A fixed class relation rule failed a shared gain test in 20 iEEG patients and 16 fMRI adults

We tested whether sound class relations from one neural measurement improve a correction from a few sounds in another. Separate public studies supplied signed human iEEG and BOLD fMRI. Four people per study supplied fixed donor matrices. A local plan and code lock preceded tests in 20 other patients and 16 other adults. Each target correction used four examples per class. Mean gains over the acoustic model were 2.735 [0.300, 5.170] and 4.821 [4.422, 5.219] points of mean-prediction error. Brackets give 95% person t limits. This shared test passed, Holm p=0.029679. Added gains over independent 0.25 class corrections were -0.333 [-1.451, 0.786] and 0.544 [0.485, 0.603] points. That shared claim failed, Holm p=0.729500. The MRI added gain remained positive across response estimates and in eight adults from another acquisition. All negative people and checks remain. An added exact calculation showed that averaging the 120 conjugate class operators gives the independent correction. Mean shuffled loss adds a nonnegative prediction-variance term. The electrical rule beat mean shuffled loss by 0.421 points while losing to the independent correction. Thus that shuffle comparison alone cannot establish useful class transfer. An unfitted scalar acoustic model shows a further interpretation limit. The first synthetic reference helper missed a divisor; we repaired it after confirmation started, with no change to the locked primary analysis. Clinical sampling, separate cohorts and sound panels, four donors, source selection and different filters limit scope. No added shared class benefit or common neural cause was established. Code, permitted inputs, source hashes and terms, sufficient loss statistics, all results, figures and repeat steps are supplied.

human neuroscience · iEEG · fMRI · auditory cortex · prediction · class relations · calibration · permutation controls