Saved precision reproduced fixed-penalty ridge refits after 948 known simulated input losses in ten offline SEEG readers

We tested a repair rule for ten offline SEEG readers from OpenNeuro ds004703 version 1.1.0. After known simulated input loss, saved regularized precision gave the same survivor weights as a direct ridge refit, with the penalty, feature transform, and intercept fixed. No calibration rows or labels were needed at repair time. A proof and 948 loss cases support this result. The largest test prediction difference was 5.33e-15 in float64. Tests covered 129 single-contact losses, 778 two-contact losses, and 41 groups defined by source channel names; physical shaft geometry was not checked. Exact repair gave higher mean target error than mean replacement on later passages, so this selected cohort does not establish an accuracy gain. The single-contact benchmark and compact inputs were reused from cited work. The package preserves the new code, proof, figures, result tables, predictions, and saved reader states. Source recordings, audio, source metadata, and derived input caches are omitted because dataset metadata state CC0 while the source README states noncommercial and no-identification restrictions. The report preserves the source URLs, hashes, and uncertainty.

ridge regression · brain-computer interface · SEEG · fault tolerance · input loss · precision matrix · speech amplitude · reproducibility