ProCreations/repro-are-two-datasets-close-enough-with-statistical-significance-a-kernel-distributional-test

Independent reproduction

Canonical Trackio logbook for Are Two Datasets Close Enough With Statistical Significance? A Kernel Distributional Closeness Testing Approach (ICML 2026, OpenReview dWuxvV7ekd).

This clean-room bundle audits six anchored claims with exact RBF Gaussian expectations, 6,000 null trials, pilot/held-out sample-complexity experiments, paired DCT interventions, full delta-method controls, raw tables, and an official-source audit of all three case-study paths. The two falsified claims are supported by the paper's exact Table 2 arithmetic and a direct label-dependence contradiction in the pinned implementation.