Veo 3.1参考图结果门禁

Applicable scenario

Use after a Veo 3.1 Ingredients to Video candidate has been generated from role-mapped reference images. Python 3.9+ is required. The skill reads redacted JSON evidence and never opens or transfers the media file.

Inputs and outputs

Input one is the role contract. Input two is a result object with status, output_ref, file_sha256, file_bytes, fingerprint, and reference_observations. Each observation states the reference id, coverage, identity_match, and scene_match.

Installation and execution

1. 1. Keep the contract and result evidence in a private working directory.
2. Fill observations only after a person or approved media tool has inspected the candidate.
3. Run python scripts/audit_reference_result.py <contract.json> <result.json> --out-dir <output-dir>.
4. Treat pass_for_human_review as a queue state, not publication approval.

Rules

The script requires completed status, non-empty output reference, a 64-character SHA256, positive file bytes, a unique output fingerprint, exact reference coverage, and true identity and scene observations for every contract reference. It always emits release_ready=false.

Recovery

Preserve the blocked report, correct the evidence or regenerate through normal provider controls, and rerun. Do not mark a missing reference as covered or turn a visual uncertainty into a pass.

Permission and privacy

Use only redacted ids, hashes, and observations. Do not place faces, customer data, source paths, API keys, cookies, or unlicensed references in the evidence. Human rights and content-policy checks remain mandatory.

Acceptance

The valid fixture passes the structural gate but remains pending human review. The invalid fixture exits nonzero for missing coverage and a false identity observation. A buyer accepts the skill only when the report and non-release boundary are present.