A NARPS prior reduced later log loss in 16 Tom study adults after one calibration run with a unit penalty

We tested transfer of a choice model between two public neuroscience studies. The source contained 54 selected NARPS adults. The target contained 16 Tom study adults. Each target model used run 1 for personal calibration and predicted recorded choices in runs 2 and 3. Both primary models used the same features and a unit penalty. Only the penalty center differed. The source-centered model reduced mean log loss from 0.3252 to 0.2936 nats per recorded choice. The reduction was 0.0316, with joint person bootstrap 95% interval [0.0233, 0.0401]. All 16 person gains were positive. The primary interval used 10,000 draws and included source refitting. All six planned checks were retained. At a lower penalty of q=0.25, the gain was 0.0050 and its interval included zero. The target-only model then scored 0.2942, close to the primary transfer score. Thus, the test does not establish an advantage after separate penalty tuning. The package preserves raw events, code, provenance, predictions, coefficients, bootstrap draws, and instructions. Independent optimization and an exact full repeat passed. The result concerns recorded choices under fixed studies and model rules. It establishes no brain mechanism or stable loss-aversion trait.

choice prediction · logistic regression · transfer learning · mixed gambles · log loss · behavioral neuroscience