A fixed covariance correction increased mean recall forecast loss in 16 SEEG session pairs after five early lists at C=0.01

We tested early covariance correction for later recall forecasts with public SEEG data from OpenNeuro ds004809 version 2.2.0. The test used 16 reused session pairs, 96 signal features, and 2880 later test words. Neural readers used five early lists to set feature scales. Each reader used the same 60 early labels for one intercept. The full map used no target labels and had covariance shrinkage 0.5. At neural penalty C=0.01, it increased mean loss by 0.009629 nats per word over separate feature normalization. Loss increased in 13 of 16 people. The descriptive 95% paired person interval was [0.002973, 0.015078]. The mean early covariance gap fell by 46.23%. Thus, a smaller early gap did not ensure better later forecasts for this fixed reader. The package preserves data, source records, code, all forecasts, three figures, and independent checks. This exploratory offline test does not establish a stimulation benefit or a result for new people.

covariance correction · SEEG · recall forecasting · domain adaptation · log loss · memory interfaces