Equal group weights gave no clear mean AUC gain in 19 offline SEEG tests with eight pairs and penalty 10

We tested equal fit weights for four direction and choice groups in public MetaRDK SEEG data. Each reader used eight bipolar pairs and fixed slope penalty 10. The primary window was 50 to 450 ms after movement onset. The primary score averaged choice AUC on correct and error trials. Nineteen people supplied 1,677 later test trials, including 459 task errors. The mean score was 0.569 with equal choice-label weights and 0.556 with equal four-group weights. The paired change was -0.012, with descriptive 95% person bootstrap interval [-0.041, +0.017]. The exact sign-flip p value was 0.432 under independent-person and sign-symmetric null assumptions. We also proved that any fixed score based only on binary direction gives primary score 0.5. Independent fits and AUC checks passed. Nine numerical files replayed exactly. Fresh source metadata matched, and a complete sub-04 source rebuild matched all seven arrays. The package includes code, compact data, frozen models, predictions, three figures, and retrieval recipes. This exploratory cohort reuse does not establish causal timing, a live interface, or stimulation effects.

SEEG · brain-computer interface · choice decoding · AUC · class weighting · MetaRDK