[
  {
    "benchmark": "bfcl",
    "retention": 1.0,
    "token_savings_pct": 79.3,
    "baseline_accuracy": 1.0,
    "baseline_ci_95": [
      0.9286,
      1.0
    ],
    "entroly_accuracy": 1.0,
    "entroly_ci_95": [
      0.9286,
      1.0
    ],
    "baseline_avg_tokens": 2849.6,
    "entroly_avg_tokens": 589.7,
    "samples": 50,
    "provenance": {
      "model": "gpt-4o-mini",
      "seed": 42,
      "budget": 500,
      "timestamp_utc": "2026-05-24T01:34:18.155208+00:00",
      "git_sha": "0a54982698ee",
      "elapsed_s": 89.7,
      "audit_jsonl": "bfcl_audit.jsonl",
      "audit_record_count": 50,
      "scoring": "bench.accuracy._check_answer (benchmark='bfcl')",
      "compressor": "entroly.qccr.select (via bench.accuracy._compress_messages_modal mode='entroly')",
      "dataset_loader": "_load_bfcl"
    }
  }
]