press_button_terra_delta README: ensembling undecided for delta

press_button_terra_delta/README.md

@@ -22,6 +22,11 @@ while True:                                                   # 10 Hz
         wait_until_next_10hz_tick()
 ```
 
+**Ensembling is undecided for this checkpoint.** On the pick-can DELTA arm temporal ensembling *helped* (0.914 vs
+1.054 deg, LEDGER 2026-11-10); here, on 5 held-out episodes, it is slightly *worse* (0.968 vs 0.918). The
+recommended n=2 / `ensemble=False` recipe was derived on the ABSOLUTE arm. Try both on the robot
+(`p.act(frame, instr)` gives the ensembled action, then `to_joint_targets` it the same way).
+
 Chunk entry k was trained relative to the joint state **at step k**, so `q` must be read every tick, not once per
 chunk. `ctrl_meta.json` carries `"action_space": "delta"`; `piper_policy.py` reads it and `to_joint_targets()`
 refuses to run without a joint state. Verified on all 1,877 dataset steps: `q + delta` reproduces the demonstrated