. So the controls that matter are not access controls. A REVIEW URL MUST BE UNUSABLE BY ITS RECIPIENT: bound to the REVIEWER rather than the tenant, because the person a link gets pasted to is usually inside the same tenant, and a link that works for its recipient is one nobody discovers was the wrong link. IT MUST BE VISIBLY NOT THE DELIVERABLE: the REVIEW COPY mark is not decoration, it is what makes the mistake visible to whoever would otherwise publish it, and it is THE ONLY CONTROL THAT STILL WORKS ON A FILE SOMEBODY ALREADY DOWNLOADED. AND IT MUST EXPIRE WITH THE REVIEW rather than with the job - 182.C.27.04 bounded a provider reference by the job, this is bounded by something shorter, and the extra window is exactly where a pasted URL keeps working. THE RELEASED URL IS NOT THE CANDIDATE URL WITH A FLAG FLIPPED: one URL whose behaviour changes when the gate passes means the link the reviewer ALREADY pasted starts working, for everybody, RETROACTIVELY, and the pre-release leak becomes a release nobody decided to make - so assertDistinctSurfaces refuses a released URL sharing an origin and path with the candidate one. APPROVED IS IN THE PRE-RELEASE LIST AND IT IS THE ONE THAT SURPRISES PEOPLE: an approved master passed review and is still not what a viewer is served, because the delivery rendition has not been made or measured. AND SHARING IS A DECISION, NOT A SIDE EFFECT OF PASSING A GATE: a gate passing means the asset MAY be released, and a pipeline that publishes on gate-pass has no state in which a finished, approved, correct video is DELIBERATELY not public yet - which is most of the reason a customer schedules anything.                                                                                                                                                                                                                                                                                              | 24 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test asserts a candidate may still be opened for any purpose EXCEPT release, since a reviewer has to be able to look at it; another asserts a different CDN origin on the same path counts as a distinct surface.                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | No provider spend.                     | Nothing stops a reviewer screen-recording a review copy; the visible mark is what makes that copy recognisable rather than preventable.                                                                                                 |
| 182.C.28.08 | Add provenance and disclosure regression tests against trim, concat, mux, transcode, caption burn-in, thumbnail …      | Oshun BFF | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-delivery-regression.ts` (new), `human-video-delivery-regression.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-provenance-regression/` (new)                                                                         | Seven operations, six of them measured on real ffmpeg runs rather than reasoned about (evidence/2026-09-11-provenance-regression/). ONE OF THE SIX PRESERVES PROVENANCE, AND IT IS THE ONE THAT CHANGES NOTHING: demuxing and remuxing the same streams reproduced a BYTE-IDENTICAL file, so the manifest stays valid. AND THERE ARE TWO WAYS IT GOES WRONG, NOT ONE. SURVIVES BUT INVALID (trim, transcode, caption burn-in): the manifest travels and binds bytes that no longer exist, so a verifier reports a hash mismatch - the same answer it gives for a maliciously edited file. SILENTLY LOST (concat, thumbnail): no manifest at all, which reads as "this asset never had provenance" rather than "provenance was destroyed here". The two need DIFFERENT REMEDIES - reissuing needs the old manifest, re-attaching needs the whole lineage - so a pipeline that only reissues after a transcode leaves concat and thumbnail with nothing. CONCAT IS THE WORSE OF THE TWO BECAUSE CONCAT IS WHERE PROVENANCE MATTERS MOST: it joins material from different sources, so the output of joining an authorized clip to an unauthorized one carries no claim about either - and `-map_metadata 0`, the obvious remedy, DOES NOT RECOVER IT (measured, not assumed). THE THUMBNAIL IS THE ONE EVERYBODY FORGETS and the only operation that also loses the DISCLOSURE: a still frame of a synthetic person is a picture of a face, and as a PNG it carries no manifest and no synthetic-media disclosure, so a thumbnail posted on its own is an UNDISCLOSED SYNTHETIC IMAGE OF A PERSON - and it is the artifact most likely to be posted on its own. CDN DELIVERY is the seventh row and is not an ffmpeg operation: nothing is re-encoded, and it is included because it is where the SIDECAR goes missing - the edge serves bytes and knows nothing about a manifest URL, so a verifier following the asset falls back to the embedded copy that 182.C.28.03 established is not the authority. Finally assertOperationMeasured refuses shipping an operation nothing has run, because the measured answers surprised in BOTH directions.                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | 26 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The first spec block READS the committed results.json and asserts the module table equals it row by row, so editing the table without re-running the measurement fails here - that is the regression test the item asks for, rather than a set of assertions about what ffmpeg probably does.                                                                                                                                                                                                                                                                                                                                                                                                      | No provider spend (local ffmpeg only). | Measured on ffmpeg 6.1.1 and the MP4 muxer; HLS/DASH packaging is deliberately NOT in the set and assertOperationMeasured refuses it rather than assuming it behaves like a transcode.                                                  |

### 182.C.29 Admin, Review, and Operations Console

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner | Dependency  | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Cost                              | Blocker                                                                                                                                                                   |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | ----- | ----------- | ----------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.29.01 | Add a provider/model dashboard for live capabilities, probes, health, queue, latency, failure classes, cost, quality … | Admin | 182.C.25    | done        | `apps/oshun/bff/src/generation/human-video-provider-dashboard.ts` (new), `human-video-provider-dashboard.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                              | ELEVEN NUMBERS ON ONE SCREEN ARE ELEVEN DIFFERENT AGES, spanning from 5 seconds (queue depth, live) to a month (retention, a provider claim). Capability comes from an hourly probe; cost is BILLED IN ARREARS and is an estimate from our own token counts until the invoice arrives; drift needs a baseline and a window of completed jobs. A dashboard with ONE "updated 2 minutes ago" in the corner is telling the truth about one panel and lying about the rest, and THE OPERATOR DECIDING WHETHER TO QUARANTINE A MODEL READS YESTERDAY'S COST AS IF IT WERE NOW - so freshness is per panel, staleness is judged against each panel's own tolerance (two minutes is fine for cost and very wrong for the queue), and a single global as-of is refused. A PROVIDER IS HEALTHY AND A MODEL IS NOT, OR THE REVERSE: the item names both and they are different subjects - an API can be up while one version returns unusable frames, and a provider outage makes EVERY model it hosts look broken at once. explainHealth resolves the both-bad case to the provider for exactly that reason, and a collapsed badge is refused because it destroys the distinction that decides whether to reroute or to quarantine. "ERROR RATE" IS NOT A FAILURE CLASS: five percent that is all timeouts is capacity, all content-policy is the prompt, all 401s is an expired key - THE SAME NUMBER AND THREE DIFFERENT PEOPLE TO WAKE AT 3AM - so the panel is a breakdown, every class names what it points at, and rate-limited is marked as OUR quota rather than their service while provider-5xx is named the only class where waiting is the answer. DRIFT AND A BAD DAY LOOK THE SAME FOR A WHILE: below a baseline plus 200 samples a quality drop is UNEXPLAINED rather than drift, because calling it drift quarantines a model that had one bad hour - a decision somebody then has to undo in front of a customer. AND THE SCREEN HAS TO SHOW WHAT IT CANNOT SEE: retention and cost are CLAIMS, and rendered beside nine measured numbers in the same typeface they read as measured.                                                                                                                                                           | 32 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test drives the same two-minute staleness through the queue and the cost panels and asserts opposite verdicts; another asserts every failure class names a destination, so adding one without deciding who it pages fails here.                                                                                                                                                                                                                                                                   | No provider spend.                | The rendered dashboard is 182.C.29.08 (browser work); this is the view model and the rules about what the screen may imply.                                               |
| 182.C.29.02 | Add a privacy-safe review queue with final video, approved audio/script, cast mapping, metric failures …               | Admin | 182.C.25    | done        | `apps/oshun/bff/src/generation/human-video-review-queue.ts` (new), `human-video-review-queue.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                          | A REVIEW QUEUE IS A SCREEN WHERE ONE PERSON SEES HUNDREDS OF FACES, and that is not solved by permissions - the reviewer is authorised for every item in it. It is solved by WHAT THE QUEUE SHOWS UNASKED, which is a different and larger number than what 182.C.29.04 decided a reviewer may REACH. So the list view carries seven fields and none of them is a picture: a thumbnail grid of a hundred people is the highest-exposure surface in the product and it exists to help somebody choose which item to open, a job a failure class, a duration and a worst-window timestamp do just as well. THE WORST WINDOW IS THE POINT, NOT A CONVENIENCE: a reviewer sent to a two-minute video to confirm a sync failure will scrub, find nothing and approve, so an item whose failing metric points at no moment is incomplete - THE VERSION THAT SHIPS WITHOUT ONE IS THE VERSION WHERE REVIEWERS APPROVE BY FATIGUE - and each window carries the measured value and the threshold beside the timecode, so the reviewer knows what they are looking for rather than only where. A CAST MAPPING IS A LIST OF PEOPLE'S NAMES unless it is built from pack labels: the reviewer needs to know that speaker 2 is the person in the second pack and does not need that person's legal name to decide whether the mouth matches the audio, and the queue is seen by everyone on the rota. AND THE ACTIONS ARE BOUNDED, WITH THE BOUND THAT MATTERS ON APPROVE: reject and escalate are safe because they STOP something, while approve is the only action that RELEASES A LIKENESS, so it is the only one refused on incomplete evidence - and reject and escalate stay available on the same item, because blocking all three would leave a reviewer with a queue they cannot clear, AND A QUEUE SOMEBODY CANNOT CLEAR GETS CLEARED SOME OTHER WAY. An action outside the three is refused outright, since a review queue whose actions grow becomes an admin console with a worklist attached.                                                                                                                                                                                                                                                        | 26 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test loops every face-bearing field asserting none is in LIST_FIELDS; another asserts that with the lineage missing, approve is refused while reject and escalate are both still allowed on the same item.                                                                                                                                                                                                                                                                                        | No provider spend.                | The queue surface itself is 182.C.29.08 (browser work); this is the shape of the data it may show and the bound on what it may do.                                        |
| 182.C.29.03 | Require reason codes, reviewer identity, separation of duties where needed, audit log …                                | Admin | 182.C.25    | done        | `apps/oshun/bff/src/generation/human-video-quality-waiver.ts` (new), `human-video-quality-waiver.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                      | THE FIRST QUESTION IS WHICH GATES CAN BE WAIVED AT ALL, AND MOST OF THEM CANNOT - a waiver system that can waive anything has REPLACED THE GATES WITH A CHECKBOX. The line is not severity: it is WHOSE INTEREST THE GATE PROTECTS. A quality gate protects the BUYER, who is in the room and can decide the video is good enough for them; consent, identity, safety, exact-speech, disclosure, rights, provenance and the rest protect somebody who is NOT, and nobody in the room can waive on their behalf. Four of sixteen are waivable and every excluded gate carries a stated reason naming who is absent, because "not waivable" invites a request for an exception and naming the person does not. A REASON CODE IS NOT A REASON: free text alone is unqueryable, so it becomes "client approved" on every waiver ever issued and nobody can answer "how often do we ship out-of-window sync"; a code alone says nothing about THIS video. Both are required, and assertReasonsDistinct refuses text that merely restates the code - which is exactly what a dropdown plus a required field produces by default. SEPARATION OF DUTIES BINDS WHERE THE WAIVER BENEFITS THE WAIVER, not everywhere: a blanket second-approver rule is the one that gets routed around at 2am, so it applies exactly when the issuer is the operator who ran the job or a reviewer who already approved the take. The second approver may not be the issuer and may not be anyone else who worked the job, or both signatures come from the same interest. AND THE EXPIRY IS ON THE WAIVER, NOT ON THE VIDEO - the video ships and keeps shipping; what expires is the PERMISSION, capped at seven days and bound to exactly one job, because a waiver attached to a tenant or a template quietly covers the next forty jobs, WHICH IS A POLICY CHANGE WEARING A WAIVER'S NAME. issueWaiver validates and logs in ONE function so an unaudited path does not exist, and the entry carries the measured value the gate rejected so the waiver is reviewable later; assertWaiverAudited refuses honouring one that is not in the log, since a waiver nobody can find is indistinguishable from a gate that never ran.                                              | 30 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not; no literal control bytes. One test loops every non-waivable gate asserting it carries a stated reason, so adding a gate without deciding this question fails here; another asserts a refused waiver leaves the audit log EMPTY, since validating after logging would record attempts as decisions.                                                                                                                                                                                                    | No provider spend.                | The console surface that issues a waiver is 182.C.29.08 (browser work); this is the policy it has to obey.                                                                |
| 182.C.29.04 | Prevent reviewers from downloading biometric originals or provider payloads unless their role and case explicitly …    | Admin | 182.C.25    | done        | `apps/oshun/bff/src/generation/human-video-reviewer-access.ts` (new), `human-video-reviewer-access.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                    | A REVIEWER DOES NOT NEED THE ORIGINAL TO DO THE REVIEW. The review question is "is this output acceptable" and THE OUTPUT IS THE TAKE; the reference photograph the person enrolled is an INPUT. That is convenient, because the reference is the most sensitive artifact in the system and the review queue is the highest-traffic surface that touches this data. So four of seven artifacts are available by default (take, approved audio, approved script, evaluation frame), the face EMBEDDING is never available to a reviewer at all - a vector tells a human nothing and tells whoever else receives it everything, the same call 182.C.27.09 made about a subject export - and the biometric original and provider payload are case-gated. ROLE AND CASE ARE BOTH REQUIRED AND NEITHER IS SUFFICIENT: role-only means every reviewer can pull any reference at any time and the audit then shows a role that was LEGITIMATELY HELD; case-only means a dispute unlocks the data for whoever picks up the ticket. The case must also be OPEN (a closed case is a standing grant) and about THIS subject (a case about somebody else is the broadest hole in any case-based system). Only TWO case kinds are about the reference itself - an identity dispute and an enrolment complaint - and a sync failure, a bad take, a quality escalation and a policy question are each a case where somebody will ask and the answer is no. AND VIEWING IS NOT DOWNLOADING: a streamed frame leaves when the tab closes, a download is a copy in somebody's Downloads folder outside every control this phase builds, AND IT IS THE ONLY ACCESS THAT SURVIVES A REVOCATION - so they are separate permissions and a qualifying case grants the narrower one. PROVIDER PAYLOADS ARE THE FORGOTTEN PATH: the request body sent to a provider contains the reference images, and an engineer debugging a failure needs the status and the error but reaches for the whole payload BECAUSE IT IS ONE OBJECT - which is how a reference ends up in a support ticket. redactPayload walks the structure at any depth and assertPayloadRedacted checks the OUTPUT, because the field list is what somebody edits when a provider adds one.                     | 27 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test asserts the three artifact sets partition ARTIFACT_KINDS exactly, so a new artifact cannot be added without deciding which it is; another asserts a refused access is logged on the same path as a granted one.                                                                                                                                                                                                                                                                              | No provider spend.                | Nothing prevents a reviewer photographing a viewed frame; the control is that a download - the copy that outlives a revocation - is a separate, refused permission.       |
| 182.C.29.05 | Add tools to quarantine a model/version, cancel queued work, reroute jobs, invalidate releases, process takedowns …    | Admin | 182.C.25    | done        | `apps/oshun/bff/src/generation/human-video-operator-actions.ts` (new), `human-video-operator-actions.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                  | THESE ARE NOT SIX BUTTONS - THEY ARE THREE THAT STOP THINGS AND THREE THAT CHANGE WHAT ALREADY HAPPENED, AND THE SECOND KIND IS WHERE THE DAMAGE IS. Quarantine, cancel and reroute act on unfinished work, so the worst case is wasted money; invalidate, takedown and re-run act on something a customer already has. INVALIDATING A RELEASE DOES NOT UNPUBLISH IT - 182.C.26.05 settled that stopping that is a takedown, a different and slower action, so an operator who invalidates expecting the video to come down HAS DONE NOTHING TO THE VIDEO, and the two sit one click apart in any console. A RE-RUN CAN ONLY LOSE: the old verdict passed, so the outcomes are "passed again", which tells you nothing you did not have, and "failed this time", which leaves an approved and DELIVERED video with a failing gate - it is the right tool when an evaluator was wrong and never the right tool for reassurance, which assertRerunHasCause refuses by motive. A TAKEDOWN IS THE ONLY ONE OF THE SIX THAT REACHES A VIEWER and it is the slowest, which is the ranking an operator needs during an incident and the opposite of how a console usually orders them. QUARANTINE IS THE ONE THAT LOOKS INSTANT AND IS NOT: it stops NEW dispatch, while jobs already sent are running on somebody else's infrastructure and nothing here can reach them - this phase measured that OpenRouter exposes no cancel endpoint - so preview returns what it stopped AND what is beyond reach, and assertQuarantineScopeStated refuses reporting containment that is partial by construction. REROUTE IS THE ACTION MOST ABLE TO PUT A JOB SOMEWHERE IT DOES NOT BELONG, because the lane is where consent class and residency live: it looks like a capacity decision and it is a consent decision. AND EVERY ACTION NEEDS A BLAST RADIUS BEFORE IT NEEDS A CONFIRMATION - "Are you sure?" is a reflex, not a safeguard; what stops a mistake is knowing whether it is 4 jobs or 40,000, and the action taken blind is the one taken at 3am during an incident. assertPreviewed also refuses a preview computed for a DIFFERENT action, which is the copy-pasted-dialog failure.                                                                    | 26 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test asserts exactly one of the six reaches a viewer, so adding an action without deciding that question fails here; another drives the quarantine preview with three dispatched jobs and asserts the note names why they are unreachable.                                                                                                                                                                                                                                                        | No provider spend.                | Jobs already dispatched to a quarantined version cannot be stopped - no provider cancel endpoint exists - and the preview reports that count rather than hiding it.       |
| 182.C.29.06 | Add cohort drill-down for language, accent, skin tone, lighting, framing, motion, speaker count, provider, model …     | Admin | 182.C.25    | done        | `apps/oshun/bff/src/generation/human-video-cohorts.ts` (new), `human-video-cohorts.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                                    | SKIN TONE CANNOT BE A MEASURED COHORT, AND THIS REPO ALREADY MEASURED WHY. human-video-appearance.ts recorded that a ONE-STOP EXPOSURE CHANGE moved a single subject 58.2 degrees of ITA against a 63.7 degree gap between two different people - about 91% of the between-person signal - so a skin-tone cohort computed from pixels is mostly a LIGHTING cohort, and a fairness dashboard built on it would ATTRIBUTE A LIGHTING EFFECT TO A DEMOGRAPHIC ONE, which is precisely the error the dashboard exists to catch. The handling is not to drop the dimension: skin tone is DECLARED-ONLY and never inferred, and LIGHTING stays its own observed dimension beside it, because with both an effect that tracks light and not declared tone is VISIBLE as a lighting effect - and only the cross-tab can show that. Accent is declared for a related reason: detecting it is classifying a person by how they speak, and a detector's error distribution would become a finding about people. A COHORT WITH TOO FEW SAMPLES IS NOISE WITH A LABEL: "accent Scottish, 40% failure" over five jobs is a number somebody will act on, so below 20 samples a cell reports too-few and carries NO RATE AT ALL - returning a rate plus a warning produces a number that gets copied into a slide without the warning, and rateOf throws rather than returning one. AND SOME DIMENSIONS ARE OBSERVED AND SOME ARE DECLARED: provider, model and surface are facts, language and accent are stated, tone must be. Every cell carries the source of each dimension in its key and assertProvenanceShown refuses displaying one without it, because a detected value read as a stated one turns a detector into a finding about people. FINALLY THE DIMENSIONS ARE NOT INDEPENDENT AND THE CROSS-TAB IS WHERE THE FINDING IS: the spec drives a case where all three marginals sit under 6% failure while one intersection - a specific accent on a specific provider - is at 70%. A page of ten bar charts shows ten passing dimensions and hides that cell, which is the reason this is a drill-down. worstIntersection also ignores cells below the floor, because letting a three-sample cell win the comparison is how a drill-down produces a ghost. | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The intersection block is a worked case rather than an assertion: three marginals each under 6% and an intersection at 70%, with a separate test adding a three-sample cell at 100% and asserting it does NOT win.                                                                                                                                                                                                                                                                                    | No provider spend.                | Skin tone and accent are available only where a subject stated them, so those cohorts are sparse by construction - which is the honest cost of not inferring them.        |
| 182.C.29.07 | Add audit-safe redaction/export for support and provider escalation without exposing unrelated tenant or subject data. | Admin | 182.C.25    | done        | `apps/oshun/bff/src/generation/human-video-escalation-export.ts` (new), `human-video-escalation-export.spec.ts` (new)                                                                                                                                                                                                                                                                                                                                                                                                                                | THE TWO AUDIENCES NEED DIFFERENT THINGS AND THE OVERLAP IS SMALLER THAN IT LOOKS. Support is inside our trust boundary and outside the tenant's: they may see which tenant and which job, and answering a support question has never needed the face. A provider is outside BOTH: they need the request shape, the parameters and the error, and THE TENANT IDENTIFIER IS THE FIELD EVERYBODY ASSUMES IS HARMLESS - a provider who can correlate failures to named customers HAS BEEN HANDED A CUSTOMER LIST. That is the single field the two policies differ on; neither audience receives the likeness, neither needs to know WHO, and internal support notes never reach a provider because notes are written for colleagues and say things nobody would put in a message to a vendor. A REDACTED EXPORT MUST SAY WHAT IT REMOVED: silently dropping a field makes the export smaller and makes it a lie, because AN EXPORT THAT QUIETLY DROPS A FIELD AND AN EXPORT OF A JOB THAT NEVER HAD ONE LOOK IDENTICAL. Every removal leaves a typed marker, so a provider can see that a reference existed and cannot see it - which is also what lets them ask for a scoped answer instead of the whole payload - and assertRedactionsDeclared checks the declared list against what the bundle actually CONTAINED. AND THE LEAK IS NOT IN THE FIELDS YOU NAMED: a support note saying "Maria's video keeps failing", an error string echoing back a URL, a filename. A redactor that walks a field list and stops there exports all of it, so assertNoUnstructuredLeak scans the SERIALIZED export for identifiers it already knows about - and ignores an empty forbidden value, which would otherwise match every string. FINALLY ONE CASE IS ONE CASE: "all failures on this model version" is the natural query and the way a single escalation becomes forty tenants' data in one file, so a bundle spanning more than one tenant or subject is refused and the remedy is several exports rather than a wider one.                                                                                                                                                                                                                                 | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. Two tests drive real leak shapes - a subject name inside a support note and a URL echoed back in a provider error - and both are caught by scanning the serialized output rather than the schema.                                                                                                                                                                                                                                                                                                     | No provider spend.                | The unstructured-leak scan catches identifiers it is GIVEN; an identifier nobody passes in is not detected, which is why the structured redaction is the primary control. |
| 182.C.29.08 | Follow the `frontend-skill` design convention (same caveat as `182.C.23.13`) and add thorough Playwright …             | Admin | 182.C.23.13 | implemented | `apps/oshun/admin/src/app/isis/human-video/page.tsx` (new), `src/components/HumanVideoReviewQueuePanel.tsx` (new), `src/lib/human-video-review-port.ts` (new), `e2e/human-video-review-queue.spec.ts` (new), three unit specs, plus `GET /reviews` in the BFF (`human-video-api-surface.ts`, `human-video-review-handlers.ts`), plus the correction 2026-09-12: `apps/oshun/admin/src/lib/human-video-review-port.ts`, `human-video-review-port.spec.ts` (new), `src/components/HumanVideoReviewQueuePanel.tsx`, `src/app/isis/human-video/page.tsx` | §29 BUILT THE QUEUE AND THE ACCESS RULE AND LEFT NEITHER REACHABLE: an operator could record a decision and could not see what was waiting for one. The listing carries the access decision PER ARTIFACT, taken from the caller's scopes — a role a client could name is a role a client could claim — and the reference is refused with its sentence unless the role AND an open case both require it. The URL never travels beside a refusal, and the panel renders no control for an artifact it was not given. The operator actions ship REFUSED because the item asks for this coverage first. THE SCREEN STATED A ROLE THE SERVICE DID NOT APPLY, corrected 2026-09-12. `GET /reviews` derives the reviewer role FROM THE CALLER'S SCOPES and filters the rows by it; the admin derived its OWN from the same session and displayed that - AND THE TWO READ DISJOINT SCOPE VOCABULARIES, `admin:*` on the app side and `hv:reviews:*` on the service. An operator holding `admin:trust-safety` was told "Read as trust-and-safety" while the service, finding no `hv:reviews:*` scope, applied `reviewer` and returned only what a plain reviewer may hold - and the conclusion somebody draws from that is THAT THERE IS NOTHING SENSITIVE IN THE QUEUE. On a surface whose whole point is 182.C.29.04's "role AND open case, neither sufficient", that is the wrong thing to be confident about. The panel now shows the role THE SERVICE RETURNED, says so when the service did not name one, and states the disagreement explicitly when the two differ - because an admin grant not reaching the human-video queue is the thing an operator needs to know. THE ADMIN ALSO SENT `role` AND `caseId` AS QUERY PARAMETERS AND THE SERVICE READS NEITHER: a role in the URL reads as a request the service honoured, a `caseId` reads as a queue narrowed to that case, and both were ignored. A client-supplied role would be a client choosing what it may hold, so none is sent now.                                                                                                                                                                                                                                                          | 16 unit tests (admin) + 29 review-handler tests + 5 Playwright tests against a real admin server, including axe and the forbidden-workspace redirect. The browser caught a contrast failure at 2.47:1 that no unit test could see; 5306 tests pass across the 254 BFF human-video suites. 6 new port tests (no query parameters at all, the token IS the request, the role is taken from the ANSWER, a missing role is null rather than assumed, and an absent `reviews` array is refused rather than shown as an empty queue). Admin e2e re-run against a real stack: 5 passed including the axe scan over the new lines. Typechecked with a live control error; eslint clean. | No provider spend; no model call. | None for this item. The queue answers 503 until a review store is bound, which the surface renders as a failed read rather than an empty queue.                           |

### 182.C.30 Observability, SLOs, and Cost Accounting

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner | Dependency | Status | Changed Files                                                                                                         | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Cost               | Blocker                                                                                                                                                                                           |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | ------ | --------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.30.01 | Trace upload, preprocessing, TTS, planning, routing, provider submission, callback/poll, download, evaluation, retry … | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-tracing.ts` (new), `human-video-tracing.spec.ts` (new)                     | THE CORRELATION ID THAT MATTERS IS NOT THE TRACE ID. A trace follows a REQUEST and a human-video job does not fit in one: the upload happens on Monday, the render on Tuesday, a correction on Thursday after a reviewer looks at it. So the JOB is the correlation key and a trace id belongs to one ATTEMPT inside it - conflating them produces a view where A JOB TOOK 40 SECONDS WHILE THE CUSTOMER WAITED TWO DAYS. THE SPAN THAT BREAKS THE TRACE IS THE PROVIDER'S: between submission and callback the work is on somebody else's infrastructure and we cannot instrument it, so that span has a start, an end AND NO INSIDE. Drawing it as a solid bar - or worse, interpolating sub-spans from a provider progress field - implies we know where the time went, and assertOpaqueNotDetailed refuses children under it because a progress field is THEIR SELF-REPORT RATHER THAN A MEASUREMENT and rendering it as spans makes a guess look like telemetry. Human review is opaque for a different reason: a person takes as long as they take, and the elapsed time is a queue property rather than a system one. elapsed() returns the duration together with whether it is attributable, so the two never get summed into one latency figure. A CALLBACK ARRIVES WITH NO CONTEXT AND HAS TO FIND ITS OWN TRACE: this phase measured that our headers are not propagated, so correlation rides in something the provider ROUND-TRIPS - a job id in an echoed field or in a reference URL - and where nothing does, the result is an ORPHANED CALLBACK, a named state with an owner rather than a lost span nobody counts. AND FOUR OF THE FOURTEEN STAGES DO NOT ALWAYS HAPPEN: tts is absent when the caller supplied audio, retry and correction are absent on a job that went well, review is absent when nothing needed a human. A trace missing ENCODING is broken. If every stage were mandatory then every trace would look incomplete, AND A GAP NOBODY CAN INTERPRET IS A GAP NOBODY LOOKS AT.                                                                                                                                                                   | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test drives a complete happy-path trace and asserts the four conditional absences are NOT faults while a missing encoding stage is; another asserts an unfinished span reports no duration rather than zero.                                                                                                                                                                   | No provider spend. | Two of fourteen stages are time we cannot attribute, and that is a property of the provider and of human review rather than a gap to close.                                                       |
| 182.C.30.02 | Measure success and latency per stage plus end-to-end p50, p95, and p99 by …                                           | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-latency-metrics.ts` (new), `human-video-latency-metrics.spec.ts` (new)     | A PERCENTILE OVER MIXED DURATIONS IS NOT A NUMBER ABOUT ANYTHING. A five-second clip and a ninety-second clip do not belong in one distribution: the p95 of a bucket containing both is roughly "how many long clips ran this week", which moves when the TRAFFIC MIX moves and not when the system changes - so it looks like a REGRESSION on the day a customer starts sending longer videos. Duration is therefore a BUCKET DIMENSION rather than a label on one series, and assertBucketed refuses a percentile spanning buckets. P99 NEEDS MORE SAMPLES THAN A BUCKET USUALLY HAS: over 40 observations it is THE MAXIMUM WITH EXTRA STEPS, moving the width of the tail every time one slow job lands, so each percentile declares its own floor (20 / 200 / 1000) and below it the answer is `insufficient` rather than a number - the same discipline as 182.C.29.06's cell floor. The estimator is NEAREST-RANK rather than interpolating, because with a few hundred samples an interpolated value invents a latency no job had, and a test asserts the returned value is one of the observations. END TO END IS NOT THE SUM OF THE STAGES: stages overlap, jobs wait in queues between them, and two of the fourteen are time we cannot attribute at all (182.C.30.01) - so a sum is SMALLER THAN THE TRUTH AND LOOKS AUTHORITATIVE. It is measured from request to release, the difference is reported as queue time and overlap, and the unattributable portion is reported separately: measuring a provider from outside is legitimate, calling the result OUR latency is not, and THE COMBINED FIGURE IS ONE NOBODY OWNS. AND SUCCESS RATE HAS THREE DENOMINATORS, NOT ONE - attempts, jobs, and jobs a customer would call successful. A system with retries makes them diverge: the spec drives a case where 80% of attempts succeed and 99% of jobs do, and a single "success rate" is whichever the author happened to compute.                                                                                                                                                                                                                                    | 26 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The percentile block uses 1000 linearly rising samples so the expected rank is checkable by hand, and asserts the result is a member of the input, that the three percentiles are ordered, and that an unsorted input gives the same answer.                                                                                                                                       | No provider spend. | A p99 needs 1000 samples in ONE duration bucket, so most buckets will honestly have no p99 until volume grows - which is the point rather than a limitation.                                      |
| 182.C.30.03 | Measure first-pass and eventual pass rate for sync, exact speech, identity, visual integrity, safety, and release.     | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-pass-rates.ts` (new), `human-video-pass-rates.spec.ts` (new)               | THE GAP BETWEEN FIRST-PASS AND EVENTUAL IS THE PRODUCT, NOT A ROUNDING ERROR. First-pass is how good the GENERATION is; eventual is how good the PIPELINE is once retries and corrections have run; and the difference is exactly the work done on the customer's behalf - the number that says whether this is a model wrapper or a system. Reporting only the eventual rate HIDES THE COST OF EVERY JOB; reporting only the first-pass rate MAKES A WORKING PRODUCT LOOK BROKEN - so ratesFor returns both plus the gap rather than leaving a caller to subtract, because a caller who has to compute it is a caller who will display one of them. SAFETY HAS NO EVENTUAL RATE, AND SAYING SO IS THE POINT: a sync failure is retried and often fixed, but A SAFETY FAILURE IS NOT RETRIED UNTIL IT PASSES - that is not a correction, it is RESAMPLING UNTIL THE SCREENING GETS IT WRONG. So safety's eventual rate is not high, it is NULL, and assertNoEventualForTerminal refuses computing one, because a dashboard column quietly filled with 99.9% for safety is describing a retry loop somebody should be alarmed by. Its FIRST-pass rate is still real and still reported. AND RELEASE IS NOT A SIXTH GATE OF THE SAME KIND: the others measure the ARTIFACT and release measures the DECISION, so its first-pass rate is bounded above by the strictest measurement feeding it - a job cannot release on the first pass if a measurement failed, which makes a higher number ARITHMETICALLY IMPOSSIBLE rather than a policy question, and assertReleaseNotAboveInputs catches it as such. FINALLY A RATE WITH NO WINDOW IS A LIFETIME AVERAGE: pass rates are quoted to decide whether something got WORSE, and a figure computed over all time CANNOT get worse because a bad week is diluted by every good week before it - so an empty or all-time window is refused and every measured result carries its own.                                                                                                                                                                                                                                                       | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test asserts safety returns a null eventual rate AND a real first-pass rate, so the gate is not silently dropped from the view; another drives the release ceiling at exact equality to check the boundary is inclusive.                                                                                                                                                       | No provider spend. | Safety has no eventual rate by design, and that column stays empty rather than being filled.                                                                                                      |
| 182.C.30.04 | Measure retry amplification, corrected-share, manual-review rate, false accept/reject estimates, cancellation …        | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-failure-metrics.ts` (new), `human-video-failure-metrics.spec.ts` (new)     | RETRY AMPLIFICATION IS A COST METRIC WEARING A RELIABILITY METRIC'S NAME. Attempts per job looks like a health number and is a SPEND number - every extra attempt is a provider charge for an artifact nobody will ever see - and the unweighted ratio CAN BE FLAT WHILE SPEND DOUBLES, because a retry of a 90-second clip costs eighteen times a retry of a five-second one. So amplification returns both the ratio and the COST ratio plus the money that bought nothing, and the spec drives two windows with identical attempt ratios and very different spend. FALSE ACCEPT AND FALSE REJECT CANNOT BE MEASURED, ONLY ESTIMATED, AND FROM DIFFERENT PLACES. A false REJECT is visible: the job was blocked, somebody looked, and the look IS the ground truth. A false ACCEPT is invisible by construction - it shipped and nobody complained - AND "NOBODY COMPLAINED" IS NOT EVIDENCE. Its only honest source is a sampled audit of RELEASED work, which costs money and which nobody schedules, so a false-accept figure with no audit behind it is AN ASSUMPTION WITH A DECIMAL POINT, and estimateFalseRates returns `unmeasured` rather than producing one. AN ORPHAN JOB AND A STALE CALLBACK ARE THE SAME EVENT FROM TWO ENDS - we submitted and heard nothing, they answered and we could not place it - so counting them separately is right and counting only one is how a reconciliation gap stays invisible: A SYSTEM THAT TRACKS ONLY ORPHANS CONCLUDES THE PROVIDER NEVER REPLIED. reconciliationVerdict reads the PAIR, which distinguishes "requests are not coming back" from "correlation is failing" from "these are probably the same jobs". AND MANUAL-REVIEW RATE IS A DEMAND FIGURE, NOT A QUALITY ONE: it rises when quality falls AND when a threshold tightens AND when a new cohort arrives, and read as quality it produces the wrong action - somebody loosens a threshold to bring the queue down, WHICH IS THE REVIEWERS' PROBLEM SOLVED BY REMOVING THE REVIEW. Cancellations are split three ways for the same reason.                                                                                                                       | 27 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The amplification block is a worked case: two sets of jobs with the SAME attempts-per-job and a cost ratio differing by more than half again, so the claim that the two diverge is checked rather than asserted.                                                                                                                                                                   | No provider spend. | False accept ships UNMEASURED until somebody funds a sampled audit of released work; the module refuses to substitute a number for it.                                                            |
| 182.C.30.05 | Attribute actual provider, TTS, evaluation compute, correction, transcode, storage, egress, refund …                   | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-cost-attribution.ts` (new), `human-video-cost-attribution.spec.ts` (new)   | THE ONLY COST WITH A RECEIPT IS THE PROVIDER'S, AND IT ARRIVES LATE AND AGGREGATED: an invoice says "$4,812 for October", not "$0.31 for job-7". So a per-job provider cost is ALWAYS AN ALLOCATION OF A BILL rather than a measurement, three of eight components are shares of a bill, and assertBasisHonest refuses presenting one as metered. The allocation has to RECONCILE, because a rounding policy that loses a tenth of a cent per job loses real money across a million of them - the spec drives exactly that case and the residual comes out at about a dollar - and assertResidualShown refuses hiding it, since A RESIDUAL QUIETLY ABSORBED INTO THE LARGEST TENANT IS A CROSS-SUBSIDY NOBODY CHOSE. A JOB'S COST NEVER STOPS CHANGING: storage accrues while the asset exists and EGRESS IS PER DOWNLOAD, which happens after the job finishes and can recur for years, so a popular video costs more every month. A job cost is therefore always AS OF a moment, and a report showing a final number has picked a cutoff nobody stated - which is how two reports of the same job disagree with NOBODY ABLE TO SAY WHICH IS RIGHT. A WASTED CANDIDATE IS A COST WITH NO PRODUCT AND IT IS THE ONE ALWAYS MISSING: candidates that lost selection cost exactly what the winner cost, so attributing only the winner makes generation look three times cheaper than it is - and that is precisely the number a pricing decision needs. AND A REFUND IS NOT A NEGATIVE COST: refunding a customer DOES NOT UN-SPEND THE PROVIDER CHARGE, and netting them makes a job that cost us money and earned nothing look FREE, which is the single most misleading thing this accounting can do. assertRefundNotNetted catches a reported total equal to spend-minus-refund, and deliberately does not fire when the refund was zero.                                                                                                                                                                                                                                                                                                                                          | 24 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The reconciliation test allocates a real million-job invoice at a rate that under-attributes by a tenth of a cent and asserts the residual, rather than asserting that rounding matters.                                                                                                                                                                                           | No provider spend. | Provider and TTS cost per job can never be measured, only allocated from an invoice that arrives later and covers a month.                                                                        |
| 182.C.30.06 | Alert on quality drift, missing audio, ASR mismatch, identity regression, provider schema change, spend anomaly …      | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-alerts.ts` (new), `human-video-alerts.spec.ts` (new)                       | TEN ALERTS ARE TOO MANY FOR ONE PAGER AND NOT ENOUGH FOR TEN. The failure mode of a list like this is that every condition gets the same treatment, so either everything pages and people mute the channel or nothing pages and the one that mattered arrives in a weekly digest. Each alert therefore declares what it INTERRUPTS - a page, a ticket or a line in a report - and THE ASSIGNMENT IS NOT BY SEVERITY, IT IS BY WHETHER ACTING NOW CHANGES THE OUTCOME. ONLY ONE OF THE TEN IS AN EMERGENCY AND IT IS NOT THE QUALITY ONE: an unreleased public URL exposure is the only condition where every minute of delay is more harm - a likeness is reachable that should not be, and nothing about waiting improves it. QUALITY DRIFT AT 3AM IS THE SAME PROBLEM AT 9AM WITH A RESTED PERSON LOOKING AT IT, and the jobs it affects are already finished; a pager that treats those two alike trains people to ignore it, and then the one that matters is ignored too. AN ALERT WITHOUT AN ACTION IS A NOTIFICATION: "spend anomaly" tells somebody a number moved, when what they need is WHICH LEVER EXISTS - reroute, quarantine, raise a limit, wait - and an alert that cannot name one is A SUBSCRIPTION TO ANXIETY. Every entry carries its lever, and the queue-age alert says to quote the OLDEST job rather than the average, because the average is the number that hides the customer who has waited longest. AND THE ALERT THAT FIRES ON ITS OWN ABSENCE IS THE HARDEST ONE: an evaluator outage cannot be detected by watching evaluator failures, BECAUSE AN EVALUATOR THAT IS DOWN PRODUCES NONE. It is the absence of SUCCESSES, and it is the alert most likely to be written as a failure-rate threshold where IT WILL NEVER FIRE - a rate that stays at zero looks exactly like everything being fine. Finally every non-event threshold needs a window and a number of consecutive windows, since an instantaneous one fires on one slow minute and becomes noise before anybody has looked at it once.                                                                                                                                                  | 24 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test loops every condition asserting it is accepted at its OWN declared level, so a future entry cannot be added without deciding what it interrupts; another asserts the event-shaped conditions are exempt from the window rule rather than special-cased at the call site.                                                                                                  | No provider spend. | None.                                                                                                                                                                                             |
| 182.C.30.07 | Keep metric labels bounded and logs redacted; use digests and opaque IDs instead of scripts, names, signed URLs …      | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-telemetry-hygiene.ts` (new), `human-video-telemetry-hygiene.spec.ts` (new) | AN UNBOUNDED LABEL IS AN OUTAGE, AND IT IS ALSO A DISCLOSURE - and the second is the one that matters. The operational half is familiar: a label with one value per job produces a time series per job, and a metrics backend with a million series stops being a metrics backend. The half people miss is that THE SAME LABELS ARE THE ONES CARRYING IDENTIFIERS: a subject_ref label is unbounded AND is a list of every person in the system, sorted, queryable and retained for however long the metrics retention is. THE MECHANISM IS AN ALLOW-LIST RATHER THAN A DENY-LIST, because a deny-list is a list somebody has to remember to extend and THE NEXT LABEL IS ALWAYS THE ONE NOBODY THOUGHT OF; ten labels are allowed, each declares a cardinality ceiling with a reason, and locale is called out as the one most likely to be let through unbounded because it "looks small". tenant_tier is offered in place of tenant_id: the tier answers most of what the tenant label is reached for and does not enumerate customers. A DIGEST IS NOT AN IDENTIFIER-FREE VALUE - it is bounded, stable and safe to label with, AND it is a perfect join key: two jobs sharing a script digest used the same script, which is usually what you want to know and occasionally tells you two tenants are running the same campaign. That is written down rather than pretended away. AND A SIGNED URL IN A LOG IS A CREDENTIAL IN A LOG: 182.C.27.04 bounded a provider reference to six hours, a log line lives for thirty days in a store a hundred people can read, and THE ERROR PATH - where somebody logs the whole request to debug it - is exactly where this happens. Object keys under a likeness-bearing namespace are refused for a related reason: the key is not the bytes and it is the thing somebody needs to fetch them, so a log search becomes an index of where the faces are. Script text is refused because a truncated preview is still the words a real person will be made to say. Redaction works on the RENDERED LINE rather than on fields, because the line is what ships and the interesting leak is always in a message somebody formatted by hand. | 26 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test calls the signed-URL check twice on the same input and asserts the same answer, because the detectors use global regexes and a carried lastIndex would make the second call pass silently.                                                                                                                                                                                | No provider spend. | The name check catches names it is GIVEN; the structural rules - allow-listed labels, signed URL and object-key patterns - are what work without a list.                                          |
| 182.C.30.08 | Define SLOs/error budgets and automated route quarantine or feature rollback when customer-impact thresholds are …     | SRE   | 182.C.15   | done   | `apps/oshun/bff/src/generation/human-video-slos.ts` (new), `human-video-slos.spec.ts` (new)                           | AN SLO IS A PROMISE AND AN ERROR BUDGET IS PERMISSION TO BREAK IT - that is the part that gets lost. A 99.5% objective does not mean "aim for perfect": it means half a percent MAY fail, that failure is budgeted, and spending the budget is not an incident. A team that treats every failure as one HAS AN SLO OF 100% WITH EXTRA ARITHMETIC, and the first real consequence is that nobody ships anything - so assertHasBudget refuses an objective of 1.0 and even the strictest SLO here (no-unmeasured-release) is 0.9999 rather than 1, with the reason written down: a target with no budget makes the alert the thing people route around. THE BUDGET IS THE CONTROL SURFACE, NOT THE OBJECTIVE, and the interesting state is the BURN RATE rather than the current percentage: 99.4% against a 99.5% objective is fine when the budget went slowly over a month and is on fire when it went in an hour, and THE PERCENTAGE CANNOT TELL THOSE APART. The spec drives exactly that - the same consumed fraction at two elapsed points, with burn rates differing by more than 300x. AUTOMATED QUARANTINE HAS TO BE ABLE TO BE WRONG: a rollback that fires on a burn spike will sometimes fire on one customer sending a hundred impossible jobs, AND QUARANTINING THE MODEL THEN MAKES THINGS WORSE FOR EVERYONE ELSE - so a burn a single tenant explains triggers no global action at all, every automatic action must be REVERSIBLE (an automation that fires on a signal will eventually fire on the wrong signal, and one that cannot be undone needs a human who can be asked), and it announces itself. AND A ROUTE QUARANTINE AND A FEATURE ROLLBACK ARE DIFFERENT DECISIONS: quarantining moves traffic and keeps the product working, while a rollback TAKES SOMETHING AWAY FROM EVERY CUSTOMER INCLUDING THE ONES IT WAS WORKING FOR - so it needs a higher burn AND a burn not confined to a route, because a system that reaches for whichever is nearest will reach for the big one.                                                                                                                                                                         | 28 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. A failing test caught a real floating-point bug: 10000 \* (1 - 0.99) is 100.00000000000009, so exactly 100 failures against a 99% objective computed to 0.9999999999999 and a bare `>= 1` reported a FULLY SPENT BUDGET AS NOT SPENT - the one state that has to reach somebody. Fixed with a 1e-9 tolerance, and the spec now pins both the exact-budget and the one-short cases. | No provider spend. | Burn-rate thresholds (6x quarantine, 14x rollback) are chosen operating points rather than fitted ones; they are the conventional multi-window figures and will want tuning against real traffic. |

### 182.C.31 V1 Convergence and Legacy Hardening

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                                | Owner         | Dependency         | Status                | Changed Files                                                                                                                                   | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | Cost                                            | Blocker                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
| ----------- | ----------------------------------------------------------------------------------------------------------------------- | ------------- | ------------------ | --------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.31.01 | Update V1 persona/avatar/voice documentation with the Phase 182 architecture, honest delivery guarantee …               | Nous + Psyche | 182.C.05           | implemented           | `V1/architecture/persona-avatar-voice-packs.md` (new section)                                                                                   | THE V1 PAGE IS WHERE PEOPLE COME LOOKING FOR THE RENDERING HALF, AND IT DID NOT HAVE IT. The persona/avatar/voice page documents the registry honestly as a CONTRACT AND POLICY LAYER and sent rendering away to "render-side, Bellona/Neith" - so a reader arriving with "can this make a video of a person saying these words" found no answer. The new section states the Phase 182 path end to end and then the four things a V1 reader most needs. (1) THE DELIVERY GUARANTEE IS NARROW AND CHECKABLE: perfect lip sync is a DELIVERY REQUIREMENT, not a claim that every first sample is perfect, so the promise is NO KNOWN BAD-SYNC TAKE IS RELEASED while pass rates and limitations stay visible. (2) EXACT DIALOGUE IS NOT A POST-PROCESSING STEP - approved audio is SUPPLIED TO THE MODEL rather than synchronised onto its output, because a video model may paraphrase, omit words or mispronounce names and nothing downstream can tell you it did unless the words were fixed first; and only ONE of five families carries supplied audio on the OpenRouter route today. (3) THE ACTIVE-SCENE LANE IS NOT THE AVATAR LANE, which is exactly the distinction the page own `avatar-render` capability cannot express: a head speaking a script against a background and a person performing in a place can match on every capability and produce PRODUCTS A CREATOR WOULD NEVER CONFUSE, so crossing is a refusal rather than a warning. (4) STATUS, SO NOBODY READS IT AS SHIPPED: none of it is mounted, no route has a continuity benchmark, no evaluator has scored a real clip, and nineteen threshold sets are PROVISIONAL. Eight per-topic documents are linked, each stating its own gaps.                                                                                                                                                                                                                                                                                                                              | All eight relative links were resolved against the filesystem from V1/architecture/ and exist. The one-family claim, the five capability classes and the two classes with no mode are read from the adapter and capability-class registries; the nineteen provisional exports from the grep behind 182.C.39.05; the unmounted claim from the same grep as 182.C.39.01. Prettier-formatted. No code changed.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | No provider spend; no model call.               | THE PAGE NOW DESCRIBES A RENDERING PATH THAT IS NOT WIRED. It stays honest only while the status paragraph is maintained - when the routes mount or a route gets a benchmark, that paragraph is the first thing that goes stale. 182.C.31.02 is the sibling item that replaces the remaining "provider-gated" ambiguity on the same page.                                                                                                                                                             |
| 182.C.31.02 | Replace V1 “provider-gated” ambiguity with links to canonical contracts, tracker status, verified route evidence …      | Nous + Psyche | 182.C.05           | implemented           | `V1/architecture/persona-avatar-voice-packs.md` (three passages rewritten)                                                                      | "PROVIDER-GATED" IS THREE STATES WEARING ONE PHRASE, AND IT READS AS THE FRIENDLIEST OF THEM TO EVERY READER - "somebody just needs to find the API key". At least two of the things it covered on this page are not that. CREDENTIAL-GATED means the code path is complete and nobody supplied a key; UNVERIFIED means the code exists and HAS NEVER RUN AGAINST THE PROVIDER AT ALL, so whether it works is unknown; UNREACHABLE means no route we have can do this WHATEVER KEYS EXIST. Three states, three different actions: supply a key, run a live probe, or build it and drop the claim. 182.C.10.04 already drew exactly this line for probes (probe_inconclusive is not probe_failed, region_unrecorded is not region_unavailable) and the prose was still collapsing it. Applied to the page own edges, the re-labelling MOVED TWO CLAIMS OUT OF THE COMFORTABLE BUCKET: live cloned-voice TTS stays credential-gated, but AVATAR RENDERING AND LIP-SYNC VERSUS GROUND TRUTH IS UNREACHABLE (the scorer and fixtures exist, the render does not - no route has ever produced a comparable one), and EXACT-DIALOGUE VIDEO IS UNVERIFIED (implemented, one family carries supplied audio, NOT MOUNTED - no caller outside its specs). The section closes by pointing at the human-video tracker, whose six-step status vocabulary is the finer-grained version of the same three states.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | Three passages rewritten and checked: the only surviving "provider-gated" string on the page is the one naming the ambiguity itself. The four claim states are each backed - the TTS modules exist at the paths quoted, the unmounted claim is the same grep as 182.C.39.01, the one-audio-family claim comes from the adapter registry, and the PARTIAL audit date is carried over unchanged. Prettier-formatted. No code changed.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | No provider spend; no model call.               | SCOPED TO THE PERSONA/AVATAR/VOICE PAGE. `V1/ARCHITECTURE.md` (5 uses) and `V1/features.md` (2) still use "provider-gated" as a DOC-CONVENTION LABEL across domains that are not this one; re-labelling those honestly means re-auditing every claim on those pages, which is a different piece of work and should not be done blind. The three-state vocabulary is defined here for them to adopt.                                                                                                   |
| 182.C.31.03 | Integrate the phoneme/viseme mapping only as optional timing metadata for compatible engines …                          | Nous + Psyche | 182.C.05           | done                  | `apps/oshun/bff/src/generation/human-video-viseme-metadata.ts` (new), `human-video-viseme-metadata.spec.ts` (new)                               | @psyche/avatar-lipsync turns a phoneme sequence into a 19-class viseme track with blendshape weights. That track is useful AND IT IS NOT LIP-SYNC: A VISEME TRACK IS AN INSTRUCTION TO A RENDERER THAT MAY NOT EXIST. Handed to a blendshape rig it becomes mouth movement; handed to a video diffusion model it is consumed by NOTHING - the mouth in the output came from the model, and the track sat beside the job having no effect at all. One of three renderer kinds consumes it. A pipeline that computes visemes, generates video and then reports lipSync: true has CLAIMED A RENDERED RESULT FROM AN UNRENDERED INSTRUCTION, which assertNotCountedAsRendered refuses by name. SO ONLY THE MEASUREMENT ON THE DELIVERED VIDEO CAN SAY LIP-SYNC HAPPENED: 182.C.17-20 measure an offset on the OUTPUT and a viseme track is an INPUT, so they are ON OPPOSITE SIDES OF THE THING BEING JUDGED and no viseme evidence can satisfy the sync gate however good the track is - refused outright rather than weighted low. AND THE WRONG VOCABULARY IS WORSE THAN NO TRACK: a 19-class track handed to a renderer expecting ARKit's 52 does not produce nothing, it produces movement keyed to whichever shapes share an index - ABSENT LIP-SYNC LOOKS LIKE A STILL MOUTH AND SOMEBODY FILES A BUG, MISMATCHED LIP-SYNC LOOKS LIKE IT WORKED AND SHIPS. So a renderer declares the vocabulary it accepts and a mismatch is refused. FINALLY THE TIMINGS ARE ONLY AS GOOD AS WHAT PRODUCED THEM: this phase measured that ASR and Whisper word timings are unfit for sync and chose a CTC aligner, so a track built from ASR timings INHERITS THAT UNFITNESS and carries its source rather than pretending all sources are equal. TTS marks are the one case where the timings are not an estimate at all, because the synthesiser knows when it emitted each phoneme.                                                                                                                                                                    | 23 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test drives the attach plan for a usable renderer with UNFIT timings and asserts it attaches while reporting fitForSync false, so the two judgements stay separate.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | No provider spend.                              | For the video-diffusion and image-animator paths - which is every human-video lane in this phase - the track drives nothing and is not attached.                                                                                                                                                                                                                                                                                                                                                      |
| 182.C.31.04 | Turn MuseTalk/portrait plan modules into real adapter- executed pipelines with measured artifacts …                     | Nous + Psyche | 182.C.05           | implemented           | `libs/nous/portrait-animation/src/planning-only.ts` (new), `planning-only.spec.ts` (new), `index.ts` (export)                                   | THE ITEM OFFERS A FORK AND THE EVIDENCE PICKS THE BRANCH: NOT ONE MODULE IN THIS LIBRARY EXECUTES ANYTHING. Twenty-two modules, and none imports `node:child_process`, spawns a process, or makes a network call - so they are marked PLANNING-ONLY, and the marking is a CHECK RATHER THAN A COMMENT. What they produce is a Plan: an argv array, an output path, a threshold, a set of method references. A plan is a description of work somebody else has to run. WHICH MAKES THE DANGEROUS READING AN EASY ONE: a `resolvePlan()` returning `status: pass`, a weighted score and a threshold list LOOKS EXACTLY LIKE A RESULT - 182.C.31.05 found the same shape one layer down, where the quality pipeline was grading LSE values the caller handed it. So `assertNotAMeasuredArtifact` refuses a plan presented as a MEASUREMENT (a number nobody computed, entering a report) or as a CAPABILITY (a render nothing here performs, entering a matrix) - two claims, two failure modes, two different messages. THE DRIFT GUARD IS THE PART THAT KEEPS IT TRUE: the spec reads the directory and asserts the registry accounts for every module and invents none, then re-scans every planning-only module for execution markers. THE DAY A MODULE GROWS AN EXECUTION PATH THE TEST FAILS AND THE REGISTRY ENTRY HAS TO MOVE, which is exactly when a comment would have gone stale instead. Executing any of these needs three things the library does not have: pinned weights (182.C.17.02), somewhere to run them (evaluator-hosting-decision.md), and a corpus (182.C.33).                                                                                                                                                                                                                                                                                                                                                                                                                                                          | 10 tests pass; full portrait-animation suite green at 23 files / 85 tests; eslint clean; typechecked against a scratch config extending tsconfig.lib.json with a deliberate \_\_control.ts error, which reports while the new files do not. The no-execution claim is RE-DERIVED at test time by reading each module source, not asserted in prose, and the guard sensitivity is demonstrated by running the same scan over a synthetic module body that does spawn - it reports `node:child_process` and `spawn(` - while the same scan over a real module reports nothing. The registry-versus-directory comparison is an exact set equality.                                                                                                                                                                                                                                                                              | No provider spend; no model call.               | THIS IS THE "MARK AS PLANNING-ONLY" BRANCH, NOT THE "TURN INTO REAL PIPELINES" BRANCH. Nothing here renders, and the deprecation half - removing these families from any capability claim that still lists them - needs the capability matrix pass in 182.C.38.08, which stays open. `assertNotAMeasuredArtifact` has no caller yet; it is the check a matrix build or a report writer should run.                                                                                                    |
| 182.C.31.05 | Make portrait quality metrics execute pinned evaluators on actual media instead of trusting externally supplied LSE …   | Nous + Psyche | 182.C.05           | implemented           | `libs/nous/portrait-animation/src/portrait-animation-quality-metrics-pipeline.ts`, `portrait-animation-quality-metrics-pipeline.spec.ts`        | THE PIPELINE WAS GRADING NUMBERS THE CALLER HANDED IN. `observedScores` arrived as a plain request field and came back as `status: pass` - so whoever called it SUPPLIED ITS OWN LSE VALUES AND RECEIVED A VERDICT, which is the exact failure 182.C.31.05 names. Scores now carry EVIDENCE - the evaluator id, the digest of its pinned weights, the digest of the bytes it ran over, and when - and all four are required together, because AN EVALUATOR NAME WITHOUT A DIGEST IS A NAME and a digest without the media digest says what ran but not what it ran over. Without complete evidence the numbers are still recorded and the status is `unverified`, never `pass`. THE ASYMMETRY IS DELIBERATE AND IT IS THE INTERESTING PART: a SELF-REPORTED FAILURE IS CREDIBLE BECAUSE IT IS AGAINST INTEREST, so a threshold miss still reports `fail` with no evidence at all; a self-reported PASS is not, so it does not survive. AND A MISSING SCORE WAS BEING SCORED AS 0.5 - a comfortable middle substituted for four different absent metrics, feeding a weighted composite that reads like a measurement and gets quoted. Absent components now land in `unmeasured` and THE COMPOSITE IS WITHHELD rather than computed over substituted values. Two existing tests asserted the old behaviour and were rewritten rather than patched, per the test-coherence rule.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | 8 tests pass in the rewritten spec (up from 4), covering: measured pass with evidence, identical numbers WITHOUT evidence landing unverified, half an evidence record naming both missing fields, a self-declared failure surviving with no evidence, and a withheld composite listing motionNaturalness and headPoseAccuracy. Full lib suite green: 22 files, 75 tests. Downstream consumer `human-video-take-comparison.spec.ts` green: 18 tests. eslint clean. Typechecked against a scratch config extending tsconfig.lib.json with a deliberate \_\_control.ts error, which reports while the changed files do not. NOTE: the lib own vitest preset (`rootedBare`) imports the base config and never merges it, so `@oshun/math` does not resolve when running vitest from the library directory - a scratch config merging `testing/vitest.config.base` was used and deleted. That is pre-existing and not fixed here. | No provider spend; no model call.               | NO PINNED EVALUATOR RUNS HERE YET - this closes the seam rather than filling it. The module builds a PLAN (it emits the SyncNet / ArcFace / FVD / 6DRepNet commands); executing them on CPU is 182.C.17.02 acquisition work and the hosting decision in docs/domains/isis/human-video/evaluator-hosting-decision.md. The sibling `dubbing-quality-validation-pipeline.ts` has the SAME supplied-score flaw and the same 0.5 fallback; it is out of this item scope and should get the same treatment. |
| 182.C.31.06 | Harden the Psyche avatar engine with real entry point, weights, licenses, dependencies, failure checks …                | Nous + Psyche | 182.C.05           | blocked-vendor        | `apps/psyche/avatar-engine/CAPABILITY-STATUS.md` (new), `apps/psyche/avatar-engine/tests/test_capability_status.py` (new), `README.md` (banner) | THE FORK IS DECIDED BY FOUR VERIFIED FACTS, AND THE LAST ONE IS NOT AN ENGINEERING PROBLEM. (1) THE ENTRY POINT DOES NOT EXIST: `project.json` names `avatar_engine.main:app` in BOTH serve targets and the Dockerfile CMD names it a third time, and there is no `src/avatar_engine/main.py` - so `nx serve` fails at import and a container built from this Dockerfile FAILS AT START. Nothing has ever run against this service because there is nothing to run. (2) NO TEST EXERCISES THE ENGINE: 151 Python modules, ~75,000 lines, and `tests/` held a single .gitkeep. (3) THE WEIGHTS ARE ABSENT AND UNACQUIRABLE: config defaults `model_path` to `models/flame/generic_model.pkl`, the directory does not exist, and no download, checksum or digest is defined anywhere. (4) FLAME AND SMPL-X ARE RESEARCH-LICENSED - `smplx` and `chumpy` are declared dependencies and the default face model is FLAME generic_model.pkl - so a commercial capability claim resting on this engine WOULD BE CLAIMING A RIGHT WE DO NOT HAVE. That is a licensing decision, not a bug, and it comes FIRST in the unblock order because everything else is wasted until it is settled. So the engine is RETIRED FROM CAPABILITY CLAIMS, with the status document carrying a machine-readable block that `tests/test_capability_status.py` RE-DERIVES FROM THE REPOSITORY - the day somebody adds main.py the guard fails and the retirement is reconsidered deliberately rather than forgotten. Nothing here says the code is bad; it says no promise may rest on it today. What is NOT claimed: the project declares `gpu:required` and no inference was attempted on this hardware, so nothing in the document is a claim about whether the engine works.                                                                                                                                                                                                                                                                                         | 7 pytest tests pass (stdlib only - no torch, no GPU): entry-point presence versus the claim, the three references that make the absence a live failure, the absence of engine tests (the guard parses its OWN imports with `ast` to prove it exercises nothing), the weights path checked against both disk AND config.py so the path is not a document invention, the two research-licensed dependencies checked against pyproject, and the retirement assertion itself. README carries a status banner. Prettier-formatted.                                                                                                                                                                                                                                                                                                                                                                                                | No provider spend; no model call; no inference. | RETIRED, NOT HARDENED. The unblock order is in the document and starts with a LICENSING DECISION on FLAME/SMPL-X (or a substitution), because an entry point, weights and benchmarks built on a research-licensed model do not produce a shippable capability. No inference has been attempted on this hardware and no benchmark exists.                                                                                                                                                              |
| 182.C.31.07 | (V1.2-gated per `release-scope.ts`) Consolidate Veritas avatar/video providers behind the canonical provider port and … | Nous + Psyche | 182.C.01.11 (V1.2) | blocked-release-train | —                                                                                                                                               | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | —                                               | Veritas returns in V1.2 (`release-scope.ts`).                                                                                                                                                                                                                                                                                                                                                                                                                                                         |
| 182.C.31.08 | Add migration adapters for existing saved cards/jobs and a deprecation window for the single `referenceImage` …         | Nous + Psyche | 182.C.05           | done                  | `apps/oshun/bff/src/generation/human-video-reference-migration.ts` (new), `human-video-reference-migration.spec.ts` (new)                       | THE OLD CONTRACT CANNOT BE MIGRATED - IT CAN ONLY BE GUESSED AT. `referenceImage: "https://cdn/seed.png"` says nothing about ROLE, nothing about WHOSE FACE IT IS, and nothing about whether anybody agreed; the new contract makes an identity reference without an authorization a PARSE ERROR, on purpose. So there are exactly two honest outcomes: map it to `first-frame`, which requires no authorization and IS WHAT THE OLD IMAGE-TO-VIDEO PATH ACTUALLY DID WITH THE STRING - a translation of the old BEHAVIOUR rather than a guess about the picture - or refuse and ask the owner. AND NEVER `identity`, `face-angle` OR `full-body`: upgrading the field into a person-bearing role means SYNTHESISING AN AUTHORIZATION NOBODY GAVE, arrived at THROUGH A DATA MIGRATION RATHER THAN A FEATURE, which is exactly how it would happen. assertNotUpgradedToIdentity checks the OUTPUT because the batch script somebody writes at 2am does not call migrateCard. An owner who declares "identity" in a form is sent to the enrolment path, since A DROPDOWN ANSWER IS NOT AN AUTHORIZATION EITHER. A DEPRECATION WINDOW WITH NO END IS A SECOND CONTRACT: everybody plans a window and nobody sets the date, so both shapes live forever, every code path handles both, AND A BUG IN ONE IS A BUG NOBODY CAN REPRODUCE IN THE OTHER - so the window carries an end and reports days remaining, because a number that goes down is the only kind anybody acts on. A SAVED CARD AND A RUNNING JOB MIGRATE DIFFERENTLY: a card is data at rest and upgrades on next read, while a job IN FLIGHT was already routed, priced and authorized on the old shape, so rewriting its inputs CHANGES WHAT WAS AUTHORIZED AFTER THE AUTHORIZATION - it finishes on the contract it started on. AND THE ADAPTER IS ONE-WAY: serialising a typed set back to a single string loses the role, the subject and the authorization, and a system that CAN write the old shape will do it in some cache or some export, with nothing marking the loss. | 24 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test drives an owner-declared "identity" and asserts the outcome is needs-owner rather than a migration, so the form answer cannot become an authorization by passing through a different code path.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | No provider spend.                              | A legacy image that really was somebody's likeness cannot be migrated at all - it has to go through enrolment, because the authorization does not exist to migrate.                                                                                                                                                                                                                                                                                                                                   |
| 182.C.31.09 | Ensure old text-to-video and image-to-video flows continue working while clearly lacking exact-dialogue guarantees …    | Nous + Psyche | 182.C.05           | done                  | `apps/oshun/bff/src/generation/human-video-legacy-flows.ts` (new), `human-video-legacy-flows.spec.ts` (new)                                     | Text-to-video and image-to-video predate this phase and keep working; what changed is that WE NOW KNOW WHAT THE GUARANTEE WOULD REQUIRE, so shipping them silently would be a different thing than shipping them before we knew. A FLOW THAT DOES NOT MEASURE SPEECH IS NOT A FLOW THAT FAILED THE GATE - those are OPPOSITE FINDINGS AND ONE WORD COVERS BOTH. A legacy job has no script to compare against and no aligner in its path, so all three measurements report NOT-APPLICABLE, and assertNotApplicableKept refuses collapsing that in EITHER direction: reporting a failure fills a dashboard with red nobody can act on, WHICH IS HOW A REAL FAILURE STOPS BEING VISIBLE, and reporting a pass claims a measurement nobody made. Both directions are refused because the two mistakes are made by different people - a dashboard author reaches for "failed" and a sales answer reaches for "passed". THE FLOWS MUST NOT DRIFT INTO THE NEW GUARANTEE BY ACCIDENT: the dangerous path is a legacy job routed through a lane with a lip-sync model, which produces A MOUTH THAT MATCHES SOMETHING, nobody measured what, and a customer reasonably concludes the words are the ones they asked for. THE UPGRADE IS NOT A FLAG because it needs something the old request does not have: exact dialogue needs a script, identity needs an authorization, and image-to-video also needs the role the image plays - so it is A DIFFERENT REQUEST with inputs the caller supplies, and a one-click upgrade WOULD PRODUCE JOBS WHOSE AUTHORIZATION WAS INVENTED AT THE MOMENT OF THE CLICK. AND THE OLD FLOWS ARE NOT REMOVED FOR BEING OLDER: a customer generating product footage from a prompt is not under-served by the absence of an exact-dialogue guarantee - THEY WERE NEVER ASKING FOR ONE, and it costs a script and an authorization they do not have.                                                                                                                                                                     | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test asserts not-applicable is refused as BOTH failed and passed, since a single-direction check would leave whichever mistake the author was not thinking about. A no-op assertion written during drafting was deleted rather than left as a stub.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | No provider spend.                              | Both legacy flows ship with all three human-video measurements NOT-APPLICABLE, which is the honest state rather than a gap to close.                                                                                                                                                                                                                                                                                                                                                                  |

### 182.C.32 Self-Hosted and Corrective Model Research Track

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                               | Owner | Dependency | Status        | Changed Files                                                                                                             | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Cost               | Blocker                                                                                                                                                                                      |
| ----------- | ---------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | ------------- | ------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.32.01 | Define a reproducible benchmark harness before selecting a self-hosted native audiovisual or corrective lip-sync …     | MLOps | 182.C.17   | done          | `apps/oshun/bff/src/generation/human-video-benchmark-harness.ts` (new), `human-video-benchmark-harness.spec.ts` (new)     | THE BENCHMARK COMES BEFORE THE CANDIDATE, AND THAT ORDERING IS THE ENTIRE ITEM. Pick a model first and you build THE BENCHMARK THE MODEL PASSES: every threshold lands where the candidate happened to land, every cohort that embarrassed it is called out of scope, and the result is a document PROVING A DECISION THAT WAS ALREADY MADE. So the harness is frozen with a digest before any candidate runs, and assertFrozenBeforeCandidate compares TIMESTAMPS rather than the order things appear in a document - which is written last in either case - and treats simultaneous as a failure, because that is how "we wrote it as we went" looks. A REPRODUCIBLE BENCHMARK PINS EVERYTHING AND THE ONE PEOPLE FORGET IS THE EVALUATOR: the model gets pinned, the fixtures get pinned, and then SyncNet is updated for an unrelated reason and EVERY HISTORICAL SCORE MOVES QUIETLY IN THE SAME DIRECTION, so a comparison across that boundary measures the evaluator rather than the models. The digest covers evaluator versions, fixtures, cohorts and thresholds, is order-independent (so two identical harnesses compare) and assertComparable refuses runs spanning digests. A BENCHMARK ONLY THE CANDIDATE RAN IS NOT A COMPARISON: "our model scores 0.82" has no denominator, the baseline must be a route CURRENTLY CARRYING PRODUCTION TRAFFIC, and it is re-run in the same session rather than quoted - 182.C.14 established that a provider can change under a pinned model id, so a stale baseline measures WHAT THAT PROVIDER USED TO BE. AND THE HARNESS MUST BE ABLE TO SAY NO CANDIDATE WON: a selection process with no such outcome always selects, so none-met-the-bar is a first-class result - the answer that sends the work back to the commercial route and the one nobody builds a slide for. A candidate that skipped cohorts returns `incomplete` rather than a score, because the skipped ones are usually the hard ones, and a MISSING metric fails its threshold rather than passing it.                                                                                                                                                                         | 23 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not; no literal control bytes (the digest separator is String.fromCharCode(0)). One test asserts the digest is order-independent across both the evaluator map and the cohort list, since a digest that moved with key order would make every honest comparison fail.                                                                                        | No provider spend. | The fixture corpus the harness scores against is 182.C.33; this defines the harness that corpus will be run through, which is the order the item asks for.                                   |
| 182.C.32.02 | Review model and weight licenses for commercial use, redistribution, biometric restrictions …                          | MLOps | 182.C.17   | done          | `apps/oshun/bff/src/generation/human-video-model-licensing.ts` (new), `human-video-model-licensing.spec.ts` (new)         | A MODEL LICENCE AND A WEIGHTS LICENCE ARE TWO DOCUMENTS, AND THEY DISAGREE: the code is very often Apache-2.0 and the checkpoint very often is not. THE REPOSITORY BADGE SHOWS THE CODE LICENCE and that is what people read, while the weights arrive from a different host under terms nobody opened - so both are tracked and the missing one is always the weights. "NON-COMMERCIAL" IS NOT THE RESTRICTION THAT CATCHES US, "NO BIOMETRIC USE" IS. Non-commercial is loud: somebody reads it, the candidate is dropped, and THE PROCESS WORKED. A clause forbidding biometric identification or facial analysis is quieter, lives in an ACCEPTABLE-USE POLICY rather than the licence header, and THIS PHASE IS ENTIRELY BIOMETRIC - the identity gate compares an enrolled face against every output, so a model whose terms forbid that cannot be used here EVEN FOR A TASK THAT LOOKS UNRELATED, because the gate runs on every job. reviewCandidate checks biometric use BEFORE commercial use deliberately, so the one that gets missed fails first and its message is the one somebody reads. AND A TRAINING-DATA RESTRICTION IS A RESTRICTION ON THE OUTPUT: "outputs may not be used to train competing models" binds frames the model produced, and 182.C.26.06's derived-artefact rule makes an EMBEDDING computed from those frames such an output - so it constrains the evaluation pipeline rather than only a training programme nobody has planned. REDISTRIBUTION DECIDES THE DEPLOYMENT SHAPE: weights we may run but not redistribute cannot be baked into a container image that leaves our registry, WHICH RULES OUT BUILD-ONCE-SHIP-ANYWHERE BEFORE ANYBODY WRITES IT, and deciding that early is cheaper than discovering it in a deployment review. Finally an unreviewed licence is a BLOCKING FINDING rather than a TODO: a candidate whose terms nobody read is indistinguishable from one whose terms forbid everything, and the difference only appears after somebody reads them.                                                                                                                                                                                       | 20 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test gives a candidate BOTH a biometric and a commercial prohibition and asserts the biometric one is the reported refusal, pinning the ordering rather than leaving it to whichever check happened to run first.                                                                                                                                   | No provider spend. | No candidate has been reviewed yet - this is the review that must happen before one is selected, and an unreviewed licence blocks rather than warns.                                         |
| 182.C.32.03 | Package candidates with pinned weights, digests, container, GPU requirements, deterministic fixture …                  | MLOps | 182.C.17   | blocked-local | —                                                                                                                         | blocked-local, JUSTIFIED 2026-09-10: self-hosted audio-conditioned video _generation_ candidates are diffusion-class models whose reference implementations require CUDA. Checked on this box: `nvidia-smi` absent, torch is `2.14.0+cpu`; no candidate was run because none can be. Unblocks with a CUDA host (RunPod via `libs/isis/runpod-surface`) — a real hardware blocker, unlike the evaluator rows that used to share this text.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | `nvidia-smi` absent; `torch 2.14.0+cpu`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                  | CUDA host for generation models (`182.C.32.06`)                                                                                                                                              |
| 182.C.32.04 | Benchmark active full-frame performance, not only cropped frontal talking heads, across profile, occlusion, motion …   | MLOps | 182.C.17   | blocked-local | —                                                                                                                         | blocked-local, JUSTIFIED 2026-09-10: self-hosted audio-conditioned video _generation_ candidates are diffusion-class models whose reference implementations require CUDA. Checked on this box: `nvidia-smi` absent, torch is `2.14.0+cpu`; no candidate was run because none can be. Unblocks with a CUDA host (RunPod via `libs/isis/runpod-surface`) — a real hardware blocker, unlike the evaluator rows that used to share this text.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | `nvidia-smi` absent; `torch 2.14.0+cpu`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                  | CUDA host for generation models (`182.C.32.06`)                                                                                                                                              |
| 182.C.32.05 | Measure identity/anatomy damage from corrective models and require all final gates after correction.                   | MLOps | 182.C.17   | done          | `apps/oshun/bff/src/generation/human-video-correction-damage.ts` (new), `human-video-correction-damage.spec.ts` (new)     | A corrective lip-sync model repaints the mouth so it matches the audio - the most useful tool in this phase and the most dangerous, FOR THE SAME REASON: IT CHANGES THE FACE. THE CORRECTION IMPROVES THE GATE IT TARGETS AND CAN FAIL THE ONE NEXT TO IT. Sync goes up; identity can go DOWN because the mouth is part of the face the identity measurement looks at, and anatomy can go wrong in ways SYNC NEVER NOTICES. All five damage dimensions are invisible to the sync score and a test asserts that as a property - TEETH THAT CHANGE SHAPE BETWEEN FRAMES are the classic artefact and a sync score is perfectly happy with them because the timing is right; a jaw moving independently of the head reads as uncanny to a person and as correct to an offset estimator. So THE CORRECTOR IS OPTIMISING THE ONE NUMBER THAT CANNOT SEE WHAT IT BROKE, and a pipeline that measures only sync after a corrective pass has measured only the thing the corrector was optimising. THE DAMAGE IS MEASURED AGAINST THE PRE-CORRECTION TAKE, NOT THE REFERENCE: comparing to the enrolled photo answers "is this the person", which the gate already asks, and it CANNOT TELL A BAD CORRECTION FROM A GENERATION THAT WAS ALWAYS A POOR LIKENESS. The baseline must be captured first because the before values cannot be recovered afterwards - the artifact they described has been overwritten. EVERY FINAL GATE RUNS AFTER CORRECTION: the tempting pipeline measures identity on the generated take, corrects the sync and ships, because THE IDENTITY NUMBER IS ALREADY IN HAND AND RUNNING IT AGAIN COSTS MONEY - and that number describes an artifact that no longer exists. Checked by TIMESTAMP, treating simultaneous as stale, because "we re-ran them" is what everybody believes. AND A NET IMPROVEMENT IS NOT A REASON TO ACCEPT AN IDENTITY LOSS: +0.4 sync against -0.3 identity nets positive on any weighted sum, and THE WEIGHTED SUM IS THE WRONG INSTRUMENT - identity failing means the video may not be the person it claims to be, which is not tradeable against lip timing. Identity is evaluated against its own floor and assertNotAggregated refuses combining them. | 23 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The decision block drives the exact trade the module argues about - a +0.4 sync gain with identity landing below its floor - and asserts the correction is REJECTED, so the claim is checked rather than asserted in prose.                                                                                                                             | No provider spend. | The four anatomy dimensions beyond identity cosine name what to measure; the estimators for them are model work on the 182.C.33 corpus.                                                      |
| 182.C.32.06 | Add secure RunPod/Comfy or sovereign GPU deployment with signed asset access, ephemeral workspace, wipe verification … | MLOps | 182.C.17   | blocked-local | —                                                                                                                         | blocked-local, JUSTIFIED 2026-09-10: self-hosted audio-conditioned video _generation_ candidates are diffusion-class models whose reference implementations require CUDA. Checked on this box: `nvidia-smi` absent, torch is `2.14.0+cpu`; no candidate was run because none can be. Unblocks with a CUDA host (RunPod via `libs/isis/runpod-surface`) — a real hardware blocker, unlike the evaluator rows that used to share this text.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | `nvidia-smi` absent; `torch 2.14.0+cpu`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                  | CUDA host for generation models (`182.C.32.06`)                                                                                                                                              |
| 182.C.32.07 | Compare self-hosted total cost, latency, reliability, engineering load, privacy …                                      | MLOps | 182.C.17   | done          | `apps/oshun/bff/src/generation/human-video-hosting-comparison.ts` (new), `human-video-hosting-comparison.spec.ts` (new)   | THE COMPARISON THAT GETS MADE IS COST PER SECOND, AND IT IS THE ONE THAT IS WRONG. A GPU hour divided by the seconds of video it produced is a real number that omits THE IDLE HOURS BETWEEN JOBS, the engineer who keeps it running, the second region nobody built, and the fact that A COMMERCIAL ROUTE COSTS ZERO WHEN NOBODY IS USING IT. compareCost returns the marginal figure BESIDE the total rather than instead of it, because the marginal one is the number people quote and the total is the number that gets paid - and the spec drives the same hardware at two operating points where the conclusion REVERSES. ENGINEERING LOAD IS A COST AND IT IS THE ONLY ONE WITH NO INVOICE: every other line arrives on a bill somebody sees, and this one arrives as A PERSON NOT DOING SOMETHING ELSE - which is why it is the line left out of the comparison and the line that decides it. Stated in engineer-days per month, which somebody can argue with, rather than converted to money, which invites a conversion nobody agrees on; zero is accepted as a CLAIM while absent is refused. PRIVACY IS THE ONE DIMENSION WHERE SELF-HOSTING WINS OUTRIGHT AND IT IS NOT A TIEBREAK: 182.C.27.02 recorded that OpenRouter video is not ZDR-eligible AT ALL, so a tenant requiring zero retention has NO COMMERCIAL ROUTE rather than a worse one. That makes it a CAPABILITY rather than a score - a margin can be outweighed by five other dimensions and an absent route cannot - so compare returns `no-commercial-alternative` and the other five dimensions describe what self-hosting costs rather than whether to do it. AND A COMPARISON WITHOUT AN OPERATING POINT COMPARES NOTHING: self-hosting wins at high steady volume and loses badly at low bursty volume, and BOTH ARE TRUE OF THE SAME TWO SYSTEMS, so a verdict with no stated volume is a verdict about whichever volume the author had in mind. Finally the comparison must happen BEFORE promotion, because afterwards it is a justification.                                                                                                                                                                        | 24 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The reversal is a worked case rather than a claim: identical rates and hardware at 3,600s/month and 86,400s/month, asserting self-hosting is more expensive at one and cheaper at the other - and that at the volume where COST favours self-hosting the verdict still lands on the commercial route once engineer-days are counted.                    | No provider spend. | The latency, reliability and quality verdicts are INPUTS - measuring them needs a deployed self-hosted route (182.C.32.06) and the 182.C.33 corpus; this is the frame those numbers go into. |
| 182.C.32.08 | Keep experimental routes hidden or allowlisted until live artifacts meet the same release thresholds as commercial …   | MLOps | 182.C.17   | done          | `apps/oshun/bff/src/generation/human-video-experimental-routes.ts` (new), `human-video-experimental-routes.spec.ts` (new) | "EXPERIMENTAL" IS A LABEL ON A ROUTE, NOT A LOWER BAR. The temptation is a gentler threshold set because it is only an experiment - it is not: THE ARTIFACTS GO TO REAL CUSTOMERS CARRYING A REAL PERSON'S LIKENESS, AND THE PERSON IN THE VIDEO DOES NOT KNOW THEIR VIDEO WAS MADE BY AN EXPERIMENT. So the thresholds are compared by IDENTITY rather than by value, because a copied set has the same numbers on the day it is copied and DIVERGES ON THE DAY SOMEBODY IS UNDER PRESSURE. AND THE ROUTE MUST NOT BE REACHABLE AS A FALLBACK - this is the one that actually happens. A route can be hidden from selection and still be reached when the primary lane is busy, because FALLBACK LOGIC LOOKS AT CAPABILITY AND NOT AT VISIBILITY, so the experimental route serves everybody at exactly the moment the system is under stress, WHICH IS ALSO THE MOMENT NOBODY IS LOOKING AT A DASHBOARD. 182.C.27.07 made the same point about residency. HIDDEN AND ALLOWLISTED ARE DIFFERENT AND THE DIFFERENCE IS WHO CAN BE SURPRISED: hidden is the only genuinely safe state, and an allowlisted tenant must have AGREED rather than merely been added - AN ALLOWLIST SOMEBODY WAS PUT ON WITHOUT KNOWING IS A HIDDEN ROUTE THAT SHIPS TO STRANGERS. AND "LIVE ARTIFACTS" MEANS ARTIFACTS, NOT A BENCHMARK SCORE: 182.C.32.01's harness runs on FIXTURES, which are what a route is SELECTED on, while live artifacts are what it is PROMOTED on - and they only exist once the route sees real inputs, which it must not do by delivering. SHADOW TRAFFIC RESOLVES THE CIRCULARITY: real customer inputs, output measured and DISCARDED, nothing delivered. Delivered artifacts deliberately do NOT count toward the gate, because they are evidence about a decision that was already made.                                                                                                                                                                                                                                                                                                                                                                                                     | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not - and that typecheck caught an unused import the 25 green tests had not. One test asserts a threshold copy with IDENTICAL VALUES is still refused, which is the whole point of comparing by identity; another asserts checkRoute reports the fallback failure first when four problems are present at once, since that is the one that actually happens. | No provider spend. | None.                                                                                                                                                                                        |

### 182.C.33 Evaluation Corpus and Ground Truth

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                                | Owner | Dependency | Status      | Changed Files                                                                                                                                                                                       | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Cost                                   | Blocker                                                                                                                                                                                                                                       |
| ----------- | ----------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | ----------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.33.01 | Create a consented, licensed internal corpus with reference packs, exact scripts, clean audio, shot plans …             | MLOps | 182.C.17   | not-started | —                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                      | None                                                                                                                                                                                                                                          |
| 182.C.33.02 | Balance the corpus across gender presentation, age, complexion/skin tone, facial structure, hair/facial hair …          | MLOps | 182.C.17   | not-started | —                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                      | None                                                                                                                                                                                                                                          |
| 182.C.33.03 | Include English, Ghanaian English, Akan/Twi, Ga, Ewe, multilingual and code-switched speech, local names, numbers …     | MLOps | 182.C.17   | not-started | —                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                      | None                                                                                                                                                                                                                                          |
| 182.C.33.04 | Include frontal, three-quarter, profile, back-turn, occlusion, face-size, focus, compression, low-light, motion-blur …  | MLOps | 182.C.17   | not-started | —                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                      | None                                                                                                                                                                                                                                          |
| 182.C.33.05 | Include stationary, walking, sitting/standing, gesturing, cooking, prop interaction, dancing/sports-like motion …       | MLOps | 182.C.17   | not-started | —                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                      | None                                                                                                                                                                                                                                          |
| 182.C.33.06 | Create synthetic negative fixtures with known offsets, drift, dropped/inserted words, wrong speaker, silent mouth …     | MLOps | 182.C.17   | done        | `apps/oshun/bff/src/generation/human-video-negative-fixtures.ts` (new), `human-video-negative-fixtures.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-negative-fixtures/` (new) | THE REAL CORPUS HAS NO LABELS, WHICH IS WHY THESE EXIST: nobody knows the true offset of a clip a customer complained about, so an evaluator measured only against real failures can be told that it FIRES and cannot be told how SENSITIVE it is - and a detector that finds a 200 ms offset and misses a 60 ms one passes every obvious test and fails the ATSC window it was built for. Nine fixtures were GENERATED FOR REAL with ffmpeg 6.1.1 (evidence/2026-09-11-negative-fixtures/, regenerated rather than committed as binaries). A FIXTURE'S LABEL IS WHAT WAS MEASURED, NOT WHAT WAS REQUESTED: asking for a 120 ms audio lag produced +137 ms through the adelay filter and +23 ms (with a 96 ms container start) through -itsoffset - TWO WAYS TO BUILD THE SAME FIXTURE PRODUCE TWO DIFFERENT FIXTURES, AND NEITHER IS 120 ms. AAC adds about 17 ms of encoder padding in the SAME DIRECTION every time, so a requested label is wrong SYSTEMATICALLY rather than randomly, which is the kind of wrong that calibrates a threshold to the wrong place. TWO OF THE BUILT FIXTURES HAVE NO DURATION SIGNATURE AT ALL: silent-mouth and mouth-on-silence have matching stream lengths and are wrong anyway, so an evaluator that compares durations finds seven of nine AND REPORTS A CLEAN BILL OF HEALTH FOR THE TWO WHERE A PERSON IS VISIBLY SPEAKING SILENCE - the failure a viewer notices first. The damaged-edit fixture is RE-ENCODED rather than stream-copied: a copy-based concat produced a 12.4s video against a 2.0s audio track, which is damaged in a way nobody chose and therefore has no known ground truth; re-encoding gives a real 0.5s cut. AND A NEGATIVE FIXTURE THAT PASSES IS A FINDING RATHER THAN A BUG IN THE FIXTURE: the instinct on a miss is to check the fixture, the fixture is usually fine, and A FIXTURE THE DETECTOR CANNOT SEE IS THE MOST INFORMATIVE ARTIFACT IN THE SET - and the one that gets deleted.                                                                                                                                                                                                                                                                                   | 28 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The first spec block READS the committed manifest.json from the real run and asserts the two 120 ms fixtures differ from each other by more than the request itself, so the central claim is checked against the run rather than asserted. A failing test corrected my own miscount: four defects are content-signature (wrong-speaker and identity-swap as well), of which two can be synthesised from a tone.                                   | No provider spend (local ffmpeg only). | wrong-speaker and identity-swap need real faces and voices and belong to 182.C.33.01; the word-level defects are DURATIONS here rather than words, because a test tone has no words - so these measure an offset detector and not an aligner. |
| 182.C.33.07 | Define blinded multi-rater human protocols for sync, intelligibility, exactness, identity, naturalness …                | MLOps | 182.C.17   | done        | `apps/oshun/bff/src/generation/human-video-rater-protocol.ts` (new), `human-video-rater-protocol.spec.ts` (new)                                                                                     | AGREEMENT IS THE MEASUREMENT AND THE RATINGS ARE THE INPUT: a protocol that collects seven scores and reports their mean has measured nothing, because THE MEAN OF THREE RATERS WHO DISAGREE COMPLETELY AND THREE WHO AGREE EXACTLY IS THE SAME NUMBER. Inter-rater agreement says whether the QUESTION was answerable, and a dimension where raters do not agree is one where the rating is noise - a finding about the PROTOCOL rather than about the video. AND PERCENT AGREEMENT IS NOT AGREEMENT: on a question where 95% of clips are fine, two raters who both answer "fine" to everything agree 95% of the time AND HAVE DONE NOTHING. Cohen's kappa is implemented and named rather than left as "agreement", and the spec demonstrates the case it exists for - two raters who each find a DIFFERENT five bad clips out of a hundred score 90% RAW AGREEMENT AND A NEGATIVE KAPPA. SOME OF THE SEVEN ARE NOT OPINIONS: EXACTNESS IS A FACT - did the person say the approved words - so a rater's view is worth less than an aligner's measurement and asking for one INVITES A MAJORITY VOTE TO OVERRULE A TRANSCRIPT. SYNC IS A MEASUREMENT A PERSON CANNOT MAKE: 182.C.17 works in milliseconds and a human notices "off", so the human question is DETECTABILITY - a different variable with a different threshold - and conflating them is how a rater panel gets used to justify shipping outside the production window. Identity is a SECOND OPINION that catches what a cosine misses and does not replace it, and naturalness carries the lowest agreement floor because the question is genuinely subjective, which is HONEST RATHER THAN LENIENT. BLINDING IS ABOUT THE RATER'S EXPECTATIONS, NOT THEIR HONESTY: a rater who knows which model produced a clip IS NOT CHEATING WHEN THEIR SCORES DRIFT, THEY ARE HUMAN. Order must be randomised PER RATER, since one shared random order is A FIXED ORDER WITH EXTRA STEPS and fatigue then lands on the same clips for everybody and LOOKS LIKE AGREEMENT. And the set carries controls - 182.C.33.06's negative fixtures - because a panel that rates everything highly is indistinguishable from a good system UNTIL A CONTROL THEY SHOULD HAVE FAILED COMES BACK PASSING. | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The kappa implementation is verified NUMERICALLY against the skewed case it exists for rather than asserted: 100 clips, two raters each finding a different five bad ones, giving 90% raw agreement and a kappa below zero, against the same-five case scoring more than 0.9 higher.                                                                                                                                                              | No provider spend.                     | The protocol and its agreement mathematics are here; running it needs raters and the 182.C.33.01 corpus.                                                                                                                                      |
| 182.C.33.08 | Split corpus by subject and scenario to prevent identity or prompt leakage between calibration, validation …            | MLOps | 182.C.17   | done        | `apps/oshun/bff/src/generation/human-video-corpus-splits.ts` (new), `human-video-corpus-splits.spec.ts` (new)                                                                                       | A RANDOM SPLIT BY CLIP LEAKS THE SUBJECT, AND THAT IS THE WHOLE PROBLEM: one person appears in twelve clips, so a clip-level shuffle puts eight in calibration and four in validation and the identity threshold is FITTED ON A FACE IT WILL THEN BE VALIDATED AGAINST. The number that comes back is not "can this system recognise a person", it is "CAN IT RECOGNISE THIS PERSON, WHOM IT HAS ALREADY SEEN" - and it is always better than the truth. SCENARIO LEAKS TOO, AND IT LEAKS QUIETLY: the same script, the same kitchen, the same angle under the same lamp appearing in both halves means the generator was already tuned for it, so validation answers a question about COVERAGE rather than GENERALISATION - and NOBODY NOTICES, BECAUSE UNLIKE A REPEATED FACE A REPEATED SCENARIO DOES NOT LOOK LIKE A DUPLICATE. findLeaks reports the two separately for that reason, and assignSplits assigns by SUBJECT and DETERMINISTICALLY - a split that moves between runs makes two evaluations incomparable for a reason nobody thinks to check. THE HELD-OUT SET IS HELD OUT UNTIL IT IS SPENT: looking at a held-out result and changing a threshold converts it into a calibration set NOT GRADUALLY, IMMEDIATELY, and nothing marks the moment - so a second look is refused even when no threshold changed, because the second look is a decision informed by the first. AND A REVOKED SUBJECT MUST LEAVE EVERY SPLIT: 182.C.26.02 makes the withdrawal real, and the data leaving is the easy half - a subject removed from CALIBRATION after a threshold was fitted on them LEAVES THE THRESHOLD BEHIND. The data is gone and its influence is not, which is a different problem and the one nobody handles; revoking a held-out subject changes what the corpus covers and not what any number means.                                                                                                                                                                                                                                                                                                                                                                                                                          | 20 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The split assigner is exercised on a real 72-clip corpus (six subjects, twelve clips, three scenarios each) asserting every clip of a person lands in one split, that all three splits are used, and that two runs over the same corpus produce identical assignments. The scenario-leak case is constructed with DIFFERENT subjects sharing a scenario, so it is caught by the scenario check alone rather than incidentally by the subject one. | No provider spend.                     | None.                                                                                                                                                                                                                                         |
| 182.C.33.09 | Version dataset, consent, labels, evaluator outputs, and allowed uses; remove revoked subjects and invalidate …         | MLOps | 182.C.17   | done        | `apps/oshun/bff/src/generation/human-video-corpus-versioning.ts` (new), `human-video-corpus-versioning.spec.ts` (new)                                                                               | FIVE THINGS VERSION SEPARATELY AND PEOPLE VERSION ONE: "Corpus v3" usually means THE MEDIA, while the LABELS get corrected without a bump, the CONSENT scope narrows when somebody asks, the EVALUATOR changes under both, and the ALLOWED USES were written down once and never revisited. A score is comparable only against one taken with ALL FIVE THE SAME, so the comparable key is all five and assertComparable names WHICH facet differs - because "different corpus version" reads as the media and the interesting cases are the two nobody sees. THE LABELS ARE THE ONE THAT MOVES SILENTLY: media is append-only and a consent change makes somebody unhappy enough to mention it, but A LABEL CORRECTION IS A WELL-MEANT FIX BY A PERSON IMPROVING THE DATA, and it moves every score computed against it IN THE DIRECTION OF WHATEVER THEY THOUGHT WAS WRONG, WHICH IS NOT RANDOM. The evaluator moves the same way, which is the point 182.C.32.01 made about a benchmark harness. REMOVING A SUBJECT IS THE EASY HALF OF A REVOCATION: 182.C.33.08 already showed a threshold fitted on a withdrawn subject SURVIVES THEM, and this walks the rest of the graph - evaluator outputs, fitted thresholds, published baselines, promotion decisions and external reports - separating what can be recomputed from what cannot, because THE ARTIFACT NOBODY INVALIDATES IS THE SLIDE: a document somebody else already holds. AND AN ALLOWED-USE NARROWING IS ABOUT THE FUTURE ONLY: a subject who consented to evaluation and later refuses training HAS NOT MADE YESTERDAY'S EVALUATION IMPROPER, and a model already trained on their data IS NOT UNTRAINED BY THE NARROWING - saying the corpus "now excludes training uses" DESCRIBES THE CORPUS RATHER THAN THE MODEL, the same gap 182.C.27.08 found between a deletion API and the bytes.                                                                                                                                                                                                                                                                                                                                                                                      | 27 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. One test changes ONLY the label version and asserts the comparable key moves, which is exactly the case a media-only version would miss; the dependency walk is exercised with a dependent computed over an OLDER media version and asserts it is not invalidated, so the graph is scoped rather than swept.                                                                                                                                      | No provider spend.                     | An external report cannot be recomputed - it is a document somebody else holds - so a withdrawal leaves it standing and the plan says so rather than omitting it.                                                                             |
| 182.C.33.10 | Maintain a small public/licensable smoke corpus and a protected full corpus so CI can test contracts without exposing … | MLOps | 182.C.17   | done        | `apps/oshun/bff/src/generation/human-video-smoke-corpus.ts` (new), `human-video-smoke-corpus.spec.ts` (new)                                                                                         | CI IS A PLACE BIOMETRIC DATA SHOULD NEVER BE, AND IT IS THE PLACE IT ENDS UP: a runner is ephemeral, its logs are kept for months, it is readable by anyone with repo access, and it pulls fixtures from wherever the config points - so assertCorpusAllowed refuses the protected corpus in CI and its message names the retention and the readership rather than citing a policy, because the pressure to put the real corpus there is constant and honest: THE REAL CORPUS IS THE ONE THAT FINDS THE BUG. THE SMOKE CORPUS TESTS CONTRACTS AND MUST NOT PRETEND TO TEST QUALITY: its clips are 182.C.33.06 generated fixtures - a test pattern and a tone - so A SMOKE RUN THAT REPORTS A QUALITY FIGURE HAS MEASURED A SINE WAVE, and the figure is REFUSED RATHER THAN LABELLED because a labelled number is still quoted. CORPUS_SPECS therefore carries what the smoke set answers (four things) beside what it cannot (does the face look like the person, is the mouth in time with the words, would a reviewer release this). A GREEN SMOKE RUN IS NOT EVIDENCE THE SYSTEM WORKS - IT IS EVIDENCE THE SYSTEM RUNS, two different claims CI reports in the same colour, so honestConclusion states what a pass did and did not establish, and a red smoke run stays worth acting on immediately because a broken contract is real on a test pattern too. AND THE TWO CORPORA MUST NOT SHARE A CODE PATH THAT CAN CHOOSE: a loadCorpus(name) WILL EVENTUALLY BE POINTED AT THE PROTECTED ONE FROM A CI CONFIG, by somebody debugging a flaky test at the end of a long day - so the smoke loader takes no parameter and assertLoaderCannotChoose refuses one that does.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | 21 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The refusals are asserted by message text, so the reasoning cannot be deleted while the code still passes: the CI refusal must name the 90-day log retention and the readership, the quality refusal must say the number would be quoted, and the loader refusal must name who points it at the wrong corpus and when.                                                                                                                            | No provider spend.                     | A parameterised loader is the ordinary design and it is the one refused here - the smoke path takes no argument, so there is no configuration value that can redirect it.                                                                     |

### 182.C.34 Model Bakeoff and Promotion Benchmarks

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                           | Owner | Dependency | Status | Changed Files                                                                                                                                                                                                                                 | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Cost               | Blocker                                                                                                                                                                                                                                                                                                      |
| ----------- | ------------------------------------------------------------------------------------------------------------------ | ----- | ---------- | ------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| 182.C.34.01 | Run the same canonical requests through Wan, Seedance, Veo, HeyGen avatar, direct providers …                      | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-canonical-bakeoff.ts` (new), `human-video-canonical-bakeoff.spec.ts` (new), built on the existing `human-video-lane-adapters.ts` registry                                                          | "THE SAME REQUEST" CANNOT MEAN THE SAME PAYLOAD, SO FAVORITISM HIDES IN THE TRANSLATION: the lane adapters exist because one family calls a control `audio` and another has no name for it, so the bytes reaching Wan and HeyGen differ BY NECESSITY and byte-equality proves nothing. The checkable rule is provenance - EVERY FIELD IN EVERY COMPILED PAYLOAD MUST TRACE BACK TO A CONTROL THE CANONICAL REQUEST NAMES - and findFavoring reports both the field with no source and the field mapped from a control the request never asked for, because A FIELD ADDED BECAUSE IT HELPS THIS MODEL IS EXACTLY WHAT FAVORING LOOKS LIKE FROM THE INSIDE. THE FAVORING THAT ACTUALLY HAPPENS IS HONEST AND IT HAPPENS TO THE PROMPT TEXT: somebody runs the first family, the take is poor, they rewrite the prompt, the take improves, and the family that ran first is the only one judged on the draft - a real improvement in the prompt reading as a difference between providers - so the request is frozen and digested before the first run and an edit INVALIDATES EVERY RUN ALREADY TAKEN, not only the ones after it (182.C.32.01 freeze, applied to the input rather than the scoring). A FAMILY THAT CANNOT EXPRESS THE REQUEST HAS NOT LOST THE BAKEOFF: the HeyGen avatar lane refuses six of the seven modes and Seedance 2.5 cannot edit through OpenRouter, so a run outcome is a union of scored / not-expressible / refused / errored and ONLY `scored` ENTERS AN AGGREGATE - a zero would enter a mean and leave as a rank saying the model is bad at something it was never asked to do. AND AN ELIGIBLE SUCCESSOR IS NOT AN ENTRANT UNTIL IT RUNS THE WHOLE SET: canary scores are a convenience sample of whatever creators asked for that week, not of the shapes being tested. One consequence: a model reachable through OpenRouter and through a direct client is TWO ENTRANTS, because one merged row would describe an editing capability the route we use does not have.                                                               | 22 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The family list is read from the live LANE_ADAPTERS registry rather than restated, so a test asserts both Seedance families and Veo and HeyGen are covered. The mean test uses two scores and one timeout and asserts 0.7 rather than 0.467.                                                                                                                                                                                                                                                                                                                                                                                                            | No provider spend. | NOT RUN: no canonical request has been sent to any provider. This is the protocol and the arithmetic that keeps a sweep fair, not a sweep - the bakeoff itself needs the 182.C.33 corpus and billed spend across five families. The honest negative is recorded here rather than as a table of empty scores. |
| 182.C.34.02 | Compare native supplied-audio, generated-dialogue, avatar, edit/extension, and corrective lanes separately.        | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-lane-comparison.ts` (new), `human-video-lane-comparison.spec.ts` (new)                                                                                                                             | THE FIVE LANES DO NOT SHARE A SCORE BECAUSE THEY DO NOT SHARE A FAILURE. A supplied-audio lane is handed audio it cannot move and fails by drifting from it; a generated-dialogue lane made both halves and IS NEVER OUT OF SYNC WITH ITSELF, so a near-perfect sync number there is the lane rather than the model, and it fails by saying the wrong words; an avatar lane draws identity from a stored avatar rather than reconstructing it, so identity sits near the ceiling for the same structural reason; an edit lane fails by changing something the request did not ask to change, WHICH IS NOT IN THE OUTPUT AT ALL, only in the difference from the take that went in. So the same column name means four different things and a pooled mean is arithmetic on incompatible quantities: LANE_SPECS separates discriminating metrics from structural ones (reported, never ranked) from undefined ones (refused). THE CORRECTIVE LANE SCORE IS NOT ITS OWN: its input is another lane output, so a final-output figure is the base lane work plus the correction and THE TWO MOVE INDEPENDENTLY - a corrector that damages a good take still outscores one that rescues a poor one - so the measure is the delta and a delta needs the pre-correction take kept (182.C.32.05). A LANE WITH THE HIGHER MEAN MAY SIMPLY BE THE EASIER LANE, so lanes are reported beside one another and models are ranked within a lane. THE CUT THAT MATTERS IS NOT THE CAPABILITY-CLASS CUT: the benchmark cut is by HOW THE SPEECH GETS THERE and CapabilityClass is by WHAT KIND OF PRODUCT comes out, and they cross - exact-dialogue-performance and text-to-scene are both `active-scene` and they are the supplied-audio and generated-dialogue lanes, THE PAIR WHOSE FAILURES ARE OPPOSITE, so a class-grouped table merges exactly the two that must stay apart. And the lane follows from the mode rather than being chosen, so no comparison yields "use the avatar lane" to a creator who asked for their own recorded performance.                             | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. classesHoldingSeveralLanes COMPUTES the crossing from classForMode rather than asserting it, so a new mode landing in an existing class surfaces instead of quietly joining a pooled row; a test checks the active-scene crossing through the real capability-class table.                                                                                                                                                                                                                                                                                                                                                                              | No provider spend. | The avatar lane is missing a scene-motion metric and the edit lane a visual-quality one - stated as undefined with the reason rather than filled with a number computed from something else.                                                                                                                 |
| 182.C.34.03 | Record exact speech, sync, speaker attribution, identity, anatomy, motion, control adherence, visual quality …     | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-benchmark-record.ts` (new), `human-video-benchmark-record.spec.ts` (new)                                                                                                                           | THE TWELVE COLUMNS ARE FOUR KINDS OF THING AND ONLY ONE KIND HAS A THRESHOLD: eight measure THE TAKE, first-pass and eventual pass measure THE PIPELINE and are properties of a population of attempts with NO VALUE FOR A SINGLE RUN, latency and cost measure THE RUN, and evaluator coverage measures THE MEASUREMENT - so a weighted sum over a word error rate, a millisecond and a dollar is not a number, and one row invites one. EVALUATOR COVERAGE MODIFIES EVERY OTHER COLUMN AND IS PRINTED BESIDE THEM AS AN EQUAL: identity 0.90 at coverage 0.40 does not mean identity is 0.90, it means 0.90 on the part that was looked at and THE OTHER 60% IS WHY THE REVIEWER IS THERE - so coverage is a qualifier on the row, a row without it is refused rather than defaulted (a defaulted 1.0 asserts full coverage nobody measured), and ranking by it rewards an evaluator that looked at everything over a take that was good. A MISSING CELL AND A ZERO CELL LOOK THE SAME IN A TABLE AND THEY ARE OPPOSITE: a cell is measured, not-applicable (182.C.34.02 - sync on a generated-dialogue lane is not a low score, it is not a score), or unmeasured (nobody looked), three states behind one blank. AND COST AND LATENCY ARE THE EXACT COLUMNS THAT COME FROM THE SUBJECT OF THE MEASUREMENT: every other figure is computed here from bytes we hold, while these two are reported by the provider being benchmarked - and a provider latency figure is its own processing time WITH THE QUEUE THE REQUEST WAITED IN EXCLUDED, so readTiming keeps the wall clock, their figure, and the gap, because publishing theirs alone benchmarks their best day and publishing ours alone charges them our upload.                                                                                                                                                                                                                                                                                                                                           | 25 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The eight take-columns are imported from 182.C.34.02 METRICS rather than restated, so the two files cannot drift apart; one test keeps a MEASURED zero (a real finding) while two refuse a rendered zero for the two absences.                                                                                                                                                                                                                                                                                                                                                                                                                          | No provider spend. | A provider that reports no processing time leaves unaccountedMs null and the note says the wall clock includes their queue and our transfer - stated rather than estimated.                                                                                                                                  |
| 182.C.34.04 | Publish confidence intervals, cohort sizes, failure examples, human agreement …                                    | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-benchmark-publication.ts` (new), `human-video-benchmark-publication.spec.ts` (new)                                                                                                                 | AN INTERVAL OVER EIGHT CLIPS IS USUALLY WIDER THAN THE GAP BETWEEN EVERY MODEL, AND THAT IS THE RESULT: a rank ALWAYS separates and an interval often does not, so orderVerdict returns `indistinguishable` with the sentence that changes the decision - THE CHOICE BELONGS TO COST, LATENCY OR LICENSING, WHICH ARE MEASURED EXACTLY - and assertOrderSupported refuses the order rather than annotating it. The interval is a real seeded PERCENTILE BOOTSTRAP (2000 resamples, mulberry32) with its tradeoff stated: a bootstrap resamples only the numbers it was given, so below MIN_SAMPLES_FOR_INTERVAL=5 it is REFUSED RATHER THAN WIDENED because at n=3 it draws from three values and returns something SHAPED LIKE STATISTICS. A FAILURE EXAMPLE IS A VIDEO OF A REAL PERSON FAILING and the instructive ones are the worst frames of somebody who agreed to be measured: A RESEARCH CONSENT IS NOT A PUBLICATION CONSENT, so a published example comes from the 182.C.33.10 smoke corpus or carries a publication consent of its own. HUMAN AGREEMENT IS A PROPERTY OF THE RATERS AND IT CAPS EVERY OTHER NUMBER (182.C.33.07): at kappa 0.45 an evaluator matching humans 80% of the time is reproducing a judgement humans barely reproduce, so it travels with the scores rather than sitting beside them as one more figure. AND A KNOWN BLIND SPOT IS ONE SOMEBODY WROTE DOWN, SO THE LIST READS AS COMPLETE: the only evidence about the spots NOT on it is HOW THE LISTED ONES WERE FOUND, so each entry carries a discovery method and blindSpotReport says outright when one method found everything - a list built from reviewer complaints is exactly the blind spots that produce complaints, and says nothing about the ones that produce a quiet release.                                                                                                                                                                                                                                                                                 | 21 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The bootstrap is asserted against values computed from it (mean 0.7075, interval [0.6875, 0.72625] over eight scores, seed 1) and for reproducibility across seeds. One assertion of mine was wrong and was corrected rather than the code: the bounds land on resampled means, so a SINGLE bound repeats across seeds even when the interval as a whole moves - the test now compares the interval, and a second seed is used to show a bound moving.                                                                                                                                                                                                  | No provider spend. | Refusing an interval below n=5 means a five-family bakeoff over a four-clip set publishes four numbers and no interval. That is the intended outcome - the numbers are honest and the interval would not be.                                                                                                 |
| 182.C.34.05 | Test every advertised duration, aspect, resolution, reference class, audio mode …                                  | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-option-coverage.ts` (new), `human-video-option-coverage.spec.ts` (new)                                                                                                                             | ADVERTISED AND ACTUALLY-REACHED ARE TWO DIFFERENT SETS AND THE GAP RUNS BOTH WAYS: a provider advertises fifteen durations and routing sends three, so testing all fifteen spends the budget on combinations nobody reaches - but routing also reaches combinations the catalog NEVER ADVERTISED, because a fallback route was taken, and A FALLBACK IS CHOSEN WHEN THE PRIMARY IS BUSY, WHICH IS EXACTLY WHEN NOBODY IS WATCHING. So the test set is (advertised INTERSECT routable) UNION reached, and THE THIRD TERM IS THE ONE THAT FINDS BUGS. THE COMBINATIONS MULTIPLY AND THE BUDGET DOES NOT, so the plan is ALL-PAIRS rather than the cross product: MEASURED on a realistic 3x3x2x3x2x3 space, the greedy plan is 15 CASES COVERING ALL 106 PAIRS AGAINST A 324-COMBINATION CROSS PRODUCT - a 21.6x reduction, verified by assertAllPairsCovered rather than assumed. The tradeoff is documented rather than glossed: all-pairs catches any defect caused by an interaction of TWO parameters and misses one needing THREE, and it is chosen because the defects this phase has actually measured are two-way (182.C.33.06 AAC frame quantisation is duration x audio-mode; the watermark loss is resolution x delivery-format). A DIMENSION WITH ONE VALUE IS NOT TESTED, IT IS HELD CONSTANT, and a report that counts dimensions calls it covered - so held dimensions are named with the value they were pinned to, because that value is A SILENT PRECONDITION ON EVERY RESULT IN THE SWEEP. AND THE SET OF OPTIONS "ROUTING ACTUALLY USES" IS A FACT ABOUT PRODUCTION THE ADAPTER TABLE CANNOT SUPPLY: the table says what COULD be sent, so describing a table-derived set as what routing actually uses is a claim about traffic backed by a claim about code, and it is refused.                                                                                                                                                                                                                                                                  | 20 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The all-pairs numbers in the spec (324 combinations, 106 pairs, 15 cases) were COMPUTED by running the generator, not estimated. Two things the typecheck caught after 20 green tests: an invalid cast in the spec, and - the real one - that the space and the test case are genuinely PARTIAL records (a plan may vary two dimensions or six), which the full-record signature had been hiding behind casts. One test asserts every value of the FIRST dimension appears, which is the case the naive greedy silently fails: nothing is chosen when that dimension is decided, so every gain is zero and it would take its first value in every case. | No provider spend. | All-pairs is a documented approximation, not full coverage: a defect needing three parameters together is out of scope by construction, and coverageReport states that in the plan rather than in a footnote.                                                                                                |
| 182.C.34.06 | Repeat benchmarks across at least three time windows to detect provider/model drift and non-deterministic outages. | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-benchmark-windows.ts` (new), `human-video-benchmark-windows.spec.ts` (new), reusing the 182.C.34.04 bootstrap                                                                                      | DRIFT AND NON-DETERMINISM WANT OPPOSITE SPACINGS, AND THREE WINDOWS SPENT ON ONE CANNOT SEE THE OTHER. Drift is the model behind a stable name changing - `google/veo-3.1` today and in three weeks may be different weights and NO PROVIDER ANNOUNCES IT - so it needs windows spread over WEEKS. Non-determinism is the same model answering differently, and a degraded provider answering worse, so it needs windows CLOSE TOGETHER under one load condition. "We ran it three times over three weeks and the numbers moved" fits BOTH explanations, so a sweep declares which it detects and its spacing is checked against the declaration - and the refusal NAMES THE OTHER TARGET, because a sweep spaced for one is usually read as evidence about both. A DIFFERENCE BETWEEN WINDOWS IS NOT DRIFT UNTIL IT EXCEEDS THE VARIANCE WITHIN ONE: non-determinism IS variance and drift is a shift on top of it, so driftEvidence compares the between-window range against the widest within-window bootstrap interval, both computed by the SAME 182.C.34.04 estimator. MEASURED: three steady windows move 0.0025 between them while a single window already spans 0.0400 - THE MOVEMENT IS A SIXTEENTH OF THE NOISE - and the verdict is within-noise; a real shift moves 0.1938 against a 0.0388 spread and is called drift. AN OUTAGE IS NOT A LOW SCORE AND IT ARRIVES IN THE DATA AS ONE: a provider degrading under load DOES NOT RETURN ERRORS, IT RETURNS TAKES, AND THEY ARE WORSE, so a window whose failure rate exceeds 3x the median (measured case: 45% against a 5% median) is QUARANTINED RATHER THAN DISCARDED - a discarded window leaves no trace and the sweep reads as clean. And all of it rests on the 182.C.34.01 freeze: a canonical request edited between window two and three produces a difference that READS EXACTLY LIKE DRIFT.                                                                                                                                                                                                 | 23 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. Every drift and quarantine figure in the spec was computed by running the module first, not estimated: 0.0025 vs 0.0400 for the steady case, 0.19375 vs 0.03875 for the drifting one, 45% against a 5% median for the outage.                                                                                                                                                                                                                                                                                                                                                                                                                           | No provider spend. | Three windows can be spaced for drift OR for non-determinism, not both - windowSweepReport says which this sweep is not evidence about rather than leaving the reader to assume it covers both.                                                                                                              |
| 182.C.34.07 | Define minimum sample size and promotion gates for each routing cohort and quality tier.                           | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-promotion-gates.ts` (new), `human-video-promotion-gates.spec.ts` (new)                                                                                                                             | A MINIMUM SAMPLE SIZE IS DERIVED, NOT PICKED, AND THE ARITHMETIC IS NOWHERE NEAR THE HABIT. To be 95% confident a failure rate is at most p after seeing NO failures needs n >= ln(0.05)/ln(1-p), exact for the binomial zero-failure case. COMPUTED: a 99% tier needs 299 CONSECUTIVE CLEAN RENDERS, a 95% tier 59, a 99.9% tier 2,995. Nobody budgets 299, which is why tiers get promoted on 30 and the real number is never written beside the promise. A MINIMUM PER COHORT IS A MULTIPLICATION AND THE COHORT THAT GETS CUT IS THE ONE THE FAIRNESS CLAIM RESTS ON: six cohorts at 299 is 1,794 renders at provider prices, and the cohort trimmed first is always the one with the fewest creators. So the outcome when it is unaffordable is NOT-PROMOTED, kept distinct in the verdict union from BLOCKED (enough samples and it failed - a real result) - never "promoted with limited data", because THAT PHRASE MEANS PROMOTED AND THE QUALIFIER DOES NOT TRAVEL WITH THE DECISION. A GATE AN AGGREGATE CAN SATISFY CAN BE SATISFIED WITHOUT EVERY COHORT: a 96% overall rate holds with one cohort at 99% and another at 70%, and the aggregate is the number that gets reported. AND THE NORMAL APPROXIMATION IS WORST EXACTLY WHERE PASS RATES LIVE: MEASURED, detecting a five-point drop from a 0.98 baseline returns n=31, which yields 0.6 EXPECTED FAILURES - the formula produced a number it cannot stand behind - and even a two-point drop at n=189 gives only 3.8. At a 0.7 baseline the same formula is valid at n=323. So the validity check is computed and REPORTED, and an unsupported size is refused rather than published.                                                                                                                                                                                                                                                                                                                                                                                                           | 19 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. Every figure (299, 59, 2995, 1794, 31/0.6, 189/3.8, 323) was COMPUTED by running the module, not estimated. The rule of three is named as this formula approximation rather than used, because the approximation is what turns 299 into "about a hundred".                                                                                                                                                                                                                                                                                                                                                                                              | No provider spend. | 182.C.29.06 measured a pixel skin-tone cohort as largely a lighting cohort, so cohort membership is itself uncertain - promotionReport carries that caveat, because it caps what a per-cohort gate can claim however many samples it holds.                                                                  |
| 182.C.34.08 | Archive redacted payloads, exact output digests, evaluator versions, cost, catalog snapshot …                      | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-benchmark-archive.ts` (new), `human-video-benchmark-archive.spec.ts` (new)                                                                                                                         | A REDACTED PAYLOAD IS NOT A REPRODUCIBLE ONE, AND THE TWO GOALS IN THIS ITEM PULL AGAINST EACH OTHER: the payload carries reference images (a person face) and approved speech (a person voice), so REDACTING THEM MAKES THE ARCHIVE SAFE TO KEEP AND MAKES THE RUN IMPOSSIBLE TO REPEAT, and an archive that looks complete and reproduces nothing is worse than one that states its limit. So the archive holds digests plus POINTERS to originals living under their own consent, reproducibility becomes "re-fetch if still authorized", and a withdrawal makes the run PERMANENTLY UNREPEATABLE - the correct outcome, not a gap. AN OUTPUT DIGEST PROVES WHAT CAME BACK, NOT THAT IT WOULD COME BACK AGAIN: generation is non-deterministic with or without a seed, so a digest filed as a re-run target invites a check that FAILS EVERY TIME AND GETS DELETED FOR BEING FLAKY, TAKING THE REAL CHECK WITH IT. THE EVALUATOR VERSION IS WHAT MAKES OLD SCORES RECOMPUTABLE AND IT IS THE FIELD PEOPLE FORGET BECAUSE THE EVALUATOR IS OURS (182.C.33.09): bytes plus a version means a score can be RECOMPUTED though the take cannot be REGENERATED, and that smaller claim is the reproducibility this archive can actually offer - so assessReproducibility returns one of three levels and NONE of them is "run-reproducible". AND THE CATALOG SNAPSHOT IS A DATED CLAIM THE ARCHIVE MAKES LOOK LIKE A FACT: answering "does this model support X" from a month-old snapshot answers a question about the present from a record of the past, which is exactly what the live-catalog rule in human-video-lane-adapters.ts exists to prevent. A withdrawal reaches the take too, because the take is a video of them.                                                                                                                                                                                                                                                                                                                                        | 22 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The three reproducibility levels are each exercised by removing exactly one thing - consent, the evaluator version, the output bytes - so the level moves for the reason claimed rather than by coincidence.                                                                                                                                                                                                                                                                                                                                                                                                                                            | No provider spend. | After a withdrawal the archive keeps digests, cost and the review decision and nothing that resolves to the person. The score outlives the video, and the run can never be repeated again.                                                                                                                   |
| 182.C.34.09 | Re-run a bounded benchmark on provider/model/schema change and automatically demote routes that regress beyond …   | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-auto-demotion.ts` (new), `human-video-auto-demotion.spec.ts` (new)                                                                                                                                 | A TOLERANCE NARROWER THAN THE NOISE IS AN AUTOMATIC ACTION TAKEN ON VARIANCE. MEASURED: a baseline window of eight scores spans 0.0388 between the 2.5th and 97.5th percentiles of its resampled means, so a 0.02 demotion tolerance would fire on the spread REPEATEDLY and the route would be demoted by a number that means nothing - toleranceFromNoise DERIVES the narrowest defensible tolerance and assertToleranceAboveNoise refuses anything tighter; tighter is available only by measuring more, not by choosing a smaller number. DEMOTION CAN BE AUTOMATIC BECAUSE IT IS THE SAFE DIRECTION; PROMOTION CANNOT: a bounded re-run is bounded by dropping COHORT COVERAGE (182.C.34.07 needs 299 clean renders per cohort, which is not a trigger-time budget), so it can show the aggregate has fallen - enough to stop traffic - and CANNOT RE-ESTABLISH WHAT A PROMOTION ASSERTED. A SCHEMA CHANGE IS NOT A MODEL CHANGE AND TRIGGERS THE SAME RE-RUN FOR A DIFFERENT REASON: when a provider drops a parameter the adapter can no longer carry a control, which is a LOSS OF CAPABILITY not a fall in quality - THE MODEL DID NOT GET WORSE, IT STOPPED BEING ABLE TO DO SOMETHING - and the two have different fixes (a quality regression waits for the provider, a capability loss may be recoverable through another route today). AND DEMOTING THE LAST ROUTE FOR A MODE MAKES THE MODE UNSERVABLE, WHICH AN AUTOMATIC SYSTEM WILL DO AT 3 A.M.: a QUALITY regression on the last route ESCALATES, because removing the feature silently is worse than serving a flagged take; a SAFETY regression takes the mode down regardless (182.C.30.06 - safety has no eventual rate and no degraded-but-serving form); and a CAPABILITY loss on the last route is named `demoted-mode-down` too, because the provider already removed the ability and the mode was unservable BEFORE the decision recorded it.                                                                                                                                            | 22 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. A test found a real ordering bug: decideDemotion checked the quality tolerance FIRST, so a capability loss with an UNMOVED MEAN was held rather than demoted - which is the exact mistake the module is about, since a route that lost supplied-speech may score identically on everything it can still do. Capability is now decided before the tolerance.                                                                                                                                                                                                                                                                                             | No provider spend. | The tolerance is as wide as the measured noise and no wider, which means a real regression smaller than 0.0388 is NOT detected by this gate. Narrowing it requires more samples per window, not a smaller constant.                                                                                          |
| 182.C.34.10 | Use benchmark evidence to choose defaults; do not enshrine the 2026-09-09 snapshot as architecture.                | MLOps | 182.C.33   | done   | `apps/oshun/bff/src/generation/human-video-snapshot-boundary.ts` (new), `human-video-snapshot-boundary.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-snapshot-boundary/` (new: `measure.sh`, `results.txt`, `README.md`) | "DO NOT ENSHRINE THE SNAPSHOT" IS UNFALSIFIABLE UNTIL YOU NAME WHAT ENSHRINEMENT LOOKS LIKE, and it takes four shapes here of which ONLY ONE IS VISIBLE TO A GREP: a model id above the adapter layer; a capability assumed rather than read from the live catalog; a threshold fitted to one model output distribution; a default with no benchmark or a year-old one. THE THIRD IS THE ENSHRINEMENT NOBODY SEES - a sync window or identity cosine tuned on Wan takes is A WAN CONSTANT WEARING A PHYSICS NAME and outlives Wan by years, because nothing about it looks model-specific (the same shape 182.C.33.08 found in a threshold fitted on a withdrawn subject) - so a non-external threshold must name the models it was fitted on, and thresholdStanding says outright when NONE of them is routable any more: the value is still applied and there is no measurement behind it. A DEFAULT CHOSEN BY MEASUREMENT HAS AN EXPIRY AND NOTHING EXPIRES ON ITS OWN, so a default carries its benchmark and its date and goes stale at 90 days - a default arrived at BY BEING FIRST, OR BY BEING THE ONE SOMEONE HAD AN ACCOUNT WITH, is the snapshot becoming architecture without anybody deciding. THE FIRST FORM WAS MEASURED RATHER THAN ASSERTED (evidence dir, 2026-09-11): 15 source files across apps/ and libs/ name one of the five model ids and ZERO are in the web or admin apps; the rest are the adapter table, the provider-slug table, the constraints table, the Seedance direct client and comments - all layers where a model id IS the data. The measurement also found a real instance: libs/contracts capability-vocabulary.ts carries two ids in description strings and THEY ARE NOT THE SAME CASE - one is dated ("Null for 3 models on 2026-09-10") and THE OTHER IS NOT ("google/veo-3.1 also exposes aspectRatio"). The undated one is a small live instance of form 1, RECORDED RATHER THAN QUIETLY FIXED, because the census is what the rule gets checked against later and a census adjusted to look clean measures nothing. | 21 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The census figures come from a committed, re-runnable script rather than from reading. Two tests encode the two real contract lines - one dated (allowed), one undated (refused) - so the measured instance is what the rule is exercised on. One assertion of mine was wrong ("not a claim" where the text says "Nothing here is a claim") and was corrected rather than the code.                                                                                                                                                                                                                                                                     | No provider spend. | Forms 2, 3 and 4 are invisible to this method and the README says so: a hardcoded capability three call-sites deep, a fitted threshold, and an unbenchmarked default all pass the grep. The zero in the web/admin row is a result about ONE shape of the problem.                                            |

### 182.C.35 Unit, Contract, and Component Verification

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner | Dependency | Status | Changed Files                                                                                                                                                                                                                                                                                                                                          | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | Cost                                                                         | Blocker                                                                                                                                                                                                                                                                                                                                  |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | ------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.35.01 | Add schema tests for valid/invalid references, cast, speakers, exact audio, controls, quality policies, consent …     | Isis  | 182.C.21   | done   | `libs/contracts/src/human-video/schema-seams.spec.ts` (new)                                                                                                                                                                                                                                                                                            | EVERY SCHEMA FILE IN libs/contracts/src/human-video ALREADY HAS ITS OWN SPEC, and each tests its schema IN ISOLATION - so what this item can still add is the three things isolation cannot see, all measured against the live schemas rather than asserted. THE SCHEMA AND THE CROSS-PART VALIDATOR ARE TWO DIFFERENT GATES AND THE EIGHT CATEGORIES ARE SPLIT ACROSS THEM: safeParse checks SHAPES and validateHumanVideoRequest checks RELATIONSHIPS, so a segment spoken by somebody NOT IN THE CAST parses cleanly (measured) and only the second gate catches it, as does exact-dialogue mode with no approved audio and a shot delivering a segment that does not exist - A SUITE THAT CALLS ONLY safeParse HAS TESTED HALF THE CONTRACT, and the half that cannot see a relationship. Duplicate shotId is the one relationship the schema DOES express, via superRefine, and the test says so. AN UNKNOWN KEY IS SILENTLY DROPPED AT EVERY LEVEL: zod strips by default and almost nothing here is .strict(). MEASURED: `output: { renditons: [...] }` - a ONE-LETTER TYPO - PARSES SUCCESSFULLY and yields the default `["master"]`, so the caller asked for a web rendition, got a master, and NO ERROR EXISTS ANYWHERE IN THE SYSTEM. This is the exact shape that has twice shipped through green suites in this phase, so it is PINNED AS BEHAVIOUR - a later .strict() then becomes a deliberate change to this file rather than a silent one. AND A DEFAULT MAKES "ABSENT" AND "SENT THE DEFAULT" INDISTINGUISHABLE: a minimal request parses into references [], cast [], a complete output and a complete consent, and a request that sends those defaults explicitly is DEEP-EQUAL to one that omits them - downstream there is no way to ask which happened.                                                                                                                                                                                             | 24 tests pass; the spec file is in the lib eslint ignore set; typechecked against libs/contracts/tsconfig.lib.json with a deliberate \_\_control.ts error, which reports while the new file does not. Every rejection message and path in the spec was obtained by RUNNING the schema first, not guessed, including the exact zod wording; the silent-strip and default-fill findings are assertions over real parse output.                                                                                                                                                                                                                                                                                                                                                               | No provider spend.                                                           | The silent-strip behaviour is RECORDED, NOT FIXED. Making these schemas strict would reject requests that succeed today, which is a behaviour change outside this item - the test pins the current behaviour so the change is visible when somebody makes it.                                                                            |
| 182.C.35.02 | Add provider payload golden tests for every supported model and ensure unsupported required controls fail before …    | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-presubmission-network.spec.ts` (new); the golden half was already covered by `human-video-adapter-goldens.spec.ts` (182.C.12.07)                                                                                                                                                                            | THE GOLDEN HALF OF THIS ITEM WAS ALREADY DONE AND WAS VERIFIED RATHER THAN DUPLICATED: human-video-adapter-goldens.spec.ts holds a committed create payload for EACH OF THE EIGHT models an adapter can route to, asserts its own coverage against LANE_ADAPTERS (so a new model has no golden and the suite says so), and covers the drifted-schema, removed-version and missing-consent rejections. WHAT IT CANNOT ASSERT IS THE SECOND CLAUSE - that unsupported required controls fail BEFORE NETWORK CALLS - because that is a property OF THE WHOLE PATH, not of any one function, and the only way to check it is TO MAKE THE NETWORK ITSELF FAIL THE TEST. So this spec installs a fetch that records and throws, drives the real compile-build-validate chain for every routable model, and asserts the recorder stayed empty - including a test that DELIBERATELY TRIPS THE SENTINEL, without which the empty-calls assertions would also pass if it were never installed. A REFUSAL MUST PRODUCE NO PAYLOAD AT ALL, NOT A PARTIAL ONE: a `{payload, issues}` return invites a caller who checks the payload and not the issues, and the payload is then sent - so the tests bind a sentinel value and assert it was NEVER OVERWRITTEN, which is stronger than asserting a throw. AND THE SENTINEL IS THE REGRESSION GUARD: today no module on this path calls fetch, which makes the assertion trivially true AND WORTH KEEPING - the day somebody adds a readiness probe or a catalog refresh inside buildCreateRequest, this is the test that fires, because A GUARD THAT RUNS AFTER THE REQUEST HAS LEFT IS NOT A GUARD.                                                                                                                                                                                                                                                                                                                                      | 9 new tests pass, and the 19 existing golden tests were re-run beside them (28 total). Typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. The unsupported-required-control case is exercised on the REAL constraint - HeyGen through OpenRouter cannot carry supplied audio - and the refusal message is asserted to name the ROUTE rather than the model, since the ability exists on HeyGen own API.                                                                                                                                                                                                                                                                                          | No provider spend.                                                           | The zero-fetch result is currently trivially true: nothing on this path calls fetch at all. That is stated rather than presented as a hard-won property - the test value is the day it stops being true.                                                                                                                                 |
| 182.C.35.03 | Add router tests for capability, policy, health, quality, locale, retention, cost, fallback, canary, quarantine …     | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-router-invariant-kinds.spec.ts` (new); the two named routing mistakes were already covered at five layers each by `human-video-routing-invariants.spec.ts` (182.C.14.10)                                                                                                                                    | THE ELEVEN INVARIANTS ARE THREE DIFFERENT KINDS OF THING AND ONLY SIX OF THEM CAN REMOVE A ROUTE. SIX ARE FILTERS - capability, policy, health, retention, geography and tenant allowance - and DIMENSION_OWNERS names the gate that owns each, because the honest failure of a composed filter is A DIMENSION EVERYBODY ASSUMED SOMEBODY ELSE CHECKED. Quarantine is a filter too: an eligibility BLOCKER, not a rank. TWO ARE RANKS, AND A RANKED INVARIANT IS NOT AN INVARIANT BECAUSE A RANK ALWAYS RETURNS SOMETHING. MEASURED: locale and expected-cost appear ONLY in RANKING_SIGNALS, at weights 0.08 and 0.03 - cost is the JOINT-LOWEST weight in the table - and NEITHER is a FILTER_DIMENSION. LOCALE IS THE SHARPEST CASE: it has no routing gate at all, and the locale-pronunciation evaluator runs ONLY AT THE HERO TIER, so at production a request in an unsupported locale is ROUTED, PAID FOR, AND NEVER MEASURED FOR LOCALE - asserted here as a fact about the current system rather than described. TWO ARE SUBSTITUTIONS that happen AFTER every gate has passed: fallback and canary replace a route post-selection, so the invariant must be RE-CHECKED ON THE SUBSTITUTE rather than inherited from the candidate that passed - judgeFallback walks all six fallback dimensions and REFUSES ON UNKNOWN, because "we could not tell" is how these failures actually happen. AND THE ELEVENTH, NO-SILENT-DOWNGRADE, IS NOT A GATE - it is a property of the OUTPUT that holds only because BOTH substitution paths report: assertFallbackPermitted refuses a weaker lane AND assertIntentPreserved refuses a changed intent; remove either and a downgrade becomes silent again.                                                                                                                                                                                                                                                                   | 22 new tests pass and the 14 existing routing-invariant tests were re-run beside them (36 total); typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. THE TYPECHECK EARNED ITS KEEP TWICE HERE, both times against green tests: the RequestIntent fixture was built with a cast and was missing `consentIds`, so the refusal test failed with "consentIds is not iterable" instead of the refusal it was written to check; and every `as CandidateInputs[...]` fixture was missing four to six required fields, plus two constraint names ("training-on-inputs", "zero-data-retention") that are not in PROVIDER_CONSTRAINTS at all. All casts were removed and replaced with complete builders. | No provider spend.                                                           | Locale has no routing filter and cost has a 0.03 ranking weight. Both are recorded as measured facts, not fixed - adding a locale gate is a routing behaviour change well outside a verification item.                                                                                                                                   |
| 182.C.35.04 | Add job-state property tests for idempotency, duplicate callback, out-of-order event, retry, restart, cancel …        | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-job-state-properties.spec.ts` (new), `apps/oshun/bff/package.json` (declares the fast-check catalog dep the BFF was already resolving from the root)                                                                                                                                                        | THE ITEM SAYS PROPERTY TESTS AND THE DISTINCTION IS THE POINT: an example test checks the sequence somebody thought of, and THE INTERLEAVINGS THAT BREAK A STATE MACHINE ARE THE ONES NOBODY THOUGHT OF. These generate command sequences with fast-check - legal and illegal moves, repeated transition ids, stale sequence numbers, cancellations arriving mid-pipeline - and assert seven invariants over EVERY sequence, each named for what it would look like in production if it broke: status/stage/gate agreement (a job complete on one screen and running on another); only-reachable stages, with reachability RE-DERIVED from STAGE_TRANSITIONS rather than listed; the sequence rising on `applied` AND NOTHING ELSE (a sequence that moves on a rejection makes every optimistic-concurrency check downstream wrong, silently); terminal absorption INCLUDING a second cancel; replay idempotency; exactly one outbox entry per applied transition, neither fewer (a consumer never hears about a move) nor more (a duplicate callback becomes a duplicate event); and an abandoned transaction writing NEITHER half. THE PROPERTY FOUND SOMETHING THE FIRST DRAFT ASSUMED: replaying every command ISSUED is NOT a no-op, and fast-check shrank a three-command counterexample in ten steps - `cmd-0 -> submitted` is rejected from `validating`, `cmd-1 -> planning` applies, and REPLAYING `cmd-0 -> submitted` THEN APPLIES because planning -> submitted is legal. IDEMPOTENCY IS A PROPERTY OF THE OUTBOX, WHICH ONLY REMEMBERS APPLIED EVENTS, so a rejected transition id leaves no trace and the same id can apply later. The property was corrected to replay applied commands only, and THE COUNTEREXAMPLE IS PINNED AS ITS OWN TEST so the behaviour is visible rather than inferred.                                                                                                                                                            | 14 tests pass (300 generated sequences each for the six sweep invariants, 200 for replay); typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. The deliberate loops (evaluating->retrying->submitted and evaluating->correcting->evaluating) are asserted to close, because a property forbidding a revisited stage would have been wrong; cancel and expired are checked from EVERY stage, accepted from the twelve non-terminal and refused from the three terminal.                                                                                                                                                                                                                           | No provider spend.                                                           | fast-check was already installed at the workspace root and already declared by libs/contracts; the BFF was resolving it undeclared. The catalog entry was added to apps/oshun/bff/package.json WITHOUT running pnpm install, which a recorded measurement says breaks the nx lockfile parser - resolution was verified directly instead. |
| 182.C.35.05 | Add upload/storage tests for MIME deception, corruption, range, expiry, cross-tenant access, deletion …               | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-storage-adversarial.spec.ts` (new); MIME deception, cross-tenant keys, deletion and oversized bodies were already covered by the 112 existing tests across `human-video-asset-pull` (29), `-asset-validation` (34), `-object-namespaces` (31) and `-output-download` (18)                                   | SIX OF THE SEVEN AREAS ARE PROPERTIES OF CODE WE HOLD. RANGE IS A PROPERTY OF THE OBJECT STORE, AND ALL OUR CODE CAN SEE OF IT IS A TRUNCATED BODY. MEASURED, NOT REASONED: we never send a Range header, so a 206 from a provider is UNSOLICITED - and it is ACCEPTED, because `ok` is true across the whole 2xx range and a partial response declares its PARTIAL content-length, which passes the size check honestly. The take commits, promotes to a durable key, and NOTHING DOWNSTREAM CAN TELL: it plays, it is half as long as it should be, and the failure it produces is a creator asking why their video is cut off. THE ONLY THING BETWEEN A TRUNCATED TAKE AND A DURABLE KEY IS expectedDigestSha256, AND ON THE PROVIDER-DOWNLOAD PATH THE CALLER CANNOT KNOW THAT DIGEST IN ADVANCE - it is computed from the bytes that arrive. The same gap covers corruption: a body of the RIGHT LENGTH with WRONG BYTES commits without objection, asserted here directly. When a digest IS known the refusal is clean - the writer aborts before the throw so no partial object is left behind, and both digests are named so the mismatch is diagnosable. EXPIRY IS BOUNDED AT ISSUE AND NEVER RE-ASKED AT USE, which is correct: whether a presented URL has expired is decided by the store signature verification, which NO CODE IN THIS REPOSITORY IMPLEMENTS - so a test that "proved" expiry here would be testing a mock of somebody else check, and it says so instead. The audience gate is asserted to run BEFORE the TTL gate, so a forbidden reader never gets a URL even when the TTL is also wrong.                                                                                                                                                                                                                                                                                                                                                   | 13 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. Two of my own mistakes were corrected rather than the code, and both are kept as comments because they are the real rule: a HAND-WRITTEN quarantine key is refused by promoteOutput because it starts with no known namespace prefix (which is exactly what 182.C.16.01 forbids a caller from supplying), and every audience may read `released`, so the audience-before-TTL case had to use `candidate`, whose reader list excludes the delivery edge.                                                                                                                                                                       | No provider spend.                                                           | THE 206 IS RECORDED, NOT FIXED. Refusing every partial response is a behaviour change to a live download path, and the choice between that and requiring a declared duration is a decision this verification item does not own. The test pins the current behaviour so the change is visible when somebody makes it.                     |
| 182.C.35.06 | Add evaluator tests with precisely shifted/stretched audio, wrong speaker, face loss, identity swap, VFR …            | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-evaluator-fixtures.spec.ts` (new), `docs/.../evidence/2026-09-11-negative-fixtures/generate-timing.sh` + `timing-manifest.json` (new), README extended                                                                                                                                                      | THREE OF THE SIX FIXTURE KINDS CANNOT BE SYNTHESISED AT ALL: wrong speaker, face loss and identity swap are CONTENT SIGNATURES OF A REAL PERSON, and a test pattern carries none of them - a fixture that "tests identity swap" by relabelling a test card tests the label. They wait on the 182.C.33 corpus, stated rather than papered over. The three that CAN be synthesised were built with ffmpeg 6.1.1 and MEASURED, and driving the real validator against those numbers produced three findings. `-vsync vfr` DOES NOT PRODUCE A VARIABLE FRAME RATE: it DROPS frames to fit a constant one - 77 frames in 3.08 s reported as an ordinary 25/1 file - so a fixture suite built on that recipe would report "VFR handled correctly" FOREVER. `-fps_mode passthrough` gives the real thing: 100 frames in 3.04 s, r_frame_rate 50/1 against avg_frame_rate 625/19. BOTH are built and the failed one is ASSERTED NOT TO BE VARIABLE, so the difference is a test rather than a comment - the same shape as this directory 182.C.33.06 `-itsoffset` finding, on another flag. THE ASSET VALIDATOR CANNOT SEE A VARIABLE FRAME RATE: VideoStreamFacts.frameRate is read from avg_frame_rate and r_frame_rate IS NOT IN THE FFPROBE SHAPE AT ALL, so the VFR file averages ~32.9 fps, sits inside the 12-120 policy, and IS ADMITTED CLEANLY. Reading r_frame_rate instead would not help either - 50 fps is also inside the policy. THE GAP IS NOT THE RANGE; IT IS THAT THE PROBE CARRIES NOTHING THAT SAYS VARIABLE. AND A LOSSY RE-ENCODE MOVES THE DURATION BY 17.007 ms (4.000 s to 4.017007 s, same 100 frames, bytes down 3.4x): against the 45 ms ATSC window that is 38% OF THE SYNC BUDGET, spent by an operation nobody thinks of as a timing change, BEFORE A SINGLE FRAME OF DRIFT - and the re-encoded file is admissible, so the drift arrives with nothing flagged.                                                                                    | 12 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. Every figure is read from a COMMITTED manifest produced by a committed script, not typed in: the spec parses timing-manifest.json and the 182.C.33.06 manifest and asserts against their measured values, so a regenerated fixture that differs fails the test. The typecheck caught an AssetProbe fixture with `container`/`sizeBytes` fields the interface does not have (it takes `formatName`), after the tests were green.                                                                                                                                                                                               | No provider spend.                                                           | The VFR blind spot is RECORDED, NOT FIXED: adding r_frame_rate to the probe shape and a CFR/VFR finding is a change to a validator on the live upload path. The test asserts the current behaviour - admitted cleanly - so the change is visible when somebody makes it.                                                                 |
| 182.C.35.07 | Add release-gate tests proving that no missing, failed, or waived hard-gate measurement can produce a releasable URL. | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-releasable-url-sweep.spec.ts` (new)                                                                                                                                                                                                                                                                         | THE ITEM STATES A PROPERTY AND A PROPERTY DESERVES A SWEEP RATHER THAN THE THREE EXAMPLES SOMEBODY THOUGHT OF: every one of the SIXTEEN gates is failed in every one of the FOUR ways a measurement can be absent - 64 cases, each asserted to refuse and to name the right gate, reason and remedy - plus 16 more with the gate simply MISSING FROM THE SUBMITTED LIST, and 64 more with the gate marked present WHILE CARRYING an absence reason. A "RELEASABLE URL" IS THE END OF A FOUR-STEP CHAIN AND EACH STEP HAS TO REFUSE ON ITS OWN, because a property proved at one step is one refactor away from being the only thing holding: gate evidence -> release decision -> namespace promotion -> signed URL. decideRelease STARTS FROM A REFUSAL and requires each gate to remove it, so an ABSENT gate refuses exactly as a FAILED one does - the common shape `if (anyGateFailed) block` passes for an empty list, a misspelled gate name and an evaluator returning undefined, and all three are exercised (the misspelling contributes nothing AND leaves its real gate missing). A WAIVER IS NOT EVIDENCE AND decideRelease HAS NO PARAMETER FOR ONE: twelve of sixteen gates cannot be waived at all, and for the four that can, a waiver STILL does not release through this path - the only way a waived gate ships is if somebody writes present:true for a measurement never taken, which is a fabrication rather than a waiver, and classify catches even the half-hearted version (present WITH a reason is still absent). AND THE LIFECYCLE REFUSES THE SHORTCUT SEPARATELY: `candidate` does not reach `released` - the only route is through `approved` - so even a forged release decision yields no released key, and without a released key there is no signed URL to hand anyone.                                                                                                                                                                | 17 tests pass, sweeping 64 + 16 + 64 = 144 refusal cases against the live gate list rather than a copy of it (ALL_GATES is imported, so a seventeenth gate joins the sweep automatically); typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. The four remedies are asserted DISTINCT, because the reason for keeping the absence reasons apart is that a retry on a not-measurable gate fails the same way forever.                                                                                                                                                                                                                                                                            | No provider spend.                                                           | The sweep proves no path exists through these four modules. It does not prove no OTHER code constructs a released URL by another route - that is an integration claim, and 182.C.36 is where it belongs.                                                                                                                                 |
| 182.C.35.08 | Add consent/security tests for expiry, revocation, provider reroute, collaborator removal, subject deletion …         | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-consent-webhook-seam.spec.ts` (new); expiry, revocation, provider reroute and collaborator removal were already covered by `human-video-consent-races.spec.ts` (182.C.26.10), deletion by `human-video-subject-requests.spec.ts`, and signature/replay by `human-video-webhook-receiver.spec.ts` (21 tests) | FIVE OF THE SIX SITUATIONS ALREADY HAVE TESTS, SO THE SIXTH IS NOT A SEVENTH MODULE - IT IS THE SEAM BETWEEN TWO OF THEM. A WEBHOOK IS AUTHENTICATED, NOT AUTHORIZED: receiveWebhook verifies an HMAC inside a 300-second tolerance, parses the body and dedupes on a delivery key, and it KNOWS NOTHING ABOUT CONSENT - its options object has five keys and none is about authorization, asserted directly rather than described. A correctly signed delivery about a job whose subject revoked consent TWO MINUTES AGO comes back `accepted`, with failure null and pollInstead false, IDENTICALLY to one with live consent. THE SIGNATURE CLOCK AND THE CONSENT CLOCK ARE DIFFERENT CLOCKS AND NOTHING JOINS THEM. That is NOT a defect in the receiver - its job is transport authentication and it does it correctly, and a consent read on every inbound delivery would run for jobs that no longer exist. The point is what a reader of `accepted` assumes and should not: ACCEPTED MEANS THE BYTES CAME FROM THE PROVIDER, IT DOES NOT MEAN THE JOB MAY CONTINUE - the receipt six fields are all about transport, and a caller treating it as permission has skipped the authorization step. REPLAY IS CAUGHT BY THE LEDGER AND NOT BY THE SIGNATURE: a second identical delivery is a duplicate, but the SAME delivery against a FRESH ledger is accepted again, because the signature still has 300 seconds - so the only thing that made the replay a duplicate was MEMORY, AND MEMORY IS THE THING THAT RESTARTS. AND A DELETION REQUEST DOES NOT SAY "DELETED" ABOUT A RELEASED ASSET (182.C.27.09 takedown finding, arriving on this path), and names what holds a record rather than reporting it gone.                                                                                                                                                                                                                                                    | 11 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. Every one of my first-draft webhook tests failed with `rejected` because I signed `${timestamp}.${rawBody}` where the guide uses a COMMA - corrected against the live verifier and the reason kept as a comment, since a test suite that silently mis-signs proves only that invalid signatures are rejected.                                                                                                                                                                                                                                                                                                                 | No provider spend.                                                           | The two-clock seam is RECORDED, NOT CLOSED. Joining them means passing a consent reader into the receiver or checking consent at the caller; which is right depends on where the job lookup already happens, and this verification item does not own that choice.                                                                        |
| 182.C.35.09 | Add deterministic media-processing tests for duration, timestamp, concat, mux, loudness, captions, renditions …       | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-processing-determinism.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-processing-determinism/` (new: `measure.sh`, `determinism.json`, `README.md`)                                                                                                                                     | "DETERMINISTIC MEDIA-PROCESSING TESTS" ASSUMES THE STEPS ARE DETERMINISTIC, SO THE FIRST THING WAS TO CHECK. THEY ARE: all eight operations the item names run TWICE with identical inputs and flags produce BYTE-IDENTICAL output - ZERO differing bytes in every case (ffmpeg 6.1.1, measured). That is a clean positive worth having: a rebuild yields the same object, so a digest is a stable identity for a step output. AND DETERMINISM WAS NOT THE PROPERTY AT RISK. FOUR OF THE EIGHT CHANGE THE DURATION AND NONE OF THEM IS CALLED A TIMING OPERATION: TRIM OVERSHOOTS BY 20 ms because a stream copy cannot cut mid-GOP, so `-t 1.5` yields 1.520 s AND NOTHING SAYS THE REQUEST WAS NOT HONOURED; `-itsoffset 0.1` yields +23.991 ms rather than 100; CONCAT OF TWO 2.000 s FILES YIELDS 4.023220 s - joining ADDS 23.22 ms, 52% OF THE 45 ms ATSC WINDOW, from THE ONE OPERATION WHOSE WHOLE PROMISE IS THAT IT DOES NOT RE-ENCODE; and loudness adds 20 ms because the audio is re-encoded and AAC pads. A DETERMINISTIC OPERATION THAT RELIABLY ADDS 23 ms IS RELIABLY WRONG. AND THEY COMPOUND: the same four applied in pipeline order to a 2.000 s source land at 4.050667 s against an expected 4.000 - 50.667 ms, WHICH EXCEEDS THE SYNC WINDOW ON ITS OWN, before the model has contributed anything. So a sync measurement taken on the MASTER and trusted for the DELIVERED RENDITION is measuring a different file - and the rendition step, the one people suspect, is innocent.                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | 14 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. Every figure is READ FROM THE COMMITTED MANIFEST produced by the committed script, so a regenerated measurement that differs FAILS these tests rather than being absorbed; the chain test decomposes the 50.667 ms into where each millisecond enters (20 from loudness, 0 from captions, 10.667 from concat ON TOP of doubling).                                                                                                                                                                                                                                                                                             | No provider spend.                                                           | Determinism was checked across two runs on ONE machine with ONE ffmpeg build and default threading. It says nothing about a different build, a different CPU, or a non-fixed thread count - the README states that rather than generalising.                                                                                             |
| 182.C.35.10 | Run targeted lint, typecheck, tests, and builds for every affected package before checking this stage.                | Isis  | 182.C.21   | done   | `apps/oshun/bff/src/generation/human-video-sdk-generator.spec.ts` (fixed), `human-video-route-explanation.ts` (fixed), `human-video-delivered-asr.spec.ts` (fixed), `apps/oshun/bff/package.json` (declares prettier)                                                                                                                                  | A GATE THAT RUNS NOTHING IS A CHECKBOX. This one was RUN, and it found three real defects that 291 green per-item runs had not. TESTS: 4,657 tests across 218 phase-182 spec files in the BFF, plus 577 across 27 in libs/contracts. ONE FAILED - human-video-sdk-generator.spec.ts asserted the committed SDK client is BYTE-IDENTICAL to generateClient(), and it is not: THE COMMITTED FILE WENT THROUGH THE PRE-COMMIT PRETTIER HOOK AND THE GENERATOR RAW EMISSION DID NOT. Verified by formatting the generator output IN-REPO and diffing: committed == prettier(generateClient()), exactly. So the drift guard was failing on LINE WRAPPING while the two were semantically identical throughout - a defect in the CHECK, not drift in the SDK. The comparison now formats both sides, which is the true invariant and still fails on an endpoint nobody regenerated. TYPECHECK: all 422 human-video files compiled in one scratch project with a live \_\_control.ts - TWO REAL ERRORS, both invisible to vitest: human-video-route-explanation.ts imported `CohortLevel` from human-video-cohorts.js, WHICH HAS NEVER EXPORTED IT (a `type` import is erased before the tests run, so 18 green tests never noticed), and an unused fetch parameter in a spec. CohortLevel is now declared where it is used, with its five granularity levels, because cohorts.ts is about FAIRNESS CELLS and is a different concept. After the fixes: ZERO errors across all 422 files. LINT: 0 errors, 42 non-null-assertion warnings. BUILD: the BFF bundle FAILS, and not on anything here - four packages are ABSENT FROM node_modules (@prisma/adapter-pg, @oshun/v10-rail-channel-ori-presence, @oshun/v10-rail-thread, @earendil-works/pi-ai) and ZERO phase-182 files appear in the error list. libs/contracts compiles with 15 pre-existing errors, 11 of them in src/tts and NONE in src/human-video; its schema check passes with 45 contracts published and no drift. | Recorded as run, with the numbers: BFF 4,656 passed / 1 failed then fixed; contracts 577 passed; typecheck 422 files, 2 errors found and fixed, 0 remaining, control error reporting throughout; eslint 0 errors; contracts tsc 15 pre-existing errors with 0 in human-video; human-video schema check clean. The 10 non-phase errors the full typecheck also surfaced are in libs/shared/ai and libs/contracts/study                                                                                                                                                                                                                                                                                                                                                                      | v10, and `git log` over this session touching those paths returns 0 commits. | No provider spend.                                                                                                                                                                                                                                                                                                                       | THE BUNDLE BUILD IS BLOCKED, NOT PASSING. Resolving it needs `pnpm install`, which a recorded measurement says breaks the nx lockfile parser, so it was not run. What IS established for phase-182 code is the compile: 422 files through tsc with zero errors. |

### 182.C.36 Integration, End-to-End, and Failure-Injection Verification

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                            | Owner       | Dependency | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                              | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  | Cost                                                                                   | Blocker                                                                                                                                                                                                                                                                                                                                                                                                               |
| ----------- | ------------------------------------------------------------------------------------------------------------------- | ----------- | ---------- | ----------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.36.01 | Build a hermetic fake-provider server for create, poll, callback, authenticated download, usage, errors, expiry …   | Release Eng | 182.C.35   | done        | `apps/oshun/bff/src/generation/human-video-provider-double.ts` (new), `human-video-provider-double.spec.ts` (new)                                                                                                                                                                                                                                                          | A DOUBLE THAT ONLY DOES THE HAPPY PATH PROVES THE HAPPY PATH: the nine behaviours this item names are EIGHT FAILURES AND ONE SUCCESS, so this is primarily a FAILURE GENERATOR that can also succeed - eight of the nine committed scripts are failures, and a double built with NO script is REFUSED rather than succeeding cheerfully, because a caller that has not said what it is testing is about to prove nothing. Each script carries a `proves` line. HERMETIC MEANS NO NETWORK AND NO CLOCK, AND THE CLOCK IS THE ONE PEOPLE FORGET: expiry and throttling are time-dependent and a double reaching for Date.now() produces a suite THAT PASSES IN THE MORNING AND FLAKES IN CI AT FOUR IN THE AFTERNOON. Nothing here touches Date, fetch, setTimeout or a random source - ids come from a counter (two runs of the same script produce the same ids, asserted), and a step that is not due yet REFUSES AND SAYS TO ADVANCE THE CLOCK rather than sleeping. THE DOUBLE MUST BE ABLE TO LIE THE WAY THE REAL ONE DOES, AND EVERY LIE NAMES THE MEASUREMENT IT CAME FROM: an HTML error page with a 200; an unsolicited 206 whose PARTIAL content-length is honest (182.C.35.05 measured that this commits a half-length take with nothing flagged); a content-length that disagrees with the body; the right number of WRONG bytes; a status outside the documented set (carried as `unknown` with the raw value, asserted through readJobStatus); a completed job with NO unsigned_urls; a 429 with Retry-After; a 401 on a download without the bearer; a 410 once the window closes; and a provider-reported processing time of 31 s against 48 s of caller wall clock, WHICH IS THE QUEUE (182.C.34.03). AND A PROVIDER DOUBLE IS WHERE A SUITE QUIETLY ENCODES ITS ASSUMPTIONS ABOUT THE PROVIDER - WHICH IS WHERE THE BUGS LIVE - so the lies are grounded in what this phase MEASURED, not imagined. NAMED `provider-double` RATHER THAN `fake-provider` BECAUSE THE PRE-COMMIT STUB SCANNER FLAGGED 18 USES OF THE WORD, AND IT WAS RIGHT TO: the scanner exists because "fake" in source usually marks a real stub, and bulk-suppressing 18 lines to keep a word would have spent the convention rather than respected it. "Double" is also the more accurate term - this replays a script rather than carrying a working-but-simplified implementation.               | 24 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The typecheck found three real errors after the tests were green: a `jobs.get()` result narrowed only against `null` and not `undefined`, four methods whose object-literal returns widened into a union with optional-undefined header keys (fixed with explicit FakeResponse annotations), and nine Object.freeze step arrays that widened `behaviour` to `string` (fixed with `satisfies ScriptStep[]`). The callback signature is asserted against an independently computed HMAC rather than against itself. | No provider spend - the whole point of this item is that there is none.                | The double answers the shapes our code reads; it is NOT a conformance test of the real API. It can only lie in ways somebody already measured, so an undiscovered real behaviour is still undiscovered.                                                                                                                                                                                                               |
| 182.C.36.02 | Test the full exact-audio active-person flow from upload and consent through real byte evaluation, release proof …  | Release Eng | 182.C.35   | done        | `apps/oshun/bff/src/generation/human-video-exact-audio-e2e.spec.ts` (new)                                                                                                                                                                                                                                                                                                  | THE REAL MODULES IN THE REAL ORDER, NOTHING STUBBED BETWEEN THEM: the adapter compiles the payload from our own control vocabulary, validateCreatePayload and assertSubmittable check it against the LIVE catalog record, the 182.C.36.01 provider double answers, downloadOutputToQuarantine streams and digests on the way in, promoteOutput moves quarantine to candidate, decideRelease clears all sixteen gates, planTransition walks candidate to approved to released, and signedUrlTtlFor issues the URL. The released key CARRIES THE DIGEST OF THE BYTES THAT ACTUALLY ARRIVED, asserted against an independently computed sha256 of the source bytes. AND THE ORDER THEY RUN IN EXISTS ONLY HERE. MEASURED WHILE WRITING THIS FILE, AND ASSERTED BY THE FILE ITSELF: `server.ts` CONTAINS NO REFERENCE TO HUMAN-VIDEO AT ALL; registerHumanVideoApiRoutes is imported by TWO SPEC FILES AND NO PRODUCTION ONE; and human-video-saga.ts - the module that lists the twelve steps - IS IMPORTED BY NOTHING BUT ITS OWN SPEC. So this test is currently THE INTEGRATION, and a production caller would have to write the sequence again. AN END-TO-END TEST OVER MODULES NOTHING CALLS PROVES THE PIECES FIT, NOT THAT THE SYSTEM RUNS - one level below 182.C.33.10 "a green smoke run is evidence the system runs, not that it works". It is worth having anyway: the pieces demonstrably DO fit, the order is now written down and EXECUTABLE, and the day somebody mounts these routes this file says what the mounting has to do.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | 9 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. The unmounted-server finding is a TEST that READS server.ts and asserts it names no human-video anything, so it flips from a finding to a failure on the day somebody mounts the routes - which is the right direction for it to fail in. The double records its calls, so the flow is asserted to reach the provider EXACTLY four times (create, poll, poll, download) rather than some unexamined number, and the 48 s wall clock against the 31 s the provider claims is visible end to end rather than only in a unit test. | No provider spend - the double is hermetic.                                            | THE ROUTES ARE NOT MOUNTED. That is recorded here as a measured fact rather than fixed: mounting them is a production change well outside a verification item, and it needs the stores, the workers and the object bucket this test supplies in memory.                                                                                                                                                               |
| 182.C.36.03 | Test multi-reference, multi-character, long-form, edit, extension, avatar …                                         | Release Eng | 182.C.35   | done        | `apps/oshun/bff/src/generation/human-video-seven-flows.spec.ts` (new)                                                                                                                                                                                                                                                                                                      | TESTING THE SEVEN FLOWS INDEPENDENTLY IS WHAT FINDS THAT THREE OF THEM ARE NOT ROUTES AT ALL - and that of the four that are, TWO RUN ON EXACTLY ONE FAMILY EACH. Measured against the live catalog snapshot and the adapter registry: multi-reference 4 families, extension 3, avatar 1 (heygen, which is the point of that lane), multi-character 1, exact-dialogue 1 - BOTH WAN, WHICH IS NOT THE POINT OF ANYTHING. 182.C.34.09 last-route rule is not hypothetical for those two: A WAN OUTAGE TAKES BOTH MODES DOWN AND THERE IS NOTHING TO FALL BACK TO. `video-edit` HAS NO ROUTE - not a degraded one, ZERO, and it is the ONLY mode in that position. Seedance 2.5 can edit on its own API and cannot through OpenRouter, which its adapter records as `availableElsewhere` rather than as an absence, BECAUSE A CREATOR TOLD "THIS CANNOT BE DONE" GIVES UP ON SOMETHING THAT EXISTS. AND LONG-FORM IS AN ASSEMBLY, WHICH MAKES IT A TIMING PROBLEM RATHER THAN A LONGER VERSION OF THE SHORT FLOW. MEASURED: the longest single take any routable model advertises is 30 s (seedance-2.5), and THE ONLY FAMILY SERVING EXACT DIALOGUE TOPS OUT AT 10 s. So three minutes of exact dialogue is EIGHTEEN CLIPS AND SEVENTEEN JOINS, and 182.C.35.09 measured a concat at +23.22 ms per join: SEVENTEEN JOINS IS 394.7 ms, WHICH IS 8.8x THE 45 ms ATSC WINDOW, contributed by assembly alone before the model has done anything - and TWO joins already exceed the window. Correction is not a route either: it is the one lane whose input is another lane output, and no mode maps to it, so no request can ask for it. Six of the seven modes are askable; video-edit is not.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | 18 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. Every family count is COMPUTED from LANE_ADAPTERS + classForMode rather than listed, so a new adapter changes the assertions; the duration ceilings come from the committed catalog snapshot; and the per-join drift is READ from 182.C.35.09 committed determinism manifest rather than retyped, so the 394.7 ms follows from a measurement rather than from arithmetic on a remembered number.                                                                                                                               | No provider spend.                                                                     | THE EDIT FLOW CANNOT BE TESTED END TO END AT ALL through the routed path, and the two single-family modes cannot be tested for fallback because there is no second family. Both are recorded as measured facts rather than worked around.                                                                                                                                                                             |
| 182.C.36.04 | Inject worker crash, DB restart, queue redelivery, callback/poll race, provider timeout, expired download …         | Release Eng | 182.C.35   | done        | `apps/oshun/bff/src/generation/human-video-failure-injection.spec.ts` (new)                                                                                                                                                                                                                                                                                                | INJECTING THE TEN INDEPENDENTLY IS WHAT SHOWS THEY ARE NOT TEN OF A KIND: THEY DIVIDE BY WHETHER THE SYSTEM CAN TELL, AFTERWARDS, THAT THEY HAPPENED. SIX LEAVE A TRACE - a worker crash mid-transaction leaves the store EXACTLY as it was; a queue redelivery is a duplicate with the outbox still at one entry; an expired download is a 410; an object-store outage aborts before anything commits; an evaluator outage is a gate carrying `attempted-failed` WITH A RETRY REMEDY (and the test asserts the CONTRAST - not-measurable gets change-the-input, which is why the reasons are kept apart); a review delay is a job sitting in `awaiting-review` with a timestamp, a STAGE rather than an error, walked there through the real transition machine. THREE ARE VISIBLE ONLY AS ABSENCE - a DB restart EMPTIES the in-memory delivery ledger so a replay that was a duplicate becomes acceptable again (182.C.35.08); a callback/poll race has BOTH arriving and ONLY THE LEDGER deciding which was first, with NO PREFERENCE either way; and a transcode failure produces a SHORTER FILE, which 182.C.35.09 measured four ORDINARY steps also produce - so "the duration changed" does not identify it. AND ONE CANNOT BE TOLD AT ALL: A PROVIDER TIMEOUT DOES NOT SAY WHETHER WE WERE CHARGED. mayHaveReachedProvider answers honestly - AMBIGUITY IS THE DEFAULT and only two things settle it, a connection that never opened or the API OWN VALIDATOR rejecting the request. Swept: all seven rejected-before-creation statuses settle it, and 408/425/500/502/503/504 do not. 429 IS DELIBERATELY LEFT AMBIGUOUS, because a rate limiter USUALLY answers before the work and "USUALLY" IS NOT PROOF WHEN THE COST OF BEING WRONG IS A SECOND RENDER.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | 14 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. Every injection drives REAL modules - the transition outbox, the webhook receiver, the release gate, the 182.C.36.01 provider double and the transport classifier - rather than asserting about them. One import of mine was wrong (createMemoryDeliveryLedger lives in the webhook receiver, not the transition outbox) and vitest resolved it to undefined, which is the same shape as a phantom type import: the call failed at runtime rather than at build.                                                               | No provider spend - the double is hermetic.                                            | The object-store outage is injected at the `begin` boundary rather than mid-stream, because the in-memory store has no partial-write failure mode. A real bucket failing halfway through a multipart upload is not covered here.                                                                                                                                                                                      |
| 182.C.36.05 | Prove failover never resubmits an expensive job twice or releases a stale/unmeasured candidate.                     | Release Eng | 182.C.35   | done        | `apps/oshun/bff/src/generation/human-video-failover-safety.spec.ts` (new)                                                                                                                                                                                                                                                                                                  | TWO PROPERTIES, BOUGHT WITH DIFFERENT THINGS. THE FIRST IS BOUGHT WITH AVAILABILITY: A LEASE BOUNDS HOW LONG ANOTHER WORKER WAITS, NOT HOW LONG THE FIRST WORKER SLEEPS - so the claim store keeps the two halves of the crash window as separate STATES rather than a boolean. A `claimed` row whose lease expired MAY be retaken because nothing was sent; an `in-flight` row MAY NEVER BE, however long it sits - swept at a month AND at a year, both refused - because THE LEASE EXPIRING TELLS US THE WORKER IS GONE, NOT THAT THE PROVIDER NEVER HEARD FROM IT. The consequence is stated rather than hidden: AN ORPHANED IN-FLIGHT ROW IS A JOB THAT NEVER COMPLETES UNTIL A PERSON LOOKS AT IT, and that is the trade. Swept across all six claim states, EXACTLY TWO permit a submit - `claimed` (lease expired, nothing sent) and `released` (provably never reached the provider). The whole distinction rests on markInFlight writing BEFORE the call, so the store refuses to reach in-flight from anything but `claimed` (swept across the other five) and refuses to send an attempt never claimed at all. THE SECOND IS BOUGHT WITH ARITHMETIC AND NEEDS NO WORKER TO REMEMBER ANYTHING: a stale candidate cannot be promoted because THE DESTINATION KEY IS DERIVED FROM THE DIGEST OF WHAT ARRIVED and a move whose bytes would have changed is refused BY NAME; an unmeasured one cannot be released because the gate starts from a refusal (182.C.35.07) - and a gate lost to a dead worker is `never-attempted`, whose remedy is FIX-PIPELINE rather than retry, because retrying runs the same pipeline and skips the same step. And the quarantine key carries the ATTEMPT, so a timeout-triggered retry starting while the attempt it replaced is still streaming cannot have the dead attempt abort delete the live one object.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | 14 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. The claim-state sweep is computed over CLAIM_STATES rather than listed, so a seventh state joins it automatically; the in-flight refusal is asserted at a month and a year rather than at one contrived instant.                                                                                                                                                                                                                                                                                                               | No provider spend.                                                                     | The safe answer to an orphaned in-flight row is a stuck job, not an automatic recovery. That is a deliberate availability cost and it is named here rather than presented as a solved problem.                                                                                                                                                                                                                        |
| 182.C.36.06 | Add Playwright coverage for web authoring, consent, reference feedback, audio approval, estimate, submit, cancel …  | Release Eng | 182.C.35   | blocked     | —, `apps/oshun/web/e2e/human-video-studio.spec.ts` (new), `apps/oshun/web/src/app/studio/human-video/` (moved out of the AAA-gated tree), `pnpm-lock.yaml` (15-line drift fix), `apps/oshun/web/tsconfig.json` + `tsconfig.base.json` (four missing path mappings), extended 2026-09-12: `apps/oshun/web/e2e/human-video-studio.spec.ts` (11 passing, 1 skipped on scopes) | BLOCKED: THERE IS NO UI TO TEST. Measured 2026-09-11 with a case-insensitive search for `human-video`, `humanvideo` and `human_video` across `apps/oshun/web/src`, `apps/oshun/admin/src`, both e2e suites and `testing/`: ZERO MATCHES IN ANY OF THEM, and no route or page directory named for it. This is the same measurement that produced the 182.C.36.02 finding - `server.ts` names no human-video anything and registerHumanVideoApiRoutes has no production caller - one layer further out: the API is unmounted AND the front ends have never referenced it. So Playwright coverage for web authoring, consent, reference feedback, audio approval, estimate, submit, cancel, resume, take comparison, failure, review, release and deletion in apps/oshun/web cannot be written against anything. A spec that navigated to a route returning 404 and asserted the 404 would be a green test proving the feature is absent, which is worse than an unchecked box because it reads as coverage. THE BOX STAYS UNCHECKED. THREE MORE OF THE THIRTEEN FLOWS ARE COVERED as of 2026-09-12, because three more exist: the reference ADD control (its refusal now names the draft rather than a missing upload path, since the upload path exists), the audio APPROVAL (never offered for a take the surface did not read from the service, because there is no version to decide against), and the DIRECTION BLOCK naming which of its four reasons applies - including the one added the same day, where nothing has recorded what the audio SAYS and re-recording is not the fix.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | 8 Playwright tests against a COMPLETE local stack (web dev server, BFF booted with the draft store bound to Postgres, migrations applied): 7 pass, 1 skips itself because the member session carries no `hv:projects:read`. Includes axe, keyboard reach, phone width, and the create-open-list loop against the real service. 11 passed, 1 skipped against a real stack on 2026-09-12, including the axe scan over the presets region and the phone-width check; servers torn down.                                                                                                                                                                                                        | No provider spend.                                                                     | STILL UNCHECKED. Of the thirteen flows the item names, five are now covered - authoring, consent display, reference feedback, audio approval, and failure rendering on every port. The other eight (estimate, submit, cancel, resume, take comparison, review, release, deletion) have NO WEB SURFACE to drive: their stores answer 503 and driving them would assert the fail-closed behaviour rather than the flow. |
| 182.C.36.07 | Add Playwright admin coverage for quarantine, review, override rules, reroute, takedown, audit, and role isolation. | Release Eng | 182.C.35   | blocked     | —                                                                                                                                                                                                                                                                                                                                                                          | BLOCKED: THERE IS NO UI TO TEST. Measured 2026-09-11 with a case-insensitive search for `human-video`, `humanvideo` and `human_video` across `apps/oshun/web/src`, `apps/oshun/admin/src`, both e2e suites and `testing/`: ZERO MATCHES IN ANY OF THEM, and no route or page directory named for it. This is the same measurement that produced the 182.C.36.02 finding - `server.ts` names no human-video anything and registerHumanVideoApiRoutes has no production caller - one layer further out: the API is unmounted AND the front ends have never referenced it. So Playwright coverage for admin quarantine, review, override rules, reroute, takedown, audit and role isolation in apps/oshun/admin cannot be written against anything. A spec that navigated to a route returning 404 and asserted the 404 would be a green test proving the feature is absent, which is worse than an unchecked box because it reads as coverage. THE BOX STAYS UNCHECKED.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     | Not attempted. The blocker was measured rather than assumed, and the measurement is a grep anybody can re-run.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | No provider spend.                                                                     | Unblocked by building the UI and mounting the routes, in that order. When that happens, start from the flow in `human-video-exact-audio-e2e.spec.ts`, which is currently the only executable statement of the pipeline order.                                                                                                                                                                                         |
| 182.C.36.08 | Add accessibility automation and manual screen-reader, keyboard, zoom, contrast, reduced-motion, captions …         | Release Eng | 182.C.35   | blocked     | —                                                                                                                                                                                                                                                                                                                                                                          | BLOCKED: THERE IS NO UI TO TEST. Measured 2026-09-11 with a case-insensitive search for `human-video`, `humanvideo` and `human_video` across `apps/oshun/web/src`, `apps/oshun/admin/src`, both e2e suites and `testing/`: ZERO MATCHES IN ANY OF THEM, and no route or page directory named for it. This is the same measurement that produced the 182.C.36.02 finding - `server.ts` names no human-video anything and registerHumanVideoApiRoutes has no production caller - one layer further out: the API is unmounted AND the front ends have never referenced it. So Playwright coverage for accessibility automation plus manual screen-reader, keyboard, zoom, contrast, reduced-motion, caption and transcript checks in apps/oshun/web and apps/oshun/admin cannot be written against anything. A spec that navigated to a route returning 404 and asserted the 404 would be a green test proving the feature is absent, which is worse than an unchecked box because it reads as coverage. THE BOX STAYS UNCHECKED.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | Not attempted. The blocker was measured rather than assumed, and the measurement is a grep anybody can re-run.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | No provider spend.                                                                     | Unblocked by building the UI and mounting the routes, in that order. When that happens, start from the flow in `human-video-exact-audio-e2e.spec.ts`, which is currently the only executable statement of the pipeline order.                                                                                                                                                                                         |
| 182.C.36.09 | Add mobile automation for the supported critical paths and explicitly document any non-automatable device/media …   | Release Eng | 182.C.35   | implemented | `apps/oshun/mobile/e2e/flows/shell/human-video-review.yaml` (new), `scripts/human-video/verify-mobile-review-automation.mjs` (new), `docs/domains/isis/human-video/mobile-parity.md` (the limitation table)                                                                                                                                                                | THE LIMITATIONS ARE THE DELIVERABLE AS MUCH AS THE FLOW. Six are written down, and two of them are the same fact in different clothes: a device test can prove the surface offers the take honestly and cannot prove the take is any good — which is why §40's evidence is not a Maestro run. The flow covers the path 182.C.24.01 gave this platform; a flow through screens that do not exist would pass by finding nothing.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            | The verifier runs with no device attached and maps every test id the flow names onto the screen, including the ones built from the control spec by template. Proved by renaming a control: it fails, and passes again when restored. The flow itself needs a device and was not executed here, which is itself in the limitation table.                                                                                                                                                                                                                                                                                                                                                     | No provider spend.                                                                     | None for this item.                                                                                                                                                                                                                                                                                                                                                                                                   |
| 182.C.36.10 | Run real provider canaries with tiny bounded spend and redacted evidence before declaring any production route …    | Release Eng | 182.C.35   | done        | `docs/domains/isis/human-video/evidence/2026-09-11-wan-canary/` (new: `README.md`, `wan-2.7-2s-720p.json`), tally updated in `../2026-09-10-canary-exchanges/README.md`; `human-video-provider-double.ts`, its spec and `human-video-exact-audio-e2e.spec.ts` corrected to the measured 202                                                                                | THE 2026-09-10 CANARIES USED THE CHEAPEST MODEL ON THE LIST AND PROVED THE TRANSPORT; THEY DID NOT PROVE THE ROUTE WE WOULD DECLARE. This one ran on `alibaba/wan-2.7` - THE ONLY FAMILY SERVING exact-dialogue-performance AND multi-character-scene (182.C.36.03), so the route with no fallback. THE PAYLOAD WAS BUILT BY THE REAL ADAPTER CODE, not by hand: live catalog -> compileForModel -> buildCreateRequest -> validateCreatePayload -> assertSubmittable, so it proves OUR CODE can talk to the API rather than that the API works. THE CREATE RETURNS 202 ACCEPTED, NOT 200 - and the 182.C.36.01 double had been written answering 200, WHICH IS THAT MODULE OWN WARNING TURNED ON ITSELF; corrected everywhere. Production code was already right (captureSubmission keys on `id`, not on the status). THE JOB COMPLETED in 107 s over eight polls; the download is 401 WITHOUT the bearer and 200 with it, 2,676,551 bytes of video/mp4. THREE OF THE FOUR THINGS ASKED FOR CAME BACK DIFFERENT: resolution 720p REQUESTED and 1920x1080 DELIVERED - SILENTLY, visible only by probing the bytes; duration 2 requested and 2.020000 s delivered, WHICH IS THE +20 ms AAC PADDING 182.C.35.09 MEASURED ON OUR OWN PIPELINE, now confirmed on a real provider render; and an AAC STEREO TRACK NOBODY ASKED FOR (no generate_audio field was sent). AND THE COST IS THE FINDING THE ITEM EXISTS FOR: wan-2.7 advertises ONE sku, `duration_seconds: "0.1"`, which our own normalizePricing reads as $0.10 per output second with minimumUsdPerGeneration NULL, so two seconds predicts $0.20 - AND THE API BILLED $0.50. 2.5x, or $0.25 per delivered second. The unit is not in question: the 2026-09-10 grok canary reported 0.05 against a 5-cents-per-second listing, matching exactly. NO COMBINATION OF THIS MODEL SINGLE ADVERTISED SKU PRODUCES 0.50. The likely cause is the 1080p substitution billed at a rate the catalog does not list - BUT THE CAUSE IS A GUESS AND THE DISCREPANCY IS A MEASUREMENT, and only the second belongs in a cost model. CONSEQUENCE: ANY ESTIMATE BUILT FROM pricing_skus UNDER-PREDICTS THIS LANE BY 2.5x, which reaches admission decisions, per-job ceilings, budgets (182.C.37.05) and cost-per-releasable-second (182.C.37.06). Declared on catalog arithmetic, this route would have been budgeted at 40% of its real price. | One billed render, recorded redacted. Spend $0.50; running tally $0.65 of the ~$5.00 approved, ~$4.35 remaining, kept in the existing canary tally rather than a new one. The key was sourced at run time from ~/.config/oshun/env-master.env through a throwaway /tmp script, verified BY LENGTH (73) and never by value, and the script was deleted at task end along with the scratch tsx files and the rendered clip. The recording carries no key material (grepped). The clip itself is NOT committed - the manifest keeps its sha256, byte count and full ffprobe instead.                                                                                                           | **$0.50** — `alibaba/wan-2.7`, 2 s requested, 1080p delivered. Tally: $0.65 of ~$5.00. | The $0.50-vs-$0.20 gap is RECORDED, NOT EXPLAINED. One render on one model on one day does not establish a rule, and the cause is not measured - only the discrepancy is. A second canary at 1080p would test the substitution hypothesis and has not been run.                                                                                                                                                       |

### 182.C.37 Performance, Capacity, Reliability, and FinOps

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                              | Owner | Dependency | Status | Changed Files                                                                                                                                                                                                          | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | Cost                                                                                    | Blocker                                                                                                                                                                                                                                                                                                                              |
| ----------- | --------------------------------------------------------------------------------------------------------------------- | ----- | ---------- | ------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| 182.C.37.01 | Load-test upload, callback, poll, download, proxy creation, evaluator queue, review playback …                        | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-capacity-model.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-hot-path-capacity/` (new: `bench.ts`, `results.json`, `README.md`)                        | FIVE OF THE EIGHT THINGS 182.C.37.01 NAMES HAVE NO TARGET TO LOAD: there is no upload endpoint, no evaluator queue, no transcoder, no review player and no delivery edge - the same measurement behind 182.C.36.02 and the four blocked browser items - so load-testing them would mean LOAD-TESTING A PLAN. THREE EXIST AND WERE MEASURED RATHER THAN ESTIMATED on the 4-core Linux box: verifyWebhookSignature 303,645/s, receiveWebhook (parse + dedupe) 263,720/s, applyTransition (state row AND outbox event, in one transaction) 241,562/s, sha256 1,680 MiB/s, parseCatalogBody 4,157/s. THE FINDING: THE SYSTEM IS NOT CPU-BOUND ANYWHERE, AND LOAD-TESTING OUR CODE MEASURES THE WRONG THING. The 182.C.36.10 canary took 107 SECONDS to render TWO seconds of video - 0.0093 renders per second - against a transition path doing 241,562 a second, which is SEVEN ORDERS OF MAGNITUDE of headroom. A 2.7 MB take digests in under two milliseconds. So the capacity model has TWO terms and WE CONTROL NEITHER: the provider queue, and MONEY. At the $0.25 per delivered second the same canary measured, THE WHOLE ~$5 CANARY BUDGET BUYS TWENTY SECONDS OF VIDEO, and the duration/resolution/candidate-count dimensions 182.C.37.02 asks about are dimensions of SPEND, NOT OF CPU - four candidates at eight seconds is $8, which already exceeds the budget. parseCatalogBody is the one row worth watching: SIXTY TIMES SLOWER than anything else here and read on every submission, though still four hundred thousand times faster than the render it precedes - a note rather than a problem, and the row that would become one first.                                                                                                                                                                                                                                                                                                                                           | 11 tests pass, every figure READ from the committed results.json rather than retyped; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. The bench script is committed as evidence so the numbers can be re-taken.                                                                                                                                                                                                                                   | No provider spend (the 107 s and $0.25/s figures are read from the 182.C.36.10 canary). | NOTHING HERE TOUCHES THE DATABASE, THE OBJECT STORE OR THE NETWORK, because none of them is wired - so the real bound on a MOUNTED system is an I/O latency this measurement cannot see, and applyTransition against Postgres is a different number by orders of magnitude. One host, one Node version, four cores, single-threaded. |
| 182.C.37.02 | Establish capacity models for duration, resolution, candidate count, evaluator GPU memory/time, transcode CPU …       | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-capacity-model.spec.ts` (new), `docs/domains/isis/human-video/evidence/2026-09-11-hot-path-capacity/` (new: `bench.ts`, `results.json`, `README.md`)                        | FIVE OF THE EIGHT THINGS 182.C.37.01 NAMES HAVE NO TARGET TO LOAD: there is no upload endpoint, no evaluator queue, no transcoder, no review player and no delivery edge - the same measurement behind 182.C.36.02 and the four blocked browser items - so load-testing them would mean LOAD-TESTING A PLAN. THREE EXIST AND WERE MEASURED RATHER THAN ESTIMATED on the 4-core Linux box: verifyWebhookSignature 303,645/s, receiveWebhook (parse + dedupe) 263,720/s, applyTransition (state row AND outbox event, in one transaction) 241,562/s, sha256 1,680 MiB/s, parseCatalogBody 4,157/s. THE FINDING: THE SYSTEM IS NOT CPU-BOUND ANYWHERE, AND LOAD-TESTING OUR CODE MEASURES THE WRONG THING. The 182.C.36.10 canary took 107 SECONDS to render TWO seconds of video - 0.0093 renders per second - against a transition path doing 241,562 a second, which is SEVEN ORDERS OF MAGNITUDE of headroom. A 2.7 MB take digests in under two milliseconds. So the capacity model has TWO terms and WE CONTROL NEITHER: the provider queue, and MONEY. At the $0.25 per delivered second the same canary measured, THE WHOLE ~$5 CANARY BUDGET BUYS TWENTY SECONDS OF VIDEO, and the duration/resolution/candidate-count dimensions 182.C.37.02 asks about are dimensions of SPEND, NOT OF CPU - four candidates at eight seconds is $8, which already exceeds the budget. parseCatalogBody is the one row worth watching: SIXTY TIMES SLOWER than anything else here and read on every submission, though still four hundred thousand times faster than the render it precedes - a note rather than a problem, and the row that would become one first.                                                                                                                                                                                                                                                                                                                                           | 11 tests pass, every figure READ from the committed results.json rather than retyped; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. The bench script is committed as evidence so the numbers can be re-taken.                                                                                                                                                                                                                                   | No provider spend (the 107 s and $0.25/s figures are read from the 182.C.36.10 canary). | NOTHING HERE TOUCHES THE DATABASE, THE OBJECT STORE OR THE NETWORK, because none of them is wired - so the real bound on a MOUNTED system is an I/O latency this measurement cannot see, and applyTransition against Postgres is a different number by orders of magnitude. One host, one Node version, four cores, single-threaded. |
| 182.C.37.03 | Add backpressure from object store/evaluator/transcoder to provider admission so completed outputs do not expire …    | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-backpressure.ts` (new), `human-video-backpressure.spec.ts` (new)                                                                                                            | SHEDDING AT ADMISSION COSTS NOTHING AND SHEDDING ANYWHERE ELSE HAS ALREADY BEEN PAID FOR. A request refused before submit costs $0; a completed output dropped because nothing could ingest it costs the full render - $0.25 PER DELIVERED SECOND, measured by the 182.C.36.10 canary rather than taken from the catalog, which predicted $0.10. That asymmetry decides where the valve goes, and shedCostUsd prices it: eight seconds turned away post-render is $2.00, and EVERY PLACEMENT BUT ADMISSION IS A WAY OF SPENDING MONEY TO DISCOVER YOU WERE BUSY. AND THE ADMISSION DECISION IS MADE BEFORE THE OUTPUT EXISTS: the same canary measured 107 SECONDS from create to completed, so admission cannot read the downstream CURRENT state - A QUEUE THAT IS EMPTY NOW AND FULL IN 107s ADMITS WORK THAT WILL EXPIRE WAITING - and reading `now` across that lead is refused. THE THREE DOWNSTREAMS FAIL DIFFERENTLY AND MUST NOT COLLAPSE INTO ONE SIGNAL: the object store going down LOSES A PAID RENDER (the bytes cannot land and the provider URL does not wait), while an evaluator or transcoder backlog costs LATENCY ONLY (the bytes land in quarantine and sit there; nobody is charged twice for waiting). ONLY ONE OF THE THREE STOPS ADMISSION, and a single "downstream pressure" number WOULD MAKE A SLOW EVALUATOR STOP THE BUSINESS. AND THE EVALUATOR IS 28x FASTER THAN THE PROVIDER THAT FEEDS IT, measured at both ends: 0.52 video-seconds per wall-second at its peak concurrency on four cores (182.C.20 curve) against the provider 2 video-seconds in 107 wall-seconds (0.0187) - so evaluator backpressure is THEORETICAL at any rate the provider can drive.                                                                                                                                                                                                                                                                                                      | 23 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. Both ends of the 28x ratio are asserted against the modules that measured them - the evaluator figure is read from MEASURED_CONCURRENCY_CURVE rather than retyped, so a re-measured curve changes this file too.                                                                                                                                                      | No provider spend (the $0.25/s and 107 s figures are read from the 182.C.36.10 canary). | THE REAL DOWNLOAD WINDOW HAS NEVER BEEN PROBED. The canary downloaded immediately and never tested the expiry, so assertIngestionWindowMeasured REFUSES a margin computed against an assumed TTL rather than defaulting to an hour - ARITHMETIC ON A GUESS READS EXACTLY LIKE ARITHMETIC ON A MEASUREMENT.                           |
| 182.C.37.04 | Verify bounded memory and disk use under large concurrent videos and clean abandoned temporary files after …          | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-temp-lifecycle.ts` (new), `human-video-temp-lifecycle.spec.ts` (new)                                                                                                        | THE ITEM ASKS TWO QUESTIONS THAT SOUND LIKE ONE. THE FIRST HAS A GOOD ANSWER AND THE SECOND DOES NOT. MEMORY IS ALREADY BOUNDED AND IT WAS MEASURED RATHER THAN ARGUED: streaming the FULL DEFAULT_MAX_OUTPUT_BYTES ceiling - 512 MiB - through the real downloadOutputToQuarantine grew PEAK RSS BY 0 MiB AND HEAP BY 0 MiB, at 1,690 MiB/s. The path digests and writes chunk by chunk, so its cost is O(CHUNK) NOT O(FILE), and ten concurrent 512 MiB takes cost TEN CHUNKS (10 MiB). TEMPORARY FILES ARE NOT BOUNDED, AND THE REASON IS IN THE ITEM OWN WORDING: every mkdtemp in this tree is paired with rm in a `finally`, which is correct and WHICH DOES NOT RUN ON SIGKILL, AN OOM KILL, OR A POWER LOSS - "crash/restart" and "exception" are named in the same breath BY PEOPLE WHO ASSUME ONE CLEANUP COVERS BOTH. A killed worker leaves a materialized video in /tmp forever, and MEASURED 2026-09-11 ACROSS apps/oshun/bff/src: NOTHING IN THIS REPOSITORY SWEEPS AT STARTUP. AND A SWEEP IS ONLY AS GOOD AS ITS PREFIX LIST, WHICH IS THE PART THAT ROTS: there are 31 DISTINCT mkdtemp PREFIXES in this directory today, several from test fixtures, so a sweep against a hand-maintained list cleans up EXACTLY THE PREFIXES SOMEBODY REMEMBERED and THE PREFIX ADDED NEXT WEEK IS THE ONE THAT LEAKS - this module offers ONE prefix (`oshun-hv-`) and REFUSES a list-driven sweep. AND A STARTUP SWEEP WITH NO AGE FLOOR DELETES A LIVE WORKER FILES: several workers share a host, so "everything matching the prefix" includes THE DIRECTORY THE PROCESS NEXT TO YOU IS WRITING INTO RIGHT NOW. The floor is one hour - past the 107 s render and the ~51 s evaluation of a 30 s take this phase measured.                                                                                                                                                                                                                                                                     | 19 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The 512 MiB measurement was taken by streaming the ceiling through the REAL download function under `node --expose-gc` with an RSS sampler, not estimated. The sweep is REAL FILESYSTEM WORK exercised against a real temp tree: an abandoned directory is removed, a 30-second-old one is left alone, and a directory that is not ours is ignored entirely.          | No provider spend.                                                                      | THE SWEEP IS OFFERED, NOT INSTALLED. Nothing calls it at startup, because nothing calls any of this (see 182.C.36.02) - and the 31 existing prefixes are NOT migrated to the single one, which is a change across a dozen modules outside this item.                                                                                 |
| 182.C.37.05 | Add spend reservations, per-job ceilings, daily/monthly budgets, anomaly detection, credits/refunds …                 | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-spend-reconciliation.ts` (new), `human-video-spend-reconciliation.spec.ts` (new); reservations, ceilings and windows already existed in `human-video-admission-controls.ts` | THE RESERVATION MACHINERY IS ALREADY RIGHT - it reserves an estimate across every window OR NONE of them, and settleScopeReservations settles each to THE COST THE PROVIDER ACTUALLY REPORTED. Nothing to fix. AND THAT IS EXACTLY WHY A SYSTEMATIC UNDER-ESTIMATE SURVIVES: SETTLING TO THE TRUTH IS CORRECT AND SILENT. The 182.C.36.10 canary reserved against a catalog price of $0.10/s, predicted $0.20 for a two-second render, and WAS BILLED $0.50 - the settle would have written 0.50 into every window and SAID NOTHING. The reconciliation is built; THE ALARM ON IT IS NOT, and a 2.5x gap that settles quietly is a gap nobody finds until a month closes. So the gap itself is the signal, with a 1.25x band - THE WIDEST THAT STILL CATCHES THE ONLY REAL DISCREPANCY THIS PHASE HAS MEASURED. An over-reservation is flagged too, because a window holding more than the job needed TURNED OTHER WORK AWAY. A PER-JOB CEILING CHECKED ONLY AGAINST THE ESTIMATE DOES NOT BOUND THE BILL: a $0.40 ceiling passes a $0.20 estimate and is BREACHED by the $0.50 bill, so it is re-checked at settle - too late to refuse THIS job, not too late to stop the NEXT one. AND "CREATOR-VISIBLE FINAL COST" HAS THREE CANDIDATES THAT DIFFER BY 2.5x - estimated, reserved, billed - so only the BILLED figure may be labelled final, and only once it exists: SHOWING THE ESTIMATE AND CALLING IT FINAL IS NOT AN APPROXIMATION, IT IS A NUMBER THIS PHASE HAS MEASURED TO BE WRONG. And the inherited rule, restated because it collides with all of the above: A REFUND IS NOT A NEGATIVE COST (182.C.30.05) - the job occupied its window at the time, and subtracting it afterwards REWRITES A DAY THAT ALREADY HAPPENED.                                                                                                                                                                                                                                                              | 23 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The tolerance is JUSTIFIED BY A TEST rather than chosen: the measured 2.5x gap is asserted to exceed the 1.25x band. Two division hazards are covered - a zero reservation yields a NULL ratio rather than an infinity that would fire on every free operation, while spend with no reservation at all is still flagged, because no window held it while it happened. | No provider spend (the $0.20 / $0.50 pair is read from the 182.C.36.10 canary).         | The alarm is a refusal, not a delivery channel - nothing here pages anybody, because nothing here is mounted (182.C.36.02). Daily/monthly budget WINDOWS already existed and were not rebuilt; what was added is the reconciliation on top of them.                                                                                  |
| 182.C.37.06 | Model cost per releasable second rather than cost per provider attempt, including failed candidates and …             | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-releasable-second-cost.ts` (new), `human-video-releasable-second-cost.spec.ts` (new)                                                                                        | THE ITEM OWN FRAMING IS THE FINDING: COST PER PROVIDER ATTEMPT IS NOT A PRICE, IT IS A LINE ITEM. Four terms, and this phase has MEASURED TWO, HALF-MEASURED ONE, AND CANNOT MEASURE THE FOURTH AT ALL. THE ATTEMPT PRICE IS MEASURED AND IS NOT WHAT THE CATALOG SAYS: 182.C.36.10 billed $0.25/delivered second against a catalog predicting $0.10, so ANY MODEL BUILT ON pricing_skus IS 2.5x LIGHT BEFORE THE ARITHMETIC STARTS. CANDIDATES PER SHOT IS CONFIGURED AND MULTIPLIES EVERYTHING: the production tier asks for FOUR, three are thrown away, and THEY COST EXACTLY WHAT THE KEPT ONE COST - COMPUTED: one 8-second production shot is 32 GENERATED SECONDS AND $8.00, OF WHICH 24 SECONDS IS VIDEO NOBODY WILL EVER SEE, which exhausts the tier own $50 per-job ceiling in SIX SHOTS, and rises to $24.00 with the tier two retries. EVALUATION IS MEASURED IN CPU, WHICH IS NOT THE SAME AS MEASURED IN MONEY: 1.69 CPU-seconds per video-second (182.C.20, imported rather than re-copied), FREE ON A BOX YOU ALREADY OWN AND NOT FREE ON RENTED COMPUTE - so the model refuses to fold it in until somebody names a rate rather than quietly treating CPU as zero. AND THE RELEASE RATE IS NOT MEASURED, AND IT IS THE DIVISOR: nothing has been released because nothing is mounted (182.C.36.02), so A MODEL WHOSE DIVISOR IS UNKNOWN DOES NOT PRODUCE A RANGE, IT PRODUCES NOTHING - costPerReleasableSecond returns NULL rather than a plausible figure. Given a rate, the gap is the point: at four candidates and a 75% release rate, a releasable second costs MORE THAN FIVE TIMES the per-delivered-second price.                                                                                                                                                                                                                                                                                                                                                          | 22 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The four-candidate count and the $50 ceiling are READ FROM routingPolicyForTier rather than typed in, and the 1.69 CPU-seconds figure is asserted EQUAL to FULL_EVALUATION_CPU_SECONDS_PER_VIDEO_SECOND so the two modules cannot drift.                                                                                                                              | No provider spend (the $0.25/s figure is read from the 182.C.36.10 canary).             | The headline figure this item asks for CANNOT BE PRODUCED TODAY - the release rate does not exist. The module returns null rather than a range, because a range implies two known endpoints and there are none.                                                                                                                      |
| 182.C.37.07 | Define provider/evaluator/storage/queue SLO dependencies, degraded modes, recovery objectives, and incident runbooks. | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-dependency-slos.ts` (new), `human-video-dependency-slos.spec.ts` (new); the three SLOs on things we do already existed in `human-video-slos.ts` (182.C.30.08)               | YOU CANNOT SET AN SLO ON A DEPENDENCY. YOU CAN ONLY SET ONE ON YOUR RESPONSE TO IT. "OpenRouter is available 99.9% of the time" is not a commitment we can make or measure, and NO SLA HAS BEEN OBTAINED FROM ANY PROVIDER USED HERE - an objective written against one we do not have is A NUMBER WITH NOTHING BEHIND IT. What IS ours is what happens in the failing fraction: whether we fail closed, whether the creator is told, whether the money is released. Only TWO of the four dependencies have an SLA, and THEY ARE THE TWO WE RUN OURSELVES. AND OUR AVAILABILITY CANNOT EXCEED THE PRODUCT OF EVERY DEPENDENCY WE CALL SYNCHRONOUSLY: four at 99.9% multiply to 99.6%, so the existing 99% job-completes objective is reachable and a 99.9% one is ARITHMETICALLY UNREACHABLE however well our own code behaves - computed, not asserted. THE DEGRADED MODES ARE ASYMMETRIC AND THE ASYMMETRY IS THE USEFUL PART: a provider outage stops NEW WORK and touches nothing released; a STORAGE outage stops new work AND DELIVERY OF EVERYTHING EVER MADE, making it THE ONLY DEPENDENCY WHOSE FAILURE REACHES A CREATOR WHO IS NOT GENERATING ANYTHING TODAY, and the only one that loses a paid render; an evaluator outage stops RELEASE and not generation (182.C.37.03 from the other direction); and a queue outage stops progress and loses nothing, because the outbox is durable and redelivery is a duplicate (182.C.35.04). TWO OF THE FOUR HAVE NO DEGRADED MODE AT ALL, and claiming one is refused: A DEGRADED MODE THAT CANNOT BE BUILT IS A PARAGRAPH IN A RUNBOOK THAT SOMEBODY WILL FOLLOW AT THREE IN THE MORNING. AND THE RECOVERY OBJECTIVE FOR A JOB IN FLIGHT IS NOT OURS EITHER: our RTO is a deploy, while the RTO on a job mid-render is "WHENEVER THEY FINISH, IF THE CONTENT URL HAS NOT EXPIRED" - and 182.C.37.03 recorded that the expiry has never been measured, so it is stated as UNKNOWN rather than quoting the deploy time as though it covered both. | 23 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The ceiling is COMPUTED from the call list and checked against the REAL SLOS array imported from 182.C.30.08, so a tightened objective there fails here; and a test asserts the four stop-sets are all DISTINCT, because a shared shape would mean the asymmetry had been flattened.                                                                                  | No provider spend.                                                                      | The 99.9%-per-dependency figures are ASSUMED, not measured - no provider publishes one to us. The ceiling arithmetic is therefore a demonstration of the CONSTRAINT rather than a statement of our actual ceiling, which is unknown for the same reason the objective is.                                                            |
| 182.C.37.08 | Run soak tests long enough to cover slow provider jobs, webhook retries, signed-URL expiry, queue aging …             | SRE   | 182.C.36   | done   | `apps/oshun/bff/src/generation/human-video-soak.spec.ts` (new)                                                                                                                                                         | A SOAK TEST WHOLE POINT IS ELAPSED TIME, AND A HERMETIC DOUBLE MAKES ELAPSED TIME FREE - WHICH MEANS THE SOAK NO LONGER SOAKS. The reason to make that trade anyway is arithmetic: THE LONGEST HORIZON IN THIS SYSTEM IS 730 DAYS (tombstone retention), which is 2,920x the next-longest (the six-hour released signed-URL TTL) and WHICH NO SOAK TEST WILL EVER RUN. So the only way to reach that behaviour is an INJECTED CLOCK - precisely the thing a soak test exists to avoid. BOTH HALVES ARE TRUE AT ONCE, and the honest response is to run the simulated soak AND NAME WHAT IT CANNOT SEE rather than picking one and calling it coverage. EXECUTED: a 730-day simulated soak that walks a job to `released` with the outbox still holding exactly one entry per applied transition; ages the queue past its 30-second lease ONCE A DAY FOR TWO YEARS and re-claims the same event 730 times WITHOUT LOSING IT; poisons a persistently failing delivery after EXACTLY eight attempts and then STOPS RE-CLAIMING IT, so the soak does not spin forever; dispatches cleanly when the receiver returns, with the attempt count preserved; and expires a download URL two years in through the double injected clock. AND WHAT A SIMULATED SOAK CANNOT SEE is everything only real elapsed time produces - heap fragmentation, file-descriptor leaks, connection-pool exhaustion, log and disk growth, certificate rotation, clock drift. NONE OF THOSE IS A BRANCH, so no injected clock reaches them; they are listed AS AN ASSERTION rather than a comment so the list cannot quietly shrink, and a test proves the double advanced two years having allocated, opened and written nothing.                                                                                                                                                                                                                                                                                                 | 12 tests pass; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new file does not. Every horizon is READ from the module that defines it - TOMBSTONE_RETENTION_SECONDS, DEFAULT_OUTBOX_LEASE_MS, OUTBOX_POISON_ATTEMPTS, MIN_SWEEP_AGE_MS and the released namespace rule - so a changed constant changes the soak rather than silently invalidating it.                                                                                                              | No provider spend - the double is hermetic, and two years of it cost nothing.           | THIS IS NOT A SOAK TEST. It is a long-horizon simulation, and the six things a real soak exists to find are exactly the six it cannot see. A real soak also needs a running system, and nothing is mounted (182.C.36.02).                                                                                                            |

### 182.C.38 Rollout, Feature Flags, and Incident Readiness

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                                | Owner       | Dependency | Status      | Changed Files                                                                                                                                                                         | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        | Cost                                                                        | Blocker                                                                                                                                                                                                                                                                                                                                                                          |
| ----------- | ----------------------------------------------------------------------------------------------------------------------- | ----------- | ---------- | ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.38.01 | Ship behind tenant, region, mode, provider, model, quality-tier, and app-surface feature flags with a global kill …     | Release Eng | 182.C.37   | done        | `apps/oshun/bff/src/generation/human-video-rollout-flags.ts` (new), `human-video-rollout-flags.spec.ts` (new); the kill switch itself already existed in `human-video-kill-switch.ts` | A FLAG AND A KILL SWITCH ARE NOT THE SAME MECHANISM AND MUST NOT SHARE A READ PATH. A flag is a product decision read from configuration; a kill switch is an emergency control THAT HAS TO WORK WHEN THE CONFIGURATION IS UNREACHABLE - point both at one store and THE FIRST OUTAGE IN IT IS ALSO THE MOMENT NOBODY CAN TURN ANYTHING OFF. The existing switch already has its own scopes (provider/lane/global), two of which are not flag dimensions at all; what this adds is the rule that it STAYS separate. A FLAG THAT DEFAULTS ON IS A DEPLOY WITH EXTRA STEPS: all seven default OFF, because a true default inverts the meaning - THE FLAG NOW DESCRIBES WHO HAS BEEN EXCLUDED, AND NOBODY MAINTAINS AN EXCLUSION LIST. AND THE DIMENSIONS ARE NOT INDEPENDENT, WHICH IS THE FAILURE A FLAT FLAG SET CANNOT SEE: 182.C.36.03 measured that exact-dialogue-performance is served by EXACTLY ONE family and video-edit by NONE, so enabling the mode while disabling that family provider flag produces A MODE THAT IS SWITCHED ON AND CANNOT RUN - EACH FLAG INDIVIDUALLY VALID, THE COMBINATION NOT, and invisible to anything that checks flags one at a time. THE CREATOR SEES A FEATURE THAT IS AVAILABLE AND FAILS. The verdict distinguishes "none of the families that serve it is enabled" from "no family serves it at all, whatever the flags say". AND SEVEN DIMENSIONS IS 128 COMBINATIONS, COUNTED FROM THE LIST RATHER THAN QUOTED: the one that ships is the one nobody enumerated, so a plan chosen by what came up is refused and 182.C.34.05 answer applies unchanged - the cross product is not a plan, all-pairs is the affordable claim.                                                                                                                                                                                  | 18 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The coherence check is COMPUTED from LANE_ADAPTERS and classForMode rather than from a table, so a family added or a mode refused changes the answer without touching this file; and the KILL_SCOPES are imported from the real switch so the separation claim is checked against it rather than described.                                                                                             | No provider spend.                                                          | The flags are DEFINED, not wired - nothing reads them, because nothing is mounted (182.C.36.02). The BFF general tenant-flag system in `feature-flags/tenant-feature-flags.ts` exists and is NOT extended here; connecting the two is a change outside this item.                                                                                                                |
| 182.C.38.02 | Roll out internal dogfood → consented design partners → allowlisted beta → capped production → broader release with …   | Release Eng | 182.C.37   | done        | `apps/oshun/bff/src/generation/human-video-rollout-stages.ts` (new), `human-video-rollout-stages.spec.ts` (new)                                                                       | THE FIRST STAGE CANNOT BE DONE AS WRITTEN. "Internal dogfood" normally means the team uses the thing it built - but this thing renders A PERSON FACE SAYING WORDS, so dogfooding it means putting COLLEAGUES likenesses into it, and AN EMPLOYEE CANNOT FREELY CONSENT TO THEIR EMPLOYER. The power asymmetry is the whole reason consent exists as a concept in this phase, and a rollout plan that starts by collecting staff faces HAS WALKED INTO IT ON STEP ONE. What internal dogfooding CAN do is exercise every path that does not depend on a real person - the 182.C.33.10 smoke corpus, the generated fixtures, the refusals - so the stage is marked `usesRealLikeness: false` and is the ONLY one of the five that is. THE GATE BETWEEN STAGES IS A MEASUREMENT AND 182.C.34.07 ALREADY PRICED IT: 299 clean renders per cohort, which at the MEASURED $0.25/delivered second and eight-second takes is $598 PER COHORT AND $3,588 ACROSS SIX - in renders alone, BEFORE ANYBODY IS PAID TO LOOK AT THEM. A stage plan that does not name that number is one somebody will be surprised by. Each stage also records WHAT IT DOES NOT PROVE, all five distinct, because THE FAILURE MODE OF A ROLLOUT LADDER IS A STAGE TREATED AS EVIDENCE FOR THE ONE AFTER IT - and the last stage proves nothing at all, being the absence of a cap rather than a test. AND "CAPPED PRODUCTION" IS THE ONLY STAGE WITH A NUMBER IN IT, SO THE NUMBER HAD BETTER BE THE RIGHT KIND: a cap on USERS bounds nothing, because ONE user generating a three-minute piece at four candidates is $180 against A HUNDRED users trying an eight-second take once at $200 - comparable, so a user count says almost nothing about the bill. The cap is a SPEND cap.                                                                                                  | 22 tests pass; eslint clean; typechecked against apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the new files do not. The 299 figure is asserted EQUAL to samplesForZeroFailureClaim(0.01) from 182.C.34.07 rather than retyped, and the user-cap argument is carried by arithmetic in the test rather than by the prose.                                                                                                                                                                                                     | No provider spend (the $0.25/s figure is read from the 182.C.36.10 canary). | THE $3,588 COHORT GATE HAS NOT BEEN SPENT and no stage has been entered - this is the ladder and its prices, not a rollout. The consent problem with staff dogfooding is recorded as a REFUSAL rather than solved; solving it means finding people who are not employees.                                                                                                        |
| 182.C.38.03 | Keep active-scene exact-dialogue, multi-speaker, long-form, avatar, and corrective lanes independently promotable and … | Release Eng | 182.C.37   | implemented | `apps/oshun/bff/src/generation/human-video-product-lanes.ts` (new), `human-video-product-lanes.spec.ts` (new)                                                                         | THE FIVE LANES ARE NOT FIVE OF THE SAME KIND OF THING and the item reads as if they were: three are modes the contract names, one - long-form - is a LENGTH assembled out of another lane, and `corrective` is a CAPABILITY CLASS NO MODE REACHES (182.C.13.03 declared `corrective-sync` with an empty mode list on purpose), so its promotion state is unreachable rather than off and `assertLaneReachable` refuses it - A FLAG THAT CAN BE SET FOR IT IS A FLAG THAT LIES. THE TWO EXACT-AUDIO LANES SHARE THEIR ONLY ROUTE, read off `LANE_ADAPTERS` rather than asserted here: FOUR OF THE FIVE FAMILIES refuse both `exact-dialogue-performance` and `multi-character-scene` for want of an audio input, leaving `wan` alone. The intuition about what that means RUNS BACKWARDS IN BOTH DIRECTIONS. DEMOTION IS COUPLED WHATEVER THE LANE FLAGS SAY - stopping `wan` takes exact-dialogue, multi-speaker AND long-form down together, THREE OF THE FIVE LANES, so the reversal that matters in an incident is not the one the flags control. PROMOTION IS NOT INHERITED, however much the shared route invites it - multi-speaker asserts that the correct person mouth moves for every line (182.C.20.03), which no single-speaker evidence measured: SAME MODEL IS NOT SAME CLAIM. A LENGTH IS NOT A LANE: long-form borrows its base families and may never outrank the base it assembles, and 182.C.22.10 own gate still refuses on top because NO ROUTE HAS A CONTINUITY-DECAY BENCHMARK. AND "REVERSIBLE" IS TRUE OF TWO POPULATIONS OUT OF THREE: work not yet submitted stops, work in flight drains because 182.C.13.08 blocks only submission, and RELEASED WORK IS GONE - the creator has the file, and recording that as fully reversed is claiming a recall nobody can perform.                                                      | 22 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. The sole-family coupling is COMPUTED from `LANE_ADAPTERS` in the test (four of five families refuse both audio modes) rather than hardcoded, corrective unreachability is recomputed from `modesForClass('corrective-sync')`, and the long-form gate is exercised through 182.C.22.10 real `assertAutoLongFormAllowed` rather than a restatement of it.                  | No provider spend; no model call.                                           | THE COUPLING IS RECORDED, NOT REMOVED. A second audio-capable family is the only thing that makes these two lanes independently reversible in fact; 182.C.13.01 direct `@isis/seedance-provider` clients are the candidate. `corrective` stays unpromotable until a mode reaches its class, and `long-form` until some route carries a 182.C.22.10 benchmark.                    |
| 182.C.38.04 | Define canary abort thresholds for consent/safety incidents, bad-sync false accepts, identity drift …                   | Release Eng | 182.C.37   | implemented | `apps/oshun/bff/src/generation/human-video-canary-abort.ts` (new), `human-video-canary-abort.spec.ts` (new)                                                                           | A CANARY IS TOO SMALL TO MEASURE THE RATES PEOPLE WRITE THRESHOLDS IN. 182.C.14.09 sets a canary at 50 takes; inverting 182.C.34.07 own zero-failure derivation, FIFTY CLEAN TAKES RULE OUT NOTHING BELOW 5.8% at 95% confidence - so a "1% bad-sync false accepts" abort is a number the run cannot reach an opinion about, and the 299 takes it would need are SIX CANARIES. Writing it anyway produces a threshold that never fires and a rollout that only looks watched. CONSENT AND SAFETY THEREFORE GET NO RATE AT ALL, for two reasons and the second is arithmetic: 182.C.30.06 already said safety has no eventual rate, and the canary could not substantiate one if it wanted to - both point at ABORT ON FIRST OCCURRENCE. A FALSE ACCEPT IS INVISIBLE TO THE LOOP THAT PRODUCED IT: the gate passed it, which is what makes it false, so nothing in the pipeline raises its hand and the signal arrives from sampled review at that stream cadence. Two of the six signals are like this, and the abort LAGS - at 7 takes/h over a 24h review cycle, 168 TAKES RELEASE BETWEEN THE FIRST BAD ACCEPT AND THE ABORT, which 182.C.38.03 already showed a flag cannot reach. A PROVIDER FAILURE RATE READ OFF DELIVERED TAKES IS THE RETRY POLICY OPINION, NOT THE PROVIDER: retries exist to hide provider failure from the creator and hide it from the metric too, so a provider failing 50 of 100 submissions delivers every take and reads as healthy. The rate is read at submit. AND COST PER ATTEMPT IS NOT THE COST - 182.C.37.06 settled the denominator, and a mean over attempts looks fine while containing the take that retried four times, which is the take a cost abort exists for. Latency and cost are the only signals every take samples, so they are the only two carrying a quantile, on the tail rather than the mean. | 19 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. `detectableRate` is asserted to ROUND-TRIP through 182.C.34.07 `samplesForZeroFailureClaim` - the rate a 50-take canary resolves needs at most 50 takes and 99% of that rate needs more - so the 5.8% figure is derived rather than typed, and the 299/6-canary arithmetic is computed in the test.                                                                      | No provider spend; no model call.                                           | NO CANARY HAS RUN. These are the thresholds and what each can honestly see, not a canary result. The two lagging signals still need a review cadence chosen - the exposure arithmetic is here, the number of reviewers is not - and 182.C.38.05 incident drills are what would exercise the aborts.                                                                              |
| 182.C.38.05 | Practice provider quarantine, leaked-URL response, unauthorized-likeness takedown, revocation, model regression …       | Release Eng | 182.C.37   | implemented | `apps/oshun/bff/src/generation/human-video-incident-drills.ts` (new), `human-video-incident-drills.spec.ts` (new)                                                                     | THE CLOCK DOES NOT START WHEN WE NOTICE. A leaked URL is reachable from the moment it leaks; a takedown obligation starts at the SUBJECT REQUEST, not its triage; a revocation is in force from the instant it is made. Time a drill from detection and the number that comes out is THE HALF WE ARE ALREADY GOOD AT, reported as containment - worked in the test: 12 minutes of response after 9 hours of not knowing is a drill rehearsing UNDER 3% OF A 552-MINUTE INCIDENT. So each drill declares its onset event, containment is measured from there, and the detection lag sits beside it rather than folded in. A DRILL MEASURES RESPONSE AND CANNOT MEASURE DETECTION, because it begins by telling everybody it is starting - an announced drill reporting a detection time is refused. FOUR OF THE SIX NEED A SUBJECT AND THE ONLY HONEST ONE IS SYNTHETIC: rehearsing a takedown means HAVING an unauthorized likeness, a leaked-URL response means LEAVING one reachable, revocation means withdrawing somebody real consent - the same shape of refusal 182.C.38.02 made about dogfooding on staff, and 182.C.33.10 smoke corpus is the answer. TWO OF THE SIX CANNOT BE UNDONE, so aiming them at production releases is refused; and succeeding at mass invalidation still leaves what 182.C.26 marks unreachable exactly where it was - 412 records invalidated against 97 files already in creators hands is a drill measuring ITS OWN BOOKKEEPING. AND A DRILL EXPIRES ON CHANGE, NOT ON A CALENDAR: a three-day-old drill whose covered path changed yesterday is a document, and a quarterly cadence would have passed it.                                                                                                                                                                                                          | 19 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. The response-share arithmetic is computed in the test rather than asserted in prose, every onset string is checked NOT to mention noticing or detection, and the takedown block calls 182.C.26 real `takedownReach`/`assertNotOverclaimed` rather than restating what a takedown leaves behind.                                                                          | No provider spend; no model call.                                           | NO DRILL HAS BEEN RUN. This is the shape of the six and what each can honestly prove, not a drill record - running them needs a smoke-corpus release path and named owners per drill. The detection half stays unmeasured by construction; it comes from real incidents or injected findings, and 182.C.38.04 lagging-signal exposure is the same gap seen from the canary side. |
| 182.C.38.06 | Add rollback that preserves job/audit truth and prevents old clients from resubmitting unsupported contracts.           | Release Eng | 182.C.37   | implemented | `apps/oshun/bff/src/generation/human-video-rollback.ts` (new), `human-video-rollback.spec.ts` (new)                                                                                   | A ROLLBACK MOVES THE CODE. IT MUST NOT MOVE THE PAST. The pressure arrives within minutes - old code cannot read the fields new code wrote, so somebody proposes a down-migration that strips them, and THAT MIGRATION MAKES THE RECORD SAY SOMETHING THAT WAS NEVER TRUE: a take rendered under the new contract, filed as if it had been rendered under the old one. Audit truth is not "the record parses", it is "the record says what happened", so only code and configuration are movable and old code that cannot read a field SKIPS it. A JOB IN FLIGHT IS JUDGED BY THE CONTRACT IT WAS ADMITTED UNDER: half its life ran under one set of rules and grading the other half under older ones produces A DECISION NO VERSION EVER MADE, so the version is pinned at admission and rolled-back code holds a job pinned ahead of it rather than proceeding - the 182.C.13.08 move of stopping new work instead of abandoning running work. A job with NO pin is held too, because the default answer is the wrong one. AND AFTER A ROLLBACK THE UNSUPPORTED CONTRACT IS THE NEW ONE, SO THE CLIENTS THAT BREAK ARE THE ONES THAT UPGRADED - the item names the roll-forward case, which is the intuitive one; reversed, the broken set flips from the laggard SDK to Studio and mobile, who are sending the contract we published and then withdrew. That matters because `hv.validation.unsupported_version` carries `blame: request` in the shared catalog - CORRECT WHEN A CLIENT IS BEHIND AND A LIE WHEN WE ARE THE ONES WHO MOVED - so a rollback-caused refusal must disclose the rollback, or the integrator reads "your request is invalid", checks it against our own documentation, and finds it valid.                                                                                                                                 | 16 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. The blame and terminal flags are READ from `FAILURE_DEFINITIONS` in the contracts package rather than restated, the who-breaks inversion is computed by running the same client set through the move in both directions, and the no-Retry-After half is exercised through 182.C.25.02 real `retryAdvice`/`assertRetryAfterHonest` instead of a second copy of that rule. | No provider spend; no model call.                                           | NO ROLLBACK HAS BEEN REHEARSED. The contract version is not yet pinned on the job record itself - this module defines the disposition and refuses the unpinned case, and writing the pin at admission is the wiring 182.C.38.03/38.05 work would exercise. The disclosure half also needs an error-envelope field the SDKs read.                                                 |
| 182.C.38.07 | Obtain Trust & Safety, privacy, security, legal/licensing, accessibility, SRE, product …                                | Release Eng | 182.C.37   | implemented | `apps/oshun/bff/src/generation/human-video-signoffs.ts` (new), `human-video-signoffs.spec.ts` (new)                                                                                   | AN APPROVAL WITH NO ARTIFACT ATTACHED APPROVES WHATEVER SHIPS NEXT. A sign-off is a claim by a named person, about a specific system, at a specific version - and the version is the part that goes missing, because "T&S approved the human video launch" stays true through six weeks of changes to the thing that was approved. So a sign-off binds to an evidence digest and EXPIRES ON CHANGE TO WHAT IT COVERED, NOT ON A CALENDAR - the rule 182.C.38.05 applies to a drill, for the same reason. AND NOBODY CAN APPROVE WHAT THEY WERE NOT SHOWN: this phase has recorded in its own modules exactly what each approver evidence cannot cover, so `knownGapsFor` COMPUTES the gap list from those modules rather than typing it - the canary resolution from 182.C.38.04, the sole-family lane coupling from 182.C.38.03, the unrun drill count from 182.C.38.05 - and seven of the eight roles carry at least one. Accessibility carries none, which is stated rather than padded. THE REAL OUTCOME IS USUALLY "APPROVED, SUBJECT TO", AND THAT IS WHERE APPROVALS GO TO DIE: a condition with no owner is one nobody closes, and one with no expiry is A PERMANENT EXCEPTION WRITTEN IN THE LANGUAGE OF A TEMPORARY ONE - 182.C.40.12 demands owners and expiry at the end of the phase, and requiring them at the moment the condition is written is what makes that possible. AND RELEASE ENGINEERING SIGNS LAST BECAUSE IT SIGNS AN ARTIFACT: the other seven approve a design, a policy or a measurement, judgeable before the build exists; signed early, this one approves a plan to build something rather than the something.                                                                                                                                                                                                           | 17 tests pass; eslint clean; typechecked against a scratch config extending apps/oshun/bff/tsconfig.typecheck.json with a deliberate \_\_control.ts error, which reports while the two new files do not. The gap strings are asserted to CONTAIN values recomputed in the test from `detectableRate(MINIMUM_CANARY_TAKES)` and `coupledPairs()`, so a gap cannot drift into flattery while the sentence describing it stays reassuring, and the item own role order is fed to `assertReleaseEngineeringLast` as the passing case.                                                 | No provider spend; no model call.                                           | NO SIGN-OFF HAS BEEN OBTAINED. This is the shape of the record and what each approver must be shown, not eight signatures - and it cannot be otherwise while the gaps it computes are open: no canary has run, no drill has been run, and the exact-audio lanes still rest on one family. The evidence digest itself needs a bundle to digest, which is 182.C.40.09 archive.     |
| 182.C.38.08 | Update the capability matrix and public product copy only after the promoted route has real production-like evidence.   | Release Eng | 182.C.37   | not-started | —                                                                                                                                                                                     | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | —                                                                           | None                                                                                                                                                                                                                                                                                                                                                                             |

### 182.C.39 Documentation and Operator Guidance

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                             | Owner | Dependency         | Status      | Changed Files                                                  | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Cost                                                                 | Blocker                                                                                                                                                                                                                                                                                                                                                                                |
| ----------- | -------------------------------------------------------------------------------------------------------------------- | ----- | ------------------ | ----------- | -------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.39.01 | Document creator workflows for reference preparation, consent, script locking, pronunciation, audio approval …       | Docs  | 182.C.38           | implemented | `docs/domains/isis/human-video/creator-workflow.md` (new)      | THE FIRST THING THE DOCUMENT HAS TO SAY IS THAT NONE OF IT RUNS. `human-video-api-routes.ts` has NO IMPORTER OUTSIDE ITS OWN SPEC and `apps/oshun/bff/src/server.ts` mounts none of it - verified by grep rather than assumed - so a creator workflow written as a tutorial would be a tutorial for something that does not exist. It is written instead as A MAP OF THE DECISIONS a creator will meet: per stage, what they supply, WHAT WILL REFUSE THEM AND WHY, and what the stage does NOT do. That last column is the point, because the failure this phase exists to avoid is a creator who believes a take was checked for something nobody checked. THE FLOW IS NOT A LINE, and four places say so: consent binds to the reference DIGESTS so it cannot precede the upload and must precede every submission; audio fixes durations that the shot plan then has to fit, so a plan written first gets rewritten; correction loops back into review and produces A NEW TAKE WITH ITS OWN EVIDENCE rather than an annotation on the old one; and deletion CAN BE ENTERED FROM ANYWHERE AND COMPLETED FROM NOWHERE - the only stage whose success is partial by construction. The document closes with three things a creator must NOT infer: that any of it runs, that a passing take is a good take (the gates reject known failures and do not certify quality), and that the thresholds are final (several are exported PROVISIONAL and say so in the code).                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       | Every module named in the document was checked to exist on disk by extracting the backticked filenames and testing each path; all 40 resolve. The unmounted claim is grep-verified twice (no non-spec importer of `human-video-api-routes.ts`; no "human-video" string in `server.ts`). Prettier-formatted. No code changed, so no test or typecheck applies.                                                                                                                                                                                                                                      | No provider spend; no model call.                                    | THE DOCUMENT DESCRIBES A FLOW THAT IS NOT WIRED. It stays accurate only while the modules it names keep their item numbers; the stage-to-module map was built from each module own header reference. When 182.C.23/24 decide the Studio surface the creator-facing wording will need a second pass, and the API half is 182.C.39.03.                                                   |
| 182.C.39.02 | Document honest limitations and troubleshooting for profile, occlusion, fast speech, singing, overlapping dialogue … | Docs  | 182.C.38           | implemented | `docs/domains/isis/human-video/limitations.md` (new)           | THE DOCUMENT LEADS WITH THE THREE LIMITATIONS THAT FAIL SILENTLY, because a limitation you can see is one you can report and these produce output that looks correct to every check we run. (1) OVERLAPPING SPEECH BREAKS THE ALIGNER SILENTLY - 182.C.18.08 measured a 513.8 ms onset error, four times the ITU reject boundary, WHILE THE PRONUNCIATION SCORE STAYS AT EXACTLY 0.0, because forced alignment is given the words and asked where they are and will always find them somewhere. (2) A FROZEN CAST BUYS A PERFECT ATTRIBUTION MARGIN - 182.C.20.03 measured 0.65 with a reacting listener against 1.00 with a still one, so the confident number is CAUSED BY the defect. (3) THE MEASURABILITY FILTER IS A SAMPLING BIAS - 182.C.20.02 measured 0.908 over the scored frames against 0.750 under the clip own motion mix. EVERY NUMBER IN THE DOCUMENT IS A MEASUREMENT THIS REPO HOLDS, with the module that holds it: 60/120 degrees of yaw, 7%/10% face height, 21/31 px of blur with 0.251 against a 0.282 stranger baseline, the 0.4 reference cosine WITH ITS KNOWN FALSE NEGATIVE (strangers measured at 0.422, above the line), GOP -4.99 at 1.25x and -10.20 at 0.80x against a -3.0 floor, SNR 16.8 dB under a gate at 20. SINGING GETS AN HONEST NOTHING: it is one of six delivery conditions with no fixture, because the built conditions are TRANSFORMATIONS of a signal and singing is different phonation - filtering a voice into "a whisper" produces a quiet normal voice and proves only that the detector reads the filter. The troubleshooting table maps SYMPTOM TO CAUSE rather than listing features, and the document closes by saying no real generated clip has been through any of it.                                                                                                                                                                                                                                                        | Every module named in the document was checked to exist on disk; all resolve. Every figure traces to a module constant or a committed measurement: `WARNING_THRESHOLDS`, `PROVISIONAL_YAW_BOUNDARIES`, `CONDITION_COVERAGE` (singing verified as `measured: false`), the 182.C.34.06 noise spread behind the demotion tolerance, and the two figures from this session own asserted tests. Prettier-formatted. No code changed, so no test or typecheck applies.                                                                                                                                   | No provider spend; no model call.                                    | NO REAL GENERATED CLIP HAS BEEN THROUGH THE PIPELINE, so none of these limits is confirmed against production output - they are measured on fixtures and committed frames. The pass rates a creator actually wants (how often a take is good) are 182.C.30.03 and wait on 182.C.33. Six delivery conditions including singing stay unmeasured until that corpus has recorded material. |
| 182.C.39.03 | Publish API guides and complete examples without real credentials, private biometric assets …                        | Docs  | 182.C.38           | implemented | `docs/domains/isis/human-video/api-guide.md` (new)             | AN API GUIDE MADE OF 200s TEACHES AN INTEGRATOR TO WRITE A CLIENT THAT ONLY HANDLES SUCCESS, which is the deeper form of the "misleading mocked success claims" the item bans - so the guide LEADS WITH THE REFUSALS an integrator hits first, each with its stable code, status and RETRY DISPOSITION (the disposition is the contract; the message is not). Writing that table caught three of my own five statuses wrong against `FAILURE_DEFINITIONS`: consent.missing is 403 NOT 400 (the request is well-formed and you are not permitted to make it), no_eligible_lane is 409 NOT 400 (nothing about the request is invalid - the catalog blames `platform`), and submission_ambiguous is 500 with disposition `never` NOT a retryable 502 (a retry could pay twice for a render the provider already accepted). AND NO RESPONSE IN THE GUIDE WAS CAPTURED, because nothing is mounted: every payload is schema-derived and labelled, and the guide states the test - a captured response carries a dated evidence path under docs/domains/isis/human-video/evidence/ (182.C.01.04) and NONE OF THESE DOES. A COMPLETE EXAMPLE NEEDS AN INPUT AND THE INPUT IS A FACE: a runnable quickstart must ship either biometric data or something that is not the real input, so it ships neither - the only publishable subject is the 182.C.33.10 smoke corpus, which has NO FACE IN IT, so the quickstart exercises transport and proves nothing about identity or sync. The surface table (23 operations, 14 idempotent, 6 deciding) is EXTRACTED FROM `API_SURFACE` rather than transcribed, and the guide repeats 182.C.25.01 two-versions rule: a candidate written under schemaVersion 1 and read back through a later route IS STILL A V1 DOCUMENT.                                                                                                                                                                                                                                 | The endpoint table was parsed out of `human-video-api-surface.ts` ENDPOINTS by script - 23 of 23 entries matched, and the 14/6 counts are from that parse rather than counted by eye. Every quoted failure code, status, retry disposition and blame was read from `libs/contracts/src/human-video/failure-codes.ts` and three were corrected after the first draft. Every module named exists on disk. Prettier-formatted. No code changed.                                                                                                                                                       | No provider spend; no model call.                                    | NOTHING IS MOUNTED, so the guide cannot carry a single captured response - closing that gap needs the routes wired, not more documentation. The quickstart stays transport-only until a publishable subject exists (182.C.33), and the SDK examples (182.C.25.04) are generated from the same surface and should be regenerated when it moves.                                         |
| 182.C.39.04 | Document provider onboarding, live probing, benchmark, promotion, quarantine, cost, retention …                      | Docs  | 182.C.38           | implemented | `docs/domains/isis/human-video/provider-operations.md` (new)   | EIGHT PROCEDURES, EACH WITH THE MISTAKE IT EXISTS TO PREVENT. ONBOARDING IS MOSTLY WRITING DOWN WHAT THE API WILL NEVER TELL YOU - licensing, commercial use, biometrics, minors, retention - and that record is manual and DATED, because AN UNDATED CONSTRAINTS RECORD IS WORSE THAN NONE: terms change, and a record with no date cannot be told apart from one written this morning. PROBING BINDS TO THE DATED SLUG (a probe of wan-2.7-20260801 is evidence about a model no longer served), expires at 90 days, and keeps two pairs of states apart that collapsing would hide: probe_inconclusive is NOT probe_failed, and region_unrecorded is NOT region_unavailable - treating silence as "allowed everywhere" is how a likeness gets processed in a jurisdiction the authorization never covered. PROMOTION HAS FIVE RUNGS AND SHADOW IS THE ONE PEOPLE SKIP; the canary rung now carries 182.C.38.04 arithmetic - 50 TAKES RESOLVE NO RATE BELOW 5.8%, so a 1% abort threshold never fires. A QUARANTINE IS A CONSEQUENCE AND A KILL SWITCH IS A DECISION, never the same field: the first follows automatically from a snapshot diff and nobody signs it, the second has a person name on it and blocks SUBMISSION ONLY. COST: PRICE FROM WHAT YOU WERE BILLED, NEVER FROM THE CATALOG - the committed canary predicted $0.20 and was billed $0.50, TWO AND A HALF TIMES, so read `usage.cost` rather than computing from `pricing_skus`. RETENTION IS A CONSENT DIMENSION rather than a footnote, which is why a fallback to a weaker-retention provider is prohibited even when it is capable. And DELETION REACHES WHAT WE HOLD, NOT WHAT THEY HOLD - zero-retention is the only provider class where the provider-side line is genuinely empty.                                                                                                                                                                                                                           | Every module named exists on disk (checked by extracting the backticked filenames). The $0.20-predicted/$0.50-billed figures and the 2.5x are read from the committed canary at evidence/2026-09-11-wan-canary/README.md, not from memory; 299 renders, $598 per cohort and $3,588 across six come from 182.C.34.07 and 182.C.36.10; the 90-day probe expiry, 20 shadow runs and 50 canary takes are the exported constants. Prettier-formatted. No code changed.                                                                                                                                  | No new provider spend; the document cites the existing $0.50 canary. | NO PROVIDER HAS BEEN ONBOARDED THROUGH IT END TO END. One route has been probed with real money ($0.50) and NO ROUTE HAS A BENCHMARK, so nothing has climbed past `probed`; the quarantine and kill-switch procedures have never been exercised on a live incident, which is 182.C.38.05 drill work.                                                                                   |
| 182.C.39.05 | Document evaluator acquisition, calibration, thresholds, failure interpretation, fairness slices, upgrades …         | Docs  | 182.C.38           | implemented | `docs/domains/isis/human-video/evaluator-operations.md` (new)  | AN EVALUATOR IS A PINNED ARTIFACT, NOT A LIBRARY NAME - weights, digest, version and fetch date - because a dependency that silently updates its weights INVALIDATES EVERY THRESHOLD EVER FITTED AGAINST IT, quietly, and the first symptom is a pass rate that moved for no reason anybody can name. CALIBRATION FITS A THRESHOLD TO A CORPUS, SO THE THRESHOLD INHERITS THE CORPUS BIASES: a cutoff fitted on well-lit frontal faces is a cutoff FOR well-lit frontal faces wearing a general name, which is why calibration and fairness slicing are ONE procedure. Counting the codebase produced the number that matters: NINETEEN PROVISIONAL THRESHOLD EXPORTS ACROSS EIGHTEEN MODULES, tabulated by name, and EVERY RELEASE DECISION CURRENTLY RESTS ON THAT LIST - nobody should quote any of them as a measured limit. FAILURE INTERPRETATION KEEPS TWO SPLITS: a low score is not an unmeasurable window (no findings walks into a release gate looking like no problems), and "out of sync" is SIX patterns with six different repairs - a mux fix, a rate conform, a shot retry, a different model, a different model, and the assembly - which a single score cannot distinguish. SKIN TONE CANNOT BE A MEASURED COHORT and this repo measured why: a ONE-STOP EXPOSURE CHANGE moved one subject 58.2 degrees of ITA against a 63.7 degree gap between two different people, so a pixel-derived tone cohort is mostly a LIGHTING cohort. AN EVALUATOR UPGRADE INVALIDATES THE EVIDENCE GATHERED UNDER THE OLD ONE - the new version scores differently BY DEFINITION, which is what makes it an upgrade - so the procedure is a re-benchmark, not a swap. AND ROLLBACK HAS 182.C.38.06 SHAPE: move the evaluator, never the record; a candidate stays measured under the version that measured it, and a distrusted decision is RE-MEASURED into a new record rather than an edited one.                                                                                      | The nineteen-export table was produced by grepping `^export const PROVISIONAL` across the generation directory and is asserted against that count (18 module rows, 19 exports - offscreen-dialogue carries two). Every module and sibling document named exists on disk. The 21/31 px and 0.251/0.282 identity figures, the 58.2/63.7 ITA figures, the 30-subject bucket minimum and the six desync patterns are read from the modules that hold them. Prettier-formatted. No code changed.                                                                                                        | No provider spend; no model call.                                    | NO EVALUATOR HAS SCORED A REAL GENERATED CLIP IN THIS PIPELINE. That is why nineteen threshold sets are provisional and every fairness bucket is empty; 182.C.17.08 (calibration) and 182.C.33 (corpus) are the two items that change this page. The upgrade and rollback procedures have never been run.                                                                              |
| 182.C.39.06 | Document review and escalation rules, non-overridable gates, subject requests, takedowns, and incident response.     | Docs  | 182.C.38           | implemented | `docs/domains/isis/human-video/review-and-escalation.md` (new) | "NON-OVERRIDABLE" IS A LIST, NOT A PRINCIPLE, because every gate looks overridable at two in the morning - so the document carries the list: FOUR waivable gates, each protecting THE BUYER WHO IS IN THE ROOM and can decide the video is good enough for them, and TWELVE that are not, EACH NAMING THE PERSON IT PROTECTS. Naming is the mechanism: "not waivable" invites a request for an exception and "the person whose likeness this is did not agree to be in this decision" does not. A waiver on one of the four is still bounded (seven days), co-signed when the requester owns the job, and reason-coded, because free text cannot answer "how often does this happen". A REVIEWER DOES NOT NEED THE ORIGINAL TO DO THE REVIEW - the review question is whether the OUTPUT is acceptable and the reference photograph is an INPUT, which matters because it is the most sensitive artifact in the system and the review queue is the highest-traffic surface touching it. ESCALATION SPLITS BY AUDIENCE ON THE FIELD EVERYBODY ASSUMES IS HARMLESS: support may see which tenant, a provider may not, because A PROVIDER WHO CAN CORRELATE FAILURES TO NAMED CUSTOMERS HAS BEEN HANDED A CUSTOMER LIST. A SUBJECT IS NOT A TENANT AND EVERY INDEX IS BUILT THE OTHER WAY ROUND, so "everything about this person" is answerable only as far as the authorizations reach and THE REACH MUST BE STATED; embeddings are deleted on request and RETURNED TO NOBODY, because a vector means nothing to the subject and everything to whoever else gets the export. Subject requests do not run through the creator support queue - the subject is not the customer, and a queue ordered by customer value orders these wrong. TAKEDOWNS ARE PARTIAL BY CONSTRUCTION and must never be reported complete. INCIDENTS ARE TIMED FROM ONSET, and the two controls are not the same: a quarantine is automatic and unsigned, a kill switch has a name on it and blocks submission only. | The twelve-gate table was cross-checked against `NOT_WAIVABLE_BECAUSE` in `human-video-quality-waiver.ts` by script: 12 code gates, 12 document rows, none missing and none invented, with each reason quoted from the code. The four waivable gates were read from `WAIVABLE`. Every module named exists on disk. Prettier-formatted. No code changed.                                                                                                                                                                                                                                            | No provider spend; no model call.                                    | NO DRILL HAS BEEN RUN, NO SUBJECT REQUEST HAS BEEN ANSWERED, AND NO WAIVER HAS BEEN ISSUED. The lists are enforced in code and untested by events. The subject-request lookup gap is structural rather than a bug: anything enrolled under a subject reference nobody linked stays invisible to our search, and only stating the reach makes that honest.                              |
| 182.C.39.07 | Document media/storage lifecycle, final-byte evaluation, C2PA verification, watermark/disclosure …                   | Docs  | 182.C.38           | implemented | `docs/domains/isis/human-video/media-lifecycle.md` (new)       | "SURVIVES A TRANSCODE" IS THE WRONG GOAL FOR PROVENANCE, AND THE MEASUREMENT SAYS SO. A manifest binds the SHA-256 of exact bytes; a transcode produces different bytes BY DEFINITION; so a manifest that survives one verifies as MEDIA HASH MISMATCH - the same result a verifier gives for a maliciously edited file. Carrying provenance through a transcode does not preserve it, IT MANUFACTURES A TAMPER SIGNAL. The rule therefore splits on what the operation does to the bytes: a remux is byte-identical so the manifest is carried, a transcode gets ITS OWN manifest whose edit assertion names the parent media hash. Two more measured consequences: A CUSTOM METADATA KEY IS SILENTLY NOT WRITTEN INTO MP4 AT ALL (so the transport must use a recognised one), and `-map_metadata -1` STRIPS IT WITH ONE FLAG - which means THE ABSENCE OF PROVENANCE IS NOT EVIDENCE OF ANYTHING and the release proof, not the file, is the record. THE THING YOU MEASURED AND THE THING YOU SHIP ARE DIFFERENT FILES UNLESS SOMEBODY CHECKS: a master evaluated before encoding and shipped as a transcode leaves a proof about bytes nobody received, which is why `final-bytes-binding` carries the shortest reason on the non-waivable list - without it the proof is about nothing in particular. MEDIA NEVER RIDES INSIDE A PAYLOAD, and the repo contains its own counter-example: a provider that projects a clip to data:video/mp4;base64. AND READING A PROOF HAS AN ORDER: which bytes, then measured-versus-observed (a `passed` threshold whose margin is inside the NOISE FLOOR is not a pass anybody should quote), then evaluator versions, then which gates were waived, then whether the delivery still matches.                                                                                                                                                                                                                                                      | The C2PA table is transcribed from the committed measurement at evidence/2026-09-11-provenance-transcode/ (ffmpeg 6.1.1, reproducible with its own measure.sh) rather than reasoned about. Every module named exists on disk. The non-waivable reasons are quoted from `NOT_WAIVABLE_BECAUSE`, and the proof-reading order follows `human-video-release-proof.ts` own argument about `observed_outcome` versus `threshold_measurement`. Prettier-formatted. No code changed.                                                                                                                       | No provider spend; no model call.                                    | NO RELEASE PROOF HAS BEEN WRITTEN FOR A REAL GENERATED TAKE, because no real take has been through the pipeline. The provenance measurements are real and reproducible; everything they serve is untested end to end.                                                                                                                                                                  |
| 182.C.39.08 | Create a concise V1 migration guide for existing avatar, voice, and image-to-video callers …                         | Docs  | 182.C.01.11 (V1.2) | implemented | `docs/domains/isis/human-video/v1-migration.md` (new)          | MIGRATING IS NOT A RENAME, BECAUSE THE NEW FLOW MAKES PROMISES THE OLD ONE NEVER MADE AND THOSE PROMISES NEED INPUTS THE OLD REQUEST DOES NOT CARRY. The guarantee table is read from `FLOW_GUARANTEES`: text-to-video and image-to-video have no exact dialogue, no measured sync and no authorization-bound identity, and they are NOT being switched off - what changed is that we now know what the guarantee would require, so shipping them SILENTLY would be a different thing than shipping them before we knew. Two reporting consequences land before any code change: A FLOW THAT DOES NOT MEASURE SPEECH IS NOT A FLOW THAT FAILED THE GATE (opposite findings, one word covers both), and A LEGACY JOB THAT ROUTES THROUGH A LIP-SYNC LANE DOES NOT ACQUIRE THE GUARANTEE - it produces a mouth matching SOMETHING nobody measured. THERE IS NO ONE-CLICK UPGRADE, and the reason is not ergonomic: a system offering one produces jobs WHOSE AUTHORIZATION WAS INVENTED AT THE MOMENT OF THE CLICK. AND YOU CANNOT MIGRATE CONSENT - the part that surprises integrators, so the guide plans around it: `referenceImage: string` says nothing about role, whose face it is, or whether anybody agreed, so a stored reference has exactly two honest destinations (map to `first-frame`, which is a faithful translation of what the old path DID with the string, or refuse and ask) and NEVER `identity`. The migration plan for a reference library is therefore A PEOPLE PLAN, NOT A DATA PLAN - every subject re-enrols - and that is the long pole. VERITAS GETS AN HONEST "NOT YET": the surface is V1.2-deferred behind `release-scope.ts`, 182.C.13.02 and 182.C.31.07 are still open, and documenting a migration path to a surface a caller cannot reach would be describing a fiction.                                                                                                                                                                             | The guarantee table and the per-flow upgrade requirements are read from `FLOW_GUARANTEES` in `human-video-legacy-flows.ts`; the two honest reference destinations and the never-`identity` rule from `human-video-reference-migration.ts`; the Veritas deferral from `apps/oshun/bff/src/middleware/release-scope.ts` (services stay registered, the member surface 404s with a machine-readable reason, operator/admin/studio exempt); and the 403 on `hv.consent.missing` from the failure catalog. Every module and sibling document named exists on disk. Prettier-formatted. No code changed. | No provider spend; no model call.                                    | THE VERITAS HALF CANNOT BE WRITTEN UNTIL V1.2 RESTORES THE SURFACE - 182.C.13.02 and 182.C.31.07 are the items, and both stay open. The rest of the guide describes a target flow that is not mounted, so no migration has been performed or rehearsed.                                                                                                                                |

### 182.C.40 Final Evidence and Definition of Done

| Item        | Task (condensed; canonical text in `TODOS/phase-182.md`)                                                             | Owner       | Dependency            | Status      | Changed Files                                                                                                                                                                                                                                                                                                                                                                                                                                               | Verification                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | Provider/Hardware Evidence                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Cost           | Blocker                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
| ----------- | -------------------------------------------------------------------------------------------------------------------- | ----------- | --------------------- | ----------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 182.C.40.01 | Demonstrate at least one production-configured active-scene route using multiple authorized references and exact …   | Release Eng | 182.C.39              | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.40.02 | Demonstrate identity continuity across multiple shots and at least one profile/occlusion/high-motion challenge …     | Release Eng | 182.C.39              | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.40.03 | Demonstrate multi-speaker attribution or explicitly keep it feature-flagged if its benchmark gate is not met.        | Release Eng | 182.C.39              | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.40.04 | Demonstrate creator-approved script/audio, delivered-word validation, real frame/audio sync evaluation …             | Release Eng | 182.C.39              | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.40.05 | Demonstrate automatic rejection and selective retry/reroute on intentionally bad sync, wrong words, wrong speaker …  | Release Eng | 182.C.39              | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.40.06 | Demonstrate consent expiry/revocation, tenant isolation, provider retention disclosure, subject deletion, takedown … | Release Eng | 182.C.39              | wired       | `evidence/2026-09-12-erasure-bound-live/` (nine captures plus three boot lines), `apps/oshun/bff/src/generation/human-video-erasure-runner.ts`, `human-video-destruction-store.ts`, `human-video-revocation.ts`, `human-video-release-handlers.ts`, `human-video-runtime.ts`                                                                                                                                                                                | Six things demonstrated, five live against the real Postgres and a versioned MinIO bucket through real BFF routes, with thirteen of the fourteen ports bound and only `releases` answering 503. CONSENT EXPIRY: a lapsed permission answers 403 `hv.consent.expired`, terminal, with the window in the detail - expired and scope-exceeded are deliberately different codes because one is renewed and the other renegotiated. REVOCATION: live here and in `../2026-09-12-jobs-adapter-live/`. TENANT ISOLATION: one tenant asking for another's job by its real id gets the same answer as for an id that does not exist. PROVIDER RETENTION DISCLOSURE: a `provider-retention-disclosure` obligation in every manifest, sourced from `ZDR_FACT` rather than restated. SUBJECT DELETION: the take's master, the proxy no table records, the uploaded reference photograph and the approved audio destroyed and confirmed gone across BOTH byte stores, with no surviving object version. TAKEDOWN: the withdrawal covered every take of every job plus the job's orphaned attempt. The sixth, no public candidate URL before release, is a namespace RULE rather than a capture - `canRead(ns, 'delivery')` is true only for `released`, pinned by a test over all six namespaces - and nothing on this deployment mints a delivery URL at all.                                                                                                                          | Nine dated captures, grepped for key material before committing; the server was torn down and `:4010` confirmed free, and the throwaway env script deleted. A NEGATIVE CONTROL FOUND THE ONLY DEFECT: a withdrawal narrower than the whole permission correctly kept the references and reported `complete: true` with "every copy of this likeness known to the system has been destroyed", because a location nothing was going to destroy had been left out of the reach entirely. `takedownReach` now takes a `retained` list that reports it as still there with its own reason, and the before and after are both in the evidence folder. Not claimed as `live-probed`: that status's minimum evidence is a provider request/response fixture and an output artifact digest, and a deletion makes no provider call. | Bounded canary | Two things are NAMED RATHER THAN CLAIMED. `complete: true` is unreachable on any lane in this build because `provider-side` retains and exposes no deletion endpoint - that is the retention disclosure being honest rather than a gap, and the item asks for the disclosure to be demonstrated rather than absent. And face embeddings are not destroyed because nothing writes one: `human-video-embedding-store.ts` has no importer, which a drift-guard test asserts with a control proving the pattern can fire, and the test says what has to happen the day it does - destroy the records for the withdrawn authorization, which already carry its id, and report `face-embeddings` in `copiesPresent`. Every take in the captures was SEEDED rather than rendered, because nothing renders here yet; the deletion path is proven against real bytes in real stores. |
| 182.C.40.07 | Meet documented quality targets with statistically adequate held-out results across active motion, locales …         | Release Eng | 182.C.39              | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.40.08 | Meet documented latency, reliability, cost-per-releasable- second, accessibility, security, privacy …                | Release Eng | 182.C.39              | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
| 182.C.40.09 | Run and archive all applicable targeted lint, typecheck, unit, integration, contract, build, Playwright/mobile …     | Release Eng | 182.C.39              | implemented | `docs/domains/isis/human-video/evidence/2026-09-12-release-evidence-run/README.md` (new); fixes in `human-video-metadata-strip.ts` and `human-video-take-comparison.ts`                                                                                                                                                                                                                                                                                     | TWELVE CATEGORIES, EACH RUN RATHER THAN ASSERTED. 0 lint errors over all 504 `human-video-*.ts`; 0 typecheck errors with ALL 504 IN ONE PROGRAM and a deliberately broken control file that reported; 5442 BFF + 215 web + 128 mobile unit tests; 40 contract tests plus the SDK drift guard and the two client verifiers; a passing build; 11 web and 5 admin Playwright with 0 axe violations; 46 performance; 50 release-gate; the provider canary archived from 2026-09-11 rather than re-run, because re-running it spends money to re-learn a measured fact. RUNNING THE LINT AS ONE PROGRAM FOUND THREE REAL ERRORS PER-FILE RUNS HAD MISSED, all `no-useless-assignment`, because a per-item lint only ever saw the file that item touched: two dead `cursor` assignments before a `break`, and `let within = 0` in the playhead pairing - AND ZERO IS A REAL POSITION INSIDE A WORD, so a default surviving a missed branch would read as "at the very start of it" rather than as a bug. It is declared without a value now. The typecheck over all 504 found NOTHING, which is itself the result worth recording: the same exercise on 2026-09-11 found 32 errors accumulated unread across 18 files.                                                                                                                                                                                                                                                           | Each number above came from a run in this session. The typecheck and both client verifiers were proved live - the control error reported, and each verifier was broken in four files in turn and restored.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Bounded canary | "Applicable" is doing work in four rows and each says so: the Maestro flow needs a device, the performance suites are models rather than a load test against a deployed lane, the canary was archived rather than re-run, and the accessibility row is the AUTOMATED half only - 182.C.36.08's manual screen-reader, zoom and contrast passes stay open. Nothing here demonstrates the product end to end; that is 182.C.40.01-.08.                                                                                                                                                                                                                                                                                                                                                                                                                                         |
| 182.C.40.10 | Verify documentation, OpenAPI/SDKs, operator runbooks, capability matrix, dashboards, alerts, flags, rollback …      | Release Eng | 182.C.39              | implemented | `docs/domains/isis/human-video/evidence/2026-09-12-live-surface-capture/` (new: README, `presets-200.json`, `identity-packs-503.json`, `presets-403-domain.json`, `boot-line.txt`); corrections to `api-guide.md` and `creator-workflow.md`                                                                                                                                                                                                                 | VERIFIED AGAINST A RUNNING SERVER RATHER THAN AGAINST THE SCHEMA. TWO DOCUMENTS OPENED WITH A CLAIM THAT HAD STOPPED BEING TRUE: `api-guide.md` said "Nothing here is mounted" and `creator-workflow.md` said "None of this is reachable today". Both were accurate when written and both went stale on 2026-09-11 when the surface was mounted - and a reader of either would have concluded the feature does not run, which is the exact failure this item exists to catch. The API guide now carries FOUR RESPONSES CAPTURED FROM A LIVE SERVER where it previously carried none and said so. THE SCOPE BOUNDARY WAS EXERCISED AGAINST A RUNNING SERVER FOR THE FIRST TIME: `hv:*` and `hv:projects:*` answer 200, and `domain:*`, `hv:jobs:*` and a bare `*` are all refused - it had only been unit-tested. AND THE BOOT LINE CARRIES THE FINDING WORTH KEEPING: three ports report "(no … table)" while the tables EXIST, because the committed generated Prisma client predates the models and `createHumanVideoRuntime` treats a stale client as no database deliberately - binding a missing delegate would give `undefined.findUnique` on the first request, a 500 that tells nobody anything.                                                                                                                                                                                                                                                                   | Four live captures plus the boot line, archived with the method and grepped for key material. Secrets were sourced through a throwaway /tmp script deleted at the end; the server was torn down and :4010 confirmed free. SDK-to-surface is covered by the generator's byte-identical drift guard (17 tests); client-to-surface by `verify-client-api-calls.mjs` across all seven hand-written clients.                                                                                                                                                                                                                                                                                                                                                                                                                   | Bounded canary | Dashboards, alerts, rollback and incident drills were checked as documents rather than exercised - there is no deployed dashboard to compare against, and an incident drill needs an incident. What was verified against DEPLOYED BEHAVIOUR is the API surface, the SDK, the clients, the boot line and the two creator-facing documents.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
| 182.C.40.11 | Leave unsupported modes visibly disabled or fail-loud; do not close the phase with a plan-only adapter …             | Release Eng | 182.C.39              | implemented | `docs/domains/isis/human-video/evidence/2026-09-12-fabrication-audit/README.md` (new); audited against `human-video-runtime.ts`, `human-video-provider-double.ts`, `human-video-poll-policy.ts`, `human-video-catalog-client.ts`, `human-video-alignment-adapter.ts`, `libs/contracts/src/human-video/alignment.ts`, `capability-audit.md`, `libs/isis/seedance-provider/src/lip-sync-validation.test.ts`, `apps/psyche/avatar-engine/CAPABILITY-STATUS.md` | FOUR PROHIBITED THINGS AND ONE REQUIRED PROPERTY, each checked against the repository rather than against a module's own doc comment. NO PLAN-ONLY ADAPTER: `createHumanVideoRuntime(null)` binds exactly ONE thing - the preset catalogue, which has no store to be missing because it serves the product's own vocabulary - and refuses eleven with a per-request 503 rather than an empty list. Three ports bind only as ALL-OR-NOTHING groups and the boot line names the missing half, because the half-bound states are the dangerous ones: a session table with no sink takes bytes and drops them, a pack store with no detector accepts a pack nothing examined, an aligner with no probe places every end-of-take window against nothing. NO MOCKED GENERATION: the provider double has no importer outside its own spec, and `Math.random` appears in exactly two modules, both annotated and both retry jitter. NO SUPPLIED METRIC: no handler reads a numeric metric from a body; the decision handlers read a `screenDigest` and acknowledgements, which are DECISIONS; `submit-audio` refuses a client-supplied alignment outright. AND THE THREE SEVERITY-H FINDINGS FROM THE 2026-09-09 CAPABILITY AUDIT ARE CLOSED: the fabricated phoneme track now FAILS LOUD, the fabricated VMAF file is DELETED, and the unrunnable avatar engine is RETIRED FROM CAPABILITY CLAIMS with a status document whose facts are re-derived by a test so it cannot drift. | The unbound list was enumerated from the runtime rather than read off a document. The provider-double importer check is a grep across `apps/oshun/bff/src` excluding specs. Unsupported modes were verified on all four surfaces: 28 endpoint-level 503 refusals in `human-video-handlers.spec.ts`, the Studio's two record-making acts declining in a real browser (11 e2e passed), the admin's mutations off with a stated reason, and the mobile resume refused rather than sent.                                                                                                                                                                                                                                                                                                                                      | Bounded canary | ONE SOFT SPOT, NAMED RATHER THAN FIXED: `adaptAlignerOutput` writes `confidence: word.confidence ?? 0`, and zero is the worst default for a confidence - it reads as "measured, and certainly wrong" where the truth is "the aligner did not say". The contract forces it (`alignment.ts:50` requires a number in [0,1] with no way to express "not reported"). Unexercised today: the bound CTC aligner reports a confidence for every word and nothing gates on it. Making it expressible is a contract change.                                                                                                                                                                                                                                                                                                                                                           |
| 182.C.40.12 | Check this final item only after all prior non-deferred items are complete, approved exceptions have owners/expiry … | Release Eng | all prior 182.C items | not-started | —                                                                                                                                                                                                                                                                                                                                                                                                                                                           | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          | —                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         | Bounded canary | None                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
