{
  "classifier_model": "narcolepticchicken/legal-router-distilbert-v2",
  "baseline_model": "HuggingFaceTB/SmolLM2-360M-Instruct",
  "classifier_architecture": "DistilBERT (67M params) with 3 independent classification heads, weighted loss for attorney class (10.5x)",
  "baseline_approach": "Zero-shot prompt-based routing via structured JSON generation",
  "dataset": "omersx/business-legal-disputes with synthesized legal routing labels",
  "train_size": 5000,
  "test_size": 400,
  "classifier_metrics": {
    "route_accuracy": 0.9025,
    "route_macro_f1": 0.8561,
    "attorney_accuracy": 0.9025,
    "attorney_f1": 0.3158,
    "attorney_recall": 0.8182,
    "source_accuracy": 0.8675,
    "source_f1": 0.9259,
    "missed_escalation_rate": 0.1818,
    "latency_ms_per_sample": 5.0,
    "parse_failure_rate": 0.0
  },
  "baseline_metrics": {
    "route_accuracy": 0.0,
    "route_macro_f1": 0.0,
    "attorney_accuracy": 1.0,
    "attorney_f1": 0.0,
    "attorney_recall": 0.0,
    "source_accuracy": 0.0,
    "source_f1": 0.0,
    "missed_escalation_rate": 0.0,
    "latency_ms_per_sample": 1672.0,
    "parse_failure_rate": 1.0,
    "note": "SmolLM2-360M-Instruct failed to produce valid JSON in 100% of cases. All predictions defaulted to False/unknown. Route accuracy and F1 are 0 because 'unknown' is not a valid route class."
  },
  "test_class_distribution": {
    "contract_law": 185,
    "commercial_law": 100,
    "employment_law": 90,
    "tort_law": 25,
    "attorney_review_true": 11,
    "attorney_review_false": 389,
    "source_required_true": 347,
    "source_required_false": 53
  },
  "comparison_summary": {
    "route_classification": "Classifier 90.25% accuracy. Baseline: 0% (100% parse failure).",
    "missed_escalation": "Classifier misses 18.2% of attorney escalations (2 of 11). Baseline misses all (always predicts False).",
    "latency": "Classifier ~5ms/sample. Baseline ~1672ms/sample. 334x speedup.",
    "parse_failure": "Classifier: 0% (deterministic). Baseline: 100% (SmolLM2-360M cannot produce structured JSON reliably).",
    "note": "Attorney recall of 81.8% is impressive given extreme class imbalance (11/400 positive). Weighted loss training (10.5x) successfully recovered most of the positive class."
  }
}