A distance gate rejected a mean of 99.96% of inputs with a simulated 1 mV, 10 ms contact pulse in eight SEEG recordings

We tested input rejection for fixed SEEG state decoders in public OpenNeuro ds004473 version 1.0.2. Eight recordings supplied 1332 model-fitting trials, 336 recorded calibration trials, and 330 later evaluation trials. Each gate used earlier data only. We added one simulated contact pulse to each later input and kept the fault location hidden from the gates. For the planned 1 mV, 10 ms pulse, confidence accepted 98.45% of inputs. A covariance distance gate accepted 0.038%, and rejected 99.96%. The gates produced wrong outputs on 42.79% and 0.013% of all pulse inputs, respectively. They produced correct outputs on 55.66% and 0.026%. The distance gate retained 95.96% of later recorded inputs. It did not recover useful outputs during the pulse. This is an exploratory software test on a reused epilepsy cohort. It does not establish natural fault detection or live BCI benefit. The package preserves inputs, source versions, predictions, code, three figures, and exact repeat checks. Fresh scalar source reads rebuilt all 30720 audited fault feature values exactly.

stereoelectroencephalography · brain-computer interfaces · input rejection · Mahalanobis distance · contact pulse · power features