The error forecast gain interval included zero for five prior decision times in 21 calibrated ds004119 people under fixed models

We tested whether five prior decision times improved later recorded error forecasts in BCIT Basic Guard Duty, OpenNeuro ds004119 v1.0.0. We used 21 people. Blocks 1 to 5 gave personal calibration. Blocks 6 to 10 gave 2,961 test decisions with 159 recorded errors. The reference logistic model used task labels, recent decisions, personal intercepts, and time controls. The added feature was the median of five prior log decision times. Mean binary log loss fell from 0.186669 to 0.186588 nats. The relative gain was 0.043 percent. A 2,000 draw person bootstrap that refitted both models gave a 95 percent interval from -0.998 to 0.495 percent. This interval included zero. It did not establish added forecast value under these models. We report all six planned sensitivity checks. Separate event and fit code checked the calculations. A repeat from fresh source files matched every numerical output byte. The package preserves source events, hashes, code, and features. It also preserves exclusions, forecasts, coefficients, bootstrap draws, figures, and repeat instructions. This result is a behavioral reference for future EEG tests. It does not test a neural cause, equal model performance, or forecasts for new people without calibration.

error forecasting · decision time · behavioral prediction · logistic regression · bootstrap · EEG · OpenNeuro · Guard Duty