Four sound classes gave fMRI prediction gains but failed a shared test in 20 iEEG patients and 16 fMRI adults

We tested whether four sound classes improve response prediction after fixed acoustic models in separate public human studies. Development used five iEEG patients and four MRI adults. A local code lock preceded tests in 20 other patients and 16 other adults. Nested sound folds held out all repeats of each target identity. Mean class gains were -0.049% and +2.873% of constant-model error. Their 95% person intervals were [-0.988%,+0.891%] and [+2.398%,+3.348%]. Positive counts were 6/20 and 16/16. The fixed shared rule failed, p=1. Full-model mean skills were positive in each modality. Real classes exceeded 199 refitted random-label references, p=0.045 and 0.005, but the electrical result gave no net mean gain over acoustics. An audit found 21 changed annotations among the original 165 MRI sounds. A post-result check with newer labels gave a smaller positive MRI mean gain. Source replay rebuilt all 110,682 electrical window means exactly and verified patient and parcel metadata. Independent direct solves reproduced every reserved gain within 8.5e-16. The package preserves code, source pins and terms, plans, all outcomes and repeat instructions. The result is a failed cross-modal prediction claim, a positive MRI component and a source-label audit. Sparse clinical sampling, different cohorts and acoustic descriptors, source screens, fixed sounds and unequal measurement filters limit scope. No common semantic mechanism or causal modality effect is established.

human neuroscience · intracranial EEG · fMRI · auditory cortex · sound categories · acoustic models · nested cross-validation · negative result · data provenance