Protected attributes and fair-lending check

Measure bad rate and later model approval rate by sex, age band and foreign-worker status; decide which attributes may be used as features. Concepts: Fairness; protected attributes; proxies; regulation Project A: personal_status_sex mixes sex and marital status; is_foreign_worker is nationality-related; age is often allowed but must be justified. Your notebook keeps sex_merged as a feature (Cramér's V only 0.07, IV 0.03) → drop it from the model and keep it for the fairness audit. Project B: Not applicable (no people). Equivalent check: does error differ by region or time of day so that some customers are systematically under-served?