{
  "authentication_receipt_sha256": "921699b45e585e4990defb3c81fd95334c52d71f953446c537d21a3f640bcac7",
  "backend": "vllm",
  "context_adherence": {
    "mechanics-00000": {
      "checks": {
        "deranged_supplied_adherence": true,
        "registered_delta": true,
        "systematic_candidate_adherence": true,
        "systematic_execution": true
      },
      "deranged_registered_count": 0,
      "deranged_registered_rate": 0.0,
      "deranged_supplied_count": 21,
      "deranged_supplied_rate": 0.875,
      "pass": true,
      "rows": 24,
      "systematic_registered_count": 21,
      "systematic_registered_minus_deranged_registered": 0.875,
      "systematic_registered_rate": 0.875,
      "systematic_supplied_count": 21,
      "systematic_supplied_rate": 0.875,
      "thresholds": {
        "candidate_adherence_min": 0.6,
        "deranged_adherence_gain_min": 0.35,
        "deranged_own_execution_min": 0.6,
        "execution_rate_min": 0.75
      }
    },
    "mechanics-00001": {
      "checks": {
        "deranged_supplied_adherence": true,
        "registered_delta": true,
        "systematic_candidate_adherence": true,
        "systematic_execution": true
      },
      "deranged_registered_count": 0,
      "deranged_registered_rate": 0.0,
      "deranged_supplied_count": 22,
      "deranged_supplied_rate": 0.9166666666666666,
      "pass": true,
      "rows": 24,
      "systematic_registered_count": 22,
      "systematic_registered_minus_deranged_registered": 0.9166666666666666,
      "systematic_registered_rate": 0.9166666666666666,
      "systematic_supplied_count": 22,
      "systematic_supplied_rate": 0.9166666666666666,
      "thresholds": {
        "candidate_adherence_min": 0.6,
        "deranged_adherence_gain_min": 0.35,
        "deranged_own_execution_min": 0.6,
        "execution_rate_min": 0.75
      }
    },
    "mechanics-00002": {
      "checks": {
        "deranged_supplied_adherence": true,
        "registered_delta": true,
        "systematic_candidate_adherence": true,
        "systematic_execution": true
      },
      "deranged_registered_count": 0,
      "deranged_registered_rate": 0.0,
      "deranged_supplied_count": 22,
      "deranged_supplied_rate": 0.9166666666666666,
      "pass": true,
      "rows": 24,
      "systematic_registered_count": 20,
      "systematic_registered_minus_deranged_registered": 0.8333333333333334,
      "systematic_registered_rate": 0.8333333333333334,
      "systematic_supplied_count": 20,
      "systematic_supplied_rate": 0.8333333333333334,
      "thresholds": {
        "candidate_adherence_min": 0.6,
        "deranged_adherence_gain_min": 0.35,
        "deranged_own_execution_min": 0.6,
        "execution_rate_min": 0.75
      }
    },
    "mechanics-00003": {
      "checks": {
        "deranged_supplied_adherence": true,
        "registered_delta": true,
        "systematic_candidate_adherence": true,
        "systematic_execution": true
      },
      "deranged_registered_count": 0,
      "deranged_registered_rate": 0.0,
      "deranged_supplied_count": 19,
      "deranged_supplied_rate": 0.7916666666666666,
      "pass": true,
      "rows": 24,
      "systematic_registered_count": 21,
      "systematic_registered_minus_deranged_registered": 0.875,
      "systematic_registered_rate": 0.875,
      "systematic_supplied_count": 21,
      "systematic_supplied_rate": 0.875,
      "thresholds": {
        "candidate_adherence_min": 0.6,
        "deranged_adherence_gain_min": 0.35,
        "deranged_own_execution_min": 0.6,
        "execution_rate_min": 0.75
      }
    }
  },
  "decision": "INVALID_INTERFACE_PARSE",
  "gates": {
    "context_adherence_valid": true,
    "correct_hypothesis_ceiling_valid": false,
    "hypothesis_adherence_valid": true,
    "interface_valid": false,
    "primary_interface_valid": false,
    "program_interface_valid": true
  },
  "metrics": {
    "deranged": {
      "answer_limit_contact_rate": 0.10416666666666667,
      "contexts": {
        "mechanics-00000": {
          "answer_limit_contact_count": 2,
          "answer_limit_contact_rate": 0.08333333333333333,
          "parse_count": 22,
          "parse_rate": 0.9166666666666666,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 21,
          "supplied_execution_rate": 0.875
        },
        "mechanics-00001": {
          "answer_limit_contact_count": 2,
          "answer_limit_contact_rate": 0.08333333333333333,
          "parse_count": 22,
          "parse_rate": 0.9166666666666666,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 22,
          "supplied_execution_rate": 0.9166666666666666
        },
        "mechanics-00002": {
          "answer_limit_contact_count": 2,
          "answer_limit_contact_rate": 0.08333333333333333,
          "parse_count": 23,
          "parse_rate": 0.9583333333333334,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 22,
          "supplied_execution_rate": 0.9166666666666666
        },
        "mechanics-00003": {
          "answer_limit_contact_count": 4,
          "answer_limit_contact_rate": 0.16666666666666666,
          "parse_count": 20,
          "parse_rate": 0.8333333333333334,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 19,
          "supplied_execution_rate": 0.7916666666666666
        }
      },
      "logical_model_tokens": 63776,
      "parse_rate": 0.90625,
      "registered_execution_rate": 0.0,
      "rows": 96,
      "sampled_tokens": 16017,
      "supplied_execution_rate": 0.875
    },
    "duplicate": {
      "answer_limit_contact_rate": 0.22916666666666666,
      "contexts": {
        "mechanics-00000": {
          "answer_limit_contact_count": 6,
          "answer_limit_contact_rate": 0.25,
          "parse_count": 14,
          "parse_rate": 0.5833333333333334,
          "registered_execution_count": 1,
          "registered_execution_rate": 0.041666666666666664,
          "rows": 24,
          "supplied_execution_count": 14,
          "supplied_execution_rate": 0.5833333333333334
        },
        "mechanics-00001": {
          "answer_limit_contact_count": 6,
          "answer_limit_contact_rate": 0.25,
          "parse_count": 18,
          "parse_rate": 0.75,
          "registered_execution_count": 1,
          "registered_execution_rate": 0.041666666666666664,
          "rows": 24,
          "supplied_execution_count": 18,
          "supplied_execution_rate": 0.75
        },
        "mechanics-00002": {
          "answer_limit_contact_count": 3,
          "answer_limit_contact_rate": 0.125,
          "parse_count": 21,
          "parse_rate": 0.875,
          "registered_execution_count": 1,
          "registered_execution_rate": 0.041666666666666664,
          "rows": 24,
          "supplied_execution_count": 21,
          "supplied_execution_rate": 0.875
        },
        "mechanics-00003": {
          "answer_limit_contact_count": 7,
          "answer_limit_contact_rate": 0.2916666666666667,
          "parse_count": 17,
          "parse_rate": 0.7083333333333334,
          "registered_execution_count": 1,
          "registered_execution_rate": 0.041666666666666664,
          "rows": 24,
          "supplied_execution_count": 17,
          "supplied_execution_rate": 0.7083333333333334
        }
      },
      "logical_model_tokens": 74147,
      "parse_rate": 0.7291666666666666,
      "registered_execution_rate": 0.041666666666666664,
      "rows": 96,
      "sampled_tokens": 18182,
      "supplied_execution_rate": 0.7291666666666666
    },
    "placebo": {
      "answer_limit_contact_rate": 0.40625,
      "contexts": {
        "mechanics-00000": {
          "answer_limit_contact_count": 12,
          "answer_limit_contact_rate": 0.5,
          "parse_count": 9,
          "parse_rate": 0.375,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 0,
          "supplied_execution_rate": 0.0
        },
        "mechanics-00001": {
          "answer_limit_contact_count": 9,
          "answer_limit_contact_rate": 0.375,
          "parse_count": 14,
          "parse_rate": 0.5833333333333334,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 0,
          "supplied_execution_rate": 0.0
        },
        "mechanics-00002": {
          "answer_limit_contact_count": 8,
          "answer_limit_contact_rate": 0.3333333333333333,
          "parse_count": 15,
          "parse_rate": 0.625,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 0,
          "supplied_execution_rate": 0.0
        },
        "mechanics-00003": {
          "answer_limit_contact_count": 10,
          "answer_limit_contact_rate": 0.4166666666666667,
          "parse_count": 13,
          "parse_rate": 0.5416666666666666,
          "registered_execution_count": 0,
          "registered_execution_rate": 0.0,
          "rows": 24,
          "supplied_execution_count": 0,
          "supplied_execution_rate": 0.0
        }
      },
      "logical_model_tokens": 87943,
      "parse_rate": 0.53125,
      "registered_execution_rate": 0.0,
      "rows": 96,
      "sampled_tokens": 22852,
      "supplied_execution_rate": 0.0
    },
    "systematic": {
      "answer_limit_contact_rate": 0.09375,
      "contexts": {
        "mechanics-00000": {
          "answer_limit_contact_count": 2,
          "answer_limit_contact_rate": 0.08333333333333333,
          "parse_count": 22,
          "parse_rate": 0.9166666666666666,
          "registered_execution_count": 21,
          "registered_execution_rate": 0.875,
          "rows": 24,
          "supplied_execution_count": 21,
          "supplied_execution_rate": 0.875
        },
        "mechanics-00001": {
          "answer_limit_contact_count": 1,
          "answer_limit_contact_rate": 0.041666666666666664,
          "parse_count": 23,
          "parse_rate": 0.9583333333333334,
          "registered_execution_count": 22,
          "registered_execution_rate": 0.9166666666666666,
          "rows": 24,
          "supplied_execution_count": 22,
          "supplied_execution_rate": 0.9166666666666666
        },
        "mechanics-00002": {
          "answer_limit_contact_count": 4,
          "answer_limit_contact_rate": 0.16666666666666666,
          "parse_count": 20,
          "parse_rate": 0.8333333333333334,
          "registered_execution_count": 20,
          "registered_execution_rate": 0.8333333333333334,
          "rows": 24,
          "supplied_execution_count": 20,
          "supplied_execution_rate": 0.8333333333333334
        },
        "mechanics-00003": {
          "answer_limit_contact_count": 2,
          "answer_limit_contact_rate": 0.08333333333333333,
          "parse_count": 22,
          "parse_rate": 0.9166666666666666,
          "registered_execution_count": 21,
          "registered_execution_rate": 0.875,
          "rows": 24,
          "supplied_execution_count": 21,
          "supplied_execution_rate": 0.875
        }
      },
      "logical_model_tokens": 62613,
      "parse_rate": 0.90625,
      "registered_execution_rate": 0.875,
      "rows": 96,
      "sampled_tokens": 15442,
      "supplied_execution_rate": 0.875
    }
  },
  "model_id": "Qwen/Qwen3.5-4B",
  "passed": false,
  "program_ceiling": {
    "answer_limit_contact_rate": 0.0,
    "first_operations": [
      "reverse",
      "running_sum",
      "adjacent_diff",
      "sort_desc",
      "add_k(-2)",
      "mul_k(3)",
      "take_k(3)",
      "rotate_k(2)"
    ],
    "parameterized_rows": 4,
    "parameterized_visible_pass_rate": 0.5,
    "parse_rate": 1.0,
    "rows": 8,
    "scope": "noncausal_reachability_only",
    "visible_pass_rate": 0.375
  },
  "qualification_authorized": false,
  "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
  "schema_version": 1,
  "scored_files": {
    "deranged": "2bcacd5f3834c6b21a7f1ce7d6d49538e01f8232dc2019cd5aa5ce2b1d4ec0d0",
    "duplicate": "be543e1ca2a65a7b707407e25744bdc230e7c0984b5160849eab3c160d906656",
    "placebo": "886ca31ed7319ef2c07ed0646b7b51caef247d511026c4d6769839bced2927f3",
    "program_ceiling": "93a2155712e880588f9943764a4c0f15b2998e285664aeea326030dafd1e373b",
    "systematic": "2e9bfd1c158931ca91c07c3ca71dfa89ca5621e2c183f53e66270011f6d53679"
  },
  "stage": "mechanics",
  "systematic_context_support": 4,
  "systematic_operation_support": 24,
  "systematic_registered_minus_deranged_registered": 0.875
}