{
  "accuracy_percent": 95.0,
  "categories": [
    "sales_lead",
    "support",
    "spam",
    "sensitive_review",
    "manual_review"
  ],
  "dataset": "synthetic_only",
  "failures": [
    {
      "expected": "sales_lead",
      "id": "synthetic-001",
      "predicted": "manual_review",
      "reason": "Deliberately ambiguous synthetic wording falls to manual review."
    },
    {
      "expected": "spam",
      "id": "synthetic-018",
      "predicted": "manual_review",
      "reason": "Deliberately ambiguous synthetic wording falls to manual review."
    },
    {
      "expected": "support",
      "id": "synthetic-052",
      "predicted": "manual_review",
      "reason": "Deliberately ambiguous synthetic wording falls to manual review."
    },
    {
      "expected": "sensitive_review",
      "id": "synthetic-069",
      "predicted": "manual_review",
      "reason": "Deliberately ambiguous synthetic wording falls to manual review."
    },
    {
      "expected": "sales_lead",
      "id": "synthetic-086",
      "predicted": "manual_review",
      "reason": "Deliberately ambiguous synthetic wording falls to manual review."
    }
  ],
  "inquiries": 100,
  "limitations": [
    "Synthetic inquiries do not establish production accuracy.",
    "No customer data or customer attribution is used.",
    "Time savings are not claimed because handling time was not measured.",
    "Sensitive and ambiguous cases remain under human review."
  ],
  "manual_review_count": 44,
  "manual_review_percent": 44.0,
  "method": "Deterministic keyword rules with confidence below 0.75 routed to human review.",
  "predicted_counts": {
    "manual_review": 25,
    "sales_lead": 18,
    "sensitive_review": 19,
    "spam": 19,
    "support": 19
  },
  "schema_version": 1
}