Audit: chr20/21 holdout holds in all 18 tasks; README H3K4me3 train count is 12,532 high
I audited revision 851f994 (18 tasks, 532,064 rows) and found no within-task train/test reuse: 0 of 38,822 test rows share an exact sequence, reverse complement, parsed locus or interval with their own task's train split.
The chromosome holdout holds. Test rows are chr20/21 only; train uses the other 21 chromosomes; every row's coordinates parsed. The README lists 30,000 H3K4me3 train rows; the pinned file has 17,468. That 12,532 difference is the entire gap between the README total (544,596) and the Hub listing (532,064). Test counts match everywhere.
One promoter_all test window is 0.98 identical over 299 aligned bases to a train window on another chromosome. Across tasks, 113 of 38,822 test rows reach identity 0.95 or better, all at different-chromosome loci. Inspected pairs share no locus; repeat homology is plausible, though no RepeatMasker/Dfam annotation was run.
Within splits, 2,589 rows are excess occurrences of duplicated sequences, with no duplicate group carrying conflicting labels. All 34 .fna/Parquet pairs are multiset-equal; H2AFZ ships Parquet only.
Method: pinned revision, both representations, edlib verification, mutation-recall probes, independent DuckDB recomputation of exact and coordinate counters (90/90).
Caveats: the assembly is not stated in the files, and near-tier candidate recall is 0.88-0.95 at a 0.05 substitution rate.
Thanks to InstaDeepAI for the benchmark and revision. The chr20/21 design is documented in the card and discussion #1; this audit tests that claim rather than discovering it.