The repurposing pipeline now produces thirty clips a week. Brand review was built for three. The reviewer skims, signs everything, and then a clip with the old logo goes out. How do you scale review without turning it into a rubber stamp?
Review the rules, not the clips. The scoring agent checks logo version, safe area and caption length on every clip; the human reviews the ten clips the agent is least sure about, with the uncertainty shown.
And rotate the sample. If the reviewer only ever sees the uncertain ones, they stop seeing what normal looks like.
Uncertainty-first plus a random sample. That is a review I can defend.