The 95 percent interval for SART loss gain from prior response time included zero in 24 external people after source model refits, assuming independent people.

We trained two SART error forecast models on 189 people from OpenNeuro ds003823 v1.3.5. We tested fixed models on 24 people from ds004350 v2.0.0. The primary test contained 755 eligible no-go trials. Adding prior go response time reduced mean person log loss from 0.628237 to 0.604372. The 95 percent gain interval was -0.002243 to 0.048497 after source model refits and resampling of both cohorts. This interval assumes independent people within age or feedback groups. With source models fixed, the interval was 0.000474 to 0.045554. The speed model predicted 42.7 percent errors; 28.3 percent were observed. A later exploratory forecast of the past error rate had lower measured loss. Its relative gain interval also included zero. The analysis was exploratory because earlier results from these datasets were known. The package contains all 400 source files, hashes, retrieval code, trial selection records, forecasts, 2,000 model resamples, figures, and repeat instructions. A repeat with fresh source files reproduced all 13 primary output files exactly. Secondary results concern later sessions in the same people. The data do not identify a brain mechanism or treatment effect.

Sustained Attention to Response Task · response time · error forecasting · cross-dataset transfer · logistic regression · calibration · OpenNeuro