Datasets:

rca-umb
/

five-desks-code

@@ -59,6 +59,6 @@ The untrained 4/40 reference is the wrong comparison. naklitechie's null control
 reward, so every group has zero advantage (Trackio: loss 0 and grad_norm 0 on all 236 steps): the model leaves the
 pipeline unchanged. Those runs, with the other runs that trained with no updates, scored 0, 1, 2 and 2 (mean 1.25),
 against about 2.9 for trained runs. Inside this pipeline an unchanged model scores well below 4/40, so training helps on
-average, and "change the model less" is not supported. The 0–25-step runs above trained with gradients on most of
-their steps; their 3.75 (n=4) says nothing about zero-gradient runs, and terminal-habits-v1 (93% zero-gradient steps)
-scored 2/40 (#238).

 reward, so every group has zero advantage (Trackio: loss 0 and grad_norm 0 on all 236 steps): the model leaves the
 pipeline unchanged. Those runs, with the other runs that trained with no updates, scored 0, 1, 2 and 2 (mean 1.25),
 against about 2.9 for trained runs. Inside this pipeline an unchanged model scores well below 4/40, so training helps on
+average, and "change the model less" is not supported. The 0–25-step runs above are few steps, not
+zero-gradient steps (their gradient share is not public), so their 3.75 (n=4) says nothing about unchanged models;
+terminal-habits-v1, with 93% zero-gradient steps, scored 2/40 (#238).