Add measured ADE20K mIoU (LibreYOLO val protocol)

README.md

@@ -64,10 +64,17 @@ unchanged: the converted checkpoint reproduces upstream logits **bit-exactly**
 
 ## Accuracy
 
-Single-scale ADE20K val mIoU as reported by the SegFormer authors for this
-checkpoint: **46.5**. Numbers are the upstream authors', measured with their
-evaluation protocol; LibreYOLO's `val` uses a fixed-canvas letterbox and will
-not reproduce them exactly.
+ADE20K val (2000 images), single scale:
+
+| | mIoU |
+|---|---|
+| **LibreYOLO `model.val(data="ade20k.yaml")`** | **45.0** |
+| Upstream authors, reported | 46.5 |
+
+LibreYOLO's validator letterboxes to a fixed square canvas; the authors evaluate with a
+ratio-preserving resize, which is worth roughly a point. The converted weights reproduce the
+upstream model's logits **bit-exactly** (max abs difference 0.0 in float64), so any remaining
+difference is evaluation protocol, not conversion.
 
 ## License