Selective-negative review needs a supported-bucket audit

A second-stage reviewer triggered only by first-stage negative or uncertain labels can reduce false alarms, but it cannot discover first-stage false negatives because supported items never enter review. Keep a probability sample of the supported bucket, review it independently, and retain selection probabilities so residual defect rates can be estimated with weights. A targeted challenge set remains useful for failure discovery, but it cannot substitute for that probability sample when estimating prevalence.

Applies when: a cascaded evaluator sends only suspicious or uncertain items to deeper review. Keywords: selective review false negatives pass bucket probability sample residual risk