With fixed references, model error profiles repeated in sound fMRI and late image MEG; iEEG transfer was not confirmed

We tested whether model error size repeats across people in separate human studies. Models predicted held-out sensory conditions. Class, fold and input-distance terms were removed from log error profiles. Fixed references used four people per study. Reserved sound fMRI in 12 adults gave r* 0.6649 [0.5961, 0.7240]. Reserved image MEG at 140-160 ms in 11 adults gave r* 0.1150 [0.0427, 0.1862]. Here r* is tanh of the mean Fisher score; brackets give 95% person Student limits. The shared fMRI-MEG rule passed two-claim Holm correction, p=0.0364. The original iEEG-fMRI claim failed. Sixteen iEEG patients gave r* -0.0176 [-0.0648, 0.0296]. MEG ranges crossed zero when reference people or conditions were resampled. Thus the positive result is conditional on the fixed references and panels. It does not confirm transfer to arbitrary reference groups or new condition samples. A cross-estimate fMRI check stayed positive. Exact noise and coefficient-weight examples can give repeated errors without a common missing neural computation. Cohort differences, tasks, feature sets, sampling, source screens and measurement filters limit scope. No causal or shared-generator claim is made. The package gives code, permitted inputs, source pins and terms, all negative results, figures, plans, error statistics and repeat instructions.

human neuroscience · model error · cross-modal generalization · iEEG · fMRI · MEG · auditory cortex · visual responses · reproducibility