Error maps from other people improved forecasts in 14 fMRI adults but failed a shared test with 18 iEEG patients

We tested if sound error maps from other people improve forecasts of errors from a clipped acoustic response model. Separate public studies supplied signed iEEG LFP and BOLD fMRI. Six people per study supplied development and reference maps. A local code lock preceded tests in 18 other patients and 14 other adults. Nested sound folds excluded each target response from both response fitting and error calibration. Base forecasts controlled voice class, acoustic distance, feature norm and predicted response norm. The mean log-error forecast gain was -0.991% in iEEG and +37.057% in fMRI. Pointwise 95% person intervals were [-1.702%, -0.281%] and [32.426%, 41.687%]. Gains were positive in 3/18 and 14/14 people. The shared positive test failed, with person sign conjunction p=0.999344. Common sound controls, donor and sound sensitivity ranges, site-count checks and repeat estimates remain. Exact source extraction matched all retained electrical arrays. Independent fits and a clean primary replay checked the results. At three sampled sites per person, the fMRI mean gain fell to 0.837%, with an interval across zero. The useful forecast is limited to full selected BOLD panels with target calibration. Source selection, clinical sampling, unequal tissue and measurement filters prevent a causal modality claim. No cross-modal generalization or shared neural cause is established. The package preserves code, permitted data, source versions and terms, all results, figures, plans and repeat instructions.

human neuroscience · intracranial EEG · fMRI · auditory encoding · model error · cross participant transfer · negative results