A bank uses an ML model for loan approvals. Accuracy: 92%. But analysis shows: approval rate for Group A: 78%, Group B: 45%. The bank says 'the model doesn't use race as a feature.' (1) Explain how the model can be discriminatory without using race directly. (2) What proxy variables might cause this? (3) Is equalizing approval rates the right fix? What are the tradeoffs? (4) Design an audit procedure.