Four low-repeatability guides failed a shared prediction test in six SEEG patients and 16 fMRI adults

We tested whether four sites with low repeat correlation improve prediction beyond four sites with high repeat correlation. Separate public studies supplied human SEEG high-gamma activity and BOLD fMRI. Training sounds set site ranks, class means, scales, regression maps and correction weights. Other sounds tested predictions at other sites in different repeats. A local code lock preceded the reserved tests. Mean gains were 0.020% in six patients and 0.349% in 16 B2021 adults. The mean was 0.467% in eight NH2015 adults. Gains refer to target energy. Positive counts were 3/6, 15/16, 7/8. The fixed shared rule failed, p=0.863166. We retain all negative results and fixed controls. A comparison used the next four most repeatable guides. Its gains are descriptive. A Gaussian model shows how weak guides can lower error with one latent signal. Thus a positive gain would not count neural sources. We supply code, source pins and terms, plans, person scores, figures and repeat instructions. An independent solver reproduced losses for all 30 reserved people. Clinical sampling, source selection, separate cohorts, shared schedules and unequal measurement filters limit scope. The test does not establish absent weak-site information, a common neural cause, or positive cross-modal generalization.

auditory cortex · SEEG · fMRI · repeatability · response prediction · negative result