You can’t watch a candidate work with AI in an hour. You can only test whether their account of it holds up under follow-up. A four-level rubric that grades the answer, the questions that make a thin account run out of detail, and the two places the levels stop telling you anything.