Model Interpretability

Each training row gets a weight equal to the count its (group, outcome) cell would have if group and label were independent, divided by the count actually observed. The weighted data carries no group-label association, and scoring never needs the protected column.

A generalized additive model fits one curve per feature and sums them, so a feature's whole effect is a single readable plot: non-linear in shape, never tangled with the other features.

Precision is the share of rows satisfying the rule that the model still labels the same way as the explained row; coverage is the share of the population that satisfies the rule at all. Higher precision usually buys narrower coverage.

Rank at the level of business reason codes, not raw features: map correlated inputs to one code and sum their contributions inside it. Add a deterministic tie-break so equal contributions produce the same order every run.

Partial dependence sets one feature to each grid value in every row, manufacturing combinations the data never contains, such as a tiny engine in a very heavy car. Accumulated local effects compare predictions only within each row's own neighbourhood.

Start from the harm. Extra document checks burden a legitimate seller, so equalizing the false-positive rate across groups is the commitment to state publicly. Prefer equal flag rates instead when the fraud labels are themselves untrustworthy.

Spend is a proxy for need, and the two come apart by group. A group with equal illness but less access spends less, so it scores as lower risk. The bias sits in the target definition, where no feature audit will find it.

Only model-facing ones: monitoring priority, data-collection spend, sanity checks against domain expectation. It measures one model's reliance on a column under one dataset and metric, not evidence that changing the quantity would change outcomes.

Treat the explanation as a product surface with its own bar: a faithfulness test the team runs, a measured stability budget across seeds and retrains, a stated purpose, and a fallback when checks fail.

The background set defines what the prediction is compared against, so it decides the whole explanation: attributions sum to the prediction minus the background's prediction, and a feature already sitting at its background value gets zero credit.