It is 2019, and a lender proudly reports that its credit model never sees race — the attribute was deleted from the training data, and compliance has signed off. The model sees zip code, shopping patterns, and phone metadata instead, and in a country whose geography was shaped by decades of redlining, those are race wearing a thin coat. Audit the model properly: measure disparate impact on outcomes rather than inputs, trace which proxies carry the protected information and how much, and then face the uncomfortable design question — whether fairness requires ignoring the attribute or explicitly using it to correct the proxies. Get it wrong and "we don't collect race" becomes the industry's alibi: discrimination laundered through correlated features, at scale, with a clean audit trail.
Wasserman's broad statistical synthesis and his career-long bridge-building between classical statistics and machine learning give him genuine general fluency with the kind of rigorous hypothesis testing this problem's audit requires, distinguishing a genuine absence of disparate impact from a merely superficial absence of a protected attribute in the input data. His minimax-rate instincts and general statistical rigor transfer reasonably well to designing a proper disparate-impact test. But he is a broad synthesizer rather than a specialist in algorithmic fairness or proxy-discrimination auditing specifically, so his score reflects strong general statistical competence rather than direct engagement with this problem's specific fairness literature. That leaves Larry Wasserman as a credible secondary consultant on this problem, useful for framing and adjacent technique but not the first name anyone would call to build the solution itself.
Breiman's 'two cultures' essay, warning that algorithmic models can produce accurate-seeming predictions through mechanisms the modeler does not fully understand or control, offers a genuinely relevant cautionary instinct for this problem's concern that a model 'laundering discrimination through correlated features' can look perfectly innocent from the algorithm-design side while still causing real harm. His random forests and variable-importance measures also offer a technical tool for detecting which features, including proxy variables, actually drive a model's predictions. But his own major technical contributions predate the specific algorithmic-fairness literature this problem centers on, so his relevance is a general cautionary and diagnostic instinct rather than direct engagement with fairness-auditing methodology. Leo Breiman would recognize the shape of this problem immediately from adjacent work, even without having personally published the specific technique it calls for.
Battle #157 · 8/10/2026, 11:41:04 AM · this result is deterministic: the same two personas on this problem always resolve the same way.