Mean choice prediction gains from expectation ratings change sign across three fixed logistic penalties under ds004295 v1.0.0 shape codes

We tested whether expectation reports improve later shape choice predictions after recent history is known. We used 23 retained people and 12,880 trials from ds004295 version 1.0.0. Separate logistic models used early trials to predict 6,440 later choices. The primary model added the previous rating and its interaction with previous feedback. Mean log loss gain was 0.001226 nats per choice. Its 95% participant bootstrap interval was -0.004978 to 0.008230. Only eight people had positive mean gains. The mean gain was negative at the fixed penalty 0.1 and positive at penalties 1 and 10. All eight fixed sensitivity checks are reported. The result assumes stable shape identities in the archived action codes. It concerns later choices within the same people and tasks. It does not establish a neural mechanism or performance in new people. The package preserves source files, corrected trial links, code, features, every prediction, coefficients, bootstrap draws, and instructions. All 1,242 fits passed an independent optimizer check. Fresh downloads matched all source hashes. A full repeat matched every derived data file, numerical result, report, and figure byte for byte.

expectation ratings · choice prediction · reinforcement learning · logistic regression · cross-validation