A housing waitlist model uses proxy features strongly correlated with protected classes. What should audit emphasize?

A housing waitlist model that uses proxy features strongly correlated with protected classes can recreate discriminatory patterns even without explicit sensitive labels. Audit should emphasize assessing feature appropriateness and bias risk before continuing use of proxy inputs in civic eligibility AI because proxies such as geography, network, or historical outcome features may encode protected-class differentials and undermine fair access to housing assistance. Removing protected attribute labels while retaining correlated proxies as neutral eligibility signals fails by concept: dropping a sensitive label does not stop correlated inputs from encoding protected-class differentials in civic eligibility scoring. Documenting proxy definitions in the model card and continuing deployment with current input selection fails by concept: merely recording what the proxies are does not assess whether they encode protected-class risk or stop their use before bias is evaluated. Prioritizing validation accuracy and calibration when deciding whether proxy features remain in production scoring fails by concept: strong accuracy and calibration do not show that proxies are free of disparate impact, and they cannot substitute for a bias assessment before continued use. Exam caveat: AAIA data-management/fairness items on civic eligibility AI require feature appropriateness and bias assessment for proxies—not dropping labels while keeping proxies as neutral signals, not merely documenting proxies and continuing deployment, and not prioritizing accuracy and calibration over bias review. Operational check: inventory proxy features, run bias and correlation analyses with fairness reviewers, document accept/mitigate/remove decisions, and constrain production use until remediation is approved.