Proxy Discrimination: Algorithmic Redlining Lab (2D)

A lending model that never sees a protected attribute can still discriminate against it — through a "neutral" proxy feature, like a zip code, that quietly re-encodes group membership. This 2D simulator generates a synthetic population of loan applicants, trains a real logistic-regression classifier by gradient descent on only a proxy feature and a legitimate feature (the protected attribute is excluded from training entirely), and evaluates the trained model on a held-out test set. A live audit panel reports each group's measured approval rate, the legal four-fifths disparate-impact ratio, true test accuracy against a hidden ground-truth outcome, and the model's own learned weight on the proxy feature — showing concretely why deleting the protected attribute alone ("fairness through unawareness") does not make a model fair.