Your render finished with no errors. That is not evidence it worked.

A corrupt encode reports the exact right frame count, resolution, frame rate and duration, because those numbers come from the container header. A loudness check passes on the same file, because it only reads the audio stream.

AI Video QA is a skill for your AI agent, plus ten scripts you can run yourself. It is the verification pass for video you generate from code: Remotion, Motion Canvas, Manim, a headless browser, After Effects scripting, or your own ffmpeg chain.

Seventeen defect classes, grouped by what actually finds them. Every one of them shipped at least once in a real production and survived a review that was looking for it.

What you get

Ten scripts. Decode verification with per-scene frame hashing, encoded-peak audio measurement, a mastering script, a frozen-frame gate, a dead-opening sweep, a safe-area gate, beat-aware frame extraction, and a browser-side overlap and overflow gate.

ffmpeg and bash only. No Python, no numpy, no paid API, no framework assumptions.