A personal reward and transition interaction coefficient reduced mean choice loss after a rule change in 61 ds007474 test people under fixed models

We tested whether a person's first-rule reward and transition coefficient improved choice prediction after a reward-rule change. We used public behavior data for ds007474. Of 183 eligible people, 122 trained the population models and 61 tested them. The base model used six other personal coefficients. The added model also used the reward and transition coefficient. Mean test log loss fell from 0.55122 to 0.52735 nats per choice. The gain was 0.02387, with a 95 percent bootstrap interval of [0.00672, 0.04374]. All 2,000 main draws refit the population models. Six planned sensitivity checks and a full numerical replay are preserved. State-to-stimulus transfer had stronger evidence. The opposite direction's interval included zero. The result applies to one visit, supplied task labels, the selected people, and the fixed models. It does not identify a planning or brain mechanism. The local archive retains unchanged source data, provenance, code, trial decisions, predictions, and all bootstrap fits. The SCI payload omits the OSF behavior CSVs because their CC BY-NC-SA 4.0 terms conflict with SCI distribution terms.

two-step task · choice prediction · reward and transition · individual differences · transfer learning · logistic regression