Update README.md

README.md

@@ -24,24 +24,12 @@ clf("She reached out and touched the rough bark.")
 # -> [{'label': 'haptic_material', ...}]
 ```
 
-The classifier takes the sentence alone - no surrounding context. Human annotators
-saw a ~200-word context window, so borderline cases the model gets wrong are often
-ones a reader would resolve from context.
 
 ## Performance
 
 Held-out test set (n=580): micro-F1 **0.805** [0.772, 0.835], macro-F1 0.797.
 5-fold CV: 0.808 +/- 0.022. Human inter-annotator agreement ceiling: Cohen's kappa 0.760.
 
-**Read before citing:** these are post-adjudication figures. A review pass revised
-387 gold labels (~10% of the dataset), but it only surfaced cases where the model
-disagreed with the original label, so the gain over the pre-cleanup model
-(test 0.750, CV 0.755) is partly selection bias. Report both.
-
-Known limitation: `segmentation_error` is style-correlated - Dickens and Gissing
-top the per-book rate at 15-20% - so it is not ignorable noise in author-level
-aggregates.
-
 ## Data and code
 
 Codebook, gold annotations, and reproduction notebooks: