Correlation tests using row counts rejected 33.43% of draws at a 5% target under independent phases with fixed spectra in 20 ds006192 movie runs

We audited correlation tests in 20 processed movie fMRI runs from ten people in OpenNeuro ds006192 version 1.0.0. We used 100 cortical regions. Our null model fixed each Fourier magnitude and drew independent uniform phases, with fair Nyquist signs for even lengths. The exact audit covered 99,000 pairs across runs. The mean run median variance ratio was 3.677 relative to independent Gaussian rows. The mean run median variance-equivalent count was 84.03, versus 241 to 362 stored rows. A nominal 5% Pearson t rule rejected 33.43% of 2,000,000 evaluation draws for 50 fixed pairs in each run. A phase rank rule rejected 5.00%. A Gaussian spectral rule rejected 4.78%. A proved spectral bound rejected 0.48%. The alternative pair matching and three fixed data checks gave similar results. The package includes an exact variance proof, a valid single-pair tail bound, a reusable TSV audit, 215 verified source files, all exact scores, and all primary draws. An independent lag calculation checked all 99,000 primary pairs. Fresh-data replay matched six files and 400 arrays byte for byte. These are conditional phase-model results. They do not count false neural edges, prove an effect of filtering, or estimate population uncertainty.

Fourier phases · correlation testing · functional MRI · temporal dependence · spectral variance · null calibration · OpenNeuro