An early scale reduced log loss for fixed Gaussian readers in 10 SEEG patients and 6 fMRI adults

We tested whether an early confidence scale reduces later probability error in human neural readouts. Separate public studies supplied SEEG memory-load features and fMRI object responses. The Gaussian readers assumed independence across sites. Earlier records fit class means, then one positive scale. Later tests used 32 sessions in ten patients and 184 blocks in six adults. Mean log-loss gains were 0.9435 nats in SEEG and 29.1132 nats in fMRI. Their 95% person t intervals were 0.5660 to 1.3210 and 20.1970 to 38.0293. All ten and all six person gains were positive. The added shared rule had nominal three-claim Holm p=0.046875. Class choices stayed fixed. Yet corrected mean SEEG loss was 0.7114, versus 0.6931 for uniform labels. Its gain over uniform labels was uncertain. One fMRI Brier score became worse. The SEEG addition was fixed after two fMRI outcomes and before new SEEG scores. It reused the same fMRI test and a known related SEEG archive. An original visual ECoG confirmation was not completed after a scope change during source connection failures. All three development records and both incomplete claims remain. The result supports a narrow method prediction in reused human records. It does not establish perfect calibration, useful SEEG decoding, a common neural cause, or causality. The package gives compact permitted SEEG inputs, source versions and rights, all scores, plans, code, figures, and repeat instructions.

human neuroscience · SEEG · fMRI · confidence · temperature scaling · log loss · Brier score · neural decoding · probability calibration · reproducibility