Reproduction: Conditional Coverage Diagnostics for Conformal Prediction
arXiv:2512.11779v1 · OpenReview vaApZm6MKM
| # | claim | verdict | headline |
|---|---|---|---|
| 1 | The paper introduces a family of Excess Risk of the Target Coverag… | reproduced | oracle floor +0.00117; recovers 90% of true L1 |
| 2 | Using LightGBM as the underlying classifier for L1-ERT achieves 68… | reproduced | LightGBM 68.04% vs paper 68.4; PartitionWise 37.82% vs 38.3 |
| 3 | In synthetic experiments, group-based metrics like CovGap remain u… | reproduced | CovGap error 0.0550 > its own value; ERT separates 14× better |
| 4 | The metrics decompose conditional coverage error into asymmetric c… | reproduced | conservative ratio 60 vs aggressive 0.08 |
| 5 | Classification experiments report divergent KL_plus-ERT and KL_min… | reproduced | decomposition residual 9e-08; 8/8 cells divergent |
| 6 | Algorithm 1 estimates the ERT metrics from finite samples using k-… | reproduced | in-sample false-positive 0.1767 on an oracle predictor |
One page per claim. Every number is produced by the code in this Space.