Document research findings (corrected eval, capability-bound, anchored fusion)

README.md

@@ -67,3 +67,21 @@ same instruction string as `train_grpo.py`.
 ## Build
 
 Built by `build_brill_dataset.py`. Label cap: 20 codes/image.
+
+## Research context & key finding
+
+This dataset was built to test whether the iconclass classifier's ~25% recall ceiling was
+caused by **truncated training labels** (the original `iconclass-vlm-sft` was capped at 3.54
+codes/image; this restores the full Brill labels at ~4.36).
+
+**Re-SFT on these fuller labels did _not_ improve the model.** Training converged well
+(eval_loss 0.47) but on the contamination-safe `test` split it scored H-F1 45.3 /
+hier-recall 46.4 — recall unchanged. The bottleneck is model **capability** (identifying the
+right codes), not label completeness. The lever that *did* work was **anchored fusion** (the
+fine-tuned model as a precision anchor + a graded VLM-judge gating in semantic-retrieval
+recall → H-F1 47.5 / hier-recall 57.6, with no extra training).
+
+### Splits & contamination
+- `train` (86,216) / `test` (788), split deterministically by image filename hash (disjoint).
+- The `test` split is clean for models trained on this dataset's `train` split. Older
+  checkpoints trained on the overlapping `iconclass-vlm-sft` images are **contaminated** on it.