Every fairness metric you report rests on skin-tone labels assumed to be correct. On the field's standard benchmark, expert sources fully agree only 47.9% of the time. Check yours — using the same published method behind our audits.
| record_id | tone_a | tone_b (optional) | outcome | prob (optional) |
|---|---|---|---|---|
| img_0001 | 3 | 4 | benign | 0.12 |
| img_0002 | 6 | 5 | malignant | 0.81 |
prob column (your model's predicted probability of the positive class) and the tool also reports AUROC, sensitivity/specificity at the operating point, and calibration error per skin-tone band, with a bootstrap disparity interval. tone_a / tone_b are two independent skin-tone label sources (e.g., two annotators, or annotator vs. algorithm). With one source you still get representation and precision ceilings; with two you get the full reliability analysis.
Method: Radiant Revive Equity Audit framework — preprint DOI 10.21203/rs.3.rs-10670899/v1; preregistration OSF.IO/FMSNE. Bands: Fitzpatrick I–II / III–IV / V–VI, or Monk 1–3 / 4–7 / 8–10. Precision ceiling = 95% CI half-width for sensitivity at 0.85, given the positive cases in that band.
The full audit adds patient-level leakage checks, per-stratum representation against clinical prevalence, a written report with provenance you can hand to procurement or a regulator, and a findings call with the team that published the method.
Book a 30-Minute Scoping Call → See engagement tiersThis free tool is provided for informational purposes and is not an audit, certification, or regulatory determination. Results depend entirely on the labels you supply. Radiant Revive does not receive, store, or transmit your data through this page.