A model can be blind to a protected attribute and still produce disparate impact through correlated features. The actuarial answer is not to inspect the model but to cut its realized error stream by the protected class and test the outcome. This page runs three escalating tests on the same synthetic book the tool uses — the four-fifths / adverse-impact ratio, a significance test, and the actuarial contribution: a decomposition of the disparity into the part explained by a legitimate rating factor and the unexplained residual.
The protected-class labels here are a synthetic stand-in for any ECOA / Title VII basis (race, ethnicity, sex, age ≥ 40, …). The model never uses them; they only reveal how the already-drawn errors fall across groups. The math is identical whatever the real basis.
Adverse rate, disadvantaged vs advantaged group, by cohort.
Within each cohort, A/E ≈ 1.0 means the group is treated as validated. A residual A/E gap inside a cohort is disparity the cohort factor cannot explain.