Before a model decides anything, test it on a diverse cohort. Compare performance across groups. Reject anything with >5% disparity.