A second calculation gives the same 35 sign differences in 910 ds000102 run-2 model pairs.

A separately written calculation tested whether different code reproduces the contrast-gain sign differences reported for ds000102 run-2 event models. It used event tables pinned to OpenNeuroDatasets/ds000102 Git commit 288d722ba88d93eb4e59070303598b081eee8323 and motion files from the OpenfMRI ds000102 R1.0.0 archive. The dataset release correspondence for these two sources is not established. Across 26 participants, seven motion settings, and five assumed noise-correlation inputs (rho=0, 0.2, 0.4, 0.6, 0.8), the second calculation reproduced all 35 run-2 contrast-gain sign differences in 910 model pairs. All 1,820 primary rows, 780 scan-index-reset comparison rows, 53,144 sample-mask rows, and event counts agreed with the archived results; the largest primary numerical difference was 2.3092638912203256e-14. Resetting scan indices after sample removal changed gain signs in six of 312 positive-rho run-2 pairs. Five participants had negative gains for all 35 motion and rho settings. The pairs reuse 26 participants and are not independent observations. The rho values are model inputs, not measured noise estimates. No brain signals were measured, and the calculation does not identify the correct condition labels or verify image timing. The result is a same-source computational reproduction, not independent biological verification.

generalized least squares · contrast gain · noise covariance · fMRI · replication · OpenNeuro · ds000102 · condition labels · framewise displacement