Video Success Criteria (binary checks, written before you render)

The convention for EVERY video ad: before you produce it, write the concrete pass/fail checks THAT video must meet. vid-qa then tests the render against this list and returns pass/fail per check. The video is not handed to a human until every check passes (or a human explicitly waives one).

This is the video-specific peer of /success-criteria (which turns a plan into a machine-checkable checklist). The difference: a video's "done" lives in pixels and motion and sound over time, so the checks must be verifiable on frames, motion windows, the beat grid, OCR, and side-by-side against the reference, not just on a spec doc.

When it runs: at the storyboard + criteria gate (the first human gate in vid-ad / ai-ad-pipeline), right after the storyboard stills are approved and BEFORE any animation spend. The criteria file is an artifact the human approves alongside the storyboard. Re-checked after every re-roll.

The cardinal rule (the reason this skill exists)

Criteria come from the REFERENCE plus the core creative qualities, NOT from the client's literal feedback words. A checklist built from literal instructions can pass a version that is actually worse.

The failure this prevents (real, 2026-06-24, Gold Cash "Pass It On"): the criteria were written from the owner's literal words ("two hands, flat note, 1mg endcard, captions readable"). The render passed all of them. the owner rejected it as "getting worse and worse," because the criteria missed the two things that actually mattered:

1. Does the passing match the reference's dynamic energy? (the Apple-style hand-off / note in motion, not a stiff two-hand hold)
2. Does the note read light-as-paper? (floats and flutters, not a rigid card)