Covariance from remaining inputs increased mean SEEG accuracy for eight sites from 27.1% to 34.7% after known simulated shaft loss in three offline sessions at 100 to 500 ms

We tested three readers after known simulated loss of one recording shaft. We used public SEEG data from OpenNeuro ds004457 version 1.0.2. The readers shared four fit pulses per candidate site, eight selected sites, and four test pulses per selected site. The primary window was 100 to 500 ms. Three of five records retained inputs after every single-shaft loss. In these records, covariance from remaining inputs gave 34.72% equal-person mean loss accuracy, compared with 27.08% for mean filling. All three records improved. The mean gain was 7.64 percentage points. A precision-block reader gave 35.42%, so the covariance reader was not empirically best. The sign-flip reference was 0.25 for three people. Two other records lost every input and gave only 12.5% guess accuracy. A Gaussian identity explains how mean filling adds a class-dependent penalty for unobserved values. The proof assumes shared covariance and a label-independent loss mask. We checked 110 fresh source files, rebuilt all binned voltage fields, checked all 9,792 predicted labels, and replayed 27 numerical files exactly. The package preserves code, data, source hashes, model matrices, figures, and instructions. These reused clinical records support an offline method comparison. They do not establish a population benefit, physical fault tolerance, or a working brain-computer interface.

stereoelectroencephalography · missing data · Gaussian covariance · site classification · shaft loss · brain-computer interface