# Benchmark Protocol

1. Load a processed HDF5 dataset.
2. Select an input component, such as `X_raw`, `X_wake_raw`, `X_potential_raw`, `X_external_raw`, `X_self_raw`, or `X_total_raw`.
3. Construct temporal windows of length `H` using saved wake phases for each pose.
4. Split by held-out test pose: all phases of each test pose are withheld. Split the remaining non-test samples into training and validation.
5. Standardize inputs using training-set statistics only.
6. Train the temporal MLP baseline or another model.
7. Report global and close-wake position RMSE; report circular `theta_rel` MAE for full-state runs.

The quick training command is for functionality testing only. Manuscript numbers come from stored summaries in `data/results/`.