Keeping site labels improved fixed vocal class readers in six SEEG patients and 24 fMRI adults

We tested whether stable site labels improve a fixed reader of vocal sound class after measured response mean and length are removed. Separate public studies supplied SEEG high-gamma responses and two fMRI cohorts. Both readers used the same 16-site samples and mean-centroid rule. One kept site labels; the other sorted signed values. Each test excluded its two sound identities and its repeat from the class fit. Development used two patients and four MRI adults. Reserved evaluation used six other patients, 16 other B2021 adults, and eight NH2015 adults. Mean labelled-minus-sorted accuracy gains were 25.86, 20.52, and 20.10 percentage points. Their 95% person intervals were [18.81,32.90], [17.59,23.46], and [16.40,23.81]. The fixed shared interval rule passed at nominal p=0.000225513. Independent direct calculation reproduced every primary pair score. An exact model, added after human scoring, gave 100% labelled and 50% sorted accuracy from one varying spatial pattern without noise. Calibration changed gains below and above its reference mean into opposite vectors with the same sorted values. This is an interpretation limit, not a fitted human cause. Thus the result supports this reader comparison, not a source count, a necessary role for site identity in every reader, or a cause. Prior results on the same public cohorts helped select the question; the local lock is not an external preregistration. Clinical sampling, full-panel source selection, common sounds, acoustic differences, and unequal spatial and temporal scales limit scope. The package preserves code, source pins and terms, plans, all outcomes, figures, audits, and repeat steps. Original neural arrays remain external.

human neuroscience · SEEG · fMRI · vocal sound class · site identity · response distribution · centroid reader · fixed test · reproducibility