Card: soften jev rounding claim to hypothesis, per fact-check

README.md

@@ -59,8 +59,9 @@ Question wording and criteria are identical across engines and frozen in the man
 
 ## Scoring conventions (matters for calibration metrics)
 
-- jev and laya return display-rounded probabilities (sums off by up to 0.01); all vectors are
-  renormalized before scoring. jev affected 416 rows.
+- jev returns probability maps with sum drift up to 0.01 (display rounding is the working
+  hypothesis — see the run's error notes); 416 rows affected. laya probabilities pass through
+  display rounding (round(4)). All vectors are renormalized before scoring.
 - julia Banking77 probabilities are conditional on the final tournament candidate set
   (`probability_scope=final_candidates`) — the vendor's documented Router behavior.
 - NLL is clipped; ECE is 15 equal-width bins.