How Many Labels Does It Take to Verify Safe AI Triage? Distribution-Free Site-Acceptance Testing of Skin-Lesion Foundation-Model Ensembles under Dataset Shift

Preregistration of the confirmatory analysis of a methodological study. We measure how many target-site labels are needed to verify that an AI skin-lesion triage rule misses at most 5% of cancers, and how much referral remains after verification. The study uses five public dermatology datasets and frozen foundation-model ensembles. Dev-only analyses were run first and are disclosed below. The protocol is frozen as git tag protocol-v1-triage.