ICML 2026 · CPU reproduction · Algorithm 1

Cross-fitting removes a false conditional-coverage violation

Five seeds, three scales, four fold counts, an exact partition audit, and a no-cross-validation control make the anti-overfitting mechanism directly visible.

270recorded summary fields independently recomputed

12 / 12cross-fitted size/fold cells below |0.01|

0.09735largest no-CV inflation, observed at n=10,000

3 / 3deliberately broken checker controls rejected

Partitions match exactly Five recorded test folds equal an independent seeded KFold split, are disjoint, cover every row, and never score a training row. Claim 6 · partition audit ↗ The negative control fails Fitting and scoring the same rows creates 0.0407–0.0974 spurious L1-ERT even though the oracle truth is exactly zero. Claim 6 · no-CV control ↗ Cross-fitting repairs it Every k in {2,3,5,10} stays below |0.01| at all three scales and improves by more than fivefold. Claim 6 · five-seed result ↗

OpenReview vaApZm6MKM · seeds 0–4 · n=2,000/10,000/50,000 · 161.9 s full CPU stage · no GPU · Claims 2 and 5 remain honestly blocked by >2h CPU · poster workflow: Chenruishuo/posterly.