You Can Be Perfectly Calibrated and Useless

A green verification dashboard can hide a base-rate collapse that makes calibrated reviewers less useful every day. When AI floods a queue with cheap plausible outputs, prevalence falls, boundary cases concentrate, and human predicates drift from absolute standards toward best-of-batch comparison. The fix starts upstream: codify the predicate, preserve prevalence for expensive verifiers, and sample by predicted risk before lower escalations become the wrong success story.