Algorithmic Fairness Auditor

A fairness audit of the ProPublica COMPAS recidivism dataset (6,172 cases, ProPublica's own analytic-sample filters). Pick two groups and a risk-score threshold to see selection rate, true positive rate, and false positive rate side by side — the same test as Legal-1's Gradio app, ported to static JS for a free demo.

Reproduces ProPublica's 2016 finding: at the default threshold, COMPAS's false positive rate for African-American defendants runs roughly double that for Caucasian defendants, despite similar overall accuracy. Race categories are COMPAS's own dataset labels, used only for this audit.

Group A

Group B

Flag as high-risk if decile score ≥ 5

Group A Group B

Part of AI Governance Case Files — a four-project portfolio series. Data: ProPublica compas-analysis. Metric definitions: fairlearn. This static page reproduces fairness_metrics.py's calculations in client-side JS on the same 6,172-row filtered dataset — see this repo's 02-fairness-auditor/ folder for the Python source and the full local Gradio app.