part 2 summary: correct the interpretation (inconclusive, not equivalence)

chess-sim/reports/part2/summary.md

@@ -57,12 +57,15 @@ Paired comparisons (exact McNemar, episodes only one model completed):
 
 1. **The normalization override was a real, large harm.** It is now controlled: keep the
    baseline's normalization in every fine-tune.
-2. **The correction data makes no detectable difference** once normalization is kept (A vs B).
-3. **Neither fine-tune measurably changes the baseline.** Both scored a few episodes lower,
-   within noise.
+2. **The effect of the correction data is inconclusive** once normalization is kept (A vs B,
+   p = 0.82). This test cannot detect small or moderate effects, so it is not evidence that the
+   corrections have no effect.
+3. **Both fine-tunes scored lower than the baseline** (21 and 23 against 28-31 of 60). The
+   differences are not significant (p = 0.14 and 0.36), so it is not established whether the
+   decline is real; it does not show equivalence.
 4. **The failure is unchanged.** The target is knocked over while grasping in 26-28 of 60 for
-   every model, and closing errors stay at 4.7-5.4 mm. Fine-tuning on this data does not improve
-   grasp precision.
+   every model, and closing errors stay at 4.7-5.4 mm. Fixing the normalization did not solve
+   contact precision.
 5. **Closed-loop tests are noisy.** Two runs of the same model on the same seeds differ in 13
    of 60 episodes, from tiny GPU numeric differences amplified over a 20 s rollout. Sixty
    episodes can only detect large effects (about 15 points).