Describe local LLM human-in-loop annotation workflow

Remove stale owner-ruling and document-domain claims; retain revised holdout counts.

README.md

@@ -30,9 +30,7 @@ configs:
 ## Annotation workflow
 
 Labels were produced with a locally run LLM through the passage-annotation harness, with
-human-in-the-loop passage review and adjudication under the project's data-use doctrine. The
-holdout was subsequently revised through a passage-level FP/FN disagreement audit at score
-cutoff 0.55. This revision is post-hoc and is not an independent benchmark.
+human-in-the-loop passage review and adjudication under the project's data-use doctrine. The same passage-level annotation workflow was used for train, validation, and holdout.
 
 ## Labels
 
@@ -81,7 +79,3 @@ from datasets import load_dataset
 ds = load_dataset("rafmacalaba/datause-agents-v3", "gliner")   # ds['train'] / ds['val'] / ds['holdout']
 ```
 
-
-## Holdout audit revision (2026-09-27)
-
-The holdout labels were revised after a passage-level, doctrine-based review of model/gold disagreement cases using `gliner_agents_v3` predictions at score cutoff 0.55. Train and validation splits are unchanged. The revised holdout was adjudicated through the passage-annotation harness with human-in-the-loop review. Because review was triggered by this model's own predictions, its scores are a post-hoc diagnostic, not an independent accuracy benchmark. Do not train on this holdout or tune a deployment threshold from it. Use a fresh independently annotated split for independent evaluation.