validate-generated-media

Validate Generated Media

When to use

• Reviewing generated image or audio candidates before handoff or promotion.
• Verifying a generation manifest and staged file hashes.
• Comparing an exact asset or audio coverage plan against delivered outputs.

Workflow

1. Validate the generation manifest structure, explicit unavailable markers, staged paths and hashes. For direct Bubo LOCAL, retain the provider-neutral manifest plus exact native output/settings, installation, build and host-policy ancestry.
2. Compare delivered files against the planned matrix for exact names, states, variants, dimensions, durations, and destinations.
3. Review images in contact sheets at full, thumbnail, grayscale, and representative runtime scale.
4. Review audio in a listen-through sheet for duration, loudness, peaks, silence, loop boundaries, artifacts, and semantic fit.
5. Check rights, consent, provenance, tool and model revisions, workflow digest, parameters, seed and runtime budgets. For LOCAL, compare actual execution/settings and measured cleanup with the request; preserve failed/rejected attempts and confirm shared approval-state/candidate consumption has not reset across tasks or workspace changes. Use Bubo's catalog/API for model identities rather than duplicating pins here.
6. Record each candidate as accepted or rejected with reasons; promote only accepted, human-reviewed outputs in a separate action.

Use iterate-art-to-sanity for image defects within the already authorized allowance; record findings per required item rather than treating an aggregate score as acceptance. For sprites, tiles or layered art, use create-game-art-sheets content review and retain its current frame/cell and assembled-layout evidence. Header dimensions, an alpha-capable format or endpoint similarity do not establish transparency, semantic anchors or usable animation. Missing rights/provenance and exhausted budgets remain blockers regardless of visual quality.

Output

• A manifest validation result and hash report.
• Contact-sheet or listen-through evidence.
• Exact coverage, technical-budget, provenance, and promotion verdicts.

Guardrails

• Never invent missing reproducibility, rights, consent, or review evidence.
• Do not silently repair or replace rejected artifacts.
• Do not accept a visually or aurally plausible file when coverage, geometry, budget, or provenance fails.
• Do not promote staged outputs without an explicit human review status.
• Keep historical CLOUD/LOCAL manifests and receipts unchanged. New runtime or policy verification does not relabel earlier outputs as newly executed or authorize another candidate, private voice disclosure or promotion.