{
  "easy": {
    "a2|review_check|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 0,
        "D2": 0,
        "D3": 0,
        "D4": 0,
        "D5": 0,
        "D6": 0,
        "D7": 0,
        "D8": 0,
        "D9": 0,
        "D10": 0
      },
      "mean_defects": 0.0,
      "decoy_flags_mean": 0.0
    },
    "a2|review_facts|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 0,
        "D2": 0,
        "D3": 0,
        "D4": 0,
        "D5": 0,
        "D6": 0,
        "D7": 0,
        "D8": 0,
        "D9": 0,
        "D10": 0
      },
      "mean_defects": 0.0,
      "decoy_flags_mean": 0.0
    },
    "a2|review_plain|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 0,
        "D2": 0,
        "D3": 0,
        "D4": 0,
        "D5": 0,
        "D6": 5,
        "D7": 0,
        "D8": 0,
        "D9": 0,
        "D10": 0
      },
      "mean_defects": 0.62,
      "decoy_flags_mean": 0.0
    },
    "a|review_check|claude-opus-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 10.0,
      "decoy_flags_mean": 0.25
    },
    "a|review_check|claude-sonnet-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 10.0,
      "decoy_flags_mean": 0.12
    },
    "a|review_check|gpt-6-astra": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 10.0,
      "decoy_flags_mean": 0.0
    },
    "a|review_facts|claude-opus-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 10.0,
      "decoy_flags_mean": 0.12
    },
    "a|review_plain|claude-opus-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 6,
        "D5": 0,
        "D6": 8,
        "D7": 7,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 8.62,
      "decoy_flags_mean": 0.75
    },
    "a|rewrite_plain|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 1,
        "D2": 0,
        "D3": 2,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 7,
        "D8": 0,
        "D9": 0,
        "D10": 8
      },
      "mean_defects": 5.25,
      "decoy_flags_mean": 0.0
    }
  },
  "hard": {
    "a2|review_check|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 0,
        "D2": 0,
        "D3": 0,
        "D4": 0,
        "D5": 0,
        "D6": 0,
        "D7": 0,
        "D8": 0,
        "D9": 0,
        "D10": 0
      },
      "mean_defects": 0.0,
      "decoy_flags_mean": 0.0
    },
    "a2|review_facts|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 0,
        "D2": 0,
        "D3": 0,
        "D4": 0,
        "D5": 0,
        "D6": 0,
        "D7": 0,
        "D8": 0,
        "D9": 0,
        "D10": 8
      },
      "mean_defects": 1.0,
      "decoy_flags_mean": 0.0
    },
    "a2|review_plain|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 0,
        "D2": 0,
        "D3": 0,
        "D4": 0,
        "D5": 0,
        "D6": 0,
        "D7": 0,
        "D8": 0,
        "D9": 0,
        "D10": 4
      },
      "mean_defects": 0.5,
      "decoy_flags_mean": 0.0
    },
    "a|review_check|claude-haiku-4-5-20251001": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 3,
        "D3": 8,
        "D4": 6,
        "D5": 6,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 3,
        "D10": 8
      },
      "mean_defects": 8.25,
      "decoy_flags_mean": 0.25
    },
    "a|review_check|claude-opus-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 10.0,
      "decoy_flags_mean": 1.0
    },
    "a|review_check|claude-sonnet-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 10.0,
      "decoy_flags_mean": 1.62
    },
    "a|review_check|gpt-6-astra": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 10.0,
      "decoy_flags_mean": 0.88
    },
    "a|review_facts|claude-opus-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 8,
        "D9": 8,
        "D10": 1
      },
      "mean_defects": 9.12,
      "decoy_flags_mean": 0.0
    },
    "a|review_plain|claude-opus-5-5": {
      "mode": "review",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 8,
        "D3": 1,
        "D4": 8,
        "D5": 8,
        "D6": 2,
        "D7": 2,
        "D8": 8,
        "D9": 0,
        "D10": 8
      },
      "mean_defects": 6.62,
      "decoy_flags_mean": 3.12
    },
    "a|rewrite_facts|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 0,
        "D2": 0,
        "D3": 0,
        "D4": 0,
        "D5": 0,
        "D6": 0,
        "D7": 0,
        "D8": 0,
        "D9": 0,
        "D10": 8
      },
      "mean_defects": 1.0,
      "decoy_flags_mean": 0.0
    },
    "a|rewrite_plain|claude-opus-5-5": {
      "mode": "rewrite",
      "n": 8,
      "per_defect": {
        "D1": 8,
        "D2": 4,
        "D3": 8,
        "D4": 8,
        "D5": 8,
        "D6": 8,
        "D7": 8,
        "D8": 7,
        "D9": 8,
        "D10": 8
      },
      "mean_defects": 9.38,
      "decoy_flags_mean": 0.0
    }
  },
  "fabrication": {
    "self_plain": {
      "n": 8,
      "mean_claims": 11.1,
      "counts": [7, 9, 10, 11, 12, 12, 13, 15],
      "chars_mean": 925,
      "kakunin_mean": 0
    },
    "twostep_same": {
      "n": 8,
      "mean_claims": 5.4,
      "counts": [3, 3, 3, 5, 6, 7, 7, 9],
      "chars_mean": 681,
      "kakunin_mean": 12
    },
    "separate": {
      "n": 8,
      "mean_claims": 7.9,
      "counts": [4, 5, 6, 6, 7, 9, 9, 17],
      "chars_mean": 774,
      "kakunin_mean": 12.5
    },
    "draft": {
      "n": 8,
      "mean_claims": 13.1,
      "counts": [12, 12, 13, 13, 13, 14, 14, 14],
      "chars_mean": 902,
      "kakunin_mean": 0
    },
    "separate_gpt": {
      "n": 8,
      "mean_claims": 0.2,
      "counts": [0, 0, 0, 0, 0, 0, 1, 1],
      "chars_mean": 1072,
      "kakunin_mean": 6.6
    },
    "self_crit": {
      "n": 8,
      "mean_claims": 9.4,
      "counts": [5, 7, 7, 9, 10, 12, 12, 13],
      "chars_mean": 842,
      "kakunin_mean": 7.5
    }
  }
}
