Fixed confidence rank rules increased later log loss from ds002158 v1.0.2 to ds002739 v1.0.0 and reduced it in the reverse test against source constants

We fitted logistic rules to early confidence ranks in OpenNeuro ds002158 v1.0.2 and ds002739 v1.0.0. We tested later choices in the other source. The samples had 20 dot-task people and 24 motion-task people. Each comparison used a constant fitted to the same source people. Dots to motion increased mean log loss from 0.544316 to 0.782979 nats per trial. Its gain was -0.238664, with a 97.5% interval from -0.434239 to -0.112184. Motion to dots reduced loss from 0.591243 to 0.399664. Its gain was 0.191579, with an interval from 0.152558 to 0.233973. Gain means constant loss minus rule loss. We refitted the rules in 4,999 bootstrap draws from both person samples. The two intervals use an adjustment for two primary tests. All three planned checks kept the same directions. Confidence ranks remove numeric scale units under fixed row selection. They did not give a useful transfer rule in both tested directions. Early target labels defined eligibility but did not fit the transfer rules. The result does not separate task, rating method, tie structure, selection, and person effects. It does not identify a neural cause or clinical performance. The package preserves all 224 source files, hashes, code, the local plan, predictions, draws, figures, and repeat instructions. Independent checks passed. A full repeat from fresh downloads matched all numerical outputs.

confidence · perceptual decision · accuracy prediction · cross-task transfer · log loss