Two-state fits with expected error updates gave fast-state shares of at most 10% and at least 90% within 1% loss in 9 of 31 ds004021 v1.0.1 day 1 records

We tested whether rotation fits could assign most of the final correction to different hidden model states. We used OpenNeuro ds004021 version 1.0.1, with expected error updates, zero initial states, and one state pair shared across target directions. In 9 of 31 eligible day 1 records, two fits had loss within 1% of the lowest stored candidate loss. One fit had a final fast-state share of at most 10%. The other had a share of at least 90%. Each fit had a retention gap of at least 0.01 and slow learning of at least 0.0001, within the saved numerical tolerance. Day 2 had pairs in 14 of 32 records from the same people. Neither day had pairs at a 0.1% loss tolerance. The package preserves 134 source files, 12,033 candidates, 189 separately checked fits, code, and instructions. The reference is the lowest loss in the stored candidate pool. The search does not prove a global minimum. The fit tolerance is not a confidence region. The result concerns model states in the selected public sample.

visuomotor adaptation · two-state model · hidden-state inference · parameter identifiability · motor learning · model fitting · OpenNeuro ds004021