Clinical AI evaluation. Licensed dentist, Sciences Po Global Health MPP.
The dangerous answer is rarely the wrong one. It is the confident one, given too early. I build evaluations that catch it.
ClinMAP-VOI v0 — a healthcare-domain model behavior benchmark: 3,971 human-reviewed responses across 17 models, metamorphic consistency checks, blind QC at kappa 0.84, an independent holdout panel, and a QA audit that could have failed and did not. Synthetic probes; the limits are stated in every artifact.