Mean sound class scores from repeat spread were above chance in six SEEG patients and 16 fMRI adults; a fixed shared test failed

We tested whether the spatial pattern of spread across three repeat measurements can rank speech/song sounds above source nonvoice sounds. Each test removed overall spread and fitted on other sounds. Separate public studies supplied human SEEG high-gamma activity and BOLD fMRI. Development used two patients and four adults. A local code lock preceded tests in six other patients and 16 other adults. Mean AUCs were .632 and .695, with 95% person Student intervals [.526,.738] and [.647,.742]. Common label-reference p values were .001 in both. But only five of six SEEG patients scored above chance. The fixed shared mean, sign and label rule failed, p=.109375. Eight adults in a second fMRI study had mean .543 [.434,.652], with label p=.060. Quadratic presentation-order removal also left an uncertain SEEG interval. All failures remain. An exact control gave AUC 1.0 from fixed mean maps and one repeat gain, with no added intrinsic noise. Thus normalized spread decoding alone does not identify a separate variability code. The contribution is a class-prediction benchmark, failed confirmation and a specific model limit. Separate cohorts, source screens, shared sounds, task schedules, unequal sampling and different repeat precision limit scope. No positive cross-modal generalization or common cause is established. The package preserves code, source pins and terms, plans, all outcomes, figures, numerical audits and repeat instructions. Original human neural arrays remain external.

SEEG · fMRI · repeat spread · sound class · neural variability · gain · negative result · reproducibility