{
  "schema_version": 3,
  "_note": "Top-level headline fields mirror the canonical run (openai x sfdc-to-ns-customers in --matrix mode; the single run otherwise); runs[] carries every provider x pair cell of this invocation. v3 adds accuracy_top1_ci95 (Wilson 95% score interval) per run + headline mirror. accuracy_top1 (fraction 0..1) is mirrored into metrics.json:ai_accuracy.latest by scripts/generate-metrics.mjs; README templates substitute METRIC:ai_accuracy.latest_pct. See docs/review/ai-accuracy-benchmark.md for the runbook.",
  "run_mode": "live",
  "provider": "openai",
  "model": "gpt-5.4-mini",
  "fixture": "scripts/golden/fixtures/sfdc-to-ns-customers.yaml",
  "target_system": "netsuite",
  "fixture_cases": 22,
  "fixture_mappings": 61,
  "accuracy_top1": 0.9508,
  "accuracy_top1_ci95": {
    "low": 0.8651,
    "high": 0.9831
  },
  "hallucination_count": 0,
  "manual_edit_rate": 0.0492,
  "estimated_cost_usd": 0.039914,
  "max_cost_usd": 5,
  "duration_ms": 25860,
  "generated_at": "2026-06-12T05:24:57.355Z",
  "total_estimated_cost_usd": 0.114174,
  "runs": [
    {
      "pair": "sfdc-to-ns-customers",
      "fixture": "scripts/golden/fixtures/sfdc-to-ns-customers.yaml",
      "target_system": "netsuite",
      "provider": "openai",
      "model": "gpt-5.4-mini",
      "run_mode": "live",
      "fixture_cases": 22,
      "fixture_mappings": 61,
      "accuracy_top1": 0.9508,
      "accuracy_top1_ci95": {
        "low": 0.8651,
        "high": 0.9831
      },
      "hallucination_count": 0,
      "manual_edit_rate": 0.0492,
      "estimated_cost_usd": 0.039914,
      "duration_ms": 25860
    },
    {
      "pair": "sfdc-to-bc-customers",
      "fixture": "scripts/golden/fixtures/sfdc-to-bc-customers.yaml",
      "target_system": "businesscentral",
      "provider": "openai",
      "model": "gpt-5.4-mini",
      "run_mode": "live",
      "fixture_cases": 5,
      "fixture_mappings": 11,
      "accuracy_top1": 1,
      "accuracy_top1_ci95": {
        "low": 0.7412,
        "high": 1
      },
      "hallucination_count": 0,
      "manual_edit_rate": 0,
      "estimated_cost_usd": 0.002869,
      "duration_ms": 3451
    },
    {
      "pair": "sfdc-to-ns-customers",
      "fixture": "scripts/golden/fixtures/sfdc-to-ns-customers.yaml",
      "target_system": "netsuite",
      "provider": "anthropic",
      "model": "claude-haiku-4-5",
      "run_mode": "live",
      "fixture_cases": 22,
      "fixture_mappings": 61,
      "accuracy_top1": 0.9672,
      "accuracy_top1_ci95": {
        "low": 0.8881,
        "high": 0.991
      },
      "hallucination_count": 0,
      "manual_edit_rate": 0.0328,
      "estimated_cost_usd": 0.066225,
      "duration_ms": 34012
    },
    {
      "pair": "sfdc-to-bc-customers",
      "fixture": "scripts/golden/fixtures/sfdc-to-bc-customers.yaml",
      "target_system": "businesscentral",
      "provider": "anthropic",
      "model": "claude-haiku-4-5",
      "run_mode": "live",
      "fixture_cases": 5,
      "fixture_mappings": 11,
      "accuracy_top1": 1,
      "accuracy_top1_ci95": {
        "low": 0.7412,
        "high": 1
      },
      "hallucination_count": 0,
      "manual_edit_rate": 0,
      "estimated_cost_usd": 0.005166,
      "duration_ms": 5209
    }
  ]
}