{
  "e6": {
    "specified_weights": [
      0.25,
      0.5,
      0.25
    ],
    "counterexample_permutation": [
      1,
      0,
      2
    ],
    "counterexample_resulting_weights": [
      0.5,
      0.25,
      0.25
    ],
    "weight_preserving_permutations": [
      [
        0,
        1,
        2
      ],
      [
        2,
        1,
        0
      ]
    ],
    "valid_constructions": 16,
    "enumerated_constructions": 1296,
    "conclusion": "Published range is not established as a fixed-marginal coupling range."
  },
  "e7b": {
    "calibration_comparison": [
      {
        "judge": "groq70b_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": 1.0,
        "shared_pool_threshold": 1.0,
        "recorded_threshold": 1.0
      },
      {
        "judge": "mistral_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": 1.0,
        "shared_pool_threshold": 1.0,
        "recorded_threshold": 1.0
      },
      {
        "judge": "nim_google_gemma-4-31b-it_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": null,
        "shared_pool_threshold": null,
        "recorded_threshold": null
      },
      {
        "judge": "nim_minimaxai_minimax-m3_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": 0.95,
        "shared_pool_threshold": 0.95,
        "recorded_threshold": 0.95
      },
      {
        "judge": "nim_mistralai_mistral-large-3-675b-instruct-2512_bench",
        "own_benign_n": 75,
        "shared_benign_n": 75,
        "own_pool_threshold": null,
        "shared_pool_threshold": null,
        "recorded_threshold": null
      },
      {
        "judge": "nim_mistralai_mistral-medium-3_5-128b_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": 0.95,
        "shared_pool_threshold": 0.95,
        "recorded_threshold": 0.95
      },
      {
        "judge": "nim_nvidia_nemotron-3-super-120b-a12b_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": 0.95,
        "shared_pool_threshold": 1.0,
        "recorded_threshold": 1.0
      },
      {
        "judge": "nim_qwen_qwen3_5-122b-a10b_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": null,
        "shared_pool_threshold": null,
        "recorded_threshold": null
      },
      {
        "judge": "nim_stepfun-ai_step-3_7-flash_bench",
        "own_benign_n": 97,
        "shared_benign_n": 75,
        "own_pool_threshold": 0.95,
        "shared_pool_threshold": 0.95,
        "recorded_threshold": 0.95
      }
    ],
    "changed_thresholds": 1,
    "conclusion": "Implementation calibrates on all-nine shared pool, contrary to own-pool prereg wording. No frozen outputs changed."
  },
  "input_sha256": {
    "experiments/e7b/PREREG.md": "9d97f550daa0199e5bce72358761595e5da5ee976e3edf450ad4592559076ffc",
    "experiments/e7b/run/measure.py": "d47078392678aaf5c9430eaa8d9233e44cd90c6020d2ae0fb868f7ae2d354116",
    "experiments/e7b/results/e7_result.json": "ab6b7287caa3fa155bc2ff85fe587a1779fab404b6086c80294ba8334b9c1869",
    "experiments/e6/freeze/sources.json": "ad4568cb1df854c2a7b5d892edb37206ae5a075968fe171c73b715f0c90f3236",
    "scripts/verify_e6.py": "0b4dc002d2b18f83d6e51cb1c586450c99257dcb22aebd0fcf8ee8296eed4854"
  }
}