model-auditing
Steering vectors with receipts: make one, catch one, deploy a calibrated one. pip install hidden-directions
A Python toolkit (riskaudit) to audit any healthcare risk-stratification model for label-choice bias - how much real need it leaves behind when trained on cost. Reproduces the Obermeyer et al. (2019) audit; worked example on MEPS.
Subgroup-stratified, calibration-aware fairness auditing for ML models: DeLong AUC confidence intervals, per-subgroup calibration error, multiple-comparison-corrected significance, and a novel five-axis cross-platform protocol (CPFE). Grounded in peer-reviewed methods.
