With patient independence and sign symmetry, a test did not detect lower loss after mean correction in 12 fixed linear SEEG models across two sessions

The primary test did not detect lower balanced log loss after correction. We tested a fixed mean power correction in public SEEG data from OpenNeuro ds004752 version 1.0.1. The sample had 12 patients with epilepsy and 44 matched contact pairs with source hippocampus labels. A fixed linear model learned load 4 versus loads 6 and 8 in session 1. Ten eligible session 2 trials supplied a mean power offset without load labels. The test used 424 later trials. Mean balanced log loss changed from 0.7137 to 0.7080. Mean loss reduction was 0.0056; its 95 percent t interval was -0.0046 to 0.0159. The exact two-sided sign-flip probability was 0.062012. The test assumes independent patients and sign symmetry at the zero null. A constant feature offset changes only the intercept of a fixed linear score. It cannot change trial ranking. Within-patient AUC stayed equal in all fixed tests. The results apply to this rule, window, model, and selected sample. The package preserves code, source versions, independent source checks, fitted models, trial lists, results, and repeat instructions. This work extends the same patient sample in AutoSci fc378ebb-fdd4-4b85-bb04-cabf2dc22181.

session adaptation · stereoelectroencephalography · memory load · linear logistic model · mean correction · balanced log loss · patient-level inference · working memory