Random list splits gave smaller mean high gamma Brier gains than later list tests in 18 SEEG sessions at penalty 0.1

We tested whether random list splits give larger recall prediction gains from study high gamma than tests on later lists. We used public OpenNeuro ds005059, release 1.0.6, with one SEEG session from each of 18 patients. We reused 2,334 trial features and labels from the cited benchmark. The new comparison used 200 random list splits per patient and matched training trial counts. At fixed L2 penalty 0.1, random list splits gave smaller mean neural Brier gains. The mean contrast was -0.015013 Brier units. Its 95 percent patient bootstrap interval was -0.027612 to -0.001960. Twelve of 18 patient contrasts were negative. The exact sign-flip p value was 0.040710 under independent-patient and symmetric-null assumptions. A stronger penalty reduced the contrast to -0.002667. A second seed gave a similar mean contrast. Separate score calculations, a separate optimizer, source range checks, and a full analysis repeat passed. The package contains code, features, source versions, split masks, primary predictions, results, and reproduction instructions. The two rules test different trials and different prediction uses. This reused cohort does not isolate a neural cause or test clinical benefit.

SEEG · memory recall · high gamma · Brier score · random splits · chronological evaluation · logistic regression · prediction evaluation