A correction with weight 0.5 reduced mean cortical connectivity prediction error by 6 percent between two movies in ten ds006192 people

We tested a fixed personal fMRI connectivity correction in ds006192 version 1.0.0. Ten people had processed signals from Bao and Your Friend the Rat. The primary test used 100 cortical regions and 4,950 Fisher z edges. The target reference was the movie mean from the other nine people. The prediction added 0.5 times the source movie deviation from its own reference. We averaged squared error across both directions and people. Mean error fell from 0.087926 to 0.082626, a 6.03 percent reduction. Six of ten people improved. The 95 percent person bootstrap gain interval was -0.000420 to 0.010949 and included zero. The true assignment ranked first among all 3,628,800 person matches. Its conditional p value was 1/10! under an exchangeable matching null. This test measures matching association, not positive prediction gain in a new cohort. Weight 1 increased error by 37.94 percent. Including all 156 regions increased error by 1.61 percent at weight 0.5. Both retained strong matching evidence. A gain formula shows why these results can coexist. An existing same-cohort report already established strong matching with another score. The new contribution is the fixed prediction-error test and its weight and region limits. The package preserves 215 public source files, hashes, code, proof, bootstrap draws, all planned results, and instructions. Independent calculations verified the scores and every exact tail count. Offline and fresh-source builds reproduced all 11 numeric files exactly. These same-session results do not establish neural specificity or performance for new people or visits.

functional connectivity · fMRI · personal prediction · cross-movie prediction · brain connectivity · prediction error · permutation test