Mean AUC lower bounds exceeded 0.5 for site errors capped at 5% of training response SD in 20 iEEG patients and 20 fMRI adults

We tested whether sound category transfer survives bounded errors at recording sites. Separate human studies supplied iEEG local field potentials and fMRI BOLD estimates. A model used response ranks across sites. It learned voices without speech and then ranked speech against animal sounds. A pairwise rank identity gave conservative AUC lower bounds for every allowed change to test responses. Each site error was capped at 5% of its training response standard deviation. Errors could differ across all sites and sounds. Four patients supplied development. The fixed reserved test used 20 other patients and 20 fMRI adults. Mean lower bounds were 0.62978 and 0.84088. Their 95% person Student intervals were 0.51688 to 0.74267 and 0.74848 to 0.93329. Both modality mean and sign tests passed; two-claim Holm p was 0.02647. The initial acoustic-error test failed its fixed speech-coverage rule. At a 10% cap, neither modality established a mean bound above 0.5. A score-equivalent common weight change can make separate rank intervals arbitrarily loose. The pair-difference bound avoids this problem. This is a method test on reused cohorts, not new acquisition replication. Different cohorts, sounds, clinical coverage, source selection and measurement filters limit scope. The result does not identify a shared neural cause or a non-acoustic code. The package gives a proof, code, permitted electrical inputs, source versions and terms, all outcomes, figures and repeat steps.

human auditory cortex · intracranial EEG · fMRI · bounded response error · rank decoding · sound category transfer · robustness certificate