Glossary

The words the documentation uses with a fixed meaning.

Factual. The instance being explained: one row, in the model's feature order, as the model saw it. Written x throughout.

Counterfactual. A full feature vector x_cf the model scores inside the target, differing from the factual only in the features the plan changes. Counterfactual is also the result type that carries it, with its cost, its verified score, and its proof.

Plan. The changes a counterfactual asks for — changes, a mapping from feature to (factual value, counterfactual value). "Plan" and "counterfactual" name the same object from two sides: the vector, and what to do to reach it.

Recourse. The plan read as advice: what the applicant can change to be accepted. A recourse region, a recourse menu, and a recourse map are all views of plans.

Lever. A feature a plan is allowed to change. The set of levers a plan actually changes is its lever set; two plans are different ways to reach the target when their lever sets differ.

Target. The interval the model's output must land in, on the raw score, the model probability, or a calibrated probability; a Target.bands ladder is several intervals.

Constraint. A rule every plan must satisfy — Freeze, Monotone, Range, linear inter-feature rules, one-hot consistency — compiled once and enforced by every backend.

Cell. For one feature, an interval between two adjacent split thresholds of the ensemble, inside which every tree routes identically. The search works over cells, not real numbers; a counterfactual value is the point of its cell nearest to the factual.

Category block. The categorical counterpart of a cell: a set of codes every tree routes the same way.

Backend. The search engine: "genetic" (the default, heuristic, on the Rust core), "python" (its numpy reference), or "exact" (branch-and-bound with proofs).

Proof. The claim a result makes: "heuristic" (no optimality claim), "optimal", "optimal_within_gap", or for an Infeasible result "certified" versus "search_exhausted".

Region. A per-feature box around a counterfactual, every point of which is verified in-target and constraint-feasible. Certified, not necessarily maximal, not monotone in the target; (data-limited) marks a side that stopped at the range of the background data.

Menu. Every lever set up to a size solved as its own coalition, with the minimal frontier and the sets certified unable to reach the target.

Certificate. A self-contained JSON record of one result — fingerprints of the model and constraints, the solve parameters, and a fresh verification — that check_certificate re-verifies later.