Creative Release Workbench | Media QA Gate

Every AI video workflow has the same hole at the end: something ships that should not have. Wrong aspect for the slot. Silent audio on a platform that needs sound. A file that parses but does not decode. Nobody catches it because catching it is boring work.

The Creative Release Workbench is the boring work, automated. It is a local, fail-closed instrument that takes one media file and one delivery policy, then measures the actual bytes.

Media in, then ffprobe metadata, a full ffmpeg decode, and a SHA-256. Then the policy gates. Then a technical verdict. Then human review. Then a JSON receipt.

Three policies ship with it. Web loop: landscape, audio optional, 640x360 minimum, 30-second cap. Vertical social: 9:16, audio required, so a silent file fails loudly. Vertical web loop: 9:16, audio optional, for pages that supply their own sound.

Ten gates, every one computed from a measurement rather than assumed: media parses, full decode, codec and pixel format, orientation, aspect ratio, minimum dimensions, frame rate, duration, file size, audio presence. Unreadable media fails closed instead of passing by absence. An unknown policy identifier throws rather than silently substituting a default.

A technical pass releases nothing. It reads AWAITING_HUMAN_REVIEW until a named reviewer checks the three things software cannot check: visual continuity, provenance and rights, caption legibility. A technical fail can never be reviewed around, and stale or duplicate reviews are rejected.

Under the hood: Node 22, zero dependencies, no network calls, loopback only. Non-loopback Host headers get a 421. 23 tests pass, and they are real runs: actual ffprobe and ffmpeg over generated fixtures, strict full-decode corruption, API behaviour, and the security boundary including filename traversal, MIME mismatch, oversize bodies, cross-origin writes and concurrency limits.

Stated limits, because they matter more than the feature list. The technical checks say nothing about taste or brand fit. Human review here is a self-attestation, not authentication. Runs live in memory, and the downloaded receipt is the durable record.

Independent work sample by Supreme at GiveMePrompts, built for the release step of AI video production.