{
  "Julia-1": {
    "verified_rows": 48,
    "argmax_matches_serialized_choice": true,
    "label_ids_match_fixed_candidates": true,
    "expected_unchanged": true,
    "confusion": [
      {
        "task": "evidence_relation",
        "expected": "insufficient",
        "choice": "insufficient",
        "n": 4
      },
      {
        "task": "evidence_relation",
        "expected": "insufficient",
        "choice": "supports",
        "n": 4
      },
      {
        "task": "evidence_relation",
        "expected": "refutes",
        "choice": "insufficient",
        "n": 5
      },
      {
        "task": "evidence_relation",
        "expected": "refutes",
        "choice": "supports",
        "n": 3
      },
      {
        "task": "evidence_relation",
        "expected": "supports",
        "choice": "insufficient",
        "n": 5
      },
      {
        "task": "evidence_relation",
        "expected": "supports",
        "choice": "supports",
        "n": 3
      },
      {
        "task": "intent_routing",
        "expected": "access",
        "choice": "access",
        "n": 7
      },
      {
        "task": "intent_routing",
        "expected": "access",
        "choice": "shipping",
        "n": 1
      },
      {
        "task": "intent_routing",
        "expected": "billing",
        "choice": "access",
        "n": 3
      },
      {
        "task": "intent_routing",
        "expected": "billing",
        "choice": "billing",
        "n": 5
      },
      {
        "task": "intent_routing",
        "expected": "shipping",
        "choice": "access",
        "n": 1
      },
      {
        "task": "intent_routing",
        "expected": "shipping",
        "choice": "shipping",
        "n": 7
      }
    ]
  },
  "Kev-0.8B": {
    "verified_rows": 48,
    "argmax_matches_serialized_choice": true,
    "label_ids_match_fixed_candidates": true,
    "expected_unchanged": true,
    "confusion": [
      {
        "task": "evidence_relation",
        "expected": "insufficient",
        "choice": "insufficient",
        "n": 4
      },
      {
        "task": "evidence_relation",
        "expected": "insufficient",
        "choice": "refutes",
        "n": 4
      },
      {
        "task": "evidence_relation",
        "expected": "refutes",
        "choice": "insufficient",
        "n": 1
      },
      {
        "task": "evidence_relation",
        "expected": "refutes",
        "choice": "refutes",
        "n": 5
      },
      {
        "task": "evidence_relation",
        "expected": "refutes",
        "choice": "supports",
        "n": 2
      },
      {
        "task": "evidence_relation",
        "expected": "supports",
        "choice": "refutes",
        "n": 3
      },
      {
        "task": "evidence_relation",
        "expected": "supports",
        "choice": "supports",
        "n": 5
      },
      {
        "task": "intent_routing",
        "expected": "access",
        "choice": "access",
        "n": 8
      },
      {
        "task": "intent_routing",
        "expected": "billing",
        "choice": "billing",
        "n": 8
      },
      {
        "task": "intent_routing",
        "expected": "shipping",
        "choice": "shipping",
        "n": 8
      }
    ]
  },
  "mDeBERTa": {
    "verified_rows": 24,
    "argmax_matches_serialized_choice": true,
    "label_ids_match_fixed_candidates": true,
    "expected_unchanged": true,
    "confusion": [
      {
        "task": "evidence_relation",
        "expected": "insufficient",
        "choice": "insufficient",
        "n": 4
      },
      {
        "task": "evidence_relation",
        "expected": "insufficient",
        "choice": "refutes",
        "n": 3
      },
      {
        "task": "evidence_relation",
        "expected": "insufficient",
        "choice": "supports",
        "n": 1
      },
      {
        "task": "evidence_relation",
        "expected": "refutes",
        "choice": "refutes",
        "n": 5
      },
      {
        "task": "evidence_relation",
        "expected": "refutes",
        "choice": "supports",
        "n": 3
      },
      {
        "task": "evidence_relation",
        "expected": "supports",
        "choice": "refutes",
        "n": 1
      },
      {
        "task": "evidence_relation",
        "expected": "supports",
        "choice": "supports",
        "n": 7
      }
    ]
  },
  "Qwen3.5-4B": {
    "verified_rows": 48,
    "actual_generated_json_parse_matches": true,
    "duration_unit": "Ollama nanoseconds converted to seconds; client wall time includes request overhead",
    "score_not_probability": true
  },
  "dataset_sha256": "ff2a2f979ad3aa849af10488dd506e6036cc5de43272dc1d7792c9b87a2bc03b",
  "routing_difficulty": {
    "n": 24,
    "labels": 3,
    "examples_per_label": 8,
    "note": "Small explicit consumer-support intents; mostly strong lexical cues, some competing-domain mentions. No unknown/multi-intent/abstain classes; 24/24 does not establish general routing quality."
  },
  "evaluation_date_kst": "2026-10-07"
}