{
  "experiment_id": "experiment-20260811T041432Z-3ff8ac0a",
  "comparison_fingerprint": "6d9ef73080714dfbb12d5b9346b07ca62c69e96b8fae43e7e47352acbb8ba058",
  "experiment_complete": true,
  "model": {
    "name": "ollama-qwen2.5-coder:7b",
    "digest": "dae161e27b0e90dd1856c8bb3209201fd6736d8eb66298e75ed87571486f4364",
    "parameter_size": "7.6B",
    "quantization": "Q4_K_M",
    "format": "gguf"
  },
  "training": {
    "episodes": 8,
    "interactions": 32,
    "domain_feedback_records": 32,
    "memory_records": 7,
    "queries": 64,
    "ambiguous_writes": 0,
    "memory_reasons": {
      "recurring_failure": 7
    }
  },
  "evaluation": {
    "runs": 40,
    "interactions": 160,
    "queries": 160,
    "records": 0,
    "operational_failures": 0,
    "ambiguous_writes": 0,
    "arms": {
      "baseline": {
        "n": 10,
        "pass_rate": 0.0,
        "mean_final_score": 4.3,
        "mean_score_auc": 9.55,
        "execution_error_rate": 0.525,
        "adapt_selection_rate": 0.0
      },
      "warm_frozen": {
        "n": 10,
        "pass_rate": 0.0,
        "mean_final_score": 0.0,
        "mean_score_auc": 0.0,
        "execution_error_rate": 0.325,
        "adapt_selection_rate": 1.0
      },
      "domain_only": {
        "n": 10,
        "pass_rate": 0.0,
        "mean_final_score": 0.9,
        "mean_score_auc": 1.35,
        "execution_error_rate": 0.35,
        "adapt_selection_rate": 1.0
      },
      "memory_only": {
        "n": 10,
        "pass_rate": 0.0,
        "mean_final_score": 0.9,
        "mean_score_auc": 1.95,
        "execution_error_rate": 0.55,
        "adapt_selection_rate": 0.0
      }
    },
    "task_final_score_means": {
      "iron_ore_throughput": {
        "baseline": 8.6,
        "warm_frozen": 0.0,
        "domain_only": 1.8,
        "memory_only": 1.8
      },
      "iron_gear_wheel_throughput": {
        "baseline": 0.0,
        "warm_frozen": 0.0,
        "domain_only": 0.0,
        "memory_only": 0.0
      }
    },
    "policy_counts": {
      "baseline": {
        "inspect": 13,
        "debug": 27
      },
      "warm_frozen": {
        "inspect": 40
      },
      "domain_only": {
        "inspect": 40
      },
      "memory_only": {
        "inspect": 11,
        "debug": 29
      }
    }
  },
  "adapt_usage_for_study": {
    "records": 39,
    "queries": 224
  }
}
