When is a different score threshold per protected group defensible in a deployed model?

## What the technique is The model is fitted once and left alone. It emits a score; the decision rule turns the score into an accept or reject by comparing it with a cutoff. A per-group threshold rule uses a different cutoff depending on the protected group the case belongs to - 0.52 for one group, 0.61 for another on a scholarship-award or student-support model. The reason it works so well is that the quantity being audited and the quantity being adjusted are the same quantity. If the gap you are repairing is stated in terms of selection rates or of error rates at the operating point, then moving the operating point per group can close it essentially exactly, and the optimal per-group cutoffs can be read off each group's error-rate curve directly. Group-specific cutoffs can be derived so that the groups meet at a common point on their respective curves; that shared point generally sits below where the stronger group alone could operate, which is where the accuracy cost comes from, and in some cases the exact match requires randomising decisions inside a band rather than a clean cutoff. ## Why it is nevertheless the hardest one to ship **1. It uses the protected attribute at decision time.** Pre- and in-processing consume the attribute during training and hand you a group-blind scorer. A per-group threshold cannot be group-blind - the group *is* an input to the decision. Two applicants with identical scores get different outcomes because of a protected characteristic. In domains covered by anti-discrimination law that is a materially different legal object from a model trained to be fair, and the analysis is jurisdiction- and domain-specific. The correct engineering behaviour is to bring the option to counsel with the numbers attached, not to decide it in a notebook and not to assert what is lawful. **2. The attribute has to exist, at decision time, and be right.** Protected attributes are frequently self-reported, optional, missing for a large share of records, or inferred. A rule that switches on a field which is blank a third of the time is not the rule you evaluated. Decide up front what happens to a case with no recorded group - almost always, apply the stricter or the default threshold and record that this happened - and measure how often it happens. **3. The two numbers are estimates.** Each cutoff is fitted on that group's held-out data. If the smaller group contributes a few hundred rows, the cutoff carries a wide interval, and a threshold tuned to the last decimal is tuned to sampling noise. Report the cutoffs with intervals, and prefer round, defensible numbers over precise ones. **4. It drifts.** Base rates and populations move. A pair of cutoffs that equalised something last quarter will not this quarter. Per-group thresholds are a standing commitment to re-measure and re-fit on a schedule, with an owner. ## The reject-option alternative Instead of deciding differently, decide *less*. Define a band around the decision boundary - the region where the model is least confident and therefore where its errors concentrate. Inside the band, route disadvantaged-group cases to a human reviewer rather than auto-rejecting them; outside the band, the model's decision stands for everyone. This keeps a human in the loop exactly where the model is weakest, and it is often easier to explain and to defend than a numeric cutoff difference, because the machine is not making a different decision - it is declining to make one. The cost is concrete and budgetable: band width times case volume equals extra reviews per week, at a known cost and turnaround per review. Widen the band and you close more of the gap and buy more review hours; that is the whole tuning problem, and it is a staffing conversation as much as a modelling one. ## What good sounds like in an interview Say that the technique is mechanically the most effective, name the decision-time use of the protected attribute as the reason it is the hardest to deploy, insist that the legal call belongs to counsel, raise attribute availability and quality, raise the noise in the cutoffs for small groups, and offer the two escape routes: move the repair upstream into training so the shipped rule is group-blind, or use a reject-option band that routes rather than decides. Candidates who present the dial as a quick fix, and candidates who dismiss it as illegal everywhere without qualification, both miss.