Joint calibration reduced rest windows with any call across intact and known simulated single contact loss cases in eight offline SEEG tests

We tested rest calibration for known simulated contact loss. We used eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. Training used 998 early trials. Calibration used 326 separate rest windows. Testing used 674 later trials. Each patient model used eight bipolar pairs. The test covered intact input and 105 single contact losses. Missing features became their training mean. The contact identity was known, and other signals stayed unchanged. A joint limit used the largest calibration score across all allowed cases. The mean fraction of rest windows with any allowed call fell from 8.9% to 3.9%. Each patient had equal weight. Mean response sensitivity across lost contacts fell from 24.3% to 15.6%. Three patients had joint rest fractions above 5%. The uniform mean across lost contacts was 3.5% with the intact limit. Thus, the main outcome is a stress envelope, not a measured natural fault rate. A standard rank proof bounds one rest window across all allowed cases under exchangeability. Chronological task changes make this assumption uncertain. The later trials were used in earlier studies. This is an exploratory offline test. It does not establish a live BCI error limit. The package includes features, source versions, range hashes, code, model weights, all scores and counts, three figures, and reproduction instructions. Independent numerical checks and an exact repeat run passed.

stereo-EEG · brain-computer interfaces · contact loss · rest calibration · logistic regression · offline evaluation · signal processing