Conditional Coverage Diagnostics for Conformal Prediction — reproduction

arXiv:2512.11779v1 · OpenReview vaApZm6MKM

The synthetic design is chosen so the true conditional coverage p(x) has a closed form, so every estimate is scored against ground truth rather than against another estimate. The oracle interval gives p(x)=1−α exactly and is the negative control each metric must return ≈0 on.

ClaimVerdictHeadline
1 ERT family + constant-predictor floorreproducedoracle L1-ERT +0.0012; recovers 90% of true L1
2 LightGBM 68.4% vs PartitionWise 38.3%reproduced68.04% / 37.82% from released rows
3 ERT converges, CovGap does notreproducedCovGap error grows with n; ERT separates 14× better
4 Over/under-coverage decompositionreproduced60× ratio on a known-conservative predictor
5 Divergent KL± in classificationreproducedresidual 8.9e−8; 8/8 cells divergent
6 Algorithm 1 cross-fittingreproducedin-sample fabricates 0.177 on a valid predictor