Check against the brief. The shot-brief is the spec (references/qa-checklist.md). The clip either does its story job or it doesn't. Everything else is secondary. Name failures from the taxonomy (references/failure-taxonomy.md) so each defect maps to its fix.
Pull frame evidence. Contact sheets at clip boundaries and key beats (references/frame-evidence.md, scripts/contact_sheet.py). For volume, run the embedding triage first (references/automated-identity-qa.md, scripts/identity_score.py). Humans review the failures plus a sample of passes. Eyes miss drift that stills catch.
Check the joins. Continuity across every cut (references/continuity.md): geography, light, wardrobe, screen direction.
Write the verdict. Ship, fix (with the one variable to change), or cut. No fourth option, no "maybe it'll work in the edit."

Rules that don't bend

The brief is the spec. "It's a nice clip" is not a pass. Does it do what the shot list row says? That's the only question.
Check the tails. AI clips degrade toward their ends. The last two seconds get the hardest look, not the first two.
One variable per fix. A failed clip goes back with exactly one slot changed. Two failures with two different changes teach you nothing.
Two strikes, then replan. If a clip fails twice, the problem is the shot, not the prompt: simplify it, shorten it, switch to image-to-video, or cut it.
Write it down. Clip ID, verdict, what was wrong, what changed. This log becomes your project's failure-mode list. The most valuable document you own by the end.

video-qa