Nominal 90 percent point arcs covered every offline theta target in 4.8 percent of one second groups in 17 SEEG rest files

We tested phase arcs for rolling 50 ms theta forecasts in 17 public SEEG rest files from seven patients. The target used offline 4 to 10 Hz filtering and a Hilbert transform. Training, calibration, and later test periods were separate. Nominal 90 percent point arcs contained 89.8 percent of individual targets. They contained every target in only 4.8 percent of complete one second groups. Their mean full width was 278.1 degrees. Arcs calibrated from the largest group error contained every target in 91.2 percent of groups. Their mean width was 358.4 degrees. These are equal patient means from 2910 complete test groups. Descriptive patient bootstrap intervals and fixed secondary checks are saved. All 102 file and configuration results passed an independent numerical audit. An offline repeat matched all 22 analysis outputs. Fresh public range checks reproduced all 17 native inputs. The SCI package preserves native inputs, source pins, scores, models, results, and retrieval instructions; GPL-licensed scripts are omitted, so the exact pipeline cannot be rerun from this package. Contiguous scores were not assumed to be exchangeable. The measured coverage does not establish a prospective or neural phase guarantee.

SEEG · theta phase · phase forecasting · conformal prediction · coverage · uncertainty