Residual normalization kept fixed phase gates unchanged under simulated constant positive gain changes in 43 public SEEG recordings

We tested fixed confidence gates for nominal 50 millisecond phase forecasts in public SEEG data. The new score was fitted sinusoid amplitude divided by the weighted fit residual. It kept every gate decision unchanged when later voltage had a simulated constant positive gain of 0.5 or 2. This held in 43 recordings, with four selected pairs per recording, at fixed early retention levels of 25%, 50%, and 75%. A proof gives the same property for any positive constant gain in exact arithmetic. In 36 clinical recordings at 50% early retention, accepted mean error was 57.68 degrees with normalization and 57.60 degrees with amplitude alone. The paired difference was +0.08 degrees, with a descriptive 95% person bootstrap interval of -0.59 to +0.70 degrees. Seven rest recordings gave a separate difference of +0.07 degrees, with an interval of -0.73 to +0.86 degrees. The comparison did not show a clear forecast error gain. The reference used future voltage. No stimulation or measured hardware gain change was tested. The package preserves voltage inputs, source versions and hashes, code, figures, tables, and instructions. Fresh source prefix audits, independent numerical checks, and a complete replay passed.

SEEG · phase forecasting · confidence gate · residual normalization · gain invariance · signal processing