Clarify SegFormer-B0 evaluation caveats

segformer-b0/README.md

@@ -55,6 +55,14 @@ requires spatial dimensions divisible by 32, so this model was trained and
 evaluated at 416 × 416. The comparison is therefore not resolution-aligned, as
 already flagged on the GitHub PR.
 
+## Caveats
+
+- The IGN test evaluation contains positive images only; false positives on
+  negative scenes are not measured.
+- This is a single run with seed 42, so seed-to-seed variability is unknown.
+- The test metrics are pixel-aggregated; the reference-protocol row above is
+  included separately for comparison with the published DeepLab baseline.
+
 ## Evaluate
 
 ```bash