Blank correlation correction failed a shared face-house error test in six ECoG patients and six fMRI adults

We tested whether blank-period correlations improve later face-house prediction beyond independent-site scaling. Separate public studies supplied surface intracranial EEG and BOLD fMRI. One electrical patient supplied development. A local plan and code lock preceded tests in six other patients and all six MRI adults. The fixed reader used equal-window log power or single BOLD volumes, 50% diagonal shrinkage and training-only scalar calibration. Mean error gains over a diagonal reader were 23.246 [-2.737, +49.229] and 12.258 [+3.395, +21.122] points of constant-reference error. Brackets give 95% person Student limits. The shared rule also required gain over site permutations and skill over a constant in each modality. It failed, p=0.09375. All negative people and fixed controls remain. Descriptive checks compare task covariance and a correction without site identities. An exact permutation mean needs only the common spatial direction and its complement. One MRI label storage error was repaired before any MRI reader outcome; all numeric values and labels stayed exact. Code, source pins and terms, results, figures and repeat steps are supplied. Small cohorts, source region selection, short blanks, separate tasks and measurement filters limit scope. No common neural cause or broad clinical transfer is established.

human neuroscience · ECoG · fMRI · blank covariance · spatial whitening · negative results · reproducible benchmark