Five of seven local SEEG votes kept the clean majority under one shaft fault in eight recordings

We tested a SEEG reader with seven local voters in eight public recordings from OpenNeuro ds004473, version 1.0.2. Each voter used one pair on a separate electrode shaft. The gate required at least five matching votes. Across 1,348 later test windows, every output after one arbitrary shaft-vote replacement matched the clean ungated local majority. A complete proof and all 128 vote patterns confirmed this bound. The gate gave output on 55.3% of windows and correct output on 43.1%, with equal weight for each person. One fault could still expose wrong output from 249 of 583 clean withheld windows. Only 18.3% of windows kept correct output under every allowed fault. The guarantee requires isolated shaft processing and correct vote software. It does not prove that an event label is correct. The package preserves code, features, pinned source records, raw audit ranges, model states, fault cases, checks, and three figures. The scope is an event-aligned test in one session per person.

stereoelectroencephalography · brain-computer interfaces · fault tolerance · voting classifiers · selective prediction · robust classification