piper README: null baselines -- the policy beats the trajectory baselines 3-4x and loses to persistence, so do not rank it by action L1

piper_pick_can_10k/README.md

@@ -48,6 +48,33 @@ protocol but it is **not** the error you will see. The table above is. Under tha
 curve is flat (0.0972 → 0.0930), which is why this checkpoint is the last one and not an earlier one; the open-loop
 error of the earlier snapshots has not been measured.
 
+## Is that error good or bad?  The null baselines, same 8 episodes
+
+Mean absolute error per joint, degrees:
+
+| | mean per joint |
+|---|---|
+| constant = the training mean action | 14.70 |
+| the average demonstration, time-warped to each episode | 11.02 |
+| **this policy** | **3.24** |
+| repeat the last command (persistence) | 2.06 |
+
+Read it this way. The policy **beats both trajectory-level baselines by 3-4x**, so it is following each episode's
+own trajectory rather than replaying an average — it has learned the task. It **loses to doing nothing**, because
+the action is an ABSOLUTE joint target at 10 Hz and the arm only moves ~2 deg per step (joint 5: 0.08 deg), which
+makes persistence a very strong L1 baseline that would of course fail every episode on the robot. **Do not rank this
+policy by action L1** — the metric rewards standing still. What the number does say is that its precision (~3 deg)
+is of the same order as the per-step motion, and that it adds ~1 deg of error on joints that barely move.
+
+Averaging samples will not help: the flow head is stochastic, but 8 samples per observation only improve the error
+from 0.1483 to 0.1452 (normalised), and more Euler steps are slightly worse — about 0.145 of the 0.148 is
+systematic bias, not sampling noise. If the policy is imprecise on the robot, the levers are training-side (more
+than 142 demonstrations of one task, delta instead of absolute actions, a higher control rate), not inference-side.
+
+Temporal ensembling (on by default, as in the LIBERO evaluator) is slightly *worse* here offline, 3.69 vs 3.24 deg:
+it was tuned where the action is a delta and its lag is free. Pass `ensemble=False` to try the other way — offline
+cannot decide between them.
+
 ## What has NOT been established
 
 **This policy has never been run on a physical robot.** Everything above is offline. Action error on held-out