Add the LLM judge's answers and update the card

README.md

@@ -151,6 +151,8 @@ The labels come from the construction, so we checked that people see the same th
 
 The anonymized judgments and the annotation tool are released [on GitHub](https://github.com/idiap/dialfork/blob/main/docs/human_study.md).
 
+An LLM judge calibrated on these judgments (`qwen3.6:35b`, which agrees with each annotator about as well as two annotators agree with each other) extends the same checks to every dialogue: it confirms 84% of the injected instructions, against 2% of foils, and places 84% of onsets within one turn of `fork_assistant_turn`. Its answers are in `extra/dialfork_llm_judge.csv.gz`, one row per dialogue (same `id` as the dataset); its prompts and code are [on GitHub](https://github.com/idiap/dialfork/blob/main/docs/llm_judge.md).
+
 ## 🚀 Usage
 
 ```python
@@ -240,7 +242,7 @@ task3_train = ds.filter(lambda r: r["task3_split"] == "train" and r["kind"] == "
 task3_test  = ds.filter(lambda r: r["task3_split"] == "test")    # what Task 3 is scored on
 ```
 
-The split is for Task 3 only. **Tasks 1 and 2** measure distances, learn nothing from the data, and use every dialogue; methods for them must not train on or tune with the labels (`fork_position`, `fork_assistant_turn`, `fork_turn_index`, `hallucination_type`).
+The split is for Task 3 only. **Tasks 1 and 2** compare forks with their conformant reference, which is part of each task's input and never itself scored, so they use every dialogue; methods for them must not train on or tune with the labels (`fork_position`, `fork_assistant_turn`, `fork_turn_index`, `hallucination_type`).
 
 The split is also in `extra/task3_splits.csv` (one row per source dialogue), which `scripts/download_data.py` writes as `datasets/dialfork/task3_splits.csv`.
 

extra/dialfork_llm_judge.csv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17e5ac42fd7a7342b5ae39de93080f297c9125f91d45a611abe5810be11f1150
+size 328985