Final Model Readiness Report

• Readiness status: READY
• Dataset root: data/generated
• Checkpoint: model_output/best_checkpoint.pt
• Eval report: reports/eval_report.json

Dataset Summary

• Splits: [{"split": "train", "samples": 3500, "shards": 70}, {"split": "val", "samples": 750, "shards": 15}, {"split": "test", "samples": 750, "shards": 15}]

Validation/Test Metrics

• Metrics: {"macro_f1_relevance": 1.0, "macro_f1_scope": 0.9819635502193986, "macro_f1_intent": 0.9393803679117839, "dispute_leakage_matrix": [[495, 0, 0], [0, 54, 0], [0, 0, 74]], "extraction_exact_match": 1.0, "extraction_partial_match": 1.0, "hard_negative_fp_rate": 0.0, "by_thread_depth": {"3": 1.0, "1": 1.0}, "by_evidence_location": {"subject": 1.0, "history": 1.0, "signature": 1.0, "attachment": 1.0, "body": 1.0}}

Known Risks

• Synthetic data only — shadow-real evaluation pending production corpus
• All samples AR-relevant — non-AR-relevant class not yet generated