For 128 fixed sounds, scales from repeat differences gave higher mean matching AUC than variance scales in six SEEG patients and 24 fMRI adults

We tested whether site scales from repeat differences improve sound identity matching over total variance scales. Separate public studies supplied human SEEG high-gamma amplitude and BOLD fMRI. Calibration and test sounds were separate. Each query estimate was matched to templates from two other estimates. Matching AUC is the fraction of wrong templates ranked below the correct one, with half credit for a tie. A fixed half mixture of total variance and repeat-difference variance set the new scales. A local code lock preceded new scores in six SEEG patients, 16 B2021 fMRI adults and eight NH2015 fMRI adults. Mean matching AUC rose by 0.949, 1.197 and 1.018 percentage points. Pointwise 95% person Student limits were [0.084,1.813], [1.048,1.347] and [0.829,1.207]. All 30 person means increased. The fixed test required positive mean gains and AUC above 0.5 in all three groups. It passed, p=0.018515. A symmetry-based mean sign-flip check gave p=0.015625. Gains stayed positive in all 20 added sound splits in each cohort. The result applies to 128 fixed source nonvoice sounds and source-selected sites. The new rule lost to two simpler scales in SEEG, and SEEG top-choice accuracy did not improve. The small-site electrical control was uncertain. These discordant results remain. The method is a bounded test of an established normalization idea, not a new best readout or a causal neural model. Clinical sampling, source selection, shared task order, baseline treatment, fixed sounds, separate cohorts and unequal measurement filters limit scope. The package supplies code, all person and cell ranks, source pins and terms, local plans, figures, checks and repeat instructions. Original neural arrays stay at their public sources.

human auditory cortex · SEEG · fMRI · natural sounds · noise normalization · sound identity · cross-modal generalization · reproducibility