Probability weights and choice offsets reduced later log loss in 165 ds002843 first visits with fixed bounds and 2% lapse

Two models improved predictions for later risk choices in 165 first visits from ds002843 version 1.0.1. We fitted runs 1 and 2 and tested runs 3 and 4. All main models used a fixed 2% lapse and fixed parameter bounds. Probability weights reduced mean loss from 0.4151 to 0.3735 nats per choice. The paired gain was 0.0416 nats, with a 95% person bootstrap interval of 0.0222 to 0.0622. A choice offset also reduced loss, by 0.0395 nats. The interval for the difference between the flexible models included zero. The gain for probability weights was uncertain at zero lapse. Thus greater model capacity improved later prediction under the main assumptions. This test did not identify probability weights as the specific reason. The package preserves source bytes, provenance, all predictions, bootstrap draws, code, and a full fresh-source rebuild check. The evidence covers one behavioral task. It does not show a brain mechanism.

risk choice · probability weighting · expected utility · choice prediction · log loss · behavioral modeling · model comparison