{
  "benchmark_id": "OTM-OOS-Replay",
  "version": "v1.0",
  "run_id": "2026-05-25T08:06:05.524946+00:00",
  "completed_at": "2026-05-25T08:06:05.524946+00:00",
  "agent_version": "v0.2.0-oos-clean",
  "n_samples": 110,
  "seeds": [
    42
  ],
  "metrics": {
    "primary": {
      "name": "directional_accuracy",
      "value": 0.5,
      "ci_low": 0.4091,
      "ci_high": 0.6,
      "n": 110
    },
    "secondary": [
      {
        "name": "sharpe_ratio",
        "value": 1.0526
      },
      {
        "name": "cumulative_pnl_bps",
        "value": 1512.88
      },
      {
        "name": "max_drawdown_bps",
        "value": 1203.99
      },
      {
        "name": "win_rate",
        "value": 0.5
      },
      {
        "name": "total_trading_days",
        "value": 110
      },
      {
        "name": "mean_confidence_correct",
        "value": 0.5069
      },
      {
        "name": "mean_confidence_wrong",
        "value": 0.4971
      },
      {
        "name": "total_agent_runs",
        "value": 110
      },
      {
        "name": "total_cost_usd",
        "value": 13.13
      },
      {
        "name": "accuracy_2025-05",
        "value": 0.45,
        "n": 20
      },
      {
        "name": "accuracy_2025-06",
        "value": 0.6316,
        "n": 19
      },
      {
        "name": "accuracy_2025-07",
        "value": 0.4762,
        "n": 21
      },
      {
        "name": "accuracy_2025-08",
        "value": 0.45,
        "n": 20
      },
      {
        "name": "accuracy_2025-09",
        "value": 0.4762,
        "n": 21
      },
      {
        "name": "accuracy_2025-10",
        "value": 0.5556,
        "n": 9
      }
    ]
  },
  "baselines": [
    {
      "name": "random_coin_flip",
      "primary_value": 0.5,
      "description": "50/50 random direction"
    },
    {
      "name": "majority_class",
      "primary_value": 0.529,
      "description": "Always predict majority direction (estimated from historical WTI up-day frequency)"
    }
  ],
  "notes": "OOS replay with causal leakage filter (query_ts) threading through all 8 memory retrieval paths. Price data source: BigQuery/CSV (production). Total LLM cost: $13.13. Directional accuracy is the primary metric; Sharpe is secondary.",
  "data_filter": "OOS window: 2025-05-01 to 2026-05-01, weekdays only. Causal filter: query_ts={day} for each day."
}