Radiant Revive AI Equity Audit: Methods Document, Version 1.1 (RR-STEM v1.1)

Versioned public methods document for the Radiant Revive AI Equity Audit, an independent statistical audit of how dermatology AI models perform across skin tones (engine release RR-STEM v1.1). It fixes in advance what the audit measures, how it measures it, what counts as a finding, and how findings are judged, so that nothing is chosen after a client's data has been opened. Contents: inputs and pre-processing (fingerprinting, validation, deduplication); two primary endpoints (sensitivity disparity between the lightest and darkest reportable Fitzpatrick bands, and an ordered trend in sensitivity across tone levels) with a Bonferroni family-wise policy; secondary endpoints; operating-point rule; interval methods (Wilson, Newcombe hybrid score, patient-resampled bootstrap, DeLong); patient-level clustering and design effects; equivalence testing (TOST); precision ceilings and the no-claim rule; severity thresholds; verdict rule; controlled reporting language; reproducibility and provenance; validation; change control. Version 1.1 (4 October 2026) supersedes version 1.0 (19 September 2026). It adds a small-cluster correction to the Wilson and Newcombe intervals when patients contribute repeated images (larger of bootstrap and ANOVA design effects, a k/(k-1) factor, and a Student t critical value with k-1 degrees of freedom), and uses the exact standard intervals when no repeated images are present. Point estimates are unchanged. An audit report is a statement of findings. It is not a certification, clearance or approval of any product.