{
  "benchmark": "maatwork-lab-documents-v2",
  "corpus": {
    "documents": 100,
    "composition": {
      "invoices": 50,
      "purchase_orders": 50
    },
    "formats": [
      "ISO and DD/MM/YYYY dates",
      "decimal point and Argentine decimal comma",
      "subtotal, tax and final total",
      "buyer and supplier labels",
      "present and absent purchase-order references"
    ],
    "synthetic": true,
    "customer_data": false
  },
  "baseline": {
    "documents": 100,
    "classification_accuracy_pct": 50,
    "field_exact_match_pct": 84.57,
    "structured_output_valid_pct": 50,
    "human_review_rate_pct": 50,
    "latency_ms_p50": 56970,
    "latency_ms_p95": 62236
  },
  "specialized": {
    "documents": 100,
    "classification_accuracy_pct": 100,
    "field_exact_match_pct": 100,
    "structured_output_valid_pct": 100,
    "human_review_rate_pct": 0,
    "latency_ms_p50": 56278,
    "latency_ms_p95": 64898
  },
  "publication_gate": {
    "passed": true,
    "rule": "field_exact_match must improve and structured_output_valid must not regress"
  },
  "run_metadata": {
    "evaluated_at": "2026-09-17T07:58:54.621Z",
    "model_alias": "maat/general",
    "model_build": "Qwen3.8-27B MLX mxfp4 (internal build)",
    "provider": "Maat MLX gateway",
    "hardware": "Apple M2 Max · 64 GB unified memory",
    "temperature": 0,
    "batch_size": 10,
    "synthetic_holdout": true
  },
  "limitations": [
    "Synthetic text-only corpus; no OCR or document layout evaluation.",
    "Confidence threshold is a review-gate proxy, not a calibrated probability.",
    "Batch latency is local hardware performance and is not a production SLA.",
    "The result is not a customer outcome and must not be presented as one."
  ]
}