The model stayed home.
Its reasoning travels.

A 90-day care-utilization risk model (gradient-boosted, 0.898 holdout AUC) distilled onto fully synthetic data into an additive surrogate whose every prediction decomposes into exact per-feature contributions — no approximations, no server, and not one real record in this page.

De-identified academic demonstration. The competition dataset is under a no-sharing agreement: this page ships a surrogate trained on synthetic samples only (fidelity to the real model: Spearman 0.92, attribution agreement 98%). Nothing here is medical advice.