Prehistoric Animal Onboarding

Turn a licensed source into a small, hash-bound review packet. The agent should find defects before the owner sees the asset; the owner should decide tradeoffs, taste and release boundaries, not manually rediscover every engineering error.

Authority boundaries

Keep these states separate:

1. Machine pass proves rights fields, file structure, budgets, clip timing, hashes and evidence completeness.
2. Agent visual pass proves the contract's anatomy, deformation, full-loop, presentation and child-comfort checks against the exact GLB.
3. Owner decision may accept L3 investment, lock a model for finishing, and later approve public distribution. Never infer or auto-record it.

A model lock authorizes derivative and narration work for one GLB SHA. It does not authorize publication. Production approval remains a later separate record.

Load only what the task needs

Always read references/workflow-contract.md before running commands.

Also read references/failure-regressions.md when repairing an existing asset, handling L2/L3 work, defining a review contract, or explaining a budget/Idle deviation. Reuse its regression checks rather than relying on memory.

The tracked tool and this Skill are the portable source of truth. Do not depend on ignored docs/specification/ files or another worktree's local scripts.

Stage order and locks

Run from the prehistoric-animal-museum repository root. Every animal gets its own candidate directory and .handoff/animal-onboarding-runs/<animal-id>/. Create both from the tracked source model with run init; its deliberately blocking placeholders must be replaced with verified source, science, model and presentation facts before their respective gates can pass.

1. Environment lock. Run doctor, baseline verification and golden regression. On a fresh worktree, capture the local golden record only after production baseline verification passes, then regress it. Stop on a doctor or baseline failure.
2. Requirement lock. Resolve all four source-record.json blockers, then write and validate review-contract.json before model edits. Add real task-specific issues, invariants and evidence beyond the starter template. Name target defects, unchanged invariants, Given/When/Then interaction semantics and the exact evidence that closes each item. Every asset-inspection.json known issue must bind one exact contract subject and its complete evidence set.
3. Source/risk lock. Record asset-inspection.json; route it to L0-L3. Rights or identity blockers stop work. L3 and owner-requested parallel work may proceed only with explicit acceptance, an isolated animal workspace and an independent stage lock. An explicit owner instruction to advance named L3 animals in parallel already accepts that named L3 investment; record it without asking the same question again. A generic preference for concurrency does not.
4. Model lock candidate. Inspect the source rig and animation, then select the processing strategy explicitly. The shared Blender normalizer implements only replace-with-project-morph; it may discard a source rig or animation only when that destructive replacement is recorded. Use a dedicated L3 operation for preserve-source-rig-retime or custom-rebuild. Retain structure/rig/log/landmark evidence and run model-only QA. The model input must be the exact rights-verified source path; derived revisions belong in the output path and hash-bound processing log. Profile, normalization log, risk route and accepted L3 operations must name the same processing strategy. Use the smallest bounded operation consistent with the contract. Complete risk-evidence-manifest.json for every route-required proof before owner review. Do not create backgrounds, posters, narration or editorial packages yet.
5. Runtime evidence lock. Create a review-efficient capture plan and execute it through the already-open headed Browser or Chrome control Skill. Load that Skill at this stage, start the documented local review server, and execute every emitted request in order. Never launch Playwright, Puppeteer, Chrome, Chromium or another headless browser. The collector must record a name, Codex task identity and the exact collector attestation. The local validator proves internal consistency, state effects and hashes; because Browser/Chrome control does not provide a signed receipt, it must describe capture origin as collector-attested, never cryptographically verified. Ingest the evidence so every screenshot is bound to the final and actually loaded GLB SHA, viewport, observed camera, state, exact paused animation time and file digest. Set the base camera before applying interaction actions; a closest-state capture must show a smaller observed distance and different pixels than its paired initial state. In review-efficient mode, every primary and auxiliary angle covers 0/2/4/6/8 seconds; the validator proves each contract perspective × time combination rather than counting files. A transparency full-cycle contract requires at least two angles and no sampling gap above 0.25 seconds. The production golden baseline is mandatory and hash-bound in the persisted plan; it cannot be omitted after planning.
6. Agent visual lock. Generate the agent-review template, inspect the actual evidence and fill every category with a concrete finding and evidence path. For animated animals inspect 0/2/4/6/8 seconds at every declared primary and auxiliary angle. Static stills cannot prove a full loop, and PNG sequences cannot be labelled as continuous video. For L3, use an agent/task other than the model author and the capture collector. Record both reviewer name and reviewer task identity; the machine rejects reuse of the collector task.
7. Owner model lock. Generate the one-page owner packet only after machine and agent checks pass. Owner preparation reruns strict model-only QA into owner-model-qa.json and shows measured budgets, warnings, L3 acceptance, route-evidence completion, capture-provenance limits and key agent findings. Record model-lock.json only from an explicit owner decision. The lock binds that fresh QA decision plus normalization log, normalized Blend, landmarks, validator, risk manifest and capture/review evidence. Any bound change reopens the model stage.
8. Finishing lock. After the model lock, generate landscape/portrait backgrounds, poster, portrait poster, thumbnail, localized editorial records and independent zh-CN/en Qwen3-TTS Serena narration. Run complete QA and prepare the promotion manifest.
9. Publication lock. Record final approval only after the owner accepts science, visual quality, motion, both complete narration listening reviews and public distribution. Dry-run the whole batch, promote atomically, then verify baseline, installed hashes and original collection order.

Do not advance downstream artifacts while the preceding lock is missing or stale. A failure in one parallel animal never unlocks or invalidates another.

Model contract

Use the target, normal review and absolute exception budgets from ANIMAL_AUTHORING_GUIDE.md; the executable values live in tools/animal-onboarding/src/budget-policy.ts. A profile may tighten a limit. Raising a normal ceiling requires a metric-scoped budgetException with reason, risk owner and date, and may never exceed the absolute ceiling.

Every animated runtime has exactly one self-contained Idle lasting eight seconds at 24 fps, continuous at 0/8 seconds and without root drift. Retiming a 7.5-second clip is mandatory; it is not a budget exception. Clip metadata alone does not prove visible or natural motion.

For mouth motion, use source-rig only with a verified weighted jaw/tongue chain, or curated-components with an exact per-model hinge and bounded vertex selection. Otherwise record disabled. Never move a guessed broad head region.

Repair rule

On failure, repair the smallest deterministic region and rerun that stage plus all downstream locks whose hashes changed. Do not weaken shared gates to make a candidate pass. Convert recurring visual failures into contract invariants and evidence requirements using failure-regressions.md.

If a repair needs new rigging, full rebind, anatomy reconstruction, complex transparency work or mouth reconstruction, route it as L3. Owner acceptance allows the investment; it does not make the result acceptable.

Handoff

Give the owner one compact packet per animal with:

• risk level, route and any explicitly accepted exception;
• exact model SHA and machine gate result;
• agent findings with full-loop and multi-view evidence links;
• unresolved blockers only;
• the precise decision requested: accept model for finishing, or approve final public promotion.

Say “machine pass”, “agent visual pass”, “model locked for finishing”, or “owner-approved for promotion” precisely. Never collapse them into “approved”.