Plausibility

An optional hard constraint keeps counterfactuals inside the data manifold: an isolation forest, parsed through the same tree IR, bounds the anomaly score.

from sklearn.ensemble import IsolationForest
from treecf import Explainer, Plausibility

iso = IsolationForest(n_estimators=100).fit(X_train)
exp = Explainer(model, background=X_train,
                plausibility=Plausibility.isolation_forest(iso, max_anomaly_score=0.55))

The bound s(x') <= theta is equivalent to a single linear constraint on the sum of depth-adjusted path lengths, so it composes exactly with everything else. The genetic backend evaluates the same score directly.

Cost transparency: the forest's trees join cell construction and add one boolean per forest leaf — roughly doubling model size for a typical 100-tree forest. plausibility=None costs nothing.

v0.1 restriction: plausibility cannot be combined with AllowMissing or NaN factual values (isolation forests define no NaN routing).