BV-BRC genome_amr records for six antituberculosis drugs: a faulty Mycobacterium tuberculosis extract used in a provenance audit (2026-09-22), a complete API snapshot (2026-09-27), and an integrity check

Version 2. Version 1 archived the 2026-09-22 extract and a 2026-09-27 snapshot for isoniazid and rifampin. Version 2 documents an error in the 2026-09-22 extract, adds bedaquiline, clofazimine, delamanid and linezolid to both sets, and adds a script that reproduces the finding. The error. In every amr_*_full.csv file of the 2026-09-22 extract, genome_id was stored as a number, which drops trailing zeros (1733.9420 becomes 1733.942). The shortened identifier is usually that of a different, real genome, so about 10% of records were attached to the wrong genome. The isoniazid and rifampin files also lack about 6,800 laboratory records each, almost all citing PMID 30280646. This is an error in how the authors produced the extract; there is no evidence of any change to BV-BRC. After the same number conversion is applied to the 2026-09-27 identifiers, every record in the 2026-09-22 files, for all six drugs and both evidence types, is found in the 2026-09-27 snapshot. Consequences for the audit built on the extract. 3,114 of 31,664 isoniazid and rifampin laboratory records (9.8%) belong to another genome. Of 747 genome×drug groups carrying both a Resistant and a Susceptible laboratory record, 668 were created by records of another genome; the complete snapshot has 88, all isoniazid pairs at 0.1 (Resistant) and 0.4 (Susceptible). All 76 records classed as confirmed discrepancies, and all 754 laboratory records of the 399-genome 'orphan batch', belong to other genomes. Contents. frozen_extract_2026-09-22/: the extract as analysed, unchanged and including the error, for six drugs, with genome metadata. pre_reply_snapshot_2026-09-27/: every genome_amr record for the six drugs across all taxa, with all API fields including record identifiers and insertion/modification dates (row counts equal the API totals), genome metadata, pull scripts and manifests. The snapshot was taken on 2026-09-27, the day the authors sent factual-verification requests to the maintainers of BV-BRC, the CRyPTIC Consortium Dataset and the Hu et al. (2024) AMR benchmark. integrity_check_2026-09-27/: check_extract.py, its outputs, and the audit objects it traces. The 2026-09-22 files are kept only so that the audit and the check can be reproduced; do not use them as a copy of BV-BRC. SHA256SUMS.txt gives a SHA-256 checksum for every other file. Zenodo stores files without folders, so each file name carries its folder as a prefix (folder__file); restore the folders before running the scripts.