De-emphasize in-distribution table: lead with real-world headline, label synthetic holdout

README.md

@@ -108,7 +108,9 @@ see the Honest evaluation update at the top of this card.
 
 ## Performance
 
-### Overall Metrics (643-sample held-out test set)
+**Real-world headline first.** On *real* legitimate-research queries this line's over-refusal is roughly 18-19% (not the 0% below), and on real bio response-harm (n=554) its successor response head reaches recall 0.921 / AUROC 0.952 but is **Pareto-dominated by the openly-available Qwen3Guard-0.6B**. Everything below is **in-distribution, on a synthetic holdout generated by the same pipeline as the training data** — kept for completeness, not as real-world performance. See the honest summary at the top of this card.
+
+### In-distribution metrics — synthetic holdout, not real-world (643-sample)
 
 | Metric | Value |
 |--------|-------|