Six votes from eight shaft readers gave 5.04 percentage points more mean correct output than agreement among three groups in eight offline SEEG records

We tested two output gates on the same eight-shaft input budget. The public source was OpenNeuro ds004473 version 1.0.2. Eight reused clinical records supplied 1,332 fit trials and 674 later test trials. Each test trial supplied a rest window and a pre-response window. Six votes from eight separate shaft readers gave 32.49% mean correct output over all offered windows. Full agreement among three separate shaft groups gave 27.45%. The paired advantage for eight readers was 5.04 percentage points. Its descriptive 95% person bootstrap interval was 3.17 to 6.98 points. All eight records favored the eight-reader rule. The exact sign-flip value was 0.0078125. Mean wrong output was 8.38% and 6.24%, respectively. A proof shows that one arbitrary shaft feature change cannot make an accepted label differ from that rule's clean majority. The proof covers clean withheld inputs. It does not certify a true brain state. Wrong clean majorities can still give newly admitted wrong outputs. All 21,568 saved finite fault cases passed replay. Fresh source ranges rebuilt 16 trials exactly. The package preserves features, code, fits, outputs, figures, source records, and repeat instructions. This is an offline method test with known event times and future-dependent eligibility. It does not test live control or stimulation.

stereoelectroencephalography · brain interfaces · fault tolerance · voting gates · selective classification · offline evaluation