Two programs reproduce the 304 run-2 label differences in the tested ds000102 files

We reproduced a complete file audit of the OpenNeuro ds000102 flanker-task files and compared it with a separate program. We retrieved 325 public source files on 2026-10-09 UTC; every file matched the recorded size and SHA256 checksum. The BIDS files are pinned to snapshot 00001 and Git commit fdc1ce49d524fc1e77e5b763f10acb02f403c581. The legacy OpenfMRI ds102 URLs do not identify an acquisition version. Across 26 participants, 52 runs, and 1,248 trials, all 11,232 comparisons between nine BIDS fields and their specified legacy sources agreed. Both programs found no condition-label differences in run 1 and 304 of 624 differences in run 2 (48.7179487179%), affecting all participants. Both identified the same 90 all-zero model placeholder rows and kept all 52 actual onset-zero trials. The reference program reproduced all nine archived outputs byte for byte; the independent Fraction and SQLite program produced the same exact values in all nine. Its strict label check exits with code 1 because the labels disagree; the reproduction driver expects that result. These finite-file counts do not identify which arrows were presented. Original presentation records remain unavailable, and the assumed conversion of legacy response times from milliseconds to BIDS seconds is not established by the legacy header. One researcher wrote and ran the programs without independent review. No brain signal model or new participant measurements were used.

flanker task · condition labels · data audit · reproducibility · BIDS · OpenfMRI · trial records · exact arithmetic