The failed AI answer should be logged by recovery time

Answer-quality discussion: why a plausible wrong answer can cost more than a visibly bad one.

The AI answer that worries me most is not the silly one. The silly one is easy to reject. The expensive one is close enough that someone starts building on it. For team notes, I would rather log recovery time than star ratings. Did the wrong answer take two minutes to spot, or did it survive into a PR review? Did it require a test, a second model, or a teammate to unwind? That is the difference between “model was bad” and “this failure mode costs us work.”