Reproducibility records
=====================

These are the derived records behind every number reported in the manuscript.
They are sufficient to check the reported statistics without rerunning any
physics.

results/analysis/
    per_window_rollouts.csv   one row per simulated rollout in the analysed
                              panel: window, checkpoint, prediction error,
                              both conditions' instability, and match quality
    participant_primary.csv   the 80 panel windows with their participants
    checkpoint_correlation.json   per-checkpoint, within-window, pooled and
                              accuracy-level results
    statistical_summary.json  the participant-level paired tests
    sensitivity.json          the split and both slopes under equal
                              participant weight, with truncated windows
                              removed, and with the outcome as a rate
    provenance.json           the commit the published code sits at, the
                              hash of each physics source as published and
                              as it stood at the reported run, the
                              controller constants, and package versions

results/prediction_confirmation/, prediction_development/
    the held-out prediction scores and the development selection

results/semg_controls/, temporal_control/, kinematic_input_check/
    the surrogate controls, the 500 ms timing control, and the knee-input check

results/physics/
    the panel manifest, the matching summary, the zero-error preflight, and the
    physics protocol the reported rollouts were run under

results/training_path/
    the descent protocol and the checkpoint manifest

Absolute paths in these records have been replaced with logical roots.
