- On a fixed synthetic tenant-screening fixture, an audit found a 0.694 minimum disparate-impact ratio (DIR). A bounded search raised it to 0.875 at a measured 0.36% AUC cost. The audit was only the starting line. 🧵
- A red metric documents a result. It does not show which lower-disparity alternative preserves useful model performance, how it was chosen, or what evidence should follow the decision.
- We set a reproducible rule first: candidates must clear the demo's configured four-fifths policy gate and remain within a 0.03 AUC budget. That 0.80 gate is a demo policy threshold, not a housing-law conclusion.
- The Least Discriminatory Alternative (LDA) search evaluates 480 explicit, facially neutral configurations at equal selectivity. This is a bounded grid, not a claim that every possible model was searched.
- The grid varies optional feature subsets, four credit-score caps, whether guaranteed voucher income is credited, and six regularization values. Every candidate is measured against the same synthetic fixture.
- Of those 480 configurations, 240 clear the configured gate within the accuracy budget. The selection rule then takes the qualifying candidate with the smallest measured AUC loss.
- The selected candidate credits guaranteed income, caps credit-score influence at 640, and uses L2 10.0. Minimum DIR moves from 0.694 to 0.875 while AUC moves from 0.7823 to 0.7788.
- The output is an accuracy/fairness Pareto frontier plus the evaluated configurations and selection rule. A model-risk team can inspect why this candidate won, replay the search, and retain the tradeoff evidence for review.
- For housing model-risk teams: should a bounded alternative search become a release prerequisite, or stay an escalation only after an audit misses its policy gate? #ModelRisk #FairHousing #ResponsibleAI #AICompliance
- We built Equora to make that search inspectable without hiding the synthetic scope or measured cost. See the full breakdown: https://veriprajna.com/demos/housing-ai-compliance
Published on X · September 21, 2026
On social media
See this post on its original platform
In our archive