Atlas forecasts lowered error for fixed sounds in SEEG and fMRI; an added SEEG baseline control gave no clear gain

We tested whether a small reference atlas improves response forecasts from 32 calibration sounds. Separate public studies supplied human SEEG high-gamma amplitude and BOLD fMRI. Four reference response axes and a 0.25 forecast multiplier predicted contrasts among 96 other sounds in each fixed 128-sound nonvocal library. Donor development selected the multiplier after an initial electrical forecast failed. A local code lock preceded new endpoints in six patients and 16 MRI adults. Mean error decreases were 2.681% and 6.898%, with 95% person Student limits [0.939,4.423]% and [5.960,7.836]%. Both gains and gains over shuffled atlas labels passed the fixed shared rule, p=0.015625. The unchanged MRI atlas gave a 5.285% decrease [4.258,6.312]% in eight adults from another acquisition of the same clips. All 30 person means were positive. Fixed controls limit the finding. Added 100 ms electrical baseline subtraction gave -0.048% [-0.341,0.245]%. First-24-site means were negative in all three studies. The unscaled electrical forecast also failed. Later probes found loss of gain after adding or shuffling the short baseline. Those probes are exploratory and change noise. Thus the result is a conditional measured forecast benefit, not evidence for a robust common neural code. Source selection, shared electrical order, sparse clinical coverage, fixed reference people, different cohorts and filters, and small-sample inference assumptions limit scope. Code, provenance, all negative results, figures and repeat instructions are preserved. Original neural arrays remain external under source terms.

human neuroscience · SEEG · fMRI · auditory cortex · calibration · shared response · prediction error · baseline sensitivity · negative controls · reproducibility