For 53 people in ds004086 v1.2.0, separate confidence curves gave a later-run prediction gain with a 95% interval that included zero

We tested whether confidence curves that differ for old and new answers improve recognition accuracy prediction. The source was OpenNeuro ds004086 version 1.2.0. The common model already had an answer offset. The separate model added three terms for different confidence curve shapes. For each of 53 people, both models used early runs from other people and predicted that person's later runs. The test included 6,070 valid late trials. Mean log loss fell from 0.654265 to 0.652835 nats per trial. The gain was 0.001430 nats per trial. A 1,999-sample person bootstrap refitted both models and removed all copies of each test person from training. Its 95% interval was -0.000683 to 0.003533. All planned checks also had intervals that included zero. A secondary confidence ranking measure was higher for old answers, but the primary test gave no clear prediction benefit from the extra terms. We repaired only exact complete table copies. The package preserves original source bytes, trial records, models, predictions, bootstrap samples, code, and figures. An independent optimizer matched all primary fits. A full repeat from new source downloads matched all result files. The result concerns behavior in one fMRI task. It does not establish model equivalence or a neural cause.

confidence · recognition memory · accuracy prediction · logistic regression · OpenNeuro · cross-validation · bootstrap