A margin guard certified audit sign stability under bounded additive error in five public SEEG sessions

We tested a sign guard for one audit pulse from a known stimulation site. We used five public SEEG sessions from OpenNeuro ds004457, version 1.0.2. The guard uses a metric fixed from calibration pulses. Each template projection must exceed a stated additive error bound in magnitude. The proof shows that accepted signs must agree with the error-free audit under that bound. All 13,526 accepted main window cases with simulated nonzero error met this condition. Directed errors at the planned radius changed signs in four selected audits. The guard rejected all four. A site rule based on calibration template length accepted 61.09% of cases at the planned metric radius of 8. The inherited consistency rule accepted 59.49%. The mean paired difference was 1.60 percentage points across five sessions. An exact descriptive sign-flip test gave p = 0.50. These data do not give clear evidence of a general selection benefit. The certificate does not prove physical wiring signs or correct site predictions. A valid device error bound remains necessary. Only 27 of 549 source targets followed all fit pulses and the selected audit. No live interface was tested. The package preserves the proof, code, voltage bins, source hashes, results, figures, and a checked raw retrieval recipe.

SEEG · audit sign stability · bounded additive error · brain-computer interface · intracranial EEG · robust signal processing