Logistic maps trained on other people reduced Brier loss in later runs of 20 people in ds002158 v1.0.2 when confidence was used as probability

We tested confidence as a probability of correct choices in 20 people from OpenNeuro ds002158 version 1.0.2. Each logistic map used early records from the other 19 people. It scored the excluded person on later runs. Mean Brier loss fell from 0.09282 to 0.06505. The gain was 0.02777, with a refitted person-bootstrap 95% interval of 0.01926 to 0.03549. All 20 people had positive mean gains. All 40 primary maps preserved confidence order within each person and condition. Seven planned checks also gave positive gains. A separate parser and optimizer checked the results. A full replay matched 18 numerical files. The package preserves original source rows, hashes, the local plan, code, fitted maps, predictions, sample counts, and instructions. This result assumes that raw confidence values can be used as probabilities. It concerns this small public group and later runs of the same visit. It does not establish transfer to a new task or a clinical group.

confidence calibration · Brier loss · probability prediction · cross-validation · bootstrap · metacognition · OpenNeuro