Coding agents used to fix software rely on feedback; unit tests give only pass/fail signals that are often noisy or missing.
This paper teaches video-making AI models to say how sure they are about each tiny part of every frame they create.