Clarify significance claim in baselines

DEEPARITHMARK_DESIGN.md

@@ -280,7 +280,8 @@ a definable ceiling. It also shows the benchmark is not degenerate — it is not
 ### The 2026 band: the first sub-1B gradient
 
 Qwen3-0.6B-Base (596 M) is the first sub-1B base model in the table to clear
-chance: **29.20 %**, p = 0.0027 against 25 %. More importantly it is the first to
+chance significantly: **29.20 %**, p = 0.0027 against 25 % (GPT-X3-150M's 25.2 %
+is nominally above chance but p = 0.44). More importantly it is the first to
 show the *shape* of partial skill rather than a flat line — the depth curve
 decreases monotonically and is significant at both ends (n = 250 per tier):