“At Anthropic, a small set of test cases, sometimes as few as dozens, can be sufficient to prove a flaw in a model and justify an intervention.”