Eve library

Phase 182: Controlled Human Video Generation — Multi-Reference Identity, Active Performance, Exact Dialogue, Native Audiovisual Lip Sync, Quality-Gated Delivery, and OpenRouter Routing

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Build the production video system that lets a creator provide one or more authorized reference images of a person, specify exactly what that person must say, direct how they move and perform in an arbitrary scene, and receive a coherent video whose visible speech is synchronized to the approved audio. The system must support active people in cinematic and everyday situations—not just front-facing talking heads—and it must preserve identity, voice, wardrobe, scene, camera, motion, and dialogue intent across shots.

This phase treats “perfect lip sync” as a delivery requirement, not an honest claim that every model's first sample will be perfect. Oshun may use a native audio-conditioned model so generation and synchronization happen in one render, but it must still inspect the actual output frames and audio. A take that misses words, assigns speech to the wrong person, drifts out of sync, hides the mouth for an unjustified duration, damages identity, or crosses a release threshold must be rejected, retried, rerouted, or sent for explicitly disclosed corrective processing. The product promise is therefore: no known bad-sync take is released, while measured pass rates and limitations remain visible.

The primary path for exact wording is:

authorized references → approved script → controlled speech audio → shot plan → native audio-conditioned video candidates → audiovisual/identity/safety evaluation → select or retry → encode/provenance → release

Supplying speech audio to the video model is not a separate post-generation lip-sync step. Prompt-only generated dialogue remains an optional creative mode, not the exact-dialogue mode, because a video model may paraphrase, omit words, mispronounce names, or change timing.

Primary owning domains: Isis, Oshun BFF, Nous, Psyche, Veritas, Contracts, Studio Web/Desktop/Mobile, Admin, Trust & Safety, Media Infrastructure, MLOps, and Release Engineering.

Existing anchor points to extend rather than duplicate:

  • libs/isis/curated-cards/src/card-types.ts
  • apps/oshun/bff/src/generation/video-executor.ts
  • apps/oshun/bff/src/generation/openrouter-video-provider.ts
  • apps/oshun/bff/src/generation/video-provider-env.ts
  • apps/oshun/bff/src/generation/provider-quality-router.ts
  • apps/oshun/bff/src/generation/comfy-video-provider.ts
  • apps/oshun/bff/src/generation/comfy-video-longform.ts
  • apps/oshun/bff/src/generation/video-probe.ts
  • apps/oshun/bff/src/generation/video-governance.ts
  • apps/oshun/bff/src/generation/video-hero-gate.ts
  • apps/oshun/bff/src/generation/generated-artifact-object-store.ts
  • apps/oshun/bff/src/generation/release-gate.ts
  • libs/isis/ai-video/
  • libs/isis/ai-providers/src/providers/video-generation/
  • libs/isis/workflows/src/consistency/
  • libs/nous/portrait-animation/
  • libs/oshun/embodiment-psyche/src/avatar-sync.ts
  • apps/psyche/avatar-engine/
  • apps/veritas/video/
  • apps/oshun/bff/src/studio/voice-consent-store.ts
  • apps/oshun/bff/src/admin/persona-lifecycle-store.ts
  • V1/architecture/persona-avatar-voice-packs.md
  • CONTENT_GENERATION_CAPABILITY_MATRIX.md
  • libs/isis/seedance-provider/ (about 40 modules: fal, PiAPI, and Volcengine Ark clients, audio-reference.ts, reference-assets.ts, character-consistency.ts, video-extension.ts, video-inpainting.ts, voiceover-control.ts, lip-sync-validation.ts, expiry-manager.ts)
  • apps/oshun/bff/src/generation/jobs-route.ts (the durable job envelope and its GenerationJobStatus enum: queued, running, complete, failed, blocked, needs_review)
  • apps/oshun/bff/src/generation/generation-budget-gate.ts, generation-quota-gate.ts, request-validation.ts
  • apps/oshun/bff/src/generation/generated-artifact-erasure.ts, media-releasability.ts, promotion-gate.ts, provider-measurement.ts
  • apps/oshun/bff/src/generation/synthesis-disclosure.ts, video-watermark.ts, audio-c2pa-signer.ts, audio-probe.ts
  • apps/oshun/bff/src/generation/dub-faithfulness.ts (an injectable ASR-to-source faithfulness seam that fails loud when unconfigured)
  • apps/oshun/bff/src/generation/narration-executor.ts, narration-provider-env.ts (the existing TTS path and its disclosure gate)
  • apps/oshun/bff/src/generation/runpod-comfy-client.ts, libs/isis/runpod-surface/
  • apps/oshun/bff/src/generation/release-proof.test-fixtures.ts and libs/contracts/src/common/proof-types.ts (ProofRecordSchema, nine named release gates)
  • libs/oshun/generation-control-isis/ (the release gate the worker already enforces)
  • apps/oshun/bff/src/aja/upload-consent-store.ts and libs/aja/consent-management/
  • libs/oshun/trust-safety/src/policy-taxonomy/categories.ts
  • apps/oshun/bff/src/isis/voice-cloning-store.ts (an explicitly unbound projection with no consent ledger or biometric store)
  • libs/isis/portrait-studio/, libs/isis/face-synthesis/, libs/isis/video-edit-studio/, libs/isis/video-enhancement/
  • apps/veritas/video/src/providers/{whisper,assemblyai,deepgram}.ts (the only ASR clients in the repo)
  • docs/agents/model-cost-openrouter.md (has no video pricing yet)
  • libs/oshun/navigation/src/release-scope.ts (V1.0 ships four rooms; Veritas and Metis return in V1.2)

Current-State Finding#

The existing general-video card and BFF request expose a text or motion prompt, one optional reference image, duration, aspect ratio, style, camera movement, resolution, locale, and a hero flag. They do not yet carry a script, speech audio, multiple typed references, speaker/timeline data, performance direction, native-audio requirements, lip-sync thresholds, or an output-quality decision. The OpenRouter adapter currently implements create/poll/download with a single first-frame image and returns the full MP4 as a base64 data URI; it does not yet cover typed image references, provider passthrough options (the only way audio or video inputs reach a model on OpenRouter), final-frame guidance, generated-audio controls, signed callbacks, capability discovery, or durable storage. Per-provider quality routing does already exist: provider-quality-router.ts runs Beta–Thompson selection over each provider's measured release rate, and Phase 182 must extend it to per-model, per-cohort keys rather than replace it.

Oshun V1 contains useful consent, persona, voice, identity, provider, and viseme concepts, but it does not currently prove this flow end to end. The existing 39-phoneme-to-11-viseme mapping is animation data rather than a video renderer. Several portrait-animation modules produce plans or accept externally supplied quality scores rather than executing a measured inference/evaluation pipeline. The Psyche avatar engine and Veritas avatar providers must be audited and either hardened, integrated behind the canonical contracts, or explicitly retired; their existence must not be represented as proof of active-scene lip sync.

The repo also already contains @isis/seedance-provider, a roughly 40-module library with fal, PiAPI, and Volcengine Ark clients plus audio-reference, reference-asset, character-consistency, video-extension, video-inpainting, voiceover-control, and lip-sync-validation modules. It is the natural direct lane for the reference and editing capabilities OpenRouter does not expose, and it must be audited before anything is built beside it: its lip-sync-validation.ts synthesizes per-utterance confidence from the frame index (0.7 + (index % 3) * 0.08 at line 324), which is a fabricated metric, not a measurement. The BFF likewise already has a durable generation job envelope with a six-state status enum, budget and quota gates, artifact erasure, media releasability, a promotion gate, provider measurement, synthesis disclosure, video watermarking, audio C2PA signing, an ffprobe-based video probe, an ASR-to-source faithfulness seam, a TTS narration path, and a ProofRecordSchema with nine named release gates. Every one of those is an extension point for this phase, not something to rebuild.

OpenRouter Routing Snapshot (Verify Dynamically Before Implementation)#

The OpenRouter video catalog is mutable. The table below was re-verified on 2026-09-09 against the live GET /api/v1/videos/models response and the official video-generation guide. It corrects an earlier draft that assumed audio and video references travel in input_references; they do not. Re-checked against the live catalog on 2026-09-10: every lane row below is unchanged.

Lane Candidate What OpenRouter actually exposes today Required proof before promotion
Active human + supplied audio + references alibaba/wan-2.7 First/last frame, 2–10 s, 720p/1080p, generate_audio, seed; passthrough audio, video, videos, images, last_image, ratio, negative_prompt, prompt_extend. Already the repo default. Live probe that the audio passthrough accepts approved speech and yields measured sync; identity gates on images/videos
Active human + supplied audio (fallback) alibaba/wan-2.6 First frame only, 5–10 s, 720p/1080p; passthrough audio, shot_type, size, negative_prompt, enable_prompt_expansion. No image/video reference passthrough. Same audio probe; identity must come from the first frame plus input_references images
Native (prompt-only) dialogue, multi-speaker bytedance/seedance-1-5-pro First/last frame, 4–12 s, 480p–1080p, generate_audio, seed; passthrough watermark, req_key only. No audio input, so creative-dialogue mode only. Speaker attribution, ASR word accuracy, profile/occlusion, cross-language benchmarks; never labeled exact-dialogue
Longer takes + image references bytedance/seedance-2.5 Image-only input_references, first/last frame, 4–30 s, 480p/720p, seed; passthrough watermark, req_key, output_format. No audio/video references, edit, or extend via OpenRouter; those need the direct fal/PiAPI/Ark clients in @isis/seedance-provider. Live reference-count probe; continuity benchmarks; direct-lane parity before any edit/extend claim
Premium short hero shot (prompt-native audio) google/veo-3.1 First/last frame, 4–8 s, 720p–4K, generate_audio, seed; passthrough personGeneration, aspectRatio, negativePrompt, conditioningScale, enhancePrompt. No audio input; never an exact-dialogue lane. Price, latency, resolution, control adherence; prompt-native dialogue accuracy reported, not promised
Talking-head-only lane heygen/avatar-iv Script (prompt) plus passthrough voice_id, voice_settings, motion_prompt, expressiveness, fit, remove_background, background, caption, title; no seed, no generate_audio, duration follows the script. Supplied audio is not exposed on OpenRouter. Keep separate from active-scene claims; benchmark crop, gesture, and background constraints; bind voice_id to voice consent

Other catalog entries worth a bounded probe, not yet lanes: kwaivgi/kling-v3.0-pro and -std (first/last frame, negative_prompt, cfg_scale), minimax/hailuo-3 (instruction-guided video editing, 2K), runway/aleph-2 (in-context editing with keyframes), openai/sora-2-pro (multi-shot, no frame images), and alibaba/wan-3.0 and -prime (2–30 s, no passthrough at all).

Wire facts verified on 2026-09-09 that the checklist below depends on:

  • Endpoints: POST /api/v1/videos, GET /api/v1/videos/{id}, GET /api/v1/videos/{id}/content?index=0 (Bearer-authenticated download), GET /api/v1/videos/models.
  • Request fields: model, prompt, duration, resolution, aspect_ratio, size, frame_images, input_references, generate_audio, seed, callback_url, provider. input_references is documented as "reference images for style guidance" and each entry is { type: 'image_url', image_url }. There is no audio or video reference type; audio and video inputs exist only as model-specific provider passthrough parameters named in the catalog's allowed_passthrough_parameters. frame_images takes precedence over input_references when both are present.
  • generate_audio defaults to true on audio-capable models; an exact-audio request must set it deliberately per adapter.
  • Catalog record fields: id, a dated canonical_slug (pin this as the model version), supported_durations, supported_resolutions, supported_aspect_ratios, supported_sizes, supported_frame_images, generate_audio, seed, pricing_skus, allowed_passthrough_parameters. The catalog carries no reference-count limit, retention, region, or data-handling fields; those must be maintained as dated manual records.
  • pricing_skus are heterogeneous: per second by resolution, per second with or without audio, per generation mode, or per video_tokens. Cost estimation needs a normalizer, not a single per-second rate.
  • Webhooks: callback_url must be HTTPS and overrides a workspace-level default; payloads are HMAC-SHA256 signed in X-OpenRouter-Signature; event types are video.generation.completed, .failed, .cancelled, and .expired (expired means the job exceeded its maximum time to live; generation_id and model may be null on early failures).
  • Zero data retention: video generation is not eligible for ZDR, and an account with ZDR enforcement enabled receives no video routing at all. A tenant that requires ZDR is therefore ineligible for every OpenRouter video lane, not merely warned.

Never make this table a permanent hard-coded truth. Fetch and cache the live model catalog, intersect requested capabilities with observed provider support, and fail closed when a required feature is missing. Official references to track:

Non-Negotiable Product Rules#

  1. Active-scene generation is a first-class lane; a talking-head avatar is not an acceptable fallback unless the creator explicitly selects that lane.
  2. Exact dialogue mode uses creator-approved audio or deterministic TTS derived from the approved script; prompt-native dialogue alone cannot claim verbatim wording.
  3. Provider marketing, request acceptance, job completion, metadata timestamps, and the presence of an audio track are not quality evidence. Quality gates must inspect decoded frames and audio from the delivered bytes.
  4. No model or provider is the permanent global default. Routing is based on required capabilities, live health, benchmarked quality, locale, policy, latency, and budget.
  5. Identity and voice inputs require authorization, revocation, retention, and impersonation controls before provider submission.
  6. Candidate generation, evaluation, selective retry, and release proof are one workflow. A completed provider job is not automatically a releasable asset.
  7. Provider-specific power remains available through typed options, but the creator experience uses stable, provider-neutral creative controls.
  8. Every released asset retains its request, model/version, provider, seed, references, approved script/audio digest, measurements, provenance, and release decision without storing more biometric material than necessary.

Repo Constraints That Bind This Phase#

  • Model binding for tests and harnesses. Every LLM-shaped step this phase adds (brief-to-shot planning, prompt compilation, semantic-substitution checks, review triage) 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. Video generation canaries bind tiny bounded spend. docs/agents/model-cost-openrouter.md carries the video pricing_skus normalization and per-lane prices (fetched 2026-09-09, max age 24 h); 182.C.14.06 reads that record rather than re-deriving it.
  • Hardware (corrected 2026-09-10; read before marking anything blocked-local). The phase was drafted against a 16 GB MacBook Air. The worktree that implements it runs on Linux, 15 GiB RAM, 4 cores, no GPU (nvidia-smi absent; torch is the +cpu build). CPU probes on 2026-09-10 ran every evaluator class the draft called hardware-blocked: YuNet face detection with SQPnP head pose, SFace identity embeddings, a SyncNet v2 forward pass at about 1 s of CPU per second of 25 fps video, and a wav2vec2 CTC forced aligner at about 0.2 s of CPU per second of audio (evidence under docs/domains/isis/human-video/evidence/2026-09-10-*). So 182.C.17, 182.C.19, and 182.C.20 are built and probed on CPU here first; a hosted GPU worker (RunPod via libs/isis/runpod-surface and runpod-comfy-client.ts) is a throughput decision taken at 182.C.17.13 from the numbers measured at 182.C.17.12, not a prerequisite. Only self-hosted video generation (182.C.32.03–.08) is genuinely GPU-bound. A blocked-local row must name what was run and how it failed; a blocker copied across a section is a claim about none of its items. Model weights are staged in ~/.cache/oshun-human-video/ (digests in the evidence records); Python packages go in a venv and never touch the pnpm lockfile. The memory checkpoint before any heavy run is free -m on this box (memory_pressure is macOS-only); below about 3 GB available, tear down first.
  • ASR (decided and funded 2026-09-10, 182.C.02.12). Transcription binds OpenRouter audio-input models: mistralai/voxtral-small-24b-2507 primary, google/gemini-3.1-flash-lite as the cheap default, with 46 audio-input models on the account for ensembles (182.C.18.07). Their word timestamps are disqualified for any timing measurement (three models disagreed by up to 1013 ms on one 3.95 s clip), and so are Whisper's cross-attention timestamps (up to 600 ms onset error and merged words on a known-onset clip). Word timing comes only from a CTC forced aligner run on our own CPU (wav2vec2 CTC: max 81 ms raw onset error with a systematic late bias, on the known-onset fixture). Veritas's Whisper, AssemblyAI, and Deepgram wrappers are not used by this phase.
  • Release train. V1.0 ships four rooms; Veritas and Metis return in V1.2 (libs/oshun/navigation/src/release-scope.ts). Items that touch Veritas (182.C.13.02, 182.C.31.07, 182.C.39.08) are gated on V1.2, and the creator flow must declare its train before 182.C.38.
  • Secrets on the Linux box. Keys live in ~/.config/oshun/env-master.env (mode 600; the Mac's drawer file does not exist here, and the worktree .env holds only Tara settings). Source it through a throwaway script (set -a; . ~/.config/oshun/env-master.env; set +a), verify by length (echo ${#OPENROUTER_API_KEY}), and never put a key in a committed file or in command text.
  • Surfaces. No Oshun studio desktop app exists. The studio lives in apps/oshun/web routes (studio, lilith-studio, mawu-studio, egbe-studio); the desktop apps are apps/yemaya/studio-desktop (Electron, film/game), apps/lilith/desktop, and apps/iris/desktop. 182.C.24 must name a host or drop desktop from scope.
  • frontend-skill. Phases 72 and 139 cite it as mandatory, but it is not installed in every working environment. On the Linux box it resolves at /root/.codex/skills/frontend-skill/SKILL.md (a Codex skill, absent from the Claude skill list); read that file before 182.C.23 and 182.C.29.

What the 30 unchecked items are waiting on (swept 2026-09-12)#

Every remaining item was walked in one pass and each one traces to one of a short list of root blockers. The list is short on purpose: pick the blocker that has become available and the items it releases are named here rather than rediscovered — which means an entry that has been discharged must be REWRITTEN rather than left standing, or the next session re-derives work that is already done. Three were rewritten on 2026-09-13 for exactly that reason, and one of them had never been written down at all.

  1. Consented human subjects (182.C.33.01–.05). Releases 182.C.17.08, 18.08, 19.05, 19.08, 19.09, 20.01, 20.07, 32.04 and most of 182.C.40. The machinery for all of them is built and refuses to report a result it does not have; what is missing is people who agreed to be filmed. 19.09's module says so in its own header, which is why that box is not ticked on the strength of the slices existing.

  2. A PROVIDER-REACHABLE OBJECT ROOT — what is left of the blocker that used to read "NOTHING RUNS THE PIPELINE". That framing was correct on 2026-09-12 and is now out of date in three of its four parts, so it is restated rather than repeated. A RUNNER EXISTS: human-video-job-runner.ts on the durable envelope advances one stage per claim and was measured live reaching planning. A LANE REGISTRY EXISTS (2026-09-13): CatalogueLaneRegistry registers every route from the live catalogue and runs it through evaluateLaneEligibility — it answers NONE today, which is a fact about the request rather than about the deployment. A QUARANTINE STORE EXISTS (2026-09-13): FilesystemQuarantineStore, chunked writes, rename-on-commit, promotion decided by planTransition.

    What is left is the one piece that cannot be closed by writing code here: placeReferences hands the provider a URL it must RETRIEVE, and this box's object store is on a private port. Until a reference is fetchable by OpenRouter — and a CONSENTED one exists to fetch, which is blocker 1 — no candidate can be produced. That still releases take comparison, review and release in 182.C.36.06, delete-asset deletion (plansFor resolves an asset id to a TAKE), 36.07's quarantine and takedown screens, 182.C.24.06 (whose web half is 36.06), and most of 182.C.40. Note what it does NOT release on its own: 40.02 additionally needs 19.08's re-fit, and the content faults in 40.05 need labelled examples.

  3. DISCHARGED 2026-09-13 — was "a bound model registry". It read: releases the ESTIMATE half of 182.C.36.06 and 36.07's reroute screen, because buildStudioLaneView needs a ModelVersionRecord per lane and nothing registered one. Both halves are done. The estimate is POST /jobs/estimate, the fortieth operation, covered in a real browser against the live price list; the registry is CatalogueLaneRegistry, and the thing that had always been missing underneath both was an adapter for the catalogue, which human-video-catalogue-reader.ts now is.

    ONE PIECE OF THE OLD ENTRY SURVIVES AND IS NOT A BLOCKER: human-video-pre-submission-summary.ts is still consumed by nothing. It is 182.C.23.06's deliverable and buildStudioLaneView can feed it now, so what stands between it and a creator is a Studio region rather than a missing input. Same shape as human-video-take-comparison.ts and human-video-export-controls.ts below: client/surface drift, not a blocker.

  4. A GPU host (182.C.32.03, 32.06, and 32.04 with blocker 1).

  5. A person with a screen reader (182.C.36.08's manual half; the automation half is done and passing).

  6. The V1.2 release train (182.C.13.02, 31.07, gated by release-scope.ts) and evidence that follows from blockers 1 and 2 (182.C.38.08, which is a PROHIBITION that currently holds).

  7. DISCHARGED 2026-09-13 — a blocker nobody had noticed. THE ADMIN SURFACE ANSWERED 403 TO EVERY OPERATOR and had since it was written: an admin session carries admin:*, satisfiesScope stops the wildcard at the hv: namespace on purpose, and nothing in the service ever issued a staff scope. So 182.C.36.07's claim that review and role isolation were covered was false in the same way the web side's was on 2026-09-12 — the e2e spec tolerated a failed read, and an unreachable surface and one failing closed are indistinguishable from the client. human-video-staff-grant.ts closes it, and the queue is a PLATFORM read now: hv:reviews:all-tenants is what drops the tenant filter, demanded by buildQuery, carried on the cursor and refused to hv:*. Anything else built on the admin surface would have hit this first.

THE "NOT A BLOCKER" LIST, CORRECTED 2026-09-13. It used to say override rules and audit were "missing their whole stack, not a page". Both have a stack AND a page now, and so does reroute: /isis/human-video/routing over GET /operator/lanes/:mode and /isis/human-video/trust-and-safety over GET /operator/trust-and-safety, both read-only because the operator grant issues no scope that writes until 182.C.29.08's coverage exists, both covered in a real browser. Three of 36.07's five flows now have screens; quarantine and takedown are the two that remain and they are blocker 2.

WHAT SURVIVES, AND IT IS THE RECURRING SHAPE BEHIND THIS PHASE'S REAL DEFECTS: CLIENT/SURFACE DRIFT. 182.C.23.08, 23.09 and 23.12 all shipped as logic libraries no browser ever called, and binding the first of them on 2026-09-12 found three defects in a day. human-video-pre-submission-summary.ts was wired on 2026-09-13 and cost two more findings: buildStudioLaneView under it had never been called by anything but its own spec either, and it could not describe a multi-shot job at all — it priced ONE CostRequest, and a per-generation minimum applies per shot. human-video-take-comparison.ts and human-video-export-controls.ts are STILL consumed by nothing, and both need a candidate (blocker 2); check them before believing a surface is finished. Its sibling, found three times on 2026-09-13: A FIX THAT UPDATES ONE OF A MODULE'S TWO SPEC FILES. Three specs were sitting red against behaviour that had deliberately changed — a duplicate admin port spec, the review panel's own spec, and an older __tests__ copy of the first. Grep for every spec that names a module before believing the suite is green about it.


182.C Canonical Sequential Implementation Checklist#

This is the executable checklist for Phase 182. Work from top to bottom unless a dependency is explicitly recorded. An item is checked only when implementation, automated verification, documentation, and evidence are committed. A typed seam or a mocked provider response is not evidence that a real capability works.

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

  • 182.C.01.01 Create docs/domains/isis/human-video/implementation-tracker.md with one row per 182.C item, owner, dependency, status, changed files, verification, provider/hardware evidence, cost, and blocker.
  • 182.C.01.02 Record the stage order and prohibit checking downstream release work before contracts, consent, real evaluation, and storage are complete.
  • 182.C.01.03 Define “implemented,” “wired,” “live-probed,” “benchmark-passed,” “release-enabled,” and “blocked” so plan-only modules cannot be mistaken for working inference.
  • 182.C.01.04 Require every claimed provider capability to link a dated, redacted request/response fixture and a real generated artifact digest.
  • 182.C.01.05 Require every quality claim to link measurements computed from the exact released bytes, not a provider preview or metadata summary.
  • 182.C.01.06 Define a revalidation cadence for mutable provider models, pricing, duration limits, reference limits, option names, and retention.
  • 182.C.01.07 Add explicit owner and expiry fields for every temporary exception, disabled evaluator, manual review, or model-specific waiver.
  • 182.C.01.08 Record targeted lint, typecheck, unit, integration, provider-sandbox, Playwright, accessibility, performance, and release-gate commands as packages are created.
  • 182.C.01.09 Bind every test, harness, probe, and local stack to the cheap OpenRouter model per CLAUDE.md (OPENROUTER_PROVIDER_SORT=price), and add the video pricing_skus normalization to docs/agents/model-cost-openrouter.md.
  • 182.C.01.10 Add a blocked-local tracker status for GPU-bound evaluator and ASR work, record the hosting route for each, and require a memory_pressure checkpoint before any local ffmpeg or evaluator run.
  • 182.C.01.11 Record the target release train, the V1.2 gating of every Veritas-touching item, and the desktop-surface decision from 182.C.24.05 in the tracker before any Studio work starts.

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

  • 182.C.02.01 Audit the curated video card types and validation path; record every field needed for references, dialogue, audio, speakers, direction, quality policy, and output preferences.
  • 182.C.02.02 Audit the BFF video executor, provider environment, router, queue, catalog, storage, governance, and release gates; distinguish real code paths from fixtures and seams.
  • 182.C.02.03 Audit @isis/ai-video, @isis/ai-providers, and @isis/seedance-provider for duplicate clients, stale claimed models (the Wan provider names fal-style alibaba/wan-2-6-* ids), committed .js/.d.ts artifacts beside source, contract mismatches, reusable workflow primitives, and fabricated measurements (start with the index-derived confidence in seedance-provider/src/lip-sync-validation.ts line 324).
  • 182.C.02.04 Audit identity, scene, shot, wardrobe, and temporal consistency modules and map them to the new human-video contract.
  • 182.C.02.05 Audit every @nous/portrait-animation export and label it executable, adapter-backed, planning-only, score-ingestion-only, or stale.
  • 182.C.02.06 Audit apps/psyche/avatar-engine for entry point, dependencies, model weights, license, GPU/CPU behavior, failure behavior, tests, and whether it should be hardened or retired.
  • 182.C.02.07 Audit Veritas HeyGen, Tavus, D-ID, Runway, and unified providers; identify which are actually constructed and reachable at runtime.
  • 182.C.02.08 Audit V1 persona/avatar/voice documentation against current code and create a fact table of delivered, partial, provider-gated, and missing capabilities.
  • 182.C.02.09 Update CONTENT_GENERATION_CAPABILITY_MATRIX.md only from verified findings, leaving unprobed generation and encoding unchecked.
  • 182.C.02.10 Create a deletion/consolidation list across the five provider abstractions that already exist (@isis/ai-video, @isis/ai-providers/video-generation, @isis/seedance-provider, the Veritas unified provider, and the BFF video-provider-env.ts seam) so Phase 182 does not add a sixth.
  • 182.C.02.11 Audit the BFF generation primitives named in the anchor list (job envelope and status enum, budget/quota gates, erasure, releasability, promotion gate, provider measurement, synthesis disclosure, watermark, C2PA signer, probes, dub faithfulness, narration TTS, release-proof schema) and record an extend-or-replace decision for each before any parallel implementation is written.
  • 182.C.02.12 Decide and fund the ASR vendor for exact-speech evaluation (OpenAI has no credits and there is no Deepgram account; options are a self-hosted Whisper worker on RunPod, AssemblyAI, or an OpenRouter audio-input model) and record the decision as a blocker on 182.C.18 until it is made.

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

  • 182.C.03.01 Define supported modes: text-to-scene with generated cast, reference-to-video, exact-dialogue performance, multi-character scene, continuation/extension, video edit, and explicit talking-head avatar.
  • 182.C.03.02 Define active-person benchmark scenarios: walking, cooking, presenting, driving while stationary/safely staged, dancing, sports, working, gesturing, interacting with props, and moving through camera angles while speaking.
  • 182.C.03.03 Define out-of-scope behavior for launch: live impersonation, unauthorized public figures, unsafe stunts, invisible/off-screen dialogue represented as visible sync, and unsupported crowd dialogue.
  • 182.C.03.04 Define exact-dialogue semantics: canonical script, approved pronunciation, allowed disfluencies, expected language, word-error target, timing tolerance, and whether nonverbal vocalizations are permitted.
  • 182.C.03.05 Define “perfect delivered sync” as thresholded audiovisual alignment plus exact speech, correct speaker attribution, identity preservation, visible-mouth coverage, and absence of critical artifacts.
  • 182.C.03.06 Define separate quality tiers for draft, standard, production, and hero delivery, including candidate count and retry budget.
  • 182.C.03.07 Define a deterministic failure message for requests no currently healthy provider can satisfy; never silently downgrade an active scene to an avatar or exact audio to prompt-only dialogue.
  • 182.C.03.08 Define creator-visible limitations for side profiles, occlusion, very fast speech, singing, shouting, whispering, overlapping dialogue, extreme motion blur, masks, and tiny faces.

182.C.04 Architecture Decisions and Ownership#

  • 182.C.04.01 Write an ADR naming one canonical human-video request, provider capability, job state, measurement, and release-proof contract.
  • 182.C.04.02 Decide which types belong in shared contracts versus Isis implementation libraries versus the Oshun BFF boundary.
  • 182.C.04.03 Decide whether @isis/ai-video or @isis/ai-providers/video-generation is the provider SDK of record and migrate/retire the other path without breaking consumers.
  • 182.C.04.04 Keep provider credentials, webhooks, downloads, and biometric uploads at server boundaries; prohibit browser/mobile access to provider secrets.
  • 182.C.04.05 Define ports for asset ingestion, speech production, shot planning, capability discovery, generation, evaluation, correction, encoding, provenance, and release.
  • 182.C.04.06 Define durable event ownership and compatibility rules for BFF, workers, Studio clients, Veritas, and admin tools.
  • 182.C.04.07 Define how direct providers, OpenRouter, Comfy/RunPod, and future self-hosted models implement the same capabilities without lowest- common-denominator payloads.
  • 182.C.04.08 Produce a dependency diagram and data-flow threat model covering every copy of face images, voice audio, intermediate clips, and evaluation artifacts.

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

  • 182.C.05.01 Add a versioned HumanVideoGenerationRequest schema with tenant/project, mode, creative brief, references, cast, dialogue, audio, shots, controls, output, quality policy, budget, consent, and callback.
  • 182.C.05.02 Add typed reference roles: identity, face-angle, full-body, wardrobe, pose, composition, first-frame, last-frame, motion, performance, environment, prop, voice, and source video.
  • 182.C.05.03 Allow multiple references per role with priority, subject binding, temporal range, strength, notes, and provider-safe transformation metadata.
  • 182.C.05.04 Add cast and speaker IDs so every dialogue segment, reference, blocking instruction, and evaluator result binds to the correct person.
  • 182.C.05.05 Add canonical dialogue segments with exact text, locale, pronunciation lexicon, emotion, delivery, start constraint, overlap rule, and audio asset/digest.
  • 182.C.05.06 Add provider-neutral controls for duration, aspect, resolution, FPS, camera, lens, framing, shot size, motion, blocking, expression, gaze, gesture, pace, lighting, style, negative constraints, seed, and continuity.
  • 182.C.05.07 Add requiredCapabilities distinct from preferences so the router can fail closed instead of silently discarding native audio, multi-reference, multi-speaker, extension, or resolution requirements.
  • 182.C.05.08 Add a quality policy with metric thresholds, candidate and retry budgets, allowed fallback classes, human-review rules, and maximum spend/latency.
  • 182.C.05.09 Add a response contract containing durable asset URLs, provider/model/version, seed, timing, cost, measurements, warnings, candidate lineage, provenance, and release state.
  • 182.C.05.10 Extend the existing GenerationJobStatus in jobs-route.ts (queued, running, complete, failed, blocked, needs_review) with validating, preparing audio, planning, submitted, rendering, downloading, evaluating, retrying, correcting, encoding, awaiting review, releasable, released, cancelled, and expired; map blocked and needs_review onto the release decision instead of adding a parallel enum.
  • 182.C.05.11 Version schemas with forward/backward compatibility tests and reject unknown security- or consent-critical fields.
  • 182.C.05.12 Generate OpenAPI/JSON Schema/SDK types from the canonical source and prohibit hand-copied client types.

182.C.06 Reference Asset Ingestion and Validation#

  • 182.C.06.01 Build resumable, signed image/audio/video upload sessions with tenant-scoped object keys, quotas, MIME sniffing, digesting, and idempotency. Bound to real storage 2026-09-12, and doing it fixed something worse than it added: the running SHA-256 lived in a module-level Map and completeUploadSession fell back to createHash('sha256').digest('hex') — THE DIGEST OF AN EMPTY FILE — whenever that 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. The digest is now computed at completion by READING THE STORED OBJECT BACK. Three operations were added too, because a session was not an upload: the surface answered a nextOffset and there was nowhere to send the chunk.
  • 182.C.06.02 Validate dimensions, duration, codecs, sample rate, channels, bit depth, frame rate, orientation, color profile, alpha, and corruption before admission.
  • 182.C.06.03 Strip unneeded EXIF/GPS/device metadata while retaining color and orientation data required for correct processing.
  • 182.C.06.04 Scan uploads for malware, decompression bombs, malformed containers, adversarial payloads, and policy-prohibited content.
  • 182.C.06.05 Detect faces/people and reject ambiguous identity packs until the creator maps each face to a cast member.
  • 182.C.06.06 Compute face visibility, pose/yaw/pitch/roll, blur, exposure, occlusion, crop, resolution, full-body visibility, and duplicate scores; provide actionable reference-quality feedback. All nine delivered on CPU 2026-09-10: ffmpeg blur/exposure, YuNet box and SQPnP pose, dHash duplicates, segmenter-based occlusion, and pose-landmark full-body visibility (the pose model is also the person detector).
  • 182.C.06.07 Normalize provider-bound assets to accepted formats without silently changing aspect, speech timing, channel alignment, or identity.
  • 182.C.06.08 Generate low-risk thumbnails/proxies separately from biometric originals and enforce the same authorization on both.
  • 182.C.06.09 Support authorized HTTPS pull with SSRF protection, allowlisted redirects, size/time limits, content validation, and no access to private network ranges.
  • 182.C.06.10 Record an immutable asset lineage from original through crop, normalization, provider upload, candidate, correction, and release.

182.C.07 Identity Pack and Character Continuity#

  • 182.C.07.01 Extend the existing identity-profile contract into a versioned character pack with subject ID, authorized references, appearance traits, allowed transformations, wardrobe, voice binding, and consent version.
  • 182.C.07.02 Build guidance for a useful reference set: clean face, three-quarter/profile angles, full body, expression range, lighting range, and intended wardrobe.
  • 182.C.07.03 Detect mixed identities, face swaps, age mismatches, and severe reference contradictions before spending provider credits. Mixed identity (SFace pairwise, uncalibrated until 182.C.19.08) and reference contradictions delivered 2026-09-10. Face swap delivered 2026-09-11 by identity-attribute inconsistency (a real swap fixture: the face embeds as A at 0.737 while the hair around it is B's), not by a manipulation detector, with its blind spots named. Age mismatch resolved as human review BY DECISION per 182.C.19.02: estimating age from face geometry skews by ethnicity, so a number would be worse than none.
  • 182.C.07.04 Support immutable identity traits and mutable shot traits so camera, makeup, hairstyle, wardrobe, and scene changes are intentional.
  • 182.C.07.05 Bind each person to stable cast tokens and provider-specific reference ordering without leaking provider semantics into the UI.
  • 182.C.07.06 Integrate existing scene, shot, wardrobe, identity-drift, temporal-drift, and golden-output modules behind the canonical workflow.
  • 182.C.07.07 Define cross-shot continuity handoff using reference packs, seeds where meaningful, selected terminal frames, source clips, and shot state—not seed reuse alone.
  • 182.C.07.08 Support creator-approved reference updates with versioned invalidation, re-consent, benchmark impact, and reproducible old projects.
  • 182.C.07.09 Define identity-score thresholds by framing and visibility; do not penalize shots where a face is legitimately not measurable. Thresholds are per embedding model: under SFace two different people scored cosine 0.422, above the inherited minIdentityCosine of 0.35, so the cutoff is re-fit on a measured ROC for the model that ships (182.C.19.08).

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

  • 182.C.08.01 Add a script editor contract that preserves exact text, punctuation intent, stage directions, nonverbal sounds, speaker turns, and locked versus editable spans.
  • 182.C.08.02 Add a pronunciation lexicon supporting phonetic spelling, IPA/provider phonemes where available, names, acronyms, numbers, multilingual words, and project-level reuse.
  • 182.C.08.03 Support creator-uploaded speech, consented voice synthesis (reusing the narration executor/provider-env TTS path and its synthesis-disclosure gate), and provider-native voice IDs while keeping a single approved-audio asset as the exact timing source.
  • 182.C.08.04 Integrate the existing voice-consent and persona lifecycle stores; reject expired, revoked, scope-mismatched, or unverified voice use.
  • 182.C.08.05 Add voice preview and approval before expensive video generation, including word transcript, duration, pronunciation, emotion, pace, and locale.
  • 182.C.08.06 Normalize loudness, peak, silence, sample rate, channels, and leading/trailing padding without shifting word timing.
  • 182.C.08.07 Produce word- and phoneme-level alignment with confidence, uncertainty, pause boundaries, viseme hints, and speaker IDs.
  • 182.C.08.08 Split long speech only at safe phrase/breath boundaries and preserve sample-accurate offsets for later stitching.
  • 182.C.08.09 Validate script-to-audio exactness before rendering and require explicit creator approval for any substitution or disfluency.
  • 182.C.08.10 Support Ghanaian English, Akan/Twi, Ga, Ewe, and code-switching fixtures with names and local pronunciations; report when a TTS provider lacks adequate locale quality.
  • 182.C.08.11 Keep prompt-native dialogue as a separately labeled mode with mandatory post-render ASR and no exact-word promise.

182.C.09 Shot Planning and Performance Direction#

  • 182.C.09.01 Build a shot-plan schema containing cast, speaker, action, blocking, beat timing, camera, lens, framing, gaze, expression, gesture, props, environment, continuity, and dialogue segments.
  • 182.C.09.02 Translate a creator brief into editable shots without changing locked dialogue or consented identities.
  • 182.C.09.03 Estimate speech duration before planning and prevent shot durations that cannot contain the approved audio plus transition handles.
  • 182.C.09.04 Mark visible-speech windows, off-screen speech, reaction shots, cutaways, overlaps, and deliberate occlusions for evaluation.
  • 182.C.09.05 Add blocking primitives for walking, sitting, standing, turning, object interaction, position marks, path, speed, and endpoint.
  • 182.C.09.06 Add performance primitives for emotion, intensity, posture, gesture frequency, eye line, attention target, conversational turn, breath, and pause.
  • 182.C.09.07 Add camera primitives for static, pan, tilt, dolly, truck, crane, orbit, handheld, tracking, rack focus, focal length, depth of field, and motion intensity.
  • 182.C.09.08 Add feasibility validation for contradictory actions, impossible timing, too-small speaking faces, unsafe activity, and provider limits.
  • 182.C.09.09 Compile provider-neutral direction into model-specific prompts/options and store the compilation for reproducibility.
  • 182.C.09.10 Provide manual shot locking so retries can change only the failed dimension—sync, identity, motion, framing, or artifact—without needlessly regenerating approved choices.

182.C.10 Live Provider Capability Registry#

  • 182.C.10.01 Define a capability vocabulary covering text/image/video/ audio inputs, first/last frames, reference counts/types, native audio, supplied audio, exact-script lane, multi-speaker, edit, extend, durations, aspect ratios, resolutions, seeds, callbacks, and provider options, and mark which entries the live catalog supplies (durations, resolutions, aspects, sizes, frame-image slots, generate_audio, seed, passthrough names) versus which must be derived from passthrough names, probes, or manual records (supplied audio, reference limits, edit, extend).
  • 182.C.10.02 Implement an OpenRouter model-catalog client with strict schema validation, ETag/cache handling, timeout, retry, and fail-closed behavior for required fields.
  • 182.C.10.03 Persist dated catalog snapshots and diffs so removed or changed capabilities automatically quarantine affected routes.
  • 182.C.10.04 Combine advertised capability with live-probe status, benchmark promotion, regional availability, policy status, and health; advertised support alone must not make a model eligible.
  • 182.C.10.05 Record the dated canonical_slug as the model version, a normalized pricing basis derived from the heterogeneous pricing_skus, max duration, output formats, and, as dated manual records because the catalog does not carry them, reference limits, data handling, retention, and geographic restrictions.
  • 182.C.10.06 Add an operator override with scope, owner, reason, expiry, audit log, and automatic rollback; never hide the override from release evidence.
  • 182.C.10.07 Expose a sanitized capability view to Studio for live UI enablement, estimates, and warnings without exposing credentials or internal health details.
  • 182.C.10.08 Test catalog additions, removals, partial fields, unknown options, stale caches, incompatible changes, and total outage.

182.C.11 Complete the OpenRouter Video Transport#

  • 182.C.11.01 Extend the create payload with image-only input_references (the documented shape is { type: 'image_url', image_url }) and route audio and video references through the provider passthrough object only where the selected model's allowed_passthrough_parameters name them (audio on Wan 2.6/2.7; video, videos, images, last_image on Wan 2.7); preserve cast/role ordering and fail closed when a required reference class has no passthrough on the selected model.
  • 182.C.11.02 Support first- and last-frame frame_images independently of general identity/scene references.
  • 182.C.11.03 Support generate_audio (it defaults to true on audio-capable models, so exact-audio requests must set it deliberately), seed, HTTPS callback_url, output preferences, and a provider object validated against the model's allowed_passthrough_parameters.
  • 182.C.11.04 Validate each payload against the selected live capability record and reject unsupported fields before provider submission.
  • 182.C.11.05 Preserve OpenRouter request ID, generation ID, polling URL, model, resolved upstream provider where available, usage, cost, and raw redacted status evidence.
  • 182.C.11.06 Implement callback/webhook receipt that verifies the HMAC-SHA256 X-OpenRouter-Signature, handles the four event types (video.generation.completed, .failed, .cancelled, .expired), tolerates null generation_id/model on early failures, and adds replay defense, idempotency, and poll fallback.
  • 182.C.11.07 Honor retry/backoff headers, bound polling, classify retryable versus permanent errors, treat expired (job exceeded its maximum time to live) as terminal, and support cancellation where the API exposes it.
  • 182.C.11.08 Replace base64 MP4 data URIs with streamed authenticated download to quarantine object storage, incremental digesting, size limits, and atomic promotion.
  • 182.C.11.09 Extend the existing ffprobe/ffmpeg measurement in video-probe.ts to reject missing, truncated, HTML/error, wrong MIME, silent, zero-frame, or unsupported outputs.
  • 182.C.11.10 Store only redacted payloads; never log API keys, signed reference URLs, biometric bytes, full private scripts, or provider download tokens.
  • 182.C.11.11 Add contract tests from official examples plus real, redacted canary fixtures for create, poll, callback, download, usage, failure, and timeout.
  • 182.C.11.12 Add integration tests proving temporary provider URLs can expire after durable ingestion without breaking the released asset.

182.C.12 Model-Specific OpenRouter Adapters#

  • 182.C.12.01 Implement a Wan-family adapter with alibaba/wan-2.7 as the primary route (first/last frame, 2–10 s; passthrough audio, video, videos, images, last_image, ratio, negative_prompt, prompt_extend) and alibaba/wan-2.6 as the fallback (first frame only, 5–10 s; passthrough audio, shot_type, size, negative_prompt, enable_prompt_expansion), mapping each control only where the live catalog names it.
  • 182.C.12.02 Implement a Seedance 1.5 Pro-family adapter for prompt-native audiovisual generation only (first/last frames, 4–12 s, 480p–1080p; passthrough watermark, req_key), multi-character dialogue, camera direction, and multilingual prompts; it has no audio input on OpenRouter and must refuse exact-dialogue mode.
  • 182.C.12.03 Implement a Seedance 2.5-family adapter for what OpenRouter exposes (image-only input_references, first/last frame, 4–30 s, 480p/720p, generate_audio, seed; passthrough watermark, req_key, output_format); audio/video references, edit, and extend are not available on this route and must go through the direct @isis/seedance-provider clients under 182.C.13.01, with the OpenRouter adapter failing closed when a request needs them.
  • 182.C.12.04 Implement a Veo 3.1-family adapter (google/veo-3.1, -fast, -lite) for prompt-native audio, first/last frames, 4–8 s, up to 4K, and passthrough personGeneration, aspectRatio, negativePrompt, conditioningScale, enhancePrompt; it has no audio input and must refuse exact-dialogue mode.
  • 182.C.12.05 Implement HeyGen Avatar IV as an explicitly avatar-scoped adapter supporting script plus passthrough voice_id, voice_settings, motion_prompt, expressiveness, fit, remove_background, background, caption, title, and 720p/1080p; supplied audio is not exposed on OpenRouter, so exact-audio requests fail closed on this route, and every voice_id must bind to an active voice-consent record.
  • 182.C.12.06 Ensure every adapter returns an honest consumedControls and ignoredOptionalControls report; required controls may never be ignored.
  • 182.C.12.07 Add golden payload tests and rejection tests per adapter, including unknown live-schema changes and removed model versions.
  • 182.C.12.08 Keep adapter selection data-driven so successor versions can be canaried without changing the canonical creator contract.

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

  • 182.C.13.01 Bring the direct fal/RunPod/Comfy video adapters and the existing @isis/seedance-provider fal, PiAPI, and Volcengine Ark clients under the same contract and capability registry as the direct lane for audio/video references, edit, and extend; remove special-case executor behavior.
  • 182.C.13.02 (V1.2-gated per release-scope.ts) Wire the Veritas unified provider deliberately or retire it; do not leave unreachable HeyGen/Tavus/D-ID adapters presented as runtime capabilities. Verified 2026-09-12: the PROHIBITION half holds. heygen/avatar-iv is in the live catalogue capture, so HEYGEN_AVATAR_FAMILY is a reachable adapter rather than an unreachable one; Tavus and D-ID appear nowhere in the human-video code; none of the three is named in web, admin or mobile. The VERB is gated: veritas is in V1_DEFERRED_DOMAIN_IDS in libs/oshun/navigation/src/release-scope.ts, so wiring or retiring it is a V1.2 decision and the box stays unchecked until then. blocked:release
  • 182.C.13.03 Separate active-scene, avatar, video-edit, dubbing, and corrective-sync capability classes so routing cannot cross them silently.
  • 182.C.13.04 Define provider credential/region health checks and deployment readiness independent of TypeScript module import success.
  • 182.C.13.05 Add a self-hosted-provider port for future audio-conditioned video models without claiming parity before inference and benchmarks exist.
  • 182.C.13.06 Record each provider's licensing, acceptable-use, commercial-use, biometric, public-figure, minors, retention, and deletion constraints.
  • 182.C.13.07 Prohibit a provider fallback that weakens consent, retention, required inputs, exact-dialogue, resolution, or quality policy.
  • 182.C.13.08 Add a provider kill switch that stops new submission while allowing safe ingestion/evaluation of already completed jobs.

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

  • 182.C.14.01 Filter first by hard capabilities, consent/policy, geography, retention, tenant allowance, and live health.
  • 182.C.14.02 Rank eligible routes by benchmarked lip sync, exact speech, identity, motion, visual quality, locale, control adherence, success rate, latency, and expected total cost including retries, extending the Beta–Thompson release-rate router in provider-quality-router.ts to per-model, per-cohort keys rather than replacing it.
  • 182.C.14.03 Maintain separate rankings for active single-speaker, active multi-speaker, profile/occlusion, high-motion, talking-head, extension/edit, locale, resolution, and duration cohorts.
  • 182.C.14.04 Treat confidence intervals and sample size as routing data; prevent a tiny benchmark from displacing a proven model.
  • 182.C.14.05 Support draft/standard/production/hero policies with different candidate counts, evaluators, providers, budgets, and review.
  • 182.C.14.06 Estimate end-to-end cost before admission, including TTS, multiple candidates, provider retries, downloads, evaluation GPU time, correction, encoding, storage, and egress, using a normalizer over the heterogeneous catalog pricing_skus (per second by resolution, with or without audio, per mode, or per video_tokens) published in docs/agents/model-cost-openrouter.md.
  • 182.C.14.07 Re-route on capability regression, quality drift, outage, throttling, or budget pressure while preserving immutable request intent.
  • 182.C.14.08 Return a machine-readable explanation of selected route, rejected routes, tradeoffs, and any optional controls not available.
  • 182.C.14.09 Add shadow evaluation and canary weights for new models; never promote straight from catalog discovery to customer default.
  • 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 mode.

182.C.15 Durable Job Orchestration#

  • 182.C.15.01 Extend the generation job envelope with request version, project, shot, candidate, attempt, provider correlation, asset digests, policy snapshot, and idempotency key.
  • 182.C.15.02 Implement a durable workflow/saga for validation, consent, audio, planning, submission, polling/callback, download, evaluation, selective retry, encoding, review, and release.
  • 182.C.15.03 Guarantee idempotent provider submission and callback handling across process restart, queue redelivery, timeout, and race.
  • 182.C.15.04 Persist state transitions transactionally with an outbox so Studio and admin cannot observe impossible job sequences.
  • 182.C.15.05 Add cancellation semantics for queued, submitted, downloading, evaluating, retrying, and review states, including what upstream spend cannot be recovered.
  • 182.C.15.06 Add per-tenant/provider/model concurrency, rate, quota, spend, and GPU controls with fair scheduling and priority classes, extending generation-budget-gate.ts and generation-quota-gate.ts.
  • 182.C.15.07 Classify validation, consent, capability, provider, transport, corrupt-output, quality, safety, budget, cancellation, and internal failures with stable error codes.
  • 182.C.15.08 Add bounded retries with jitter, attempt lineage, circuit breakers, poison-job quarantine, and no infinite provider polling.
  • 182.C.15.09 Allow failed shots to resume without regenerating approved audio, references, or already released shots.
  • 182.C.15.10 Add progress events that reflect real stages and avoid fake percentages where provider progress is unavailable.

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

  • 182.C.16.01 Define quarantine, intermediate, candidate, approved, released, and deleted object namespaces with lifecycle and access rules. Six mutually non-prefixing key prefixes, so one prefix-scoped bucket grant covers exactly one namespace; transitions refuse stage-skips and digest changes, and nothing derives from quarantine.
  • 182.C.16.02 Stream uploads/downloads through bounded memory; prohibit base64 video payloads in jobs, databases, logs, and client responses. Detection sniffs container signatures from a bounded prefix rather than flagging long strings, so a JWT or a signature is not a false positive; the log redaction profile bounded no string length at all and now does. Bounded memory is proved by streaming every chunk from one reused buffer: a retained chunk changes the digest.
  • 182.C.16.03 Probe and normalize container, codec, dimensions, FPS, duration, time base, rotation, pixel format, color space, transfer, audio codec, channels, sample rate, and loudness. Untagged colour is reported as untagged rather than defaulted to BT.709, ffprobe's rotation sign is flipped once so display geometry is derived rather than inferred from storage, and the conform keeps tagging separate from converting. Proved by a round trip through real ffmpeg and a re-probe of the output.
  • 182.C.16.04 Preserve the untouched provider output for audit while deriving evaluation proxies and delivery renditions from immutable hashes. The master is registered once and a retry reusing its identity with different bytes is refused; every derivative is keyed by the master's digest and by a recipe digest that includes the tool version; and "untouched" is verified against a freshly measured digest, not the record's own copy of it.
  • 182.C.16.05 Generate thumbnails, contact sheets, scrub proxies, waveforms, captions, transcript overlays for review, and frame-accurate timecode metadata. Drop-frame and non-drop counts both implemented against the standard checkpoints and round-tripped; captions come from the approved alignment rather than from ASR of the delivered video; the transcript overlay is a sidecar bound to frame ranges, never burned into the pixels an evaluator measures.
  • 182.C.16.06 Implement deterministic muxing so approved speech audio remains sample-aligned after normalization, stitching, and transcoding. Verified with a click at a known sample rather than by duration comparison, which found a real defect: -avoid_negative_ts make_zero destroys the mp4 edit list carrying the AAC priming delay and moved the speech 1024 samples (21.3 ms) later. It is now a forbidden flag/value pair, and -shortest is banned because truncating drops approved speech.
  • 182.C.16.07 Add HLS/DASH or signed progressive delivery as appropriate, range requests, CDN cache controls, expiry, and tenant authorization. The signature covers key, tenant, audience and expiry together and is checked before the clock; TTLs come from the namespace rules rather than the caller; cache max-age is clamped to what is left of the signature so bytes never outlive their authorization at the edge; and HLS segments carry the playlist's expiry so playback cannot die halfway.
  • 182.C.16.08 Verify decoded duration and A/V timestamps after every transcode; reject drift introduced by variable frame rate or time-base changes. Measured: the same 29.97 source at a 1 kHz timescale becomes variable-rate (33/34 tick alternation) where 90 kHz gives exactly 3003 ticks every frame, with no ffmpeg error either way. Exactness is now computed and asserted at conform-plan time, and the before/after check compares decoded counts and the A/V start offset rather than durations.
  • 182.C.16.09 Make deletion propagate to derivatives, provider copies where supported, CDN, caches, search indexes, and evaluation assets while retaining only allowed audit tombstones, extending generated-artifact-erasure.ts. The edge is purged before and after origin deletion, the tombstone is written last so a crash looks un-erased, a lane with no deletion API yields a retention disclosure rather than a claimed deletion, and the tombstone is refused if it carries inline media or a numeric run the shape of a face embedding. BOUND AND RUN LIVE 2026-09-12: the plan had no adapter until now, so nothing behind it could remove a byte. human-video-object-store.ts can, and proves it — a HEAD cannot confirm a deletion (HeadObject on an absent key and on an absent BUCKET come back byte-identical from MinIO, Error/NotFound/404/no Code, because a HEAD has no body to parse a code out of), so the check is a LIST of the exact key; and a plain DeleteObject on a versioned bucket leaves every byte version behind a delete marker, measured at deleted 3 version(s). human-video-erasure-runner.ts binds it to the database. Evidence: evidence/2026-09-12-erasure-bound-live/.
  • 182.C.16.10 Load-test large references, long clips, concurrent downloads, and evaluator fan-out without exhausting BFF memory or local disk. The constraint is temp disk and child-process memory, not the BFF heap: 384 MiB streamed across six concurrent downloads holds nothing (proved by digest, not by RSS sampling), four concurrent conforms of a 30 s take run three at a time under a disk budget, and the fan-out is bounded. Admission refuses rather than queueing without bound.

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

  • 182.C.17.01 Select and license at least two complementary sync evaluators—one SyncNet/LSE-class baseline and one current robust audiovisual alignment model—plus a human calibration protocol. The measured baseline candidate is SyncNet v2 (joonson/syncnet_python, MIT; syncnet_v2.model sha256 961e8696…, 13.6 M parameters), which runs on this CPU at about 1 s per second of video (evidence 2026-09-10-cpu-sync-and-alignment). Selected with a geometric lip-envelope correlator as the complement — chosen to fail differently, since two measures sharing a blind spot are one measure. The second LEARNED model stays open with its blockers recorded: AV-HuBERT is CC-BY-NC and disqualified for a paid product, a Wav2Lip-class discriminator shares SyncNet's lineage. Calibration protocol in docs/domains/isis/human-video/sync-evaluator-selection.md, executable via assertCalibrationDesign.
  • 182.C.17.02 Implement reproducible model acquisition with pinned weights/digests, license record, device capability checks, and fail-loud startup when the evaluator is unavailable. The digest is the authority and the URL only a hint: bytes are hashed while streaming and renamed into place only on a match, which matters most for SyncNetModel.py because that artifact is code the evaluator imports. A missing or corrupt evaluator throws rather than letting an unmeasured take be reported as a measured one.
  • 182.C.17.03 Decode real frames and audio, detect/track faces, locate mouth regions, resample consistently, and preserve timestamps through the evaluator input pipeline. Runs against a real face fixture: 100 frames conformed from 30 to 25 fps with the resampling reported, a face tracked in all 100, mouths located from YuNet's own landmarks, MFCCs from the library SyncNet's reference implementation uses, and a source-time second on every crop. Gaps are recorded rather than skipped. Crop scale is a parameter awaiting calibration against SyncNet's training framing.
  • 182.C.17.04 Calculate per-frame/window sync confidence and offset, not only one clip-level LSE-C/LSE-D value. SyncNet v2 runs strictly from the pinned checkpoint over real crops. The case was measured: on a frozen face the clip-level figure read "0 ms, confidence 0.28" while per-window confidence was 0.032 and offsets scattered ±9. The offset arithmetic is proved model-independently by recovering known shifts exactly. Two measured caveats recorded for 182.C.17.08: confidence has a floor set by the search width, and it conflates "nothing aligned" with "nothing moving".
  • 182.C.17.05 Bind speech-active windows to tracked cast members and score only the expected speaker while detecting false mouth motion on listeners. Building the fixture found two defects: normalising motion by a track's own median makes a constantly-moving listener invisible (the baseline is now its motion while nobody speaks), and a pixel-difference measure has a compression floor — a pixel-identical face measured up to 8.4 units of change from h264 rate control, which the five-frame run-length guard exists for. Speech nobody is planned to be speaking in is refused rather than defaulted.
  • 182.C.17.06 Treat missing/tiny/occluded/profile faces as coverage data, not an automatic pass; apply shot-plan-aware measurability rules. Three verdicts rather than two — measurable, expected coverage, and unexpected coverage — because the same small face is a correct wide shot and a broken close-up. The measurability floor does not move with the plan; only whether falling below it was expected. Two assertions enforce the rule, including one that catches a caller scoring more windows than it could see.
  • 182.C.17.07 Detect global offset, local drift, intermittent desync, frozen mouth, excessive mouth motion during silence, and edit-boundary discontinuities. Six detectors, each tested against the other five's series, because the value of separate detectors is in their not firing on each other — a uniform offset and a drift want different repairs. A global offset needs a small spread, a drift needs the line to explain the variance, intermittent measures against the clip's own median, and the edit boundary is a real change-point search rather than a threshold on adjacent differences.
  • 182.C.17.08 Calibrate thresholds against blinded human ratings across languages, skin tones, facial hair, head poses, speaking styles, motion, compression, and shot sizes. Blocked on the 182.C.33 rated corpus and on human raters — fitting thresholds to invented ratings would be the fabrication this phase forbids. Every threshold the evaluators use is already a named, exported, provisional constant with its reasoning beside it, so calibration is a re-fit rather than a search. Audited 2026-09-12 and that claim was ALMOST true: findMotionFaults took every threshold as a parameter except the "is this sequence moving at all?" gate, which was two inline literals. It decides whether the motion checks emit anything — a take below it produces no findings, which reads as nothing wrong — so it is now movingFlowStdDev / movingMeanFlow on MotionThresholds, with the measured gap (static shot 0, smallest moving sequence 0.942) as its provenance and three tests pinning it. The re-fit precondition now holds. blocked:corpus blocked:human
  • 182.C.17.09 Return structured evidence with distributions, worst windows, annotated review frames, model/version, confidence, and failure reason. The distribution leads rather than the average, because a reviewer shown "0 ms, confidence 0.28" would have approved the take that measured it. Coverage is its own number, the search width is part of the evaluator's identity, and the failure reason is a sentence a reviewer can disagree with. A "passed" verdict carrying findings is refused.
  • 182.C.17.10 Require the evaluator to process the exact final muxed bytes before release; pre-transcode success is insufficient. Not theoretical: this phase measured a mux flag moving the audio 21.3 ms, a timescale turning a constant rate variable, and a re-tag changing how every frame is displayed — all invisible to an earlier score. The gate compares digests, refuses a proxy before it even compares, and needs every track of a multi-track release to have measured the delivery.
  • 182.C.17.11 Add adversarial tests where metadata says aligned but audio is shifted, stretched, replaced, silent, routed to the wrong speaker, or desynchronized only mid-clip. Six lying clips through the real pipeline, with the honest baseline as a control — a suite where everything fails proves only that the detector fires. Measured: honest speech/motion overlap 26 of 27 frames, shifted 0 of 27. Which pathway catches them is asserted rather than claimed: this fixture flickers, it does not talk, so the embedding has nothing to score and the speech-against-motion pathway is what fires.
  • 182.C.17.12 Benchmark evaluator throughput and schedule GPU/CPU work without starving render workloads; the starting point is the CPU figure measured 2026-09-10 (SyncNet v2 lip stream about 1 s CPU per video-second on 4 cores). Measured end to end: 1.70 CPU-seconds per video-second (0.59× realtime), linear across 10 s and 20 s clips. The concurrency curve turns over inside the measured range — four evaluators are worse than three, and three buy 24% for three times the box — so the policy is one at a time with three cores reserved, threads capped alongside processes.
  • 182.C.17.13 Decide where evaluators run for production throughput, from the 182.C.17.12 numbers: CPU inside the BFF worker, or a hosted GPU worker (RunPod through libs/isis/runpod-surface// runpod-comfy-client.ts); either way keep a CPU path that produces real measurements for CI. This is a hosting decision, not a prerequisite: 182.C.17.01–.12 are built and probed on CPU first, and nothing in this section is blocked-local without a named failed attempt. DECIDED: CPU inside the BFF worker, one evaluator at a time, no GPU worker provisioned now — the throughput is not the long pole, a GPU would buy latency the pipeline cannot spend, and the scaling lever is horizontal. Three named reversal thresholds and the RunPod route recorded in docs/domains/isis/human-video/evaluator-hosting-decision.md.

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

  • 182.C.18.01 Run independent ASR on delivered audio (extending the fail-loud transcriber seam in dub-faithfulness.ts, with the vendor decided at 182.C.02.12: OpenRouter audio-input models, mistralai/voxtral-small-24b-2507 primary and google/gemini-3.1-flash-lite default) and compare it with the locked script using word error rate, semantic substitutions, omissions, insertions, and speaker turns. Run the currency/number expansion in dialogue.ts before scoring; two recognizers spell spoken "forty two dollars" as "$42". The spec replays the real recorded transcripts from all three probed models and asserts each now scores zero content-word errors — the defect the probe surfaced. The transcriber refuses to be the generating model or a model from the same vendor.

  • 182.C.18.02 Align delivered words to approved-audio words and flag timing stretch, truncation, re-synthesis, unexpected dialogue, and altered nonverbal sounds. Word timing comes from a CTC forced aligner on CPU (wav2vec2 CTC, measured 2026-09-10: max 81 ms raw onset error with a systematic late bias; word ends about 100–140 ms early, so use onsets). Never from OpenRouter audio models' or Whisper's generated word timestamps, both measured unfit. First step: calibrate the aligner's bias on the committed known-onset fixture (evidence/2026-09-10-cpu-sync-and-alignment/known-onset-clip-16k.wav) before it gates anything.

  • 182.C.18.03 Evaluate pronunciation for names, acronyms, numbers, lexicon entries, and code-switched phrases with locale-aware confidence.

  • 182.C.18.04 Detect clipping, silence, low SNR, pumping, phasing, duplicate audio, channel inversion, excessive noise, and sample-rate/time- stretch artifacts.

  • 182.C.18.05 Verify loudness and true peak against delivery profiles without hiding provider speech defects through aggressive processing.

  • 182.C.18.06 Define exact-audio mode, exact-words mode, and creative- dialogue mode separately; expose the achieved mode in release proof.

  • 182.C.18.07 Add multilingual ASR ensembles or human review fallback where a single recognizer is not validated for the requested language.

  • 182.C.18.08 Test fast/slow speech, whisper, shouting, laughter, singing, emotional delivery, background music, accent variation, and overlapping speech. Worked 2026-09-10 and re-verified 2026-09-12: human-video-delivery-conditions.ts expands the nine phrases to TEN conditions and built the FOUR that can honestly be made from a committed speech clip — fast, slow, background music, overlapping speech — because those are TRANSFORMATIONS of a signal. The other six are different phonation or a different speaker (whisper is unvoiced, shouting moves formants as well as level, laughter and singing are not speech, an accent is another person); filtering a normal voice into "a whisper" gives a quiet normal voice, and a detector tested against that has been shown the filter. The four found two real holes, both guarded in that module: time-stretch collapses the GOP to -4.99 at 1.25x on identical words, and a second voice 0.5 s in gives a 513.8 ms onset error while the GOP stays at exactly 0.0 — the aligner is given the words and will always find them somewhere. Measurements in evidence/2026-09-10-delivery-conditions/. Remaining six need real speakers (182.C.33). blocked:corpus

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

  • 182.C.19.01 Measure subject-specific identity similarity over visible frames and report temporal drift, not only best-frame similarity.
  • 182.C.19.02 Evaluate face shape, complexion, age presentation, hair, distinctive features, body proportions, wardrobe, and permitted changes using calibrated metrics plus human review.
  • 182.C.19.03 Track identity through profile turns, partial occlusion, camera motion, lighting changes, cuts, interactions, and multiple people.
  • 182.C.19.04 Detect face merging, identity swaps, duplicate people, flicker, sudden age/skin-tone shifts, and a speaker inheriting another person's mouth or voice.
  • 182.C.19.05 Detect malformed teeth/tongue/lips, jaw discontinuity, asymmetry spikes, face boundary tearing, extra limbs/digits, fused props, and impossible object contact. Probed 2026-09-10: dense face-mesh asymmetry and lip-curvature are disqualified (the mesh scored a doubled mouth as more regular than the clean face), and hand-keypoint models cannot count digits (they emit 21 keypoints unconditionally). Candidates are a trained anomaly/artifact detector or a segmentation-consistency check; mediapipe is Apache-2.0, ultralytics is AGPL. Probed 2026-09-12: THE SEGMENTATION-CONSISTENCY CANDIDATE IS ALSO DISQUALIFIED. tools/human-video/segmentation-consistency.py measured three metrics on the same three fixtures and separated nothing — the doubled mouth and the tear score identically, both 22% from clean. YuNet GREW its face box to swallow the paste (208 px against the clean 195), so a metric relative to the face box cannot see an artifact the detector has already accepted as face. Same failure as the mesh in a different instrument: every learned model with a facial prior rationalises the artifact instead of reporting it, and both instruments that run on this hardware have one. That leaves the trained anomaly detector, which needs 182.C.33's labelled examples. Evidence: evidence/2026-09-12-segmentation-consistency-probe/. blocked:corpus
  • 182.C.19.06 Evaluate motion smoothness, temporal consistency, optical flow outliers, frame duplication, freeze, judder, warping, and background instability.
  • 182.C.19.07 Evaluate camera/control adherence, action completion, blocking, gaze, emotion, gesture, framing, and scene/wardrobe continuity.
  • 182.C.19.08 Calibrate identity and artifact thresholds by visibility, shot size, motion blur, and stylization rather than one universal cutoff, and per embedding model (SFace false accept at 0.422 against the inherited 0.35 gate); blocked on a labelled corpus (182.C.33), not on hardware. Audited 2026-09-12 and FOUND A LIVE ONE. human-video-continuity-bridge.ts graded identity with similarity >= minIdentityCosine — a single universal cutoff, no condition adjustment, no check that the point was fitted for the embedding model in use — so the 0.422 false accept this item is named after would have come back outcome: 'pass'. It now takes a FittedOperatingPoint | null instead of a raw cosine, and answers UNKNOWN (keeping the assessment, so the review record is not emptied) when the point is absent, unfitted, or fitted for a different model; a fitted point is adjusted through adjustForConditions before it grades. Six tests. The re-fit itself still waits on 182.C.33. blocked:corpus
  • 182.C.19.09 Add fairness slices for dark skin, varied lighting, facial hair, head coverings, makeup, glasses, age, and disability-related facial or motor differences. Verified 2026-09-12: all EIGHT are declared in human-video-fairness-slices.ts with assignment rules, buckets and provenance, and the machinery refuses what matters — an aggregate with no slices, and an under-powered slice called a result. The seed measurement (two subjects, nine exposures) found the OPPOSITE of the expected direction: the lighter-complexioned subject lost more complexion measurements, under brightening. Populating the slices needs subjects, which is 182.C.33. blocked:corpus
  • 182.C.19.10 Store only necessary embeddings with encryption, tenant isolation, deletion linkage, access audit, and prohibited reuse for identification outside the creator's project.

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

  • 182.C.20.01 Build active-scene tests with walking, turning, hand/prop interaction, sitting/standing, camera tracking, foreground occlusion, and varied shot sizes while speaking. Verified 2026-09-12: all six conditions are declared in human-video-active-scene-cases.ts, crossed with 182.C.17.06's shot sizes rather than a second list, and the matrix enforces the two things that make it a test — a case whose speech window barely overlaps its condition window is not a test of that condition however it is named, and two conditions that reach the evaluator as the same mechanism in the same band collapse into one, loudly. Empty cells are listed by name rather than divided out of a coverage figure. The clips are 182.C.33. blocked:corpus
  • 182.C.20.02 Measure sync and identity during motion rather than only cherry-picking stationary frontal frames.
  • 182.C.20.03 Detect whether the correct person's mouth moves for every dialogue segment and whether non-speakers remain naturally responsive.
  • 182.C.20.04 Validate conversational turn-taking, interruptions, overlaps, reaction beats, eyelines, and speaker-position continuity.
  • 182.C.20.05 Reject scenes where a speaking face is persistently hidden contrary to the shot plan or where cuts disguise failed sync.
  • 182.C.20.06 Define when off-screen dialogue is legitimate and require the plan/release evidence to distinguish it from unevaluable sync.
  • 182.C.20.07 Add challenge cases for side/back turns, masks, hands near the face, eating/drinking, wind/hair occlusion, fast camera moves, low light, shallow focus, and compression. Verified 2026-09-12: all declared in human-video-challenge-cases.ts, each with the answer that is CORRECT for it — because a challenge suite whose every expected answer is "reject" is passed by a system that rejects everything, and looks more rigorous the longer it gets. For most of these the correct answer is neither pass nor fail but DECLINE TO SCORE, and the case checks the system reaches that rather than a confident one. The footage is 182.C.33. blocked:corpus
  • 182.C.20.08 Require manual review until each active/multi-speaker cohort reaches its sample-size and agreement targets.

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

  • 182.C.21.01 Generate multiple bounded candidates for production/hero requests and retain a shared request/shot/audio lineage.
  • 182.C.21.02 Rank candidates with a Pareto-aware policy so visual beauty cannot compensate for failed words, wrong speaker, consent, or critical identity/sync defects.
  • 182.C.21.03 Define hard gates versus soft preferences and make exact speech, consent, safety, and severe sync/identity failures non-compensable.
  • 182.C.21.04 Classify failures into prompt/control, provider transport, speech, sync, speaker attribution, identity, anatomy, motion, continuity, encoding, and policy categories.
  • 182.C.21.05 Retry selectively with changed seed, clarified prompt, reference ordering, shot subdivision, stronger audio/reference binding, or a different model based on the failure class.
  • 182.C.21.06 Stop retries at spend/latency limits and return the best unreleased diagnostics rather than silently lowering thresholds.
  • 182.C.21.07 Implement optional corrective lip-sync/video-edit providers behind an explicit fallback policy for otherwise excellent candidates; never make correction an invisible default.
  • 182.C.21.08 Re-run identity, anatomy, motion, exact-speech, sync, encoding, safety, and provenance gates after any correction.
  • 182.C.21.09 Preserve before/after candidates and record which frames changed so reviewers can detect identity or texture damage.
  • 182.C.21.10 Let creators compare takes and regenerate one shot/segment without invalidating unrelated approved work.

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

  • 182.C.22.01 Segment long scripts by sentence, performance beat, provider duration limit, continuity, and safe audio boundaries.
  • 182.C.22.02 Maintain cast, wardrobe, props, environment, time of day, camera grammar, voice, and speech timing in a continuity ledger.
  • 182.C.22.03 Support selected terminal-frame, final-frame, reference- video, or extension handoff based on provider capability.
  • 182.C.22.04 Add edit handles and transitions without trimming phonemes, duplicating words, or causing A/V timestamp discontinuities.
  • 182.C.22.05 Reconcile provider clips with CFR/VFR, time base, color, resolution, grain, motion cadence, and audio ambience differences.
  • 182.C.22.06 Apply room tone, music, effects, and mix automation after dialogue approval while preserving the speech reference used for sync evaluation.
  • 182.C.22.07 Evaluate each shot, every edit boundary, and the assembled master; a per-shot pass must not automatically pass the sequence.
  • 182.C.22.08 Support extension/edit requests that lock approved frames and dialogue regions and clearly identify provider-side inpainting or regeneration scope.
  • 182.C.22.09 Provide shot-level resume, versioning, alternate takes, approvals, and rollback.
  • 182.C.22.10 Benchmark continuity decay and determine maximum reliable sequence length per route before enabling automatic long-form generation.

182.C.23 Studio Creator Experience#

  • 182.C.23.01 Design a dedicated “Directed Human Video” flow distinct from simple text-to-video and talking-head avatar cards.
  • 182.C.23.02 Add drag/drop reference sets with role, cast member, priority, quality feedback, consent state, crop, and provider-eligibility indicators. Delivered 2026-09-12. The controls shipped with the region; the last gap was that ADDING was disabled, with the reason attached, because the surface had no endpoint to send a chunk to. It does now (182.C.06.01), so uploadPort.ts opens, chunks and completes, and the reference is added ONLY AFTER THE BYTES HAVE STOPPED — the digest an authorization binds to exists only at completion, and a row written when the file was chosen would name an upload that can still change under it. THE ANSWER IS PER FILE, NOT PER DROP: somebody drops six photographs and the fourth is a screenshot of an error page, and refusing the whole drop throws away five good uploads. A new reference's quality is NOT READ rather than ok — the bytes are stored and nothing has looked at them.
  • 182.C.23.03 Add script/speaker editing, pronunciation tools, voice selection/upload, TTS preview, word timing, and explicit audio approval. ALL SIX SHIP, browser-verified 2026-09-12 (chromium, 17 passed, 1 skipped). The item was open for three different reasons on three passes, and each turned out to be something other than what it was recorded as. Script and speaker. ScriptEditor edits text and speaker, and now WRITES and REMOVES a line. ProjectEdit had no create at all — a creator could change a line's words and not write one — so every region past Cast was reachable only for a project something else had populated. Removal is refused for a line recorded audio speaks: the segment's own record of what it says would name a line nothing can find, and the check that catches a take speaking superseded words would have no script to compare against. Pronunciation. PronunciationTools, unchanged; it states that a respelling changes how a term is SAID and never how it is written, because the script is the promise the delivered-word check compares to. Voice selection. VoiceSourceChooser offers three routes, each with the consent it obliges at the moment of choosing, and a route this deployment cannot serve is shown DISABLED with its reason rather than hidden. The availability now comes from GET /audio/voices rather than a default that called everything unavailable — which was correct for a build with nothing bound and wrong the moment something was, in the direction that hides a working capability. TTS preview. Recorded as a spend-or-install decision: kokoro, misaki and espeak-ng "not installed". Two of the three were installed and the third is not needed — misaki carries its own G2P for American English — so the blocker was a measurement nobody had re-taken. Kokoro-82M (Apache-2.0) runs on this CPU: 1.725 s of 24 kHz mono PCM for "The kettle is on.", ffprobe agreeing with the script's own sample count, ~9 s wall per call. tools/human-video/speech-synthesis.py (text on STDIN, because a script is somebody's words and argv lands in the process table) and local-kokoro-voice-provider.ts behind the SAME VoiceSynthesisClient seam ElevenLabs and the Comfy endpoint use. A preview STORES NOTHING and is not a take; only the CATALOGUE route is served, because a cloned voice is produced after the consent ledger, the persona lifecycle and the clone verification are asked together and a preview endpoint that synthesised one would be a way around that gate; the duration is measured by walking the WAVE chunk table rather than believing the synthesiser's own count, and a disagreement refuses; speed is honoured and measured to be (3.15 s at 1.0, 3.975 s at 0.8) while stability, style and useSpeakerBoost have no Kokoro equivalent and are NAMED to the surface rather than absorbed. The preview omits the audible disclosure on purpose, so it is SHORTER than what ships, and says so. Upload. uploadPort opened every session as kind: 'image' and no audio control existed anywhere, so the creator-upload route was offered, consented, and had nothing behind it. It now sends a recording, and submitAudioTake registers it: measured 10.5 s on this CPU for a 3.15 s clip, with real CTC onsets. The LANGUAGE is chosen from what the service says it can align rather than defaulted — which matters because the aligner took a locale AND PASSED IT NOWHERE: the script runs an English-only character CTC model, so a take submitted as Akan, Ga or Ewe was aligned by a model that has never seen the language and the timings came back looking exactly like measurements. It refuses now, before spawning anything, and the refusal is hv.capability.locale_unsupported rather than the hv.internal.unhandled 500 it first produced, which blamed the platform and told the client to retry forever. Word timing. TakePreview.words renders each timing measured-or- estimated; the registered take carries the aligner's own per-word source, and an unplaced word is named rather than dropped. Explicit audio approval. Reachable, and it REFUSES on this deployment — both halves matter. It was unreachable for a reason nothing about this surface could fix: ETag is not CORS-safelisted, so the browser discarded the tag every If-Match needs, and the read that exists so an approval survives a reload handed the client a header it could not see. exposedHeaders now names it. The refusal is correct: no recognizer is bound, so nothing knows what the bytes SAY, and for an exact-dialogue take unknown is as firm as a deviation. A SEGMENT CAN NOW EXIST BEFORE ITS TAKE, which the web model had always allowed for (audioId: "null for a segment the surface only knows from the draft") and the wire type and the loader did not. A planned segment is unapproved, so Direction stays blocked; it warns about nothing, so a new script is not a wall of warnings; and its chosen voice route is held locally, because it is a pre-decision and a stored one outlives the reason for it. audio/plan cuts one segment per run of consecutive lines by one speaker and ONLY ADDS — an existing segment is never re-cut, because its pointer may be the only handle on bytes somebody approved. KNOWN, NOT FIXED HERE: draftReducer refuses every seed once edits > 0, and that counter is edits EVER MADE rather than edits not yet saved — so after the first keystroke the service's answer can never reach the screen again for the life of the mount. The registration path works around it by remounting at the one point it has just saved the draft itself. The guard itself is draft-conflict machinery and wants its own change.
  • 182.C.23.04 Add storyboard/timeline controls for shots, actions, blocking, camera, expression, gaze, gesture, props, dialogue, and visible- speech windows. Delivered 2026-09-12. The controls shipped with the region; the gap was that Direction was BLOCKED because no audio take could be read, and a plan placed against a duration nobody has moves under the creator. audioPort.ts reads them now, and building it found a defect in the check that guards the whole idea: spokenLines.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. "Nobody wrote down what this says" is not "this says the right thing", so it is a fourth reason now — unrecorded — and it tells the creator that RE-RECORDING IS NOT THE FIX, because the bytes may be perfectly good and something simply has to read them.
  • 182.C.23.05 Provide simple presets with an expandable advanced panel; do not expose raw provider JSON as the primary control surface. Delivered 2026-09-11: GET /v1/human-video/presets (the 29th operation) serves human-video-presets.ts — five presets, six advanced settings, and the three settings deliberately NOT offered with their measured reasons. 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 a delivered take cannot be checked against it — and a NAMED DROPDOWN NOTHING CAN SCORE IS THE SAME FAILURE IN NICER CLOTHING. Every setting therefore declares the DirectionAspect it moves, taken from the ten scores in human-video-direction-adherence.ts, and the catalogue refuses to boot with a setting whose aspect is one of the three JUDGEMENT_ASPECTS. A PRESET'S AVAILABILITY IS NOT A PROPERTY OF THE PRESET, so no route count travels with it: 0 would be a claim made from a lane registry nobody read, and it disables the control. And the panel moved OUT of the Direction region, which is blocked until an approved audio duration exists — a preset is picked first, and it was invisible on the first screen a creator sees.
  • 182.C.23.06 Show current eligible routes, capability limitations, retention warning, expected candidate count, price range, and latency before submission. REACHES A CREATOR SINCE 2026-09-13, and did not before: the module was consumed by nothing, and so was buildStudioLaneView underneath it — 182.C.10.07's whole capability view had never been built by anything but its own spec. Both are wired through POST /jobs/estimate, which is the call 182.C.23.06's own note says eligibility is answered by and the one the Studio already makes, so the summary needed no second catalogue read (two reads could disagree about which models exist between the first price and the last). THE STUDIO VIEW HAD TO LEARN WHAT A JOB IS: it priced ONE CostRequest, so a multi-shot job could not be described by it at all, and the number it would have shown is not the number a creator pays — a per-generation minimum applies PER SHOT. It takes the shots now and prices them through estimateJob, the same estimator the route prices come from, so the two cannot disagree. The summary is NULL on this deployment with the reason beside it, because no route is eligible anywhere yet, and the reason distinguishes "nothing decided which routes could serve" from "every route was refused". Covered in a real browser on the live stack against the live price list, which also found a browser assertion that had gone stale the day the lane registry shipped: it looked for the word "unproven", which is what provenFor says about a route the registry never mentioned, and a registry that answers reports its own blockers instead. Delivered 2026-09-11 in human-video-pre-submission-summary.ts. Two of the six cannot be stated as the word suggests: a price range over routes some of which cannot be priced is not a bound (182.C.10.05), and latency is not an estimate under 182.C.15.10's 20-run floor — one route's median is not a property of the set. The expected take count is the smallest of the three 182.C.21.01 bounds, with the binding constraint named.
  • 182.C.23.07 Warn about small speaking faces, excessive motion, conflicting references, unsupported language, insufficient duration, and weak identity coverage before generation. All six delivered 2026-09-11 in human-video-pre-generation-warnings.ts, each carrying the measurement that produced it. Two of the six are not warnings: a face under 7% of frame height and a take shorter than one second leave sync and identity NOT MEASURED rather than measured badly, so they refuse the submission instead of advising. Ordered by what the fix costs, not by severity.
  • 182.C.23.08 Show honest stage progress and allow safe cancellation, background completion, notification, and recovery after reload. Delivered 2026-09-11 in human-video-studio-session.ts over 182.C.15.10, 182.C.15.05 and 182.C.12.02. Two findings not in any of them: "complete" and "ready" are different events — a notification at provider-complete claims gates that have not run — and an elapsed time is only honest anchored to the job record, because a client-side counter restarts at zero on reload and reads as progress. A resume carrying a new idempotency key is refused: a second key is a second charge. AUDITED 2026-09-12 AND THE MODULE WAS NEVER RENDERED BY ANYTHING. Each of the four functions this item delivered is about what a SURFACE may show, and no surface called any of them: the Studio had no region that submitted a job, so there was no stage to show progress for, nothing to cancel, and no notification to send. Bound now by the delivery region (182.C.36.06), and binding it found the document withholding the two things this module computes — stageEnteredAt, without which the only clock a client can reach is the one assertServerAnchored refuses, and the cancel affordance, without which a browser must keep a second copy of 182.C.15.05's table. Both now ride the job document.
  • 182.C.23.09 Build a take-comparison review with synchronized playback, A/B view, per-metric explanation, worst-window navigation, transcript, frame annotations, and regenerate/correct actions. Delivered 2026-09-11 in human-video-take-comparison.ts. Two of the seven hide inside the word "synchronized": playback anchors on the shared WORD rather than the timestamp, because two takes of one line differ in their word timings by design, and an A/B needs both 182.C.21.01 comparability (same audio) and 182.C.17.09 evaluator comparability, which fail independently. Worst-window navigation takes the union of both takes.
  • 182.C.23.10 Require creator confirmation before accepting a warning- eligible take and forbid override of consent, exact-speech, critical safety, or severe identity/sync gates. Delivered 2026-09-11 in human-video-acceptance.ts. "Severe" implies a mild tier, and whether one exists was answered per gate: sync has a real one from ATSC IS-191 against ITU-R BT.1359-1, identity has none because its operating point is unfitted (182.C.19.08) and the inherited 0.35 admitted two strangers at 0.422 — so every identity failure is non-overridable, stricter than asked. The confirmation carries a digest of the exact findings screen, so an acceptance cannot survive a re-measurement the creator never saw.
  • 182.C.23.11 Support shot locking, variant naming, project history, reference/script/audio version changes, and invalidation preview. Delivered 2026-09-11 in human-video-project-history.ts. A LOCK PRESERVES BYTES AND NOT VERDICTS, so a preview has three outcomes and not two: regenerated, kept-but-void, untouched. The middle one is what a UI would render as "locked, fine" — the inherited-verdict failure 182.C.21.08 refuses. A reference change leaves the audio alone; a one-word script change does not, and the preview is where that gets said.
  • 182.C.23.12 Add export controls for aspect/rendition, clean/captioned master, audio stems where licensed, captions, transcript, provenance, and platform presets. Delivered 2026-09-11 in human-video-export-controls.ts. Two of the seven are gates: a rendition is not the thing that was gated — a scale-only downscale carries the verdicts while the face stays above 182.C.19.08's measured 64 px floor, and a reframe never does — and a clean master cannot be un-burned out of a captioned one, so with no pre-burn source it is impossible rather than degraded. Stems need the voice authorization to say so.
  • 182.C.23.13 Follow the frontend-skill design convention (confirm it resolves in the working environment or link the guidance it encodes) before implementation and create responsive loading/error/empty/review states that remain useful without color alone or fine pointer input. Delivered 2026-09-11 across frontend-convention.md, the two Lilith primitives, and the /studio/isis/human-video surface. The skill does not resolve here; the guidance it encodes is linked. Two corrections came out of building it: the four states are five — Direction before approval is not loading, not failed, not review and NOT empty, because an empty state hands over the next action and the next action is in another region — and the four empties are seven, the sharpest being a subject nobody has asked yet, which is not a wait. Chromium at 1280/768/400 px: 0 axe violations, no horizontal overflow, every region state carried by a word, three outcome shapes surviving a full colour flatten, and 15 controls found under the 24 px target minimum that 156 jsdom tests could not see.

182.C.24 Mobile and Desktop Workflow Parity#

  • 182.C.24.01 Decide which authoring controls are full-parity on mobile and which use a review/approval-focused experience.
  • 182.C.24.02 Implement resumable background uploads, constrained-device previews, notification deep links, and job recovery on mobile. Delivered 2026-09-11 in apps/oshun/mobile/src/humanVideo/uploadRuntime.ts and src/native/humanVideoCapabilities.ts, with the notification listener and the relaunch recovery started in app/_layout.tsx and the preview rule in the review screen. THE KEY IS WRITTEN BEFORE THE SESSION IS OPENED: the worst case is then a key with no session, which the server can resolve, rather than a session with no key, which nothing can — and a regenerated key opens a second session that holds a quota slot nobody will ever write to. The offset is asked for every time, because the one case where local progress and arrived bytes differ is the case the resume path exists for. sessionType: 'background' is what makes the transfer survive the app; a foreground task dies with the process and every upload then takes the resume path rather than the interrupted ones. Recovery READS and never re-submits. The preview says an approval binds to the measured original rather than to the downsampled copy on screen. KNOWN CALLER GAP: startOrResumeUpload has no screen calling it, because the screen that would 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, and unprotected caches. Delivered 2026-09-11 in apps/oshun/mobile/src/humanVideo/osContainment.ts — the one layer that touches an OS API — with the review screen as its caller and an eslint rule that makes a direct call to any of the five channels an error inside the human-video mobile paths. The policy module decided all of this in June and nothing called it; "prevent" is a construction, not a convention, so the enforcement ships with it. AN ABSENT CAPABILITY REFUSES RATHER THAN NO-OPS: this build has no clipboard, sharing or media-library module, and a function that quietly does nothing tells the caller the copy was made and starts working unreviewed the day somebody installs the module. Cache location is chosen from the material rather than from the caller, because the caller is the one place that does not know that a biometric file in the documents directory survives the withdrawal that was supposed to end it. The log redacts rather than dropping the line — a log that goes quiet when something is identifiable goes quiet exactly when something is going wrong.
  • 182.C.24.04 Add accessible touch targets, screen-reader labels, captioned previews, transcript navigation, reduced motion, and external keyboard support. Delivered 2026-09-11 in apps/oshun/mobile/src/humanVideo/HumanVideoReviewScreen.tsx, reviewSource.ts and the app/human-video-review.tsx route; 90 tests under the app's own jest. THE SCREEN RENDERS FROM THE CONTROL SPEC THE CHECKER READS, so a control that breaks one of the six rules cannot be drawn without the spec saying so — checkControls existed and nothing called it. Three things the wiring found: reduced motion must not stop the take, because 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, since "not measurable" denies a measurement that happened and a pass or fail invents the threshold 182.C.19.08 leaves unfitted; and an unread alignment is not an empty one — both render as a blank caption block, and only one of them is a take whose words nobody could place. Every OS hand-off goes through mayPlace, and a refused surface has no control rather than a disabled one.
  • 182.C.24.05 Decide whether a desktop surface is in scope and which app hosts it (no Oshun studio desktop app exists; the candidates are apps/yemaya/studio-desktop, apps/lilith/desktop, and apps/iris/desktop), and if so ensure it uses the same server contracts, capability registry, consent, quality decisions, and release proof as web.
  • 182.C.24.06 Add platform-appropriate automation for critical upload, author, submit, cancel, review, approve, and export flows. Worked 2026-09-12. "Platform-appropriate" is the load-bearing word: 182.C.24.01 decided the phone is review-and-approval rather than authoring, so upload, author, submit, cancel and export are WEB flows and belong to 182.C.36.06. The phone's two are review and approve. Review was automated on 2026-09-11; APPROVE WAS NOT — the flow asserted the Accept control was present and never pressed it, leaving the one act a phone exists for unexercised. Pressing it also had nothing to assert, because the screen looked identical afterwards: a decision that produces no visible change reads as one that did not register, and the answer to that on a phone is another tap, which here is a second decision about somebody's face. The screen now takes a decision prop, renders hv-decided, and spends both controls; the flow taps through and asserts all three; the verifier reads 13 test ids against the screen. Still unchecked because the web half is 182.C.36.06. blocked:upstream

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

  • 182.C.25.01 Add versioned endpoints for upload sessions, identity packs, voice/audio approval, human-video jobs, candidates, measurements, review, correction, release, cancellation, and deletion. Delivered 2026-09-11: 23 endpoints declared in human-video-api-surface.ts and registered by walking that table in human-video-api-routes.ts, so the set served and the set documented cannot drift. "Versioned" is two versions — a document written under schemaVersion 1 read back through /v2 is still a v1 document. Deletion returns a destruction manifest, not a 204. Declaring the surface also corrected 182.C.25.03: a precondition guards admission and must never block a safety retraction. THE MOUNTING PROGRAM FINISHED 2026-09-12: the boot line now ends "answering 503 until bound: nothing" (evidence/2026-09-12-fourteen-ports-bound/). Binding the last of them found a FOURTEENTH nobody had counted — the route reads an If-Match subject before any deciding handler runs, and the runtime passed decisionSubjects: null, so all six deciding operations including approve-candidate answered 500 inside a refusal that was right and loud and that nobody was reading. A reader for five of them had existed since the review store landed, with release deliberately omitted because its own port answered 503; what was missing was anything passing it. It is reported as its own boot-line entry because it cuts ACROSS the ports: "candidates: bound" was true and was not the whole answer. One deciding operation stays unwired for a reason no adapter changes — accept-with-warnings is declared and has no endpoint, so nothing can call it.
  • 182.C.25.02 Add cursor pagination and authorization for projects, jobs, candidates, assets, and evaluations. Delivered 2026-09-11 in human-video-list-access.ts. Four leaks in the obvious implementation: filtering after paging makes the short page itself a statement about what was hidden, a total count is the same leak as a number, an unsigned cursor is an editable offset, and a signed cursor carrying its own scope is a capability that outlives the grant — so the scope is compared against the caller's present grants, never trusted. Asset scope is separate from project access, because listing assets is listing faces.
  • 182.C.25.03 Define idempotency, optimistic concurrency, request size, rate, quota, retry, and stable error semantics. Delivered 2026-09-11 in human-video-api-semantics.ts. Three of the seven are wrong in the obvious implementation: an idempotency key must be bound to a digest of the request (a retry with a changed body must not replay), the concurrency token must cover the EVIDENCE a decision is made on rather than the row, and Retry-After must follow 182.C.15.07's disposition rather than the HTTP status — a 5xx ambiguous submission must not be retried, a 429 rate limit must be.
  • 182.C.25.04 Generate SDK clients and examples for exact-audio active scenes, multi-reference cast, multi-speaker dialogue, polling/webhooks, and shot regeneration. Delivered 2026-09-11: sdks/human-video-ts/ is EMITTED from 182.C.25.01's declared surface, and its spec regenerates and compares byte-identically, so adding an endpoint without regenerating fails a test. A deciding method takes ifMatch as a required TYPE parameter and a safety retraction takes none — proven by a deliberately broken example that fails to compile with "Property 'ifMatch' is missing … but required". All five named examples ship.
  • 182.C.25.05 Sign outbound webhooks, rotate secrets, prevent replay, retry with bounded delivery, expose delivery logs, and support endpoint disablement. Delivered 2026-09-11 in human-video-outbound-webhooks.ts, reusing studio/webhook-delivery-store.ts for the backoff rather than growing a second copy. Rotation signs with both secrets and must CLOSE; the timestamp and event id are inside the signed material; the delivery log holds a digest, because "log the full event" is where 182.C.25.07 gets undone; and a long-failing endpoint is disabled because a lapsed host can be re-registered by somebody else.
  • 182.C.25.06 Define events for every durable job/review/release/deletion state with tenant, project, request, shot, candidate, attempt, and trace correlation. Delivered 2026-09-11 in human-video-events.ts. 12 of the 15 event types do NOT correspond to a stage transition — a reviewer approving one candidate and rejecting another moves no stage, and a deletion may move none at all. attempt earns its place because without it a retried render is indistinguishable from a duplicate delivery and gets deduped away, and the trace must be captured at transition time, because the outbox dispatches later and a relay would stamp its own.
  • 182.C.25.07 Redact private script/reference/audio/provider data from events and separate creator-safe measurement summaries from internal forensic evidence. Delivered 2026-09-11 in human-video-event-redaction.ts. THREE audiences, not two: the integrator a tenant pointed a webhook at is the one everybody forgets and the one an event stream actually has. An allowlist, because a denylist fails open. And a summary is BUILT rather than trimmed out of the forensic record, because that record is forensic in its shape — a per-window series reconstructs the performance whatever the paths are called.
  • 182.C.25.08 Add consumer-driven contract tests for Studio, mobile, Veritas, admin, analytics, and external SDK clients. Delivered 2026-09-11 in human-video-consumer-contracts.ts, with all six declared in the spec. The point is not testing the provider again — it is learning what may be CHANGED, so a contract claiming a whole document is refused and whoBreaks names the consumers a removal affects. A contract also pins each field's INTERPRETATION, because a field changing meaning breaks consumers while every schema check stays green; and a contract can be unsatisfiable before any test runs, because 182.C.25.07 decides what its audience ever receives.
  • 182.C.26.01 Define explicit likeness and voice authorization scopes by subject, tenant, project, purpose, media type, geography, duration, provider class, and allowed transformations, extending the existing voice-consent store, the @aja/consent-management upload-consent manager, and the libs/oshun/trust-safety policy taxonomy rather than adding a fourth consent model. Delivered 2026-09-11 in human-video-likeness-authorization.ts, which stores no consent — liveness is an INPUT from those three. The nine dimensions behave differently and an unspecified one is a REFUSAL, never a wildcard. Provider class is a consent dimension (by what the vendor does, not its name), and the transformation allowlist is the anti-impersonation control.
  • 182.C.26.02 Require proof/attestation at reference and voice enrollment, bind it to asset digests, and surface the consent version on every job. Delivered 2026-09-11 in human-video-enrolment-proof.ts. The binding is the whole item: an unbound proof is not false, it is simply not ABOUT anything — swap the images and it still verifies. The digest covers the SET with the count inside it, so appending an asset does not fit, and adding a reference is a new enrolment. The method travels with the proof because a checkbox and a countersigned release are not the same evidence, and the consent version stamped on a job is the enrolment's, never today's.
  • 182.C.26.03 Handle third-party/public-figure likeness, parody/satire, political content, sexual content, fraud, harassment, and deceptive impersonation with jurisdiction-aware policy. Delivered 2026-09-11 in human-video-content-policy.ts, which holds the SHAPE of each decision and not the decision — every region posture ships UNSET and refuses rather than defaulting, because permissive is unsafe and prohibited silently blocks lawful work while looking responsible. Three of the seven are not jurisdictional at all. Public-figure protection varies in BOTH directions, so there is no safe default. And the jurisdiction is three places: the tenant, the audience, and the subject — the one most often left out.
  • 182.C.26.04 Add stricter minor safeguards, guardian authorization, age-uncertainty escalation, prohibited scenarios, retention, and access. Delivered 2026-09-11 in human-video-minor-safeguards.ts. There is no age detector and there will not be one — 182.C.19.02 declined a geometric estimator because it skews by ethnicity, so its errors would fall on particular children. The system therefore behaves correctly WITHOUT an answer: uncertain is its own band and resolves strict, because the two errors are not comparable. Guardian authorization is the only authorization, the minor's assent is separately required, and prohibited scenarios have no review path.
  • 182.C.26.05 Make consent revocation stop new work immediately, cancel eligible in-flight jobs, block unreleased candidates, and trigger scoped deletion/takedown workflows. Delivered 2026-09-11 in human-video-revocation.ts. The four reach different distances and only new work stops without caveat: cancelling in flight stops the RESULT, not the vendor's processing, because 182.C.15.05 found no cancel endpoint; and an already-released take needs a takedown, which links handed out and cached copies put outside this system's reach. Nothing may be reported complete while copies remain — that ends the conversation with the one person entitled to know otherwise. AUDITED 2026-09-12 AGAINST A REAL GRANT AND THE ONE ACT THIS ITEM NAMES DID NOT WORK. POST /authorizations/:id/withdrawal answered 404 for every LIVE authorization — "no authorization with that identifier is reachable from this workspace" — because withdrawConsent opened by calling revocationFor, which returns a Revocation, and a Revocation exists only once revokedAt is set. NOTHING IN THIS SERVICE EVER SET IT: the delegate was bound with findUnique and findMany only, and no code path wrote a withdrawal. So the endpoint could only answer for a permission that had already been taken back, which nothing could produce — and somebody trying to stop the use of a person's likeness was told the permission did not exist. Everything below it was built and correct: stopWork, the draft redaction, the destruction plan, the manifest. The write was the missing piece and the reach machinery is why it was hard to see. The spec had presumed the bug away: its revocationInputs double answered a revocation whose revokedAt was already stamped, so 27 tests ran against a state production could never reach. The double now starts LIVE and withdraw is what changes it. withdraw is on the revocations port and the authorization store, the handler records before it reports (a notice about a withdrawal that was not recorded describes something that did not happen), an absent projectIds takes the WHOLE permission because "all" and "none" are the two readings and only one is safe to assume, and a second withdrawal answers the one already recorded rather than restamping the moment somebody took their permission back. Measured live: the withdrawal answers 200 with a manifest naming the reference image it destroyed, and the next submission is refused hv.consent.missing. Five store tests, and the delegate now requires all four methods — create had been in use by grant since enrolment shipped and was not in the required list, which is the same rule delegateWith states about itself. AND A PERMISSION THAT CAN ONLY BE ENDED WITH curl IS NOT ONE A CREATOR CAN END. With the write in place the endpoint worked and nothing called it, which is the same client/surface drift that hid the Delivery region. consentPort.ts reads where each cast member stands and CastConsentPanel renders it in the Cast region — which had shown consent as unknown on EVERY project ever opened, because nothing on this surface read one. The standing comes from the INVITATIONS, not the authorizations: the creator's surface has no listing of those, and an invitation carries authorizationId once answered, which is the one fact a withdrawal needs. An answered invitation that names no authorization is DROPPED rather than shown as granted — it would offer a control that does nothing. "Asked" is rendered as a wait on us and never as a state the subject is in, and a lapsed ask is kept apart from a refusal: our clock running out is not their no. Ending one takes two presses and the second says what it does rather than "are you sure", because it destroys photographs. What comes back is the manifest — what was destroyed, what could not be reached, what is still running on a provider, what needs a takedown — and only the service's own complete: true is read as complete. Covered in a real browser: the panel shows the permission, the consequence appears before the control can be pressed, the account comes back complete, and a submission afterwards is refused hv.consent.missing. 10 port tests, 11 panel tests.
  • 182.C.26.06 Prevent voice/face packs from being shared across tenants, projects, purposes, or collaborators beyond their granted scope. Delivered 2026-09-11 in human-video-pack-isolation.ts. The four boundaries leak in different places: project leaks through COPYING (afterwards every read-time check passes), tenant leaks through the derived EMBEDDING, a collaborator's grant does not end when they do, and purpose has no technical edge at all — so it is re-checked against the authorization on every use. A share is derived from the authorization rather than recorded on the pack, because the pack is not the authority on who may see a person's face.
  • 182.C.26.07 Run image, audio, prompt, script, and produced-video safety checks at the appropriate stages; post-generation safety is mandatory. Delivered 2026-09-11 in human-video-safety-stages.ts. The mandatory clause is there because every pre-generation check is a check on an INPUT and the output is not a function of the inputs alone — so the refusal explicitly rejects "but the inputs passed". Two surfaces cannot be checked where people put the check: an image screened after enrolment has already been stored, embedded and sent onward. And a not-run recorded as a pass is how an unscreened video ships with a green tick beside it.
  • 182.C.26.08 Add step-up authentication and human review for high-risk likeness, political, financial, medical, legal, or mass-distribution use. Delivered 2026-09-11 in human-video-step-up-review.ts. The two controls answer different questions — step-up defends against a compromised session, review against the legitimate account holder — and most abuse of a likeness system is the second, so a strong login never substitutes for a review. Risk is a property of the COMBINATION: a political script is not high risk until a real person is made to carry it. Mass distribution gates RELEASE, because reach is not knowable at submission.
  • 182.C.26.09 Maintain appeal, takedown, abuse-report, incident, evidence-preservation, and subject-notification workflows. Delivered 2026-09-11 in human-video-appeals.ts. Every one is started by somebody who is not the tenant, and a takedown route behind a login is one the depicted person cannot use — they were never the customer. Preservation and deletion pull opposite ways and both are right, so a hold names exact artefacts with an expiry: "preserve everything relating to this incident" turns a report of misuse into a reason to retain the misused likeness. And notifying the subject is a recorded decision either way, because it is usually right and occasionally a disclosure of the reporter.
  • 182.C.26.10 Test consent expiry/revocation races, collaborator removal, job retry after revocation, provider fallback, derived assets, and cached release URLs. Delivered 2026-09-11 in human-video-consent-races.spec.ts — 17 tests across all six, because each is a question about how 26.01, 26.02, 26.05 and 26.06 interact and the failures live in the seams. Two have no clean answer and the tests say so: a provider fallback can land in a class the subject never agreed to (routing does not know, because provider class is a consent dimension), and cached URLs outlive a revocation by design — 182.C.16.07 bounds that exposure and does not end it.

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

  • 182.C.27.01 Classify face images, face embeddings, voice recordings, voice embeddings, scripts, videos, and evaluator artifacts; assign minimum retention and legal basis. Delivered 2026-09-11 in human-video-data-classes.ts. An embedding's risk GROWS while a script's does not, so retention is not one policy and a derivative may never outlive its source. Minimum and maximum retention can be UNSATISFIABLE — written by different people for different reasons — and that is refused rather than silently resolved. The legal basis decides what a withdrawal does, per class. And the evaluator artifact is classified biometric, because it is a face stored under a name that does not sound like one.
  • 182.C.27.02 Document OpenRouter asynchronous-video retention and non-ZDR status in policy, UI, tenant controls, and routing decisions, including that a ZDR-enforced account receives no video routing at all, so a tenant requiring ZDR is ineligible for every OpenRouter lane. Delivered 2026-09-11 in human-video-zdr-posture.ts. The four surfaces fail differently: the UI failure is that the control and its consequence live in different screens, so the cost is computed BEFORE the toggle; the tenant-control failure is that enabling ZDR does not un-send what was already sent; and an empty route list without an explanation reads as a bug, which is how the setting gets turned off to make the product work.
  • 182.C.27.03 Encrypt all biometric/source media in transit and at rest with tenant isolation, scoped signed URLs, short TTLs, and key rotation. Delivered 2026-09-11 in human-video-key-custody.ts, composing 182.C.16.03's namespace TTLs, 182.C.16.07's signed URLs and 182.C.19.10's per-tenant AES-256-GCM rather than duplicating them. The new part is rotation: keeping the old key so old data still opens is an ACCUMULATION, not a rotation, so completion is computed from whether any DEK still wraps under it. A compromise rotation may not re-wrap lazily — the record nobody touches is the one an attacker has time for — and a per-tenant estate cannot be reported as globally rotated.
  • 182.C.27.04 Prohibit provider access to permanent object URLs; issue single-purpose expiring references where remote pull is unavoidable. Delivered 2026-09-11 in human-video-provider-references.ts. The dangerous URL is not the permanent one — nobody ships that deliberately — it is the one with a generous TTL because nobody knew how long the provider would take, which is a permanent URL measured in hours. So the expiry is DERIVED from the job's own deadline, one reference names one asset for one job, and the module states plainly that a TTL bounds what can be FETCHED and is not a retention control, however much it reads like one.
  • 182.C.27.05 Add least-privilege service identities for ingest, render, evaluation, review, delivery, and deletion with access auditing.
  • 182.C.27.06 Threat-model SSRF, webhook spoofing, IDOR, signed-URL leak, queue tampering, prompt/reference injection, malicious containers, biometric exfiltration, and model output poisoning. Delivered 2026-09-11 in human-video-threat-model.ts. Five are prevented and cite the item that does it; the three useful rows are the others — a signed-URL leak is BOUNDED by the TTL rather than prevented by entropy, prompt injection is DETECTED afterwards because no validator separates a photo from a photo plus a payload, and model output poisoning is UNMITIGATED because the output is the product and a subtly wrong video is a video. That last one must be acknowledged at release, not resolved.
  • 182.C.27.07 Add data residency and provider allow/deny controls for regulated or enterprise tenants; fail closed when no compliant route exists.
  • 182.C.27.08 Verify provider deletion/export capabilities and record unavoidable retention gaps in the job's consent and release evidence.
  • 182.C.27.09 Add subject-access/export/deletion workflows that resolve originals, derivatives, embeddings, provider IDs, released assets, and audit exceptions.
  • 182.C.27.10 Run application, media-parser, object-storage, webhook, and authorization security testing before public release.

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

  • 182.C.28.01 Extend video C2PA/provenance to cover source references, approved script/audio digests, provider/model/version, seed, edits, evaluation, human approval, and final encoded bytes.
  • 182.C.28.02 Apply visible and/or imperceptible synthetic-media disclosure according to policy, geography, tenant, distribution channel, and provider watermark terms.
  • 182.C.28.03 Verify provenance survives supported transcodes and expose a verifier result alongside each downloadable rendition.
  • 182.C.28.04 Produce a canonical release proof binding exact final asset digest to consent, safety, speech, sync, identity, visual, encoding, provenance, and review decisions, extending ProofRecordSchema in libs/contracts/src/common/proof-types.ts and the nine named gates in release-proof.test-fixtures.ts rather than defining a second proof record.
  • 182.C.28.05 Fail closed if required measurements, evaluator versions, consent evidence, provider lineage, or final-byte digests are missing.
  • 182.C.28.06 Distinguish native one-pass generation, provider edit, corrective lip sync, audio replacement, manual edit, and re-encode in creator-visible history.
  • 182.C.28.07 Prevent a candidate or pre-transcode URL from being shared as a released artifact before the final gate passes.
  • 182.C.28.08 Add provenance and disclosure regression tests against trim, concat, mux, transcode, caption burn-in, thumbnail, and CDN delivery.

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

  • 182.C.29.01 Add a provider/model dashboard for live capabilities, probes, health, queue, latency, failure classes, cost, quality, drift, retention, and promotion state.
  • 182.C.29.02 Add a privacy-safe review queue with final video, approved audio/script, cast mapping, metric failures, worst windows, lineage, and bounded approve/reject/escalate actions.
  • 182.C.29.03 Require reason codes, reviewer identity, separation of duties where needed, audit log, and expiry for any allowed quality waiver.
  • 182.C.29.04 Prevent reviewers from downloading biometric originals or provider payloads unless their role and case explicitly require it.
  • 182.C.29.05 Add tools to quarantine a model/version, cancel queued work, reroute jobs, invalidate releases, process takedowns, and re-run evaluation.
  • 182.C.29.06 Add cohort drill-down for language, accent, skin tone, lighting, framing, motion, speaker count, provider, model, and app surface.
  • 182.C.29.07 Add audit-safe redaction/export for support and provider escalation without exposing unrelated tenant or subject data.
  • 182.C.29.08 Follow the frontend-skill design convention (same caveat as 182.C.23.13) and add thorough Playwright, accessibility, and authorization coverage before enabling admin mutations. Delivered 2026-09-11: apps/oshun/admin/src/app/isis/human-video/page.tsx with its panel and port, 16 unit tests, and 5 Playwright tests passing against a real admin server — including the axe scan and the refusal an operator without the workspace scope gets. THE MUTATIONS ARE PRESENT AND REFUSED, which is what the item asks: quarantine, cancel and invalidate name what is missing rather than doing nothing quietly. §29 had built the queue and the reviewer-access rule and left neither reachable, so GET /reviews was added — it carries the access decision PER ARTIFACT, taken from the caller's scopes, because a role a client could name is a role a client could claim; the reference comes back refused with its sentence unless the role AND an open case both require it, and the URL never travels beside a refusal. The browser caught what jsdom could not: the panel's first colours measured 2.47:1 on the admin's dark canvas.

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

  • 182.C.30.01 Trace upload, preprocessing, TTS, planning, routing, provider submission, callback/poll, download, evaluation, retry, correction, encoding, review, and release with shared correlation IDs.
  • 182.C.30.02 Measure success and latency per stage plus end-to-end p50, p95, and p99 by provider/model/mode/duration/resolution/locale.
  • 182.C.30.03 Measure first-pass and eventual pass rate for sync, exact speech, identity, visual integrity, safety, and release.
  • 182.C.30.04 Measure retry amplification, corrected-share, manual-review rate, false accept/reject estimates, cancellation, orphan jobs, and stale callbacks.
  • 182.C.30.05 Attribute actual provider, TTS, evaluation compute, correction, transcode, storage, egress, refund, and wasted-candidate cost to tenant/project/job/shot.
  • 182.C.30.06 Alert on quality drift, missing audio, ASR mismatch, identity regression, provider schema change, spend anomaly, queue age, callback failure, evaluator outage, and unreleased public URL exposure.
  • 182.C.30.07 Keep metric labels bounded and logs redacted; use digests and opaque IDs instead of scripts, names, signed URLs, or biometric paths.
  • 182.C.30.08 Define SLOs/error budgets and automated route quarantine or feature rollback when customer-impact thresholds are crossed.

182.C.31 V1 Convergence and Legacy Hardening#

  • 182.C.31.01 Update V1 persona/avatar/voice documentation with the Phase 182 architecture, honest delivery guarantee, active-scene lane, and exact-dialogue behavior.
  • 182.C.31.02 Replace V1 “provider-gated” ambiguity with links to canonical contracts, tracker status, verified route evidence, and explicit gaps.
  • 182.C.31.03 Integrate the phoneme/viseme mapping only as optional timing metadata for compatible engines; do not count it as rendered lip-sync.
  • 182.C.31.04 Turn MuseTalk/portrait plan modules into real adapter- executed pipelines with measured artifacts, or mark/deprecate them as planning-only.
  • 182.C.31.05 Make portrait quality metrics execute pinned evaluators on actual media instead of trusting externally supplied LSE values.
  • 182.C.31.06 Harden the Psyche avatar engine with real entry point, weights, licenses, dependencies, failure checks, inference tests, and benchmark evidence—or retire it from capability claims.
  • 182.C.31.07 (V1.2-gated per release-scope.ts) Consolidate Veritas avatar/video providers behind the canonical provider port and remove blank defaults, dead routes, and stale API assumptions. Verified 2026-09-12, same shape as 182.C.13.02: the VERB is gated (veritas is deferred, and the BFF middleware blocks /veritas/...), and the prohibition is enforced from the human-video side by human-video-avatar-boundary.spec.ts — six tests, all passing, holding the registry to exactly one avatar route that goes through OpenRouter and is reachable. libs/veritas/video-production's Tavus client refuses without a key rather than defaulting to a blank one, and nothing outside that library imports it. The consolidation question that IS open — two D-ID clients, one of them called by nothing — is recorded in that spec's header as a Veritas decision this phase does not answer from the outside. blocked:release
  • 182.C.31.08 Add migration adapters for existing saved cards/jobs and a deprecation window for the single referenceImage contract.
  • 182.C.31.09 Ensure old text-to-video and image-to-video flows continue working while clearly lacking exact-dialogue guarantees unless upgraded.

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

  • 182.C.32.01 Define a reproducible benchmark harness before selecting a self-hosted native audiovisual or corrective lip-sync model.
  • 182.C.32.02 Review model and weight licenses for commercial use, redistribution, biometric restrictions, training-data restrictions, and acceptable-use compatibility.
  • 182.C.32.03 Package candidates with pinned weights, digests, container, GPU requirements, deterministic fixture, health/readiness, metrics, and fail-loud behavior. Verified 2026-09-12: there is no candidate to package. human-video-selfhosted-port.ts reports no-implementation — the port for a self-hosted model that renders to a creator's approved SPEECH has no adapter, and ComfyVideoProvider is not one (grepped 2026-09-10: no audio parameter anywhere). The packaging DISCIPLINE this item asks for already exists and is exercised on the evaluator weights: human-video-model-acquisition.ts pins a digest, treats the URL as a hint and the hash as the authority, discards on mismatch rather than installing, allow-lists the fetch host before any request, checks device capability against the registration, and fails closed on a missing evaluator rather than skipping the score. Applying it to a generator candidate needs a generator candidate, which needs a GPU (uname -a → dev-server, 4 cores, no GPU).
  • 182.C.32.04 Benchmark active full-frame performance, not only cropped frontal talking heads, across profile, occlusion, motion, resolution, and speaker-count cohorts. Verified 2026-09-12: the harness is 182.C.32.01 and is done; the cohorts are 182.C.20.01's matrix and is built. What is missing is both inputs at once — a candidate model to benchmark (see 182.C.32.03, none exists) and the footage to benchmark it on (182.C.33). Two blockers, neither of them this item's own. blocked:upstream blocked:corpus
  • 182.C.32.05 Measure identity/anatomy damage from corrective models and require all final gates after correction.
  • 182.C.32.06 Add secure RunPod/Comfy or sovereign GPU deployment with signed asset access, ephemeral workspace, wipe verification, queue limits, observability, and no bundled private references. Verified 2026-09-12: hardware-blocked, and the route is already chosen and written down in docs/domains/isis/human-video/evaluator-hosting-decision.md with three named reversal thresholds (182.C.32.05, checked). Standing the deployment up and verifying a wipe both need the GPU host itself; nothing about it can be honestly measured from a box with no GPU, and claiming a wipe was verified without running one is the fabrication this phase forbids. Re-read 2026-09-18 under the install-before-you-park rule: "needs the GPU host itself" is not a blocker this project lacks — the Isis track rents RunPod GPUs on demand (infra/runpod/endpoints/desired-state.json, the spend rules of ISIS_CHROMA_RUNPOD_MVP_TODOS_2026-09-11.md). This item is actionable on a bounded RunPod pod: stand the deployment up, run one wipe and verify it from outside the workspace, tear it down, and record the measured cost. Ask before spending, as that tracker does.
  • 182.C.32.07 Compare self-hosted total cost, latency, reliability, engineering load, privacy, and quality to OpenRouter/direct routes before production promotion.
  • 182.C.32.08 Keep experimental routes hidden or allowlisted until live artifacts meet the same release thresholds as commercial providers.

182.C.33 Evaluation Corpus and Ground Truth#

  • 182.C.33.01 Create a consented, licensed internal corpus with reference packs, exact scripts, clean audio, shot plans, identity labels, speaker tracks, mouth visibility, and human quality ratings. Verified 2026-09-12: blocked on CONSENTED SUBJECTS, which is not a thing code can produce — synthesising faces to fill it would be the fabrication this phase forbids, and a corpus of people who did not agree is the harm it exists to prevent. Everything downstream of it is built and waiting: 182.C.33.08's splits refuse a subject or scenario that leaks across calibration and validation, and the seven items that name this one as their blocker (17.08, 19.05, 19.08, 19.09, 20.01, 20.07, 32.04) each have their machinery in place. blocked:corpus
  • 182.C.33.02 Balance the corpus across gender presentation, age, complexion/skin tone, facial structure, hair/facial hair, glasses, head coverings, disability, body type, wardrobe, and lighting. Verified 2026-09-12: blocked on CONSENTED SUBJECTS, which is not a thing code can produce — synthesising faces to fill it would be the fabrication this phase forbids, and a corpus of people who did not agree is the harm it exists to prevent. The buckets to balance INTO are declared: human-video-fairness-slices.ts carries all eight slices with their assignment rules, and refuses an aggregate that carries no slices. blocked:corpus
  • 182.C.33.03 Include English, Ghanaian English, Akan/Twi, Ga, Ewe, multilingual and code-switched speech, local names, numbers, acronyms, emotional styles, and pace variation. Verified 2026-09-12: blocked on CONSENTED SUBJECTS, which is not a thing code can produce — synthesising faces to fill it would be the fabrication this phase forbids, and a corpus of people who did not agree is the harm it exists to prevent. The locale machinery is built and the two conditions that CAN be made from a signal — pace variation, overlapping speech — were measured in 182.C.18.08 and found two real holes. Accent and language need speakers. blocked:corpus
  • 182.C.33.04 Include frontal, three-quarter, profile, back-turn, occlusion, face-size, focus, compression, low-light, motion-blur, and fast-camera cohorts. Verified 2026-09-12: blocked on CONSENTED SUBJECTS, which is not a thing code can produce — synthesising faces to fill it would be the fabrication this phase forbids, and a corpus of people who did not agree is the harm it exists to prevent. The cohorts exist as declarations in human-video-challenge-cases.ts, each carrying the answer that is correct for it — and for most of these that answer is DECLINE TO SCORE rather than pass or fail. blocked:corpus
  • 182.C.33.05 Include stationary, walking, sitting/standing, gesturing, cooking, prop interaction, dancing/sports-like motion, and multi-character conversational scenarios. Verified 2026-09-12: blocked on CONSENTED SUBJECTS, which is not a thing code can produce — synthesising faces to fill it would be the fabrication this phase forbids, and a corpus of people who did not agree is the harm it exists to prevent. human-video-active-scene-cases.ts holds the matrix these clips fill, and already refuses a case whose speech window barely overlaps its condition window — which is how a walking clip that walks, stops, then speaks fills a cell without testing anything. blocked:corpus
  • 182.C.33.06 Create synthetic negative fixtures with known offsets, drift, dropped/inserted words, wrong speaker, silent mouth, mouth-on- silence, identity swap, time stretch, and damaged edit boundary.
  • 182.C.33.07 Define blinded multi-rater human protocols for sync, intelligibility, exactness, identity, naturalness, control adherence, and overall releasability with inter-rater agreement.
  • 182.C.33.08 Split corpus by subject and scenario to prevent identity or prompt leakage between calibration, validation, and held-out release gates.
  • 182.C.33.09 Version dataset, consent, labels, evaluator outputs, and allowed uses; remove revoked subjects and invalidate dependent baselines.
  • 182.C.33.10 Maintain a small public/licensable smoke corpus and a protected full corpus so CI can test contracts without exposing biometrics.

182.C.34 Model Bakeoff and Promotion Benchmarks#

  • 182.C.34.01 Run the same canonical requests through Wan, Seedance, Veo, HeyGen avatar, direct providers, and eligible successors without provider-favoring prompt changes.
  • 182.C.34.02 Compare native supplied-audio, generated-dialogue, avatar, edit/extension, and corrective lanes separately.
  • 182.C.34.03 Record exact speech, sync, speaker attribution, identity, anatomy, motion, control adherence, visual quality, first-pass/eventual pass, latency, cost, and evaluator coverage.
  • 182.C.34.04 Publish confidence intervals, cohort sizes, failure examples, human agreement, and known evaluator blind spots—not only aggregate ranks.
  • 182.C.34.05 Test every advertised duration, aspect, resolution, reference class, audio mode, and important provider option actually used by routing.
  • 182.C.34.06 Repeat benchmarks across at least three time windows to detect provider/model drift and non-deterministic outages.
  • 182.C.34.07 Define minimum sample size and promotion gates for each routing cohort and quality tier.
  • 182.C.34.08 Archive redacted payloads, exact output digests, evaluator versions, cost, catalog snapshot, and review decisions for reproducibility.
  • 182.C.34.09 Re-run a bounded benchmark on provider/model/schema change and automatically demote routes that regress beyond tolerance.
  • 182.C.34.10 Use benchmark evidence to choose defaults; do not enshrine the 2026-09-09 snapshot as architecture.

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

  • 182.C.35.01 Add schema tests for valid/invalid references, cast, speakers, exact audio, controls, quality policies, consent, and outputs.
  • 182.C.35.02 Add provider payload golden tests for every supported model and ensure unsupported required controls fail before network calls.
  • 182.C.35.03 Add router tests for capability, policy, health, quality, locale, retention, cost, fallback, canary, quarantine, and no-silent- downgrade invariants.
  • 182.C.35.04 Add job-state property tests for idempotency, duplicate callback, out-of-order event, retry, restart, cancel, timeout, and resume.
  • 182.C.35.05 Add upload/storage tests for MIME deception, corruption, range, expiry, cross-tenant access, deletion, and large streamed media.
  • 182.C.35.06 Add evaluator tests with precisely shifted/stretched audio, wrong speaker, face loss, identity swap, VFR, and re-encode fixtures.
  • 182.C.35.07 Add release-gate tests proving that no missing, failed, or waived hard-gate measurement can produce a releasable URL.
  • 182.C.35.08 Add consent/security tests for expiry, revocation, provider reroute, collaborator removal, subject deletion, and webhook replay.
  • 182.C.35.09 Add deterministic media-processing tests for duration, timestamp, concat, mux, loudness, captions, renditions, and provenance.
  • 182.C.35.10 Run targeted lint, typecheck, tests, and builds for every affected package before checking this stage.

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

  • 182.C.36.01 Build a hermetic fake-provider server for create, poll, callback, authenticated download, usage, errors, expiry, throttling, and malformed media.
  • 182.C.36.02 Test the full exact-audio active-person flow from upload and consent through real byte evaluation, release proof, and download.
  • 182.C.36.03 Test multi-reference, multi-character, long-form, edit, extension, avatar, and optional correction flows independently.
  • 182.C.36.04 Inject worker crash, DB restart, queue redelivery, callback/poll race, provider timeout, expired download, object-store outage, evaluator outage, transcode failure, and review delay.
  • 182.C.36.05 Prove failover never resubmits an expensive job twice or releases a stale/unmeasured candidate.
  • 182.C.36.06 Add Playwright coverage for web authoring, consent, reference feedback, audio approval, estimate, submit, cancel, resume, take comparison, failure, review, release, and deletion. Worked 2026-09-12. Authoring and consent display are covered by human-video-studio.spec.ts — re-run today in a real browser, 11 passed and 1 skipped. THE BLOCKER FOUND WAS IN THE HARNESS, not the surface: the e2e BFF sets OSHUN_V1_DATABASE_URL and none of the human-video env, and every listing in that surface refuses outright without a cursor-signing secret, so jobs, candidates, the review queue and project listing all answered 503 and no spec could have driven them whatever it asserted. The config now passes a cursor secret and a staging root; the CPU tool paths are deliberately left out, so identity packs and audio takes stay unbound under test and say so at boot rather than resolving a machine-specific path to nothing. The remaining flows need fixture seeding through the now-bound stores. Worked again 2026-09-12 and THE SECOND BLOCKER WAS NOT THE HARNESS: no code in this service has ever issued an hv: scope. buildAuthSession mints ['domain:*'] for every customer, satisfiesScope keeps domain:* out of the hv: namespace on purpose, and grepping every issuance site found nothing else — so all thirty-two endpoints answered 403 to every signed-in creator, and the surface was UNREACHABLE rather than gated. The two are indistinguishable from the client, which is why a spec that drove them would have asserted the fail-closed behaviour and called it coverage. human-video-entitlement-grant.ts makes the entitlement decision the 182.C.25.02 comment left open: the capability rides a paid plan, the grant is re-read per request rather than minted into a token (a lapsed plan must lose reach over reference images and their embeddings at once, not at token expiry), it is enumerated scopes rather than hv:* because decodeCursor re-authorizes against the exact minting scope and a wildcard-only caller cannot page any list past its first screenful, and hv:review:write and the two reviewer roles are refused at runtime rather than merely omitted. assertGrantCoversSurface fails the boot when the surface declares a scope nobody can hold. Eleven module tests, six over a real Fastify instance; the remaining flows still need fixture seeding. AND THE GRANT WAS NOT ENOUGH: run against a real BFF, all five listings still answered 401 missing_tenant_claim, because the customer token carries no tid either — a second gate in front of the first. The tenant is now resolved at this surface's own boundary and a consumer's is THEIR OWN rather than the shared 'house' the rest of the repo falls back to, because every Caller here is built with projectIds: null and one shared consumer tenant would list every creator's projects to every other creator, with hv:assets:read reaching the faces behind them. With both closed, the surface answered a browser for the first time in this phase, and the first run found a third defect it had been hiding: presetPort, uploadPort and audioPort each passed an ABSOLUTE url to an api client that prepends its own base, so the preset region's error was a doubled host rather than the scope it was attributed to on 2026-09-11 — three of the four ports had it and the only one that did not is the only one whose region ever worked. portUrls.spec.ts reads the four sources and fails on an absolute path, because each port's own spec mocks api and asserts the result rather than the path. COVERED IN A REAL BROWSER (16 chromium tests, all green): authoring (create, open, edit, autosave at a service version), resume across a reload, consent display, the free-plan refusal naming the PLAN rather than a scope, cross-creator isolation, and every region stating what it is doing. NOT COVERED and why: reference feedback past the empty state, audio approval, estimate, submit, cancel, take comparison, review, release and deletion all hang off a CAST MEMBER, a cast member is somebody who has authorized their likeness, and the screen that records that is enrolment (182.C.26) — the known caller gap 182.C.24.02 already names. Submit additionally needs identity packs, which need the face detector this deployment deliberately leaves unbound. The box stays unchecked for those nine. RE-MEASURED LATER THE SAME DAY, after 182.C.23.03 shipped. Three of the nine moved. AUDIO APPROVAL is covered: a recording is uploaded, registered as a take with measured CTC timings, and the approval is offered and refuses with its reason — and reaching it required exposing ETag, which is not CORS-safelisted, so the tag every If-Match needs was being discarded by the browser and no decision on this surface was reachable from the web app at all. RESUME and FAILURE were already covered. AND EVERY PORT NOW BINDS ON THIS BOX: opencv-python-headless was the one missing package (the YuNet weights were already staged, sha256 matching the one face-geometry.py documents), so with the face detector, the probe, the aligner, the synthesiser and a filesystem object root set, the boot line reads "answering 503 until bound: NOTHING" for the first time in this phase. The numpy upgrade it pulled was checked against the committed alignment ground truth, which still passes. SO WHAT REMAINS IS NOT ADAPTERS, IT IS TWO MISSING SCREENS. submit-job refuses without a likeness AUTHORIZATION, and nothing in the API creates one — withdraw-consent exists and no create does, because an authorization is the subject's act and belongs to enrolment (182.C.26). And reference feedback stops at the upload: the Studio has no control that DECLARES an identity pack, so quality stays not-read on every reference and the face-mapping and eligibility controls that already exist are fed by nothing. estimate, submit, cancel, take comparison, review, release and deletion all sit behind the first of those. REFERENCE FEEDBACK IS NOW COVERED, and closing it found TWO DEFECTS BEHIND THE MISSING CONTROL rather than only the control. packPort.ts and the pack block in ReferenceSet declare the set, which is what the per-reference quality, the face picker and the eligibility line were always meant to be fed by. FIRST, THE QUESTION A PACK ASKS WAS UNANSWERABLE: a mapping names a faceId, an id that exists only inside one detection run, and the run was computed and discarded — so the permitted answers were gone, and re-detecting to recover them is not the same run, which is the whole finding of the 2026-09-10 probe (two models, 2 and 6 faces, one photograph). The detections are kept with the pack now and the document offers the candidates for the reference that is waiting, only that one, because a settled reference has no question and listing every face in every photograph hands out more about the people in the background than the decision needs. SECOND, "A LATER CALL AGAINST THIS ID" — which the create handler's own comment has promised since it was written — DID NOT EXIST, so the only way to supply a mapping was to declare the pack again: a new detector run with different ids, and a new pack id, which an authorization binds its digests to. Fixing the creator's slip would have cost the subject's consent. POST /identity-packs/:packId/mappings is the thirty-third endpoint, deciding, re-evaluated against the STORED run, a replacement per reference rather than an append (two mappings for one reference is mapping_conflicts, so appending refuses the act of correcting), and a version-conditional write because If-Match is checked before the handler runs and two callers holding the same fresh tag both clear it. On the web side the verdict is an EDIT rather than a reseed: it arrives in answer to reference/map-face, so edits are always pending when it lands, and draftReducer refuses a seed then — correctly, a refresh does not get to take the creator's work — so it has its own reducer action that applies it without counting as an edit. COVERED IN A REAL BROWSER, two chromium tests on the live stack with the detector bound: a faceless image is refused with the service's own sentence and the row stays not-read rather than being cleared by a failed declaration; and the committed multi-face evidence image from that same probe produces the real question, the picker offers the faces that answer it, each choice says WHERE the face is, and the answer clears the pack with the service's verdict rather than this browser's optimism. The first run of it also found a disclosure: the detector's refusal carried execFile's Command failed: <full argv> into the error envelope, so the interpreter path, the script path, the weights path and THE STAGING PATH OF THE UPLOADED REFERENCE rendered in the Cast region; the status or signal plus stderr says everything diagnosable without any of them. WHAT IS LEFT IS ONE SCREEN, NOT TWO: estimate, submit, cancel, take comparison, review, release and deletion all sit behind a likeness authorization, which is the subject's act and belongs to enrolment. WORKED AGAIN 2026-09-12, AND THE NINE THAT WERE LEFT WERE NOT WAITING ON A SCREEN NOBODY HAD DESIGNED — THEY WERE WAITING ON THE REGION THE OTHER SIX POINT AT, WHICH WAS NEVER ON THE PAGE. 182.C.23.08 shipped human-video-studio-session.ts on 2026-09-11 — the stage, whether a bar may be drawn, what pressing cancel would actually do — and NOTHING ON THIS SURFACE EVER CALLED IT. Same for 182.C.23.09's take comparison and 182.C.23.12's export controls: three modules, each delivered and tested, each consumed by no browser. So the Studio had six regions and not one of them submitted anything, and takes could only ever be empty — not because no provider is bound, but because no job had ever been created from a browser at all. Every endpoint behind it has answered since 182.C.25.01. SUBMIT, HONEST STAGE PROGRESS AND CANCEL ARE NOW COVERED IN A REAL BROWSER. A seventh region, delivery, sits between the plan and the takes; jobPort.ts calls submit-job, get-job, list-jobs and cancel-job. It is a FORM and not a button because a submission is checked against the nine dimensions of the authorization the subject granted and the handler is explicit that an unstated dimension is a refusal rather than a wildcard — so every chooser opens on "not stated", each says why it is being asked, and the submit control names the missing answers instead of spending a round trip to be told. 2 chromium tests on the live stack, 38 module tests, and six more over the job handlers. The 22 browser tests the three earlier human-video specs already had were re-run with the seventh region on the page and stayed green, axe included. AND BINDING IT FOUND THE JOB DOCUMENT WITHHOLDING BOTH THINGS THE SHIPPED MODULE COMPUTES. jobDocument dropped stageEnteredAt, which the stored record has carried all along — so the only clock a client could reach was the one assertServerAnchored exists to refuse: a counter started when the tab first saw the job, which restarts at zero on every reload and reads as progress for a render that has been going an hour. And it sent no cancel affordance, so a surface had to re-derive 182.C.15.05's stage table in the browser — the drift cancelAffordanceFor warns about in its own comment, and wrong in the direction that costs money. Both ride the document now, beside progressSource, barIsDeterminate, isTerminal and notifiable. No elapsedMs: a field that changes on every read of an unchanged job makes two identical versions serialize differently. THE BROWSER FOUND A THIRD THAT NO UNIT TEST COULD. Everything passed up to the last assertion, where the cancelled job rendered as "Stopped from : " with both halves empty. ONE RESOURCE WAS ANSWERING IN TWO VOCABULARIES: the cancel response composes cancelledFromStage and explanation, and the job document embedded the raw CancellationDecision, which spells the same two facts fromStage and detail. A client written against the response lost both the moment it re-read the job — which is exactly where a browser lands after a reload, and a reload is the NORMAL case in a flow that waits on a provider for minutes. Both now go through one cancelledView mapper, so the sentence a creator reads after a reload is the same sentence they were shown when they pressed the button. WHAT IS LEFT, AND WHY IT IS NOT A SCREEN THIS TIME: take comparison, review and release. No provider is bound in this deployment, so a submitted job rests at validating and produces no candidate — and a comparison of nothing is not a test of a comparison, nor is releasing nothing a test of a release. Estimate is its own gap and a real one: there is no estimate operation among the thirty-nine, human-video-cost-estimate.ts is implemented and reachable from no endpoint, and 182.C.23.06's own note says eligibility "is answered against the routes just before you submit" — which is the call that does not exist. TAKES IS FED NOW, AND FEEDING IT FOUND THE WEB UNABLE TO SAY WHAT THE PHONE CAN. notRead.takes had been true on every project since the region was written — nothing on this surface ever read a job's candidates — so Takes could only ever report itself PARTIAL: "the candidates of this project's jobs were not read here". readJobTakes reads the listing and then each candidate's MEASUREMENTS, because the listing says only whether something looked, and a region drawn from evaluated: true alone would be a review of nothing. A job that has produced nothing now says "nothing has been generated yet", which is a sentence about the work rather than about the screen. And measurement.measured is null and never [] for a candidate nobody evaluated, so a document flagged evaluated with null rows is called NOT evaluated rather than rendered as a clean review. THE FOURTH OUTCOME: Lilith's MeasuredOutcome had three and MeasuredFinding.passed on the wire is boolean | null — the null being 182.C.19.08's unfitted identity operating point, a number with no verdict. Both three-outcome answers are wrong for it: "not measurable" denies a measurement that happened and a pass or fail invents the threshold. The phone settled this on 2026-09-11 and the design system never followed, so anything the web rendered for an unfitted gate had to be one of the two lies. measured-no-verdict is now a fourth outcome with its own shape and the phone's wording verbatim, pinned by a test, because two spellings of one verdict is how a creator and a reviewer end up disagreeing about a take neither of them mis-read. DELETION IS COVERED FOR THE PATH THIS DEPLOYMENT CAN REACH, which is the withdrawal: ending a permission destroys the references it covered, and human-video-delivery.spec.ts drives it in a browser and reads the manifest that comes back. delete-asset is NOT covered and cannot be here — plansFor resolves an asset id to a TAKE, so it needs a candidate, and no provider is bound. Measured 2026-09-12 rather than assumed. WORKED AGAIN LATER THE SAME DAY, AND THE ESTIMATE GAP IS CLOSED. human-video-cost-estimate.ts had priced a job since 182.C.14.06 and was reachable from NO ENDPOINT, so the region that asks a creator for nine permission answers then offered a control that spends money with no number beside it. POST /jobs/estimate is the fortieth operation. It prices the plan against the routes LANE_ADAPTERS would actually send to — the same table LaneRenderRequestBuilder selects through — rather than the whole catalogue, because a price for a model this service would never choose is a number a creator budgets against and never pays. AND IT ASKS FOR NO AUTHORIZATION DIMENSIONS: an estimate is not a use of anybody's likeness (it reads a public price list and multiplies; no row is read and nothing is written), so requiring a purpose and a geography before showing a price would make somebody declare what a face is for in order to read a rate card, and would put the estimate after the decisions it exists to inform. THREE THINGS IT WILL NOT DO. It states a FLOOR and says so: the number of passes per released candidate comes from a measured release rate and nothing here has released anything, so the generation leg is reported unknown and the derivable half — one pass over every shot — is stated beside it rather than folded into a total. It refuses to price a route it cannot: both Seedance families price per video_tokens, whose tokenizer is the vendor's, and they come back null with that sentence rather than a guess. And NO ROUTE IS SAID TO BE AVAILABLE — eligibility needs a probe against the dated slug, a benchmark promotion, a region record, terms and health, and no lane is registered on any deployment, so every route carries proven: false with the reason next to its own figure. The first implementation of CatalogueReader came with it: the port every routing decision goes through had no adapter anywhere, and human-video-catalogue-reader.ts is the live OpenRouter video catalogue behind it, with the date and the content hash of the bytes travelling beside the prices. A catalogue that cannot be read answers a TRANSPORT code rather than store_unavailable, because an operator reading the second goes looking for a setting that does not exist. estimateJob generalises the estimator to a whole job and fixes an arithmetic trap on the way: a per-generation minimum applies PER SHOT (four one-second shots on runway/aleph-2 pay the $0.56 floor four times, where summed durations pay it once) while per-job legs such as encoding are counted ONCE, and a single unpriceable shot makes the whole leg unknown and names that shot rather than totalling the rest into a smaller number that looks complete. The surface grew a checked changesNothing: a POST that writes nothing is exempt from the idempotency rule, and an endpoint declaring itself safe while also deciding or deduplicating now fails the boot check. Measured live against the real catalogue (29 models): exact dialogue, 6s + 4s at 1080p, alibaba/wan-2.7 $1.00 and wan-2.6 $1.50; text-to-scene 5s 720p, google/veo-3.1-lite $0.15 cheapest of seven routes with both Seedance lanes unpriceable. COVERED IN A REAL BROWSER (1 chromium test on the live stack, which reads the live price list): nothing is shown before the press, the control is refused without a mode, a real figure and both caveats render, and the submit control beside it is still refused — which is the point of asking first. 57 module tests behind it: 22 over the handler, 9 over the catalogue reader, 6 more over the estimator, 11 over the web port and 9 over the region. AMENDED LATER THE SAME DAY: proven shipped as a CONSTANT — the one field in that response that was not derived — because no lane registry existed to ask. One does now, so the field reports what evaluateLaneEligibility decided with the engine's own blockers as the reason, and a test asserts a route CAN come back proven so the field cannot quietly become a constant again. The answer is still no on every route; the difference is that it is now a decision. WHAT IS LEFT OF THIS ITEM IS STILL take comparison, review and release: they need a candidate, and no provider is bound.
  • 182.C.36.07 Add Playwright admin coverage for quarantine, review, override rules, reroute, takedown, audit, and role isolation. Verified 2026-09-12: TWO of the seven are covered by apps/oshun/admin/e2e/human-video-review-queue.spec.ts (five tests) — review, and role isolation through the scoped admin session. The other five have no admin surface to drive: apps/oshun/admin/src/app/isis/ human-video/ is one page, the review queue panel, and quarantine, override rules, reroute, takedown and audit are each a screen that does not exist. A spec that tapped through them would pass by finding nothing. SWEPT AGAIN 2026-09-12 AND "A SCREEN THAT DOES NOT EXIST" UNDERSTATES IT: none of the five has a bound source behind it either. human-video-operator-override.ts is pure logic with no store and is bound nowhere in the runtime; human-video-appeals.ts has no store interface at all; quarantine rows come from a failed download and takedowns from a released take, so both wait on a provider; and reroute waits on the model registry that buildStudioLaneView needs. Building the pages first would put five screens over nothing, which is the same mistake in the other direction. UPDATED 2026-09-13: TWO OF THE FIVE HAVE MOVED, and one of them was unblocked by exactly what this note predicted. REROUTE no longer "waits on the model registry that buildStudioLaneView needs" — CatalogueLaneRegistry is that registry and it ships. OVERRIDE RULES no longer have "no store": human-video-override-store.ts and the human_video_operator_override table are the stack that was missing, and the lane registry now applies REAL overrides instead of the literal empty list it shipped with, so an operator's judgment reaches a routing decision with the un-overridden decision kept beside it. There is still NO SCREEN for either, and a spec that tapped through one would still pass by finding nothing — so the box stays unchecked, and what these two are waiting for has changed from "no stack" to "no page". Quarantine and takedown still wait on a provider (their rows come from a failed download and a released take). AUDIT NOW HAS ITS STACK TOO (2026-09-13): human-video-case-store.ts with human_video_case and human_video_legal_hold behind it. The case table has NO TENANT COLUMN on purpose — a takedown is started by the depicted person or a member of the public who never had an account, and scoping the row to a tenant would make assertOpenToUnauthenticated's refusal unenforceable because there would be no row to write. Acknowledgement and resolution are separate columns because they are separate promises, and isOverdue is asked rather than re-derived. A hold is refused unless it names its artefacts and expires, and it is LIFTED rather than deleted, because a hold taken off early is as interesting as one that ran out. So THREE of the five have moved from "no stack" to "no page": override rules, reroute and audit. Quarantine and takedown still wait on a provider, and nothing here wires a hold into the erasure path — planDeletionUnderHolds is still consumed by nothing, and adding a hold check to a destructive path is its own change rather than a side effect of giving holds somewhere to live. AMENDED LATER THE SAME DAY: NONE OF THE THREE STACKS COULD BIND. The generated Prisma client under libs/oshun/persistence/src/generated is CHECKED IN and was not regenerated for either migration, so delegateWith found no delegate for humanVideoOperatorOverride, humanVideoCase or humanVideoLegalHold, all three stores resolved to null, and the boot line listed them unbound — which is indistinguishable from a deployment that left the port out on purpose, because that is exactly what the bound/unbound list is for. So the lane registry went on applying the literal empty override list it was said to have stopped applying. Regenerated, and generated-client-drift.spec.ts now asserts that every model in the committed schema has a delegate in the committed client, proved by reverting the client rather than asserted: it names the three. AND THE TWO FLOWS THIS ITEM CALLED COVERED WERE NOT. Measured by injection 2026-09-13: an admin session carries admin:* and admin:workspace:isis, satisfiesScope stops the wildcard at the hv: namespace on purpose, and NOTHING IN THIS SERVICE EVER ISSUED A STAFF SCOPE — the entitlement grant's own comment says the staff scopes are "held by reviewers and trust-and-safety through the admin authentication path", and that path did not exist. GET /reviews answered 403 to every operator who has ever opened the review queue, and the spec passed throughout because it TOLERATED data-queue-state="failed" and asserted the refusal rendered nicely. Same confusion as the web side found on 2026-09-12, in the other direction: unreachable and correctly-failing-closed are indistinguishable from the client, so a spec written for the second passes on the first. Role isolation was the same illusion — everything was refused, which looks like isolation. human-video-staff-grant.ts is the missing half: admin:*/admin:studio/admin:workspace:isis carry the queue read and the platform reach; each elevated role needs its OWN named claim and is never covered by a wildcard, because 182.C.29.04 makes the role half of what unlocks a member of the public's photograph; and NO decision scope is issued at all, so 182.C.29.08's "coverage before mutations" now lives in the grant instead of a comment on a page. THE REACH IS A SCOPE AND NOT AN INFERENCE: a reviewer's queue is every tenant's waiting takes, so hv:reviews:all-tenants is demanded by buildQuery before it will drop the tenant filter, re-authorized when a cursor minted under it comes back, and refused to hv:*. Binding it found two more: the queue looked each row's speakers up under the CALLER's tenant, which on a platform read finds no job and drops every row as depicting nobody — an empty queue, the one answer this queue may never give by accident — and the keyset position was a bare candidate id whose tiebreak re-composed a key from the caller's tenant, which belongs to none of the rows. 26 tests: 10 over the grant, 5 over a real Fastify instance, 7 over the reach and the cursor, 3 in the store, 3 in the handler, plus the two the admin's stale duplicate port spec lost (it had been red since the role parameters were removed). AND IT IS VERIFIED IN A REAL BROWSER, which this worktree was not supposed to manage: the admin e2e needs the BFF and this worktree has no workspace link farm, so server.ts cannot boot. What CAN is the human-video surface on its own — createHumanVideoRuntime over the local Postgres behind registerHumanVideoApiRoutes, 14 ports bound, with only the surrounding app faked (/v1/admin/session/me). Against it: an Isis operator reads the queue 200, a support-workspace operator is refused 403, a creator holding hv:candidates:read sees zero rows and so does one holding hv:*, and one seeded take in ANOTHER tenant renders with data-queue-state="ok", its row id, Tenant tenant-harness-customer and the gate that failed. 8 chromium tests green, axe included. The single red on the first pass was a cold next dev compile of /unauthorized exceeding a 10 s expect timeout — it passes alone and passes warm, the dev-server artifact CLAUDE.md warns about rather than a defect. Closing it found a THIRD stale spec left red by an earlier fix: HumanVideoReviewQueuePanel.spec.tsx omitted the queue's reviewerRole entirely and undefined is not null, so the panel rendered "Read as , which is the role the SERVICE applied" — a sentence claiming the service decided and naming nobody. It coalesces now, with three tests over the three branches. WORKED AGAIN 2026-09-13 AND THE FIRST OF THE FIVE FLOWS HAS A PAGE: override rules and reroute, which are one screen because they are two questions about one decision and two screens would show two versions of it. GET /operator/lanes/:mode is the forty-first operation — every route in today's catalogue that serves a named mode, its eligibility decision, and every override applied to or refused for it. A get- and not a list-, because the rows are bounded by what the vendor publishes rather than by anything a tenant wrote, and paginating a fixed vocabulary tells a client it grows with use. THE MODE IS A PATH SEGMENT and the SDK drift guard is why: the generator emits a query string only for an endpoint with a declared lists target, so a mode in the query produced a generated method with no way to pass one — a client generated from the surface that could not call it, which is the exact drift the generator exists to prevent. Its scope hv:lanes:read is staff-only: reading who authorised an exception and why is the one thing the Studio capability view deliberately withholds from a creator, and human-video-operator-lane-view.ts is the mirror image of that file — structural, no spread of a registration anywhere, but the owner, the reason, the expiry and the real health all cross. BOTH DECISIONS TRAVEL AND THE PAGE RENDERS BOTH: a lane that is ineligible on its own terms and shipping under an exception is the row that matters most and is invisible on a screen with one verdict. A REFUSED OVERRIDE IS SHOWN TOO, because its author is the person who needs to know it did not apply. The write half is absent and says so — the operator grant issues no scope that writes, so 182.C.29.08's sequencing is enforced rather than commented. 17 module tests (9 over the handler, 8 over the admin port) and 10 chromium tests on the live stack, axe included, against a database holding three real overrides: two lifting policy_not_approved and region_unrecorded on wan:alibaba/wan-2.7 so it ships under them, and one scoped to another tenant so it comes back wrong_tenant — neither of those two tests skipped. AND THE AUDIT FLOW HAS A PAGE TOO, the third of the five. GET /operator/trust-and-safety is the forty-second operation and answers TWO lists in one read, because they are the two halves of one obligation and they pull against each other: a case is somebody asking for something to stop, and a hold is the reason part of it has not. A queue of unacknowledged reports reads as neglect until you know a hold is why one is unfinished; a list of holds reads as diligence until you know somebody has waited four days for a reply. UNPAGINATED AND CAPPED AT 200, and past the cap it REFUSES rather than truncating — the first two hundred of four hundred, with no way to know, reads as a backlog of two hundred, which is the number somebody acts on. Nothing about the report itself crosses: whether a subject is named rather than who, whether there is a contact rather than what it is, and never the statement, because a triage list read at a glance by whoever is on shift is the wrong place for the text of somebody's complaint about their own likeness. 7 handler tests, 7 over the admin port and 9 chromium tests on the live stack, against a database holding a real four-day-old unacknowledged takedown and two live holds, one lapsing within the day. WHAT IS STILL MISSING FOR THIS BOX: quarantine and takedown wait on a provider, and none of the three human-video admin routes is in the workspace nav — all are reached by URL, as the shipped review page already was. AND THE HOLD IS WIRED INTO THE ERASURE PATH, which the note above listed as deliberately not done. destroy reads the live holds before it destroys anything and planDeletionUnderHolds decides: what a hold NAMES is kept and everything else still goes, because an incident about one video must not become a reason to keep a person's whole reference pack. Four things it took. plansFor answers entries carrying each take's ASSET ID — a hold is placed by somebody reading a review queue, where a take is an id, and the plan identified its master only by digest and object key, so a withdrawal could not have honoured a hold on one of its own takes; both spellings are compared, because the one worked example in this repo names a key. The uploaded bytes are enumerated BEFORE the erasure (sourceLocators) and skipped inside it (a preserve list), because a held key reported afterwards as gone: false is indistinguishable from a delete that failed, which the handler raises on. A suspended destruction is NOT ledgered: the ledger replays, so a stored partial manifest would report the suspension as the final outcome forever and the held take would never be destroyed. And the manifest carries what was preserved, with the same never-overclaim CHECK unreachable has — proved against a throwaway database, where the row claiming completeness beside a preserved artefact is refused by the table. assertHoldsRelevant is deliberately NOT called and the reason is written down: it refuses a caller whose holds name nothing being deleted, and this path passes every live hold, so one unrelated incident would make that guard block the erasure it exists to protect. 14 new tests over four files in the app and two in the library; the human-video suite is green at 291 files / 6063 tests. None of this checks the box: the five flows still have no screen.
  • 182.C.36.08 Add accessibility automation and manual screen-reader, keyboard, zoom, contrast, reduced-motion, captions, and transcript checks. Verified 2026-09-12: the AUTOMATION half is done and passing. axe runs on the web authoring surface and on the admin review queue; keyboard reach through every region and phone-width reflow are Playwright assertions (both in today's 11-passed run); reduced motion, captions and transcript are apps/oshun/mobile/src/humanVideo/reviewAccessibility.ts with its own specs, and its position is the load-bearing one — reduced motion governs CHROME and never the take, because suppressing playback would remove the take from a review of the take, and captions come from the alignment the gates were computed against rather than a second ASR pass that would put different words on the screen from the ones anybody checked. The MANUAL half needs a person with a screen reader, and that is what remains. Board tag 2026-09-18: the automation half is done and passing; what is left is a person with a screen reader, which no agent can stand in for. blocked:human
  • 182.C.36.09 Add mobile automation for the supported critical paths and explicitly document any non-automatable device/media limitation. Delivered 2026-09-11: apps/oshun/mobile/e2e/flows/shell/human-video-review.yaml covers the path this platform has — 182.C.24.01 made the phone review-and-approval — and scripts/human-video/verify-mobile-review-automation.mjs keeps the flow's test ids and the screen in step on a machine with no device, because a flow that asserts a renamed id passes by finding nothing (proved by renaming a control: it fails, and passes again when restored). The six limitations are written down in mobile-parity.md rather than left implicit — and two of them are the same fact twice: 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.
  • 182.C.36.10 Run real provider canaries with tiny bounded spend and redacted evidence before declaring any production route available.

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

  • 182.C.37.01 Load-test upload, callback, poll, download, proxy creation, evaluator queue, review playback, and delivery separately and end to end.
  • 182.C.37.02 Establish capacity models for duration, resolution, candidate count, evaluator GPU memory/time, transcode CPU, storage, and network egress.
  • 182.C.37.03 Add backpressure from object store/evaluator/transcoder to provider admission so completed outputs do not expire before ingestion.
  • 182.C.37.04 Verify bounded memory and disk use under large concurrent videos and clean abandoned temporary files after crash/restart.
  • 182.C.37.05 Add spend reservations, per-job ceilings, daily/monthly budgets, anomaly detection, credits/refunds, and creator-visible final cost.
  • 182.C.37.06 Model cost per releasable second rather than cost per provider attempt, including failed candidates and evaluation/correction.
  • 182.C.37.07 Define provider/evaluator/storage/queue SLO dependencies, degraded modes, recovery objectives, and incident runbooks.
  • 182.C.37.08 Run soak tests long enough to cover slow provider jobs, webhook retries, signed-URL expiry, queue aging, and lifecycle cleanup.

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

  • 182.C.38.01 Ship behind tenant, region, mode, provider, model, quality-tier, and app-surface feature flags with a global kill switch.
  • 182.C.38.02 Roll out internal dogfood → consented design partners → allowlisted beta → capped production → broader release with explicit gates.
  • 182.C.38.03 Keep active-scene exact-dialogue, multi-speaker, long-form, avatar, and corrective lanes independently promotable and reversible.
  • 182.C.38.04 Define canary abort thresholds for consent/safety incidents, bad-sync false accepts, identity drift, provider failures, latency, and cost.
  • 182.C.38.05 Practice provider quarantine, leaked-URL response, unauthorized-likeness takedown, revocation, model regression, and mass release invalidation.
  • 182.C.38.06 Add rollback that preserves job/audit truth and prevents old clients from resubmitting unsupported contracts.
  • 182.C.38.07 Obtain Trust & Safety, privacy, security, legal/licensing, accessibility, SRE, product, and release-engineering sign-off.
  • 182.C.38.08 Update the capability matrix and public product copy only after the promoted route has real production-like evidence. Verified 2026-09-12: this is a PROHIBITION and it currently holds — there is no promoted route with production-like evidence (182.C.40.01), and nothing claims one. capability-audit.md is a gap ledger whose verdicts were reached by reading source, limitations.md says "where nothing was measured, the entry says so instead of estimating", and the preset catalogue deliberately carries NO availability field at all — its note explains that how many routes can serve a preset is a question about one request at one moment, and that "0 routes" would be a claim rather than a missing value. The box is checked when there IS evidence and the copy is updated, so it follows 182.C.40.01 rather than blocking on its own. Board tag 2026-09-18: follows 182.C.40.01, as the note above says. blocked:upstream

182.C.39 Documentation and Operator Guidance#

  • 182.C.39.01 Document creator workflows for reference preparation, consent, script locking, pronunciation, audio approval, scene direction, take review, correction, export, and deletion.
  • 182.C.39.02 Document honest limitations and troubleshooting for profile, occlusion, fast speech, singing, overlapping dialogue, tiny faces, rapid movement, weak references, unsupported languages, and model drift.
  • 182.C.39.03 Publish API guides and complete examples without real credentials, private biometric assets, or misleading mocked success claims.
  • 182.C.39.04 Document provider onboarding, live probing, benchmark, promotion, quarantine, cost, retention, and deletion procedures.
  • 182.C.39.05 Document evaluator acquisition, calibration, thresholds, failure interpretation, fairness slices, upgrades, and rollback.
  • 182.C.39.06 Document review and escalation rules, non-overridable gates, subject requests, takedowns, and incident response.
  • 182.C.39.07 Document media/storage lifecycle, final-byte evaluation, C2PA verification, watermark/disclosure, and release-proof inspection.
  • 182.C.39.08 Create a concise V1 migration guide for existing avatar, voice, and image-to-video callers, and for Veritas callers when V1.2 restores that surface.

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

  • 182.C.40.01 Demonstrate at least one production-configured active-scene route using multiple authorized references and exact supplied audio, with the person moving and speaking outside a talking-head setup. As of 2026-09-09 only the Wan 2.7/2.6 audio passthrough can satisfy this on OpenRouter; if that live probe fails, the item is satisfied by the direct lane from 182.C.13.01 and the evidence must name which route. Worked 2026-09-12: THE PIPELINE NOW EXISTS AND IS PROVEN HERMETICALLY. human-video-render-request-builder.ts turns a submitted job into a provider payload, human-video-provider-transport.ts is the exchange, and human-video-render-worker.ts submits, polls, downloads, probes and files a candidate; human-video-render-end-to-end.spec.ts runs the whole leg through the real stores with only the four replaceable boundaries stubbed. What blocks a REAL run is neither code nor money (~$4.35 of the approved canary budget is untouched): a CONSENTED REFERENCE (182.C.33.01) and an object store the provider can fetch from — placeReferences hands the provider a URL it must retrieve, and the local MinIO is on a private port so a presigned URL to it is unreachable from OpenRouter. A render with no references is not a likeness render and cannot demonstrate this item. Evidence: evidence/2026-09-12-jobs-adapter-live/.
    • (Dependency stated 2026-09-19: depends on phase-182-human-video:182.C.33.01.)
    • (Parked 2026-09-19: its own text says a real run needs a consented reference (182.C.33.01, which needs a corpus of consenting people that no agent can obtain) and an object store the provider can fetch from; it was on offer as ready. The hermetic pipeline it describes is built. Unpark when 182.C.33.01 is met.) blocked:corpus
  • 182.C.40.02 Demonstrate identity continuity across multiple shots and at least one profile/occlusion/high-motion challenge without lowering gates. Verified 2026-09-12: blocked TWICE OVER, and the second is sharper than "needs renders". Since today human-video-continuity-bridge.ts refuses to grade identity against an operating point nobody fitted, so even with footage in hand this item cannot pass until 182.C.19.08's re-fit happens — which is 182.C.33. "Without lowering gates" is exactly what that refusal protects: the alternative was grading against the inherited 0.35, which admitted two strangers at 0.422. Board tag 2026-09-18: waits on the re-fit of 182.C.19.08, which waits on the consented corpus of 182.C.33. blocked:corpus
  • 182.C.40.03 Demonstrate multi-speaker attribution or explicitly keep it feature-flagged if its benchmark gate is not met. Worked 2026-09-12: NEITHER ALTERNATIVE WAS TRUE. WAN_FAMILY refuses only video-edit and talking-head-avatar, so multi-character-scene was servable, and nothing has benchmarked which face says which line — servable with an unmet gate is the worst of the two states. human-video-multi-speaker-gate.ts now holds the mode back, and the flag READS cohortVerdict rather than being a boolean somebody sets, because "keep it flagged until the benchmark passes" implemented as a constant is a constant somebody flips on a one-character diff nobody can tell from the one that was earned. not-promoted and blocked stay apart in what a creator is told: one is a schedule, the other is a bug report. Found alongside it that submit-job read mode as any non-empty string while the contract has a closed seven-value enum, so an unknown mode was accepted with a 202 and failed at render time; it now reads a member. The box stays unchecked until somebody decides this counts as the flagged alternative. Decided 2026-09-18 under the owner's delegation: a mode held back by human-video-multi-speaker-gate.ts reading cohortVerdict IS the "explicitly keep it feature-flagged" alternative this item offers — the flag is earned by a benchmark, not set by hand, which is stricter than what was asked. What is left for an agent: verify with both passes that no route serves multi-character-scene while the verdict is not-promoted (submit-job, the preset catalogue, the Studio and the mobile review screen), and that the creator-facing message tells not-promoted from blocked; then this box can be checked on that evidence.
  • 182.C.40.04 Demonstrate creator-approved script/audio, delivered-word validation, real frame/audio sync evaluation, identity/artifact evaluation, final-byte recheck, and a cryptographically bound release proof. Verified 2026-09-12: the chain is built up to the release and stops there. Approval binds to the speech digest, the aligner and sync evaluators run on CPU, and the render leg files a candidate — but the RELEASE PROOF needs deps.proofs, and releases is one of the two ports still answering 503. release() already refuses without it in the right words: publishing without checking the sixteen gates would be a release that passed nothing, and a proof set silent about a gate reads as a release that passed it. WORKED AGAIN LATER THE SAME DAY: the release port is BOUND and the refusal is now about the pipeline rather than the deployment (evidence/2026-09-12-fourteen-ports-bound/). The proof set is DERIVED from what a candidate stores, so a gate is present exactly when the thing it attests to is stored, and human_video_release's CHECK requires all sixteen — a row cannot exist for a take that passed fifteen. Driven live against a measured take, the refusal names TEN: nine gates nothing in this pipeline produces (no C2PA manifest, watermark report, disclosure record, output-shape or encoding measurement, policy or rights decision, quality aggregate, or signed provider attribution is recorded against a candidate anywhere — UNPRODUCED_GATES names each with the item that closes it), plus identity-preservation, which was measured at 0.71 and is absent because passed is null: a number exists and nothing fits a verdict to it, which is where 182.C.19.08 left the operating point. So this item now needs those nine producers and a real render, and the list is a work queue rather than a missing adapter. WORKED AGAIN LATER THE SAME DAY: TWO OF THE TEN ARE NOW PRODUCED, and what was missing turned out not to be a measurement. output-shape-valid and encoding-conformance have had every number they need since 182.C.16.03 — probeMediaFacts reads the display matrix, the geometry, both frame rates, the pixel format, the colour tags and the audio layout out of real ffprobe — and NOTHING KEPT THE ANSWER AGAINST THE TAKE IT WAS ABOUT, so at release time there was nothing to check. A candidate now carries shape: the measured facts whole, plus the four numbers the shot DIRECTED, snapshotted at render time for the same reason the lineage snapshots the audio digest — reading the job's controls at release would grade a take against instructions that may have been re-directed since. The facts are stored and the VERDICT is not, so both gates are derived at release and a profile edited next month re-decides every take rather than only the ones rendered after it. output-shape-valid compares length, displayed aspect, resolution and frame rate, each answered separately because "the output is wrong" sends nobody anywhere and "it came back 24fps for a 30fps request" sends them to the lane's frame-rate mapping; the resolution tier is read off the SHORT edge, because 1080p portrait is 1080x1920 and reading the tier as a height would fail every portrait take. encoding-conformance asks a DIFFERENT question, because a candidate holds the provider's untouched master and holding that to DEFAULT_DELIVERY_PROFILE directly would fail almost every take for tagging the conform step exists to fix: it asks whether planConform can reach the house profile from these bytes, which is what a release actually needs since the delivered file is derived from this master. Three states each, and the third is the one a written verdict would have lost — nothing measured it (ABSENT, never-attempted), the measurement disagrees with what was directed (ABSENT, measured-failed), or it agrees (PRESENT). A number nobody could read is not a pass either. FfprobeTakeProbe is the producer, and it measures LOUDNESS too, because the house profile names a target and planConform refuses to invent a gain from a file nobody decoded — found by running the gate against a real file rather than by reading the profile. It reads a LOCAL path and says so: a bucket-backed deployment gets null from resolveTakeProbe and a boot line naming which setting is missing, rather than a prober that would have to invent the numbers. 26 tests, 7 of them against files real ffmpeg wrote and real ffprobe read. WHAT IS STILL TRUE: the render leg that would CALL this prober is not bound, so a candidate's shape is null on this deployment and both gates read as absent-never-attempted; and for a take whose shape IS measured the refusal names eight rather than ten — the seven with no producer, plus identity-preservation's unfitted operating point. AND A THIRD GATE GOT A PRODUCER THE SAME DAY: provider-lineage. Its entry said "the lane and model are stored and NOT SIGNED; an unkeyed digest cannot establish who issued it" — and a candidate's laneId is a row anybody with write access can edit, so a release that read it would attribute a take to whatever the database last said. What this can HONESTLY attest is stated in the module and in the field names: OpenRouter signs nothing, so this is Oshun asserting under its own Ed25519 key that these exact bytes came back from this lane, on this dated model version, through this provider slug, on this date. A manifest implying the vendor had signed would be the fabrication 182.C.28.05 is about. The digest is INSIDE the signed statement, so a signature lifted from another take fails against this one, and the gate is DERIVED BY CHECKING rather than by reading: algorithm, then digest, then signature, in that order, because a verifier that checks the signature first and the digest afterwards has already said yes to somebody else's take. Three outcomes again — nothing signed one (absent), a signature this deployment holds no key to check (absent, and the reason names the deployment rather than the take), or a checked verdict. It shares OSHUN_C2PA_SIGNING_KEY with the audio manifest signer deliberately: a deployment with two provenance keys has two answers to "who issued this" and a verifier holding one reads the other's output as a forgery. RenderRequest gained modelVersion and providerSlug to carry it, the dated canonical_slug rather than the alias, because an alias names a model and only the dated slug names the version that made the bytes. 17 tests, including a lane edited on the row after signing and a valid signature lifted onto other bytes — both refused. UNPRODUCED_GATES is now SIX: C2PA manifest, watermark coverage, disclosure record, quality aggregate, policy decision, rights decision.
  • 182.C.40.05 Demonstrate automatic rejection and selective retry/reroute on intentionally bad sync, wrong words, wrong speaker, identity swap, malformed output, provider outage, and evaluator outage. Verified 2026-09-12: SPLITS CLEANLY IN TWO. The transport faults — provider outage, malformed output, and the rejection that follows — are already demonstrable hermetically and are demonstrated: human-video-provider-double.ts scripts throttled, html-error-page, partial-download, completed-without-output and create-refused, and human-video-render-end-to-end.spec.ts drives two of them through the real stores to show nothing is filed and nothing is left in the object store. The CONTENT faults — bad sync, wrong words, wrong speaker, identity swap — need a take that is actually wrong in those ways, which is a render plus 182.C.33's labelled examples. Board tag 2026-09-18: the transport-fault half is demonstrated; the content-fault half needs takes that are wrong in known ways, which is the labelled corpus of 182.C.33. blocked:corpus
  • 182.C.40.06 Demonstrate consent expiry/revocation, tenant isolation, provider retention disclosure, subject deletion, takedown, and no public candidate URL before release. Verified 2026-09-12: FOUR OF THE SIX ARE DEMONSTRATED, two are not. Consent REVOCATION was driven live against Postgres today — a whole withdrawal answered 403 hv.consent.missing and a partial one answered 403 hv.consent.scope_exceeded while the sibling project still rendered (evidence/2026-09-12-jobs-adapter-live/). Tenant isolation is in the row key and was exercised in the same run. Provider retention disclosure is human-video-zdr-posture.ts. No public candidate URL holds by namespace: released is the only namespace the CDN can read, and a candidate URL is deliberately short-lived because one that outlives the review window is an unreleased take loose on the internet. SUBJECT DELETION and TAKEDOWN were the two that were not. CLOSED LATER THE SAME DAY in three passes, all live against Postgres and a versioned MinIO bucket (evidence/2026-09-12-erasure-bound-live/). The erasure port is bound, so thirteen of the fourteen ports are and only releases answers 503. DELETE /assets/:id destroyed one 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 failure a plain delete hides, measured at deleted 3 version(s) on a key an ordinary delete had already "removed". The withdrawal covered every take, the job's orphaned attempt, and the uploaded reference photograph and approved audio, which live in a second store with its own key scheme; a subject deletion that leaves the source photographs is not one. Expiry and tenant isolation were driven live too: a lapsed permission answers 403 hv.consent.expired with the window in the detail, and one tenant asking for another's job by its real id gets the same answer as for an id that does not exist. The sixth, no public candidate URL, is a namespace rule rather than a capture — canRead(ns, 'delivery') is true only for released, pinned over all six namespaces, and nothing on this deployment mints a delivery URL at all. TWO THINGS ARE NAMED RATHER THAN CLAIMED: complete: true is unreachable on any lane in this build because provider-side retains, which is the retention disclosure being honest rather than a gap; and face embeddings are not destroyed because nothing writes one, which a drift-guard test asserts with a control so the day something does, it fails and says what to do. A NEGATIVE CONTROL FOUND THE ONLY REAL 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.
  • 182.C.40.07 Meet documented quality targets with statistically adequate held-out results across active motion, locales, skin tones/lighting, profiles/occlusions, shot sizes, and supported speaker counts. Verified 2026-09-12: "statistically adequate" is already defined rather than left to judgement — samplesForZeroFailureClaim derives the count a zero-failure claim needs at 95% confidence, and cohortVerdict keeps not-promoted (too few samples) apart from blocked (measured and failed), so an under-powered cohort cannot be reported as a weaker pass. The held-out split refuses a subject or scenario that leaks. Everything is in place except the clips: 182.C.33. Board tag 2026-09-18: everything is in place except the clips of 182.C.33. blocked:corpus
  • 182.C.40.08 Meet documented latency, reliability, cost-per-releasable- second, accessibility, security, privacy, and provenance gates. Verified 2026-09-12: cost-per-releasable-second RETURNS NULL BY DESIGN today, and the reason is the item's own point. human-video-releasable-second-cost.ts names four terms: the attempt price is measured and is 2.5x what the catalog predicts ($0.25/s billed against $0.10 predicted on wan-2.7), candidates-per-shot is configured at four so a releasable second starts at 4x before any retry, evaluation is measured in CPU-seconds and refuses to be priced until somebody names a rate rather than treating CPU as costless, and THE RELEASE RATE — the divisor — is not measured because nothing has been released. A model whose divisor is unknown does not produce a range, it produces nothing, and the module returns null rather than a plausible figure. Latency and reliability need runs. Accessibility's automated half passes today (182.C.36.08). Gated on 182.C.40.01 and 182.C.40.04. Board tag 2026-09-18: cost per releasable second has no divisor until something is released, which is 182.C.40.01 and 182.C.40.04. blocked:upstream
  • 182.C.40.09 Run and archive all applicable targeted lint, typecheck, unit, integration, contract, build, Playwright/mobile, accessibility, performance, provider-canary, evaluator, and release-gate evidence. Run 2026-09-12 and archived at evidence/2026-09-12-release-evidence-run/. Twelve categories: 0 lint errors over all 504 files, 0 typecheck errors with all 504 in ONE program and a control that reported, 5442 + 215 + 128 unit tests, 40 contract tests plus two guards, a passing build, 11 + 5 Playwright with 0 axe violations, 46 performance, 50 release-gate, and the canary archived from 2026-09-11. RUNNING IT AS ONE PROGRAM FOUND THREE REAL LINT ERRORS THAT PER-FILE RUNS HAD MISSED — a per-item lint only ever saw the file that item touched — including let within = 0 where ZERO IS A REAL POSITION INSIDE A WORD, so a default surviving a missed branch would have read as "at the very start of it" rather than as a bug.
  • 182.C.40.10 Verify documentation, OpenAPI/SDKs, operator runbooks, capability matrix, dashboards, alerts, flags, rollback, and incident drills match deployed behavior. Verified 2026-09-12 against a RUNNING SERVER, archived at evidence/2026-09-12-live-surface-capture/. TWO DOCUMENTS OPENED WITH A CLAIM THAT WAS NO LONGER TRUE: api-guide.md said "Nothing here is mounted" and creator-workflow.md said "None of this is reachable today" — both true when written, both false since the surface was mounted on 2026-09-11, and a reader of either would have concluded the feature does not run. Both now say what runs, and the API guide carries FOUR RESPONSES CAPTURED FROM A LIVE SERVER where it previously carried none. The SDK matches the surface by a byte-identical drift guard, and all seven hand-written clients now match it too. The boot line's three "(no … table)" notes are the finding worth keeping: the tables exist and the ports are unbound, because the committed Prisma client predates the models and the runtime treats a stale client as no database on purpose.
  • 182.C.40.11 Leave unsupported modes visibly disabled or fail-loud; do not close the phase with a plan-only adapter, mocked generation, supplied metric, or provider marketing claim represented as shipped capability. Audited 2026-09-12 in docs/domains/isis/human-video/evidence/2026-09-12-fabrication-audit/. NO PORT IS BOUND TO SOMETHING THAT DOES NOT DO THE WORK: a deployment with no database binds exactly one thing — the preset catalogue, which has no store to be missing — and refuses eleven, and three ports bind only as ALL-OR-NOTHING groups because the half-bound states are the dangerous ones. The provider double has no importer outside its own spec, and the only two Math.random calls are annotated backoff jitter. No handler accepts a metric from a request body. And the three severity-H findings from capability-audit.md — the fabricated phoneme track, the fabricated VMAF, the unrunnable avatar engine — are closed: fails loud, deleted, and retired from capability claims with a drift-guarded status document. One soft spot is NAMED RATHER THAN FIXED: confidence ?? 0 is the worst default for a confidence, and the contract cannot express "not reported"; it is unexercised today and making it expressible is a contract decision.
  • 182.C.40.12 Check this final item only after all prior non-deferred items are complete, approved exceptions have owners/expiry, and origin/main contains the implementation and its evidence. Verified 2026-09-12 as the last step of a sequential pass over all 32 unchecked items. Every one was opened against its source and either fixed or given a measured reason; none became checkable, and the reasons now reduce to four: 182.C.33's CONSENTED CORPUS (17.08, 18.08 in part, 19.05, 19.08, 19.09, 20.01, 20.07, 33.01-05, 40.02, 40.05 in part, 40.07), a GPU HOST (32.03, 32.04, 32.06), the V1.2 RELEASE TRAIN (13.02, 31.07), and the two unbound ports plus a reachable object store (40.01, 40.04, 40.06, 40.08). The three that are neither: 23.03 waits on a TTS spend-or-install decision, 36.08 on a person with a screen reader, and 40.03 on whether its new benchmark flag counts as the flagged alternative. origin/main contains everything through this pass. Board tag 2026-09-18: the closing gate of the phase; it follows every item above. blocked:upstream

Delivery Milestones#

Milestone A — Honest Foundation#

Complete 182.C.01–182.C.10: audited reality, one canonical contract, authorized reference/audio ingestion, editable shot plans, and a live capability registry. No customer generation claim is enabled at this milestone.

Milestone B — First Exact-Audio Active-Scene Vertical Slice#

Complete the relevant parts of 182.C.11–182.C.21 for one promoted model: multiple references, approved audio, active performance, durable storage, actual sync/speech/identity/artifact evaluation, bounded retry, and fail-closed release. On OpenRouter only the Wan 2.7/2.6 audio passthrough can carry approved audio today, so the candidates for this slice are Wan 2.7 via OpenRouter and Seedance via the direct @isis/seedance-provider clients; the milestone records which one passed. This is the first milestone that may support an allowlisted internal demo.

Milestone C — Creator-Ready Product#

Complete 182.C.22–182.C.30: multi-shot continuity, Studio/API flows, consent/security/provenance, admin review, and production observability. Enable only benchmark-passing modes and cohorts.

Milestone D — SOTA Routing and Production Release#

Complete 182.C.31–182.C.40: converge V1/legacy systems, evaluate self-hosted options, build a representative corpus, run the model bakeoff, finish deep automation and load/security verification, canary safely, and publish evidence.

Critical Path#

  1. Canonical contracts and consent must precede provider submission.
  2. Durable object storage must replace base64 outputs before production-scale clips or real evaluation.
  3. Real decoded-frame/audio evaluators must exist before the product can claim measured lip sync.
  4. Exact speech, correct speaker, identity, anatomy, and safety gates must join sync in one non-compensable release decision.
  5. One active-scene vertical slice must pass held-out tests before building a broad provider menu.
  6. Studio and API release paths must consume the same release proof.
  7. Multi-speaker and long-form modes remain separately gated until their own cohorts pass.
  8. Model defaults are promoted from recurring evidence and are reversible.
  9. A RUNNER MUST EXIST BEFORE ANY OF 5 CAN BE DEMONSTRATED, AND ONE DOES NOT (traced 2026-09-12). A submitted job is written at validating and the only writer of a job row reachable from the running service is the cancel handler, so the single thing that can happen to a render request today is that somebody stops it. Every module the pipeline needs is written and specced — human-video-render-worker.ts, human-video-workflow.ts's stage executor, human-video-saga.ts, human-video-provider-double.ts — and nothing calls any of them outside a test. WHICH RUNNER IS AN OPEN DECISION AND IT IS NOT A WIRING DETAIL. human-video-workflow.ts says of itself that its queue is in-process and non-durable and is "not a replacement for the durable job envelope in jobs-route.ts", so the two candidates disagree about durability. And the stages between validating and planning are validation and audio preparation: a runner that jumped a job forward to produce a candidate would be recording that both had happened. The defensible reading — that submit-job already performs every check the validating stage names, so a job that survived submission has been validated — is a decision to take deliberately, not to assume while wiring. RESOLVED: the durable envelope is the runner (user decision 2026-09-12). human-video-job-runner.ts advances ONE stage per claim, the envelope persists the claim first so a crash loses at most one transition, and submit-job hands the recorded job over through a PORT rather than importing the queue. validating → planning is taken as a record of what the submission already did, and audio preparation is skipped because a named take is already the approved bytes the render is measured against. It stops at planning and throws with the binding that is missing, which the envelope files as needs_review — resolving there would mark a generation job complete for a render that never started. Measured live: submit, operator tick, validating → planning. What remained before submitted was a registered lane with a model record, an object root a provider can fetch a reference from, and a quarantine store. THE FIRST OF THE THREE IS DONE (2026-09-12). No registry existed anywhere because the one input every LaneRegistration needs — the catalogue — had no adapter until human-video-catalogue-reader.ts shipped the same day. CatalogueLaneRegistry assembles every route from today's catalogue and hands it to evaluateLaneEligibility; it does NOT decide eligibility itself, which is 182.C.10.04's whole rule — advertised support alone must not make a model eligible. Measured live against the real catalogue: 2 routes registered for exact dialogue, 7 for text-to-scene, 1 for the avatar lane, and NONE eligible — each blocked on policy_not_approved and region_unrecorded, with the exact-dialogue pair additionally denied accepts-supplied-audio as unprobed. So the runner's refusal has moved from "no lane registry is bound on this deployment" to "a registry answered, and it answered none", which is a fact about the REQUEST rather than about the deployment. Every unknown is a parameter with an honest empty default — no probes, no promotion, no region record, unreviewed policy — and the one that cannot say "unknown", health, answers zero consecutive failures with lastSuccessAt: null, because healthy in that engine means "not currently failing" and a lane nobody has called is not currently failing. The serving region is unstated unless configured, which matches no dated record, so a guessed region cannot serve a likeness into a jurisdiction nobody wrote down. AND THE SECOND OF THE THREE IS DONE. QuarantineStore had been declared since the download path was written and had no implementation anywhere — three specs held doubles and nothing else. human-video-quarantine-store.ts is a real one over a local object root: the writer appends each chunk and retains none, which is 182.C.16.02's bounded memory; the commit is a RENAME, so a half-written take is never observable under the name a reader uses; and the promotion asks planTransition rather than renaming what it was handed, so a stage skip, a digest change, a destination under another job or another tenant, and a leaf the lifecycle rules did not compute are all refused in the namespace module's own words. It is filesystem-only and says so: promoting in S3 is a copy-then-delete whose partial failure leaves the same bytes under two namespaces, and a bucket- backed deployment gets null with that sentence rather than an adapter pretending the two are the same. 17 tests against a real directory. SO WHAT REMAINS BEFORE submitted IS ONE THING: an object root a provider can fetch a reference from. 182.C.40.01 records it as the real blocker — placeReferences hands the provider a URL it must retrieve and the local store is on a private port — and it is the only one of the three that cannot be closed by writing code on this box.

Success Standard#

Phase 182 is complete when a creator can authorize a person and voice, provide multiple images and/or reference media, lock the exact words and pronunciation, direct a genuinely active performance, submit it through an eligible model, review candidates, and download only a final asset that has passed real audiovisual sync, exact-speech, speaker, identity, anatomy, motion, safety, encoding, provenance, and consent gates. The same flow must survive provider failure, job restart, cancellation, revocation, model drift, and final transcoding without misrepresenting partial or unmeasured capability as SOTA.