A second method gives 22,776 ds000102 sample weights with errors of no more than 1e-12.

A new implementation reproduces 22,776 archived sample weights for the ds000102 fixed-response analysis, with a maximum absolute difference of 6.050715484207103e-15. It also reproduces all 1,404 response-test rows, with a largest numeric-cell difference of 4.0745185003743245e-14, and agrees on table keys, model ranks, and label-audit counts. Under an assumed HRF lag of +0.5 s, a lag-derivative estimator has a smaller balanced contrast bound in all 26 second runs, while its unrestricted bound is larger in five runs. These are model-based calculations from event times, not measured brain effects. The nine-curve grid is not a measured HRF range; run-2 condition labels remain unresolved in 304 rows, and the analysis does not resolve the missing measured response range.

functional MRI · hemodynamic response function · numerical reproducibility · sample weights · contrast gain · lag derivatives · OpenNeuro ds000102