Design a bias audit for a deployed loan-approval model

A lender's approval model has been in production for a year. A regulator-facing risk team asks you to design a bias audit. Describe what data you need (including the protected attribute problem), which metrics and intervals you compute, how you look for sources of bias, what interventions you would evaluate and how you would present the trade-off, the legal and operational constraints, and the cadence. Include what you would conclude from these preliminary numbers on a 10,000-applicant test set at one threshold: group A approval 0.615, group B 0.270; TPR 0.688 and 0.364; FPR 0.386 and 0.114; PPV 0.850 and 0.841.