papperrollinggery AI Film Asset Stress Test
Authority: validation_only.
Use this Skill after DIRcreative has produced a candidate asset foundation and before the asset enters repeated shots, group compositions, or batch generation.
Ownership
- DIRcreative owns asset truth, state families, shot scope, approvals, and final project state.
- This Skill validates stability only. It does not redesign assets, generate
media, approve on behalf of a user/client, or grant
visual_assets_complete. - A producer-authored receipt cannot certify its own output. Missing real output
evidence remains
unverified.
Required packet
Read and validate the packet with:
- legacy v1:
docs/film-preproduction/schemas/ai-film-asset-stress-test.schema.json - split face/headless compatibility v2:
docs/film-preproduction/schemas/ai-film-asset-stress-test-v2.schema.json - unified character-sheet v3:
docs/film-preproduction/schemas/ai-film-asset-stress-test-v3.schema.json python3 scripts/ai_film_asset_stress_test.py validate <report> --artifact-root <root> --review-receipt <receipt> --review-signature <sig>
The packet binds one or more asset/state descriptors, exact references and rights, target model surface, intended shot classes/light, required co-cast/prop combinations, and a configurable per-asset test matrix.
certified and conditional require a detached signature over the exact review
receipt. The authority and public-key SHA-256 must already exist in the
host-owned registry selected by DIRCREATIVE_REVIEW_TRUST_REGISTRY or
~/.codex/dircreative/review-trust-registry.json; the generic
CLI cannot select a new key or pin. The package registry is an empty compatibility fallback
until a host administrator configures reviewer actor, authority kind, allowed
purpose/source, key path, and key hash. The validator fails closed if
OpenSSL, the signature, or the configured detached key is missing. A second
producer-written JSON is not a trust root.
Matrix policy
Choose cases from actual risk. High-risk characters normally cover face close-up,
full body, wide/FOV, target lighting, group composition, occlusion/prop
interaction, state variants, handedness and scale/topology. New recurring
clothed-human certification uses v3. Its headed default binds one physical master
sheet and follows the single owner contract in
skills/dircreative/references/character-master-sheet.md. This validator checks
schema, hashes, source/derivation roles, required cases, evidence and signed
review binding. It does not measure subject pixels or prove final exact-graphic
composite fidelity; those are external visual-review findings. This is a risk
profile, not a universal fixed attempt count.
Existing v1 and v2 reports remain verifiable under their original schemas but do not gain v3 unified-sheet guarantees. Migration requires a new physical master, view and derivation metadata, required v3 cases, fresh evidence and a new detached review signature. Never auto-promote an older report.
Each reference has one primary role. Planning boards cannot become canonical identity, topology, geography, or clean-frame truth.
Coverage is per asset, not a global union. Required co-cast/prop combinations must be named by passing cases, every passing case needs its own real evidence, and evidence IDs/output hashes cannot be reused across unrelated matrix cases. Canonical asset bytes/version/path are verified separately from supporting references, and a planning-only path/hash cannot be aliased under a canonical ID.
Verdicts
certified: independently reviewed evidence covers the complete intended scope with no blocked scope.conditional: evidence supports only the listed allowed scope and explicitly blocks the remainder.rejected: demonstrated deviation blocks the asset.unverified: evidence, matrix coverage, or independent review is missing.
Always return allowed/blocked shot scope, observed deviations, evidence hashes, review source, and the smallest next action. Never equate a stress-test verdict with installation, host adoption, media delivery, or customer approval.