FAIR//AUDIT

Responsible-AI audit of four income classifiers on UCI Adult · records · every number comes from the executed notebook

NOTE The numbers reported here are produced by this code on the public UCI Adult data and may differ from the values reported in the accompanying paper, due to differences in data splits, preprocessing, hyperparameters, thresholds and random seeds.

01 Test-set performance thresholds frozen on validation

02 Fairness gaps by sex 95% stratified bootstrap CI · resamples

Absolute gap between men (Group A) and women (Group B). DPD = selection-rate gap · EOD = TPR gap · FPRD = false-positive-rate gap.

03 Sex × race selection rate XGBoost

Hatched cells = exploratory (n < 50 or < 10 positives). EO is left blank where a cell has no positives.

04 Mitigation Pareto front validation

05 Explanation audit frozen XGBoost · consistency, stability, fidelity reported separately

06 Governance crosswalk Jaccard · interpretive coding

Coded by the analyst to express the described coverage; these values follow from the coding, not from data.

07 What-if predictor exported Logistic Regression

09 Repeated runs 10 seeds · mean ± std (box) · min–max (line) · test set

08 Paper vs reproduced neutral comparison