One fragment, one question

A reviewer never sees the whole run. They see a short fragment around the break and a single question with a few possible answers.

It should take less time to answer than to wait for the next match. That constraint shapes everything upstream: the model has to cut the fragment well, or the question cannot be answered quickly.

Before their answers count, reviewers work through cases where the answer is already known. Agreement with each other matters as much as agreement with the model.