A shared 10-point loss from mean correction was not confirmed over all empirical class mixtures in six ECoG and five fMRI records

We tested whether an unlabeled mean correction is sensitive to class mix in human neural decoding. Separate public studies supplied surface ECoG and BOLD fMRI. A fixed reader used early responses. Late electrical tests also excluded fitting image identities; MRI tests used new runs. We held the late responses fixed and changed only the expected correction mean over all empirical class mixtures. Development used one person per modality. A local code lock preceded six reserved patients and five reserved MRI participants. Mean balanced-to-worst losses were 15.67 and 24.17 percentage points. Their 95% person Student intervals were [6.92,24.41] and [14.91,33.42] points. The fixed shared 10-point cost and above-chance worst-accuracy claim failed, p=0.109375. One electrical cost was 4 points. All MRI readers had sample AUC 1 despite their threshold losses. Restricting mixtures to 10-90% gave mean losses of 10.33 and 20.83 points. Electrical baseline subtraction reduced mean loss to 4.67 points. An exact scalar Gaussian example has AUC tending to 1 and full-range worst accuracy tending to 0.75. It is not fitted to human physiology. We provide an exact finite-library audit, every negative result, code, source pins and terms, figures, independent arithmetic checks and a fresh source replay. Original neural arrays stay external. Small fixed test sets, source selection, separate cohorts and unequal measurement filters limit scope. We did not change stimulus probabilities, prove an online guarantee, establish a shared neural cause or confirm the planned cross-modal cost.

human neuroscience · ECoG · fMRI · neural decoding · class mixture · mean correction · method audit