document experiment_cfg + embodiment slot

README.md

@@ -41,3 +41,28 @@ decent-vla `--resume`, not `from_pretrained`:
   (the LocalRunner end-of-run save, which also records embodiment + backend metadata).
   Note there is no `server_round_0075.pt`: periodic saves ran every 10 rounds, and 75 is
   not a multiple of 10, so round 75 exists only as this end-of-run file.
+
+## Running these weights (experiment_cfg/)
+
+The checkpoint holds ONLY `action_head.*` (537 tensors, 1.6205 B). Three things GR00T
+derives from the training dataset are not in it and not in the public N1.7 base repo, so
+they are published here under `experiment_cfg/`:
+
+| file | what it is |
+|---|---|
+| `embodiment_id.json` | the projector-slot mapping used in this run |
+| `processor_config.json` | modality config (state/action index ranges, video keys) + image/state processor settings |
+| `statistics.json` | state/action normalization statistics |
+
+**Embodiment slot = 10.** `new_embodiment` is absent from the base repo's
+`embodiment_id.json`; the value comes from `EMBODIMENT_TAG_TO_PROJECTOR_INDEX` in
+`gr00t/model/gr00t_n1d7/processing_gr00t_n1d7.py`. Verified against the weights, not just
+that table: diffing this checkpoint's 14 embodiment-indexed tensors against the base head
+puts **100.0% of the drift in slot 10 and exactly 0.0 in all 31 other slots** (see
+`experiment_cfg/slot_drift.json`). Slot 0 is untrained — falling back to it runs without
+error and produces meaningless actions.
+
+Modality (from `processor_config.json`): video keys `front`, `wrist`; state and action
+`single_arm` (0-5) + `gripper` (5-6); action horizon 16; `shortest_image_edge` 256.
+
+The training config for this run is included as `experiment_cfg/train_config.yaml`.