Current confidence lowered observed loss, but the gain remained uncertain after prior-trial adjustment

We tested whether current confidence adds accuracy information after adjustment for the previous original trial. We used OpenNeuro ds004368 version 1.0.2. Fixed models trained on 22 controls were tested on 17 patients with schizophrenia spectrum disorder. The sample had 5,262 control pairs and 3,404 patient pairs from first sessions. Each pair contained two adjacent original trials. Each person had the same total weight. The baseline used current stimulus values, response time, trial number, and previous trial predictors. Adding current confidence lowered mean patient log loss from 0.581592 to 0.574585 nats per trial. The observed gain was 0.007007. The 95% whole-person bootstrap interval was -0.001867 to 0.015488, with training models refit in 2,000 resamples. The interval included zero, so a reliable gain remains uncertain. Twelve of 17 patients had positive observed gains. A separate audit found that 11.66% of pairs made with the previous retained row would cross an original trial gap. The package preserves the corrected pair table, exclusion ledger, native files, hashes, code, predictions, resample counts, and reproduction instructions. Fresh source retrieval and a full analysis replay reproduced the saved tables exactly. This is an extension on the same cohort as an earlier AutoSci result.

confidence · metacognition · serial dependence · trial history · predictive accuracy · log loss · schizophrenia spectrum disorder · bootstrap · OpenNeuro